btrfs: fix assertion failure during scrub due to block group reallocation

During a scrub, or device replace, we can race with block group removal
and allocation and trigger the following assertion failure:

[7526.385524] assertion failed: cache->start == chunk_offset, in fs/btrfs/scrub.c:3817
[7526.387351] ------------[ cut here ]------------
[7526.387373] kernel BUG at fs/btrfs/ctree.h:3599!
[7526.388001] invalid opcode: 0000 [#1] PREEMPT SMP DEBUG_PAGEALLOC PTI
[7526.388970] CPU: 2 PID: 1158150 Comm: btrfs Not tainted 5.17.0-rc8-btrfs-next-114 #4
[7526.390279] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.14.0-0-g155821a1990b-prebuilt.qemu.org 04/01/2014
[7526.392430] RIP: 0010:assertfail.constprop.0+0x18/0x1a [btrfs]
[7526.393520] Code: f3 48 c7 c7 20 (...)
[7526.396926] RSP: 0018:ffffb9154176bc40 EFLAGS: 00010246
[7526.397690] RAX: 0000000000000048 RBX: ffffa0db8a910000 RCX: 0000000000000000
[7526.398732] RDX: 0000000000000000 RSI: ffffffff9d7239a2 RDI: 00000000ffffffff
[7526.399766] RBP: ffffa0db8a911e10 R08: ffffffffa71a3ca0 R09: 0000000000000001
[7526.400793] R10: 0000000000000001 R11: 0000000000000000 R12: ffffa0db4b170800
[7526.401839] R13: 00000003494b0000 R14: ffffa0db7c55b488 R15: ffffa0db8b19a000
[7526.402874] FS:  00007f6c99c40640(0000) GS:ffffa0de6d200000(0000) knlGS:0000000000000000
[7526.404038] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[7526.405040] CR2: 00007f31b0882160 CR3: 000000014b38c004 CR4: 0000000000370ee0
[7526.406112] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
[7526.407148] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
[7526.408169] Call Trace:
[7526.408529]  <TASK>
[7526.408839]  scrub_enumerate_chunks.cold+0x11/0x79 [btrfs]
[7526.409690]  ? do_wait_intr_irq+0xb0/0xb0
[7526.410276]  btrfs_scrub_dev+0x226/0x620 [btrfs]
[7526.410995]  ? preempt_count_add+0x49/0xa0
[7526.411592]  btrfs_ioctl+0x1ab5/0x36d0 [btrfs]
[7526.412278]  ? __fget_files+0xc9/0x1b0
[7526.412825]  ? kvm_sched_clock_read+0x14/0x40
[7526.413459]  ? lock_release+0x155/0x4a0
[7526.414022]  ? __x64_sys_ioctl+0x83/0xb0
[7526.414601]  __x64_sys_ioctl+0x83/0xb0
[7526.415150]  do_syscall_64+0x3b/0xc0
[7526.415675]  entry_SYSCALL_64_after_hwframe+0x44/0xae
[7526.416408] RIP: 0033:0x7f6c99d34397
[7526.416931] Code: 3c 1c e8 1c ff (...)
[7526.419641] RSP: 002b:00007f6c99c3fca8 EFLAGS: 00000246 ORIG_RAX: 0000000000000010
[7526.420735] RAX: ffffffffffffffda RBX: 00005624e1e007b0 RCX: 00007f6c99d34397
[7526.421779] RDX: 00005624e1e007b0 RSI: 00000000c400941b RDI: 0000000000000003
[7526.422820] RBP: 0000000000000000 R08: 00007f6c99c40640 R09: 0000000000000000
[7526.423906] R10: 00007f6c99c40640 R11: 0000000000000246 R12: 00007fff746755de
[7526.424924] R13: 00007fff746755df R14: 0000000000000000 R15: 00007f6c99c40640
[7526.425950]  </TASK>

That assertion is relatively new, introduced with commit d04fbe19
("btrfs: scrub: cleanup the argument list of scrub_chunk()").

The block group we get at scrub_enumerate_chunks() can actually have a
start address that is smaller then the chunk offset we extracted from a
device extent item we got from the commit root of the device tree.
This is very rare, but it can happen due to a race with block group
removal and allocation. For example, the following steps show how this
can happen:

1) We are at transaction T, and we have the following blocks groups,
   sorted by their logical start address:

   [ bg A, start address A, length 1G (data) ]
   [ bg B, start address B, length 1G (data) ]
   (...)
   [ bg W, start address W, length 1G (data) ]

     --> logical address space hole of 256M,
         there used to be a 256M metadata block group here

   [ bg Y, start address Y, length 256M (metadata) ]

      --> Y matches W's end offset + 256M

   Block group Y is the block group with the highest logical address in
   the whole filesystem;

2) Block group Y is deleted and its extent mapping is removed by the call
   to remove_extent_mapping() made from btrfs_remove_block_group().

   So after this point, the last element of the mapping red black tree,
   its rightmost node, is the mapping for block group W;

3) While still at transaction T, a new data block group is allocated,
   with a length of 1G. When creating the block group we do a call to
   find_next_chunk(), which returns the logical start address for the
   new block group. This calls returns X, which corresponds to the
   end offset of the last block group, the rightmost node in the mapping
   red black tree (fs_info->mapping_tree), plus one.

   So we get a new block group that starts at logical address X and with
   a length of 1G. It spans over the whole logical range of the old block
   group Y, that was previously removed in the same transaction.

   However the device extent allocated to block group X is not the same
   device extent that was used by block group Y, and it also does not
   overlap that extent, which must be always the case because we allocate
   extents by searching through the commit root of the device tree
   (otherwise it could corrupt a filesystem after a power failure or
   an unclean shutdown in general), so the extent allocator is behaving
   as expected;

4) We have a task running scrub, currently at scrub_enumerate_chunks().
   There it searches for device extent items in the device tree, using
   its commit root. It finds a device extent item that was used by
   block group Y, and it extracts the value Y from that item into the
   local variable 'chunk_offset', using btrfs_dev_extent_chunk_offset();

   It then calls btrfs_lookup_block_group() to find block group for
   the logical address Y - since there's currently no block group that
   starts at that logical address, it returns block group X, because
   its range contains Y.

   This results in triggering the assertion:

      ASSERT(cache->start == chunk_offset);

   right before calling scrub_chunk(), as cache->start is X and
   chunk_offset is Y.

This is more likely to happen of filesystems not larger than 50G, because
for these filesystems we use a 256M size for metadata block groups and
a 1G size for data block groups, while for filesystems larger than 50G,
we use a 1G size for both data and metadata block groups (except for
zoned filesystems). It could also happen on any filesystem size due to
the fact that system block groups are always smaller (32M) than both
data and metadata block groups, but these are not frequently deleted, so
much less likely to trigger the race.

So make scrub skip any block group with a start offset that is less than
the value we expect, as that means it's a new block group that was created
in the current transaction. It's pointless to continue and try to scrub
its extents, because scrub searches for extents using the commit root, so
it won't find any. For a device replace, skip it as well for the same
reasons, and we don't need to worry about the possibility of extents of
the new block group not being to the new device, because we have the write
duplication setup done through btrfs_map_block().

Fixes: d04fbe19 ("btrfs: scrub: cleanup the argument list of scrub_chunk()")
CC: stable@vger.kernel.org # 5.17
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>

fs/btrfs/dev-replace.c

@@ -734,7 +734,12 @@ static int btrfs_dev_replace_start(struct btrfs_fs_info *fs_info,
 
 	btrfs_wait_ordered_roots(fs_info, U64_MAX, 0, (u64)-1);
 
-	/* Commit dev_replace state and reserve 1 item for it. */
+	/*
+	 * Commit dev_replace state and reserve 1 item for it.
+	 * This is crucial to ensure we won't miss copying extents for new block
+	 * groups that are allocated after we started the device replace, and
+	 * must be done after setting up the device replace state.
+	 */
 	trans = btrfs_start_transaction(root, 1);
 	if (IS_ERR(trans)) {
 		ret = PTR_ERR(trans);

fs/btrfs/scrub.c

@@ -3699,6 +3699,31 @@ int scrub_enumerate_chunks(struct scrub_ctx *sctx,
 		if (!cache)
 			goto skip;
 
+		ASSERT(cache->start <= chunk_offset);
+		/*
+		 * We are using the commit root to search for device extents, so
+		 * that means we could have found a device extent item from a
+		 * block group that was deleted in the current transaction. The
+		 * logical start offset of the deleted block group, stored at
+		 * @chunk_offset, might be part of the logical address range of
+		 * a new block group (which uses different physical extents).
+		 * In this case btrfs_lookup_block_group() has returned the new
+		 * block group, and its start address is less than @chunk_offset.
+		 *
+		 * We skip such new block groups, because it's pointless to
+		 * process them, as we won't find their extents because we search
+		 * for them using the commit root of the extent tree. For a device
+		 * replace it's also fine to skip it, we won't miss copying them
+		 * to the target device because we have the write duplication
+		 * setup through the regular write path (by btrfs_map_block()),
+		 * and we have committed a transaction when we started the device
+		 * replace, right after setting up the device replace state.
+		 */
+		if (cache->start < chunk_offset) {
+			btrfs_put_block_group(cache);
+			goto skip;
+		}
+
 		if (sctx->is_dev_replace && btrfs_is_zoned(fs_info)) {
 			spin_lock(&cache->lock);
 			if (!cache->to_copy) {
@@ -3822,7 +3847,6 @@ int scrub_enumerate_chunks(struct scrub_ctx *sctx,
 		dev_replace->item_needs_writeback = 1;
 		up_write(&dev_replace->rwsem);
 
-		ASSERT(cache->start == chunk_offset);
 		ret = scrub_chunk(sctx, cache, scrub_dev, found_key.offset,
 				  dev_extent_len);