- 24 Aug, 2004 26 commits
-
-
Ingo Molnar authored
while looking at HT scheduler bugreports and boot failures i discovered a bad assumption in most of the HT scheduling code: that resched_task() can be called without holding the task's runqueue. This is most definitely not valid - doing it without locking can lead to the task on that CPU exiting, and this CPU corrupting the (ex-) task_info struct. It can also lead to HT-wakeup races with task switching on that other CPU. (this_CPU marking the wrong task on that_CPU as need_resched - resulting in e.g. idle wakeups not working.) The attached patch against fixes it all up. Changes: - resched_task() needs to touch the task so the runqueue lock of that CPU must be held: resched_task() now enforces this rule. - wake_priority_sleeper() was called without holding the runqueue lock. - wake_sleeping_dependent() needs to hold the runqueue locks of all siblings (2 typically). Effects of this ripples back to schedule() as well - in the non-SMT case it gets compiled out so it's fine. - dependent_sleeper() needs the runqueue locks too - and it's slightly harder because it wants to know the 'next task' info which might change during the lock-drop/reacquire. Ripple effect on schedule() => compiled out on non-SMT so fine. - resched_task() was disabling preemption for no good reason - all paths that called this function had either a spinlock held or irqs disabled. Compiled & booted on x86 SMP and UP, with and without SMT. Booted the SMT kernel on a real SMP+HT box as well. (Unpatched kernel wouldn't even boot with the resched_task() assert in place.) Signed-off-by:
Ingo Molnar <mingo@elte.hu> Signed-off-by:
Andrew Morton <akpm@osdl.org> Signed-off-by:
Linus Torvalds <torvalds@osdl.org>
-
Ingo Molnar authored
disable preemption in the self-reap codepath, as such tasks may not be on the tasklist anymore and CPU-hotplug relies on the tasklist to migrate tasks. Signed-off-by:
-
Ingo Molnar authored
release_task() calls proc_pid_flush() call dput(), which can sleep. But that's a late-in-exit no-preempt path with CONFIG_PREEMPT. Signed-off-by:
-
Ingo Molnar authored
Rusty noticed that we update the parent ->avg_sleep without holding the runqueue lock. Also the code needed cleanups. Signed-off-by:
-
Ingo Molnar authored
* Nick Piggin <nickpiggin@yahoo.com.au> wrote: > Increasing priority (negative nice) doesn't have much impact. -20 CPU > hog only gets about double the CPU of a 0 priority CPU hog and only > about 120% the CPU time of a nice -10 hog. this is a property of the base scheduler as well. We can do a nonlinear timeslice distribution trivially - the attached patch implements the following timeslice distribution ontop of 2.6.8-rc3-mm1: [ -20 ... 0 ... 19 ] => [800ms ... 100ms ... 5ms] the nice-20/nice+19 ratio is now 1:160 - sufficient for all aspects. Signed-off-by:
-
Ingo Molnar authored
- whitespace and style cleanups Signed-off-by:
-
Andrew Morton authored
SMP fix -- for_each_domain() is not defined if not CONFIG_SMP, so show_schedstat needed a couple of extra ifdefs. Signed-off-by:Rick Lindsley <ricklind@us.ibm.com> Signed-off-by:
-
William Lee Irwin III authored
Fix up sparc32 properly. Signed-off-by:
-
William Lee Irwin III authored
It appears that init_idle() and fork_by_hand() could be combined into a single method that calls init_idle() on behalf of the caller. Signed-off-by:
-
Nathan Lynch authored
Otherwise it shows up under "iSeries device drivers", which doesn't seem right. Signed-off-by:
Nathan Lynch <nathanl@austin.ibm.com> Signed-off-by:
-
Rick Lindsley authored
It adds lots of CPU scheduler stats in /proc/pid/stat. They are described in the new Documentation//sched-stats.txt We were carrying this patch offline for some time, but as there's still considerable ongoing work in this area, and as the new stats are a configuration option, I think it's best that this capability be in the base kernel. Nick removed a fair amount of statistics that he wasn't using. The full patch gathers more information. In particular, his patch doesn't include the code to measure the latency between the time a process is made runnable and the time it hits a processor which will be key to measuring interactivity changes. He passed his changes back to me and I got finished merging his changes with the current statistics patches just before OLS. I believe this is largely a superset of the patch you grabbed and should port relatively easily too. Versions also exist for 2.6.8-rc2 2.6.8-rc2-mm1 2.6.8-rc2-mm2 at http://eaglet.rain.com/rick/linux/schedstat/patches/ and within 24 hours at http://oss.software.ibm.com/linux/patches/?patch_id=730&show=all The version below is for 2.6.8-rc2-mm2 without the staircase code and has been compiled cleanly but not yet run. From: Ingo Molnar <mingo@elte.hu> this code needs a couple of cleanups before it can go into mainline: fs/proc/array.c, fs/proc/base.c, fs/proc/proc_misc.c: - moved the new /proc/<PID>/stat fields to /proc/<PID>/schedstat, because the new fields break older procps. It's cleaner this way anyway. This moving of fields necessiated a bump to version 10. Documentation/sched-stats.txt: - updated sched-stats.txt for version 10 - wake_up_forked_thread() => wake_up_new_task() - updated the per-process field description Kconfig: - removed the default y and made the option dependent on DEBUG_KERNEL. This is really for scheduler analysis, normal users dont need the overhead. include/linux/sched.h: - moved the definitions into kernel/sched.c - this fixes UP compilation and is cleaner. - also moved the sched-domain definitions to sched.c - now that the sched-domains internals are not exposed to architectures this is doable. It's also necessary due to the previous change. kernel/fork.c: - moved the ->sched_info init to sched_fork() where it belongs. kernel/sched.c: - wake_up_forked_thread() -> wake_up_new_task(), wuft_cnt -> wunt_cnt, wuft_moved -> wunt_moved. - wunt_cnt and wunt_moved were defined by never updated - added the missing code to wake_up_new_task(). - whitespace/style police - removed whitespace changes done to code not related to schedstats - i'll send a separate patch for these (and more). Signed-off-by: -
Con Kolivas authored
The smt-nice handling is a little too aggressive by not estimating the per cpu gain as high enough for pentium4 hyperthread. This patch changes the per sibling cpu gain from 15% to 25%. The true per cpu gain is entirely dependant on the workload but overall the 2 species of Pentium4 that support hyperthreading have about 20-30% gain. P.S: Anton - For the power processors that are now using this SMT nice infrastructure it would be worth setting this value separately at 40%. Signed-off-by:
Con Kolivas <kernel@kolivas.org> Signed-off-by:
-
Matthew Dobson authored
In light of some proposed changes in the sched_domains code, I coded up this little ditty that simply creates and populates a cpu_sibling_map for PPC64 machines. The patch just checks the CPU flags to determine if the CPU supports SMT (aka Hyper-Threading aka Multi-Threading aka ...) and fills in a mask of the siblings for each CPU in the system. This should allow us to build sched_domains for PPC64 with generic code in kernel/sched.c for the SMT systems. SMT is becoming more popular and is turning up in more and more architectures. I don't think it will be too long until this feature is supported by most arches... Signed-off-by:
Matthew Dobson <colpatch@us.ibm.com> Signed-off-by:
-
Dimitri Sivanich authored
Here's a version of the isolated scheduler domain code that I mentioned in an RFC on 7/22. This patch applies on top of 2.6.8-rc2-mm1 (to include all of the new arch_init_sched_domain code). This patch also contains the 2 line fix to remove the check of first_cpu(sd->groups->cpumask)) that Jesse sent in earlier. Note that this has not been tested with CONFIG_SCHED_SMT. I hope that my handling of those instances is OK. Signed-off-by:
Dimitri Sivanich <sivanich@sgi.com> Signed-off-by:
-
Jesse Barnes authored
This patch limits the cpu span of each node's scheduler domain to prevent balancing across too many cpus. The cpus included in a node's domain are determined by the SD_NODES_PER_DOMAIN define and the arch specific sched_domain_node_span routine if ARCH_HAS_SCHED_DOMAIN is defined. If ARCH_HAS_SCHED_DOMAIN is not defined, behavior is unchanged--all possible cpus will be included in each node's scheduling domain. Currently, only ia64 provides an arch specific sched_domain_node_span routine. From: Jesse Barnes <jbarnes@engr.sgi.com> This patch adds some more NUMA specific logic to the creation of scheduler domains. Domains spanning all CPUs in a large system are too large to schedule across efficiently, leading to livelocks and inordinate amounts of time being spent in scheduler routines. With this patch applied, the node scheduling domains for NUMA platforms will only contain a specified number of nearby CPUs, based on the value of SD_NODES_PER_DOMAIN. It also allows arches to override SD_NODE_INIT, which sets the domain scheduling parameters for each node's domain. This is necessary especially for large systems. Possible future directions: o multilevel node hierarchy (e.g. node domains could contain 4 nodes worth of CPUs, supernode domains could contain 32 nodes worth, etc. each with their own SD_NODE_INIT values) o more tweaking of SD_NODE_INIT values for good load balancing vs. overhead tradeoffs From: mita akinobu <amgta@yacht.ocn.ne.jp> Compile fix Signed-off-by:Jesse Barnes <jbarnes@sgi.com> Signed-off-by:
-
Nick Piggin authored
Teach the generic domains builder about SMT, and consolidate all architecture specific domain code into that. Also, the SD_*_INIT macros can now be redefined by arch code without duplicating the entire setup code. This can be done by defining ARCH_HASH_SCHED_TUNE. The generic builder has been simplified with the addition of a helper macro which will probably prove to be useful to arch specific code as well and should be exported if that is the case. Signed-off-by:
Nick Piggin <nickpiggin@yahoo.com.au> From: Matthew Dobson <colpatch@us.ibm.com> The attached patch is against 2.6.8-rc2-mm2, and removes Nick's conditional definition & population of cpu_sibling_map[] in favor of my unconditional ones. This does not affect how cpu_sibling_map is used, just gives it broader scope. From: Nick Piggin <nickpiggin@yahoo.com.au> Small fix to sched-consolidate-domains.patch picked up by From: Suresh <suresh.b.siddha@intel.com> another sched consolidate domains fix From: Nick Piggin <nickpiggin@yahoo.com.au> Don't use cpu_sibling_map if !CONFIG_SCHED_SMT This one spotted by Dimitri Sivanich <sivanich@sgi.com> Signed-off-by:
-
Ingo Molnar authored
- remove the hotplug lock from around much of fork(), and re-copy the cpus_allowed mask to solve the hotplug race cleanly. Signed-off-by:
Srivatsa Vaddagiri <vatsa@in.ibm.com> Signed-off-by:
-
Nick Piggin authored
This removes balance on clone capability altogether. I told Andi we wouldn't remove it yet, but provided it is in a single small patch, he mightn't get too upset. Signed-off-by:
-
Nick Piggin authored
Don't balance on clone by default. Balance on clone has a number of trivial performance failure cases, but it was needed to get decent OpenMP performance on NUMA (Opteron) systems. Not doing child-runs-first for new threads also solves this problem in a nicer way (implemented in a previous patch). Signed-off-by:
-
Nick Piggin authored
Add some likely/unliklies, a for_each_cpu => for_each_cpu_online, and close the sched_exit race. From: Ingo Molnar <mingo@elte.hu> fix a typo in a previous patch breaking RT scheduling & interactivity. Signed-off-by:
-
Nick Piggin authored
From: Ingo Molnar <mingo@elte.hu> RT tasks are unlikely, move this into rt_task() instead of open-coding it. Signed-off-by:
-
Ingo Molnar authored
- fix two stale comments - cleanup Signed-off-by:
-
Nick Piggin authored
Now that init_idle does not remove tasks from the runqueue, those architectures that use kernel_thread instead of copy_process for the idle task will break. To fix, ensure that CLONE_IDLETASK tasks are not put on the runqueue in the first place. Signed-off-by:
-
Nick Piggin authored
Move balancing and child-runs-first logic from fork.c into sched.c where it belongs. * Consolidate wake_up_forked_process and wake_up_forked_thread into wake_up_new_process, and pass in clone_flags as suggested by Linus. This removes a lot of code duplication and allows all logic to be handled in that function. * Don't do balance-on-clone balancing for vfork'ed threads. * Don't do set_task_cpu or balance one clone in wake_up_new_process. Instead do it in sched_fork to fix set_cpus_allowed races. * Don't do child-runs-first for CLONE_VM processes, as there is obviously no COW benifit to be had. This is a big one, it enables Andi's workload to run well without clone balancing, because the OpenMP child threads can get balanced off to other nodes *before* they start running and allocating memory. * Rename sched_balance_exec to sched_exec: hide the policy from the API. From: Ingo Molnar <mingo@elte.hu> rename wake_up_new_process -> wake_up_new_task. in sched.c we are gradually moving away from the overloaded 'process' or 'thread' notion to the traditional task (or context) naming. Signed-off-by:
-
Nick Piggin authored
Clean up init_idle to not use wake_up_forked_process, then undo all the stuff that call does. Instead, do everything in init_idle. Make double_rq_lock depend on CONFIG_SMP because it is no longer used on UP. Signed-off-by:
-
Ingo Molnar authored
The main benefit is that with the default HZ=1000 nice +19 tasks now get 5 msecs of timeslices, so the ratio of CPU use is linear. (nice 0 task gets 20 times more CPU time than a nice 19 task. Prior this change the ratio was 1:10) another effect is that nice 0 tasks now get a round 100 msecs of timeslices (as intended), instead of 102 msecs. here's a table of old/new timeslice values, for HZ=1000 and 100: HZ=1000 ( HZ=100 ) old new ( old new ) nice -20: 200 200 ( 200 200 ) nice -19: 195 195 ( 190 190 ) ... nice 0: 102 100 ( 100 100 ) nice 1: 97 95 ( 90 90 ) nice 2: 92 90 ( 90 90 ) ... nice 17: 19 15 ( 10 10 ) nice 18: 14 10 ( 10 10 ) nice 19: 10 5 ( 10 10 ) i've tested the patch on x86. Signed-off-by:
-
- 23 Aug, 2004 14 commits
-
-
Linus Torvalds authored
-
Paul Mackerras authored
This patch changes hose_list from a simple linked list to a "list.h"-style list. This is in preparation for the runtime addition/removal of PCI Host Bridges. Signed-off-by:
John Rose <johnrose@austin.ibm.com> Signed-off-by:
Paul Mackerras <paulus@samba.org> Signed-off-by:
-
Nathan Fontenot authored
On some platforms (notably power5) you can't enable surveillance (firmware/service processor watchdog) from the kernel - you have to do it in the firmware. This patch changes enable_surveillance() to make the message that is printed in this situation more informative. Additionaly, the rtas_call was changed to rtas_set_indicator so as to avoid having to handle RTAS_BUSY returns. Signed-off-by:
Nathan Fontenot <nfont@austin.ibm.com> Signed-off-by:
-
bk://ppc.bkbits.net/for-linus-ppc64Linus Torvalds authored
into ppc970.osdl.org:/home/torvalds/v2.6/linux
-
Linus Torvalds authored
The code doesn't actually _care_ about 32/64-bit issues, only about F_SETLK vs F_SETLKW, and the F_SETLK64 doesn't exist except as a compatibility thing on 64-bit architectures (since the regular one already _is_ 64-bit, of course).
-
http://nfsclient.bkbits.net/linux-2.6Trond Myklebust authored
into fys.uio.no:/home/linux/bitkeeper/nfsclient-2.6
-
Trond Myklebust authored
server are not allowed to be interrupted as that may result in the client and server disagreeing.
-
Trond Myklebust authored
recall ability. Signed-off-by:Trond Myklebust <trond.myklebust@fys.uio.no>
-
Trond Myklebust authored
NFS4ERR_CB_PATH_DOWN error.
-
Trond Myklebust authored
Signed-off-by: -
Trond Myklebust authored
Signed-off-by: -
Trond Myklebust authored
Signed-off-by: -
Trond Myklebust authored
Signed-off-by: -
Trond Myklebust authored
Signed-off-by:
-
