Merge tag 'md-3.5' of git://neil.brown.name/md

Pull md updates from NeilBrown:
"It's been a busy cycle for md - lots of fun stuff here.. if you like
this kind of thing :-)

Main features:
- RAID10 arrays can be reshaped - adding and removing devices and
changing chunks (not 'far' array though)
- allow RAID5 arrays to be reshaped with a backup file (not tested
yet, but the priciple works fine for RAID10).
- arrays can be reshaped while a bitmap is present - you no longer
need to remove it first
- SSSE3 support for RAID6 syndrome calculations

and of course a number of minor fixes etc."

* tag 'md-3.5' of git://neil.brown.name/md: (56 commits)
md/bitmap: record the space available for the bitmap in the superblock.
md/raid10: Remove extras after reshape to smaller number of devices.
md/raid5: improve removal of extra devices after reshape.
md: check the return of mddev_find()
MD RAID1: Further conditionalize 'fullsync'
DM RAID: Use md_error() in place of simply setting Faulty bit
DM RAID: Record and handle missing devices
DM RAID: Set recovery flags on resume
md/raid5: Allow reshape while a bitmap is present.
md/raid10: resize bitmap when required during reshape.
md: allow array to be resized while bitmap is present.
md/bitmap: make sure reshape request are reflected in superblock.
md/bitmap: add bitmap_resize function to allow bitmap resizing.
md/bitmap: use DIV_ROUND_UP instead of open-code
md/bitmap: create a 'struct bitmap_counts' substructure of 'struct bitmap'
md/bitmap: make bitmap bitops atomic.
md/bitmap: make _page_attr bitops atomic.
md/bitmap: merge bitmap_file_unmap and bitmap_file_put.
md/bitmap: remove async freeing of bitmap file.
md/bitmap: convert some spin_lock_irqsave to spin_lock_irq
...

arch/x86/Makefile

···
 
 # does binutils support specific instructions?
 asinstr := $(call as-instr,fxsaveq (%rax),-DCONFIG_AS_FXSAVEQ=1)
+avx_instr := $(call as-instr,vxorps %ymm0$(comma)%ymm1$(comma)%ymm2,-DCONFIG_AS_AVX=1)
 
-KBUILD_AFLAGS += $(cfi) $(cfi-sigframe) $(cfi-sections) $(asinstr)
-KBUILD_CFLAGS += $(cfi) $(cfi-sigframe) $(cfi-sections) $(asinstr)
+KBUILD_AFLAGS += $(cfi) $(cfi-sigframe) $(cfi-sections) $(asinstr) $(avx_instr)
+KBUILD_CFLAGS += $(cfi) $(cfi-sigframe) $(cfi-sections) $(asinstr) $(avx_instr)
 
 LDFLAGS := -m elf_$(UTS_MACHINE)
 

arch/x86/include/asm/xor_32.h

···
 	.do_5 = xor_sse_5,
 };
 
+/* Also try the AVX routines */
+#include "xor_avx.h"
+
 /* Also try the generic routines.  */
 #include <asm-generic/xor.h>
 
···
 	xor_speed(&xor_block_8regs_p);			\
 	xor_speed(&xor_block_32regs);			\
 	xor_speed(&xor_block_32regs_p);			\
+	AVX_XOR_SPEED;					\
 	if (cpu_has_xmm)				\
 		xor_speed(&xor_block_pIII_sse);		\
 	if (cpu_has_mmx) {				\
···
    We may also be able to load into the L1 only depending on how the cpu
    deals with a load to a line that is being prefetched.  */
 #define XOR_SELECT_TEMPLATE(FASTEST)			\
-	(cpu_has_xmm ? &xor_block_pIII_sse : FASTEST)
+	AVX_SELECT(cpu_has_xmm ? &xor_block_pIII_sse : FASTEST)
 
 #endif /* _ASM_X86_XOR_32_H */

arch/x86/include/asm/xor_64.h

···
 	.do_5 = xor_sse_5,
 };
 
+
+/* Also try the AVX routines */
+#include "xor_avx.h"
+
 #undef XOR_TRY_TEMPLATES
 #define XOR_TRY_TEMPLATES			\
 do {						\
+	AVX_XOR_SPEED;				\
 	xor_speed(&xor_block_sse);		\
 } while (0)
 
 /* We force the use of the SSE xor block because it can write around L2.
    We may also be able to load into the L1 only depending on how the cpu
    deals with a load to a line that is being prefetched.  */
-#define XOR_SELECT_TEMPLATE(FASTEST) (&xor_block_sse)
+#define XOR_SELECT_TEMPLATE(FASTEST) \
+	AVX_SELECT(&xor_block_sse)
 
 #endif /* _ASM_X86_XOR_64_H */

arch/x86/include/asm/xor_avx.h

···
+#ifndef _ASM_X86_XOR_AVX_H
+#define _ASM_X86_XOR_AVX_H
+
+/*
+ * Optimized RAID-5 checksumming functions for AVX
+ *
+ * Copyright (C) 2012 Intel Corporation
+ * Author: Jim Kukunas <[e-mail]>
+ *
+ * Based on Ingo Molnar and Zach Brown's respective MMX and SSE routines
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License
+ * as published by the Free Software Foundation; version 2
+ * of the License.
+ */
+
+#ifdef CONFIG_AS_AVX
+
+#include <linux/compiler.h>
+#include <asm/i387.h>
+
+#define ALIGN32 __aligned(32)
+
+#define YMM_SAVED_REGS 4
+
+#define YMMS_SAVE \
+do { \
+	preempt_disable(); \
+	cr0 = read_cr0(); \
+	clts(); \
+	asm volatile("vmovaps %%ymm0, %0" : "=m" (ymm_save[0]) : : "memory"); \
+	asm volatile("vmovaps %%ymm1, %0" : "=m" (ymm_save[32]) : : "memory"); \
+	asm volatile("vmovaps %%ymm2, %0" : "=m" (ymm_save[64]) : : "memory"); \
+	asm volatile("vmovaps %%ymm3, %0" : "=m" (ymm_save[96]) : : "memory"); \
+} while (0);
+
+#define YMMS_RESTORE \
+do { \
+	asm volatile("sfence" : : : "memory"); \
+	asm volatile("vmovaps %0, %%ymm3" : : "m" (ymm_save[96])); \
+	asm volatile("vmovaps %0, %%ymm2" : : "m" (ymm_save[64])); \
+	asm volatile("vmovaps %0, %%ymm1" : : "m" (ymm_save[32])); \
+	asm volatile("vmovaps %0, %%ymm0" : : "m" (ymm_save[0])); \
+	write_cr0(cr0); \
+	preempt_enable(); \
+} while (0);
+
+#define BLOCK4(i) \
+		BLOCK(32 * i, 0) \
+		BLOCK(32 * (i + 1), 1) \
+		BLOCK(32 * (i + 2), 2) \
+		BLOCK(32 * (i + 3), 3)
+
+#define BLOCK16() \
+		BLOCK4(0) \
+		BLOCK4(4) \
+		BLOCK4(8) \
+		BLOCK4(12)
+
+static void xor_avx_2(unsigned long bytes, unsigned long *p0, unsigned long *p1)
+{
+	unsigned long cr0, lines = bytes >> 9;
+	char ymm_save[32 * YMM_SAVED_REGS] ALIGN32;
+
+	YMMS_SAVE
+
+	while (lines--) {
+#undef BLOCK
+#define BLOCK(i, reg) \
+do { \
+	asm volatile("vmovdqa %0, %%ymm" #reg : : "m" (p1[i / sizeof(*p1)])); \
+	asm volatile("vxorps %0, %%ymm" #reg ", %%ymm"  #reg : : \
+		"m" (p0[i / sizeof(*p0)])); \
+	asm volatile("vmovdqa %%ymm" #reg ", %0" : \
+		"=m" (p0[i / sizeof(*p0)])); \
+} while (0);
+
+		BLOCK16()
+
+		p0 = (unsigned long *)((uintptr_t)p0 + 512);
+		p1 = (unsigned long *)((uintptr_t)p1 + 512);
+	}
+
+	YMMS_RESTORE
+}
+
+static void xor_avx_3(unsigned long bytes, unsigned long *p0, unsigned long *p1,
+	unsigned long *p2)
+{
+	unsigned long cr0, lines = bytes >> 9;
+	char ymm_save[32 * YMM_SAVED_REGS] ALIGN32;
+
+	YMMS_SAVE
+
+	while (lines--) {
+#undef BLOCK
+#define BLOCK(i, reg) \
+do { \
+	asm volatile("vmovdqa %0, %%ymm" #reg : : "m" (p2[i / sizeof(*p2)])); \
+	asm volatile("vxorps %0, %%ymm" #reg ", %%ymm" #reg : : \
+		"m" (p1[i / sizeof(*p1)])); \
+	asm volatile("vxorps %0, %%ymm" #reg ", %%ymm" #reg : : \
+		"m" (p0[i / sizeof(*p0)])); \
+	asm volatile("vmovdqa %%ymm" #reg ", %0" : \
+		"=m" (p0[i / sizeof(*p0)])); \
+} while (0);
+
+		BLOCK16()
+
+		p0 = (unsigned long *)((uintptr_t)p0 + 512);
+		p1 = (unsigned long *)((uintptr_t)p1 + 512);
+		p2 = (unsigned long *)((uintptr_t)p2 + 512);
+	}
+
+	YMMS_RESTORE
+}
+
+static void xor_avx_4(unsigned long bytes, unsigned long *p0, unsigned long *p1,
+	unsigned long *p2, unsigned long *p3)
+{
+	unsigned long cr0, lines = bytes >> 9;
+	char ymm_save[32 * YMM_SAVED_REGS] ALIGN32;
+
+	YMMS_SAVE
+
+	while (lines--) {
+#undef BLOCK
+#define BLOCK(i, reg) \
+do { \
+	asm volatile("vmovdqa %0, %%ymm" #reg : : "m" (p3[i / sizeof(*p3)])); \
+	asm volatile("vxorps %0, %%ymm" #reg ", %%ymm" #reg : : \
+		"m" (p2[i / sizeof(*p2)])); \
+	asm volatile("vxorps %0, %%ymm" #reg ", %%ymm" #reg : : \
+		"m" (p1[i / sizeof(*p1)])); \
+	asm volatile("vxorps %0, %%ymm" #reg ", %%ymm" #reg : : \
+		"m" (p0[i / sizeof(*p0)])); \
+	asm volatile("vmovdqa %%ymm" #reg ", %0" : \
+		"=m" (p0[i / sizeof(*p0)])); \
+} while (0);
+
+		BLOCK16();
+
+		p0 = (unsigned long *)((uintptr_t)p0 + 512);
+		p1 = (unsigned long *)((uintptr_t)p1 + 512);
+		p2 = (unsigned long *)((uintptr_t)p2 + 512);
+		p3 = (unsigned long *)((uintptr_t)p3 + 512);
+	}
+
+	YMMS_RESTORE
+}
+
+static void xor_avx_5(unsigned long bytes, unsigned long *p0, unsigned long *p1,
+	unsigned long *p2, unsigned long *p3, unsigned long *p4)
+{
+	unsigned long cr0, lines = bytes >> 9;
+	char ymm_save[32 * YMM_SAVED_REGS] ALIGN32;
+
+	YMMS_SAVE
+
+	while (lines--) {
+#undef BLOCK
+#define BLOCK(i, reg) \
+do { \
+	asm volatile("vmovdqa %0, %%ymm" #reg : : "m" (p4[i / sizeof(*p4)])); \
+	asm volatile("vxorps %0, %%ymm" #reg ", %%ymm" #reg : : \
+		"m" (p3[i / sizeof(*p3)])); \
+	asm volatile("vxorps %0, %%ymm" #reg ", %%ymm" #reg : : \
+		"m" (p2[i / sizeof(*p2)])); \
+	asm volatile("vxorps %0, %%ymm" #reg ", %%ymm" #reg : : \
+		"m" (p1[i / sizeof(*p1)])); \
+	asm volatile("vxorps %0, %%ymm" #reg ", %%ymm" #reg : : \
+		"m" (p0[i / sizeof(*p0)])); \
+	asm volatile("vmovdqa %%ymm" #reg ", %0" : \
+		"=m" (p0[i / sizeof(*p0)])); \
+} while (0);
+
+		BLOCK16()
+
+		p0 = (unsigned long *)((uintptr_t)p0 + 512);
+		p1 = (unsigned long *)((uintptr_t)p1 + 512);
+		p2 = (unsigned long *)((uintptr_t)p2 + 512);
+		p3 = (unsigned long *)((uintptr_t)p3 + 512);
+		p4 = (unsigned long *)((uintptr_t)p4 + 512);
+	}
+
+	YMMS_RESTORE
+}
+
+static struct xor_block_template xor_block_avx = {
+	.name = "avx",
+	.do_2 = xor_avx_2,
+	.do_3 = xor_avx_3,
+	.do_4 = xor_avx_4,
+	.do_5 = xor_avx_5,
+};
+
+#define AVX_XOR_SPEED \
+do { \
+	if (cpu_has_avx) \
+		xor_speed(&xor_block_avx); \
+} while (0)
+
+#define AVX_SELECT(FASTEST) \
+	(cpu_has_avx ? &xor_block_avx : FASTEST)
+
+#else
+
+#define AVX_XOR_SPEED {}
+
+#define AVX_SELECT(FASTEST) (FASTEST)
+
+#endif
+#endif

crypto/xor.c

···
 #include <linux/gfp.h>
 #include <linux/raid/xor.h>
 #include <linux/jiffies.h>
+#include <linux/preempt.h>
 #include <asm/xor.h>
 
 /* The xor routines to use.  */
···
 do_xor_speed(struct xor_block_template *tmpl, void *b1, void *b2)
 {
 	int speed;
-	unsigned long now;
+	unsigned long now, j;
 	int i, count, max;
 
 	tmpl->next = template_list;
 	template_list = tmpl;
+
+	preempt_disable();
 
 	/*
 	 * Count the number of XORs done during a whole jiffy, and use
···
 	 */
 	max = 0;
 	for (i = 0; i < 5; i++) {
-		now = jiffies;
+		j = jiffies;
 		count = 0;
-		while (jiffies == now) {
+		while ((now = jiffies) == j)
+			cpu_relax();
+		while (time_before(jiffies, now + 1)) {
 			mb(); /* prevent loop optimzation */
 			tmpl->do_2(BENCH_SIZE, b1, b2);
 			mb();
···
 		if (count > max)
 			max = count;
 	}
+
+	preempt_enable();
 
 	speed = max * (HZ * BENCH_SIZE / 1024);
 	tmpl->speed = speed;

drivers/md/bitmap.c

···
  * if we find our page, we increment the page's refcount so that it stays
  * allocated while we're using it
  */
-static int bitmap_checkpage(struct bitmap *bitmap,
+static int bitmap_checkpage(struct bitmap_counts *bitmap,
 			    unsigned long page, int create)
 __releases(bitmap->lock)
 __acquires(bitmap->lock)
···
 	spin_lock_irq(&bitmap->lock);
 
 	if (mappage == NULL) {
-		pr_debug("%s: bitmap map page allocation failed, hijacking\n",
-			 bmname(bitmap));
+		pr_debug("md/bitmap: map page allocation failed, hijacking\n");
 		/* failed - set the hijacked flag so that we can use the
 		 * pointer as a counter */
 		if (!bitmap->bp[page].map)
···
 /* if page is completely empty, put it back on the free list, or dealloc it */
 /* if page was hijacked, unmark the flag so it might get alloced next time */
 /* Note: lock should be held when calling this */
-static void bitmap_checkfree(struct bitmap *bitmap, unsigned long page)
+static void bitmap_checkfree(struct bitmap_counts *bitmap, unsigned long page)
 {
 	char *ptr;
 
···
  */
 
 /* IO operations when bitmap is stored near all superblocks */
-static struct page *read_sb_page(struct mddev *mddev, loff_t offset,
-				 struct page *page,
-				 unsigned long index, int size)
+static int read_sb_page(struct mddev *mddev, loff_t offset,
+			struct page *page,
+			unsigned long index, int size)
 {
 	/* choose a good rdev and read the page from there */
 
 	struct md_rdev *rdev;
 	sector_t target;
-	int did_alloc = 0;
-
-	if (!page) {
-		page = alloc_page(GFP_KERNEL);
-		if (!page)
-			return ERR_PTR(-ENOMEM);
-		did_alloc = 1;
-	}
 
 	rdev_for_each(rdev, mddev) {
 		if (! test_bit(In_sync, &rdev->flags)
···
 				 roundup(size, bdev_logical_block_size(rdev->bdev)),
 				 page, READ, true)) {
 			page->index = index;
-			attach_page_buffers(page, NULL); /* so that free_buffer will
-							  * quietly no-op */
-			return page;
+			return 0;
 		}
 	}
-	if (did_alloc)
-		put_page(page);
-	return ERR_PTR(-EIO);
-
+	return -EIO;
 }
 
 static struct md_rdev *next_active_rdev(struct md_rdev *rdev, struct mddev *mddev)
···
 	struct md_rdev *rdev = NULL;
 	struct block_device *bdev;
 	struct mddev *mddev = bitmap->mddev;
+	struct bitmap_storage *store = &bitmap->storage;
 
 	while ((rdev = next_active_rdev(rdev, mddev)) != NULL) {
 		int size = PAGE_SIZE;
···
 
 		bdev = (rdev->meta_bdev) ? rdev->meta_bdev : rdev->bdev;
 
-		if (page->index == bitmap->file_pages-1)
-			size = roundup(bitmap->last_page_size,
+		if (page->index == store->file_pages-1) {
+			int last_page_size = store->bytes & (PAGE_SIZE-1);
+			if (last_page_size == 0)
+				last_page_size = PAGE_SIZE;
+			size = roundup(last_page_size,
 				       bdev_logical_block_size(bdev));
+		}
 		/* Just make sure we aren't corrupting data or
 		 * metadata
 		 */
···
 {
 	struct buffer_head *bh;
 
-	if (bitmap->file == NULL) {
+	if (bitmap->storage.file == NULL) {
 		switch (write_sb_page(bitmap, page, wait)) {
 		case -EINVAL:
-			bitmap->flags |= BITMAP_WRITE_ERROR;
+			set_bit(BITMAP_WRITE_ERROR, &bitmap->flags);
 		}
 	} else {
 
···
 			wait_event(bitmap->write_wait,
 				   atomic_read(&bitmap->pending_writes)==0);
 	}
-	if (bitmap->flags & BITMAP_WRITE_ERROR)
+	if (test_bit(BITMAP_WRITE_ERROR, &bitmap->flags))
 		bitmap_file_kick(bitmap);
 }
 
 static void end_bitmap_write(struct buffer_head *bh, int uptodate)
 {
 	struct bitmap *bitmap = bh->b_private;
-	unsigned long flags;
 
-	if (!uptodate) {
-		spin_lock_irqsave(&bitmap->lock, flags);
-		bitmap->flags |= BITMAP_WRITE_ERROR;
-		spin_unlock_irqrestore(&bitmap->lock, flags);
-	}
+	if (!uptodate)
+		set_bit(BITMAP_WRITE_ERROR, &bitmap->flags);
 	if (atomic_dec_and_test(&bitmap->pending_writes))
 		wake_up(&bitmap->write_wait);
 }
···
 }
 static void free_buffers(struct page *page)
 {
-	struct buffer_head *bh = page_buffers(page);
+	struct buffer_head *bh;
 
+	if (!PagePrivate(page))
+		return;
+
+	bh = page_buffers(page);
 	while (bh) {
 		struct buffer_head *next = bh->b_this_page;
 		free_buffer_head(bh);
···
  * This usage is similar to how swap files are handled, and allows us
  * to write to a file with no concerns of memory allocation failing.
  */
-static struct page *read_page(struct file *file, unsigned long index,
-			      struct bitmap *bitmap,
-			      unsigned long count)
+static int read_page(struct file *file, unsigned long index,
+		     struct bitmap *bitmap,
+		     unsigned long count,
+		     struct page *page)
 {
-	struct page *page = NULL;
+	int ret = 0;
 	struct inode *inode = file->f_path.dentry->d_inode;
 	struct buffer_head *bh;
 	sector_t block;
···
 	pr_debug("read bitmap file (%dB @ %llu)\n", (int)PAGE_SIZE,
 		 (unsigned long long)index << PAGE_SHIFT);
 
-	page = alloc_page(GFP_KERNEL);
-	if (!page)
-		page = ERR_PTR(-ENOMEM);
-	if (IS_ERR(page))
-		goto out;
-
 	bh = alloc_page_buffers(page, 1<<inode->i_blkbits, 0);
 	if (!bh) {
-		put_page(page);
-		page = ERR_PTR(-ENOMEM);
+		ret = -ENOMEM;
 		goto out;
 	}
 	attach_page_buffers(page, bh);
···
 			bh->b_blocknr = bmap(inode, block);
 			if (bh->b_blocknr == 0) {
 				/* Cannot use this file! */
-				free_buffers(page);
-				page = ERR_PTR(-EINVAL);
+				ret = -EINVAL;
 				goto out;
 			}
 			bh->b_bdev = inode->i_sb->s_bdev;
···
 
 	wait_event(bitmap->write_wait,
 		   atomic_read(&bitmap->pending_writes)==0);
-	if (bitmap->flags & BITMAP_WRITE_ERROR) {
-		free_buffers(page);
-		page = ERR_PTR(-EIO);
-	}
+	if (test_bit(BITMAP_WRITE_ERROR, &bitmap->flags))
+		ret = -EIO;
 out:
-	if (IS_ERR(page))
-		printk(KERN_ALERT "md: bitmap read error: (%dB @ %llu): %ld\n",
+	if (ret)
+		printk(KERN_ALERT "md: bitmap read error: (%dB @ %llu): %d\n",
 			(int)PAGE_SIZE,
 			(unsigned long long)index << PAGE_SHIFT,
-			PTR_ERR(page));
-	return page;
+			ret);
+	return ret;
 }
 
 /*
···
 		return;
 	if (bitmap->mddev->bitmap_info.external)
 		return;
-	if (!bitmap->sb_page) /* no superblock */
+	if (!bitmap->storage.sb_page) /* no superblock */
 		return;
-	sb = kmap_atomic(bitmap->sb_page);
+	sb = kmap_atomic(bitmap->storage.sb_page);
 	sb->events = cpu_to_le64(bitmap->mddev->events);
 	if (bitmap->mddev->events < bitmap->events_cleared)
 		/* rocking back to read-only */
···
 	/* Just in case these have been changed via sysfs: */
 	sb->daemon_sleep = cpu_to_le32(bitmap->mddev->bitmap_info.daemon_sleep/HZ);
 	sb->write_behind = cpu_to_le32(bitmap->mddev->bitmap_info.max_write_behind);
+	/* This might have been changed by a reshape */
+	sb->sync_size = cpu_to_le64(bitmap->mddev->resync_max_sectors);
+	sb->chunksize = cpu_to_le32(bitmap->mddev->bitmap_info.chunksize);
+	sb->sectors_reserved = cpu_to_le32(bitmap->mddev->
+					   bitmap_info.space);
 	kunmap_atomic(sb);
-	write_page(bitmap, bitmap->sb_page, 1);
+	write_page(bitmap, bitmap->storage.sb_page, 1);
 }
 
 /* print out the bitmap file superblock */
···
 {
 	bitmap_super_t *sb;
 
-	if (!bitmap || !bitmap->sb_page)
+	if (!bitmap || !bitmap->storage.sb_page)
 		return;
-	sb = kmap_atomic(bitmap->sb_page);
+	sb = kmap_atomic(bitmap->storage.sb_page);
 	printk(KERN_DEBUG "%s: bitmap file superblock:\n", bmname(bitmap));
 	printk(KERN_DEBUG "         magic: %08x\n", le32_to_cpu(sb->magic));
 	printk(KERN_DEBUG "       version: %d\n", le32_to_cpu(sb->version));
···
 	unsigned long chunksize, daemon_sleep, write_behind;
 	int err = -EINVAL;
 
-	bitmap->sb_page = alloc_page(GFP_KERNEL);
-	if (IS_ERR(bitmap->sb_page)) {
-		err = PTR_ERR(bitmap->sb_page);
-		bitmap->sb_page = NULL;
+	bitmap->storage.sb_page = alloc_page(GFP_KERNEL);
+	if (IS_ERR(bitmap->storage.sb_page)) {
+		err = PTR_ERR(bitmap->storage.sb_page);
+		bitmap->storage.sb_page = NULL;
 		return err;
 	}
-	bitmap->sb_page->index = 0;
+	bitmap->storage.sb_page->index = 0;
 
-	sb = kmap_atomic(bitmap->sb_page);
+	sb = kmap_atomic(bitmap->storage.sb_page);
 
 	sb->magic = cpu_to_le32(BITMAP_MAGIC);
 	sb->version = cpu_to_le32(BITMAP_MAJOR_HI);
···
 
 	memcpy(sb->uuid, bitmap->mddev->uuid, 16);
 
-	bitmap->flags |= BITMAP_STALE;
-	sb->state |= cpu_to_le32(BITMAP_STALE);
+	set_bit(BITMAP_STALE, &bitmap->flags);
+	sb->state = cpu_to_le32(bitmap->flags);
 	bitmap->events_cleared = bitmap->mddev->events;
 	sb->events_cleared = cpu_to_le64(bitmap->mddev->events);
 
···
 	bitmap_super_t *sb;
 	unsigned long chunksize, daemon_sleep, write_behind;
 	unsigned long long events;
+	unsigned long sectors_reserved = 0;
 	int err = -EINVAL;
+	struct page *sb_page;
 
+	if (!bitmap->storage.file && !bitmap->mddev->bitmap_info.offset) {
+		chunksize = 128 * 1024 * 1024;
+		daemon_sleep = 5 * HZ;
+		write_behind = 0;
+		set_bit(BITMAP_STALE, &bitmap->flags);
+		err = 0;
+		goto out_no_sb;
+	}
 	/* page 0 is the superblock, read it... */
-	if (bitmap->file) {
-		loff_t isize = i_size_read(bitmap->file->f_mapping->host);
+	sb_page = alloc_page(GFP_KERNEL);
+	if (!sb_page)
+		return -ENOMEM;
+	bitmap->storage.sb_page = sb_page;
+
+	if (bitmap->storage.file) {
+		loff_t isize = i_size_read(bitmap->storage.file->f_mapping->host);
 		int bytes = isize > PAGE_SIZE ? PAGE_SIZE : isize;
 
-		bitmap->sb_page = read_page(bitmap->file, 0, bitmap, bytes);
+		err = read_page(bitmap->storage.file, 0,
+				bitmap, bytes, sb_page);
 	} else {
-		bitmap->sb_page = read_sb_page(bitmap->mddev,
-					       bitmap->mddev->bitmap_info.offset,
-					       NULL,
-					       0, sizeof(bitmap_super_t));
+		err = read_sb_page(bitmap->mddev,
+				   bitmap->mddev->bitmap_info.offset,
+				   sb_page,
+				   0, sizeof(bitmap_super_t));
 	}
-	if (IS_ERR(bitmap->sb_page)) {
-		err = PTR_ERR(bitmap->sb_page);
-		bitmap->sb_page = NULL;
+	if (err)
 		return err;
-	}
 
-	sb = kmap_atomic(bitmap->sb_page);
+	sb = kmap_atomic(sb_page);
 
 	chunksize = le32_to_cpu(sb->chunksize);
 	daemon_sleep = le32_to_cpu(sb->daemon_sleep) * HZ;
 	write_behind = le32_to_cpu(sb->write_behind);
+	sectors_reserved = le32_to_cpu(sb->sectors_reserved);
 
 	/* verify that the bitmap-specific fields are valid */
 	if (sb->magic != cpu_to_le32(BITMAP_MAGIC))
···
 			       "-- forcing full recovery\n",
 			       bmname(bitmap), events,
 			       (unsigned long long) bitmap->mddev->events);
-			sb->state |= cpu_to_le32(BITMAP_STALE);
+			set_bit(BITMAP_STALE, &bitmap->flags);
 		}
 	}
 
 	/* assign fields using values from superblock */
-	bitmap->mddev->bitmap_info.chunksize = chunksize;
-	bitmap->mddev->bitmap_info.daemon_sleep = daemon_sleep;
-	bitmap->mddev->bitmap_info.max_write_behind = write_behind;
 	bitmap->flags |= le32_to_cpu(sb->state);
 	if (le32_to_cpu(sb->version) == BITMAP_MAJOR_HOSTENDIAN)
-		bitmap->flags |= BITMAP_HOSTENDIAN;
+		set_bit(BITMAP_HOSTENDIAN, &bitmap->flags);
 	bitmap->events_cleared = le64_to_cpu(sb->events_cleared);
-	if (bitmap->flags & BITMAP_STALE)
-		bitmap->events_cleared = bitmap->mddev->events;
 	err = 0;
 out:
 	kunmap_atomic(sb);
+out_no_sb:
+	if (test_bit(BITMAP_STALE, &bitmap->flags))
+		bitmap->events_cleared = bitmap->mddev->events;
+	bitmap->mddev->bitmap_info.chunksize = chunksize;
+	bitmap->mddev->bitmap_info.daemon_sleep = daemon_sleep;
+	bitmap->mddev->bitmap_info.max_write_behind = write_behind;
+	if (bitmap->mddev->bitmap_info.space == 0 ||
+	    bitmap->mddev->bitmap_info.space > sectors_reserved)
+		bitmap->mddev->bitmap_info.space = sectors_reserved;
 	if (err)
 		bitmap_print_sb(bitmap);
 	return err;
-}
-
-enum bitmap_mask_op {
-	MASK_SET,
-	MASK_UNSET
-};
-
-/* record the state of the bitmap in the superblock.  Return the old value */
-static int bitmap_mask_state(struct bitmap *bitmap, enum bitmap_state bits,
-			     enum bitmap_mask_op op)
-{
-	bitmap_super_t *sb;
-	int old;
-
-	if (!bitmap->sb_page) /* can't set the state */
-		return 0;
-	sb = kmap_atomic(bitmap->sb_page);
-	old = le32_to_cpu(sb->state) & bits;
-	switch (op) {
-	case MASK_SET:
-		sb->state |= cpu_to_le32(bits);
-		bitmap->flags |= bits;
-		break;
-	case MASK_UNSET:
-		sb->state &= cpu_to_le32(~bits);
-		bitmap->flags &= ~bits;
-		break;
-	default:
-		BUG();
-	}
-	kunmap_atomic(sb);
-	return old;
 }
 
 /*
···
  * file a page at a time. There's a superblock at the start of the file.
  */
 /* calculate the index of the page that contains this bit */
-static inline unsigned long file_page_index(struct bitmap *bitmap, unsigned long chunk)
+static inline unsigned long file_page_index(struct bitmap_storage *store,
+					    unsigned long chunk)
 {
-	if (!bitmap->mddev->bitmap_info.external)
+	if (store->sb_page)
 		chunk += sizeof(bitmap_super_t) << 3;
 	return chunk >> PAGE_BIT_SHIFT;
 }
 
 /* calculate the (bit) offset of this bit within a page */
-static inline unsigned long file_page_offset(struct bitmap *bitmap, unsigned long chunk)
+static inline unsigned long file_page_offset(struct bitmap_storage *store,
+					     unsigned long chunk)
 {
-	if (!bitmap->mddev->bitmap_info.external)
+	if (store->sb_page)
 		chunk += sizeof(bitmap_super_t) << 3;
 	return chunk & (PAGE_BITS - 1);
 }
···
  * 1 page (e.g., x86) or less than 1 page -- so the bitmap might start on page
  * 0 or page 1
  */
-static inline struct page *filemap_get_page(struct bitmap *bitmap,
+static inline struct page *filemap_get_page(struct bitmap_storage *store,
 					    unsigned long chunk)
 {
-	if (file_page_index(bitmap, chunk) >= bitmap->file_pages)
+	if (file_page_index(store, chunk) >= store->file_pages)
 		return NULL;
-	return bitmap->filemap[file_page_index(bitmap, chunk)
-			       - file_page_index(bitmap, 0)];
+	return store->filemap[file_page_index(store, chunk)
+			      - file_page_index(store, 0)];
 }
 
-static void bitmap_file_unmap(struct bitmap *bitmap)
+static int bitmap_storage_alloc(struct bitmap_storage *store,
+				unsigned long chunks, int with_super)
+{
+	int pnum;
+	unsigned long num_pages;
+	unsigned long bytes;
+
+	bytes = DIV_ROUND_UP(chunks, 8);
+	if (with_super)
+		bytes += sizeof(bitmap_super_t);
+
+	num_pages = DIV_ROUND_UP(bytes, PAGE_SIZE);
+
+	store->filemap = kmalloc(sizeof(struct page *)
+				 * num_pages, GFP_KERNEL);
+	if (!store->filemap)
+		return -ENOMEM;
+
+	if (with_super && !store->sb_page) {
+		store->sb_page = alloc_page(GFP_KERNEL|__GFP_ZERO);
+		if (store->sb_page == NULL)
+			return -ENOMEM;
+		store->sb_page->index = 0;
+	}
+	pnum = 0;
+	if (store->sb_page) {
+		store->filemap[0] = store->sb_page;
+		pnum = 1;
+	}
+	for ( ; pnum < num_pages; pnum++) {
+		store->filemap[pnum] = alloc_page(GFP_KERNEL|__GFP_ZERO);
+		if (!store->filemap[pnum]) {
+			store->file_pages = pnum;
+			return -ENOMEM;
+		}
+		store->filemap[pnum]->index = pnum;
+	}
+	store->file_pages = pnum;
+
+	/* We need 4 bits per page, rounded up to a multiple
+	 * of sizeof(unsigned long) */
+	store->filemap_attr = kzalloc(
+		roundup(DIV_ROUND_UP(num_pages*4, 8), sizeof(unsigned long)),
+		GFP_KERNEL);
+	if (!store->filemap_attr)
+		return -ENOMEM;
+
+	store->bytes = bytes;
+
+	return 0;
+}
+
+static void bitmap_file_unmap(struct bitmap_storage *store)
 {
 	struct page **map, *sb_page;
-	unsigned long *attr;
 	int pages;
-	unsigned long flags;
+	struct file *file;
 
-	spin_lock_irqsave(&bitmap->lock, flags);
-	map = bitmap->filemap;
-	bitmap->filemap = NULL;
-	attr = bitmap->filemap_attr;
-	bitmap->filemap_attr = NULL;
-	pages = bitmap->file_pages;
-	bitmap->file_pages = 0;
-	sb_page = bitmap->sb_page;
-	bitmap->sb_page = NULL;
-	spin_unlock_irqrestore(&bitmap->lock, flags);
+	file = store->file;
+	map = store->filemap;
+	pages = store->file_pages;
+	sb_page = store->sb_page;
 
 	while (pages--)
 		if (map[pages] != sb_page) /* 0 is sb_page, release it below */
 			free_buffers(map[pages]);
 	kfree(map);
-	kfree(attr);
+	kfree(store->filemap_attr);
 
 	if (sb_page)
 		free_buffers(sb_page);
-}
-
-static void bitmap_file_put(struct bitmap *bitmap)
-{
-	struct file *file;
-	unsigned long flags;
-
-	spin_lock_irqsave(&bitmap->lock, flags);
-	file = bitmap->file;
-	bitmap->file = NULL;
-	spin_unlock_irqrestore(&bitmap->lock, flags);
-
-	if (file)
-		wait_event(bitmap->write_wait,
-			   atomic_read(&bitmap->pending_writes)==0);
-	bitmap_file_unmap(bitmap);
 
 	if (file) {
 		struct inode *inode = file->f_path.dentry->d_inode;
···
 {
 	char *path, *ptr = NULL;
 
-	if (bitmap_mask_state(bitmap, BITMAP_STALE, MASK_SET) == 0) {
+	if (!test_and_set_bit(BITMAP_STALE, &bitmap->flags)) {
 		bitmap_update_sb(bitmap);
 
-		if (bitmap->file) {
+		if (bitmap->storage.file) {
 			path = kmalloc(PAGE_SIZE, GFP_KERNEL);
 			if (path)
-				ptr = d_path(&bitmap->file->f_path, path,
-					     PAGE_SIZE);
+				ptr = d_path(&bitmap->storage.file->f_path,
+					     path, PAGE_SIZE);
 
 			printk(KERN_ALERT
 			      "%s: kicking failed bitmap file %s from array!\n",
···
 			       "%s: disabling internal bitmap due to errors\n",
 			       bmname(bitmap));
 	}
-
-	bitmap_file_put(bitmap);
-
-	return;
 }
 
 enum bitmap_page_attr {
···
 	BITMAP_PAGE_NEEDWRITE = 2, /* there are cleared bits that need to be synced */
 };
 
-static inline void set_page_attr(struct bitmap *bitmap, struct page *page,
-				enum bitmap_page_attr attr)
+static inline void set_page_attr(struct bitmap *bitmap, int pnum,
+				 enum bitmap_page_attr attr)
 {
-	__set_bit((page->index<<2) + attr, bitmap->filemap_attr);
+	set_bit((pnum<<2) + attr, bitmap->storage.filemap_attr);
 }
 
-static inline void clear_page_attr(struct bitmap *bitmap, struct page *page,
-				enum bitmap_page_attr attr)
+static inline void clear_page_attr(struct bitmap *bitmap, int pnum,
+				   enum bitmap_page_attr attr)
 {
-	__clear_bit((page->index<<2) + attr, bitmap->filemap_attr);
+	clear_bit((pnum<<2) + attr, bitmap->storage.filemap_attr);
 }
 
-static inline unsigned long test_page_attr(struct bitmap *bitmap, struct page *page,
+static inline int test_page_attr(struct bitmap *bitmap, int pnum,
+				 enum bitmap_page_attr attr)
+{
+	return test_bit((pnum<<2) + attr, bitmap->storage.filemap_attr);
+}
+
+static inline int test_and_clear_page_attr(struct bitmap *bitmap, int pnum,
 					   enum bitmap_page_attr attr)
 {
-	return test_bit((page->index<<2) + attr, bitmap->filemap_attr);
+	return test_and_clear_bit((pnum<<2) + attr,
+				  bitmap->storage.filemap_attr);
 }
-
 /*
  * bitmap_file_set_bit -- called before performing a write to the md device
  * to set (and eventually sync) a particular bit in the bitmap file
···
 	unsigned long bit;
 	struct page *page;
 	void *kaddr;
-	unsigned long chunk = block >> bitmap->chunkshift;
+	unsigned long chunk = block >> bitmap->counts.chunkshift;
 
-	if (!bitmap->filemap)
-		return;
-
-	page = filemap_get_page(bitmap, chunk);
+	page = filemap_get_page(&bitmap->storage, chunk);
 	if (!page)
 		return;
-	bit = file_page_offset(bitmap, chunk);
+	bit = file_page_offset(&bitmap->storage, chunk);
 
 	/* set the bit */
 	kaddr = kmap_atomic(page);
-	if (bitmap->flags & BITMAP_HOSTENDIAN)
+	if (test_bit(BITMAP_HOSTENDIAN, &bitmap->flags))
 		set_bit(bit, kaddr);
 	else
-		__set_bit_le(bit, kaddr);
+		test_and_set_bit_le(bit, kaddr);
 	kunmap_atomic(kaddr);
 	pr_debug("set file bit %lu page %lu\n", bit, page->index);
 	/* record page number so it gets flushed to disk when unplug occurs */
-	set_page_attr(bitmap, page, BITMAP_PAGE_DIRTY);
+	set_page_attr(bitmap, page->index, BITMAP_PAGE_DIRTY);
+}
+
+static void bitmap_file_clear_bit(struct bitmap *bitmap, sector_t block)
+{
+	unsigned long bit;
+	struct page *page;
+	void *paddr;
+	unsigned long chunk = block >> bitmap->counts.chunkshift;
+
+	page = filemap_get_page(&bitmap->storage, chunk);
+	if (!page)
+		return;
+	bit = file_page_offset(&bitmap->storage, chunk);
+	paddr = kmap_atomic(page);
+	if (test_bit(BITMAP_HOSTENDIAN, &bitmap->flags))
+		clear_bit(bit, paddr);
+	else
+		test_and_clear_bit_le(bit, paddr);
+	kunmap_atomic(paddr);
+	if (!test_page_attr(bitmap, page->index, BITMAP_PAGE_NEEDWRITE)) {
+		set_page_attr(bitmap, page->index, BITMAP_PAGE_PENDING);
+		bitmap->allclean = 0;
+	}
 }
 
 /* this gets called when the md device is ready to unplug its underlying
···
  * sync the dirty pages of the bitmap file to disk */
 void bitmap_unplug(struct bitmap *bitmap)
 {
-	unsigned long i, flags;
+	unsigned long i;
 	int dirty, need_write;
-	struct page *page;
 	int wait = 0;
 
-	if (!bitmap)
+	if (!bitmap || !bitmap->storage.filemap ||
+	    test_bit(BITMAP_STALE, &bitmap->flags))
 		return;
 
 	/* look at each page to see if there are any set bits that need to be
 	 * flushed out to disk */
-	for (i = 0; i < bitmap->file_pages; i++) {
-		spin_lock_irqsave(&bitmap->lock, flags);
-		if (!bitmap->filemap) {
-			spin_unlock_irqrestore(&bitmap->lock, flags);
+	for (i = 0; i < bitmap->storage.file_pages; i++) {
+		if (!bitmap->storage.filemap)
 			return;
+		dirty = test_and_clear_page_attr(bitmap, i, BITMAP_PAGE_DIRTY);
+		need_write = test_and_clear_page_attr(bitmap, i,
+						      BITMAP_PAGE_NEEDWRITE);
+		if (dirty || need_write) {
+			clear_page_attr(bitmap, i, BITMAP_PAGE_PENDING);
+			write_page(bitmap, bitmap->storage.filemap[i], 0);
 		}
-		page = bitmap->filemap[i];
-		dirty = test_page_attr(bitmap, page, BITMAP_PAGE_DIRTY);
-		need_write = test_page_attr(bitmap, page, BITMAP_PAGE_NEEDWRITE);
-		clear_page_attr(bitmap, page, BITMAP_PAGE_DIRTY);
-		clear_page_attr(bitmap, page, BITMAP_PAGE_NEEDWRITE);
 		if (dirty)
 			wait = 1;
-		spin_unlock_irqrestore(&bitmap->lock, flags);
-
-		if (dirty || need_write)
-			write_page(bitmap, page, 0);
 	}
 	if (wait) { /* if any writes were performed, we need to wait on them */
-		if (bitmap->file)
+		if (bitmap->storage.file)
 			wait_event(bitmap->write_wait,
 				   atomic_read(&bitmap->pending_writes)==0);
 		else
 			md_super_wait(bitmap->mddev);
 	}
-	if (bitmap->flags & BITMAP_WRITE_ERROR)
+	if (test_bit(BITMAP_WRITE_ERROR, &bitmap->flags))
 		bitmap_file_kick(bitmap);
 }
 EXPORT_SYMBOL(bitmap_unplug);
···
 static int bitmap_init_from_disk(struct bitmap *bitmap, sector_t start)
 {
 	unsigned long i, chunks, index, oldindex, bit;
-	struct page *page = NULL, *oldpage = NULL;
-	unsigned long num_pages, bit_cnt = 0;
+	struct page *page = NULL;
+	unsigned long bit_cnt = 0;
 	struct file *file;
-	unsigned long bytes, offset;
+	unsigned long offset;
 	int outofdate;
 	int ret = -ENOSPC;
 	void *paddr;
+	struct bitmap_storage *store = &bitmap->storage;
 
-	chunks = bitmap->chunks;
-	file = bitmap->file;
+	chunks = bitmap->counts.chunks;
+	file = store->file;
 
-	BUG_ON(!file && !bitmap->mddev->bitmap_info.offset);
+	if (!file && !bitmap->mddev->bitmap_info.offset) {
+		/* No permanent bitmap - fill with '1s'. */
+		store->filemap = NULL;
+		store->file_pages = 0;
+		for (i = 0; i < chunks ; i++) {
+			/* if the disk bit is set, set the memory bit */
+			int needed = ((sector_t)(i+1) << (bitmap->counts.chunkshift)
+				      >= start);
+			bitmap_set_memory_bits(bitmap,
+					       (sector_t)i << bitmap->counts.chunkshift,
+					       needed);
+		}
+		return 0;
+	}
 
-	outofdate = bitmap->flags & BITMAP_STALE;
+	outofdate = test_bit(BITMAP_STALE, &bitmap->flags);
 	if (outofdate)
 		printk(KERN_INFO "%s: bitmap file is out of date, doing full "
 			"recovery\n", bmname(bitmap));
 
-	bytes = DIV_ROUND_UP(bitmap->chunks, 8);
-	if (!bitmap->mddev->bitmap_info.external)
-		bytes += sizeof(bitmap_super_t);
-
-	num_pages = DIV_ROUND_UP(bytes, PAGE_SIZE);
-
-	if (file && i_size_read(file->f_mapping->host) < bytes) {
+	if (file && i_size_read(file->f_mapping->host) < store->bytes) {
 		printk(KERN_INFO "%s: bitmap file too short %lu < %lu\n",
-			bmname(bitmap),
-			(unsigned long) i_size_read(file->f_mapping->host),
-			bytes);
+		       bmname(bitmap),
+		       (unsigned long) i_size_read(file->f_mapping->host),
+		       store->bytes);
 		goto err;
 	}
 
-	ret = -ENOMEM;
-
-	bitmap->filemap = kmalloc(sizeof(struct page *) * num_pages, GFP_KERNEL);
-	if (!bitmap->filemap)
-		goto err;
-
-	/* We need 4 bits per page, rounded up to a multiple of sizeof(unsigned long) */
-	bitmap->filemap_attr = kzalloc(
-		roundup(DIV_ROUND_UP(num_pages*4, 8), sizeof(unsigned long)),
-		GFP_KERNEL);
-	if (!bitmap->filemap_attr)
-		goto err;
-
 	oldindex = ~0L;
+	offset = 0;
+	if (!bitmap->mddev->bitmap_info.external)
+		offset = sizeof(bitmap_super_t);
 
 	for (i = 0; i < chunks; i++) {
 		int b;
-		index = file_page_index(bitmap, i);
-		bit = file_page_offset(bitmap, i);
+		index = file_page_index(&bitmap->storage, i);
+		bit = file_page_offset(&bitmap->storage, i);
 		if (index != oldindex) { /* this is a new page, read it in */
 			int count;
 			/* unmap the old page, we're done with it */
-			if (index == num_pages-1)
-				count = bytes - index * PAGE_SIZE;
+			if (index == store->file_pages-1)
+				count = store->bytes - index * PAGE_SIZE;
 			else
 				count = PAGE_SIZE;
-			if (index == 0 && bitmap->sb_page) {
-				/*
-				 * if we're here then the superblock page
-				 * contains some bits (PAGE_SIZE != sizeof sb)
-				 * we've already read it in, so just use it
-				 */
-				page = bitmap->sb_page;
-				offset = sizeof(bitmap_super_t);
-				if (!file)
-					page = read_sb_page(
-						bitmap->mddev,
-						bitmap->mddev->bitmap_info.offset,
-						page,
-						index, count);
-			} else if (file) {
-				page = read_page(file, index, bitmap, count);
-				offset = 0;
-			} else {
-				page = read_sb_page(bitmap->mddev,
-						    bitmap->mddev->bitmap_info.offset,
-						    NULL,
-						    index, count);
-				offset = 0;
-			}
-			if (IS_ERR(page)) { /* read error */
-				ret = PTR_ERR(page);
+			page = store->filemap[index];
+			if (file)
+				ret = read_page(file, index, bitmap,
+						count, page);
+			else
+				ret = read_sb_page(
+					bitmap->mddev,
+					bitmap->mddev->bitmap_info.offset,
+					page,
+					index, count);
+
+			if (ret)
 				goto err;
-			}
 
 			oldindex = index;
-			oldpage = page;
-
-			bitmap->filemap[bitmap->file_pages++] = page;
-			bitmap->last_page_size = count;
 
 			if (outofdate) {
 				/*
···
 				write_page(bitmap, page, 1);
 
 				ret = -EIO;
-				if (bitmap->flags & BITMAP_WRITE_ERROR)
+				if (test_bit(BITMAP_WRITE_ERROR,
+					     &bitmap->flags))
 					goto err;
 			}
 		}
 		paddr = kmap_atomic(page);
-		if (bitmap->flags & BITMAP_HOSTENDIAN)
+		if (test_bit(BITMAP_HOSTENDIAN, &bitmap->flags))
 			b = test_bit(bit, paddr);
 		else
 			b = test_bit_le(bit, paddr);
 		kunmap_atomic(paddr);
 		if (b) {
 			/* if the disk bit is set, set the memory bit */
-			int needed = ((sector_t)(i+1) << bitmap->chunkshift
+			int needed = ((sector_t)(i+1) << bitmap->counts.chunkshift
 				      >= start);
 			bitmap_set_memory_bits(bitmap,
-					       (sector_t)i << bitmap->chunkshift,
+					       (sector_t)i << bitmap->counts.chunkshift,
 					       needed);
 			bit_cnt++;
 		}
-	}
-
-	/* everything went OK */
-	ret = 0;
-	bitmap_mask_state(bitmap, BITMAP_STALE, MASK_UNSET);
-
-	if (bit_cnt) { /* Kick recovery if any bits were set */
-		set_bit(MD_RECOVERY_NEEDED, &bitmap->mddev->recovery);
-		md_wakeup_thread(bitmap->mddev->thread);
+		offset = 0;
 	}
 
 	printk(KERN_INFO "%s: bitmap initialized from disk: "
-	       "read %lu/%lu pages, set %lu of %lu bits\n",
-	       bmname(bitmap), bitmap->file_pages, num_pages, bit_cnt, chunks);
+	       "read %lu pages, set %lu of %lu bits\n",
+	       bmname(bitmap), store->file_pages,
+	       bit_cnt, chunks);
 
 	return 0;
 
···
 	 */
 	int i;
 
-	spin_lock_irq(&bitmap->lock);
-	for (i = 0; i < bitmap->file_pages; i++)
-		set_page_attr(bitmap, bitmap->filemap[i],
+	if (!bitmap || !bitmap->storage.filemap)
+		return;
+	if (bitmap->storage.file)
+		/* Only one copy, so nothing needed */
+		return;
+
+	for (i = 0; i < bitmap->storage.file_pages; i++)
+		set_page_attr(bitmap, i,
 			      BITMAP_PAGE_NEEDWRITE);
 	bitmap->allclean = 0;
-	spin_unlock_irq(&bitmap->lock);
 }
 
-static void bitmap_count_page(struct bitmap *bitmap, sector_t offset, int inc)
+static void bitmap_count_page(struct bitmap_counts *bitmap,
+			      sector_t offset, int inc)
 {
 	sector_t chunk = offset >> bitmap->chunkshift;
 	unsigned long page = chunk >> PAGE_COUNTER_SHIFT;
 	bitmap->bp[page].count += inc;
 	bitmap_checkfree(bitmap, page);
 }
-static bitmap_counter_t *bitmap_get_counter(struct bitmap *bitmap,
+
+static void bitmap_set_pending(struct bitmap_counts *bitmap, sector_t offset)
+{
+	sector_t chunk = offset >> bitmap->chunkshift;
+	unsigned long page = chunk >> PAGE_COUNTER_SHIFT;
+	struct bitmap_page *bp = &bitmap->bp[page];
+
+	if (!bp->pending)
+		bp->pending = 1;
+}
+
+static bitmap_counter_t *bitmap_get_counter(struct bitmap_counts *bitmap,
 					    sector_t offset, sector_t *blocks,
 					    int create);
 
···
 {
 	struct bitmap *bitmap;
 	unsigned long j;
-	unsigned long flags;
-	struct page *page = NULL, *lastpage = NULL;
+	unsigned long nextpage;
 	sector_t blocks;
-	void *paddr;
+	struct bitmap_counts *counts;
 
 	/* Use a mutex to guard daemon_work against
 	 * bitmap_destroy.
···
 	}
 	bitmap->allclean = 1;
 
-	spin_lock_irqsave(&bitmap->lock, flags);
-	for (j = 0; j < bitmap->chunks; j++) {
-		bitmap_counter_t *bmc;
-		if (!bitmap->filemap)
-			/* error or shutdown */
-			break;
+	/* Any file-page which is PENDING now needs to be written.
+	 * So set NEEDWRITE now, then after we make any last-minute changes
+	 * we will write it.
+	 */
+	for (j = 0; j < bitmap->storage.file_pages; j++)
+		if (test_and_clear_page_attr(bitmap, j,
+					     BITMAP_PAGE_PENDING))
+			set_page_attr(bitmap, j,
+				      BITMAP_PAGE_NEEDWRITE);
 
-		page = filemap_get_page(bitmap, j);
-
-		if (page != lastpage) {
-			/* skip this page unless it's marked as needing cleaning */
-			if (!test_page_attr(bitmap, page, BITMAP_PAGE_PENDING)) {
-				int need_write = test_page_attr(bitmap, page,
-								BITMAP_PAGE_NEEDWRITE);
-				if (need_write)
-					clear_page_attr(bitmap, page, BITMAP_PAGE_NEEDWRITE);
-
-				spin_unlock_irqrestore(&bitmap->lock, flags);
-				if (need_write)
-					write_page(bitmap, page, 0);
-				spin_lock_irqsave(&bitmap->lock, flags);
-				j |= (PAGE_BITS - 1);
-				continue;
-			}
-
-			/* grab the new page, sync and release the old */
-			if (lastpage != NULL) {
-				if (test_page_attr(bitmap, lastpage,
-						   BITMAP_PAGE_NEEDWRITE)) {
-					clear_page_attr(bitmap, lastpage,
-							BITMAP_PAGE_NEEDWRITE);
-					spin_unlock_irqrestore(&bitmap->lock, flags);
-					write_page(bitmap, lastpage, 0);
-				} else {
-					set_page_attr(bitmap, lastpage,
-						      BITMAP_PAGE_NEEDWRITE);
-					bitmap->allclean = 0;
-					spin_unlock_irqrestore(&bitmap->lock, flags);
-				}
-			} else
-				spin_unlock_irqrestore(&bitmap->lock, flags);
-			lastpage = page;
-
-			/* We are possibly going to clear some bits, so make
-			 * sure that events_cleared is up-to-date.
-			 */
-			if (bitmap->need_sync &&
-			    mddev->bitmap_info.external == 0) {
-				bitmap_super_t *sb;
-				bitmap->need_sync = 0;
-				sb = kmap_atomic(bitmap->sb_page);
-				sb->events_cleared =
-					cpu_to_le64(bitmap->events_cleared);
-				kunmap_atomic(sb);
-				write_page(bitmap, bitmap->sb_page, 1);
-			}
-			spin_lock_irqsave(&bitmap->lock, flags);
-			if (!bitmap->need_sync)
-				clear_page_attr(bitmap, page, BITMAP_PAGE_PENDING);
-			else
-				bitmap->allclean = 0;
-		}
-		bmc = bitmap_get_counter(bitmap,
-					 (sector_t)j << bitmap->chunkshift,
-					 &blocks, 0);
-		if (!bmc)
-			j |= PAGE_COUNTER_MASK;
-		else if (*bmc) {
-			if (*bmc == 1 && !bitmap->need_sync) {
-				/* we can clear the bit */
-				*bmc = 0;
-				bitmap_count_page(bitmap,
-						  (sector_t)j << bitmap->chunkshift,
-						  -1);
-
-				/* clear the bit */
-				paddr = kmap_atomic(page);
-				if (bitmap->flags & BITMAP_HOSTENDIAN)
-					clear_bit(file_page_offset(bitmap, j),
-						  paddr);
-				else
-					__clear_bit_le(
-						file_page_offset(bitmap,
-								 j),
-						paddr);
-				kunmap_atomic(paddr);
-			} else if (*bmc <= 2) {
-				*bmc = 1; /* maybe clear the bit next time */
-				set_page_attr(bitmap, page, BITMAP_PAGE_PENDING);
-				bitmap->allclean = 0;
-			}
+	if (bitmap->need_sync &&
+	    mddev->bitmap_info.external == 0) {
+		/* Arrange for superblock update as well as
+		 * other changes */
+		bitmap_super_t *sb;
+		bitmap->need_sync = 0;
+		if (bitmap->storage.filemap) {
+			sb = kmap_atomic(bitmap->storage.sb_page);
+			sb->events_cleared =
+				cpu_to_le64(bitmap->events_cleared);
+			kunmap_atomic(sb);
+			set_page_attr(bitmap, 0,
+				      BITMAP_PAGE_NEEDWRITE);
 		}
 	}
-	spin_unlock_irqrestore(&bitmap->lock, flags);
+	/* Now look at the bitmap counters and if any are '2' or '1',
+	 * decrement and handle accordingly.
+	 */
+	counts = &bitmap->counts;
+	spin_lock_irq(&counts->lock);
+	nextpage = 0;
+	for (j = 0; j < counts->chunks; j++) {
+		bitmap_counter_t *bmc;
+		sector_t  block = (sector_t)j << counts->chunkshift;
 
-	/* now sync the final page */
-	if (lastpage != NULL) {
-		spin_lock_irqsave(&bitmap->lock, flags);
-		if (test_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE)) {
-			clear_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
-			spin_unlock_irqrestore(&bitmap->lock, flags);
-			write_page(bitmap, lastpage, 0);
-		} else {
-			set_page_attr(bitmap, lastpage, BITMAP_PAGE_NEEDWRITE);
+		if (j == nextpage) {
+			nextpage += PAGE_COUNTER_RATIO;
+			if (!counts->bp[j >> PAGE_COUNTER_SHIFT].pending) {
+				j |= PAGE_COUNTER_MASK;
+				continue;
+			}
+			counts->bp[j >> PAGE_COUNTER_SHIFT].pending = 0;
+		}
+		bmc = bitmap_get_counter(counts,
+					 block,
+					 &blocks, 0);
+
+		if (!bmc) {
+			j |= PAGE_COUNTER_MASK;
+			continue;
+		}
+		if (*bmc == 1 && !bitmap->need_sync) {
+			/* We can clear the bit */
+			*bmc = 0;
+			bitmap_count_page(counts, block, -1);
+			bitmap_file_clear_bit(bitmap, block);
+		} else if (*bmc && *bmc <= 2) {
+			*bmc = 1;
+			bitmap_set_pending(counts, block);
 			bitmap->allclean = 0;
-			spin_unlock_irqrestore(&bitmap->lock, flags);
+		}
+	}
+	spin_unlock_irq(&counts->lock);
+
+	/* Now start writeout on any page in NEEDWRITE that isn't DIRTY.
+	 * DIRTY pages need to be written by bitmap_unplug so it can wait
+	 * for them.
+	 * If we find any DIRTY page we stop there and let bitmap_unplug
+	 * handle all the rest.  This is important in the case where
+	 * the first blocking holds the superblock and it has been updated.
+	 * We mustn't write any other blocks before the superblock.
+	 */
+	for (j = 0;
+	     j < bitmap->storage.file_pages
+		     && !test_bit(BITMAP_STALE, &bitmap->flags);
+	     j++) {
+
+		if (test_page_attr(bitmap, j,
+				   BITMAP_PAGE_DIRTY))
+			/* bitmap_unplug will handle the rest */
+			break;
+		if (test_and_clear_page_attr(bitmap, j,
+					     BITMAP_PAGE_NEEDWRITE)) {
+			write_page(bitmap, bitmap->storage.filemap[j], 0);
 		}
 	}
 
···
 	mutex_unlock(&mddev->bitmap_info.mutex);
 }
 
-static bitmap_counter_t *bitmap_get_counter(struct bitmap *bitmap,
+static bitmap_counter_t *bitmap_get_counter(struct bitmap_counts *bitmap,
 					    sector_t offset, sector_t *blocks,
 					    int create)
 __releases(bitmap->lock)
···
 		sector_t blocks;
 		bitmap_counter_t *bmc;
 
-		spin_lock_irq(&bitmap->lock);
-		bmc = bitmap_get_counter(bitmap, offset, &blocks, 1);
+		spin_lock_irq(&bitmap->counts.lock);
+		bmc = bitmap_get_counter(&bitmap->counts, offset, &blocks, 1);
 		if (!bmc) {
-			spin_unlock_irq(&bitmap->lock);
+			spin_unlock_irq(&bitmap->counts.lock);
 			return 0;
 		}
 
···
 			 */
 			prepare_to_wait(&bitmap->overflow_wait, &__wait,
 					TASK_UNINTERRUPTIBLE);
-			spin_unlock_irq(&bitmap->lock);
+			spin_unlock_irq(&bitmap->counts.lock);
 			io_schedule();
 			finish_wait(&bitmap->overflow_wait, &__wait);
 			continue;
···
 		switch (*bmc) {
 		case 0:
 			bitmap_file_set_bit(bitmap, offset);
-			bitmap_count_page(bitmap, offset, 1);
+			bitmap_count_page(&bitmap->counts, offset, 1);
 			/* fall through */
 		case 1:
 			*bmc = 2;
···
 
 		(*bmc)++;
 
-		spin_unlock_irq(&bitmap->lock);
+		spin_unlock_irq(&bitmap->counts.lock);
 
 		offset += blocks;
 		if (sectors > blocks)
···
 		unsigned long flags;
 		bitmap_counter_t *bmc;
 
-		spin_lock_irqsave(&bitmap->lock, flags);
-		bmc = bitmap_get_counter(bitmap, offset, &blocks, 0);
+		spin_lock_irqsave(&bitmap->counts.lock, flags);
+		bmc = bitmap_get_counter(&bitmap->counts, offset, &blocks, 0);
 		if (!bmc) {
-			spin_unlock_irqrestore(&bitmap->lock, flags);
+			spin_unlock_irqrestore(&bitmap->counts.lock, flags);
 			return;
 		}
 
···
 
 		(*bmc)--;
 		if (*bmc <= 2) {
-			set_page_attr(bitmap,
-				      filemap_get_page(
-					      bitmap,
-					      offset >> bitmap->chunkshift),
-				      BITMAP_PAGE_PENDING);
+			bitmap_set_pending(&bitmap->counts, offset);
 			bitmap->allclean = 0;
 		}
-		spin_unlock_irqrestore(&bitmap->lock, flags);
+		spin_unlock_irqrestore(&bitmap->counts.lock, flags);
 		offset += blocks;
 		if (sectors > blocks)
 			sectors -= blocks;
···
 		*blocks = 1024;
 		return 1; /* always resync if no bitmap */
 	}
-	spin_lock_irq(&bitmap->lock);
-	bmc = bitmap_get_counter(bitmap, offset, blocks, 0);
+	spin_lock_irq(&bitmap->counts.lock);
+	bmc = bitmap_get_counter(&bitmap->counts, offset, blocks, 0);
 	rv = 0;
 	if (bmc) {
 		/* locked */
···
 			}
 		}
 	}
-	spin_unlock_irq(&bitmap->lock);
+	spin_unlock_irq(&bitmap->counts.lock);
 	return rv;
 }
 
···
 		*blocks = 1024;
 		return;
 	}
-	spin_lock_irqsave(&bitmap->lock, flags);
-	bmc = bitmap_get_counter(bitmap, offset, blocks, 0);
+	spin_lock_irqsave(&bitmap->counts.lock, flags);
+	bmc = bitmap_get_counter(&bitmap->counts, offset, blocks, 0);
 	if (bmc == NULL)
 		goto unlock;
 	/* locked */
···
 			*bmc |= NEEDED_MASK;
 		else {
 			if (*bmc <= 2) {
-				set_page_attr(bitmap,
-					      filemap_get_page(bitmap, offset >> bitmap->chunkshift),
-					      BITMAP_PAGE_PENDING);
+				bitmap_set_pending(&bitmap->counts, offset);
 				bitmap->allclean = 0;
 			}
 		}
 	}
  unlock:
-	spin_unlock_irqrestore(&bitmap->lock, flags);
+	spin_unlock_irqrestore(&bitmap->counts.lock, flags);
 }
 EXPORT_SYMBOL(bitmap_end_sync);
 
···
 
 	bitmap->mddev->curr_resync_completed = sector;
 	set_bit(MD_CHANGE_CLEAN, &bitmap->mddev->flags);
-	sector &= ~((1ULL << bitmap->chunkshift) - 1);
+	sector &= ~((1ULL << bitmap->counts.chunkshift) - 1);
 	s = 0;
 	while (s < sector && s < bitmap->mddev->resync_max_sectors) {
 		bitmap_end_sync(bitmap, s, &blocks, 0);
···
 static void bitmap_set_memory_bits(struct bitmap *bitmap, sector_t offset, int needed)
 {
 	/* For each chunk covered by any of these sectors, set the
-	 * counter to 1 and set resync_needed.  They should all
+	 * counter to 2 and possibly set resync_needed.  They should all
 	 * be 0 at this point
 	 */
 
 	sector_t secs;
 	bitmap_counter_t *bmc;
-	spin_lock_irq(&bitmap->lock);
-	bmc = bitmap_get_counter(bitmap, offset, &secs, 1);
+	spin_lock_irq(&bitmap->counts.lock);
+	bmc = bitmap_get_counter(&bitmap->counts, offset, &secs, 1);
 	if (!bmc) {
-		spin_unlock_irq(&bitmap->lock);
+		spin_unlock_irq(&bitmap->counts.lock);
 		return;
 	}
 	if (!*bmc) {
-		struct page *page;
 		*bmc = 2 | (needed ? NEEDED_MASK : 0);
-		bitmap_count_page(bitmap, offset, 1);
-		page = filemap_get_page(bitmap, offset >> bitmap->chunkshift);
-		set_page_attr(bitmap, page, BITMAP_PAGE_PENDING);
+		bitmap_count_page(&bitmap->counts, offset, 1);
+		bitmap_set_pending(&bitmap->counts, offset);
 		bitmap->allclean = 0;
 	}
-	spin_unlock_irq(&bitmap->lock);
+	spin_unlock_irq(&bitmap->counts.lock);
 }
 
 /* dirty the memory and file bits for bitmap chunks "s" to "e" */
···
 	unsigned long chunk;
 
 	for (chunk = s; chunk <= e; chunk++) {
-		sector_t sec = (sector_t)chunk << bitmap->chunkshift;
+		sector_t sec = (sector_t)chunk << bitmap->counts.chunkshift;
 		bitmap_set_memory_bits(bitmap, sec, 1);
-		spin_lock_irq(&bitmap->lock);
 		bitmap_file_set_bit(bitmap, sec);
-		spin_unlock_irq(&bitmap->lock);
 		if (sec < bitmap->mddev->recovery_cp)
 			/* We are asserting that the array is dirty,
 			 * so move the recovery_cp address back so
···
 	if (!bitmap) /* there was no bitmap */
 		return;
 
-	/* release the bitmap file and kill the daemon */
-	bitmap_file_put(bitmap);
+	/* Shouldn't be needed - but just in case.... */
+	wait_event(bitmap->write_wait,
+		   atomic_read(&bitmap->pending_writes) == 0);
 
-	bp = bitmap->bp;
-	pages = bitmap->pages;
+	/* release the bitmap file  */
+	bitmap_file_unmap(&bitmap->storage);
+
+	bp = bitmap->counts.bp;
+	pages = bitmap->counts.pages;
 
 	/* free all allocated memory */
 
···
 {
 	struct bitmap *bitmap;
 	sector_t blocks = mddev->resync_max_sectors;
-	unsigned long chunks;
-	unsigned long pages;
 	struct file *file = mddev->bitmap_info.file;
 	int err;
 	struct sysfs_dirent *bm = NULL;
 
 	BUILD_BUG_ON(sizeof(bitmap_super_t) != 256);
-
-	if (!file
-	    && !mddev->bitmap_info.offset) /* bitmap disabled, nothing to do */
-		return 0;
 
 	BUG_ON(file && mddev->bitmap_info.offset);
 
···
 	if (!bitmap)
 		return -ENOMEM;
 
-	spin_lock_init(&bitmap->lock);
+	spin_lock_init(&bitmap->counts.lock);
 	atomic_set(&bitmap->pending_writes, 0);
 	init_waitqueue_head(&bitmap->write_wait);
 	init_waitqueue_head(&bitmap->overflow_wait);
···
 	} else
 		bitmap->sysfs_can_clear = NULL;
 
-	bitmap->file = file;
+	bitmap->storage.file = file;
 	if (file) {
 		get_file(file);
 		/* As future accesses to this file will use bmap,
···
 		goto error;
 
 	bitmap->daemon_lastrun = jiffies;
-	bitmap->chunkshift = (ffz(~mddev->bitmap_info.chunksize)
-			      - BITMAP_BLOCK_SHIFT);
-
-	chunks = (blocks + (1 << bitmap->chunkshift) - 1) >>
-			bitmap->chunkshift;
-	pages = (chunks + PAGE_COUNTER_RATIO - 1) / PAGE_COUNTER_RATIO;
-
-	BUG_ON(!pages);
-
-	bitmap->chunks = chunks;
-	bitmap->pages = pages;
-	bitmap->missing_pages = pages;
-
-	bitmap->bp = kzalloc(pages * sizeof(*bitmap->bp), GFP_KERNEL);
-
-	err = -ENOMEM;
-	if (!bitmap->bp)
+	err = bitmap_resize(bitmap, blocks, mddev->bitmap_info.chunksize, 1);
+	if (err)
 		goto error;
 
 	printk(KERN_INFO "created bitmap (%lu pages) for device %s\n",
-		pages, bmname(bitmap));
+	       bitmap->counts.pages, bmname(bitmap));
 
 	mddev->bitmap = bitmap;
-
-
-	return (bitmap->flags & BITMAP_WRITE_ERROR) ? -EIO : 0;
+	return test_bit(BITMAP_WRITE_ERROR, &bitmap->flags) ? -EIO : 0;
 
  error:
 	bitmap_free(bitmap);
···
 
 	if (err)
 		goto out;
+	clear_bit(BITMAP_STALE, &bitmap->flags);
+
+	/* Kick recovery in case any bits were set */
+	set_bit(MD_RECOVERY_NEEDED, &bitmap->mddev->recovery);
 
 	mddev->thread->timeout = mddev->bitmap_info.daemon_sleep;
 	md_wakeup_thread(mddev->thread);
 
 	bitmap_update_sb(bitmap);
 
-	if (bitmap->flags & BITMAP_WRITE_ERROR)
+	if (test_bit(BITMAP_WRITE_ERROR, &bitmap->flags))
 		err = -EIO;
 out:
 	return err;
···
 void bitmap_status(struct seq_file *seq, struct bitmap *bitmap)
 {
 	unsigned long chunk_kb;
-	unsigned long flags;
+	struct bitmap_counts *counts;
 
 	if (!bitmap)
 		return;
 
-	spin_lock_irqsave(&bitmap->lock, flags);
+	counts = &bitmap->counts;
+
 	chunk_kb = bitmap->mddev->bitmap_info.chunksize >> 10;
 	seq_printf(seq, "bitmap: %lu/%lu pages [%luKB], "
 		   "%lu%s chunk",
-		   bitmap->pages - bitmap->missing_pages,
-		   bitmap->pages,
-		   (bitmap->pages - bitmap->missing_pages)
+		   counts->pages - counts->missing_pages,
+		   counts->pages,
+		   (counts->pages - counts->missing_pages)
 		   << (PAGE_SHIFT - 10),
 		   chunk_kb ? chunk_kb : bitmap->mddev->bitmap_info.chunksize,
 		   chunk_kb ? "KB" : "B");
-	if (bitmap->file) {
+	if (bitmap->storage.file) {
 		seq_printf(seq, ", file: ");
-		seq_path(seq, &bitmap->file->f_path, " \t\n");
+		seq_path(seq, &bitmap->storage.file->f_path, " \t\n");
 	}
 
 	seq_printf(seq, "\n");
-	spin_unlock_irqrestore(&bitmap->lock, flags);
 }
+
+int bitmap_resize(struct bitmap *bitmap, sector_t blocks,
+		  int chunksize, int init)
+{
+	/* If chunk_size is 0, choose an appropriate chunk size.
+	 * Then possibly allocate new storage space.
+	 * Then quiesce, copy bits, replace bitmap, and re-start
+	 *
+	 * This function is called both to set up the initial bitmap
+	 * and to resize the bitmap while the array is active.
+	 * If this happens as a result of the array being resized,
+	 * chunksize will be zero, and we need to choose a suitable
+	 * chunksize, otherwise we use what we are given.
+	 */
+	struct bitmap_storage store;
+	struct bitmap_counts old_counts;
+	unsigned long chunks;
+	sector_t block;
+	sector_t old_blocks, new_blocks;
+	int chunkshift;
+	int ret = 0;
+	long pages;
+	struct bitmap_page *new_bp;
+
+	if (chunksize == 0) {
+		/* If there is enough space, leave the chunk size unchanged,
+		 * else increase by factor of two until there is enough space.
+		 */
+		long bytes;
+		long space = bitmap->mddev->bitmap_info.space;
+
+		if (space == 0) {
+			/* We don't know how much space there is, so limit
+			 * to current size - in sectors.
+			 */
+			bytes = DIV_ROUND_UP(bitmap->counts.chunks, 8);
+			if (!bitmap->mddev->bitmap_info.external)
+				bytes += sizeof(bitmap_super_t);
+			space = DIV_ROUND_UP(bytes, 512);
+			bitmap->mddev->bitmap_info.space = space;
+		}
+		chunkshift = bitmap->counts.chunkshift;
+		chunkshift--;
+		do {
+			/* 'chunkshift' is shift from block size to chunk size */
+			chunkshift++;
+			chunks = DIV_ROUND_UP_SECTOR_T(blocks, 1 << chunkshift);
+			bytes = DIV_ROUND_UP(chunks, 8);
+			if (!bitmap->mddev->bitmap_info.external)
+				bytes += sizeof(bitmap_super_t);
+		} while (bytes > (space << 9));
+	} else
+		chunkshift = ffz(~chunksize) - BITMAP_BLOCK_SHIFT;
+
+	chunks = DIV_ROUND_UP_SECTOR_T(blocks, 1 << chunkshift);
+	memset(&store, 0, sizeof(store));
+	if (bitmap->mddev->bitmap_info.offset || bitmap->mddev->bitmap_info.file)
+		ret = bitmap_storage_alloc(&store, chunks,
+					   !bitmap->mddev->bitmap_info.external);
+	if (ret)
+		goto err;
+
+	pages = DIV_ROUND_UP(chunks, PAGE_COUNTER_RATIO);
+
+	new_bp = kzalloc(pages * sizeof(*new_bp), GFP_KERNEL);
+	ret = -ENOMEM;
+	if (!new_bp) {
+		bitmap_file_unmap(&store);
+		goto err;
+	}
+
+	if (!init)
+		bitmap->mddev->pers->quiesce(bitmap->mddev, 1);
+
+	store.file = bitmap->storage.file;
+	bitmap->storage.file = NULL;
+
+	if (store.sb_page && bitmap->storage.sb_page)
+		memcpy(page_address(store.sb_page),
+		       page_address(bitmap->storage.sb_page),
+		       sizeof(bitmap_super_t));
+	bitmap_file_unmap(&bitmap->storage);
+	bitmap->storage = store;
+
+	old_counts = bitmap->counts;
+	bitmap->counts.bp = new_bp;
+	bitmap->counts.pages = pages;
+	bitmap->counts.missing_pages = pages;
+	bitmap->counts.chunkshift = chunkshift;
+	bitmap->counts.chunks = chunks;
+	bitmap->mddev->bitmap_info.chunksize = 1 << (chunkshift +
+						     BITMAP_BLOCK_SHIFT);
+
+	blocks = min(old_counts.chunks << old_counts.chunkshift,
+		     chunks << chunkshift);
+
+	spin_lock_irq(&bitmap->counts.lock);
+	for (block = 0; block < blocks; ) {
+		bitmap_counter_t *bmc_old, *bmc_new;
+		int set;
+
+		bmc_old = bitmap_get_counter(&old_counts, block,
+					     &old_blocks, 0);
+		set = bmc_old && NEEDED(*bmc_old);
+
+		if (set) {
+			bmc_new = bitmap_get_counter(&bitmap->counts, block,
+						     &new_blocks, 1);
+			if (*bmc_new == 0) {
+				/* need to set on-disk bits too. */
+				sector_t end = block + new_blocks;
+				sector_t start = block >> chunkshift;
+				start <<= chunkshift;
+				while (start < end) {
+					bitmap_file_set_bit(bitmap, block);
+					start += 1 << chunkshift;
+				}
+				*bmc_new = 2;
+				bitmap_count_page(&bitmap->counts,
+						  block, 1);
+				bitmap_set_pending(&bitmap->counts,
+						   block);
+			}
+			*bmc_new |= NEEDED_MASK;
+			if (new_blocks < old_blocks)
+				old_blocks = new_blocks;
+		}
+		block += old_blocks;
+	}
+
+	if (!init) {
+		int i;
+		while (block < (chunks << chunkshift)) {
+			bitmap_counter_t *bmc;
+			bmc = bitmap_get_counter(&bitmap->counts, block,
+						 &new_blocks, 1);
+			if (bmc) {
+				/* new space.  It needs to be resynced, so
+				 * we set NEEDED_MASK.
+				 */
+				if (*bmc == 0) {
+					*bmc = NEEDED_MASK | 2;
+					bitmap_count_page(&bitmap->counts,
+							  block, 1);
+					bitmap_set_pending(&bitmap->counts,
+							   block);
+				}
+			}
+			block += new_blocks;
+		}
+		for (i = 0; i < bitmap->storage.file_pages; i++)
+			set_page_attr(bitmap, i, BITMAP_PAGE_DIRTY);
+	}
+	spin_unlock_irq(&bitmap->counts.lock);
+
+	if (!init) {
+		bitmap_unplug(bitmap);
+		bitmap->mddev->pers->quiesce(bitmap->mddev, 0);
+	}
+	ret = 0;
+err:
+	return ret;
+}
+EXPORT_SYMBOL_GPL(bitmap_resize);
 
 static ssize_t
 location_show(struct mddev *mddev, char *page)
···
 
 static struct md_sysfs_entry bitmap_location =
 __ATTR(location, S_IRUGO|S_IWUSR, location_show, location_store);
+
+/* 'bitmap/space' is the space available at 'location' for the
+ * bitmap.  This allows the kernel to know when it is safe to
+ * resize the bitmap to match a resized array.
+ */
+static ssize_t
+space_show(struct mddev *mddev, char *page)
+{
+	return sprintf(page, "%lu\n", mddev->bitmap_info.space);
+}
+
+static ssize_t
+space_store(struct mddev *mddev, const char *buf, size_t len)
+{
+	unsigned long sectors;
+	int rv;
+
+	rv = kstrtoul(buf, 10, &sectors);
+	if (rv)
+		return rv;
+
+	if (sectors == 0)
+		return -EINVAL;
+
+	if (mddev->bitmap &&
+	    sectors < (mddev->bitmap->storage.bytes + 511) >> 9)
+		return -EFBIG; /* Bitmap is too big for this small space */
+
+	/* could make sure it isn't too big, but that isn't really
+	 * needed - user-space should be careful.
+	 */
+	mddev->bitmap_info.space = sectors;
+	return len;
+}
+
+static struct md_sysfs_entry bitmap_space =
+__ATTR(space, S_IRUGO|S_IWUSR, space_show, space_store);
 
 static ssize_t
 timeout_show(struct mddev *mddev, char *page)
···
 
 static struct attribute *md_bitmap_attrs[] = {
 	&bitmap_location.attr,
+	&bitmap_space.attr,
 	&bitmap_timeout.attr,
 	&bitmap_backlog.attr,
 	&bitmap_chunksize.attr,

drivers/md/bitmap.h

···
 
 /* use these for bitmap->flags and bitmap->sb->state bit-fields */
 enum bitmap_state {
-	BITMAP_STALE  = 0x002,  /* the bitmap file is out of date or had -EIO */
-	BITMAP_WRITE_ERROR = 0x004, /* A write error has occurred */
-	BITMAP_HOSTENDIAN = 0x8000,
+	BITMAP_STALE	   = 1,  /* the bitmap file is out of date or had -EIO */
+	BITMAP_WRITE_ERROR = 2, /* A write error has occurred */
+	BITMAP_HOSTENDIAN  =15,
 };
 
 /* the superblock at the front of the bitmap file -- little endian */
···
 	__le32 chunksize;    /* 52  the bitmap chunk size in bytes */
 	__le32 daemon_sleep; /* 56  seconds between disk flushes */
 	__le32 write_behind; /* 60  number of outstanding write-behind writes */
+	__le32 sectors_reserved; /* 64 number of 512-byte sectors that are
+				  * reserved for the bitmap. */
 
-	__u8  pad[256 - 64]; /* set to zero */
+	__u8  pad[256 - 68]; /* set to zero */
 } bitmap_super_t;
 
 /* notes:
···
 	 */
 	unsigned int hijacked:1;
 	/*
+	 * If any counter in this page is '1' or '2' - and so could be
+	 * cleared then that page is marked as 'pending'
+	 */
+	unsigned int pending:1;
+	/*
 	 * count of dirty bits on the page
 	 */
-	unsigned int  count:31;
+	unsigned int  count:30;
 };
 
 /* the main bitmap structure - one per mddev */
 struct bitmap {
-	struct bitmap_page *bp;
-	unsigned long pages; /* total number of pages in the bitmap */
-	unsigned long missing_pages; /* number of pages not yet allocated */
+
+	struct bitmap_counts {
+		spinlock_t lock;
+		struct bitmap_page *bp;
+		unsigned long pages;		/* total number of pages
+						 * in the bitmap */
+		unsigned long missing_pages;	/* number of pages
+						 * not yet allocated */
+		unsigned long chunkshift;	/* chunksize = 2^chunkshift
+						 * (for bitops) */
+		unsigned long chunks;		/* Total number of data
+						 * chunks for the array */
+	} counts;
 
 	struct mddev *mddev; /* the md device that the bitmap is for */
-
-	/* bitmap chunksize -- how much data does each bit represent? */
-	unsigned long chunkshift; /* chunksize = 2^(chunkshift+9) (for bitops) */
-	unsigned long chunks; /* total number of data chunks for the array */
 
 	__u64	events_cleared;
 	int need_sync;
 
-	/* bitmap spinlock */
-	spinlock_t lock;
-
-	struct file *file; /* backing disk file */
-	struct page *sb_page; /* cached copy of the bitmap file superblock */
-	struct page **filemap; /* list of cache pages for the file */
-	unsigned long *filemap_attr; /* attributes associated w/ filemap pages */
-	unsigned long file_pages; /* number of pages in the file */
-	int last_page_size; /* bytes in the last page */
+	struct bitmap_storage {
+		struct file *file;		/* backing disk file */
+		struct page *sb_page;		/* cached copy of the bitmap
+						 * file superblock */
+		struct page **filemap;		/* list of cache pages for
+						 * the file */
+		unsigned long *filemap_attr;	/* attributes associated
+						 * w/ filemap pages */
+		unsigned long file_pages;	/* number of pages in the file*/
+		unsigned long bytes;		/* total bytes in the bitmap */
+	} storage;
 
 	unsigned long flags;
 
···
 
 void bitmap_unplug(struct bitmap *bitmap);
 void bitmap_daemon_work(struct mddev *mddev);
+
+int bitmap_resize(struct bitmap *bitmap, sector_t blocks,
+		  int chunksize, int init);
 #endif
 
 #endif

drivers/md/dm-raid.c

···
 	for (i = 0; i < rs->md.raid_disks; i++) {
 		if (rs->dev[i].meta_dev)
 			dm_put_device(rs->ti, rs->dev[i].meta_dev);
-		if (rs->dev[i].rdev.sb_page)
-			put_page(rs->dev[i].rdev.sb_page);
-		rs->dev[i].rdev.sb_page = NULL;
-		rs->dev[i].rdev.sb_loaded = 0;
+		md_rdev_clear(&rs->dev[i].rdev);
 		if (rs->dev[i].data_dev)
 			dm_put_device(rs->ti, rs->dev[i].data_dev);
 	}
···
 	if (!sync_page_io(rdev, 0, size, rdev->sb_page, READ, 1)) {
 		DMERR("Failed to read superblock of device at position %d",
 		      rdev->raid_disk);
-		set_bit(Faulty, &rdev->flags);
+		md_error(rdev->mddev, rdev);
 		return -EINVAL;
 	}
 
···
 
 static void super_sync(struct mddev *mddev, struct md_rdev *rdev)
 {
-	struct md_rdev *r;
+	int i;
 	uint64_t failed_devices;
 	struct dm_raid_superblock *sb;
+	struct raid_set *rs = container_of(mddev, struct raid_set, md);
 
 	sb = page_address(rdev->sb_page);
 	failed_devices = le64_to_cpu(sb->failed_devices);
 
-	rdev_for_each(r, mddev)
-		if ((r->raid_disk >= 0) && test_bit(Faulty, &r->flags))
-			failed_devices |= (1ULL << r->raid_disk);
+	for (i = 0; i < mddev->raid_disks; i++)
+		if (!rs->dev[i].data_dev ||
+		    test_bit(Faulty, &(rs->dev[i].rdev.flags)))
+			failed_devices |= (1ULL << i);
 
 	memset(sb, 0, sizeof(*sb));
 
···
 {
 	struct raid_set *rs = ti->private;
 
+	set_bit(MD_CHANGE_DEVS, &rs->md.flags);
 	if (!rs->bitmap_loaded) {
 		bitmap_load(&rs->md);
 		rs->bitmap_loaded = 1;
-	} else
-		md_wakeup_thread(rs->md.thread);
+	}
 
+	clear_bit(MD_RECOVERY_FROZEN, &rs->md.recovery);
 	mddev_resume(&rs->md);
 }
 

drivers/md/md.c

···
 	wake_up(&mddev->sb_wait);
 	mddev->pers->quiesce(mddev, 0);
 
+	set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
 	md_wakeup_thread(mddev->thread);
 	md_wakeup_thread(mddev->sync_thread); /* possibly kick off a reshape */
 }
···
 			atomic_inc(&rdev->nr_pending);
 			atomic_inc(&rdev->nr_pending);
 			rcu_read_unlock();
-			bi = bio_alloc_mddev(GFP_KERNEL, 0, mddev);
+			bi = bio_alloc_mddev(GFP_NOIO, 0, mddev);
 			bi->bi_end_io = md_end_flush;
 			bi->bi_private = rdev;
 			bi->bi_bdev = rdev->bdev;
···
 	init_waitqueue_head(&mddev->sb_wait);
 	init_waitqueue_head(&mddev->recovery_wait);
 	mddev->reshape_position = MaxSector;
+	mddev->reshape_backwards = 0;
 	mddev->resync_min = 0;
 	mddev->resync_max = MaxSector;
 	mddev->level = LEVEL_NONE;
···
 	return 0;
 }
 
-static void free_disk_sb(struct md_rdev * rdev)
+void md_rdev_clear(struct md_rdev *rdev)
 {
 	if (rdev->sb_page) {
 		put_page(rdev->sb_page);
···
 		put_page(rdev->bb_page);
 		rdev->bb_page = NULL;
 	}
+	kfree(rdev->badblocks.page);
+	rdev->badblocks.page = NULL;
 }
-
+EXPORT_SYMBOL_GPL(md_rdev_clear);
 
 static void super_written(struct bio *bio, int error)
 {
···
 		rdev->meta_bdev : rdev->bdev;
 	if (metadata_op)
 		bio->bi_sector = sector + rdev->sb_start;
+	else if (rdev->mddev->reshape_position != MaxSector &&
+		 (rdev->mddev->reshape_backwards ==
+		  (sector >= rdev->mddev->reshape_position)))
+		bio->bi_sector = sector + rdev->new_data_offset;
 	else
 		bio->bi_sector = sector + rdev->data_offset;
 	bio_add_page(bio, page, size, 0);
···
 struct super_type  {
 	char		    *name;
 	struct module	    *owner;
-	int		    (*load_super)(struct md_rdev *rdev, struct md_rdev *refdev,
+	int		    (*load_super)(struct md_rdev *rdev,
+					  struct md_rdev *refdev,
 					  int minor_version);
-	int		    (*validate_super)(struct mddev *mddev, struct md_rdev *rdev);
-	void		    (*sync_super)(struct mddev *mddev, struct md_rdev *rdev);
+	int		    (*validate_super)(struct mddev *mddev,
+					      struct md_rdev *rdev);
+	void		    (*sync_super)(struct mddev *mddev,
+					  struct md_rdev *rdev);
 	unsigned long long  (*rdev_size_change)(struct md_rdev *rdev,
 						sector_t num_sectors);
+	int		    (*allow_new_offset)(struct md_rdev *rdev,
+						unsigned long long new_offset);
 };
 
 /*
···
 
 	rdev->preferred_minor = sb->md_minor;
 	rdev->data_offset = 0;
+	rdev->new_data_offset = 0;
 	rdev->sb_size = MD_SB_BYTES;
 	rdev->badblocks.shift = -1;
 
···
 		mddev->dev_sectors = ((sector_t)sb->size) * 2;
 		mddev->events = ev1;
 		mddev->bitmap_info.offset = 0;
+		mddev->bitmap_info.space = 0;
+		/* bitmap can use 60 K after the 4K superblocks */
 		mddev->bitmap_info.default_offset = MD_SB_BYTES >> 9;
+		mddev->bitmap_info.default_space = 64*2 - (MD_SB_BYTES >> 9);
+		mddev->reshape_backwards = 0;
 
 		if (mddev->minor_version >= 91) {
 			mddev->reshape_position = sb->reshape_position;
···
 			mddev->new_level = sb->new_level;
 			mddev->new_layout = sb->new_layout;
 			mddev->new_chunk_sectors = sb->new_chunk >> 9;
+			if (mddev->delta_disks < 0)
+				mddev->reshape_backwards = 1;
 		} else {
 			mddev->reshape_position = MaxSector;
 			mddev->delta_disks = 0;
···
 		mddev->max_disks = MD_SB_DISKS;
 
 		if (sb->state & (1<<MD_SB_BITMAP_PRESENT) &&
-		    mddev->bitmap_info.file == NULL)
+		    mddev->bitmap_info.file == NULL) {
 			mddev->bitmap_info.offset =
 				mddev->bitmap_info.default_offset;
+			mddev->bitmap_info.space =
+				mddev->bitmap_info.space;
+		}
 
 	} else if (mddev->pers == NULL) {
 		/* Insist on good event counter while assembling, except
···
 	return num_sectors;
 }
 
+static int
+super_90_allow_new_offset(struct md_rdev *rdev, unsigned long long new_offset)
+{
+	/* non-zero offset changes not possible with v0.90 */
+	return new_offset == 0;
+}
 
 /*
  * version 1 superblock
···
 	struct mdp_superblock_1 *sb;
 	int ret;
 	sector_t sb_start;
+	sector_t sectors;
 	char b[BDEVNAME_SIZE], b2[BDEVNAME_SIZE];
 	int bmask;
 
···
 		       bdevname(rdev->bdev,b));
 		return -EINVAL;
 	}
+	if (sb->pad0 ||
+	    sb->pad3[0] ||
+	    memcmp(sb->pad3, sb->pad3+1, sizeof(sb->pad3) - sizeof(sb->pad3[1])))
+		/* Some padding is non-zero, might be a new feature */
+		return -EINVAL;
 
 	rdev->preferred_minor = 0xffff;
 	rdev->data_offset = le64_to_cpu(sb->data_offset);
+	rdev->new_data_offset = rdev->data_offset;
+	if ((le32_to_cpu(sb->feature_map) & MD_FEATURE_RESHAPE_ACTIVE) &&
+	    (le32_to_cpu(sb->feature_map) & MD_FEATURE_NEW_OFFSET))
+		rdev->new_data_offset += (s32)le32_to_cpu(sb->new_offset);
 	atomic_set(&rdev->corrected_errors, le32_to_cpu(sb->cnt_corrected_read));
 
 	rdev->sb_size = le32_to_cpu(sb->max_dev) * 2 + 256;
···
 
 	if (minor_version
 	    && rdev->data_offset < sb_start + (rdev->sb_size/512))
+		return -EINVAL;
+	if (minor_version
+	    && rdev->new_data_offset < sb_start + (rdev->sb_size/512))
 		return -EINVAL;
 
 	if (sb->level == cpu_to_le32(LEVEL_MULTIPATH))
···
 		else
 			ret = 0;
 	}
-	if (minor_version)
-		rdev->sectors = (i_size_read(rdev->bdev->bd_inode) >> 9) -
-			le64_to_cpu(sb->data_offset);
-	else
-		rdev->sectors = rdev->sb_start;
-	if (rdev->sectors < le64_to_cpu(sb->data_size))
+	if (minor_version) {
+		sectors = (i_size_read(rdev->bdev->bd_inode) >> 9);
+		sectors -= rdev->data_offset;
+	} else
+		sectors = rdev->sb_start;
+	if (sectors < le64_to_cpu(sb->data_size))
 		return -EINVAL;
 	rdev->sectors = le64_to_cpu(sb->data_size);
-	if (le64_to_cpu(sb->size) > rdev->sectors)
-		return -EINVAL;
 	return ret;
 }
 
···
 		mddev->dev_sectors = le64_to_cpu(sb->size);
 		mddev->events = ev1;
 		mddev->bitmap_info.offset = 0;
+		mddev->bitmap_info.space = 0;
+		/* Default location for bitmap is 1K after superblock
+		 * using 3K - total of 4K
+		 */
 		mddev->bitmap_info.default_offset = 1024 >> 9;
-		
+		mddev->bitmap_info.default_space = (4096-1024) >> 9;
+		mddev->reshape_backwards = 0;
+
 		mddev->recovery_cp = le64_to_cpu(sb->resync_offset);
 		memcpy(mddev->uuid, sb->set_uuid, 16);
 
 		mddev->max_disks =  (4096-256)/2;
 
 		if ((le32_to_cpu(sb->feature_map) & MD_FEATURE_BITMAP_OFFSET) &&
-		    mddev->bitmap_info.file == NULL )
+		    mddev->bitmap_info.file == NULL) {
 			mddev->bitmap_info.offset =
 				(__s32)le32_to_cpu(sb->bitmap_offset);
+			/* Metadata doesn't record how much space is available.
+			 * For 1.0, we assume we can use up to the superblock
+			 * if before, else to 4K beyond superblock.
+			 * For others, assume no change is possible.
+			 */
+			if (mddev->minor_version > 0)
+				mddev->bitmap_info.space = 0;
+			else if (mddev->bitmap_info.offset > 0)
+				mddev->bitmap_info.space =
+					8 - mddev->bitmap_info.offset;
+			else
+				mddev->bitmap_info.space =
+					-mddev->bitmap_info.offset;
+		}
 
 		if ((le32_to_cpu(sb->feature_map) & MD_FEATURE_RESHAPE_ACTIVE)) {
 			mddev->reshape_position = le64_to_cpu(sb->reshape_position);
···
 			mddev->new_level = le32_to_cpu(sb->new_level);
 			mddev->new_layout = le32_to_cpu(sb->new_layout);
 			mddev->new_chunk_sectors = le32_to_cpu(sb->new_chunk);
+			if (mddev->delta_disks < 0 ||
+			    (mddev->delta_disks == 0 &&
+			     (le32_to_cpu(sb->feature_map)
+			      & MD_FEATURE_RESHAPE_BACKWARDS)))
+				mddev->reshape_backwards = 1;
 		} else {
 			mddev->reshape_position = MaxSector;
 			mddev->delta_disks = 0;
···
 	sb->feature_map = 0;
 	sb->pad0 = 0;
 	sb->recovery_offset = cpu_to_le64(0);
-	memset(sb->pad1, 0, sizeof(sb->pad1));
 	memset(sb->pad3, 0, sizeof(sb->pad3));
 
 	sb->utime = cpu_to_le64((__u64)mddev->utime);
···
 		sb->devflags |= WriteMostly1;
 	else
 		sb->devflags &= ~WriteMostly1;
+	sb->data_offset = cpu_to_le64(rdev->data_offset);
+	sb->data_size = cpu_to_le64(rdev->sectors);
 
 	if (mddev->bitmap && mddev->bitmap_info.file == NULL) {
 		sb->bitmap_offset = cpu_to_le32((__u32)mddev->bitmap_info.offset);
···
 		sb->delta_disks = cpu_to_le32(mddev->delta_disks);
 		sb->new_level = cpu_to_le32(mddev->new_level);
 		sb->new_chunk = cpu_to_le32(mddev->new_chunk_sectors);
+		if (mddev->delta_disks == 0 &&
+		    mddev->reshape_backwards)
+			sb->feature_map
+				|= cpu_to_le32(MD_FEATURE_RESHAPE_BACKWARDS);
+		if (rdev->new_data_offset != rdev->data_offset) {
+			sb->feature_map
+				|= cpu_to_le32(MD_FEATURE_NEW_OFFSET);
+			sb->new_offset = cpu_to_le32((__u32)(rdev->new_data_offset
+							     - rdev->data_offset));
+		}
 	}
 
 	if (rdev->badblocks.count == 0)
···
 	sector_t max_sectors;
 	if (num_sectors && num_sectors < rdev->mddev->dev_sectors)
 		return 0; /* component must fit device */
+	if (rdev->data_offset != rdev->new_data_offset)
+		return 0; /* too confusing */
 	if (rdev->sb_start < rdev->data_offset) {
 		/* minor versions 1 and 2; superblock before data */
 		max_sectors = i_size_read(rdev->bdev->bd_inode) >> 9;
···
 		       rdev->sb_page);
 	md_super_wait(rdev->mddev);
 	return num_sectors;
+
+}
+
+static int
+super_1_allow_new_offset(struct md_rdev *rdev,
+			 unsigned long long new_offset)
+{
+	/* All necessary checks on new >= old have been done */
+	struct bitmap *bitmap;
+	if (new_offset >= rdev->data_offset)
+		return 1;
+
+	/* with 1.0 metadata, there is no metadata to tread on
+	 * so we can always move back */
+	if (rdev->mddev->minor_version == 0)
+		return 1;
+
+	/* otherwise we must be sure not to step on
+	 * any metadata, so stay:
+	 * 36K beyond start of superblock
+	 * beyond end of badblocks
+	 * beyond write-intent bitmap
+	 */
+	if (rdev->sb_start + (32+4)*2 > new_offset)
+		return 0;
+	bitmap = rdev->mddev->bitmap;
+	if (bitmap && !rdev->mddev->bitmap_info.file &&
+	    rdev->sb_start + rdev->mddev->bitmap_info.offset +
+	    bitmap->storage.file_pages * (PAGE_SIZE>>9) > new_offset)
+		return 0;
+	if (rdev->badblocks.sector + rdev->badblocks.size > new_offset)
+		return 0;
+
+	return 1;
 }
 
 static struct super_type super_types[] = {
···
 		.validate_super	    = super_90_validate,
 		.sync_super	    = super_90_sync,
 		.rdev_size_change   = super_90_rdev_size_change,
+		.allow_new_offset   = super_90_allow_new_offset,
 	},
 	[1] = {
 		.name	= "md-1",
···
 		.validate_super	    = super_1_validate,
 		.sync_super	    = super_1_sync,
 		.rdev_size_change   = super_1_rdev_size_change,
+		.allow_new_offset   = super_1_allow_new_offset,
 	},
 };
 
···
 	sysfs_remove_link(&rdev->kobj, "block");
 	sysfs_put(rdev->sysfs_state);
 	rdev->sysfs_state = NULL;
-	kfree(rdev->badblocks.page);
 	rdev->badblocks.count = 0;
-	rdev->badblocks.page = NULL;
 	/* We need to delay this, otherwise we can deadlock when
 	 * writing to 'remove' to "dev/state".  We also need
 	 * to delay it due to rcu usage.
···
 		bdevname(rdev->bdev,b));
 	if (rdev->mddev)
 		MD_BUG();
-	free_disk_sb(rdev);
+	md_rdev_clear(rdev);
 #ifndef MODULE
 	if (test_bit(AutoDetected, &rdev->flags))
 		md_autodetect_dev(rdev->bdev->bd_dev);
···
 static ssize_t
 offset_store(struct md_rdev *rdev, const char *buf, size_t len)
 {
-	char *e;
-	unsigned long long offset = simple_strtoull(buf, &e, 10);
-	if (e==buf || (*e && *e != '\n'))
+	unsigned long long offset;
+	if (strict_strtoull(buf, 10, &offset) < 0)
 		return -EINVAL;
 	if (rdev->mddev->pers && rdev->raid_disk >= 0)
 		return -EBUSY;
···
 
 static struct rdev_sysfs_entry rdev_offset =
 __ATTR(offset, S_IRUGO|S_IWUSR, offset_show, offset_store);
+
+static ssize_t new_offset_show(struct md_rdev *rdev, char *page)
+{
+	return sprintf(page, "%llu\n",
+		       (unsigned long long)rdev->new_data_offset);
+}
+
+static ssize_t new_offset_store(struct md_rdev *rdev,
+				const char *buf, size_t len)
+{
+	unsigned long long new_offset;
+	struct mddev *mddev = rdev->mddev;
+
+	if (strict_strtoull(buf, 10, &new_offset) < 0)
+		return -EINVAL;
+
+	if (mddev->sync_thread)
+		return -EBUSY;
+	if (new_offset == rdev->data_offset)
+		/* reset is always permitted */
+		;
+	else if (new_offset > rdev->data_offset) {
+		/* must not push array size beyond rdev_sectors */
+		if (new_offset - rdev->data_offset
+		    + mddev->dev_sectors > rdev->sectors)
+				return -E2BIG;
+	}
+	/* Metadata worries about other space details. */
+
+	/* decreasing the offset is inconsistent with a backwards
+	 * reshape.
+	 */
+	if (new_offset < rdev->data_offset &&
+	    mddev->reshape_backwards)
+		return -EINVAL;
+	/* Increasing offset is inconsistent with forwards
+	 * reshape.  reshape_direction should be set to
+	 * 'backwards' first.
+	 */
+	if (new_offset > rdev->data_offset &&
+	    !mddev->reshape_backwards)
+		return -EINVAL;
+
+	if (mddev->pers && mddev->persistent &&
+	    !super_types[mddev->major_version]
+	    .allow_new_offset(rdev, new_offset))
+		return -E2BIG;
+	rdev->new_data_offset = new_offset;
+	if (new_offset > rdev->data_offset)
+		mddev->reshape_backwards = 1;
+	else if (new_offset < rdev->data_offset)
+		mddev->reshape_backwards = 0;
+
+	return len;
+}
+static struct rdev_sysfs_entry rdev_new_offset =
+__ATTR(new_offset, S_IRUGO|S_IWUSR, new_offset_show, new_offset_store);
 
 static ssize_t
 rdev_size_show(struct md_rdev *rdev, char *page)
···
 
 	if (strict_blocks_to_sectors(buf, &sectors) < 0)
 		return -EINVAL;
+	if (rdev->data_offset != rdev->new_data_offset)
+		return -EINVAL; /* too confusing */
 	if (my_mddev->pers && rdev->raid_disk >= 0) {
 		if (my_mddev->persistent) {
 			sectors = super_types[my_mddev->major_version].
···
 	&rdev_errors.attr,
 	&rdev_slot.attr,
 	&rdev_offset.attr,
+	&rdev_new_offset.attr,
 	&rdev_size.attr,
 	&rdev_recovery_start.attr,
 	&rdev_bad_blocks.attr,
···
 	rdev->raid_disk = -1;
 	rdev->flags = 0;
 	rdev->data_offset = 0;
+	rdev->new_data_offset = 0;
 	rdev->sb_events = 0;
 	rdev->last_read_error.tv_sec  = 0;
 	rdev->last_read_error.tv_nsec = 0;
···
 abort_free:
 	if (rdev->bdev)
 		unlock_rdev(rdev);
-	free_disk_sb(rdev);
-	kfree(rdev->badblocks.page);
+	md_rdev_clear(rdev);
 	kfree(rdev);
 	return ERR_PTR(err);
 }
···
 		mddev->new_chunk_sectors = mddev->chunk_sectors;
 		mddev->raid_disks -= mddev->delta_disks;
 		mddev->delta_disks = 0;
+		mddev->reshape_backwards = 0;
 		module_put(pers->owner);
 		printk(KERN_WARNING "md: %s: %s would not accept array\n",
 		       mdname(mddev), clevel);
···
 	mddev->layout = mddev->new_layout;
 	mddev->chunk_sectors = mddev->new_chunk_sectors;
 	mddev->delta_disks = 0;
+	mddev->reshape_backwards = 0;
 	mddev->degraded = 0;
 	if (mddev->pers->sync_request == NULL) {
 		/* this is now an array without redundancy, so
···
 		del_timer_sync(&mddev->safemode_timer);
 	}
 	pers->run(mddev);
-	mddev_resume(mddev);
 	set_bit(MD_CHANGE_DEVS, &mddev->flags);
-	set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
-	md_wakeup_thread(mddev->thread);
+	mddev_resume(mddev);
 	sysfs_notify(&mddev->kobj, NULL, "level");
 	md_new_event(mddev);
 	return rv;
···
 	if (mddev->pers)
 		rv = update_raid_disks(mddev, n);
 	else if (mddev->reshape_position != MaxSector) {
+		struct md_rdev *rdev;
 		int olddisks = mddev->raid_disks - mddev->delta_disks;
+
+		rdev_for_each(rdev, mddev) {
+			if (olddisks < n &&
+			    rdev->data_offset < rdev->new_data_offset)
+				return -EINVAL;
+			if (olddisks > n &&
+			    rdev->data_offset > rdev->new_data_offset)
+				return -EINVAL;
+		}
 		mddev->delta_disks = n - olddisks;
 		mddev->raid_disks = n;
+		mddev->reshape_backwards = (mddev->delta_disks < 0);
 	} else
 		mddev->raid_disks = n;
 	return rv ? rv : len;
···
 	if (!test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
 		return sprintf(page, "none\n");
 
-	if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
+	if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
+	    test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
 		max_sectors = mddev->resync_max_sectors;
 	else
 		max_sectors = mddev->dev_sectors;
···
 static ssize_t
 reshape_position_store(struct mddev *mddev, const char *buf, size_t len)
 {
+	struct md_rdev *rdev;
 	char *e;
 	unsigned long long new = simple_strtoull(buf, &e, 10);
 	if (mddev->pers)
···
 		return -EINVAL;
 	mddev->reshape_position = new;
 	mddev->delta_disks = 0;
+	mddev->reshape_backwards = 0;
 	mddev->new_level = mddev->level;
 	mddev->new_layout = mddev->layout;
 	mddev->new_chunk_sectors = mddev->chunk_sectors;
+	rdev_for_each(rdev, mddev)
+		rdev->new_data_offset = rdev->data_offset;
 	return len;
 }
 
 static struct md_sysfs_entry md_reshape_position =
 __ATTR(reshape_position, S_IRUGO|S_IWUSR, reshape_position_show,
        reshape_position_store);
+
+static ssize_t
+reshape_direction_show(struct mddev *mddev, char *page)
+{
+	return sprintf(page, "%s\n",
+		       mddev->reshape_backwards ? "backwards" : "forwards");
+}
+
+static ssize_t
+reshape_direction_store(struct mddev *mddev, const char *buf, size_t len)
+{
+	int backwards = 0;
+	if (cmd_match(buf, "forwards"))
+		backwards = 0;
+	else if (cmd_match(buf, "backwards"))
+		backwards = 1;
+	else
+		return -EINVAL;
+	if (mddev->reshape_backwards == backwards)
+		return len;
+
+	/* check if we are allowed to change */
+	if (mddev->delta_disks)
+		return -EBUSY;
+
+	if (mddev->persistent &&
+	    mddev->major_version == 0)
+		return -EINVAL;
+
+	mddev->reshape_backwards = backwards;
+	return len;
+}
+
+static struct md_sysfs_entry md_reshape_direction =
+__ATTR(reshape_direction, S_IRUGO|S_IWUSR, reshape_direction_show,
+       reshape_direction_store);
 
 static ssize_t
 array_size_show(struct mddev *mddev, char *page)
···
 	&md_safe_delay.attr,
 	&md_array_state.attr,
 	&md_reshape_position.attr,
+	&md_reshape_direction.attr,
 	&md_array_size.attr,
 	&max_corr_read_errors.attr,
 	NULL,
···
 		err = -EINVAL;
 		mddev->pers->stop(mddev);
 	}
-	if (err == 0 && mddev->pers->sync_request) {
+	if (err == 0 && mddev->pers->sync_request &&
+	    (mddev->bitmap_info.file || mddev->bitmap_info.offset)) {
 		err = bitmap_create(mddev);
 		if (err) {
 			printk(KERN_ERR "%s: failed to create bitmap (%d)\n",
···
 	mddev->events = 0;
 	mddev->can_decrease_events = 0;
 	mddev->delta_disks = 0;
+	mddev->reshape_backwards = 0;
 	mddev->new_level = LEVEL_NONE;
 	mddev->new_layout = 0;
 	mddev->new_chunk_sectors = 0;
···
 	mddev->merge_check_needed = 0;
 	mddev->bitmap_info.offset = 0;
 	mddev->bitmap_info.default_offset = 0;
+	mddev->bitmap_info.default_space = 0;
 	mddev->bitmap_info.chunksize = 0;
 	mddev->bitmap_info.daemon_sleep = 0;
 	mddev->bitmap_info.max_write_behind = 0;
···
 		goto out;
 
 	/* bitmap disabled, zero the first byte and copy out */
-	if (!mddev->bitmap || !mddev->bitmap->file) {
+	if (!mddev->bitmap || !mddev->bitmap->storage.file) {
 		file->pathname[0] = '\0';
 		goto copy_out;
 	}
···
 	if (!buf)
 		goto out;
 
-	ptr = d_path(&mddev->bitmap->file->f_path, buf, sizeof(file->pathname));
+	ptr = d_path(&mddev->bitmap->storage.file->f_path,
+		     buf, sizeof(file->pathname));
 	if (IS_ERR(ptr))
 		goto out;
 
···
 	set_bit(MD_CHANGE_DEVS, &mddev->flags);
 
 	mddev->bitmap_info.default_offset = MD_SB_BYTES >> 9;
+	mddev->bitmap_info.default_space = 64*2 - (MD_SB_BYTES >> 9);
 	mddev->bitmap_info.offset = 0;
 
 	mddev->reshape_position = MaxSector;
···
 	mddev->new_chunk_sectors = mddev->chunk_sectors;
 	mddev->new_layout = mddev->layout;
 	mddev->delta_disks = 0;
+	mddev->reshape_backwards = 0;
 
 	return 0;
 }
···
 	 */
 	if (mddev->sync_thread)
 		return -EBUSY;
-	if (mddev->bitmap)
-		/* Sorry, cannot grow a bitmap yet, just remove it,
-		 * grow, and re-add.
-		 */
-		return -EBUSY;
+
 	rdev_for_each(rdev, mddev) {
 		sector_t avail = rdev->sectors;
 
···
 static int update_raid_disks(struct mddev *mddev, int raid_disks)
 {
 	int rv;
+	struct md_rdev *rdev;
 	/* change the number of raid disks */
 	if (mddev->pers->check_reshape == NULL)
 		return -EINVAL;
···
 		return -EINVAL;
 	if (mddev->sync_thread || mddev->reshape_position != MaxSector)
 		return -EBUSY;
+
+	rdev_for_each(rdev, mddev) {
+		if (mddev->raid_disks < raid_disks &&
+		    rdev->data_offset < rdev->new_data_offset)
+			return -EINVAL;
+		if (mddev->raid_disks > raid_disks &&
+		    rdev->data_offset > rdev->new_data_offset)
+			return -EINVAL;
+	}
+
 	mddev->delta_disks = raid_disks - mddev->raid_disks;
+	if (mddev->delta_disks < 0)
+		mddev->reshape_backwards = 1;
+	else if (mddev->delta_disks > 0)
+		mddev->reshape_backwards = 0;
 
 	rv = mddev->pers->check_reshape(mddev);
-	if (rv < 0)
+	if (rv < 0) {
 		mddev->delta_disks = 0;
+		mddev->reshape_backwards = 0;
+	}
 	return rv;
 }
 
···
 				return -EINVAL;
 			mddev->bitmap_info.offset =
 				mddev->bitmap_info.default_offset;
+			mddev->bitmap_info.space =
+				mddev->bitmap_info.default_space;
 			mddev->pers->quiesce(mddev, 1);
 			rv = bitmap_create(mddev);
 			if (!rv)
···
 			/* remove the bitmap */
 			if (!mddev->bitmap)
 				return -ENOENT;
-			if (mddev->bitmap->file)
+			if (mddev->bitmap->storage.file)
 				return -EINVAL;
 			mddev->pers->quiesce(mddev, 1);
 			bitmap_destroy(mddev);
···
 	struct mddev *mddev = mddev_find(bdev->bd_dev);
 	int err;
 
+	if (!mddev)
+		return -ENODEV;
+
 	if (mddev->gendisk != bdev->bd_disk) {
 		/* we are racing with mddev_put which is discarding this
 		 * bd_disk.
···
 
 	resync = mddev->curr_resync - atomic_read(&mddev->recovery_active);
 
-	if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
+	if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
+	    test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
 		max_sectors = mddev->resync_max_sectors;
 	else
 		max_sectors = mddev->dev_sectors;
···
 			j = mddev->recovery_cp;
 
 	} else if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
-		max_sectors = mddev->dev_sectors;
+		max_sectors = mddev->resync_max_sectors;
 	else {
 		/* recovery follows the physical size of devices */
 		max_sectors = mddev->dev_sectors;
···
 			goto unlock;
 
 		if (mddev->pers->sync_request) {
-			if (spares && mddev->bitmap && ! mddev->bitmap->file) {
+			if (spares) {
 				/* We are adding a device or devices to an array
 				 * which has the bitmap stored on all devices.
 				 * So make sure all bitmap pages get written
···
 }
 EXPORT_SYMBOL(md_wait_for_blocked_rdev);
 
+void md_finish_reshape(struct mddev *mddev)
+{
+	/* called be personality module when reshape completes. */
+	struct md_rdev *rdev;
+
+	rdev_for_each(rdev, mddev) {
+		if (rdev->data_offset > rdev->new_data_offset)
+			rdev->sectors += rdev->data_offset - rdev->new_data_offset;
+		else
+			rdev->sectors -= rdev->new_data_offset - rdev->data_offset;
+		rdev->data_offset = rdev->new_data_offset;
+	}
+}
+EXPORT_SYMBOL(md_finish_reshape);
 
 /* Bad block management.
  * We can record which blocks on each device are 'bad' and so just
···
 }
 
 int rdev_set_badblocks(struct md_rdev *rdev, sector_t s, int sectors,
-		       int acknowledged)
+		       int is_new)
 {
-	int rv = md_set_badblocks(&rdev->badblocks,
-				  s + rdev->data_offset, sectors, acknowledged);
+	int rv;
+	if (is_new)
+		s += rdev->new_data_offset;
+	else
+		s += rdev->data_offset;
+	rv = md_set_badblocks(&rdev->badblocks,
+			      s, sectors, 0);
 	if (rv) {
 		/* Make sure they get written out promptly */
 		sysfs_notify_dirent_safe(rdev->sysfs_state);
···
 	return rv;
 }
 
-int rdev_clear_badblocks(struct md_rdev *rdev, sector_t s, int sectors)
+int rdev_clear_badblocks(struct md_rdev *rdev, sector_t s, int sectors,
+			 int is_new)
 {
+	if (is_new)
+		s += rdev->new_data_offset;
+	else
+		s += rdev->data_offset;
 	return md_clear_badblocks(&rdev->badblocks,
-				  s + rdev->data_offset,
-				  sectors);
+				  s, sectors);
 }
 EXPORT_SYMBOL_GPL(rdev_clear_badblocks);
 

drivers/md/md.h

···
 	int		sb_loaded;
 	__u64		sb_events;
 	sector_t	data_offset;	/* start of data in array */
+	sector_t	new_data_offset;/* only relevant while reshaping */
 	sector_t 	sb_start;	/* offset of the super block (in 512byte sectors) */
 	int		sb_size;	/* bytes in the superblock */
 	int		preferred_minor;	/* autorun support */
···
 	return 0;
 }
 extern int rdev_set_badblocks(struct md_rdev *rdev, sector_t s, int sectors,
-			      int acknowledged);
-extern int rdev_clear_badblocks(struct md_rdev *rdev, sector_t s, int sectors);
+			      int is_new);
+extern int rdev_clear_badblocks(struct md_rdev *rdev, sector_t s, int sectors,
+				int is_new);
 extern void md_ack_all_badblocks(struct badblocks *bb);
 
 struct mddev {
···
 	sector_t			reshape_position;
 	int				delta_disks, new_level, new_layout;
 	int				new_chunk_sectors;
+	int				reshape_backwards;
 
 	atomic_t			plug_cnt;	/* If device is expecting
 							 * more bios soon.
···
 						 * For external metadata, offset
 						 * from start of device. 
 						 */
+		unsigned long		space; /* space available at this offset */
 		loff_t			default_offset; /* this is the offset to use when
 							 * hot-adding a bitmap.  It should
 							 * eventually be settable by sysfs.
 							 */
+		unsigned long		default_space; /* space available at
+							* default offset */
 		struct mutex		mutex;
 		unsigned long		chunksize;
 		unsigned long		daemon_sleep; /* how many jiffies between updates? */
···
 extern void md_write_end(struct mddev *mddev);
 extern void md_done_sync(struct mddev *mddev, int blocks, int ok);
 extern void md_error(struct mddev *mddev, struct md_rdev *rdev);
+extern void md_finish_reshape(struct mddev *mddev);
 
 extern int mddev_congested(struct mddev *mddev, int bits);
 extern void md_flush_request(struct mddev *mddev, struct bio *bio);
···
 extern void md_stop(struct mddev *mddev);
 extern void md_stop_writes(struct mddev *mddev);
 extern int md_rdev_init(struct md_rdev *rdev);
+extern void md_rdev_clear(struct md_rdev *rdev);
 
 extern void mddev_suspend(struct mddev *mddev);
 extern void mddev_resume(struct mddev *mddev);

drivers/md/raid1.c

···
 
 			rdev = conf->mirrors[d].rdev;
 			if (rdev &&
-			    test_bit(In_sync, &rdev->flags) &&
+			    (test_bit(In_sync, &rdev->flags) ||
+			     (!test_bit(Faulty, &rdev->flags) &&
+			      rdev->recovery_offset >= sect + s)) &&
 			    is_badblock(rdev, sect, s,
 					&first_bad, &bad_sectors) == 0 &&
 			    sync_page_io(rdev, sect, s<<9,
···
 			continue;
 		if (test_bit(BIO_UPTODATE, &bio->bi_flags) &&
 		    test_bit(R1BIO_MadeGood, &r1_bio->state)) {
-			rdev_clear_badblocks(rdev, r1_bio->sector, s);
+			rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
 		}
 		if (!test_bit(BIO_UPTODATE, &bio->bi_flags) &&
 		    test_bit(R1BIO_WriteError, &r1_bio->state)) {
···
 			struct md_rdev *rdev = conf->mirrors[m].rdev;
 			rdev_clear_badblocks(rdev,
 					     r1_bio->sector,
-					     r1_bio->sectors);
+					     r1_bio->sectors, 0);
 			rdev_dec_pending(rdev, conf->mddev);
 		} else if (r1_bio->bios[m] != NULL) {
 			/* This drive got a write error.  We need to
···
 		if (!disk->rdev ||
 		    !test_bit(In_sync, &disk->rdev->flags)) {
 			disk->head_position = 0;
-			if (disk->rdev)
+			if (disk->rdev &&
+			    (disk->rdev->saved_raid_disk < 0))
 				conf->fullsync = 1;
 		} else if (conf->last_used < 0)
 			/*
···
 	 * any io in the removed space completes, but it hardly seems
 	 * worth it.
 	 */
-	md_set_array_sectors(mddev, raid1_size(mddev, sectors, 0));
-	if (mddev->array_sectors > raid1_size(mddev, sectors, 0))
+	sector_t newsize = raid1_size(mddev, sectors, 0);
+	if (mddev->external_size &&
+	    mddev->array_sectors > newsize)
 		return -EINVAL;
+	if (mddev->bitmap) {
+		int ret = bitmap_resize(mddev->bitmap, newsize, 0, 0);
+		if (ret)
+			return ret;
+	}
+	md_set_array_sectors(mddev, newsize);
 	set_capacity(mddev->gendisk, mddev->array_sectors);
 	revalidate_disk(mddev->gendisk);
 	if (sectors > mddev->dev_sectors &&

drivers/md/raid10.c

···
 #include <linux/module.h>
 #include <linux/seq_file.h>
 #include <linux/ratelimit.h>
+#include <linux/kthread.h>
 #include "md.h"
 #include "raid10.h"
 #include "raid0.h"
···
 static void allow_barrier(struct r10conf *conf);
 static void lower_barrier(struct r10conf *conf);
 static int enough(struct r10conf *conf, int ignore);
+static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
+				int *skipped);
+static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio);
+static void end_reshape_write(struct bio *bio, int error);
+static void end_reshape(struct r10conf *conf);
 
 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
 {
···
 	if (!r10_bio)
 		return NULL;
 
-	if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
+	if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
+	    test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
 		nalloc = conf->copies; /* resync */
 	else
 		nalloc = 2; /* recovery */
···
 		struct bio *rbio = r10_bio->devs[j].repl_bio;
 		bio = r10_bio->devs[j].bio;
 		for (i = 0; i < RESYNC_PAGES; i++) {
-			if (j == 1 && !test_bit(MD_RECOVERY_SYNC,
-						&conf->mddev->recovery)) {
-				/* we can share bv_page's during recovery */
+			if (j > 0 && !test_bit(MD_RECOVERY_SYNC,
+					       &conf->mddev->recovery)) {
+				/* we can share bv_page's during recovery
+				 * and reshape */
 				struct bio *rbio = r10_bio->devs[0].bio;
 				page = rbio->bi_io_vec[i].bv_page;
 				get_page(page);
···
 	while (j--)
 		for (i = 0; i < RESYNC_PAGES ; i++)
 			safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
-	j = -1;
+	j = 0;
 out_free_bio:
-	while (++j < nalloc) {
-		bio_put(r10_bio->devs[j].bio);
+	for ( ; j < nalloc; j++) {
+		if (r10_bio->devs[j].bio)
+			bio_put(r10_bio->devs[j].bio);
 		if (r10_bio->devs[j].repl_bio)
 			bio_put(r10_bio->devs[j].repl_bio);
 	}
···
  * sector offset to a virtual address
  */
 
-static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
+static void __raid10_find_phys(struct geom *geo, struct r10bio *r10bio)
 {
 	int n,f;
 	sector_t sector;
 	sector_t chunk;
 	sector_t stripe;
 	int dev;
-
 	int slot = 0;
 
 	/* now calculate first sector/dev */
-	chunk = r10bio->sector >> conf->chunk_shift;
-	sector = r10bio->sector & conf->chunk_mask;
+	chunk = r10bio->sector >> geo->chunk_shift;
+	sector = r10bio->sector & geo->chunk_mask;
 
-	chunk *= conf->near_copies;
+	chunk *= geo->near_copies;
 	stripe = chunk;
-	dev = sector_div(stripe, conf->raid_disks);
-	if (conf->far_offset)
-		stripe *= conf->far_copies;
+	dev = sector_div(stripe, geo->raid_disks);
+	if (geo->far_offset)
+		stripe *= geo->far_copies;
 
-	sector += stripe << conf->chunk_shift;
+	sector += stripe << geo->chunk_shift;
 
 	/* and calculate all the others */
-	for (n=0; n < conf->near_copies; n++) {
+	for (n = 0; n < geo->near_copies; n++) {
 		int d = dev;
 		sector_t s = sector;
 		r10bio->devs[slot].addr = sector;
 		r10bio->devs[slot].devnum = d;
 		slot++;
 
-		for (f = 1; f < conf->far_copies; f++) {
-			d += conf->near_copies;
-			if (d >= conf->raid_disks)
-				d -= conf->raid_disks;
-			s += conf->stride;
+		for (f = 1; f < geo->far_copies; f++) {
+			d += geo->near_copies;
+			if (d >= geo->raid_disks)
+				d -= geo->raid_disks;
+			s += geo->stride;
 			r10bio->devs[slot].devnum = d;
 			r10bio->devs[slot].addr = s;
 			slot++;
 		}
 		dev++;
-		if (dev >= conf->raid_disks) {
+		if (dev >= geo->raid_disks) {
 			dev = 0;
-			sector += (conf->chunk_mask + 1);
+			sector += (geo->chunk_mask + 1);
 		}
 	}
-	BUG_ON(slot != conf->copies);
+}
+
+static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
+{
+	struct geom *geo = &conf->geo;
+
+	if (conf->reshape_progress != MaxSector &&
+	    ((r10bio->sector >= conf->reshape_progress) !=
+	     conf->mddev->reshape_backwards)) {
+		set_bit(R10BIO_Previous, &r10bio->state);
+		geo = &conf->prev;
+	} else
+		clear_bit(R10BIO_Previous, &r10bio->state);
+
+	__raid10_find_phys(geo, r10bio);
 }
 
 static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
 {
 	sector_t offset, chunk, vchunk;
+	/* Never use conf->prev as this is only called during resync
+	 * or recovery, so reshape isn't happening
+	 */
+	struct geom *geo = &conf->geo;
 
-	offset = sector & conf->chunk_mask;
-	if (conf->far_offset) {
+	offset = sector & geo->chunk_mask;
+	if (geo->far_offset) {
 		int fc;
-		chunk = sector >> conf->chunk_shift;
-		fc = sector_div(chunk, conf->far_copies);
-		dev -= fc * conf->near_copies;
+		chunk = sector >> geo->chunk_shift;
+		fc = sector_div(chunk, geo->far_copies);
+		dev -= fc * geo->near_copies;
 		if (dev < 0)
-			dev += conf->raid_disks;
+			dev += geo->raid_disks;
 	} else {
-		while (sector >= conf->stride) {
-			sector -= conf->stride;
-			if (dev < conf->near_copies)
-				dev += conf->raid_disks - conf->near_copies;
+		while (sector >= geo->stride) {
+			sector -= geo->stride;
+			if (dev < geo->near_copies)
+				dev += geo->raid_disks - geo->near_copies;
 			else
-				dev -= conf->near_copies;
+				dev -= geo->near_copies;
 		}
-		chunk = sector >> conf->chunk_shift;
+		chunk = sector >> geo->chunk_shift;
 	}
-	vchunk = chunk * conf->raid_disks + dev;
-	sector_div(vchunk, conf->near_copies);
-	return (vchunk << conf->chunk_shift) + offset;
+	vchunk = chunk * geo->raid_disks + dev;
+	sector_div(vchunk, geo->near_copies);
+	return (vchunk << geo->chunk_shift) + offset;
 }
 
 /**
···
 	struct r10conf *conf = mddev->private;
 	sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
 	int max;
-	unsigned int chunk_sectors = mddev->chunk_sectors;
+	unsigned int chunk_sectors;
 	unsigned int bio_sectors = bvm->bi_size >> 9;
+	struct geom *geo = &conf->geo;
 
-	if (conf->near_copies < conf->raid_disks) {
+	chunk_sectors = (conf->geo.chunk_mask & conf->prev.chunk_mask) + 1;
+	if (conf->reshape_progress != MaxSector &&
+	    ((sector >= conf->reshape_progress) !=
+	     conf->mddev->reshape_backwards))
+		geo = &conf->prev;
+
+	if (geo->near_copies < geo->raid_disks) {
 		max = (chunk_sectors - ((sector & (chunk_sectors - 1))
 					+ bio_sectors)) << 9;
 		if (max < 0)
···
 	if (mddev->merge_check_needed) {
 		struct r10bio r10_bio;
 		int s;
+		if (conf->reshape_progress != MaxSector) {
+			/* Cannot give any guidance during reshape */
+			if (max <= biovec->bv_len && bio_sectors == 0)
+				return biovec->bv_len;
+			return 0;
+		}
 		r10_bio.sector = sector;
 		raid10_find_phys(conf, &r10_bio);
 		rcu_read_lock();
···
 	struct md_rdev *rdev, *best_rdev;
 	int do_balance;
 	int best_slot;
+	struct geom *geo = &conf->geo;
 
 	raid10_find_phys(conf, r10_bio);
 	rcu_read_lock();
···
 		 * sequential read speed for 'far copies' arrays.  So only
 		 * keep it for 'near' arrays, and review those later.
 		 */
-		if (conf->near_copies > 1 && !atomic_read(&rdev->nr_pending))
+		if (geo->near_copies > 1 && !atomic_read(&rdev->nr_pending))
 			break;
 
 		/* for far > 1 always use the lowest address */
-		if (conf->far_copies > 1)
+		if (geo->far_copies > 1)
 			new_distance = r10_bio->devs[slot].addr;
 		else
 			new_distance = abs(r10_bio->devs[slot].addr -
···
 	if (mddev_congested(mddev, bits))
 		return 1;
 	rcu_read_lock();
-	for (i = 0; i < conf->raid_disks && ret == 0; i++) {
+	for (i = 0;
+	     (i < conf->geo.raid_disks || i < conf->prev.raid_disks)
+		     && ret == 0;
+	     i++) {
 		struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
 		if (rdev && !test_bit(Faulty, &rdev->flags)) {
 			struct request_queue *q = bdev_get_queue(rdev->bdev);
···
 	spin_unlock_irq(&conf->resync_lock);
 }
 
+static sector_t choose_data_offset(struct r10bio *r10_bio,
+				   struct md_rdev *rdev)
+{
+	if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
+	    test_bit(R10BIO_Previous, &r10_bio->state))
+		return rdev->data_offset;
+	else
+		return rdev->new_data_offset;
+}
+
 static void make_request(struct mddev *mddev, struct bio * bio)
 {
 	struct r10conf *conf = mddev->private;
 	struct r10bio *r10_bio;
 	struct bio *read_bio;
 	int i;
-	int chunk_sects = conf->chunk_mask + 1;
+	sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
+	int chunk_sects = chunk_mask + 1;
 	const int rw = bio_data_dir(bio);
 	const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
 	const unsigned long do_fua = (bio->bi_rw & REQ_FUA);
···
 	int plugged;
 	int sectors_handled;
 	int max_sectors;
+	int sectors;
 
 	if (unlikely(bio->bi_rw & REQ_FLUSH)) {
 		md_flush_request(mddev, bio);
···
 	/* If this request crosses a chunk boundary, we need to
 	 * split it.  This will only happen for 1 PAGE (or less) requests.
 	 */
-	if (unlikely( (bio->bi_sector & conf->chunk_mask) + (bio->bi_size >> 9)
-		      > chunk_sects &&
-		    conf->near_copies < conf->raid_disks)) {
+	if (unlikely((bio->bi_sector & chunk_mask) + (bio->bi_size >> 9)
+		     > chunk_sects
+		     && (conf->geo.near_copies < conf->geo.raid_disks
+			 || conf->prev.near_copies < conf->prev.raid_disks))) {
 		struct bio_pair *bp;
 		/* Sanity check -- queue functions should prevent this happening */
 		if (bio->bi_vcnt != 1 ||
···
 	 */
 	wait_barrier(conf);
 
+	sectors = bio->bi_size >> 9;
+	while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
+	    bio->bi_sector < conf->reshape_progress &&
+	    bio->bi_sector + sectors > conf->reshape_progress) {
+		/* IO spans the reshape position.  Need to wait for
+		 * reshape to pass
+		 */
+		allow_barrier(conf);
+		wait_event(conf->wait_barrier,
+			   conf->reshape_progress <= bio->bi_sector ||
+			   conf->reshape_progress >= bio->bi_sector + sectors);
+		wait_barrier(conf);
+	}
+	if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
+	    bio_data_dir(bio) == WRITE &&
+	    (mddev->reshape_backwards
+	     ? (bio->bi_sector < conf->reshape_safe &&
+		bio->bi_sector + sectors > conf->reshape_progress)
+	     : (bio->bi_sector + sectors > conf->reshape_safe &&
+		bio->bi_sector < conf->reshape_progress))) {
+		/* Need to update reshape_position in metadata */
+		mddev->reshape_position = conf->reshape_progress;
+		set_bit(MD_CHANGE_DEVS, &mddev->flags);
+		set_bit(MD_CHANGE_PENDING, &mddev->flags);
+		md_wakeup_thread(mddev->thread);
+		wait_event(mddev->sb_wait,
+			   !test_bit(MD_CHANGE_PENDING, &mddev->flags));
+
+		conf->reshape_safe = mddev->reshape_position;
+	}
+
 	r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
 
 	r10_bio->master_bio = bio;
-	r10_bio->sectors = bio->bi_size >> 9;
+	r10_bio->sectors = sectors;
 
 	r10_bio->mddev = mddev;
 	r10_bio->sector = bio->bi_sector;
···
 		r10_bio->devs[slot].rdev = rdev;
 
 		read_bio->bi_sector = r10_bio->devs[slot].addr +
-			rdev->data_offset;
+			choose_data_offset(r10_bio, rdev);
 		read_bio->bi_bdev = rdev->bdev;
 		read_bio->bi_end_io = raid10_end_read_request;
 		read_bio->bi_rw = READ | do_sync;
···
 		r10_bio->devs[i].bio = mbio;
 
 		mbio->bi_sector	= (r10_bio->devs[i].addr+
-				   conf->mirrors[d].rdev->data_offset);
+				   choose_data_offset(r10_bio,
+						      conf->mirrors[d].rdev));
 		mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
 		mbio->bi_end_io	= raid10_end_write_request;
 		mbio->bi_rw = WRITE | do_sync | do_fua;
···
 		 * so it cannot disappear, so the replacement cannot
 		 * become NULL here
 		 */
-		mbio->bi_sector	= (r10_bio->devs[i].addr+
-				   conf->mirrors[d].replacement->data_offset);
+		mbio->bi_sector	= (r10_bio->devs[i].addr +
+				   choose_data_offset(
+					   r10_bio,
+					   conf->mirrors[d].replacement));
 		mbio->bi_bdev = conf->mirrors[d].replacement->bdev;
 		mbio->bi_end_io	= raid10_end_write_request;
 		mbio->bi_rw = WRITE | do_sync | do_fua;
···
 	struct r10conf *conf = mddev->private;
 	int i;
 
-	if (conf->near_copies < conf->raid_disks)
+	if (conf->geo.near_copies < conf->geo.raid_disks)
 		seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
-	if (conf->near_copies > 1)
-		seq_printf(seq, " %d near-copies", conf->near_copies);
-	if (conf->far_copies > 1) {
-		if (conf->far_offset)
-			seq_printf(seq, " %d offset-copies", conf->far_copies);
+	if (conf->geo.near_copies > 1)
+		seq_printf(seq, " %d near-copies", conf->geo.near_copies);
+	if (conf->geo.far_copies > 1) {
+		if (conf->geo.far_offset)
+			seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
 		else
-			seq_printf(seq, " %d far-copies", conf->far_copies);
+			seq_printf(seq, " %d far-copies", conf->geo.far_copies);
 	}
-	seq_printf(seq, " [%d/%d] [", conf->raid_disks,
-					conf->raid_disks - mddev->degraded);
-	for (i = 0; i < conf->raid_disks; i++)
+	seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
+					conf->geo.raid_disks - mddev->degraded);
+	for (i = 0; i < conf->geo.raid_disks; i++)
 		seq_printf(seq, "%s",
 			      conf->mirrors[i].rdev &&
 			      test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
···
  * Don't consider the device numbered 'ignore'
  * as we might be about to remove it.
  */
-static int enough(struct r10conf *conf, int ignore)
+static int _enough(struct r10conf *conf, struct geom *geo, int ignore)
 {
 	int first = 0;
 
···
 			if (conf->mirrors[first].rdev &&
 			    first != ignore)
 				cnt++;
-			first = (first+1) % conf->raid_disks;
+			first = (first+1) % geo->raid_disks;
 		}
 		if (cnt == 0)
 			return 0;
 	} while (first != 0);
 	return 1;
+}
+
+static int enough(struct r10conf *conf, int ignore)
+{
+	return _enough(conf, &conf->geo, ignore) &&
+		_enough(conf, &conf->prev, ignore);
 }
 
 static void error(struct mddev *mddev, struct md_rdev *rdev)
···
 	       "md/raid10:%s: Disk failure on %s, disabling device.\n"
 	       "md/raid10:%s: Operation continuing on %d devices.\n",
 	       mdname(mddev), bdevname(rdev->bdev, b),
-	       mdname(mddev), conf->raid_disks - mddev->degraded);
+	       mdname(mddev), conf->geo.raid_disks - mddev->degraded);
 }
 
 static void print_conf(struct r10conf *conf)
···
 		printk(KERN_DEBUG "(!conf)\n");
 		return;
 	}
-	printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
-		conf->raid_disks);
+	printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
+		conf->geo.raid_disks);
 
-	for (i = 0; i < conf->raid_disks; i++) {
+	for (i = 0; i < conf->geo.raid_disks; i++) {
 		char b[BDEVNAME_SIZE];
 		tmp = conf->mirrors + i;
 		if (tmp->rdev)
···
 	 * Find all non-in_sync disks within the RAID10 configuration
 	 * and mark them in_sync
 	 */
-	for (i = 0; i < conf->raid_disks; i++) {
+	for (i = 0; i < conf->geo.raid_disks; i++) {
 		tmp = conf->mirrors + i;
 		if (tmp->replacement
 		    && tmp->replacement->recovery_offset == MaxSector
···
 	int err = -EEXIST;
 	int mirror;
 	int first = 0;
-	int last = conf->raid_disks - 1;
+	int last = conf->geo.raid_disks - 1;
 	struct request_queue *q = bdev_get_queue(rdev->bdev);
 
 	if (mddev->recovery_cp < MaxSector)
···
 		 * very different from resync
 		 */
 		return -EBUSY;
-	if (rdev->saved_raid_disk < 0 && !enough(conf, -1))
+	if (rdev->saved_raid_disk < 0 && !_enough(conf, &conf->prev, -1))
 		return -EINVAL;
 
 	if (rdev->raid_disk >= 0)
···
 	if (!test_bit(Faulty, &rdev->flags) &&
 	    mddev->recovery_disabled != p->recovery_disabled &&
 	    (!p->replacement || p->replacement == rdev) &&
+	    number < conf->geo.raid_disks &&
 	    enough(conf, -1)) {
 		err = -EBUSY;
 		goto abort;
···
 	struct r10conf *conf = r10_bio->mddev->private;
 	int d;
 
-	d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
+	if (bio == r10_bio->master_bio) {
+		/* this is a reshape read */
+		d = r10_bio->read_slot; /* really the read dev */
+	} else
+		d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
 
 	if (test_bit(BIO_UPTODATE, &bio->bi_flags))
 		set_bit(R10BIO_Uptodate, &r10_bio->state);
···
 				       " (%d sectors at %llu on %s)\n",
 				       mdname(mddev), s,
 				       (unsigned long long)(
-					       sect + rdev->data_offset),
+					       sect +
+					       choose_data_offset(r10_bio,
+								  rdev)),
 				       bdevname(rdev->bdev, b));
 				printk(KERN_NOTICE "md/raid10:%s: %s: failing "
 				       "drive\n",
···
 				       " (%d sectors at %llu on %s)\n",
 				       mdname(mddev), s,
 				       (unsigned long long)(
-					       sect + rdev->data_offset),
+					       sect +
+					       choose_data_offset(r10_bio, rdev)),
 				       bdevname(rdev->bdev, b));
 				printk(KERN_NOTICE "md/raid10:%s: %s: failing "
 				       "drive\n",
···
 				       " (%d sectors at %llu on %s)\n",
 				       mdname(mddev), s,
 				       (unsigned long long)(
-					       sect + rdev->data_offset),
+					       sect +
+					       choose_data_offset(r10_bio, rdev)),
 				       bdevname(rdev->bdev, b));
 				atomic_add(s, &rdev->corrected_errors);
 			}
···
 		wbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
 		md_trim_bio(wbio, sector - bio->bi_sector, sectors);
 		wbio->bi_sector = (r10_bio->devs[i].addr+
-				   rdev->data_offset+
+				   choose_data_offset(r10_bio, rdev) +
 				   (sector - r10_bio->sector));
 		wbio->bi_bdev = rdev->bdev;
 		if (submit_bio_wait(WRITE, wbio) == 0)
···
 	r10_bio->devs[slot].bio = bio;
 	r10_bio->devs[slot].rdev = rdev;
 	bio->bi_sector = r10_bio->devs[slot].addr
-		+ rdev->data_offset;
+		+ choose_data_offset(r10_bio, rdev);
 	bio->bi_bdev = rdev->bdev;
 	bio->bi_rw = READ | do_sync;
 	bio->bi_private = r10_bio;
···
 				rdev_clear_badblocks(
 					rdev,
 					r10_bio->devs[m].addr,
-					r10_bio->sectors);
+					r10_bio->sectors, 0);
 			} else {
 				if (!rdev_set_badblocks(
 					    rdev,
···
 				rdev_clear_badblocks(
 					rdev,
 					r10_bio->devs[m].addr,
-					r10_bio->sectors);
+					r10_bio->sectors, 0);
 			} else {
 				if (!rdev_set_badblocks(
 					    rdev,
···
 				rdev_clear_badblocks(
 					rdev,
 					r10_bio->devs[m].addr,
-					r10_bio->sectors);
+					r10_bio->sectors, 0);
 				rdev_dec_pending(rdev, conf->mddev);
 			} else if (bio != NULL &&
 				   !test_bit(BIO_UPTODATE, &bio->bi_flags)) {
···
 				rdev_clear_badblocks(
 					rdev,
 					r10_bio->devs[m].addr,
-					r10_bio->sectors);
+					r10_bio->sectors, 0);
 				rdev_dec_pending(rdev, conf->mddev);
 			}
 		}
···
 		if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
 		    test_bit(R10BIO_WriteError, &r10_bio->state))
 			handle_write_completed(conf, r10_bio);
+		else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
+			reshape_request_write(mddev, r10_bio);
 		else if (test_bit(R10BIO_IsSync, &r10_bio->state))
 			sync_request_write(mddev, r10_bio);
 		else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
···
 	buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
 	BUG_ON(conf->r10buf_pool);
 	conf->have_replacement = 0;
-	for (i = 0; i < conf->raid_disks; i++)
+	for (i = 0; i < conf->geo.raid_disks; i++)
 		if (conf->mirrors[i].replacement)
 			conf->have_replacement = 1;
 	conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
···
 	sector_t sync_blocks;
 	sector_t sectors_skipped = 0;
 	int chunks_skipped = 0;
+	sector_t chunk_mask = conf->geo.chunk_mask;
 
 	if (!conf->r10buf_pool)
 		if (init_resync(conf))
···
 
  skipped:
 	max_sector = mddev->dev_sectors;
-	if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
+	if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
+	    test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
 		max_sector = mddev->resync_max_sectors;
 	if (sector_nr >= max_sector) {
 		/* If we aborted, we need to abort the
···
 		 * we need to convert that to several
 		 * virtual addresses.
 		 */
+		if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
+			end_reshape(conf);
+			return 0;
+		}
+
 		if (mddev->curr_resync < max_sector) { /* aborted */
 			if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
 				bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
 						&sync_blocks, 1);
-			else for (i=0; i<conf->raid_disks; i++) {
+			else for (i = 0; i < conf->geo.raid_disks; i++) {
 				sector_t sect =
 					raid10_find_virt(conf, mddev->curr_resync, i);
 				bitmap_end_sync(mddev->bitmap, sect,
···
 				/* Completed a full sync so the replacements
 				 * are now fully recovered.
 				 */
-				for (i = 0; i < conf->raid_disks; i++)
+				for (i = 0; i < conf->geo.raid_disks; i++)
 					if (conf->mirrors[i].replacement)
 						conf->mirrors[i].replacement
 							->recovery_offset
···
 		*skipped = 1;
 		return sectors_skipped;
 	}
-	if (chunks_skipped >= conf->raid_disks) {
+
+	if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
+		return reshape_request(mddev, sector_nr, skipped);
+
+	if (chunks_skipped >= conf->geo.raid_disks) {
 		/* if there has been nothing to do on any drive,
 		 * then there is nothing to do at all..
 		 */
···
 	/* make sure whole request will fit in a chunk - if chunks
 	 * are meaningful
 	 */
-	if (conf->near_copies < conf->raid_disks &&
-	    max_sector > (sector_nr | conf->chunk_mask))
-		max_sector = (sector_nr | conf->chunk_mask) + 1;
+	if (conf->geo.near_copies < conf->geo.raid_disks &&
+	    max_sector > (sector_nr | chunk_mask))
+		max_sector = (sector_nr | chunk_mask) + 1;
 	/*
 	 * If there is non-resync activity waiting for us then
 	 * put in a delay to throttle resync.
···
 		int j;
 		r10_bio = NULL;
 
-		for (i=0 ; i<conf->raid_disks; i++) {
+		for (i = 0 ; i < conf->geo.raid_disks; i++) {
 			int still_degraded;
 			struct r10bio *rb2;
 			sector_t sect;
···
 			/* Need to check if the array will still be
 			 * degraded
 			 */
-			for (j=0; j<conf->raid_disks; j++)
+			for (j = 0; j < conf->geo.raid_disks; j++)
 				if (conf->mirrors[j].rdev == NULL ||
 				    test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
 					still_degraded = 1;
···
 		r10_bio->sector = sector_nr;
 		set_bit(R10BIO_IsSync, &r10_bio->state);
 		raid10_find_phys(conf, r10_bio);
-		r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1;
+		r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
 
-		for (i=0; i<conf->copies; i++) {
+		for (i = 0; i < conf->copies; i++) {
 			int d = r10_bio->devs[i].devnum;
 			sector_t first_bad, sector;
 			int bad_sectors;
···
 	struct r10conf *conf = mddev->private;
 
 	if (!raid_disks)
-		raid_disks = conf->raid_disks;
+		raid_disks = min(conf->geo.raid_disks,
+				 conf->prev.raid_disks);
 	if (!sectors)
 		sectors = conf->dev_sectors;
 
-	size = sectors >> conf->chunk_shift;
-	sector_div(size, conf->far_copies);
+	size = sectors >> conf->geo.chunk_shift;
+	sector_div(size, conf->geo.far_copies);
 	size = size * raid_disks;
-	sector_div(size, conf->near_copies);
+	sector_div(size, conf->geo.near_copies);
 
-	return size << conf->chunk_shift;
+	return size << conf->geo.chunk_shift;
 }
 
 static void calc_sectors(struct r10conf *conf, sector_t size)
···
 	 * conf->stride
 	 */
 
-	size = size >> conf->chunk_shift;
-	sector_div(size, conf->far_copies);
-	size = size * conf->raid_disks;
-	sector_div(size, conf->near_copies);
+	size = size >> conf->geo.chunk_shift;
+	sector_div(size, conf->geo.far_copies);
+	size = size * conf->geo.raid_disks;
+	sector_div(size, conf->geo.near_copies);
 	/* 'size' is now the number of chunks in the array */
 	/* calculate "used chunks per device" */
 	size = size * conf->copies;
···
 	/* We need to round up when dividing by raid_disks to
 	 * get the stride size.
 	 */
-	size = DIV_ROUND_UP_SECTOR_T(size, conf->raid_disks);
+	size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
 
-	conf->dev_sectors = size << conf->chunk_shift;
+	conf->dev_sectors = size << conf->geo.chunk_shift;
 
-	if (conf->far_offset)
-		conf->stride = 1 << conf->chunk_shift;
+	if (conf->geo.far_offset)
+		conf->geo.stride = 1 << conf->geo.chunk_shift;
 	else {
-		sector_div(size, conf->far_copies);
-		conf->stride = size << conf->chunk_shift;
+		sector_div(size, conf->geo.far_copies);
+		conf->geo.stride = size << conf->geo.chunk_shift;
 	}
+}
+
+enum geo_type {geo_new, geo_old, geo_start};
+static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
+{
+	int nc, fc, fo;
+	int layout, chunk, disks;
+	switch (new) {
+	case geo_old:
+		layout = mddev->layout;
+		chunk = mddev->chunk_sectors;
+		disks = mddev->raid_disks - mddev->delta_disks;
+		break;
+	case geo_new:
+		layout = mddev->new_layout;
+		chunk = mddev->new_chunk_sectors;
+		disks = mddev->raid_disks;
+		break;
+	default: /* avoid 'may be unused' warnings */
+	case geo_start: /* new when starting reshape - raid_disks not
+			 * updated yet. */
+		layout = mddev->new_layout;
+		chunk = mddev->new_chunk_sectors;
+		disks = mddev->raid_disks + mddev->delta_disks;
+		break;
+	}
+	if (layout >> 17)
+		return -1;
+	if (chunk < (PAGE_SIZE >> 9) ||
+	    !is_power_of_2(chunk))
+		return -2;
+	nc = layout & 255;
+	fc = (layout >> 8) & 255;
+	fo = layout & (1<<16);
+	geo->raid_disks = disks;
+	geo->near_copies = nc;
+	geo->far_copies = fc;
+	geo->far_offset = fo;
+	geo->chunk_mask = chunk - 1;
+	geo->chunk_shift = ffz(~chunk);
+	return nc*fc;
 }
 
 static struct r10conf *setup_conf(struct mddev *mddev)
 {
 	struct r10conf *conf = NULL;
-	int nc, fc, fo;
 	int err = -EINVAL;
+	struct geom geo;
+	int copies;
 
-	if (mddev->new_chunk_sectors < (PAGE_SIZE >> 9) ||
-	    !is_power_of_2(mddev->new_chunk_sectors)) {
+	copies = setup_geo(&geo, mddev, geo_new);
+
+	if (copies == -2) {
 		printk(KERN_ERR "md/raid10:%s: chunk size must be "
 		       "at least PAGE_SIZE(%ld) and be a power of 2.\n",
 		       mdname(mddev), PAGE_SIZE);
 		goto out;
 	}
 
-	nc = mddev->new_layout & 255;
-	fc = (mddev->new_layout >> 8) & 255;
-	fo = mddev->new_layout & (1<<16);
-
-	if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks ||
-	    (mddev->new_layout >> 17)) {
+	if (copies < 2 || copies > mddev->raid_disks) {
 		printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
 		       mdname(mddev), mddev->new_layout);
 		goto out;
···
 	if (!conf)
 		goto out;
 
-	conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
+	/* FIXME calc properly */
+	conf->mirrors = kzalloc(sizeof(struct mirror_info)*(mddev->raid_disks +
+							    max(0,mddev->delta_disks)),
 				GFP_KERNEL);
 	if (!conf->mirrors)
 		goto out;
···
 	if (!conf->tmppage)
 		goto out;
 
-
-	conf->raid_disks = mddev->raid_disks;
-	conf->near_copies = nc;
-	conf->far_copies = fc;
-	conf->copies = nc*fc;
-	conf->far_offset = fo;
-	conf->chunk_mask = mddev->new_chunk_sectors - 1;
-	conf->chunk_shift = ffz(~mddev->new_chunk_sectors);
-
+	conf->geo = geo;
+	conf->copies = copies;
 	conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
 					   r10bio_pool_free, conf);
 	if (!conf->r10bio_pool)
 		goto out;
 
 	calc_sectors(conf, mddev->dev_sectors);
-
+	if (mddev->reshape_position == MaxSector) {
+		conf->prev = conf->geo;
+		conf->reshape_progress = MaxSector;
+	} else {
+		if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
+			err = -EINVAL;
+			goto out;
+		}
+		conf->reshape_progress = mddev->reshape_position;
+		if (conf->prev.far_offset)
+			conf->prev.stride = 1 << conf->prev.chunk_shift;
+		else
+			/* far_copies must be 1 */
+			conf->prev.stride = conf->dev_sectors;
+	}
 	spin_lock_init(&conf->device_lock);
 	INIT_LIST_HEAD(&conf->retry_list);
 
···
 	return conf;
 
  out:
-	printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
-	       mdname(mddev));
+	if (err == -ENOMEM)
+		printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
+		       mdname(mddev));
 	if (conf) {
 		if (conf->r10bio_pool)
 			mempool_destroy(conf->r10bio_pool);
···
 	struct mirror_info *disk;
 	struct md_rdev *rdev;
 	sector_t size;
-
-	/*
-	 * copy the already verified devices into our private RAID10
-	 * bookkeeping area. [whatever we allocate in run(),
-	 * should be freed in stop()]
-	 */
+	sector_t min_offset_diff = 0;
+	int first = 1;
 
 	if (mddev->private == NULL) {
 		conf = setup_conf(mddev);
···
 
 	chunk_size = mddev->chunk_sectors << 9;
 	blk_queue_io_min(mddev->queue, chunk_size);
-	if (conf->raid_disks % conf->near_copies)
-		blk_queue_io_opt(mddev->queue, chunk_size * conf->raid_disks);
+	if (conf->geo.raid_disks % conf->geo.near_copies)
+		blk_queue_io_opt(mddev->queue, chunk_size * conf->geo.raid_disks);
 	else
 		blk_queue_io_opt(mddev->queue, chunk_size *
-				 (conf->raid_disks / conf->near_copies));
+				 (conf->geo.raid_disks / conf->geo.near_copies));
 
 	rdev_for_each(rdev, mddev) {
+		long long diff;
 
 		disk_idx = rdev->raid_disk;
-		if (disk_idx >= conf->raid_disks
-		    || disk_idx < 0)
+		if (disk_idx < 0)
+			continue;
+		if (disk_idx >= conf->geo.raid_disks &&
+		    disk_idx >= conf->prev.raid_disks)
 			continue;
 		disk = conf->mirrors + disk_idx;
 
···
 				goto out_free_conf;
 			disk->rdev = rdev;
 		}
+		diff = (rdev->new_data_offset - rdev->data_offset);
+		if (!mddev->reshape_backwards)
+			diff = -diff;
+		if (diff < 0)
+			diff = 0;
+		if (first || diff < min_offset_diff)
+			min_offset_diff = diff;
 
 		disk_stack_limits(mddev->gendisk, rdev->bdev,
 				  rdev->data_offset << 9);
 
 		disk->head_position = 0;
 	}
+
 	/* need to check that every block has at least one working mirror */
 	if (!enough(conf, -1)) {
 		printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n",
···
 		goto out_free_conf;
 	}
 
+	if (conf->reshape_progress != MaxSector) {
+		/* must ensure that shape change is supported */
+		if (conf->geo.far_copies != 1 &&
+		    conf->geo.far_offset == 0)
+			goto out_free_conf;
+		if (conf->prev.far_copies != 1 &&
+		    conf->geo.far_offset == 0)
+			goto out_free_conf;
+	}
+
 	mddev->degraded = 0;
-	for (i = 0; i < conf->raid_disks; i++) {
+	for (i = 0;
+	     i < conf->geo.raid_disks
+		     || i < conf->prev.raid_disks;
+	     i++) {
 
 		disk = conf->mirrors + i;
 
···
 		       mdname(mddev));
 	printk(KERN_INFO
 		"md/raid10:%s: active with %d out of %d devices\n",
-		mdname(mddev), conf->raid_disks - mddev->degraded,
-		conf->raid_disks);
+		mdname(mddev), conf->geo.raid_disks - mddev->degraded,
+		conf->geo.raid_disks);
 	/*
 	 * Ok, everything is just fine now
 	 */
···
 	 * maybe...
 	 */
 	{
-		int stripe = conf->raid_disks *
+		int stripe = conf->geo.raid_disks *
 			((mddev->chunk_sectors << 9) / PAGE_SIZE);
-		stripe /= conf->near_copies;
-		if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
-			mddev->queue->backing_dev_info.ra_pages = 2* stripe;
+		stripe /= conf->geo.near_copies;
+		if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
+			mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
 	}
 
 	blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
 
 	if (md_integrity_register(mddev))
 		goto out_free_conf;
+
+	if (conf->reshape_progress != MaxSector) {
+		unsigned long before_length, after_length;
+
+		before_length = ((1 << conf->prev.chunk_shift) *
+				 conf->prev.far_copies);
+		after_length = ((1 << conf->geo.chunk_shift) *
+				conf->geo.far_copies);
+
+		if (max(before_length, after_length) > min_offset_diff) {
+			/* This cannot work */
+			printk("md/raid10: offset difference not enough to continue reshape\n");
+			goto out_free_conf;
+		}
+		conf->offset_diff = min_offset_diff;
+
+		conf->reshape_safe = conf->reshape_progress;
+		clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
+		clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
+		set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
+		set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
+		mddev->sync_thread = md_register_thread(md_do_sync, mddev,
+							"reshape");
+	}
 
 	return 0;
 
···
 	struct r10conf *conf = mddev->private;
 	sector_t oldsize, size;
 
-	if (conf->far_copies > 1 && !conf->far_offset)
+	if (mddev->reshape_position != MaxSector)
+		return -EBUSY;
+
+	if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
 		return -EINVAL;
 
 	oldsize = raid10_size(mddev, 0, 0);
 	size = raid10_size(mddev, sectors, 0);
-	md_set_array_sectors(mddev, size);
-	if (mddev->array_sectors > size)
+	if (mddev->external_size &&
+	    mddev->array_sectors > size)
 		return -EINVAL;
+	if (mddev->bitmap) {
+		int ret = bitmap_resize(mddev->bitmap, size, 0, 0);
+		if (ret)
+			return ret;
+	}
+	md_set_array_sectors(mddev, size);
 	set_capacity(mddev->gendisk, mddev->array_sectors);
 	revalidate_disk(mddev->gendisk);
 	if (sectors > mddev->dev_sectors &&
···
 	return ERR_PTR(-EINVAL);
 }
 
+static int raid10_check_reshape(struct mddev *mddev)
+{
+	/* Called when there is a request to change
+	 * - layout (to ->new_layout)
+	 * - chunk size (to ->new_chunk_sectors)
+	 * - raid_disks (by delta_disks)
+	 * or when trying to restart a reshape that was ongoing.
+	 *
+	 * We need to validate the request and possibly allocate
+	 * space if that might be an issue later.
+	 *
+	 * Currently we reject any reshape of a 'far' mode array,
+	 * allow chunk size to change if new is generally acceptable,
+	 * allow raid_disks to increase, and allow
+	 * a switch between 'near' mode and 'offset' mode.
+	 */
+	struct r10conf *conf = mddev->private;
+	struct geom geo;
+
+	if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
+		return -EINVAL;
+
+	if (setup_geo(&geo, mddev, geo_start) != conf->copies)
+		/* mustn't change number of copies */
+		return -EINVAL;
+	if (geo.far_copies > 1 && !geo.far_offset)
+		/* Cannot switch to 'far' mode */
+		return -EINVAL;
+
+	if (mddev->array_sectors & geo.chunk_mask)
+			/* not factor of array size */
+			return -EINVAL;
+
+	if (!enough(conf, -1))
+		return -EINVAL;
+
+	kfree(conf->mirrors_new);
+	conf->mirrors_new = NULL;
+	if (mddev->delta_disks > 0) {
+		/* allocate new 'mirrors' list */
+		conf->mirrors_new = kzalloc(
+			sizeof(struct mirror_info)
+			*(mddev->raid_disks +
+			  mddev->delta_disks),
+			GFP_KERNEL);
+		if (!conf->mirrors_new)
+			return -ENOMEM;
+	}
+	return 0;
+}
+
+/*
+ * Need to check if array has failed when deciding whether to:
+ *  - start an array
+ *  - remove non-faulty devices
+ *  - add a spare
+ *  - allow a reshape
+ * This determination is simple when no reshape is happening.
+ * However if there is a reshape, we need to carefully check
+ * both the before and after sections.
+ * This is because some failed devices may only affect one
+ * of the two sections, and some non-in_sync devices may
+ * be insync in the section most affected by failed devices.
+ */
+static int calc_degraded(struct r10conf *conf)
+{
+	int degraded, degraded2;
+	int i;
+
+	rcu_read_lock();
+	degraded = 0;
+	/* 'prev' section first */
+	for (i = 0; i < conf->prev.raid_disks; i++) {
+		struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
+		if (!rdev || test_bit(Faulty, &rdev->flags))
+			degraded++;
+		else if (!test_bit(In_sync, &rdev->flags))
+			/* When we can reduce the number of devices in
+			 * an array, this might not contribute to
+			 * 'degraded'.  It does now.
+			 */
+			degraded++;
+	}
+	rcu_read_unlock();
+	if (conf->geo.raid_disks == conf->prev.raid_disks)
+		return degraded;
+	rcu_read_lock();
+	degraded2 = 0;
+	for (i = 0; i < conf->geo.raid_disks; i++) {
+		struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
+		if (!rdev || test_bit(Faulty, &rdev->flags))
+			degraded2++;
+		else if (!test_bit(In_sync, &rdev->flags)) {
+			/* If reshape is increasing the number of devices,
+			 * this section has already been recovered, so
+			 * it doesn't contribute to degraded.
+			 * else it does.
+			 */
+			if (conf->geo.raid_disks <= conf->prev.raid_disks)
+				degraded2++;
+		}
+	}
+	rcu_read_unlock();
+	if (degraded2 > degraded)
+		return degraded2;
+	return degraded;
+}
+
+static int raid10_start_reshape(struct mddev *mddev)
+{
+	/* A 'reshape' has been requested. This commits
+	 * the various 'new' fields and sets MD_RECOVER_RESHAPE
+	 * This also checks if there are enough spares and adds them
+	 * to the array.
+	 * We currently require enough spares to make the final
+	 * array non-degraded.  We also require that the difference
+	 * between old and new data_offset - on each device - is
+	 * enough that we never risk over-writing.
+	 */
+
+	unsigned long before_length, after_length;
+	sector_t min_offset_diff = 0;
+	int first = 1;
+	struct geom new;
+	struct r10conf *conf = mddev->private;
+	struct md_rdev *rdev;
+	int spares = 0;
+	int ret;
+
+	if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
+		return -EBUSY;
+
+	if (setup_geo(&new, mddev, geo_start) != conf->copies)
+		return -EINVAL;
+
+	before_length = ((1 << conf->prev.chunk_shift) *
+			 conf->prev.far_copies);
+	after_length = ((1 << conf->geo.chunk_shift) *
+			conf->geo.far_copies);
+
+	rdev_for_each(rdev, mddev) {
+		if (!test_bit(In_sync, &rdev->flags)
+		    && !test_bit(Faulty, &rdev->flags))
+			spares++;
+		if (rdev->raid_disk >= 0) {
+			long long diff = (rdev->new_data_offset
+					  - rdev->data_offset);
+			if (!mddev->reshape_backwards)
+				diff = -diff;
+			if (diff < 0)
+				diff = 0;
+			if (first || diff < min_offset_diff)
+				min_offset_diff = diff;
+		}
+	}
+
+	if (max(before_length, after_length) > min_offset_diff)
+		return -EINVAL;
+
+	if (spares < mddev->delta_disks)
+		return -EINVAL;
+
+	conf->offset_diff = min_offset_diff;
+	spin_lock_irq(&conf->device_lock);
+	if (conf->mirrors_new) {
+		memcpy(conf->mirrors_new, conf->mirrors,
+		       sizeof(struct mirror_info)*conf->prev.raid_disks);
+		smp_mb();
+		kfree(conf->mirrors_old); /* FIXME and elsewhere */
+		conf->mirrors_old = conf->mirrors;
+		conf->mirrors = conf->mirrors_new;
+		conf->mirrors_new = NULL;
+	}
+	setup_geo(&conf->geo, mddev, geo_start);
+	smp_mb();
+	if (mddev->reshape_backwards) {
+		sector_t size = raid10_size(mddev, 0, 0);
+		if (size < mddev->array_sectors) {
+			spin_unlock_irq(&conf->device_lock);
+			printk(KERN_ERR "md/raid10:%s: array size must be reduce before number of disks\n",
+			       mdname(mddev));
+			return -EINVAL;
+		}
+		mddev->resync_max_sectors = size;
+		conf->reshape_progress = size;
+	} else
+		conf->reshape_progress = 0;
+	spin_unlock_irq(&conf->device_lock);
+
+	if (mddev->delta_disks && mddev->bitmap) {
+		ret = bitmap_resize(mddev->bitmap,
+				    raid10_size(mddev, 0,
+						conf->geo.raid_disks),
+				    0, 0);
+		if (ret)
+			goto abort;
+	}
+	if (mddev->delta_disks > 0) {
+		rdev_for_each(rdev, mddev)
+			if (rdev->raid_disk < 0 &&
+			    !test_bit(Faulty, &rdev->flags)) {
+				if (raid10_add_disk(mddev, rdev) == 0) {
+					if (rdev->raid_disk >=
+					    conf->prev.raid_disks)
+						set_bit(In_sync, &rdev->flags);
+					else
+						rdev->recovery_offset = 0;
+
+					if (sysfs_link_rdev(mddev, rdev))
+						/* Failure here  is OK */;
+				}
+			} else if (rdev->raid_disk >= conf->prev.raid_disks
+				   && !test_bit(Faulty, &rdev->flags)) {
+				/* This is a spare that was manually added */
+				set_bit(In_sync, &rdev->flags);
+			}
+	}
+	/* When a reshape changes the number of devices,
+	 * ->degraded is measured against the larger of the
+	 * pre and  post numbers.
+	 */
+	spin_lock_irq(&conf->device_lock);
+	mddev->degraded = calc_degraded(conf);
+	spin_unlock_irq(&conf->device_lock);
+	mddev->raid_disks = conf->geo.raid_disks;
+	mddev->reshape_position = conf->reshape_progress;
+	set_bit(MD_CHANGE_DEVS, &mddev->flags);
+
+	clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
+	clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
+	set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
+	set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
+
+	mddev->sync_thread = md_register_thread(md_do_sync, mddev,
+						"reshape");
+	if (!mddev->sync_thread) {
+		ret = -EAGAIN;
+		goto abort;
+	}
+	conf->reshape_checkpoint = jiffies;
+	md_wakeup_thread(mddev->sync_thread);
+	md_new_event(mddev);
+	return 0;
+
+abort:
+	mddev->recovery = 0;
+	spin_lock_irq(&conf->device_lock);
+	conf->geo = conf->prev;
+	mddev->raid_disks = conf->geo.raid_disks;
+	rdev_for_each(rdev, mddev)
+		rdev->new_data_offset = rdev->data_offset;
+	smp_wmb();
+	conf->reshape_progress = MaxSector;
+	mddev->reshape_position = MaxSector;
+	spin_unlock_irq(&conf->device_lock);
+	return ret;
+}
+
+/* Calculate the last device-address that could contain
+ * any block from the chunk that includes the array-address 's'
+ * and report the next address.
+ * i.e. the address returned will be chunk-aligned and after
+ * any data that is in the chunk containing 's'.
+ */
+static sector_t last_dev_address(sector_t s, struct geom *geo)
+{
+	s = (s | geo->chunk_mask) + 1;
+	s >>= geo->chunk_shift;
+	s *= geo->near_copies;
+	s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
+	s *= geo->far_copies;
+	s <<= geo->chunk_shift;
+	return s;
+}
+
+/* Calculate the first device-address that could contain
+ * any block from the chunk that includes the array-address 's'.
+ * This too will be the start of a chunk
+ */
+static sector_t first_dev_address(sector_t s, struct geom *geo)
+{
+	s >>= geo->chunk_shift;
+	s *= geo->near_copies;
+	sector_div(s, geo->raid_disks);
+	s *= geo->far_copies;
+	s <<= geo->chunk_shift;
+	return s;
+}
+
+static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
+				int *skipped)
+{
+	/* We simply copy at most one chunk (smallest of old and new)
+	 * at a time, possibly less if that exceeds RESYNC_PAGES,
+	 * or we hit a bad block or something.
+	 * This might mean we pause for normal IO in the middle of
+	 * a chunk, but that is not a problem was mddev->reshape_position
+	 * can record any location.
+	 *
+	 * If we will want to write to a location that isn't
+	 * yet recorded as 'safe' (i.e. in metadata on disk) then
+	 * we need to flush all reshape requests and update the metadata.
+	 *
+	 * When reshaping forwards (e.g. to more devices), we interpret
+	 * 'safe' as the earliest block which might not have been copied
+	 * down yet.  We divide this by previous stripe size and multiply
+	 * by previous stripe length to get lowest device offset that we
+	 * cannot write to yet.
+	 * We interpret 'sector_nr' as an address that we want to write to.
+	 * From this we use last_device_address() to find where we might
+	 * write to, and first_device_address on the  'safe' position.
+	 * If this 'next' write position is after the 'safe' position,
+	 * we must update the metadata to increase the 'safe' position.
+	 *
+	 * When reshaping backwards, we round in the opposite direction
+	 * and perform the reverse test:  next write position must not be
+	 * less than current safe position.
+	 *
+	 * In all this the minimum difference in data offsets
+	 * (conf->offset_diff - always positive) allows a bit of slack,
+	 * so next can be after 'safe', but not by more than offset_disk
+	 *
+	 * We need to prepare all the bios here before we start any IO
+	 * to ensure the size we choose is acceptable to all devices.
+	 * The means one for each copy for write-out and an extra one for
+	 * read-in.
+	 * We store the read-in bio in ->master_bio and the others in
+	 * ->devs[x].bio and ->devs[x].repl_bio.
+	 */
+	struct r10conf *conf = mddev->private;
+	struct r10bio *r10_bio;
+	sector_t next, safe, last;
+	int max_sectors;
+	int nr_sectors;
+	int s;
+	struct md_rdev *rdev;
+	int need_flush = 0;
+	struct bio *blist;
+	struct bio *bio, *read_bio;
+	int sectors_done = 0;
+
+	if (sector_nr == 0) {
+		/* If restarting in the middle, skip the initial sectors */
+		if (mddev->reshape_backwards &&
+		    conf->reshape_progress < raid10_size(mddev, 0, 0)) {
+			sector_nr = (raid10_size(mddev, 0, 0)
+				     - conf->reshape_progress);
+		} else if (!mddev->reshape_backwards &&
+			   conf->reshape_progress > 0)
+			sector_nr = conf->reshape_progress;
+		if (sector_nr) {
+			mddev->curr_resync_completed = sector_nr;
+			sysfs_notify(&mddev->kobj, NULL, "sync_completed");
+			*skipped = 1;
+			return sector_nr;
+		}
+	}
+
+	/* We don't use sector_nr to track where we are up to
+	 * as that doesn't work well for ->reshape_backwards.
+	 * So just use ->reshape_progress.
+	 */
+	if (mddev->reshape_backwards) {
+		/* 'next' is the earliest device address that we might
+		 * write to for this chunk in the new layout
+		 */
+		next = first_dev_address(conf->reshape_progress - 1,
+					 &conf->geo);
+
+		/* 'safe' is the last device address that we might read from
+		 * in the old layout after a restart
+		 */
+		safe = last_dev_address(conf->reshape_safe - 1,
+					&conf->prev);
+
+		if (next + conf->offset_diff < safe)
+			need_flush = 1;
+
+		last = conf->reshape_progress - 1;
+		sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
+					       & conf->prev.chunk_mask);
+		if (sector_nr + RESYNC_BLOCK_SIZE/512 < last)
+			sector_nr = last + 1 - RESYNC_BLOCK_SIZE/512;
+	} else {
+		/* 'next' is after the last device address that we
+		 * might write to for this chunk in the new layout
+		 */
+		next = last_dev_address(conf->reshape_progress, &conf->geo);
+
+		/* 'safe' is the earliest device address that we might
+		 * read from in the old layout after a restart
+		 */
+		safe = first_dev_address(conf->reshape_safe, &conf->prev);
+
+		/* Need to update metadata if 'next' might be beyond 'safe'
+		 * as that would possibly corrupt data
+		 */
+		if (next > safe + conf->offset_diff)
+			need_flush = 1;
+
+		sector_nr = conf->reshape_progress;
+		last  = sector_nr | (conf->geo.chunk_mask
+				     & conf->prev.chunk_mask);
+
+		if (sector_nr + RESYNC_BLOCK_SIZE/512 <= last)
+			last = sector_nr + RESYNC_BLOCK_SIZE/512 - 1;
+	}
+
+	if (need_flush ||
+	    time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
+		/* Need to update reshape_position in metadata */
+		wait_barrier(conf);
+		mddev->reshape_position = conf->reshape_progress;
+		if (mddev->reshape_backwards)
+			mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
+				- conf->reshape_progress;
+		else
+			mddev->curr_resync_completed = conf->reshape_progress;
+		conf->reshape_checkpoint = jiffies;
+		set_bit(MD_CHANGE_DEVS, &mddev->flags);
+		md_wakeup_thread(mddev->thread);
+		wait_event(mddev->sb_wait, mddev->flags == 0 ||
+			   kthread_should_stop());
+		conf->reshape_safe = mddev->reshape_position;
+		allow_barrier(conf);
+	}
+
+read_more:
+	/* Now schedule reads for blocks from sector_nr to last */
+	r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
+	raise_barrier(conf, sectors_done != 0);
+	atomic_set(&r10_bio->remaining, 0);
+	r10_bio->mddev = mddev;
+	r10_bio->sector = sector_nr;
+	set_bit(R10BIO_IsReshape, &r10_bio->state);
+	r10_bio->sectors = last - sector_nr + 1;
+	rdev = read_balance(conf, r10_bio, &max_sectors);
+	BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
+
+	if (!rdev) {
+		/* Cannot read from here, so need to record bad blocks
+		 * on all the target devices.
+		 */
+		// FIXME
+		set_bit(MD_RECOVERY_INTR, &mddev->recovery);
+		return sectors_done;
+	}
+
+	read_bio = bio_alloc_mddev(GFP_KERNEL, RESYNC_PAGES, mddev);
+
+	read_bio->bi_bdev = rdev->bdev;
+	read_bio->bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
+			       + rdev->data_offset);
+	read_bio->bi_private = r10_bio;
+	read_bio->bi_end_io = end_sync_read;
+	read_bio->bi_rw = READ;
+	read_bio->bi_flags &= ~(BIO_POOL_MASK - 1);
+	read_bio->bi_flags |= 1 << BIO_UPTODATE;
+	read_bio->bi_vcnt = 0;
+	read_bio->bi_idx = 0;
+	read_bio->bi_size = 0;
+	r10_bio->master_bio = read_bio;
+	r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
+
+	/* Now find the locations in the new layout */
+	__raid10_find_phys(&conf->geo, r10_bio);
+
+	blist = read_bio;
+	read_bio->bi_next = NULL;
+
+	for (s = 0; s < conf->copies*2; s++) {
+		struct bio *b;
+		int d = r10_bio->devs[s/2].devnum;
+		struct md_rdev *rdev2;
+		if (s&1) {
+			rdev2 = conf->mirrors[d].replacement;
+			b = r10_bio->devs[s/2].repl_bio;
+		} else {
+			rdev2 = conf->mirrors[d].rdev;
+			b = r10_bio->devs[s/2].bio;
+		}
+		if (!rdev2 || test_bit(Faulty, &rdev2->flags))
+			continue;
+		b->bi_bdev = rdev2->bdev;
+		b->bi_sector = r10_bio->devs[s/2].addr + rdev2->new_data_offset;
+		b->bi_private = r10_bio;
+		b->bi_end_io = end_reshape_write;
+		b->bi_rw = WRITE;
+		b->bi_flags &= ~(BIO_POOL_MASK - 1);
+		b->bi_flags |= 1 << BIO_UPTODATE;
+		b->bi_next = blist;
+		b->bi_vcnt = 0;
+		b->bi_idx = 0;
+		b->bi_size = 0;
+		blist = b;
+	}
+
+	/* Now add as many pages as possible to all of these bios. */
+
+	nr_sectors = 0;
+	for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
+		struct page *page = r10_bio->devs[0].bio->bi_io_vec[s/(PAGE_SIZE>>9)].bv_page;
+		int len = (max_sectors - s) << 9;
+		if (len > PAGE_SIZE)
+			len = PAGE_SIZE;
+		for (bio = blist; bio ; bio = bio->bi_next) {
+			struct bio *bio2;
+			if (bio_add_page(bio, page, len, 0))
+				continue;
+
+			/* Didn't fit, must stop */
+			for (bio2 = blist;
+			     bio2 && bio2 != bio;
+			     bio2 = bio2->bi_next) {
+				/* Remove last page from this bio */
+				bio2->bi_vcnt--;
+				bio2->bi_size -= len;
+				bio2->bi_flags &= ~(1<<BIO_SEG_VALID);
+			}
+			goto bio_full;
+		}
+		sector_nr += len >> 9;
+		nr_sectors += len >> 9;
+	}
+bio_full:
+	r10_bio->sectors = nr_sectors;
+
+	/* Now submit the read */
+	md_sync_acct(read_bio->bi_bdev, r10_bio->sectors);
+	atomic_inc(&r10_bio->remaining);
+	read_bio->bi_next = NULL;
+	generic_make_request(read_bio);
+	sector_nr += nr_sectors;
+	sectors_done += nr_sectors;
+	if (sector_nr <= last)
+		goto read_more;
+
+	/* Now that we have done the whole section we can
+	 * update reshape_progress
+	 */
+	if (mddev->reshape_backwards)
+		conf->reshape_progress -= sectors_done;
+	else
+		conf->reshape_progress += sectors_done;
+
+	return sectors_done;
+}
+
+static void end_reshape_request(struct r10bio *r10_bio);
+static int handle_reshape_read_error(struct mddev *mddev,
+				     struct r10bio *r10_bio);
+static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
+{
+	/* Reshape read completed.  Hopefully we have a block
+	 * to write out.
+	 * If we got a read error then we do sync 1-page reads from
+	 * elsewhere until we find the data - or give up.
+	 */
+	struct r10conf *conf = mddev->private;
+	int s;
+
+	if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
+		if (handle_reshape_read_error(mddev, r10_bio) < 0) {
+			/* Reshape has been aborted */
+			md_done_sync(mddev, r10_bio->sectors, 0);
+			return;
+		}
+
+	/* We definitely have the data in the pages, schedule the
+	 * writes.
+	 */
+	atomic_set(&r10_bio->remaining, 1);
+	for (s = 0; s < conf->copies*2; s++) {
+		struct bio *b;
+		int d = r10_bio->devs[s/2].devnum;
+		struct md_rdev *rdev;
+		if (s&1) {
+			rdev = conf->mirrors[d].replacement;
+			b = r10_bio->devs[s/2].repl_bio;
+		} else {
+			rdev = conf->mirrors[d].rdev;
+			b = r10_bio->devs[s/2].bio;
+		}
+		if (!rdev || test_bit(Faulty, &rdev->flags))
+			continue;
+		atomic_inc(&rdev->nr_pending);
+		md_sync_acct(b->bi_bdev, r10_bio->sectors);
+		atomic_inc(&r10_bio->remaining);
+		b->bi_next = NULL;
+		generic_make_request(b);
+	}
+	end_reshape_request(r10_bio);
+}
+
+static void end_reshape(struct r10conf *conf)
+{
+	if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
+		return;
+
+	spin_lock_irq(&conf->device_lock);
+	conf->prev = conf->geo;
+	md_finish_reshape(conf->mddev);
+	smp_wmb();
+	conf->reshape_progress = MaxSector;
+	spin_unlock_irq(&conf->device_lock);
+
+	/* read-ahead size must cover two whole stripes, which is
+	 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
+	 */
+	if (conf->mddev->queue) {
+		int stripe = conf->geo.raid_disks *
+			((conf->mddev->chunk_sectors << 9) / PAGE_SIZE);
+		stripe /= conf->geo.near_copies;
+		if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
+			conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
+	}
+	conf->fullsync = 0;
+}
+
+
+static int handle_reshape_read_error(struct mddev *mddev,
+				     struct r10bio *r10_bio)
+{
+	/* Use sync reads to get the blocks from somewhere else */
+	int sectors = r10_bio->sectors;
+	struct r10bio r10b;
+	struct r10conf *conf = mddev->private;
+	int slot = 0;
+	int idx = 0;
+	struct bio_vec *bvec = r10_bio->master_bio->bi_io_vec;
+
+	r10b.sector = r10_bio->sector;
+	__raid10_find_phys(&conf->prev, &r10b);
+
+	while (sectors) {
+		int s = sectors;
+		int success = 0;
+		int first_slot = slot;
+
+		if (s > (PAGE_SIZE >> 9))
+			s = PAGE_SIZE >> 9;
+
+		while (!success) {
+			int d = r10b.devs[slot].devnum;
+			struct md_rdev *rdev = conf->mirrors[d].rdev;
+			sector_t addr;
+			if (rdev == NULL ||
+			    test_bit(Faulty, &rdev->flags) ||
+			    !test_bit(In_sync, &rdev->flags))
+				goto failed;
+
+			addr = r10b.devs[slot].addr + idx * PAGE_SIZE;
+			success = sync_page_io(rdev,
+					       addr,
+					       s << 9,
+					       bvec[idx].bv_page,
+					       READ, false);
+			if (success)
+				break;
+		failed:
+			slot++;
+			if (slot >= conf->copies)
+				slot = 0;
+			if (slot == first_slot)
+				break;
+		}
+		if (!success) {
+			/* couldn't read this block, must give up */
+			set_bit(MD_RECOVERY_INTR,
+				&mddev->recovery);
+			return -EIO;
+		}
+		sectors -= s;
+		idx++;
+	}
+	return 0;
+}
+
+static void end_reshape_write(struct bio *bio, int error)
+{
+	int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
+	struct r10bio *r10_bio = bio->bi_private;
+	struct mddev *mddev = r10_bio->mddev;
+	struct r10conf *conf = mddev->private;
+	int d;
+	int slot;
+	int repl;
+	struct md_rdev *rdev = NULL;
+
+	d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
+	if (repl)
+		rdev = conf->mirrors[d].replacement;
+	if (!rdev) {
+		smp_mb();
+		rdev = conf->mirrors[d].rdev;
+	}
+
+	if (!uptodate) {
+		/* FIXME should record badblock */
+		md_error(mddev, rdev);
+	}
+
+	rdev_dec_pending(rdev, mddev);
+	end_reshape_request(r10_bio);
+}
+
+static void end_reshape_request(struct r10bio *r10_bio)
+{
+	if (!atomic_dec_and_test(&r10_bio->remaining))
+		return;
+	md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
+	bio_put(r10_bio->master_bio);
+	put_buf(r10_bio);
+}
+
+static void raid10_finish_reshape(struct mddev *mddev)
+{
+	struct r10conf *conf = mddev->private;
+
+	if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
+		return;
+
+	if (mddev->delta_disks > 0) {
+		sector_t size = raid10_size(mddev, 0, 0);
+		md_set_array_sectors(mddev, size);
+		if (mddev->recovery_cp > mddev->resync_max_sectors) {
+			mddev->recovery_cp = mddev->resync_max_sectors;
+			set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
+		}
+		mddev->resync_max_sectors = size;
+		set_capacity(mddev->gendisk, mddev->array_sectors);
+		revalidate_disk(mddev->gendisk);
+	} else {
+		int d;
+		for (d = conf->geo.raid_disks ;
+		     d < conf->geo.raid_disks - mddev->delta_disks;
+		     d++) {
+			struct md_rdev *rdev = conf->mirrors[d].rdev;
+			if (rdev)
+				clear_bit(In_sync, &rdev->flags);
+			rdev = conf->mirrors[d].replacement;
+			if (rdev)
+				clear_bit(In_sync, &rdev->flags);
+		}
+	}
+	mddev->layout = mddev->new_layout;
+	mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
+	mddev->reshape_position = MaxSector;
+	mddev->delta_disks = 0;
+	mddev->reshape_backwards = 0;
+}
+
 static struct md_personality raid10_personality =
 {
 	.name		= "raid10",
···
 	.size		= raid10_size,
 	.resize		= raid10_resize,
 	.takeover	= raid10_takeover,
+	.check_reshape	= raid10_check_reshape,
+	.start_reshape	= raid10_start_reshape,
+	.finish_reshape	= raid10_finish_reshape,
 };
 
 static int __init raid_init(void)

drivers/md/raid10.h

···
 struct r10conf {
 	struct mddev		*mddev;
 	struct mirror_info	*mirrors;
-	int			raid_disks;
+	struct mirror_info	*mirrors_new, *mirrors_old;
 	spinlock_t		device_lock;
 
 	/* geometry */
-	int			near_copies;  /* number of copies laid out
+	struct geom {
+		int		raid_disks;
+		int		near_copies;  /* number of copies laid out
 					       * raid0 style */
-	int 			far_copies;   /* number of copies laid out
+		int		far_copies;   /* number of copies laid out
 					       * at large strides across drives
 					       */
-	int			far_offset;   /* far_copies are offset by 1
+		int		far_offset;   /* far_copies are offset by 1
 					       * stripe instead of many
 					       */
-	int			copies;	      /* near_copies * far_copies.
-					       * must be <= raid_disks
-					       */
-	sector_t		stride;	      /* distance between far copies.
+		sector_t	stride;	      /* distance between far copies.
 					       * This is size / far_copies unless
 					       * far_offset, in which case it is
 					       * 1 stripe.
 					       */
+		int		chunk_shift; /* shift from chunks to sectors */
+		sector_t	chunk_mask;
+	} prev, geo;
+	int			copies;	      /* near_copies * far_copies.
+					       * must be <= raid_disks
+					       */
 
 	sector_t		dev_sectors;  /* temp copy of
 					       * mddev->dev_sectors */
-
-	int			chunk_shift; /* shift from chunks to sectors */
-	sector_t		chunk_mask;
+	sector_t		reshape_progress;
+	sector_t		reshape_safe;
+	unsigned long		reshape_checkpoint;
+	sector_t		offset_diff;
 
 	struct list_head	retry_list;
 	/* queue pending writes and submit them on unplug */
···
 	R10BIO_Uptodate,
 	R10BIO_IsSync,
 	R10BIO_IsRecover,
+	R10BIO_IsReshape,
 	R10BIO_Degraded,
 /* Set ReadError on bios that experience a read error
  * so that raid10d knows what to do with them.
···
  */
 	R10BIO_MadeGood,
 	R10BIO_WriteError,
+/* During a reshape we might be performing IO on the
+ * 'previous' part of the array, in which case this
+ * flag is set
+ */
+	R10BIO_Previous,
 };
 #endif

drivers/md/raid5.c

···
 	return sh;
 }
 
+/* Determine if 'data_offset' or 'new_data_offset' should be used
+ * in this stripe_head.
+ */
+static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
+{
+	sector_t progress = conf->reshape_progress;
+	/* Need a memory barrier to make sure we see the value
+	 * of conf->generation, or ->data_offset that was set before
+	 * reshape_progress was updated.
+	 */
+	smp_rmb();
+	if (progress == MaxSector)
+		return 0;
+	if (sh->generation == conf->generation - 1)
+		return 0;
+	/* We are in a reshape, and this is a new-generation stripe,
+	 * so use new_data_offset.
+	 */
+	return 1;
+}
+
 static void
 raid5_end_read_request(struct bio *bi, int error);
 static void
···
 			replace_only = 1;
 		} else
 			continue;
+		if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
+			rw |= REQ_SYNC;
 
 		bi = &sh->dev[i].req;
 		rbi = &sh->dev[i].rreq; /* For writing to replacement */
···
 				__func__, (unsigned long long)sh->sector,
 				bi->bi_rw, i);
 			atomic_inc(&sh->count);
-			bi->bi_sector = sh->sector + rdev->data_offset;
+			if (use_new_offset(conf, sh))
+				bi->bi_sector = (sh->sector
+						 + rdev->new_data_offset);
+			else
+				bi->bi_sector = (sh->sector
+						 + rdev->data_offset);
 			bi->bi_flags = 1 << BIO_UPTODATE;
 			bi->bi_idx = 0;
 			bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
···
 				__func__, (unsigned long long)sh->sector,
 				rbi->bi_rw, i);
 			atomic_inc(&sh->count);
-			rbi->bi_sector = sh->sector + rrdev->data_offset;
+			if (use_new_offset(conf, sh))
+				rbi->bi_sector = (sh->sector
+						  + rrdev->new_data_offset);
+			else
+				rbi->bi_sector = (sh->sector
+						  + rrdev->data_offset);
 			rbi->bi_flags = 1 << BIO_UPTODATE;
 			rbi->bi_idx = 0;
 			rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
···
 				dev->sector + STRIPE_SECTORS) {
 				if (wbi->bi_rw & REQ_FUA)
 					set_bit(R5_WantFUA, &dev->flags);
+				if (wbi->bi_rw & REQ_SYNC)
+					set_bit(R5_SyncIO, &dev->flags);
 				tx = async_copy_data(1, wbi, dev->page,
 					dev->sector, tx);
 				wbi = r5_next_bio(wbi, dev->sector);
···
 	int pd_idx = sh->pd_idx;
 	int qd_idx = sh->qd_idx;
 	int i;
-	bool fua = false;
+	bool fua = false, sync = false;
 
 	pr_debug("%s: stripe %llu\n", __func__,
 		(unsigned long long)sh->sector);
 
-	for (i = disks; i--; )
+	for (i = disks; i--; ) {
 		fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
+		sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
+	}
 
 	for (i = disks; i--; ) {
 		struct r5dev *dev = &sh->dev[i];
···
 			set_bit(R5_UPTODATE, &dev->flags);
 			if (fua)
 				set_bit(R5_WantFUA, &dev->flags);
+			if (sync)
+				set_bit(R5_SyncIO, &dev->flags);
 		}
 	}
 
···
 	int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
 	char b[BDEVNAME_SIZE];
 	struct md_rdev *rdev = NULL;
-
+	sector_t s;
 
 	for (i=0 ; i<disks; i++)
 		if (bi == &sh->dev[i].req)
···
 	if (!rdev)
 		rdev = conf->disks[i].rdev;
 
+	if (use_new_offset(conf, sh))
+		s = sh->sector + rdev->new_data_offset;
+	else
+		s = sh->sector + rdev->data_offset;
 	if (uptodate) {
 		set_bit(R5_UPTODATE, &sh->dev[i].flags);
 		if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
···
 				"md/raid:%s: read error corrected"
 				" (%lu sectors at %llu on %s)\n",
 				mdname(conf->mddev), STRIPE_SECTORS,
-				(unsigned long long)(sh->sector
-						     + rdev->data_offset),
+				(unsigned long long)s,
 				bdevname(rdev->bdev, b));
 			atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
 			clear_bit(R5_ReadError, &sh->dev[i].flags);
···
 				"md/raid:%s: read error on replacement device "
 				"(sector %llu on %s).\n",
 				mdname(conf->mddev),
-				(unsigned long long)(sh->sector
-						     + rdev->data_offset),
+				(unsigned long long)s,
 				bdn);
 		else if (conf->mddev->degraded >= conf->max_degraded)
 			printk_ratelimited(
···
 				"md/raid:%s: read error not correctable "
 				"(sector %llu on %s).\n",
 				mdname(conf->mddev),
-				(unsigned long long)(sh->sector
-						     + rdev->data_offset),
+				(unsigned long long)s,
 				bdn);
 		else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
 			/* Oh, no!!! */
···
 				"md/raid:%s: read error NOT corrected!! "
 				"(sector %llu on %s).\n",
 				mdname(conf->mddev),
-				(unsigned long long)(sh->sector
-						     + rdev->data_offset),
+				(unsigned long long)s,
 				bdn);
 		else if (atomic_read(&rdev->read_errors)
 			 > conf->max_nr_stripes)
···
 			if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
 				rdev = conf->disks[i].rdev;
 				rdev_clear_badblocks(rdev, sh->sector,
-						     STRIPE_SECTORS);
+						     STRIPE_SECTORS, 0);
 				rdev_dec_pending(rdev, conf->mddev);
 			}
 			if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
···
 					/* rdev have been moved down */
 					rdev = conf->disks[i].rdev;
 				rdev_clear_badblocks(rdev, sh->sector,
-						     STRIPE_SECTORS);
+						     STRIPE_SECTORS, 0);
 				rdev_dec_pending(rdev, conf->mddev);
 			}
 		}
···
 		raid_bio->bi_next = (void*)rdev;
 		align_bi->bi_bdev =  rdev->bdev;
 		align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
+		/* No reshape active, so we can trust rdev->data_offset */
 		align_bi->bi_sector += rdev->data_offset;
 
 		if (!bio_fits_rdev(align_bi) ||
···
 	plugged = mddev_check_plugged(mddev);
 	for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
 		DEFINE_WAIT(w);
-		int disks, data_disks;
 		int previous;
 
 	retry:
 		previous = 0;
-		disks = conf->raid_disks;
 		prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
 		if (unlikely(conf->reshape_progress != MaxSector)) {
 			/* spinlock is needed as reshape_progress may be
···
 			 * to check again.
 			 */
 			spin_lock_irq(&conf->device_lock);
-			if (mddev->delta_disks < 0
+			if (mddev->reshape_backwards
 			    ? logical_sector < conf->reshape_progress
 			    : logical_sector >= conf->reshape_progress) {
-				disks = conf->previous_raid_disks;
 				previous = 1;
 			} else {
-				if (mddev->delta_disks < 0
+				if (mddev->reshape_backwards
 				    ? logical_sector < conf->reshape_safe
 				    : logical_sector >= conf->reshape_safe) {
 					spin_unlock_irq(&conf->device_lock);
···
 			}
 			spin_unlock_irq(&conf->device_lock);
 		}
-		data_disks = disks - conf->max_degraded;
 
 		new_sector = raid5_compute_sector(conf, logical_sector,
 						  previous,
···
 				 */
 				int must_retry = 0;
 				spin_lock_irq(&conf->device_lock);
-				if (mddev->delta_disks < 0
+				if (mddev->reshape_backwards
 				    ? logical_sector >= conf->reshape_progress
 				    : logical_sector < conf->reshape_progress)
 					/* mismatch, need to try again */
···
 
 	if (sector_nr == 0) {
 		/* If restarting in the middle, skip the initial sectors */
-		if (mddev->delta_disks < 0 &&
+		if (mddev->reshape_backwards &&
 		    conf->reshape_progress < raid5_size(mddev, 0, 0)) {
 			sector_nr = raid5_size(mddev, 0, 0)
 				- conf->reshape_progress;
-		} else if (mddev->delta_disks >= 0 &&
+		} else if (!mddev->reshape_backwards &&
 			   conf->reshape_progress > 0)
 			sector_nr = conf->reshape_progress;
 		sector_div(sector_nr, new_data_disks);
···
 	else
 		reshape_sectors = mddev->chunk_sectors;
 
-	/* we update the metadata when there is more than 3Meg
-	 * in the block range (that is rather arbitrary, should
-	 * probably be time based) or when the data about to be
-	 * copied would over-write the source of the data at
-	 * the front of the range.
-	 * i.e. one new_stripe along from reshape_progress new_maps
-	 * to after where reshape_safe old_maps to
+	/* We update the metadata at least every 10 seconds, or when
+	 * the data about to be copied would over-write the source of
+	 * the data at the front of the range.  i.e. one new_stripe
+	 * along from reshape_progress new_maps to after where
+	 * reshape_safe old_maps to
 	 */
 	writepos = conf->reshape_progress;
 	sector_div(writepos, new_data_disks);
···
 	sector_div(readpos, data_disks);
 	safepos = conf->reshape_safe;
 	sector_div(safepos, data_disks);
-	if (mddev->delta_disks < 0) {
+	if (mddev->reshape_backwards) {
 		writepos -= min_t(sector_t, reshape_sectors, writepos);
 		readpos += reshape_sectors;
 		safepos += reshape_sectors;
···
 		safepos -= min_t(sector_t, reshape_sectors, safepos);
 	}
 
+	/* Having calculated the 'writepos' possibly use it
+	 * to set 'stripe_addr' which is where we will write to.
+	 */
+	if (mddev->reshape_backwards) {
+		BUG_ON(conf->reshape_progress == 0);
+		stripe_addr = writepos;
+		BUG_ON((mddev->dev_sectors &
+			~((sector_t)reshape_sectors - 1))
+		       - reshape_sectors - stripe_addr
+		       != sector_nr);
+	} else {
+		BUG_ON(writepos != sector_nr + reshape_sectors);
+		stripe_addr = sector_nr;
+	}
+
 	/* 'writepos' is the most advanced device address we might write.
 	 * 'readpos' is the least advanced device address we might read.
 	 * 'safepos' is the least address recorded in the metadata as having
 	 *     been reshaped.
-	 * If 'readpos' is behind 'writepos', then there is no way that we can
+	 * If there is a min_offset_diff, these are adjusted either by
+	 * increasing the safepos/readpos if diff is negative, or
+	 * increasing writepos if diff is positive.
+	 * If 'readpos' is then behind 'writepos', there is no way that we can
 	 * ensure safety in the face of a crash - that must be done by userspace
 	 * making a backup of the data.  So in that case there is no particular
 	 * rush to update metadata.
···
 	 * Maybe that number should be configurable, but I'm not sure it is
 	 * worth it.... maybe it could be a multiple of safemode_delay???
 	 */
-	if ((mddev->delta_disks < 0
+	if (conf->min_offset_diff < 0) {
+		safepos += -conf->min_offset_diff;
+		readpos += -conf->min_offset_diff;
+	} else
+		writepos += conf->min_offset_diff;
+
+	if ((mddev->reshape_backwards
 	     ? (safepos > writepos && readpos < writepos)
 	     : (safepos < writepos && readpos > writepos)) ||
 	    time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
···
 		sysfs_notify(&mddev->kobj, NULL, "sync_completed");
 	}
 
-	if (mddev->delta_disks < 0) {
-		BUG_ON(conf->reshape_progress == 0);
-		stripe_addr = writepos;
-		BUG_ON((mddev->dev_sectors &
-			~((sector_t)reshape_sectors - 1))
-		       - reshape_sectors - stripe_addr
-		       != sector_nr);
-	} else {
-		BUG_ON(writepos != sector_nr + reshape_sectors);
-		stripe_addr = sector_nr;
-	}
 	INIT_LIST_HEAD(&stripes);
 	for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
 		int j;
···
 		list_add(&sh->lru, &stripes);
 	}
 	spin_lock_irq(&conf->device_lock);
-	if (mddev->delta_disks < 0)
+	if (mddev->reshape_backwards)
 		conf->reshape_progress -= reshape_sectors * new_data_disks;
 	else
 		conf->reshape_progress += reshape_sectors * new_data_disks;
···
 	struct md_rdev *rdev;
 	sector_t reshape_offset = 0;
 	int i;
+	long long min_offset_diff = 0;
+	int first = 1;
 
 	if (mddev->recovery_cp != MaxSector)
 		printk(KERN_NOTICE "md/raid:%s: not clean"
 		       " -- starting background reconstruction\n",
 		       mdname(mddev));
+
+	rdev_for_each(rdev, mddev) {
+		long long diff;
+		if (rdev->raid_disk < 0)
+			continue;
+		diff = (rdev->new_data_offset - rdev->data_offset);
+		if (first) {
+			min_offset_diff = diff;
+			first = 0;
+		} else if (mddev->reshape_backwards &&
+			 diff < min_offset_diff)
+			min_offset_diff = diff;
+		else if (!mddev->reshape_backwards &&
+			 diff > min_offset_diff)
+			min_offset_diff = diff;
+	}
+
 	if (mddev->reshape_position != MaxSector) {
 		/* Check that we can continue the reshape.
-		 * Currently only disks can change, it must
-		 * increase, and we must be past the point where
-		 * a stripe over-writes itself
+		 * Difficulties arise if the stripe we would write to
+		 * next is at or after the stripe we would read from next.
+		 * For a reshape that changes the number of devices, this
+		 * is only possible for a very short time, and mdadm makes
+		 * sure that time appears to have past before assembling
+		 * the array.  So we fail if that time hasn't passed.
+		 * For a reshape that keeps the number of devices the same
+		 * mdadm must be monitoring the reshape can keeping the
+		 * critical areas read-only and backed up.  It will start
+		 * the array in read-only mode, so we check for that.
 		 */
 		sector_t here_new, here_old;
 		int old_disks;
···
 		/* here_old is the first stripe that we might need to read
 		 * from */
 		if (mddev->delta_disks == 0) {
+			if ((here_new * mddev->new_chunk_sectors !=
+			     here_old * mddev->chunk_sectors)) {
+				printk(KERN_ERR "md/raid:%s: reshape position is"
+				       " confused - aborting\n", mdname(mddev));
+				return -EINVAL;
+			}
 			/* We cannot be sure it is safe to start an in-place
-			 * reshape.  It is only safe if user-space if monitoring
+			 * reshape.  It is only safe if user-space is monitoring
 			 * and taking constant backups.
 			 * mdadm always starts a situation like this in
 			 * readonly mode so it can take control before
 			 * allowing any writes.  So just check for that.
 			 */
-			if ((here_new * mddev->new_chunk_sectors != 
-			     here_old * mddev->chunk_sectors) ||
-			    mddev->ro == 0) {
-				printk(KERN_ERR "md/raid:%s: in-place reshape must be started"
-				       " in read-only mode - aborting\n",
+			if (abs(min_offset_diff) >= mddev->chunk_sectors &&
+			    abs(min_offset_diff) >= mddev->new_chunk_sectors)
+				/* not really in-place - so OK */;
+			else if (mddev->ro == 0) {
+				printk(KERN_ERR "md/raid:%s: in-place reshape "
+				       "must be started in read-only mode "
+				       "- aborting\n",
 				       mdname(mddev));
 				return -EINVAL;
 			}
-		} else if (mddev->delta_disks < 0
-		    ? (here_new * mddev->new_chunk_sectors <=
+		} else if (mddev->reshape_backwards
+		    ? (here_new * mddev->new_chunk_sectors + min_offset_diff <=
 		       here_old * mddev->chunk_sectors)
 		    : (here_new * mddev->new_chunk_sectors >=
-		       here_old * mddev->chunk_sectors)) {
+		       here_old * mddev->chunk_sectors + (-min_offset_diff))) {
 			/* Reading from the same stripe as writing to - bad */
 			printk(KERN_ERR "md/raid:%s: reshape_position too early for "
 			       "auto-recovery - aborting.\n",
···
 	if (IS_ERR(conf))
 		return PTR_ERR(conf);
 
+	conf->min_offset_diff = min_offset_diff;
 	mddev->thread = conf->thread;
 	conf->thread = NULL;
 	mddev->private = conf;
···
 		blk_queue_io_opt(mddev->queue, chunk_size *
 				 (conf->raid_disks - conf->max_degraded));
 
-		rdev_for_each(rdev, mddev)
+		rdev_for_each(rdev, mddev) {
 			disk_stack_limits(mddev->gendisk, rdev->bdev,
 					  rdev->data_offset << 9);
+			disk_stack_limits(mddev->gendisk, rdev->bdev,
+					  rdev->new_data_offset << 9);
+		}
 	}
 
 	return 0;
···
 	 * any io in the removed space completes, but it hardly seems
 	 * worth it.
 	 */
+	sector_t newsize;
 	sectors &= ~((sector_t)mddev->chunk_sectors - 1);
-	md_set_array_sectors(mddev, raid5_size(mddev, sectors,
-					       mddev->raid_disks));
-	if (mddev->array_sectors >
-	    raid5_size(mddev, sectors, mddev->raid_disks))
+	newsize = raid5_size(mddev, sectors, mddev->raid_disks);
+	if (mddev->external_size &&
+	    mddev->array_sectors > newsize)
 		return -EINVAL;
+	if (mddev->bitmap) {
+		int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
+		if (ret)
+			return ret;
+	}
+	md_set_array_sectors(mddev, newsize);
 	set_capacity(mddev->gendisk, mddev->array_sectors);
 	revalidate_disk(mddev->gendisk);
 	if (sectors > mddev->dev_sectors &&
···
 	    mddev->new_layout == mddev->layout &&
 	    mddev->new_chunk_sectors == mddev->chunk_sectors)
 		return 0; /* nothing to do */
-	if (mddev->bitmap)
-		/* Cannot grow a bitmap yet */
-		return -EBUSY;
 	if (has_failed(conf))
 		return -EINVAL;
 	if (mddev->delta_disks < 0) {
···
 	if (!check_stripe_cache(mddev))
 		return -ENOSPC;
 
-	rdev_for_each(rdev, mddev)
+	if (has_failed(conf))
+		return -EINVAL;
+
+	rdev_for_each(rdev, mddev) {
 		if (!test_bit(In_sync, &rdev->flags)
 		    && !test_bit(Faulty, &rdev->flags))
 			spares++;
+	}
 
 	if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
 		/* Not enough devices even to make a degraded array
···
 	conf->chunk_sectors = mddev->new_chunk_sectors;
 	conf->prev_algo = conf->algorithm;
 	conf->algorithm = mddev->new_layout;
-	if (mddev->delta_disks < 0)
+	conf->generation++;
+	/* Code that selects data_offset needs to see the generation update
+	 * if reshape_progress has been set - so a memory barrier needed.
+	 */
+	smp_mb();
+	if (mddev->reshape_backwards)
 		conf->reshape_progress = raid5_size(mddev, 0, 0);
 	else
 		conf->reshape_progress = 0;
 	conf->reshape_safe = conf->reshape_progress;
-	conf->generation++;
 	spin_unlock_irq(&conf->device_lock);
 
 	/* Add some new drives, as many as will fit.
···
 		mddev->recovery = 0;
 		spin_lock_irq(&conf->device_lock);
 		mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
+		rdev_for_each(rdev, mddev)
+			rdev->new_data_offset = rdev->data_offset;
+		smp_wmb();
 		conf->reshape_progress = MaxSector;
 		mddev->reshape_position = MaxSector;
 		spin_unlock_irq(&conf->device_lock);
···
 {
 
 	if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
+		struct md_rdev *rdev;
 
 		spin_lock_irq(&conf->device_lock);
 		conf->previous_raid_disks = conf->raid_disks;
+		rdev_for_each(rdev, conf->mddev)
+			rdev->data_offset = rdev->new_data_offset;
+		smp_wmb();
 		conf->reshape_progress = MaxSector;
 		spin_unlock_irq(&conf->device_lock);
 		wake_up(&conf->wait_for_overlap);
···
 			     d < conf->raid_disks - mddev->delta_disks;
 			     d++) {
 				struct md_rdev *rdev = conf->disks[d].rdev;
-				if (rdev &&
-				    raid5_remove_disk(mddev, rdev) == 0) {
-					sysfs_unlink_rdev(mddev, rdev);
-					rdev->raid_disk = -1;
-				}
+				if (rdev)
+					clear_bit(In_sync, &rdev->flags);
+				rdev = conf->disks[d].replacement;
+				if (rdev)
+					clear_bit(In_sync, &rdev->flags);
 			}
 		}
 		mddev->layout = conf->algorithm;
 		mddev->chunk_sectors = conf->chunk_sectors;
 		mddev->reshape_position = MaxSector;
 		mddev->delta_disks = 0;
+		mddev->reshape_backwards = 0;
 	}
 }
 

drivers/md/raid5.h

···
 			 */
 	R5_Wantdrain,	/* dev->towrite needs to be drained */
 	R5_WantFUA,	/* Write should be FUA */
+	R5_SyncIO,	/* The IO is sync */
 	R5_WriteError,	/* got a write error - need to record it */
 	R5_MadeGood,	/* A bad block has been fixed by writing to it */
 	R5_ReadRepl,	/* Will/did read from replacement rather than orig */
···
 	short			generation; /* increments with every reshape */
 	unsigned long		reshape_checkpoint; /* Time we last updated
 						     * metadata */
+	long long		min_offset_diff; /* minimum difference between
+						  * data_offset and
+						  * new_data_offset across all
+						  * devices.  May be negative,
+						  * but is closest to zero.
+						  */
 
 	struct list_head	handle_list; /* stripes needing handling */
 	struct list_head	hold_list; /* preread ready stripes */

include/linux/raid/md_p.h

···
 	__le32	delta_disks;	/* change in number of raid_disks		*/
 	__le32	new_layout;	/* new layout					*/
 	__le32	new_chunk;	/* new chunk size (512byte sectors)		*/
-	__u8	pad1[128-124];	/* set to 0 when written */
+	__le32  new_offset;	/* signed number to add to data_offset in new
+				 * layout.  0 == no-change.  This can be
+				 * different on each device in the array.
+				 */
 
 	/* constant this-device information - 64 bytes */
 	__le64	data_offset;	/* sector start of data, often 0 */
···
 					    * active device with same 'role'.
 					    * 'recovery_offset' is also set.
 					    */
+#define	MD_FEATURE_RESHAPE_BACKWARDS	32 /* Reshape doesn't change number
+					    * of devices, but is going
+					    * backwards anyway.
+					    */
+#define	MD_FEATURE_NEW_OFFSET		64 /* new_offset must be honoured */
 #define	MD_FEATURE_ALL			(MD_FEATURE_BITMAP_OFFSET	\
 					|MD_FEATURE_RECOVERY_OFFSET	\
 					|MD_FEATURE_RESHAPE_ACTIVE	\
 					|MD_FEATURE_BAD_BLOCKS		\
-					|MD_FEATURE_REPLACEMENT)
+					|MD_FEATURE_REPLACEMENT		\
+					|MD_FEATURE_RESHAPE_BACKWARDS	\
+					|MD_FEATURE_NEW_OFFSET		\
+					)
 
 #endif 

include/linux/raid/pq.h

···
 extern const struct raid6_calls raid6_altivec4;
 extern const struct raid6_calls raid6_altivec8;
 
+struct raid6_recov_calls {
+	void (*data2)(int, size_t, int, int, void **);
+	void (*datap)(int, size_t, int, void **);
+	int  (*valid)(void);
+	const char *name;
+	int priority;
+};
+
+extern const struct raid6_recov_calls raid6_recov_intx1;
+extern const struct raid6_recov_calls raid6_recov_ssse3;
+
 /* Algorithm list */
 extern const struct raid6_calls * const raid6_algos[];
+extern const struct raid6_recov_calls *const raid6_recov_algos[];
 int raid6_select_algo(void);
 
 /* Return values from chk_syndrome */
···
 
 /* Galois field tables */
 extern const u8 raid6_gfmul[256][256] __attribute__((aligned(256)));
+extern const u8 raid6_vgfmul[256][32] __attribute__((aligned(256)));
 extern const u8 raid6_gfexp[256]      __attribute__((aligned(256)));
 extern const u8 raid6_gfinv[256]      __attribute__((aligned(256)));
 extern const u8 raid6_gfexi[256]      __attribute__((aligned(256)));
 
 /* Recovery routines */
-void raid6_2data_recov(int disks, size_t bytes, int faila, int failb,
+extern void (*raid6_2data_recov)(int disks, size_t bytes, int faila, int failb,
 		       void **ptrs);
-void raid6_datap_recov(int disks, size_t bytes, int faila, void **ptrs);
+extern void (*raid6_datap_recov)(int disks, size_t bytes, int faila,
+			void **ptrs);
 void raid6_dual_recov(int disks, size_t bytes, int faila, int failb,
 		      void **ptrs);
 

lib/raid6/Makefile

···
 obj-$(CONFIG_RAID6_PQ)	+= raid6_pq.o
 
-raid6_pq-y	+= algos.o recov.o tables.o int1.o int2.o int4.o \
+raid6_pq-y	+= algos.o recov.o recov_ssse3.o tables.o int1.o int2.o int4.o \
 		   int8.o int16.o int32.o altivec1.o altivec2.o altivec4.o \
 		   altivec8.o mmx.o sse1.o sse2.o
 hostprogs-y	+= mktables

lib/raid6/algos.c

···
  */
 
 #include <linux/raid/pq.h>
-#include <linux/module.h>
 #ifndef __KERNEL__
 #include <sys/mman.h>
 #include <stdio.h>
 #else
+#include <linux/module.h>
 #include <linux/gfp.h>
 #if !RAID6_USE_EMPTY_ZERO_PAGE
 /* In .bss so it's zeroed */
···
 EXPORT_SYMBOL_GPL(raid6_call);
 
 const struct raid6_calls * const raid6_algos[] = {
-	&raid6_intx1,
-	&raid6_intx2,
-	&raid6_intx4,
-	&raid6_intx8,
 #if defined(__ia64__)
 	&raid6_intx16,
 	&raid6_intx32,
···
 	&raid6_altivec4,
 	&raid6_altivec8,
 #endif
+	&raid6_intx1,
+	&raid6_intx2,
+	&raid6_intx4,
+	&raid6_intx8,
+	NULL
+};
+
+void (*raid6_2data_recov)(int, size_t, int, int, void **);
+EXPORT_SYMBOL_GPL(raid6_2data_recov);
+
+void (*raid6_datap_recov)(int, size_t, int, void **);
+EXPORT_SYMBOL_GPL(raid6_datap_recov);
+
+const struct raid6_recov_calls *const raid6_recov_algos[] = {
+#if (defined(__i386__) || defined(__x86_64__)) && !defined(__arch_um__)
+	&raid6_recov_ssse3,
+#endif
+	&raid6_recov_intx1,
 	NULL
 };
 
···
 #define time_before(x, y) ((x) < (y))
 #endif
 
-/* Try to pick the best algorithm */
-/* This code uses the gfmul table as convenient data set to abuse */
-
-int __init raid6_select_algo(void)
+static inline const struct raid6_recov_calls *raid6_choose_recov(void)
 {
-	const struct raid6_calls * const * algo;
-	const struct raid6_calls * best;
-	char *syndromes;
-	void *dptrs[(65536/PAGE_SIZE)+2];
-	int i, disks;
-	unsigned long perf, bestperf;
-	int bestprefer;
-	unsigned long j0, j1;
+	const struct raid6_recov_calls *const *algo;
+	const struct raid6_recov_calls *best;
 
-	disks = (65536/PAGE_SIZE)+2;
-	for ( i = 0 ; i < disks-2 ; i++ ) {
-		dptrs[i] = ((char *)raid6_gfmul) + PAGE_SIZE*i;
-	}
+	for (best = NULL, algo = raid6_recov_algos; *algo; algo++)
+		if (!best || (*algo)->priority > best->priority)
+			if (!(*algo)->valid || (*algo)->valid())
+				best = *algo;
 
-	/* Normal code - use a 2-page allocation to avoid D$ conflict */
-	syndromes = (void *) __get_free_pages(GFP_KERNEL, 1);
+	if (best) {
+		raid6_2data_recov = best->data2;
+		raid6_datap_recov = best->datap;
 
-	if ( !syndromes ) {
-		printk("raid6: Yikes!  No memory available.\n");
-		return -ENOMEM;
-	}
+		printk("raid6: using %s recovery algorithm\n", best->name);
+	} else
+		printk("raid6: Yikes! No recovery algorithm found!\n");
 
-	dptrs[disks-2] = syndromes;
-	dptrs[disks-1] = syndromes + PAGE_SIZE;
+	return best;
+}
 
-	bestperf = 0;  bestprefer = 0;  best = NULL;
+static inline const struct raid6_calls *raid6_choose_gen(
+	void *(*const dptrs)[(65536/PAGE_SIZE)+2], const int disks)
+{
+	unsigned long perf, bestperf, j0, j1;
+	const struct raid6_calls *const *algo;
+	const struct raid6_calls *best;
 
-	for ( algo = raid6_algos ; *algo ; algo++ ) {
-		if ( !(*algo)->valid || (*algo)->valid() ) {
+	for (bestperf = 0, best = NULL, algo = raid6_algos; *algo; algo++) {
+		if (!best || (*algo)->prefer >= best->prefer) {
+			if ((*algo)->valid && !(*algo)->valid())
+				continue;
+
 			perf = 0;
 
 			preempt_disable();
 			j0 = jiffies;
-			while ( (j1 = jiffies) == j0 )
+			while ((j1 = jiffies) == j0)
 				cpu_relax();
 			while (time_before(jiffies,
 					    j1 + (1<<RAID6_TIME_JIFFIES_LG2))) {
-				(*algo)->gen_syndrome(disks, PAGE_SIZE, dptrs);
+				(*algo)->gen_syndrome(disks, PAGE_SIZE, *dptrs);
 				perf++;
 			}
 			preempt_enable();
 
-			if ( (*algo)->prefer > bestprefer ||
-			     ((*algo)->prefer == bestprefer &&
-			      perf > bestperf) ) {
-				best = *algo;
-				bestprefer = best->prefer;
+			if (perf > bestperf) {
 				bestperf = perf;
+				best = *algo;
 			}
 			printk("raid6: %-8s %5ld MB/s\n", (*algo)->name,
 			       (perf*HZ) >> (20-16+RAID6_TIME_JIFFIES_LG2));
···
 	} else
 		printk("raid6: Yikes!  No algorithm found!\n");
 
+	return best;
+}
+
+
+/* Try to pick the best algorithm */
+/* This code uses the gfmul table as convenient data set to abuse */
+
+int __init raid6_select_algo(void)
+{
+	const int disks = (65536/PAGE_SIZE)+2;
+
+	const struct raid6_calls *gen_best;
+	const struct raid6_recov_calls *rec_best;
+	char *syndromes;
+	void *dptrs[(65536/PAGE_SIZE)+2];
+	int i;
+
+	for (i = 0; i < disks-2; i++)
+		dptrs[i] = ((char *)raid6_gfmul) + PAGE_SIZE*i;
+
+	/* Normal code - use a 2-page allocation to avoid D$ conflict */
+	syndromes = (void *) __get_free_pages(GFP_KERNEL, 1);
+
+	if (!syndromes) {
+		printk("raid6: Yikes!  No memory available.\n");
+		return -ENOMEM;
+	}
+
+	dptrs[disks-2] = syndromes;
+	dptrs[disks-1] = syndromes + PAGE_SIZE;
+
+	/* select raid gen_syndrome function */
+	gen_best = raid6_choose_gen(&dptrs, disks);
+
+	/* select raid recover functions */
+	rec_best = raid6_choose_recov();
+
 	free_pages((unsigned long)syndromes, 1);
 
-	return best ? 0 : -EINVAL;
+	return gen_best && rec_best ? 0 : -EINVAL;
 }
 
 static void raid6_exit(void)

lib/raid6/mktables.c

···
 	printf("EXPORT_SYMBOL(raid6_gfmul);\n");
 	printf("#endif\n");
 
+	/* Compute vector multiplication table */
+	printf("\nconst u8  __attribute__((aligned(256)))\n"
+		"raid6_vgfmul[256][32] =\n"
+		"{\n");
+	for (i = 0; i < 256; i++) {
+		printf("\t{\n");
+		for (j = 0; j < 16; j += 8) {
+			printf("\t\t");
+			for (k = 0; k < 8; k++)
+				printf("0x%02x,%c", gfmul(i, j + k),
+				       (k == 7) ? '\n' : ' ');
+		}
+		for (j = 0; j < 16; j += 8) {
+			printf("\t\t");
+			for (k = 0; k < 8; k++)
+				printf("0x%02x,%c", gfmul(i, (j + k) << 4),
+				       (k == 7) ? '\n' : ' ');
+		}
+		printf("\t},\n");
+	}
+	printf("};\n");
+	printf("#ifdef __KERNEL__\n");
+	printf("EXPORT_SYMBOL(raid6_vgfmul);\n");
+	printf("#endif\n");
+
 	/* Compute power-of-2 table (exponent) */
 	v = 1;
 	printf("\nconst u8 __attribute__((aligned(256)))\n"

lib/raid6/recov.c

···
 #include <linux/raid/pq.h>
 
 /* Recover two failed data blocks. */
-void raid6_2data_recov(int disks, size_t bytes, int faila, int failb,
+void raid6_2data_recov_intx1(int disks, size_t bytes, int faila, int failb,
 		       void **ptrs)
 {
 	u8 *p, *q, *dp, *dq;
···
 		p++; q++;
 	}
 }
-EXPORT_SYMBOL_GPL(raid6_2data_recov);
 
 /* Recover failure of one data block plus the P block */
-void raid6_datap_recov(int disks, size_t bytes, int faila, void **ptrs)
+void raid6_datap_recov_intx1(int disks, size_t bytes, int faila, void **ptrs)
 {
 	u8 *p, *q, *dq;
 	const u8 *qmul;		/* Q multiplier table */
···
 		q++; dq++;
 	}
 }
-EXPORT_SYMBOL_GPL(raid6_datap_recov);
+
+
+const struct raid6_recov_calls raid6_recov_intx1 = {
+	.data2 = raid6_2data_recov_intx1,
+	.datap = raid6_datap_recov_intx1,
+	.valid = NULL,
+	.name = "intx1",
+	.priority = 0,
+};
 
 #ifndef __KERNEL__
 /* Testing only */

lib/raid6/recov_ssse3.c

···
+/*
+ * Copyright (C) 2012 Intel Corporation
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License
+ * as published by the Free Software Foundation; version 2
+ * of the License.
+ */
+
+#if (defined(__i386__) || defined(__x86_64__)) && !defined(__arch_um__)
+
+#include <linux/raid/pq.h>
+#include "x86.h"
+
+static int raid6_has_ssse3(void)
+{
+	return boot_cpu_has(X86_FEATURE_XMM) &&
+		boot_cpu_has(X86_FEATURE_XMM2) &&
+		boot_cpu_has(X86_FEATURE_SSSE3);
+}
+
+void raid6_2data_recov_ssse3(int disks, size_t bytes, int faila, int failb,
+		       void **ptrs)
+{
+	u8 *p, *q, *dp, *dq;
+	const u8 *pbmul;	/* P multiplier table for B data */
+	const u8 *qmul;		/* Q multiplier table (for both) */
+	static const u8 __aligned(16) x0f[16] = {
+		 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f,
+		 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f};
+
+	p = (u8 *)ptrs[disks-2];
+	q = (u8 *)ptrs[disks-1];
+
+	/* Compute syndrome with zero for the missing data pages
+	   Use the dead data pages as temporary storage for
+	   delta p and delta q */
+	dp = (u8 *)ptrs[faila];
+	ptrs[faila] = (void *)raid6_empty_zero_page;
+	ptrs[disks-2] = dp;
+	dq = (u8 *)ptrs[failb];
+	ptrs[failb] = (void *)raid6_empty_zero_page;
+	ptrs[disks-1] = dq;
+
+	raid6_call.gen_syndrome(disks, bytes, ptrs);
+
+	/* Restore pointer table */
+	ptrs[faila]   = dp;
+	ptrs[failb]   = dq;
+	ptrs[disks-2] = p;
+	ptrs[disks-1] = q;
+
+	/* Now, pick the proper data tables */
+	pbmul = raid6_vgfmul[raid6_gfexi[failb-faila]];
+	qmul  = raid6_vgfmul[raid6_gfinv[raid6_gfexp[faila] ^
+		raid6_gfexp[failb]]];
+
+	kernel_fpu_begin();
+
+	asm volatile("movdqa %0,%%xmm7" : : "m" (x0f[0]));
+
+#ifdef CONFIG_X86_64
+	asm volatile("movdqa %0,%%xmm6" : : "m" (qmul[0]));
+	asm volatile("movdqa %0,%%xmm14" : : "m" (pbmul[0]));
+	asm volatile("movdqa %0,%%xmm15" : : "m" (pbmul[16]));
+#endif
+
+	/* Now do it... */
+	while (bytes) {
+#ifdef CONFIG_X86_64
+		/* xmm6, xmm14, xmm15 */
+
+		asm volatile("movdqa %0,%%xmm1" : : "m" (q[0]));
+		asm volatile("movdqa %0,%%xmm9" : : "m" (q[16]));
+		asm volatile("movdqa %0,%%xmm0" : : "m" (p[0]));
+		asm volatile("movdqa %0,%%xmm8" : : "m" (p[16]));
+		asm volatile("pxor   %0,%%xmm1" : : "m" (dq[0]));
+		asm volatile("pxor   %0,%%xmm9" : : "m" (dq[16]));
+		asm volatile("pxor   %0,%%xmm0" : : "m" (dp[0]));
+		asm volatile("pxor   %0,%%xmm8" : : "m" (dp[16]));
+
+		/* xmm0/8 = px */
+
+		asm volatile("movdqa %xmm6,%xmm4");
+		asm volatile("movdqa %0,%%xmm5" : : "m" (qmul[16]));
+		asm volatile("movdqa %xmm6,%xmm12");
+		asm volatile("movdqa %xmm5,%xmm13");
+		asm volatile("movdqa %xmm1,%xmm3");
+		asm volatile("movdqa %xmm9,%xmm11");
+		asm volatile("movdqa %xmm0,%xmm2"); /* xmm2/10 = px */
+		asm volatile("movdqa %xmm8,%xmm10");
+		asm volatile("psraw  $4,%xmm1");
+		asm volatile("psraw  $4,%xmm9");
+		asm volatile("pand   %xmm7,%xmm3");
+		asm volatile("pand   %xmm7,%xmm11");
+		asm volatile("pand   %xmm7,%xmm1");
+		asm volatile("pand   %xmm7,%xmm9");
+		asm volatile("pshufb %xmm3,%xmm4");
+		asm volatile("pshufb %xmm11,%xmm12");
+		asm volatile("pshufb %xmm1,%xmm5");
+		asm volatile("pshufb %xmm9,%xmm13");
+		asm volatile("pxor   %xmm4,%xmm5");
+		asm volatile("pxor   %xmm12,%xmm13");
+
+		/* xmm5/13 = qx */
+
+		asm volatile("movdqa %xmm14,%xmm4");
+		asm volatile("movdqa %xmm15,%xmm1");
+		asm volatile("movdqa %xmm14,%xmm12");
+		asm volatile("movdqa %xmm15,%xmm9");
+		asm volatile("movdqa %xmm2,%xmm3");
+		asm volatile("movdqa %xmm10,%xmm11");
+		asm volatile("psraw  $4,%xmm2");
+		asm volatile("psraw  $4,%xmm10");
+		asm volatile("pand   %xmm7,%xmm3");
+		asm volatile("pand   %xmm7,%xmm11");
+		asm volatile("pand   %xmm7,%xmm2");
+		asm volatile("pand   %xmm7,%xmm10");
+		asm volatile("pshufb %xmm3,%xmm4");
+		asm volatile("pshufb %xmm11,%xmm12");
+		asm volatile("pshufb %xmm2,%xmm1");
+		asm volatile("pshufb %xmm10,%xmm9");
+		asm volatile("pxor   %xmm4,%xmm1");
+		asm volatile("pxor   %xmm12,%xmm9");
+
+		/* xmm1/9 = pbmul[px] */
+		asm volatile("pxor   %xmm5,%xmm1");
+		asm volatile("pxor   %xmm13,%xmm9");
+		/* xmm1/9 = db = DQ */
+		asm volatile("movdqa %%xmm1,%0" : "=m" (dq[0]));
+		asm volatile("movdqa %%xmm9,%0" : "=m" (dq[16]));
+
+		asm volatile("pxor   %xmm1,%xmm0");
+		asm volatile("pxor   %xmm9,%xmm8");
+		asm volatile("movdqa %%xmm0,%0" : "=m" (dp[0]));
+		asm volatile("movdqa %%xmm8,%0" : "=m" (dp[16]));
+
+		bytes -= 32;
+		p += 32;
+		q += 32;
+		dp += 32;
+		dq += 32;
+#else
+		asm volatile("movdqa %0,%%xmm1" : : "m" (*q));
+		asm volatile("movdqa %0,%%xmm0" : : "m" (*p));
+		asm volatile("pxor   %0,%%xmm1" : : "m" (*dq));
+		asm volatile("pxor   %0,%%xmm0" : : "m" (*dp));
+
+		/* 1 = dq ^ q
+		 * 0 = dp ^ p
+		 */
+		asm volatile("movdqa %0,%%xmm4" : : "m" (qmul[0]));
+		asm volatile("movdqa %0,%%xmm5" : : "m" (qmul[16]));
+
+		asm volatile("movdqa %xmm1,%xmm3");
+		asm volatile("psraw  $4,%xmm1");
+		asm volatile("pand   %xmm7,%xmm3");
+		asm volatile("pand   %xmm7,%xmm1");
+		asm volatile("pshufb %xmm3,%xmm4");
+		asm volatile("pshufb %xmm1,%xmm5");
+		asm volatile("pxor   %xmm4,%xmm5");
+
+		asm volatile("movdqa %xmm0,%xmm2"); /* xmm2 = px */
+
+		/* xmm5 = qx */
+
+		asm volatile("movdqa %0,%%xmm4" : : "m" (pbmul[0]));
+		asm volatile("movdqa %0,%%xmm1" : : "m" (pbmul[16]));
+		asm volatile("movdqa %xmm2,%xmm3");
+		asm volatile("psraw  $4,%xmm2");
+		asm volatile("pand   %xmm7,%xmm3");
+		asm volatile("pand   %xmm7,%xmm2");
+		asm volatile("pshufb %xmm3,%xmm4");
+		asm volatile("pshufb %xmm2,%xmm1");
+		asm volatile("pxor   %xmm4,%xmm1");
+
+		/* xmm1 = pbmul[px] */
+		asm volatile("pxor   %xmm5,%xmm1");
+		/* xmm1 = db = DQ */
+		asm volatile("movdqa %%xmm1,%0" : "=m" (*dq));
+
+		asm volatile("pxor   %xmm1,%xmm0");
+		asm volatile("movdqa %%xmm0,%0" : "=m" (*dp));
+
+		bytes -= 16;
+		p += 16;
+		q += 16;
+		dp += 16;
+		dq += 16;
+#endif
+	}
+
+	kernel_fpu_end();
+}
+
+
+void raid6_datap_recov_ssse3(int disks, size_t bytes, int faila, void **ptrs)
+{
+	u8 *p, *q, *dq;
+	const u8 *qmul;		/* Q multiplier table */
+	static const u8 __aligned(16) x0f[16] = {
+		 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f,
+		 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f};
+
+	p = (u8 *)ptrs[disks-2];
+	q = (u8 *)ptrs[disks-1];
+
+	/* Compute syndrome with zero for the missing data page
+	   Use the dead data page as temporary storage for delta q */
+	dq = (u8 *)ptrs[faila];
+	ptrs[faila] = (void *)raid6_empty_zero_page;
+	ptrs[disks-1] = dq;
+
+	raid6_call.gen_syndrome(disks, bytes, ptrs);
+
+	/* Restore pointer table */
+	ptrs[faila]   = dq;
+	ptrs[disks-1] = q;
+
+	/* Now, pick the proper data tables */
+	qmul  = raid6_vgfmul[raid6_gfinv[raid6_gfexp[faila]]];
+
+	kernel_fpu_begin();
+
+	asm volatile("movdqa %0, %%xmm7" : : "m" (x0f[0]));
+
+	while (bytes) {
+#ifdef CONFIG_X86_64
+		asm volatile("movdqa %0, %%xmm3" : : "m" (dq[0]));
+		asm volatile("movdqa %0, %%xmm4" : : "m" (dq[16]));
+		asm volatile("pxor %0, %%xmm3" : : "m" (q[0]));
+		asm volatile("movdqa %0, %%xmm0" : : "m" (qmul[0]));
+
+		/* xmm3 = q[0] ^ dq[0] */
+
+		asm volatile("pxor %0, %%xmm4" : : "m" (q[16]));
+		asm volatile("movdqa %0, %%xmm1" : : "m" (qmul[16]));
+
+		/* xmm4 = q[16] ^ dq[16] */
+
+		asm volatile("movdqa %xmm3, %xmm6");
+		asm volatile("movdqa %xmm4, %xmm8");
+
+		/* xmm4 = xmm8 = q[16] ^ dq[16] */
+
+		asm volatile("psraw $4, %xmm3");
+		asm volatile("pand %xmm7, %xmm6");
+		asm volatile("pand %xmm7, %xmm3");
+		asm volatile("pshufb %xmm6, %xmm0");
+		asm volatile("pshufb %xmm3, %xmm1");
+		asm volatile("movdqa %0, %%xmm10" : : "m" (qmul[0]));
+		asm volatile("pxor %xmm0, %xmm1");
+		asm volatile("movdqa %0, %%xmm11" : : "m" (qmul[16]));
+
+		/* xmm1 = qmul[q[0] ^ dq[0]] */
+
+		asm volatile("psraw $4, %xmm4");
+		asm volatile("pand %xmm7, %xmm8");
+		asm volatile("pand %xmm7, %xmm4");
+		asm volatile("pshufb %xmm8, %xmm10");
+		asm volatile("pshufb %xmm4, %xmm11");
+		asm volatile("movdqa %0, %%xmm2" : : "m" (p[0]));
+		asm volatile("pxor %xmm10, %xmm11");
+		asm volatile("movdqa %0, %%xmm12" : : "m" (p[16]));
+
+		/* xmm11 = qmul[q[16] ^ dq[16]] */
+
+		asm volatile("pxor %xmm1, %xmm2");
+
+		/* xmm2 = p[0] ^ qmul[q[0] ^ dq[0]] */
+
+		asm volatile("pxor %xmm11, %xmm12");
+
+		/* xmm12 = p[16] ^ qmul[q[16] ^ dq[16]] */
+
+		asm volatile("movdqa %%xmm1, %0" : "=m" (dq[0]));
+		asm volatile("movdqa %%xmm11, %0" : "=m" (dq[16]));
+
+		asm volatile("movdqa %%xmm2, %0" : "=m" (p[0]));
+		asm volatile("movdqa %%xmm12, %0" : "=m" (p[16]));
+
+		bytes -= 32;
+		p += 32;
+		q += 32;
+		dq += 32;
+
+#else
+		asm volatile("movdqa %0, %%xmm3" : : "m" (dq[0]));
+		asm volatile("movdqa %0, %%xmm0" : : "m" (qmul[0]));
+		asm volatile("pxor %0, %%xmm3" : : "m" (q[0]));
+		asm volatile("movdqa %0, %%xmm1" : : "m" (qmul[16]));
+
+		/* xmm3 = *q ^ *dq */
+
+		asm volatile("movdqa %xmm3, %xmm6");
+		asm volatile("movdqa %0, %%xmm2" : : "m" (p[0]));
+		asm volatile("psraw $4, %xmm3");
+		asm volatile("pand %xmm7, %xmm6");
+		asm volatile("pand %xmm7, %xmm3");
+		asm volatile("pshufb %xmm6, %xmm0");
+		asm volatile("pshufb %xmm3, %xmm1");
+		asm volatile("pxor %xmm0, %xmm1");
+
+		/* xmm1 = qmul[*q ^ *dq */
+
+		asm volatile("pxor %xmm1, %xmm2");
+
+		/* xmm2 = *p ^ qmul[*q ^ *dq] */
+
+		asm volatile("movdqa %%xmm1, %0" : "=m" (dq[0]));
+		asm volatile("movdqa %%xmm2, %0" : "=m" (p[0]));
+
+		bytes -= 16;
+		p += 16;
+		q += 16;
+		dq += 16;
+#endif
+	}
+
+	kernel_fpu_end();
+}
+
+const struct raid6_recov_calls raid6_recov_ssse3 = {
+	.data2 = raid6_2data_recov_ssse3,
+	.datap = raid6_datap_recov_ssse3,
+	.valid = raid6_has_ssse3,
+#ifdef CONFIG_X86_64
+	.name = "ssse3x2",
+#else
+	.name = "ssse3x1",
+#endif
+	.priority = 1,
+};
+
+#endif

lib/raid6/test/Makefile

···
 all:	raid6.a raid6test
 
 raid6.a: int1.o int2.o int4.o int8.o int16.o int32.o mmx.o sse1.o sse2.o \
-	 altivec1.o altivec2.o altivec4.o altivec8.o recov.o algos.o \
+	 altivec1.o altivec2.o altivec4.o altivec8.o recov.o recov_ssse3.o algos.o \
 	 tables.o
 	 rm -f $@
 	 $(AR) cq $@ $^

lib/raid6/test/test.c

···
 int main(int argc, char *argv[])
 {
 	const struct raid6_calls *const *algo;
+	const struct raid6_recov_calls *const *ra;
 	int i, j;
 	int err = 0;
 
 	makedata();
 
-	for (algo = raid6_algos; *algo; algo++) {
-		if (!(*algo)->valid || (*algo)->valid()) {
-			raid6_call = **algo;
+	for (ra = raid6_recov_algos; *ra; ra++) {
+		if ((*ra)->valid  && !(*ra)->valid())
+			continue;
+		raid6_2data_recov = (*ra)->data2;
+		raid6_datap_recov = (*ra)->datap;
 
-			/* Nuke syndromes */
-			memset(data[NDISKS-2], 0xee, 2*PAGE_SIZE);
+		printf("using recovery %s\n", (*ra)->name);
 
-			/* Generate assumed good syndrome */
-			raid6_call.gen_syndrome(NDISKS, PAGE_SIZE,
-						(void **)&dataptrs);
+		for (algo = raid6_algos; *algo; algo++) {
+			if (!(*algo)->valid || (*algo)->valid()) {
+				raid6_call = **algo;
 
-			for (i = 0; i < NDISKS-1; i++)
-				for (j = i+1; j < NDISKS; j++)
-					err += test_disks(i, j);
+				/* Nuke syndromes */
+				memset(data[NDISKS-2], 0xee, 2*PAGE_SIZE);
+
+				/* Generate assumed good syndrome */
+				raid6_call.gen_syndrome(NDISKS, PAGE_SIZE,
+							(void **)&dataptrs);
+
+				for (i = 0; i < NDISKS-1; i++)
+					for (j = i+1; j < NDISKS; j++)
+						err += test_disks(i, j);
+			}
 		}
 		printf("\n");
 	}

lib/raid6/x86.h

···
 {
 }
 
+#define __aligned(x) __attribute__((aligned(x)))
+
 #define X86_FEATURE_MMX		(0*32+23) /* Multimedia Extensions */
 #define X86_FEATURE_FXSR	(0*32+24) /* FXSAVE and FXRSTOR instructions
 					   * (fast save and restore) */
 #define X86_FEATURE_XMM		(0*32+25) /* Streaming SIMD Extensions */
 #define X86_FEATURE_XMM2	(0*32+26) /* Streaming SIMD Extensions-2 */
+#define X86_FEATURE_XMM3	(4*32+ 0) /* "pni" SSE-3 */
+#define X86_FEATURE_SSSE3	(4*32+ 9) /* Supplemental SSE-3 */
+#define X86_FEATURE_AVX	(4*32+28) /* Advanced Vector Extensions */
 #define X86_FEATURE_MMXEXT	(1*32+22) /* AMD MMX extensions */
 
 /* Should work well enough on modern CPUs for testing */
 static inline int boot_cpu_has(int flag)
 {
-	u32 eax = (flag >> 5) ? 0x80000001 : 1;
-	u32 edx;
+	u32 eax = (flag & 0x20) ? 0x80000001 : 1;
+	u32 ecx, edx;
 
 	asm volatile("cpuid"
-		     : "+a" (eax), "=d" (edx)
-		     : : "ecx", "ebx");
+		     : "+a" (eax), "=d" (edx), "=c" (ecx)
+		     : : "ebx");
 
-	return (edx >> (flag & 31)) & 1;
+	return ((flag & 0x80 ? ecx : edx) >> (flag & 31)) & 1;
 }
 
 #endif /* ndef __KERNEL__ */