On Thu, Feb 1, 2024 at 1:08 AM Yosry Ahmed <yosryahmed@xxxxxxxxxx> wrote:
>
> Hey Nhat,
>
> I have a few more comments, sorry for not catching everything the first
> time around.
No worries :)
>
> Adding Roman to CC.
>
> On Wed, Jan 31, 2024 at 07:27:18PM -0800, Nhat Pham wrote:
> > Add a selftest to cover the zswapin code path, allocating more memory
> > than the cgroup limit to trigger swapout/zswapout, then reading the
> > pages back in memory several times. This is inspired by a recently
> > encountered kernel crash on the zswapin path in our internal kernel,
> > which went undetected because of a lack of test coverage for this path.
> >
> > Add a selftest to verify that when memory.zswap.max = 0, no pages can go
> > to the zswap pool for the cgroup.
> >
> > Suggested-by: Rik van Riel <riel@xxxxxxxxxxx>
> > Suggested-by: Yosry Ahmed <yosryahmed@xxxxxxxxxx>
> > Signed-off-by: Nhat Pham <nphamcs@xxxxxxxxx>
> > ---
> > tools/testing/selftests/cgroup/test_zswap.c | 97 +++++++++++++++++++++
> > 1 file changed, 97 insertions(+)
> >
> > diff --git a/tools/testing/selftests/cgroup/test_zswap.c b/tools/testing/selftests/cgroup/test_zswap.c
> > index 32ce975b21d1..14d1f18f1098 100644
> > --- a/tools/testing/selftests/cgroup/test_zswap.c
> > +++ b/tools/testing/selftests/cgroup/test_zswap.c
> > @@ -60,6 +60,27 @@ static long get_zswpout(const char *cgroup)
> > return cg_read_key_long(cgroup, "memory.stat", "zswpout ");
> > }
> >
> > +static int allocate_bytes_and_read(const char *cgroup, void *arg)
>
> I think allocate_and_read_bytes() is easier to read, but I don't feel
> strongly about it.
Ah you're right. I'll fix that in the next version.
>
> > +{
> > + size_t size = (size_t)arg;
> > + char *mem = (char *)malloc(size);
> > + int ret = 0;
> > +
> > + if (!mem)
> > + return -1;
> > + for (int i = 0; i < size; i += 4095)
> > + mem[i] = 'a';
>
> cgroup_util.h defines PAGE_SIZE, see alloc_anon() for example.
>
> On that note, alloc_anon() is awfully close to allocate_bytes() below,
> perhaps we should consolidate them. The only difference I see is that
> alloc_anon() does not check for the allocation failure, but a lot of
> functions in cgroup_helpers.c don't, so it seems intentional for
> simplification.
Hmm I didn't know about this function. I think it was Domenico who
added allocate_bytes() for the initial zswap tests, and I've just been
piggybacking on it ever since:
https://github.com/torvalds/linux/commit/d9cfaf405b8ffe2c716b1ce4c82e0a19d50951da
I can send a separate patch to clean this up later :) Doesn't seem that bad.
>
> > +
> > + /* go through the allocated memory to (z)swap in and out pages */
> > + for (int i = 0; i < size; i += 4095) {
> > + if (mem[i] != 'a')
> > + ret = -1;
> > + }
> > +
> > + free(mem);
> > + return ret;
> > +}
> > +
> > static int allocate_bytes(const char *cgroup, void *arg)
> > {
> > size_t size = (size_t)arg;
> > @@ -133,6 +154,80 @@ static int test_zswap_usage(const char *root)
> > return ret;
> > }
> >
> > +/*
> > + * Check that when memory.zswap.max = 0, no pages can go to the zswap pool for
> > + * the cgroup.
> > + */
> > +static int test_swapin_nozswap(const char *root)
> > +{
> > + int ret = KSFT_FAIL;
> > + char *test_group;
> > + long zswpout;
> > +
> > + /* Set up */
>
> I think this comment is unnecessary.
Fair!
>
> > + test_group = cg_name(root, "no_zswap_test");
> > +
> > + if (!test_group)
> > + goto out;
> > + if (cg_create(test_group))
> > + goto out;
> > + if (cg_write(test_group, "memory.max", "8M"))
> > + goto out;
> > + /* Disable zswap */
>
> I think this comment is unnecessary.
Fair!
>
> > + if (cg_write(test_group, "memory.zswap.max", "0"))
> > + goto out;
> > +
> > + /* Allocate and read more than memory.max to trigger swapin */
> > + if (cg_run(test_group, allocate_bytes_and_read, (void *)MB(32)))
> > + goto out;
> > +
> > + /* Verify that no zswap happened */
>
> If we want to be really meticulous, we can verify that we did swap out,
> but not to zswap. IOW, we can check memory.swap.current or something.
Hmm would memory.swap.current go back to 0 once the memory-in-swap is
freed? It doesn't seem like we have any counters at the cgroup level
for swapout/swapin events. Maybe such counters were not useful enough
to justify the extra overhead of maintaining them? :)
Anyway, I think checking zswpout should probably be enough here.
That's the spirit of the test anyway - make absolutely sure that no
zswap-out happened.
>
> > + zswpout = get_zswpout(test_group);
> > + if (zswpout < 0) {
> > + ksft_print_msg("Failed to get zswpout\n");
> > + goto out;
> > + } else if (zswpout > 0) {
>
> nit: This can be a separate if condition, I think it would be more
> inline with the style of separate consecutive if blocks we are
> following.
Yeah now that you point out the inconsistency, it starts to bug my
brain too :) I'll fix it in the next version.
>
> > + ksft_print_msg(
> > + "Pages should not go to zswap when memory.zswap.max = 0\n");
>
> We can probably avoid the line break with something more concise, for
> example:
> "zswapout > 0 when zswap is disabled"
> or "zswapout > 0 when memory.zswap.max = 0"
>
> > + goto out;
> > + }
> > + ret = KSFT_PASS;
> > +
> > +out:
> > + cg_destroy(test_group);
> > + free(test_group);
> > + return ret;
> > +}
> > +
> > +/* Simple test to verify the (z)swapin code paths */
> > +static int test_zswapin_no_limit(const char *root)
>
> I think test_zswapin() is enough to be distinct from
> test_swapin_nozswap(). The limit is not a factor here AFAICT.
Fair.
>
> > +{
> > + int ret = KSFT_FAIL;
> > + char *test_group;
> > +
> > + /* Set up */
>
> I think this comment is unnecessary.
Fair.
>
> > + test_group = cg_name(root, "zswapin_test");
> > + if (!test_group)
> > + goto out;
> > + if (cg_create(test_group))
> > + goto out;
> > + if (cg_write(test_group, "memory.max", "8M"))
> > + goto out;
> > + if (cg_write(test_group, "memory.zswap.max", "max"))
> > + goto out;
> > +
> > + /* Allocate and read more than memory.max to trigger (z)swap in */
> > + if (cg_run(test_group, allocate_bytes_and_read, (void *)MB(32)))
> > + goto out;
>
> We should probably check for a positive zswapin here, no?
Oh right. I'll just do a quick check here:
zswpin = cg_read_key_long(test_group, "memory.stat", "zswpin ");
if (zswpin < 0) {
ksft_print_msg("Failed to get zswpin\n");
goto out;
}
if (zswpin == 0) {
ksft_print_msg("zswpin should not be 0\n");
goto out;
}
>
> > +
> > + ret = KSFT_PASS;
> > +
> > +out:
> > + cg_destroy(test_group);
> > + free(test_group);
> > + return ret;
> > +}
> > +
> > /*
> > * When trying to store a memcg page in zswap, if the memcg hits its memory
> > * limit in zswap, writeback should affect only the zswapped pages of that
> > @@ -309,6 +404,8 @@ struct zswap_test {
> > const char *name;
> > } tests[] = {
> > T(test_zswap_usage),
> > + T(test_swapin_nozswap),
> > + T(test_zswapin_no_limit),
> > T(test_no_kmem_bypass),
> > T(test_no_invasive_cgroup_shrink),
> > };
> > --
> > 2.39.3
