And last but not least we need to make sure the scheduler tick never
preempts a nosched context.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Arnd Bergmann <arnd.bergmann@de.ibm.com>
Signed-off-by: Jeremy Kerr <jk@ozlabs.org>
Signed-off-by: Paul Mackerras <paulus@samba.org>
spu_deactivate should never be called for nosched contets. Put in
a check so we can print a stacktrace and exit early in case it
happes erroneously.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Arnd Bergmann <arnd.bergmann@de.ibm.com>
Signed-off-by: Jeremy Kerr <jk@ozlabs.org>
Signed-off-by: Paul Mackerras <paulus@samba.org>
Add a cpus_allowed allowed filed to struct spu_context so that we always
use the cpu mask of the owning thread instead of the one happening to
call into the scheduler. Also use this information in
grab_runnable_context to avoid spurious wakeups.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Arnd Bergmann <arnd.bergmann@de.ibm.com>
Signed-off-by: Jeremy Kerr <jk@ozlabs.org>
Signed-off-by: Paul Mackerras <paulus@samba.org>
Update scheduling information on every spu_run to allow for setting
threads to realtime priority just before running them. This requires
some slightly ugly code in spufs_run_spu because we can just update
the information unlocked if the spu is not runnable, but we need to
acquire the active_mutex when it is runnable to protect against
find_victim. This locking scheme requires opencoding
spu_acquire_runnable in spufs_run_spu which actually is a nice cleanup
all by itself.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Arnd Bergmann <arnd.bergmann@de.ibm.com>
Signed-off-by: Jeremy Kerr <jk@ozlabs.org>
Signed-off-by: Paul Mackerras <paulus@samba.org>
Print out a few scheduler tuning parameters when we've compiled
with DEBUG defined.
Signed-off-by: Jeremy Kerr <jk@ozlabs.org>
Signed-off-by: Paul Mackerras <paulus@samba.org>
The current timeslice code mixes 'jiffies' up with 'spesched ticks'. This
change correctly defines the number of time slices each SPE contexts is
given, and clarifies the comment.
This brings the default timeslice for SPE contexts into a reasonable
range.
Signed-off-by: Jeremy Kerr <jk@ozlabs.org>
Signed-off-by: Paul Mackerras <paulus@samba.org>
Enable preemptive scheduling for non-RT contexts.
We use the same algorithms as the CPU scheduler to calculate the time
slice length, and for now we also use the same timeslice length as the
CPU scheduler. This might be not enough for good performance and can be
changed after some benchmarking.
Note that currently we do not boost the priority for contexts waiting
on the runqueue for a long time, so contexts with a higher nice value
could starve ones with less priority. This could easily be fixed once
the rework of the spu lists that Luke and I discussed is done.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Arnd Bergmann <arnd.bergmann@de.ibm.com>
Signed-off-by: Jeremy Kerr <jk@ozlabs.org>
Signed-off-by: Paul Mackerras <paulus@samba.org>
Get rid of the scheduler workqueues that complicated things a lot to
a dedicated spu scheduler thread that gets woken by a traditional
scheduler tick. By default this scheduler tick runs a HZ * 10, aka
one spu scheduler tick for every 10 cpu ticks.
Currently the tick is not disabled when we have less context than
available spus, but I will implement this later.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Arnd Bergmann <arnd.bergmann@de.ibm.com>
Signed-off-by: Jeremy Kerr <jk@ozlabs.org>
Signed-off-by: Paul Mackerras <paulus@samba.org>
Nosched context sould never be scheduled out, thus we must not
deactivate them in spu_yield ever.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jeremy Kerr <jk@ozlabs.org>
Signed-off-by: Paul Mackerras <paulus@samba.org>
Fix the race between checking for contexts on the runqueue and actually
waking them in spu_deactive and spu_yield.
The guts of spu_reschedule are split into a new helper called
grab_runnable_context which shows if there is a runnable thread below
a specified priority and if yes removes if from the runqueue and uses
it. This function is used by the new __spu_deactivate hepler shared
by preemption and spu_yield to grab a new context before deactivating
a specified priority and if yes removes if from the runqueue and uses
it. This function is used by the new __spu_deactivate hepler shared
by preemption and spu_yield to grab a new context before deactivating
the old one.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Arnd Bergmann <arnd.bergmann@de.ibm.com>
Signed-off-by: Jeremy Kerr <jk@ozlabs.org>
Signed-off-by: Paul Mackerras <paulus@samba.org>
Remove includes of <linux/smp_lock.h> where it is not used/needed.
Suggested by Al Viro.
Builds cleanly on x86_64, i386, alpha, ia64, powerpc, sparc,
sparc64, and arm (all 59 defconfigs).
Signed-off-by: Randy Dunlap <randy.dunlap@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Addition to stop_wq needs to happen before adding to the runqeueue and
under the same lock so that we don't have a race window for a lost
wake up in the spu scheduler.
Signed-off-by: Luke Browning <lukebrowning@us.ibm.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Arnd Bergmann <arnd.bergmann@de.ibm.com>
A single context should only be woken once, and we should not have
more wakeups for a given priority than the number of contexts on
that runqueue position.
Also add some asserts to trap future problems in this area more
easily.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Arnd Bergmann <arnd.bergmann@de.ibm.com>
set_bit does not guarantee ordering on powerpc, so using it
for communication between threads requires explicit
mb() calls.
Signed-off-by: Arnd Bergmann <arnd.bergmann@de.ibm.com>
To not lose a spu thread we need to make sure it always gets put back
on the runqueue. In find_victim aswell as in the scheduler tick as done
in the previous patch.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Arnd Bergmann <arnd.bergmann@de.ibm.com>
To not lose a spu thread we need to make sure it always gets put back
on the runqueue.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Acked-by: Jeremy Kerr <jk@ozlabs.org>
Signed-off-by: Arnd Bergmann <arnd.bergmann@de.ibm.com>
The scheduler workqueue may rearm itself and deadlock when we try to stop
it. Put a flag in place to avoid skip the work if we're tearing down
the context.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Arnd Bergmann <arnd.bergmann@de.ibm.com>
There is no reason to yield the CPU in spu_yield - if the backing
thread reenters spu_run it gets added to the end of the runqueue for
it's priority. So the yield is just a slowdown for the case where
we have higher priority contexts waiting.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Acked-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Paul Mackerras <paulus@samba.org>
The SPU code doesn't properly invalidate SPUs SLBs when necessary,
for example when changing a segment size from the hugetlbfs code. In
addition, it saves and restores the SLB content on context switches
which makes it harder to properly handle those invalidations.
This patch removes the saving & restoring for now, something more
efficient might be found later on. It also adds a spu_flush_all_slbs(mm)
that can be used by the core mm code to flush the SLBs of all SPEs that
are running a given mm at the time of the flush.
In order to do that, it adds a spinlock to the list of all SPEs and move
some bits & pieces from spufs to spu_base.c
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
For SCHED_RR tasks we can do some really trivial timeslicing. Basically
we fire up a time for every scheduler tick that searches for a higher
or same priority thread that is on the runqueue and if there is one
context switches to it. Because we can't lock spus from timer context
we actually run this from a delayed runqueue instead of a timer.
A nice optimization would be to skip the actual priority bitmap search
when there are less contexts than physical spus available. To implement
this I need a so far unpublished patch from Andre, and it will be added
after we have that patch in.
Note that right now we only do the time slicing for SCHED_RR tasks.
The code would work for SCHED_OTHER tasks aswell, but their prio
value is defered from the one the PPU thread has at time of spu_run,
and using this for spu scheduling decisions would make the code very
unfair. SCHED_OTHER support will be enabled once we the spu scheduler
knows how to calculcate cpu_context.prio (very soon)
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Arnd Bergmann <arnd.bergmann@de.ibm.com>
use DECLARE_BITMAP in the spu scheduler instead of reimplementing it.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Arnd Bergmann <arnd.bergmann@de.ibm.com>
If we start a spu context with realtime priority we want it to run
immediately and not wait until some other lower priority thread has
finished. Try to find a suitable victim and use it's spu in this
case.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Arnd Bergmann <arnd.bergmann@de.ibm.com>
Give spu_yield a kerneldoc comment and remove the old comment
documenting spu_activate, spu_deactive and spu_yield as all of them
now have descriptive kerneldoc comments of their own.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Arnd Bergmann <arnd.bergmann@de.ibm.com>
If we call spu_remove_from_active_list that spu is always guaranteed
to be on the active list and in runnable state, so we can simply
do a list_del to remove it and unconditionally take the was_active
codepath.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Arnd Bergmann <arnd.bergmann@de.ibm.com>
There is no need to directly wake up contexts in spu_activate when
called from spu_run, so add a flag to surpress this wakeup.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Arnd Bergmann <arnd.bergmann@de.ibm.com>
This is the biggest patch in this series, and it reworks the guts of
the spu scheduler runqueue mechanism:
- instead of embedding a waitqueue in the runqueue there is now a
simple doubly-linked list, the actual wakeups happen by reusing
the stop_wq in the spu context (maybe we should rename it one day)
- spu_free and spu_prio_wakeup are merged into a single spu_reschedule
function
- various functionality is split out into small helpers, and kerneldoc
comments are added in various places to document what's going on.
- spu_activate is rewritten into a tight loop by removing test for
various impossible conditions and using the infrastructure in this
patch.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Arnd Bergmann <arnd.bergmann@de.ibm.com>
It doesn't make any sense to have a priority field in the physical spu
structure. Move it into the spu context instead.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Arnd Bergmann <arnd.bergmann@de.ibm.com>
The r/w semaphore to lock the spus was overkill and can be replaced
with a mutex to make it faster, simpler and easier to debug. It also
helps to allow making most spufs interruptible in future patches.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Arnd Bergmann <arnd.bergmann@de.ibm.com>
Various cleanups to sched.c that don't change the global control flow:
- add kerneldoc comments to various functions
- add spu_ prefixes to various functions
- add/remove context from the runqueue in bind/unbind_context as
it's part of the logical operation
- add a call to put_active_spu to spu_unbind_contex as it's logically
part of the unbind operation
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Arnd Bergmann <arnd.bergmann@de.ibm.com>
Only bind_context/unbind_context change the spu context state. Thus
we can move all assignents of SPU_STATE_RUNNABLE into bind_context,
which parallels the unbind side aswell.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Arnd Bergmann <arnd.bergmann@de.ibm.com>
unbind_context already sets the context state to SPU_STATE_SAVED, thus
the spu_deactivate callers don't need to do it again.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Arnd Bergmann <arnd.bergmann@de.ibm.com>
This adds an 'object-id' file that the spe library can
use to store a pointer to its ELF object. This was
originally meant for use by oprofile, but is now
also used by the GNU debugger, if available.
In order for oprofile to find the location in an spu-elf
binary where an event counter triggered, we need a way
to identify the binary in the first place.
Unfortunately, that binary itself can be embedded in a
powerpc ELF binary. Since we can assume it is mapped into
the effective address space of the running process,
have that one write the pointer value into a new spufs
file.
When a context switch occurs, pass the user value to
the profiler so that can look at the mapped file (with
some care).
Signed-off-by: Arnd Bergmann <arnd.bergmann@de.ibm.com>
Signed-off-by: Paul Mackerras <paulus@samba.org>
This tries to fix spufs so we have an interface closer to what is
specified in the man page for events returned in the third argument of
spu_run.
Fortunately, libspe has never been using the returned contents of that
register, as they were the same as the return code of spu_run (duh!).
Unlike the specification that we never implemented correctly, we now
require a SPU_CREATE_EVENTS_ENABLED flag passed to spu_create, in
order to get the new behavior. When this flag is not passed, spu_run
will simply ignore the third argument now.
Signed-off-by: Arnd Bergmann <arnd.bergmann@de.ibm.com>
Signed-off-by: Paul Mackerras <paulus@samba.org>
This patch adds NUMA support to the the spufs scheduler.
The new arch/powerpc/platforms/cell/spufs/sched.c is greatly
simplified, in an attempt to reduce complexity while adding
support for NUMA scheduler domains. SPUs are allocated starting
from the calling thread's node, moving to others as supported by
current->cpus_allowed. Preemption is gone as it was buggy, but
should be re-enabled in another patch when stable.
The new arch/powerpc/platforms/cell/spu_base.c maintains idle
lists on a per-node basis, and allows caller to specify which
node(s) an SPU should be allocated from, while passing -1 tells
spu_alloc() that any node is allowed.
Since the patch removes the currently implemented preemptive
scheduling, it is technically a regression, but practically
all users have since migrated to this version, as it is
part of the IBM SDK and the yellowdog distribution, so there
is not much point holding it back while the new preemptive
scheduling patch gets delayed further.
Signed-off-by: Arnd Bergmann <arnd.bergmann@de.ibm.com>
Signed-off-by: Paul Mackerras <paulus@samba.org>
This changes the hypervisor abstraction of setting cpu affinity to a
higher level to avoid platform dependent interrupt controller
routines. I replaced spu_priv1_ops:spu_int_route_set() with a
new routine spu_priv1_ops:spu_cpu_affinity_set().
As a by-product, this change eliminated what looked like an
existing bug in the set affinity code where spu_int_route_set()
mistakenly called int_stat_get().
Signed-off-by: Geoff Levand <geoffrey.levand@am.sony.com>
Signed-off-by: Arnd Bergmann <arnd.bergmann@de.ibm.com>
Signed-off-by: Paul Mackerras <paulus@samba.org>
This patch adds a new file called 'mfc' to each spufs directory.
The file accepts DMA commands that are a subset of what would
be legal DMA commands for problem state register access. Upon
reading the file, a bitmask is returned with the completed
tag groups set.
The file is meant to be used from an abstraction in libspe
that is added by a different patch.
From the kernel perspective, this means a process can now
offload a memory copy from or into an SPE local store
without having to run code on the SPE itself.
The transfer will only be performed while the SPE is owned
by one thread that is waiting in the spu_run system call
and the data will be transferred into that thread's
address space, independent of which thread started the
transfer.
Signed-off-by: Arnd Bergmann <arnd.bergmann@de.ibm.com>
Signed-off-by: Paul Mackerras <paulus@samba.org>
For far, all SPU triggered interrupts always end up on
the first SMT thread, which is a bad solution.
This patch implements setting the affinity to the
CPU that was running last when entering execution on
an SPU. This should result in a significant reduction
in IPI calls and better cache locality for SPE thread
specific data.
Signed-off-by: Arnd Bergmann <arndb@de.ibm.com>
Signed-off-by: Paul Mackerras <paulus@samba.org>
checking bits manually might not be synchonized with
the use of set_bit/clear_bit. Make sure we always use
the correct bitops by removing the unnecessary
identifiers.
Signed-off-by: Arnd Bergmann <arndb@de.ibm.com>
Signed-off-by: Paul Mackerras <paulus@samba.org>
One of my last patches contained a broken line
from splitting out some other changes, this
restores a working version.
Signed-off-by: Arnd Bergmann <arndb@de.ibm.com>
Signed-off-by: Paul Mackerras <paulus@samba.org>
This patch reduces lock complexity of SPU scheduler, particularly
for involuntary preemptive switches. As a result the new code
does a better job of mapping the highest priority tasks to SPUs.
Lock complexity is reduced by using the system default workqueue
to perform involuntary saves. In this way we avoid nasty lock
ordering problems that the previous code had. A "minimum timeslice"
for SPU contexts is also introduced. The intent here is to avoid
thrashing.
While the new scheduler does a better job at prioritization it
still does nothing for fairness.
From: Mark Nutter <mnutter@us.ibm.com>
Signed-off-by: Arnd Bergmann <arndb@de.ibm.com>
Signed-off-by: Paul Mackerras <paulus@samba.org>
This patch makes it easier to preempt an SPU context by
having the scheduler hold ctx->state_sema for much shorter
periods of time.
As part of this restructuring, the control logic for the "run"
operation is moved from arch/ppc64/kernel/spu_base.c to
fs/spufs/file.c. Of course the base retains "bottom half"
handlers for class{0,1} irqs. The new run loop will re-acquire
an SPU if preempted.
From: Mark Nutter <mnutter@us.ibm.com>
Signed-off-by: Arnd Bergmann <arndb@de.ibm.com>
Signed-off-by: Paul Mackerras <paulus@samba.org>
spufs is rather noisy when debugging is enabled, this
turns off the messages for production use.
Signed-off-by: Arnd Bergmann <arndb@de.ibm.com>
Signed-off-by: Paul Mackerras <paulus@samba.org>
This adds a scheduler for SPUs to make it possible to use
more logical SPUs than physical ones are present in the
system.
Currently, there is no support for preempting a running
SPU thread, they have to leave the SPU by either triggering
an event on the SPU that causes it to return to the
owning thread or by sending a signal to it.
This patch also adds operations that enable accessing an SPU
in either runnable or saved state. We use an RW semaphore
to protect the state of the SPU from changing underneath
us, while we are holding it readable. In order to change
the state, it is acquired writeable and a context save
or restore is executed before downgrading the semaphore
to read-only.
From: Mark Nutter <mnutter@us.ibm.com>,
Uli Weigand <Ulrich.Weigand@de.ibm.com>
Signed-off-by: Arnd Bergmann <arndb@de.ibm.com>
Signed-off-by: Paul Mackerras <paulus@samba.org>
Christoph Hellwig