[yocto] [PATCH 1/1] mm: msync: fix the behaviour msync on tmpfs

[Zumeng Chen]

于 2011年12月21日 23:38, Bruce Ashfield 写道:
> On 11-12-21 10:24 AM, zumeng.chen wrote:
>> This patch has been validated by the following commands
>> on both CPUs with or without cache alias, which is for
>>
>> http://bugzilla.pokylinux.org/show_bug.cgi?id=1429
>
> Glad to see this. I was composing an email that had just this
> question. Which bug, and which kernel version is the issue.
For all versions, including the latest mainline kernel

>>
>> root at routerstationpro:~# dmesg|grep alias
>> Primary data cache 32kB, 4-way, VIPT, cache aliases, linesize 32 bytes
>> root at routerstationpro:~# for i in `seq 1 1000`; do ./msync01 ;done
>>
>>
>> root at localhost:/tmp> dmesg|grep alias
>> Primary data cache 32kB, 4-way, PIPT, no aliases, linesize 32 bytes
>> root at localhost:/tmp> zcat /proc/config.gz |grep WINPATH
>> CONFIG_WINTEGRA_WINPATH=y
>> CONFIG_WINTEGRA_WINPATH3=y
>> CONFIG_SERIAL_8250_WINPATH=y
>> root at localhost:/tmp> for i in `seq 1 1000`; do ./msync01 ;done
>>
>> Passed.
>>
>> On 2011年12月21日 23:17, Zumeng Chen wrote:
>>> For tmpfs, no matter which flag(ASYNC or SYNC) gets from msync,
>>> there is not so much thing to do for CPUs without cache alias,
>>> But for some CPUs with cache alias, msync has to flush cache
>>> explicitly, which makes sure the data coherency between memory
>>> file and cache.
>
> Is it just coherency, or are there corruption concerns as well ?
There are two issues:
1 ) for TMPFS, nothing need to be done in sys_msync,
      we just return for all arches.

2 ) But for MIPS CPUs with cache alias(dmesg|grep alias),
      it maybe has the issue which reported by 1429 while
      the memory of memset used by msycn01 has other
      alias, then failure will be report.
      So, in this situation, we need to flush the related vma
      to make sure read correctly.
>
> In the commit message, it would be useful to put an example
> interaction that can trigger this issue, since it isn't entirely
> obvious from the message. More information is always better.
>
> Is this for the 3.0 and 2.6.37 kernels .. or just one, or the
> other ?
For both and the latest mainline.
>
>
>>>
>>> Signed-off-by: Zumeng Chen<zumeng.chen at windriver.com>
>>> ---
>>> include/linux/mm.h | 1 +
>>> mm/msync.c | 20 +++++++++++++++++++-
>>> mm/shmem.c | 5 +++++
>>> 3 files changed, 25 insertions(+), 1 deletions(-)
>>>
>>> diff --git a/include/linux/mm.h b/include/linux/mm.h
>>> index f59179b..858db90 100644
>>> --- a/include/linux/mm.h
>>> +++ b/include/linux/mm.h
>>> @@ -868,6 +868,7 @@ extern void pagefault_out_of_memory(void);
>>> extern void show_free_areas(unsigned int flags);
>>> extern bool skip_free_areas_node(unsigned int flags, int nid);
>>>
>>> +int is_shmem_file(struct file *file);
>
> If there aren't any more users of this other than your
> msync.c below, we shouldn't need this exposed in mm.h. I
> realized that it is implemented in shmem.c, but isn't there
> a less global place to add this export (consider mm/internal.h)?
> Also shouldn't it be extern, like the other defines in the file ?
Yes, I'll export it in include/linux/shmem_fs.h V2.
>
>>> int shmem_lock(struct file *file, int lock, struct user_struct *user);
>>> struct file *shmem_file_setup(const char *name, loff_t size, unsigned
>>> long flags);
>>> void shmem_set_file(struct vm_area_struct *vma, struct file *file);
>>> diff --git a/mm/msync.c b/mm/msync.c
>>> index 632df45..2f0d8fa 100644
>>> --- a/mm/msync.c
>>> +++ b/mm/msync.c
>>> @@ -13,6 +13,7 @@
>>> #include<linux/file.h>
>>> #include<linux/syscalls.h>
>>> #include<linux/sched.h>
>>> +#include<asm/cacheflush.h>
>>>
>>> /*
>>> * MS_SYNC syncs the entire file - including mappings.
>>> @@ -33,6 +34,7 @@ SYSCALL_DEFINE3(msync, unsigned long, start, size_t,
>>> len, int, flags)
>>> unsigned long end;
>>> struct mm_struct *mm = current->mm;
>>> struct vm_area_struct *vma;
>>> + struct file *file;
>>> int unmapped_error = 0;
>>> int error = -EINVAL;
>>>
>>> @@ -56,8 +58,24 @@ SYSCALL_DEFINE3(msync, unsigned long, start,
>>> size_t, len, int, flags)
>>> */
>>> down_read(&mm->mmap_sem);
>>> vma = find_vma(mm, start);
>>> +
>>> +#ifdef CONFIG_TMPFS
>>> + /*
>>> + * For tmpfs, no matter which flag(ASYNC or SYNC) gets from msync,
>>> + * there is not so much thing to do for CPUs without cache alias,
>>> + * But for some CPUs with cache alias, msync has to flush cache
>>> + * explicitly, which makes sure the data coherency between memory
>>> + * file and cache.
>>> + */
>>> + file = vma->vm_file;
>>> + if (is_shmem_file(file)) {
>>> + flush_cache_range(vma, start, start+len);
>>> + error = 0;
>
> Hasn't error already been set to 0 ? .. I see it in the
> original function being set right before this chunk of code
> will run.
Yes, see above item1
>
>>> + goto out_unlock;
>
> Do we really want this for all architectures/all boards ? This
> is common/global code, so we need to tread carefully.
Yes, done on powerpc 8513, and PMC without cache alias.
>
> Is there a way that we could do this by arch hooks .. or is
> flush_cache_range() sufficiently benign that it is little overhead
> and safe everywhere ?
Yeah, I have already taken this into accounts, and it
seems should be OK since only msync trigger this.
And only for TMPFS, for others, we'll go through and
skip flush_cache_range.
I'll add CPU_has_alias check in V2 too.
>
> In the end, I'm not a mm expert, so we can fix this up, think
> about it, and then run it past the people that really know :)
>
>>> + }
>>> +#endif
>>> +
>>> for (;;) {
>>> - struct file *file;
>
> There's an assignment:
>
>   file = vma->vm_file;
>
> Later in this routine, you've already done that above, so can't
> it be dropped ?
No, we'll goto the end if TMPFS.

Regards,
Zumeng
>
> Bruce
>
>>>
>>> /* Still start< end. */
>>> error = -ENOMEM;
>>> diff --git a/mm/shmem.c b/mm/shmem.c
>>> index 92e5c15..4fabfac 100644
>>> --- a/mm/shmem.c
>>> +++ b/mm/shmem.c
>>> @@ -2793,6 +2793,11 @@ static struct file_system_type tmpfs_fs_type = {
>>> .kill_sb = kill_litter_super,
>>> };
>>>
>>> +int is_shmem_file(struct file *file)
>>> +{
>>> + return (file->f_op ==&shmem_file_operations)? 1 : 0;
>>> +}
>>> +
>>> int __init init_tmpfs(void)
>>> {
>>> int error;
>>
>