Re: delayed allocatiou result in Oops

[Alex Tomas <alex@clusterfs.com>]

Mingming Cao wrote:
>>From the comments it says the page->private is set to 1 to letting
> commit_write know that it needs block reservation, but I don't see the
> page->private value being checked in ext4_wb_commit_write(). Instead,
> the PageMappedToDisk(page) flag is being checked.
> 
> Alex, can you clarify the use of page->private and PagePrivate flag
> here? Do we still need the page->private for delayed allocation, with
> PageBooked flag and PageMappedToDisk page flag?

sorry for confusion, we need PagePrivate so that truncate calls our
->invalidatepage(). in turn this call is needed to drop unused reservation.

block_read_full_page() must not happen to pages being delayed allocated
in the first place - it's uptodate by definition. I think the problem
began when ext4_wb_commit_write() exited due to -ENOSPC but left not
uptodate *and* with PG_private. then subsequent access to page turned to
block_read_full_page() which relies on PG_private and meaningful private
field. Dmitry, could you repeat the test with SetPagePrivate(page) moved
to after that if() with ext4_wb_reserve_space_page(), please?


as i'm here ... status update: I've been reworking delayed allocation
patches to support blocksize < PAGE_CACHE_SIZE (and address akpm's request
for more generic implementation). the patch isn't ready for review, but
hopefully will be in few days.


thanks, Alex