Re: [PATCH 1/2] mm: introduce put_user_page*(), placeholder versions

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On 12/4/18 4:40 PM, Dan Williams wrote:
> On Tue, Dec 4, 2018 at 4:37 PM Jerome Glisse <jglisse@xxxxxxxxxx> wrote:
>>
>> On Tue, Dec 04, 2018 at 03:03:02PM -0800, Dan Williams wrote:
>>> On Tue, Dec 4, 2018 at 1:56 PM John Hubbard <jhubbard@xxxxxxxxxx> wrote:
>>>>
>>>> On 12/4/18 12:28 PM, Dan Williams wrote:
>>>>> On Mon, Dec 3, 2018 at 4:17 PM <john.hubbard@xxxxxxxxx> wrote:
>>>>>>
>>>>>> From: John Hubbard <jhubbard@xxxxxxxxxx>
>>>>>>
>>>>>> Introduces put_user_page(), which simply calls put_page().
>>>>>> This provides a way to update all get_user_pages*() callers,
>>>>>> so that they call put_user_page(), instead of put_page().
>>>>>>
>>>>>> Also introduces put_user_pages(), and a few dirty/locked variations,
>>>>>> as a replacement for release_pages(), and also as a replacement
>>>>>> for open-coded loops that release multiple pages.
>>>>>> These may be used for subsequent performance improvements,
>>>>>> via batching of pages to be released.
>>>>>>
>>>>>> This is the first step of fixing the problem described in [1]. The steps
>>>>>> are:
>>>>>>
>>>>>> 1) (This patch): provide put_user_page*() routines, intended to be used
>>>>>>    for releasing pages that were pinned via get_user_pages*().
>>>>>>
>>>>>> 2) Convert all of the call sites for get_user_pages*(), to
>>>>>>    invoke put_user_page*(), instead of put_page(). This involves dozens of
>>>>>>    call sites, and will take some time.
>>>>>>
>>>>>> 3) After (2) is complete, use get_user_pages*() and put_user_page*() to
>>>>>>    implement tracking of these pages. This tracking will be separate from
>>>>>>    the existing struct page refcounting.
>>>>>>
>>>>>> 4) Use the tracking and identification of these pages, to implement
>>>>>>    special handling (especially in writeback paths) when the pages are
>>>>>>    backed by a filesystem. Again, [1] provides details as to why that is
>>>>>>    desirable.
>>>>>
>>>>> I thought at Plumbers we talked about using a page bit to tag pages
>>>>> that have had their reference count elevated by get_user_pages()? That
>>>>> way there is no need to distinguish put_page() from put_user_page() it
>>>>> just happens internally to put_page(). At the conference Matthew was
>>>>> offering to free up a page bit for this purpose.
>>>>>
>>>>
>>>> ...but then, upon further discussion in that same session, we realized that
>>>> that doesn't help. You need a reference count. Otherwise a random put_page
>>>> could affect your dma-pinned pages, etc, etc.
>>>
>>> Ok, sorry, I mis-remembered. So, you're effectively trying to capture
>>> the end of the page pin event separate from the final 'put' of the
>>> page? Makes sense.
>>>
>>>> I was not able to actually find any place where a single additional page
>>>> bit would help our situation, which is why this still uses LRU fields for
>>>> both the two bits required (the RFC [1] still applies), and the dma_pinned_count.
>>>
>>> Except the LRU fields are already in use for ZONE_DEVICE pages... how
>>> does this proposal interact with those?
>>>
>>>> [1] https://lore.kernel.org/r/20181110085041.10071-7-jhubbard@xxxxxxxxxx
>>>>
>>>>>> [1] https://lwn.net/Articles/753027/ : "The Trouble with get_user_pages()"
>>>>>>
>>>>>> Reviewed-by: Jan Kara <jack@xxxxxxx>
>>>>>
>>>>> Wish, you could have been there Jan. I'm missing why it's safe to
>>>>> assume that a single put_user_page() is paired with a get_user_page()?
>>>>>
>>>>
>>>> A put_user_page() per page, or a put_user_pages() for an array of pages. See
>>>> patch 0002 for several examples.
>>>
>>> Yes, however I was more concerned about validation and trying to
>>> locate missed places where put_page() is used instead of
>>> put_user_page().
>>>
>>> It would be interesting to see if we could have a debug mode where
>>> get_user_pages() returned dynamically allocated pages from a known
>>> address range and catch drivers that operate on a user-pinned page
>>> without using the proper helper to 'put' it. I think we might also
>>> need a ref_user_page() for drivers that may do their own get_page()
>>> and expect the dma_pinned_count to also increase.

Good idea about a new ref_user_page() call. It's going to hard to find 
those places at all of the call sites, btw.

>>
>> Total crazy idea for this, but this is the right time of day
>> for this (for me at least it is beer time :)) What about mapping
>> all struct page in two different range of kernel virtual address
>> and when get user space is use it returns a pointer from the second
>> range of kernel virtual address to the struct page. Then in put_page
>> you know for sure if the code putting the page got it from GUP or
>> from somewhere else. page_to_pfn() would need some trickery to
>> handle that.
> 
> Yes, exactly what I was thinking, if only as a debug mode since
> instrumenting every pfn/page translation would be expensive.
> 

That does sound viable as a debug mode. I'll try it out. A reliable way
(in both directions) of sorting out put_page() vs. put_user_page() 
would be a huge improvement, even if just in debug mode.

>> Dunno if we are running out of kernel virtual address (outside
>> 32bits that i believe we are trying to shot down quietly behind
>> the bar).
> 
> There's room, KASAN is in a roughly similar place.
> 

Looks like I'd better post a new version of the entire RFC, rather than just
these two patches. It's still less fully-baked than I'd hoped. :)

thanks,
-- 
John Hubbard
NVIDIA



[Index of Archives]     [Linux Ext4 Filesystem]     [Union Filesystem]     [Filesystem Testing]     [Ceph Users]     [Ecryptfs]     [AutoFS]     [Kernel Newbies]     [Share Photos]     [Security]     [Netfilter]     [Bugtraq]     [Yosemite News]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Linux Cachefs]     [Reiser Filesystem]     [Linux RAID]     [Samba]     [Device Mapper]     [CEPH Development]

  Powered by Linux