On Wed 2020-10-14 16:05:20, Petr Mladek wrote:
On Wed 2020-10-14 16:19:18, Naresh Kamboju wrote:
While testing LTP controllers testing on x86_64 KASAN enabled linux next 20201013 tag kernel this kernel BUG noticed. The oom-killer log also noticed while running tests.
metadata: git branch: master git repo: https://gitlab.com/Linaro/lkft/mirrors/next/linux-next git commit: f2fb1afc57304f9dd68c20a08270e287470af2eb git describe: next-20201013 make_kernelversion: 5.9.0 kernel-config: https://builds.tuxbuild.com/TXP6cqokP8BCYJrv7zf6kw/kernel.config
This is the last string stored in the ring buffer before KASAN trigerred.
[ 221.922108] ================================================================== [ 221.922111] BUG: KASAN: global-out-of-bounds in vprintk_store+0x362/0x3d0 [ 221.922112] Write of size 2 at addr ffffffffba51dbcd by task memcg_test_1/11282
[ 221.922114] CPU: 1 PID: 11282 Comm: memcg_test_1 Not tainted 5.9.0-next-20201013 #1 [ 221.922116] Hardware name: Supermicro SYS-5019S-ML/X11SSH-F, BIOS 2.0b 07/27/2017 [ 221.922116] Call Trace: [ 221.922117] dump_stack+0xa4/0xd9 [ 221.922118] print_address_description.constprop.0+0x21/0x210 [ 221.922119] ? _raw_write_lock_bh+0xe0/0xe0 [ 221.922120] ? vprintk_store+0x362/0x3d0 [ 221.922121] kasan_report.cold+0x37/0x7c [ 221.922122] ? vprintk_store+0x362/0x3d0 [ 221.922123] check_memory_region+0x18c/0x1f0 [ 221.922124] memcpy+0x3c/0x60 [ 221.922125] vprintk_store+0x362/0x3d0
It seems that vprintk() store was able to extend the last string by the two characters. So this is likely:
static size_t log_output(int facility, int level, enum log_flags lflags, const struct dev_printk_info *dev_info, char *text, size_t text_len) { [...] if (prb_reserve_in_last(&e, prb, &r, caller_id, LOG_LINE_MAX)) { memcpy(&r.text_buf[r.info->text_len], text, text_len);
But very likely the two characters were copied to wrong location. There are many similar lines in the full log and they always contain
task_memcg=/0It means that the size of the path is 2 characters but it should be "/0". I guess that "in" was in the log buffer from the previous wrap.
So, it seems that prb_reserve_in_last() correctly updated the size of the extended record but it returned wrong pointer to the buffer or wrong current length.
Anyway, prb_commit(&e) moved the buffer back to consistent state.
[ 213O[ 221.922182] The buggy address belongs to the variable: [ 221.922183] clear_seq+0x2d/0x40
The address ffffffffba51dbcd seems to be from the range of "module mapping space" adresses.
"clear_seq" is static variable from kernel/printk/printk.c.
There is a bug in data_realloc(). We return wrong pointer when there is no need to resize the record and it has been wrapped.
It might happen when we add only few characters and there was a space because of alignment.
It would cause writing outside the buffer into another statically defined variables. So, it might be it.
The following change patch should fix it:
diff --git a/kernel/printk/printk_ringbuffer.c b/kernel/printk/printk_ringbuffer.c index 2493348a1631..24a960a89aa8 100644 --- a/kernel/printk/printk_ringbuffer.c +++ b/kernel/printk/printk_ringbuffer.c @@ -1125,7 +1125,10 @@ static char *data_realloc(struct printk_ringbuffer *rb,
/* If the data block does not increase, there is nothing to do. */ if (head_lpos - next_lpos < DATA_SIZE(data_ring)) { - blk = to_block(data_ring, blk_lpos->begin); + if (wrapped) + blk = to_block(data_ring, 0); + else + blk = to_block(data_ring, blk_lpos->begin); return &blk->data[0]; }