arch, x86: pmem api for ensuring durability of persistent memory updates

Based on an original patch by Ross Zwisler [1].

Writes to persistent memory have the potential to be posted to cpu
cache, cpu write buffers, and platform write buffers (memory controller)
before being committed to persistent media.  Provide apis,
memcpy_to_pmem(), wmb_pmem(), and memremap_pmem(), to write data to
pmem and assert that it is durable in PMEM (a persistent linear address
range).  A '__pmem' attribute is added so sparse can track proper usage
of pointers to pmem.

This continues the status quo of pmem being x86 only for 4.2, but
reworks to ioremap, and wider implementation of memremap() will enable
other archs in 4.3.

[1]: https://lists.01.org/pipermail/linux-nvdimm/2015-May/000932.html

Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com>
[djbw: various reworks]
Signed-off-by: Dan Williams <dan.j.williams@intel.com>

7 files changed, 257 insertions, 13 deletions

diff --git a/arch/x86/Kconfig b/arch/x86/Kconfig
index 1a2cbf641667..62564ddf7f78 100644
--- a/arch/x86/Kconfig
+++ b/arch/x86/Kconfig
@@ -27,6 +27,7 @@ config X86
 	select ARCH_HAS_DEBUG_STRICT_USER_COPY_CHECKS
 	select ARCH_HAS_FAST_MULTIPLIER
 	select ARCH_HAS_GCOV_PROFILE_ALL
+	select ARCH_HAS_PMEM_API
 	select ARCH_MIGHT_HAVE_PC_PARPORT
 	select ARCH_MIGHT_HAVE_PC_SERIO
 	select HAVE_AOUT if X86_32
diff --git a/arch/x86/include/asm/cacheflush.h b/arch/x86/include/asm/cacheflush.h
index 47c8e32f621a..ec23bb753a3e 100644
--- a/arch/x86/include/asm/cacheflush.h
+++ b/arch/x86/include/asm/cacheflush.h
@@ -4,6 +4,7 @@
 /* Caches aren't brain-dead on the intel. */
 #include <asm-generic/cacheflush.h>
 #include <asm/special_insns.h>
+#include <asm/uaccess.h>
 
 /*
  * The set_memory_* API can be used to change various attributes of a virtual
@@ -104,4 +105,75 @@ static inline int rodata_test(void)
 }
 #endif
 
+#ifdef ARCH_HAS_NOCACHE_UACCESS
+
+/**
+ * arch_memcpy_to_pmem - copy data to persistent memory
+ * @dst: destination buffer for the copy
+ * @src: source buffer for the copy
+ * @n: length of the copy in bytes
+ *
+ * Copy data to persistent memory media via non-temporal stores so that
+ * a subsequent arch_wmb_pmem() can flush cpu and memory controller
+ * write buffers to guarantee durability.
+ */
+static inline void arch_memcpy_to_pmem(void __pmem *dst, const void *src,
+		size_t n)
+{
+	int unwritten;
+
+	/*
+	 * We are copying between two kernel buffers, if
+	 * __copy_from_user_inatomic_nocache() returns an error (page
+	 * fault) we would have already reported a general protection fault
+	 * before the WARN+BUG.
+	 */
+	unwritten = __copy_from_user_inatomic_nocache((void __force *) dst,
+			(void __user *) src, n);
+	if (WARN(unwritten, "%s: fault copying %p <- %p unwritten: %d\n",
+				__func__, dst, src, unwritten))
+		BUG();
+}
+
+/**
+ * arch_wmb_pmem - synchronize writes to persistent memory
+ *
+ * After a series of arch_memcpy_to_pmem() operations this drains data
+ * from cpu write buffers and any platform (memory controller) buffers
+ * to ensure that written data is durable on persistent memory media.
+ */
+static inline void arch_wmb_pmem(void)
+{
+	/*
+	 * wmb() to 'sfence' all previous writes such that they are
+	 * architecturally visible to 'pcommit'.  Note, that we've
+	 * already arranged for pmem writes to avoid the cache via
+	 * arch_memcpy_to_pmem().
+	 */
+	wmb();
+	pcommit_sfence();
+}
+
+static inline bool __arch_has_wmb_pmem(void)
+{
+#ifdef CONFIG_X86_64
+	/*
+	 * We require that wmb() be an 'sfence', that is only guaranteed on
+	 * 64-bit builds
+	 */
+	return static_cpu_has(X86_FEATURE_PCOMMIT);
+#else
+	return false;
+#endif
+}
+#else /* ARCH_HAS_NOCACHE_UACCESS i.e. ARCH=um */
+extern void arch_memcpy_to_pmem(void __pmem *dst, const void *src, size_t n);
+extern void arch_wmb_pmem(void);
+
+static inline bool __arch_has_wmb_pmem(void)
+{
+	return false;
+}
+#endif
+
 #endif /* _ASM_X86_CACHEFLUSH_H */
diff --git a/arch/x86/include/asm/io.h b/arch/x86/include/asm/io.h
index 34a5b93704d3..c60c3f3b0183 100644
--- a/arch/x86/include/asm/io.h
+++ b/arch/x86/include/asm/io.h
@@ -247,6 +247,12 @@ static inline void flush_write_buffers(void)
 #endif
 }
 
+static inline void __pmem *arch_memremap_pmem(resource_size_t offset,
+	unsigned long size)
+{
+	return (void __force __pmem *) ioremap_cache(offset, size);
+}
+
 #endif /* __KERNEL__ */
 
 extern void native_io_delay(void);
diff --git a/drivers/nvdimm/pmem.c b/drivers/nvdimm/pmem.c
index 42b766f33e59..ade9eb917a4d 100644
--- a/drivers/nvdimm/pmem.c
+++ b/drivers/nvdimm/pmem.c
@@ -23,6 +23,7 @@
 #include <linux/module.h>
 #include <linux/moduleparam.h>
 #include <linux/slab.h>
+#include <linux/pmem.h>
 #include <linux/nd.h>
 #include "nd.h"
 
@@ -32,7 +33,7 @@ struct pmem_device {
 
 	/* One contiguous memory region per device */
 	phys_addr_t		phys_addr;
-	void			*virt_addr;
+	void __pmem		*virt_addr;
 	size_t			size;
 };
 
@@ -44,13 +45,14 @@ static void pmem_do_bvec(struct pmem_device *pmem, struct page *page,
 {
 	void *mem = kmap_atomic(page);
 	size_t pmem_off = sector << 9;
+	void __pmem *pmem_addr = pmem->virt_addr + pmem_off;
 
 	if (rw == READ) {
-		memcpy(mem + off, pmem->virt_addr + pmem_off, len);
+		memcpy_from_pmem(mem + off, pmem_addr, len);
 		flush_dcache_page(page);
 	} else {
 		flush_dcache_page(page);
-		memcpy(pmem->virt_addr + pmem_off, mem + off, len);
+		memcpy_to_pmem(pmem_addr, mem + off, len);
 	}
 
 	kunmap_atomic(mem);
@@ -71,6 +73,10 @@ static void pmem_make_request(struct request_queue *q, struct bio *bio)
 				bio_data_dir(bio), iter.bi_sector);
 	if (do_acct)
 		nd_iostat_end(bio, start);
+
+	if (bio_data_dir(bio))
+		wmb_pmem();
+
 	bio_endio(bio, 0);
 }
 
@@ -94,7 +100,8 @@ static long pmem_direct_access(struct block_device *bdev, sector_t sector,
 	if (!pmem)
 		return -ENODEV;
 
-	*kaddr = pmem->virt_addr + offset;
+	/* FIXME convert DAX to comprehend that this mapping has a lifetime */
+	*kaddr = (void __force *) pmem->virt_addr + offset;
 	*pfn = (pmem->phys_addr + offset) >> PAGE_SHIFT;
 
 	return pmem->size - offset;
@@ -118,6 +125,8 @@ static struct pmem_device *pmem_alloc(struct device *dev,
 
 	pmem->phys_addr = res->start;
 	pmem->size = resource_size(res);
+	if (!arch_has_pmem_api())
+		dev_warn(dev, "unable to guarantee persistence of writes\n");
 
 	if (!request_mem_region(pmem->phys_addr, pmem->size, dev_name(dev))) {
 		dev_warn(dev, "could not reserve region [0x%pa:0x%zx]\n",
@@ -126,11 +135,7 @@ static struct pmem_device *pmem_alloc(struct device *dev,
 		return ERR_PTR(-EBUSY);
 	}
 
-	/*
-	 * Map the memory as non-cachable, as we can't write back the contents
-	 * of the CPU caches in case of a crash.
-	 */
-	pmem->virt_addr = ioremap_nocache(pmem->phys_addr, pmem->size);
+	pmem->virt_addr = memremap_pmem(pmem->phys_addr, pmem->size);
 	if (!pmem->virt_addr) {
 		release_mem_region(pmem->phys_addr, pmem->size);
 		kfree(pmem);
@@ -195,16 +200,18 @@ static int pmem_rw_bytes(struct nd_namespace_common *ndns,
 	}
 
 	if (rw == READ)
-		memcpy(buf, pmem->virt_addr + offset, size);
-	else
-		memcpy(pmem->virt_addr + offset, buf, size);
+		memcpy_from_pmem(buf, pmem->virt_addr + offset, size);
+	else {
+		memcpy_to_pmem(pmem->virt_addr + offset, buf, size);
+		wmb_pmem();
+	}
 
 	return 0;
 }
 
 static void pmem_free(struct pmem_device *pmem)
 {
-	iounmap(pmem->virt_addr);
+	memunmap_pmem(pmem->virt_addr);
 	release_mem_region(pmem->phys_addr, pmem->size);
 	kfree(pmem);
 }
diff --git a/include/linux/compiler.h b/include/linux/compiler.h
index 867722591be2..9a528d945498 100644
--- a/include/linux/compiler.h
+++ b/include/linux/compiler.h
@@ -21,6 +21,7 @@
 # define __rcu		__attribute__((noderef, address_space(4)))
 #else
 # define __rcu
+# define __pmem		__attribute__((noderef, address_space(5)))
 #endif
 extern void __chk_user_ptr(const volatile void __user *);
 extern void __chk_io_ptr(const volatile void __iomem *);
@@ -42,6 +43,7 @@ extern void __chk_io_ptr(const volatile void __iomem *);
 # define __cond_lock(x,c) (c)
 # define __percpu
 # define __rcu
+# define __pmem
 #endif
 
 /* Indirect macros required for expanded argument pasting, eg. __LINE__. */
diff --git a/include/linux/pmem.h b/include/linux/pmem.h
new file mode 100644
index 000000000000..f6481a0b1d4f
--- /dev/null
+++ b/include/linux/pmem.h
@@ -0,0 +1,153 @@
+/*
+ * Copyright(c) 2015 Intel Corporation. All rights reserved.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of version 2 of the GNU General Public License as
+ * published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ */
+#ifndef __PMEM_H__
+#define __PMEM_H__
+
+#include <linux/io.h>
+
+#ifdef CONFIG_ARCH_HAS_PMEM_API
+#include <asm/cacheflush.h>
+#else
+static inline void arch_wmb_pmem(void)
+{
+	BUG();
+}
+
+static inline bool __arch_has_wmb_pmem(void)
+{
+	return false;
+}
+
+static inline void __pmem *arch_memremap_pmem(resource_size_t offset,
+		unsigned long size)
+{
+	return NULL;
+}
+
+static inline void arch_memcpy_to_pmem(void __pmem *dst, const void *src,
+		size_t n)
+{
+	BUG();
+}
+#endif
+
+/*
+ * Architectures that define ARCH_HAS_PMEM_API must provide
+ * implementations for arch_memremap_pmem(), arch_memcpy_to_pmem(),
+ * arch_wmb_pmem(), and __arch_has_wmb_pmem().
+ */
+
+static inline void memcpy_from_pmem(void *dst, void __pmem const *src, size_t size)
+{
+	memcpy(dst, (void __force const *) src, size);
+}
+
+static inline void memunmap_pmem(void __pmem *addr)
+{
+	iounmap((void __force __iomem *) addr);
+}
+
+/**
+ * arch_has_wmb_pmem - true if wmb_pmem() ensures durability
+ *
+ * For a given cpu implementation within an architecture it is possible
+ * that wmb_pmem() resolves to a nop.  In the case this returns
+ * false, pmem api users are unable to ensure durability and may want to
+ * fall back to a different data consistency model, or otherwise notify
+ * the user.
+ */
+static inline bool arch_has_wmb_pmem(void)
+{
+	if (IS_ENABLED(CONFIG_ARCH_HAS_PMEM_API))
+		return __arch_has_wmb_pmem();
+	return false;
+}
+
+static inline bool arch_has_pmem_api(void)
+{
+	return IS_ENABLED(CONFIG_ARCH_HAS_PMEM_API) && arch_has_wmb_pmem();
+}
+
+/*
+ * These defaults seek to offer decent performance and minimize the
+ * window between i/o completion and writes being durable on media.
+ * However, it is undefined / architecture specific whether
+ * default_memremap_pmem + default_memcpy_to_pmem is sufficient for
+ * making data durable relative to i/o completion.
+ */
+static void default_memcpy_to_pmem(void __pmem *dst, const void *src,
+		size_t size)
+{
+	memcpy((void __force *) dst, src, size);
+}
+
+static void __pmem *default_memremap_pmem(resource_size_t offset,
+		unsigned long size)
+{
+	/* TODO: convert to ioremap_wt() */
+	return (void __pmem __force *)ioremap_nocache(offset, size);
+}
+
+/**
+ * memremap_pmem - map physical persistent memory for pmem api
+ * @offset: physical address of persistent memory
+ * @size: size of the mapping
+ *
+ * Establish a mapping of the architecture specific memory type expected
+ * by memcpy_to_pmem() and wmb_pmem().  For example, it may be
+ * the case that an uncacheable or writethrough mapping is sufficient,
+ * or a writeback mapping provided memcpy_to_pmem() and
+ * wmb_pmem() arrange for the data to be written through the
+ * cache to persistent media.
+ */
+static inline void __pmem *memremap_pmem(resource_size_t offset,
+		unsigned long size)
+{
+	if (arch_has_pmem_api())
+		return arch_memremap_pmem(offset, size);
+	return default_memremap_pmem(offset, size);
+}
+
+/**
+ * memcpy_to_pmem - copy data to persistent memory
+ * @dst: destination buffer for the copy
+ * @src: source buffer for the copy
+ * @n: length of the copy in bytes
+ *
+ * Perform a memory copy that results in the destination of the copy
+ * being effectively evicted from, or never written to, the processor
+ * cache hierarchy after the copy completes.  After memcpy_to_pmem()
+ * data may still reside in cpu or platform buffers, so this operation
+ * must be followed by a wmb_pmem().
+ */
+static inline void memcpy_to_pmem(void __pmem *dst, const void *src, size_t n)
+{
+	if (arch_has_pmem_api())
+		arch_memcpy_to_pmem(dst, src, n);
+	else
+		default_memcpy_to_pmem(dst, src, n);
+}
+
+/**
+ * wmb_pmem - synchronize writes to persistent memory
+ *
+ * After a series of memcpy_to_pmem() operations this drains data from
+ * cpu write buffers and any platform (memory controller) buffers to
+ * ensure that written data is durable on persistent memory media.
+ */
+static inline void wmb_pmem(void)
+{
+	if (arch_has_pmem_api())
+		arch_wmb_pmem();
+}
+#endif /* __PMEM_H__ */
diff --git a/lib/Kconfig b/lib/Kconfig
index 601965a948e8..d27c13a91c28 100644
--- a/lib/Kconfig
+++ b/lib/Kconfig
@@ -522,4 +522,7 @@ source "lib/fonts/Kconfig"
 config ARCH_HAS_SG_CHAIN
 	def_bool n
 
+config ARCH_HAS_PMEM_API
+	bool
+
 endmenu