lkml.org 
[lkml]   [2021]   [Sep]   [28]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
SubjectRe: [PATCH v2 1/2] mm, thp: check page mapping when truncating page cache
From


On 9/28/21 6:24 AM, Song Liu wrote:
> On Fri, Sep 24, 2021 at 12:12 AM Rongwei Wang
> <rongwei.wang@linux.alibaba.com> wrote:
>>
>>
>>
>> On 9/24/21 10:43 AM, Andrew Morton wrote:
>>> On Thu, 23 Sep 2021 01:04:54 +0800 Rongwei Wang <rongwei.wang@linux.alibaba.com> wrote:
>>>
>>>>
>>>>
>>>>> On Sep 22, 2021, at 7:37 PM, Matthew Wilcox <willy@infradead.org> wrote:
>>>>>
>>>>> On Wed, Sep 22, 2021 at 03:06:44PM +0800, Rongwei Wang wrote:
>>>>>> Transparent huge page has supported read-only non-shmem files. The file-
>>>>>> backed THP is collapsed by khugepaged and truncated when written (for
>>>>>> shared libraries).
>>>>>>
>>>>>> However, there is race in two possible places.
>>>>>>
>>>>>> 1) multiple writers truncate the same page cache concurrently;
>>>>>> 2) collapse_file rolls back when writer truncates the page cache;
>>>>>
>>>>> As I've said before, the bug here is that somehow there is a writable fd
>>>>> to a file with THPs. That's what we need to track down and fix.
>>>> Hi, Matthew
>>>> I am not sure get your means. We know “mm, thp: relax the VM_DENYWRITE constraint on file-backed THPs"
>>>> Introduced file-backed THPs for DSO. It is possible {very rarely} for DSO to be opened in writeable way.
>>>>
>>>> ...
>>>>
>>>>> https://lore.kernel.org/linux-mm/YUdL3lFLFHzC80Wt@casper.infradead.org/
>>>> All in all, what you mean is that we should solve this race at the source?
>>>
>>> Matthew is being pretty clear here: we shouldn't be permitting
>>> userspace to get a writeable fd for a thp-backed file.
>>>
>>> Why are we permitting the DSO to be opened writeably? If there's a
>>> legitimate case for doing this then presumably "mm, thp: relax the
>> There is a use case to stress file-backed THP within attachment.
>> I test this case in a system which has enabled CONFIG_READ_ONLY_THP_FOR_FS:
>>
>> $ gcc -Wall -g -o stress_madvise_dso stress_madvise_dso.c
>> $ ulimit -s unlimited
>> $ ./stress_madvise_dso 10000 <libtest.so>
>>
>> the meaning of above parameters:
>> 10000: the max test time;
>> <libtest.so>: the DSO that will been mapped into file-backed THP by
>> madvise. It recommended that the text segment of DSO to be tested is
>> greater than 2M.
>>
>> The crash will been triggered at once in the latest kernel. And this
>> case also can used to trigger the bug that mentioned in our another patch.
>
> Hmm.. I am not able to use the repro program to crash the system. Not
> sure what I did wrong.
>
Hi
I have tried to check my test case again. Can you make sure the DSO that
you test have THP mapping?

If you are willing to try again, I can send my libtest.c which is used
to test by myself (actually, it shouldn't be target DSO problem).

Thanks very much!
> OTOH, does it make sense to block writes within khugepaged, like:
>
> diff --git i/mm/khugepaged.c w/mm/khugepaged.c
> index 045cc579f724e..ad7c41ec15027 100644
> --- i/mm/khugepaged.c
> +++ w/mm/khugepaged.c
> @@ -51,6 +51,7 @@ enum scan_result {
> SCAN_CGROUP_CHARGE_FAIL,
> SCAN_TRUNCATED,
> SCAN_PAGE_HAS_PRIVATE,
> + SCAN_BUSY_WRITE,
> };
>
> #define CREATE_TRACE_POINTS
> @@ -1652,6 +1653,11 @@ static void collapse_file(struct mm_struct *mm,
> /* Only allocate from the target node */
> gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
>
> + if (deny_write_access(file)) {
> + result = SCAN_BUSY_WRITE;
> + return;
> + }
> +
This can indeed avoid some possible races from source.

But, I am thinking about whether this will lead to DDoS attack?
I remember the reason of DSO has ignored MAP_DENYWRITE in kernel
is that DDoS attack. In addition, 'deny_write_access' will change
the behavior, such as user will get 'Text file busy' during
collapse_file. I am not sure whether the behavior changing is acceptable
in user space.

If it is acceptable, I am very willing to fix the races like your way.

Thanks!
> new_page = khugepaged_alloc_page(hpage, gfp, node);
> if (!new_page) {
> result = SCAN_ALLOC_HUGE_PAGE_FAIL;
> @@ -1863,19 +1869,6 @@ static void collapse_file(struct mm_struct *mm,
> else {
> __mod_lruvec_page_state(new_page, NR_FILE_THPS, nr);
> filemap_nr_thps_inc(mapping);
> - /*
> - * Paired with smp_mb() in do_dentry_open() to ensure
> - * i_writecount is up to date and the update to nr_thps is
> - * visible. Ensures the page cache will be truncated if the
> - * file is opened writable.
> - */
> - smp_mb();
> - if (inode_is_open_for_write(mapping->host)) {
> - result = SCAN_FAIL;
> - __mod_lruvec_page_state(new_page, NR_FILE_THPS, -nr);
> - filemap_nr_thps_dec(mapping);
> - goto xa_locked;
> - }
> }
>
> if (nr_none) {
> @@ -1976,6 +1969,7 @@ static void collapse_file(struct mm_struct *mm,
> VM_BUG_ON(!list_empty(&pagelist));
> if (!IS_ERR_OR_NULL(*hpage))
> mem_cgroup_uncharge(*hpage);
> + allow_write_access(file);
> /* TODO: tracepoints */
> }
> #ifndef _DEFAULT_SOURCE
#define _DEFAULT_SOURCE
#endif

#include <fcntl.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/mman.h>
#include <unistd.h>

/* "volatile" to forbid compiler optimization */
volatile static int x;

#define DO_0 ++x;
#define DO_1 {DO_0; DO_0; DO_0; DO_0; DO_0; DO_0; DO_0; DO_0; DO_0}
#define DO_2 {DO_1; DO_1; DO_1; DO_1; DO_1; DO_1; DO_1; DO_1; DO_1}
#define DO_3 {DO_2; DO_2; DO_2; DO_2; DO_2; DO_2; DO_2; DO_2; DO_2}
#define DO_4 {DO_3; DO_3; DO_3; DO_3; DO_3; DO_3; DO_3; DO_3; DO_3}
#define DO_5 {DO_4; DO_4; DO_4; DO_4; DO_4; DO_4; DO_4; DO_4; DO_4}
#define DO_6 {DO_5; DO_5; DO_5; DO_5; DO_5; DO_5; DO_5; DO_5; DO_5}
#define DO_7 {DO_6; DO_6; DO_6; DO_6; DO_6; DO_6; DO_6; DO_6; DO_6}

void libtest_work1(void)
{
printf("work 1\n");
DO_0;
}

void libtest_work2(void)
{
printf("work 2\n");
DO_2;
}

void libtest_work3(void)
{
printf("work 3\n");
DO_4;
}

void libtest_work4(void)
{
printf("work 4\n");
DO_6;
}
\
 
 \ /
  Last update: 2021-09-28 18:20    [W:0.316 / U:1.992 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site