From: Stefan Hajnoczi <stefanha@redhat.com>
To: "Lluís Vilanova" <vilanova@ac.upc.edu>
Cc: qemu-devel@nongnu.org, Eric Blake <eblake@redhat.com>,
Eduardo Habkost <ehabkost@redhat.com>,
Paolo Bonzini <pbonzini@redhat.com>,
Peter Crosthwaite <crosthwaite.peter@gmail.com>,
Richard Henderson <rth@twiddle.net>
Subject: Re: [Qemu-devel] [PATCH v6 3/7] trace: [tcg] Delay changes to dynamic state when translating
Date: Mon, 9 Jan 2017 17:01:56 +0000 [thread overview]
Message-ID: <20170109170156.GL30228@stefanha-x1.localdomain> (raw)
In-Reply-To: <148295047061.19871.11792107348459066542.stgit@fimbulvetr.bsc.es>
[-- Attachment #1: Type: text/plain, Size: 7601 bytes --]
On Wed, Dec 28, 2016 at 07:41:10PM +0100, Lluís Vilanova wrote:
> This keeps consistency across all decisions taken during translation
> when the dynamic state of a vCPU is changed in the middle of translating
> some guest code.
>
> Signed-off-by: Lluís Vilanova <vilanova@ac.upc.edu>
> ---
> cpu-exec.c | 26 ++++++++++++++++++++++++++
> include/qom/cpu.h | 7 +++++++
> qom/cpu.c | 4 ++++
> trace/control-target.c | 11 +++++++++--
> 4 files changed, 46 insertions(+), 2 deletions(-)
>
> diff --git a/cpu-exec.c b/cpu-exec.c
> index 4188fed3c6..1b7366efb0 100644
> --- a/cpu-exec.c
> +++ b/cpu-exec.c
> @@ -33,6 +33,7 @@
> #include "hw/i386/apic.h"
> #endif
> #include "sysemu/replay.h"
> +#include "trace/control.h"
>
> /* -icount align implementation. */
>
> @@ -451,9 +452,21 @@ static inline bool cpu_handle_exception(CPUState *cpu, int *ret)
> #ifndef CONFIG_USER_ONLY
> } else if (replay_has_exception()
> && cpu->icount_decr.u16.low + cpu->icount_extra == 0) {
> + /* delay changes to this vCPU's dstate during translation */
> + atomic_set(&cpu->trace_dstate_delayed_req, false);
> + atomic_set(&cpu->trace_dstate_must_delay, true);
> +
> /* try to cause an exception pending in the log */
> cpu_exec_nocache(cpu, 1, tb_find(cpu, NULL, 0), true);
> *ret = -1;
> +
> + /* apply and disable delayed dstate changes */
> + atomic_set(&cpu->trace_dstate_must_delay, false);
> + if (unlikely(atomic_read(&cpu->trace_dstate_delayed_req))) {
> + bitmap_copy(cpu->trace_dstate, cpu->trace_dstate_delayed,
> + trace_get_vcpu_event_count());
> + }
> +
> return true;
> #endif
> }
> @@ -634,8 +647,21 @@ int cpu_exec(CPUState *cpu)
>
> for(;;) {
> cpu_handle_interrupt(cpu, &last_tb);
> +
> + /* delay changes to this vCPU's dstate during translation */
> + atomic_set(&cpu->trace_dstate_delayed_req, false);
> + atomic_set(&cpu->trace_dstate_must_delay, true);
> +
> tb = tb_find(cpu, last_tb, tb_exit);
> cpu_loop_exec_tb(cpu, tb, &last_tb, &tb_exit, &sc);
> +
> + /* apply and disable delayed dstate changes */
> + atomic_set(&cpu->trace_dstate_must_delay, false);
> + if (unlikely(atomic_read(&cpu->trace_dstate_delayed_req))) {
> + bitmap_copy(cpu->trace_dstate, cpu->trace_dstate_delayed,
> + trace_get_vcpu_event_count());
> + }
> +
> /* Try to align the host and virtual clocks
> if the guest is in advance */
> align_clocks(&sc, cpu);
> diff --git a/include/qom/cpu.h b/include/qom/cpu.h
> index 3f79a8e955..58255d06fa 100644
> --- a/include/qom/cpu.h
> +++ b/include/qom/cpu.h
> @@ -295,6 +295,10 @@ struct qemu_work_item;
> * @kvm_fd: vCPU file descriptor for KVM.
> * @work_mutex: Lock to prevent multiple access to queued_work_*.
> * @queued_work_first: First asynchronous work pending.
> + * @trace_dstate_must_delay: Whether a change to trace_dstate must be delayed.
> + * @trace_dstate_delayed_req: Whether a change to trace_dstate was delayed.
> + * @trace_dstate_delayed: Delayed changes to trace_dstate (includes all changes
> + * to @trace_dstate).
> * @trace_dstate: Dynamic tracing state of events for this vCPU (bitmask).
> *
> * State of one CPU core or thread.
> @@ -370,6 +374,9 @@ struct CPUState {
> * Dynamically allocated based on bitmap requried to hold up to
> * trace_get_vcpu_event_count() entries.
> */
> + bool trace_dstate_must_delay;
> + bool trace_dstate_delayed_req;
> + unsigned long *trace_dstate_delayed;
> unsigned long *trace_dstate;
>
> /* TODO Move common fields from CPUArchState here. */
> diff --git a/qom/cpu.c b/qom/cpu.c
> index 03d9190f8c..d56496d28d 100644
> --- a/qom/cpu.c
> +++ b/qom/cpu.c
> @@ -367,6 +367,9 @@ static void cpu_common_initfn(Object *obj)
> QTAILQ_INIT(&cpu->breakpoints);
> QTAILQ_INIT(&cpu->watchpoints);
>
> + cpu->trace_dstate_must_delay = false;
> + cpu->trace_dstate_delayed_req = false;
> + cpu->trace_dstate_delayed = bitmap_new(trace_get_vcpu_event_count());
> cpu->trace_dstate = bitmap_new(trace_get_vcpu_event_count());
>
> cpu_exec_initfn(cpu);
> @@ -375,6 +378,7 @@ static void cpu_common_initfn(Object *obj)
> static void cpu_common_finalize(Object *obj)
> {
> CPUState *cpu = CPU(obj);
> + g_free(cpu->trace_dstate_delayed);
> g_free(cpu->trace_dstate);
> }
>
> diff --git a/trace/control-target.c b/trace/control-target.c
> index 7ebf6e0bcb..aba8db55de 100644
> --- a/trace/control-target.c
> +++ b/trace/control-target.c
> @@ -69,13 +69,20 @@ void trace_event_set_vcpu_state_dynamic(CPUState *vcpu,
> if (state_pre != state) {
> if (state) {
> trace_events_enabled_count++;
> - set_bit(vcpu_id, vcpu->trace_dstate);
> + set_bit(vcpu_id, vcpu->trace_dstate_delayed);
> + if (!atomic_read(&vcpu->trace_dstate_must_delay)) {
> + set_bit(vcpu_id, vcpu->trace_dstate);
> + }
> (*ev->dstate)++;
> } else {
> trace_events_enabled_count--;
> - clear_bit(vcpu_id, vcpu->trace_dstate);
> + clear_bit(vcpu_id, vcpu->trace_dstate_delayed);
> + if (!atomic_read(&vcpu->trace_dstate_must_delay)) {
> + clear_bit(vcpu_id, vcpu->trace_dstate);
> + }
> (*ev->dstate)--;
> }
> + atomic_set(&vcpu->trace_dstate_delayed_req, true);
> }
> }
This lock-free scheme looks broken to me. Consider the following case
with threads A and B:
A: atomic_set(&cpu->trace_dstate_delayed_req, false);
A: atomic_set(&cpu->trace_dstate_must_delay, true);
B: if (!atomic_read(&vcpu->trace_dstate_must_delay)) { /* false */
A: atomic_set(&cpu->trace_dstate_must_delay, false);
A: if (unlikely(atomic_read(&cpu->trace_dstate_delayed_req))) { /* false */
B: atomic_set(&vcpu->trace_dstate_delayed_req, true);
Oops, we missed the delayed update.
Now when A runs the next iteration we forget there was a delayed req:
A: atomic_set(&cpu->trace_dstate_delayed_req, false);
As a result even the next iteration may not copy the delayed bitmap.
Perhaps you should use RCU.
Or use a simpler scheme:
struct CPUState {
...
uint32_t dstate_update_count;
};
In trace_event_set_vcpu_state_dynamic():
if (state) {
trace_events_enabled_count++;
set_bit(vcpu_id, vcpu->trace_dstate_delayed);
atomic_inc(&vcpu->dstate_update_count, 1);
(*ev->dstate)++;
} ...
In cpu_exec() and friends:
last_dstate_update_count = atomic_read(&vcpu->dstate_update_count);
tb = tb_find(cpu, last_tb, tb_exit);
cpu_loop_exec_tb(cpu, tb, &last_tb, &tb_exit, &sc);
/* apply and disable delayed dstate changes */
if (unlikely(atomic_read(&cpu->dstate_update_count) != last_dstate_update_count)) {
bitmap_copy(cpu->trace_dstate, cpu->trace_dstate_delayed,
trace_get_vcpu_event_count());
}
(You'll need to adjust the details but the update counter approach
should be workable.)
Stefan
[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 455 bytes --]
next prev parent reply other threads:[~2017-01-10 14:43 UTC|newest]
Thread overview: 21+ messages / expand[flat|nested] mbox.gz Atom feed top
2016-12-28 18:40 [Qemu-devel] [PATCH v6 0/7] trace: [tcg] Optimize per-vCPU tracing states with separate TB caches Lluís Vilanova
2016-12-28 18:41 ` [Qemu-devel] [PATCH v6 1/7] exec: [tcg] Refactor flush of per-CPU virtual TB cache Lluís Vilanova
2017-01-10 20:07 ` Richard Henderson
2016-12-28 18:41 ` [Qemu-devel] [PATCH v6 2/7] trace: Make trace_get_vcpu_event_count() inlinable Lluís Vilanova
2017-01-10 20:08 ` Richard Henderson
2017-01-12 18:14 ` Lluís Vilanova
2016-12-28 18:41 ` [Qemu-devel] [PATCH v6 3/7] trace: [tcg] Delay changes to dynamic state when translating Lluís Vilanova
2017-01-09 17:01 ` Stefan Hajnoczi [this message]
2017-01-10 16:31 ` Paolo Bonzini
2017-01-11 16:16 ` Stefan Hajnoczi
2017-01-12 19:37 ` Lluís Vilanova
2017-01-12 21:25 ` Paolo Bonzini
2017-01-13 20:08 ` Lluís Vilanova
2016-12-28 18:41 ` [Qemu-devel] [PATCH v6 4/7] exec: [tcg] Use different TBs according to the vCPU's dynamic tracing state Lluís Vilanova
2017-01-10 20:10 ` Richard Henderson
2016-12-28 18:41 ` [Qemu-devel] [PATCH v6 5/7] trace: [tcg] Do not generate TCG code to trace dinamically-disabled events Lluís Vilanova
2016-12-28 18:41 ` [Qemu-devel] [PATCH v6 6/7] trace: [tcg, trivial] Re-align generated code Lluís Vilanova
2017-01-12 11:19 ` Michael Tokarev
2017-01-12 18:46 ` Lluís Vilanova
2016-12-28 18:41 ` [Qemu-devel] [PATCH v6 7/7] trace: [trivial] Statically enable all guest events Lluís Vilanova
2017-01-09 17:04 ` [Qemu-devel] [PATCH v6 0/7] trace: [tcg] Optimize per-vCPU tracing states with separate TB caches Stefan Hajnoczi
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=20170109170156.GL30228@stefanha-x1.localdomain \
--to=stefanha@redhat.com \
--cc=crosthwaite.peter@gmail.com \
--cc=eblake@redhat.com \
--cc=ehabkost@redhat.com \
--cc=pbonzini@redhat.com \
--cc=qemu-devel@nongnu.org \
--cc=rth@twiddle.net \
--cc=vilanova@ac.upc.edu \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).