[v2] sysdeps/posix/posix_fallocate*: Make emulated posix_fallocate() work properly
Commit Message
Emulated posix_fallocate() only writes data in one block if block
size is 4096, offset is 4095 and len is 2. The emulated code should
write data in two blocks in the case because it actually crosses two
blocks.
Signed-off-by: Xiao Yang <yangx.jy@cn.fujitsu.com>
---
sysdeps/posix/posix_fallocate.c | 17 +++++++++++++++++
sysdeps/posix/posix_fallocate64.c | 17 +++++++++++++++++
2 files changed, 34 insertions(+)
Comments
On 1/15/20 7:01 AM, Xiao Yang wrote:
> Emulated posix_fallocate() only writes data in one block if block
> size is 4096, offset is 4095 and len is 2. The emulated code should
> write data in two blocks in the case because it actually crosses two
> blocks.
>
> Signed-off-by: Xiao Yang <yangx.jy@cn.fujitsu.com>
Could you please clarify your copyright assignment status? I suppose
you ought to be covered by any agreement Fujitsu may have with the FSF
but I'm not a steward and hence do not have a way to check. I'll review
anyway given that there are patches from @cn.fujitsu.com in the repo.
> ---
> sysdeps/posix/posix_fallocate.c | 17 +++++++++++++++++
> sysdeps/posix/posix_fallocate64.c | 17 +++++++++++++++++
> 2 files changed, 34 insertions(+)
Please file a bug report describing the issue and also add a test case
to protect from regressions.
> diff --git a/sysdeps/posix/posix_fallocate.c b/sysdeps/posix/posix_fallocate.c
> index e7fccfc1c8..22e5fea091 100644
> --- a/sysdeps/posix/posix_fallocate.c
> +++ b/sysdeps/posix/posix_fallocate.c
> @@ -93,6 +93,23 @@ posix_fallocate (int fd, __off_t offset, __off_t len)
> increment = 4096;
> }
>
> + if (offset % increment + len % increment > increment)
> + {
> + if (offset < st.st_size)
> + {
> + unsigned char b;
> + ssize_t rdsize = __pread (fd, &b, 1, offset);
> + if (rdsize < 0)
> + return errno;
> + if (rdsize == 1 && b != 0)
> + goto next;
> + }
> +
> + if (__pwrite (fd, "", 1, offset) != 1)
> + return errno;
> + }
A better fix may be to fix the loop conditions to fold this corner case in.
Thanks,
Siddhesh
On 12/20/20 11:13 PM, Siddhesh Poyarekar wrote:
> On 1/15/20 7:01 AM, Xiao Yang wrote:
>> Emulated posix_fallocate() only writes data in one block if block
>> size is 4096, offset is 4095 and len is 2. The emulated code
>> should write data in two blocks in the case because it actually
>> crosses two blocks.
>>
>> Signed-off-by: Xiao Yang <yangx.jy@cn.fujitsu.com>
>
> Could you please clarify your copyright assignment status? I suppose
> you ought to be covered by any agreement Fujitsu may have with the
> FSF but I'm not a steward and hence do not have a way to check. I'll
> review anyway given that there are patches from @cn.fujitsu.com in
> the repo.
There is no assignment on file with the FSF for Fujitsu regarding glibc.
On 12/21/20 9:53 AM, Carlos O'Donell wrote:
> On 12/20/20 11:13 PM, Siddhesh Poyarekar wrote:
>> On 1/15/20 7:01 AM, Xiao Yang wrote:
>>> Emulated posix_fallocate() only writes data in one block if block
>>> size is 4096, offset is 4095 and len is 2. The emulated code
>>> should write data in two blocks in the case because it actually
>>> crosses two blocks.
>>>
>>> Signed-off-by: Xiao Yang <yangx.jy@cn.fujitsu.com>
>>
>> Could you please clarify your copyright assignment status? I suppose
>> you ought to be covered by any agreement Fujitsu may have with the
>> FSF but I'm not a steward and hence do not have a way to check. I'll
>> review anyway given that there are patches from @cn.fujitsu.com in
>> the repo.
>
> There is no assignment on file with the FSF for Fujitsu regarding glibc.
>
Thanks for confirming Carlos, I've marked the patch as such in patchwork.
Siddhesh
@@ -93,6 +93,23 @@ posix_fallocate (int fd, __off_t offset, __off_t len)
increment = 4096;
}
+ if (offset % increment + len % increment > increment)
+ {
+ if (offset < st.st_size)
+ {
+ unsigned char b;
+ ssize_t rdsize = __pread (fd, &b, 1, offset);
+ if (rdsize < 0)
+ return errno;
+ if (rdsize == 1 && b != 0)
+ goto next;
+ }
+
+ if (__pwrite (fd, "", 1, offset) != 1)
+ return errno;
+ }
+
+next:
/* Write a null byte to every block. This is racy; we currently
lack a better option. Compare-and-swap against a file mapping
might additional local races, but requires interposition of a
@@ -93,6 +93,23 @@ __posix_fallocate64_l64 (int fd, __off64_t offset, __off64_t len)
increment = 4096;
}
+ if (offset % increment + len % increment > increment)
+ {
+ if (offset < st.st_size)
+ {
+ unsigned char b;
+ ssize_t rdsize = __libc_pread64 (fd, &b, 1, offset);
+ if (rdsize < 0)
+ return errno;
+ if (rdsize == 1 && b != 0)
+ goto next;
+ }
+
+ if (__libc_pwrite64 (fd, "", 1, offset) != 1)
+ return errno;
+ }
+
+next:
/* Write a null byte to every block. This is racy; we currently
lack a better option. Compare-and-swap against a file mapping
might address local races, but requires interposition of a signal