Re: [PATCH 5/5] contrib/plugins: add ips plugin example for cost modeling

["Alex Bennée" <alex.bennee@linaro.org>]

Pierrick Bouvier <pierrick.bouvier@linaro.org> writes:

(Added Philip to CC)

> On 5/28/24 12:57, Alex Bennée wrote:
>> Pierrick Bouvier <pierrick.bouvier@linaro.org> writes:
>> 
>>> On 5/28/24 12:14, Alex Bennée wrote:
>>>> Pierrick Bouvier <pierrick.bouvier@linaro.org> writes:
>>>>
>>>>> This plugin uses the new time control interface to make decisions
>>>>> about the state of time during the emulation. The algorithm is
>>>>> currently very simple. The user specifies an ips rate which applies
>>>>> per core.

<snip>

>>>>> +static void vcpu_init(qemu_plugin_id_t id, unsigned int cpu_index)
>>>>> +{
>>>>> +    vCPUTime *vcpu = qemu_plugin_scoreboard_find(vcpus, cpu_index);
>>>>> +    /* ensure start time is set first */
>>>>> +    set_start_time();
>>>>> +    /* start counter from absolute time reference */
>>>>> +    vcpu->counter = num_insn_during(uptime_ns());
>>>>> +    vcpu_set_state(vcpu, EXECUTING);
>>>>> +}
>>>>> +
>>>>> +static void vcpu_idle(qemu_plugin_id_t id, unsigned int cpu_index)
>>>>> +{
>>>>> +    vCPUTime *vcpu = qemu_plugin_scoreboard_find(vcpus, cpu_index);
>>>>> +    vcpu_set_state(vcpu, IDLE);
>>>>> +}
>>>>> +
>>>>> +static void vcpu_resume(qemu_plugin_id_t id, unsigned int cpu_index)
>>>>> +{
>>>>> +    vCPUTime *vcpu = qemu_plugin_scoreboard_find(vcpus, cpu_index);
>>>>> +    g_assert(vcpu->state == IDLE);
>>>> I'm triggering a weird race here:
>>>>     (gdb) b vcpu_init
>>>>     Breakpoint 1 at 0x7ffff7fa15f7: file /home/alex/lsrc/qemu.git/contrib/plugins/ips.c, line 127.
>>>>     (gdb) r
>>>>     The program being debugged has been started already.
>>>>     Start it from the beginning? (y or n) y
>>>>     Starting program:
>>>> /home/alex/lsrc/qemu.git/builds/arm.debug/qemu-system-aarch64
>>>> -machine type=virt,virtualization=on,pflash0=rom,pflash1=efivars
>>>> -cpu cortex-a57 -smp 32 -accel tcg -device
>>>> virtio-net-pci,netdev=unet -device virtio-scsi-pci -device
>>>> scsi-hd,drive=hd -netdev user,id=unet,hostfwd=tcp::2222-:22
>>>> -blockdev
>>>> driver=raw,node-name=hd,file.driver=host_device,file.filename=/dev/zen-ssd2/trixie-arm64,discard=unmap
>>>> -serial mon:stdio -blockdev
>>>> node-name=rom,driver=file,filename=/home/alex/lsrc/qemu.git/builds/arm.debug/pc-bios/edk2-aarch64-code.fd,read-only=true
>>>> -blockdev
>>>> node-name=efivars,driver=file,filename=/home/alex/images/qemu-arm64-efivars
>>>> -m 8192 -object memory-backend-memfd,id=mem,size=8G,share=on -kernel
>>>> /home/alex/lsrc/linux.git/builds/arm64/arch/arm64/boot/Image -append
>>>> root=/dev/sda2\ console=ttyAMA0 -plugin
>>>> contrib/plugins/libips.so,ips=1000000000
>>>>     [Thread debugging using libthread_db enabled]
>>>>     Using host libthread_db library "/lib/x86_64-linux-gnu/libthread_db.so.1".
>>>>     [New Thread 0x7fffe72006c0 (LWP 360538)]
>>>>     [New Thread 0x7fffe68006c0 (LWP 360540)]
>>>>     [New Thread 0x7fffe5e006c0 (LWP 360541)]
>>>>     [New Thread 0x7fffe54006c0 (LWP 360542)]
>>>>     [New Thread 0x7fffe4a006c0 (LWP 360543)]
>>>>     [New Thread 0x7fffdfe006c0 (LWP 360544)]
>>>>     [New Thread 0x7fffdf4006c0 (LWP 360545)]
>>>>     [New Thread 0x7fffdea006c0 (LWP 360546)]
>>>>     [Switching to Thread 0x7fffdf4006c0 (LWP 360545)]
>>>>     Thread 8 "qemu-system-aar" hit Breakpoint 1, vcpu_init
>>>> (id=10457908569352202058, cpu_index=0) at
>>>> /home/alex/lsrc/qemu.git/contrib/plugins/ips.c:127
>>>>     127         vCPUTime *vcpu = qemu_plugin_scoreboard_find(vcpus, cpu_index);
>>>>     (gdb) c
>>>>     Continuing.
>>>>     [New Thread 0x7fffde0006c0 (LWP 360548)]
>>>>     [Switching to Thread 0x7fffdea006c0 (LWP 360546)]
>>>>     Thread 9 "qemu-system-aar" hit Breakpoint 1, vcpu_init
>>>> (id=10457908569352202058, cpu_index=1) at
>>>> /home/alex/lsrc/qemu.git/contrib/plugins/ips.c:127
>>>>     127         vCPUTime *vcpu = qemu_plugin_scoreboard_find(vcpus, cpu_index);
>>>>     (gdb)
>>>>     Continuing.
>>>>     [New Thread 0x7fffdd6006c0 (LWP 360549)]
>>>>     [Switching to Thread 0x7fffde0006c0 (LWP 360548)]
>>>>     Thread 10 "qemu-system-aar" hit Breakpoint 1, vcpu_init
>>>> (id=10457908569352202058, cpu_index=2) at
>>>> /home/alex/lsrc/qemu.git/contrib/plugins/ips.c:127
>>>>     127         vCPUTime *vcpu = qemu_plugin_scoreboard_find(vcpus, cpu_index);
>>>>     (gdb)
>>>>     Continuing.
>>>>     [New Thread 0x7fffdcc006c0 (LWP 360550)]
>>>>     [Switching to Thread 0x7fffdd6006c0 (LWP 360549)]
>>>>     Thread 11 "qemu-system-aar" hit Breakpoint 1, vcpu_init
>>>> (id=10457908569352202058, cpu_index=3) at
>>>> /home/alex/lsrc/qemu.git/contrib/plugins/ips.c:127
>>>>     127         vCPUTime *vcpu = qemu_plugin_scoreboard_find(vcpus, cpu_index);
>>>>     (gdb)
>>>>     Continuing.
>>>>     [New Thread 0x7fffd3e006c0 (LWP 360551)]
>>>>     [Switching to Thread 0x7fffdcc006c0 (LWP 360550)]
>>>>     Thread 12 "qemu-system-aar" hit Breakpoint 1, vcpu_init
>>>> (id=10457908569352202058, cpu_index=4) at
>>>> /home/alex/lsrc/qemu.git/contrib/plugins/ips.c:127
>>>>     127         vCPUTime *vcpu = qemu_plugin_scoreboard_find(vcpus, cpu_index);
>>>>     (gdb) n
>>>>     129         set_start_time();
>>>>     (gdb)
>>>>     131         vcpu->counter = num_insn_during(uptime_ns());
>>>>     (gdb)
>>>>     132         vcpu_set_state(vcpu, EXECUTING);
>>>>     (gdb)
>>>>     133     }
>>>>     (gdb) p vcpu->state
>>>>     $1 = EXECUTING
>>>>     (gdb) p &vcpu->state
>>>>     $2 = (vCPUState *) 0x555557c6b5d0
>>>>     (gdb) watch *(vCPUState *) 0x555557c6b5d0
>>>>     Hardware watchpoint 2: *(vCPUState *) 0x555557c6b5d0
>>>>     (gdb) c
>>>>     Continuing.
>>>>     [Thread 0x7fffdfe006c0 (LWP 360544) exited]
>>>>     [Thread 0x7fffe5e006c0 (LWP 360541) exited]
>>>>     Thread 12 "qemu-system-aar" hit Hardware watchpoint 2:
>>>> *(vCPUState *) 0x555557c6b5d0
>>>>     Old value = EXECUTING
>>>>     New value = IDLE
>>>>     vcpu_set_state (vcpu=0x555557c6b5c0, new_state=IDLE) at /home/alex/lsrc/qemu.git/contrib/plugins/ips.c:83
>>>>     83      }
>>>>     (gdb) c
>>>>     Continuing.
>>>>     [New Thread 0x7fffdfe006c0 (LWP 360591)]
>>>>     [New Thread 0x7fffe5e006c0 (LWP 360592)]
>>>>     [New Thread 0x7fffd34006c0 (LWP 360593)]
>>>>     [Switching to Thread 0x7fffdfe006c0 (LWP 360591)]
>>>>     Thread 14 "qemu-system-aar" hit Breakpoint 1, vcpu_init
>>>> (id=10457908569352202058, cpu_index=6) at
>>>> /home/alex/lsrc/qemu.git/contrib/plugins/ips.c:127
>>>>     127         vCPUTime *vcpu = qemu_plugin_scoreboard_find(vcpus, cpu_index);
>>>>     (gdb)
>>>>     Continuing.
>>>>     [New Thread 0x7fffd2a006c0 (LWP 360594)]
>>>>     **
>>>>     ERROR:/home/alex/lsrc/qemu.git/contrib/plugins/ips.c:144:vcpu_resume: assertion failed: (vcpu->state == IDLE)
>>>>     Bail out! ERROR:/home/alex/lsrc/qemu.git/contrib/plugins/ips.c:144:vcpu_resume: assertion failed: (vcpu->state == IDLE)
>>>>     Thread 13 "qemu-system-aar" received signal SIGABRT, Aborted.
>>>>     [Switching to Thread 0x7fffd3e006c0 (LWP 360551)]
>>>>     __pthread_kill_implementation (threadid=<optimized out>, signo=signo@entry=6, no_tid=no_tid@entry=0) at ./nptl/pthread_kill.c:44
>>>>     44      ./nptl/pthread_kill.c: No such file or directory.
>>>>     (gdb) bt
>>>>     #0  __pthread_kill_implementation (threadid=<optimized out>, signo=signo@entry=6, no_tid=no_tid@entry=0) at ./nptl/pthread_kill.c:44
>>>>     #1  0x00007ffff4ca9e8f in __pthread_kill_internal (signo=6, threadid=<optimized out>) at ./nptl/pthread_kill.c:78
>>>>     #2  0x00007ffff4c5afb2 in __GI_raise (sig=sig@entry=6) at ../sysdeps/posix/raise.c:26
>>>>     #3  0x00007ffff4c45472 in __GI_abort () at ./stdlib/abort.c:79
>>>>     #4  0x00007ffff6e46ec8 in  () at /lib/x86_64-linux-gnu/libglib-2.0.so.0
>>>>     #5  0x00007ffff6ea6e1a in g_assertion_message_expr () at /lib/x86_64-linux-gnu/libglib-2.0.so.0
>>>>     #6  0x00007ffff7fa16e2 in vcpu_resume (id=10457908569352202058, cpu_index=5) at /home/alex/lsrc/qemu.git/contrib/plugins/ips.c:144
>>>>     #7  0x00005555562c3202 in plugin_vcpu_cb__simple (cpu=0x5555583d9540, ev=QEMU_PLUGIN_EV_VCPU_RESUME) at ../../plugins/core.c:111
>>>>     #8  0x00005555562c43f7 in qemu_plugin_vcpu_resume_cb (cpu=0x5555583d9540) at ../../plugins/core.c:523
>>>>     #9  0x0000555555d9150e in qemu_wait_io_event (cpu=0x5555583d9540) at ../../system/cpus.c:465
>>>>     #10 0x00005555562c67d0 in mttcg_cpu_thread_fn (arg=0x5555583d9540) at ../../accel/tcg/tcg-accel-ops-mttcg.c:118
>>>>     #11 0x00005555565120ff in qemu_thread_start (args=0x5555584716e0) at ../../util/qemu-thread-posix.c:541
>>>>     #12 0x00007ffff4ca8134 in start_thread (arg=<optimized out>) at ./nptl/pthread_create.c:442
>>>>     #13 0x00007ffff4d287dc in clone3 () at ../sysdeps/unix/sysv/linux/x86_64/clone3.S:81
>>>>     (gdb)
>>>> But I don't understand how we can ever hit the idle callback without
>>>> first hitting the init callback.
>>>>
>>>
>>> More exactly, the assert we hit means that the idle callback, was not
>>> called before the resume callback.
>>> Any chance you can check what is the current vcpu->state value? I
>>> wonder if it's not an exited cpu, that gets resumed after.
>> No when I looked at it it was set as UNKNOWN. Unfortunately it
>> doesn't
>> trigger if I stick breakpoints in or run under rr.
>>
>
> By using rr record --chaos mode maybe?

Ahh yes that triggered it.

It looks like we have a race:

  (rr) c
  Continuing.

  Thread 1 hit Hardware watchpoint 7: *(int *) 0x5559ba6899b0

  Old value = -1
  New value = 10
  machvirt_init (machine=0x5559b9ac3c00) at ../../hw/arm/virt.c:2214
  2214            numa_cpu_pre_plug(&possible_cpus->cpus[cs->cpu_index], DEVICE(cpuobj),
  (rr) c
  Continuing.
  [New Thread 650125.650142]
  [Switching to Thread 650125.650142]

  Thread 44 hit Breakpoint 8, mttcg_cpu_thread_fn (arg=0x5559ba6896e0) at ../../accel/tcg/tcg-accel-ops-mttcg.c:70
  70          assert(tcg_enabled());
  (rr) p cpu->cpu_index
  $24 = 10
  (rr) c
  Continuing.
  [Switching to Thread 650125.650125]

  Thread 1 hit Breakpoint 10, cpu_common_realizefn (dev=0x5559ba6896e0, errp=0x7fff02322c10) at ../../hw/core/cpu-common.c:205
  205         Object *machine = qdev_get_machine();
  (rr) p cpu->cpu_index
  $25 = 10
  (rr) bt
  #0  cpu_common_realizefn (dev=0x5559ba6896e0, errp=0x7fff02322c10) at ../../hw/core/cpu-common.c:205
  #1  0x00005559b606e6a5 in arm_cpu_realizefn (dev=0x5559ba6896e0, errp=0x7fff02322c10) at ../../target/arm/cpu.c:2552
  #2  0x00005559b63dd5d0 in device_set_realized (obj=0x5559ba6896e0, value=true, errp=0x7fff02322d20) at ../../hw/core/qdev.c:510
  #3  0x00005559b63e84e7 in property_set_bool (obj=0x5559ba6896e0, v=0x5559ba681540, name=0x5559b68fafe1 "realized", opaque=0x5559b97f73c0, errp=0x7fff02322d20)
      at ../../qom/object.c:2354
  #4  0x00005559b63e6065 in object_property_set (obj=0x5559ba6896e0, name=0x5559b68fafe1 "realized", v=0x5559ba681540, errp=0x7fff02322d20) at ../../qom/object.c:1463
  #5  0x00005559b63ead1d in object_property_set_qobject (obj=0x5559ba6896e0, name=0x5559b68fafe1 "realized", value=0x5559ba681420, errp=0x5559b7810320 <error_fatal>)
      at ../../qom/qom-qobject.c:28
  #6  0x00005559b63e640a in object_property_set_bool (obj=0x5559ba6896e0, name=0x5559b68fafe1 "realized", value=true, errp=0x5559b7810320 <error_fatal>)
      at ../../qom/object.c:1533
  #7  0x00005559b63dccbd in qdev_realize (dev=0x5559ba6896e0, bus=0x0, errp=0x5559b7810320 <error_fatal>) at ../../hw/core/qdev.c:291
  #8  0x00005559b5fe4243 in machvirt_init (machine=0x5559b9ac3c00) at ../../hw/arm/virt.c:2295
  #9  0x00005559b5ada5d6 in machine_run_board_init (machine=0x5559b9ac3c00, mem_path=0x0, errp=0x7fff02322f30) at ../../hw/core/machine.c:1576
  #10 0x00005559b5eb416f in qemu_init_board () at ../../system/vl.c:2621
  #11 0x00005559b5eb4482 in qmp_x_exit_preconfig (errp=0x5559b7810320 <error_fatal>) at ../../system/vl.c:2713
  #12 0x00005559b5eb7088 in qemu_init (argc=35, argv=0x7fff02323268) at ../../system/vl.c:3759
  #13 0x00005559b63d7ff2 in main (argc=35, argv=0x7fff02323268) at ../../system/main.c:47

Because looking at the order things happen in the specific CPU realisation:

    qemu_init_vcpu(cs);
    cpu_reset(cs);

    acc->parent_realize(dev, errp);

And the common realize function doesn't get to do:

    /* Plugin initialization must wait until the cpu start executing code */
#ifdef CONFIG_PLUGIN
    if (tcg_enabled()) {
        cpu->plugin_state = qemu_plugin_create_vcpu_state();
        async_run_on_cpu(cpu, qemu_plugin_vcpu_init__async, RUN_ON_CPU_NULL);
    }
#endif

before we've started executing.....

-- 
Alex Bennée
Virtualisation Tech Lead @ Linaro