[v3,5/5] cgroup: Add memcontrol02

This test appears to compare the overall "current" counter with anon
and file (page cache) memory counters. The file test assumes much more
memory will be consumed by the page cache. This is certainly true on
XFS and BTRFS where little or no memory outside the page cache is
used. However on ext4, ntfs and especially exfat, we can see more
memory being used.

This seems to be related to fs/buffer.c and buffer_head usage. exfat
in particular allocates a lot of memory. This is possibly due to the
buffer_head being allocated in cont_write_begin:

 Children      Self       Samples         bytes_alloc  Parent symbol
 ........  ........  ............  ..................  .................

    97.66%    97.66%        102401                 168  exfat_write_begin
            |
            ---0xf7eec549
               entry_SYSENTER_compat_after_hwframe
               do_fast_syscall_32
               __noinstr_text_start
               ksys_write
               vfs_write
               new_sync_write
               generic_file_write_iter
               __generic_file_write_iter
               generic_perform_write
               exfat_write_begin
               cont_write_begin
               __block_write_begin_int
               create_page_buffers
               create_empty_buffers
               alloc_page_buffers
               alloc_buffer_head
               kmem_cache_alloc
               kmem_cache_alloc

     0.20%     0.20%           205                 584  exfat_write_begin
     0.06%     0.06%            64                 168  exfat_fill_super
     0.05%     0.05%            49                 256  exfat_fill_super
     0.00%     0.00%             4                 584  exfat_fill_super
     0.00%     0.00%             1                 256  exfat_create
     0.00%     0.00%             1                1528  exfat_create
     0.00%     0.00%             1                1528  exfat_fill_super
     0.00%     0.00%             1                 312  exfat_fill_super

We can see (using slub_debug=U,buffer_head) that these buffer_head
objects are not freed until the file is unlinked:

 Outputting '/sys/kernel/debug/slab/buffer_head/alloc_traces' ...
 102480 alloc_buffer_head+0x1b/0x90 age=0/268/2629529 pid=248-430 cpus=0,2

 Outputting '/sys/kernel/debug/slab/buffer_head/free_traces' ...
 102400 <not-available> age=4297543659 pid=0 cpus=0
     80 free_buffer_head+0x21/0x60 age=147/266217/2629538 pid=337-427 cpus=1-2

ext4 and ntfs also use some of the "buffer" code, but don't seem to
allocate quite as much. Although ext4 begins to fail when slub debug
is enabled due to the extra memory debugging uses.

In any case, it appears that the CGroup code is correct, so I have
increased the error margin when exfat or ext234 is detected.

Signed-off-by: Richard Palethorpe <rpalethorpe@suse.com>
---
 runtest/controllers                           |   1 +
 testcases/kernel/controllers/memcg/.gitignore |   1 +
 .../kernel/controllers/memcg/memcontrol02.c   | 183 ++++++++++++++++++
 3 files changed, 185 insertions(+)
 create mode 100644 testcases/kernel/controllers/memcg/memcontrol02.c

Hi!
> This test appears to compare the overall "current" counter with anon
> and file (page cache) memory counters. The file test assumes much more
> memory will be consumed by the page cache. This is certainly true on
> XFS and BTRFS where little or no memory outside the page cache is
> used. However on ext4, ntfs and especially exfat, we can see more
> memory being used.
> 
> This seems to be related to fs/buffer.c and buffer_head usage. exfat
> in particular allocates a lot of memory. This is possibly due to the
> buffer_head being allocated in cont_write_begin:
> 
>  Children      Self       Samples         bytes_alloc  Parent symbol
>  ........  ........  ............  ..................  .................
> 
>     97.66%    97.66%        102401                 168  exfat_write_begin
>             |
>             ---0xf7eec549
>                entry_SYSENTER_compat_after_hwframe
>                do_fast_syscall_32
>                __noinstr_text_start
>                ksys_write
>                vfs_write
>                new_sync_write
>                generic_file_write_iter
>                __generic_file_write_iter
>                generic_perform_write
>                exfat_write_begin
>                cont_write_begin
>                __block_write_begin_int
>                create_page_buffers
>                create_empty_buffers
>                alloc_page_buffers
>                alloc_buffer_head
>                kmem_cache_alloc
>                kmem_cache_alloc
> 
>      0.20%     0.20%           205                 584  exfat_write_begin
>      0.06%     0.06%            64                 168  exfat_fill_super
>      0.05%     0.05%            49                 256  exfat_fill_super
>      0.00%     0.00%             4                 584  exfat_fill_super
>      0.00%     0.00%             1                 256  exfat_create
>      0.00%     0.00%             1                1528  exfat_create
>      0.00%     0.00%             1                1528  exfat_fill_super
>      0.00%     0.00%             1                 312  exfat_fill_super
> 
> We can see (using slub_debug=U,buffer_head) that these buffer_head
> objects are not freed until the file is unlinked:
> 
>  Outputting '/sys/kernel/debug/slab/buffer_head/alloc_traces' ...
>  102480 alloc_buffer_head+0x1b/0x90 age=0/268/2629529 pid=248-430 cpus=0,2
> 
>  Outputting '/sys/kernel/debug/slab/buffer_head/free_traces' ...
>  102400 <not-available> age=4297543659 pid=0 cpus=0
>      80 free_buffer_head+0x21/0x60 age=147/266217/2629538 pid=337-427 cpus=1-2
> 
> ext4 and ntfs also use some of the "buffer" code, but don't seem to
> allocate quite as much. Although ext4 begins to fail when slub debug
> is enabled due to the extra memory debugging uses.
> 
> In any case, it appears that the CGroup code is correct, so I have
> increased the error margin when exfat or ext234 is detected.
> 
> Signed-off-by: Richard Palethorpe <rpalethorpe@suse.com>
> ---
>  runtest/controllers                           |   1 +
>  testcases/kernel/controllers/memcg/.gitignore |   1 +
>  .../kernel/controllers/memcg/memcontrol02.c   | 183 ++++++++++++++++++
>  3 files changed, 185 insertions(+)
>  create mode 100644 testcases/kernel/controllers/memcg/memcontrol02.c
> 
> diff --git a/runtest/controllers b/runtest/controllers
> index 2b41a94d3..09e0107e4 100644
> --- a/runtest/controllers
> +++ b/runtest/controllers
> @@ -18,6 +18,7 @@ memcg_control		memcg_control_test.sh
>  
>  # kselftest ports
>  memcontrol01 memcontrol01
> +memcontrol02 memcontrol02
>  
>  cgroup_fj_function_debug cgroup_fj_function.sh debug
>  cgroup_fj_function_cpuset cgroup_fj_function.sh cpuset
> diff --git a/testcases/kernel/controllers/memcg/.gitignore b/testcases/kernel/controllers/memcg/.gitignore
> index c3565f85c..f7de40d53 100644
> --- a/testcases/kernel/controllers/memcg/.gitignore
> +++ b/testcases/kernel/controllers/memcg/.gitignore
> @@ -6,3 +6,4 @@
>  /regression/memcg_test_4
>  /stress/memcg_process_stress
>  memcontrol01
> +memcontrol02
> diff --git a/testcases/kernel/controllers/memcg/memcontrol02.c b/testcases/kernel/controllers/memcg/memcontrol02.c
> new file mode 100644
> index 000000000..f236e6cdc
> --- /dev/null
> +++ b/testcases/kernel/controllers/memcg/memcontrol02.c
> @@ -0,0 +1,183 @@
> +// SPDX-License-Identifier: GPL-2.0
> +/*\
> + *
> + * [Description]
> + *
> + * Conversion of second kself test in cgroup/test_memcontrol.c.
> + *
> + * Original description:
> + * "This test creates a memory cgroup, allocates some anonymous memory
> + * and some pagecache and check memory.current and some memory.stat
> + * values."
> + *
> + * Note that the V1 rss and cache counters were renamed to anon and
> + * file in V2. Besides error reporting, this test differs from the
> + * kselftest in the following ways:
> + *
> + * . It supports V1.
> + * . It writes instead of reads to fill the page cache. Because no
> + *   pages were allocated on tmpfs.

Shouldn't we actually run the test both for read/write and skip the read
part of tmpfs?

Well I guess that the pages are put into the page cache the same way
regardless if they came from userspace write or as a request for data to
be read from the disk, so probably not I guess.

> + * . It runs on most filesystems available
> + * . On EXFAT and extN we change the margine of error between all and file
                                           ^
					   margin
> + *   memory to 50%. Because these allocate non-page-cache memory during writes.
> + */
> +#define _GNU_SOURCE
> +
> +#include <stdlib.h>
> +#include <stdio.h>

Do we really need stdio here?

> +#include "tst_test.h"
> +#include "tst_cgroup.h"
> +
> +#define TMPDIR "mntdir"
> +#define MB(x) (x << 20)
> +
> +static size_t page_size;
> +static const struct tst_cgroup_group *cg_test;
> +static int is_v1_memcg;
> +static struct tst_cgroup_group *cg_child;
> +static int fd;
> +static int file_to_all_error = 10;
> +
> +/*
> + * Checks if two given values differ by less than err% of their sum.
> + */
> +static inline int values_close(const ssize_t a,
> +			       const ssize_t b,
> +			       const ssize_t err)
> +{
> +	return labs(a - b) <= (a + b) / 100 * err;
> +}

I guess that this crude integer version works only since we allocate
reasonable large memory sizes.

> +static void alloc_anon_50M_check(void)
> +{
> +	const ssize_t size = MB(50);
> +	char *buf, *ptr;
> +	ssize_t anon, current;
> +	const char *const anon_key_fmt = is_v1_memcg ? "rss %zd" : "anon %zd";
> +
> +	buf = SAFE_MALLOC(size);
> +	for (ptr = buf; ptr < buf + size; ptr += page_size)
> +		*ptr = 0;
> +
> +	SAFE_CGROUP_SCANF(cg_child, "memory.current", "%zd", &current);
> +	TST_EXP_EXPR(current >= size,
> +		     "(memory.current=%zd) >= (size=%zd)", current, size);
> +
> +	SAFE_CGROUP_LINES_SCANF(cg_child, "memory.stat", anon_key_fmt, &anon);
> +
> +	TST_EXP_EXPR(anon > 0, "(memory.stat.anon=%zd) > 0", anon);
> +	TST_EXP_EXPR(values_close(size, current, 3),
> +		     "(size=%zd) ~= (memory.stat.anon=%zd)", size, current);
> +	TST_EXP_EXPR(values_close(anon, current, 3),
> +		     "(memory.current=%zd) ~= (memory.stat.anon=%zd)",
> +		     current, anon);
> +}
> +
> +static void alloc_pagecache(const int fd, size_t size)
> +{
> +	char buf[BUFSIZ];
> +	size_t i;

We may as well fill the buffer with something so that this function will
not trigger static analyzers.

> +	for (i = 0; i < size; i += sizeof(buf))
> +		SAFE_WRITE(1, fd, buf, sizeof(buf));
> +}
> +
> +static void alloc_pagecache_50M_check(void)
> +{
> +	const size_t size = MB(50);
> +	size_t current, file;
> +	const char *const file_key_fmt = is_v1_memcg ? "cache %zd" : "file %zd";

This may be a global initialized in the test setup.

> +	TEST(open(TMPDIR"/tmpfile", O_RDWR | O_CREAT, 0600));
> +
> +	if (TST_RET < 0) {
> +		if (TST_ERR == EOPNOTSUPP)
> +			tst_brk(TCONF, "O_TMPFILE not supported by FS");
> +
> +		tst_brk(TBROK | TTERRNO,
> +			"open(%s, O_TMPFILE | O_RDWR | O_EXCL", TMPDIR"/.");
> +	}
> +	fd = TST_RET;
> +
> +	SAFE_CGROUP_SCANF(cg_child, "memory.current", "%zu", &current);
> +	tst_res(TINFO, "Created temp file: memory.current=%zu", current);
> +
> +	alloc_pagecache(fd, size);
> +
> +	SAFE_CGROUP_SCANF(cg_child, "memory.current", "%zu", &current);
> +	TST_EXP_EXPR(current >= size,
> +			 "(memory.current=%zu) >= (size=%zu)", current, size);
> +
> +	SAFE_CGROUP_LINES_SCANF(cg_child, "memory.stat", file_key_fmt, &file);
> +	TST_EXP_EXPR(file > 0, "(memory.stat.file=%zd) > 0", file);
> +
> +	TST_EXP_EXPR(values_close(file, current, file_to_all_error),
> +			 "(memory.current=%zd) ~= (memory.stat.file=%zd)",
> +			 current, file);
> +
> +	SAFE_CLOSE(fd);
> +}
> +
> +static void test_memcg_current(unsigned int n)
> +{
> +	size_t current;
> +
> +	cg_child = tst_cgroup_group_mk(cg_test, "child");
> +	SAFE_CGROUP_SCANF(cg_child, "memory.current", "%zu", &current);
> +	TST_EXP_EXPR(current == 0, "(current=%zu) == 0", current);
> +
> +	if (!SAFE_FORK()) {
> +		SAFE_CGROUP_PRINTF(cg_child, "cgroup.procs", "%d", getpid());
> +
> +		SAFE_CGROUP_SCANF(cg_child, "memory.current", "%zu", &current);
> +		tst_res(TINFO, "Added proc to memcg: memory.current=%zu",
> +			current);
> +
> +		if (!n)
> +			alloc_anon_50M_check();
> +		else
> +			alloc_pagecache_50M_check();
> +	} else {
> +		tst_reap_children();
> +		cg_child = tst_cgroup_group_rm(cg_child);
> +	}
> +}
> +
> +static void setup(void)
> +{
> +	page_size = SAFE_SYSCONF(_SC_PAGESIZE);
> +
> +	tst_cgroup_require("memory", NULL);
> +	cg_test = tst_cgroup_get_test_group();
> +
> +	is_v1_memcg = TST_CGROUP_VER(cg_test, "memory") == TST_CGROUP_V1;

I find this statement a bit confusing, maybe TST_CGGROUP_IS_V1() and
TST_CGROUP_IS_V2() macros in the library would make it slightly better.

> +	switch (tst_fs_type(TMPDIR)) {
> +	case TST_EXFAT_MAGIC:
> +	case TST_EXT234_MAGIC:
> +		file_to_all_error = 50;
> +		break;
> +	}
> +}
> +
> +static void cleanup(void)
> +{
> +	if (cg_child)
> +		cg_child = tst_cgroup_group_rm(cg_child);
> +
> +	tst_cgroup_cleanup();
> +}
> +
> +static struct tst_test test = {
> +	.setup = setup,
> +	.cleanup = cleanup,
> +	.tcnt = 2,
> +	.test = test_memcg_current,
> +	.mount_device = 1,
> +	.dev_min_size = 256,
> +	.mntpoint = TMPDIR,
> +	.all_filesystems = 1,
> +	.forks_child = 1,
> +	.needs_root = 1,
> +};

Generally minus typos and minor things it looks good, I guess that it
can as well go in if you fix the typos as it is.

Reviewed-by: Cyril Hrubis <chrubis@susec.cz>

Hello,

Cyril Hrubis <chrubis@suse.cz> writes:

>> + *
>> + * [Description]
>> + *
>> + * Conversion of second kself test in cgroup/test_memcontrol.c.
>> + *
>> + * Original description:
>> + * "This test creates a memory cgroup, allocates some anonymous memory
>> + * and some pagecache and check memory.current and some memory.stat
>> + * values."
>> + *
>> + * Note that the V1 rss and cache counters were renamed to anon and
>> + * file in V2. Besides error reporting, this test differs from the
>> + * kselftest in the following ways:
>> + *
>> + * . It supports V1.
>> + * . It writes instead of reads to fill the page cache. Because no
>> + *   pages were allocated on tmpfs.
>
> Shouldn't we actually run the test both for read/write and skip the read
> part of tmpfs?
>
> Well I guess that the pages are put into the page cache the same way
> regardless if they came from userspace write or as a request for data to
> be read from the disk, so probably not I guess.

I reckon there are a lot of ways to fill the page cache from
userland. mmap and madvise also come to mind. I don't know how many ways
there are to get/allocate a page from the page cache internally. I guess
it's possible to circumvent the accountancy.

I think for now just writing is good enough. This doesn't appear to be
the only test which measures the page cache. So I think we should look
at what the other tests do first. 

>
>> + * . It runs on most filesystems available
>> + * . On EXFAT and extN we change the margine of error between all and file
>                                            ^
> 					   margin
>> + *   memory to 50%. Because these allocate non-page-cache memory during writes.
>> + */
>> +#define _GNU_SOURCE
>> +
>> +#include <stdlib.h>
>> +#include <stdio.h>
>
> Do we really need stdio here?

Sorry, nope.

>
>> +#include "tst_test.h"
>> +#include "tst_cgroup.h"
>> +
>> +#define TMPDIR "mntdir"
>> +#define MB(x) (x << 20)
>> +
>> +static size_t page_size;
>> +static const struct tst_cgroup_group *cg_test;
>> +static int is_v1_memcg;
>> +static struct tst_cgroup_group *cg_child;
>> +static int fd;
>> +static int file_to_all_error = 10;
>> +
>> +/*
>> + * Checks if two given values differ by less than err% of their sum.
>> + */
>> +static inline int values_close(const ssize_t a,
>> +			       const ssize_t b,
>> +			       const ssize_t err)
>> +{
>> +	return labs(a - b) <= (a + b) / 100 * err;
>> +}
>
> I guess that this crude integer version works only since we allocate
> reasonable large memory sizes.

I suppose I could multiply both sides of the inequality by
100... perhaps we could even use the FPU, I hear it's quite usable in
userland. :-p

>
>> +static void alloc_anon_50M_check(void)
>> +{
>> +	const ssize_t size = MB(50);
>> +	char *buf, *ptr;
>> +	ssize_t anon, current;
>> +	const char *const anon_key_fmt = is_v1_memcg ? "rss %zd" : "anon %zd";
>> +
>> +	buf = SAFE_MALLOC(size);
>> +	for (ptr = buf; ptr < buf + size; ptr += page_size)
>> +		*ptr = 0;
>> +
>> +	SAFE_CGROUP_SCANF(cg_child, "memory.current", "%zd", &current);
>> +	TST_EXP_EXPR(current >= size,
>> +		     "(memory.current=%zd) >= (size=%zd)", current, size);
>> +
>> +	SAFE_CGROUP_LINES_SCANF(cg_child, "memory.stat", anon_key_fmt, &anon);
>> +
>> +	TST_EXP_EXPR(anon > 0, "(memory.stat.anon=%zd) > 0", anon);
>> +	TST_EXP_EXPR(values_close(size, current, 3),
>> +		     "(size=%zd) ~= (memory.stat.anon=%zd)", size, current);
>> +	TST_EXP_EXPR(values_close(anon, current, 3),
>> +		     "(memory.current=%zd) ~= (memory.stat.anon=%zd)",
>> +		     current, anon);
>> +}
>> +
>> +static void alloc_pagecache(const int fd, size_t size)
>> +{
>> +	char buf[BUFSIZ];
>> +	size_t i;
>
> We may as well fill the buffer with something so that this function will
> not trigger static analyzers.

+1
>
>> +	for (i = 0; i < size; i += sizeof(buf))
>> +		SAFE_WRITE(1, fd, buf, sizeof(buf));
>> +}
>> +
>> +static void alloc_pagecache_50M_check(void)
>> +{
>> +	const size_t size = MB(50);
>> +	size_t current, file;
>> +	const char *const file_key_fmt = is_v1_memcg ? "cache %zd" : "file %zd";
>
> This may be a global initialized in the test setup.

Is it not easier to read when set in the context it is used? Otherwise
the key name is hidden in setup.

>
>> +	TEST(open(TMPDIR"/tmpfile", O_RDWR | O_CREAT, 0600));
>> +
>> +	if (TST_RET < 0) {
>> +		if (TST_ERR == EOPNOTSUPP)
>> +			tst_brk(TCONF, "O_TMPFILE not supported by FS");
>> +
>> +		tst_brk(TBROK | TTERRNO,
>> +			"open(%s, O_TMPFILE | O_RDWR | O_EXCL", TMPDIR"/.");
>> +	}
>> +	fd = TST_RET;
>> +
>> +	SAFE_CGROUP_SCANF(cg_child, "memory.current", "%zu", &current);
>> +	tst_res(TINFO, "Created temp file: memory.current=%zu", current);
>> +
>> +	alloc_pagecache(fd, size);
>> +
>> +	SAFE_CGROUP_SCANF(cg_child, "memory.current", "%zu", &current);
>> +	TST_EXP_EXPR(current >= size,
>> +			 "(memory.current=%zu) >= (size=%zu)", current, size);
>> +
>> +	SAFE_CGROUP_LINES_SCANF(cg_child, "memory.stat", file_key_fmt, &file);
>> +	TST_EXP_EXPR(file > 0, "(memory.stat.file=%zd) > 0", file);
>> +
>> +	TST_EXP_EXPR(values_close(file, current, file_to_all_error),
>> +			 "(memory.current=%zd) ~= (memory.stat.file=%zd)",
>> +			 current, file);
>> +
>> +	SAFE_CLOSE(fd);
>> +}
>> +
>> +static void test_memcg_current(unsigned int n)
>> +{
>> +	size_t current;
>> +
>> +	cg_child = tst_cgroup_group_mk(cg_test, "child");
>> +	SAFE_CGROUP_SCANF(cg_child, "memory.current", "%zu", &current);
>> +	TST_EXP_EXPR(current == 0, "(current=%zu) == 0", current);
>> +
>> +	if (!SAFE_FORK()) {
>> +		SAFE_CGROUP_PRINTF(cg_child, "cgroup.procs", "%d", getpid());
>> +
>> +		SAFE_CGROUP_SCANF(cg_child, "memory.current", "%zu", &current);
>> +		tst_res(TINFO, "Added proc to memcg: memory.current=%zu",
>> +			current);
>> +
>> +		if (!n)
>> +			alloc_anon_50M_check();
>> +		else
>> +			alloc_pagecache_50M_check();
>> +	} else {
>> +		tst_reap_children();
>> +		cg_child = tst_cgroup_group_rm(cg_child);
>> +	}
>> +}
>> +
>> +static void setup(void)
>> +{
>> +	page_size = SAFE_SYSCONF(_SC_PAGESIZE);
>> +
>> +	tst_cgroup_require("memory", NULL);
>> +	cg_test = tst_cgroup_get_test_group();
>> +
>> +	is_v1_memcg = TST_CGROUP_VER(cg_test, "memory") == TST_CGROUP_V1;
>
> I find this statement a bit confusing, maybe TST_CGGROUP_IS_V1() and
> TST_CGROUP_IS_V2() macros in the library would make it slightly
> better.

Yeah, I suppose this is proving to be tiresome.

>
>> +	switch (tst_fs_type(TMPDIR)) {
>> +	case TST_EXFAT_MAGIC:
>> +	case TST_EXT234_MAGIC:
>> +		file_to_all_error = 50;
>> +		break;
>> +	}
>> +}
>> +
>> +static void cleanup(void)
>> +{
>> +	if (cg_child)
>> +		cg_child = tst_cgroup_group_rm(cg_child);
>> +
>> +	tst_cgroup_cleanup();
>> +}
>> +
>> +static struct tst_test test = {
>> +	.setup = setup,
>> +	.cleanup = cleanup,
>> +	.tcnt = 2,
>> +	.test = test_memcg_current,
>> +	.mount_device = 1,
>> +	.dev_min_size = 256,
>> +	.mntpoint = TMPDIR,
>> +	.all_filesystems = 1,
>> +	.forks_child = 1,
>> +	.needs_root = 1,
>> +};
>
> Generally minus typos and minor things it looks good, I guess that it
> can as well go in if you fix the typos as it is.
>
> Reviewed-by: Cyril Hrubis <chrubis@susec.cz>

Hello,

Richard Palethorpe <rpalethorpe@suse.de> writes:

> Hello,
>
> Cyril Hrubis <chrubis@suse.cz> writes:
>
>>> + *
>>> + * [Description]
>>> + *
>>> + * Conversion of second kself test in cgroup/test_memcontrol.c.
>>> + *
>>> + * Original description:
>>> + * "This test creates a memory cgroup, allocates some anonymous memory
>>> + * and some pagecache and check memory.current and some memory.stat
>>> + * values."
>>> + *
>>> + * Note that the V1 rss and cache counters were renamed to anon and
>>> + * file in V2. Besides error reporting, this test differs from the
>>> + * kselftest in the following ways:
>>> + *
>>> + * . It supports V1.
>>> + * . It writes instead of reads to fill the page cache. Because no
>>> + *   pages were allocated on tmpfs.
>>
>> Shouldn't we actually run the test both for read/write and skip the read
>> part of tmpfs?
>>
>> Well I guess that the pages are put into the page cache the same way
>> regardless if they came from userspace write or as a request for data to
>> be read from the disk, so probably not I guess.
>
> I reckon there are a lot of ways to fill the page cache from
> userland. mmap and madvise also come to mind. I don't know how many ways
> there are to get/allocate a page from the page cache internally. I guess
> it's possible to circumvent the accountancy.
>
> I think for now just writing is good enough. This doesn't appear to be
> the only test which measures the page cache. So I think we should look
> at what the other tests do first. 
>
>>
>>> + * . It runs on most filesystems available
>>> + * . On EXFAT and extN we change the margine of error between all and file
>>                                            ^
>> 					   margin
>>> + *   memory to 50%. Because these allocate non-page-cache memory during writes.
>>> + */
>>> +#define _GNU_SOURCE
>>> +
>>> +#include <stdlib.h>
>>> +#include <stdio.h>
>>
>> Do we really need stdio here?
>
> Sorry, nope.
>

Actually it is needed otherwise BUFSIZ is undefined...