| Message ID | 20241126085808.14616-1-liwang@redhat.com |
|---|---|
| State | Accepted |
| Headers | show |
| Series | process_madvise01: running the test in mem_cg | expand |
Hi Li, > The MADV_PAGEOUT behavior in the kernel is advisory and may skip > swapping if the system has sufficient free RAM, even when the > advice is explicitly requested. This causes sporadic false positives > in our CI, particularly on systems with large amounts of RAM: > process_madvise01.c:38: TINFO: Allocate memory: 1048576 bytes > process_madvise01.c:99: TINFO: Reclaim memory using MADV_PAGEOUT > process_madvise01.c:62: TFAIL: Expect: Most of the memory has been swapped out: 0kB out of 1024kB > To address this, the patch confines the test to a memory cgroup > with configured limits for memory.max and memory.swap.max, improving > control over memory and swap usage. This reduces the likelihood of > false positives caused by system-wide memory conditions. Out of curiosity, on how many RAM does it fail? And is it arch specific? ... > +++ b/testcases/kernel/syscalls/process_madvise/process_madvise01.c > @@ -23,7 +23,9 @@ > #include "lapi/syscalls.h" > #include "process_madvise.h" > -#define MEM_CHILD (1 * TST_MB) > +#define MEM_LIMIT (100 * TST_MB) > +#define MEMSW_LIMIT (200 * TST_MB) > +#define MEM_CHILD (1 * TST_MB) ... > @@ -123,7 +131,9 @@ static struct tst_test test = { > .min_kver = "5.10", > .needs_checkpoints = 1, > .needs_root = 1, > - .min_swap_avail = MEM_CHILD / TST_MB, > + .min_mem_avail = 2 * MEM_LIMIT / TST_MB, Requiring 200 MB for test looks LGTM. Reviewed-by: Petr Vorel <pvorel@suse.cz> Kind regards, Petr > + .min_swap_avail = 2 * MEM_CHILD / TST_MB, > + .needs_cgroup_ctrls = (const char *const []){ "memory", NULL },
On Thu, Jan 2, 2025 at 10:59 PM Petr Vorel <pvorel@suse.cz> wrote: > Hi Li, > > > The MADV_PAGEOUT behavior in the kernel is advisory and may skip > > swapping if the system has sufficient free RAM, even when the > > advice is explicitly requested. This causes sporadic false positives > > in our CI, particularly on systems with large amounts of RAM: > > > process_madvise01.c:38: TINFO: Allocate memory: 1048576 bytes > > process_madvise01.c:99: TINFO: Reclaim memory using MADV_PAGEOUT > > process_madvise01.c:62: TFAIL: Expect: Most of the memory has been > swapped out: 0kB out of 1024kB > > > To address this, the patch confines the test to a memory cgroup > > with configured limits for memory.max and memory.swap.max, improving > > control over memory and swap usage. This reduces the likelihood of > > false positives caused by system-wide memory conditions. > > Out of curiosity, on how many RAM does it fail? And is it arch specific? > > Ah, I completely forgot the context of this patch, my memory is bad now :/. > ... > > +++ b/testcases/kernel/syscalls/process_madvise/process_madvise01.c > > @@ -23,7 +23,9 @@ > > #include "lapi/syscalls.h" > > #include "process_madvise.h" > > > -#define MEM_CHILD (1 * TST_MB) > > +#define MEM_LIMIT (100 * TST_MB) > > +#define MEMSW_LIMIT (200 * TST_MB) > > +#define MEM_CHILD (1 * TST_MB) > ... > > @@ -123,7 +131,9 @@ static struct tst_test test = { > > .min_kver = "5.10", > > .needs_checkpoints = 1, > > .needs_root = 1, > > - .min_swap_avail = MEM_CHILD / TST_MB, > > + .min_mem_avail = 2 * MEM_LIMIT / TST_MB, > Requiring 200 MB for test looks LGTM. > > Reviewed-by: Petr Vorel <pvorel@suse.cz> > Thanks for the review, after reading the patch summary I guess it still makes sense to apply it, since MADV_PAGEOUT is indeed not a mandatory option and still may ignored by kernel in the test.
Hi Li, ... > Thanks for the review, after reading the patch summary I guess it > still makes sense to apply it, since MADV_PAGEOUT is indeed not > a mandatory option and still may ignored by kernel in the test. Sure, please go ahead and merge the patchset. Kind regards, Petr
On Fri, Jan 3, 2025 at 4:50 PM Petr Vorel <pvorel@suse.cz> wrote: > Hi Li, > > ... > > Thanks for the review, after reading the patch summary I guess it > > still makes sense to apply it, since MADV_PAGEOUT is indeed not > > a mandatory option and still may ignored by kernel in the test. > > Sure, please go ahead and merge the patchset. > Merged, thanks!
diff --git a/testcases/kernel/syscalls/process_madvise/process_madvise01.c b/testcases/kernel/syscalls/process_madvise/process_madvise01.c index 0fd3c1ef4..ca314c4da 100644 --- a/testcases/kernel/syscalls/process_madvise/process_madvise01.c +++ b/testcases/kernel/syscalls/process_madvise/process_madvise01.c @@ -23,7 +23,9 @@ #include "lapi/syscalls.h" #include "process_madvise.h" -#define MEM_CHILD (1 * TST_MB) +#define MEM_LIMIT (100 * TST_MB) +#define MEMSW_LIMIT (200 * TST_MB) +#define MEM_CHILD (1 * TST_MB) static void **data_ptr; @@ -67,6 +69,12 @@ static void child_alloc(void) static void setup(void) { + SAFE_CG_PRINTF(tst_cg, "memory.max", "%d", MEM_LIMIT); + if (SAFE_CG_HAS(tst_cg, "memory.swap.max")) + SAFE_CG_PRINTF(tst_cg, "memory.swap.max", "%d", MEMSW_LIMIT); + + SAFE_CG_PRINTF(tst_cg, "cgroup.procs", "%d", getpid()); + data_ptr = SAFE_MMAP(NULL, sizeof(void *), PROT_READ | PROT_WRITE, MAP_SHARED | MAP_ANONYMOUS, -1, 0); @@ -123,7 +131,9 @@ static struct tst_test test = { .min_kver = "5.10", .needs_checkpoints = 1, .needs_root = 1, - .min_swap_avail = MEM_CHILD / TST_MB, + .min_mem_avail = 2 * MEM_LIMIT / TST_MB, + .min_swap_avail = 2 * MEM_CHILD / TST_MB, + .needs_cgroup_ctrls = (const char *const []){ "memory", NULL }, .needs_kconfigs = (const char *[]) { "CONFIG_SWAP=y", NULL
The MADV_PAGEOUT behavior in the kernel is advisory and may skip swapping if the system has sufficient free RAM, even when the advice is explicitly requested. This causes sporadic false positives in our CI, particularly on systems with large amounts of RAM: process_madvise01.c:38: TINFO: Allocate memory: 1048576 bytes process_madvise01.c:99: TINFO: Reclaim memory using MADV_PAGEOUT process_madvise01.c:62: TFAIL: Expect: Most of the memory has been swapped out: 0kB out of 1024kB To address this, the patch confines the test to a memory cgroup with configured limits for memory.max and memory.swap.max, improving control over memory and swap usage. This reduces the likelihood of false positives caused by system-wide memory conditions. Signed-off-by: Li Wang <liwang@redhat.com> --- .../syscalls/process_madvise/process_madvise01.c | 14 ++++++++++++-- 1 file changed, 12 insertions(+), 2 deletions(-)