[Patch] blk-wbt: fix a divide-by-zero error in rwb_arm_timer()

[Patch] blk-wbt: fix a divide-by-zero error in rwb_arm_timer()

[Cong Wang]

From: Cong Wang <cong.wang@bytedance.com>

We hit a divide error in rwb_arm_timer() and crash dump shows
rqd->scale_step is 16777215 (0xffffff in hex), so the expression
"(rqd->scale_step + 1) << 8)" is 0x100000000, which is just beyond
32-bit integer range, hence it is truncated to 0 and int_sqrt(0)
returns 0 too, so we end up passing 0 as a divisor to div_u64().

Looking at the assembly code generated:

    add    $0x1,%edi
    shl    $0x8,%edi
    movslq %edi,%rdi
    mov    0x10(%rbx),%rdi
    xor    %edx,%edx
    mov    %eax,%ecx
    shl    $0x4,%rdi
    mov    %rdi,%rax
    div    %rcx

we notice that the left shift is still using 32 bit register %edi,
because the type of rqd->scale_step is 'int'. But actually int_sqrt()
takes 'long' as a parameter, so the temporary result should fit well
at least on x86_64. Fix this by explicitly casting the expression to
u64 and call int_sqrt64() to avoid any ambiguity on 32 bit.

After this patch, the assembly code looks correct:

    add    $0x1,%edi
    movslq %edi,%rdi
    shl    $0x8,%rdi
    mov    0x10(%rbx),%rdi
    xor    %edx,%edx
    mov    %eax,%ecx
    shl    $0x4,%rdi
    mov    %rdi,%rax
    div    %rcx

Fixes: e34cbd307477 ("blk-wbt: add general throttling mechanism")
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Fam Zheng <fam.zheng@bytedance.com>
Cc: Xiongchun Duan <duanxiongchun@bytedance.com>
Signed-off-by: Cong Wang <cong.wang@bytedance.com>
---
 block/blk-wbt.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/block/blk-wbt.c b/block/blk-wbt.c
index 42aed0160f86..5157ca86574f 100644
--- a/block/blk-wbt.c
+++ b/block/blk-wbt.c
@@ -337,7 +337,7 @@ static void rwb_arm_timer(struct rq_wb *rwb)
 		 * though.
 		 */
 		rwb->cur_win_nsec = div_u64(rwb->win_nsec << 4,
-					int_sqrt((rqd->scale_step + 1) << 8));
+					int_sqrt64((u64)(rqd->scale_step + 1) << 8));
 	} else {
 		/*
 		 * For step < 0, we don't want to increase/decrease the
-- 
2.25.1

Re: [Patch] blk-wbt: fix a divide-by-zero error in rwb_arm_timer()

[Cong Wang]

On Sat, Apr 17, 2021 at 9:41 PM Cong Wang <xiyou.wangcong@gmail.com> wrote:
>
> From: Cong Wang <cong.wang@bytedance.com>
>
> We hit a divide error in rwb_arm_timer() and crash dump shows
> rqd->scale_step is 16777215 (0xffffff in hex), so the expression
> "(rqd->scale_step + 1) << 8)" is 0x100000000, which is just beyond
> 32-bit integer range, hence it is truncated to 0 and int_sqrt(0)
> returns 0 too, so we end up passing 0 as a divisor to div_u64().
>

Never mind. rqd->scale_step should be capped by
rq_depth_scale_down(), so should never be so large. In the old
calc_wb_limits() implementation, rwb->wb_max was set to zero
accidentally.

Thanks.