linux-ext4.vger.kernel.org archive mirror
 help / color / mirror / Atom feed
From: Zheng Liu <gnehzuil.liu@gmail.com>
To: Dan Carpenter <dan.carpenter@oracle.com>
Cc: wenqing.lz@taobao.com, linux-ext4@vger.kernel.org, kbuild@01.org,
	Theodore Ts'o <tytso@mit.edu>
Subject: Re: ext4: fold ext4_generic_write_end() into ext4_write_end()
Date: Tue, 2 Apr 2013 19:09:18 +0800	[thread overview]
Message-ID: <20130402110918.GA10495@gmail.com> (raw)
In-Reply-To: <20130402081144.GA11441@longonot.mountain>

[Add Ted into cc list]

Hi Dan,

Thanks for pointing it out.

On Tue, Apr 02, 2013 at 11:11:44AM +0300, Dan Carpenter wrote:
> Hello Zheng Liu,
> 
> The patch 31e4045fda81: "ext4: fold ext4_generic_write_end() into 
> ext4_write_end()" from Mar 26, 2013, leads to the following warning:
> "fs/ext4/inode.c:1169 ext4_write_end()
> 	 warn: unsigned 'copied' is never less than zero."
> 
> 
>   1110  static int ext4_write_end(struct file *file,
>   1111                            struct address_space *mapping,
>   1112                            loff_t pos, unsigned len, unsigned copied,
> 
> copied is unsigned.
> 
>   1113                            struct page *page, void *fsdata)
>   1114  {
>   1115          handle_t *handle = ext4_journal_current_handle();
>   1116          struct inode *inode = mapping->host;
>   1117          int ret = 0, ret2;
>   1118          int i_size_changed = 0;
>   1119  
>   1120          trace_ext4_write_end(inode, pos, len, copied);
>   1121          if (ext4_test_inode_state(inode, EXT4_STATE_ORDERED_MODE)) {
>   1122                  ret = ext4_jbd2_file_inode(handle, inode);
>   1123                  if (ret) {
>   1124                          unlock_page(page);
>   1125                          page_cache_release(page);
>   1126                          goto errout;
>   1127                  }
>   1128          }
>   1129  
>   1130          if (ext4_has_inline_data(inode))
>   1131                  copied = ext4_write_inline_data_end(inode, pos, len,
>   1132                                                      copied, page);
>   1133          else
>   1134                  copied = block_write_end(file, mapping, pos,
>   1135                                           len, copied, page, fsdata);
> 
> I don't think these functions return negative error codes.

Yes, I check these two functions and they don't return a negative error
code.

> 
>   1136  
>   1137          /*
>   1138           * No need to use i_size_read() here, the i_size
>   1139           * cannot change under us because we hole i_mutex.
>   1140           *
>   1141           * But it's important to update i_size while still holding page lock:
>   1142           * page writeout could otherwise come in and zero beyond i_size.
>   1143           */
>   1144          if (pos + copied > inode->i_size) {
>   1145                  i_size_write(inode, pos + copied);
>   1146                  i_size_changed = 1;
>   1147          }
>   1148  
>   1149          if (pos + copied > EXT4_I(inode)->i_disksize) {
>   1150                  /* We need to mark inode dirty even if
>   1151                   * new_i_size is less that inode->i_size
>   1152                   * but greater than i_disksize. (hint delalloc)
>   1153                   */
>   1154                  ext4_update_i_disksize(inode, (pos + copied));
>   1155                  i_size_changed = 1;
>   1156          }
>   1157          unlock_page(page);
>   1158          page_cache_release(page);
>   1159  
>   1160          /*
>   1161           * Don't mark the inode dirty under page lock. First, it unnecessarily
>   1162           * makes the holding time of page lock longer. Second, it forces lock
>   1163           * ordering of page lock and transaction start for journaling
>   1164           * filesystems.
>   1165           */
>   1166          if (i_size_changed)
>   1167                  ext4_mark_inode_dirty(handle, inode);
>   1168  
>   1169          if (copied < 0)
>   1170                  ret = copied;
> 
> copied is never less than zero because it's unsigned.

Yes, it should be removed.

Ted, please remove this check.

Thanks,
                                                - Zheng

      reply	other threads:[~2013-04-02 10:53 UTC|newest]

Thread overview: 2+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2013-04-02  8:11 ext4: fold ext4_generic_write_end() into ext4_write_end() Dan Carpenter
2013-04-02 11:09 ` Zheng Liu [this message]

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=20130402110918.GA10495@gmail.com \
    --to=gnehzuil.liu@gmail.com \
    --cc=dan.carpenter@oracle.com \
    --cc=kbuild@01.org \
    --cc=linux-ext4@vger.kernel.org \
    --cc=tytso@mit.edu \
    --cc=wenqing.lz@taobao.com \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).