[LTP] [RFC PATCH 2/2] hugemmap33: Test to detect bug with migrating gigantic pages

Cyril Hrubis chrubis@suse.cz
Thu Sep 7 13:30:18 CEST 2023


On Wed, May 24, 2023 at 05:39:30PM +0800, Li Wang wrote:
> Signed-off-by: Li Wang <liwang@redhat.com>
> ---
>  runtest/hugetlb                               |   1 +
>  testcases/kernel/mem/.gitignore               |   1 +
>  .../kernel/mem/hugetlb/hugemmap/hugemmap33.c  | 143 ++++++++++++++++++
>  3 files changed, 145 insertions(+)
>  create mode 100644 testcases/kernel/mem/hugetlb/hugemmap/hugemmap33.c
> 
> diff --git a/runtest/hugetlb b/runtest/hugetlb
> index 299c07ac9..3576e063d 100644
> --- a/runtest/hugetlb
> +++ b/runtest/hugetlb
> @@ -35,6 +35,7 @@ hugemmap29 hugemmap29
>  hugemmap30 hugemmap30
>  hugemmap31 hugemmap31
>  hugemmap32 hugemmap32
> +hugemmap32 hugemmap33
>  hugemmap05_1 hugemmap05 -m
>  hugemmap05_2 hugemmap05 -s
>  hugemmap05_3 hugemmap05 -s -m
> diff --git a/testcases/kernel/mem/.gitignore b/testcases/kernel/mem/.gitignore
> index 7258489ed..d130d4dcd 100644
> --- a/testcases/kernel/mem/.gitignore
> +++ b/testcases/kernel/mem/.gitignore
> @@ -34,6 +34,7 @@
>  /hugetlb/hugemmap/hugemmap30
>  /hugetlb/hugemmap/hugemmap31
>  /hugetlb/hugemmap/hugemmap32
> +/hugetlb/hugemmap/hugemmap33
>  /hugetlb/hugeshmat/hugeshmat01
>  /hugetlb/hugeshmat/hugeshmat02
>  /hugetlb/hugeshmat/hugeshmat03
> diff --git a/testcases/kernel/mem/hugetlb/hugemmap/hugemmap33.c b/testcases/kernel/mem/hugetlb/hugemmap/hugemmap33.c
> new file mode 100644
> index 000000000..12a470193
> --- /dev/null
> +++ b/testcases/kernel/mem/hugetlb/hugemmap/hugemmap33.c
> @@ -0,0 +1,143 @@
> +// SPDX-License-Identifier: GPL-2.0-or-later
> +/*
> + * Copyright (c) Linux Test Project, 2023
> + * Copyright (C) 2023, Red Hat, Inc.
> + *
> + * Author: David Hildenbrand <david@redhat.com>
> + * Port-to-LTP: Li Wang <liwang@redhat.com>
> + */
> +
> +/*\
> + * [Description]
> + *
> + * Migration code will first unmap the old page and replace the present PTE
> + * by a migration entry. Then migrate the page. Once that succeeded (and there
> + * are no unexpected page references), we replace the migration entries by
> + * proper present PTEs pointing at the new page.
> + *
> + * For ordinary pages we handle PTEs. For 2 MiB hugetlb/THP, it's PMDs.
> + * For 1 GiB hugetlb, it's PUDs.
> + *
> + * So without below commit, GUP-fast code was simply not aware that we could
> + * have migration entries stored in PUDs. Migration + GUP-fast code should be
> + * able to handle any such races.
> + *
> + * For example, GUP-fast will re-verify the PUD after pinning to make sure it
> + * didn't change. If it did change, it backs off.
> + *
> + * Migration code should detect the additional page references and back off
> + * as well.
> + *
> + *  commit 15494520b776aa2eadc3e2fefae524764cab9cea
> + *  Author: Qiujun Huang <hqjagain@gmail.com>
> + *  Date:   Thu Jan 30 22:12:10 2020 -0800
> + *
> + *      mm: fix gup_pud_range
> + *
> + */
> +
> +#define _GNU_SOURCE
> +#include <stdio.h>
> +#include <pthread.h>
> +#include <unistd.h>
> +#include <stdbool.h>
> +#include <sys/mman.h>
> +#include <sys/syscall.h>
> +#include <sys/ioctl.h>
> +#include <linux/mempolicy.h>
> +#include <linux/mman.h>
> +
> +#include "lapi/syscalls.h"
> +#include "numa_helper.h"
> +#include "hugetlb.h"
> +#if HAVE_NUMA_H
> +#include <numa.h>
> +#endif

This is a bit strange, do we actually use anything from numa.h?

Because:

- if we do it should be ifdefed as well, otherwise the test will not
  build without libnuma-devel

- if we do not, why do we include that header in the first place

> +static char *mem;
> +static size_t pagesize;
> +static size_t hugetlbsize;
> +static volatile int looping = 1;
> +
> +static void *migration_thread_fn(void *arg LTP_ATTRIBUTE_UNUSED)
> +{
> +	while (looping) {
> +		TST_EXP_PASS_SILENT(syscall(__NR_mbind, mem, hugetlbsize,
> +			MPOL_LOCAL, NULL, 0x7fful, MPOL_MF_MOVE));
> +	}
> +
> +	return NULL;
> +}
> +
> +static void run_test(void)
> +{
> +	ssize_t transferred;
> +	struct iovec iov;
> +	int fds[2];
> +
> +	pthread_t migration_thread;
> +
> +	if (!is_numa(NULL, NH_MEMS, 1))
> +		tst_brk(TCONF, "requires NUMA with at least 1 node");

I guess that everything as at least 1 numa node, unless CONFIG_NUMA is
off, so I wonder if we just need needs_kconfig CONFIG_NUMA=y instead.

> +	pagesize = getpagesize();
> +	hugetlbsize = 1 * 1024 * 1024 * 1024u;
> +
> +	mem = mmap(NULL, hugetlbsize, PROT_READ|PROT_WRITE,
> +		   MAP_PRIVATE|MAP_ANON|MAP_HUGETLB|MAP_HUGE_1GB,
> +		   -1, 0);
> +	if (mem == MAP_FAILED)
> +		tst_brk(TBROK | TERRNO, "mmap() failed");
> +
> +	memset(mem, 1, hugetlbsize);
> +
> +	/* Keep migrating the page around ... */
> +	TEST(pthread_create(&migration_thread, NULL, migration_thread_fn, NULL));
> +	if (TST_RET)
> +		tst_brk(TBROK | TRERRNO, "pthread_create failed");

We do have SAFE_PTHREAD_CREATE()

> +	while (looping) {
> +		if (pipe(fds) < 0)
> +			tst_brk(TBROK | TERRNO, "pipe() failed");

SAFE_PIPE()

> +		iov.iov_base = mem;
> +		iov.iov_len = pagesize;
> +		transferred = vmsplice(fds[1], &iov, 1, 0);
> +		if (transferred <= 0)
> +			tst_brk(TBROK | TERRNO, "vmsplice() failed");
> +
> +		close(fds[0]);
> +		close(fds[1]);

SAFE_CLOSE() ?

> +		if (!tst_remaining_runtime()) {
> +			tst_res(TINFO, "Runtime exhausted, exiting");
> +			looping = 0;
> +		}
> +	}
> +
> +	TEST(pthread_join(migration_thread, NULL));
> +	if (TST_RET)
> +		tst_brk(TBROK | TRERRNO, "pthread_join failed");

SAFE_PTHREAD_...()

> +	if (tst_taint_check())
> +		tst_res(TFAIL, "Test resulted in kernel tainted");
> +	else
> +		tst_res(TPASS, "Test completed successfully");

tst_test.taint_check ?


> +}
> +
> +static struct tst_test test = {
> +	.needs_root = 1,
> +	.test_all = run_test,
> +	.max_runtime = 60,
> +	.hugepages = {2, TST_NEEDS, 1048576},
                                    ^
				    So this is 1GB hugepage?

				    I wonder if it would be bettter to
				    define constants for these, or at
				    least write it as 1024 * 1024
> +	.taint_check = TST_TAINT_W | TST_TAINT_D,
> +	.tags = (struct tst_tag[]) {
> +	    {"linux-git", "15494520b776"},
> +	    {}
> +	},
> +	.supported_archs = (const char *const []) {
> +		"x86",
> +		"x86_64",
> +		NULL

I do not see anything x86 specific in the test code, or did I miss
something?

I guess that we can leave the test enabled for all architectures as long
as we check for the existente of the hugepages of the requested size in
the library.

> +	},
> +};
> -- 
> 2.40.0
> 
> 
> -- 
> Mailing list info: https://lists.linux.it/listinfo/ltp

-- 
Cyril Hrubis
chrubis@suse.cz


More information about the ltp mailing list