[PATCH v3 RESEND] perf: Fix refcount warning on event->mmap_count increment

[Will Rosenberg]

When calling refcount_inc(&event->mmap_count) inside perf_mmap_rb(), the
following warning is triggered:

        refcount_t: addition on 0; use-after-free.
        WARNING: lib/refcount.c:25

PoC:

    struct perf_event_attr attr = {0};
    int fd = syscall(__NR_perf_event_open, &attr, 0, -1, -1, 0);
    mmap(NULL, 0x3000, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
    int victim = syscall(__NR_perf_event_open, &attr, 0, -1, fd,
                         PERF_FLAG_FD_OUTPUT);
    mmap(NULL, 0x3000, PROT_READ | PROT_WRITE, MAP_SHARED, victim, 0);

This occurs when creating a group member event with the flag
PERF_FLAG_FD_OUTPUT. The group leader should be mmap-ed and then mmap-ing
the event triggers the warning.

Since the event has copied the output_event in perf_event_set_output(),
event->rb is set. As a result, perf_mmap_rb() calls
refcount_inc(&event->mmap_count) when event->mmap_count = 0.

Disallow the case when event->mmap_count = 0. This also prevents two
events from updating the same user_page.

Fixes: 448f97fba901 ("perf: Convert mmap() refcounts to refcount_t")
Suggested-by: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
Signed-off-by: Will Rosenberg <whrosenb@xxxxxxx>
---

Notes:
    v2 -> v3: Update patch to error out instead of incrementing.
    
    Thank you, this is a much better solution. I was not thinking
    that the mmap itself was unintended.
    
    I believe you are missing a "!" in your patch. After adding
    that, I tested the patch, and it fixed the bug.
    
    I also wanted to check my understanding of the race with
    perf_mmap_close() to double check this patch will not cause
    an issue. perf_mmap_rb() should always hold the
    event->mmap_mutex, so there should be no race on
    event->mmap_count with perf_mmap_close()'s
    refcount_dec_and_mutex_lock(). If there was a race, we would
    risk returning -EBUSY when we should "continue as if !event->rb."
    
    Thank you for your help.

 kernel/events/core.c | 9 +++++++++
 1 file changed, 9 insertions(+)

diff --git a/kernel/events/core.c b/kernel/events/core.c
index 3c2a491200c6..ac7f12560172 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -7273,6 +7273,15 @@ static int perf_mmap_rb(struct vm_area_struct *vma, struct perf_event *event,
 		if (data_page_nr(event->rb) != nr_pages)
 			return -EINVAL;
 
+		/*
+		 * If this event doesn't have mmap_count, we're attempting to
+		 * create an alias of another event's mmap(); this would mean
+		 * both events will end up scribbling the same user_page;
+		 * which makes no sense.
+		 */
+		if (!refcount_read(&event->mmap_count))
+			return -EBUSY;
+
 		if (refcount_inc_not_zero(&event->rb->mmap_count)) {
 			/*
 			 * Success -- managed to mmap() the same buffer

base-commit: 5d3b0106245d467fd5ba0bd9a373a13356684f6e
-- 
2.34.1