The fastpath in __synchronize_srcu() is designed to handle cases where
there are a large number of concurrent calls for the same srcu_struct
structure. However, the Linux kernel currently does not use SRCU in
this manner, so remove the fastpath checks for simplicity.
Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
The current implementation of synchronize_srcu_expedited() can cause
severe OS jitter due to its use of synchronize_sched(), which in turn
invokes try_stop_cpus(), which causes each CPU to be sent an IPI.
This can result in severe performance degradation for real-time workloads
and especially for short-interation-length HPC workloads. Furthermore,
because only one instance of try_stop_cpus() can be making forward progress
at a given time, only one instance of synchronize_srcu_expedited() can
make forward progress at a time, even if they are all operating on
distinct srcu_struct structures.
This commit, inspired by an earlier implementation by Peter Zijlstra
(https://lkml.org/lkml/2012/1/31/211) and by further offline discussions,
takes a strictly algorithmic bits-in-memory approach. This has the
disadvantage of requiring one explicit memory-barrier instruction in
each of srcu_read_lock() and srcu_read_unlock(), but on the other hand
completely dispenses with OS jitter and furthermore allows SRCU to be
used freely by CPUs that RCU believes to be idle or offline.
The update-side implementation handles the single read-side memory
barrier by rechecking the per-CPU counters after summing them and
by running through the update-side state machine twice.
This implementation has passed moderate rcutorture testing on both
x86 and Power. Also updated to use this_cpu_ptr() instead of per_cpu_ptr(),
as suggested by Peter Zijlstra.
Reported-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Reviewed-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Although rcutorture does invoke rcu_barrier() and friends, it cannot
really be called a torture test given that it invokes them only once
at the end of the test. This commit therefore introduces heavy-duty
rcutorture testing for rcu_barrier(), which may be carried out
concurrently with normal rcutorture testing.
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Pull power management fixes from Rafael J. Wysocki:
"Fix for an issue causing hibernation to hang on systems with highmem
(that practically means i386) due to broken memory management (bug
introduced in 3.2, so -stable material) and PM documentation update
making the freezer documentation follow the code again after some
recent updates."
* tag 'pm-for-3.4-rc5' of git://git.kernel.org/pub/scm/linux/kernel/git/rafael/linux-pm:
PM / Freezer / Docs: Update documentation about freezing of tasks
PM / Hibernate: fix the number of pages used for hibernate/thaw buffering
Pull perf fixes from Ingo Molnar.
* 'perf-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
perf: Fix perf_event_for_each() to use sibling
perf symbols: Read plt symbols from proper symtab_type binary
tracing: Fix stacktrace of latency tracers (irqsoff and friends)
perf tools: Add 'G' and 'H' modifiers to event parsing
tracing: Fix regression with tracing_on
perf tools: Drop CROSS_COMPILE from flex and bison calls
perf report: Fix crash showing warning related to kernel maps
tracing: Fix build breakage without CONFIG_PERF_EVENTS (again)
Pull build fixes for less mainstream architectures from Paul Gortmaker:
"These are fixes for frv(1), blackfin(2), powerpc(1) and xtensa(4).
Fortunately the touches are nearly all specific to files just used by
the arch in question. The two touches to shared/common files
[kernel/irq/debug.h and drivers/pci/Makefile] are trivial to assess as
no risk to anyone.
Half of them relate to xtensa directly. It was only when I fixed the
last xtensa issue that I realized that the arch has been broken for a
significant time, and isn't a specific v3.4 regression. So if you
wanted, we could leave xtensa lying bleeding in the street for a
couple more weeks and queue those for 3.5. But given they are no risk
to anyone outside of xtensa, I figured to just leave them in.
If you are OK with taking the xtensa fixes, then please pull to get:
- one last implicit include uncovered by system.h that is in a file
specific to just one powerpc defconfig. (I'd sync'd with BenH).
- fix an oversight in the PCI makefile where shared code wasn't being
compiled for ARCH=frv
- fix a missing include for GPIO in blackfin framebuffer.
- audit and tag endif in blackfin ezkit board file, in order to find
and fix the misplaced endif masking a block of code.
- fix irq/debug.h choice of temporary macro names to be more internal
so they don't conflict with names used by xtensa.
- fix a reference to an undeclared local var in xtensa's signal.c
- fix an implicit bug.h usage in xtensa's asm/io.h uncovered by my
removing bug.h from kernel.h
- fix xtensa to properly indicate it is using asm-generic/hardirq.h
in order to resolve the link error - undefined ack_bad_irq
The xtensa still fails final link as my latest binutils does something
evil when ld forward-relocates unlikely() blocks, but in theory people
who have older/valid toolchains could now use the thing."
* 'for-v3.4-rc5' of git://git.kernel.org/pub/scm/linux/kernel/git/paulg/linux:
xtensa: fix build fail on undefined ack_bad_irq
blackfin: fix ifdef fustercluck in mach-bf538/boards/ezkit.c
blackfin: fix compile error in bfin-lq035q1-fb.c
pci: frv architecture needs generic setup-bus infrastructure
irq: hide debug macros so they don't collide with others.
xtensa: fix build error in xtensa/include/asm/io.h
xtensa: fix build failure in xtensa/kernel/signal.c
powerpc: fix system.h fallout in sysdev/scom.c [chroma_defconfig]
Updating max_usage is something one would expect when we reach
a new maximum usage value even when we do this by forcing through
the limit with res_counter_charge_nofail().
(Whether we want to account failcnt when we force through the limit
is another debate).
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Acked-by: Glauber Costa <glommer@parallels.com>
Acked-by: Kirill A. Shutemov <kirill@shutemov.name>
Cc: Li Zefan <lizefan@huawei.com>
These two functions do almost the same thing and duplicate some code.
Merge their implementation into a single common function.
res_counter_charge_locked() takes one more parameter but it doesn't seem
to be used outside res_counter.c yet anyway.
There is no (intended) change in the behaviour.
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Acked-by: Glauber Costa <glommer@parallels.com>
Acked-by: Kirill A. Shutemov <kirill@shutemov.name>
Cc: Li Zefan <lizefan@huawei.com>
The mod_timer_pinned() header comment states that it prevents timers
from being migrated to a different CPU. This is not the case, instead,
it ensures that the timer is posted to the current CPU, but does nothing
to prevent CPU-hotplug operations from migrating the timer.
This commit therefore brings the comment header into alignment with
reality.
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Acked-by: Steven Rostedt <rostedt@goodmis.org>
RCU_FAST_NO_HZ uses a timer to limit the time that a CPU with callbacks
can remain in dyntick-idle mode. This timer is cancelled when the CPU
exits idle, and therefore should never fire. However, if the timer
were migrated to some other CPU for whatever reason (1) the timer could
actually fire and (2) firing on some other CPU would fail to wake up the
CPU with callbacks, possibly resulting in sluggishness or a system hang.
This commit therfore adds a WARN_ON_ONCE() to the timer handler in order
to detect this condition.
Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
In perf_event_for_each() we call a function on an event, and then
iterate over the siblings of the event.
However we don't call the function on the siblings, we call it
repeatedly on the original event - it seems "obvious" that we should
be calling it with sibling as the argument.
It looks like this broke in commit 75f937f24b ("Fix ctx->mutex
vs counter->mutex inversion").
The only effect of the bug is that the PERF_IOC_FLAG_GROUP parameter
to the ioctls doesn't work.
Signed-off-by: Michael Ellerman <michael@ellerman.id.au>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/1334109253-31329-1-git-send-email-michael@ellerman.id.au
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Under extreme memory used up situations, percpu allocation
might fail. We hit it when system goes to suspend-to-ram,
causing a kworker panic:
EIP: [<c124411a>] build_sched_domains+0x23a/0xad0
Kernel panic - not syncing: Fatal exception
Pid: 3026, comm: kworker/u:3
3.0.8-137473-gf42fbef #1
Call Trace:
[<c18cc4f2>] panic+0x66/0x16c
[...]
[<c1244c37>] partition_sched_domains+0x287/0x4b0
[<c12a77be>] cpuset_update_active_cpus+0x1fe/0x210
[<c123712d>] cpuset_cpu_inactive+0x1d/0x30
[...]
With this fix applied build_sched_domains() will return -ENOMEM and
the suspend attempt fails.
Signed-off-by: he, bo <bo.he@intel.com>
Reviewed-by: Zhang, Yanmin <yanmin.zhang@intel.com>
Reviewed-by: Srivatsa S. Bhat <srivatsa.bhat@linux.vnet.ibm.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: <stable@kernel.org>
Link: http://lkml.kernel.org/r/1335355161.5892.17.camel@hebo
[ So, we fail to deallocate a CPU because we cannot allocate RAM :-/
I don't like that kind of sad behavior but nevertheless it should
not crash under high memory load. ]
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Commits 367456c756 ("sched: Ditch per cgroup task lists for
load-balancing") and 5d6523ebd ("sched: Fix load-balance wreckage")
left some more wreckage.
By setting loop_max unconditionally to ->nr_running load-balancing
could take a lot of time on very long runqueues (hackbench!). So keep
the sysctl as max limit of the amount of tasks we'll iterate.
Furthermore, the min load filter for migration completely fails with
cgroups since inequality in per-cpu state can easily lead to such
small loads :/
Furthermore the change to add new tasks to the tail of the queue
instead of the head seems to have some effect.. not quite sure I
understand why.
Combined these fixes solve the huge hackbench regression reported by
Tim when hackbench is ran in a cgroup.
Reported-by: Tim Chen <tim.c.chen@linux.intel.com>
Acked-by: Tim Chen <tim.c.chen@linux.intel.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Link: http://lkml.kernel.org/r/1335365763.28150.267.camel@twins
[ got rid of the CONFIG_PREEMPT tuning and made small readability edits ]
Signed-off-by: Ingo Molnar <mingo@kernel.org>
- Convert the old uid mapping functions into compatibility wrappers
- Add a uid/gid mapping layer from user space uid and gids to kernel
internal uids and gids that is extent based for simplicty and speed.
* Working with number space after mapping uids/gids into their kernel
internal version adds only mapping complexity over what we have today,
leaving the kernel code easy to understand and test.
- Add proc files /proc/self/uid_map /proc/self/gid_map
These files display the mapping and allow a mapping to be added
if a mapping does not exist.
- Allow entering the user namespace without a uid or gid mapping.
Since we are starting with an existing user our uids and gids
still have global mappings so are still valid and useful they just don't
have local mappings. The requirement for things to work are global uid
and gid so it is odd but perfectly fine not to have a local uid
and gid mapping.
Not requiring global uid and gid mappings greatly simplifies
the logic of setting up the uid and gid mappings by allowing
the mappings to be set after the namespace is created which makes the
slight weirdness worth it.
- Make the mappings in the initial user namespace to the global
uid/gid space explicit. Today it is an identity mapping
but in the future we may want to twist this for debugging, similar
to what we do with jiffies.
- Document the memory ordering requirements of setting the uid and
gid mappings. We only allow the mappings to be set once
and there are no pointers involved so the requirments are
trivial but a little atypical.
Performance:
In this scheme for the permission checks the performance is expected to
stay the same as the actuall machine instructions should remain the same.
The worst case I could think of is ls -l on a large directory where
all of the stat results need to be translated with from kuids and
kgids to uids and gids. So I benchmarked that case on my laptop
with a dual core hyperthread Intel i5-2520M cpu with 3M of cpu cache.
My benchmark consisted of going to single user mode where nothing else
was running. On an ext4 filesystem opening 1,000,000 files and looping
through all of the files 1000 times and calling fstat on the
individuals files. This was to ensure I was benchmarking stat times
where the inodes were in the kernels cache, but the inode values were
not in the processors cache. My results:
v3.4-rc1: ~= 156ns (unmodified v3.4-rc1 with user namespace support disabled)
v3.4-rc1-userns-: ~= 155ns (v3.4-rc1 with my user namespace patches and user namespace support disabled)
v3.4-rc1-userns+: ~= 164ns (v3.4-rc1 with my user namespace patches and user namespace support enabled)
All of the configurations ran in roughly 120ns when I performed tests
that ran in the cpu cache.
So in summary the performance impact is:
1ns improvement in the worst case with user namespace support compiled out.
8ns aka 5% slowdown in the worst case with user namespace support compiled in.
Acked-by: Serge Hallyn <serge.hallyn@canonical.com>
Signed-off-by: Eric W. Biederman <ebiederm@xmission.com>
- Transform userns->creator from a user_struct reference to a simple
kuid_t, kgid_t pair.
In cap_capable this allows the check to see if we are the creator of
a namespace to become the classic suser style euid permission check.
This allows us to remove the need for a struct cred in the mapping
functions and still be able to dispaly the user namespace creators
uid and gid as 0.
- Remove the now unnecessary delayed_work in free_user_ns.
All that is left for free_user_ns to do is to call kmem_cache_free
and put_user_ns. Those functions can be called in any context
so call them directly from free_user_ns removing the need for delayed work.
Acked-by: Serge Hallyn <serge.hallyn@canonical.com>
Signed-off-by: Eric W. Biederman <ebiederm@xmission.com>
The rcutorture initialization code ignored the error returns from
rcu_torture_onoff_init() and rcu_torture_stall_init(). The rcutorture
cleanup code failed to NULL out a number of pointers. These bugs will
normally have no effect, but this commit fixes them nevertheless.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
