On 7/23/19 11:17 PM, Christoph Hellwig wrote:
> On Tue, Jul 23, 2019 at 09:25:06PM -0700, john.hubbard@xxxxxxxxx wrote:
>> * Store, in the iov_iter, a "came from gup (get_user_pages)" parameter.
>> Then, use the new iov_iter_get_pages_use_gup() to retrieve it when
>> it is time to release the pages. That allows choosing between put_page()
>> and put_user_page*().
>>
>> * Pass in one more piece of information to bio_release_pages: a "from_gup"
>> parameter. Similar use as above.
>>
>> * Change the block layer, and several file systems, to use
>> put_user_page*().
>
> I think we can do this in a simple and better way. We have 5 ITER_*
> types. Of those ITER_DISCARD as the name suggests never uses pages, so
> we can skip handling it. ITER_PIPE is rejected іn the direct I/O path,
> which leaves us with three.
>
> Out of those ITER_BVEC needs a user page reference, so we want to call
^ ITER_IOVEC, I hope. Otherwise I'm hopeless lost. :)
> put_user_page* on it. ITER_BVEC always already has page reference,
> which means in the block direct I/O path path we alread don't take
> a page reference. We should extent that handling to all other calls
> of iov_iter_get_pages / iov_iter_get_pages_alloc. I think we should
> just reject ITER_KVEC for direct I/O as well as we have no users and
> it is rather pointless. Alternatively if we see a use for it the
> callers should always have a life page reference anyway (or might
> be on kmalloc memory), so we really should not take a reference either.
>
> In other words: the only time we should ever have to put a page in
> this patch is when they are user pages. We'll need to clean up
> various bits of code for that, but that can be done gradually before
> even getting to the actual put_user_pages conversion.
>
Sounds great. I'm part way into it and it doesn't look too bad. The main
question is where to scatter various checks and assertions, to keep
the kvecs out of direct I/0. Or at least keep the gups away from
direct I/0.
thanks,
--
John Hubbard
NVIDIA
