- 04 Apr, 2017 2 commits
-
-
Xunlei Pang authored
We should deboost before waking the high-priority task, such that we don't run two tasks with the same "state" (priority, deadline, sched_class, etc). In order to make sure the boosting task doesn't start running between unlock and deboost (due to 'spurious' wakeup), we move the deboost under the wait_lock, that way its serialized against the wait loop in __rt_mutex_slowlock(). Doing the deboost early can however lead to priority-inversion if current would get preempted after the deboost but before waking our high-prio task, hence we disable preemption before doing deboost, and enabling it after the wake up is over. This gets us the right semantic order, but most importantly however; this change ensures pointer stability for the next patch, where we have rt_mutex_setprio() cache a pointer to the top-most waiter task. If we, as before this change, do the wakeup first and then deboost, this pointer might point into thin air. [peterz: Changelog + patch munging] Suggested-by:
Peter Zijlstra <peterz@infradead.org> Signed-off-by:
Xunlei Pang <xlpang@redhat.com> Signed-off-by:
Steven Rostedt <rostedt@goodmis.org> Cc: juri.lelli@arm.com Cc: bigeasy@linutronix.de Cc: mathieu.desnoyers@efficios.com Cc: jdesfossez@efficios.com Cc: bristot@redhat.com Link: http://lkml.kernel.org/r/20170323150216.110065320@infradead.orgSigned-off-by:
Thomas Gleixner <tglx@linutronix.de>
-
Thomas Gleixner authored
Required for the rtmutex/sched_deadline patches which depend on both branches
-
- 30 Mar, 2017 4 commits
-
-
Chris Wilson authored
Use a timeout rather than a fixed number of loops to avoid running for very long periods, such as under the kbuilder VMs. Reported-by:
kernel test robot <xiaolong.ye@intel.com> Signed-off-by:
Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by:
Ingo Molnar <mingo@kernel.org>
-
Yuyang Du authored
The main PELT function ___update_load_avg(), which implements the accumulation and progression of the geometric average series, is implemented along the following lines for the scenario where the time delta spans all 3 possible sections (see figure below): 1. add the remainder of the last incomplete period 2. decay old sum 3. accumulate new sum in full periods since last_update_time 4. accumulate the current incomplete period 5. update averages Or: d1 d2 d3 ^ ^ ^ | | | |<->|<----------------->|<--->| ... |---x---|------| ... |------|-----x (now) load_sum' = (load_sum + weight * scale * d1) * y^(p+1) + (1,2) p weight * scale * 1024 * \Sum y^n + (3) n=1 weight * scale * d3 * y^0 (4) load_avg' = load_sum' / LOAD_AVG_MAX (5) Where: d1 - is the delta part completing the remainder of the last incomplete period, d2 - is the delta part spannind complete periods, and d3 - is the delta part starting the current incomplete period. We can simplify the code in two steps; the first step is to separate the first term into new and old parts like: (load_sum + weight * scale * d1) * y^(p+1) = load_sum * y^(p+1) + weight * scale * d1 * y^(p+1) Once we've done that, its easy to see that all new terms carry the common factors: weight * scale If we factor those out, we arrive at the form: load_sum' = load_sum * y^(p+1) + weight * scale * (d1 * y^(p+1) + p 1024 * \Sum y^n + n=1 d3 * y^0) Which results in a simpler, smaller and faster implementation. Signed-off-by:Yuyang Du <yuyang.du@intel.com> Signed-off-by:
-
Peter Zijlstra authored
The __update_load_avg() function is an __always_inline because its used with constant propagation to generate different variants of the code without having to duplicate it (which would be prone to bugs). Explicitly instantiate the 3 variants. Note that most of this is called from rather hot paths, so reducing branches is good. Signed-off-by:
-
Peter Zijlstra authored
Dmitry noted that the new atomic_try_cmpxchg() primitive is broken when the old pointer doesn't point to the local stack. He writes: "Consider a classical lock-free stack push: node->next = atomic_read(&head); do { } while (!atomic_try_cmpxchg(&head, &node->next, node)); This code is broken with the current implementation, the problem is with unconditional update of *__po. In case of success it writes the same value back into *__po, but in case of cmpxchg success we might have lose ownership of some memory locations and potentially over what __po has pointed to. The same holds for the re-read of *__po. " He also points out that this makes it surprisingly different from the similar C/C++ atomic operation. After investigating the code-gen differences caused by this patch; and a number of alternatives (Linus dislikes this interface lots), we arrived at these results (size x86_64-defconfig/vmlinux): GCC-6.3.0: 10735757 cmpxchg 10726413 try_cmpxchg 10730509 try_cmpxchg + patch 10730445 try_cmpxchg-linus GCC-7 (20170327): 10709514 cmpxchg 10704266 try_cmpxchg 10704266 try_cmpxchg + patch 10704394 try_cmpxchg-linus From this we see that the patch has the advantage of better code-gen on GCC-7 and keeps the interface roughly consistent with the C language variant. Reported-by:Dmitry Vyukov <dvyukov@google.com> Signed-off-by:
-
- 27 Mar, 2017 1 commit
-
-
Srikar Dronamraju authored
If the child domain prefers tasks to go siblings, the local group could end up pulling tasks to itself even if the local group is almost equally loaded as the source group. Lets assume a 4 core,smt==2 machine running 5 thread ebizzy workload. Everytime, local group has capacity and source group has atleast 2 threads, local group tries to pull the task. This causes the threads to constantly move between different cores. This is even more profound if the cores have more threads, like in Power 8, smt 8 mode. Fix this by only allowing local group to pull a task, if the source group has more number of tasks than the local group. Here are the relevant perf stat numbers of a 22 core,smt 8 Power 8 machine. Without patch: Performance counter stats for 'ebizzy -t 22 -S 100' (5 runs): 1,440 context-switches # 0.001 K/sec ( +- 1.26% ) 366 cpu-migrations # 0.000 K/sec ( +- 5.58% ) 3,933 page-faults # 0.002 K/sec ( +- 11.08% ) Performance counter stats for 'ebizzy -t 48 -S 100' (5 runs): 6,287 context-switches # 0.001 K/sec ( +- 3.65% ) 3,776 cpu-migrations # 0.001 K/sec ( +- 4.84% ) 5,702 page-faults # 0.001 K/sec ( +- 9.36% ) Performance counter stats for 'ebizzy -t 96 -S 100' (5 runs): 8,776 context-switches # 0.001 K/sec ( +- 0.73% ) 2,790 cpu-migrations # 0.000 K/sec ( +- 0.98% ) 10,540 page-faults # 0.001 K/sec ( +- 3.12% ) With patch: Performance counter stats for 'ebizzy -t 22 -S 100' (5 runs): 1,133 context-switches # 0.001 K/sec ( +- 4.72% ) 123 cpu-migrations # 0.000 K/sec ( +- 3.42% ) 3,858 page-faults # 0.002 K/sec ( +- 8.52% ) Performance counter stats for 'ebizzy -t 48 -S 100' (5 runs): 2,169 context-switches # 0.000 K/sec ( +- 6.19% ) 189 cpu-migrations # 0.000 K/sec ( +- 12.75% ) 5,917 page-faults # 0.001 K/sec ( +- 8.09% ) Performance counter stats for 'ebizzy -t 96 -S 100' (5 runs): 5,333 context-switches # 0.001 K/sec ( +- 5.91% ) 506 cpu-migrations # 0.000 K/sec ( +- 3.35% ) 10,792 page-faults # 0.001 K/sec ( +- 7.75% ) Which show that in these workloads CPU migrations get reduced significantly. Signed-off-by:Srikar Dronamraju <srikar@linux.vnet.ibm.com> Signed-off-by:
-
- 26 Mar, 2017 1 commit
-
-
Peter Zijlstra authored
Since commit 383776fa ("locking/lockdep: Handle statically initialized PER_CPU locks properly") we try to collapse per-cpu locks into a single class by giving them all the same key. For this key we choose the canonical address of the per-cpu object, which would be the offset into the per-cpu area. This has two problems: - there is a case where we run !0 lock->key through static_obj() and expect this to pass; it doesn't for canonical pointers. - 0 is a valid canonical address. Cure both issues by redefining the canonical address as the address of the per-cpu variable on the boot CPU. Since I didn't want to rely on CPU0 being the boot-cpu, or even existing at all, track the boot CPU in a variable. Fixes: 383776fa ("locking/lockdep: Handle statically initialized PER_CPU locks properly") Reported-by:
kernel test robot <fengguang.wu@intel.com> Signed-off-by:
Borislav Petkov <bp@suse.de> Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de> Cc: linux-mm@kvack.org Cc: wfg@linux.intel.com Cc: kernel test robot <fengguang.wu@intel.com> Cc: LKP <lkp@01.org> Link: http://lkml.kernel.org/r/20170320114108.kbvcsuepem45j5cr@hirez.programming.kicks-ass.netSigned-off-by:
-
- 23 Mar, 2017 18 commits
-
-
Peter Zijlstra authored
When PREEMPT_RT_FULL does the spinlock -> rt_mutex substitution the PI chain code will (falsely) report a deadlock and BUG. The problem is that it hold hb->lock (now an rt_mutex) while doing task_blocks_on_rt_mutex on the futex's pi_state::rtmutex. This, when interleaved just right with futex_unlock_pi() leads it to believe to see an AB-BA deadlock. Task1 (holds rt_mutex, Task2 (does FUTEX_LOCK_PI) does FUTEX_UNLOCK_PI) lock hb->lock lock rt_mutex (as per start_proxy) lock hb->lock Which is a trivial AB-BA. It is not an actual deadlock, because it won't be holding hb->lock by the time it actually blocks on the rt_mutex, but the chainwalk code doesn't know that and it would be a nightmare to handle this gracefully. To avoid this problem, do the same as in futex_unlock_pi() and drop hb->lock after acquiring wait_lock. This still fully serializes against futex_unlock_pi(), since adding to the wait_list does the very same lock dance, and removing it holds both locks. Aside of solving the RT problem this makes the lock and unlock mechanism symetric and reduces the hb->lock held time. Reported-and-tested-by:Sebastian Andrzej Siewior <bigeasy@linutronix.de> Suggested-by:
-
Peter Zijlstra authored
The problem with returning -EAGAIN when the waiter state mismatches is that it becomes very hard to proof a bounded execution time on the operation. And seeing that this is a RT operation, this is somewhat important. While in practise; given the previous patch; it will be very unlikely to ever really take more than one or two rounds, proving so becomes rather hard. However, now that modifying wait_list is done while holding both hb->lock and wait_lock, the scenario can be avoided entirely by acquiring wait_lock while still holding hb-lock. Doing a hand-over, without leaving a hole. Signed-off-by:
-
Peter Zijlstra authored
By changing futex_lock_pi() to use rt_mutex_*_proxy_lock() all wait_list modifications are done under both hb->lock and wait_lock. This closes the obvious interleave pattern between futex_lock_pi() and futex_unlock_pi(), but not entirely so. See below: Before: futex_lock_pi() futex_unlock_pi() unlock hb->lock lock hb->lock unlock hb->lock lock rt_mutex->wait_lock unlock rt_mutex_wait_lock -EAGAIN lock rt_mutex->wait_lock list_add unlock rt_mutex->wait_lock schedule() lock rt_mutex->wait_lock list_del unlock rt_mutex->wait_lock <idem> -EAGAIN lock hb->lock After: futex_lock_pi() futex_unlock_pi() lock hb->lock lock rt_mutex->wait_lock list_add unlock rt_mutex->wait_lock unlock hb->lock schedule() lock hb->lock unlock hb->lock lock hb->lock lock rt_mutex->wait_lock list_del unlock rt_mutex->wait_lock lock rt_mutex->wait_lock unlock rt_mutex_wait_lock -EAGAIN unlock hb->lock It does however solve the earlier starvation/live-lock scenario which got introduced with the -EAGAIN since unlike the before scenario; where the -EAGAIN happens while futex_unlock_pi() doesn't hold any locks; in the after scenario it happens while futex_unlock_pi() actually holds a lock, and then it is serialized on that lock. Signed-off-by:
-
Peter Zijlstra authored
With the ultimate goal of keeping rt_mutex wait_list and futex_q waiters consistent it's necessary to split 'rt_mutex_futex_lock()' into finer parts, such that only the actual blocking can be done without hb->lock held. Split split_mutex_finish_proxy_lock() into two parts, one that does the blocking and one that does remove_waiter() when the lock acquire failed. When the rtmutex was acquired successfully the waiter can be removed in the acquisiton path safely, since there is no concurrency on the lock owner. This means that, except for futex_lock_pi(), all wait_list modifications are done with both hb->lock and wait_lock held. [bigeasy@linutronix.de: fix for futex_requeue_pi_signal_restart] Signed-off-by:
-
Peter Zijlstra authored
Since there's already two copies of this code, introduce a helper now before adding a third one. Signed-off-by:
-
Peter Zijlstra authored
There's a number of 'interesting' problems, all caused by holding hb->lock while doing the rt_mutex_unlock() equivalient. Notably: - a PI inversion on hb->lock; and, - a SCHED_DEADLINE crash because of pointer instability. The previous changes: - changed the locking rules to cover {uval,pi_state} with wait_lock. - allow to do rt_mutex_futex_unlock() without dropping wait_lock; which in turn allows to rely on wait_lock atomicity completely. - simplified the waiter conundrum. It's now sufficient to hold rtmutex::wait_lock and a reference on the pi_state to protect the state consistency, so hb->lock can be dropped before calling rt_mutex_futex_unlock(). Signed-off-by: -
Peter Zijlstra authored
There is a weird state in the futex_unlock_pi() path when it interleaves with a concurrent futex_lock_pi() at the point where it drops hb->lock. In this case, it can happen that the rt_mutex wait_list and the futex_q disagree on pending waiters, in particular rt_mutex will find no pending waiters where futex_q thinks there are. In this case the rt_mutex unlock code cannot assign an owner. The futex side fixup code has to cleanup the inconsistencies with quite a bunch of interesting corner cases. Simplify all this by changing wake_futex_pi() to return -EAGAIN when this situation occurs. This then gives the futex_lock_pi() code the opportunity to continue and the retried futex_unlock_pi() will now observe a coherent state. The only problem is that this breaks RT timeliness guarantees. That is, consider the following scenario: T1 and T2 are both pinned to CPU0. prio(T2) > prio(T1) CPU0 T1 lock_pi() queue_me() <- Waiter is visible preemption T2 unlock_pi() loops with -EAGAIN forever Which is undesirable for PI primitives. Future patches will rectify this. Signed-off-by: -
Peter Zijlstra authored
Add a put_pit_state() as counterpart for get_pi_state() so the refcounting becomes consistent. Signed-off-by:
-
Peter Zijlstra authored
Currently futex-pi relies on hb->lock to serialize everything. But hb->lock creates another set of problems, especially priority inversions on RT where hb->lock becomes a rt_mutex itself. The rt_mutex::wait_lock is the most obvious protection for keeping the futex user space value and the kernel internal pi_state in sync. Rework and document the locking so rt_mutex::wait_lock is held accross all operations which modify the user space value and the pi state. This allows to invoke rt_mutex_unlock() (including deboost) without holding hb->lock as a next step. Nothing yet relies on the new locking rules. Signed-off-by:
-
Peter Zijlstra authored
Part of what makes futex_unlock_pi() intricate is that rt_mutex_futex_unlock() -> rt_mutex_slowunlock() can drop rt_mutex::wait_lock. This means it cannot rely on the atomicy of wait_lock, which would be preferred in order to not rely on hb->lock so much. The reason rt_mutex_slowunlock() needs to drop wait_lock is because it can race with the rt_mutex fastpath, however futexes have their own fast path. Since futexes already have a bunch of separate rt_mutex accessors, complete that set and implement a rt_mutex variant without fastpath for them. Signed-off-by:
-
Peter Zijlstra authored
These are unused and clutter up the code. Signed-off-by:
-
Peter Zijlstra authored
Since the futex_q can dissapear the instruction after assigning NULL, this really should be a RELEASE barrier. That stops loads from hitting dead memory too. Signed-off-by:
-
Peter Zijlstra authored
futex_top_waiter() returns the top-waiter on the pi_mutex. Assinging this to a variable 'match' totally obscures the code. Signed-off-by:
-
Peter Zijlstra authored
Better code generation: text data bss name 10665111 4530096 843776 defconfig-build/vmlinux.3 10655703 4530096 843776 defconfig-build/vmlinux.4 Signed-off-by: -
Peter Zijlstra authored
Generates better code (GCC-6.2.1): text filename 1576 defconfig-build/lib/refcount.o.pre 1488 defconfig-build/lib/refcount.o.post Signed-off-by:
-
Peter Zijlstra authored
Add a new cmpxchg interface: bool try_cmpxchg(u{8,16,32,64} *ptr, u{8,16,32,64} *val, u{8,16,32,64} new); Where the boolean returns the result of the compare; and thus if the exchange happened; and in case of failure, the new value of *ptr is returned in *val. This allows simplification/improvement of loops like: for (;;) { new = val $op $imm; old = cmpxchg(ptr, val, new); if (old == val) break; val = old; } into: do { } while (!try_cmpxchg(ptr, &val, val $op $imm)); while also generating better code (GCC6 and onwards). Signed-off-by: -
Vincent Guittot authored
A regression of the FTQ noise has been reported by Ying Huang, on the following hardware: 8 threads Intel(R) Core(TM)i7-4770 CPU @ 3.40GHz with 8G memory ... which was caused by this commit: commit 4e516076 ("sched/fair: Propagate asynchrous detach") The only part of the patch that can increase the noise is the update of blocked load of group entity in update_blocked_averages(). We can optimize this call and skip the update of group entity if its load and utilization are already null and there is no pending propagation of load in the task group. This optimization partly restores the noise score. A more agressive optimization has been tried but has shown worse score. Reported-by: ying.huang@linux.intel.com Signed-off-by:
Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by:
-
Wanpeng Li authored
This can be reproduced by running rt-migrate-test: WARNING: CPU: 2 PID: 2195 at kernel/locking/lockdep.c:3670 lock_unpin_lock() unpinning an unpinned lock ... Call Trace: dump_stack() __warn() warn_slowpath_fmt() lock_unpin_lock() __balance_callback() __schedule() schedule() futex_wait_queue_me() futex_wait() do_futex() SyS_futex() do_syscall_64() entry_SYSCALL64_slow_path() Revert the rq_lock_irqsave() usage here, the whole point of the balance_callback() was to allow dropping rq->lock. Reported-by:
Wanpeng Li <wanpeng.li@hotmail.com> Signed-off-by:
-
- 16 Mar, 2017 14 commits
-
-
Peter Zijlstra authored
Currently each thread starts an acquire context only once, and performs all its loop iterations under it. This means that the Wound/Wait relations between threads are fixed. To make things a little more realistic and cover more of the functionality with the test, open a new acquire context for each loop. Signed-off-by:
-
Thomas Gleixner authored
If a PER_CPU struct which contains a spin_lock is statically initialized via: DEFINE_PER_CPU(struct foo, bla) = { .lock = __SPIN_LOCK_UNLOCKED(bla.lock) }; then lockdep assigns a seperate key to each lock because the logic for assigning a key to statically initialized locks is to use the address as the key. With per CPU locks the address is obvioulsy different on each CPU. That's wrong, because all locks should have the same key. To solve this the following modifications are required: 1) Extend the is_kernel/module_percpu_addr() functions to hand back the canonical address of the per CPU address, i.e. the per CPU address minus the per CPU offset. 2) Check the lock address with these functions and if the per CPU check matches use the returned canonical address as the lock key, so all per CPU locks have the same key. 3) Move the static_obj(key) check into look_up_lock_class() so this check can be avoided for statically initialized per CPU locks. That's required because the canonical address fails the static_obj(key) check for obvious reasons. Reported-by:Mike Galbraith <efault@gmx.de> Signed-off-by:
-
J. R. Okajima authored
Commit: f8319483 ("locking/lockdep: Provide a type check for lock_is_held") didn't fully cover rwsems as downgrade_write() was left out. Introduce lock_downgrade() and use it to add new checks. See-also: http://marc.info/?l=linux-kernel&m=148581164003149&w=2Originally-written-by:
J. R. Okajima <hooanon05g@gmail.com> Signed-off-by:
-
J. R. Okajima authored
Behaviour should not change. Signed-off-by:
-
J. R. Okajima authored
A simple consolidataion to factor out repeated patterns. The behaviour should not change. Signed-off-by:
-
Ingo Molnar authored
Signed-off-by: -
Peter Zijlstra authored
Address this case: WARNING: CPU: 0 PID: 2070 at ../kernel/sched/core.c:109 update_rq_clock+0x74/0x80 rq->clock_update_flags & RQCF_UPDATED Call Trace: update_rq_clock() move_queued_task() __set_cpus_allowed_ptr() ... Signed-off-by:
-
Peter Zijlstra authored
Signed-off-by:
-
Peter Zijlstra authored
Add DEQUEUE_NOCLOCK to all places where we just did an update_rq_clock() already. Signed-off-by:
-
Peter Zijlstra authored
Instead of relying on deactivate_task() to call update_rq_clock() and handling the case where it didn't happen (task_on_rq_queued), unconditionally do update_rq_clock() and skip any further updates. This also avoids a double update on deactivate_task() + ttwu_local(). Signed-off-by:
-
Peter Zijlstra authored
Since all tasks on the wake_list are woken under a single rq->lock avoid calling update_rq_clock() for each task. Signed-off-by:
-
Peter Zijlstra authored
In all cases, ENQUEUE_RESTORE should also have ENQUEUE_NOCLOCK because DEQUEUE_SAVE will have done an update_rq_clock(). Signed-off-by:
-
Peter Zijlstra authored
Currently {en,de}queue_task() do an unconditional update_rq_clock(). However since we want to avoid duplicate updates, so that each rq->lock section appears atomic in time, we need to be able to skip these clock updates. Signed-off-by: -
Peter Zijlstra authored
The missing update_rq_clock() check can work with partial rq->lock wrappery, since a missing wrapper can cause the warning to not be emitted when it should have, but cannot cause the warning to trigger when it should not have. The duplicate update_rq_clock() check however can cause false warnings to trigger. Therefore add more comprehensive rq->lock wrappery. Signed-off-by:
-
