- Sep 05, 2008
-
-
Jeremy Kerr authored
We currently have a race for a free SPE. With one thread doing a spu_yield(), and another doing a spu_activate(): thread 1 thread 2 spu_yield(oldctx) spu_activate(ctx) __spu_deactivate(oldctx) spu_unschedule(oldctx, spu) spu->alloc_state = SPU_FREE spu = spu_get_idle(ctx) - searches for a SPE in state SPU_FREE, gets the context just freed by thread 1 spu_schedule(ctx, spu) spu->alloc_state = SPU_USED spu_schedule(newctx, spu) - assumes spu is still free - tries to schedule context on already-used spu This change introduces a 'free_spu' flag to spu_unschedule, to indicate whether or not the function should free the spu after descheduling the context. We only set this flag if we're not going to re-schedule another context on this SPU. Add a comment to document this behaviour. Signed-off-by:Jeremy Kerr <jk@ozlabs.org>
-
Jeremy Kerr authored
Commit 8d5636fb introduced a reference count on SPU contexts during find_victim, but this may cause a leak in the reference count if we later find a better contender for a context to unschedule. Change the reference to after we've found our victim context, so we don't do the extra get_spu_context(). Signed-off-by:
-
- Aug 19, 2008
-
-
Ilpo Järvinen authored
Signed-off-by:
Ilpo Järvinen <ilpo.jarvinen@helsinki.fi> Signed-off-by:
-
- Aug 14, 2008
-
-
Jeremy Kerr authored
Based on an original patch from Christoph Hellwig <hch@lst.de>. Currently, there is a possible reference-after-free in the spusched code - contexts may be freed after we have released their state_mutex in spusched_tick and find_victim. This change takes a reference to the context before releasing the mutex, so that the context doesn't get destroyed. Signed-off-by:
-
- Jul 24, 2008
-
-
Andre Detsch authored
This patch adjusts the placement of a reference context from a spu affinity chain. The reference context can now be placed only on nodes that have enough spus not intended to be used by another gang (already running on the node). Signed-off-by:
Andre Detsch <adetsch@br.ibm.com> Signed-off-by:
-
Andre Detsch authored
Currenlt,, it is possible to lock aff_mutex and cbe_spu_info[n].list_mutex in different orders, allowing a deadlock to occur. With this change, aff_mutex is not taken within a list_mutex critical section anymore. Signed-off-by:
-
- Jul 09, 2008
-
-
Maxim Shchetynin authored
As nr_active counter includes also spus waiting for syscalls to return we need a seperate counter that only counts spus that are currently running on spu side. This counter shall be used by a cpufreq governor that targets a frequency dependent from the number of running spus. Signed-off-by:
Christian Krafft <krafft@de.ibm.com> Acked-by:
Benjamin Herrenschmidt <benh@kernel.crashing.org>
-
Luke Browning authored
An spu context shouldn't get an extra tick if the time slice code couldn't find something else to run. This means contexts that are not within spu_run (ie, SPU_SCHED_SPU_RUN is cleared) will not receive extra ticks while we have no other contexts waiting. Signed-off-by:
Luke Browning <lukebrowning@us.ibm.com> Signed-off-by:
-
- Jun 16, 2008
-
-
Luke Browning authored
There is a delay in the transition to the stopped state for class 2 interrupts. In some cases, the controlling thread detects the state of the spu as running, and goes back to sleep resulting in a hung application as the event is missed. This change detects the stop condition and re-generates the wakeup event after a context save. Signed-off-by:
-
Luke Browning authored
Time slicing can occur at the same time as spu exception handling resulting in the wakeup of the wrong thread. This change uses the the spu's register_lock to enforce synchronization between bind/unbind and spu exception handling so that they are mutually exclusive. Signed-off-by:
-
- May 15, 2008
-
-
Luke Browning authored
If victim (not ctx) is in spu_run, add victim to rq. Signed-off-by:
Christoph Hellwig <hch@lst.de> Signed-off-by:
Paul Mackerras <paulus@samba.org>
-
- May 08, 2008
-
-
Christoph Hellwig authored
We should not requeue the victim context in find_victim if the owner is not in spu_run. It's first not needed because leaving the context on the spu is an optimization and second is harmful because it means the owner could re-enter spu_run when the context is on the runqueue and trip the BUG_ON in __spu_update_sched_info. Signed-off-by:
-
- May 05, 2008
-
-
Luke Browning authored
Currently, we re-route SPU interrupts to the current cpu, which may be on a remote node. In the case of time slicing, all spu interrupts will end up routed to the same cpu, where the spusched_tick occurs. This change routes mfc interrupts to the cpu where the controlling thread last ran, provided that cpu is on the same node as the spu (otherwise don't reroute interrupts). This should improve performance and provide a more predictable environment for processing spu exceptions. In the past we have seen concurrent delivery of spu exceptions to two cpus. This eliminates that concern. Signed-off-by:
Luke Browning <lukebr@linux.vnet.ibm.com> Signed-off-by:
-
- Apr 30, 2008
-
-
Julio M. Merino Vidal authored
Fix a typo in the marker for the find_victim function, which prevented it from being traced. It previously read find_vitim. Signed-off-by:
Julio M. Merino Vidal <jmerino@ac.upc.edu> Signed-off-by:
-
Christoph Hellwig authored
There are userspace instrumentation tools that need to monitor spu context switches. This patch adds a new file called 'switch_log' to each spufs context directory that can be used to monitor the context switches. Context switch in/out and exit from spu_run are monitored after the file was first opened and can be read from it. Signed-off-by:
-
- Apr 29, 2008
-
-
Denis V. Lunev authored
Use proc_create()/proc_create_data() to make sure that ->proc_fops and ->data be setup before gluing PDE to main tree. Add correct ->owner to proc_fops to fix reading/module unloading race. Signed-off-by:
Denis V. Lunev <den@openvz.org> Cc: Paul Mackerras <paulus@samba.org> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Alexey Dobriyan <adobriyan@gmail.com> Cc: "Eric W. Biederman" <ebiederm@xmission.com> Signed-off-by:
Andrew Morton <akpm@linux-foundation.org> Signed-off-by:
Linus Torvalds <torvalds@linux-foundation.org>
-
- Mar 11, 2008
-
-
Jeremy Kerr authored
commit 4ef11014 introduced a usage of SCHED_IDLE to detect when a context is within spu_run. Instead of SCHED_IDLE (which has other meaning), add a flag to sched_flags to tell if a context should be running. Signed-off-by:
-
- Feb 29, 2008
-
-
Andre Detsch authored
The spu_runcntl_RW register is restored within spu_restore function. So, at the end of spu_bind_context, the SPU context is not just loaded, but running. This change corrects the state switch to account the time as USER. Signed-off-by:
-
- Feb 18, 2008
-
-
Jeremy Kerr authored
2.6.25 has a regression where we can starve the scheduler by creating (N_SPES+1) contexts, then running them one at a time. The final context will never be run, as the other contexts are loaded on the SPEs, none of which are repoted as free (ie, spu->alloc_state != SPU_FREE), so spu_get_idle() doesn't give us a spu to run on. Because all of the contexts are stopped, none are descheduled by the scheduler tick, as spusched_tick returns if spu_stopped(ctx). This change replaces the spu_stopped() check with checking for SCHED_IDLE in ctx->policy. We set a context's policy to SCHED_IDLE when we're not in spu_run(). We also favour SCHED_IDLE contexts when looking for contexts to unbind, but leave their timeslice intact for later resumption. This patch fixes the following test in the spufs-testsuite: tests/20-scheduler/02-yield-starvation Signed-off-by:
-
- Feb 06, 2008
-
-
Christoph Hellwig authored
This adds markers two important points in the spufs code and a new module (sputrace.ko) that allows reading these out through a proc file. Long-term I'd rather see something like lttng extended to use the spufs instrumentation, but for now I think this is a good enough quick solution. We'll probably want to add various addition event in addition to that ones I have already. Signed-off-by:
-
- Jan 02, 2008
-
-
Paul Mackerras authored
Commit aed3a8c9 introduced a definition of notify_spus_active in .../cell/spu_syscalls.c, and another definition under #ifndef MODULE in .../cell/spufs/sched.c. The latter is not necessary and causes the build to fail when CONFIG_SPU_FS=y, so this removes it. It also removes the export of do_notify_spus_active, which is unnecessary. Signed-off-by:
Arnd Bergmann <arnd@arndb.de> Acked-by:
-
- Dec 28, 2007
-
-
Bob Nelson authored
This removes an OProfile dependency on the spufs module. This dependency was causing a problem for multiplatform systems that are built with support for Oprofile on Cell but try to load the oprofile module on a non-Cell system. Signed-off-by:
Bob Nelson <rrnelson@us.ibm.com> Signed-off-by:
Arnd Bergmann <arnd.bergmann@de.ibm.com> Acked-by:
-
- Dec 21, 2007
-
-
Aegis Lin authored
The original spusched_timer was designed to take effect only when a context is waiting in the runqueue. This change adds an additional lower-freq timer has been added to purely handle the spu_load updates. The new timer will be triggered per LOAD_FREQ ticks. Signed-off-by:
Aegis Lin <aegislin@gmail.com> Signed-off-by:
-
Christoph Hellwig authored
Make most places that use spu_acquire/spu_acquire_saved interruptible, this allows getting out of the spufs code when e.g. pressing ctrl+c. There are a few places where we get called e.g. from spufs teardown routines were we can't simply err out so these are left with a comment. For now I've also not touched the poll routines because it's open what libspe would expect in terms of interrupted system calls. Acked-by:
-
Luke Browning authored
Change spufs_spu_run so that the context is queued directly to the scheduler and the controlling thread advances directly to spufs_wait() for spe errors and exceptions. nosched contexts are treated the same as before. Fixes from Christoph Hellwig <hch@lst.de> Signed-off-by:
-
Luke Browning authored
Need to re-check priority after dropping lock. Otherwise, a more favored context may be preempted. Signed-off-by:
-
Luke Browning authored
This cleans up spu_run_init so that it does all of the spu initialization for spufs_run_spu. It initializes the spu context as much as possible before it activates the spu and writes the runcntl register. Signed-off-by:
-
Jeremy Kerr authored
Based on original patches from Arnd Bergmann <arnd.bergman@de.ibm.com>; and Luke Browning <lukebr@linux.vnet.ibm.com> Currently, spu contexts need to be loaded to the SPU in order to take class 0 and class 1 exceptions. This change makes the actual interrupt-handlers much simpler (ie, set the exception information in the context save area), and defers the handling code to the spufs_handle_class[01] functions, called from spufs_run_spu. This should improve the concurrency of the spu scheduling leading to greater SPU utilization when SPUs are overcommited. Signed-off-by:
-
Arnd Bergmann authored
This change disables the logic that faults-in spu contexts under the covers from the page fault handler. When a fault requires a runnable context, the handler will block until the context is scheduled by other means. Signed-off-by:
-
Jeremy Kerr authored
Currently, part of the spufs code (switch.o, lscsa_alloc.o and fault.o) is compiled directly into the kernel. This change moves these components of spufs into the kernel. The lscsa and switch objects are fairly straightforward to move in. For the fault.o module, we split the fault-handling code into two parts: a/p/p/c/spu_fault.c and a/p/p/c/spufs/fault.c. The former is for the in-kernel spu_handle_mm_fault function, and we move the rest of the fault-handling code into spufs. Signed-off-by:
-
Julio M. Merino Vidal authored
Fix a few typos in the spufs scheduler comments Signed-off-by:
-
- Sep 19, 2007
-
-
Christoph Hellwig authored
find_victim can dereference a NULL pointer when iterating over the list of victim spus because list_mutex only guarantees spu->ct to be stable, but of course not to be non-NULL. Also fix find_victim to not call spu_unbind_context without list_mutex because that violates the above guarantee. Signed-off-by:
-
Andre Detsch authored
Affinity reference point location (gang->aff_ref_spu) is reset when the whole gang is descheduled. However, the last member of a gang can be descheduled while we are trying to schedule another member of the gang. This was leading to a race condition, and the code was using gang->aff_ref_spu in an unsafe manner. By holding the gang->aff_mutex a little bit longer, and increment gang->aff_sched_count (which controls when gang->aff_ref_spu should be reset) a little bit earlier, the problem is fixed. Signed-off-by:
-
- Aug 03, 2007
-
-
Andre Detsch authored
This patch fixes affinity reference point placement, which was not being done in some situations, after the introduction of node_allowed() calls. The previously used parameter, 'ctx', is just the iterator of the previous list_for_each_entry_reverse loop, and its value might be invalid at the end of the loop. Also, the right context to seek for information when defining the reference ctx location _is_ the reference ctx. Signed-off-by:
-
- Jul 26, 2007
-
-
Masato Noguchi authored
We currently initialize cbe_spu_info[].spus in both init_spu_base and spu_sched_init. The initialise in spu_sched_init clears the SPU list, so we end up with no physical SPUs. Because of this, the spu_run syscall will block forever. This change removes the unnecessary initialization in spu_sched_init. Signed-off-by:
Masato Noguchi <Masato.Noguchi@jp.sony.com> Signed-off-by:
-
- Jul 20, 2007
-
-
Christoph Hellwig authored
This sorts out the various lists and related locks in the spu code. In detail: - the per-node free_spus and active_list are gone. Instead struct spu gained an alloc_state member telling whether the spu is free or not - the per-node spus array is now locked by a per-node mutex, which takes over from the global spu_lock and the per-node active_mutex - the spu_alloc* and spu_free function are gone as the state change is now done inline in the spufs code. This allows some more sharing of code for the affinity vs normal case and more efficient locking - some little refactoring in the affinity code for this locking scheme Signed-off-by:
-
Bob Nelson authored
From: Maynard Johnson <mpjohn@us.ibm.com> This patch updates the existing arch/powerpc/oprofile/op_model_cell.c to add in the SPU profiling capabilities. In addition, a 'cell' subdirectory was added to arch/powerpc/oprofile to hold Cell-specific SPU profiling code. Exports spu_set_profile_private_kref and spu_get_profile_private_kref which are used by OProfile to store private profile information in spufs data structures. Also incorporated several fixes from other patches (rrn). Check pointer returned from kzalloc. Eliminated unnecessary cast. Better error handling and cleanup in the related area. 64-bit unsigned long parameter was being demoted to 32-bit unsigned int and eventually promoted back to unsigned long. Signed-off-by:
Carl Love <carll@us.ibm.com> Signed-off-by:
Maynard Johnson <mpjohn@us.ibm.com> Signed-off-by:
-
Bob Nelson authored
From: Maynard Johnson <mpjohn@us.ibm.com> This patch adds to the capability of spu_switch_event_register so that the caller is also notified of currently active SPU tasks. Exports spu_switch_event_register and spu_switch_event_unregister so that OProfile can get access to the notifications provided. Signed-off-by:
-
Arnd Bergmann authored
This patch makes the scheduller honor affinity information for each context being scheduled. If the context has no affinity information, behaviour is unchanged. If there are affinity information, context is schedulled to be run on the exact spu recommended by the affinity placement algorithm. Signed-off-by:
-
Arnd Bergmann authored
This patch provides the spu affinity placement logic for the spufs scheduler. Each time a gang is going to be scheduled, the placement of a reference context is defined. The placement of all other contexts with affinity from the gang is defined based on this reference context location and on a precomputed displacement offset. Signed-off-by:
-
