| Age | Commit message (Collapse) | Author |
|---|
|
No functional change in this patch. A helper is added to find if
vcpu is dispatched by hypervisor. Use that instead of opencoding.
Also clarify some of the comments.
Signed-off-by: "Aneesh Kumar K.V" <aneesh.kumar@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://msgid.link/20231114071219.198222-1-aneesh.kumar@linux.ibm.com
|
|
PowerVM Hypervisor dispatches on a whole core basis. In a shared LPAR, a
CPU from a core that is CEDED or preempted may have a larger latency. In
such a scenario, its preferable to choose a different CPU to run.
If one of the CPUs in the core is active, i.e neither CEDED nor
preempted, then consider this CPU as not preempted.
Also if any of the CPUs in the core has yielded but OS has not requested
CEDE or CONFER, then consider this CPU to be preempted.
Correct detection of preempted CPUs is important for detecting idle
CPUs/cores in task scheduler.
Tested-by: Aboorva Devarajan <aboorvad@linux.vnet.ibm.com>
Reviewed-by: Shrikanth Hegde <sshegde@linux.vnet.ibm.com>
Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://msgid.link/20231019091452.95260-1-srikar@linux.vnet.ibm.com
|
|
By adding a forward declaration for struct lppaca we can untangle paca.h
and lppaca.h. Also move get_lppaca() into lppaca.h for consistency.
Add includes of lppaca.h to some files that need it.
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://msgid.link/20230823055317.751786-3-mpe@ellerman.id.au
|
|
CONFIG_VIRT_CPU_ACCOUNTING_GEN under pseries does not provide stolen
time accounting unless CONFIG_PARAVIRT_TIME_ACCOUNTING is enabled.
Implement this using the VPA accumulated wait counters.
Note this will not work on current KVM hosts because KVM does not
implement the VPA dispatch counters (yet). It could be implemented
with the dispatch trace log as it is for VIRT_CPU_ACCOUNTING_NATIVE,
but that is not necessary for the more limited accounting provided
by PARAVIRT_TIME_ACCOUNTING, and it is more expensive, complex, and
has downsides like potential log wrap.
From Shrikanth:
[...] it was tested on Power10 [PowerVM] Shared LPAR. system has two
LPAR. we will call first one LPAR1 and second one as LPAR2. Test was
carried out in SMT=1. Similar observation was seen in SMT=8 as well.
LPAR config header from each LPAR is below. LPAR1 is twice as big as
LPAR2. Since Both are sharing the same underlying hardware, work
stealing will happen when both the LPAR's are contending for the same
resource.
LPAR1:
type=Shared mode=Uncapped smt=Off lcpu=40 cpus=40 ent=20.00
LPAR2:
type=Shared mode=Uncapped smt=Off lcpu=20 cpus=40 ent=10.00
mpstat was used to check for the utilization. stress-ng has been used
as the workload. Few cases are tested. when the both LPAR are idle
there is no steal time. when LPAR1 starts running at 100% which
consumes all of the physical resource, steal time starts to get
accounted. With LPAR1 running at 100% and LPAR2 starts running, steal
time starts increasing. This is as expected. When the LPAR2 Load is
increased further, steal time increases further.
Case 1: 0% LPAR1; 0% LPAR2
%usr %nice %sys %iowait %irq %soft %steal %guest %gnice %idle
0.00 0.00 0.05 0.00 0.00 0.00 0.00 0.00 0.00 99.95
Case 2: 100% LPAR1; 0% LPAR2
%usr %nice %sys %iowait %irq %soft %steal %guest %gnice %idle
97.68 0.00 0.00 0.00 0.00 0.00 2.32 0.00 0.00 0.00
Case 3: 100% LPAR1; 50% LPAR2
%usr %nice %sys %iowait %irq %soft %steal %guest %gnice %idle
86.34 0.00 0.10 0.00 0.00 0.03 13.54 0.00 0.00 0.00
Case 4: 100% LPAR1; 100% LPAR2
%usr %nice %sys %iowait %irq %soft %steal %guest %gnice %idle
78.54 0.00 0.07 0.00 0.00 0.02 21.36 0.00 0.00 0.00
Case 5: 50% LPAR1; 100% LPAR2
%usr %nice %sys %iowait %irq %soft %steal %guest %gnice %idle
49.37 0.00 0.00 0.00 0.00 0.00 1.17 0.00 0.00 49.47
Patch is accounting for the steal time and basic tests are holding
good.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Tested-by: Shrikanth Hegde <sshegde@linux.ibm.com>
[mpe: Add SPDX tag to new paravirt_api_clock.h]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20220902085316.2071519-3-npiggin@gmail.com
|
|
vcpu_is_preempted() can be used outside of preempt-disabled critical
sections, yielding warnings such as:
BUG: using smp_processor_id() in preemptible [00000000] code: systemd-udevd/185
caller is rwsem_spin_on_owner+0x1cc/0x2d0
CPU: 1 PID: 185 Comm: systemd-udevd Not tainted 5.15.0-rc2+ #33
Call Trace:
[c000000012907ac0] [c000000000aa30a8] dump_stack_lvl+0xac/0x108 (unreliable)
[c000000012907b00] [c000000001371f70] check_preemption_disabled+0x150/0x160
[c000000012907b90] [c0000000001e0e8c] rwsem_spin_on_owner+0x1cc/0x2d0
[c000000012907be0] [c0000000001e1408] rwsem_down_write_slowpath+0x478/0x9a0
[c000000012907ca0] [c000000000576cf4] filename_create+0x94/0x1e0
[c000000012907d10] [c00000000057ac08] do_symlinkat+0x68/0x1a0
[c000000012907d70] [c00000000057ae18] sys_symlink+0x58/0x70
[c000000012907da0] [c00000000002e448] system_call_exception+0x198/0x3c0
[c000000012907e10] [c00000000000c54c] system_call_common+0xec/0x250
The result of vcpu_is_preempted() is always used speculatively, and the
function does not access per-cpu resources in a (Linux) preempt-unsafe way.
Use raw_smp_processor_id() to avoid such warnings, adding explanatory
comments.
Fixes: ca3f969dcb11 ("powerpc/paravirt: Use is_kvm_guest() in vcpu_is_preempted()")
Signed-off-by: Nathan Lynch <nathanl@linux.ibm.com>
Reviewed-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20210928214147.312412-3-nathanl@linux.ibm.com
|
|
Add comments more clearly documenting that this function determines whether
hypervisor-level preemption of the VM has occurred.
Signed-off-by: Nathan Lynch <nathanl@linux.ibm.com>
Reviewed-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20210928214147.312412-2-nathanl@linux.ibm.com
|
|
The paravit queued spinlock slow path adds itself to the queue then
calls pv_wait to wait for the lock to become free. This is implemented
by calling H_CONFER to donate cycles.
When hcall tracing is enabled, this H_CONFER call can lead to a spin
lock being taken in the tracing code, which will result in the lock to
be taken again, which will also go to the slow path because it queues
behind itself and so won't ever make progress.
An example trace of a deadlock:
__pv_queued_spin_lock_slowpath
trace_clock_global
ring_buffer_lock_reserve
trace_event_buffer_lock_reserve
trace_event_buffer_reserve
trace_event_raw_event_hcall_exit
__trace_hcall_exit
plpar_hcall_norets_trace
__pv_queued_spin_lock_slowpath
trace_clock_global
ring_buffer_lock_reserve
trace_event_buffer_lock_reserve
trace_event_buffer_reserve
trace_event_raw_event_rcu_dyntick
rcu_irq_exit
irq_exit
__do_irq
call_do_irq
do_IRQ
hardware_interrupt_common_virt
Fix this by introducing plpar_hcall_norets_notrace(), and using that to
make SPLPAR virtual processor dispatching hcalls by the paravirt
spinlock code.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Reviewed-by: Naveen N. Rao <naveen.n.rao@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20210508101455.1578318-2-npiggin@gmail.com
|
|
./arch/powerpc/include/asm/paravirt.h:83:44: error: implicit declaration
of function 'smp_processor_id'; did you mean 'raw_smp_processor_id'?
smp_processor_id is defined in linux/smp.h but it is not included.
The build error happens only when the patch is applied to 5.3 kernel but
it only works by chance in mainline.
Fixes: ca3f969dcb11 ("powerpc/paravirt: Use is_kvm_guest() in vcpu_is_preempted()")
Signed-off-by: Michal Suchanek <msuchanek@suse.de>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20210120132838.15589-1-msuchanek@suse.de
|
|
If its a shared LPAR but not a KVM guest, then see if the vCPU is
related to the calling vCPU. On PowerVM, only cores can be preempted.
So if one vCPU is a non-preempted state, we can decipher that all
other vCPUs sharing the same core are in non-preempted state.
Performance results:
$ perf stat -r 5 -a perf bench sched pipe -l 10000000 (lesser time is better)
powerpc/next
35,107,951.20 msec cpu-clock # 255.898 CPUs utilized ( +- 0.31% )
23,655,348 context-switches # 0.674 K/sec ( +- 3.72% )
14,465 cpu-migrations # 0.000 K/sec ( +- 5.37% )
82,463 page-faults # 0.002 K/sec ( +- 8.40% )
1,127,182,328,206 cycles # 0.032 GHz ( +- 1.60% ) (66.67%)
78,587,300,622 stalled-cycles-frontend # 6.97% frontend cycles idle ( +- 0.08% ) (50.01%)
654,124,218,432 stalled-cycles-backend # 58.03% backend cycles idle ( +- 1.74% ) (50.01%)
834,013,059,242 instructions # 0.74 insn per cycle
# 0.78 stalled cycles per insn ( +- 0.73% ) (66.67%)
132,911,454,387 branches # 3.786 M/sec ( +- 0.59% ) (50.00%)
2,890,882,143 branch-misses # 2.18% of all branches ( +- 0.46% ) (50.00%)
137.195 +- 0.419 seconds time elapsed ( +- 0.31% )
powerpc/next + patchset
29,981,702.64 msec cpu-clock # 255.881 CPUs utilized ( +- 1.30% )
40,162,456 context-switches # 0.001 M/sec ( +- 0.01% )
1,110 cpu-migrations # 0.000 K/sec ( +- 5.20% )
62,616 page-faults # 0.002 K/sec ( +- 3.93% )
1,430,030,626,037 cycles # 0.048 GHz ( +- 1.41% ) (66.67%)
83,202,707,288 stalled-cycles-frontend # 5.82% frontend cycles idle ( +- 0.75% ) (50.01%)
744,556,088,520 stalled-cycles-backend # 52.07% backend cycles idle ( +- 1.39% ) (50.01%)
940,138,418,674 instructions # 0.66 insn per cycle
# 0.79 stalled cycles per insn ( +- 0.51% ) (66.67%)
146,452,852,283 branches # 4.885 M/sec ( +- 0.80% ) (50.00%)
3,237,743,996 branch-misses # 2.21% of all branches ( +- 1.18% ) (50.01%)
117.17 +- 1.52 seconds time elapsed ( +- 1.30% )
This is around 14.6% improvement in performance.
Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com>
Acked-by: Waiman Long <longman@redhat.com>
[mpe: Fold in performance results from cover letter]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20201202050456.164005-5-srikar@linux.vnet.ibm.com
|
|
This implements the generic paravirt qspinlocks using H_PROD and
H_CONFER to kick and wait.
This uses an un-directed yield to any CPU rather than the directed
yield to a pre-empted lock holder that paravirtualised simple
spinlocks use, that requires no kick hcall. This is something that
could be investigated and improved in future.
Performance results can be found in the commit which added queued
spinlocks.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Waiman Long <longman@redhat.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20200724131423.1362108-5-npiggin@gmail.com
|
|
These functions will be used by the queued spinlock implementation,
and may be useful elsewhere too, so move them out of spinlock.h.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Acked-by: Waiman Long <longman@redhat.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20200724131423.1362108-2-npiggin@gmail.com
|
