[Qemu-devel] [RFC v3 56/56] cputlb: queue async flush jobs without the BQL

["Emilio G. Cota" <cota@braap.org>]

This yields sizable scalability improvements, as the below results show.

Host: Two Intel E5-2683 v3 14-core CPUs at 2.00 GHz (Haswell)

Workload: Ubuntu 18.04 ppc64 compiling the linux kernel with
"make -j N", where N is the number of cores in the guest.

                      Speedup vs a single thread (higher is better):

         14 +---------------------------------------------------------------+
            |       +    +       +      +       +      +      $$$$$$  +     |
            |                                            $$$$$              |
            |                                      $$$$$$                   |
         12 |-+                                $A$$                       +-|
            |                                $$                             |
            |                             $$$                               |
         10 |-+                         $$    ##D#####################D   +-|
            |                        $$$ #####**B****************           |
            |                      $$####*****                   *****      |
            |                    A$#*****                             B     |
          8 |-+                $$B**                                      +-|
            |                $$**                                           |
            |               $**                                             |
          6 |-+           $$*                                             +-|
            |            A**                                                |
            |           $B                                                  |
            |           $                                                   |
          4 |-+        $*                                                 +-|
            |          $                                                    |
            |         $                                                     |
          2 |-+      $                                                    +-|
            |        $                                 +cputlb-no-bql $$A$$ |
            |       A                                   +per-cpu-lock ##D## |
            |       +    +       +      +       +      +     baseline **B** |
          0 +---------------------------------------------------------------+
                    1    4       8      12      16     20      24     28
                                       Guest vCPUs
  png: https://imgur.com/zZRvS7q

Some notes:
- baseline corresponds to the commit before this series

- per-cpu-lock is the commit that converts the CPU loop to per-cpu locks.

- cputlb-no-bql is this commit.

- I'm using taskset to assign cores to threads, favouring locality whenever
  possible but not using SMT. When N=1, I'm using a single host core, which
  leads to superlinear speedups (since with more cores the I/O thread can execute
  while vCPU threads sleep). In the future I might use N+1 host cores for N
  guest cores to avoid this, or perhaps pin guest threads to cores one-by-one.

- Scalability is not good at 64 cores, where the BQL for handling
  interrupts dominates. I got this from another machine (a 64-core one),
  that unfortunately is much slower than this 28-core one, so I don't have
  the numbers for 1-16 cores. The plot is normalized at 16-core baseline
  performance, and therefore very ugly :-) https://imgur.com/XyKGkAw
  See below for an example of the *huge* amount of waiting on the BQL:

(qemu) info sync-profile
Type               Object  Call site                             Wait Time (s)         Count  Average (us)
----------------------------------------------------------------------------------------------------------
BQL mutex  0x55ba286c9800  accel/tcg/cpu-exec.c:545                 2868.85676      14872596        192.90
BQL mutex  0x55ba286c9800  hw/ppc/ppc.c:70                           539.58924       3666820        147.15
BQL mutex  0x55ba286c9800  target/ppc/helper_regs.h:105              323.49283       2544959        127.11
mutex      [           2]  util/qemu-timer.c:426                     181.38420       3666839         49.47
condvar    [          61]  cpus.c:1327                               136.50872         15379       8876.31
BQL mutex  0x55ba286c9800  accel/tcg/cpu-exec.c:516                   86.14785        946301         91.04
condvar    0x55ba286eb6a0  cpus-common.c:196                          78.41010           126     622302.35
BQL mutex  0x55ba286c9800  util/main-loop.c:236                       28.14795        272940        103.13
mutex      [          64]  include/qom/cpu.h:514                      17.87662      75139413          0.24
BQL mutex  0x55ba286c9800  target/ppc/translate_init.inc.c:8665        7.04738         36528        192.93
----------------------------------------------------------------------------------------------------------

Single-threaded performance is affected very lightly. Results
below for debian aarch64 bootup+test for the entire series
on an Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz host:

- Before:

 Performance counter stats for 'taskset -c 0 ../img/aarch64/die.sh' (10 runs):

       7269.033478      task-clock (msec)         #    0.998 CPUs utilized            ( +-  0.06% )
    30,659,870,302      cycles                    #    4.218 GHz                      ( +-  0.06% )
    54,790,540,051      instructions              #    1.79  insns per cycle          ( +-  0.05% )
     9,796,441,380      branches                  # 1347.695 M/sec                    ( +-  0.05% )
       165,132,201      branch-misses             #    1.69% of all branches          ( +-  0.12% )

       7.287011656 seconds time elapsed                                          ( +-  0.10% )

- After:

       7375.924053      task-clock (msec)         #    0.998 CPUs utilized            ( +-  0.13% )
    31,107,548,846      cycles                    #    4.217 GHz                      ( +-  0.12% )
    55,355,668,947      instructions              #    1.78  insns per cycle          ( +-  0.05% )
     9,929,917,664      branches                  # 1346.261 M/sec                    ( +-  0.04% )
       166,547,442      branch-misses             #    1.68% of all branches          ( +-  0.09% )

       7.389068145 seconds time elapsed                                          ( +-  0.13% )

That is, a 1.37% slowdown.

Cc: Peter Crosthwaite <crosthwaite.peter@gmail.com>
Cc: Richard Henderson <rth@twiddle.net>
Signed-off-by: Emilio G. Cota <cota@braap.org>
---
 accel/tcg/cputlb.c | 19 ++++++++++---------
 1 file changed, 10 insertions(+), 9 deletions(-)

diff --git a/accel/tcg/cputlb.c b/accel/tcg/cputlb.c
index 353d76d6a5..e3582f2f1d 100644
--- a/accel/tcg/cputlb.c
+++ b/accel/tcg/cputlb.c
@@ -212,7 +212,7 @@ static void flush_all_helper(CPUState *src, run_on_cpu_func fn,
 
     CPU_FOREACH(cpu) {
         if (cpu != src) {
-            async_run_on_cpu(cpu, fn, d);
+            async_run_on_cpu_no_bql(cpu, fn, d);
         }
     }
 }
@@ -280,8 +280,8 @@ void tlb_flush(CPUState *cpu)
     if (cpu->created && !qemu_cpu_is_self(cpu)) {
         if (atomic_mb_read(&cpu->pending_tlb_flush) != ALL_MMUIDX_BITS) {
             atomic_mb_set(&cpu->pending_tlb_flush, ALL_MMUIDX_BITS);
-            async_run_on_cpu(cpu, tlb_flush_global_async_work,
-                             RUN_ON_CPU_NULL);
+            async_run_on_cpu_no_bql(cpu, tlb_flush_global_async_work,
+                                    RUN_ON_CPU_NULL);
         }
     } else {
         tlb_flush_nocheck(cpu);
@@ -341,8 +341,8 @@ void tlb_flush_by_mmuidx(CPUState *cpu, uint16_t idxmap)
             tlb_debug("reduced mmu_idx: 0x%" PRIx16 "\n", pending_flushes);
 
             atomic_or(&cpu->pending_tlb_flush, pending_flushes);
-            async_run_on_cpu(cpu, tlb_flush_by_mmuidx_async_work,
-                             RUN_ON_CPU_HOST_INT(pending_flushes));
+            async_run_on_cpu_no_bql(cpu, tlb_flush_by_mmuidx_async_work,
+                                    RUN_ON_CPU_HOST_INT(pending_flushes));
         }
     } else {
         tlb_flush_by_mmuidx_async_work(cpu,
@@ -442,8 +442,8 @@ void tlb_flush_page(CPUState *cpu, target_ulong addr)
     tlb_debug("page :" TARGET_FMT_lx "\n", addr);
 
     if (!qemu_cpu_is_self(cpu)) {
-        async_run_on_cpu(cpu, tlb_flush_page_async_work,
-                         RUN_ON_CPU_TARGET_PTR(addr));
+        async_run_on_cpu_no_bql(cpu, tlb_flush_page_async_work,
+                                RUN_ON_CPU_TARGET_PTR(addr));
     } else {
         tlb_flush_page_async_work(cpu, RUN_ON_CPU_TARGET_PTR(addr));
     }
@@ -514,8 +514,9 @@ void tlb_flush_page_by_mmuidx(CPUState *cpu, target_ulong addr, uint16_t idxmap)
     addr_and_mmu_idx |= idxmap;
 
     if (!qemu_cpu_is_self(cpu)) {
-        async_run_on_cpu(cpu, tlb_check_page_and_flush_by_mmuidx_async_work,
-                         RUN_ON_CPU_TARGET_PTR(addr_and_mmu_idx));
+        async_run_on_cpu_no_bql(cpu,
+                                tlb_check_page_and_flush_by_mmuidx_async_work,
+                                RUN_ON_CPU_TARGET_PTR(addr_and_mmu_idx));
     } else {
         tlb_check_page_and_flush_by_mmuidx_async_work(
             cpu, RUN_ON_CPU_TARGET_PTR(addr_and_mmu_idx));
-- 
2.17.1