| Message ID | 20221126121458.3564942-1-lijinlin3@huawei.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | blk-iocost: fix shift-out-of-bounds in iocg_hick_delay() | expand |
On Sat, Nov 26, 2022 at 08:14:58PM +0800, Li Jinlin wrote: > if (iocg->delay) > - delay = iocg->delay >> div64_u64(tdelta, USEC_PER_SEC); > + delay = iocg->delay >> min(div64_u64(tdelta, USEC_PER_SEC), 63); Let's just set delay to 0 if the shift is too big. Thanks.
diff --git a/block/blk-iocost.c b/block/blk-iocost.c index 07c1a31dd495..2b837ac4b2ba 100644 --- a/block/blk-iocost.c +++ b/block/blk-iocost.c @@ -1332,7 +1332,7 @@ static bool iocg_kick_delay(struct ioc_gq *iocg, struct ioc_now *now) /* calculate the current delay in effect - 1/2 every second */ tdelta = now->now - iocg->delay_at; if (iocg->delay) - delay = iocg->delay >> div64_u64(tdelta, USEC_PER_SEC); + delay = iocg->delay >> min(div64_u64(tdelta, USEC_PER_SEC), 63); else delay = 0;
We got the following UBSAN report: ==================================================================== UBSAN: shift-out-of-bounds in block/blk-iocost.c:1294:23 shift exponent 18446744073709 is too large for 64-bit type ...... CPU: 1 PID: 1088217 Comm: fsstress Kdump: loaded Not tainted ...... Hardware name: QEMU Standard PC (i440FX + PIIX, 1996) ...... Call Trace: dump_stack+0x9c/0xd3 ubsan_epilogue+0xa/0x4e __ubsan_handle_shift_out_of_bounds.cold+0x87/0x137 iocg_kick_delay.cold+0x18/0x60 ioc_rqos_throttle+0x7f8/0x870 __rq_qos_throttle+0x40/0x60 blk_mq_submit_bio+0x24d/0xd60 __submit_bio_noacct_mq+0x10b/0x270 submit_bio_noacct+0x13d/0x150 submit_bio+0xbf/0x280 submit_bh_wbc+0x3aa/0x450 ext4_read_bh_nowait+0xdb/0x180 [ext4] ext4_read_bh_lock+0x6d/0x90 [ext4] ext4_bread_batch+0x24c/0x2e0 [ext4] __ext4_find_entry+0x2d2/0x880 [ext4] ext4_lookup.part.0+0xbf/0x370 [ext4] ext4_lookup+0x3e/0x60 [ext4] lookup_open.isra.0+0x343/0x630 open_last_lookups+0x1f2/0x750 path_openat+0x133/0x330 do_filp_open+0x122/0x270 do_sys_openat2+0x3a8/0x550 __x64_sys_creat+0xae/0xe0 do_syscall_64+0x33/0x40 entry_SYSCALL_64_after_hwframe+0x61/0xc6 =================================================================== The result of E1 >> E2 is E1 right-shifted E2 bit positions. From the report, we know E2 is greater than the width of E1. In the C99 standard, if the value of the E2 is negative or is greater than or equal to the width of E1, the behavior is undefined. In the actual test, if the E2 is greater than or equal to the width of E1, the result of E1 >> E2 is E1 >> (E2 %/ E1width), which is not what we want. So letting the value of the right operand be less than the width of u64 in this expression. Signed-off-by: Li Jinlin <lijinlin3@huawei.com> --- block/blk-iocost.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)