lkml.org 
[lkml]   [2012]   [Aug]   [23]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
SubjectRe: [PATCH 3/3] HWPOISON: prevent inode cache removal to keep AS_HWPOISON sticky
Date
On Thu, Aug 23, 2012 at 05:11:25PM +0800, Fengguang Wu wrote:
> On Wed, Aug 22, 2012 at 11:17:35AM -0400, Naoya Horiguchi wrote:
> > "HWPOISON: report sticky EIO for poisoned file" still has a corner case
> > where we have possibilities of data lost. This is because in this fix
> > AS_HWPOISON is cleared when the inode cache is dropped.
> >
> > For example, consider an application in which a process periodically
> > (every 10 minutes) writes some logs on a file (and closes it after
> > each writes,) and at the end of each day some batch programs run using
> > the log file. If a memory error hits on dirty pagecache of this log file
> > just after periodic write/close and the inode cache is cleared before the
> > next write, then this application is not aware of the error and the batch
> > programs will work wrongly.
> >
> > To avoid this, this patch makes us pin the hwpoisoned inode on memory
> > until we remove or completely truncate the hwpoisoned file.
>
> Good point!
>
> > Signed-off-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
> > ---
> > fs/inode.c | 12 ++++++++++++
> > include/linux/pagemap.h | 11 +++++++++++
> > mm/memory-failure.c | 2 +-
> > mm/truncate.c | 2 ++
> > 4 files changed, 26 insertions(+), 1 deletion(-)
> >
> > diff --git v3.6-rc1.orig/fs/inode.c v3.6-rc1/fs/inode.c
> > index ac8d904..8742397 100644
> > --- v3.6-rc1.orig/fs/inode.c
> > +++ v3.6-rc1/fs/inode.c
> > @@ -717,6 +717,15 @@ void prune_icache_sb(struct super_block *sb, int nr_to_scan)
> > }
> >
> > /*
> > + * Keep inode caches on memory for user processes to certainly
> > + * be aware of memory errors.
> > + */
> > + if (unlikely(mapping_hwpoison(inode->i_mapping))) {
> > + spin_unlock(&inode->i_lock);
> > + continue;
> > + }
>
> That chunk prevents reclaiming all the cached pages. However the intention
> is only to keep the struct inode together with the hwpoison bit?

Yes, we can not reclaim pagecaches from shrink_slab(), but we can do from
shrink_zone(). So it shouldn't happen that cached pages on hwpoisoned file
remain for long under high memory pressure.

> > + /*
> > * Referenced or dirty inodes are still in use. Give them
> > * another pass through the LRU as we canot reclaim them now.
> > */
> > @@ -1405,6 +1414,9 @@ static void iput_final(struct inode *inode)
> > inode->i_state &= ~I_WILL_FREE;
> > }
> >
> > + if (unlikely(mapping_hwpoison(inode->i_mapping) && drop))
> > + mapping_clear_hwpoison(inode->i_mapping);
>
> Is that clear necessary? Because the bit will be gone with the inode
> struct: it's going to be de-allocated anyway.

With the chunk in prune_icache_sb() we keep the inode struct with
AS_HWPOISON set on memory, so in order to remove it, we need explicitly
clear the bit.
Without this clear, the inode remains until system reboot.

> > inode->i_state |= I_FREEING;
> > if (!list_empty(&inode->i_lru))
> > inode_lru_list_del(inode);
> > diff --git v3.6-rc1.orig/include/linux/pagemap.h v3.6-rc1/include/linux/pagemap.h
> > index 4d8d821..9fce4e4 100644
> > --- v3.6-rc1.orig/include/linux/pagemap.h
> > +++ v3.6-rc1/include/linux/pagemap.h
> > @@ -59,11 +59,22 @@ static inline int mapping_hwpoison(struct address_space *mapping)
> > {
> > return test_bit(AS_HWPOISON, &mapping->flags);
> > }
> > +static inline void mapping_set_hwpoison(struct address_space *mapping)
> > +{
> > + set_bit(AS_HWPOISON, &mapping->flags);
> > +}
> > +static inline void mapping_clear_hwpoison(struct address_space *mapping)
> > +{
> > + clear_bit(AS_HWPOISON, &mapping->flags);
> > +}
> > #else
> > static inline int mapping_hwpoison(struct address_space *mapping)
> > {
> > return 0;
> > }
> > +static inline void mapping_clear_hwpoison(struct address_space *mapping)
> > +{
> > +}
> > #endif
> >
> > static inline gfp_t mapping_gfp_mask(struct address_space * mapping)
> > diff --git v3.6-rc1.orig/mm/memory-failure.c v3.6-rc1/mm/memory-failure.c
> > index a1e7e00..ca064c6 100644
> > --- v3.6-rc1.orig/mm/memory-failure.c
> > +++ v3.6-rc1/mm/memory-failure.c
> > @@ -652,7 +652,7 @@ static int me_pagecache_dirty(struct page *p, unsigned long pfn)
> > * the first EIO, but we're not worse than other parts
> > * of the kernel.
> > */
> > - set_bit(AS_HWPOISON, &mapping->flags);
> > + mapping_set_hwpoison(mapping);
> > }
> >
> > return me_pagecache_clean(p, pfn);
> > diff --git v3.6-rc1.orig/mm/truncate.c v3.6-rc1/mm/truncate.c
> > index 75801ac..82a994f 100644
> > --- v3.6-rc1.orig/mm/truncate.c
> > +++ v3.6-rc1/mm/truncate.c
> > @@ -574,6 +574,8 @@ void truncate_setsize(struct inode *inode, loff_t newsize)
> >
> > oldsize = inode->i_size;
> > i_size_write(inode, newsize);
> > + if (unlikely(mapping_hwpoison(inode->i_mapping) && !newsize))
>
> It might be a bit better to test !newsize first.

Ah, OK.

Thanks,
Naoya


\
 
 \ /
  Last update: 2012-08-23 23:21    [W:0.144 / U:0.012 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site