target: pscsi: avoid OOM in pscsi_map_sg()

Commit Message

Martin Wilck March 23, 2021, 4:22 p.m. UTC

From: Martin Wilck <[e-mail]>

pscsi_map_sg() uses the variable nr_pages as a hint for bio_kmalloc()
how many vector elements to allocate. If nr_pages is < BIO_MAX_PAGES,
it will be reset to 0 after successful allocation of the bio.

If bio_add_pc_page() fails later for whatever reason, pscsi_map_sg()
tries to allocate another bio, passing nr_vecs=0. This causes
bio_add_pc_page() to fail immediately in the next call. pci_map_sg()
continues to allocate zero-length bios until memory is exhausted and
the kernel crashes with OOM. This can be easily observed by exporting
a SATA DVD drive via pscsi. The target crashes as soon as the client
tries to access the DVD LUN. In the case I analyzed, bio_add_pc_page()
would fail because the DVD device's max_sectors_kb (128) was
exceeded.

Avoid this by simply not resetting nr_pages to 0 after allocating the
bio. This way, the client receives an IO error when it tries to send
requests exceeding the devices max_sectors_kb, and eventually gets
it right. The client must still limit max_sectors_kb e.g. by an udev
rule if (like in my case) the driver doesn't report valid block
limits, otherwise it encounters I/O errors.

Signed-off-by: Martin Wilck <[e-mail]>
---
 drivers/target/target_core_pscsi.c | 1 -
 1 file changed, 1 deletion(-)

Comments

Christoph Hellwig March 23, 2021, 4:29 p.m. UTC | #1

On Tue, Mar 23, 2021 at 05:22:03PM +0100, [e-mail] wrote:
> Avoid this by simply not resetting nr_pages to 0 after allocating the
> bio. This way, the client receives an IO error when it tries to send
> requests exceeding the devices max_sectors_kb, and eventually gets
> it right. The client must still limit max_sectors_kb e.g. by an udev
> rule if (like in my case) the driver doesn't report valid block
> limits, otherwise it encounters I/O errors.

FYI, I think the what you did here is correct, but not enough.
When pscsi_get_bio (that is bio_kmalloc) fails, this function needs
to unwind and return an error insted of blindly retrying the allocation,
else we can't recover from a memory shortage.

Chaitanya Kulkarni March 23, 2021, 6:07 p.m. UTC | #2

Martin,

On 3/23/21 09:23, [e-mail] wrote:
> From: Martin Wilck <[e-mail]>
>
> pscsi_map_sg() uses the variable nr_pages as a hint for bio_kmalloc()
> how many vector elements to allocate. If nr_pages is < BIO_MAX_PAGES,
> it will be reset to 0 after successful allocation of the bio.

I think BIO_MAX_PAGES is replaced by BIO_MAX_VECS with
commit a8affc03a9b3 ("block: rename BIO_MAX_PAGES to BIO_MAX_VECS").

Martin Wilck March 23, 2021, 8:01 p.m. UTC | #3

On Tue, 2021-03-23 at 16:29 +0000, Christoph Hellwig wrote:
> On Tue, Mar 23, 2021 at 05:22:03PM +0100, [e-mail] wrote:
> > Avoid this by simply not resetting nr_pages to 0 after allocating the
> > bio. This way, the client receives an IO error when it tries to send
> > requests exceeding the devices max_sectors_kb, and eventually gets
> > it right. The client must still limit max_sectors_kb e.g. by an udev
> > rule if (like in my case) the driver doesn't report valid block
> > limits, otherwise it encounters I/O errors.
> 
> FYI, I think the what you did here is correct, but not enough.
> When pscsi_get_bio (that is bio_kmalloc) fails, this function needs
> to unwind and return an error insted of blindly retrying the
> allocation,
> else we can't recover from a memory shortage.

I can try to do that, but in the tests I ran, I never observed
bio_kmalloc() failing. I saw it eat all memory, and various processes
being killed by the OOM killer, before the system eventually panicked
with OOM. It takes only fractions of a second until this happens:

[   57.356238] oom-kill:constraint=CONSTRAINT_NONE,nodemask=(null),cpuset=/,mems_allowed=0,global_oom,task_memcg=/,task=lightdm-gtk-gre,pid=4586,uid=481
[   57.369635] Out of memory: Killed process 4586 (lightdm-gtk-gre) total-vm:1035752kB, anon-rss:0kB, file-rss:2936kB, shmem-rss:0kB
...
[   57.698656] Node 0 Normal free:54828kB min:55008kB low:68760kB high:82512kB active_anon:4kB inactive_anon:0kB active_file:936kB inactive_file:1164kB unevictable:58036kB writepending:720kB present:13631488kB managed:13310252kB mlocked:58036kB kernel_stack:5808kB pagetables:8972kB bounce:0kB free_pcp:564kB local_pcp:0kB free_cma:0kB
...
[   57.818978] Unreclaimable slab info:
[   58.254160] kmalloc-192         15683546KB   15683546KB

(system has 16GiB memory).

Martin

Martin Wilck March 23, 2021, 9:20 p.m. UTC | #4

On Tue, 2021-03-23 at 18:07 +0000, Chaitanya Kulkarni wrote:
> Martin,
> 
> On 3/23/21 09:23, [e-mail] wrote:
> > From: Martin Wilck <[e-mail]>
> > 
> > pscsi_map_sg() uses the variable nr_pages as a hint for
> > bio_kmalloc()
> > how many vector elements to allocate. If nr_pages is <
> > BIO_MAX_PAGES,
> > it will be reset to 0 after successful allocation of the bio.
> 
> I think BIO_MAX_PAGES is replaced by BIO_MAX_VECS with
> commit a8affc03a9b3 ("block: rename BIO_MAX_PAGES to BIO_MAX_VECS").

Right. I made my patch against mkp/queue, which doesn't include this
commit yet. As this is just in the description, I don't think it
matters much, does it?

Martin

Chaitanya Kulkarni March 23, 2021, 9:25 p.m. UTC | #5

On 3/23/21 14:21, Martin Wilck wrote:
>> I think BIO_MAX_PAGES is replaced by BIO_MAX_VECS with
>> commit a8affc03a9b3 ("block: rename BIO_MAX_PAGES to BIO_MAX_VECS").
> Right. I made my patch against mkp/queue, which doesn't include this
> commit yet. As this is just in the description, I don't think it
> matters much, does it?
>
> Martin
>
>
>

I don't think it does, I'll let you decide that.

Patch

diff --git a/drivers/target/target_core_pscsi.c b/drivers/target/target_core_pscsi.c
index 7b1035e..977362d 100644
--- a/drivers/target/target_core_pscsi.c
+++ b/drivers/target/target_core_pscsi.c
@@ -881,7 +881,6 @@  pscsi_map_sg(struct se_cmd *cmd, struct scatterlist *sgl, u32 sgl_nents,
 			if (!bio) {
 new_bio:
 				nr_vecs = bio_max_segs(nr_pages);
-				nr_pages -= nr_vecs;
 				/*
 				 * Calls bio_kmalloc() and sets bio->bi_end_io()
 				 */