Re: [PATCH 3/6] CIFS: Make write call work with strict cache mode (try #2)

[Suresh Jayaraman <sjayaraman-l3A5Bk7waGM@public.gmane.org>]

On 11/29/2010 05:07 PM, Jeff Layton wrote:
> On Mon, 29 Nov 2010 13:58:05 +0300
> Pavel Shilovsky <piastryyy-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org> wrote:
> 
>> 2010/11/28 Jeff Layton <jlayton-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>:
>>> On Sun, 28 Nov 2010 06:36:04 -0500
>>> Jeff Layton <jlayton-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org> wrote:
>>>
>>>> On Sun, 28 Nov 2010 11:12:49 +0300
>>>> Pavel Shilovsky <piastryyy-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org> wrote:
>>>>
>>>>> On strict cache mode if we don't have Exclusive oplock we write a data to
>>>>> the server through cifs_user_write. Then if we Level II oplock store it in
>>>>> the cache, otherwise - invalidate inode pages affected by this writing.
>>>>>
>>>>> Signed-off-by: Pavel Shilovsky <piastryyy-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org>
>>>>> ---
>>>>> �fs/cifs/cifsfs.c | � 40 ++++++++++++++++++++++++++++++++++++----
>>>>> �1 files changed, 36 insertions(+), 4 deletions(-)
>>>>>
>>>>> diff --git a/fs/cifs/cifsfs.c b/fs/cifs/cifsfs.c
>>>>> index bbb5294..901c82b 100644
>>>>> --- a/fs/cifs/cifsfs.c
>>>>> +++ b/fs/cifs/cifsfs.c
>>>>> @@ -598,12 +598,44 @@ static ssize_t cifs_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
>>>>> �static ssize_t cifs_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
>>>>> � � � � � � � � � � � � � � � �unsigned long nr_segs, loff_t pos)
>>>>> �{
>>>>> - � struct inode *inode = iocb->ki_filp->f_path.dentry->d_inode;
>>>>> + � struct inode *inode;
>>>>> + � struct cifs_sb_info *cifs_sb;
>>>>> � � ssize_t written;
>>>>>
>>>>> - � written = generic_file_aio_write(iocb, iov, nr_segs, pos);
>>>>> - � if (!CIFS_I(inode)->clientCanCacheAll)
>>>>> - � � � � � filemap_fdatawrite(inode->i_mapping);
>>>>> + � inode = iocb->ki_filp->f_path.dentry->d_inode;
>>>>> +
>>>>> + � if (CIFS_I(inode)->clientCanCacheAll)
>>>>> + � � � � � return generic_file_aio_write(iocb, iov, nr_segs, pos);
>>>>> +
>>>>> + � cifs_sb = CIFS_SB(iocb->ki_filp->f_path.dentry->d_sb);
>>>>> +
>>>>> + � if ((cifs_sb->mnt_cifs_flags & CIFS_MOUNT_STRICT_IO) == 0) {
>>>>> + � � � � � int rc;
>>>>> +
>>>>> + � � � � � written = generic_file_aio_write(iocb, iov, nr_segs, pos);
>>>>> +
>>>>> + � � � � � rc = filemap_fdatawrite(inode->i_mapping);
>>>>> + � � � � � if (rc)
>>>>> + � � � � � � � � � cFYI(1, "cifs_file_aio_write: %d rc on %p inode",
>>>>> + � � � � � � � � � � � �rc, inode);
>>>>> + � � � � � return written;
>>>>> + � }
>>>>> +
>>>>> + � /* in strict cache mode we need to write the data to the server exactly
>>>>> + � � �from the pos to pos+len-1 rather than flush all affected pages
>>>>> + � � �because it may cause a error with mandatory locks on these pages but
>>>>> + � � �not on the region from pos to ppos+len-1 */
>>>>
>>>> � � � Again, please fix the comment style. Here: ^^^^
>>>>
>>>>> + � written = cifs_user_write(iocb->ki_filp, iov->iov_base,
>>>>> + � � � � � � � � � � � � � � iov->iov_len, &pos);
>>>>> +
>>>>> + � iocb->ki_pos = pos;
>>>>> +
>>>>> + � /* if we were successful - invalidate inode pages the write affected */
>>>>> + � if (written > 0)
>>>>> + � � � � � invalidate_mapping_pages(inode->i_mapping,
>>>>> + � � � � � � � � � � � � � � � � � � � � (pos-written) >> PAGE_CACHE_SHIFT,
>>>>> + � � � � � � � � � � � � � � � � � � � � (pos-1) >> PAGE_CACHE_SHIFT);
>>>>> +
>>>>> � � return written;
>>>>> �}
>>>>>
>>>>
>>>> May god have mercy on anyone who tries to mix strictcache and mmap.
>>>>
>>>> Reviewed-by: Jeff Layton <jlayton-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
>>>
>>> (cc'ing Suresh so he can comment)
>>>
>>> Actually...I'm going to withdraw my Reviewed-by tag here for now. This
>>> bare invalidate_mapping_pages doesn't deal with fscache.
>>>
>>> I think I need to understand what's intended when someone specifies
>>> strictcache and fsc before I can ack this. The simple answer would be
>>> that they are mutually exclusive, but if that's the case then the patch
>>> that adds the mount option needs to deal with that appropriately.
>>
>>
>> I don't think they can live together. I think we should do smth like a
>> following in mount options parsing:
>>
>> ...
>> if (opt == fscache) {
>>    vol->fscahe = 1;
>>    vol->strictcache = 0;
>> }
>> ...
>> if (opt == strictcache) {
>>   vol->strictcache = 1;
>>   vol->fscache = 0;
>> }
>>
>> So, if user specify both only the last will affect the client
>> behavior. Also we should add this information into cifs manpage.
>> Thoughts?
>>
> 
> That would one way to deal with it.
> 
> On the other hand though...fscache allows you to keep more data cached
> than you have RAM. This could be useful in a strictcache situation as
> well. Consider the case of an application on a client that has a lot of
> large files open for read. The server may grant oplocks on all of them.
> fscache would allow for fewer round trips to the server in such a case.
> 
> So, another way to deal with it would be to simply invalidate the
> fscache whenever you'd invalidate the in-ram cache. I'm not sure what
> to do for the cifs_file_aio_write case where you're invalidating just a
> small range however.
> 

Yes, we seem to have those two options while the later has the advantage
that Jeff mentioned. I think we could do the later without much of a
problme. We just need to retire the cookies and get new ones
(relinquishing with retire set to 1) i.e. by calling
cifs_fscache_reset_inode_cookie(). FS-Cache does not provide data
invalidation by itself. For the cifs_file_aio_write case too we could
set invalid_mapping and get fresh cookies.


Thanks,


-- 
Suresh Jayaraman