the bitrig patrick_smpns branch has been created by patrick.

it is 192 commits behind master, and 14 commits ahead.

commit a749e70cdb0a094a202988bee485f181180d9815
diff: https://github.com/bitrig/bitrig/commit/a749e70
author: Patrick Wildt <[email protected]>
date: Sat Aug 23 22:52:57 2014 +0200

Revert to nearly the old softintr_ API, but keep ithreads for it.

M       sys/arch/amd64/include/intr.h
M       sys/conf/files
M       sys/dev/ic/com.c
M       sys/dev/usb/usb.c
M       sys/kern/kern_clock.c
M       sys/kern/kern_ithread.c
M       sys/kern/kern_sig.c
A       sys/kern/kern_softintr.c
M       sys/net/netisr.c
M       sys/net/netisr.h
M       sys/net/pipex.c
M       sys/sys/ithread.h
A       sys/sys/softintr.h

commit a195dde40cc19de073d95b38d8936f2840459049
diff: https://github.com/bitrig/bitrig/commit/a195dde
author: Christiano Haesbaert <[email protected]>
date: Fri Jun 27 16:27:14 2014 +0200

Turnstile lock implementation, turn kernel lock into one.

BMTX stands for "blocking mutex", and it's implemented as a turnstile
exclusive recursive lock.

The Kernel Lock has been made a bmtx.

Background
~~~~~~~~~~

I've been tracking the performance impacts of the smpns changes and
the one thing who really hurt performance was turning kernel lock into
a rrwlock. The number of IPIs when building the kernel or world would
go over 5k/s, much higher than the <1k/s of the original mplock
implementation. That should be no surprise, since rrwlock is an always
sleeping lock which goes through wakeup/tsleep on every contention.

Performance also degraded considerably, baseline would build an image
in ~4m, taking ~5m25s after the change to rrwlock.

Therefore, the motivation for a turnstile lock was pure performance
wise, as they hold the same semantics as a sleeping lock like rwlock.

Semantics
~~~~~~~~~

 o Bmtx uses no critical section and doesn't block interrupts on any
way, they are resilient to preemption. They are adequate to interlock
against interrupt threads, so that should become the normal lock
throughout the kernel. This is the attempt of having synchronous
locking across interrupts and kernel threads.
 o Recursion is allowed, options for forbiding it in the future should
be discussed.
 o Interleaving on unlocking is allowed, the following is legal:

bmtx_lock(a);
bmtx_lock(b);
bmtx_unlock(a);     OR     bmtx_unlock(b);
bmtx_unlock(b);            bmtx_unlock(a);

 o Sleeping with a bmtx held is legal.

Implementation
~~~~~~~~~~~~~~

 o The cost for an uncontested operation is never higher than the
maximum cost of a single compare-and-set operation, average cost on a
cached situation is less than 50cycles per pair of lock/unlock on my
haswell i5 2.4ghz, considering function call overhead to be 0. We can
easily cut the function call overhead by making the function into a
macro that just tries the fast operation outside (one compare-and-set
call), if it fails, it calls into the function.

o Bmtx synchronizes across one atomic word called bmtx_lock, in which we
store:

>From bits 2-31 or 2-63: the curproc address structure.
The first two bits are abused to store two flags which are not mutually 
exclusive.
BMTX_RECURSE - signals that this bmtx was recursed.
BMTX_WAITERS - signals that there are sleepers waiting for the acquisition of 
the lock.

o Bmtx is a turnstile lock since it tries to spin _as long_ as the
owner of the lock is currently _running_ on another cpu. Under all
other conditions, the caller blocks. Blocking means setting the
BMTX_WAITERS in the atomic word and inserting itself in the bmtx's
waitqueue.

o To be able to block on the lock, we can't depend only on the atomic
word, we need a way to interlock with the other cpu, so that we don't
miss a wakeup. This is implemented using SCHED_LOCK as an interlock,
this allows us to overcome the lost-wakeup race. The algorithm is
something like this:

1 spin on the lock until holder is not SONPROC (running).  The test
for SONPROC is done on a racy fashion, which follows the process
structure and checks p->p_stat.

2 If we think the owner appears to not be SONPROC, we interlock with
the SCHED_LOCK and _retest_ the condition, therefore, recovering from
the race.

3 Assuming the owner is indeed blocked, we set BMTX_WAITERS and put
ourselves in the queue.

On the unlock path, the fast path only attempts to clear ownership if
BMTX_WAITERS is _not_ set, if it is, we need the slower path, where we
interlock with SCHED_LOCK, saving ourselves from the race.

In the future, when the scheduler doesn't suck horse-ass, we will
interlock with an arbitrary lock, something like a stripelock, instead
of a global SCHED_LOCK. We could do it with mutexes and msleep now, but
this is so ridiculous as it's just another lock to pass ownership for
the SCHED_LOCK. A performant implementation requires the scheduler to
expose the underlying interlock to the caller.

CAVEATS
~~~~~~~

On the lock path, when testing if the owner is SONPROC, we have the
following comment:
        /* XXX if we are preempted here, high chance p might have gone and
         * we're fucked. This is an "accepted" race in freebsd because of the
         * way they recycle the td structures, it might not be to us. We could
         * think of a crit_enter()/crit_leave().

We can not safely test p and deference p->p_stat, the process might have
gone through sched_exit() and p might be freed, the timing would have
to be very very very evil, I've never hit this situation and I'm
buying it for now.

Alternative solutions is changing the way we alocate proc{} (like
freebsd), or keeping track of all held mutexes and updating a state on
the mutex itself saying "OWNER_IS_RUNNING".

Results
~~~~~~~

My building times went to be as performant as baseline, sometimes
slightly faster, I could only test this in vmware so far, so I can't
have precise numbers. I would go as far as saying that it is _no_
worse than the baseline kernel.

Building time improved around 25% against the rrwlock implementation.

IPIs were reduced from 5000-6000 to 200-1200/s.

Statistics
~~~~~~~~~~

I've collected the number of fast vs slow paths by defining BMTX_STATS
on the lock, as shown below, the number of slow paths are
insignificant compared to the fast paths:

bmtx_fast_locks              21017802
bmtx_fast_unlocks            31133706
bmtx_recursed_locks          524294
bmtx_recursed_unlocks        524279
bmtx_lock_blocks             33174
bmtx_lock_blocks_cancelled   29742
bmtx_lock_blocks_slept       3862
bmtx_spun_locks              10117383
bmtx_blocked_unlocks         3866
bmtx_unlock_has_waiter       3866

M       sys/conf/files
A       sys/kern/kern_bmtx.c
M       sys/kern/kern_lock.c
A       sys/sys/bmtx.h
M       sys/sys/mplock.h
M       sys/sys/systm.h

commit 4361c2402504b2ad65199bdb3ec31605a8de1520
diff: https://github.com/bitrig/bitrig/commit/4361c24
author: Christiano Haesbaert <[email protected]>
date: Fri Jun 6 11:42:01 2014 +0200

Stop using the sleepqueue for interrupt threads.

There is no need for it, it cuts some hacks and it's less
instructions. Since we always wakeup from ithread_run(), we always
know the proc to awake, the solely point of the sleepqueue is to be a
storage to be found via ident hashing.

M       sys/kern/kern_ithread.c

commit 3db5da72d79896955c1a7cbf7a57785a025ed5d0
diff: https://github.com/bitrig/bitrig/commit/3db5da7
author: Christiano Haesbaert <[email protected]>
date: Sun May 11 17:22:11 2014 +0200

No need for running interrupt handlers in critical sections anymore.

Preemption correctly relies on the priority to do preemption, having a
ithread->ithread preemption due to ithread sleeping at a lower
priority should be ok.

M       sys/kern/kern_ithread.c

commit f8f0166a31320a6d367595411286945f9af5975b
diff: https://github.com/bitrig/bitrig/commit/f8f0166
author: Christiano Haesbaert <[email protected]>
date: Sat Mar 1 11:08:55 2014 +0100

Initial kernel preemption.

This diff introduces a rudimentary form of kernel preemption, it
allows interrupt threads to preempt other kernel processing, like
userland doing syscalls or other normal kthreads.

Preemption is deferred if the current process is in a critical section.

When curproc preempts:
  o It must not be stolen to run on another cpu, it
is still SRUNable, but must therefore be pinned to the curcpu()
runqueue.
  o It must continue to hold all the locks it had, in our case:
rwlocks, including kernel lock.
  o In case the _preempting_ process (not the preempted) tries to
acquire the kernel lock, but the holder was preempted, the
_preempting_ process blocks so that the _preempted_ might be resumed.

CAVEATS:

  o The way we test for preemption is still primitive, I'll only be
able to attack this properly once I introduce proper APIs into the
scheduler, but for that I want to make the scheduler modular, so that
we are able to play with different schedulers. So we need MORE APIs,
not LESS as _some_ people believe.
  o ci_want_resched which is responsible for preempting userland is
still present as at this point we need it, once I'm able to hack the
scheduler there should be only one way of preemption, so it will look
a bit funky until there.

Also introduce a sysctl (kern.preemption) which enables us to activate
or deactivate kernel preemption.

M       sys/kern/kern_crit.c
M       sys/kern/kern_exit.c
M       sys/kern/kern_ithread.c
M       sys/kern/kern_ktrace.c
M       sys/kern/kern_sched.c
M       sys/kern/kern_subr.c
M       sys/kern/kern_sysctl.c
M       sys/kern/sched_bsd.c
M       sys/kern/subr_xxx.c
M       sys/sys/proc.h
M       sys/sys/sched.h
M       sys/sys/syscall_mi.h
M       sys/sys/sysctl.h
M       sys/uvm/uvm_glue.c

commit 9d227a155c0dd0412eff26fb73725c15689274e8
diff: https://github.com/bitrig/bitrig/commit/9d227a1
author: Christiano Haesbaert <[email protected]>
date: Wed Feb 26 23:24:37 2014 +0100

Newly formed processes should start in a critical section.

They have the schedlock, and release on proc_trampoline_mp(), if the
newly formed process is not in a critical section, a clock interrupt
might attempt to grab the SCHED LOCK, or worse, we might preempt to
early in the future.

This paves the way for preemption and allows SCHED LOCK to be a mutex,
we only need guenther's fix for the recursive tsleep.

M       sys/kern/kern_fork.c
M       sys/sys/proc.h

commit 701ffe8997e35347fb31fec8d8eff1d6069bde20
diff: https://github.com/bitrig/bitrig/commit/701ffe8
author: Christiano Haesbaert <[email protected]>
date: Sun Feb 23 22:24:24 2014 +0100

Make sure we have a process context early and kill crit_escaped.

This was lost in a merge.

M       sys/arch/amd64/amd64/machdep.c
M       sys/kern/kern_crit.c

commit fafc84c4ae54c636b54e7b949f83cf739782bfad
diff: https://github.com/bitrig/bitrig/commit/fafc84c
author: Christiano F. Haesbaert <[email protected]>
date: Sat Feb 1 16:20:50 2014 +0100

Convert kernel lock to a rrwlock.

Introduce rrw_exit_all(), rrw_enter_cnt() and rw_held()/rrw_held().

Also, make the assert unlocked functions make sense, now we assert
that _we_ don't hold the lock, asserting that no one has the lock is
impossible.

We can't sleep on proc_trampoline_mp for the idle process, since it
should never be in a sleepqueue, I had to introduce a hack to make it
not acquire the kernel lock.

The correct fix is to pass a flag to fork1() that gets propagated to
proc_trampoline_mp().

M       sys/arch/amd64/amd64/intr.c
M       sys/kern/kern_fork.c
M       sys/kern/kern_ithread.c
M       sys/kern/kern_ktrace.c
M       sys/kern/kern_lock.c
M       sys/kern/kern_rwlock.c
M       sys/kern/kern_sched.c
M       sys/kern/kern_sig.c
M       sys/kern/kern_synch.c
M       sys/sys/mplock.h
M       sys/sys/rwlock.h
M       sys/sys/systm.h

commit 44cf47dad9d897a7dfe64e0bc0744c75c9fc9ddf
diff: https://github.com/bitrig/bitrig/commit/44cf47d
author: Christiano F. Haesbaert <[email protected]>
date: Tue Dec 31 02:23:56 2013 +0100

Fix compilation of !MULTIPROCESSOR.

Armani had fixed this a while ago in the old Openbsd tree, this is
just an updated version of his diff.

spotted by aalm.

M       sys/arch/amd64/amd64/intr.c
M       sys/kern/sched_bsd.c
M       sys/sys/systm.h

commit 1632b05fc6da65557879325be6b8484ab3fe1858
diff: https://github.com/bitrig/bitrig/commit/1632b05
author: Christiano F. Haesbaert <[email protected]>
date: Mon Dec 30 01:32:35 2013 +0100

Remove __mp_release_all_but_one().

M       sys/arch/amd64/amd64/lock_machdep.c
M       sys/arch/amd64/include/mplock.h
M       sys/sys/mplock.h

commit baf92e9ee9095ca97e0a8cfe57c1515ea0fab5f1
diff: https://github.com/bitrig/bitrig/commit/baf92e9
author: Christiano F. Haesbaert <[email protected]>
date: Mon Dec 30 01:27:11 2013 +0100

Make so that mi_switch() won't relock sched_lock.

The caller needs to reacquire it if desired. More complex paths like
proc_stop() were left like before, so they reacquire the lock.

Almost every user of mi_switch() just do a SCHED_UNLOCK() after
return, so this commit prevents a LOCK/UNLOCK on a contented lock.

For this to work, you can never mi_switch with a recursive sched_lock,
and this is asserted.

Thanks to grunk@ for questioning me on the diff and finding a bug
which should fix suspend/resume, it was a double acquisition of
SCHED_LOCK on sched_idle(). natano@ reported the bug.

M       sys/kern/kern_ithread.c
M       sys/kern/kern_sched.c
M       sys/kern/kern_sig.c
M       sys/kern/kern_synch.c
M       sys/kern/sched_bsd.c

commit 56d99599d28ebc0fb49c32738b1fe9f4a3223940
diff: https://github.com/bitrig/bitrig/commit/56d9959
author: Christiano F. Haesbaert <[email protected]>
date: Mon Dec 30 01:21:49 2013 +0100

Make __mp_lock_held() return the lock depth.

M       sys/arch/amd64/amd64/lock_machdep.c

commit a26efc0574c5723f4d7c421bc4a7a6a25e43b252
diff: https://github.com/bitrig/bitrig/commit/a26efc0
author: Christiano F. Haesbaert <[email protected]>
date: Mon Dec 30 00:39:49 2013 +0100

Rework the SCHED_LOCK vs KERNEL_LOCK dance in mi_switch().

This avoids the lock/relock to fix lock ordering on mi_switch for all
the cases except one.

Before the idea was:
        - mi_switch() releases all kernel locks before context switching.
        - save the count in the stack.
        - context switch with SCHED_LOCK held.
        - wakeup, but to assure lock ordering it needs to:
                - release SCHED_LOCK
                - reacquire KERNEL_LOCK
                - acquire SCHED_LOCK

With this diff, the caller is responsible for doing this, so you can't
enter mi_switch with kernel lock now, you must release/reacquire
yourself, being careful to always grab SCHED_LOCK before releasing the
kernel locks, if not you lose atomicity before you can.

Since we can grab kernel lock within a critical section,
KERNEL_RELOCK_ALL must do the release/reenter dance.

Next step is to make mi_switch() return with SCHED_LOCK unlocked, if
the caller wants it is his job, he _already_ lost atomicity. Most of
the cases mi_switch() relocks SCHED_LOCK only for the caller to unlock
it, which is pretty stupid.

M       sys/kern/kern_ithread.c
M       sys/kern/kern_lock.c
M       sys/kern/kern_sched.c
M       sys/kern/kern_sig.c
M       sys/kern/kern_synch.c
M       sys/kern/sched_bsd.c
M       sys/sys/proc.h
M       sys/sys/systm.h

commit 4287c0b4b9fe7d1583df20adf9f1224bb3e0aaad
diff: https://github.com/bitrig/bitrig/commit/4287c0b
author: Christiano F. Haesbaert <[email protected]>
date: Sat May 25 13:19:17 2013 +0200

New interrupt model, move away from IPLs.

This diff changes the interrupt model to something very similar to
what other modern unixes like Solaris, FreeBSD, DragonflyBSD and
linux. It also introduces critical sections.

Interrupts except clock and ipi are handled by interrupt threads, when
an interrupt fires, the only job for the small interrupt stub is to
schedule the corresponding interrupt thread, when the ithread gets
scheduled, it services the interrupt.

The kernel must be made preemptive, so that interrupt threads may
preempt the current running code.

In this model, you normally never block interrupts, you rely solely on
locks to protect the code, if the ithread preempts the running code,
and it tries to acquire a contested lock, it blocks (sleeps) and gets
resumed only when that lock is released.

This allows us to have lower interrupt latency, as instead of blocking
interrupts for a long section, you can fine grain that section in a
specific lock. Instead of raising to IPL_BIO and preventing softclock
from running for example, we can make softclock preempt the old
IPL_BIO and only block if there is an actual lock contention.

This was first implemented in Solaris 2 (Kleiman & Eykholt circa 93),
they demonstrated that this model, after properly locked, changed the
worst case latency from 1s to <1ms on a single-core 40mhz machine.

It also allows for us to properly fight livelocks in the future, as
the scheduler will make sure "everything runs at some point".

The the kernel can now be synchronous with regard to locking, as you
can use the same lock to interlock against interrupts or other normal
kthreads.

You can only block on locks if you have a process context, so you
still need a way to block interrupts from normal (read: not ithread
interrupts), like clock and ipi. For that we introduce critical
sections, which block everything, in practice they're only used to
protect scheduler data and on mutexes, so they are very short.

In the future critical sections will also be the only thing that
prevents kernel preemption.

In order to prevent deadlocks, you must never be preempted while
holding a spinlock, so a critical section is implied there, this is
also akin to what every system does.

In this present state, kernel preemption has not been implemented, all
threads <IPL_CLOCK were moved to ithreads, there is no observable loss
of performance, it's been stable since the last half a year.

All spl calls <IPL_CLOCK were made no-ops, while >= IPL_CLOCK were
replaced by critical sections.

Most of the old assembly code has been rewritten in C, just because I
refuse to maintain unecessary asm blocks.

The next steps are:
        o Turn kernel lock into a rrwlock.
        o Enable kernel preemption, at this point all the interrupt
        interlocking will be done through kernel lock.
        o Decide on which subsystem to release first, having a wide subsytem 
lock.

M       sys/arch/amd64/amd64/cpu.c
M       sys/arch/amd64/amd64/db_interface.c
M       sys/arch/amd64/amd64/fpu.c
M       sys/arch/amd64/amd64/genassym.cf
M       sys/arch/amd64/amd64/intr.c
M       sys/arch/amd64/amd64/ipi.c
M       sys/arch/amd64/amd64/lapic.c
M       sys/arch/amd64/amd64/locore.S
M       sys/arch/amd64/amd64/machdep.c
M       sys/arch/amd64/amd64/mp_setperf.c
D       sys/arch/amd64/amd64/softintr.c
M       sys/arch/amd64/amd64/spl.S
M       sys/arch/amd64/amd64/trap.c
M       sys/arch/amd64/amd64/vector.S
M       sys/arch/amd64/amd64/via.c
M       sys/arch/amd64/conf/files.amd64
M       sys/arch/amd64/include/atomic.h
M       sys/arch/amd64/include/cpu.h
M       sys/arch/amd64/include/frame.h
M       sys/arch/amd64/include/frameasm.h
M       sys/arch/amd64/include/intr.h
M       sys/arch/amd64/include/intrdefs.h
M       sys/arch/amd64/isa/clock.c
M       sys/arch/arm/arm/db_interface.c
M       sys/arch/arm/arm/pmap.c
M       sys/conf/files
M       sys/dev/acpi/acpi.c
M       sys/dev/cardbus/cardbus.c
M       sys/dev/ic/com.c
M       sys/dev/ic/comvar.h
M       sys/dev/ic/elink3.c
M       sys/dev/ic/i82365.c
M       sys/dev/ic/tcic2.c
M       sys/dev/ic/vga.c
M       sys/dev/ic/vga_subr.c
M       sys/dev/isa/if_ef_isapnp.c
M       sys/dev/isa/pcppi.c
M       sys/dev/onewire/onewire_bitbang.c
M       sys/dev/pci/drm/drm_atomic.h
M       sys/dev/pci/drm/radeon/radeon_kms.c
M       sys/dev/pci/mbg.c
M       sys/dev/pci/pccbb.c
M       sys/dev/pci/pci.c
M       sys/dev/pci/pci_map.c
M       sys/dev/pci/pciide.c
M       sys/dev/pcmcia/if_xe.c
M       sys/dev/rasops/rasops.c
M       sys/dev/sdmmc/sdmmc_io.c
M       sys/dev/usb/ehci.c
M       sys/dev/usb/ohci.c
M       sys/dev/usb/uhci.c
M       sys/dev/usb/umidi.c
M       sys/dev/usb/usb.c
M       sys/dev/usb/usbdivar.h
M       sys/dev/usb/usbf_subr.c
M       sys/dev/usb/usbfvar.h
M       sys/dev/wsfont/wsfont.c
M       sys/kern/init_main.c
M       sys/kern/kern_clock.c
A       sys/kern/kern_crit.c
M       sys/kern/kern_event.c
M       sys/kern/kern_exec.c
M       sys/kern/kern_fork.c
A       sys/kern/kern_ithread.c
M       sys/kern/kern_lock.c
M       sys/kern/kern_mutex.c
M       sys/kern/kern_proc.c
M       sys/kern/kern_resource.c
M       sys/kern/kern_sched.c
M       sys/kern/kern_sensors.c
M       sys/kern/kern_sig.c
M       sys/kern/kern_synch.c
M       sys/kern/kern_time.c
M       sys/kern/kern_timeout.c
M       sys/kern/sched_bsd.c
M       sys/kern/subr_disk.c
M       sys/kern/subr_evcount.c
M       sys/kern/subr_hibernate.c
M       sys/kern/subr_log.c
M       sys/kern/subr_pool.c
M       sys/kern/subr_prf.c
M       sys/kern/subr_prof.c
M       sys/kern/subr_xxx.c
M       sys/kern/sys_generic.c
M       sys/kern/sys_process.c
M       sys/kern/vfs_subr.c
M       sys/kern/vfs_sync.c
M       sys/lib/libkern/libkern.h
M       sys/net/if_tun.c
M       sys/net/netisr.c
M       sys/net/netisr.h
M       sys/net/pipex.c
A       sys/sys/ithread.h
M       sys/sys/proc.h
M       sys/sys/sched.h
M       sys/sys/systm.h
M       sys/sys/timeout.h
M       sys/uvm/uvm_map.c

Reply via email to