On 2/10/22 11:29, Minchan Kim wrote:
On Thu, Feb 10, 2022 at 11:20:31AM -0800, John Hubbard wrote:
On 2/10/22 11:17, Minchan Kim wrote:
* pin_user_pages() - pin user pages in memory for use by other devices
< snip >
* @pages: array that receives pointers to the pages pinned.
* Should be at least nr_pages long. Or NULL, if caller
* only intends to ensure the pages are faulted in.
pin_user_pages(,, pages = NULL, );
gup_flags |= FOLL_PIN
__get_user_pages_locked
__get_user_pages
..
VM_BUG_ON(!!pages != !!(gup_flags & (FOLL_GET | FOLL_PIN)));
Only FOLL_GET or FOLL_PIN are supposed to fill in the **pages array. So
if a caller passes a null **pages arg, then that caller must not also
set FOLL_GET or FOLL_PIN. That's what the VM_BUG_ON() is expressing.
Yub, but pin_user_pages adds FOLL_PIN unconditinally and the comments
says it supports NUU pages argument. Isn't it conflict?
Oh right, that is a conflict. The documentation should *not* say that a
NULL **pages arg is supported. Because the whole point of the FOLL_PIN
APIs is to actually pin struct pages. The NULL cases are only useful for
get_user_pages*().
So removing that last sentence is appropriate, plus also looking around
for similar documentation claims, including in pin_user_pages.rst. I
don't see anything from a very quick scan, though.
Sending out a fix will also trigger the observation that both the
kerneldoc headers mm/gup.c, and the writings in pin_user_pages.rst,
could use some updating. However, it is also true that this can be
reasonably treated as a documentation bug fix, and therefore allowed to
be limited to just this change.
Were you going to send out a formal patch? If not, I can include it in
an upcoming gup series, with your Reported-by tag. Up to you.
Perhaps that should be part of the documentation. It sort of is already,
for get_user_pages*().
I expected it was just copied from get_user_pages.
Yes it was.
thanks,
--
John Hubbard
NVIDIA
