ARC: dma [IOC] Enable per device io coherency

So far the IOC treatment was global on ARC, being turned on (or off)
for all devices in the system. With this patch, this can now be done
per device using the "dma-coherent" DT property; IOW with this patch
we can use both HW-coherent and regular DMA peripherals simultaneously.

The changes involved are too many so enlisting the summary below:

1. common code calls ARC arch_setup_dma_ops() per device.

2. For coherent dma (IOC) it plugs in generic @dma_direct_ops which
doesn't need any arch specific backend: No need for any explicit
cache flushes or MMU mappings to provide for uncached access

- dma_(map|sync)_single* return early as corresponding dma ops callbacks
are NULL in generic code.
So arch_sync_dma_*() -> dma_cache_*() need not handle the coherent
dma case, hence drop ARC __dma_cache_*_ioc() which were no-op anyways

3. For noncoherent dma (non IOC) generic @dma_noncoherent_ops is used
which in turns calls ARC specific routines

- arch_dma_alloc() no longer checks for @ioc_enable since this is
called only for !IOC case.

Reviewed-by: Christoph Hellwig <[e-mail]>
Signed-off-by: Eugeniy Paltsev <[e-mail]>
Signed-off-by: Vineet Gupta <[e-mail]>
[vgupta: rewrote changelog]

arch/arc/include/asm/dma-mapping.h

···
+// SPDX-License-Identifier:  GPL-2.0
+// (C) 2018 Synopsys, Inc. (www.synopsys.com)
+
+#ifndef ASM_ARC_DMA_MAPPING_H
+#define ASM_ARC_DMA_MAPPING_H
+
+#include <asm-generic/dma-mapping.h>
+
+void arch_setup_dma_ops(struct device *dev, u64 dma_base, u64 size,
+			const struct iommu_ops *iommu, bool coherent);
+#define arch_setup_dma_ops arch_setup_dma_ops
+
+#endif

arch/arc/mm/cache.c

···
 
 	n += scnprintf(buf + n, len - n, "Peripherals\t: %#lx%s%s\n",
 		       perip_base,
-		       IS_AVAIL3(ioc_exists, ioc_enable, ", IO-Coherency "));
+		       IS_AVAIL3(ioc_exists, ioc_enable, ", IO-Coherency (per-device) "));
 
 	return buf;
 }
···
 }
 
 /*
- * DMA ops for systems with IOC
- * IOC hardware snoops all DMA traffic keeping the caches consistent with
- * memory - eliding need for any explicit cache maintenance of DMA buffers
- */
-static void __dma_cache_wback_inv_ioc(phys_addr_t start, unsigned long sz) {}
-static void __dma_cache_inv_ioc(phys_addr_t start, unsigned long sz) {}
-static void __dma_cache_wback_ioc(phys_addr_t start, unsigned long sz) {}
-
-/*
  * Exported DMA API
  */
 void dma_cache_wback_inv(phys_addr_t start, unsigned long sz)
···
 	if (is_isa_arcv2() && ioc_enable)
 		arc_ioc_setup();
 
-	if (is_isa_arcv2() && ioc_enable) {
-		__dma_cache_wback_inv = __dma_cache_wback_inv_ioc;
-		__dma_cache_inv = __dma_cache_inv_ioc;
-		__dma_cache_wback = __dma_cache_wback_ioc;
-	} else if (is_isa_arcv2() && l2_line_sz && slc_enable) {
+	if (is_isa_arcv2() && l2_line_sz && slc_enable) {
 		__dma_cache_wback_inv = __dma_cache_wback_inv_slc;
 		__dma_cache_inv = __dma_cache_inv_slc;
 		__dma_cache_wback = __dma_cache_wback_slc;
···
 		__dma_cache_inv = __dma_cache_inv_l1;
 		__dma_cache_wback = __dma_cache_wback_l1;
 	}
+	/*
+	 * In case of IOC (say IOC+SLC case), pointers above could still be set
+	 * but end up not being relevant as the first function in chain is not
+	 * called at all for @dma_direct_ops
+	 *     arch_sync_dma_for_cpu() -> dma_cache_*() -> __dma_cache_*()
+	 */
 }
 
 void __ref arc_cache_init(void)

arch/arc/mm/dma.c

···
  * published by the Free Software Foundation.
  */
 
-/*
- * DMA Coherent API Notes
- *
- * I/O is inherently non-coherent on ARC. So a coherent DMA buffer is
- * implemented by accessing it using a kernel virtual address, with
- * Cache bit off in the TLB entry.
- *
- * The default DMA address == Phy address which is 0x8000_0000 based.
- */
-
 #include <linux/dma-noncoherent.h>
 #include <asm/cache.h>
 #include <asm/cacheflush.h>
 
+/*
+ * ARCH specific callbacks for generic noncoherent DMA ops (dma/noncoherent.c)
+ *  - hardware IOC not available (or "dma-coherent" not set for device in DT)
+ *  - But still handle both coherent and non-coherent requests from caller
+ *
+ * For DMA coherent hardware (IOC) generic code suffices
+ */
 void *arch_dma_alloc(struct device *dev, size_t size, dma_addr_t *dma_handle,
 		gfp_t gfp, unsigned long attrs)
 {
···
 	if (!page)
 		return NULL;
 
-	/*
-	 * IOC relies on all data (even coherent DMA data) being in cache
-	 * Thus allocate normal cached memory
-	 *
-	 * The gains with IOC are two pronged:
-	 *   -For streaming data, elides need for cache maintenance, saving
-	 *    cycles in flush code, and bus bandwidth as all the lines of a
-	 *    buffer need to be flushed out to memory
-	 *   -For coherent data, Read/Write to buffers terminate early in cache
-	 *   (vs. always going to memory - thus are faster)
-	 */
-	if ((is_isa_arcv2() && ioc_enable) ||
-	    (attrs & DMA_ATTR_NON_CONSISTENT))
+	if (attrs & DMA_ATTR_NON_CONSISTENT)
 		need_coh = 0;
 
 	/*
···
 	struct page *page = virt_to_page(paddr);
 	int is_non_coh = 1;
 
-	is_non_coh = (attrs & DMA_ATTR_NON_CONSISTENT) ||
-			(is_isa_arcv2() && ioc_enable);
+	is_non_coh = (attrs & DMA_ATTR_NON_CONSISTENT);
 
 	if (PageHighMem(page) || !is_non_coh)
 		iounmap((void __force __iomem *)vaddr);
···
 
 	default:
 		break;
+	}
+}
+
+/*
+ * Plug in coherent or noncoherent dma ops
+ */
+void arch_setup_dma_ops(struct device *dev, u64 dma_base, u64 size,
+			const struct iommu_ops *iommu, bool coherent)
+{
+	/*
+	 * IOC hardware snoops all DMA traffic keeping the caches consistent
+	 * with memory - eliding need for any explicit cache maintenance of
+	 * DMA buffers - so we can use dma_direct cache ops.
+	 */
+	if (is_isa_arcv2() && ioc_enable && coherent) {
+		set_dma_ops(dev, &dma_direct_ops);
+		dev_info(dev, "use dma_direct_ops cache ops\n");
+	} else {
+		set_dma_ops(dev, &dma_noncoherent_ops);
+		dev_info(dev, "use dma_noncoherent_ops cache ops\n");
 	}
 }