Re: [PATCH v2 2/3] xen/privcmd: Mark pages as dirty

[John Hubbard <jhubbard@nvidia.com>]

On 2020-07-07 04:43, Jürgen Groß wrote:
> On 07.07.20 13:30, Souptick Joarder wrote:
>> On Tue, Jul 7, 2020 at 3:08 PM Jürgen Groß <jgross@suse.com> wrote:
...
>>>> diff --git a/drivers/xen/privcmd.c b/drivers/xen/privcmd.c
>>>> index 33677ea..f6c1543 100644
>>>> --- a/drivers/xen/privcmd.c
>>>> +++ b/drivers/xen/privcmd.c
>>>> @@ -612,8 +612,11 @@ static void unlock_pages(struct page *pages[], unsigned int nr_pages)
>>>>    {
>>>>        unsigned int i;
>>>>
>>>> -     for (i = 0; i < nr_pages; i++)
>>>> +     for (i = 0; i < nr_pages; i++) {
>>>> +             if (!PageDirty(pages[i]))
>>>> +                     set_page_dirty_lock(pages[i]);
>>>
>>> With put_page() directly following I think you should be able to use
>>> set_page_dirty() instead, as there is obviously a reference to the page
>>> existing.
>>
>> Patch [3/3] will convert above codes to use unpin_user_pages_dirty_lock()
>> which internally do the same check. So I thought to keep linux-stable and
>> linux-next code in sync. John had a similar concern [1] and later agreed to keep
>> this check.
>>
>> Shall I keep this check ?  No ?

It doesn't matter *too* much, because patch 3/3 fixes up everything by
changing it all to unpin_user_pages_dirty_lock(). However, there is something
to be said for having correct interim patches, too. :)  Details:

>>
>> [1] https://lore.kernel.org/xen-devel/a750e5e5-fd5d-663b-c5fd-261d7c939ba7@nvidia.com/
> 
> I wasn't referring to checking PageDirty(), but to the use of
> set_page_dirty_lock().
> 
> Looking at the comment just before the implementation of
> set_page_dirty_lock() suggests that it is fine to use set_page_dirty()
> instead (so not calling lock_page()).


no no, that's a misreading of the comment. Unless this xen/privcmd code has
somehow taken a reference on page->mapping->host (which I do *not* think is
the case), then it is still racy to call set_page_dirty() here. Instead,
set_page_dirty_lock() should be used.


> 
> Only the transition from get_user_pages_fast() to pin_user_pages_fast()
> requires to use the locked version IMO.
> 

That's a different misunderstanding. :) pin_user_pages*() APIs are meant to be
functionally drop-in replacements for get_user_pages*(). Internally,
pin_user_pages*() functions do some additional tracking, but from a caller's
perspective, it should look the same. In other words, there is nothing
about pin_user_pages_fast() that requires set_page_dirty_lock() upon release.
The reason set_page_dirty_lock() was chosen is that there are very few
(none at all?) call sites that need to release and dirty a page, that also meet
the requirements to safely call set_page_dirty().

That's why there is a unpin_user_pages_dirty_lock(), but there is not a
corresponding unpin_user_pages_dirty() call: the latter has not been required
so far, even though the call site conversions are nearly done.


thanks,
-- 
John Hubbard
NVIDIA