On 2/28/20 7:43 AM, Claudio Imbrenda wrote: > In case pin fails, we need to unpin, a simple put_page will not be enough > > Signed-off-by: Claudio Imbrenda <imbrenda@xxxxxxxxxxxxx> > --- > mm/gup.c | 5 ++++- > 1 file changed, 4 insertions(+), 1 deletion(-) > > diff --git a/mm/gup.c b/mm/gup.c > index f589299b0d4a..0b9a806898f3 100644 > --- a/mm/gup.c > +++ b/mm/gup.c > @@ -2134,7 +2134,10 @@ static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end, > goto pte_unmap; > > if (unlikely(pte_val(pte) != pte_val(*ptep))) { > - put_page(head); > + if (flags & FOLL_GET) > + put_page(head); > + else if (flags & FOLL_PIN) > + unpin_user_page(head); Hi Claudio, Instead, I think that should actually be: put_compound_head(page, 1, flags); which does a bit more (bug checks and /proc/vmstat instrumentation) than your diff, but has the same basic idea: call the right "put" function. ...oh, actually, I see you have the commit hash in the subject line. Instead, it should be in the commit description. Let's maybe change the subject line to approx: mm/gup: Fix a missing put_compound_head() call in gup_pte_range() And the write up...how about something like this, if you like it: try_grab_compound_head() must be undone via put_compound_head(), not put_page(). This was missed in the original implementation of the gup/dma tracking patch, so fix it now. Fixes: 0ea2781c3de4 ("mm/gup: track FOLL_PIN pages") (Aside: I'm using the linux-next commit hash. How does one get the correct hash before it goes to mainline? I guess maintainer scripts fix all those up?) It's also good to Cc some reviewers in case I'm overlooking something, so I'm adding Jan and Kirill. thanks, -- John Hubbard NVIDIA > goto pte_unmap; > } > >