sched_throw() can no longer take a NULL thread, APs enter through
sched_ap_entry() instead. This completely removes branching in the
common case and cleans up both paths. No functional change intended.
Reviewed by: kib, markj
Differential Revision: https://reviews.freebsd.org/D32829
schedinit_ap() sets up an AP for a later call to sched_throw(NULL).
Currently, ULE sets up some pcpu bits and fixes the idlethread lock with
a call to sched_throw(NULL); this results in a window where curthread is
setup in platforms' init_secondary(), but it has the wrong td_lock.
Typical platform AP startup procedure looks something like:
- Setup curthread
- ... other stuff, including cpu_initclocks_ap()
- Signal smp_started
- sched_throw(NULL) to enter the scheduler
cpu_initclocks_ap() may have callouts to process (e.g., nvme) and
attempt to sched_add() for this AP, but this attempt fails because
of the noted violated assumption leading to locking heartburn in
sched_setpreempt().
Interrupts are still disabled until cpu_throw() so we're not really at
risk of being preempted -- just let the scheduler in on it a little
earlier as part of setting up curthread.
Reviewed by: alfredo, kib, markj
Triage help from: andrew, markj
Smoke-tested by: alfredo (ppc), kevans (arm64, x86), mhorne (arm)
Differential Revision: https://reviews.freebsd.org/D32797
Before this change kern.sched.interact sysctl setting above 32 gave
all interactive threads identical priority of PRI_MIN_INTERACT due to
((PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) / sched_interact) turning
zero. Setting the sysctl lower reduced the range of used priority
levels up to half, that is not great either.
Change of the operations order should fix the issue, always using full
range of priorities, while overflow is impossible there since both
score and priority values are small. While there, make the variables
unsigned as they really are.
MFC after: 1 month
e745d729be caused infinite loop with interrupts disabled in load
stealing code if steal_thresh set below 2. Such configuration should
not generally be used, but appeared some people are using it to
workaround some problems.
To fix the problem explicitly pass to sched_highest() minimum number
of transferrable threads, supported by the caller, instead of guessing.
MFC after: 25 days
Depending on hardware, NUMA nodes may match last level caches, or
they may be above them (AMD Zen 2/3) or below (Intel Xeon w/ SNC).
This information is provided by ACPI instead of CPUID, and it is
provided for each CPU individually instead of mask widths, but
this code should be able to properly handle all the above cases.
This change should immediately allow idle stealing in sched_ule(4)
to prefer load from NUMA-local CPUs to remote ones when the node
does not match LLC. Later we may think of how to better handle it
on sched_pickcpu() side.
MFC after: 1 month
Before this change long-term load balancer was unable to migrate
running threads, only ones waiting on run queues. But with growing
number of CPU cores it is quite typical now for system to not have
many waiting threads. But same time if due to some coincidence two
long-running CPU-bound threads ended up sharing same physical CPU
core, they could suffer from the SMT penalty indefinitely, and the
load balancer couldn't help.
Improve that by teaching the load balancer to hint running threads
to migrate by marking them with TDF_NEEDRESCHED and new TDF_PICKCPU
flag, making sched_pickcpu() to search for better CPU later, when
it is convenient.
Fix CPU search logic when balancing to limit round-robin migrations
in case of almost equal load to the group of physical cores. The
previous code bounced threads across all the system, that should be
pretty bad for caches and NUMA affinity, while additional fairness
was almost invisible, diminishing with number of cores in the group.
MFC after: 1 month
In scenarios when first thread in the queue can migrate to specified
CPU, but later ones can't runq_steal_from() incorrectly returned NULL.
MFC after: 2 weeks
I don't think it changes anything, but why not.
While there, make cpu_search_highest() use all 8 lower load bits for
noise, since it does not use cs_prefer and the code is not shared
with cpu_search_lowest() any more.
MFC after: 1 month
On some load patterns it is possible for several CPUs to try steal
thread from the same CPU despite randomization introduced. It may
cause significant lock contention when holding one queue lock idle
thread tries to acquire another one. Use of trylock on the remote
queue allows both reduce the contention and handle lock ordering
easier. If we can't get lock inside tdq_trysteal() we just return,
allowing tdq_idled() handle it. If it happens in tdq_idled(), then
we repeat search for load skipping this CPU.
On 2-socket 80-thread Xeon system I am observing dramatic reduction
of the lock spinning time when doing random uncached 4KB reads from
12 ZVOLs, while IOPS increase from 327K to 403K.
MFC after: 1 month
When sched_highest() called for some CPU group returns nothing, idle
thread calls it for the parent CPU group. But the parent CPU group
also includes the CPU group we've just searched, and unless there is
a race going on, it is unlikely we find anything new this time.
Avoid the double search in case of parent group having only two sub-
groups (the most prominent case). Instead of escalating to the parent
group run the next search over the sibling subgroup and escalate two
levels up after if that fail too. In case of more than two siblings
the difference is less significant, while searching the parent group
can result in better decision if we find several candidate CPUs.
On 2-socket 40-core Xeon system I am measuring ~25% reduction of CPU
time spent inside cpu_search_highest() in both SMT (2x20x2) and non-
SMT (2x20) cases.
MFC after: 1 month
To prevent umtx.h polluting by future changes split it on two headers:
umtx.h - ABI header for userspace;
umtxvar.h - the kernel staff.
While here fix umtx_key_match style.
Reviewed by: kib
Differential Revision: https://reviews.freebsd.org/D31248
MFC after: 2 weeks
Remove cpu_search_both(), unused for many years. Without it there is
less sense for the trick of compiling common cpu_search() into separate
cpu_search_lowest() and cpu_search_highest(), so split them completely,
making code more readable. While there, split iteration over children
groups and CPUs, complicating code for very small deduplication.
Stop passing cpuset_t arguments by value and avoid some manipulations.
Since MAXCPU bump from 64 to 256, what was a single register turned
into 32-byte memory array, requiring memory allocation and accesses.
Splitting struct cpu_search into parameter and result parts allows to
even more reduce stack usage, since the first can be passed through
on recursion.
Remove CPU_FFS() from the hot paths, precalculating first and last CPU
for each CPU group in advance during initialization. Again, it was
not a problem for 64 CPUs before, but for 256 FFS needs much more code.
With these changes on 80-thread system doing ~260K uncached ZFS reads
per second I observe ~30% reduction of time spent in cpu_search_*().
MFC after: 1 month
r357614 added CTLFLAG_NEEDGIANT to make it easier to find nodes that are
still not MPSAFE (or already are but aren’t properly marked).
Use it in preparation for a general review of all nodes.
This is non-functional change that adds annotations to SYSCTL_NODE and
SYSCTL_PROC nodes using one of the soon-to-be-required flags.
Mark all obvious cases as MPSAFE. All entries that haven't been marked
as MPSAFE before are by default marked as NEEDGIANT
Approved by: kib (mentor, blanket)
Commented by: kib, gallatin, melifaro
Differential Revision: https://reviews.freebsd.org/D23718
If the thread's lock is already that of the runqueue, don't recurse on
the queue lock.
Reviewed by: jeff, kib
Sponsored by: The FreeBSD Foundation
Differential Revision: https://reviews.freebsd.org/D23492
After r355784 we no longer hold a thread's thread lock when switching it
out. Preserve the previous synchronization protocol for td_oncpu by
setting it together with td_state, before dropping the thread lock
during a switch.
Reported and tested by: pho
Reviewed by: kib
Discussed with: jeff
Differential Revision: https://reviews.freebsd.org/D23270
it. The introduction of lockless switch in r355784 created a race to
re-use the exiting thread that was only possible to hit on a hypervisor.
Reported/Tested by: rlibby
Discussed with: rlibby, jhb
Don't hold the scheduler lock while doing context switches. Instead we
unlock after selecting the new thread and switch within a spinlock
section leaving interrupts and preemption disabled to prevent local
concurrency. This means that mi_switch() is entered with the thread
locked but returns without. This dramatically simplifies scheduler
locking because we will not hold the schedlock while spinning on
blocked lock in switch.
This change has not been made to 4BSD but in principle it would be
more straightforward.
Discussed with: markj
Reviewed by: kib
Tested by: pho
Differential Revision: https://reviews.freebsd.org/D22778
Eliminate recursion from most thread_lock consumers. Return from
sched_add() without the thread_lock held. This eliminates unnecessary
atomics and lock word loads as well as reducing the hold time for
scheduler locks. This will eventually allow for lockless remote adds.
Discussed with: kib
Reviewed by: jhb
Tested by: pho
Differential Revision: https://reviews.freebsd.org/D22626
s/BIT_NAND/BIT_ANDNOT/, and for CPU and DOMAINSET too. The actual
implementation is "and not" (or "but not"), i.e. A but not B.
Fortunately this does appear to be what all existing callers want.
Don't supply a NAND (not (A and B)) operation at this time.
Discussed with: jeff
Reviewed by: cem
Sponsored by: Dell EMC Isilon
Differential Revision: https://reviews.freebsd.org/D22791
This fixes a regression after r355311. Specifically, sched_preempt()
may trigger a context switch by calling thread_lock(), since
thread_lock() calls critical_exit() in its slow path and the interrupted
thread may have already been marked for preemption. This would happen
before tdq_ipipending is cleared, blocking further preemption IPIs. The
CPU can be left in this state indefinitely if the interrupted thread
migrates.
Rename tdq_ipipending to tdq_owepreempt. Any switch satisfies a remote
preemption request, so clear tdq_owepreempt in sched_switch() instead of
sched_preempt() to avoid subtle problems of the sort described above.
Reviewed by: jeff, kib
Sponsored by: The FreeBSD Foundation
Differential Revision: https://reviews.freebsd.org/D22758
Use of CPU_FFS() to implement CPUSET_FOREACH() allows to save up to ~0.5%
of CPU time on 72-thread SMT system doing 80K IOPS to NVMe from one thread.
MFC after: 1 month
Sponsored by: iXsystems, Inc.
I've noticed that I missed intr check at one more SCHED_AFFINITY(),
so instead of adding one more branching I prefer to remove few.
Profiler shows the function CPU time reduction from 0.24% to 0.16%.
MFC after: 1 month
Sponsored by: iXsystems, Inc.
Doing some tests with very high interrupt rates I've noticed that one of
conditions I added in r232207 to make interrupt threads in most cases
run on local CPU never worked as expected (worked only if previous time
it was executed on some other CPU, that is quite opposite). It caused
additional CPU usage to run full CPU search and could schedule interrupt
threads to some other CPU.
This patch removes that code and instead reuses existing non-interrupt
code path with some tweaks for interrupt case:
- On SMT systems, if current thread is idle, don't look on other threads.
Even if they are busy, it may take more time to do fill search and bounce
the interrupt thread to other core then execute it locally, even sharing
CPU resources. It is other threads should migrate, not bound interrupts.
- Try hard to keep interrupt threads within LLC of their original CPU.
This improves scheduling cost and supposedly cache and memory locality.
On a test system with 72 threads doing 2.2M IOPS to NVMe this saves few
percents of CPU time while adding few percents to IOPS.
MFC after: 1 month
Sponsored by: iXsystems, Inc.
- there is no need to take the process lock to iterate the thread
list after single-threading is enforced
- typically there are no mutexes to clean up (testable without taking
the global umtx lock)
- typically there is no need to adjust the priority (testable without
taking thread lock)
Reviewed by: kib
Sponsored by: The FreeBSD Foundation
Differential Revision: https://reviews.freebsd.org/D20160
initial call to sched_balance() during startup is meant to initialize
balance_ticks, but does not actually do that since smp_started is
still zero at that time. Since balance_ticks does not get set,
there are no further calls to sched_balance(). Fix this by setting
balance_ticks in sched_initticks() since we know the value of
balance_interval at that time, and eliminate the useless startup
call to sched_balance(). We don't need to randomize the intial
value of balance_ticks.
Since there is now only one call to sched_balance(), we can hoist
the tests at the top of this function out to the caller and avoid
the overhead of the function call when running a SMP kernel on UP
hardware.
PR: 223914
Reviewed by: kib
MFC after: 2 weeks
On arm64 (and possible other architectures) we are unable to use static
DPCPU data in kernel modules. This is because the compiler will generate
PC-relative accesses, however the runtime-linker expects to be able to
relocate these.
In preparation to fix this create two macros depending on if the data is
global or static.
Reviewed by: bz, emaste, markj
Sponsored by: ABT Systems Ltd
Differential Revision: https://reviews.freebsd.org/D16140
search for a thread to steal inside a critical section. Since this
allows the search to be preempted, restart the search if preemption
happens since the search results found earlier may no longer be
valid.
Decrease the latency of starting a thread that may be assigned to
this CPU during the search by polling for incoming threads during
the search and switching to that thread instead of continuing the
search.
Test for stale search results and restart the search before going
through the expense of calling tdq_lock_pair(). Retry some tests
after grabbing the locks since things may have changed while waiting
to get both locks.
Eliminate special case handling for stealing from an SMT peer that
uses 1 as the steal threshold. This can only succeed if a thread
has been assigned but our SMT peer has not yet started executing
it. This is quite rare and when it happens the other SMT thread
is generally waiting for the same tdq lock that we hold. Basically
both SMT threads are racing to grab the same spin lock.
Add the kern.sched.always_steal knob from a ULE patch by jeff@.
Incorporate another idea from Jeff's ULE patch. If the sched_switch()
detects that the CPU is about to go idle, try to steal a thread
before switching to the idle thread. Since the search for a thread
to steal has to be done inside a critical section in this context,
limit the impact on latency by adding the knob kern.sched.trysteal_limit
to limit the topological distance of the search and don't restart
the search if we detect stale results. If this search can't find
an stealable thread, the idle loop can do a more complete search.
Also poll for threads being assigned to this CPU during the search
and switch to them instead of continuing the search. This change
is responsibile for the majority of the improvement in parallel
buildworld times.
In sched_balance_group() change the minimum threshold from stealing
a thread from 1 to 2. Poaching a newly assigned thread from a CPU
that is waking up hasn't yet switched to that thread from idle is
likely very rare and is likely to have the same lock race as is
seen when stealing threads in the idle loop. Also use tdq_notify()
to kick the destintation CPU instead of always sending an IPI.
Update a stale comment, the number of transferable threads is not
calculated.
Reviewed by: kib (earlier version)
Comments by: avg, jeff, mav
MFC after: 1 month
Differential Revision: https://reviews.freebsd.org/D12130
Fix a bug when the system has no CPU 0. When created, threads were implicitly assigned to CPU 0.
This had no practical effect since a real CPU was chosen immediately by the scheduler. However,
on systems without a CPU 0, sched_ule attempted to access the scheduler queue of the "old" CPU
when assigned the initial choice of the old one. This caused an attempt to use illegal memory
and a crash (or, more usually, a deadlock). Fix this by assigned new threads to the BSP
explicitly and add some asserts to see that this problem does not recur.
Authored by: Nathan Whitehorn <nwhitehorn@freebsd.org>
Submitted by: Wojciech Macek <wma@semihalf.com>
Obtained from: Semihalf
Differential revision: https://reviews.freebsd.org/D13932
userspace to control NUMA policy administratively and programmatically.
Implement domainset based iterators in the page layer.
Remove the now legacy numa_* syscalls.
Cleanup some header polution created by having seq.h in proc.h.
Reviewed by: markj, kib
Discussed with: alc
Tested by: pho
Sponsored by: Netflix, Dell/EMC Isilon
Differential Revision: https://reviews.freebsd.org/D13403
started, but also by the turnstiles to mark a thread as runnable for
all locks, for instance sleepqueues do:
setrunnable()->sched_wakeup()->sched_add()
In r326218 code was added to allow booting from non-zero CPU numbers
by setting the ts_cpu field inside the ULE scheduler's sched_add()
function. This had an undesired side-effect that prior sched_pin() and
sched_bind() calls got disregarded. This patch fixes the
initialization of the ts_cpu field for the ULE scheduler to only
happen once when the initial thread is constructed during system
init. Forking will then later on ensure that a valid ts_cpu value gets
copied to all children.
Reviewed by: jhb, kib
Discussed with: nwhitehorn
MFC after: 1 month
Differential revision: https://reviews.freebsd.org/D13298
Sponsored by: Mellanox Technologies
Mainly focus on files that use BSD 2-Clause license, however the tool I
was using misidentified many licenses so this was mostly a manual - error
prone - task.
The Software Package Data Exchange (SPDX) group provides a specification
to make it easier for automated tools to detect and summarize well known
opensource licenses. We are gradually adopting the specification, noting
that the tags are considered only advisory and do not, in any way,
superceed or replace the license texts.
