| Message ID | 1461686131-22999-8-git-send-email-axboe@fb.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
> + return rwb && rwb->wb_normal != 0; > +} > + > +/* > + * Increment 'v', if 'v' is below 'below'. Returns true if we succeeded, > + * false if 'v' + 1 would be bigger than 'below'. > + */ > +static bool atomic_inc_below(atomic_t *v, int below) > +{ > + int cur = atomic_read(v); > + > + for (;;) { > + int old; > + > + if (cur >= below) > + return false; > + old = atomic_cmpxchg(v, cur, cur + 1); > + if (old == cur) > + break; > + cur = old; > + } > + > + return true; > +} > + > +static void wb_timestamp(struct rq_wb *rwb, unsigned long *var) > +{ > + if (rwb_enabled(rwb)) { > + const unsigned long cur = jiffies; > + > + if (cur != *var) > + *var = cur; > + } > +} > + > +void __wbt_done(struct rq_wb *rwb) > +{ > + int inflight, limit = rwb->wb_normal; > + > + /* > + * If the device does write back caching, drop further down > + * before we wake people up. > + */ > + if (rwb->wc && !atomic_read(&rwb->bdi->wb.dirty_sleeping)) > + limit = 0; > + else > + limit = rwb->wb_normal; > + > + /* > + * Don't wake anyone up if we are above the normal limit. If > + * throttling got disabled (limit == 0) with waiters, ensure > + * that we wake them up. > + */ > + inflight = atomic_dec_return(&rwb->inflight); > + if (limit && inflight >= limit) { > + if (!rwb->wb_max) > + wake_up_all(&rwb->wait); > + return; > + } > + Hi Jens, Just a little confused about this. The rwb->wb_max can't be 0 if the variable 'limit' does not equal to 0. So the if (!rwb->wb_max) branch maybe does not make sense. > + if (waitqueue_active(&rwb->wait)) { > + int diff = limit - inflight; > + > + if (!inflight || diff >= rwb->wb_background / 2) > + wake_up_nr(&rwb->wait, 1); > + } > +} > + > +/* > + * Called on completion of a request. Note that it's also called when > + * a request is merged, when the request gets freed. > + */ > +void wbt_done(struct rq_wb *rwb, struct wb_issue_stat *stat) > +{ > + if (!rwb) > + return; > + > + if (!wbt_tracked(stat)) { > + if (rwb->sync_cookie == stat) { > + rwb->sync_issue = 0; > + rwb->sync_cookie = NULL; > + } > + > + wb_timestamp(rwb, &rwb->last_comp); > + } else { > + WARN_ON_ONCE(stat == rwb->sync_cookie); > + __wbt_done(rwb); > + wbt_clear_tracked(stat); > + } > +} > + > +static void calc_wb_limits(struct rq_wb *rwb) > +{ > + unsigned int depth; > + > + if (!rwb->min_lat_nsec) { > + rwb->wb_max = rwb->wb_normal = rwb->wb_background = 0; > + return; > + } > + > + depth = min_t(unsigned int, RWB_MAX_DEPTH, rwb->queue_depth); > + > + /* > + * Reduce max depth by 50%, and re-calculate normal/bg based on that > + */ > + rwb->wb_max = 1 + ((depth - 1) >> min(31U, rwb->scale_step)); > + rwb->wb_normal = (rwb->wb_max + 1) / 2; > + rwb->wb_background = (rwb->wb_max + 3) / 4; > +} > + > +static bool inline stat_sample_valid(struct blk_rq_stat *stat) > +{ > + /* > + * We need at least one read sample, and a minimum of > + * RWB_MIN_WRITE_SAMPLES. We require some write samples to know > + * that it's writes impacting us, and not just some sole read on > + * a device that is in a lower power state. > + */ > + return stat[0].nr_samples >= 1 && > + stat[1].nr_samples >= RWB_MIN_WRITE_SAMPLES; > +} > + -- To unsubscribe from this list: send the line "unsubscribe linux-block" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Tue 26-04-16 09:55:30, Jens Axboe wrote: > We can hook this up to the block layer, to help throttle buffered > writes. Or NFS can tap into it, to accomplish the same. > > wbt registers a few trace points that can be used to track what is > happening in the system: > > wbt_lat: 259:0: latency 2446318 > wbt_stat: 259:0: rmean=2446318, rmin=2446318, rmax=2446318, rsamples=1, > wmean=518866, wmin=15522, wmax=5330353, wsamples=57 > wbt_step: 259:0: step down: step=1, window=72727272, background=8, normal=16, max=32 > > This shows a sync issue event (wbt_lat) that exceeded it's time. wbt_stat > dumps the current read/write stats for that window, and wbt_step shows a > step down event where we now scale back writes. Each trace includes the > device, 259:0 in this case. I have some comments below... > +struct rq_wb { > + /* > + * Settings that govern how we throttle > + */ > + unsigned int wb_background; /* background writeback */ > + unsigned int wb_normal; /* normal writeback */ > + unsigned int wb_max; /* max throughput writeback */ > + unsigned int scale_step; > + > + u64 win_nsec; /* default window size */ > + u64 cur_win_nsec; /* current window size */ > + > + unsigned int unknown_cnt; It would be useful to have a comment here explaining that 'unknown_cnt' is a number of consecutive periods in which we didn't have enough data to decide about queue scaling (at least this is what I understood from the code). > + > + struct timer_list window_timer; > + > + s64 sync_issue; > + void *sync_cookie; So I'm somewhat wondering: What is protecting consistency of this structure? The limits, scale_step, cur_win_nsec, unknown_cnt are updated only from timer so those should be safe. However sync_issue & sync_cookie are accessed from IO submission and completion path and there we need some protection to keep those two in sync. It seems q->queue_lock should mostly achieve those except for blk-mq submission path calling wbt_wait() which doesn't hold queue_lock. It seems you were aware of the possible races and the code handles them mostly fine (although I wouldn't bet too much there is not some weird corner case). However it would be good to comment on this somewhere and explain what the rules for these two fields are. > + > + unsigned int wc; > + unsigned int queue_depth; > + > + unsigned long last_issue; /* last non-throttled issue */ > + unsigned long last_comp; /* last non-throttled comp */ > + unsigned long min_lat_nsec; > + struct backing_dev_info *bdi; > + struct request_queue *q; > + wait_queue_head_t wait; > + atomic_t inflight; > + > + struct wb_stat_ops *stat_ops; > + void *ops_data; > +}; ... > diff --git a/lib/wbt.c b/lib/wbt.c > new file mode 100644 > index 000000000000..650da911f24f > --- /dev/null > +++ b/lib/wbt.c > @@ -0,0 +1,524 @@ > +/* > + * buffered writeback throttling. losely based on CoDel. We can't drop > + * packets for IO scheduling, so the logic is something like this: > + * > + * - Monitor latencies in a defined window of time. > + * - If the minimum latency in the above window exceeds some target, increment > + * scaling step and scale down queue depth by a factor of 2x. The monitoring > + * window is then shrunk to 100 / sqrt(scaling step + 1). > + * - For any window where we don't have solid data on what the latencies > + * look like, retain status quo. > + * - If latencies look good, decrement scaling step. I'm wondering about two things: 1) There is a logic somewhat in this direction in blk_queue_start_tag(). Probably it should be removed after your patches land? 2) As far as I can see in patch 8/8, you have plugged the throttling above the IO scheduler. When there are e.g. multiple cgroups with different IO limits operating, this throttling can lead to strange results (like a cgroup with low limit using up all available background "slots" and thus effectively stopping background writeback for other cgroups)? So won't it make more sense to plug this below the IO scheduler? Now I understand there may be other problems with this but I think we should put more though to that and provide some justification in changelogs. > +static void calc_wb_limits(struct rq_wb *rwb) > +{ > + unsigned int depth; > + > + if (!rwb->min_lat_nsec) { > + rwb->wb_max = rwb->wb_normal = rwb->wb_background = 0; > + return; > + } > + > + depth = min_t(unsigned int, RWB_MAX_DEPTH, rwb->queue_depth); > + > + /* > + * Reduce max depth by 50%, and re-calculate normal/bg based on that > + */ The comment looks a bit out of place here since we don't reduce max depth here. We just use whatever is set in scale_step... > + rwb->wb_max = 1 + ((depth - 1) >> min(31U, rwb->scale_step)); > + rwb->wb_normal = (rwb->wb_max + 1) / 2; > + rwb->wb_background = (rwb->wb_max + 3) / 4; > +} > + > +static bool inline stat_sample_valid(struct blk_rq_stat *stat) > +{ > + /* > + * We need at least one read sample, and a minimum of > + * RWB_MIN_WRITE_SAMPLES. We require some write samples to know > + * that it's writes impacting us, and not just some sole read on > + * a device that is in a lower power state. > + */ > + return stat[0].nr_samples >= 1 && > + stat[1].nr_samples >= RWB_MIN_WRITE_SAMPLES; > +} > + > +static u64 rwb_sync_issue_lat(struct rq_wb *rwb) > +{ > + u64 now, issue = ACCESS_ONCE(rwb->sync_issue); > + > + if (!issue || !rwb->sync_cookie) > + return 0; > + > + now = ktime_to_ns(ktime_get()); > + return now - issue; > +} > + > +enum { > + LAT_OK, > + LAT_UNKNOWN, > + LAT_EXCEEDED, > +}; > + > +static int __latency_exceeded(struct rq_wb *rwb, struct blk_rq_stat *stat) > +{ > + u64 thislat; > + > + /* > + * If our stored sync issue exceeds the window size, or it > + * exceeds our min target AND we haven't logged any entries, > + * flag the latency as exceeded. > + */ > + thislat = rwb_sync_issue_lat(rwb); > + if (thislat > rwb->cur_win_nsec || > + (thislat > rwb->min_lat_nsec && !stat[0].nr_samples)) { > + trace_wbt_lat(rwb->bdi, thislat); > + return LAT_EXCEEDED; > + } So I'm trying to wrap my head around this. If I read the code right, rwb_sync_issue_lat() this returns time that has passed since issuing sync request that is still running. We basically randomly pick which sync request we track as we always start tracking a sync request when some is issued and we are not tracking any at that moment. This is to detect the case when latency of sync IO is very large compared to measurement window so we would not get enough samples to make it valid? Probably the comment could explain more of "why we do this?" than pure "what we do". > + > + if (!stat_sample_valid(stat)) > + return LAT_UNKNOWN; > + > + /* > + * If the 'min' latency exceeds our target, step down. > + */ > + if (stat[0].min > rwb->min_lat_nsec) { > + trace_wbt_lat(rwb->bdi, stat[0].min); > + trace_wbt_stat(rwb->bdi, stat); > + return LAT_EXCEEDED; > + } > + > + if (rwb->scale_step) > + trace_wbt_stat(rwb->bdi, stat); > + > + return LAT_OK; > +} > + Honza
On 04/28/2016 05:05 AM, Jan Kara wrote: > I have some comments below... > >> +struct rq_wb { >> + /* >> + * Settings that govern how we throttle >> + */ >> + unsigned int wb_background; /* background writeback */ >> + unsigned int wb_normal; /* normal writeback */ >> + unsigned int wb_max; /* max throughput writeback */ >> + unsigned int scale_step; >> + >> + u64 win_nsec; /* default window size */ >> + u64 cur_win_nsec; /* current window size */ >> + >> + unsigned int unknown_cnt; > > It would be useful to have a comment here explaining that 'unknown_cnt' is > a number of consecutive periods in which we didn't have enough data to > decide about queue scaling (at least this is what I understood from the > code). Agree, I'll add that comment. >> + >> + struct timer_list window_timer; >> + >> + s64 sync_issue; >> + void *sync_cookie; > > So I'm somewhat wondering: What is protecting consistency of this > structure? The limits, scale_step, cur_win_nsec, unknown_cnt are updated only > from timer so those should be safe. However sync_issue & sync_cookie are > accessed from IO submission and completion path and there we need some > protection to keep those two in sync. It seems q->queue_lock should mostly > achieve those except for blk-mq submission path calling wbt_wait() which > doesn't hold queue_lock. Right, it's designed such that only the timer will be updating these values, and that part is serialized. For sync_issue and sync_cookie, the important part there is that we never dereference sync_cookie. That's why it's a void * now. So we just use it as a hint. And yes, if the IO happens to complete at just the time we are looking at it, we could get a false positive or false negative. That's going to be noise, and nothing we need to worry about. It's deliberate that I don't do any locking for that, the only reason we pass in the queue_lock is to be able to drop it for sleeping. > It seems you were aware of the possible races and the code handles them > mostly fine (although I wouldn't bet too much there is not some weird > corner case). However it would be good to comment on this somewhere and > explain what the rules for these two fields are. Agree, it does warrant a good code comment. If we look at the edge cases, one would be: We look at sync_issue and decide that we're now too late, at the same time as the sync_cookie gets cleared. For this case, we'll count it as an exceed and scale down. In reality we were late, so it doesn't matter. Even if it was the exact time, it's still prudent to scale down as we're going to miss soon. A more worrying case would be two issues that happen at the same time, and only one gets set. Let's assume the one that doesn't get set is the one that ends up taking a long time to complete. We'll miss scaling down in this case, we'll only notice when it completes and shows up in the stats. Not idea, but it's still being handled in the fashion that was originally intended, at completion time. >> diff --git a/lib/wbt.c b/lib/wbt.c >> new file mode 100644 >> index 000000000000..650da911f24f >> --- /dev/null >> +++ b/lib/wbt.c >> @@ -0,0 +1,524 @@ >> +/* >> + * buffered writeback throttling. losely based on CoDel. We can't drop >> + * packets for IO scheduling, so the logic is something like this: >> + * >> + * - Monitor latencies in a defined window of time. >> + * - If the minimum latency in the above window exceeds some target, increment >> + * scaling step and scale down queue depth by a factor of 2x. The monitoring >> + * window is then shrunk to 100 / sqrt(scaling step + 1). >> + * - For any window where we don't have solid data on what the latencies >> + * look like, retain status quo. >> + * - If latencies look good, decrement scaling step. > > I'm wondering about two things: > > 1) There is a logic somewhat in this direction in blk_queue_start_tag(). > Probably it should be removed after your patches land? You're referring to the read/write separation in the legacy tagging? Yes agree, we can kill that once this goes in. > 2) As far as I can see in patch 8/8, you have plugged the throttling above > the IO scheduler. When there are e.g. multiple cgroups with different IO > limits operating, this throttling can lead to strange results (like a > cgroup with low limit using up all available background "slots" and thus > effectively stopping background writeback for other cgroups)? So won't > it make more sense to plug this below the IO scheduler? Now I understand > there may be other problems with this but I think we should put more > though to that and provide some justification in changelogs. One complexity is that we have to do this early for blk-mq, since once you get a request, you're already sitting on the hw tag. CoDel should actually work fine at each hop, so hopefully this will as well. But yes, fairness is something that we have to pay attention to. Right now the wait queue has no priority associated with it, that should probably be improved to be able to wakeup in a more appropriate order. Needs testing, but hopefully it works out since if you do run into starvation, then you'll go to the back of the queue for the next attempt. >> +static void calc_wb_limits(struct rq_wb *rwb) >> +{ >> + unsigned int depth; >> + >> + if (!rwb->min_lat_nsec) { >> + rwb->wb_max = rwb->wb_normal = rwb->wb_background = 0; >> + return; >> + } >> + >> + depth = min_t(unsigned int, RWB_MAX_DEPTH, rwb->queue_depth); >> + >> + /* >> + * Reduce max depth by 50%, and re-calculate normal/bg based on that >> + */ > > The comment looks a bit out of place here since we don't reduce max depth > here. We just use whatever is set in scale_step... True, it does get called for both scaling up and down now. I'll update the comment. >> +static int __latency_exceeded(struct rq_wb *rwb, struct blk_rq_stat *stat) >> +{ >> + u64 thislat; >> + >> + /* >> + * If our stored sync issue exceeds the window size, or it >> + * exceeds our min target AND we haven't logged any entries, >> + * flag the latency as exceeded. >> + */ >> + thislat = rwb_sync_issue_lat(rwb); >> + if (thislat > rwb->cur_win_nsec || >> + (thislat > rwb->min_lat_nsec && !stat[0].nr_samples)) { >> + trace_wbt_lat(rwb->bdi, thislat); >> + return LAT_EXCEEDED; >> + } > > So I'm trying to wrap my head around this. If I read the code right, > rwb_sync_issue_lat() this returns time that has passed since issuing sync > request that is still running. We basically randomly pick which sync > request we track as we always start tracking a sync request when some is > issued and we are not tracking any at that moment. This is to detect the > case when latency of sync IO is very large compared to measurement window > so we would not get enough samples to make it valid? Right, that's pretty close. Since wbt uses the completion latencies to make decisions, if an IO hasn't completed, we don't know about it. If the device is flooded with writes, and we then issue a read, maybe that read won't complete for multiple monitoring windows. During that time, we keep thinking everything is fine. But in reality, it's not completing because of the write load. So this logic attempts to track the single sync IO request case. If that exceeds a monitoring window of time and we saw no other sync IO in that window, then treat that case as if it had completed but exceeded the min latency. And then scale back. We'll always treat a state sample with 1 read as valuable, but for this case, we don't have that sample until it completes. Does that make more sense? > Probably the comment could explain more of "why we do this?" than pure > "what we do". Agree, if you find it confusing, then it needs updating. I'll update the comment.
On 04/28/2016 12:53 PM, Jens Axboe wrote: > >> Probably the comment could explain more of "why we do this?" than pure >> "what we do". > > Agree, if you find it confusing, then it needs updating. I'll update the > comment. http://git.kernel.dk/cgit/linux-block/commit/?h=wb-buf-throttle This should address your review comments, I believe.
On Thu 28-04-16 12:53:50, Jens Axboe wrote: > >2) As far as I can see in patch 8/8, you have plugged the throttling above > > the IO scheduler. When there are e.g. multiple cgroups with different IO > > limits operating, this throttling can lead to strange results (like a > > cgroup with low limit using up all available background "slots" and thus > > effectively stopping background writeback for other cgroups)? So won't > > it make more sense to plug this below the IO scheduler? Now I understand > > there may be other problems with this but I think we should put more > > though to that and provide some justification in changelogs. > > One complexity is that we have to do this early for blk-mq, since once you > get a request, you're already sitting on the hw tag. CoDel should actually > work fine at each hop, so hopefully this will as well. OK, I see. But then this suggests that any IO scheduling and / or cgroup-related throttling should happen before we get a request for blk-mq as well? And then we can still do writeback throttling below that layer? > But yes, fairness is something that we have to pay attention to. Right now > the wait queue has no priority associated with it, that should probably be > improved to be able to wakeup in a more appropriate order. > Needs testing, but hopefully it works out since if you do run into > starvation, then you'll go to the back of the queue for the next attempt. Yeah, once I'll hunt down that regression with old disk, I can have a look into how writeback throttling plays together with blkio-controller. > >>+static int __latency_exceeded(struct rq_wb *rwb, struct blk_rq_stat *stat) > >>+{ > >>+ u64 thislat; > >>+ > >>+ /* > >>+ * If our stored sync issue exceeds the window size, or it > >>+ * exceeds our min target AND we haven't logged any entries, > >>+ * flag the latency as exceeded. > >>+ */ > >>+ thislat = rwb_sync_issue_lat(rwb); > >>+ if (thislat > rwb->cur_win_nsec || > >>+ (thislat > rwb->min_lat_nsec && !stat[0].nr_samples)) { > >>+ trace_wbt_lat(rwb->bdi, thislat); > >>+ return LAT_EXCEEDED; > >>+ } > > > >So I'm trying to wrap my head around this. If I read the code right, > >rwb_sync_issue_lat() this returns time that has passed since issuing sync > >request that is still running. We basically randomly pick which sync > >request we track as we always start tracking a sync request when some is > >issued and we are not tracking any at that moment. This is to detect the > >case when latency of sync IO is very large compared to measurement window > >so we would not get enough samples to make it valid? > > Right, that's pretty close. Since wbt uses the completion latencies to make > decisions, if an IO hasn't completed, we don't know about it. If the device > is flooded with writes, and we then issue a read, maybe that read won't > complete for multiple monitoring windows. During that time, we keep thinking > everything is fine. But in reality, it's not completing because of the write > load. So this logic attempts to track the single sync IO request case. If > that exceeds a monitoring window of time and we saw no other sync IO in that > window, then treat that case as if it had completed but exceeded the min > latency. And then scale back. > > We'll always treat a state sample with 1 read as valuable, but for this > case, we don't have that sample until it completes. > > Does that make more sense? OK, makes sense. Thanks for explanation. Honza
On 05/03/2016 03:34 AM, Jan Kara wrote: > On Thu 28-04-16 12:53:50, Jens Axboe wrote: >>> 2) As far as I can see in patch 8/8, you have plugged the throttling above >>> the IO scheduler. When there are e.g. multiple cgroups with different IO >>> limits operating, this throttling can lead to strange results (like a >>> cgroup with low limit using up all available background "slots" and thus >>> effectively stopping background writeback for other cgroups)? So won't >>> it make more sense to plug this below the IO scheduler? Now I understand >>> there may be other problems with this but I think we should put more >>> though to that and provide some justification in changelogs. >> >> One complexity is that we have to do this early for blk-mq, since once you >> get a request, you're already sitting on the hw tag. CoDel should actually >> work fine at each hop, so hopefully this will as well. > > OK, I see. But then this suggests that any IO scheduling and / or > cgroup-related throttling should happen before we get a request for blk-mq > as well? And then we can still do writeback throttling below that layer? Not necessarily. For IO scheduling, basically we care about two parts: 1) Are you allowed to allocate the resources to queue some IO 2) Are you allowed to dispatch The latter part can still be handled independently, and the former as well of course, wbt just deals with throttling back #1 for buffered writes. >> But yes, fairness is something that we have to pay attention to. Right now >> the wait queue has no priority associated with it, that should probably be >> improved to be able to wakeup in a more appropriate order. >> Needs testing, but hopefully it works out since if you do run into >> starvation, then you'll go to the back of the queue for the next attempt. > > Yeah, once I'll hunt down that regression with old disk, I can have a look > into how writeback throttling plays together with blkio-controller. Thanks!
On Tue 03-05-16 08:23:27, Jens Axboe wrote: > On 05/03/2016 03:34 AM, Jan Kara wrote: > >On Thu 28-04-16 12:53:50, Jens Axboe wrote: > >>>2) As far as I can see in patch 8/8, you have plugged the throttling above > >>> the IO scheduler. When there are e.g. multiple cgroups with different IO > >>> limits operating, this throttling can lead to strange results (like a > >>> cgroup with low limit using up all available background "slots" and thus > >>> effectively stopping background writeback for other cgroups)? So won't > >>> it make more sense to plug this below the IO scheduler? Now I understand > >>> there may be other problems with this but I think we should put more > >>> though to that and provide some justification in changelogs. > >> > >>One complexity is that we have to do this early for blk-mq, since once you > >>get a request, you're already sitting on the hw tag. CoDel should actually > >>work fine at each hop, so hopefully this will as well. > > > >OK, I see. But then this suggests that any IO scheduling and / or > >cgroup-related throttling should happen before we get a request for blk-mq > >as well? And then we can still do writeback throttling below that layer? > > Not necessarily. For IO scheduling, basically we care about two parts: > > 1) Are you allowed to allocate the resources to queue some IO > 2) Are you allowed to dispatch But then it seems suboptimal to waste a relatively scarce resource (which HW tag is AFAIU) just because you happen to run from a cgroup that is bandwidth limited and thus are not allowed to dispatch? Honza
On 05/03/2016 09:22 AM, Jan Kara wrote: > On Tue 03-05-16 08:23:27, Jens Axboe wrote: >> On 05/03/2016 03:34 AM, Jan Kara wrote: >>> On Thu 28-04-16 12:53:50, Jens Axboe wrote: >>>>> 2) As far as I can see in patch 8/8, you have plugged the throttling above >>>>> the IO scheduler. When there are e.g. multiple cgroups with different IO >>>>> limits operating, this throttling can lead to strange results (like a >>>>> cgroup with low limit using up all available background "slots" and thus >>>>> effectively stopping background writeback for other cgroups)? So won't >>>>> it make more sense to plug this below the IO scheduler? Now I understand >>>>> there may be other problems with this but I think we should put more >>>>> though to that and provide some justification in changelogs. >>>> >>>> One complexity is that we have to do this early for blk-mq, since once you >>>> get a request, you're already sitting on the hw tag. CoDel should actually >>>> work fine at each hop, so hopefully this will as well. >>> >>> OK, I see. But then this suggests that any IO scheduling and / or >>> cgroup-related throttling should happen before we get a request for blk-mq >>> as well? And then we can still do writeback throttling below that layer? >> >> Not necessarily. For IO scheduling, basically we care about two parts: >> >> 1) Are you allowed to allocate the resources to queue some IO >> 2) Are you allowed to dispatch > > But then it seems suboptimal to waste a relatively scarce resource (which > HW tag is AFAIU) just because you happen to run from a cgroup that is > bandwidth limited and thus are not allowed to dispatch? For some cases, you are absolutely right, and #1 is the main one. For your case of QD=1, that's obviously the case. For SATA, it's a bit more grey zone, and for others (nvme, scsi, etc), it's not really a scarce resource so #2 is the bigger part of it.
On Tue 03-05-16 11:34:10, Jan Kara wrote: > Yeah, once I'll hunt down that regression with old disk, I can have a look > into how writeback throttling plays together with blkio-controller. So I've tried the following script (note that you need cgroup v2 for writeback IO to be throttled): --- mkdir /sys/fs/cgroup/group1 echo 1000 >/sys/fs/cgroup/group1/io.weight dd if=/dev/zero of=/mnt/file1 bs=1M count=10000& DD1=$! echo $DD1 >/sys/fs/cgroup/group1/cgroup.procs mkdir /sys/fs/cgroup/group2 echo 100 >/sys/fs/cgroup/group2/io.weight #echo "259:65536 wbps=5000000" >/sys/fs/cgroup/group2/io.max echo "259:65536 wbps=max" >/sys/fs/cgroup/group2/io.max dd if=/dev/zero of=/mnt/file2 bs=1M count=10000& DD2=$! echo $DD2 >/sys/fs/cgroup/group2/cgroup.procs while true; do sleep 1 kill -USR1 $DD1 kill -USR1 $DD2 echo '=======================================================' done --- and watched the progress of the dd processes in different cgroups. The 1/10 weight difference has no effect with your writeback patches - the situation after one minute: 3120+1 records in 3120+1 records out 3272392704 bytes (3.3 GB) copied, 63.7119 s, 51.4 MB/s 3217+1 records in 3217+1 records out 3374010368 bytes (3.4 GB) copied, 63.5819 s, 53.1 MB/s I should add that even without your patches the progress doesn't quite correspond to the weight ratio: ... but still there is noticeable difference to cgroups with different weights. OTOH blk-throttle combines well with your patches: Limiting one cgroup to 5 M/s results in numbers like: 3883+2 records in 3883+2 records out 4072091648 bytes (4.1 GB) copied, 36.6713 s, 111 MB/s 413+0 records in 413+0 records out 433061888 bytes (433 MB) copied, 36.8939 s, 11.7 MB/s which is fine and comparable with unpatched kernel. Higher throughput number is because we do buffered writes and dd reports what it wrote into page cache. And there is no wonder blk-throttle combines fine - it throttles bios which happens before we reach writeback throttling mechanism. So I belive this demonstrates that your writeback throttling just doesn't work well with selective scheduling policy that happens below it because it can essentially lead to IO priority inversion issues... Honza
On Tue 03-05-16 17:40:32, Jan Kara wrote: > On Tue 03-05-16 11:34:10, Jan Kara wrote: > > Yeah, once I'll hunt down that regression with old disk, I can have a look > > into how writeback throttling plays together with blkio-controller. > > So I've tried the following script (note that you need cgroup v2 for > writeback IO to be throttled): > > --- > mkdir /sys/fs/cgroup/group1 > echo 1000 >/sys/fs/cgroup/group1/io.weight > dd if=/dev/zero of=/mnt/file1 bs=1M count=10000& > DD1=$! > echo $DD1 >/sys/fs/cgroup/group1/cgroup.procs > > mkdir /sys/fs/cgroup/group2 > echo 100 >/sys/fs/cgroup/group2/io.weight > #echo "259:65536 wbps=5000000" >/sys/fs/cgroup/group2/io.max > echo "259:65536 wbps=max" >/sys/fs/cgroup/group2/io.max > dd if=/dev/zero of=/mnt/file2 bs=1M count=10000& > DD2=$! > echo $DD2 >/sys/fs/cgroup/group2/cgroup.procs > > while true; do > sleep 1 > kill -USR1 $DD1 > kill -USR1 $DD2 > echo '=======================================================' > done > --- > > and watched the progress of the dd processes in different cgroups. The 1/10 > weight difference has no effect with your writeback patches - the situation > after one minute: > > 3120+1 records in > 3120+1 records out > 3272392704 bytes (3.3 GB) copied, 63.7119 s, 51.4 MB/s > 3217+1 records in > 3217+1 records out > 3374010368 bytes (3.4 GB) copied, 63.5819 s, 53.1 MB/s > > I should add that even without your patches the progress doesn't quite > correspond to the weight ratio: Forgot to fill in corresponding data for unpatched kernel here: 5962+2 records in 5962+2 records out 6252281856 bytes (6.3 GB) copied, 64.1719 s, 97.4 MB/s 1502+0 records in 1502+0 records out 1574961152 bytes (1.6 GB) copied, 64.207 s, 24.5 MB/s > but still there is noticeable difference to cgroups with different weights. > > OTOH blk-throttle combines well with your patches: Limiting one cgroup to > 5 M/s results in numbers like: > > 3883+2 records in > 3883+2 records out > 4072091648 bytes (4.1 GB) copied, 36.6713 s, 111 MB/s > 413+0 records in > 413+0 records out > 433061888 bytes (433 MB) copied, 36.8939 s, 11.7 MB/s > > which is fine and comparable with unpatched kernel. Higher throughput > number is because we do buffered writes and dd reports what it wrote into > page cache. And there is no wonder blk-throttle combines fine - it > throttles bios which happens before we reach writeback throttling > mechanism. > > So I belive this demonstrates that your writeback throttling just doesn't > work well with selective scheduling policy that happens below it because it > can essentially lead to IO priority inversion issues... > > Honza > -- > Jan Kara <jack@suse.com> > SUSE Labs, CR
On 05/03/2016 09:48 AM, Jan Kara wrote: > On Tue 03-05-16 17:40:32, Jan Kara wrote: >> On Tue 03-05-16 11:34:10, Jan Kara wrote: >>> Yeah, once I'll hunt down that regression with old disk, I can have a look >>> into how writeback throttling plays together with blkio-controller. >> >> So I've tried the following script (note that you need cgroup v2 for >> writeback IO to be throttled): >> >> --- >> mkdir /sys/fs/cgroup/group1 >> echo 1000 >/sys/fs/cgroup/group1/io.weight >> dd if=/dev/zero of=/mnt/file1 bs=1M count=10000& >> DD1=$! >> echo $DD1 >/sys/fs/cgroup/group1/cgroup.procs >> >> mkdir /sys/fs/cgroup/group2 >> echo 100 >/sys/fs/cgroup/group2/io.weight >> #echo "259:65536 wbps=5000000" >/sys/fs/cgroup/group2/io.max >> echo "259:65536 wbps=max" >/sys/fs/cgroup/group2/io.max >> dd if=/dev/zero of=/mnt/file2 bs=1M count=10000& >> DD2=$! >> echo $DD2 >/sys/fs/cgroup/group2/cgroup.procs >> >> while true; do >> sleep 1 >> kill -USR1 $DD1 >> kill -USR1 $DD2 >> echo '=======================================================' >> done >> --- >> >> and watched the progress of the dd processes in different cgroups. The 1/10 >> weight difference has no effect with your writeback patches - the situation >> after one minute: >> >> 3120+1 records in >> 3120+1 records out >> 3272392704 bytes (3.3 GB) copied, 63.7119 s, 51.4 MB/s >> 3217+1 records in >> 3217+1 records out >> 3374010368 bytes (3.4 GB) copied, 63.5819 s, 53.1 MB/s >> >> I should add that even without your patches the progress doesn't quite >> correspond to the weight ratio: > > Forgot to fill in corresponding data for unpatched kernel here: > > 5962+2 records in > 5962+2 records out > 6252281856 bytes (6.3 GB) copied, 64.1719 s, 97.4 MB/s > 1502+0 records in > 1502+0 records out > 1574961152 bytes (1.6 GB) copied, 64.207 s, 24.5 MB/s Thanks for testing this, I'll see what we can do about that. It stands to reason that we'll throttle a heavier writer more, statistically. But I'm assuming this above test was run basically with just the writes going, so no real competition? And hence we end up throttling them equally much, destroying the weighting in the process. But for both cases, we basically don't pay any attention to cgroup weights. >> but still there is noticeable difference to cgroups with different weights. >> >> OTOH blk-throttle combines well with your patches: Limiting one cgroup to >> 5 M/s results in numbers like: >> >> 3883+2 records in >> 3883+2 records out >> 4072091648 bytes (4.1 GB) copied, 36.6713 s, 111 MB/s >> 413+0 records in >> 413+0 records out >> 433061888 bytes (433 MB) copied, 36.8939 s, 11.7 MB/s >> >> which is fine and comparable with unpatched kernel. Higher throughput >> number is because we do buffered writes and dd reports what it wrote into >> page cache. And there is no wonder blk-throttle combines fine - it >> throttles bios which happens before we reach writeback throttling >> mechanism. OK, that's good, at least that part works fine. And yes, the throttle path is hit before we end up in the make_request_fn, which is where wbt drops in. >> So I belive this demonstrates that your writeback throttling just doesn't >> work well with selective scheduling policy that happens below it because it >> can essentially lead to IO priority inversion issues... It this testing still done on the QD=1 ATA disk? Not too surprising that this falls apart, since we have very little room to maneuver. I wonder if a normal SATA with NCQ would behave better in this regard. I'll have to test a bit and think about how we can best handle this case.
On 05/03/2016 10:59 AM, Jens Axboe wrote: > On 05/03/2016 09:48 AM, Jan Kara wrote: >> On Tue 03-05-16 17:40:32, Jan Kara wrote: >>> On Tue 03-05-16 11:34:10, Jan Kara wrote: >>>> Yeah, once I'll hunt down that regression with old disk, I can have >>>> a look >>>> into how writeback throttling plays together with blkio-controller. >>> >>> So I've tried the following script (note that you need cgroup v2 for >>> writeback IO to be throttled): >>> >>> --- >>> mkdir /sys/fs/cgroup/group1 >>> echo 1000 >/sys/fs/cgroup/group1/io.weight >>> dd if=/dev/zero of=/mnt/file1 bs=1M count=10000& >>> DD1=$! >>> echo $DD1 >/sys/fs/cgroup/group1/cgroup.procs >>> >>> mkdir /sys/fs/cgroup/group2 >>> echo 100 >/sys/fs/cgroup/group2/io.weight >>> #echo "259:65536 wbps=5000000" >/sys/fs/cgroup/group2/io.max >>> echo "259:65536 wbps=max" >/sys/fs/cgroup/group2/io.max >>> dd if=/dev/zero of=/mnt/file2 bs=1M count=10000& >>> DD2=$! >>> echo $DD2 >/sys/fs/cgroup/group2/cgroup.procs >>> >>> while true; do >>> sleep 1 >>> kill -USR1 $DD1 >>> kill -USR1 $DD2 >>> echo '=======================================================' >>> done >>> --- >>> >>> and watched the progress of the dd processes in different cgroups. >>> The 1/10 >>> weight difference has no effect with your writeback patches - the >>> situation >>> after one minute: >>> >>> 3120+1 records in >>> 3120+1 records out >>> 3272392704 bytes (3.3 GB) copied, 63.7119 s, 51.4 MB/s >>> 3217+1 records in >>> 3217+1 records out >>> 3374010368 bytes (3.4 GB) copied, 63.5819 s, 53.1 MB/s >>> >>> I should add that even without your patches the progress doesn't quite >>> correspond to the weight ratio: >> >> Forgot to fill in corresponding data for unpatched kernel here: >> >> 5962+2 records in >> 5962+2 records out >> 6252281856 bytes (6.3 GB) copied, 64.1719 s, 97.4 MB/s >> 1502+0 records in >> 1502+0 records out >> 1574961152 bytes (1.6 GB) copied, 64.207 s, 24.5 MB/s > > Thanks for testing this, I'll see what we can do about that. It stands > to reason that we'll throttle a heavier writer more, statistically. But > I'm assuming this above test was run basically with just the writes > going, so no real competition? And hence we end up throttling them > equally much, destroying the weighting in the process. But for both > cases, we basically don't pay any attention to cgroup weights. > >>> but still there is noticeable difference to cgroups with different >>> weights. >>> >>> OTOH blk-throttle combines well with your patches: Limiting one >>> cgroup to >>> 5 M/s results in numbers like: >>> >>> 3883+2 records in >>> 3883+2 records out >>> 4072091648 bytes (4.1 GB) copied, 36.6713 s, 111 MB/s >>> 413+0 records in >>> 413+0 records out >>> 433061888 bytes (433 MB) copied, 36.8939 s, 11.7 MB/s >>> >>> which is fine and comparable with unpatched kernel. Higher throughput >>> number is because we do buffered writes and dd reports what it wrote >>> into >>> page cache. And there is no wonder blk-throttle combines fine - it >>> throttles bios which happens before we reach writeback throttling >>> mechanism. > > OK, that's good, at least that part works fine. And yes, the throttle > path is hit before we end up in the make_request_fn, which is where wbt > drops in. > >>> So I belive this demonstrates that your writeback throttling just >>> doesn't >>> work well with selective scheduling policy that happens below it >>> because it >>> can essentially lead to IO priority inversion issues... > > It this testing still done on the QD=1 ATA disk? Not too surprising that > this falls apart, since we have very little room to maneuver. I wonder > if a normal SATA with NCQ would behave better in this regard. I'll have > to test a bit and think about how we can best handle this case. I think what we'll do for now is just disable wbt IFF we have a non-root cgroup attached to CFQ. Done here: http://git.kernel.dk/cgit/linux-block/commit/?h=wb-buf-throttle&id=7315756efe76bbdf83076fc9dbc569bbb4da5d32 We don't have a strong need for wbt (supposedly) since CFQ should take care of most of it, if you have policies set for proportional sharing. Longer term it's not a concern either, as we'll move away from that model anyway.
On 05/03/2016 12:14 PM, Jens Axboe wrote: > On 05/03/2016 10:59 AM, Jens Axboe wrote: >> On 05/03/2016 09:48 AM, Jan Kara wrote: >>> On Tue 03-05-16 17:40:32, Jan Kara wrote: >>>> On Tue 03-05-16 11:34:10, Jan Kara wrote: >>>>> Yeah, once I'll hunt down that regression with old disk, I can have >>>>> a look >>>>> into how writeback throttling plays together with blkio-controller. >>>> >>>> So I've tried the following script (note that you need cgroup v2 for >>>> writeback IO to be throttled): >>>> >>>> --- >>>> mkdir /sys/fs/cgroup/group1 >>>> echo 1000 >/sys/fs/cgroup/group1/io.weight >>>> dd if=/dev/zero of=/mnt/file1 bs=1M count=10000& >>>> DD1=$! >>>> echo $DD1 >/sys/fs/cgroup/group1/cgroup.procs >>>> >>>> mkdir /sys/fs/cgroup/group2 >>>> echo 100 >/sys/fs/cgroup/group2/io.weight >>>> #echo "259:65536 wbps=5000000" >/sys/fs/cgroup/group2/io.max >>>> echo "259:65536 wbps=max" >/sys/fs/cgroup/group2/io.max >>>> dd if=/dev/zero of=/mnt/file2 bs=1M count=10000& >>>> DD2=$! >>>> echo $DD2 >/sys/fs/cgroup/group2/cgroup.procs >>>> >>>> while true; do >>>> sleep 1 >>>> kill -USR1 $DD1 >>>> kill -USR1 $DD2 >>>> echo >>>> '=======================================================' >>>> done >>>> --- >>>> >>>> and watched the progress of the dd processes in different cgroups. >>>> The 1/10 >>>> weight difference has no effect with your writeback patches - the >>>> situation >>>> after one minute: >>>> >>>> 3120+1 records in >>>> 3120+1 records out >>>> 3272392704 bytes (3.3 GB) copied, 63.7119 s, 51.4 MB/s >>>> 3217+1 records in >>>> 3217+1 records out >>>> 3374010368 bytes (3.4 GB) copied, 63.5819 s, 53.1 MB/s >>>> >>>> I should add that even without your patches the progress doesn't quite >>>> correspond to the weight ratio: >>> >>> Forgot to fill in corresponding data for unpatched kernel here: >>> >>> 5962+2 records in >>> 5962+2 records out >>> 6252281856 bytes (6.3 GB) copied, 64.1719 s, 97.4 MB/s >>> 1502+0 records in >>> 1502+0 records out >>> 1574961152 bytes (1.6 GB) copied, 64.207 s, 24.5 MB/s >> >> Thanks for testing this, I'll see what we can do about that. It stands >> to reason that we'll throttle a heavier writer more, statistically. But >> I'm assuming this above test was run basically with just the writes >> going, so no real competition? And hence we end up throttling them >> equally much, destroying the weighting in the process. But for both >> cases, we basically don't pay any attention to cgroup weights. >> >>>> but still there is noticeable difference to cgroups with different >>>> weights. >>>> >>>> OTOH blk-throttle combines well with your patches: Limiting one >>>> cgroup to >>>> 5 M/s results in numbers like: >>>> >>>> 3883+2 records in >>>> 3883+2 records out >>>> 4072091648 bytes (4.1 GB) copied, 36.6713 s, 111 MB/s >>>> 413+0 records in >>>> 413+0 records out >>>> 433061888 bytes (433 MB) copied, 36.8939 s, 11.7 MB/s >>>> >>>> which is fine and comparable with unpatched kernel. Higher throughput >>>> number is because we do buffered writes and dd reports what it wrote >>>> into >>>> page cache. And there is no wonder blk-throttle combines fine - it >>>> throttles bios which happens before we reach writeback throttling >>>> mechanism. >> >> OK, that's good, at least that part works fine. And yes, the throttle >> path is hit before we end up in the make_request_fn, which is where wbt >> drops in. >> >>>> So I belive this demonstrates that your writeback throttling just >>>> doesn't >>>> work well with selective scheduling policy that happens below it >>>> because it >>>> can essentially lead to IO priority inversion issues... >> >> It this testing still done on the QD=1 ATA disk? Not too surprising that >> this falls apart, since we have very little room to maneuver. I wonder >> if a normal SATA with NCQ would behave better in this regard. I'll have >> to test a bit and think about how we can best handle this case. > > I think what we'll do for now is just disable wbt IFF we have a non-root > cgroup attached to CFQ. Done here: > > http://git.kernel.dk/cgit/linux-block/commit/?h=wb-buf-throttle&id=7315756efe76bbdf83076fc9dbc569bbb4da5d32 That was a bit too untested.. This should be better, it taps into where cfq normally notices a difference in blkcg: http://git.kernel.dk/cgit/linux-block/commit/?h=wb-buf-throttle&id=9b89e1bb666bd036a4cb1313479435087fb86ba0
diff --git a/include/linux/wbt.h b/include/linux/wbt.h new file mode 100644 index 000000000000..c8a12795416b --- /dev/null +++ b/include/linux/wbt.h @@ -0,0 +1,95 @@ +#ifndef WB_THROTTLE_H +#define WB_THROTTLE_H + +#include <linux/atomic.h> +#include <linux/wait.h> +#include <linux/timer.h> +#include <linux/ktime.h> + +#define ISSUE_STAT_MASK (1ULL << 63) +#define ISSUE_STAT_TIME_MASK ~ISSUE_STAT_MASK + +struct wb_issue_stat { + u64 time; +}; + +static inline void wbt_issue_stat_set_time(struct wb_issue_stat *stat) +{ + stat->time = (stat->time & ISSUE_STAT_MASK) | + (ktime_to_ns(ktime_get()) & ISSUE_STAT_TIME_MASK); +} + +static inline u64 wbt_issue_stat_get_time(struct wb_issue_stat *stat) +{ + return stat->time & ISSUE_STAT_TIME_MASK; +} + +static inline void wbt_mark_tracked(struct wb_issue_stat *stat) +{ + stat->time |= ISSUE_STAT_MASK; +} + +static inline void wbt_clear_tracked(struct wb_issue_stat *stat) +{ + stat->time &= ~ISSUE_STAT_MASK; +} + +static inline bool wbt_tracked(struct wb_issue_stat *stat) +{ + return (stat->time & ISSUE_STAT_MASK) != 0; +} + +struct wb_stat_ops { + void (*get)(void *, struct blk_rq_stat *); + void (*clear)(void *); +}; + +struct rq_wb { + /* + * Settings that govern how we throttle + */ + unsigned int wb_background; /* background writeback */ + unsigned int wb_normal; /* normal writeback */ + unsigned int wb_max; /* max throughput writeback */ + unsigned int scale_step; + + u64 win_nsec; /* default window size */ + u64 cur_win_nsec; /* current window size */ + + unsigned int unknown_cnt; + + struct timer_list window_timer; + + s64 sync_issue; + void *sync_cookie; + + unsigned int wc; + unsigned int queue_depth; + + unsigned long last_issue; /* last non-throttled issue */ + unsigned long last_comp; /* last non-throttled comp */ + unsigned long min_lat_nsec; + struct backing_dev_info *bdi; + struct request_queue *q; + wait_queue_head_t wait; + atomic_t inflight; + + struct wb_stat_ops *stat_ops; + void *ops_data; +}; + +struct backing_dev_info; + +void __wbt_done(struct rq_wb *); +void wbt_done(struct rq_wb *, struct wb_issue_stat *); +bool wbt_wait(struct rq_wb *, unsigned int, spinlock_t *); +struct rq_wb *wbt_init(struct backing_dev_info *, struct wb_stat_ops *, void *); +void wbt_exit(struct rq_wb *); +void wbt_update_limits(struct rq_wb *); +void wbt_requeue(struct rq_wb *, struct wb_issue_stat *); +void wbt_issue(struct rq_wb *, struct wb_issue_stat *); + +void wbt_set_queue_depth(struct rq_wb *, unsigned int); +void wbt_set_write_cache(struct rq_wb *, bool); + +#endif diff --git a/include/trace/events/wbt.h b/include/trace/events/wbt.h new file mode 100644 index 000000000000..a4b8b2e57bb1 --- /dev/null +++ b/include/trace/events/wbt.h @@ -0,0 +1,122 @@ +#undef TRACE_SYSTEM +#define TRACE_SYSTEM wbt + +#if !defined(_TRACE_WBT_H) || defined(TRACE_HEADER_MULTI_READ) +#define _TRACE_WBT_H + +#include <linux/tracepoint.h> +#include <linux/wbt.h> + +/** + * wbt_stat - trace stats for blk_wb + * @stat: array of read/write stats + */ +TRACE_EVENT(wbt_stat, + + TP_PROTO(struct backing_dev_info *bdi, struct blk_rq_stat *stat), + + TP_ARGS(bdi, stat), + + TP_STRUCT__entry( + __array(char, name, 32) + __field(s64, rmean) + __field(u64, rmin) + __field(u64, rmax) + __field(s64, rnr_samples) + __field(s64, rtime) + __field(s64, wmean) + __field(u64, wmin) + __field(u64, wmax) + __field(s64, wnr_samples) + __field(s64, wtime) + ), + + TP_fast_assign( + strncpy(__entry->name, dev_name(bdi->dev), 32); + __entry->rmean = stat[0].mean; + __entry->rmin = stat[0].min; + __entry->rmax = stat[0].max; + __entry->rnr_samples = stat[0].nr_samples; + __entry->wmean = stat[1].mean; + __entry->wmin = stat[1].min; + __entry->wmax = stat[1].max; + __entry->wnr_samples = stat[1].nr_samples; + ), + + TP_printk("%s: rmean=%llu, rmin=%llu, rmax=%llu, rsamples=%llu, " + "wmean=%llu, wmin=%llu, wmax=%llu, wsamples=%llu\n", + __entry->name, __entry->rmean, __entry->rmin, __entry->rmax, + __entry->rnr_samples, __entry->wmean, __entry->wmin, + __entry->wmax, __entry->wnr_samples) +); + +/** + * wbt_lat - trace latency event + * @lat: latency trigger + */ +TRACE_EVENT(wbt_lat, + + TP_PROTO(struct backing_dev_info *bdi, unsigned long lat), + + TP_ARGS(bdi, lat), + + TP_STRUCT__entry( + __array(char, name, 32) + __field(unsigned long, lat) + ), + + TP_fast_assign( + strncpy(__entry->name, dev_name(bdi->dev), 32); + __entry->lat = lat; + ), + + TP_printk("%s: latency %llu\n", __entry->name, + (unsigned long long) __entry->lat) +); + +/** + * wbt_step - trace wb event step + * @msg: context message + * @step: the current scale step count + * @window: the current monitoring window + * @bg: the current background queue limit + * @normal: the current normal writeback limit + * @max: the current max throughput writeback limit + */ +TRACE_EVENT(wbt_step, + + TP_PROTO(struct backing_dev_info *bdi, const char *msg, + unsigned int step, unsigned long window, unsigned int bg, + unsigned int normal, unsigned int max), + + TP_ARGS(bdi, msg, step, window, bg, normal, max), + + TP_STRUCT__entry( + __array(char, name, 32) + __field(const char *, msg) + __field(unsigned int, step) + __field(unsigned long, window) + __field(unsigned int, bg) + __field(unsigned int, normal) + __field(unsigned int, max) + ), + + TP_fast_assign( + strncpy(__entry->name, dev_name(bdi->dev), 32); + __entry->msg = msg; + __entry->step = step; + __entry->window = window; + __entry->bg = bg; + __entry->normal = normal; + __entry->max = max; + ), + + TP_printk("%s: %s: step=%u, window=%lu, background=%u, normal=%u, max=%u\n", + __entry->name, __entry->msg, __entry->step, __entry->window, + __entry->bg, __entry->normal, __entry->max) +); + +#endif /* _TRACE_WBT_H */ + +/* This part must be outside protection */ +#include <trace/define_trace.h> diff --git a/lib/Kconfig b/lib/Kconfig index 3cca1222578e..01da47cb9766 100644 --- a/lib/Kconfig +++ b/lib/Kconfig @@ -540,4 +540,7 @@ config STACKDEPOT bool select STACKTRACE +config WBT + bool + endmenu diff --git a/lib/Makefile b/lib/Makefile index 7bd6fd436c97..15366777a1d4 100644 --- a/lib/Makefile +++ b/lib/Makefile @@ -180,6 +180,7 @@ obj-$(CONFIG_GENERIC_NET_UTILS) += net_utils.o obj-$(CONFIG_SG_SPLIT) += sg_split.o obj-$(CONFIG_STMP_DEVICE) += stmp_device.o obj-$(CONFIG_IRQ_POLL) += irq_poll.o +obj-$(CONFIG_WBT) += wbt.o obj-$(CONFIG_STACKDEPOT) += stackdepot.o KASAN_SANITIZE_stackdepot.o := n diff --git a/lib/wbt.c b/lib/wbt.c new file mode 100644 index 000000000000..650da911f24f --- /dev/null +++ b/lib/wbt.c @@ -0,0 +1,524 @@ +/* + * buffered writeback throttling. losely based on CoDel. We can't drop + * packets for IO scheduling, so the logic is something like this: + * + * - Monitor latencies in a defined window of time. + * - If the minimum latency in the above window exceeds some target, increment + * scaling step and scale down queue depth by a factor of 2x. The monitoring + * window is then shrunk to 100 / sqrt(scaling step + 1). + * - For any window where we don't have solid data on what the latencies + * look like, retain status quo. + * - If latencies look good, decrement scaling step. + * + * Copyright (C) 2016 Jens Axboe + * + * Things that (may) need changing: + * + * - Different scaling of background/normal/high priority writeback. + * We may have to violate guarantees for max. + * - We can have mismatches between the stat window and our window. + * + */ +#include <linux/kernel.h> +#include <linux/blk_types.h> +#include <linux/slab.h> +#include <linux/backing-dev.h> +#include <linux/wbt.h> + +#define CREATE_TRACE_POINTS +#include <trace/events/wbt.h> + +enum { + /* + * Might need to be higher + */ + RWB_MAX_DEPTH = 64, + + /* + * 100msec window + */ + RWB_WINDOW_NSEC = 100 * 1000 * 1000ULL, + + /* + * Disregard stats, if we don't meet these minimums + */ + RWB_MIN_WRITE_SAMPLES = 3, + RWB_MIN_READ_SAMPLES = 1, + + RWB_UNKNOWN_BUMP = 5, +}; + +static inline bool rwb_enabled(struct rq_wb *rwb) +{ + return rwb && rwb->wb_normal != 0; +} + +/* + * Increment 'v', if 'v' is below 'below'. Returns true if we succeeded, + * false if 'v' + 1 would be bigger than 'below'. + */ +static bool atomic_inc_below(atomic_t *v, int below) +{ + int cur = atomic_read(v); + + for (;;) { + int old; + + if (cur >= below) + return false; + old = atomic_cmpxchg(v, cur, cur + 1); + if (old == cur) + break; + cur = old; + } + + return true; +} + +static void wb_timestamp(struct rq_wb *rwb, unsigned long *var) +{ + if (rwb_enabled(rwb)) { + const unsigned long cur = jiffies; + + if (cur != *var) + *var = cur; + } +} + +void __wbt_done(struct rq_wb *rwb) +{ + int inflight, limit = rwb->wb_normal; + + /* + * If the device does write back caching, drop further down + * before we wake people up. + */ + if (rwb->wc && !atomic_read(&rwb->bdi->wb.dirty_sleeping)) + limit = 0; + else + limit = rwb->wb_normal; + + /* + * Don't wake anyone up if we are above the normal limit. If + * throttling got disabled (limit == 0) with waiters, ensure + * that we wake them up. + */ + inflight = atomic_dec_return(&rwb->inflight); + if (limit && inflight >= limit) { + if (!rwb->wb_max) + wake_up_all(&rwb->wait); + return; + } + + if (waitqueue_active(&rwb->wait)) { + int diff = limit - inflight; + + if (!inflight || diff >= rwb->wb_background / 2) + wake_up_nr(&rwb->wait, 1); + } +} + +/* + * Called on completion of a request. Note that it's also called when + * a request is merged, when the request gets freed. + */ +void wbt_done(struct rq_wb *rwb, struct wb_issue_stat *stat) +{ + if (!rwb) + return; + + if (!wbt_tracked(stat)) { + if (rwb->sync_cookie == stat) { + rwb->sync_issue = 0; + rwb->sync_cookie = NULL; + } + + wb_timestamp(rwb, &rwb->last_comp); + } else { + WARN_ON_ONCE(stat == rwb->sync_cookie); + __wbt_done(rwb); + wbt_clear_tracked(stat); + } +} + +static void calc_wb_limits(struct rq_wb *rwb) +{ + unsigned int depth; + + if (!rwb->min_lat_nsec) { + rwb->wb_max = rwb->wb_normal = rwb->wb_background = 0; + return; + } + + depth = min_t(unsigned int, RWB_MAX_DEPTH, rwb->queue_depth); + + /* + * Reduce max depth by 50%, and re-calculate normal/bg based on that + */ + rwb->wb_max = 1 + ((depth - 1) >> min(31U, rwb->scale_step)); + rwb->wb_normal = (rwb->wb_max + 1) / 2; + rwb->wb_background = (rwb->wb_max + 3) / 4; +} + +static bool inline stat_sample_valid(struct blk_rq_stat *stat) +{ + /* + * We need at least one read sample, and a minimum of + * RWB_MIN_WRITE_SAMPLES. We require some write samples to know + * that it's writes impacting us, and not just some sole read on + * a device that is in a lower power state. + */ + return stat[0].nr_samples >= 1 && + stat[1].nr_samples >= RWB_MIN_WRITE_SAMPLES; +} + +static u64 rwb_sync_issue_lat(struct rq_wb *rwb) +{ + u64 now, issue = ACCESS_ONCE(rwb->sync_issue); + + if (!issue || !rwb->sync_cookie) + return 0; + + now = ktime_to_ns(ktime_get()); + return now - issue; +} + +enum { + LAT_OK, + LAT_UNKNOWN, + LAT_EXCEEDED, +}; + +static int __latency_exceeded(struct rq_wb *rwb, struct blk_rq_stat *stat) +{ + u64 thislat; + + /* + * If our stored sync issue exceeds the window size, or it + * exceeds our min target AND we haven't logged any entries, + * flag the latency as exceeded. + */ + thislat = rwb_sync_issue_lat(rwb); + if (thislat > rwb->cur_win_nsec || + (thislat > rwb->min_lat_nsec && !stat[0].nr_samples)) { + trace_wbt_lat(rwb->bdi, thislat); + return LAT_EXCEEDED; + } + + if (!stat_sample_valid(stat)) + return LAT_UNKNOWN; + + /* + * If the 'min' latency exceeds our target, step down. + */ + if (stat[0].min > rwb->min_lat_nsec) { + trace_wbt_lat(rwb->bdi, stat[0].min); + trace_wbt_stat(rwb->bdi, stat); + return LAT_EXCEEDED; + } + + if (rwb->scale_step) + trace_wbt_stat(rwb->bdi, stat); + + return LAT_OK; +} + +static int latency_exceeded(struct rq_wb *rwb) +{ + struct blk_rq_stat stat[2]; + + rwb->stat_ops->get(rwb->ops_data, stat); + return __latency_exceeded(rwb, stat); +} + +static void rwb_trace_step(struct rq_wb *rwb, const char *msg) +{ + trace_wbt_step(rwb->bdi, msg, rwb->scale_step, rwb->cur_win_nsec, + rwb->wb_background, rwb->wb_normal, rwb->wb_max); +} + +static void scale_up(struct rq_wb *rwb) +{ + /* + * If we're at 0, we can't go lower. + */ + if (!rwb->scale_step) + return; + + rwb->scale_step--; + rwb->unknown_cnt = 0; + rwb->stat_ops->clear(rwb->ops_data); + calc_wb_limits(rwb); + + if (waitqueue_active(&rwb->wait)) + wake_up_all(&rwb->wait); + + rwb_trace_step(rwb, "step up"); +} + +static void scale_down(struct rq_wb *rwb) +{ + /* + * Stop scaling down when we've hit the limit. This also prevents + * ->scale_step from going to crazy values, if the device can't + * keep up. + */ + if (rwb->wb_max == 1) + return; + + rwb->scale_step++; + rwb->unknown_cnt = 0; + rwb->stat_ops->clear(rwb->ops_data); + calc_wb_limits(rwb); + rwb_trace_step(rwb, "step down"); +} + +static void rwb_arm_timer(struct rq_wb *rwb) +{ + unsigned long expires; + + /* + * We should speed this up, using some variant of a fast integer + * inverse square root calculation. Since we only do this for + * every window expiration, it's not a huge deal, though. + */ + rwb->cur_win_nsec = div_u64(rwb->win_nsec << 4, + int_sqrt((rwb->scale_step + 1) << 8)); + expires = jiffies + nsecs_to_jiffies(rwb->cur_win_nsec); + mod_timer(&rwb->window_timer, expires); +} + +static void wb_timer_fn(unsigned long data) +{ + struct rq_wb *rwb = (struct rq_wb *) data; + int status; + + /* + * If we exceeded the latency target, step down. If we did not, + * step one level up. If we don't know enough to say either exceeded + * or ok, then don't do anything. + */ + status = latency_exceeded(rwb); + switch (status) { + case LAT_EXCEEDED: + scale_down(rwb); + break; + case LAT_OK: + scale_up(rwb); + break; + case LAT_UNKNOWN: + /* + * We had no read samples, start bumping up the write + * depth slowly + */ + if (++rwb->unknown_cnt >= RWB_UNKNOWN_BUMP) + scale_up(rwb); + break; + default: + break; + } + + /* + * Re-arm timer, if we have IO in flight + */ + if (rwb->scale_step || atomic_read(&rwb->inflight)) + rwb_arm_timer(rwb); +} + +void wbt_update_limits(struct rq_wb *rwb) +{ + rwb->scale_step = 0; + calc_wb_limits(rwb); + + if (waitqueue_active(&rwb->wait)) + wake_up_all(&rwb->wait); +} + +static bool close_io(struct rq_wb *rwb) +{ + const unsigned long now = jiffies; + + return time_before(now, rwb->last_issue + HZ / 10) || + time_before(now, rwb->last_comp + HZ / 10); +} + +#define REQ_HIPRIO (REQ_SYNC | REQ_META | REQ_PRIO) + +static inline unsigned int get_limit(struct rq_wb *rwb, unsigned long rw) +{ + unsigned int limit; + + /* + * At this point we know it's a buffered write. If REQ_SYNC is + * set, then it's WB_SYNC_ALL writeback, and we'll use the max + * limit for that. If the write is marked as a background write, + * then use the idle limit, or go to normal if we haven't had + * competing IO for a bit. + */ + if ((rw & REQ_HIPRIO) || atomic_read(&rwb->bdi->wb.dirty_sleeping)) + limit = rwb->wb_max; + else if ((rw & REQ_BG) || close_io(rwb)) { + /* + * If less than 100ms since we completed unrelated IO, + * limit us to half the depth for background writeback. + */ + limit = rwb->wb_background; + } else + limit = rwb->wb_normal; + + return limit; +} + +static inline bool may_queue(struct rq_wb *rwb, unsigned long rw) +{ + /* + * inc it here even if disabled, since we'll dec it at completion. + * this only happens if the task was sleeping in __wbt_wait(), + * and someone turned it off at the same time. + */ + if (!rwb_enabled(rwb)) { + atomic_inc(&rwb->inflight); + return true; + } + + return atomic_inc_below(&rwb->inflight, get_limit(rwb, rw)); +} + +/* + * Block if we will exceed our limit, or if we are currently waiting for + * the timer to kick off queuing again. + */ +static void __wbt_wait(struct rq_wb *rwb, unsigned long rw, spinlock_t *lock) +{ + DEFINE_WAIT(wait); + + if (may_queue(rwb, rw)) + return; + + do { + prepare_to_wait_exclusive(&rwb->wait, &wait, + TASK_UNINTERRUPTIBLE); + + if (may_queue(rwb, rw)) + break; + + if (lock) + spin_unlock_irq(lock); + + io_schedule(); + + if (lock) + spin_lock_irq(lock); + } while (1); + + finish_wait(&rwb->wait, &wait); +} + +static inline bool wbt_should_throttle(struct rq_wb *rwb, unsigned int rw) +{ + /* + * If not a WRITE (or a discard), do nothing + */ + if (!(rw & REQ_WRITE) || (rw & REQ_DISCARD)) + return false; + + /* + * Don't throttle WRITE_ODIRECT + */ + if ((rw & (REQ_SYNC | REQ_NOIDLE)) == REQ_SYNC) + return false; + + return true; +} + +/* + * Returns true if the IO request should be accounted, false if not. + * May sleep, if we have exceeded the writeback limits. Caller can pass + * in an irq held spinlock, if it holds one when calling this function. + * If we do sleep, we'll release and re-grab it. + */ +bool wbt_wait(struct rq_wb *rwb, unsigned int rw, spinlock_t *lock) +{ + if (!rwb_enabled(rwb)) + return false; + + if (!wbt_should_throttle(rwb, rw)) { + wb_timestamp(rwb, &rwb->last_issue); + return false; + } + + __wbt_wait(rwb, rw, lock); + + if (!timer_pending(&rwb->window_timer)) + rwb_arm_timer(rwb); + + return true; +} + +void wbt_issue(struct rq_wb *rwb, struct wb_issue_stat *stat) +{ + if (!rwb_enabled(rwb)) + return; + + wbt_issue_stat_set_time(stat); + + if (!wbt_tracked(stat) && !rwb->sync_issue) { + rwb->sync_cookie = stat; + rwb->sync_issue = wbt_issue_stat_get_time(stat); + } +} + +void wbt_requeue(struct rq_wb *rwb, struct wb_issue_stat *stat) +{ + if (!rwb_enabled(rwb)) + return; + if (stat == rwb->sync_cookie) { + rwb->sync_issue = 0; + rwb->sync_cookie = NULL; + } +} + +void wbt_set_queue_depth(struct rq_wb *rwb, unsigned int depth) +{ + if (rwb) { + rwb->queue_depth = depth; + wbt_update_limits(rwb); + } +} + +void wbt_set_write_cache(struct rq_wb *rwb, bool write_cache_on) +{ + if (rwb) + rwb->wc = write_cache_on; +} + +struct rq_wb *wbt_init(struct backing_dev_info *bdi, struct wb_stat_ops *ops, + void *ops_data) +{ + struct rq_wb *rwb; + + rwb = kzalloc(sizeof(*rwb), GFP_KERNEL); + if (!rwb) + return ERR_PTR(-ENOMEM); + + atomic_set(&rwb->inflight, 0); + init_waitqueue_head(&rwb->wait); + setup_timer(&rwb->window_timer, wb_timer_fn, (unsigned long) rwb); + rwb->wc = 1; + rwb->queue_depth = RWB_MAX_DEPTH; + rwb->last_comp = rwb->last_issue = jiffies; + rwb->bdi = bdi; + rwb->win_nsec = RWB_WINDOW_NSEC; + rwb->stat_ops = ops, + rwb->ops_data = ops_data; + wbt_update_limits(rwb); + return rwb; +} + +void wbt_exit(struct rq_wb *rwb) +{ + if (rwb) { + del_timer_sync(&rwb->window_timer); + kfree(rwb); + } +}
We can hook this up to the block layer, to help throttle buffered writes. Or NFS can tap into it, to accomplish the same. wbt registers a few trace points that can be used to track what is happening in the system: wbt_lat: 259:0: latency 2446318 wbt_stat: 259:0: rmean=2446318, rmin=2446318, rmax=2446318, rsamples=1, wmean=518866, wmin=15522, wmax=5330353, wsamples=57 wbt_step: 259:0: step down: step=1, window=72727272, background=8, normal=16, max=32 This shows a sync issue event (wbt_lat) that exceeded it's time. wbt_stat dumps the current read/write stats for that window, and wbt_step shows a step down event where we now scale back writes. Each trace includes the device, 259:0 in this case. Signed-off-by: Jens Axboe <axboe@fb.com> --- include/linux/wbt.h | 95 ++++++++ include/trace/events/wbt.h | 122 +++++++++++ lib/Kconfig | 3 + lib/Makefile | 1 + lib/wbt.c | 524 +++++++++++++++++++++++++++++++++++++++++++++ 5 files changed, 745 insertions(+) create mode 100644 include/linux/wbt.h create mode 100644 include/trace/events/wbt.h create mode 100644 lib/wbt.c