lkml.org 
[lkml]   [2007]   [Dec]   [12]   [last100]   RSS Feed
Views: [more markup]  [less markup]  [headers]  [forward] 
 
Messages in this thread
/
DateWed, 12 Dec 2007 16:51:24 +0900
FromHisashi Hifumi <>
SubjectRe: [PATCH] dio: falling through to buffered I/O when invalidationof a page fails
 >
 >>
 >> Past discussion about this issue is as follows.
 >> http://marc.info/?t=119343431200004&r=1&w=2
 >> http://marc.info/?t=112656762800002&r=1&w=2
 >>
 >> 2, invalidate_inode_pages2_range() sets ret=-EIO when
 >> invalidate_complete_page2()
 >> fails, but this ret is cleared if do_launder_page() succeed on a page of
 >> next index.
 >
 >Oops.  That's too bad.  So maybe we should fix it by not stomping on
 >that return code?
 >
 >	ret2 = do_launder()
 >	if (ret2 == 0)
 >		ret2 = invalidate()
 >	if (ret == 0)
 >		ret = ret2
 >
 >I'd be surprised if we ever wanted to mask an -EIO when later pages
 >laundered successfully.

This can preserve ret of -EIO. But it cannot be distinguished between
the error of invalidation and other error only by this fix.


 >> I solved problems above by introducing invalidate_inode_pages3_range()
 >> and falling
 >> through to buffered I/O when invalidation of a page failed.
 >
 >Well, I like the idea of more intelligently dealing with the known
 >problem between dio and ext3.  I'm not sure that falling back to
 >buffered is right.

In the past discussion above, it was said that dio was a sort of best
effort function. So dio falls back to buffered IO when writing to hall, I think.
My idea is that this notion applies to this issue.

 >
 >> We can distinguish between failure of page invalidation and other errors
 >> with the return value of invalidate_inode_pages3_range().
 >
 >I'm not sure duplicating the invalidation loop into a new function is
 >the right thing.  Maybe we'd just tweak inode_pages2 to indicate to the
 >caller the specific failing circumstances somehow.  Maybe.

It will be better that adding one parameter to invalidate_inode_pages2_range so
that caller can deal with a number of cases.



\
 
 \ /
  Last update: 2007-12-12 08:55    [from the cache]
©2003-2008