debian/patches/pve/0053-Revert-block-rbd-implement-bdrv_co_block_status.patch

   1 From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
   2 From: Fabian Ebner <f.ebner@proxmox.com>
   3 Date: Tue, 17 May 2022 09:46:02 +0200
   4 Subject: [PATCH] Revert "block/rbd: implement bdrv_co_block_status"
   5
   6 During backup, bdrv_co_block_status is called for each block copy
   7 chunk. When RBD is used, the current implementation with
   8 rbd_diff_iterate2() using whole_object=true takes about linearly more
   9 time, depending on the image size. Since there are linearly more
  10 chunks, the slowdown is quadratic, becoming unacceptable for large
  11 images (starting somewhere between 500-1000 GiB in my testing).
  12
  13 This reverts commit 0347a8fd4c3faaedf119be04c197804be40a384b as a
  14 stop-gap measure, until it's clear how to make the implemenation
  15 more efficient.
  16
  17 Upstream bug report:
  18 https://gitlab.com/qemu-project/qemu/-/issues/1026
  19
  20 Signed-off-by: Fabian Ebner <f.ebner@proxmox.com>
  21 ---
  22  block/rbd.c | 112 ----------------------------------------------------
  23  1 file changed, 112 deletions(-)
  24
  25 diff --git a/block/rbd.c b/block/rbd.c
  26 index a4b8fb482c..3393b06a4e 100644
  27 --- a/block/rbd.c
  28 +++ b/block/rbd.c
  29 @@ -97,12 +97,6 @@ typedef struct RBDTask {
  30      int64_t ret;
  31  } RBDTask;
  32
  33 -typedef struct RBDDiffIterateReq {
  34 -    uint64_t offs;
  35 -    uint64_t bytes;
  36 -    bool exists;
  37 -} RBDDiffIterateReq;
  38 -
  39  static int qemu_rbd_connect(rados_t *cluster, rados_ioctx_t *io_ctx,
  40                              BlockdevOptionsRbd *opts, bool cache,
  41                              const char *keypairs, const char *secretid,
  42 @@ -1267,111 +1261,6 @@ static ImageInfoSpecific *qemu_rbd_get_specific_info(BlockDriverState *bs,
  43      return spec_info;
  44  }
  45
  46 -/*
  47 - * rbd_diff_iterate2 allows to interrupt the exection by returning a negative
  48 - * value in the callback routine. Choose a value that does not conflict with
  49 - * an existing exitcode and return it if we want to prematurely stop the
  50 - * execution because we detected a change in the allocation status.
  51 - */
  52 -#define QEMU_RBD_EXIT_DIFF_ITERATE2 -9000
  53 -
  54 -static int qemu_rbd_diff_iterate_cb(uint64_t offs, size_t len,
  55 -                                    int exists, void *opaque)
  56 -{
  57 -    RBDDiffIterateReq *req = opaque;
  58 -
  59 -    assert(req->offs + req->bytes <= offs);
  60 -    /*
  61 -     * we do not diff against a snapshot so we should never receive a callback
  62 -     * for a hole.
  63 -     */
  64 -    assert(exists);
  65 -
  66 -    if (!req->exists && offs > req->offs) {
  67 -        /*
  68 -         * we started in an unallocated area and hit the first allocated
  69 -         * block. req->bytes must be set to the length of the unallocated area
  70 -         * before the allocated area. stop further processing.
  71 -         */
  72 -        req->bytes = offs - req->offs;
  73 -        return QEMU_RBD_EXIT_DIFF_ITERATE2;
  74 -    }
  75 -
  76 -    if (req->exists && offs > req->offs + req->bytes) {
  77 -        /*
  78 -         * we started in an allocated area and jumped over an unallocated area,
  79 -         * req->bytes contains the length of the allocated area before the
  80 -         * unallocated area. stop further processing.
  81 -         */
  82 -        return QEMU_RBD_EXIT_DIFF_ITERATE2;
  83 -    }
  84 -
  85 -    req->bytes += len;
  86 -    req->exists = true;
  87 -
  88 -    return 0;
  89 -}
  90 -
  91 -static int coroutine_fn qemu_rbd_co_block_status(BlockDriverState *bs,
  92 -                                                 bool want_zero, int64_t offset,
  93 -                                                 int64_t bytes, int64_t *pnum,
  94 -                                                 int64_t *map,
  95 -                                                 BlockDriverState **file)
  96 -{
  97 -    BDRVRBDState *s = bs->opaque;
  98 -    int status, r;
  99 -    RBDDiffIterateReq req = { .offs = offset };
 100 -    uint64_t features, flags;
 101 -
 102 -    assert(offset + bytes <= s->image_size);
 103 -
 104 -    /* default to all sectors allocated */
 105 -    status = BDRV_BLOCK_DATA | BDRV_BLOCK_OFFSET_VALID;
 106 -    *map = offset;
 107 -    *file = bs;
 108 -    *pnum = bytes;
 109 -
 110 -    /* check if RBD image supports fast-diff */
 111 -    r = rbd_get_features(s->image, &features);
 112 -    if (r < 0) {
 113 -        return status;
 114 -    }
 115 -    if (!(features & RBD_FEATURE_FAST_DIFF)) {
 116 -        return status;
 117 -    }
 118 -
 119 -    /* check if RBD fast-diff result is valid */
 120 -    r = rbd_get_flags(s->image, &flags);
 121 -    if (r < 0) {
 122 -        return status;
 123 -    }
 124 -    if (flags & RBD_FLAG_FAST_DIFF_INVALID) {
 125 -        return status;
 126 -    }
 127 -
 128 -    r = rbd_diff_iterate2(s->image, NULL, offset, bytes, true, true,
 129 -                          qemu_rbd_diff_iterate_cb, &req);
 130 -    if (r < 0 && r != QEMU_RBD_EXIT_DIFF_ITERATE2) {
 131 -        return status;
 132 -    }
 133 -    assert(req.bytes <= bytes);
 134 -    if (!req.exists) {
 135 -        if (r == 0) {
 136 -            /*
 137 -             * rbd_diff_iterate2 does not invoke callbacks for unallocated
 138 -             * areas. This here catches the case where no callback was
 139 -             * invoked at all (req.bytes == 0).
 140 -             */
 141 -            assert(req.bytes == 0);
 142 -            req.bytes = bytes;
 143 -        }
 144 -        status = BDRV_BLOCK_ZERO | BDRV_BLOCK_OFFSET_VALID;
 145 -    }
 146 -
 147 -    *pnum = req.bytes;
 148 -    return status;
 149 -}
 150 -
 151  static int64_t qemu_rbd_getlength(BlockDriverState *bs)
 152  {
 153      BDRVRBDState *s = bs->opaque;
 154 @@ -1607,7 +1496,6 @@ static BlockDriver bdrv_rbd = {
 155  #ifdef LIBRBD_SUPPORTS_WRITE_ZEROES
 156      .bdrv_co_pwrite_zeroes  = qemu_rbd_co_pwrite_zeroes,
 157  #endif
 158 -    .bdrv_co_block_status   = qemu_rbd_co_block_status,
 159
 160      .bdrv_snapshot_create   = qemu_rbd_snap_create,
 161      .bdrv_snapshot_delete   = qemu_rbd_snap_remove,