[PATCH v2 0/7] dax: I/O path enhancements

[PATCH v2 0/7] dax: I/O path enhancements

[Ross Zwisler]

The goal of this series is to enhance the DAX I/O path so that all operations
that store data (I/O writes, zeroing blocks, punching holes, etc.) properly
synchronize the stores to media using the PMEM API.  This ensures that the data
DAX is writing is durable on media before the operation completes.

Patches 1-4 are a few random cleanups.

Changes from v1:
 - Removed patches to PMEM for the "read flush" _DSM flag.  These are different
   enough that they deserve their own series, and they have a separate baseline
   which is currently moving (Dan's memremap() series).
 - Added clear_pmem() PMEM API to zero DAX memory and flush it in one call.
   (Dave)
 - Open coded flushing in arch_wb_cache_pmem() instead of adding a generic
   clwb_flush_range().  This allowed me to avoid having extra memory barriers
   and instead rely completely on arch_wmb_pmem() for ordering. (Dave)
 - Moved the arch implementation of the PMEM API into it's own arch header
   (Christoph).

Ross Zwisler (7):
  brd: make rd_size static
  pmem, x86: move x86 PMEM API to new pmem.h header
  pmem: remove layer when calling arch_has_wmb_pmem()
  pmem, x86: clean up conditional pmem includes
  pmem: add wb_cache_pmem() and clear_pmem()
  dax: update I/O path to do proper PMEM flushing
  pmem, dax: have direct_access use __pmem annotation

 Documentation/filesystems/Locking |   3 +-
 MAINTAINERS                       |   1 +
 arch/powerpc/sysdev/axonram.c     |   7 ++-
 arch/x86/include/asm/cacheflush.h |  71 ----------------------
 arch/x86/include/asm/pmem.h       | 123 ++++++++++++++++++++++++++++++++++++++
 drivers/block/brd.c               |   6 +-
 drivers/nvdimm/pmem.c             |   4 +-
 drivers/s390/block/dcssblk.c      |  10 ++--
 fs/block_dev.c                    |   2 +-
 fs/dax.c                          |  73 ++++++++++++++--------
 include/linux/blkdev.h            |   8 +--
 include/linux/pmem.h              |  66 ++++++++++++++++----
 12 files changed, 247 insertions(+), 127 deletions(-)
 create mode 100644 arch/x86/include/asm/pmem.h

-- 
2.1.0

[PATCH v2 6/7] dax: update I/O path to do proper PMEM flushing

[Ross Zwisler]

Update the DAX I/O path so that all operations that store data (I/O
writes, zeroing blocks, punching holes, etc.) properly synchronize the
stores to media using the PMEM API.  This ensures that the data DAX is
writing is durable on media before the operation completes.

Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com>
---
 fs/dax.c | 45 ++++++++++++++++++++++++++++++++-------------
 1 file changed, 32 insertions(+), 13 deletions(-)

diff --git a/fs/dax.c b/fs/dax.c
index b6769ce..ea1b2c8 100644
--- a/fs/dax.c
+++ b/fs/dax.c
@@ -17,12 +17,14 @@
 #include <linux/atomic.h>
 #include <linux/blkdev.h>
 #include <linux/buffer_head.h>
+#include <linux/dax.h>
 #include <linux/fs.h>
 #include <linux/genhd.h>
 #include <linux/highmem.h>
 #include <linux/memcontrol.h>
 #include <linux/mm.h>
 #include <linux/mutex.h>
+#include <linux/pmem.h>
 #include <linux/sched.h>
 #include <linux/uio.h>
 #include <linux/vmstat.h>
@@ -46,10 +48,7 @@ int dax_clear_blocks(struct inode *inode, sector_t block, long size)
 			unsigned pgsz = PAGE_SIZE - offset_in_page(addr);
 			if (pgsz > count)
 				pgsz = count;
-			if (pgsz < PAGE_SIZE)
-				memset(addr, 0, pgsz);
-			else
-				clear_page(addr);
+			clear_pmem((void __pmem *)addr, pgsz);
 			addr += pgsz;
 			size -= pgsz;
 			count -= pgsz;
@@ -59,6 +58,7 @@ int dax_clear_blocks(struct inode *inode, sector_t block, long size)
 		}
 	} while (size);
 
+	wmb_pmem();
 	return 0;
 }
 EXPORT_SYMBOL_GPL(dax_clear_blocks);
@@ -70,15 +70,16 @@ static long dax_get_addr(struct buffer_head *bh, void **addr, unsigned blkbits)
 	return bdev_direct_access(bh->b_bdev, sector, addr, &pfn, bh->b_size);
 }
 
+/* the clear_pmem() calls are ordered by a wmb_pmem() in the caller */
 static void dax_new_buf(void *addr, unsigned size, unsigned first, loff_t pos,
 			loff_t end)
 {
 	loff_t final = end - pos + first; /* The final byte of the buffer */
 
 	if (first > 0)
-		memset(addr, 0, first);
+		clear_pmem((void __pmem *)addr, first);
 	if (final < size)
-		memset(addr + final, 0, size - final);
+		clear_pmem((void __pmem *)addr + final, size - final);
 }
 
 static bool buffer_written(struct buffer_head *bh)
@@ -108,6 +109,7 @@ static ssize_t dax_io(struct inode *inode, struct iov_iter *iter,
 	loff_t bh_max = start;
 	void *addr;
 	bool hole = false;
+	bool need_wmb = false;
 
 	if (iov_iter_rw(iter) != WRITE)
 		end = min(end, i_size_read(inode));
@@ -145,18 +147,27 @@ static ssize_t dax_io(struct inode *inode, struct iov_iter *iter,
 				retval = dax_get_addr(bh, &addr, blkbits);
 				if (retval < 0)
 					break;
-				if (buffer_unwritten(bh) || buffer_new(bh))
+				if (buffer_unwritten(bh) || buffer_new(bh)) {
 					dax_new_buf(addr, retval, first, pos,
 									end);
+					need_wmb = true;
+				}
 				addr += first;
 				size = retval - first;
 			}
 			max = min(pos + size, end);
 		}
 
-		if (iov_iter_rw(iter) == WRITE)
+		if (iov_iter_rw(iter) == WRITE) {
 			len = copy_from_iter_nocache(addr, max - pos, iter);
-		else if (!hole)
+			/*
+			 * copy_from_iter_nocache() uses non-temporal stores
+			 * for iovec iterators so we can skip the write back.
+			 */
+			if (!iter_is_iovec(iter))
+				wb_cache_pmem((void __pmem *)addr, max - pos);
+			need_wmb = true;
+		} else if (!hole)
 			len = copy_to_iter(addr, max - pos, iter);
 		else
 			len = iov_iter_zero(max - pos, iter);
@@ -168,6 +179,9 @@ static ssize_t dax_io(struct inode *inode, struct iov_iter *iter,
 		addr += len;
 	}
 
+	if (need_wmb)
+		wmb_pmem();
+
 	return (pos == start) ? retval : pos - start;
 }
 
@@ -300,8 +314,10 @@ static int dax_insert_mapping(struct inode *inode, struct buffer_head *bh,
 		goto out;
 	}
 
-	if (buffer_unwritten(bh) || buffer_new(bh))
-		clear_page(addr);
+	if (buffer_unwritten(bh) || buffer_new(bh)) {
+		clear_pmem((void __pmem *)addr, PAGE_SIZE);
+		wmb_pmem();
+	}
 
 	error = vm_insert_mixed(vma, vaddr, pfn);
 
@@ -608,7 +624,9 @@ int __dax_pmd_fault(struct vm_area_struct *vma, unsigned long address,
 		if (buffer_unwritten(&bh) || buffer_new(&bh)) {
 			int i;
 			for (i = 0; i < PTRS_PER_PMD; i++)
-				clear_page(kaddr + i * PAGE_SIZE);
+				clear_pmem((void __pmem *)kaddr + i*PAGE_SIZE,
+						PAGE_SIZE);
+			wmb_pmem();
 			count_vm_event(PGMAJFAULT);
 			mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
 			result |= VM_FAULT_MAJOR;
@@ -720,7 +738,8 @@ int dax_zero_page_range(struct inode *inode, loff_t from, unsigned length,
 		err = dax_get_addr(&bh, &addr, inode->i_blkbits);
 		if (err < 0)
 			return err;
-		memset(addr + offset, 0, length);
+		clear_pmem((void __pmem *)addr + offset, length);
+		wmb_pmem();
 	}
 
 	return 0;
-- 
2.1.0

[PATCH v2 7/7] pmem, dax: have direct_access use __pmem annotation

[Ross Zwisler]

Update the annotation for the kaddr pointer returned by direct_access()
so that it is a __pmem pointer.  This is consistent with the PMEM driver
and with how this direct_access() pointer is used in the DAX code.

Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com>
---
 Documentation/filesystems/Locking |  3 ++-
 arch/powerpc/sysdev/axonram.c     |  7 ++++---
 drivers/block/brd.c               |  4 ++--
 drivers/nvdimm/pmem.c             |  4 ++--
 drivers/s390/block/dcssblk.c      | 10 +++++----
 fs/block_dev.c                    |  2 +-
 fs/dax.c                          | 44 +++++++++++++++++++++------------------
 include/linux/blkdev.h            |  8 +++----
 8 files changed, 45 insertions(+), 37 deletions(-)

diff --git a/Documentation/filesystems/Locking b/Documentation/filesystems/Locking
index 6a34a0f..06d4434 100644
--- a/Documentation/filesystems/Locking
+++ b/Documentation/filesystems/Locking
@@ -397,7 +397,8 @@ prototypes:
 	int (*release) (struct gendisk *, fmode_t);
 	int (*ioctl) (struct block_device *, fmode_t, unsigned, unsigned long);
 	int (*compat_ioctl) (struct block_device *, fmode_t, unsigned, unsigned long);
-	int (*direct_access) (struct block_device *, sector_t, void **, unsigned long *);
+	int (*direct_access) (struct block_device *, sector_t, void __pmem **,
+				unsigned long *);
 	int (*media_changed) (struct gendisk *);
 	void (*unlock_native_capacity) (struct gendisk *);
 	int (*revalidate_disk) (struct gendisk *);
diff --git a/arch/powerpc/sysdev/axonram.c b/arch/powerpc/sysdev/axonram.c
index ee90db1..a2be2a6 100644
--- a/arch/powerpc/sysdev/axonram.c
+++ b/arch/powerpc/sysdev/axonram.c
@@ -141,13 +141,14 @@ axon_ram_make_request(struct request_queue *queue, struct bio *bio)
  */
 static long
 axon_ram_direct_access(struct block_device *device, sector_t sector,
-		       void **kaddr, unsigned long *pfn, long size)
+		       void __pmem **kaddr, unsigned long *pfn, long size)
 {
 	struct axon_ram_bank *bank = device->bd_disk->private_data;
 	loff_t offset = (loff_t)sector << AXON_RAM_SECTOR_SHIFT;
+	void *addr = (void *)(bank->ph_addr + offset);
 
-	*kaddr = (void *)(bank->ph_addr + offset);
-	*pfn = virt_to_phys(*kaddr) >> PAGE_SHIFT;
+	*kaddr = (void __pmem *)addr;
+	*pfn = virt_to_phys(addr) >> PAGE_SHIFT;
 
 	return bank->size - offset;
 }
diff --git a/drivers/block/brd.c b/drivers/block/brd.c
index 5750b39..2691bb6 100644
--- a/drivers/block/brd.c
+++ b/drivers/block/brd.c
@@ -371,7 +371,7 @@ static int brd_rw_page(struct block_device *bdev, sector_t sector,
 
 #ifdef CONFIG_BLK_DEV_RAM_DAX
 static long brd_direct_access(struct block_device *bdev, sector_t sector,
-			void **kaddr, unsigned long *pfn, long size)
+			void __pmem **kaddr, unsigned long *pfn, long size)
 {
 	struct brd_device *brd = bdev->bd_disk->private_data;
 	struct page *page;
@@ -381,7 +381,7 @@ static long brd_direct_access(struct block_device *bdev, sector_t sector,
 	page = brd_insert_page(brd, sector);
 	if (!page)
 		return -ENOSPC;
-	*kaddr = page_address(page);
+	*kaddr = (void __pmem *)page_address(page);
 	*pfn = page_to_pfn(page);
 
 	/*
diff --git a/drivers/nvdimm/pmem.c b/drivers/nvdimm/pmem.c
index ade9eb9..68f6a6a 100644
--- a/drivers/nvdimm/pmem.c
+++ b/drivers/nvdimm/pmem.c
@@ -92,7 +92,7 @@ static int pmem_rw_page(struct block_device *bdev, sector_t sector,
 }
 
 static long pmem_direct_access(struct block_device *bdev, sector_t sector,
-			      void **kaddr, unsigned long *pfn, long size)
+		      void __pmem **kaddr, unsigned long *pfn, long size)
 {
 	struct pmem_device *pmem = bdev->bd_disk->private_data;
 	size_t offset = sector << 9;
@@ -101,7 +101,7 @@ static long pmem_direct_access(struct block_device *bdev, sector_t sector,
 		return -ENODEV;
 
 	/* FIXME convert DAX to comprehend that this mapping has a lifetime */
-	*kaddr = (void __force *) pmem->virt_addr + offset;
+	*kaddr = pmem->virt_addr + offset;
 	*pfn = (pmem->phys_addr + offset) >> PAGE_SHIFT;
 
 	return pmem->size - offset;
diff --git a/drivers/s390/block/dcssblk.c b/drivers/s390/block/dcssblk.c
index da21281..2c5a397 100644
--- a/drivers/s390/block/dcssblk.c
+++ b/drivers/s390/block/dcssblk.c
@@ -29,7 +29,7 @@ static int dcssblk_open(struct block_device *bdev, fmode_t mode);
 static void dcssblk_release(struct gendisk *disk, fmode_t mode);
 static void dcssblk_make_request(struct request_queue *q, struct bio *bio);
 static long dcssblk_direct_access(struct block_device *bdev, sector_t secnum,
-				 void **kaddr, unsigned long *pfn, long size);
+			 void __pmem **kaddr, unsigned long *pfn, long size);
 
 static char dcssblk_segments[DCSSBLK_PARM_LEN] = "\0";
 
@@ -879,18 +879,20 @@ fail:
 
 static long
 dcssblk_direct_access (struct block_device *bdev, sector_t secnum,
-			void **kaddr, unsigned long *pfn, long size)
+			void __pmem **kaddr, unsigned long *pfn, long size)
 {
 	struct dcssblk_dev_info *dev_info;
 	unsigned long offset, dev_sz;
+	void *addr;
 
 	dev_info = bdev->bd_disk->private_data;
 	if (!dev_info)
 		return -ENODEV;
 	dev_sz = dev_info->end - dev_info->start;
 	offset = secnum * 512;
-	*kaddr = (void *) (dev_info->start + offset);
-	*pfn = virt_to_phys(*kaddr) >> PAGE_SHIFT;
+	addr = (void *) (dev_info->start + offset);
+	*pfn = virt_to_phys(addr) >> PAGE_SHIFT;
+	*kaddr = (void __pmem *) addr;
 
 	return dev_sz - offset;
 }
diff --git a/fs/block_dev.c b/fs/block_dev.c
index 9be2d7e..4ab366d 100644
--- a/fs/block_dev.c
+++ b/fs/block_dev.c
@@ -442,7 +442,7 @@ EXPORT_SYMBOL_GPL(bdev_write_page);
  * accessible at this address.
  */
 long bdev_direct_access(struct block_device *bdev, sector_t sector,
-			void **addr, unsigned long *pfn, long size)
+			void __pmem **addr, unsigned long *pfn, long size)
 {
 	long avail;
 	const struct block_device_operations *ops = bdev->bd_disk->fops;
diff --git a/fs/dax.c b/fs/dax.c
index ea1b2c8..633a1ba 100644
--- a/fs/dax.c
+++ b/fs/dax.c
@@ -36,7 +36,7 @@ int dax_clear_blocks(struct inode *inode, sector_t block, long size)
 
 	might_sleep();
 	do {
-		void *addr;
+		void __pmem *addr;
 		unsigned long pfn;
 		long count;
 
@@ -48,7 +48,7 @@ int dax_clear_blocks(struct inode *inode, sector_t block, long size)
 			unsigned pgsz = PAGE_SIZE - offset_in_page(addr);
 			if (pgsz > count)
 				pgsz = count;
-			clear_pmem((void __pmem *)addr, pgsz);
+			clear_pmem(addr, pgsz);
 			addr += pgsz;
 			size -= pgsz;
 			count -= pgsz;
@@ -63,7 +63,8 @@ int dax_clear_blocks(struct inode *inode, sector_t block, long size)
 }
 EXPORT_SYMBOL_GPL(dax_clear_blocks);
 
-static long dax_get_addr(struct buffer_head *bh, void **addr, unsigned blkbits)
+static long dax_get_addr(struct buffer_head *bh, void __pmem **addr,
+		unsigned blkbits)
 {
 	unsigned long pfn;
 	sector_t sector = bh->b_blocknr << (blkbits - 9);
@@ -71,15 +72,15 @@ static long dax_get_addr(struct buffer_head *bh, void **addr, unsigned blkbits)
 }
 
 /* the clear_pmem() calls are ordered by a wmb_pmem() in the caller */
-static void dax_new_buf(void *addr, unsigned size, unsigned first, loff_t pos,
-			loff_t end)
+static void dax_new_buf(void __pmem *addr, unsigned size, unsigned first,
+		loff_t pos, loff_t end)
 {
 	loff_t final = end - pos + first; /* The final byte of the buffer */
 
 	if (first > 0)
-		clear_pmem((void __pmem *)addr, first);
+		clear_pmem(addr, first);
 	if (final < size)
-		clear_pmem((void __pmem *)addr + final, size - final);
+		clear_pmem(addr + final, size - final);
 }
 
 static bool buffer_written(struct buffer_head *bh)
@@ -107,7 +108,7 @@ static ssize_t dax_io(struct inode *inode, struct iov_iter *iter,
 	loff_t pos = start;
 	loff_t max = start;
 	loff_t bh_max = start;
-	void *addr;
+	void __pmem *addr;
 	bool hole = false;
 	bool need_wmb = false;
 
@@ -159,16 +160,18 @@ static ssize_t dax_io(struct inode *inode, struct iov_iter *iter,
 		}
 
 		if (iov_iter_rw(iter) == WRITE) {
-			len = copy_from_iter_nocache(addr, max - pos, iter);
+			len = copy_from_iter_nocache((void __force *)addr,
+					max - pos, iter);
 			/*
 			 * copy_from_iter_nocache() uses non-temporal stores
 			 * for iovec iterators so we can skip the write back.
 			 */
 			if (!iter_is_iovec(iter))
-				wb_cache_pmem((void __pmem *)addr, max - pos);
+				wb_cache_pmem(addr, max - pos);
 			need_wmb = true;
 		} else if (!hole)
-			len = copy_to_iter(addr, max - pos, iter);
+			len = copy_to_iter((void __force *)addr, max - pos,
+					iter);
 		else
 			len = iov_iter_zero(max - pos, iter);
 
@@ -274,11 +277,13 @@ static int dax_load_hole(struct address_space *mapping, struct page *page,
 static int copy_user_bh(struct page *to, struct buffer_head *bh,
 			unsigned blkbits, unsigned long vaddr)
 {
-	void *vfrom, *vto;
+	void __pmem *vfrom;
+	void *vto;
+
 	if (dax_get_addr(bh, &vfrom, blkbits) < 0)
 		return -EIO;
 	vto = kmap_atomic(to);
-	copy_user_page(vto, vfrom, vaddr, to);
+	copy_user_page(vto, (void __force *)vfrom, vaddr, to);
 	kunmap_atomic(vto);
 	return 0;
 }
@@ -288,7 +293,7 @@ static int dax_insert_mapping(struct inode *inode, struct buffer_head *bh,
 {
 	sector_t sector = bh->b_blocknr << (inode->i_blkbits - 9);
 	unsigned long vaddr = (unsigned long)vmf->virtual_address;
-	void *addr;
+	void __pmem *addr;
 	unsigned long pfn;
 	pgoff_t size;
 	int error;
@@ -315,7 +320,7 @@ static int dax_insert_mapping(struct inode *inode, struct buffer_head *bh,
 	}
 
 	if (buffer_unwritten(bh) || buffer_new(bh)) {
-		clear_pmem((void __pmem *)addr, PAGE_SIZE);
+		clear_pmem(addr, PAGE_SIZE);
 		wmb_pmem();
 	}
 
@@ -530,7 +535,7 @@ int __dax_pmd_fault(struct vm_area_struct *vma, unsigned long address,
 	unsigned long pmd_addr = address & PMD_MASK;
 	bool write = flags & FAULT_FLAG_WRITE;
 	long length;
-	void *kaddr;
+	void __pmem *kaddr;
 	pgoff_t size, pgoff;
 	sector_t block, sector;
 	unsigned long pfn;
@@ -624,8 +629,7 @@ int __dax_pmd_fault(struct vm_area_struct *vma, unsigned long address,
 		if (buffer_unwritten(&bh) || buffer_new(&bh)) {
 			int i;
 			for (i = 0; i < PTRS_PER_PMD; i++)
-				clear_pmem((void __pmem *)kaddr + i*PAGE_SIZE,
-						PAGE_SIZE);
+				clear_pmem(kaddr + i * PAGE_SIZE, PAGE_SIZE);
 			wmb_pmem();
 			count_vm_event(PGMAJFAULT);
 			mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
@@ -734,11 +738,11 @@ int dax_zero_page_range(struct inode *inode, loff_t from, unsigned length,
 	if (err < 0)
 		return err;
 	if (buffer_written(&bh)) {
-		void *addr;
+		void __pmem *addr;
 		err = dax_get_addr(&bh, &addr, inode->i_blkbits);
 		if (err < 0)
 			return err;
-		clear_pmem((void __pmem *)addr + offset, length);
+		clear_pmem(addr + offset, length);
 		wmb_pmem();
 	}
 
diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
index d4068c1..c401ecd 100644
--- a/include/linux/blkdev.h
+++ b/include/linux/blkdev.h
@@ -1555,8 +1555,8 @@ struct block_device_operations {
 	int (*rw_page)(struct block_device *, sector_t, struct page *, int rw);
 	int (*ioctl) (struct block_device *, fmode_t, unsigned, unsigned long);
 	int (*compat_ioctl) (struct block_device *, fmode_t, unsigned, unsigned long);
-	long (*direct_access)(struct block_device *, sector_t,
-					void **, unsigned long *pfn, long size);
+	long (*direct_access)(struct block_device *, sector_t, void __pmem **,
+			unsigned long *pfn, long size);
 	unsigned int (*check_events) (struct gendisk *disk,
 				      unsigned int clearing);
 	/* ->media_changed() is DEPRECATED, use ->check_events() instead */
@@ -1574,8 +1574,8 @@ extern int __blkdev_driver_ioctl(struct block_device *, fmode_t, unsigned int,
 extern int bdev_read_page(struct block_device *, sector_t, struct page *);
 extern int bdev_write_page(struct block_device *, sector_t, struct page *,
 						struct writeback_control *);
-extern long bdev_direct_access(struct block_device *, sector_t, void **addr,
-						unsigned long *pfn, long size);
+extern long bdev_direct_access(struct block_device *, sector_t,
+		void __pmem **addr, unsigned long *pfn, long size);
 #else /* CONFIG_BLOCK */
 
 struct block_device;
-- 
2.1.0

Re: [PATCH v2 6/7] dax: update I/O path to do proper PMEM flushing

[Dan Williams]

On Thu, Aug 13, 2015 at 9:51 AM, Ross Zwisler
<ross.zwisler@linux.intel.com> wrote:
> Update the DAX I/O path so that all operations that store data (I/O
> writes, zeroing blocks, punching holes, etc.) properly synchronize the
> stores to media using the PMEM API.  This ensures that the data DAX is
> writing is durable on media before the operation completes.
>
> Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com>
[..]
> @@ -145,18 +147,27 @@ static ssize_t dax_io(struct inode *inode, struct iov_iter *iter,
>                                 retval = dax_get_addr(bh, &addr, blkbits);
>                                 if (retval < 0)
>                                         break;
> -                               if (buffer_unwritten(bh) || buffer_new(bh))
> +                               if (buffer_unwritten(bh) || buffer_new(bh)) {
>                                         dax_new_buf(addr, retval, first, pos,
>                                                                         end);
> +                                       need_wmb = true;
> +                               }
>                                 addr += first;
>                                 size = retval - first;
>                         }
>                         max = min(pos + size, end);
>                 }
>
> -               if (iov_iter_rw(iter) == WRITE)
> +               if (iov_iter_rw(iter) == WRITE) {
>                         len = copy_from_iter_nocache(addr, max - pos, iter);
> -               else if (!hole)
> +                       /*
> +                        * copy_from_iter_nocache() uses non-temporal stores
> +                        * for iovec iterators so we can skip the write back.
> +                        */
> +                       if (!iter_is_iovec(iter))
> +                               wb_cache_pmem((void __pmem *)addr, max - pos);
> +                       need_wmb = true;

I think this should become copy_from_iter_pmem() and hide the
wb_cache_pmem() as an internal arch detail.  I.e. wb_cache_pmem()
should not be a global api when its usage is architecture specific.
Otherwise are you asserting that all architecture implementations of
copy_from_iter_nocache() are pmem safe?

Re: [PATCH v2 7/7] pmem, dax: have direct_access use __pmem annotation

[Dan Williams]

On Thu, Aug 13, 2015 at 9:51 AM, Ross Zwisler
<ross.zwisler@linux.intel.com> wrote:
> Update the annotation for the kaddr pointer returned by direct_access()
> so that it is a __pmem pointer.  This is consistent with the PMEM driver
> and with how this direct_access() pointer is used in the DAX code.
>
> Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com>
> ---
>  Documentation/filesystems/Locking |  3 ++-
>  arch/powerpc/sysdev/axonram.c     |  7 ++++---
>  drivers/block/brd.c               |  4 ++--
>  drivers/nvdimm/pmem.c             |  4 ++--
>  drivers/s390/block/dcssblk.c      | 10 +++++----
>  fs/block_dev.c                    |  2 +-
>  fs/dax.c                          | 44 +++++++++++++++++++++------------------
>  include/linux/blkdev.h            |  8 +++----
>  8 files changed, 45 insertions(+), 37 deletions(-)
>
> diff --git a/Documentation/filesystems/Locking b/Documentation/filesystems/Locking
> index 6a34a0f..06d4434 100644
> --- a/Documentation/filesystems/Locking
> +++ b/Documentation/filesystems/Locking
> @@ -397,7 +397,8 @@ prototypes:
>         int (*release) (struct gendisk *, fmode_t);
>         int (*ioctl) (struct block_device *, fmode_t, unsigned, unsigned long);
>         int (*compat_ioctl) (struct block_device *, fmode_t, unsigned, unsigned long);
> -       int (*direct_access) (struct block_device *, sector_t, void **, unsigned long *);
> +       int (*direct_access) (struct block_device *, sector_t, void __pmem **,
> +                               unsigned long *);

So this collides with the __pfn_t work.  I think the we have a
reasonable chance of getting that in to 4.3, so I'd wait to see if we
hit any major roadblocks with that set [1] before merging these.

[1]: https://lists.01.org/pipermail/linux-nvdimm/2015-August/001803.html

Re: [PATCH v2 6/7] dax: update I/O path to do proper PMEM flushing

[Ross Zwisler]

On Thu, 2015-08-13 at 14:11 -0700, Dan Williams wrote:
> On Thu, Aug 13, 2015 at 9:51 AM, Ross Zwisler
> <ross.zwisler@linux.intel.com> wrote:
> > Update the DAX I/O path so that all operations that store data (I/O
> > writes, zeroing blocks, punching holes, etc.) properly synchronize the
> > stores to media using the PMEM API.  This ensures that the data DAX is
> > writing is durable on media before the operation completes.
> >
> > Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com>
> [..]
> > @@ -145,18 +147,27 @@ static ssize_t dax_io(struct inode *inode, struct iov_iter *iter,
> >                                 retval = dax_get_addr(bh, &addr, blkbits);
> >                                 if (retval < 0)
> >                                         break;
> > -                               if (buffer_unwritten(bh) || buffer_new(bh))
> > +                               if (buffer_unwritten(bh) || buffer_new(bh)) {
> >                                         dax_new_buf(addr, retval, first, pos,
> >                                                                         end);
> > +                                       need_wmb = true;
> > +                               }
> >                                 addr += first;
> >                                 size = retval - first;
> >                         }
> >                         max = min(pos + size, end);
> >                 }
> >
> > -               if (iov_iter_rw(iter) == WRITE)
> > +               if (iov_iter_rw(iter) == WRITE) {
> >                         len = copy_from_iter_nocache(addr, max - pos, iter);
> > -               else if (!hole)
> > +                       /*
> > +                        * copy_from_iter_nocache() uses non-temporal stores
> > +                        * for iovec iterators so we can skip the write back.
> > +                        */
> > +                       if (!iter_is_iovec(iter))
> > +                               wb_cache_pmem((void __pmem *)addr, max - pos);
> > +                       need_wmb = true;
> 
> I think this should become copy_from_iter_pmem() and hide the
> wb_cache_pmem() as an internal arch detail.  I.e. wb_cache_pmem()
> should not be a global api when its usage is architecture specific.
> Otherwise are you asserting that all architecture implementations of
> copy_from_iter_nocache() are pmem safe?

Great point.  Nope, copy_from_iter_nocache() uses __copy_from_user_nocache(),
which just defaults to __copy_from_user() on non-x86.  Dang, the PMEM API just
keeps growing...  :(

Re: [PATCH v2 7/7] pmem, dax: have direct_access use __pmem annotation

[Ross Zwisler]

On Thu, 2015-08-13 at 14:20 -0700, Dan Williams wrote:
> On Thu, Aug 13, 2015 at 9:51 AM, Ross Zwisler
> <ross.zwisler@linux.intel.com> wrote:
> > Update the annotation for the kaddr pointer returned by direct_access()
> > so that it is a __pmem pointer.  This is consistent with the PMEM driver
> > and with how this direct_access() pointer is used in the DAX code.
> >
> > Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com>
> > ---
> >  Documentation/filesystems/Locking |  3 ++-
> >  arch/powerpc/sysdev/axonram.c     |  7 ++++---
> >  drivers/block/brd.c               |  4 ++--
> >  drivers/nvdimm/pmem.c             |  4 ++--
> >  drivers/s390/block/dcssblk.c      | 10 +++++----
> >  fs/block_dev.c                    |  2 +-
> >  fs/dax.c                          | 44 +++++++++++++++++++++------------------
> >  include/linux/blkdev.h            |  8 +++----
> >  8 files changed, 45 insertions(+), 37 deletions(-)
> >
> > diff --git a/Documentation/filesystems/Locking b/Documentation/filesystems/Locking
> > index 6a34a0f..06d4434 100644
> > --- a/Documentation/filesystems/Locking
> > +++ b/Documentation/filesystems/Locking
> > @@ -397,7 +397,8 @@ prototypes:
> >         int (*release) (struct gendisk *, fmode_t);
> >         int (*ioctl) (struct block_device *, fmode_t, unsigned, unsigned long);
> >         int (*compat_ioctl) (struct block_device *, fmode_t, unsigned, unsigned long);
> > -       int (*direct_access) (struct block_device *, sector_t, void **, unsigned long *);
> > +       int (*direct_access) (struct block_device *, sector_t, void __pmem **,
> > +                               unsigned long *);
> 
> So this collides with the __pfn_t work.  I think the we have a
> reasonable chance of getting that in to 4.3, so I'd wait to see if we
> hit any major roadblocks with that set [1] before merging these.
> 
> [1]: https://lists.01.org/pipermail/linux-nvdimm/2015-August/001803.html

Fair enough.  Yea, I hadn't merged with that series yet a) because I didn't
know when its review cycle would settle down and b) because that series hadn't
pulled in changes from Matthew for PMD support, which I was originally using
as a baseline.

I'll merge with your code for v3.

Re: [PATCH v2 7/7] pmem, dax: have direct_access use __pmem annotation

[Dan Williams]

On Fri, Aug 14, 2015 at 9:55 AM, Ross Zwisler
<ross.zwisler@linux.intel.com> wrote:
> On Thu, 2015-08-13 at 14:20 -0700, Dan Williams wrote:
>> On Thu, Aug 13, 2015 at 9:51 AM, Ross Zwisler
>> <ross.zwisler@linux.intel.com> wrote:
>> > Update the annotation for the kaddr pointer returned by direct_access()
>> > so that it is a __pmem pointer.  This is consistent with the PMEM driver
>> > and with how this direct_access() pointer is used in the DAX code.
>> >
>> > Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com>
>> > ---
>> >  Documentation/filesystems/Locking |  3 ++-
>> >  arch/powerpc/sysdev/axonram.c     |  7 ++++---
>> >  drivers/block/brd.c               |  4 ++--
>> >  drivers/nvdimm/pmem.c             |  4 ++--
>> >  drivers/s390/block/dcssblk.c      | 10 +++++----
>> >  fs/block_dev.c                    |  2 +-
>> >  fs/dax.c                          | 44 +++++++++++++++++++++------------------
>> >  include/linux/blkdev.h            |  8 +++----
>> >  8 files changed, 45 insertions(+), 37 deletions(-)
>> >
>> > diff --git a/Documentation/filesystems/Locking b/Documentation/filesystems/Locking
>> > index 6a34a0f..06d4434 100644
>> > --- a/Documentation/filesystems/Locking
>> > +++ b/Documentation/filesystems/Locking
>> > @@ -397,7 +397,8 @@ prototypes:
>> >         int (*release) (struct gendisk *, fmode_t);
>> >         int (*ioctl) (struct block_device *, fmode_t, unsigned, unsigned long);
>> >         int (*compat_ioctl) (struct block_device *, fmode_t, unsigned, unsigned long);
>> > -       int (*direct_access) (struct block_device *, sector_t, void **, unsigned long *);
>> > +       int (*direct_access) (struct block_device *, sector_t, void __pmem **,
>> > +                               unsigned long *);
>>
>> So this collides with the __pfn_t work.  I think the we have a
>> reasonable chance of getting that in to 4.3, so I'd wait to see if we
>> hit any major roadblocks with that set [1] before merging these.
>>
>> [1]: https://lists.01.org/pipermail/linux-nvdimm/2015-August/001803.html
>
> Fair enough.  Yea, I hadn't merged with that series yet a) because I didn't
> know when its review cycle would settle down and b) because that series hadn't
> pulled in changes from Matthew for PMD support, which I was originally using
> as a baseline.
>
> I'll merge with your code for v3.

Sounds, let me go rebase the __pfn_t patches on -mm so we'all lined up
and collision free.

Re: [PATCH v2 7/7] pmem, dax: have direct_access use __pmem annotation

[Christoph Hellwig]

On Fri, Aug 14, 2015 at 09:58:16AM -0700, Dan Williams wrote:
> > I'll merge with your code for v3.
> 
> Sounds, let me go rebase the __pfn_t patches on -mm so we'all lined up
> and collision free.

I'm doubt that we'll have PFN mapping ready for 4.3.  I'd rather see
Ross series goes first, and move the patch to remove the size argument
from ->direct access [1] over to this series as well.

[1] https://git.kernel.org/cgit/linux/kernel/git/djbw/nvdimm.git/commit/?h=pfn&id=8e15e69fb9e61ac563c5a7ffd9dd9a7b545cced3

Re: [PATCH v2 7/7] pmem, dax: have direct_access use __pmem annotation

[Christoph Hellwig]

On Thu, Aug 13, 2015 at 10:51:11AM -0600, Ross Zwisler wrote:
> Update the annotation for the kaddr pointer returned by direct_access()
> so that it is a __pmem pointer.  This is consistent with the PMEM driver
> and with how this direct_access() pointer is used in the DAX code.

IFF we stick to the __pmem annotations this looks good.

That beeing said I start to really dislike them.  We don't special
accesors to read/write from pmem, we just need to explicitly commit
it if we want to make it persistent.  So I really don't see the need
to treat it special and require all the force casts to and from the
attribute.

Re: [PATCH v2 7/7] pmem, dax: have direct_access use __pmem annotation

[Dan Williams]

On Sat, Aug 15, 2015 at 2:19 AM, Christoph Hellwig <hch@lst.de> wrote:
> On Thu, Aug 13, 2015 at 10:51:11AM -0600, Ross Zwisler wrote:
>> Update the annotation for the kaddr pointer returned by direct_access()
>> so that it is a __pmem pointer.  This is consistent with the PMEM driver
>> and with how this direct_access() pointer is used in the DAX code.
>
> IFF we stick to the __pmem annotations this looks good.
>
> That beeing said I start to really dislike them.  We don't special
> accesors to read/write from pmem, we just need to explicitly commit
> it if we want to make it persistent.  So I really don't see the need
> to treat it special and require all the force casts to and from the
> attribute.

I'm not going to put up much of a fight if it's really getting in the way....

That said, while we don't need special accessors we do need guarantees
that anything that has written to a persistent memory address has done
so in a way that wmb_pmem() is able to flush it.  It's more of a "I've
audited this code path for wmb_pmem() compatibility so use this api to
write to pmem."

Perhaps a better way to statically check for missed flushes might be
to have acquire_pmem_for_write() + release() annotations and the final
release does a wmb_pmem(), but as far as I can tell the sparse
acquire/release annotations don't stack.

Re: [PATCH v2 7/7] pmem, dax: have direct_access use __pmem annotation

[Dan Williams]

On Sat, Aug 15, 2015 at 2:11 AM, Christoph Hellwig <hch@lst.de> wrote:
> On Fri, Aug 14, 2015 at 09:58:16AM -0700, Dan Williams wrote:
>> > I'll merge with your code for v3.
>>
>> Sounds, let me go rebase the __pfn_t patches on -mm so we'all lined up
>> and collision free.
>
> I'm doubt that we'll have PFN mapping ready for 4.3.  I'd rather see
> Ross series goes first, and move the patch to remove the size argument
> from ->direct access [1] over to this series as well.
>
> [1] https://git.kernel.org/cgit/linux/kernel/git/djbw/nvdimm.git/commit/?h=pfn&id=8e15e69fb9e61ac563c5a7ffd9dd9a7b545cced3

Yes, let's do it.  The need for DAX persistence guarantees is a higher
priority than solving the DAX vs PMEM unbind bug.  Especially if we
can defer __pfn_t and get some basic struct page mapping into 4.3 as
well.

Re: [PATCH v2 7/7] pmem, dax: have direct_access use __pmem annotation

[Christoph Hellwig]

On Sat, Aug 15, 2015 at 08:44:27AM -0700, Dan Williams wrote:
> That said, while we don't need special accessors we do need guarantees
> that anything that has written to a persistent memory address has done
> so in a way that wmb_pmem() is able to flush it.  It's more of a "I've
> audited this code path for wmb_pmem() compatibility so use this api to
> write to pmem."

I'm more worried about things where don't just do plain loads and stores
to a pmem region but DMA, which will end up as a nightmare of casts.

But we can wait and see how that evolves in the end..

Re: [PATCH v2 7/7] pmem, dax: have direct_access use __pmem annotation

[Dan Williams]

On Sat, Aug 15, 2015 at 9:00 AM, Christoph Hellwig <hch@lst.de> wrote:
> On Sat, Aug 15, 2015 at 08:44:27AM -0700, Dan Williams wrote:
>> That said, while we don't need special accessors we do need guarantees
>> that anything that has written to a persistent memory address has done
>> so in a way that wmb_pmem() is able to flush it.  It's more of a "I've
>> audited this code path for wmb_pmem() compatibility so use this api to
>> write to pmem."
>
> I'm more worried about things where don't just do plain loads and stores
> to a pmem region but DMA, which will end up as a nightmare of casts.
>
> But we can wait and see how that evolves in the end..

It's already not possible to do something like dma_map_single() on an
ioremap()'d address, so there currently are't any __iomem/DMA
collisions.  As long as DMA setup is relative to a physical address
resource I think we're ok.

Making sure a DMA is both complete and persistent though is a different problem.

Re: [PATCH v2 7/7] pmem, dax: have direct_access use __pmem annotation

[Ross Zwisler]

On Sat, 2015-08-15 at 08:44 -0700, Dan Williams wrote:
> On Sat, Aug 15, 2015 at 2:19 AM, Christoph Hellwig <hch@lst.de> wrote:
> > On Thu, Aug 13, 2015 at 10:51:11AM -0600, Ross Zwisler wrote:
> >> Update the annotation for the kaddr pointer returned by direct_access()
> >> so that it is a __pmem pointer.  This is consistent with the PMEM driver
> >> and with how this direct_access() pointer is used in the DAX code.
> >
> > IFF we stick to the __pmem annotations this looks good.
> >
> > That beeing said I start to really dislike them.  We don't special
> > accesors to read/write from pmem, we just need to explicitly commit
> > it if we want to make it persistent.  So I really don't see the need
> > to treat it special and require all the force casts to and from the
> > attribute.
> 
> I'm not going to put up much of a fight if it's really getting in the way....
> 
> That said, while we don't need special accessors we do need guarantees
> that anything that has written to a persistent memory address has done
> so in a way that wmb_pmem() is able to flush it.  It's more of a "I've
> audited this code path for wmb_pmem() compatibility so use this api to
> write to pmem."
> 
> Perhaps a better way to statically check for missed flushes might be
> to have acquire_pmem_for_write() + release() annotations and the final
> release does a wmb_pmem(), but as far as I can tell the sparse
> acquire/release annotations don't stack.

FWIW I've been on the fence about the __pmem annotations, but my current
thought is that we really do need a way of saying that stores to these
pointers need special care for wmb_pmem() to do its thing and that __pmem does
a reasonably good job of that.  If we can figure out a cooler way, such as the
write() + release() flow Dan is talking about, great.  But I think we need
something to keep us from making errors by storing to PMEM pointers and
leaving data in the processor cache.