Skip to contentSkip to conversation

io_uring/rsrc: fix slab-out-of-bounds in io_buffer_register_bvec

PATCH Superseded veygax 6 participants 13 messages Dec 17, 2025 – Dec 22, 2025 io_uring

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore

veygax

author
#0 lore raw

PATCH 1/1 io_uring/rsrc: fix slab-out-of-bounds in io_buffer_register_bvec

From: Evan Lambert <[email protected]>

The function io_buffer_register_bvec() calculates the allocation size for the io_mapped_ubuf based on blk_rq_nr_phys_segments(rq). This function calculates the number of scatter-gather elements after megine physically contiguous pages.

However, the subsequent loop uses rq_for_each_bvec() to populate the array, which iterates over every individual bio_vec in the request, regardless of physical contiguity.

If a request has multiple bio_vec entries that are physically contiguous, blk_rq_nr_phys_segments() returns a value smaller than the total number of bio_vecs. This leads to a slab-out-of-bounds write.

The path is reachable from userspace via the ublk driver when a server issues a UBLK_IO_REGISTER_IO_BUF command. This requires the UBLK_F_SUPPORT_ZERO_COPY flag which is protected by CAP_NET_ADMIN.

Fix this by calculating the total number of bio_vecs by iterating over the request's bios and summing their bi_vcnt.

KASAN report:

[18:01:50] BUG: KASAN: slab-out-of-bounds in io_buffer_register_bvec+0x813/0xb80 [18:01:50] Write of size 8 at addr ffff88800223b238 by task kunit_try_catch/27 [18:01:50]
[18:01:50] CPU: 0 UID: 0 PID: 27 Comm: kunit_try_catch Tainted: G N 6.19.0-rc1-g346af1a0c65a-dirty #44 PREEMPT(none)
[18:01:50] Tainted: [N]=TEST
[18:01:50] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 0.1 11/11/2019 [18:01:50] Call Trace:

[18:01:50]  <TASK>
[18:01:50]  dump_stack_lvl+0x4d/0x70
[18:01:50]  print_report+0x151/0x4c0
[18:01:50]  ? __pfx__raw_spin_lock_irqsave+0x10/0x10
[18:01:50]  ? io_buffer_register_bvec+0x813/0xb80
[18:01:50]  kasan_report+0xec/0x120
[18:01:50]  ? io_buffer_register_bvec+0x813/0xb80
[18:01:50]  io_buffer_register_bvec+0x813/0xb80
[18:01:50]  io_buffer_register_bvec_overflow_test+0x4e6/0x9b0
[18:01:50]  ? __pfx_io_buffer_register_bvec_overflow_test+0x10/0x10
[18:01:50]  ? __pfx_pick_next_task_fair+0x10/0x10
[18:01:50]  ? _raw_spin_lock+0x7e/0xd0
[18:01:50]  ? finish_task_switch.isra.0+0x19a/0x650
[18:01:50]  ? __pfx_read_tsc+0x10/0x10
[18:01:50]  ? ktime_get_ts64+0x79/0x240
[18:01:50]  kunit_try_run_case+0x19b/0x2c0
[18:01:50]  ? __pfx_kunit_try_run_case+0x10/0x10
[18:01:50]  ? __pfx_kunit_generic_run_threadfn_adapter+0x10/0x10
[18:01:50]  kunit_generic_run_threadfn_adapter+0x80/0xf0
[18:01:50]  kthread+0x323/0x670
[18:01:50]  ? __pfx_kthread+0x10/0x10
[18:01:50]  ? __pfx__raw_spin_lock_irq+0x10/0x10
[18:01:50]  ? __pfx_kthread+0x10/0x10
[18:01:50]  ret_from_fork+0x329/0x420
[18:01:50]  ? __pfx_ret_from_fork+0x10/0x10
[18:01:50]  ? __switch_to+0xa0f/0xd40
[18:01:50]  ? __pfx_kthread+0x10/0x10
[18:01:50]  ret_from_fork_asm+0x1a/0x30
[18:01:50]  </TASK>
[18:01:50]
[18:01:50] Allocated by task 27:
[18:01:50]  kasan_save_stack+0x30/0x50
[18:01:50]  kasan_save_track+0x14/0x30
[18:01:50]  __kasan_kmalloc+0x7f/0x90
[18:01:50]  io_cache_alloc_new+0x35/0xc0
[18:01:50]  io_buffer_register_bvec+0x196/0xb80
[18:01:50]  io_buffer_register_bvec_overflow_test+0x4e6/0x9b0
[18:01:50]  kunit_try_run_case+0x19b/0x2c0
[18:01:50]  kunit_generic_run_threadfn_adapter+0x80/0xf0
[18:01:50]  kthread+0x323/0x670
[18:01:50]  ret_from_fork+0x329/0x420
[18:01:50]  ret_from_fork_asm+0x1a/0x30

[18:01:50]
[18:01:50] The buggy address belongs to the object at ffff88800223b000 [18:01:50] which belongs to the cache kmalloc-1k of size 1024 [18:01:50] The buggy address is located 0 bytes to the right of [18:01:50] allocated 568-byte region [ffff88800223b000, ffff88800223b238) [18:01:50]
[18:01:50] The buggy address belongs to the physical page:
[18:01:50] page: refcount:0 mapcount:0 mapping:0000000000000000 index:0x0 pfn:0x2238 [18:01:50] head: order:2 mapcount:0 entire_mapcount:0 nr_pages_mapped:0 pincount:0 [18:01:50] flags: 0x4000000000000040(head|zone=1)
[18:01:50] page_type: f5(slab)
[18:01:50] raw: 4000000000000040 ffff888001041dc0 dead000000000122 0000000000000000 [18:01:50] raw: 0000000000000000 0000000080080008 00000000f5000000 0000000000000000 [18:01:50] head: 4000000000000040 ffff888001041dc0 dead000000000122 0000000000000000 [18:01:50] head: 0000000000000000 0000000080080008 00000000f5000000 0000000000000000 [18:01:50] head: 4000000000000002 ffffea0000088e01 00000000ffffffff 00000000ffffffff [18:01:50] head: 0000000000000000 0000000000000000 00000000ffffffff 0000000000000000 [18:01:50] page dumped because: kasan: bad access detected
[18:01:50]
[18:01:50] Memory state around the buggy address:
[18:01:50] ffff88800223b100: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 [18:01:50] ffff88800223b180: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 [18:01:50] >ffff88800223b200: 00 00 00 00 00 00 00 fc fc fc fc fc fc fc fc fc
[18:01:50] ^
[18:01:50] ffff88800223b280: fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc [18:01:50] ffff88800223b300: fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc [18:01:50] ================================================================== [18:01:50] Disabling lock debugging due to kernel taint

Fixes 27cb27b6d5ea ("io_uring: add support for kernel registered bvecs")
---
 io_uring/rsrc.c | 11 ++++++-----
 1 file changed, 6 insertions(+), 5 deletions(-)
io_uring/rsrc.c +6 -5
index a63474b331bf..7602b71543e0 100644
@@ -946,6 +946,7 @@ int io_buffer_register_bvec(struct io_uring_cmd *cmd, struct request *rq,
 	struct io_mapped_ubuf *imu;
 	struct io_rsrc_node *node;
 	struct bio_vec bv;
+	struct bio *bio;
 	unsigned int nr_bvecs = 0;
 	int ret = 0;
 
@@ -967,11 +968,10 @@ int io_buffer_register_bvec(struct io_uring_cmd *cmd, struct request *rq,
 		goto unlock;
 	}
 
-	/*
-	 * blk_rq_nr_phys_segments() may overestimate the number of bvecs
-	 * but avoids needing to iterate over the bvecs
-	 */
-	imu = io_alloc_imu(ctx, blk_rq_nr_phys_segments(rq));
+	__rq_for_each_bio(bio, rq)
+		nr_bvecs += bio->bi_vcnt;
+
+	imu = io_alloc_imu(ctx, nr_bvecs);
 	if (!imu) {
 		kfree(node);
 		ret = -ENOMEM;
@@ -988,6 +988,7 @@ int io_buffer_register_bvec(struct io_uring_cmd *cmd, struct request *rq,
 	imu->is_kbuf = true;
 	imu->dir = 1 << rq_data_dir(rq);
 
+	nr_bvecs = 0;
 	rq_for_each_bvec(bv, rq, rq_iter)
 		imu->bvec[nr_bvecs++] = bv;
 	imu->nr_bvecs = nr_bvecs;
Signature · 1 line
2.52.0
#1 lore raw
Quoted 6 lines · veygax

From: Evan Lambert <[email protected]>

The function io_buffer_register_bvec() calculates the allocation size for the io_mapped_ubuf based on blk_rq_nr_phys_segments(rq). This function calculates the number of scatter-gather elements after megine

"merging"?

Quoted 5 lines

physically contiguous pages.

However, the subsequent loop uses rq_for_each_bvec() to populate the array, which iterates over every individual bio_vec in the request, regardless of physical contiguity.

Hmm, I would have thought that physically contiguous bio_vecs would have been merged by the block layer? But that's definitely beyond my expertise.

Quoted 8 lines

If a request has multiple bio_vec entries that are physically contiguous, blk_rq_nr_phys_segments() returns a value smaller than the total number of bio_vecs. This leads to a slab-out-of-bounds write.

The path is reachable from userspace via the ublk driver when a server issues a UBLK_IO_REGISTER_IO_BUF command. This requires the UBLK_F_SUPPORT_ZERO_COPY flag which is protected by CAP_NET_ADMIN.

"CAP_SYS_ADMIN"?

Quoted 13 lines

Fix this by calculating the total number of bio_vecs by iterating over the request's bios and summing their bi_vcnt.

KASAN report:

[18:01:50] BUG: KASAN: slab-out-of-bounds in io_buffer_register_bvec+0x813/0xb80 [18:01:50] Write of size 8 at addr ffff88800223b238 by task kunit_try_catch/27 [18:01:50]
[18:01:50] CPU: 0 UID: 0 PID: 27 Comm: kunit_try_catch Tainted: G N 6.19.0-rc1-g346af1a0c65a-dirty #44 PREEMPT(none)
[18:01:50] Tainted: [N]=TEST
[18:01:50] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 0.1 11/11/2019 [18:01:50] Call Trace:

Quoted 16 lines
[18:01:50]  <TASK>
[18:01:50]  dump_stack_lvl+0x4d/0x70
[18:01:50]  print_report+0x151/0x4c0
[18:01:50]  ? __pfx__raw_spin_lock_irqsave+0x10/0x10
[18:01:50]  ? io_buffer_register_bvec+0x813/0xb80
[18:01:50]  kasan_report+0xec/0x120
[18:01:50]  ? io_buffer_register_bvec+0x813/0xb80
[18:01:50]  io_buffer_register_bvec+0x813/0xb80
[18:01:50]  io_buffer_register_bvec_overflow_test+0x4e6/0x9b0
[18:01:50]  ? __pfx_io_buffer_register_bvec_overflow_test+0x10/0x10
[18:01:50]  ? __pfx_pick_next_task_fair+0x10/0x10
[18:01:50]  ? _raw_spin_lock+0x7e/0xd0
[18:01:50]  ? finish_task_switch.isra.0+0x19a/0x650
[18:01:50]  ? __pfx_read_tsc+0x10/0x10
[18:01:50]  ? ktime_get_ts64+0x79/0x240
[18:01:50]  kunit_try_run_case+0x19b/0x2c0

This doesn't look like an actual ublk zero-copy buffer registration. Where does the struct request come from?

Quoted 90 lines
[18:01:50]  ? __pfx_kunit_try_run_case+0x10/0x10
[18:01:50]  ? __pfx_kunit_generic_run_threadfn_adapter+0x10/0x10
[18:01:50]  kunit_generic_run_threadfn_adapter+0x80/0xf0
[18:01:50]  kthread+0x323/0x670
[18:01:50]  ? __pfx_kthread+0x10/0x10
[18:01:50]  ? __pfx__raw_spin_lock_irq+0x10/0x10
[18:01:50]  ? __pfx_kthread+0x10/0x10
[18:01:50]  ret_from_fork+0x329/0x420
[18:01:50]  ? __pfx_ret_from_fork+0x10/0x10
[18:01:50]  ? __switch_to+0xa0f/0xd40
[18:01:50]  ? __pfx_kthread+0x10/0x10
[18:01:50]  ret_from_fork_asm+0x1a/0x30
[18:01:50]  </TASK>
[18:01:50]
[18:01:50] Allocated by task 27:
[18:01:50]  kasan_save_stack+0x30/0x50
[18:01:50]  kasan_save_track+0x14/0x30
[18:01:50]  __kasan_kmalloc+0x7f/0x90
[18:01:50]  io_cache_alloc_new+0x35/0xc0
[18:01:50]  io_buffer_register_bvec+0x196/0xb80
[18:01:50]  io_buffer_register_bvec_overflow_test+0x4e6/0x9b0
[18:01:50]  kunit_try_run_case+0x19b/0x2c0
[18:01:50]  kunit_generic_run_threadfn_adapter+0x80/0xf0
[18:01:50]  kthread+0x323/0x670
[18:01:50]  ret_from_fork+0x329/0x420
[18:01:50]  ret_from_fork_asm+0x1a/0x30

[18:01:50]
[18:01:50] The buggy address belongs to the object at ffff88800223b000 [18:01:50] which belongs to the cache kmalloc-1k of size 1024 [18:01:50] The buggy address is located 0 bytes to the right of [18:01:50] allocated 568-byte region [ffff88800223b000, ffff88800223b238) [18:01:50]
[18:01:50] The buggy address belongs to the physical page:
[18:01:50] page: refcount:0 mapcount:0 mapping:0000000000000000 index:0x0 pfn:0x2238 [18:01:50] head: order:2 mapcount:0 entire_mapcount:0 nr_pages_mapped:0 pincount:0 [18:01:50] flags: 0x4000000000000040(head|zone=1)
[18:01:50] page_type: f5(slab)
[18:01:50] raw: 4000000000000040 ffff888001041dc0 dead000000000122 0000000000000000 [18:01:50] raw: 0000000000000000 0000000080080008 00000000f5000000 0000000000000000 [18:01:50] head: 4000000000000040 ffff888001041dc0 dead000000000122 0000000000000000 [18:01:50] head: 0000000000000000 0000000080080008 00000000f5000000 0000000000000000 [18:01:50] head: 4000000000000002 ffffea0000088e01 00000000ffffffff 00000000ffffffff [18:01:50] head: 0000000000000000 0000000000000000 00000000ffffffff 0000000000000000 [18:01:50] page dumped because: kasan: bad access detected
[18:01:50]
[18:01:50] Memory state around the buggy address:
[18:01:50] ffff88800223b100: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 [18:01:50] ffff88800223b180: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 [18:01:50] >ffff88800223b200: 00 00 00 00 00 00 00 fc fc fc fc fc fc fc fc fc
[18:01:50] ^
[18:01:50] ffff88800223b280: fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc [18:01:50] ffff88800223b300: fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc [18:01:50] ================================================================== [18:01:50] Disabling lock debugging due to kernel taint

Fixes 27cb27b6d5ea ("io_uring: add support for kernel registered bvecs")
---
 io_uring/rsrc.c | 11 ++++++-----
 1 file changed, 6 insertions(+), 5 deletions(-)
io_uring/rsrc.c +5 -5
index a63474b331bf..7602b71543e0 100644
@@ -946,6 +946,7 @@ int io_buffer_register_bvec(struct io_uring_cmd *cmd, struct request *rq,
        struct io_mapped_ubuf *imu;
        struct io_rsrc_node *node;
        struct bio_vec bv;
+       struct bio *bio;
        unsigned int nr_bvecs = 0;
        int ret = 0;
 
@@ -967,11 +968,10 @@ int io_buffer_register_bvec(struct io_uring_cmd *cmd, struct request *rq,
                goto unlock;
        }
 
-       /*
-        * blk_rq_nr_phys_segments() may overestimate the number of bvecs
-        * but avoids needing to iterate over the bvecs
-        */
-       imu = io_alloc_imu(ctx, blk_rq_nr_phys_segments(rq));
+       __rq_for_each_bio(bio, rq)
+               nr_bvecs += bio->bi_vcnt;
+
+       imu = io_alloc_imu(ctx, nr_bvecs);
        if (!imu) {
                kfree(node);
                ret = -ENOMEM;

io_uring/rsrc.c int io_buffer_register_bvec(struct io_uring_cmd *cmd, struct request *rq, · line 993

Quoted 6 lines
io_uring/rsrc.c +1 -0
@@ -988,6 +988,7 @@ int io_buffer_register_bvec(struct io_uring_cmd *cmd, struct request *rq,
        imu->is_kbuf = true;
        imu->dir = 1 << rq_data_dir(rq);
 
+       nr_bvecs = 0;
        rq_for_each_bvec(bv, rq, rq_iter)
                imu->bvec[nr_bvecs++] = bv;

Could alternatively check for mergability with the previous bvec here. That would avoid needing to allocate extra memory for physically contiguous bvecs.

Best,
Caleb

Quoted 5 lines
io_uring/rsrc.c +0 -1
        imu->nr_bvecs = nr_bvecs;
--

2.52.0

#2 lore raw
In reply to veygax
Quoted 19 lines · veygax

From: Evan Lambert <[email protected]>

The function io_buffer_register_bvec() calculates the allocation size for the io_mapped_ubuf based on blk_rq_nr_phys_segments(rq). This function calculates the number of scatter-gather elements after megine physically contiguous pages.

However, the subsequent loop uses rq_for_each_bvec() to populate the array, which iterates over every individual bio_vec in the request, regardless of physical contiguity.

If a request has multiple bio_vec entries that are physically contiguous, blk_rq_nr_phys_segments() returns a value smaller than the total number of bio_vecs. This leads to a slab-out-of-bounds write.

The path is reachable from userspace via the ublk driver when a server issues a UBLK_IO_REGISTER_IO_BUF command. This requires the UBLK_F_SUPPORT_ZERO_COPY flag which is protected by CAP_NET_ADMIN.

Quoted 7 lines

Fix this by calculating the total number of bio_vecs by iterating over the request's bios and summing their bi_vcnt.

KASAN report:

[18:01:50] BUG: KASAN: slab-out-of-bounds in io_buffer_register_bvec+0x813/0xb80 [18:01:50] Write of size 8 at addr ffff88800223b238 by task kunit_try_catch/27

Can you share the test case so that we can understand why page isn't merged to last bvec? Maybe there is chance to improve block layer(bio add page related code)

Quoted 95 lines

[18:01:50]
[18:01:50] CPU: 0 UID: 0 PID: 27 Comm: kunit_try_catch Tainted: G N 6.19.0-rc1-g346af1a0c65a-dirty #44 PREEMPT(none)
[18:01:50] Tainted: [N]=TEST
[18:01:50] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 0.1 11/11/2019 [18:01:50] Call Trace:

[18:01:50]  <TASK>
[18:01:50]  dump_stack_lvl+0x4d/0x70
[18:01:50]  print_report+0x151/0x4c0
[18:01:50]  ? __pfx__raw_spin_lock_irqsave+0x10/0x10
[18:01:50]  ? io_buffer_register_bvec+0x813/0xb80
[18:01:50]  kasan_report+0xec/0x120
[18:01:50]  ? io_buffer_register_bvec+0x813/0xb80
[18:01:50]  io_buffer_register_bvec+0x813/0xb80
[18:01:50]  io_buffer_register_bvec_overflow_test+0x4e6/0x9b0
[18:01:50]  ? __pfx_io_buffer_register_bvec_overflow_test+0x10/0x10
[18:01:50]  ? __pfx_pick_next_task_fair+0x10/0x10
[18:01:50]  ? _raw_spin_lock+0x7e/0xd0
[18:01:50]  ? finish_task_switch.isra.0+0x19a/0x650
[18:01:50]  ? __pfx_read_tsc+0x10/0x10
[18:01:50]  ? ktime_get_ts64+0x79/0x240
[18:01:50]  kunit_try_run_case+0x19b/0x2c0
[18:01:50]  ? __pfx_kunit_try_run_case+0x10/0x10
[18:01:50]  ? __pfx_kunit_generic_run_threadfn_adapter+0x10/0x10
[18:01:50]  kunit_generic_run_threadfn_adapter+0x80/0xf0
[18:01:50]  kthread+0x323/0x670
[18:01:50]  ? __pfx_kthread+0x10/0x10
[18:01:50]  ? __pfx__raw_spin_lock_irq+0x10/0x10
[18:01:50]  ? __pfx_kthread+0x10/0x10
[18:01:50]  ret_from_fork+0x329/0x420
[18:01:50]  ? __pfx_ret_from_fork+0x10/0x10
[18:01:50]  ? __switch_to+0xa0f/0xd40
[18:01:50]  ? __pfx_kthread+0x10/0x10
[18:01:50]  ret_from_fork_asm+0x1a/0x30
[18:01:50]  </TASK>
[18:01:50]
[18:01:50] Allocated by task 27:
[18:01:50]  kasan_save_stack+0x30/0x50
[18:01:50]  kasan_save_track+0x14/0x30
[18:01:50]  __kasan_kmalloc+0x7f/0x90
[18:01:50]  io_cache_alloc_new+0x35/0xc0
[18:01:50]  io_buffer_register_bvec+0x196/0xb80
[18:01:50]  io_buffer_register_bvec_overflow_test+0x4e6/0x9b0
[18:01:50]  kunit_try_run_case+0x19b/0x2c0
[18:01:50]  kunit_generic_run_threadfn_adapter+0x80/0xf0
[18:01:50]  kthread+0x323/0x670
[18:01:50]  ret_from_fork+0x329/0x420
[18:01:50]  ret_from_fork_asm+0x1a/0x30

[18:01:50]
[18:01:50] The buggy address belongs to the object at ffff88800223b000 [18:01:50] which belongs to the cache kmalloc-1k of size 1024 [18:01:50] The buggy address is located 0 bytes to the right of [18:01:50] allocated 568-byte region [ffff88800223b000, ffff88800223b238) [18:01:50]
[18:01:50] The buggy address belongs to the physical page:
[18:01:50] page: refcount:0 mapcount:0 mapping:0000000000000000 index:0x0 pfn:0x2238 [18:01:50] head: order:2 mapcount:0 entire_mapcount:0 nr_pages_mapped:0 pincount:0 [18:01:50] flags: 0x4000000000000040(head|zone=1)
[18:01:50] page_type: f5(slab)
[18:01:50] raw: 4000000000000040 ffff888001041dc0 dead000000000122 0000000000000000 [18:01:50] raw: 0000000000000000 0000000080080008 00000000f5000000 0000000000000000 [18:01:50] head: 4000000000000040 ffff888001041dc0 dead000000000122 0000000000000000 [18:01:50] head: 0000000000000000 0000000080080008 00000000f5000000 0000000000000000 [18:01:50] head: 4000000000000002 ffffea0000088e01 00000000ffffffff 00000000ffffffff [18:01:50] head: 0000000000000000 0000000000000000 00000000ffffffff 0000000000000000 [18:01:50] page dumped because: kasan: bad access detected
[18:01:50]
[18:01:50] Memory state around the buggy address:
[18:01:50] ffff88800223b100: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 [18:01:50] ffff88800223b180: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 [18:01:50] >ffff88800223b200: 00 00 00 00 00 00 00 fc fc fc fc fc fc fc fc fc
[18:01:50] ^
[18:01:50] ffff88800223b280: fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc [18:01:50] ffff88800223b300: fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc [18:01:50] ================================================================== [18:01:50] Disabling lock debugging due to kernel taint

Fixes 27cb27b6d5ea ("io_uring: add support for kernel registered bvecs")
---
 io_uring/rsrc.c | 11 ++++++-----
 1 file changed, 6 insertions(+), 5 deletions(-)
io_uring/rsrc.c +1 -0
index a63474b331bf..7602b71543e0 100644
@@ -946,6 +946,7 @@ int io_buffer_register_bvec(struct io_uring_cmd *cmd, struct request *rq,
 	struct io_mapped_ubuf *imu;
 	struct io_rsrc_node *node;
 	struct bio_vec bv;
+	struct bio *bio;
 	unsigned int nr_bvecs = 0;
 	int ret = 0;
 

io_uring/rsrc.c int io_buffer_register_bvec(struct io_uring_cmd *cmd, struct request *rq, · line 972

Quoted 10 lines
io_uring/rsrc.c +2 -5
@@ -967,11 +968,10 @@ int io_buffer_register_bvec(struct io_uring_cmd *cmd, struct request *rq,
 		goto unlock;
 	}
 
-	/*
-	 * blk_rq_nr_phys_segments() may overestimate the number of bvecs
-	 * but avoids needing to iterate over the bvecs
-	 */
-	imu = io_alloc_imu(ctx, blk_rq_nr_phys_segments(rq));
+	__rq_for_each_bio(bio, rq)
+		nr_bvecs += bio->bi_vcnt;

This way is wrong, bio->bi_vcnt can't be trusted for this purpose, you may have to use rq_for_each_bvec() for calculating the real nr_bvecs.

Thanks,
Ming

author
#3 lore raw
Quoted 3 lines · Ming Lei

Can you share the test case so that we can understand why page isn't merged to last bvec? Maybe there is chance to improve block layer(bio add page related code)

Sure, this is how i triggered it:

#include <kunit/test.h>
#include <linux/io_uring.h>
#include <linux/io_uring_types.h>
#include <linux/io_uring/cmd.h>
#include <linux/blk-mq.h>
#include <linux/bio.h>
#include <linux/bvec.h>
#include <linux/mm.h>
#include <linux/slab.h>

#include "io_uring.h"
#include "rsrc.h"

static void dummy_release(void *priv)
{
}

static void io_buffer_register_bvec_overflow_test(struct kunit *test)

{
	struct io_ring_ctx *ctx;
	struct io_uring_cmd *cmd;
	struct io_kiocb *req;
	struct request *rq;
	struct bio *bio;
	struct page *page;
	int i, ret;

	/*
	 * IO_CACHED_BVECS_SEGS is 32.
	 * We want more than 32 bvecs to trigger overflow if allocation uses 32.
	 */
	int num_bvecs = 40;
	
	ctx = kunit_kzalloc(test, sizeof(*ctx), GFP_KERNEL);
	KUNIT_ASSERT_NOT_NULL(test, ctx);
	
	/* Initialize caches so io_alloc_imu works and knows the size */
	if (io_rsrc_cache_init(ctx))
		kunit_skip(test, "failed to init rsrc cache");

	/* Initialize buf_table so index check passes */
	ret = io_rsrc_data_alloc(&ctx->buf_table, 1);
	KUNIT_ASSERT_EQ(test, ret, 0);

	req = kunit_kzalloc(test, sizeof(*req), GFP_KERNEL);
	KUNIT_ASSERT_NOT_NULL(test, req);
	req->ctx = ctx;
	cmd = io_kiocb_to_cmd(req, struct io_uring_cmd);

	rq = kunit_kzalloc(test, sizeof(*rq), GFP_KERNEL);
	KUNIT_ASSERT_NOT_NULL(test, rq);
	
	/* Allocate bio with enough slots */
	bio = bio_kmalloc(num_bvecs, GFP_KERNEL);
	KUNIT_ASSERT_NOT_NULL(test, bio);
	bio_init(bio, NULL, bio_inline_vecs(bio), num_bvecs, REQ_OP_WRITE);
	rq->bio = bio;
	
	page = alloc_pages(GFP_KERNEL | __GFP_COMP | __GFP_ZERO, 6);
	KUNIT_ASSERT_NOT_NULL(test, page);
	
	/*
	 * Add pages to bio manually.
	 * We use physically contiguous pages to trick blk_rq_nr_phys_segments
	 * into returning 1 segment.
	 * We use multiple bvec entries to trick the loop in
io_buffer_register_bvec
	 * into writing out of bounds.
	 */
	for (i = 0; i < num_bvecs; i++) {
		struct bio_vec *bv = &bio->bi_io_vec[i];
		bv->bv_page = page + i;
		bv->bv_len = PAGE_SIZE;
		bv->bv_offset = 0;
		bio->bi_vcnt++;
		bio->bi_iter.bi_size += PAGE_SIZE;
	}
	
	/* Trigger */
	ret = io_buffer_register_bvec(cmd, rq, dummy_release, 0, 0);
	
	/* this should not be reachable */
	__free_pages(page, 6);
	kfree(bio);
	io_rsrc_data_free(ctx, &ctx->buf_table);
	io_rsrc_cache_free(ctx);
}
static struct kunit_case io_uring_rsrc_test_cases[] = {
	KUNIT_CASE(io_buffer_register_bvec_overflow_test),
	{}
};
static struct kunit_suite io_uring_rsrc_test_suite = {
	.name = "io_uring_rsrc_test",
	.test_cases = io_uring_rsrc_test_cases,
};

kunit_test_suite(io_uring_rsrc_test_suite);
MODULE_LICENSE("GPL");

Signature · 1 line
- Evan Lambert / veygax
#4 lore raw
Quoted 7 lines · veygax
	/*
	 * Add pages to bio manually.
	 * We use physically contiguous pages to trick blk_rq_nr_phys_segments
	 * into returning 1 segment.
	 * We use multiple bvec entries to trick the loop in io_buffer_register_bvec
	 * into writing out of bounds.
	 */
Quoted 8 lines
	for (i = 0; i < num_bvecs; i++) {
		struct bio_vec *bv = &bio->bi_io_vec[i];
		bv->bv_page = page + i;
		bv->bv_len = PAGE_SIZE;
		bv->bv_offset = 0;
		bio->bi_vcnt++;
		bio->bi_iter.bi_size += PAGE_SIZE;
	}

I believe you're supposed to use the bio_add_page() API rather than open code the bvec setup.

author
#5 lore raw
Quoted 2 lines · Keith Busch

I believe you're supposed to use the bio_add_page() API rather than open code the bvec setup.

True, but I wanted fine control to prove my theory

Signature · 1 line
- Evan Lambert / veygax
#6 lore raw
Quoted 6 lines · veygax
Quoted 2 lines · Keith Busch

I believe you're supposed to use the bio_add_page() API rather than open code the bvec setup.

True, but I wanted fine control to prove my theory

But almost there aren't such cases for in-tree uses, :-)

Thanks,

#7 lore raw
In reply to veygax
Quoted 5 lines · veygax
Quoted 2 lines · Keith Busch

I believe you're supposed to use the bio_add_page() API rather than open code the bvec setup.

True, but I wanted fine control to prove my theory

But doesn't that just prove misusing the interface breaks things? Is there currently a legit way to get this error without the misuse? Or is there existing mis-use in the kernel that should be fixed instead?

maintainer
#8 lore raw
Quoted 10 lines · Keith Busch
Quoted 5 lines · veygax
Quoted 2 lines · Keith Busch

I believe you're supposed to use the bio_add_page() API rather than open code the bvec setup.

True, but I wanted fine control to prove my theory

But doesn't that just prove misusing the interface breaks things? Is there currently a legit way to get this error without the misuse? Or is there existing mis-use in the kernel that should be fixed instead?

This is the big question, and also why I originally rejected the posted poc as it's not a valid use case.

veygax, please make a real reproducer or detail how this can actually happen with the exposed APIs.

Signature · 1 line
Jens Axboe
#9 lore raw
In reply to Keith Busch
Quoted 17 lines · Keith Busch
Quoted 7 lines · veygax
	/*
	 * Add pages to bio manually.
	 * We use physically contiguous pages to trick blk_rq_nr_phys_segments
	 * into returning 1 segment.
	 * We use multiple bvec entries to trick the loop in io_buffer_register_bvec
	 * into writing out of bounds.
	 */
Quoted 8 lines
	for (i = 0; i < num_bvecs; i++) {
		struct bio_vec *bv = &bio->bi_io_vec[i];
		bv->bv_page = page + i;
		bv->bv_len = PAGE_SIZE;
		bv->bv_offset = 0;
		bio->bi_vcnt++;
		bio->bi_iter.bi_size += PAGE_SIZE;
	}
Quoted 2 lines

I believe you're supposed to use the bio_add_page() API rather than open code the bvec setup.

The above is simply an open coded version of doing repeated __bio_add_page calls. Which would be rather suboptimal, but perfectly valid.

#10 lore raw
Quoted 18 lines · Christoph Hellwig
Quoted 17 lines · Keith Busch
Quoted 15 lines · veygax
	/*
	 * Add pages to bio manually.
	 * We use physically contiguous pages to trick blk_rq_nr_phys_segments
	 * into returning 1 segment.
	 * We use multiple bvec entries to trick the loop in io_buffer_register_bvec
	 * into writing out of bounds.
	 */
	for (i = 0; i < num_bvecs; i++) {
		struct bio_vec *bv = &bio->bi_io_vec[i];
		bv->bv_page = page + i;
		bv->bv_len = PAGE_SIZE;
		bv->bv_offset = 0;
		bio->bi_vcnt++;
		bio->bi_iter.bi_size += PAGE_SIZE;
	}
Quoted 6 lines
Quoted 2 lines

I believe you're supposed to use the bio_add_page() API rather than open code the bvec setup.

The above is simply an open coded version of doing repeated __bio_add_page calls. Which would be rather suboptimal, but perfectly valid.

Yeah, there's nothing stopping someone from using it that way, but a quick survey of __bio_add_page() users appear to be special cases that allocate a single vector bio, so its existing use is a short-cut that bio_add_page() will inevitiably reach anyway. Did you intend for it to be called directly for multiple vector uses too? It is suboptimal as you said, so it still feels like a misuse if someone did that.

#11 lore raw
Quoted 10 lines · Keith Busch
Quoted 3 lines

The above is simply an open coded version of doing repeated __bio_add_page calls. Which would be rather suboptimal, but perfectly valid.

Yeah, there's nothing stopping someone from using it that way, but a quick survey of __bio_add_page() users appear to be special cases that allocate a single vector bio, so its existing use is a short-cut that bio_add_page() will inevitiably reach anyway. Did you intend for it to be called directly for multiple vector uses too? It is suboptimal as you said, so it still feels like a misuse if someone did that.

We can't even force users to use __bio_add_page. Take a look at drivers/md/bcache/util.c:bch_bio_map() for a real-life example for something that could create this bvec pattern.

author
#12 lore raw
In reply to Christoph Hellwig
Quoted 3 lines · Christoph Hellwig

The above is simply an open coded version of doing repeated __bio_add_page calls. Which would be rather suboptimal, but perfectly valid.

I can confirm iterating over num_bvecs in __bio_add_page produces the same KASAN output.

	for (i = 0; i < num_bvecs; i++)
		__bio_add_page(bio, page + i, PAGE_SIZE, 0);

Also (to all CC's), please disregard this version of the patch. V2 has been posted with typos fixed and a more performance efficient approach.
https://lore.kernel.org/io-uring/[email protected]/

Signature · 1 line
- Evan Lambert / veygax