llvm can optimize the 'if (ptr > data_end)' checks to be in the order
slightly different than the original C code which will confuse verifier.
Like:
if (ptr + 16 > data_end)
return TC_ACT_SHOT;
// may be followed by
if (ptr + 14 > data_end)
return TC_ACT_SHOT;
while llvm can see that 'ptr' is valid for all 16 bytes,
the verifier could not.
Fix verifier logic to account for such case and add a test.
Reported-by: Huapeng Zhou <hzhou@fb.com>
Fixes: 969bf05eb3 ("bpf: direct packet access")
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Martin KaFai Lau <kafai@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently trace_handle_return() looks like this:
static inline enum print_line_t trace_handle_return(struct trace_seq *s)
{
return trace_seq_has_overflowed(s) ?
TRACE_TYPE_PARTIAL_LINE : TRACE_TYPE_HANDLED;
}
Where trace_seq_overflowed(s) is:
static inline bool trace_seq_has_overflowed(struct trace_seq *s)
{
return s->full || seq_buf_has_overflowed(&s->seq);
}
And seq_buf_has_overflowed(&s->seq) is:
static inline bool
seq_buf_has_overflowed(struct seq_buf *s)
{
return s->len > s->size;
}
Making trace_handle_return() into:
return (s->full || (s->seq->len > s->seq->size)) ?
TRACE_TYPE_PARTIAL_LINE :
TRACE_TYPE_HANDLED;
One would think this is not an issue to keep as an inline. But because this
is used in the TRACE_EVENT() macro, it is extended for every tracepoint in
the system. Taking a look at a single tracepoint x86_irq_vector (was the
first one I randomly chosen). As trace_handle_return is used in the
TRACE_EVENT() macro of trace_raw_output_##call() we disassemble
trace_raw_output_x86_irq_vector and do a diff:
- is the original
+ is the out-of-line code
I removed identical lines that were different just due to different
addresses.
--- /tmp/irq-vec-orig 2017-03-16 09:12:48.569384851 -0400
+++ /tmp/irq-vec-ool 2017-03-16 09:13:39.378153385 -0400
@@ -6,27 +6,23 @@
53 push %rbx
48 89 fb mov %rdi,%rbx
4c 8b a7 c0 20 00 00 mov 0x20c0(%rdi),%r12
e8 f7 72 13 00 callq ffffffff81155c80 <trace_raw_output_prep>
83 f8 01 cmp $0x1,%eax
74 05 je ffffffff8101e993 <trace_raw_output_x86_irq_vector+0x23>
5b pop %rbx
41 5c pop %r12
5d pop %rbp
c3 retq
41 8b 54 24 08 mov 0x8(%r12),%edx
- 48 8d bb 98 10 00 00 lea 0x1098(%rbx),%rdi
+ 48 81 c3 98 10 00 00 add $0x1098,%rbx
- 48 c7 c6 7b 8a a0 81 mov $0xffffffff81a08a7b,%rsi
+ 48 c7 c6 ab 8a a0 81 mov $0xffffffff81a08aab,%rsi
- e8 c5 85 13 00 callq ffffffff81156f70 <trace_seq_printf>
=== here's the start of the main difference ===
+ 48 89 df mov %rbx,%rdi
+ e8 62 7e 13 00 callq ffffffff81156810 <trace_seq_printf>
- 8b 93 b8 20 00 00 mov 0x20b8(%rbx),%edx
- 31 c0 xor %eax,%eax
- 85 d2 test %edx,%edx
- 75 11 jne ffffffff8101e9c8 <trace_raw_output_x86_irq_vector+0x58>
- 48 8b 83 a8 20 00 00 mov 0x20a8(%rbx),%rax
- 48 39 83 a0 20 00 00 cmp %rax,0x20a0(%rbx)
- 0f 93 c0 setae %al
+ 48 89 df mov %rbx,%rdi
+ e8 4a c5 12 00 callq ffffffff8114af00 <trace_handle_return>
5b pop %rbx
- 0f b6 c0 movzbl %al,%eax
=== end ===
41 5c pop %r12
5d pop %rbp
c3 retq
If you notice, the original has 22 bytes of text more than the out of line
version. As this is for every TRACE_EVENT() defined in the system, this can
become quite large.
text data bss dec hex filename
8690305 5450490 1298432 15439227 eb957b vmlinux-orig
8681725 5450490 1298432 15430647 eb73f7 vmlinux-handle
This change has a total of 8580 bytes in savings.
$ objdump -dr /tmp/vmlinux-orig | grep '^[0-9a-f]* <trace_raw_output' | wc -l
324
That's 324 tracepoints. But this does not include modules (which contain
many more tracepoints). For an allyesconfig build:
$ objdump -dr vmlinux-allyes-orig | grep '^[0-9a-f]* <trace_raw_output' | wc -l
1401
That's 1401 tracepoints giving us:
text data bss dec hex filename
137920629 140221067 53264384 331406080 13c0db00 vmlinux-allyes-orig
137827709 140221067 53264384 331313160 13bf7008 vmlinux-allyes-handle
92920 bytes in savings!!!
Link: http://lkml.kernel.org/r/20170315021431.13107-2-andi@firstfloor.org
Reported-by: Andi Kleen <andi@firstfloor.org>
Signed-off-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
Register the function tracer right after the tracing buffers are initialized
in early boot up. This will allow function tracing to begin early if it is
enabled via the kernel command line.
Signed-off-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
As tracing can now be enabled very early in boot up, even before some
critical system services (like scheduling), do not run the tracer selftests
until after early_initcall() is performed. If a tracer is registered before
such time, it is saved off in a list and the test is run when the system is
able to handle more diverse functions.
Signed-off-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
Create an early_trace_init() function that will initialize the buffers and
allow for ealier use of trace_printk(). This will also allow for future work
to have function tracing start earlier at boot up.
Signed-off-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
There is no need to always call blocking console_lock() in
console_cpu_notify(), it's quite possible that console_sem can
be locked by other CPU on the system, either already printing
or soon to begin printing the messages. console_lock() in this
case can simply block CPU hotplug for unknown period of time
(console_unlock() is time unbound). Not that hotplug is very
fast, but still, with other CPUs being online and doing
printk() console_cpu_notify() can stuck.
Use console_trylock() instead and opt-out if console_sem is
already acquired from another CPU, since that CPU will do
the printing for us.
Link: http://lkml.kernel.org/r/20170121104729.8585-1-sergey.senozhatsky@gmail.com
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Sergey Senozhatsky <sergey.senozhatsky@gmail.com>
Signed-off-by: Petr Mladek <pmladek@suse.com>
This patch fix spelling typos found in
Documentation/output/xml/driver-api/basics.xml.
It is because the xml file was generated from comments in source,
so I had to fix the comments.
Signed-off-by: Masanari Iida <standby24x7@gmail.com>
Signed-off-by: Jiri Kosina <jkosina@suse.cz>
Under extremely heavy uses of padata, crashes occur, and with list
debugging turned on, this happens instead:
[87487.298728] WARNING: CPU: 1 PID: 882 at lib/list_debug.c:33
__list_add+0xae/0x130
[87487.301868] list_add corruption. prev->next should be next
(ffffb17abfc043d0), but was ffff8dba70872c80. (prev=ffff8dba70872b00).
[87487.339011] [<ffffffff9a53d075>] dump_stack+0x68/0xa3
[87487.342198] [<ffffffff99e119a1>] ? console_unlock+0x281/0x6d0
[87487.345364] [<ffffffff99d6b91f>] __warn+0xff/0x140
[87487.348513] [<ffffffff99d6b9aa>] warn_slowpath_fmt+0x4a/0x50
[87487.351659] [<ffffffff9a58b5de>] __list_add+0xae/0x130
[87487.354772] [<ffffffff9add5094>] ? _raw_spin_lock+0x64/0x70
[87487.357915] [<ffffffff99eefd66>] padata_reorder+0x1e6/0x420
[87487.361084] [<ffffffff99ef0055>] padata_do_serial+0xa5/0x120
padata_reorder calls list_add_tail with the list to which its adding
locked, which seems correct:
spin_lock(&squeue->serial.lock);
list_add_tail(&padata->list, &squeue->serial.list);
spin_unlock(&squeue->serial.lock);
This therefore leaves only place where such inconsistency could occur:
if padata->list is added at the same time on two different threads.
This pdata pointer comes from the function call to
padata_get_next(pd), which has in it the following block:
next_queue = per_cpu_ptr(pd->pqueue, cpu);
padata = NULL;
reorder = &next_queue->reorder;
if (!list_empty(&reorder->list)) {
padata = list_entry(reorder->list.next,
struct padata_priv, list);
spin_lock(&reorder->lock);
list_del_init(&padata->list);
atomic_dec(&pd->reorder_objects);
spin_unlock(&reorder->lock);
pd->processed++;
goto out;
}
out:
return padata;
I strongly suspect that the problem here is that two threads can race
on reorder list. Even though the deletion is locked, call to
list_entry is not locked, which means it's feasible that two threads
pick up the same padata object and subsequently call list_add_tail on
them at the same time. The fix is thus be hoist that lock outside of
that block.
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Acked-by: Steffen Klassert <steffen.klassert@secunet.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Pull power management fixes from Rafael Wysocki:
"One of these is an intel_pstate regression fix and it is not a small
change, but it mostly removes code that shouldn't be there. That code
was acquired by mistake and has been a source of constant pain since
then, so the time has come to get rid of it finally. We have not seen
problems with this change in the lab, so fingers crossed.
The rest is more usual: one more intel_pstate commit removing useless
code, a cpufreq core fix to make it restore policy limits on CPU
online (which prevents the limits from being reset over system
suspend/resume), a schedutil cpufreq governor initialization fix to
make it actually work as advertised on all systems and an extra sanity
check in the cpuidle core to prevent crashes from happening if the
arch code messes things up.
Specifics:
- Make intel_pstate use one set of global P-state limits in the
active mode regardless of the scaling_governor settings for
individual CPUs instead of switching back and forth between two of
them in a way that is hard to control (Rafael Wysocki).
- Drop a useless function from intel_pstate to prevent it from
modifying the maximum supported frequency value unexpectedly which
may confuse the cpufreq core (Rafael Wysocki).
- Fix the cpufreq core to restore policy limits on CPU online so that
the limits are not reset over system suspend/resume, among other
things (Viresh Kumar).
- Fix the initialization of the schedutil cpufreq governor to make
the IO-wait boosting mechanism in it actually work on systems with
one CPU per cpufreq policy (Rafael Wysocki).
- Add a sanity check to the cpuidle core to prevent crashes from
happening if the architecture code initialization fails to set up
things as expected (Vaidyanathan Srinivasan)"
* tag 'pm-4.11-rc4' of git://git.kernel.org/pub/scm/linux/kernel/git/rafael/linux-pm:
cpufreq: Restore policy min/max limits on CPU online
cpuidle: Validate cpu_dev in cpuidle_add_sysfs()
cpufreq: intel_pstate: Fix policy data management in passive mode
cpufreq: schedutil: Fix per-CPU structure initialization in sugov_start()
cpufreq: intel_pstate: One set of global limits in active mode
sugov_update_commit() calls trace_cpu_frequency() to record the
current CPU frequency if it has not changed in the fast switch case
to prevent utilities from getting confused (they may report that the
CPU is idle if the frequency has not been recorded for too long, for
example).
However, that may cause the tracepoint to be triggered quite often
for no real reason (if the frequency doesn't change, we will not
modify the last update time stamp and governor computations may
run again shortly when that happens), so don't do that (arguably, it
is done to work around a utilities bug anyway).
That allows code duplication in sugov_update_commit() to be reduced
somewhat too.
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>
Conflicts:
drivers/net/ethernet/broadcom/genet/bcmmii.c
drivers/net/hyperv/netvsc.c
kernel/bpf/hashtab.c
Almost entirely overlapping changes.
Signed-off-by: David S. Miller <davem@davemloft.net>
On systems with a large number of CPUs, running sysrq-<q> can cause
watchdog timeouts. There are two slow sections of code in the sysrq-<q>
path in timer_list.c.
1. print_active_timers() - This function is called by print_cpu() and
contains a slow goto loop. On a machine with hundreds of CPUs, this
loop took approximately 100ms for the first CPU in a NUMA node.
(Subsequent CPUs in the same node ran much quicker.) The total time
to print all of the CPUs is ultimately long enough to trigger the
soft lockup watchdog.
2. print_tickdevice() - This function outputs a large amount of textual
information. This function also took approximately 100ms per CPU.
Since sysrq-<q> is not a performance critical path, there should be no
harm in touching the nmi watchdog in both slow sections above. Touching
it in just one location was insufficient on systems with hundreds of
CPUs as occasional timeouts were still observed during testing.
This issue was observed on an Oracle T7 machine with 128 CPUs, but I
anticipate it may affect other systems with similarly large numbers of
CPUs.
Signed-off-by: Tom Hromatka <tom.hromatka@oracle.com>
Reviewed-by: Rob Gardner <rob.gardner@oracle.com>
Signed-off-by: John Stultz <john.stultz@linaro.org>
The scheduler clock framework may not use the correct timeout for the clock
wrap. This happens when a new clock driver calls sched_clock_register()
after the kernel called sched_clock_postinit(). In this case the clock wrap
timeout is too long thus sched_clock_poll() is called too late and the clock
already wrapped.
On my ARM system the scheduler was no longer scheduling any other task than
the idle task because the sched_clock() wrapped.
Signed-off-by: David Engraf <david.engraf@sysgo.com>
Signed-off-by: John Stultz <john.stultz@linaro.org>
A clockevent device's rate should be configured before or at registration
and changed afterwards through clockevents_update_freq() only.
For the configuration at registration, we already have
clockevents_config_and_register().
Right now, there are no clockevents_config() users outside of the
clockevents core.
To mitigiate the risk of drivers errorneously reconfiguring their rates
through clockevents_config() *after* device registration, make
clockevents_config() static.
Signed-off-by: Nicolai Stange <nicstange@gmail.com>
Signed-off-by: John Stultz <john.stultz@linaro.org>
Pull networking fixes from David Miller:
1) Several netfilter fixes from Pablo and the crew:
- Handle fragmented packets properly in netfilter conntrack, from
Florian Westphal.
- Fix SCTP ICMP packet handling, from Ying Xue.
- Fix big-endian bug in nftables, from Liping Zhang.
- Fix alignment of fake conntrack entry, from Steven Rostedt.
2) Fix feature flags setting in fjes driver, from Taku Izumi.
3) Openvswitch ipv6 tunnel source address not set properly, from Or
Gerlitz.
4) Fix jumbo MTU handling in amd-xgbe driver, from Thomas Lendacky.
5) sk->sk_frag.page not released properly in some cases, from Eric
Dumazet.
6) Fix RTNL deadlocks in nl80211, from Johannes Berg.
7) Fix erroneous RTNL lockdep splat in crypto, from Herbert Xu.
8) Cure improper inflight handling during AF_UNIX GC, from Andrey
Ulanov.
9) sch_dsmark doesn't write to packet headers properly, from Eric
Dumazet.
10) Fix SCM_TIMESTAMPING_OPT_STATS handling in TCP, from Soheil Hassas
Yeganeh.
11) Add some IDs for Motorola qmi_wwan chips, from Tony Lindgren.
12) Fix nametbl deadlock in tipc, from Ying Xue.
13) GRO and LRO packets not counted correctly in mlx5 driver, from Gal
Pressman.
14) Fix reset of internal PHYs in bcmgenet, from Doug Berger.
15) Fix hashmap allocation handling, from Alexei Starovoitov.
16) nl_fib_input() needs stronger netlink message length checking, from
Eric Dumazet.
17) Fix double-free of sk->sk_filter during sock clone, from Daniel
Borkmann.
18) Fix RX checksum offloading in aquantia driver, from Pavel Belous.
* git://git.kernel.org/pub/scm/linux/kernel/git/davem/net: (85 commits)
net:ethernet:aquantia: Fix for RX checksum offload.
amd-xgbe: Fix the ECC-related bit position definitions
sfc: cleanup a condition in efx_udp_tunnel_del()
Bluetooth: btqcomsmd: fix compile-test dependency
inet: frag: release spinlock before calling icmp_send()
tcp: initialize icsk_ack.lrcvtime at session start time
genetlink: fix counting regression on ctrl_dumpfamily()
socket, bpf: fix sk_filter use after free in sk_clone_lock
ipv4: provide stronger user input validation in nl_fib_input()
bpf: fix hashmap extra_elems logic
enic: update enic maintainers
net: bcmgenet: remove bcmgenet_internal_phy_setup()
ipv6: make sure to initialize sockc.tsflags before first use
fjes: Do not load fjes driver if extended socket device is not power on.
fjes: Do not load fjes driver if system does not have extended socket device.
net/mlx5e: Count LRO packets correctly
net/mlx5e: Count GSO packets correctly
net/mlx5: Increase number of max QPs in default profile
net/mlx5e: Avoid supporting udp tunnel port ndo for VF reps
net/mlx5e: Use the proper UAPI values when offloading TC vlan actions
...
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: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: juri.lelli@arm.com
Cc: xlpang@redhat.com
Cc: rostedt@goodmis.org
Cc: mathieu.desnoyers@efficios.com
Cc: jdesfossez@efficios.com
Cc: dvhart@infradead.org
Cc: bristot@redhat.com
Link: http://lkml.kernel.org/r/20170322104152.161341537@infradead.org
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
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 (Intel) <peterz@infradead.org>
Cc: juri.lelli@arm.com
Cc: bigeasy@linutronix.de
Cc: xlpang@redhat.com
Cc: rostedt@goodmis.org
Cc: mathieu.desnoyers@efficios.com
Cc: jdesfossez@efficios.com
Cc: dvhart@infradead.org
Cc: bristot@redhat.com
Link: http://lkml.kernel.org/r/20170322104152.062785528@infradead.org
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
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 (Intel) <peterz@infradead.org>
Cc: juri.lelli@arm.com
Cc: bigeasy@linutronix.de
Cc: xlpang@redhat.com
Cc: rostedt@goodmis.org
Cc: mathieu.desnoyers@efficios.com
Cc: jdesfossez@efficios.com
Cc: dvhart@infradead.org
Cc: bristot@redhat.com
Link: http://lkml.kernel.org/r/20170322104152.001659630@infradead.org
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
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 (Intel) <peterz@infradead.org>
Cc: juri.lelli@arm.com
Cc: bigeasy@linutronix.de
Cc: xlpang@redhat.com
Cc: rostedt@goodmis.org
Cc: mathieu.desnoyers@efficios.com
Cc: jdesfossez@efficios.com
Cc: dvhart@infradead.org
Cc: bristot@redhat.com
Link: http://lkml.kernel.org/r/20170322104151.850383690@infradead.org
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
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 4e5160766f ("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: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: dietmar.eggemann@arm.com
Cc: ying.huang@intel.com
Fixes: 4e5160766f ("sched/fair: Propagate asynchrous detach")
Link: http://lkml.kernel.org/r/1489758442-2877-1-git-send-email-vincent.guittot@linaro.org
[ Fixed typos, improved layout. ]
Signed-off-by: Ingo Molnar <mingo@kernel.org>
The way the schedutil governor uses the PELT metric causes it to
underestimate the CPU utilization in some cases.
That can be easily demonstrated by running kernel compilation on
a Sandy Bridge Intel processor, running turbostat in parallel with
it and looking at the values written to the MSR_IA32_PERF_CTL
register. Namely, the expected result would be that when all CPUs
were 100% busy, all of them would be requested to run in the maximum
P-state, but observation shows that this clearly isn't the case.
The CPUs run in the maximum P-state for a while and then are
requested to run slower and go back to the maximum P-state after
a while again. That causes the actual frequency of the processor to
visibly oscillate below the sustainable maximum in a jittery fashion
which clearly is not desirable.
That has been attributed to CPU utilization metric updates on task
migration that cause the total utilization value for the CPU to be
reduced by the utilization of the migrated task. If that happens,
the schedutil governor may see a CPU utilization reduction and will
attempt to reduce the CPU frequency accordingly right away. That
may be premature, though, for example if the system is generally
busy and there are other runnable tasks waiting to be run on that
CPU already.
This is unlikely to be an issue on systems where cpufreq policies are
shared between multiple CPUs, because in those cases the policy
utilization is computed as the maximum of the CPU utilization values
over the whole policy and if that turns out to be low, reducing the
frequency for the policy most likely is a good idea anyway. On
systems with one CPU per policy, however, it may affect performance
adversely and even lead to increased energy consumption in some cases.
On those systems it may be addressed by taking another utilization
metric into consideration, like whether or not the CPU whose
frequency is about to be reduced has been idle recently, because if
that's not the case, the CPU is likely to be busy in the near future
and its frequency should not be reduced.
To that end, use the counter of idle calls in the timekeeping code.
Namely, make the schedutil governor look at that counter for the
current CPU every time before its frequency is about to be reduced.
If the counter has not changed since the previous iteration of the
governor computations for that CPU, the CPU has been busy for all
that time and its frequency should not be decreased, so if the new
frequency would be lower than the one set previously, the governor
will skip the frequency update.
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>
Reviewed-by: Joel Fernandes <joelaf@google.com>
This patch adds hash of maps support (hashmap->bpf_map).
BPF_MAP_TYPE_HASH_OF_MAPS is added.
A map-in-map contains a pointer to another map and lets call
this pointer 'inner_map_ptr'.
Notes on deleting inner_map_ptr from a hash map:
1. For BPF_F_NO_PREALLOC map-in-map, when deleting
an inner_map_ptr, the htab_elem itself will go through
a rcu grace period and the inner_map_ptr resides
in the htab_elem.
2. For pre-allocated htab_elem (!BPF_F_NO_PREALLOC),
when deleting an inner_map_ptr, the htab_elem may
get reused immediately. This situation is similar
to the existing prealloc-ated use cases.
However, the bpf_map_fd_put_ptr() calls bpf_map_put() which calls
inner_map->ops->map_free(inner_map) which will go
through a rcu grace period (i.e. all bpf_map's map_free
currently goes through a rcu grace period). Hence,
the inner_map_ptr is still safe for the rcu reader side.
This patch also includes BPF_MAP_TYPE_HASH_OF_MAPS to the
check_map_prealloc() in the verifier. preallocation is a
must for BPF_PROG_TYPE_PERF_EVENT. Hence, even we don't expect
heavy updates to map-in-map, enforcing BPF_F_NO_PREALLOC for map-in-map
is impossible without disallowing BPF_PROG_TYPE_PERF_EVENT from using
map-in-map first.
Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch adds a few helper funcs to enable map-in-map
support (i.e. outer_map->inner_map). The first outer_map type
BPF_MAP_TYPE_ARRAY_OF_MAPS is also added in this patch.
The next patch will introduce a hash of maps type.
Any bpf map type can be acted as an inner_map. The exception
is BPF_MAP_TYPE_PROG_ARRAY because the extra level of
indirection makes it harder to verify the owner_prog_type
and owner_jited.
Multi-level map-in-map is not supported (i.e. map->map is ok
but not map->map->map).
When adding an inner_map to an outer_map, it currently checks the
map_type, key_size, value_size, map_flags, max_entries and ops.
The verifier also uses those map's properties to do static analysis.
map_flags is needed because we need to ensure BPF_PROG_TYPE_PERF_EVENT
is using a preallocated hashtab for the inner_hash also. ops and
max_entries are needed to generate inlined map-lookup instructions.
For simplicity reason, a simple '==' test is used for both map_flags
and max_entries. The equality of ops is implied by the equality of
map_type.
During outer_map creation time, an inner_map_fd is needed to create an
outer_map. However, the inner_map_fd's life time does not depend on the
outer_map. The inner_map_fd is merely used to initialize
the inner_map_meta of the outer_map.
Also, for the outer_map:
* It allows element update and delete from syscall
* It allows element lookup from bpf_prog
The above is similar to the current fd_array pattern.
Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
Fix in verifier:
For the same bpf_map_lookup_elem() instruction (i.e. "call 1"),
a broken case is "a different type of map could be used for the
same lookup instruction". For example, an array in one case and a
hashmap in another. We have to resort to the old dynamic call behavior
in this case. The fix is to check for collision on insn_aux->map_ptr.
If there is collision, don't inline the map lookup.
Please see the "do_reg_lookup()" in test_map_in_map_kern.c in the later
patch for how-to trigger the above case.
Simplifications on array_map_gen_lookup():
1. Calculate elem_size from map->value_size. It removes the
need for 'struct bpf_array' which makes the later map-in-map
implementation easier.
2. Remove the 'elem_size == 1' test
Fixes: 81ed18ab30 ("bpf: add helper inlining infra and optimize map_array lookup")
Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
In both kmalloc and prealloc mode the bpf_map_update_elem() is using
per-cpu extra_elems to do atomic update when the map is full.
There are two issues with it. The logic can be misused, since it allows
max_entries+num_cpus elements to be present in the map. And alloc_extra_elems()
at map creation time can fail percpu alloc for large map values with a warn:
WARNING: CPU: 3 PID: 2752 at ../mm/percpu.c:892 pcpu_alloc+0x119/0xa60
illegal size (32824) or align (8) for percpu allocation
The fixes for both of these issues are different for kmalloc and prealloc modes.
For prealloc mode allocate extra num_possible_cpus elements and store
their pointers into extra_elems array instead of actual elements.
Hence we can use these hidden(spare) elements not only when the map is full
but during bpf_map_update_elem() that replaces existing element too.
That also improves performance, since pcpu_freelist_pop/push is avoided.
Unfortunately this approach cannot be used for kmalloc mode which needs
to kfree elements after rcu grace period. Therefore switch it back to normal
kmalloc even when full and old element exists like it was prior to
commit 6c90598174 ("bpf: pre-allocate hash map elements").
Add tests to check for over max_entries and large map values.
Reported-by: Dave Jones <davej@codemonkey.org.uk>
Fixes: 6c90598174 ("bpf: pre-allocate hash map elements")
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Martin KaFai Lau <kafai@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
What started as a rather straightforward race condition reported by
Dmitry using the syzkaller fuzzer ended up revealing some major
problems with how the audit subsystem managed its netlink sockets and
its connection with the userspace audit daemon. Fixing this properly
had quite the cascading effect and what we are left with is this rather
large and complicated patch. My initial goal was to try and decompose
this patch into multiple smaller patches, but the way these changes
are intertwined makes it difficult to split these changes into
meaningful pieces that don't break or somehow make things worse for
the intermediate states.
The patch makes a number of changes, but the most significant are
highlighted below:
* The auditd tracking variables, e.g. audit_sock, are now gone and
replaced by a RCU/spin_lock protected variable auditd_conn which is
a structure containing all of the auditd tracking information.
* We no longer track the auditd sock directly, instead we track it
via the network namespace in which it resides and we use the audit
socket associated with that namespace. In spirit, this is what the
code was trying to do prior to this patch (at least I think that is
what the original authors intended), but it was done rather poorly
and added a layer of obfuscation that only masked the underlying
problems.
* Big backlog queue cleanup, again. In v4.10 we made some pretty big
changes to how the audit backlog queues work, here we haven't changed
the queue design so much as cleaned up the implementation. Brought
about by the locking changes, we've simplified kauditd_thread() quite
a bit by consolidating the queue handling into a new helper function,
kauditd_send_queue(), which allows us to eliminate a lot of very
similar code and makes the looping logic in kauditd_thread() clearer.
* All netlink messages sent to auditd are now sent via
auditd_send_unicast_skb(). Other than just making sense, this makes
the lock handling easier.
* Change the audit_log_start() sleep behavior so that we never sleep
on auditd events (unchanged) or if the caller is holding the
audit_cmd_mutex (changed). Previously we didn't sleep if the caller
was auditd or if the message type fell between a certain range; the
type check was a poor effort of doing what the cmd_mutex check now
does. Richard Guy Briggs originally proposed not sleeping the
cmd_mutex owner several years ago but his patch wasn't acceptable
at the time. At least the idea lives on here.
* A problem with the lost record counter has been resolved. Steve
Grubb and I both happened to notice this problem and according to
some quick testing by Steve, this problem goes back quite some time.
It's largely a harmless problem, although it may have left some
careful sysadmins quite puzzled.
Cc: <stable@vger.kernel.org> # 4.10.x-
Reported-by: Dmitry Vyukov <dvyukov@google.com>
Signed-off-by: Paul Moore <paul@paul-moore.com>
sugov_start() only initializes struct sugov_cpu per-CPU structures
for shared policies, but it should do that for single-CPU policies too.
That in particular makes the IO-wait boost mechanism work in the
cases when cpufreq policies correspond to individual CPUs.
Fixes: 21ca6d2c52 (cpufreq: schedutil: Add iowait boosting)
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Acked-by: Viresh Kumar <viresh.kumar@linaro.org>
Cc: 4.9+ <stable@vger.kernel.org> # 4.9+
Pull CPU hotplug fix from Thomas Gleixner:
"A single fix preventing the concurrent execution of the CPU hotplug
callback install/invocation machinery. Long standing bug caused by a
massive brain slip of that Gleixner dude, which went unnoticed for
almost a year"
* 'smp-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
cpu/hotplug: Serialize callback invocations proper
Pull perf fixes from Thomas Gleixner:
"A set of perf related fixes:
- fix a CR4.PCE propagation issue caused by usage of mm instead of
active_mm and therefore propagated the wrong value.
- perf core fixes, which plug a use-after-free issue and make the
event inheritance on fork more robust.
- a tooling fix for symbol handling"
* 'perf-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
perf symbols: Fix symbols__fixup_end heuristic for corner cases
x86/perf: Clarify why x86_pmu_event_mapped() isn't racy
x86/perf: Fix CR4.PCE propagation to use active_mm instead of mm
perf/core: Better explain the inherit magic
perf/core: Simplify perf_event_free_task()
perf/core: Fix event inheritance on fork()
perf/core: Fix use-after-free in perf_release()
Pull scheduler fixes from Thomas Gleixner:
"From the scheduler departement:
- a bunch of sched deadline related fixes which deal with various
buglets and corner cases.
- two fixes for the loadavg spikes which are caused by the delayed
NOHZ accounting"
* 'sched-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
sched/deadline: Use deadline instead of period when calculating overflow
sched/deadline: Throttle a constrained deadline task activated after the deadline
sched/deadline: Make sure the replenishment timer fires in the next period
sched/loadavg: Use {READ,WRITE}_ONCE() for sample window
sched/loadavg: Avoid loadavg spikes caused by delayed NO_HZ accounting
sched/deadline: Add missing update_rq_clock() in dl_task_timer()
Pull locking fixes from Thomas Gleixner:
"Three fixes related to locking:
- fix a SIGKILL issue for RWSEM_GENERIC_SPINLOCK which has been fixed
for the XCHGADD variant already
- plug a potential use after free in the futex code
- prevent leaking a held spinlock in an futex error handling code
path"
* 'locking-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
locking/rwsem: Fix down_write_killable() for CONFIG_RWSEM_GENERIC_SPINLOCK=y
futex: Add missing error handling to FUTEX_REQUEUE_PI
futex: Fix potential use-after-free in FUTEX_REQUEUE_PI
Creation of a kthread goes through a couple interlocked stages between
the kthread itself and its creator. Once the new kthread starts
running, it initializes itself and wakes up the creator. The creator
then can further configure the kthread and then let it start doing its
job by waking it up.
In this configuration-by-creator stage, the creator is the only one
that can wake it up but the kthread is visible to userland. When
altering the kthread's attributes from userland is allowed, this is
fine; however, for cases where CPU affinity is critical,
kthread_bind() is used to first disable affinity changes from userland
and then set the affinity. This also prevents the kthread from being
migrated into non-root cgroups as that can affect the CPU affinity and
many other things.
Unfortunately, the cgroup side of protection is racy. While the
PF_NO_SETAFFINITY flag prevents further migrations, userland can win
the race before the creator sets the flag with kthread_bind() and put
the kthread in a non-root cgroup, which can lead to all sorts of
problems including incorrect CPU affinity and starvation.
This bug got triggered by userland which periodically tries to migrate
all processes in the root cpuset cgroup to a non-root one. Per-cpu
workqueue workers got caught while being created and ended up with
incorrected CPU affinity breaking concurrency management and sometimes
stalling workqueue execution.
This patch adds task->no_cgroup_migration which disallows the task to
be migrated by userland. kthreadd starts with the flag set making
every child kthread start in the root cgroup with migration
disallowed. The flag is cleared after the kthread finishes
initialization by which time PF_NO_SETAFFINITY is set if the kthread
should stay in the root cgroup.
It'd be better to wait for the initialization instead of failing but I
couldn't think of a way of implementing that without adding either a
new PF flag, or sleeping and retrying from waiting side. Even if
userland depends on changing cgroup membership of a kthread, it either
has to be synchronized with kthread_create() or periodically repeat,
so it's unlikely that this would break anything.
v2: Switch to a simpler implementation using a new task_struct bit
field suggested by Oleg.
Signed-off-by: Tejun Heo <tj@kernel.org>
Suggested-by: Oleg Nesterov <oleg@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Reported-and-debugged-by: Chris Mason <clm@fb.com>
Cc: stable@vger.kernel.org # v4.3+ (we can't close the race on < v4.3)
Signed-off-by: Tejun Heo <tj@kernel.org>
Optimize:
bpf_call
bpf_map_lookup_elem
map->ops->map_lookup_elem
htab_map_lookup_elem
__htab_map_lookup_elem
into:
bpf_call
__htab_map_lookup_elem
to improve performance of JITed programs.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
Optimize bpf_call -> bpf_map_lookup_elem() -> array_map_lookup_elem()
into a sequence of bpf instructions.
When JIT is on the sequence of bpf instructions is the sequence
of native cpu instructions with significantly faster performance
than indirect call and two function's prologue/epilogue.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
