Conceptually, root_iocg->hweight_donating must be less than WEIGHT_ONE but
all hweight calculations round up and thus it may end up >= WEIGHT_ONE
triggering divide-by-zero and other issues. Bound the value to avoid
surprises.
Signed-off-by: Tejun Heo <[email protected]>
Fixes: e08d02aa5fc9 ("blk-iocost: implement Andy's method for donation weight updates")
---
block/blk-iocost.c | 14 ++++++++++----
1 file changed, 10 insertions(+), 4 deletions(-)
diff --git a/block/blk-iocost.c b/block/blk-iocost.c
index fc897bb142bcd..978753335370a 100644
--- a/block/blk-iocost.c
+++ b/block/blk-iocost.c
@@ -1881,15 +1881,21 @@ static void transfer_surpluses(struct list_head *surpluses, struct ioc_now *now)
/*
* Calculate the global donation rate (gamma) - the rate to adjust
- * non-donating budgets by. No need to use 64bit multiplication here as
- * the first operand is guaranteed to be smaller than WEIGHT_ONE
- * (1<<16).
+ * non-donating budgets by.
+ *
+ * No need to use 64bit multiplication here as the first operand is
+ * guaranteed to be smaller than WEIGHT_ONE (1<<16).
+ *
+ * We know that there are beneficiary nodes and the sum of the donating
+ * hweights can't be whole; however, due to the round-ups during hweight
+ * calculations, root_iocg->hweight_donating might still end up equal to
+ * or greater than whole. Limit the range when calculating the divider.
*
* gamma = (1 - t_r') / (1 - t_r)
*/
gamma = DIV_ROUND_UP(
(WEIGHT_ONE - root_iocg->hweight_after_donation) * WEIGHT_ONE,
- WEIGHT_ONE - root_iocg->hweight_donating);
+ WEIGHT_ONE - min_t(root_iocg->hweight_donating, WEIGHT_ONE - 1));
/*
* Calculate adjusted hwi, child_adjusted_sum and inuse for the inner
On 9/11/20 11:07 AM, Tejun Heo wrote:
> Conceptually, root_iocg->hweight_donating must be less than WEIGHT_ONE but
> all hweight calculations round up and thus it may end up >= WEIGHT_ONE
> triggering divide-by-zero and other issues. Bound the value to avoid
> surprises.
Applied, thanks.
--
Jens Axboe
Conceptually, root_iocg->hweight_donating must be less than WEIGHT_ONE but
all hweight calculations round up and thus it may end up >= WEIGHT_ONE
triggering divide-by-zero and other issues. Bound the value to avoid
surprises.
Signed-off-by: Tejun Heo <[email protected]>
Fixes: e08d02aa5fc9 ("blk-iocost: implement Andy's method for donation weight updates")
Signed-off-by: Tejun Heo <[email protected]>
---
Jens, I was flipping between doing max_t(, 1) over the whole divider and
doing min_t(, WEIGHT_ONE - 1) for hweight_donating. I thought that I as
testing after the last change but it obviously wasn't and the previous patch
doesn't compile due to missing type argument. Can you please apply this
patch instead? I can send an incremental patch if that'd be better. My
apologies.
Thanks.
block/blk-iocost.c | 14 ++++++++++----
1 file changed, 10 insertions(+), 4 deletions(-)
diff --git a/block/blk-iocost.c b/block/blk-iocost.c
index fc897bb142bcd..6e29b4dcf3566 100644
--- a/block/blk-iocost.c
+++ b/block/blk-iocost.c
@@ -1881,15 +1881,21 @@ static void transfer_surpluses(struct list_head *surpluses, struct ioc_now *now)
/*
* Calculate the global donation rate (gamma) - the rate to adjust
- * non-donating budgets by. No need to use 64bit multiplication here as
- * the first operand is guaranteed to be smaller than WEIGHT_ONE
- * (1<<16).
+ * non-donating budgets by.
+ *
+ * No need to use 64bit multiplication here as the first operand is
+ * guaranteed to be smaller than WEIGHT_ONE (1<<16).
+ *
+ * We know that there are beneficiary nodes and the sum of the donating
+ * hweights can't be whole; however, due to the round-ups during hweight
+ * calculations, root_iocg->hweight_donating might still end up equal to
+ * or greater than whole. Limit the range when calculating the divider.
*
* gamma = (1 - t_r') / (1 - t_r)
*/
gamma = DIV_ROUND_UP(
(WEIGHT_ONE - root_iocg->hweight_after_donation) * WEIGHT_ONE,
- WEIGHT_ONE - root_iocg->hweight_donating);
+ WEIGHT_ONE - min_t(u32, root_iocg->hweight_donating, WEIGHT_ONE - 1));
/*
* Calculate adjusted hwi, child_adjusted_sum and inuse for the inner
On 9/11/20 4:40 PM, Tejun Heo wrote:
> Conceptually, root_iocg->hweight_donating must be less than WEIGHT_ONE but
> all hweight calculations round up and thus it may end up >= WEIGHT_ONE
> triggering divide-by-zero and other issues. Bound the value to avoid
> surprises.
>
> Signed-off-by: Tejun Heo <[email protected]>
> Fixes: e08d02aa5fc9 ("blk-iocost: implement Andy's method for donation weight updates")
>
> Signed-off-by: Tejun Heo <[email protected]>
> ---
> Jens, I was flipping between doing max_t(, 1) over the whole divider and
> doing min_t(, WEIGHT_ONE - 1) for hweight_donating. I thought that I as
> testing after the last change but it obviously wasn't and the previous patch
> doesn't compile due to missing type argument. Can you please apply this
> patch instead? I can send an incremental patch if that'd be better. My
> apologies.
Sure, I replaced it. BTW, you had two signed-off-by's in there.
--
Jens Axboe
On Fri, Sep 11, 2020 at 04:43:18PM -0600, Jens Axboe wrote:
> On 9/11/20 4:40 PM, Tejun Heo wrote:
> > Conceptually, root_iocg->hweight_donating must be less than WEIGHT_ONE but
> > all hweight calculations round up and thus it may end up >= WEIGHT_ONE
> > triggering divide-by-zero and other issues. Bound the value to avoid
> > surprises.
> >
> > Signed-off-by: Tejun Heo <[email protected]>
> > Fixes: e08d02aa5fc9 ("blk-iocost: implement Andy's method for donation weight updates")
> >
> > Signed-off-by: Tejun Heo <[email protected]>
> > ---
> > Jens, I was flipping between doing max_t(, 1) over the whole divider and
> > doing min_t(, WEIGHT_ONE - 1) for hweight_donating. I thought that I as
> > testing after the last change but it obviously wasn't and the previous patch
> > doesn't compile due to missing type argument. Can you please apply this
> > patch instead? I can send an incremental patch if that'd be better. My
> > apologies.
>
> Sure, I replaced it. BTW, you had two signed-off-by's in there.
Yeah, sloppy all around. Sorry about that. Gotta slow down more when posting
oh-shit-I-fucked-up stuff.
Thanks.
--
tejun