raid5-ppl: load and recover the log
authorArtur Paszkiewicz <artur.paszkiewicz@intel.com>
Thu, 9 Mar 2017 09:00:01 +0000 (10:00 +0100)
committerShaohua Li <shli@fb.com>
Thu, 16 Mar 2017 23:55:55 +0000 (16:55 -0700)
Load the log from each disk when starting the array and recover if the
array is dirty.

The initial empty PPL is written by mdadm. When loading the log we
verify the header checksum and signature. For external metadata arrays
the signature is verified in userspace, so here we read it from the
header, verifying only if it matches on all disks, and use it later when
writing PPL.

In addition to the header checksum, each header entry also contains a
checksum of its partial parity data. If the header is valid, recovery is
performed for each entry until an invalid entry is found. If the array
is not degraded and recovery using PPL fully succeeds, there is no need
to resync the array because data and parity will be consistent, so in
this case resync will be disabled.

Due to compatibility with IMSM implementations on other systems, we
can't assume that the recovery data block size is always 4K. Writes
generated by MD raid5 don't have this issue, but when recovering PPL
written in other environments it is possible to have entries with
512-byte sector granularity. The recovery code takes this into account
and also the logical sector size of the underlying drives.

Signed-off-by: Artur Paszkiewicz <artur.paszkiewicz@intel.com>
Signed-off-by: Shaohua Li <shli@fb.com>
drivers/md/raid5-ppl.c
drivers/md/raid5.c

index db5b72b11594fd84cf54424b0cbe9f29fbf0331a..d336c024eef9d952174887dcf91e074334c58072 100644 (file)
@@ -103,6 +103,10 @@ struct ppl_conf {
        mempool_t *io_pool;
        struct bio_set *bs;
        mempool_t *meta_pool;
+
+       /* used only for recovery */
+       int recovered_entries;
+       int mismatch_count;
 };
 
 struct ppl_log {
@@ -514,6 +518,474 @@ void ppl_stripe_write_finished(struct stripe_head *sh)
                ppl_io_unit_finished(io);
 }
 
+static void ppl_xor(int size, struct page *page1, struct page *page2)
+{
+       struct async_submit_ctl submit;
+       struct dma_async_tx_descriptor *tx;
+       struct page *xor_srcs[] = { page1, page2 };
+
+       init_async_submit(&submit, ASYNC_TX_ACK|ASYNC_TX_XOR_DROP_DST,
+                         NULL, NULL, NULL, NULL);
+       tx = async_xor(page1, xor_srcs, 0, 2, size, &submit);
+
+       async_tx_quiesce(&tx);
+}
+
+/*
+ * PPL recovery strategy: xor partial parity and data from all modified data
+ * disks within a stripe and write the result as the new stripe parity. If all
+ * stripe data disks are modified (full stripe write), no partial parity is
+ * available, so just xor the data disks.
+ *
+ * Recovery of a PPL entry shall occur only if all modified data disks are
+ * available and read from all of them succeeds.
+ *
+ * A PPL entry applies to a stripe, partial parity size for an entry is at most
+ * the size of the chunk. Examples of possible cases for a single entry:
+ *
+ * case 0: single data disk write:
+ *   data0    data1    data2     ppl        parity
+ * +--------+--------+--------+           +--------------------+
+ * | ------ | ------ | ------ | +----+    | (no change)        |
+ * | ------ | -data- | ------ | | pp | -> | data1 ^ pp         |
+ * | ------ | -data- | ------ | | pp | -> | data1 ^ pp         |
+ * | ------ | ------ | ------ | +----+    | (no change)        |
+ * +--------+--------+--------+           +--------------------+
+ * pp_size = data_size
+ *
+ * case 1: more than one data disk write:
+ *   data0    data1    data2     ppl        parity
+ * +--------+--------+--------+           +--------------------+
+ * | ------ | ------ | ------ | +----+    | (no change)        |
+ * | -data- | -data- | ------ | | pp | -> | data0 ^ data1 ^ pp |
+ * | -data- | -data- | ------ | | pp | -> | data0 ^ data1 ^ pp |
+ * | ------ | ------ | ------ | +----+    | (no change)        |
+ * +--------+--------+--------+           +--------------------+
+ * pp_size = data_size / modified_data_disks
+ *
+ * case 2: write to all data disks (also full stripe write):
+ *   data0    data1    data2                parity
+ * +--------+--------+--------+           +--------------------+
+ * | ------ | ------ | ------ |           | (no change)        |
+ * | -data- | -data- | -data- | --------> | xor all data       |
+ * | ------ | ------ | ------ | --------> | (no change)        |
+ * | ------ | ------ | ------ |           | (no change)        |
+ * +--------+--------+--------+           +--------------------+
+ * pp_size = 0
+ *
+ * The following cases are possible only in other implementations. The recovery
+ * code can handle them, but they are not generated at runtime because they can
+ * be reduced to cases 0, 1 and 2:
+ *
+ * case 3:
+ *   data0    data1    data2     ppl        parity
+ * +--------+--------+--------+ +----+    +--------------------+
+ * | ------ | -data- | -data- | | pp |    | data1 ^ data2 ^ pp |
+ * | ------ | -data- | -data- | | pp | -> | data1 ^ data2 ^ pp |
+ * | -data- | -data- | -data- | | -- | -> | xor all data       |
+ * | -data- | -data- | ------ | | pp |    | data0 ^ data1 ^ pp |
+ * +--------+--------+--------+ +----+    +--------------------+
+ * pp_size = chunk_size
+ *
+ * case 4:
+ *   data0    data1    data2     ppl        parity
+ * +--------+--------+--------+ +----+    +--------------------+
+ * | ------ | -data- | ------ | | pp |    | data1 ^ pp         |
+ * | ------ | ------ | ------ | | -- | -> | (no change)        |
+ * | ------ | ------ | ------ | | -- | -> | (no change)        |
+ * | -data- | ------ | ------ | | pp |    | data0 ^ pp         |
+ * +--------+--------+--------+ +----+    +--------------------+
+ * pp_size = chunk_size
+ */
+static int ppl_recover_entry(struct ppl_log *log, struct ppl_header_entry *e,
+                            sector_t ppl_sector)
+{
+       struct ppl_conf *ppl_conf = log->ppl_conf;
+       struct mddev *mddev = ppl_conf->mddev;
+       struct r5conf *conf = mddev->private;
+       int block_size = ppl_conf->block_size;
+       struct page *page1;
+       struct page *page2;
+       sector_t r_sector_first;
+       sector_t r_sector_last;
+       int strip_sectors;
+       int data_disks;
+       int i;
+       int ret = 0;
+       char b[BDEVNAME_SIZE];
+       unsigned int pp_size = le32_to_cpu(e->pp_size);
+       unsigned int data_size = le32_to_cpu(e->data_size);
+
+       page1 = alloc_page(GFP_KERNEL);
+       page2 = alloc_page(GFP_KERNEL);
+
+       if (!page1 || !page2) {
+               ret = -ENOMEM;
+               goto out;
+       }
+
+       r_sector_first = le64_to_cpu(e->data_sector) * (block_size >> 9);
+
+       if ((pp_size >> 9) < conf->chunk_sectors) {
+               if (pp_size > 0) {
+                       data_disks = data_size / pp_size;
+                       strip_sectors = pp_size >> 9;
+               } else {
+                       data_disks = conf->raid_disks - conf->max_degraded;
+                       strip_sectors = (data_size >> 9) / data_disks;
+               }
+               r_sector_last = r_sector_first +
+                               (data_disks - 1) * conf->chunk_sectors +
+                               strip_sectors;
+       } else {
+               data_disks = conf->raid_disks - conf->max_degraded;
+               strip_sectors = conf->chunk_sectors;
+               r_sector_last = r_sector_first + (data_size >> 9);
+       }
+
+       pr_debug("%s: array sector first: %llu last: %llu\n", __func__,
+                (unsigned long long)r_sector_first,
+                (unsigned long long)r_sector_last);
+
+       /* if start and end is 4k aligned, use a 4k block */
+       if (block_size == 512 &&
+           (r_sector_first & (STRIPE_SECTORS - 1)) == 0 &&
+           (r_sector_last & (STRIPE_SECTORS - 1)) == 0)
+               block_size = STRIPE_SIZE;
+
+       /* iterate through blocks in strip */
+       for (i = 0; i < strip_sectors; i += (block_size >> 9)) {
+               bool update_parity = false;
+               sector_t parity_sector;
+               struct md_rdev *parity_rdev;
+               struct stripe_head sh;
+               int disk;
+               int indent = 0;
+
+               pr_debug("%s:%*s iter %d start\n", __func__, indent, "", i);
+               indent += 2;
+
+               memset(page_address(page1), 0, PAGE_SIZE);
+
+               /* iterate through data member disks */
+               for (disk = 0; disk < data_disks; disk++) {
+                       int dd_idx;
+                       struct md_rdev *rdev;
+                       sector_t sector;
+                       sector_t r_sector = r_sector_first + i +
+                                           (disk * conf->chunk_sectors);
+
+                       pr_debug("%s:%*s data member disk %d start\n",
+                                __func__, indent, "", disk);
+                       indent += 2;
+
+                       if (r_sector >= r_sector_last) {
+                               pr_debug("%s:%*s array sector %llu doesn't need parity update\n",
+                                        __func__, indent, "",
+                                        (unsigned long long)r_sector);
+                               indent -= 2;
+                               continue;
+                       }
+
+                       update_parity = true;
+
+                       /* map raid sector to member disk */
+                       sector = raid5_compute_sector(conf, r_sector, 0,
+                                                     &dd_idx, NULL);
+                       pr_debug("%s:%*s processing array sector %llu => data member disk %d, sector %llu\n",
+                                __func__, indent, "",
+                                (unsigned long long)r_sector, dd_idx,
+                                (unsigned long long)sector);
+
+                       rdev = conf->disks[dd_idx].rdev;
+                       if (!rdev) {
+                               pr_debug("%s:%*s data member disk %d missing\n",
+                                        __func__, indent, "", dd_idx);
+                               update_parity = false;
+                               break;
+                       }
+
+                       pr_debug("%s:%*s reading data member disk %s sector %llu\n",
+                                __func__, indent, "", bdevname(rdev->bdev, b),
+                                (unsigned long long)sector);
+                       if (!sync_page_io(rdev, sector, block_size, page2,
+                                       REQ_OP_READ, 0, false)) {
+                               md_error(mddev, rdev);
+                               pr_debug("%s:%*s read failed!\n", __func__,
+                                        indent, "");
+                               ret = -EIO;
+                               goto out;
+                       }
+
+                       ppl_xor(block_size, page1, page2);
+
+                       indent -= 2;
+               }
+
+               if (!update_parity)
+                       continue;
+
+               if (pp_size > 0) {
+                       pr_debug("%s:%*s reading pp disk sector %llu\n",
+                                __func__, indent, "",
+                                (unsigned long long)(ppl_sector + i));
+                       if (!sync_page_io(log->rdev,
+                                       ppl_sector - log->rdev->data_offset + i,
+                                       block_size, page2, REQ_OP_READ, 0,
+                                       false)) {
+                               pr_debug("%s:%*s read failed!\n", __func__,
+                                        indent, "");
+                               md_error(mddev, log->rdev);
+                               ret = -EIO;
+                               goto out;
+                       }
+
+                       ppl_xor(block_size, page1, page2);
+               }
+
+               /* map raid sector to parity disk */
+               parity_sector = raid5_compute_sector(conf, r_sector_first + i,
+                               0, &disk, &sh);
+               BUG_ON(sh.pd_idx != le32_to_cpu(e->parity_disk));
+               parity_rdev = conf->disks[sh.pd_idx].rdev;
+
+               BUG_ON(parity_rdev->bdev->bd_dev != log->rdev->bdev->bd_dev);
+               pr_debug("%s:%*s write parity at sector %llu, disk %s\n",
+                        __func__, indent, "",
+                        (unsigned long long)parity_sector,
+                        bdevname(parity_rdev->bdev, b));
+               if (!sync_page_io(parity_rdev, parity_sector, block_size,
+                               page1, REQ_OP_WRITE, 0, false)) {
+                       pr_debug("%s:%*s parity write error!\n", __func__,
+                                indent, "");
+                       md_error(mddev, parity_rdev);
+                       ret = -EIO;
+                       goto out;
+               }
+       }
+out:
+       if (page1)
+               __free_page(page1);
+       if (page2)
+               __free_page(page2);
+       return ret;
+}
+
+static int ppl_recover(struct ppl_log *log, struct ppl_header *pplhdr)
+{
+       struct ppl_conf *ppl_conf = log->ppl_conf;
+       struct md_rdev *rdev = log->rdev;
+       struct mddev *mddev = rdev->mddev;
+       sector_t ppl_sector = rdev->ppl.sector + (PPL_HEADER_SIZE >> 9);
+       struct page *page;
+       int i;
+       int ret = 0;
+
+       page = alloc_page(GFP_KERNEL);
+       if (!page)
+               return -ENOMEM;
+
+       /* iterate through all PPL entries saved */
+       for (i = 0; i < le32_to_cpu(pplhdr->entries_count); i++) {
+               struct ppl_header_entry *e = &pplhdr->entries[i];
+               u32 pp_size = le32_to_cpu(e->pp_size);
+               sector_t sector = ppl_sector;
+               int ppl_entry_sectors = pp_size >> 9;
+               u32 crc, crc_stored;
+
+               pr_debug("%s: disk: %d entry: %d ppl_sector: %llu pp_size: %u\n",
+                        __func__, rdev->raid_disk, i,
+                        (unsigned long long)ppl_sector, pp_size);
+
+               crc = ~0;
+               crc_stored = le32_to_cpu(e->checksum);
+
+               /* read parial parity for this entry and calculate its checksum */
+               while (pp_size) {
+                       int s = pp_size > PAGE_SIZE ? PAGE_SIZE : pp_size;
+
+                       if (!sync_page_io(rdev, sector - rdev->data_offset,
+                                       s, page, REQ_OP_READ, 0, false)) {
+                               md_error(mddev, rdev);
+                               ret = -EIO;
+                               goto out;
+                       }
+
+                       crc = crc32c_le(crc, page_address(page), s);
+
+                       pp_size -= s;
+                       sector += s >> 9;
+               }
+
+               crc = ~crc;
+
+               if (crc != crc_stored) {
+                       /*
+                        * Don't recover this entry if the checksum does not
+                        * match, but keep going and try to recover other
+                        * entries.
+                        */
+                       pr_debug("%s: ppl entry crc does not match: stored: 0x%x calculated: 0x%x\n",
+                                __func__, crc_stored, crc);
+                       ppl_conf->mismatch_count++;
+               } else {
+                       ret = ppl_recover_entry(log, e, ppl_sector);
+                       if (ret)
+                               goto out;
+                       ppl_conf->recovered_entries++;
+               }
+
+               ppl_sector += ppl_entry_sectors;
+       }
+
+       /* flush the disk cache after recovery if necessary */
+       ret = blkdev_issue_flush(rdev->bdev, GFP_KERNEL, NULL);
+out:
+       __free_page(page);
+       return ret;
+}
+
+static int ppl_write_empty_header(struct ppl_log *log)
+{
+       struct page *page;
+       struct ppl_header *pplhdr;
+       struct md_rdev *rdev = log->rdev;
+       int ret = 0;
+
+       pr_debug("%s: disk: %d ppl_sector: %llu\n", __func__,
+                rdev->raid_disk, (unsigned long long)rdev->ppl.sector);
+
+       page = alloc_page(GFP_NOIO | __GFP_ZERO);
+       if (!page)
+               return -ENOMEM;
+
+       pplhdr = page_address(page);
+       memset(pplhdr->reserved, 0xff, PPL_HDR_RESERVED);
+       pplhdr->signature = cpu_to_le32(log->ppl_conf->signature);
+       pplhdr->checksum = cpu_to_le32(~crc32c_le(~0, pplhdr, PAGE_SIZE));
+
+       if (!sync_page_io(rdev, rdev->ppl.sector - rdev->data_offset,
+                         PPL_HEADER_SIZE, page, REQ_OP_WRITE | REQ_FUA, 0,
+                         false)) {
+               md_error(rdev->mddev, rdev);
+               ret = -EIO;
+       }
+
+       __free_page(page);
+       return ret;
+}
+
+static int ppl_load_distributed(struct ppl_log *log)
+{
+       struct ppl_conf *ppl_conf = log->ppl_conf;
+       struct md_rdev *rdev = log->rdev;
+       struct mddev *mddev = rdev->mddev;
+       struct page *page;
+       struct ppl_header *pplhdr;
+       u32 crc, crc_stored;
+       u32 signature;
+       int ret = 0;
+
+       pr_debug("%s: disk: %d\n", __func__, rdev->raid_disk);
+
+       /* read PPL header */
+       page = alloc_page(GFP_KERNEL);
+       if (!page)
+               return -ENOMEM;
+
+       if (!sync_page_io(rdev, rdev->ppl.sector - rdev->data_offset,
+                         PAGE_SIZE, page, REQ_OP_READ, 0, false)) {
+               md_error(mddev, rdev);
+               ret = -EIO;
+               goto out;
+       }
+       pplhdr = page_address(page);
+
+       /* check header validity */
+       crc_stored = le32_to_cpu(pplhdr->checksum);
+       pplhdr->checksum = 0;
+       crc = ~crc32c_le(~0, pplhdr, PAGE_SIZE);
+
+       if (crc_stored != crc) {
+               pr_debug("%s: ppl header crc does not match: stored: 0x%x calculated: 0x%x\n",
+                        __func__, crc_stored, crc);
+               ppl_conf->mismatch_count++;
+               goto out;
+       }
+
+       signature = le32_to_cpu(pplhdr->signature);
+
+       if (mddev->external) {
+               /*
+                * For external metadata the header signature is set and
+                * validated in userspace.
+                */
+               ppl_conf->signature = signature;
+       } else if (ppl_conf->signature != signature) {
+               pr_debug("%s: ppl header signature does not match: stored: 0x%x configured: 0x%x\n",
+                        __func__, signature, ppl_conf->signature);
+               ppl_conf->mismatch_count++;
+               goto out;
+       }
+
+       /* attempt to recover from log if we are starting a dirty array */
+       if (!mddev->pers && mddev->recovery_cp != MaxSector)
+               ret = ppl_recover(log, pplhdr);
+out:
+       /* write empty header if we are starting the array */
+       if (!ret && !mddev->pers)
+               ret = ppl_write_empty_header(log);
+
+       __free_page(page);
+
+       pr_debug("%s: return: %d mismatch_count: %d recovered_entries: %d\n",
+                __func__, ret, ppl_conf->mismatch_count,
+                ppl_conf->recovered_entries);
+       return ret;
+}
+
+static int ppl_load(struct ppl_conf *ppl_conf)
+{
+       int ret = 0;
+       u32 signature = 0;
+       bool signature_set = false;
+       int i;
+
+       for (i = 0; i < ppl_conf->count; i++) {
+               struct ppl_log *log = &ppl_conf->child_logs[i];
+
+               /* skip missing drive */
+               if (!log->rdev)
+                       continue;
+
+               ret = ppl_load_distributed(log);
+               if (ret)
+                       break;
+
+               /*
+                * For external metadata we can't check if the signature is
+                * correct on a single drive, but we can check if it is the same
+                * on all drives.
+                */
+               if (ppl_conf->mddev->external) {
+                       if (!signature_set) {
+                               signature = ppl_conf->signature;
+                               signature_set = true;
+                       } else if (signature != ppl_conf->signature) {
+                               pr_warn("md/raid:%s: PPL header signature does not match on all member drives\n",
+                                       mdname(ppl_conf->mddev));
+                               ret = -EINVAL;
+                               break;
+                       }
+               }
+       }
+
+       pr_debug("%s: return: %d mismatch_count: %d recovered_entries: %d\n",
+                __func__, ret, ppl_conf->mismatch_count,
+                ppl_conf->recovered_entries);
+       return ret;
+}
+
 static void __ppl_exit_log(struct ppl_conf *ppl_conf)
 {
        clear_bit(MD_HAS_PPL, &ppl_conf->mddev->flags);
@@ -694,6 +1166,23 @@ int ppl_init_log(struct r5conf *conf)
                pr_warn("md/raid:%s: Volatile write-back cache should be disabled on all member drives when using PPL!\n",
                        mdname(mddev));
 
+       /* load and possibly recover the logs from the member disks */
+       ret = ppl_load(ppl_conf);
+
+       if (ret) {
+               goto err;
+       } else if (!mddev->pers &&
+                  mddev->recovery_cp == 0 && !mddev->degraded &&
+                  ppl_conf->recovered_entries > 0 &&
+                  ppl_conf->mismatch_count == 0) {
+               /*
+                * If we are starting a dirty array and the recovery succeeds
+                * without any issues, set the array as clean.
+                */
+               mddev->recovery_cp = MaxSector;
+               set_bit(MD_SB_CHANGE_CLEAN, &mddev->sb_flags);
+       }
+
        conf->log_private = ppl_conf;
 
        return 0;
index 6b86e0826afe1b889f19a5d8430946607771b495..78ed5748d33d7973cccb78fdd85016a4c2aa5823 100644 (file)
@@ -7357,7 +7357,10 @@ static int raid5_run(struct mddev *mddev)
 
        if (mddev->degraded > dirty_parity_disks &&
            mddev->recovery_cp != MaxSector) {
-               if (mddev->ok_start_degraded)
+               if (test_bit(MD_HAS_PPL, &mddev->flags))
+                       pr_crit("md/raid:%s: starting dirty degraded array with PPL.\n",
+                               mdname(mddev));
+               else if (mddev->ok_start_degraded)
                        pr_crit("md/raid:%s: starting dirty degraded array - data corruption possible.\n",
                                mdname(mddev));
                else {