MIPS: Add partial 32-bit huge page support

This adds initial support for huge pages to 32-bit MIPS systems.
Systems with extended addressing enabled (EVA,XPA,Alchemy/Netlogic)
are not yet supported.
With huge pages enabled, this implementation will increase page table
memory overhead to match that of a 64-bit MIPS system. However, the
cache-friendliness of page table walks is not affected significantly.

Signed-off-by: Daniel Silsby <[e-mail]>
Signed-off-by: Paul Cercueil <[e-mail]>
Signed-off-by: Paul Burton <[e-mail]>
Cc: Ralf Baechle <[e-mail]>
Cc: James Hogan <[e-mail]>
Cc: [e-mail]
Cc: [e-mail]
Cc: [e-mail]

arch/mips/include/asm/pgtable-32.h

···
 #include <asm/highmem.h>
 #endif
 
+/*
+ * Regarding 32-bit MIPS huge page support (and the tradeoff it entails):
+ *
+ *  We use the same huge page sizes as 64-bit MIPS. Assuming a 4KB page size,
+ * our 2-level table layout would normally have a PGD entry cover a contiguous
+ * 4MB virtual address region (pointing to a 4KB PTE page of 1,024 32-bit pte_t
+ * pointers, each pointing to a 4KB physical page). The problem is that 4MB,
+ * spanning both halves of a TLB EntryLo0,1 pair, requires 2MB hardware page
+ * support, not one of the standard supported sizes (1MB,4MB,16MB,...).
+ *  To correct for this, when huge pages are enabled, we halve the number of
+ * pointers a PTE page holds, making its last half go to waste. Correspondingly,
+ * we double the number of PGD pages. Overall, page table memory overhead
+ * increases to match 64-bit MIPS, but PTE lookups remain CPU cache-friendly.
+ *
+ * NOTE: We don't yet support huge pages if extended-addressing is enabled
+ *       (i.e. EVA, XPA, 36-bit Alchemy/Netlogic).
+ */
+
 extern int temp_tlb_entry;
 
 /*
···
  */
 
 /* PGDIR_SHIFT determines what a third-level page table entry can map */
-#define PGDIR_SHIFT	(2 * PAGE_SHIFT + PTE_ORDER - PTE_T_LOG2)
+#if defined(CONFIG_MIPS_HUGE_TLB_SUPPORT) && !defined(CONFIG_PHYS_ADDR_T_64BIT)
+# define PGDIR_SHIFT	(2 * PAGE_SHIFT + PTE_ORDER - PTE_T_LOG2 - 1)
+#else
+# define PGDIR_SHIFT	(2 * PAGE_SHIFT + PTE_ORDER - PTE_T_LOG2)
+#endif
+
 #define PGDIR_SIZE	(1UL << PGDIR_SHIFT)
 #define PGDIR_MASK	(~(PGDIR_SIZE-1))
 
···
  * Entries per page directory level: we use two-level, so
  * we don't really have any PUD/PMD directory physically.
  */
-#define __PGD_ORDER	(32 - 3 * PAGE_SHIFT + PGD_T_LOG2 + PTE_T_LOG2)
+#if defined(CONFIG_MIPS_HUGE_TLB_SUPPORT) && !defined(CONFIG_PHYS_ADDR_T_64BIT)
+# define __PGD_ORDER	(32 - 3 * PAGE_SHIFT + PGD_T_LOG2 + PTE_T_LOG2 + 1)
+#else
+# define __PGD_ORDER	(32 - 3 * PAGE_SHIFT + PGD_T_LOG2 + PTE_T_LOG2)
+#endif
+
 #define PGD_ORDER	(__PGD_ORDER >= 0 ? __PGD_ORDER : 0)
 #define PUD_ORDER	aieeee_attempt_to_allocate_pud
 #define PMD_ORDER	1
 #define PTE_ORDER	0
 
 #define PTRS_PER_PGD	(USER_PTRS_PER_PGD * 2)
-#define PTRS_PER_PTE	((PAGE_SIZE << PTE_ORDER) / sizeof(pte_t))
+#if defined(CONFIG_MIPS_HUGE_TLB_SUPPORT) && !defined(CONFIG_PHYS_ADDR_T_64BIT)
+# define PTRS_PER_PTE	((PAGE_SIZE << PTE_ORDER) / sizeof(pte_t) / 2)
+#else
+# define PTRS_PER_PTE	((PAGE_SIZE << PTE_ORDER) / sizeof(pte_t))
+#endif
 
 #define USER_PTRS_PER_PGD	(0x80000000UL/PGDIR_SIZE)
 #define FIRST_USER_ADDRESS	0UL
···
 
 extern void load_pgd(unsigned long pg_dir);
 
-extern pte_t invalid_pte_table[PAGE_SIZE/sizeof(pte_t)];
+extern pte_t invalid_pte_table[PTRS_PER_PTE];
 
 /*
  * Empty pgd/pmd entries point to the invalid_pte_table.
···
 	return pmd_val(pmd) == (unsigned long) invalid_pte_table;
 }
 
-#define pmd_bad(pmd)		(pmd_val(pmd) & ~PAGE_MASK)
+static inline int pmd_bad(pmd_t pmd)
+{
+#ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
+	/* pmd_huge(pmd) but inline */
+	if (unlikely(pmd_val(pmd) & _PAGE_HUGE))
+		return 0;
+#endif
+
+	if (unlikely(pmd_val(pmd) & ~PAGE_MASK))
+		return 1;
+
+	return 0;
+}
 
 static inline int pmd_present(pmd_t pmd)
 {
···
 #else
 #define pte_pfn(x)		((unsigned long)((x).pte >> _PFN_SHIFT))
 #define pfn_pte(pfn, prot)	__pte(((unsigned long long)(pfn) << _PFN_SHIFT) | pgprot_val(prot))
+#define pfn_pmd(pfn, prot)	__pmd(((unsigned long long)(pfn) << _PFN_SHIFT) | pgprot_val(prot))
 #endif
 #endif /* defined(CONFIG_PHYS_ADDR_T_64BIT) && defined(CONFIG_CPU_MIPS32) */
 
···
 #define pgd_offset_k(address) pgd_offset(&init_mm, address)
 
 #define pgd_index(address)	(((address) >> PGDIR_SHIFT) & (PTRS_PER_PGD-1))
+#define pmd_index(address)	(((address) >> PMD_SHIFT) & (PTRS_PER_PMD-1))
 
 /* to find an entry in a page-table-directory */
 #define pgd_offset(mm, addr)	((mm)->pgd + pgd_index(addr))

arch/mips/include/asm/pgtable-bits.h

···
 	_PAGE_WRITE_SHIFT,
 	_PAGE_ACCESSED_SHIFT,
 	_PAGE_MODIFIED_SHIFT,
-#if defined(CONFIG_64BIT) && defined(CONFIG_MIPS_HUGE_TLB_SUPPORT)
+#if defined(CONFIG_MIPS_HUGE_TLB_SUPPORT)
 	_PAGE_HUGE_SHIFT,
 #endif
 
···
 #define _PAGE_WRITE		(1 << _PAGE_WRITE_SHIFT)
 #define _PAGE_ACCESSED		(1 << _PAGE_ACCESSED_SHIFT)
 #define _PAGE_MODIFIED		(1 << _PAGE_MODIFIED_SHIFT)
-#if defined(CONFIG_64BIT) && defined(CONFIG_MIPS_HUGE_TLB_SUPPORT)
+#if defined(CONFIG_MIPS_HUGE_TLB_SUPPORT)
 # define _PAGE_HUGE		(1 << _PAGE_HUGE_SHIFT)
 #endif
 

arch/mips/mm/pgtable-32.c

···
 #include <asm/fixmap.h>
 #include <asm/pgtable.h>
 #include <asm/pgalloc.h>
+#include <asm/tlbflush.h>
 
 void pgd_init(unsigned long page)
 {
···
 		p[i + 7] = (unsigned long) invalid_pte_table;
 	}
 }
+
+#if defined(CONFIG_TRANSPARENT_HUGEPAGE)
+pmd_t mk_pmd(struct page *page, pgprot_t prot)
+{
+	pmd_t pmd;
+
+	pmd_val(pmd) = (page_to_pfn(page) << _PFN_SHIFT) | pgprot_val(prot);
+
+	return pmd;
+}
+
+
+void set_pmd_at(struct mm_struct *mm, unsigned long addr,
+		pmd_t *pmdp, pmd_t pmd)
+{
+	*pmdp = pmd;
+	flush_tlb_all();
+}
+#endif /* defined(CONFIG_TRANSPARENT_HUGEPAGE) */
 
 void __init pagetable_init(void)
 {