commit bc0c4d1e176eeb614dc8734fc3ace34292771f11
author Linus Torvalds <torvalds@linux-foundation.org> Fri Apr 24 11:10:58 2020 -0700
committer Linus Torvalds <torvalds@linux-foundation.org> Fri Apr 24 13:28:03 2020 -0700
tree 56feb0f1b79069da8c18c20387405a5ae63c2337
parent aee1a009c9d247756b368890e3f886174776d3db

mm: check that mm is still valid in madvise()

IORING_OP_MADVISE can end up basically doing mprotect() on the VM of
another process, which means that it can race with our crazy core dump
handling which accesses the VM state without holding the mmap_sem
(because it incorrectly thinks that it is the final user).

This is clearly a core dumping problem, but we've never fixed it the
right way, and instead have the notion of "check that the mm is still
ok" using mmget_still_valid() after getting the mmap_sem for writing in
any situation where we're not the original VM thread.

See commit 04f5866e41fb ("coredump: fix race condition between
mmget_not_zero()/get_task_mm() and core dumping") for more background on
this whole mmget_still_valid() thing.  You might want to have a barf bag
handy when you do.

We're discussing just fixing this properly in the only remaining core
dumping routines.  But even if we do that, let's make do_madvise() do
the right thing, and then when we fix core dumping, we can remove all
these mmget_still_valid() checks.

Reported-and-tested-by: Jann Horn <jannh@google.com>
Fixes: c1ca757bd6f4 ("io_uring: add IORING_OP_MADVISE")
Acked-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm/madvise.c

diff --git a/mm/madvise.c b/mm/madvise.c
index 4bb30ed..8cbd8c1 100644
--- a/mm/madvise.c
+++ b/mm/madvise.c
@@ -27,6 +27,7 @@
 #include <linux/swapops.h>
 #include <linux/shmem_fs.h>
 #include <linux/mmu_notifier.h>
+#include <linux/sched/mm.h>
 
 #include <asm/tlb.h>
 
@@ -1090,6 +1091,23 @@ int do_madvise(unsigned long start, size_t len_in, int behavior)
 	if (write) {
 		if (down_write_killable(&current->mm->mmap_sem))
 			return -EINTR;
+
+		/*
+		 * We may have stolen the mm from another process
+		 * that is undergoing core dumping.
+		 *
+		 * Right now that's io_ring, in the future it may
+		 * be remote process management and not "current"
+		 * at all.
+		 *
+		 * We need to fix core dumping to not do this,
+		 * but for now we have the mmget_still_valid()
+		 * model.
+		 */
+		if (!mmget_still_valid(current->mm)) {
+			up_write(&current->mm->mmap_sem);
+			return -EINTR;
+		}
 	} else {
 		down_read(&current->mm->mmap_sem);
 	}