lkml.org 
[lkml]   [2013]   [Nov]   [21]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Date
SubjectRe: [PATCH] mm/zswap: remove unneeded zswap_rb_erase calls
On Wed, Nov 20, 2013 at 9:52 PM, Weijie Yang <weijie.yang.kh@gmail.com> wrote:
> Hello Dan
>
> On Thu, Nov 21, 2013 at 3:47 AM, Dan Streetman <ddstreet@ieee.org> wrote:
>> Since zswap_rb_erase was added to the final (when refcount == 0)
>> zswap_put_entry, there is no need to call zswap_rb_erase before
>> calling zswap_put_entry.
>>
>> Signed-off-by: Dan Streetman <ddstreet@ieee.org>
>> ---
>> mm/zswap.c | 5 -----
>> 1 file changed, 5 deletions(-)
>>
>> diff --git a/mm/zswap.c b/mm/zswap.c
>> index e154f1e..f4fbbd5 100644
>> --- a/mm/zswap.c
>> +++ b/mm/zswap.c
>> @@ -711,8 +711,6 @@ static int zswap_frontswap_store(unsigned type, pgoff_t offset,
>> ret = zswap_rb_insert(&tree->rbroot, entry, &dupentry);
>> if (ret == -EEXIST) {
>> zswap_duplicate_entry++;
>> - /* remove from rbtree */
>> - zswap_rb_erase(&tree->rbroot, dupentry);
>> zswap_entry_put(tree, dupentry);
>> }
>> } while (ret == -EEXIST);
>
> If remove zswap_rb_erase, it would loop until free this dupentry. This
> would cause 2 proplems:

I need to get more familiar with when it's possible to hit a duplicate
entry, it seems strange to me that higher level swap code would be
trying to store a page with an already used offset.

> 1. zswap_duplicate_entry counter is not correct
> 2. trigger BUG_ON in zswap_entry_put when this dupentry is being writeback,
> because zswap_writeback_entry will call zswap_entry_put either.
>
> So, I don't think it is a good idea to remove zswap_rb_erase call.
>
>> @@ -787,9 +785,6 @@ static void zswap_frontswap_invalidate_page(unsigned type, pgoff_t offset)
>> return;
>> }
>>
>> - /* remove from rbtree */
>> - zswap_rb_erase(&tree->rbroot, entry);
>> -
>> /* drop the initial reference from entry creation */
>> zswap_entry_put(tree, entry);
>
> I think it is better not to remove the zswap_rb_erase call.
>
> From frontswap interface view, if invalidate is called, the page(and
> entry) should never visible to upper.
> If remove the zswap_rb_erase call, it is not fit this semantic.
>
> Consider the following scenario:
> 1. thread 0: entry A is being writeback
> 2. thread 1: invalidate entry A, as refcount != 0, it will still exist
> on rbtree.
> 3. thread 1: reuse entry A 's swp_entry_t, do a frontswap_store
> it will conflict with the entry A on the rbtree, it is not a
> normal duplicate store.
>
> If we place the zswap_rb_erase call in zswap_frontswap_invalidate_page,
> we can avoid the above scenario.
>
> So, I don't think it is a good idea to remove zswap_rb_erase call.

It seems to me that zswap_rb_erase shouldn't have been folded into
zswap_entry_put; if it was removed now, the only place it would need
to be added back is into the success path of writeback, i.e.:

if (entry == zswap_rb_search(&tree->rbroot, offset)) {
zswap_rb_erase(&tree->rbroot, entry);
zswap_entry_put(tree, entry);
}


\
 
 \ /
  Last update: 2013-11-22 00:01    [W:0.099 / U:0.304 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site