Now it will be possible to query any thread's scheduler state, not
only the current one. This aims at simplifying the watchdog checks
for reported threads. The operation is now a simple atomic xchg.
The watchdog timer has to go through complex operations due to not being
able to check if another thread's scheduler is still ticking. This is
simply because the scheduler status is marked as thread-local while it
could in fact also be an array. Let's do that (and align the array to
avoid false sharing) so that it's now possible to check any scheduler's
status.
When task profiling is enabled, the pool alloc/free code will measure the
time it takes to perform memory allocation after a cache miss or memory
freeing to the shared cache or OS. The time taken with the thread-local
cache is never measured as measuring that time is very expensive compared
to the pool access time. Here doing so costs around 2% performance at 2M
req/s, only when task profiling is enabled, so this remains reasonable.
The scheduler takes care of collecting that time and updating the
sched_activity entry corresponding to the current task when task profiling
is enabled.
The goal clearly is to track places that are wasting CPU time allocating
and releasing too often, or causing large evictions. This appears like
this in "show profiling tasks aggr":
Tasks activity over 11.428 sec till 0.000 sec ago:
function calls cpu_tot cpu_avg lkw_avg lkd_avg mem_avg lat_avg
process_stream 44183891 16.47m 22.36us 491.0ns 1.154us 1.000ns 101.1us
h1_io_cb 57386064 4.011m 4.193us 20.00ns 16.00ns - 29.47us
sc_conn_io_cb 42088024 49.04s 1.165us - - - 54.67us
h1_timeout_task 438171 196.5ms 448.0ns - - - 100.1us
srv_cleanup_toremove_conns 65 1.468ms 22.58us 184.0ns 87.00ns - 101.3us
task_process_applet 3 508.0us 169.3us - 107.0us 1.847us 29.67us
srv_cleanup_idle_conns 6 225.3us 37.55us 15.74us 36.84us - 49.47us
accept_queue_process 2 45.62us 22.81us - - 4.949us 54.33us
When DEBUG_THREAD > 0 and task profiling enabled, we'll now measure the
time spent with at least one lock held for each task. The time is
collected by locking operations when locks are taken raising the level
to one, or released resetting the level. An accumulator is updated in
the thread_ctx struct that is collected by the scheduler when the task
returns, and updated in the sched_activity entry of the related task.
This allows to observe figures like this one:
Tasks activity over 259.516 sec till 0.000 sec ago:
function calls cpu_tot cpu_avg lkw_avg lkd_avg lat_avg
h1_io_cb 15466589 2.574m 9.984us - - 33.45us <- sock_conn_iocb@src/sock.c:1099 tasklet_wakeup
sc_conn_io_cb 8047994 8.325s 1.034us - - 870.1us <- sc_app_chk_rcv_conn@src/stconn.c:844 tasklet_wakeup
process_stream 7734689 4.356m 33.79us 1.990us 1.641us 1.554ms <- sc_notify@src/stconn.c:1206 task_wakeup
process_stream 7734292 46.74m 362.6us 278.3us 132.2us 972.0us <- stream_new@src/stream.c:585 task_wakeup
sc_conn_io_cb 7733158 46.88s 6.061us - - 68.78us <- h1_wake_stream_for_recv@src/mux_h1.c:3633 tasklet_wakeup
task_process_applet 6603593 4.484m 40.74us 16.69us 34.00us 96.47us <- sc_app_chk_snd_applet@src/stconn.c:1043 appctx_wakeup
task_process_applet 4761796 3.420m 43.09us 18.79us 39.28us 138.2us <- __process_running_peer_sync@src/peers.c:3579 appctx_wakeup
process_table_expire 4710662 4.880m 62.16us 9.648us 53.95us 158.6us <- run_tasks_from_lists@src/task.c:671 task_queue
stktable_add_pend_updates 4171868 6.786s 1.626us - 1.487us 47.94us <- stktable_add_pend_updates@src/stick_table.c:869 tasklet_wakeup
h1_io_cb 2871683 1.198s 417.0ns 70.00ns 69.00ns 1.005ms <- h1_takeover@src/mux_h1.c:5659 tasklet_wakeup
process_peer_sync 2304957 5.368s 2.328us - 1.156us 68.54us <- stktable_add_pend_updates@src/stick_table.c:873 task_wakeup
process_peer_sync 1388141 3.174s 2.286us - 1.130us 52.31us <- run_tasks_from_lists@src/task.c:671 task_queue
stktable_add_pend_updates 463488 3.530s 7.615us 2.000ns 7.134us 771.2us <- stktable_touch_with_exp@src/stick_table.c:654 tasklet_wakeup
Here we see that almost the entirety of stktable_add_pend_updates() is
spent under a lock, that 1/3 of the execution time of process_stream()
was performed under a lock and that 2/3 of it was spent waiting for a
lock (this is related to the 10 track-sc present in this config), and
that the locking time in process_peer_sync() has now significantly
reduced. This is more visible with "show profiling tasks aggr":
Tasks activity over 475.354 sec till 0.000 sec ago:
function calls cpu_tot cpu_avg lkw_avg lkd_avg lat_avg
h1_io_cb 25742539 3.699m 8.622us 11.00ns 10.00ns 188.0us
sc_conn_io_cb 22565666 1.475m 3.920us - - 473.9us
process_stream 21665212 1.195h 198.6us 140.6us 67.08us 1.266ms
task_process_applet 16352495 11.31m 41.51us 17.98us 36.55us 112.3us
process_peer_sync 7831923 17.15s 2.189us - 1.107us 41.27us
process_table_expire 6878569 6.866m 59.89us 9.359us 51.91us 151.8us
stktable_add_pend_updates 6602502 14.77s 2.236us - 2.060us 119.8us
h1_timeout_task 801 703.4us 878.0ns - - 185.7us
srv_cleanup_toremove_conns 347 12.43ms 35.82us 240.0ns 70.00ns 1.924ms
accept_queue_process 142 1.384ms 9.743us - - 340.6us
srv_cleanup_idle_conns 74 475.0us 6.418us 896.0ns 5.667us 114.6us
When DEBUG_THREAD > 0, and if task profiling is enabled, then each
locking attempt will measure the time it takes to obtain the lock, then
add that time to a thread_ctx accumulator that the scheduler will then
retrieve to update the current task's sched_activity entry. The value
will then appear avearaged over the number of calls in the lkw_avg column
of "show profiling tasks", such as below:
Tasks activity over 48.298 sec till 0.000 sec ago:
function calls cpu_tot cpu_avg lkw_avg lat_avg
h1_io_cb 3200170 26.81s 8.377us - 32.73us <- sock_conn_iocb@src/sock.c:1099 tasklet_wakeup
sc_conn_io_cb 1657841 1.645s 992.0ns - 853.0us <- sc_app_chk_rcv_conn@src/stconn.c:844 tasklet_wakeup
process_stream 1600450 49.16s 30.71us 1.936us 1.392ms <- sc_notify@src/stconn.c:1206 task_wakeup
process_stream 1600321 7.770m 291.3us 209.1us 901.6us <- stream_new@src/stream.c:585 task_wakeup
sc_conn_io_cb 1599928 7.975s 4.984us - 65.77us <- h1_wake_stream_for_recv@src/mux_h1.c:3633 tasklet_wakeup
task_process_applet 997609 46.37s 46.48us 16.80us 113.0us <- sc_app_chk_snd_applet@src/stconn.c:1043 appctx_wakeup
process_table_expire 922074 48.79s 52.92us 7.275us 181.1us <- run_tasks_from_lists@src/task.c:670 task_queue
stktable_add_pend_updates 705423 1.511s 2.142us - 56.81us <- stktable_add_pend_updates@src/stick_table.c:869 tasklet_wakeup
task_process_applet 683511 34.75s 50.84us 18.37us 153.3us <- __process_running_peer_sync@src/peers.c:3579 appctx_wakeup
h1_io_cb 535395 198.1ms 370.0ns 72.00ns 930.4us <- h1_takeover@src/mux_h1.c:5659 tasklet_wakeup
It now makes it pretty obvious which tasks (hence call chains) spend their
time waiting on a lock and for what share of their execution time.
When task profiling is enabled, the current thread knows when the
currently running task was woken up and called, so we can calculate
how long ago it was woken up and called. This is convenient to figure
whether or not a warning or panic is caused by this task or by a
previous one, so let's report this info in thread outputs when known.
It would be useful to backport this to 3.2.
These structs are intensively used and really must not experience false
sharing, so let's declare them aligned to 64. We don't try to align the
struct themselves, as we don't want the compiler to expand them either.
This will make the pools size and alignment automatically inherit
the type declaration. It was done like this:
sed -i -e 's:DECLARE_POOL(\([^,]*,[^,]*,\s*\)sizeof(\([^)]*\))):DECLARE_TYPED_POOL(\1\2):g' $(git grep -lw DECLARE_POOL src addons)
sed -i -e 's:DECLARE_STATIC_POOL(\([^,]*,[^,]*,\s*\)sizeof(\([^)]*\))):DECLARE_STATIC_TYPED_POOL(\1\2):g' $(git grep -lw DECLARE_STATIC_POOL src addons)
81 replacements were made. The only remaining ones are those which set
their own size without depending on a structure. The few ones with an
extra size were manually handled.
It also means that the requested alignments are now checked against the
type's. Given that none is specified for now, no issue is reported.
It was verified with "show pools detailed" that the definitions are
exactly the same, and that the binaries are similar.
There's no point reading t->state to check for a tasklet after we've
atomically read the state into the local "state" variable. Not only it's
more expensive, it's also less clear whether that state is supposed to
be atomic or not. And in any case, tasks and tasklets have their type
forever and the one reflected in state is correct and stable.
After recent commit b81c9390f ("MEDIUM: tasks: Mutualize the TASK_KILLED
code between tasks and tasklets"), the task accounting was no longer
correct for killed tasks due to the decrement of tasks in list that was
no longer done, resulting in infinite loops in process_runnable_tasks().
This just illustrates that this code remains complex and should be further
cleaned up. No backport is needed, as this was in 3.2.
The code to handle a task/tasklet when it's been killed before it were
to run is mostly identical, so move it outside of task and tasklet
specific code, and inside the common code.
This commit is just cosmetic, and should have no impact.
TASK_QUEUED was used to mean "the task has been scheduled to run",
TASK_IN_LIST was used to mean "the tasklet has been scheduled to run",
remove TASK_IN_LIST and just use TASK_QUEUED for tasklets instead.
This commit is just cosmetic, and should not have any impact.
There is some code that should run no matter if the task was killed or
not, and was needlessly duplicated, so only use one instance.
This also fixes a small bug when a tasklet that got killed before it
could run would still count as a tasklet that ran, when it should not,
which just means that we'd run one less useful task before going back to
the poller.
This commit is mostly cosmetic, and should not have any impact.
The code that checks if we're currently running, and waits if so, was
identical between tasks and tasklets, so move it in code common to tasks
and tasklets.
This commit is just cosmetic, and should not have any impact.
tasklets were originally designed to alway run on only one thread, so it
was not possible to have it run on 2 threads concurrently.
The API has been extended so that another thread may wake the tasklet,
the idea was still that we wanted to have it run on one thread only.
However, the way it's been done meant that unless a tasklet was bound to
a specific tid with tasklet_set_tid(), or we explicitely used
tasklet_wakeup_on() to specify the thread for the target to run on, it
would be scheduled to run on the current thread.
This is in fact a desirable feature. There is however a race condition
in which the tasklet would be scheduled on a thread, while it is running
on another. This could lead to the same tasklet to run on multiple
threads, which we do not want.
To fix this, just do what we already do for regular tasks, set the
"TASK_RUNNING" flag, and when it's time to execute the tasklet, wait
until that flag is gone.
Only one case has been found in the current code, where the tasklet
could run on different threads depending on who wakes it up, in the
leastconn load balancer, since commit
627280e15f03755b8f59f0191cd6d6bcad5afeb3.
It should not be a problem in practice, as the function called can be
called concurrently.
If a bug is eventually found in relation to this problem, and this patch
should be backported, the following patches should be backported too :
MEDIUM: quic: Make sure we return the tasklet from quic_accept_run
MEDIUM: quic: Make sure we return NULL in quic_conn_app_io_cb if needed
MEDIUM: quic: Make sure we return the tasklet from qcc_io_cb
MEDIUM: mux_fcgi: Make sure we return the tasklet from fcgi_deferred_shut
MEDIUM: listener: Make sure w ereturn the tasklet from accept_queue_process
MEDIUM: checks: Make sure we return the tasklet from srv_chk_io_cb
This verifies that the scheduler is still ticking without having to
access the activity[] array nor keeping local copies of the ctxsw
counter. It just tests and sets a flag that is reset after each
return from a ->process() function.
There's a barrier after releasing the current task in the scheduler.
However it's improperly placed, it's done after pool_free() while in
fact it must be done immediately after resetting the current pointer.
Indeed, the purpose is to make sure that nobody sees the task as valid
when it's in the process of being released. This is something that
could theoretically happen if interrupted by a signal in the inlined
code of pool_free() if the compiler decided to postpone the write to
->current. In practice since nothing fancy is done in the inlined part
of the function, there's currently no risk of reordering. But it could
happen if the underlying __pool_free() were to be inlined for example,
and in this case we could possibly observe th_ctx->current pointing
to something currently being destroyed.
With the barrier between the two, there's no risk anymore.
Currently tasks being profiled have th_ctx->sched_call_date set to the
current nanosecond in monotonic time. But there's no other way to have
this, despite the scheduler being capable of it. Let's just declare a
new task flag, TASK_F_WANTS_TIME, that makes the scheduler take the time
just before calling the handler. This way, a task that needs nanosecond
resolution on the call date will be able to be called with an up-to-date
date without having to abuse now_mono_time() if not needed. In addition,
if CLOCK_MONOTONIC is not supported (now_mono_time() always returns 0),
the date is set to the most recently known now_ns, which is guaranteed
to be atomic and is only updated once per poll loop.
This date can be more conveniently retrieved using task_mono_time().
This can be useful, e.g. for pacing. The code was slightly adjusted so
as to merge the common parts between the profiling case and this one.
It's common to see process_stream() being woken up by wake_expired_tasks
in the profiling output, without knowing which timeout was set to cause
this. By making it possible to record the call places of task_queue()
and task_schedule(), and by making wake_expired_tasks() explicitly not
replace it, we'll be able to know which task_queue() or task_schedule()
was triggered for a given wakeup.
For example below:
process_stream 51200 311.4ms 6.081us 34.59s 675.6us <- run_tasks_from_lists@src/task.c:659 task_queue
process_stream 19227 70.00ms 3.640us 9.813m 30.62ms <- sc_notify@src/stconn.c:1136 task_wakeup
process_stream 6414 102.3ms 15.95us 8.093m 75.70ms <- stream_new@src/stream.c:578 task_wakeup
It's visible that it's the run_tasks_from_lists() which in fact applies
on the task->expire returned by the ->process() function itself.
Adjust BUG_ON() statement to allow tasklet_wakeup_after() for tasklets
with tid pinned to -1 (the current thread). This is similar to
tasklet_wakeup().
This should be backported up to 2.6.
As reported in github issue #1881, there are situations where an excess
of TLS handshakes can cause a livelock. What's happening is that normally
we process at most one TLS handshake per loop iteration to maintain the
latency low. This is done by tagging them with TASK_HEAVY, queuing these
tasklets in the TL_HEAVY queue. But if something slows down the loop, such
as a connect() call when no more ports are available, we could end up
processing no more than a few hundred or thousands handshakes per second.
If the llmit becomes lower than the rate of incoming handshakes, we will
accumulate them and at some point users will get impatient and give up or
retry. Then a new problem happens: the queue fills up with even more
handshake attempts, only one of which will be handled per iteration, so
we can end up processing only outdated handshakes at a low rate, with
basically nothing else in the queue. This can for example happen in
parallel with health checks that don't require incoming handshakes to
succeed to continue to cause some activity that could maintain the high
latency stuff active.
Here we're taking a slightly different approach. First, instead of always
allowing only one handshake per loop (and usually it's critical for
latency), we take the current situation into account:
- if configured with tune.sched.low-latency, the limit remains 1
- if there are other non-heavy tasks, we set the limit to 1 + one
per 1024 tasks, so that a heavily loaded queue of 4k handshakes
per thread will be able to drain them at ~4 per loops with a
limited impact on latency
- if there are no other tasks, the limit grows to 1 + one per 128
tasks, so that a heavily loaded queue of 4k handshakes per thread
will be able to drain them at ~32 per loop with still a very
limited impact on latency since only I/O will get delayed.
It was verified on a 56-core Xeon-8480 that this did not degrade the
latency; all requests remained below 1ms end-to-end in full close+
handshake, and even 500us under low-lat + busy-polling.
This must be backported to 2.4.
There's a per-thread "long_rq" counter that is used to indicate how
often we leave the scheduler with tasks still present in the run queue.
The purpose is to know when tune.runqueue-depth served to limit latency,
due to a large number of tasks being runnable at once.
However there's a bug there, it's not always set: if after the first
run, one heavy task was processed and later only heavy tasks remain,
we'll loop back to not_done_yet where we try to pick more tasks, but
none are eligible (since heavy ones have already run) so we directly
return without incrementing the counter. This is what causes ultra-low
values on long_rq during massive SSL handshakes, that are confusing
because they make one believe that tl_class_mask doesn't have the HEAVY
flag anymore. Let's just fix that by not returning from the middle of
the function.
This can be backported as far as 2.4.
The build with DEBUG_THREAD was broken by commit fc50b9dd1 ("BUG/MAJOR:
sched: protect task during removal from wait queue"). It took me a while
to figure how to declare and aligned and initialized rwlock that wasn't
static, but it turns out that __decl_aligned_rwlock() does exactly this,
so that we don't have to assign an integer value when a struct is expected
in case of debugging.
No backport is needed.
The issue addressed by commit fbb934da9 ("BUG/MEDIUM: stick-table: fix
a race condition when updating the expiration task") is still present
when thread groups are enabled, but this time it lies in the scheduler.
What happens is that a task configured to run anywhere might already
have been queued into one group's wait queue. When updating a stick
table entry, sometimes the task will have to be dequeued and requeued.
For this a lock is taken on the current thread group's wait queue lock,
but while this is necessary for the queuing, it's not sufficient for
dequeuing since another thread might be in the process of expiring this
task under its own group's lock which is different. This is easy to test
using 3 stick tables with 1ms expiration, 3 track-sc rules and 4 thread
groups. The process crashes almost instantly under heavy traffic.
One approach could consist in storing the group number the task was
queued under in its descriptor (we don't need 32 bits to store the
thread id, it's possible to use one short for the tid and another
one for the tgrp). Sadly, no safe way to do this was figured, because
the race remains at the moment the thread group number is checked, as
it might be in the process of being changed by another thread. It seems
that a working approach could consist in always having it associated
with one group, and only allowing to change it under this group's lock,
so that any code trying to change it would have to iterately read it
and lock its group until the value matches, confirming it really holds
the correct lock. But this seems a bit complicated, particularly with
wait_expired_tasks() which already uses upgradable locks to switch from
read state to a write state.
Given that the shared tasks are not that common (stick-table expirations,
rate-limited listeners, maybe resolvers), it doesn't seem worth the extra
complexity for now. This patch takes a simpler and safer approach
consisting in switching back to a single wq_lock, but still keeping
separate wait queues. Given that shared wait queues are almost always
empty and that otherwise they're scanned under a read lock, the
contention remains manageable and most of the time the lock doesn't
even need to be taken since such tasks are not present in a group's
queue. In essence, this patch reverts half of the aforementionned
patch. This was tested and confirmed to work fine, without observing
any performance degradation under any workload. The performance with
8 groups on an EPYC 74F3 and 3 tables remains twice the one of a
single group, with the contention remaining on the table's lock first.
No backport is needed.
Now that ->wake_date is common to tasks and tasklets, we don't need
anymore to carry a duplicate control block to read and update it for
tasks and tasklets. And given that this code was present early in the
if/else fork between tasks and tasklets, taking it out of the block
allows to move the task part into a more visible "else" branch that
also allows to factor the epilogue that resets th_ctx->current and
updates profile_entry->cpu_time, which also used to be duplicated.
Overall, doing just that saved 253 bytes in the function, or ~1/6,
which is not bad considering that it's on a hot path. And the code
got much ore readable.
It was a mistake to put these two fields in the struct task. This
was added in 1.9 via commit 9efd7456e ("MEDIUM: tasks: collect per-task
CPU time and latency"). These fields are used solely by streams in
order to report the measurements via the lat_ns* and cpu_ns* sample
fetch functions when task profiling is enabled. For the rest of the
tasks, this is pure CPU waste when profiling is enabled, and memory
waste 100% of the time, as the point where these latencies and usages
are measured is in the profiling array.
Let's move the fields to the stream instead, and have process_stream()
retrieve the relevant info from the thread's context.
The struct task is now back to 120 bytes, i.e. almost two cache lines,
with 32 bit still available.
The profile entry that corresponds to the current task/tasklet being
profiled is now stored into the thread's context. This will allow it
to be accessed from the tasks themselves. This is needed for an upcoming
fix.
When task profiling is enabled, the scheduler can measure and report
the cumulated time spent in each task and their respective latencies. But
this was wrong for tasks with few wakeups as well as for self-waking ones,
because the call date needed to measure how long it takes to process the
task is retrieved in the task itself (->wake_date was turned to the call
date), and we could face two conditions:
- a new wakeup while the task is executing would reset the ->wake_date
field before returning and make abnormally low values being reported;
that was likely the case for taskrun_applet for self-waking applets;
- when the task dies, NULL is returned and the call date couldn't be
retrieved, so that CPU time was not being accounted for. This was
particularly visible with process_stream() which is usually called
only twice per request, and whose time was systematically halved.
The cleanest solution here is to keep in mind that the scheduler already
uses quite a bit of local context in th_ctx, and place the intermediary
values there so that they cannot vanish. The wake_date has to be reset
immediately once read, and only its copy is used along the function. Note
that this must be done both for tasks and tasklet, and that until recently
tasklets were also able to report wrong values due to their sole dependency
on TH_FL_TASK_PROFILING between tests.
One nice benefit for future improvements is that such information will now
be available from the task without having to be stored into the task itself
anymore.
Since the tasklet part was computed on wrapping 32-bit arithmetics and
the task one was on 64-bit, the values were now consistently moved to
32-bit as it's already largely sufficient (4s spent in a task is more
than twice what the watchdog would tolerate). Some further cleanups might
be necessary, but the patch aimed at staying minimal.
Task profiling output after 1 million HTTP request previously looked like
this:
Tasks activity:
function calls cpu_tot cpu_avg lat_tot lat_avg
h1_io_cb 2012338 4.850s 2.410us 12.91s 6.417us
process_stream 2000136 9.594s 4.796us 34.26s 17.13us
sc_conn_io_cb 2000135 1.973s 986.0ns 30.24s 15.12us
h1_timeout_task 137 - - 2.649ms 19.34us
accept_queue_process 49 152.3us 3.107us 321.7yr 6.564yr
main+0x146430 7 5.250us 750.0ns 25.92us 3.702us
srv_cleanup_idle_conns 1 559.0ns 559.0ns 918.0ns 918.0ns
task_run_applet 1 - - 2.162us 2.162us
Now it looks like this:
Tasks activity:
function calls cpu_tot cpu_avg lat_tot lat_avg
h1_io_cb 2014194 4.794s 2.380us 13.75s 6.826us
process_stream 2000151 20.01s 10.00us 36.04s 18.02us
sc_conn_io_cb 2000148 2.167s 1.083us 32.27s 16.13us
h1_timeout_task 198 54.24us 273.0ns 3.487ms 17.61us
accept_queue_process 52 158.3us 3.044us 409.9us 7.882us
main+0x1466e0 18 16.77us 931.0ns 63.98us 3.554us
srv_cleanup_toremove_conns 8 282.1us 35.26us 546.8us 68.35us
srv_cleanup_idle_conns 3 149.2us 49.73us 8.131us 2.710us
task_run_applet 3 268.1us 89.38us 11.61us 3.871us
Note the two-fold difference on process_stream().
This feature is essentially used for debugging so it has extremely limited
impact. However it's used quite a bit more in bug reports and it would be
desirable that at least 2.6 gets this fix backported. It depends on at least
these two previous patches which will then also have to be backported:
MINOR: task: permanently enable latency measurement on tasklets
CLEANUP: task: rename ->call_date to ->wake_date
This field is misnamed because its real and important content is the
date the task was woken up, not the date it was called. It temporarily
holds the call date during execution but this remains confusing. In
fact before the latency measurements were possible it was indeed a call
date. Thus is will now be called wake_date.
This change is necessary because a subsequent fix will require the
introduction of the real call date in the thread ctx.
When tasklet latency measurement was enabled in 2.4 with commit b2285de04
("MINOR: tasks: also compute the tasklet latency when DEBUG_TASK is set"),
the feature was conditionned on DEBUG_TASK because the field would add 8
bytes to the struct tasklet.
This approach was not a very good idea because the struct ends on an int
anyway thus it does finish with a 32-bit hole regardless of the presence
of this field. What is true however is that adding it turned a 64-byte
struct to 72-byte when caller debugging is enabled.
This patch revisits this with a minor change. Now only the lowest 32
bits of the call date are stored, so they always fit in the remaining
hole, and this allows to remove the dependency on DEBUG_TASK. With
debugging off, we're now seeing a 48-byte struct, and with debugging
on it's exactly 64 bytes, thus still exactly one cache line. 32 bits
allow a latency of 4 seconds on a tasklet, which already indicates a
completely dead process, so there's no point storing the upper bits at
all. And even in the event it would happen once in a while, the lost
upper bits do not really add any value to the debug reports. Also, now
one tasklet wakeup every 4 billion will not be sampled due to the test
on the value itself. Similarly we just don't care, it's statistics and
the measurements are not 9-digit accurate anyway.
This one is only used as a hint to improve scheduling latency, so there
is no more point in keeping it global since each thread group handles
its own run q
Their migration was postponed for convenience only but now's time for
having the shared wait queues per thread group and not just per process,
otherwise the WQ lock uses a huge amount of CPU alone.
The thread flags are touched a little bit by other threads, e.g. the STUCK
flag may be set by other ones, and they're watched a little bit. As such
we need to use atomic ops only to manipulate them. Most places were already
using them, but here we generalize the practice. Only ha_thread_dump() does
not change because it's run under isolation.
Almost every call place of wake_thread() checks for sleeping threads and
clears the sleeping mask itself, while the function is solely used for
there. Let's move the check and the clearing of the bit inside the function
itself. Note that updt_fd_polling() still performs the check because its
rules are a bit different.
Since we don't mix tasks from different threads in the run queues
anymore, we don't need to use the eb32sc_ trees and we can switch
to the regular eb32 ones. This uses cheaper lookup and insert code,
and a 16-thread test on the queues shows a performance increase
from 570k RPS to 585k RPS.
This bit field used to be a per-thread cache of the result of the last
lookup of the presence of a task for each thread in the shared cache.
Since we now know that each thread has its own shared cache, a test of
emptiness is now sufficient to decide whether or not the shared tree
has a task for the current thread. Let's just remove this mask.
grq_total was only used to know how many tasks were being queued in the
global runqueue for stats purposes, and that was transferred to the per
thread rq_total counter once assigned. We don't need this anymore since
we know where they are, so let's just directly update rq_total and drop
that one.
Since we only use the shared runqueue to put tasks only assigned to
known threads, let's move that runqueue to each of these threads. The
goal will be to arrange an N*(N-1) mesh instead of a central contention
point.
The global_rqueue_ticks had to be dropped (for good) since we'll now
use the per-thread rqueue_ticks counter for both trees.
A few points to note:
- the rq_lock stlil remains the global one for now so there should not
be any gain in doing this, but should this trigger any regression, it
is important to detect whether it's related to the lock or to the tree.
- there's no more reason for using the scope-based version of the ebtree
now, we could switch back to the regular eb32_tree.
- it's worth checking if we still need TASK_GLOBAL (probably only to
delete a task in one's own shared queue maybe).
The runqueue ticks counter is per-thread and wasn't initially meant to
be shared. We'll soon have to share it so let's make it atomic. It's
only updated when waking up a task, and no performance difference was
observed. It was moved in the thread_ctx struct so that it doesn't
pollute the local cache line when it's later updated by other threads.
TASK_SHARED_WQ was set upon task creation and never changed afterwards.
Thus if a task was created to run anywhere (e.g. a check or a Lua task),
all its timers would always pass through the shared timers queue with a
lock. Now we know that tid<0 indicates a shared task, so we can use that
to decide whether or not to use the shared queue. The task might be
migrated using task_set_affinity() but it's always dequeued first so
the check will still be valid.
Not only this removes a flag that's difficult to keep synchronized with
the thread ID, but it should significantly lower the load on systems with
many checks. A quick test with 5000 servers and fast checks that were
saturating the CPU shows that the check rate increased by 20% (hence the
CPU usage dropped by 17%). It's worth noting that run_task_lists() almost
no longer appears in perf top now.
At a few places where the task's thread mask. Now we know that it's always
either one bit or all bits of all_threads_mask, so we can replace it with
either 1<<tid or all_threads_mask depending on what's expected.
It's worth noting that the global_tasks_mask is still set this way and
that it's reaching its limits. Similarly, the task_new() API would deserve
an update to stop using a thread mask and use a thread number instead.
Similarly, task_set_affinity() should be updated to directly take a
thread number.
At this point the task's thread mask is not used anymore.
At several places we need to figure the ID of the first thread allowed
to run a task. Till now this was performed using my_ffsl(t->thread_mask)
but since we now have the thread ID stored into the task, let's use it
instead. This is tagged major because it starts to assume that tid<0 is
strictly equivalent to atleast2(thread_mask), and that as such, among
the allowed threads are the current one.
We want to be able to schedule a tasklet onto a thread after the current tasklet
is done. What we have to do is to insert this tasklet at the head of the thread
task list. Furthermore, we would like to serialize the tasklets. They must be
run in the same order as the order in which they have been scheduled. This is
implemented passing a list of tasklet as parameter (see <head> parameters) which
must be reused for subsequent calls.
_tasklet_wakeup_after_on() is implemented to accomplish this job.
tasklet_wakeup_after_on() and tasklet_wake_after() are only wrapper macros around
_tasklet_wakeup_after_on(). tasklet_wakeup_after_on() does exactly the same thing
as _tasklet_wakeup_after_on() without having to pass the filename and line in the
filename as parameters (usefull when DEBUG_TASK is enabled).
tasklet_wakeup_after() hides also the usage of the thread parameter which is
<tl> tasklet thread ID.
tasklet_kill() was introduced in 2.5-dev4 with commit 7b368339a
("MEDIUM: task: implement tasklet kill"), but a comparison error
there makes tasklets killed on thread 1 assigned to the killing
thread. Fortunately, the function was finally not used so there's
no harm right now, hence the minor tag, but this must be fixed and
backported in case a later fix relies on it.
This should be backported to 2.5.
If we've stopped consulting the local wait queue due to too many tasks
(max_processed <= 0), there's no point starting to lock the shared WQ,
check the first task's expiration date, upgrading the lock just to
refrain from doing the work because of the limit. All this does is
increase contention on an already contended system.
Note that there is still a fairness issue in this WQ dequeuing code. If
each thread is busy with expired tasks, no thread will dequeue the global
ones. In practice it doesn't make much sense and should quickly resorb,
but it could be nice to have an alternating flag indicating where to
start from on next call to improve this.
This macro was used both for binding and for lookups. When binding tasks
or FDs, using all_threads_mask instead is better as it will later be per
group. For lookups, ~0UL always does the job. Thus in practice the macro
was already almost not used anymore since the rest of the code could run
fine with a constant of all ones there.
