Function graph tracing is a bit different than the function tracers, as
it is processed after either the ftrace_caller or ftrace_regs_caller
and we only have one place to modify the jump to ftrace_graph_caller,
the jump needs to happen after the restore of registeres.
The function graph tracer is dependent on the function tracer, where
even if the function graph tracing is going on by itself, the save and
restore of registers is still done for function tracing regardless of
if function tracing is happening, before it calls the function graph
code.
If there's no function tracing happening, it is possible to just call
the function graph tracer directly, and avoid the wasted effort to save
and restore regs for function tracing.
This requires adding new flags to the dyn_ftrace records:
FTRACE_FL_TRAMP
FTRACE_FL_TRAMP_EN
The first is set if the count for the record is one, and the ftrace_ops
associated to that record has its own trampoline. That way the mcount code
can call that trampoline directly.
In the future, trampolines can be added to arbitrary ftrace_ops, where you
can have two or more ftrace_ops registered to ftrace (like kprobes and perf)
and if they are not tracing the same functions, then instead of doing a
loop to check all registered ftrace_ops against their hashes, just call the
ftrace_ops trampoline directly, which would call the registered ftrace_ops
function directly.
Without this patch perf showed:
0.05% hackbench [kernel.kallsyms] [k] ftrace_caller
0.05% hackbench [kernel.kallsyms] [k] arch_local_irq_save
0.05% hackbench [kernel.kallsyms] [k] native_sched_clock
0.04% hackbench [kernel.kallsyms] [k] __buffer_unlock_commit
0.04% hackbench [kernel.kallsyms] [k] preempt_trace
0.04% hackbench [kernel.kallsyms] [k] prepare_ftrace_return
0.04% hackbench [kernel.kallsyms] [k] __this_cpu_preempt_check
0.04% hackbench [kernel.kallsyms] [k] ftrace_graph_caller
See that the ftrace_caller took up more time than the ftrace_graph_caller
did.
With this patch:
0.05% hackbench [kernel.kallsyms] [k] __buffer_unlock_commit
0.04% hackbench [kernel.kallsyms] [k] call_filter_check_discard
0.04% hackbench [kernel.kallsyms] [k] ftrace_graph_caller
0.04% hackbench [kernel.kallsyms] [k] sched_clock
The ftrace_caller is no where to be found and ftrace_graph_caller still
takes up the same percentage.
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
The usage of uprobe_buffer_enable() added by dcad1a20 is very wrong,
1. uprobe_buffer_enable() and uprobe_buffer_disable() are not balanced,
_enable() should be called only if !enabled.
2. If uprobe_buffer_enable() fails probe_event_enable() should clear
tp.flags and free event_file_link.
3. If uprobe_register() fails it should do uprobe_buffer_disable().
Link: http://lkml.kernel.org/p/20140627170146.GA18332@redhat.com
Acked-by: Namhyung Kim <namhyung@kernel.org>
Acked-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com>
Reviewed-by: Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com>
Fixes: dcad1a204f "tracing/uprobes: Fetch args before reserving a ring buffer"
Signed-off-by: Oleg Nesterov <oleg@redhat.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
I do not know why dd9fa555d7 "tracing/uprobes: Move argument fetching
to uprobe_dispatcher()" added the UPROBE_HANDLER_REMOVE, but it looks
wrong.
OK, perhaps it makes sense to avoid store_trace_args() if the tracee is
nacked by uprobe_perf_filter(). But then we should kill the same code
in uprobe_perf_func() and unify the TRACE/PROFILE filtering (we need to
do this anyway to mix perf/ftrace). Until then this code actually adds
the pessimization because uprobe_perf_filter() will be called twice and
return T in likely case.
Link: http://lkml.kernel.org/p/20140627170143.GA18329@redhat.com
Acked-by: Namhyung Kim <namhyung@kernel.org>
Acked-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com>
Signed-off-by: Oleg Nesterov <oleg@redhat.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Add WARN_ON's into uprobe_unregister() and uprobe_apply() to ensure
that nobody tries to play with the dead uprobe/consumer. This helps
to catch the bugs like the one fixed by the previous patch.
In the longer term we should fix this poorly designed interface.
uprobe_register() should return "struct uprobe *" which should be
passed to apply/unregister. Plus other semantic changes, see the
changelog in commit 41ccba029e.
Link: http://lkml.kernel.org/p/20140627170140.GA18322@redhat.com
Acked-by: Namhyung Kim <namhyung@kernel.org>
Acked-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com>
Signed-off-by: Oleg Nesterov <oleg@redhat.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
This reverts commit 43fe98913c.
This patch is very wrong. Firstly, this change leads to unbalanced
uprobe_unregister(). Just for example,
# perf probe -x /lib/libc.so.6 syscall
# echo 1 >> /sys/kernel/debug/tracing/events/probe_libc/enable
# perf record -e probe_libc:syscall whatever
after that uprobe is dead (unregistered) but the user of ftrace/perf
can't know this, and it looks as if nobody hits this probe.
This would be easy to fix, but there are other reasons why it is not
simple to mix ftrace and perf. If nothing else, they can't share the
same ->consumer.filter. This is fixable too, but probably we need to
fix the poorly designed uprobe_register() interface first. At least
"register" and "apply" should be clearly separated.
Link: http://lkml.kernel.org/p/20140627170136.GA18319@redhat.com
Cc: Tom Zanussi <tom.zanussi@linux.intel.com>
Cc: "zhangwei(Jovi)" <jovi.zhangwei@huawei.com>
Cc: stable@vger.kernel.org # v3.14
Acked-by: Namhyung Kim <namhyung@kernel.org>
Acked-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com>
Reviewed-by: Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com>
Signed-off-by: Oleg Nesterov <oleg@redhat.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
We've converted cgroup to kernfs so cgroup won't be intertwined with
vfs objects and locking, but there are dark areas.
Run two instances of this script concurrently:
for ((; ;))
{
mount -t cgroup -o cpuacct xxx /cgroup
umount /cgroup
}
After a while, I saw two mount processes were stuck at retrying, because
they were waiting for a subsystem to become free, but the root associated
with this subsystem never got freed.
This can happen, if thread A is in the process of killing superblock but
hasn't called percpu_ref_kill(), and at this time thread B is mounting
the same cgroup root and finds the root in the root list and performs
percpu_ref_try_get().
To fix this, we try to increase both the refcnt of the superblock and the
percpu refcnt of cgroup root.
v2:
- we should try to get both the superblock refcnt and cgroup_root refcnt,
because cgroup_root may have no superblock assosiated with it.
- adjust/add comments.
tj: Updated comments. Renamed @sb to @pinned_sb.
Cc: <stable@vger.kernel.org> # 3.15
Signed-off-by: Li Zefan <lizefan@huawei.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
# cat test.sh
#! /bin/bash
mount -t cgroup -o cpu xxx /cgroup
umount /cgroup
mount -t cgroup -o cpu,cpuacct xxx /cgroup
umount /cgroup
# ./test.sh
mount: xxx already mounted or /cgroup busy
mount: according to mtab, xxx is already mounted on /cgroup
It's because the cgroupfs_root of the first mount was under destruction
asynchronously.
Fix this by delaying and then retrying mount for this case.
v3:
- put the refcnt immediately after getting it. (Tejun)
v2:
- use percpu_ref_tryget_live() rather that introducing
percpu_ref_alive(). (Tejun)
- adjust comment.
tj: Updated the comment a bit.
Cc: <stable@vger.kernel.org> # 3.15
Signed-off-by: Li Zefan <lizefan@huawei.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
The ftrace dynamic record has a flags element that also has a counter.
Instead of hard coding "rec->flags & ~FTRACE_FL_MASK" all over the
place. Use a macro instead.
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
When registering a function callback for the function tracer, the ops
can specify if it wants to save full regs (like an interrupt would)
for each function that it traces, or if it does not care about regs
and just wants to have the fastest return possible.
Once a ops has registered a function, if other ops register that
function they all will receive the regs too. That's because it does
the work once, it does it for everyone.
Now if the ops wanting regs unregisters the function so that there's
only ops left that do not care about regs, those ops will still
continue getting regs and going through the work for it on that
function. This is because the disabling of the rec counter only
sees the ops registered, and does not see the ops that are still
attached, and does not know if the current ops that are still attached
want regs or not. To play it safe, it just keeps regs being processed
until no function is registered anymore.
Instead of doing that, check the ops that are still registered for that
function and if none want regs for it anymore, then disable the
processing of regs.
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Currently, a percpu_ref undoes percpu_ref_init() automatically by
freeing the allocated percpu area when the percpu_ref is killed.
While seemingly convenient, this has the following niggles.
* It's impossible to re-init a released reference counter without
going through re-allocation.
* In the similar vein, it's impossible to initialize a percpu_ref
count with static percpu variables.
* We need and have an explicit destructor anyway for failure paths -
percpu_ref_cancel_init().
This patch removes the automatic percpu counter freeing in
percpu_ref_kill_rcu() and repurposes percpu_ref_cancel_init() into a
generic destructor now named percpu_ref_exit(). percpu_ref_destroy()
is considered but it gets confusing with percpu_ref_kill() while
"exit" clearly indicates that it's the counterpart of
percpu_ref_init().
All percpu_ref_cancel_init() users are updated to invoke
percpu_ref_exit() instead and explicit percpu_ref_exit() calls are
added to the destruction path of all percpu_ref users.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Benjamin LaHaise <bcrl@kvack.org>
Cc: Kent Overstreet <kmo@daterainc.com>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Benjamin LaHaise <bcrl@kvack.org>
Cc: Nicholas A. Bellinger <nab@linux-iscsi.org>
Cc: Li Zefan <lizefan@huawei.com>
When runing with the kernel(3.15-rc7+), the follow bug occurs:
[ 9969.258987] BUG: sleeping function called from invalid context at kernel/locking/mutex.c:586
[ 9969.359906] in_atomic(): 1, irqs_disabled(): 0, pid: 160655, name: python
[ 9969.441175] INFO: lockdep is turned off.
[ 9969.488184] CPU: 26 PID: 160655 Comm: python Tainted: G A 3.15.0-rc7+ #85
[ 9969.581032] Hardware name: FUJITSU-SV PRIMEQUEST 1800E/SB, BIOS PRIMEQUEST 1000 Series BIOS Version 1.39 11/16/2012
[ 9969.706052] ffffffff81a20e60 ffff8803e941fbd0 ffffffff8162f523 ffff8803e941fd18
[ 9969.795323] ffff8803e941fbe0 ffffffff8109995a ffff8803e941fc58 ffffffff81633e6c
[ 9969.884710] ffffffff811ba5dc ffff880405c6b480 ffff88041fdd90a0 0000000000002000
[ 9969.974071] Call Trace:
[ 9970.003403] [<ffffffff8162f523>] dump_stack+0x4d/0x66
[ 9970.065074] [<ffffffff8109995a>] __might_sleep+0xfa/0x130
[ 9970.130743] [<ffffffff81633e6c>] mutex_lock_nested+0x3c/0x4f0
[ 9970.200638] [<ffffffff811ba5dc>] ? kmem_cache_alloc+0x1bc/0x210
[ 9970.272610] [<ffffffff81105807>] cpuset_mems_allowed+0x27/0x140
[ 9970.344584] [<ffffffff811b1303>] ? __mpol_dup+0x63/0x150
[ 9970.409282] [<ffffffff811b1385>] __mpol_dup+0xe5/0x150
[ 9970.471897] [<ffffffff811b1303>] ? __mpol_dup+0x63/0x150
[ 9970.536585] [<ffffffff81068c86>] ? copy_process.part.23+0x606/0x1d40
[ 9970.613763] [<ffffffff810bf28d>] ? trace_hardirqs_on+0xd/0x10
[ 9970.683660] [<ffffffff810ddddf>] ? monotonic_to_bootbased+0x2f/0x50
[ 9970.759795] [<ffffffff81068cf0>] copy_process.part.23+0x670/0x1d40
[ 9970.834885] [<ffffffff8106a598>] do_fork+0xd8/0x380
[ 9970.894375] [<ffffffff81110e4c>] ? __audit_syscall_entry+0x9c/0xf0
[ 9970.969470] [<ffffffff8106a8c6>] SyS_clone+0x16/0x20
[ 9971.030011] [<ffffffff81642009>] stub_clone+0x69/0x90
[ 9971.091573] [<ffffffff81641c29>] ? system_call_fastpath+0x16/0x1b
The cause is that cpuset_mems_allowed() try to take
mutex_lock(&callback_mutex) under the rcu_read_lock(which was hold in
__mpol_dup()). And in cpuset_mems_allowed(), the access to cpuset is
under rcu_read_lock, so in __mpol_dup, we can reduce the rcu_read_lock
protection region to protect the access to cpuset only in
current_cpuset_is_being_rebound(). So that we can avoid this bug.
This patch is a temporary solution that just addresses the bug
mentioned above, can not fix the long-standing issue about cpuset.mems
rebinding on fork():
"When the forker's task_struct is duplicated (which includes
->mems_allowed) and it races with an update to cpuset_being_rebound
in update_tasks_nodemask() then the task's mems_allowed doesn't get
updated. And the child task's mems_allowed can be wrong if the
cpuset's nodemask changes before the child has been added to the
cgroup's tasklist."
Signed-off-by: Gu Zheng <guz.fnst@cn.fujitsu.com>
Acked-by: Li Zefan <lizefan@huawei.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: stable <stable@vger.kernel.org>
Pull tracing cleanups and fixes from Steven Rostedt:
"This includes three patches from Oleg Nesterov. The first is a fix to
a race condition that happens between enabling/disabling syscall
tracepoints and new process creations (the check to go into the ptrace
path for a process can be set when it shouldn't, or not set when it
should). Not a major bug but one that should be fixed and even
applied to stable.
The other two patches are cleanup/fixes that are not that critical,
but for an -rc1 release would be nice to have. They both deal with
syscall tracepoints.
It also includes a patch to introduce a new macro for the
TRACE_EVENT() format called __field_struct(). Originally, __field()
was used to record any variable into a trace event, but with the
addition of setting the "is signed" attribute, the check causes
anything but a primitive variable to fail to compile. That is,
structs and unions can't be used as they once were. When the "is
signed" check was introduce there were only primitive variables being
recorded. But that will change soon and it was reported that
__field() causes build failures.
To solve the __field() issue, __field_struct() is introduced to allow
trace_events to be able to record complex types too"
* tag 'trace-fixes-v3.16-rc1-v2' of git://git.kernel.org/pub/scm/linux/kernel/git/rostedt/linux-trace:
tracing: Add __field_struct macro for TRACE_EVENT()
tracing: syscall_regfunc() should not skip kernel threads
tracing: Change syscall_*regfunc() to check PF_KTHREAD and use for_each_process_thread()
tracing: Fix syscall_*regfunc() vs copy_process() race
Pull RCU fixes from Paul E. McKenney:
" This series includes the following:
1. Export a pair of debug-object interfaces for RCU that will
allow the slab allocators to avoid a recursion bug located
by Sasha Levin. Strictly speaking, this is not a regression,
but it would be good to enable the fix.
2. Address a serious performance regression on an open/close
micro-benchmark located by Dave Hansen. The offending commit
is ac1bea8578 (Make cond_resched() report RCU quiescent states). "
Signed-off-by: Ingo Molnar <mingo@kernel.org>
A 'softlockup' is defined as a bug that causes the kernel to loop in
kernel mode for more than a predefined period to time, without giving
other tasks a chance to run.
Currently, upon detection of this condition by the per-cpu watchdog
task, debug information (including a stack trace) is sent to the system
log.
On some occasions, we have observed that the "victim" rather than the
actual "culprit" (i.e. the owner/holder of the contended resource) is
reported to the user. Often this information has proven to be
insufficient to assist debugging efforts.
To avoid loss of useful debug information, for architectures which
support NMI, this patch makes it possible to improve soft lockup
reporting. This is accomplished by issuing an NMI to each cpu to obtain
a stack trace.
If NMI is not supported we just revert back to the old method. A sysctl
and boot-time parameter is available to toggle this feature.
[dzickus@redhat.com: add CONFIG_SMP in certain areas]
[akpm@linux-foundation.org: additional CONFIG_SMP=n optimisations]
[mq@suse.cz: fix warning]
Signed-off-by: Aaron Tomlin <atomlin@redhat.com>
Signed-off-by: Don Zickus <dzickus@redhat.com>
Cc: David S. Miller <davem@davemloft.net>
Cc: Mateusz Guzik <mguzik@redhat.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Signed-off-by: Jan Moskyto Matejka <mq@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Oleg reports a division by zero error on zero-length write() to the
percpu_pagelist_fraction sysctl:
divide error: 0000 [#1] SMP DEBUG_PAGEALLOC
CPU: 1 PID: 9142 Comm: badarea_io Not tainted 3.15.0-rc2-vm-nfs+ #19
Hardware name: Bochs Bochs, BIOS Bochs 01/01/2011
task: ffff8800d5aeb6e0 ti: ffff8800d87a2000 task.ti: ffff8800d87a2000
RIP: 0010: percpu_pagelist_fraction_sysctl_handler+0x84/0x120
RSP: 0018:ffff8800d87a3e78 EFLAGS: 00010246
RAX: 0000000000000f89 RBX: ffff88011f7fd000 RCX: 0000000000000000
RDX: 0000000000000000 RSI: 0000000000000001 RDI: 0000000000000010
RBP: ffff8800d87a3e98 R08: ffffffff81d002c8 R09: ffff8800d87a3f50
R10: 000000000000000b R11: 0000000000000246 R12: 0000000000000060
R13: ffffffff81c3c3e0 R14: ffffffff81cfddf8 R15: ffff8801193b0800
FS: 00007f614f1e9740(0000) GS:ffff88011f440000(0000) knlGS:0000000000000000
CS: 0010 DS: 0000 ES: 0000 CR0: 000000008005003b
CR2: 00007f614f1fa000 CR3: 00000000d9291000 CR4: 00000000000006e0
Call Trace:
proc_sys_call_handler+0xb3/0xc0
proc_sys_write+0x14/0x20
vfs_write+0xba/0x1e0
SyS_write+0x46/0xb0
tracesys+0xe1/0xe6
However, if the percpu_pagelist_fraction sysctl is set by the user, it
is also impossible to restore it to the kernel default since the user
cannot write 0 to the sysctl.
This patch allows the user to write 0 to restore the default behavior.
It still requires a fraction equal to or larger than 8, however, as
stated by the documentation for sanity. If a value in the range [1, 7]
is written, the sysctl will return EINVAL.
This successfully solves the divide by zero issue at the same time.
Signed-off-by: David Rientjes <rientjes@google.com>
Reported-by: Oleg Drokin <green@linuxhacker.ru>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Peter Wu noticed the following splat on his machine when updating
/proc/sys/kernel/watchdog_thresh:
BUG: sleeping function called from invalid context at mm/slub.c:965
in_atomic(): 1, irqs_disabled(): 0, pid: 1, name: init
3 locks held by init/1:
#0: (sb_writers#3){.+.+.+}, at: [<ffffffff8117b663>] vfs_write+0x143/0x180
#1: (watchdog_proc_mutex){+.+.+.}, at: [<ffffffff810e02d3>] proc_dowatchdog+0x33/0x110
#2: (cpu_hotplug.lock){.+.+.+}, at: [<ffffffff810589c2>] get_online_cpus+0x32/0x80
Preemption disabled at:[<ffffffff810e0384>] proc_dowatchdog+0xe4/0x110
CPU: 0 PID: 1 Comm: init Not tainted 3.16.0-rc1-testing #34
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Bochs 01/01/2011
Call Trace:
dump_stack+0x4e/0x7a
__might_sleep+0x11d/0x190
kmem_cache_alloc_trace+0x4e/0x1e0
perf_event_alloc+0x55/0x440
perf_event_create_kernel_counter+0x26/0xe0
watchdog_nmi_enable+0x75/0x140
update_timers_all_cpus+0x53/0xa0
proc_dowatchdog+0xe4/0x110
proc_sys_call_handler+0xb3/0xc0
proc_sys_write+0x14/0x20
vfs_write+0xad/0x180
SyS_write+0x49/0xb0
system_call_fastpath+0x16/0x1b
NMI watchdog: disabled (cpu0): hardware events not enabled
What happened is after updating the watchdog_thresh, the lockup detector
is restarted to utilize the new value. Part of this process involved
disabling preemption. Once preemption was disabled, perf tried to
allocate a new event (as part of the restart). This caused the above
BUG_ON as you can't sleep with preemption disabled.
The preemption restriction seemed agressive as we are not doing anything
on that particular cpu, but with all the online cpus (which are
protected by the get_online_cpus lock). Remove the restriction and the
BUG_ON goes away.
Signed-off-by: Don Zickus <dzickus@redhat.com>
Acked-by: Michal Hocko <mhocko@suse.cz>
Reported-by: Peter Wu <peter@lekensteyn.nl>
Tested-by: Peter Wu <peter@lekensteyn.nl>
Acked-by: David Rientjes <rientjes@google.com>
Cc: <stable@vger.kernel.org> [3.13+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
To allow filtering of huge pages, makedumpfile must be able to identify
them in the dump. This can be done by checking the appropriate page
flag, so communicate its value to makedumpfile through the VMCOREINFO
interface.
There's only one small catch. Depending on how many page flags are
available on a given architecture, this bit can be called PG_head or
PG_compound.
I sent a similar patch back in 2012, but Eric Biederman did not like
using an #ifdef. So, this time I'm adding a common symbol
(PG_head_mask) instead.
See https://lkml.org/lkml/2012/11/28/91 for the previous version.
Signed-off-by: Petr Tesarik <ptesarik@suse.cz>
Acked-by: Vivek Goyal <vgoyal@redhat.com>
Cc: Eric Biederman <ebiederm@xmission.com>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Fengguang Wu <fengguang.wu@intel.com>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Shaohua Li <shli@kernel.org>
Cc: Alexey Kardashevskiy <aik@ozlabs.ru>
Cc: Sasha Levin <sasha.levin@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
There is a race between the CPU offline code (within stop-machine) and
the smp-call-function code, which can lead to getting IPIs on the
outgoing CPU, *after* it has gone offline.
Specifically, this can happen when using
smp_call_function_single_async() to send the IPI, since this API allows
sending asynchronous IPIs from IRQ disabled contexts. The exact race
condition is described below.
During CPU offline, in stop-machine, we don't enforce any rule in the
_DISABLE_IRQ stage, regarding the order in which the outgoing CPU and
the other CPUs disable their local interrupts. Due to this, we can
encounter a situation in which an IPI is sent by one of the other CPUs
to the outgoing CPU (while it is *still* online), but the outgoing CPU
ends up noticing it only *after* it has gone offline.
CPU 1 CPU 2
(Online CPU) (CPU going offline)
Enter _PREPARE stage Enter _PREPARE stage
Enter _DISABLE_IRQ stage
=
Got a device interrupt, and | Didn't notice the IPI
the interrupt handler sent an | since interrupts were
IPI to CPU 2 using | disabled on this CPU.
smp_call_function_single_async() |
=
Enter _DISABLE_IRQ stage
Enter _RUN stage Enter _RUN stage
=
Busy loop with interrupts | Invoke take_cpu_down()
disabled. | and take CPU 2 offline
=
Enter _EXIT stage Enter _EXIT stage
Re-enable interrupts Re-enable interrupts
The pending IPI is noted
immediately, but alas,
the CPU is offline at
this point.
This of course, makes the smp-call-function IPI handler code running on
CPU 2 unhappy and it complains about "receiving an IPI on an offline
CPU".
One real example of the scenario on CPU 1 is the block layer's
complete-request call-path:
__blk_complete_request() [interrupt-handler]
raise_blk_irq()
smp_call_function_single_async()
However, if we look closely, the block layer does check that the target
CPU is online before firing the IPI. So in this case, it is actually
the unfortunate ordering/timing of events in the stop-machine phase that
leads to receiving IPIs after the target CPU has gone offline.
In reality, getting a late IPI on an offline CPU is not too bad by
itself (this can happen even due to hardware latencies in IPI
send-receive). It is a bug only if the target CPU really went offline
without executing all the callbacks queued on its list. (Note that a
CPU is free to execute its pending smp-call-function callbacks in a
batch, without waiting for the corresponding IPIs to arrive for each one
of those callbacks).
So, fixing this issue can be broken up into two parts:
1. Ensure that a CPU goes offline only after executing all the
callbacks queued on it.
2. Modify the warning condition in the smp-call-function IPI handler
code such that it warns only if an offline CPU got an IPI *and* that
CPU had gone offline with callbacks still pending in its queue.
Achieving part 1 is straight-forward - just flush (execute) all the
queued callbacks on the outgoing CPU in the CPU_DYING stage[1],
including those callbacks for which the source CPU's IPIs might not have
been received on the outgoing CPU yet. Once we do this, an IPI that
arrives late on the CPU going offline (either due to the race mentioned
above, or due to hardware latencies) will be completely harmless, since
the outgoing CPU would have executed all the queued callbacks before
going offline.
Overall, this fix (parts 1 and 2 put together) additionally guarantees
that we will see a warning only when the *IPI-sender code* is buggy -
that is, if it queues the callback _after_ the target CPU has gone
offline.
[1]. The CPU_DYING part needs a little more explanation: by the time we
execute the CPU_DYING notifier callbacks, the CPU would have already
been marked offline. But we want to flush out the pending callbacks at
this stage, ignoring the fact that the CPU is offline. So restructure
the IPI handler code so that we can by-pass the "is-cpu-offline?" check
in this particular case. (Of course, the right solution here is to fix
CPU hotplug to mark the CPU offline _after_ invoking the CPU_DYING
notifiers, but this requires a lot of audit to ensure that this change
doesn't break any existing code; hence lets go with the solution
proposed above until that is done).
[akpm@linux-foundation.org: coding-style fixes]
Signed-off-by: Srivatsa S. Bhat <srivatsa.bhat@linux.vnet.ibm.com>
Suggested-by: Frederic Weisbecker <fweisbec@gmail.com>
Cc: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com>
Cc: Borislav Petkov <bp@suse.de>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Gautham R Shenoy <ego@linux.vnet.ibm.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Mike Galbraith <mgalbraith@suse.de>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Rafael J. Wysocki <rjw@rjwysocki.net>
Cc: Rik van Riel <riel@redhat.com>
Cc: Rusty Russell <rusty@rustcorp.com.au>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Tested-by: Sachin Kamat <sachin.kamat@samsung.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Commit ac1bea8578 (Make cond_resched() report RCU quiescent states)
fixed a problem where a CPU looping in the kernel with but one runnable
task would give RCU CPU stall warnings, even if the in-kernel loop
contained cond_resched() calls. Unfortunately, in so doing, it introduced
performance regressions in Anton Blanchard's will-it-scale "open1" test.
The problem appears to be not so much the increased cond_resched() path
length as an increase in the rate at which grace periods complete, which
increased per-update grace-period overhead.
This commit takes a different approach to fixing this bug, mainly by
moving the RCU-visible quiescent state from cond_resched() to
rcu_note_context_switch(), and by further reducing the check to a
simple non-zero test of a single per-CPU variable. However, this
approach requires that the force-quiescent-state processing send
resched IPIs to the offending CPUs. These will be sent only once
the grace period has reached an age specified by the boot/sysfs
parameter rcutree.jiffies_till_sched_qs, or once the grace period
reaches an age halfway to the point at which RCU CPU stall warnings
will be emitted, whichever comes first.
Reported-by: Dave Hansen <dave.hansen@intel.com>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Christoph Lameter <cl@gentwo.org>
Cc: Mike Galbraith <umgwanakikbuti@gmail.com>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Reviewed-by: Josh Triplett <josh@joshtriplett.org>
[ paulmck: Made rcu_momentary_dyntick_idle() as suggested by the
ktest build robot. Also fixed smp_mb() comment as noted by
Oleg Nesterov. ]
Merge with e552592e (Reduce overhead of cond_resched() checks for RCU)
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Currently, call_rcu() relies on implicit allocation and initialization
for the debug-objects handling of RCU callbacks. If you hammer the
kernel hard enough with Sasha's modified version of trinity, you can end
up with the sl*b allocators recursing into themselves via this implicit
call_rcu() allocation.
This commit therefore exports the debug_init_rcu_head() and
debug_rcu_head_free() functions, which permits the allocators to allocated
and pre-initialize the debug-objects information, so that there no longer
any need for call_rcu() to do that initialization, which in turn prevents
the recursion into the memory allocators.
Reported-by: Sasha Levin <sasha.levin@oracle.com>
Suggested-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Looks-good-to: Christoph Lameter <cl@linux.com>
In lowres mode, hrtimers are serviced by the tick instead of a clock
event. It works well as long as the tick stays periodic but we must also
make sure that the hrtimers are serviced in dynticks mode targets,
pretty much like timer list timers do.
Note that all dynticks modes are concerned: get_nohz_timer_target()
tries not to return remote idle CPUs but there is nothing to prevent
the elected target from entering dynticks idle mode until we lock its
base. It's also prefectly legal to enqueue hrtimers on full dynticks CPU.
So there are two requirements to correctly handle dynticks:
1) On target's tick stop time, we must not delay the next tick further
the next hrtimer.
2) On hrtimer queue time. If the tick of the target is stopped, we must
wake up that CPU such that it sees the new hrtimer and recalculate
the next tick accordingly.
The point 1 is well handled currently through get_nohz_timer_interrupt() and
cmp_next_hrtimer_event().
But the point 2 isn't handled at all.
Fixing this is easy though as we have the necessary API ready for that.
All we need is to call wake_up_nohz_cpu() on a target when a newly
enqueued hrtimer requires tick rescheduling, like timer list timer do.
Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Link: http://lkml.kernel.org/r/3d7ea08ce008698e26bd39fe10f55949391073ab.1403507178.git.viresh.kumar@linaro.org
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
In lowres mode, hrtimers are serviced by the tick instead of a clock
event. Now it works well as long as the tick stays periodic but we
must also make sure that the hrtimers are serviced in dynticks mode.
Part of that job consist in kicking a dynticks hrtimer target in order
to make it reconsider the next tick to schedule to correctly handle the
hrtimer's expiring time. And that part isn't handled by the hrtimers
subsystem.
To prepare for fixing this, we need __hrtimer_start_range_ns() to be
able to resolve the CPU target associated to a hrtimer's object
'cpu_base' so that the kick can be centralized there.
So lets store it in the 'struct hrtimer_cpu_base' to resolve the CPU
without overhead. It is set once at CPU's online notification.
Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Link: http://lkml.kernel.org/r/1403393357-2070-4-git-send-email-fweisbec@gmail.com
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
When a timer is enqueued or modified on a dynticks target, that CPU
must re-evaluate the next tick to service that timer.
The tick re-evaluation is performed by an IPI kick on the target.
Now while we correctly call wake_up_nohz_cpu() from add_timer_on(), the
mod_timer*() API family doesn't support so well dynticks targets.
The reason for this is likely that __mod_timer() isn't supposed to
select an idle target for a timer, unless that target is the current
CPU, in which case a dynticks idle kick isn't actually needed.
But there is a small race window lurking behind that assumption: the
elected target has all the time to turn dynticks idle between the call
to get_nohz_timer_target() and the locking of its base. Hence a risk
that we enqueue a timer on a dynticks idle destination without kicking
it. As a result, the timer might be serviced too late in the future.
Also a target elected by __mod_timer() can be in full dynticks mode
and thus require to be kicked as well. And unlike idle dynticks, this
concern both local and remote targets.
To fix this whole issue, lets centralize the dynticks kick to
internal_add_timer() so that it is well handled for all sort of timer
enqueue. Even timer migration is concerned so that a full dynticks target
is correctly kicked as needed when timers are migrating to it.
Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Link: http://lkml.kernel.org/r/1403393357-2070-3-git-send-email-fweisbec@gmail.com
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Timers are serviced by the tick. But when a timer is enqueued on a
dynticks target, we need to kick it in order to make it reconsider the
next tick to schedule to correctly handle the timer's expiring time.
Now while this kick is correctly performed for add_timer_on(), the
mod_timer*() family has been a bit neglected.
To prepare for fixing this, we need internal_add_timer() to be able to
resolve the CPU target associated to a timer's object 'base' so that the
kick can be centralized there.
This can't be passed as an argument as not all the callers know the CPU
number of a timer's base. So lets store it in the struct tvec_base to
resolve the CPU without much overhead. It is set once for good at every
CPU's first boot.
Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Link: http://lkml.kernel.org/r/1403393357-2070-2-git-send-email-fweisbec@gmail.com
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
futex_lock_pi_atomic() is a maze of retry hoops and loops.
Reduce it to simple and understandable states:
First step is to lookup existing waiters (state) in the kernel.
If there is an existing waiter, validate it and attach to it.
If there is no existing waiter, check the user space value
If the TID encoded in the user space value is 0, take over the futex
preserving the owner died bit.
If the TID encoded in the user space value is != 0, lookup the owner
task, validate it and attach to it.
Reduces text size by 128 bytes on x8664.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Davidlohr Bueso <davidlohr@hp.com>
Cc: Kees Cook <kees@outflux.net>
Cc: wad@chromium.org
Cc: Darren Hart <darren@dvhart.com>
Link: http://lkml.kernel.org/r/alpine.DEB.2.10.1406131137020.5170@nanos
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
The kernel tries to atomically unlock the futex without checking
whether there is kernel state associated to the futex.
So if user space manipulated the user space value, this will leave
kernel internal state around associated to the owner task.
For robustness sake, lookup first whether there are waiters on the
futex. If there are waiters, wake the top priority waiter with all the
proper sanity checks applied.
If there are no waiters, do the atomic release. We do not have to
preserve the waiters bit in this case, because a potentially incoming
waiter is blocked on the hb->lock and will acquire the futex
atomically. We neither have to preserve the owner died bit. The caller
is the owner and it was supposed to cleanup the mess.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Darren Hart <darren@dvhart.com>
Cc: Davidlohr Bueso <davidlohr@hp.com>
Cc: Kees Cook <kees@outflux.net>
Cc: wad@chromium.org
Link: http://lkml.kernel.org/r/20140611204237.016987332@linutronix.de
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
In case the dead lock detector is enabled we follow the lock chain to
the end in rt_mutex_adjust_prio_chain, even if we could stop earlier
due to the priority/waiter constellation.
But once we are no longer the top priority waiter in a certain step
or the task holding the lock has already the same priority then there
is no point in dequeing and enqueing along the lock chain as there is
no change at all.
So stop the queueing at this point.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Reviewed-by: Steven Rostedt <rostedt@goodmis.org>
Cc: Lai Jiangshan <laijs@cn.fujitsu.com>
Link: http://lkml.kernel.org/r/20140522031950.280830190@linutronix.de
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
The deadlock logic is only required for futexes.
Remove the extra arguments for the public functions and also for the
futex specific ones which get always called with deadlock detection
enabled.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Steven Rostedt <rostedt@goodmis.org>
Exit right away, when the removed waiter was not the top priority
waiter on the lock. Get rid of the extra indent level.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Steven Rostedt <rostedt@goodmis.org>
Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Add commentry to document the chain walk and the protection mechanisms
and their scope.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Steven Rostedt <rostedt@goodmis.org>
Add a separate local variable for the boost/deboost logic to make the
code more readable. Add comments where appropriate.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Steven Rostedt <rostedt@goodmis.org>
There is no point to keep the task ref across the check for lock
owner. Drop the ref before that, so the protection context is clear.
Found while documenting the chain walk.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Steven Rostedt <rostedt@goodmis.org>
Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
The current implementation of try_to_take_rtmutex() is correct, but
requires more than a single brain twist to understand the clever
encoded conditionals.
Untangle it and document the cases proper.
Looks less efficient at the first glance, but actually reduces the
binary code size on x8664 by 80 bytes.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Steven Rostedt <rostedt@goodmis.org>
Oleg noticed that rtmutex_slowtrylock() has a pointless check for
rt_mutex_owner(lock) != current.
To avoid calling try_to_take_rtmutex() we really want to check whether
the lock has an owner at all or whether the trylock failed because the
owner is NULL, but the RT_MUTEX_HAS_WAITERS bit is set. This covers
the lock is owned by caller situation as well.
We can actually do this check lockless. trylock is taking a chance
whether we take lock->wait_lock to do the check or not.
Add comments to the function while at it.
Reported-by: Oleg Nesterov <oleg@redhat.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Steven Rostedt <rostedt@goodmis.org>
Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Pull perf fixes from Ingo Molnar:
"This is larger than usual: the main reason are the ARM symbol lookup
speedups that came in late and were hard to resist.
There's also a kprobes fix and various tooling fixes, plus the minimal
re-enablement of the mmap2 support interface"
* 'perf-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip: (36 commits)
x86/kprobes: Fix build errors and blacklist context_track_user
perf tests: Add test for closing dso objects on EMFILE error
perf tests: Add test for caching dso file descriptors
perf tests: Allow reuse of test_file function
perf tests: Spawn child for each test
perf tools: Add dso__data_* interface descriptons
perf tools: Allow to close dso fd in case of open failure
perf tools: Add file size check and factor dso__data_read_offset
perf tools: Cache dso data file descriptor
perf tools: Add global count of opened dso objects
perf tools: Add global list of opened dso objects
perf tools: Add data_fd into dso object
perf tools: Separate dso data related variables
perf tools: Cache register accesses for unwind processing
perf record: Fix to honor user freq/interval properly
perf timechart: Reflow documentation
perf probe: Improve error messages in --line option
perf probe: Improve an error message of perf probe --vars mode
perf probe: Show error code and description in verbose mode
perf probe: Improve error message for unknown member of data structure
...
Pull rtmutex fixes from Thomas Gleixner:
"Another three patches to make the rtmutex code more robust. That's
the last urgent fallout from the big futex/rtmutex investigation"
* 'locking-urgent-for-linus.patch' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
rtmutex: Plug slow unlock race
rtmutex: Detect changes in the pi lock chain
rtmutex: Handle deadlock detection smarter
syscall_regfunc() ignores the kernel threads because "it has no effect",
see cc3b13c1 "Don't trace kernel thread syscalls" which added this check.
However, this means that a user-space task spawned by call_usermodehelper()
will run without TIF_SYSCALL_TRACEPOINT if sys_tracepoint_refcount != 0.
Remove this check. The unnecessary report from ret_from_fork path mentioned
by cc3b13c1 is no longer possible, see See commit fb45550d76 "make sure
that kernel_thread() callbacks call do_exit() themselves".
A kernel_thread() callback can only return and take the int_ret_from_sys_call
path after do_execve() succeeds, otherwise the kernel will crash. But in this
case it is no longer a kernel thread and thus is needs TIF_SYSCALL_TRACEPOINT.
Link: http://lkml.kernel.org/p/20140413185938.GD20668@redhat.com
Signed-off-by: Oleg Nesterov <oleg@redhat.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
1. Remove _irqsafe from syscall_regfunc/syscall_unregfunc,
read_lock(tasklist) doesn't need to disable irqs.
2. Change this code to avoid the deprecated do_each_thread()
and use for_each_process_thread() (stolen from the patch
from Frederic).
3. Change syscall_regfunc() to check PF_KTHREAD to skip
the kernel threads, ->mm != NULL is the common mistake.
Note: probably this check should be simply removed, needs
another patch.
[fweisbec@gmail.com: s/do_each_thread/for_each_process_thread/]
Link: http://lkml.kernel.org/p/20140413185918.GC20668@redhat.com
Signed-off-by: Oleg Nesterov <oleg@redhat.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Pull ACPI and power management fixes from Rafael Wysocki:
"These are fixes mostly (ia64 regression related to the ACPI
enumeration of devices, cpufreq regressions, fix for I2C controllers
included in Intel SoCs, mvebu cpuidle driver fix related to sysfs)
plus additional kernel command line arguments from Kees to make it
possible to build kernel images with hibernation and the kernel
address space randomization included simultaneously, a new ACPI
battery driver quirk for a system with a broken BIOS and a couple of
ACPI core cleanups.
Specifics:
- Fix for an ia64 regression introduced during the 3.11 cycle by a
commit that modified the hardware initialization ordering and made
device discovery fail on some systems.
- Fix for a build problem on systems where the cpufreq-cpu0 driver is
built-in and the cpu-thermal driver is modular from Arnd Bergmann.
- Fix for a recently introduced computational mistake in the
intel_pstate driver that leads to excessive rounding errors from
Doug Smythies.
- Fix for a failure code path in cpufreq_update_policy() that fails
to unlock the locks acquired previously from Aaron Plattner.
- Fix for the cpuidle mvebu driver to use shorter state names which
will prevent the sysfs interface from returning mangled strings.
From Gregory Clement.
- ACPI LPSS driver fix to make sure that the I2C controllers included
in BayTrail SoCs are not held in the reset state while they are
being probed from Mika Westerberg.
- New kernel command line arguments making it possible to build
kernel images with hibernation and kASLR included at the same time
and to select which of them will be used via the command line (they
are still functionally mutually exclusive, though). From Kees
Cook.
- ACPI battery driver quirk for Acer Aspire V5-573G that fails to
send battery status change notifications timely from Alexander
Mezin.
- Two ACPI core cleanups from Christoph Jaeger and Fabian Frederick"
* tag 'pm+acpi-3.16-rc2' of git://git.kernel.org/pub/scm/linux/kernel/git/rafael/linux-pm:
cpuidle: mvebu: Fix the name of the states
cpufreq: unlock when failing cpufreq_update_policy()
intel_pstate: Correct rounding in busy calculation
ACPI: use kstrto*() instead of simple_strto*()
ACPI / processor replace __attribute__((packed)) by __packed
ACPI / battery: add quirk for Acer Aspire V5-573G
ACPI / battery: use callback for setting up quirks
ACPI / LPSS: Take I2C host controllers out of reset
x86, kaslr: boot-time selectable with hibernation
PM / hibernate: introduce "nohibernate" boot parameter
cpufreq: cpufreq-cpu0: fix CPU_THERMAL dependency
ACPI / ia64 / sba_iommu: Restore the working initialization ordering
