Syscall User Dispatch (SUD) must take precedence over seccomp and
ptrace, since the use case is emulation (it can be invoked with a
different ABI) such that seccomp filtering by syscall number doesn't
make sense in the first place. In addition, either the syscall is
dispatched back to userspace, in which case there is no resource for to
trace, or the syscall will be executed, and seccomp/ptrace will execute
next.
Since SUD runs before tracepoints, it needs to be a SYSCALL_WORK_EXIT as
well, just to prevent a trace exit event when dispatch was triggered.
For that, the on_syscall_dispatch() examines context to skip the
tracepoint, audit and other work.
[ tglx: Add a comment on the exit side ]
Signed-off-by: Gabriel Krisman Bertazi <krisman@collabora.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Andy Lutomirski <luto@kernel.org>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Kees Cook <keescook@chromium.org>
Link: https://lore.kernel.org/r/20201127193238.821364-5-krisman@collabora.com
Introduce a mechanism to quickly disable/enable syscall handling for a
specific process and redirect to userspace via SIGSYS. This is useful
for processes with parts that require syscall redirection and parts that
don't, but who need to perform this boundary crossing really fast,
without paying the cost of a system call to reconfigure syscall handling
on each boundary transition. This is particularly important for Windows
games running over Wine.
The proposed interface looks like this:
prctl(PR_SET_SYSCALL_USER_DISPATCH, <op>, <off>, <length>, [selector])
The range [<offset>,<offset>+<length>) is a part of the process memory
map that is allowed to by-pass the redirection code and dispatch
syscalls directly, such that in fast paths a process doesn't need to
disable the trap nor the kernel has to check the selector. This is
essential to return from SIGSYS to a blocked area without triggering
another SIGSYS from rt_sigreturn.
selector is an optional pointer to a char-sized userspace memory region
that has a key switch for the mechanism. This key switch is set to
either PR_SYS_DISPATCH_ON, PR_SYS_DISPATCH_OFF to enable and disable the
redirection without calling the kernel.
The feature is meant to be set per-thread and it is disabled on
fork/clone/execv.
Internally, this doesn't add overhead to the syscall hot path, and it
requires very little per-architecture support. I avoided using seccomp,
even though it duplicates some functionality, due to previous feedback
that maybe it shouldn't mix with seccomp since it is not a security
mechanism. And obviously, this should never be considered a security
mechanism, since any part of the program can by-pass it by using the
syscall dispatcher.
For the sysinfo benchmark, which measures the overhead added to
executing a native syscall that doesn't require interception, the
overhead using only the direct dispatcher region to issue syscalls is
pretty much irrelevant. The overhead of using the selector goes around
40ns for a native (unredirected) syscall in my system, and it is (as
expected) dominated by the supervisor-mode user-address access. In
fact, with SMAP off, the overhead is consistently less than 5ns on my
test box.
Signed-off-by: Gabriel Krisman Bertazi <krisman@collabora.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Andy Lutomirski <luto@kernel.org>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Kees Cook <keescook@chromium.org>
Link: https://lore.kernel.org/r/20201127193238.821364-4-krisman@collabora.com
While debugging a situation where a delta for an event was calucalted wrong,
I realize there was nothing making sure that the delta of events are
correct. If a single event has an incorrect delta, then all events after it
will also have one. If the discrepency gets large enough, it could cause
the time stamps to go backwards when crossing sub buffers, that record a
full 64 bit time stamp, and the new deltas are added to that.
Add a way to validate the events at most events and when crossing a buffer
page. This will help make sure that the deltas are always correct. This test
will detect if they are ever corrupted.
The test adds a high overhead to the ring buffer recording, as it does the
audit for almost every event, and should only be used for testing the ring
buffer.
This will catch the bug that is fixed by commit 55ea4cf403 ("ring-buffer:
Update write stamp with the correct ts"), which is not applied when this
commit is applied.
Signed-off-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
Pull tracing fixes from Steven Rostedt:
- Use correct timestamp variable for ring buffer write stamp update
- Fix up before stamp and write stamp when crossing ring buffer sub
buffers
- Keep a zero delta in ring buffer in slow path if cmpxchg fails
- Fix trace_printk static buffer for archs that care
- Fix ftrace record accounting for ftrace ops with trampolines
- Fix DYNAMIC_FTRACE_WITH_DIRECT_CALLS dependency
- Remove WARN_ON in hwlat tracer that triggers on something that is OK
- Make "my_tramp" trampoline in ftrace direct sample code global
- Fixes in the bootconfig tool for better alignment management
* tag 'trace-v5.10-rc6' of git://git.kernel.org/pub/scm/linux/kernel/git/rostedt/linux-trace:
ring-buffer: Always check to put back before stamp when crossing pages
ftrace: Fix DYNAMIC_FTRACE_WITH_DIRECT_CALLS dependency
ftrace: Fix updating FTRACE_FL_TRAMP
tracing: Fix alignment of static buffer
tracing: Remove WARN_ON in start_thread()
samples/ftrace: Mark my_tramp[12]? global
ring-buffer: Set the right timestamp in the slow path of __rb_reserve_next()
ring-buffer: Update write stamp with the correct ts
docs: bootconfig: Update file format on initrd image
tools/bootconfig: Align the bootconfig applied initrd image size to 4
tools/bootconfig: Fix to check the write failure correctly
tools/bootconfig: Fix errno reference after printf()
Instead of having two structures that represent each block device with
different life time rules, merge them into a single one. This also
greatly simplifies the reference counting rules, as we can use the inode
reference count as the main reference count for the new struct
block_device, with the device model reference front ending it for device
model interaction.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Jan Kara <jack@suse.cz>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Move the start_sect field to struct block_device in preparation
of killing struct hd_struct.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Jan Kara <jack@suse.cz>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Now that the hd_struct always has a block device attached to it, there is
no need for having two size field that just get out of sync.
Additionally the field in hd_struct did not use proper serialization,
possibly allowing for torn writes. By only using the block_device field
this problem also gets fixed.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Reviewed-by: Jan Kara <jack@suse.cz>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Acked-by: Coly Li <colyli@suse.de> [bcache]
Acked-by: Chao Yu <yuchao0@huawei.com> [f2fs]
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The kernel currently uses kmem_cache to allocate shadow call stacks,
which means an overflows may not be immediately detected and can
potentially result in another task's shadow stack to be overwritten.
This change switches SCS to use virtually mapped shadow stacks for
tasks, which increases shadow stack size to a full page and provides
more robust overflow detection, similarly to VMAP_STACK.
Signed-off-by: Sami Tolvanen <samitolvanen@google.com>
Acked-by: Will Deacon <will@kernel.org>
Link: https://lore.kernel.org/r/20201130233442.2562064-2-samitolvanen@google.com
Signed-off-by: Will Deacon <will@kernel.org>
The current ring buffer logic checks to see if the updating of the event
buffer was interrupted, and if it is, it will try to fix up the before stamp
with the write stamp to make them equal again. This logic is flawed, because
if it is not interrupted, the two are guaranteed to be different, as the
current event just updated the before stamp before allocation. This
guarantees that the next event (this one or another interrupting one) will
think it interrupted the time updates of a previous event and inject an
absolute time stamp to compensate.
The correct logic is to always update the timestamps when traversing to a
new sub buffer.
Cc: stable@vger.kernel.org
Fixes: a389d86f7f ("ring-buffer: Have nested events still record running time stamp")
Signed-off-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
On powerpc, kprobe-direct.tc triggered FTRACE_WARN_ON() in
ftrace_get_addr_new() followed by the below message:
Bad trampoline accounting at: 000000004222522f (wake_up_process+0xc/0x20) (f0000001)
The set of steps leading to this involved:
- modprobe ftrace-direct-too
- enable_probe
- modprobe ftrace-direct
- rmmod ftrace-direct <-- trigger
The problem turned out to be that we were not updating flags in the
ftrace record properly. From the above message about the trampoline
accounting being bad, it can be seen that the ftrace record still has
FTRACE_FL_TRAMP set though ftrace-direct module is going away. This
happens because we are checking if any ftrace_ops has the
FTRACE_FL_TRAMP flag set _before_ updating the filter hash.
The fix for this is to look for any _other_ ftrace_ops that also needs
FTRACE_FL_TRAMP.
Link: https://lkml.kernel.org/r/56c113aa9c3e10c19144a36d9684c7882bf09af5.1606412433.git.naveen.n.rao@linux.vnet.ibm.com
Cc: stable@vger.kernel.org
Fixes: a124692b69 ("ftrace: Enable trampoline when rec count returns back to one")
Signed-off-by: Naveen N. Rao <naveen.n.rao@linux.vnet.ibm.com>
Signed-off-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
With 5.9 kernel on ARM64, I found ftrace_dump output was broken but
it had no problem with normal output "cat /sys/kernel/debug/tracing/trace".
With investigation, it seems coping the data into temporal buffer seems to
break the align binary printf expects if the static buffer is not aligned
with 4-byte. IIUC, get_arg in bstr_printf expects that args has already
right align to be decoded and seq_buf_bprintf says ``the arguments are saved
in a 32bit word array that is defined by the format string constraints``.
So if we don't keep the align under copy to temporal buffer, the output
will be broken by shifting some bytes.
This patch fixes it.
Link: https://lkml.kernel.org/r/20201125225654.1618966-1-minchan@kernel.org
Cc: <stable@vger.kernel.org>
Fixes: 8e99cf91b9 ("tracing: Do not allocate buffer in trace_find_next_entry() in atomic")
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Signed-off-by: Minchan Kim <minchan@kernel.org>
Signed-off-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
This patch reverts commit 978defee11 ("tracing: Do a WARN_ON()
if start_thread() in hwlat is called when thread exists")
.start hook can be legally called several times if according
tracer is stopped
screen window 1
[root@localhost ~]# echo 1 > /sys/kernel/tracing/events/kmem/kfree/enable
[root@localhost ~]# echo 1 > /sys/kernel/tracing/options/pause-on-trace
[root@localhost ~]# less -F /sys/kernel/tracing/trace
screen window 2
[root@localhost ~]# cat /sys/kernel/debug/tracing/tracing_on
0
[root@localhost ~]# echo hwlat > /sys/kernel/debug/tracing/current_tracer
[root@localhost ~]# echo 1 > /sys/kernel/debug/tracing/tracing_on
[root@localhost ~]# cat /sys/kernel/debug/tracing/tracing_on
0
[root@localhost ~]# echo 2 > /sys/kernel/debug/tracing/tracing_on
triggers warning in dmesg:
WARNING: CPU: 3 PID: 1403 at kernel/trace/trace_hwlat.c:371 hwlat_tracer_start+0xc9/0xd0
Link: https://lkml.kernel.org/r/bd4d3e70-400d-9c82-7b73-a2d695e86b58@virtuozzo.com
Cc: Ingo Molnar <mingo@redhat.com>
Cc: stable@vger.kernel.org
Fixes: 978defee11 ("tracing: Do a WARN_ON() if start_thread() in hwlat is called when thread exists")
Signed-off-by: Vasily Averin <vvs@virtuozzo.com>
Signed-off-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
When an interrupt allocation fails for N interrupts, it is pretty
common for the error handling code to free the same number of interrupts,
no matter how many interrupts have actually been allocated.
This may result in the domain freeing code to be unexpectedly called
for interrupts that have no mapping in that domain. Things end pretty
badly.
Instead, add some checks to irq_domain_free_irqs_hierarchy() to make sure
that thiss does not follow the hierarchy if no mapping exists for a given
interrupt.
Fixes: 6a6544e520 ("genirq/irqdomain: Remove auto-recursive hierarchy support")
Signed-off-by: Marc Zyngier <maz@kernel.org>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Link: https://lore.kernel.org/r/20201129135551.396777-1-maz@kernel.org
There is currently no way to convey the affinity of an interrupt
via irq_create_mapping(), which creates issues for devices that
expect that affinity to be managed by the kernel.
In order to sort this out, rename irq_create_mapping() to
irq_create_mapping_affinity() with an additional affinity parameter that
can be passed down to irq_domain_alloc_descs().
irq_create_mapping() is re-implemented as a wrapper around
irq_create_mapping_affinity().
No functional change.
Fixes: e75eafb9b0 ("genirq/msi: Switch to new irq spreading infrastructure")
Signed-off-by: Laurent Vivier <lvivier@redhat.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Greg Kurz <groug@kaod.org>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: stable@vger.kernel.org
Link: https://lore.kernel.org/r/20201126082852.1178497-2-lvivier@redhat.com
Pull locking fixes from Thomas Gleixner:
"Two more places which invoke tracing from RCU disabled regions in the
idle path.
Similar to the entry path the low level idle functions have to be
non-instrumentable"
* tag 'locking-urgent-2020-11-29' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
intel_idle: Fix intel_idle() vs tracing
sched/idle: Fix arch_cpu_idle() vs tracing
Trivial conflict in CAN, keep the net-next + the byteswap wrapper.
Conflicts:
drivers/net/can/usb/gs_usb.c
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The SCMI pull request for the arm/drivers branch requires v5.10-rc2
because of dependencies with other git trees, so merge that in here.
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Pull printk fixes from Petr Mladek:
- do not lose trailing newline in pr_cont() calls
- two trivial fixes for a dead store and a config description
* tag 'printk-for-5.10-rc6-fixup' of git://git.kernel.org/pub/scm/linux/kernel/git/printk/linux:
printk: finalize records with trailing newlines
printk: remove unneeded dead-store assignment
init/Kconfig: Fix CPU number in LOG_CPU_MAX_BUF_SHIFT description
Nowadays, there are increasing requirements to benchmark the performance
of dma_map and dma_unmap particually while the device is attached to an
IOMMU.
This patch enables the support. Users can run specified number of threads
to do dma_map_page and dma_unmap_page on a specific NUMA node with the
specified duration. Then dma_map_benchmark will calculate the average
latency for map and unmap.
A difficulity for this benchmark is that dma_map/unmap APIs must run on
a particular device. Each device might have different backend of IOMMU or
non-IOMMU.
So we use the driver_override to bind dma_map_benchmark to a particual
device by:
For platform devices:
echo dma_map_benchmark > /sys/bus/platform/devices/xxx/driver_override
echo xxx > /sys/bus/platform/drivers/xxx/unbind
echo xxx > /sys/bus/platform/drivers/dma_map_benchmark/bind
For PCI devices:
echo dma_map_benchmark > /sys/bus/pci/devices/0000:00:01.0/driver_override
echo 0000:00:01.0 > /sys/bus/pci/drivers/xxx/unbind
echo 0000:00:01.0 > /sys/bus/pci/drivers/dma_map_benchmark/bind
Cc: Will Deacon <will@kernel.org>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Marek Szyprowski <m.szyprowski@samsung.com>
Cc: Robin Murphy <robin.murphy@arm.com>
Signed-off-by: Barry Song <song.bao.hua@hisilicon.com>
[hch: folded in two fixes from Colin Ian King <colin.king@canonical.com>]
Signed-off-by: Christoph Hellwig <hch@lst.de>
When calling debugfs functions, there is no need to ever check the
return value. The function can work or not, but the code logic should
never do something different based on this.
Signed-off-by: Tiezhu Yang <yangtiezhu@loongson.cn>
Reviewed-by: Robin Murphy <robin.murphy@arm.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
At the moment we allow bypassing DMA ops only when we can do this for
the entire RAM. However there are configs with mixed type memory
where we could still allow bypassing IOMMU in most cases;
POWERPC with persistent memory is one example.
This adds an arch hook to determine where bypass can still work and
we invoke direct DMA API. The following patch checks the bus limit
on POWERPC to allow or disallow direct mapping.
This adds a ARCH_HAS_DMA_MAP_DIRECT config option to make the arch_xxxx
hooks no-op by default.
Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Provide a wrapper function to get the IMA hash of an inode. This helper
is useful in fingerprinting files (e.g executables on execution) and
using these fingerprints in detections like an executable unlinking
itself.
Since the ima_inode_hash can sleep, it's only allowed for sleepable
LSM hooks.
Signed-off-by: KP Singh <kpsingh@google.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/bpf/20201124151210.1081188-3-kpsingh@chromium.org
we have supplied the inline function: of_cft() in cgroup.h.
So replace the direct use 'of->kn->priv' with inline func
of_cft(), which is more readable.
Signed-off-by: Hui Su <sh_def@163.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
In realtime scenario, We do not want to have interference on the
isolated cpu cores. but when invoking alloc_workqueue() for percpu wq
on the housekeeping cpu, it kick a kworker on the isolated cpu.
alloc_workqueue
pwq_adjust_max_active
wake_up_worker
The comment in pwq_adjust_max_active() said:
"Need to kick a worker after thawed or an unbound wq's
max_active is bumped"
So it is unnecessary to kick a kworker for percpu's wq when invoking
alloc_workqueue(). this patch only kick a worker based on the actual
activation of delayed works.
Signed-off-by: Yunfeng Ye <yeyunfeng@huawei.com>
Reviewed-by: Lai Jiangshan <jiangshanlai@gmail.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Instead of using the array-of-pointers trick to avoid having gcc mess up
the built-in module-version array stride, specify type alignment when
declaring entries to prevent gcc from increasing alignment.
This is essentially an alternative (one-line) fix to the problem
addressed by commit b4bc842802 ("module: deal with alignment issues in
built-in module versions").
gcc can increase the alignment of larger objects with static extent as
an optimisation, but this can be suppressed by using the aligned
attribute when declaring variables.
Note that we have been relying on this behaviour for kernel parameters
for 16 years and it indeed hasn't changed since the introduction of the
aligned attribute in gcc-3.1.
Link: https://lore.kernel.org/lkml/20201103175711.10731-1-johan@kernel.org
Signed-off-by: Johan Hovold <johan@kernel.org>
Signed-off-by: Jessica Yu <jeyu@kernel.org>
The current implementation uses a number of gotos to implement a loop
and different paths within the loop, which makes the code less readable
than it would be with an explicit while-loop. This patch also replaces a
chain of if/if-elses keyed on the same expression with a switch
statement.
No change in behaviour is intended.
Signed-off-by: Wedson Almeida Filho <wedsonaf@google.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20201121015509.3594191-1-wedsonaf@google.com
Some unused macros could cause gcc warning:
kernel/audit.c:68:0: warning: macro "AUDIT_UNINITIALIZED" is not used
[-Wunused-macros]
kernel/auditsc.c:104:0: warning: macro "AUDIT_AUX_IPCPERM" is not used
[-Wunused-macros]
kernel/auditsc.c:82:0: warning: macro "AUDITSC_INVALID" is not used
[-Wunused-macros]
AUDIT_UNINITIALIZED and AUDITSC_INVALID are still meaningful and should
be in incorporated.
Just remove AUDIT_AUX_IPCPERM.
Thanks comments from Richard Guy Briggs and Paul Moore.
Signed-off-by: Alex Shi <alex.shi@linux.alibaba.com>
Cc: Paul Moore <paul@paul-moore.com>
Cc: Richard Guy Briggs <rgb@redhat.com>
Cc: Eric Paris <eparis@redhat.com>
Cc: linux-audit@redhat.com
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Paul Moore <paul@paul-moore.com>
Trade one atomic op for a full memory barrier.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Frederic Weisbecker <frederic@kernel.org>
Get rid of the __call_single_node union and cleanup the API a little
to avoid external code relying on the structure layout as much.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Frederic Weisbecker <frederic@kernel.org>
Get rid of the __call_single_node union and clean up the API a little
to avoid external code relying on the structure layout as much.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Frederic Weisbecker <frederic@kernel.org>
At fork time currently, a local node can be allowed to fill completely
and allow the periodic load balancer to fix the problem. This can be
problematic in cases where a task creates lots of threads that idle until
woken as part of a worker poll causing a memory bandwidth problem.
However, a "real" workload suffers badly from this behaviour. The workload
in question is mostly NUMA aware but spawns large numbers of threads
that act as a worker pool that can be called from anywhere. These need
to spread early to get reasonable behaviour.
This patch limits how much a local node can fill before spilling over
to another node and it will not be a universal win. Specifically,
very short-lived workloads that fit within a NUMA node would prefer
the memory bandwidth.
As I cannot describe the "real" workload, the best proxy measure I found
for illustration was a page fault microbenchmark. It's not representative
of the workload but demonstrates the hazard of the current behaviour.
pft timings
5.10.0-rc2 5.10.0-rc2
imbalancefloat-v2 forkspread-v2
Amean elapsed-1 46.37 ( 0.00%) 46.05 * 0.69%*
Amean elapsed-4 12.43 ( 0.00%) 12.49 * -0.47%*
Amean elapsed-7 7.61 ( 0.00%) 7.55 * 0.81%*
Amean elapsed-12 4.79 ( 0.00%) 4.80 ( -0.17%)
Amean elapsed-21 3.13 ( 0.00%) 2.89 * 7.74%*
Amean elapsed-30 3.65 ( 0.00%) 2.27 * 37.62%*
Amean elapsed-48 3.08 ( 0.00%) 2.13 * 30.69%*
Amean elapsed-79 2.00 ( 0.00%) 1.90 * 4.95%*
Amean elapsed-80 2.00 ( 0.00%) 1.90 * 4.70%*
This is showing the time to fault regions belonging to threads. The target
machine has 80 logical CPUs and two nodes. Note the ~30% gain when the
machine is approximately the point where one node becomes fully utilised.
The slower results are borderline noise.
Kernel building shows similar benefits around the same balance point.
Generally performance was either neutral or better in the tests conducted.
The main consideration with this patch is the point where fork stops
spreading a task so some workloads may benefit from different balance
points but it would be a risky tuning parameter.
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Link: https://lkml.kernel.org/r/20201120090630.3286-5-mgorman@techsingularity.net
Currently, an imbalance is only allowed when a destination node
is almost completely idle. This solved one basic class of problems
and was the cautious approach.
This patch revisits the possibility that NUMA nodes can be imbalanced
until 25% of the CPUs are occupied. The reasoning behind 25% is somewhat
superficial -- it's half the cores when HT is enabled. At higher
utilisations, balancing should continue as normal and keep things even
until scheduler domains are fully busy or over utilised.
Note that this is not expected to be a universal win. Any benchmark
that prefers spreading as wide as possible with limited communication
will favour the old behaviour as there is more memory bandwidth.
Workloads that communicate heavily in pairs such as netperf or tbench
benefit. For the tests I ran, the vast majority of workloads saw
a benefit so it seems to be a worthwhile trade-off.
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Link: https://lkml.kernel.org/r/20201120090630.3286-4-mgorman@techsingularity.net
In find_idlest_group(), the load imbalance is only relevant when the group
is either overloaded or fully busy but it is calculated unconditionally.
This patch moves the imbalance calculation to the context it is required.
Technically, it is a micro-optimisation but really the benefit is avoiding
confusing one type of imbalance with another depending on the group_type
in the next patch.
No functional change.
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Link: https://lkml.kernel.org/r/20201120090630.3286-3-mgorman@techsingularity.net
We call arch_cpu_idle() with RCU disabled, but then use
local_irq_{en,dis}able(), which invokes tracing, which relies on RCU.
Switch all arch_cpu_idle() implementations to use
raw_local_irq_{en,dis}able() and carefully manage the
lockdep,rcu,tracing state like we do in entry.
(XXX: we really should change arch_cpu_idle() to not return with
interrupts enabled)
Reported-by: Sven Schnelle <svens@linux.ibm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Mark Rutland <mark.rutland@arm.com>
Tested-by: Mark Rutland <mark.rutland@arm.com>
Link: https://lkml.kernel.org/r/20201120114925.594122626@infradead.org
