dm log writes: add support for DAX
authorRoss Zwisler <ross.zwisler@linux.intel.com>
Fri, 20 Oct 2017 05:24:04 +0000 (23:24 -0600)
committerMike Snitzer <snitzer@redhat.com>
Fri, 10 Nov 2017 20:44:51 +0000 (15:44 -0500)
Now that we have the ability log filesystem writes using a flat buffer, add
support for DAX.

The motivation for this support is the need for an xfstest that can test
the new MAP_SYNC DAX flag.  By logging the filesystem activity with
dm-log-writes we can show that the MAP_SYNC page faults are writing out
their metadata as they happen, instead of requiring an explicit
msync/fsync.

Unfortunately we can't easily track data that has been written via
mmap() now that the dax_flush() abstraction was removed by commit
c3ca015fab6d ("dax: remove the pmem_dax_ops->flush abstraction").
Otherwise we could just treat each flush as a big write, and store the
data that is being synced to media.  It may be worthwhile to add the
dax_flush() entry point back, just as a notifier so we can do this
logging.

Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
drivers/md/dm-log-writes.c

index 34b9b41..189badb 100644 (file)
 #include <linux/init.h>
 #include <linux/blkdev.h>
 #include <linux/bio.h>
+#include <linux/dax.h>
 #include <linux/slab.h>
 #include <linux/kthread.h>
 #include <linux/freezer.h>
+#include <linux/uio.h>
 
 #define DM_MSG_PREFIX "log-writes"
 
@@ -608,6 +610,51 @@ static int log_mark(struct log_writes_c *lc, char *data)
        return 0;
 }
 
+static int log_dax(struct log_writes_c *lc, sector_t sector, size_t bytes,
+                  struct iov_iter *i)
+{
+       struct pending_block *block;
+
+       if (!bytes)
+               return 0;
+
+       block = kzalloc(sizeof(struct pending_block), GFP_KERNEL);
+       if (!block) {
+               DMERR("Error allocating dax pending block");
+               return -ENOMEM;
+       }
+
+       block->data = kzalloc(bytes, GFP_KERNEL);
+       if (!block->data) {
+               DMERR("Error allocating dax data space");
+               kfree(block);
+               return -ENOMEM;
+       }
+
+       /* write data provided via the iterator */
+       if (!copy_from_iter(block->data, bytes, i)) {
+               DMERR("Error copying dax data");
+               kfree(block->data);
+               kfree(block);
+               return -EIO;
+       }
+
+       /* rewind the iterator so that the block driver can use it */
+       iov_iter_revert(i, bytes);
+
+       block->datalen = bytes;
+       block->sector = bio_to_dev_sectors(lc, sector);
+       block->nr_sectors = ALIGN(bytes, lc->sectorsize) >> lc->sectorshift;
+
+       atomic_inc(&lc->pending_blocks);
+       spin_lock_irq(&lc->blocks_lock);
+       list_add_tail(&block->list, &lc->unflushed_blocks);
+       spin_unlock_irq(&lc->blocks_lock);
+       wake_up_process(lc->log_kthread);
+
+       return 0;
+}
+
 static void log_writes_dtr(struct dm_target *ti)
 {
        struct log_writes_c *lc = ti->private;
@@ -873,9 +920,46 @@ static void log_writes_io_hints(struct dm_target *ti, struct queue_limits *limit
        limits->io_min = limits->physical_block_size;
 }
 
+static long log_writes_dax_direct_access(struct dm_target *ti, pgoff_t pgoff,
+                                        long nr_pages, void **kaddr, pfn_t *pfn)
+{
+       struct log_writes_c *lc = ti->private;
+       sector_t sector = pgoff * PAGE_SECTORS;
+       int ret;
+
+       ret = bdev_dax_pgoff(lc->dev->bdev, sector, nr_pages * PAGE_SIZE, &pgoff);
+       if (ret)
+               return ret;
+       return dax_direct_access(lc->dev->dax_dev, pgoff, nr_pages, kaddr, pfn);
+}
+
+static size_t log_writes_dax_copy_from_iter(struct dm_target *ti,
+                                           pgoff_t pgoff, void *addr, size_t bytes,
+                                           struct iov_iter *i)
+{
+       struct log_writes_c *lc = ti->private;
+       sector_t sector = pgoff * PAGE_SECTORS;
+       int err;
+
+       if (bdev_dax_pgoff(lc->dev->bdev, sector, ALIGN(bytes, PAGE_SIZE), &pgoff))
+               return 0;
+
+       /* Don't bother doing anything if logging has been disabled */
+       if (!lc->logging_enabled)
+               goto dax_copy;
+
+       err = log_dax(lc, sector, bytes, i);
+       if (err) {
+               DMWARN("Error %d logging DAX write", err);
+               return 0;
+       }
+dax_copy:
+       return dax_copy_from_iter(lc->dev->dax_dev, pgoff, addr, bytes, i);
+}
+
 static struct target_type log_writes_target = {
        .name   = "log-writes",
-       .version = {1, 0, 0},
+       .version = {1, 1, 0},
        .module = THIS_MODULE,
        .ctr    = log_writes_ctr,
        .dtr    = log_writes_dtr,
@@ -886,6 +970,8 @@ static struct target_type log_writes_target = {
        .message = log_writes_message,
        .iterate_devices = log_writes_iterate_devices,
        .io_hints = log_writes_io_hints,
+       .direct_access = log_writes_dax_direct_access,
+       .dax_copy_from_iter = log_writes_dax_copy_from_iter,
 };
 
 static int __init dm_log_writes_init(void)