Hi Chris,
On Tue, Nov 03, 2026 at 06:50:28PM +0000, Chris Wiseman wrote:
> verity_ctr() creates a dedicated dm_io_client (v->io) for every
> dm-verity target, but its only caller is verity_recheck(), which reads
> a single data block into a page from v->recheck_pool after a hash
> mismatch. Because dm_io_client_create() preallocates a mempool of 16
> bios and 16 biovec-max (4 KiB) vectors per target, each dm-verity
> target permanently pins ~70 KiB of unreclaimable slab that a single-page
> read never uses (~2.4 MiB across 35 targets on Android, or ~4.1 MiB
> across 59 targets).
>
> Replace dm_io() in verity_recheck() with bdev_rw_virt() and remove
> v->io. bdev_rw_virt() puts the bio and single bio_vec on the stack and
> waits for completion, while keeping v->recheck_pool per-target.
>
> Fixes: 9177f3c0dea6 ("dm-verity: recheck the hash after a failure")
> Link:
> https://lore.kernel.org/dm-devel/[email protected]/
> Signed-off-by: Chris Wiseman <[email protected]>
> ---
> TEST (ARM64 QEMU, 59 dm-verity targets, sha256, 4K blocks, mem_profiling=1):
>
> /proc/allocinfo growth from 0 to 35 targets (bytes):
> BASE PATCHED DELTA
> bio.c:biovec_init_pool 4,587,520 2,293,760 -2,293,760
> bio.c:bioset_init 1,196,160 1,052,800 -143,360
> dm-io.c:dm_io_client_create 107,520 53,760 -53,760
> mempool.c:mempool_init_node 36,680 23,240 -13,440
> Total (35 targets) 5,927,880 3,423,560 -2,504,320 (-71,552
> B/target)
>
> Unreclaimable slab saved by target count:
> 8 targets: 572,416 B ( 559 KiB)
> 15 targets: 1,073,280 B (1,048 KiB)
> 35 targets: 2,504,320 B (2,446 KiB) [SUnreclaim: BASE +8,436 kB vs PATCHED
> +5,960 kB]
> 59 targets: 4,221,568 B (4,123 KiB)
>
> drivers/md/dm-verity-target.c | 24 +++---------------------
> drivers/md/dm-verity.h | 2 --
> 2 files changed, 3 insertions(+), 23 deletions(-)
Looks like a nice improvement!
Reviewed-by: Sami Tolvanen <[email protected]>
Sami
