Re: [PATCH v3] blk-throttle: fix divide-by-zero on legacy iops limit of 0

在 2026/7/22 15:25, yu kuai 写道:
> Hi,
> 
> 在 2026/7/21 12:08, Tao Cui 写道:
>> From: Tao Cui <[email protected]>
>>
>> Writing a multiple of 2^32 (e.g. 4294967296) to a legacy cgroup v1
>> throttle iops file (blkio.throttle.{read,write}_iops_device) silently
>> truncates to 0: tg_set_conf() stores the sscanf-parsed u64 value into
>> an unsigned int field with no clamping. The cgroup v2 path,
>> tg_set_limit(), already clamps the same kind of value with
>> min_t(u64, val, UINT_MAX), but the legacy path never did. Note that
>> the "!v -> U64_MAX" mapping only catches an explicit zero and does not
>> catch a value that truncates to zero.
>>
>> With iops stored as 0, tg_update_has_rules() sets has_rules_iops[] and
>> the next IO reaches tg_within_iops_limit(), which computes
>>
>>      jiffy_wait = max(jiffy_wait, HZ / iops_limit + 1);
>>
>> triggering a divide-by-zero oops. The unclamped write in tg_set_conf()
>> is long-standing, but it only became a crash once the HZ / iops_limit
>> divide was added.
>>
>> Fix it in two places:
>>
>>    * tg_set_conf(): clamp the value to UINT_MAX, consistent with
>>      tg_set_limit(). This closes the truncation root cause (and the
>>      general silent truncation for any value above UINT_MAX).
>>
>>    * tg_dispatch_iops_time(): treat iops_limit == 0 as unlimited so the
>>      divide in tg_within_iops_limit() is never reached, defending
>>      against any future path that could produce a zero limit.
>>
>> Fixes: 1beabab88ece ("blk-throttle: fix lower control under super low iops limit")
> 
Hi Kuai,

Thanks for the review, and you're right on both points.

> I think the root cause is that iops_limit set to 0, and what's the behavior of throttle
> with iops=0? Since calculate_io_allowed() will always return 0, it's likely that no IO
> can be issued.

On the root cause: yes, iops=0 is the real root cause and it predates
1beabab88ece. With iops_limit = 0, calculate_io_allowed() returns 0
(0 * jiffy_elapsed), so no IO can be issued at all; 1beabab88ece only
turned that pre-existing iops=0 hang into a divide-by-zero via the new
HZ / iops_limit. Attributing it to 1beabab88ece was therefore wrong --
my apologies. I'll drop the Fixes: tag in v4.
> 
>> Signed-off-by: Tao Cui <[email protected]>
>>
>> ---
>> Changes in v3:
>> - Drop the (u64) cast on UINT_MAX: the kernel's type-checked min() accepts
>>    two unsigned types of different width (both >= 4 bytes), so
>>    min(v, UINT_MAX) compiles clean. (David Laight)
>> - Add a Fixes: tag pointing at the commit that introduced the
>>    HZ / iops_limit divide, which is also where the oops became reachable.
>>
>> Changes in v2:
>> - Use a "void *field" local for the config write so the assignment reads
>>    *(u64 *)field / *(unsigned int *)field instead of the
>>    (type *)((void *)tg + of_cft(of)->private) casts.
>> - Use min(v, UINT_MAX) instead of min_t(u64, v, UINT_MAX).
>> ---
>>   block/blk-throttle.c | 13 ++++++++++---
>>   1 file changed, 10 insertions(+), 3 deletions(-)
>>
>> diff --git a/block/blk-throttle.c b/block/blk-throttle.c
>> index ffc3b70065d4..e894852c3142 100644
>> --- a/block/blk-throttle.c
>> +++ b/block/blk-throttle.c
>> @@ -883,7 +883,12 @@ static unsigned long tg_dispatch_iops_time(struct throtl_grp *tg, struct bio *bi
>>   	u32 iops_limit = tg_iops_limit(tg, rw);
>>   	unsigned long iops_wait;
>>   
>> -	if (iops_limit == UINT_MAX || tg->flags & THROTL_TG_CANCELING)
>> +	/*
>> +	 * iops_limit == 0 is not a valid limit. Treat it as unlimited so we
>> +	 * never reach the HZ / iops_limit divide in tg_within_iops_limit().
>> +	 */
>> +	if (iops_limit == UINT_MAX || iops_limit == 0 ||
>> +	    tg->flags & THROTL_TG_CANCELING)
>>   		return 0;
> 
> Is the above still needed with the tg_set_conf() changes? 0 is not a valid limit, and
> it should not exist in the first place, so I don't like the above new checking.
On the iops_limit == 0 check: agreed, dropping it. With the tg_set_conf()
clamp in place, 0 can never be stored as a limit (explicit 0 maps to
U64_MAX, anything that would truncate to 0 is clamped to UINT_MAX, and
the init value is UINT_MAX), so the runtime check only guards an
unreachable state. v4 keeps only the tg_set_conf() clamp.

Will send v4 shortly.

Thanks,
Tao
> 
>>   
>>   	tg_update_slice(tg, rw);
>> @@ -1383,10 +1388,12 @@ static ssize_t tg_set_conf(struct kernfs_open_file *of,
>>   	tg = blkg_to_tg(ctx.blkg);
>>   	tg_update_carryover(tg);
>>   
>> +	void *field = (void *)tg + of_cft(of)->private;
>> +
>>   	if (is_u64)
>> -		*(u64 *)((void *)tg + of_cft(of)->private) = v;
>> +		*(u64 *)field = v;
>>   	else
>> -		*(unsigned int *)((void *)tg + of_cft(of)->private) = v;
>> +		*(unsigned int *)field = min(v, UINT_MAX);
>>   
>>   	tg_conf_updated(tg, false);
>>   	ret = 0;
>