Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Date:   Sat, 24 Dec 2022 00:12:02 -0800
Mime-Version: 1.0
Message-ID: <20221224081203.3193960-1-zokeefe@google.com>
Subject: [PATCH v2 1/2] mm/MADV_COLLAPSE: don't expand collapse when vm_end is
 past requested end
From:   "Zach O'Keefe" <zokeefe@google.com>
To:     linux-mm@kvack.org
Cc:     stable@vger.kernel.org, linux-kernel@vger.kernel.org,
        Andrew Morton <akpm@linux-foundation.org>,
        Hugh Dickins <hughd@google.com>,
        Yang Shi <shy828301@gmail.com>,
        "Zach O'Keefe" <zokeefe@google.com>
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

MADV_COLLAPSE acts on one hugepage-aligned/sized region at a time, until
it has collapsed all eligible memory contained within the bounds
supplied by the user.

At the top of each hugepage iteration we (re)lock mmap_lock and
(re)validate the VMA for eligibility and update variables that might
have changed while mmap_lock was dropped.  One thing that might occur,
is that the VMA could be resized, and as such, we refetch vma->vm_end
to make sure we don't collapse past the end of the VMA's new end.

However, it's possible that when refetching vma>vm_end that we expand the
region acted on by MADV_COLLAPSE if vma->vm_end is greater than size+len
supplied by the user.

The consequence here is that we may attempt to collapse more memory than
requested, possibly yielding either "too much success" or "false
failure" user-visible results.  An example of the former is if we
MADV_COLLAPSE the first 4MiB of a 2TiB mmap()'d file, the incorrect
refetch would cause the operation to block for much longer than
anticipated as we attempt to collapse the entire TiB region.  An example
of the latter is that applying MADV_COLLPSE to a 4MiB file mapped to the
start of a 6MiB VMA will successfully collapse the first 4MiB, then
incorrectly attempt to collapse the last hugepage-aligned/sized region
-- fail (since readahead/page cache lookup will fail) -- and report a
failure to the user.

Don't expand the acted-on region when refetching vma->vm_end.

Fixes: 4d24de9425f7 ("mm: MADV_COLLAPSE: refetch vm_end after reacquiring mmap_lock")
Reported-by: Hugh Dickins <hughd@google.com>
Signed-off-by: Zach O'Keefe <zokeefe@google.com>
Cc: Yang Shi <shy828301@gmail.com>
---
v1->v2 : Updated changelog to make clear what user-visible issues this
	 patch addresses, as well makes the case for backporting (Andrew
	 Morton).

While there aren't any stability risks, without this patch there exist
trivial examples where MADV_COLLAPSE won't work; as such, this should be
backported to stable 6.1.X to make MADV_COLLAPSE dependable in such
cases.

v1: https://lore.kernel.org/linux-mm/CAAa6QmRx_b2UCJWE2XZ3=3c3-_N3R4cDGX6Wm4OT7qhFC6U_SQ@mail.gmail.com/T/#m6c91da3cdbd9b1d1ebb29d415962deb158a2c658
---
 mm/khugepaged.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/mm/khugepaged.c b/mm/khugepaged.c
index 5cb401aa2b9d..b4d2ec0a94ed 100644
--- a/mm/khugepaged.c
+++ b/mm/khugepaged.c
@@ -2649,7 +2649,7 @@ int madvise_collapse(struct vm_area_struct *vma, struct vm_area_struct **prev,
 				goto out_nolock;
 			}
 
-			hend = vma->vm_end & HPAGE_PMD_MASK;
+			hend = min(hend, vma->vm_end & HPAGE_PMD_MASK);
 		}
 		mmap_assert_locked(mm);
 		memset(cc->node_load, 0, sizeof(cc->node_load));
-- 
2.39.0.314.g84b9a713c41-goog