Re: get_user_pages() on an mmap()ed file allowed? What to do if 0 < get_user_pages(..., nr_pages, ...) < nr_pages?

[Leon Woestenberg <leon.woestenberg@gmail.com>]

Hello ,

On Mon, Aug 3, 2009 at 6:30 PM, Hugh Dickins<hugh.dickins@tiscali.co.uk> wrote:
> On Mon, 3 Aug 2009, Leon Woestenberg wrote:
>>
>> - is it allowed to have a PCI device DMA-read from memory pages, that
>> belong to a file mmap()'d by userspace?
>
> Yes.
>
>> - what are valid reasons for get_user_pages() to fail?
>
> I'd hesitate to give a complete answer to that: but main reasons
> would be SIGKILL, or running out of memory (-ENOMEM), or running
> off the end of a mapping or mapped object, or no permission to it
> (-EFAULT): with a short page count returned instead of error if
> some pages were successfully gotten before hitting the error.
>
Next step will be to see where it bails out.

>> - what should a driver do when get_user_pages() returns less pages
>> than requested?
>
> Probably put_page the pages gotten then report the surprise;
> perhaps, before putting the pages gotten, try get_user_pages
> on the next alone, to see what error code is returned for that.
>
> Unless it's happy to work with fewer pages than requested,
> in which case work with them and ignore the surprise.
>
I expect a certain amount of data to be DMA'd from the PCI device to
the file mmap, so I'ld rather map the complete file before I start
DMA.

>>     BUG_ON(rc < nr_pages);
>
> When that BUG triggers, is rc a positive number of pages,
> or a negative error code - which?  (or even 0, but it shouldn't be).
> I assume from your Subject that you've already seen a positive number
> of pages.
>
Correct, undeterministicly it sometimes return the requested amount,
sometimes some part of it, or sometimes errors out.

> Code doesn't look wrong to me (except you shouldn't BUG), though I am
>
Correct. I took a snippet

> having to assume that buffer and boe and start are all the same address,

Correct.

> and count fits within buffersize; or at least that the range to which you
> apply get_user_pages really does fit within the area you have mmap'ed.
>
Yes, the file is large enough, a multiple of PAGE_SIZE, as is the mmap length.

> (I'd advise against using 1 /* do force */, I don't think you need
> that: the force is mysterious, and should only be called upon in direst
> need.  But it shouldn't actually be causing you any problem here.)
>
Thanks. I found that force did not mean "force the mapping"  but
rather enforce read plus write permissions.

> Is the file you have mmap'ed big enough?  If it's not as long as the
> last page you're trying to get_user_pages on, or gets truncated, then
> indeed that will give -EFAULT or a short count - just as trying to
> access the end of the mapping in userspace would give you SIGBUS.
>
Thanks for all the pointers, I think I have all the conditions right.

I'll see what get_user_pages() internally fails on. It's one of the
API's that does not either fail or complete, it can say: "hey I did a
part of it, now buzz off!"
I still wonder if the function can at all be wrapped with a
all-or-nothing function?

Regards,
-- 
Leon