[PATCHv5 0/4] block integrity: directly map user space addresses

[Kanchan Joshi]

On 12/1/2023 3:23 AM, Keith Busch wrote:
> From: Keith Busch<kbusch at kernel.org>

This causes a regression (existed in previous version too).
System freeze on issuing single read/write io that used to work fine 
earlier:
fio -iodepth=1 -rw=randread -ioengine=io_uring_cmd -cmd_type=nvme 
-bs=4096 -numjobs=1 -size=4096 -filename=/dev/ng0n1 -md_per_io_size=8 
-name=pt

This is because we pin one bvec during submission, but unpin 4 on 
completion. bio_integrity_unpin_bvec() uses bip->bip_max_vcnt, which is 
set to 4 (equal to BIO_INLINE_VECS) in this case.

To use bip_max_vcnt the way this series uses, we need below patch/fix:

diff --git a/block/bio-integrity.c b/block/bio-integrity.c
index 674a2c80454b..feef615e2c9c 100644
--- a/block/bio-integrity.c
+++ b/block/bio-integrity.c
@@ -69,15 +69,15 @@ struct bio_integrity_payload 
*bio_integrity_alloc(struct bio *bio,

         memset(bip, 0, sizeof(*bip));

+       /* always report as many vecs as asked explicitly, not inline 
vecs */
+       bip->bip_max_vcnt = nr_vecs;
         if (nr_vecs > inline_vecs) {
-               bip->bip_max_vcnt = nr_vecs;
                 bip->bip_vec = bvec_alloc(&bs->bvec_integrity_pool,
                                           &bip->bip_max_vcnt, gfp_mask);
                 if (!bip->bip_vec)
                         goto err;
         } else {
                 bip->bip_vec = bip->bip_inline_vecs;
-               bip->bip_max_vcnt = inline_vecs;
         }

         bip->bip_bio = bio;