[Devel] [PATCH VZ10] drivers/vhost/blk: fix vhost_blk_req_done() softirq handling

[Andrey Zhadchenko]

vhost_blk_req_done() does bio_release_pages(), which may sleep on
folio_lock() when we need to dirty pages. So we get this:

[ 9120.053244] BUG: scheduling while atomic: swapper/1/0/0x00000102
...
[ 9120.079613] Hardware name: Acronis OpenStack Compute/Virtuozzo, BIOS 1.16.1-1.vz9.2 04/01/2014
[ 9120.079629] Call Trace:
[ 9120.079632]  <IRQ>
[ 9120.079638]  dump_stack_lvl+0x4e/0x70
[ 9120.079646]  __schedule_bug.cold+0x3e/0x4a
[ 9120.079650]  schedule_debug.isra.0+0x93/0xc0
[ 9120.079655]  __schedule+0x7a/0x630
[ 9120.079659]  ? enqueue_task_fair+0x150/0x710
[ 9120.079664]  schedule+0x27/0x80
[ 9120.079666]  io_schedule+0x46/0x70
[ 9120.079669]  folio_wait_bit_common+0x13a/0x340
[ 9120.079686]  ? __pfx_wake_page_function+0x10/0x10
[ 9120.079690]  __bio_release_pages+0x25b/0x280
[ 9120.079698]  vhost_blk_req_done+0x98/0xa0 [vhost_blk]
[ 9120.079702]  blk_update_request+0x17c/0x420
[ 9120.079707]  blk_mq_end_request+0x1c/0x30
[ 9120.079711]  dm_softirq_done+0x158/0x280 [dm_mod]
[ 9120.079737]  blk_complete_reqs+0x40/0x50
[ 9120.079740]  handle_softirqs+0xe5/0x2a0
[ 9120.079744]  __irq_exit_rcu+0xbd/0xe0
[ 9120.079746]  sysvec_call_function_single+0x71/0x90
[ 9120.079750]  </IRQ>
[ 9120.079751]  <TASK>
[ 9120.079752]  asm_sysvec_call_function_single+0x1a/0x20

Kernel already has a suitable helper: bio_check_pages_dirty(). It
takes bio ownership and dirties pages/puts bio later in some
workqueue (see bio_set_pages_dirty() annotation).
Also we should manually mark the pages dirty from
vhost_blk_bio_send(), as they can be modified through DMA.

While at here, move request posting below bio_put. I don't expect
the worker could process and reuse the request before we access
req->bb and req->bi_opf, but better be safe then sorry.

https://virtuozzo.atlassian.net/browse/VSTOR-141284
Signed-off-by: Andrey Zhadchenko <andrey.zhadchenko at virtuozzo.com>
---
 drivers/vhost/blk.c | 24 +++++++++++++++---------
 1 file changed, 15 insertions(+), 9 deletions(-)

diff --git a/drivers/vhost/blk.c b/drivers/vhost/blk.c
index c03945ac04234..dac03566bfca5 100644
--- a/drivers/vhost/blk.c
+++ b/drivers/vhost/blk.c
@@ -156,19 +156,22 @@ static void vhost_blk_req_done(struct bio *bio)
 	if (err)
 		req->bio_err = err;
 
-	if (atomic_dec_and_test(&req->bio_nr)) {
-		llist_add(&req->llnode, &req->blk_vq->llhead);
-		vhost_vq_work_queue(&req->blk_vq->vq, &req->blk_vq->work);
-	}
-
 	/*
 	 * Bounce buffer adds kvec to bio (instead of user backed memory),
 	 * so there is no reference/pin to bio pages in this case.
 	 */
-	if (!req->bb)
-		bio_release_pages(bio, !req->bi_opf);
+	if (!req->bb && req->bi_opf == REQ_OP_READ) {
+		bio_check_pages_dirty(bio);
+	} else {
+		if (!req->bb)
+			bio_release_pages(bio, false);
+		bio_put(bio);
+	}
 
-	bio_put(bio);
+	if (atomic_dec_and_test(&req->bio_nr)) {
+		llist_add(&req->llnode, &req->blk_vq->llhead);
+		vhost_vq_work_queue(&req->blk_vq->vq, &req->blk_vq->work);
+	}
 }
 
 static void vhost_blk_req_cleanup(struct vhost_blk_req *req)
@@ -387,8 +390,11 @@ static inline void vhost_blk_bio_send(struct vhost_blk_req *req)
 
 	bio_nr = atomic_read(&req->bio_nr);
 	blk_start_plug(&plug);
-	for (i = 0; i < bio_nr; i++)
+	for (i = 0; i < bio_nr; i++) {
+		if (!req->bb && req->bi_opf == REQ_OP_READ)
+			bio_set_pages_dirty(req->bio[i]);
 		submit_bio(req->bio[i]);
+	}
 
 	blk_finish_plug(&plug);
 }
-- 
2.43.5