EIO caused by encryption + recursive gang

Encrypted blocks can not have 3 DVAs, because they use the space of the
3rd DVA for the IV+salt.  zio_write_gang_block() takes this into
account, setting `gbh_copies` to no more than 2 in this case.  Gang
members BP's do not have the X (encrypted) bit set (nor do they have the
DMU level and type fields set), because encryption is not handled at
this level.  The gang block is reassembled, and then encryption (and
compression) are handled.

To check if this gang block is encrypted, the code in
zio_write_gang_block() checks `pio->io_bp`.  This is normally fine,
because the block that's being ganged is typically the encrypted BP.

The problem is that if there is "recursive ganging", where a gang member
is itself a gang block, then when zio_write_gang_block() is called to
create a gang block for a gang member, `pio->io_bp` is the gang member's
BP, which doesn't have the X bit set, so the number of DVA's is not
restricted to 2.  It should instead be looking at the the "gang leader",
i.e. the top-level gang block, to determine how many DVA's can be used,
to avoid a "NDVA's inversion" (where a child has more DVA's than its
parent).

gang leader BP: X (encrypted) bit set, 2 DVA's, IV+salt in 3rd DVA's
space:
```
DVA[0]=<1:...:100400> DVA[1]=<0:...:100400> salt=... iv=...
[L0 ZFS plain file] fletcher4 uncompressed encrypted LE
gang unique double size=100000L/100000P birth=... fill=1 cksum=...
```

leader's GBH contains a BP with gang bit set and 3 DVA's:
```
DVA[0]=<1:...:55600> DVA[1]=<0:...:55600>
[L0 unallocated] fletcher4 uncompressed unencrypted LE
contiguous unique double size=55600L/55600P birth=... fill=0 cksum=...

DVA[0]=<1:...:55600> DVA[1]=<0:...:55600>
[L0 unallocated] fletcher4 uncompressed unencrypted LE
contiguous unique double size=55600L/55600P birth=... fill=0 cksum=...

DVA[0]=<1:...:55600> DVA[1]=<0:...:55600> DVA[2]=<1:...:200>
[L0 unallocated] fletcher4 uncompressed unencrypted LE
gang unique double size=55400L/55400P birth=... fill=0 cksum=...
```

On nondebug bits, having the 3rd DVA in the gang block works for the
most part, because it's true that all 3 DVA's are available in the gang
member BP (in the GBH).  However, for accounting purposes, gang block
DVA's ASIZE include all the space allocated below them, i.e. the
512-byte gang block header (GBH) as well as the gang members below that.
We see that above where the gang leader BP is 1MB logical (and after
compression: 0x`100000P`), but the ASIZE of each DVA is 2 sectors (1KB)
more than 1MB (0x`100400`).

Since thre are 3 copies of a block below it, we increment the ATIME of
the 3rd DVA of the gang leader by the space used by the 3rd DVA of the
child (1 sector, in this case).  But there isn't really a 3rd DVA of the
parent; the salt is stored in place of the 3rd DVA's ASIZE.

So when zio_write_gang_member_ready() increments the parent's BP's
`DVA[2]`'s ASIZE, it's actually incrementing the parent's salt.  When we
later try to read the encrypted recursively-ganged block, the salt
doesn't match what we used to write it, so MAC verification fails and we
get an EIO.

```
zio_encrypt():  encrypted 515/2/0/403 salt: 25 25 bb 9d ad d6 cd 89
zio_decrypt(): decrypting 515/2/0/403 salt: 26 25 bb 9d ad d6 cd 89
```

This commit addresses the problem by not increasing the number of copies
of the GBH beyond 2 (even for non-encrypted blocks).  This simplifies
the logic while maintaining the ability to traverse all metadata
(including gang blocks) even if one copy is lost.  (Note that 3 copies
of the GBH will still be created if requested, e.g. for `copies=3` or
MOS blocks.)  Additionally, the code that increments the parent's DVA's
ASIZE is made to check the parent DVA's NDVAS even on nondebug bits.  So
if there's a similar bug in the future, it will cause a panic when
trying to write, rather than corrupting the parent BP and causing an
error when reading.

Reviewed-by: Brian Behlendorf <[email protected]>
Co-authored-by: Brian Behlendorf <[email protected]>
Signed-off-by: Matthew Ahrens <[email protected]>
Caused-by: #14356
Closes #14440
Closes #14413

1 files changed, 9 insertions, 7 deletions

diff --git a/module/zfs/zio.c b/module/zfs/zio.c
index 5d7ed6d58..d888a584a 100644
--- a/module/zfs/zio.c
+++ b/module/zfs/zio.c
@@ -2778,7 +2778,7 @@ zio_write_gang_member_ready(zio_t *zio)
 	ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
 	ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
 	ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
-	ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
+	VERIFY3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
 
 	mutex_enter(&pio->io_lock);
 	for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
@@ -2816,18 +2816,20 @@ zio_write_gang_block(zio_t *pio, metaslab_class_t *mc)
 	uint64_t resid = pio->io_size;
 	uint64_t lsize;
 	int copies = gio->io_prop.zp_copies;
-	int gbh_copies;
 	zio_prop_t zp;
 	int error;
 	boolean_t has_data = !(pio->io_flags & ZIO_FLAG_NODATA);
 
 	/*
-	 * encrypted blocks need DVA[2] free so encrypted gang headers can't
-	 * have a third copy.
+	 * If one copy was requested, store 2 copies of the GBH, so that we
+	 * can still traverse all the data (e.g. to free or scrub) even if a
+	 * block is damaged.  Note that we can't store 3 copies of the GBH in
+	 * all cases, e.g. with encryption, which uses DVA[2] for the IV+salt.
 	 */
-	gbh_copies = MIN(copies + 1, spa_max_replication(spa));
-	if (BP_IS_ENCRYPTED(bp) && gbh_copies >= SPA_DVAS_PER_BP)
-		gbh_copies = SPA_DVAS_PER_BP - 1;
+	int gbh_copies = copies;
+	if (gbh_copies == 1) {
+		gbh_copies = MIN(2, spa_max_replication(spa));
+	}
 
 	int flags = METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER;
 	if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) {