md: Use REQ_FAILFAST_* on metadata writes where appropriate
authorNeilBrown <neilb@suse.com>
Fri, 18 Nov 2016 05:16:11 +0000 (16:16 +1100)
committerShaohua Li <shli@fb.com>
Tue, 22 Nov 2016 17:11:33 +0000 (09:11 -0800)
This can only be supported on personalities which ensure
that md_error() never causes an array to enter the 'failed'
state.  i.e. if marking a device Faulty would cause some
data to be inaccessible, the device is status is left as
non-Faulty.  This is true for RAID1 and RAID10.

If we get a failure writing metadata but the device doesn't
fail, it must be the last device so we re-write without
FAILFAST to improve chance of success.  We also flag the
device as LastDev so that future metadata updates don't
waste time on failfast writes.

Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Shaohua Li <shli@fb.com>
drivers/md/bitmap.c
drivers/md/md.c
drivers/md/md.h
drivers/md/raid1.c
drivers/md/raid10.c

index cf77cbf9ed22b8a38104eff941ab4c53c15e3cf1..c4621571b718e8ddccb96df475c29164a91af7d1 100644 (file)
@@ -209,11 +209,13 @@ static struct md_rdev *next_active_rdev(struct md_rdev *rdev, struct mddev *mdde
 
 static int write_sb_page(struct bitmap *bitmap, struct page *page, int wait)
 {
-       struct md_rdev *rdev = NULL;
+       struct md_rdev *rdev;
        struct block_device *bdev;
        struct mddev *mddev = bitmap->mddev;
        struct bitmap_storage *store = &bitmap->storage;
 
+restart:
+       rdev = NULL;
        while ((rdev = next_active_rdev(rdev, mddev)) != NULL) {
                int size = PAGE_SIZE;
                loff_t offset = mddev->bitmap_info.offset;
@@ -269,8 +271,8 @@ static int write_sb_page(struct bitmap *bitmap, struct page *page, int wait)
                               page);
        }
 
-       if (wait)
-               md_super_wait(mddev);
+       if (wait && md_super_wait(mddev) < 0)
+               goto restart;
        return 0;
 
  bad_alignment:
@@ -428,6 +430,13 @@ static void bitmap_wait_writes(struct bitmap *bitmap)
                wait_event(bitmap->write_wait,
                           atomic_read(&bitmap->pending_writes)==0);
        else
+               /* Note that we ignore the return value.  The writes
+                * might have failed, but that would just mean that
+                * some bits which should be cleared haven't been,
+                * which is safe.  The relevant bitmap blocks will
+                * probably get written again, but there is no great
+                * loss if they aren't.
+                */
                md_super_wait(bitmap->mddev);
 }
 
index 2cf0e89cc56c9ae051b0a92ebbc6200175001c33..62f3fb948b3e4b45aec046dfe7671a83d02334bf 100644 (file)
@@ -727,7 +727,13 @@ static void super_written(struct bio *bio)
        if (bio->bi_error) {
                pr_err("md: super_written gets error=%d\n", bio->bi_error);
                md_error(mddev, rdev);
-       }
+               if (!test_bit(Faulty, &rdev->flags)
+                   && (bio->bi_opf & MD_FAILFAST)) {
+                       set_bit(MD_NEED_REWRITE, &mddev->flags);
+                       set_bit(LastDev, &rdev->flags);
+               }
+       } else
+               clear_bit(LastDev, &rdev->flags);
 
        if (atomic_dec_and_test(&mddev->pending_writes))
                wake_up(&mddev->sb_wait);
@@ -744,7 +750,13 @@ void md_super_write(struct mddev *mddev, struct md_rdev *rdev,
         * if zero is reached.
         * If an error occurred, call md_error
         */
-       struct bio *bio = bio_alloc_mddev(GFP_NOIO, 1, mddev);
+       struct bio *bio;
+       int ff = 0;
+
+       if (test_bit(Faulty, &rdev->flags))
+               return;
+
+       bio = bio_alloc_mddev(GFP_NOIO, 1, mddev);
 
        atomic_inc(&rdev->nr_pending);
 
@@ -753,16 +765,24 @@ void md_super_write(struct mddev *mddev, struct md_rdev *rdev,
        bio_add_page(bio, page, size, 0);
        bio->bi_private = rdev;
        bio->bi_end_io = super_written;
-       bio_set_op_attrs(bio, REQ_OP_WRITE, WRITE_FLUSH_FUA);
+
+       if (test_bit(MD_FAILFAST_SUPPORTED, &mddev->flags) &&
+           test_bit(FailFast, &rdev->flags) &&
+           !test_bit(LastDev, &rdev->flags))
+               ff = MD_FAILFAST;
+       bio_set_op_attrs(bio, REQ_OP_WRITE, WRITE_FLUSH_FUA | ff);
 
        atomic_inc(&mddev->pending_writes);
        submit_bio(bio);
 }
 
-void md_super_wait(struct mddev *mddev)
+int md_super_wait(struct mddev *mddev)
 {
        /* wait for all superblock writes that were scheduled to complete */
        wait_event(mddev->sb_wait, atomic_read(&mddev->pending_writes)==0);
+       if (test_and_clear_bit(MD_NEED_REWRITE, &mddev->flags))
+               return -EAGAIN;
+       return 0;
 }
 
 int sync_page_io(struct md_rdev *rdev, sector_t sector, int size,
@@ -1334,9 +1354,10 @@ super_90_rdev_size_change(struct md_rdev *rdev, sector_t num_sectors)
        if (IS_ENABLED(CONFIG_LBDAF) && (u64)num_sectors >= (2ULL << 32) &&
            rdev->mddev->level >= 1)
                num_sectors = (sector_t)(2ULL << 32) - 2;
-       md_super_write(rdev->mddev, rdev, rdev->sb_start, rdev->sb_size,
+       do {
+               md_super_write(rdev->mddev, rdev, rdev->sb_start, rdev->sb_size,
                       rdev->sb_page);
-       md_super_wait(rdev->mddev);
+       } while (md_super_wait(rdev->mddev) < 0);
        return num_sectors;
 }
 
@@ -1877,9 +1898,10 @@ super_1_rdev_size_change(struct md_rdev *rdev, sector_t num_sectors)
        sb->data_size = cpu_to_le64(num_sectors);
        sb->super_offset = rdev->sb_start;
        sb->sb_csum = calc_sb_1_csum(sb);
-       md_super_write(rdev->mddev, rdev, rdev->sb_start, rdev->sb_size,
-                      rdev->sb_page);
-       md_super_wait(rdev->mddev);
+       do {
+               md_super_write(rdev->mddev, rdev, rdev->sb_start, rdev->sb_size,
+                              rdev->sb_page);
+       } while (md_super_wait(rdev->mddev) < 0);
        return num_sectors;
 
 }
@@ -2416,6 +2438,7 @@ repeat:
 
        if (mddev->queue)
                blk_add_trace_msg(mddev->queue, "md md_update_sb");
+rewrite:
        bitmap_update_sb(mddev->bitmap);
        rdev_for_each(rdev, mddev) {
                char b[BDEVNAME_SIZE];
@@ -2447,7 +2470,8 @@ repeat:
                        /* only need to write one superblock... */
                        break;
        }
-       md_super_wait(mddev);
+       if (md_super_wait(mddev) < 0)
+               goto rewrite;
        /* if there was a failure, MD_CHANGE_DEVS was set, and we re-write super */
 
        if (mddev_is_clustered(mddev) && ret == 0)
index bc6712ef8c8190f784eaab83d41370915a54481d..5c08f84101fad7daa002e0b5be3d570c7500fd6f 100644 (file)
 
 #define MaxSector (~(sector_t)0)
 
+/*
+ * These flags should really be called "NO_RETRY" rather than
+ * "FAILFAST" because they don't make any promise about time lapse,
+ * only about the number of retries, which will be zero.
+ * REQ_FAILFAST_DRIVER is not included because
+ * Commit: 4a27446f3e39 ("[SCSI] modify scsi to handle new fail fast flags.")
+ * seems to suggest that the errors it avoids retrying should usually
+ * be retried.
+ */
+#define        MD_FAILFAST     (REQ_FAILFAST_DEV | REQ_FAILFAST_TRANSPORT)
 /*
  * MD's 'extended' device
  */
@@ -177,6 +187,10 @@ enum flag_bits {
                                 * It is expects that no bad block log
                                 * is present.
                                 */
+       LastDev,                /* Seems to be the last working dev as
+                                * it didn't fail, so don't use FailFast
+                                * any more for metadata
+                                */
 };
 
 static inline int is_badblock(struct md_rdev *rdev, sector_t s, int sectors,
@@ -213,6 +227,11 @@ enum mddev_flags {
        MD_CLUSTER_RESYNC_LOCKED, /* cluster raid only, which means node
                                   * already took resync lock, need to
                                   * release the lock */
+       MD_FAILFAST_SUPPORTED,  /* Using MD_FAILFAST on metadata writes is
+                                * supported as calls to md_error() will
+                                * never cause the array to become failed.
+                                */
+       MD_NEED_REWRITE,        /* metadata write needs to be repeated */
 };
 #define MD_UPDATE_SB_FLAGS (BIT(MD_CHANGE_DEVS) | \
                            BIT(MD_CHANGE_CLEAN) | \
@@ -628,7 +647,7 @@ extern int mddev_congested(struct mddev *mddev, int bits);
 extern void md_flush_request(struct mddev *mddev, struct bio *bio);
 extern void md_super_write(struct mddev *mddev, struct md_rdev *rdev,
                           sector_t sector, int size, struct page *page);
-extern void md_super_wait(struct mddev *mddev);
+extern int md_super_wait(struct mddev *mddev);
 extern int sync_page_io(struct md_rdev *rdev, sector_t sector, int size,
                        struct page *page, int op, int op_flags,
                        bool metadata_op);
index d24adc50a31fc0bb7eb97df3b38454bd0e6bdb68..4006a9be2eab7033243a69fc375620c4b6192aa0 100644 (file)
@@ -2988,6 +2988,7 @@ static int raid1_run(struct mddev *mddev)
        mddev->thread = conf->thread;
        conf->thread = NULL;
        mddev->private = conf;
+       set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
 
        md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
 
index bd8c884d4596bdbeaad3ee8c9804d8c7dbde9c77..af50866f5ffbd66c086076f1afc6f0e090b32daa 100644 (file)
@@ -3729,6 +3729,7 @@ static int raid10_run(struct mddev *mddev)
        size = raid10_size(mddev, 0, 0);
        md_set_array_sectors(mddev, size);
        mddev->resync_max_sectors = size;
+       set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
 
        if (mddev->queue) {
                int stripe = conf->geo.raid_disks *