Skip to content

cloud-kernel
cloud-kernel

openanolis / cloud-kernel
1 年多 前同步成功

通知 160
Star 36
Fork 7
cloud-kernel
cloud-kernel
切换分支/标签
  1. 09 2月, 2015 1 次提交
  3. 06 2月, 2015 1 次提交
    • P
      kvm: add halt_poll_ns module parameter · f7819512
      由 Paolo Bonzini 提交于 
      This patch introduces a new module parameter for the KVM module; when it
is present, KVM attempts a bit of polling on every HLT before scheduling
itself out via kvm_vcpu_block.

This parameter helps a lot for latency-bound workloads---in particular
I tested it with O_DSYNC writes with a battery-backed disk in the host.
In this case, writes are fast (because the data doesn't have to go all
the way to the platters) but they cannot be merged by either the host or
the guest.  KVM's performance here is usually around 30% of bare metal,
or 50% if you use cache=directsync or cache=writethrough (these
parameters avoid that the guest sends pointless flush requests, and
at the same time they are not slow because of the battery-backed cache).
The bad performance happens because on every halt the host CPU decides
to halt itself too.  When the interrupt comes, the vCPU thread is then
migrated to a new physical CPU, and in general the latency is horrible
because the vCPU thread has to be scheduled back in.

With this patch performance reaches 60-65% of bare metal and, more
important, 99% of what you get if you use idle=poll in the guest.  This
means that the tunable gets rid of this particular bottleneck, and more
work can be done to improve performance in the kernel or QEMU.

Of course there is some price to pay; every time an otherwise idle vCPUs
is interrupted by an interrupt, it will poll unnecessarily and thus
impose a little load on the host.  The above results were obtained with
a mostly random value of the parameter (500000), and the load was around
1.5-2.5% CPU usage on one of the host's core for each idle guest vCPU.

The patch also adds a new stat, /sys/kernel/debug/kvm/halt_successful_poll,
that can be used to tune the parameter.  It counts how many HLT
instructions received an interrupt during the polling period; each
successful poll avoids that Linux schedules the VCPU thread out and back
in, and may also avoid a likely trip to C1 and back for the physical CPU.

While the VM is idle, a Linux 4 VCPU VM halts around 10 times per second.
Of these halts, almost all are failed polls.  During the benchmark,
instead, basically all halts end within the polling period, except a more
or less constant stream of 50 per second coming from vCPUs that are not
running the benchmark.  The wasted time is thus very low.  Things may
be slightly different for Windows VMs, which have a ~10 ms timer tick.

The effect is also visible on Marcelo's recently-introduced latency
test for the TSC deadline timer.  Though of course a non-RT kernel has
awful latency bounds, the latency of the timer is around 8000-10000 clock
cycles compared to 20000-120000 without setting halt_poll_ns.  For the TSC
deadline timer, thus, the effect is both a smaller average latency and
a smaller variance.
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>
      f7819512
  5. 23 1月, 2015 9 次提交
  7. 12 12月, 2014 2 次提交
    • A
      arch: Add lightweight memory barriers dma_rmb() and dma_wmb() · 1077fa36
      由 Alexander Duyck 提交于 
      There are a number of situations where the mandatory barriers rmb() and
wmb() are used to order memory/memory operations in the device drivers
and those barriers are much heavier than they actually need to be.  For
example in the case of PowerPC wmb() calls the heavy-weight sync
instruction when for coherent memory operations all that is really needed
is an lsync or eieio instruction.

This commit adds a coherent only version of the mandatory memory barriers
rmb() and wmb().  In most cases this should result in the barrier being the
same as the SMP barriers for the SMP case, however in some cases we use a
barrier that is somewhere in between rmb() and smp_rmb().  For example on
ARM the rmb barriers break down as follows:

  Barrier   Call     Explanation
  --------- -------- ----------------------------------
  rmb()     dsb()    Data synchronization barrier - system
  dma_rmb() dmb(osh) data memory barrier - outer sharable
  smp_rmb() dmb(ish) data memory barrier - inner sharable

These new barriers are not as safe as the standard rmb() and wmb().
Specifically they do not guarantee ordering between coherent and incoherent
memories.  The primary use case for these would be to enforce ordering of
reads and writes when accessing coherent memory that is shared between the
CPU and a device.

It may also be noted that there is no dma_mb().  Most architectures don't
provide a good mechanism for performing a coherent only full barrier without
resorting to the same mechanism used in mb().  As such there isn't much to
be gained in trying to define such a function.

Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
Cc: Michael Ellerman <michael@ellerman.id.au>
Cc: Michael Neuling <mikey@neuling.org>
Cc: Russell King <linux@arm.linux.org.uk>
Cc: Geert Uytterhoeven <geert@linux-m68k.org>
Cc: Heiko Carstens <heiko.carstens@de.ibm.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Martin Schwidefsky <schwidefsky@de.ibm.com>
Cc: Tony Luck <tony.luck@intel.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: David Miller <davem@davemloft.net>
Acked-by: NBenjamin Herrenschmidt <benh@kernel.crashing.org>
Acked-by: NWill Deacon <will.deacon@arm.com>
Signed-off-by: NAlexander Duyck <alexander.h.duyck@redhat.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>
      1077fa36
    • A
      arch: Cleanup read_barrier_depends() and comments · 8a449718
      由 Alexander Duyck 提交于 
      This patch is meant to cleanup the handling of read_barrier_depends and
smp_read_barrier_depends.  In multiple spots in the kernel headers
read_barrier_depends is defined as "do {} while (0)", however we then go
into the SMP vs non-SMP sections and have the SMP version reference
read_barrier_depends, and the non-SMP define it as yet another empty
do/while.

With this commit I went through and cleaned out the duplicate definitions
and reduced the number of definitions down to 2 per header.  In addition I
moved the 50 line comments for the macro from the x86 and mips headers that
defined it as an empty do/while to those that were actually defining the
macro, alpha and blackfin.
Signed-off-by: NAlexander Duyck <alexander.h.duyck@redhat.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>
      8a449718
  9. 11 12月, 2014 1 次提交
  11. 08 12月, 2014 5 次提交
  13. 06 12月, 2014 1 次提交
    • A
      net: sock: allow eBPF programs to be attached to sockets · 89aa0758
      由 Alexei Starovoitov 提交于 
      introduce new setsockopt() command:

setsockopt(sock, SOL_SOCKET, SO_ATTACH_BPF, &prog_fd, sizeof(prog_fd))

where prog_fd was received from syscall bpf(BPF_PROG_LOAD, attr, ...)
and attr->prog_type == BPF_PROG_TYPE_SOCKET_FILTER

setsockopt() calls bpf_prog_get() which increments refcnt of the program,
so it doesn't get unloaded while socket is using the program.

The same eBPF program can be attached to multiple sockets.

User task exit automatically closes socket which calls sk_filter_uncharge()
which decrements refcnt of eBPF program
Signed-off-by: NAlexei Starovoitov <ast@plumgrid.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>
      89aa0758
  15. 28 11月, 2014 6 次提交
  17. 19 11月, 2014 4 次提交
  19. 12 11月, 2014 1 次提交
    • E
      net: introduce SO_INCOMING_CPU · 2c8c56e1
      由 Eric Dumazet 提交于 
      Alternative to RPS/RFS is to use hardware support for multiple
queues.

Then split a set of million of sockets into worker threads, each
one using epoll() to manage events on its own socket pool.

Ideally, we want one thread per RX/TX queue/cpu, but we have no way to
know after accept() or connect() on which queue/cpu a socket is managed.

We normally use one cpu per RX queue (IRQ smp_affinity being properly
set), so remembering on socket structure which cpu delivered last packet
is enough to solve the problem.

After accept(), connect(), or even file descriptor passing around
processes, applications can use :

 int cpu;
 socklen_t len = sizeof(cpu);

 getsockopt(fd, SOL_SOCKET, SO_INCOMING_CPU, &cpu, &len);

And use this information to put the socket into the right silo
for optimal performance, as all networking stack should run
on the appropriate cpu, without need to send IPI (RPS/RFS).
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>
      2c8c56e1
  21. 10 11月, 2014 1 次提交
    • T
      /dev/mem: Use more consistent data types · 4707a341
      由 Thierry Reding 提交于 
      The xlate_dev_{kmem,mem}_ptr() functions take either a physical address
or a kernel virtual address, so data types should be phys_addr_t and
void *. They both return a kernel virtual address which is only ever
used in calls to copy_{from,to}_user(), so make variables that store it
void * rather than char * for consistency.

Also only define a weak unxlate_dev_mem_ptr() function if architectures
haven't overridden them in the asm/io.h header file.
Signed-off-by: NThierry Reding <treding@nvidia.com>
      4707a341
  23. 03 11月, 2014 3 次提交
    • M
      s390/pci: add sparse annotations · 5b9f2081
      由 Martin Schwidefsky 提交于 
      Fix the following warnings from the sparse code checker:

arch/s390/include/asm/pci_io.h:165:49: warning: cast removes address space of expression
arch/s390/pci/pci.c:476:44: warning: cast removes address space of expression
arch/s390/pci/pci.c:491:36: warning: incorrect type in argument 2 (different address spaces)
arch/s390/pci/pci.c:491:36:    expected void [noderef] <asn:2>*addr
arch/s390/pci/pci.c:491:36:    got void *<noident>
Signed-off-by: NMartin Schwidefsky <schwidefsky@de.ibm.com>
      5b9f2081
    • S
      s390/pci: improve irq number check for msix · b19148f6
      由 Sebastian Ott 提交于 
      s390s arch_setup_msi_irqs function ensures that we don't return with
more irqs than the PCI architecture allows and that a single PCI
function doesn't consume more irqs than the kernel is configured for.

At least the last check doesn't help much and should take the sum of
all irqs into account. Since that's already done by irq_alloc_desc
we can remove this check.

As for the first check we should use the value provided by the
firmware which can be less than what the PCI architecture allows.
Signed-off-by: NSebastian Ott <sebott@linux.vnet.ibm.com>
Signed-off-by: NMartin Schwidefsky <schwidefsky@de.ibm.com>
      b19148f6
    • M
      s390/cmpxchg: use compiler builtins · f318a122
      由 Martin Schwidefsky 提交于 
      The kernel build for s390 fails for gcc compilers with version 3.x,
set the minimum required version of gcc to version 4.3.

As the atomic builtins are available with all gcc 4.x compilers,
use the __sync_val_compare_and_swap and __sync_bool_compare_and_swap
functions to replace the complex macro and inline assembler magic
in include/asm/cmpxchg.h. The compiler can just-do-it and generates
better code with the builtins.

While we are at it use __sync_bool_compare_and_swap for the
_raw_compare_and_swap function in the spinlock code as well.
Signed-off-by: NMartin Schwidefsky <schwidefsky@de.ibm.com>
      f318a122
  25. 28 10月, 2014 3 次提交
  27. 27 10月, 2014 2 次提交
    • M
      s390/mm: pmdp_get_and_clear_full optimization · fcbe08d6
      由 Martin Schwidefsky 提交于 
      Analog to ptep_get_and_clear_full define a variant of the
pmpd_get_and_clear primitive which gets the full hint from the
mmu_gather struct. This allows s390 to avoid a costly instruction
when destroying an address space.
Signed-off-by: NMartin Schwidefsky <schwidefsky@de.ibm.com>
      fcbe08d6
    • H
      s390/ftrace,kprobes: allow to patch first instruction · c933146a
      由 Heiko Carstens 提交于 
      If the function tracer is enabled, allow to set kprobes on the first
instruction of a function (which is the function trace caller):

If no kprobe is set handling of enabling and disabling function tracing
of a function simply patches the first instruction. Either it is a nop
(right now it's an unconditional branch, which skips the mcount block),
or it's a branch to the ftrace_caller() function.

If a kprobe is being placed on a function tracer calling instruction
we encode if we actually have a nop or branch in the remaining bytes
after the breakpoint instruction (illegal opcode).
This is possible, since the size of the instruction used for the nop
and branch is six bytes, while the size of the breakpoint is only
two bytes.
Therefore the first two bytes contain the illegal opcode and the last
four bytes contain either "0" for nop or "1" for branch. The kprobes
code will then execute/simulate the correct instruction.

Instruction patching for kprobes and function tracer is always done
with stop_machine(). Therefore we don't have any races where an
instruction is patched concurrently on a different cpu.
Besides that also the program check handler which executes the function
trace caller instruction won't be executed concurrently to any
stop_machine() execution.

This allows to keep full fault based kprobes handling which generates
correct pt_regs contents automatically.
Signed-off-by: NHeiko Carstens <heiko.carstens@de.ibm.com>
Signed-off-by: NMartin Schwidefsky <schwidefsky@de.ibm.com>
      c933146a