| Message ID | 20240606135835.600022-1-kernel@pankajraghav.com (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | selftests/mm: use volatile keyword to not optimize mmap read variable | expand |
On 6 Jun 2024, at 6:58, Pankaj Raghav (Samsung) wrote: > From: Pankaj Raghav <p.raghav@samsung.com> > > create_pagecache_thp_and_fd() in split_huge_page_test.c used the > variable dummy to perform mmap read. > > However, this test was skipped even on XFS which has large folio > support. The issue was compiler (gcc 13.2.0) was optimizing out the > dummy variable, therefore, not creating huge page in the page cache. > > Add volatile keyword to force compiler not to optimize out the loop > where we read from the mmaped addr. > > Signed-off-by: Pankaj Raghav <p.raghav@samsung.com> > --- > tools/testing/selftests/mm/split_huge_page_test.c | 2 +- > 1 file changed, 1 insertion(+), 1 deletion(-) > Acked-by: Zi Yan <ziy@nvidia.com> Best Regards, Yan, Zi
On Thu, Jun 06, 2024 at 01:58:35PM +0000, Pankaj Raghav (Samsung) wrote: > +++ b/tools/testing/selftests/mm/split_huge_page_test.c > @@ -300,7 +300,7 @@ int create_pagecache_thp_and_fd(const char *testfile, size_t fd_size, int *fd, > char **addr) > { > size_t i; > - int __attribute__((unused)) dummy = 0; > + volatile int __attribute__((unused)) dummy = 0; The mistake made by whoever wrote this test was making 'dummy' a stack variable. That lets the compiler figure out that it's unused. If you make it a top-level variable (not static) so the compiler can't tell whether it's referenced by a different compilation unit, it can't make that deduction. And you don't need the stupid attibute or volatile on it.
On 6 Jun 2024, at 7:35, Matthew Wilcox wrote: > On Thu, Jun 06, 2024 at 01:58:35PM +0000, Pankaj Raghav (Samsung) wrote: >> +++ b/tools/testing/selftests/mm/split_huge_page_test.c >> @@ -300,7 +300,7 @@ int create_pagecache_thp_and_fd(const char *testfile, size_t fd_size, int *fd, >> char **addr) >> { >> size_t i; >> - int __attribute__((unused)) dummy = 0; >> + volatile int __attribute__((unused)) dummy = 0; > > The mistake made by whoever wrote this test was making 'dummy' a stack That was me. :( > variable. That lets the compiler figure out that it's unused. If you > make it a top-level variable (not static) so the compiler can't tell > whether it's referenced by a different compilation unit, it can't make > that deduction. And you don't need the stupid attibute or volatile on it. It is better to just move “dummy” like you suggested instead of adding more keywords to fix it. - Best Regards, Yan, Zi
On Thu, Jun 06, 2024 at 03:35:31PM +0100, Matthew Wilcox wrote: > On Thu, Jun 06, 2024 at 01:58:35PM +0000, Pankaj Raghav (Samsung) wrote: > > +++ b/tools/testing/selftests/mm/split_huge_page_test.c > > @@ -300,7 +300,7 @@ int create_pagecache_thp_and_fd(const char *testfile, size_t fd_size, int *fd, > > char **addr) > > { > > size_t i; > > - int __attribute__((unused)) dummy = 0; > > + volatile int __attribute__((unused)) dummy = 0; > > The mistake made by whoever wrote this test was making 'dummy' a stack > variable. That lets the compiler figure out that it's unused. If you > make it a top-level variable (not static) so the compiler can't tell > whether it's referenced by a different compilation unit, it can't make > that deduction. And you don't need the stupid attibute or volatile on it. I did cringe a little before adding the volatile keyword. While not a fan of global variables, that might be better than all these keywords. I will send a v2 right away! Thanks. -- Pankaj
On 06.06.24 15:58, Pankaj Raghav (Samsung) wrote: > From: Pankaj Raghav <p.raghav@samsung.com> > > create_pagecache_thp_and_fd() in split_huge_page_test.c used the > variable dummy to perform mmap read. > > However, this test was skipped even on XFS which has large folio > support. The issue was compiler (gcc 13.2.0) was optimizing out the > dummy variable, therefore, not creating huge page in the page cache. > > Add volatile keyword to force compiler not to optimize out the loop > where we read from the mmaped addr. > > Signed-off-by: Pankaj Raghav <p.raghav@samsung.com> > --- > tools/testing/selftests/mm/split_huge_page_test.c | 2 +- > 1 file changed, 1 insertion(+), 1 deletion(-) > > diff --git a/tools/testing/selftests/mm/split_huge_page_test.c b/tools/testing/selftests/mm/split_huge_page_test.c > index d3c7f5fb3e7b..c573a58f80ab 100644 > --- a/tools/testing/selftests/mm/split_huge_page_test.c > +++ b/tools/testing/selftests/mm/split_huge_page_test.c > @@ -300,7 +300,7 @@ int create_pagecache_thp_and_fd(const char *testfile, size_t fd_size, int *fd, > char **addr) > { > size_t i; > - int __attribute__((unused)) dummy = 0; > + volatile int __attribute__((unused)) dummy = 0; > > srand(time(NULL)); > > > base-commit: d97496ca23a2d4ee80b7302849404859d9058bcd The rick we do in some other tests is: char *tmp; tmp = *whatever; asm volatile("" : "+r" (tmp));
On 06.06.24 17:56, David Hildenbrand wrote: > On 06.06.24 15:58, Pankaj Raghav (Samsung) wrote: >> From: Pankaj Raghav <p.raghav@samsung.com> >> >> create_pagecache_thp_and_fd() in split_huge_page_test.c used the >> variable dummy to perform mmap read. >> >> However, this test was skipped even on XFS which has large folio >> support. The issue was compiler (gcc 13.2.0) was optimizing out the >> dummy variable, therefore, not creating huge page in the page cache. >> >> Add volatile keyword to force compiler not to optimize out the loop >> where we read from the mmaped addr. >> >> Signed-off-by: Pankaj Raghav <p.raghav@samsung.com> >> --- >> tools/testing/selftests/mm/split_huge_page_test.c | 2 +- >> 1 file changed, 1 insertion(+), 1 deletion(-) >> >> diff --git a/tools/testing/selftests/mm/split_huge_page_test.c b/tools/testing/selftests/mm/split_huge_page_test.c >> index d3c7f5fb3e7b..c573a58f80ab 100644 >> --- a/tools/testing/selftests/mm/split_huge_page_test.c >> +++ b/tools/testing/selftests/mm/split_huge_page_test.c >> @@ -300,7 +300,7 @@ int create_pagecache_thp_and_fd(const char *testfile, size_t fd_size, int *fd, >> char **addr) >> { >> size_t i; >> - int __attribute__((unused)) dummy = 0; >> + volatile int __attribute__((unused)) dummy = 0; >> >> srand(time(NULL)); >> >> >> base-commit: d97496ca23a2d4ee80b7302849404859d9058bcd > > The rick we do in some other tests is: > > char *tmp; > > tmp = *whatever; > asm volatile("" : "+r" (tmp)); char tmp; of course. See cow.c as an example.
On Thu, Jun 06, 2024 at 05:57:21PM +0200, David Hildenbrand wrote: > On 06.06.24 17:56, David Hildenbrand wrote: > > On 06.06.24 15:58, Pankaj Raghav (Samsung) wrote: > > > From: Pankaj Raghav <p.raghav@samsung.com> > > > > > > create_pagecache_thp_and_fd() in split_huge_page_test.c used the > > > variable dummy to perform mmap read. > > > > > > However, this test was skipped even on XFS which has large folio > > > support. The issue was compiler (gcc 13.2.0) was optimizing out the > > > dummy variable, therefore, not creating huge page in the page cache. > > > > > > Add volatile keyword to force compiler not to optimize out the loop > > > where we read from the mmaped addr. > > > > > > Signed-off-by: Pankaj Raghav <p.raghav@samsung.com> > > > --- > > > tools/testing/selftests/mm/split_huge_page_test.c | 2 +- > > > 1 file changed, 1 insertion(+), 1 deletion(-) > > > > > > diff --git a/tools/testing/selftests/mm/split_huge_page_test.c b/tools/testing/selftests/mm/split_huge_page_test.c > > > index d3c7f5fb3e7b..c573a58f80ab 100644 > > > --- a/tools/testing/selftests/mm/split_huge_page_test.c > > > +++ b/tools/testing/selftests/mm/split_huge_page_test.c > > > @@ -300,7 +300,7 @@ int create_pagecache_thp_and_fd(const char *testfile, size_t fd_size, int *fd, > > > char **addr) > > > { > > > size_t i; > > > - int __attribute__((unused)) dummy = 0; > > > + volatile int __attribute__((unused)) dummy = 0; > > > srand(time(NULL)); > > > > > > base-commit: d97496ca23a2d4ee80b7302849404859d9058bcd > > > > The rick we do in some other tests is: > > > > char *tmp; > > > > tmp = *whatever; > > asm volatile("" : "+r" (tmp)); > > char tmp; of course. See cow.c as an example. Thanks David! I remember also seeing this when I grepped for volatile in the selftests directory. Willy gave the idea of making it as a global variable [1]. But your trick also works :) diff --git a/tools/testing/selftests/mm/split_huge_page_test.c b/tools/testing/selftests/mm/split_huge_page_test.c index d3c7f5fb3e7b..9c957703c1f7 100644 --- a/tools/testing/selftests/mm/split_huge_page_test.c +++ b/tools/testing/selftests/mm/split_huge_page_test.c @@ -341,6 +341,7 @@ int create_pagecache_thp_and_fd(const char *testfile, size_t fd_size, int *fd, for (size_t i = 0; i < fd_size; i++) dummy += *(*addr + i); + asm volatile("" : "+r" (dummy)); if (!check_huge_file(*addr, fd_size / pmd_pagesize, pmd_pagesize)) { ksft_print_msg("No large pagecache folio generated, please provide a filesystem supporting large folio\n"); I am fine with either solutions. But using the trick asm volatile is more cleaner than making it a global variable IMO and makes it more uniform across the other mm tests. Let me know what others think. [1] https://lore.kernel.org/linux-mm/20240606154428.672643-1-kernel@pankajraghav.com/
On 06.06.24 22:18, Pankaj Raghav (Samsung) wrote: > On Thu, Jun 06, 2024 at 05:57:21PM +0200, David Hildenbrand wrote: >> On 06.06.24 17:56, David Hildenbrand wrote: >>> On 06.06.24 15:58, Pankaj Raghav (Samsung) wrote: >>>> From: Pankaj Raghav <p.raghav@samsung.com> >>>> >>>> create_pagecache_thp_and_fd() in split_huge_page_test.c used the >>>> variable dummy to perform mmap read. >>>> >>>> However, this test was skipped even on XFS which has large folio >>>> support. The issue was compiler (gcc 13.2.0) was optimizing out the >>>> dummy variable, therefore, not creating huge page in the page cache. >>>> >>>> Add volatile keyword to force compiler not to optimize out the loop >>>> where we read from the mmaped addr. >>>> >>>> Signed-off-by: Pankaj Raghav <p.raghav@samsung.com> >>>> --- >>>> tools/testing/selftests/mm/split_huge_page_test.c | 2 +- >>>> 1 file changed, 1 insertion(+), 1 deletion(-) >>>> >>>> diff --git a/tools/testing/selftests/mm/split_huge_page_test.c b/tools/testing/selftests/mm/split_huge_page_test.c >>>> index d3c7f5fb3e7b..c573a58f80ab 100644 >>>> --- a/tools/testing/selftests/mm/split_huge_page_test.c >>>> +++ b/tools/testing/selftests/mm/split_huge_page_test.c >>>> @@ -300,7 +300,7 @@ int create_pagecache_thp_and_fd(const char *testfile, size_t fd_size, int *fd, >>>> char **addr) >>>> { >>>> size_t i; >>>> - int __attribute__((unused)) dummy = 0; >>>> + volatile int __attribute__((unused)) dummy = 0; >>>> srand(time(NULL)); >>>> >>>> base-commit: d97496ca23a2d4ee80b7302849404859d9058bcd >>> >>> The rick we do in some other tests is: >>> >>> char *tmp; >>> >>> tmp = *whatever; >>> asm volatile("" : "+r" (tmp)); >> >> char tmp; of course. See cow.c as an example. > Thanks David! I remember also seeing this when I grepped for volatile in > the selftests directory. > > Willy gave the idea of making it as a global variable [1]. But your > trick also works :) > > diff --git a/tools/testing/selftests/mm/split_huge_page_test.c b/tools/testing/selftests/mm/split_huge_page_test.c > index d3c7f5fb3e7b..9c957703c1f7 100644 > --- a/tools/testing/selftests/mm/split_huge_page_test.c > +++ b/tools/testing/selftests/mm/split_huge_page_test.c > @@ -341,6 +341,7 @@ int create_pagecache_thp_and_fd(const char *testfile, size_t fd_size, int *fd, > > for (size_t i = 0; i < fd_size; i++) > dummy += *(*addr + i); > + asm volatile("" : "+r" (dummy)); > > if (!check_huge_file(*addr, fd_size / pmd_pagesize, pmd_pagesize)) { > ksft_print_msg("No large pagecache folio generated, please provide a filesystem supporting large folio\n"); > > I am fine with either solutions. But using the trick asm volatile is more > cleaner than making it a global variable IMO and makes it more uniform > across the other mm tests. You can then even the "__attribute__((unused))", because the compile must assume that it is used inside the asm statement.
On Thu, Jun 06, 2024 at 10:21:51PM +0200, David Hildenbrand wrote: > On 06.06.24 22:18, Pankaj Raghav (Samsung) wrote: > > On Thu, Jun 06, 2024 at 05:57:21PM +0200, David Hildenbrand wrote: > > > On 06.06.24 17:56, David Hildenbrand wrote: > > > > On 06.06.24 15:58, Pankaj Raghav (Samsung) wrote: > > > > > From: Pankaj Raghav <p.raghav@samsung.com> > > > > > > > > > > create_pagecache_thp_and_fd() in split_huge_page_test.c used the > > > > > variable dummy to perform mmap read. > > > > > > > > > > However, this test was skipped even on XFS which has large folio > > > > > support. The issue was compiler (gcc 13.2.0) was optimizing out the > > > > > dummy variable, therefore, not creating huge page in the page cache. > > > > > > > > > > Add volatile keyword to force compiler not to optimize out the loop > > > > > where we read from the mmaped addr. > > > > > > > > > > Signed-off-by: Pankaj Raghav <p.raghav@samsung.com> > > > > > --- > > > > > tools/testing/selftests/mm/split_huge_page_test.c | 2 +- > > > > > 1 file changed, 1 insertion(+), 1 deletion(-) > > > > > > > > > > diff --git a/tools/testing/selftests/mm/split_huge_page_test.c b/tools/testing/selftests/mm/split_huge_page_test.c > > > > > index d3c7f5fb3e7b..c573a58f80ab 100644 > > > > > --- a/tools/testing/selftests/mm/split_huge_page_test.c > > > > > +++ b/tools/testing/selftests/mm/split_huge_page_test.c > > > > > @@ -300,7 +300,7 @@ int create_pagecache_thp_and_fd(const char *testfile, size_t fd_size, int *fd, > > > > > char **addr) > > > > > { > > > > > size_t i; > > > > > - int __attribute__((unused)) dummy = 0; > > > > > + volatile int __attribute__((unused)) dummy = 0; > > > > > srand(time(NULL)); > > > > > > > > > > base-commit: d97496ca23a2d4ee80b7302849404859d9058bcd > > > > > > > > The rick we do in some other tests is: > > > > > > > > char *tmp; > > > > > > > > tmp = *whatever; > > > > asm volatile("" : "+r" (tmp)); > > > > > > char tmp; of course. See cow.c as an example. > > Thanks David! I remember also seeing this when I grepped for volatile in > > the selftests directory. > > > > Willy gave the idea of making it as a global variable [1]. But your > > trick also works :) > > > > diff --git a/tools/testing/selftests/mm/split_huge_page_test.c b/tools/testing/selftests/mm/split_huge_page_test.c > > index d3c7f5fb3e7b..9c957703c1f7 100644 > > --- a/tools/testing/selftests/mm/split_huge_page_test.c > > +++ b/tools/testing/selftests/mm/split_huge_page_test.c > > @@ -341,6 +341,7 @@ int create_pagecache_thp_and_fd(const char *testfile, size_t fd_size, int *fd, > > for (size_t i = 0; i < fd_size; i++) > > dummy += *(*addr + i); > > + asm volatile("" : "+r" (dummy)); > > if (!check_huge_file(*addr, fd_size / pmd_pagesize, pmd_pagesize)) { > > ksft_print_msg("No large pagecache folio generated, please provide a filesystem supporting large folio\n"); > > > > I am fine with either solutions. But using the trick asm volatile is more > > cleaner than making it a global variable IMO and makes it more uniform > > across the other mm tests. > > You can then even the "__attribute__((unused))", because the compile must > assume that it is used inside the asm statement. Yup! I will send a new version. Thanks David. > > -- > Cheers, > > David / dhildenb >
diff --git a/tools/testing/selftests/mm/split_huge_page_test.c b/tools/testing/selftests/mm/split_huge_page_test.c index d3c7f5fb3e7b..c573a58f80ab 100644 --- a/tools/testing/selftests/mm/split_huge_page_test.c +++ b/tools/testing/selftests/mm/split_huge_page_test.c @@ -300,7 +300,7 @@ int create_pagecache_thp_and_fd(const char *testfile, size_t fd_size, int *fd, char **addr) { size_t i; - int __attribute__((unused)) dummy = 0; + volatile int __attribute__((unused)) dummy = 0; srand(time(NULL));