Re: [bisected commit 0fc9d10] NFS-server corruption with 3.4

[Konstantin Khlebnikov <khlebnikov@openvz.org>]

Hans de Bruin wrote:
> On 06/05/2012 06:39 PM, Konstantin Khlebnikov wrote:
>> OGAWA Hirofumi wrote:
>>> Konstantin Khlebnikov<khlebnikov@openvz.org>  writes:
>>>
>>>> Proper fix in attachment.
>>>
>>> Maybe, you are going to add the stable tag for next stable?
>>>
>>> Thanks.
>>
>> Yes, this definitely must be in next stable 3.4.x , but first I'll wait
>> for confirmation.
>>
>> Guys, who can reproduce this, please check patch "radix-tree: fix
>> contiguous iterator"
>> from my previous mail in this thread.
>>
>>>
>>>> radix-tree: fix contiguous iterator
>>>>
>>>> From: Konstantin Khlebnikov<khlebnikov@openvz.org>
>>>>
>>>> This patch fixes bug in macro radix_tree_for_each_contig().
>>>> If radix_tree_next_slot() sees NULL in next slot it returns NULL, but
>>>> following
>>>> radix_tree_next_chunk() switches iterating into next chunk. As result
>>>> iterating
>>>> becomes non-contiguous and breaks vfs "splice" and all its users.
>>>
>>
>
> I patched on to off v3.4 and Firefox an Thunderbird do not segfault
> anymore. The do not start either. This was de feature on my 'production'
> server I could not reproduce on my test server. Maybe it has something
> to with the different type of nic's.
>
> For the second attempt I branched of at 0fc9d1040313047edf6a39fd and
> applied your patch on top of it. Firefox an Thunderbird where back
> again. So your patch works.
>
> Now I need some git-instructions. Apparently something else is broken. I
> branched of with:
>
> git branch debug 0fc9d1040313047edf6a3
> git checkout debug
> applied the patch.
> git commit -a (got commit 5c09c685ba2d36c3b905220d43ad1b47354e456eed back)
>
> Now I want all commits in my master branch up until ref v3.4 added to my
> debug branch so I can bisect between 5c09 and v3.4
>

No, it does not work in this way. you should apply patch before each test.
You can fix radix-tree bug in find_get_pages_contig() to avoid changing
radix-tree.h and rebuild kernel faster.

--- a/mm/filemap.c
+++ b/mm/filemap.c
@@ -927,7 +927,7 @@ repeat:
                  * otherwise we can get both false positives and false
                  * negatives, which is just confusing to the caller.
                  */
-               if (page->mapping == NULL || page->index != iter.index) {
+               if (page->mapping == NULL || page->index != index + ret) {
                         page_cache_release(page);
                         break;
                 }