On Wed, 27 Sept 2023 at 07:32, Hsin-Wei Hung <hsinweih@xxxxxxx> wrote: > > Hi, > > We found a potential memory leak in bpf_timer in v5.15.26 using a > customized syzkaller for fuzzing bpf runtime. It can happen when > an arraymap is being released. An entry that has been checked by > bpf_timer_cancel_and_free() can again be initialized by bpf_timer_init(). > Since both paths are almost identical between v5.15 and net-next, > I suspect this problem still exists. Below are kmemleak report and > some additional printks I inserted. > > [ 1364.081694] array_map_free_timers map:0xffffc900005a9000 > [ 1364.081730] ____bpf_timer_init map:0xffffc900005a9000 > timer:0xffff888001ab4080 > > *no bpf_timer_cancel_and_free that will kfree struct bpf_hrtimer* > at 0xffff888001ab4080 is called > > [ 1383.907869] kmemleak: 1 new suspected memory leaks (see > /sys/kernel/debug/kmemleak) > BUG: memory leak > unreferenced object 0xffff888001ab4080 (size 96): > comm "sshd", pid 279, jiffies 4295233126 (age 29.952s) > hex dump (first 32 bytes): > 80 40 ab 01 80 88 ff ff 00 00 00 00 00 00 00 00 .@.............. > 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ > backtrace: > [<000000009d018da0>] bpf_map_kmalloc_node+0x89/0x1a0 > [<00000000ebcb33fc>] bpf_timer_init+0x177/0x320 > [<00000000fb7e90bf>] 0xffffffffc02a0358 > [<000000000c89ec4f>] __cgroup_bpf_run_filter_skb+0xcbf/0x1110 > [<00000000fd663fc0>] ip_finish_output+0x13d/0x1f0 > [<00000000acb3205c>] ip_output+0x19b/0x310 > [<000000006b584375>] __ip_queue_xmit+0x182e/0x1ed0 > [<00000000b921b07e>] __tcp_transmit_skb+0x2b65/0x37f0 > [<0000000026104b23>] tcp_write_xmit+0xf19/0x6290 > [<000000006dc71bc5>] __tcp_push_pending_frames+0xaf/0x390 > [<00000000251b364a>] tcp_push+0x452/0x6d0 > [<000000008522b7d3>] tcp_sendmsg_locked+0x2567/0x3030 > [<0000000038c644d2>] tcp_sendmsg+0x30/0x50 > [<000000009fe3413f>] inet_sendmsg+0xba/0x140 > [<0000000034d78039>] sock_sendmsg+0x13d/0x190 > [<00000000f55b8db6>] sock_write_iter+0x296/0x3d0 > > Does this happen on bpf-next? Things have changed around timer freeing since then. Or even sharing the reproducer for this will work. I can take a look. Thanks
