Re: [LTP] [PATCH 5/5] syscalls/fanotify10: Add test cases for evictable ignore mark

[Amir Goldstein <amir73il@gmail.com>]

On Thu, Jul 7, 2022 at 3:49 PM Jan Stancek <jstancek@redhat.com> wrote:
>
> On Thu, Jun 30, 2022 at 10:21 AM Amir Goldstein <amir73il@gmail.com> wrote:
> >
> > On Thu, Jun 30, 2022 at 9:27 AM Jan Stancek <jstancek@redhat.com> wrote:
> > >
> > > On Mon, Jun 13, 2022 at 4:39 PM Amir Goldstein <amir73il@gmail.com> wrote:
> > > >
> > > > Test multiple groups with evictable mark with ignore mask
> > > >
> > > > Signed-off-by: Amir Goldstein <amir73il@gmail.com>
> > > > ---
> > > >  .../kernel/syscalls/fanotify/fanotify10.c     | 78 +++++++++++++++++++
> > > >  1 file changed, 78 insertions(+)
> > > >
> > > > diff --git a/testcases/kernel/syscalls/fanotify/fanotify10.c b/testcases/kernel/syscalls/fanotify/fanotify10.c
> > > > index b9a50672d..52277d0b7 100644
> > > > --- a/testcases/kernel/syscalls/fanotify/fanotify10.c
> > > > +++ b/testcases/kernel/syscalls/fanotify/fanotify10.c
> > > > @@ -71,6 +71,7 @@ static char event_buf[EVENT_BUF_LEN];
> > > >  static int exec_events_unsupported;
> > > >  static int fan_report_dfid_unsupported;
> > > >  static int filesystem_mark_unsupported;
> > > > +static int evictable_mark_unsupported;
> > > >
> > > >  #define MOUNT_PATH "fs_mnt"
> > > >  #define MNT2_PATH "mntpoint"
> > > > @@ -90,6 +91,10 @@ static int filesystem_mark_unsupported;
> > > >  #define FILE_EXEC_PATH2 MNT2_PATH"/"TEST_APP
> > > >  #define FILE2_EXEC_PATH2 MNT2_PATH"/"TEST_APP2
> > > >
> > > > +#define DROP_CACHES_FILE "/proc/sys/vm/drop_caches"
> > > > +#define CACHE_PRESSURE_FILE "/proc/sys/vm/vfs_cache_pressure"
> > > > +
> > > > +static int old_cache_pressure;
> > > >  static pid_t child_pid;
> > > >  static int bind_mount_created;
> > > >  static unsigned int num_classes = NUM_CLASSES;
> > > > @@ -98,12 +103,14 @@ enum {
> > > >         FANOTIFY_INODE,
> > > >         FANOTIFY_MOUNT,
> > > >         FANOTIFY_FILESYSTEM,
> > > > +       FANOTIFY_EVICTABLE,
> > > >  };
> > > >
> > > >  static struct fanotify_mark_type fanotify_mark_types[] = {
> > > >         INIT_FANOTIFY_MARK_TYPE(INODE),
> > > >         INIT_FANOTIFY_MARK_TYPE(MOUNT),
> > > >         INIT_FANOTIFY_MARK_TYPE(FILESYSTEM),
> > > > +       INIT_FANOTIFY_MARK_TYPE(EVICTABLE),
> > > >  };
> > > >
> > > >  static struct tcase {
> > > > @@ -289,14 +296,59 @@ static struct tcase {
> > > >                 0,
> > > >                 FILE_PATH, FAN_OPEN, FAN_OPEN
> > > >         },
> > > > +       /* Evictable ignore mark test cases */
> > > > +       {
> > > > +               "don't ignore mount events created on file with evicted ignore mark",
> > > > +               MOUNT_PATH, FANOTIFY_MOUNT,
> > > > +               FILE_PATH, FANOTIFY_EVICTABLE,
> > > > +               0,
> > > > +               FILE_PATH, FAN_OPEN, FAN_OPEN
> > > > +       },
> > > > +       {
> > > > +               "don't ignore fs events created on a file with evicted ignore mark",
> > > > +               MOUNT_PATH, FANOTIFY_FILESYSTEM,
> > > > +               FILE_PATH, FANOTIFY_EVICTABLE,
> > > > +               0,
> > > > +               FILE_PATH, FAN_OPEN, FAN_OPEN
> > > > +       },
> > > > +       {
> > > > +               "don't ignore mount events created inside a parent with evicted ignore mark",
> > > > +               MOUNT_PATH, FANOTIFY_MOUNT,
> > > > +               DIR_PATH, FANOTIFY_EVICTABLE,
> > > > +               FAN_EVENT_ON_CHILD,
> > > > +               FILE_PATH, FAN_OPEN, FAN_OPEN
> > > > +       },
> > > > +       {
> > > > +               "don't ignore fs events created inside a parent with evicted ignore mark",
> > > > +               MOUNT_PATH, FANOTIFY_FILESYSTEM,
> > > > +               DIR_PATH, FANOTIFY_EVICTABLE,
> > > > +               FAN_EVENT_ON_CHILD,
> > > > +               FILE_PATH, FAN_OPEN, FAN_OPEN
> > > > +       },
> > >
> > > Hi,
> > >
> > > we are seeing some sporadic failures from this last testcase, with
> > > recent upstream kernels (v5.19-rc4-14-g941e3e791269).
> > > Has anyone also ran into it and knows if it's bug on test side or kernel?

Hi Jan,

I am traveling so I cannot work on solving the problem for the next week,
but I can explain the problem and offer suggestions for short term solutions
and maybe a longer term solution.

> > >
> > > fanotify10.c:496: TINFO: Test #27: don't ignore fs events created
> > > inside a parent with evicted ignore mark
> > > fanotify10.c:338: TPASS: No fanotify inode marks as expected
> > > fanotify10.c:338: TPASS: No fanotify inode marks as expected
> > > fanotify10.c:338: TPASS: No fanotify inode marks as expected
> > > fanotify10.c:340: TFAIL: Unexpected inode mark (mflags=240,
> > > mask=8000020 ignored_mask=20)
> >
> > It is a test bug.
> > The problem is that we want to evict an inode, but there is no
> > reliable mechanism to do that.
> >
> > This is the reason for this workaround in fanotify23:
> >
> >         /* Shrinkers on other fs do not work reliably enough to
> > guarantee mark eviction on drop_caches */
> >         .dev_fs_type = "ext2",
> >
> > I did not encounter the problem with fanotify10 myself, but it should
> > be the same.
> > fanotify10 is not filesystem dependent, so if you can apply the same workaround
> > from fanotify23 and it works on your systems we can do that.
>
> Test is using default fs type, which should already be ext2.

I see. You can try xfs on your system to see if it behaves better,
because xfs has a specialized inode shriker.

> Here's a more complete log from failed test:
> https://s3.us-east-1.amazonaws.com/arr-cki-prod-datawarehouse-public/datawarehouse-public/2022/06/30/redhat:576928171/build_ppc64le_redhat:576928171_ppc64le/tests/1/results_0001/job.01/recipes/12221009/tasks/8/logs/syscalls.fail.log
>

Are the failures only in the first test iteration?

As a long shot, I would try to remove mount_cycle() from setup().
I see that fanotify23 does not have it and it is not really needed, so
it may help
fanotify10 pass the first iteration.

The problem is that the drop_caches knob is not a reliable way to evict inodes,
so when a test like fanotify10 needs to evict an inode, random factors are mixed
into the test run.

As a quick band aid, I suggest to replace the TFAIL, "Unexpected inode mark"
in show_fanotify_marks() with TCONF, because the test failed to setup an
"evictable ignored mark that gets evicted", propagate a failure return
value from
create_fanotify_groups() => show_fanotify_marks() and skip the test case
instead of failing it.

For long term, it would be nice if LTP could provide a
drop_inode_cache() library
function that tries harder to make inode eviction work using tricks
like fanotify10
and fanotify23 increases vfs_cache_pressure.
To pressure memory shrinkers to evict more inodes, need to check the amount
of total RAM in the system and dirty pages and add enough dirty memory pages
that cannot be evicted to drive the inode shrinker to work harder.
Utilizing memory cgroups with some of the new per memcg cache eviction knobs
could also be an option, but I did not look into it and not sure on
which kernels
this is available.

Bottom line is that there is a lot of black magic involved with memory shrinkers
and as your report shows, hacks and trick may impact different systems in
different ways, so it is better it the infrastructure is developed and tested in
LTP lib and not in individual tests.

Hope this helps.

Thanks,
Amir.

-- 
Mailing list info: https://lists.linux.it/listinfo/ltp