From mboxrd@z Thu Jan 1 00:00:00 1970 From: Junxiao Bi Date: Fri, 19 Jul 2013 17:16:50 +0800 Subject: [Ocfs2-devel] [PATCH] ocfs2: using i_size_read() to access i_size In-Reply-To: <51E8F93B.6050406@oracle.com> References: <1373876172-6467-1-git-send-email-junxiao.bi@oracle.com> <51E4B86F.2040703@tao.ma> <51E5D390.6040604@oracle.com> <51E8ECF3.7010905@oracle.com> <51E8F2E0.6000207@oracle.com> <51E8F4F2.7010205@oracle.com> <51E8F93B.6050406@oracle.com> Message-ID: <51E90402.8080409@oracle.com> List-Id: MIME-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit To: ocfs2-devel@oss.oracle.com Hi Jeff, On 07/19/2013 04:30 PM, Jeff Liu wrote: > On 07/19/2013 04:12 PM, Junxiao Bi wrote: > >> Hi Jeff, >> >> On 07/19/2013 04:03 PM, Jeff Liu wrote: >>> Hi Junxiao, >>> >>> On 07/19/2013 03:38 PM, Junxiao Bi wrote: >>> >>>> Hi Jeff, >>>> >>>> On 07/17/2013 07:13 AM, Jeff Liu wrote: >>>>>>> diff --git a/fs/ocfs2/file.c b/fs/ocfs2/file.c >>>>>>>>> index 3ce6a8b..7158710 100644 >>>>>>>>> --- a/fs/ocfs2/file.c >>>>>>>>> +++ b/fs/ocfs2/file.c >>>>>>>>> @@ -2650,7 +2650,7 @@ static loff_t ocfs2_file_llseek(struct file *file, loff_t offset, int whence) >>>>> We have the inode mutex around ocfs2_file_llseek(), and that is too extensive >>>>> to block the concurrent access for particular seek operations. At least, >>>>> we can get rid of this lock for SEEK_SET/SEEK_CUR. i.e, In either case, >>>>> we can fall through generic_file_llseek() directly without the mutex lock. >>>> Think about this again, I think we can't remove the mutex, as it is used >>>> to protect other inode member beside i_size, like >>>> inode->i_sb->s_maxbytes which is accessed in all SEEK request and also >>>> more in ocfs2_inode_lock(). >>> Thanks for digging into this. s_maxbytes is a numeric constant which is >>> calculated out at OCFS2 mount with filling up the super blocks, IOWs, it >>> does not need any lock protection IMO. >>> >>> Besides that, could you please pointed me out anything am missed? >> Can inode->i_sb be set to NULL during this? > No, actually when we get into ocfs2_file_llseek(), the corresponding > inode should be in VFS inode cache and it should not be a bad inode > as well, otherwise, it can not go through vfs_llseek(). :) Thank you for explaining this. So for SEEK_SET and SEEK_CUR, we may make some improvement, indeed in gereric_file_llseek(), it also use a spin lock to protect concurrent SEEK_CUR access. is there really a perf issue that deserve us to make a hack to hurt code readability? Thanks, Junxiao. > > Thanks, > -Jeff > >> Thanks, >> Junxiao. >>> Thanks, >>> -Jeff >>> >>>> Thanks, >>>> Junxiao. >>>>>>>>> case SEEK_SET: >>>>>>>>> break; >>>>>>>>> case SEEK_END: >>>>>>>>> - offset += inode->i_size; >>>>>>>>> + offset += i_size_read(inode); >