lkml.org 
[lkml]   [2018]   [Dec]   [18]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [PATCH 1/2] mm: introduce put_user_page*(), placeholder versions
From
Date
On 12/18/18 1:30 AM, Jan Kara wrote:
> On Mon 17-12-18 10:34:43, Matthew Wilcox wrote:
>> On Mon, Dec 17, 2018 at 01:11:50PM -0500, Jerome Glisse wrote:
>>> On Mon, Dec 17, 2018 at 08:58:19AM +1100, Dave Chinner wrote:
>>>> Sure, that's a possibility, but that doesn't close off any race
>>>> conditions because there can be DMA into the page in progress while
>>>> the page is being bounced, right? AFAICT this ext3+DIF/DIX case is
>>>> different in that there is no 3rd-party access to the page while it
>>>> is under IO (ext3 arbitrates all access to it's metadata), and so
>>>> nothing can actually race for modification of the page between
>>>> submission and bouncing at the block layer.
>>>>
>>>> In this case, the moment the page is unlocked, anyone else can map
>>>> it and start (R)DMA on it, and that can happen before the bio is
>>>> bounced by the block layer. So AFAICT, block layer bouncing doesn't
>>>> solve the problem of racing writeback and DMA direct to the page we
>>>> are doing IO on. Yes, it reduces the race window substantially, but
>>>> it doesn't get rid of it.
>>>
>>> So the event flow is:
>>> - userspace create object that match a range of virtual address
>>> against a given kernel sub-system (let's say infiniband) and
>>> let's assume that the range is an mmap() of a regular file
>>> - device driver do GUP on the range (let's assume it is a write
>>> GUP) so if the page is not already map with write permission
>>> in the page table than a page fault is trigger and page_mkwrite
>>> happens
>>> - Once GUP return the page to the device driver and once the
>>> device driver as updated the hardware states to allow access
>>> to this page then from that point on hardware can write to the
>>> page at _any_ time, it is fully disconnected from any fs event
>>> like write back, it fully ignore things like page_mkclean
>>>
>>> This is how it is to day, we allowed people to push upstream such
>>> users of GUP. This is a fact we have to live with, we can not stop
>>> hardware access to the page, we can not force the hardware to follow
>>> page_mkclean and force a page_mkwrite once write back ends. This is
>>> the situation we are inheriting (and i am personnaly not happy with
>>> that).
>>>
>>> >From my point of view we are left with 2 choices:
>>> [C1] break all drivers that do not abide by the page_mkclean and
>>> page_mkwrite
>>> [C2] mitigate as much as possible the issue
>>>
>>> For [C2] the idea is to keep track of GUP per page so we know if we
>>> can expect the page to be written to at any time. Here is the event
>>> flow:
>>> - driver GUP the page and program the hardware, page is mark as
>>> GUPed
>>> ...
>>> - write back kicks in on the dirty page, lock the page and every
>>> thing as usual , sees it is GUPed and inform the block layer to
>>> use a bounce page
>>
>> No. The solution John, Dan & I have been looking at is to take the
>> dirty page off the LRU while it is pinned by GUP. It will never be
>> found for writeback.
>>
>> That's not the end of the story though. Other parts of the kernel (eg
>> msync) also need to be taught to stay away from pages which are pinned
>> by GUP. But the idea is that no page gets written back to storage while
>> it's pinned by GUP. Only when the last GUP ends is the page returned
>> to the list of dirty pages.
>
> We've been through this in:
>
> https://lore.kernel.org/lkml/20180709194740.rymbt2fzohbdmpye@quack2.suse.cz/
>
> back in July. You cannot just skip pages for fsync(2). So as I wrote above -
> memory cleaning writeback can skip pinned pages. Data integrity writeback
> must be able to write pinned pages. And bouncing is one reasonable way how
> to do that.
>
> This writeback decision is pretty much independent from the mechanism by
> which we are going to identify pinned pages. Whether that's going to be
> separate counter in struct page, using page->_mapcount, or separately
> allocated data structure as you know promote.
>
> I currently like the most the _mapcount suggestion from Jerome but I'm not
> really attached to any solution as long as it performs reasonably and
> someone can make it working :) as I don't have time to implement it at
> least till January.
>

OK, so let's take another look at Jerome's _mapcount idea all by itself (using
*only* the tracking pinned pages aspect), given that it is the lightest weight
solution for that.

So as I understand it, this would use page->_mapcount to store both the real
mapcount, and the dma pinned count (simply added together), but only do so for
file-backed (non-anonymous) pages:


__get_user_pages()
{
...
get_page(page);

if (!PageAnon)
atomic_inc(page->_mapcount);
...
}

put_user_page(struct page *page)
{
...
if (!PageAnon)
atomic_dec(&page->_mapcount);

put_page(page);
...
}

...and then in the various consumers of the DMA pinned count, we use page_mapped(page)
to see if any mapcount remains, and if so, we treat it as DMA pinned. Is that what you
had in mind?


--
thanks,
John Hubbard
NVIDIA

\
 
 \ /
  Last update: 2018-12-19 00:30    [W:0.243 / U:0.760 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site