accel: Only include qdev-realized vCPUs in global &cpus_queue
cover accel: Only include qdev-realized vCPUs in global &cpus_queue
Hi,
The goal of this series is to expose vCPUs in a stable state to the accelerators, in particular the QDev 'REALIZED' step.
To do so we split the QTAILQ_INSERT/REMOVE calls from cpu_list_add() / cpu_list_remove(), by moving them to the DeviceClass::[un]wire() handlers, guaranty to be called just before a vCPU is exposed to the guest, as "realized".
First we have to modify how &first_cpu is used in TCG round robin implementation, and ensure we invalidate the TB jmpcache with &qemu_cpu_list locked.
I'm really out of my comfort zone here, so posting as RFC. At least all test suite is passing...
I expect these changes to allow CPUState::cpu_index clarifications and simplifications, but this will be addressed (and commented) in a separate series.
Regards,
Phil.
Philippe Mathieu-Daudé (9):
accel/tcg: Simplify use of &first_cpu in rr_cpu_thread_fn() accel/tcg: Invalidate TB jump cache with global vCPU queue locked
cpus: Remove cpu from global queue after UNREALIZE completed hw/qdev: Introduce DeviceClass::[un]wire() handlers
cpus: Add DeviceClass::[un]wire() stubs
cpus: Call hotplug handlers in DeviceWire()
cpus: Only expose REALIZED vCPUs to global &cpus_queue accel/kvm: Assert vCPU is created when calling kvm_dirty_ring_reap*() accel/kvm: Remove unreachable assertion in kvm_dirty_ring_reap*()
include/hw/qdev-core.h | 7 +++++++ accel/kvm/kvm-all.c | 9 --------- accel/tcg/tb-maint.c | 2 ++ accel/tcg/tcg-accel-ops-rr.c | 15 ++++++++------- cpu-common.c | 2 -- cpu-target.c | 7 ++----- hw/core/cpu-common.c | 18 +++++++++++++++++- hw/core/qdev.c | 20 +++++++++++++++++++- 8 files changed, 55 insertions(+), 25 deletions(-)
Signature · 1 line
2.47.1
RFC PATCH 1/9 accel/tcg: Simplify use of &first_cpu in rr_cpu_thread_fn()
Let vCPUs wait for themselves being ready first, then other ones. This allows the first thread to starts without the global vcpu queue (thus &first_cpu) being populated.
--- accel/tcg/tcg-accel-ops-rr.c | 15 ++++++++------- 1 file changed, 8 insertions(+), 7 deletions(-)
accel/tcg/tcg-accel-ops-rr.c +8 -7
@@ -197,20 +197,21 @@ static void *rr_cpu_thread_fn(void *arg)qemu_guest_random_seed_thread_part2 (cpu->random_seed);/* wait for initial kick-off after machine start */ -while (first_cpu->stopped) { -qemu_cond_wait_bql (first_cpu->halt_cond); +while (cpu->stopped) { +CPUState *iter_cpu; + +qemu_cond_wait_bql (cpu->halt_cond);/* process any pending work */ -CPU_FOREACH (cpu) { - current_cpu = cpu; -qemu_wait_io_event_common (cpu); +CPU_FOREACH (iter_cpu) { + current_cpu = iter_cpu; +qemu_wait_io_event_common (iter_cpu); } } +g_assert (first_cpu);rr_start_kick_timer (); - cpu = first_cpu; -/* process any pending work */ cpu->exit_request =1 ;
Signature · 1 line
2.47.1
Re: [RFC PATCH 1/9] accel/tcg: Simplify use of &first_cpu in rr_cpu_thread_fn()
Quoted 28 lines · Philippe Mathieu-Daudé
Let vCPUs wait for themselves being ready first, then other ones. This allows the first thread to starts without the global vcpu queue (thus
&first_cpu) being populated.Signed-off-by Philippe Mathieu-Daudé <[email protected]>--- accel/tcg/tcg-accel-ops-rr.c | 15 ++++++++------- 1 file changed, 8 insertions(+), 7 deletions(-)accel/tcg/tcg-accel-ops-rr.c +4
-5@@ -197,20 +197,21 @@ static void *rr_cpu_thread_fn(void *arg)qemu_guest_random_seed_thread_part2 (cpu->random_seed);/* wait for initial kick-off after machine start */ -while (first_cpu->stopped) { -qemu_cond_wait_bql (first_cpu->halt_cond); +while (cpu->stopped) { +CPUState *iter_cpu; + +qemu_cond_wait_bql (cpu->halt_cond);/* process any pending work */ -CPU_FOREACH (cpu) { - current_cpu = cpu; -qemu_wait_io_event_common (cpu);
accel/tcg/tcg-accel-ops-rr.c static void *rr_cpu_thread_fn(void *arg) · line 212
Quoted 10 lines
accel/tcg/tcg-accel-ops-rr.c +4
-1+CPU_FOREACH (iter_cpu) { + current_cpu = iter_cpu; +qemu_wait_io_event_common (iter_cpu); } } +g_assert (first_cpu);rr_start_kick_timer (); - cpu = first_cpu;
This final line seems to be unrelated.
I'm not saying it's wrong, but it's certainly a change in behaviour. We no longer execute from first_cpu first.
r~
RFC PATCH 2/9 accel/tcg: Invalidate TB jump cache with global vCPU queue locked
In reply to Philippe Mathieu-DaudéInvalidate TB with global vCPU queue locked.
See commit 4731f89b3b9 ("cpu: free cpu->tb_jmp_cache with RCU"):
Fixes the appended use-after-free. The root cause is that during tb invalidation we use CPU_FOREACH, and therefore to safely free a vCPU we must wait for an RCU grace period to elapse.
--- accel/tcg/tb-maint.c | 2 ++ 1 file changed, 2 insertions(+)
accel/tcg/tb-maint.c +2 -0
@@ -891,6 +891,8 @@ static void tb_jmp_cache_inval_tb(TranslationBlock *tb) }else {uint32_t h =tb_jmp_cache_hash_func (tb->pc); +QEMU_LOCK_GUARD (&qemu_cpu_list_lock); +CPU_FOREACH (cpu) {CPUJumpCache *jc = cpu->tb_jmp_cache;
Signature · 1 line
2.47.1
Re: [RFC PATCH 2/9] accel/tcg: Invalidate TB jump cache with global vCPU queue locked
Quoted 19 lines · Philippe Mathieu-Daudé
Invalidate TB with global vCPU queue locked.
See commit
4731f89b3b9("cpu: freecpu->tb_jmp_cachewith RCU"):Fixes the appended use-after-free. The root cause is that during tb invalidation we use
CPU_FOREACH, and therefore to safely free a vCPU we must wait for an RCU grace period to elapse.Signed-off-by Philippe Mathieu-Daudé <[email protected]>--- accel/tcg/tb-maint.c | 2 ++ 1 file changed, 2 insertions(+)accel/tcg/tb-maint.c +0
-0
accel/tcg/tb-maint.c static void tb_jmp_cache_inval_tb(TranslationBlock *tb) · line 898
Quoted 8 lines
accel/tcg/tb-maint.c +2
-0@@ -891,6 +891,8 @@ static void tb_jmp_cache_inval_tb(TranslationBlock *tb) }else {uint32_t h =tb_jmp_cache_hash_func (tb->pc); +QEMU_LOCK_GUARD (&qemu_cpu_list_lock); +CPU_FOREACH (cpu) {CPUJumpCache *jc = cpu->tb_jmp_cache;
I can see how maybe this can appear to fix the bug, because one can't remove cpus at all while the lock is held. But if the description is accurate that this is RCU related, then the proper locking is with rcu_read_lock/rcu_read_unlock.
r~
RFC PATCH 3/9 cpus: Remove cpu from global queue after UNREALIZE completed
In reply to Philippe Mathieu-DaudéPrevious commit removed the restriction on completing the full QDev UNREALIZE step before removing vCPUs from global queue, it is now safe to call cpu_list_remove() after accel_cpu_common_unrealize().
--- cpu-target.c | 7 ++----- 1 file changed, 2 insertions(+), 5 deletions(-)
cpu-target.c +2 -5
@@ -172,12 +172,9 @@ void cpu_exec_unrealizefn(CPUState *cpu) }#endif -cpu_list_remove (cpu); -/* -* Now that the vCPU has been removed from the RCU list, we can call -* accel_cpu_common_unrealize, which may free fields using call_rcu. -*/ accel_cpu_common_unrealize (cpu); + +cpu_list_remove (cpu); }/*
Signature · 1 line
2.47.1
Re: [RFC PATCH 3/9] cpus: Remove cpu from global queue after UNREALIZE completed
Quoted 16 lines · Philippe Mathieu-Daudé
Previous commit removed the restriction on completing the full QDev UNREALIZE step before removing vCPUs from global queue, it is now safe to call
cpu_list_remove()afteraccel_cpu_common_unrealize().Signed-off-by Philippe Mathieu-Daudé <[email protected]>--- cpu-target.c | 7 ++----- 1 file changed, 2 insertions(+), 5 deletions(-)cpu-target.c +0
-0@@ -172,12 +172,9 @@ void cpu_exec_unrealizefn(CPUState *cpu) }#endif
cpu-target.c void cpu_exec_unrealizefn(CPUState *cpu) · line 178
Quoted 10 lines
cpu-target.c +2
-5-cpu_list_remove (cpu); -/* -* Now that the vCPU has been removed from the RCU list, we can call -* accel_cpu_common_unrealize, which may free fields using call_rcu. -*/ accel_cpu_common_unrealize (cpu); + +cpu_list_remove (cpu); }
I don't believe this is correct. Why would we have an unrealized cpu on the list? What's wrong with removing the cpu from the list before unrealize?
r~
RFC PATCH 4/9 hw/qdev: Introduce DeviceClass::[un]wire() handlers
In reply to Philippe Mathieu-DaudéWe are trying to understand what means "a qdev is realized". One explanation was "the device is guest visible"; however many devices are realized before being mapped, thus are not "guest visible". Some devices map / wire their IRQs before being realized (such ISA devices). There is a need for devices to be "automatically" mapped/wired (see [2]) such CLI-created devices, but this apply generically to dynamic machines.
Currently the device creation steps are expected to roughly be:
(external use) (QDev core) (Device Impl)
~~~~~~~~~~~~ ~~~~~~~~~ ~~~~~~~~~~~
INIT enter
----->
+----------------------+
| Allocate state |
+----------------------+
----->
+---------------------+
| INIT children |
| |
| Alias children properties
| |
| Expose properties |
INIT exit +---------------------+
<-----------------------------------
+----------------+
| set properties |
| |
| set ClkIn |
+----------------+ REALIZE enter
---------------------------------->
+----------------------+
| Use config properties|
| |
| Realize children |
| |
| Init GPIOs/IRQs |
| |
| Init MemoryRegions |
+----------------------+
REALIZE exit
<----------------------------------- ---- "realized" / "guest visible"
+-----------------+
| Explicit wiring:|
| IRQs |
| I/O / Mem |
| ClkOut |
+-----------------+ RESET enter
--------------------------------->
+----------------------+
| Reset default values |
+----------------------+But as mentioned, various devices "wire" parts before they exit the "realize" step.
In order to clarify, I'm trying to enforce what can be done *before* and *after* realization.
*after* a device is expected to be stable (no more configurable) and fully usable.
To be able to use internal/auto wiring (such ISA devices) and keep the current external/explicit wiring, I propose to add an extra "internal wiring" step, happening after the REALIZE step as:
(external use) (QDev core) (Device Impl)
~~~~~~~~~~~~ ~~~~~~~~~ ~~~~~~~~~~~
INIT enter
----->
+----------------------+
| Allocate state |
+----------------------+
----->
+---------------------+
| INIT children |
| |
| Alias children properties
| |
| Expose properties |
INIT exit +---------------------+
<-----------------------------------
+----------------+
| set properties |
| |
| set ClkIn |
+----------------+ REALIZE enter
---------------------------------->
+----------------------+
| Use config properties|
| |
| Realize children |
| |
| Init GPIOs/IRQs |
| |
| Init MemoryRegions |
+----------------------+
REALIZE exit <---
+----------------------+
| Internal auto wiring |
| IRQs | (i.e. ISA bus)
| I/O / Mem |
| ClkOut |
+----------------------+
<--- ---- "realized"
+-----------------+
| External wiring:|
| IRQs |
| I/O / Mem |
| ClkOut |
+-----------------+ RESET enter ---- "guest visible"
--------------------------------->
+----------------------+
| Reset default values |
+----------------------+The "realized" point is not changed. "guest visible" concept only occurs *after* wiring, just before the reset phase.
This change introduces the DeviceClass::wire handler within qdev core realization code.
--- include/hw/qdev-core.h | 7 +++++++ hw/core/qdev.c | 20 +++++++++++++++++++- 2 files changed, 26 insertions(+), 1 deletion(-)
include/hw/qdev-core.h +7 -0
@@ -102,7 +102,12 @@ typedef int (*DeviceSyncConfig)(DeviceState *dev, Error **errp);* @props: Properties accessing state fields. * @realize: Callback function invoked when the #DeviceState:realized * property is changed to %true. +* @wire: Callback function called after @realize to connect IRQs, +* clocks and map memories. Can not fail. +* @unwire: Callback function to undo @wire. Called before @unrealize. +* Can not fail. * @unrealize: Callback function invoked when the #DeviceState:realized +* property is changed to %false. Can not fail. * property is changed to %false. * @sync_config: Callback function invoked when QMP command device-sync-config * is called. Should synchronize device configuration from host to guest part
@@ -171,6 +176,8 @@ struct DeviceClass {*/ DeviceReset legacy_reset;DeviceRealize realize; +void (*wire)(DeviceState *dev); +void (*unwire)(DeviceState *dev);DeviceUnrealize unrealize;DeviceSyncConfig sync_config;
hw/core/qdev.c +19 -1
@@ -554,6 +554,15 @@ static void device_set_realized(Object *obj, bool value, Error **errp) } } +if (dc->wire) { +if (!dc->unwire) { +warn_report_once ("wire() without unwire() for type '%s'" , +object_get_typename (OBJECT (dev))); + } + dc->wire (dev); + } + +/* At this point the device is "guest visible". */ qatomic_store_release (&dev->realized, value); }else if (!value && dev->realized) {
@@ -573,6 +582,15 @@ static void device_set_realized(Object *obj, bool value, Error **errp)*/ smp_wmb (); +if (dc->unwire) { +if (!dc->wire) { +error_report ("disconnect() without connect() for type '%s'" , +object_get_typename (OBJECT (dev))); +abort (); + } + dc->unwire (dev); + } + QLIST_FOREACH(bus, &dev->child_bus, sibling) {qbus_unrealize (bus); }
@@ -585,8 +603,8 @@ static void device_set_realized(Object *obj, bool value, Error **errp) dev->pending_deleted_event = true;DEVICE_LISTENER_CALL (unrealize, Reverse, dev); } -assert (local_err ==NULL ); +return ; child_realize_fail:
Signature · 1 line
2.47.1
Re: [RFC PATCH 4/9] hw/qdev: Introduce DeviceClass::[un]wire() handlers
Quoted 179 lines · Philippe Mathieu-Daudé
We are trying to understand what means "a qdev is realized". One explanation was "the device is guest visible"; however many devices are realized before being mapped, thus are not "guest visible". Some devices map / wire their IRQs before being realized (such ISA devices). There is a need for devices to be "automatically" mapped/wired (see [2]) such CLI-created devices, but this apply generically to dynamic machines.
Currently the device creation steps are expected to roughly be:
(external use) (QDev core) (Device Impl) ~~~~~~~~~~~~ ~~~~~~~~~ ~~~~~~~~~~~ INIT enter -----> +----------------------+ | Allocate state | +----------------------+ -----> +---------------------+ | INIT children | | | | Alias children properties | | | Expose properties | INIT exit +---------------------+ <----------------------------------- +----------------+ | set properties | | | | set ClkIn | +----------------+ REALIZE enter ----------------------------------> +----------------------+ | Use config properties| | | | Realize children | | | | Init GPIOs/IRQs | | | | Init MemoryRegions | +----------------------+ REALIZE exit <----------------------------------- ---- "realized" / "guest visible" +-----------------+ | Explicit wiring:| | IRQs | | I/O / Mem | | ClkOut | +-----------------+ RESET enter ---------------------------------> +----------------------+ | Reset default values | +----------------------+But as mentioned, various devices "wire" parts before they exit the "realize" step.
In order to clarify, I'm trying to enforce what can be done *before* and *after* realization.*after* a device is expected to be stable (no more configurable) and fully usable.
To be able to use internal/auto wiring (such ISA devices) and keep the current external/explicit wiring, I propose to add an extra "internal wiring" step, happening after the REALIZE step as:
(external use) (QDev core) (Device Impl) ~~~~~~~~~~~~ ~~~~~~~~~ ~~~~~~~~~~~ INIT enter -----> +----------------------+ | Allocate state | +----------------------+ -----> +---------------------+ | INIT children | | | | Alias children properties | | | Expose properties | INIT exit +---------------------+ <----------------------------------- +----------------+ | set properties | | | | set ClkIn | +----------------+ REALIZE enter ----------------------------------> +----------------------+ | Use config properties| | | | Realize children | | | | Init GPIOs/IRQs | | | | Init MemoryRegions | +----------------------+ REALIZE exit <--- +----------------------+ | Internal auto wiring | | IRQs | (i.e. ISA bus) | I/O / Mem | | ClkOut | +----------------------+ <--- ---- "realized" +-----------------+ | External wiring:| | IRQs | | I/O / Mem | | ClkOut | +-----------------+ RESET enter ---- "guest visible" ---------------------------------> +----------------------+ | Reset default values | +----------------------+The "realized" point is not changed. "guest visible" concept only occurs *after* wiring, just before the reset phase.
This change introduces the
DeviceClass::wirehandler within qdev core realization code.Signed-off-by Philippe Mathieu-Daudé <[email protected]>--- include/hw/qdev-core.h | 7 +++++++ hw/core/qdev.c | 20 +++++++++++++++++++- 2 files changed, 26 insertions(+), 1 deletion(-)include/hw/qdev-core.h +7
-0@@ -102,7 +102,12 @@ typedef int (*DeviceSyncConfig)(DeviceState *dev, Error **errp);* @props: Properties accessing state fields. * @realize: Callback function invoked when the #DeviceState:realized * property is changed to %true. +* @wire: Callback function called after @realize to connect IRQs, +* clocks and map memories. Can not fail. +* @unwire: Callback function to undo @wire. Called before @unrealize. +* Can not fail. * @unrealize: Callback function invoked when the #DeviceState:realized +* property is changed to %false. Can not fail. * property is changed to %false. * @sync_config: Callback function invoked when QMP command device-sync-config * is called. Should synchronize device configuration from host to guest part @@ -171,6 +176,8 @@ struct DeviceClass {*/ DeviceReset legacy_reset;DeviceRealize realize; +void (*wire)(DeviceState *dev); +void (*unwire)(DeviceState *dev);DeviceUnrealize unrealize;DeviceSyncConfig sync_config;hw/core/qdev.c +9
-0@@ -554,6 +554,15 @@ static void device_set_realized(Object *obj, bool value, Error **errp) } } +if (dc->wire) { +if (!dc->unwire) { +warn_report_once ("wire() without unwire() for type '%s'" , +object_get_typename (OBJECT (dev))); + } + dc->wire (dev); + } + +/* At this point the device is "guest visible". */ qatomic_store_release (&dev->realized, value); }else if (!value && dev->realized) {@@ -573,6 +582,15 @@ static void device_set_realized(Object *obj, bool value, Error **errp)*/
hw/core/qdev.c static void device_set_realized(Object *obj, bool value, Error **errp) · line 592
Quoted 10 lines
hw/core/qdev.c +8
-0smp_wmb (); +if (dc->unwire) { +if (!dc->wire) { +error_report ("disconnect() without connect() for type '%s'" , +object_get_typename (OBJECT (dev))); +abort (); + } + dc->unwire (dev); + }
Mismatched error messages (wire vs connect).
But, really, just check both directions properly at startup. There's probably lots of places where devices are never unrealized.
Otherwise, this seems sane, having a kind of post_init on the realize path, running after all superclass realization is done.
r~
Re: [RFC PATCH 4/9] hw/qdev: Introduce DeviceClass::[un]wire() handlers
Quoted 61 lines · On Tue, 28 Jan 2025 12:52:39 -0800 Richard Henderson…
Quoted 60 lines · Philippe Mathieu-Daudé
We are trying to understand what means "a qdev is realized". One explanation was "the device is guest visible"; however many devices are realized before being mapped, thus are not "guest visible". Some devices map / wire their IRQs before being realized (such ISA devices). There is a need for devices to be "automatically" mapped/wired (see [2]) such CLI-created devices, but this apply generically to dynamic machines.
Currently the device creation steps are expected to roughly be:
(external use) (QDev core) (Device Impl) ~~~~~~~~~~~~ ~~~~~~~~~ ~~~~~~~~~~~ INIT enter -----> +----------------------+ | Allocate state | +----------------------+ -----> +---------------------+ | INIT children | | | | Alias children properties | | | Expose properties | INIT exit +---------------------+ <----------------------------------- +----------------+ | set properties | | | | set ClkIn | +----------------+ REALIZE enter ----------------------------------> +----------------------+ | Use config properties| | | | Realize children | | | | Init GPIOs/IRQs | | | | Init MemoryRegions | +----------------------+ REALIZE exit <----------------------------------- ---- "realized" / "guest visible" +-----------------+ | Explicit wiring:| | IRQs | | I/O / Mem | | ClkOut | +-----------------+ RESET enter ---------------------------------> +----------------------+ | Reset default values | +----------------------+But as mentioned, various devices "wire" parts before they exit the "realize" step.
In order to clarify, I'm trying to enforce what can be done *before* and *after* realization.
Quoted 6 lines
Quoted 6 lines
*after* a device is expected to be stable (no more configurable) and fully usable.
To be able to use internal/auto wiring (such ISA devices) and keep the current external/explicit wiring, I propose to add an extra "internal wiring" step, happening after the REALIZE step
we partially have such wiring in place
hotplug_handler_plug
'hotplug_' prefix is legacy remnant to confuse every, it really should be renamed to just plug_handler or 'wire_foo'.
For bus-less CPUs it's used both on hot- and cold-plug paths.
It does take care of wiring after realize part is completed (basically exposing device to the external users).
And I it also handles 'bus' based plug workflow
see qdev_get_hotplug_handler() assuming bus owner has provided a callback. It's likely no-one had cared about ISA, as it was out of hotplug scope when the interface was introduced.
Unplug part is however is not wired directly into unrealize() stage, but rather to an external caller (mgmt or guest), which essentially unwinds into the same hotplug_handler chain as plug, to unwire device and then the same external caller does call unrealize on device. (see:hotplug_handler_unplug() )
As for more direct hack, there is also DEVICE_LISTENER_CALL() callback, which is more close to this approach (modulo users are hidden as there is not explicit wiring).
Unrealize part is called too late (which probably is not right), and realize part is called before subtree is realized (might work for current users, but likely should be placed at the same site as wire() call.
Hence we should think if already existing methods could be reused/adapted to, before adding similar wire/unwire mechanism.
Quoted 134 lines
Quoted 113 lines
as:
(external use) (QDev core) (Device Impl) ~~~~~~~~~~~~ ~~~~~~~~~ ~~~~~~~~~~~ INIT enter -----> +----------------------+ | Allocate state | +----------------------+ -----> +---------------------+ | INIT children | | | | Alias children properties | | | Expose properties | INIT exit +---------------------+ <----------------------------------- +----------------+ | set properties | | | | set ClkIn | +----------------+ REALIZE enter ----------------------------------> +----------------------+ | Use config properties| | | | Realize children | | | | Init GPIOs/IRQs | | | | Init MemoryRegions | +----------------------+ REALIZE exit <--- +----------------------+ | Internal auto wiring | | IRQs | (i.e. ISA bus) | I/O / Mem | | ClkOut | +----------------------+ <--- ---- "realized" +-----------------+ | External wiring:| | IRQs | | I/O / Mem | | ClkOut | +-----------------+ RESET enter ---- "guest visible" ---------------------------------> +----------------------+ | Reset default values | +----------------------+The "realized" point is not changed. "guest visible" concept only occurs *after* wiring, just before the reset phase.
This change introduces the
DeviceClass::wirehandler within qdev core realization code.Signed-off-by Philippe Mathieu-Daudé <[email protected]>--- include/hw/qdev-core.h | 7 +++++++ hw/core/qdev.c | 20 +++++++++++++++++++- 2 files changed, 26 insertions(+), 1 deletion(-)include/hw/qdev-core.h +7
-0@@ -102,7 +102,12 @@ typedef int (*DeviceSyncConfig)(DeviceState *dev, Error **errp);* @props: Properties accessing state fields. * @realize: Callback function invoked when the #DeviceState:realized * property is changed to %true. +* @wire: Callback function called after @realize to connect IRQs, +* clocks and map memories. Can not fail. +* @unwire: Callback function to undo @wire. Called before @unrealize. +* Can not fail. * @unrealize: Callback function invoked when the #DeviceState:realized +* property is changed to %false. Can not fail. * property is changed to %false. * @sync_config: Callback function invoked when QMP command device-sync-config * is called. Should synchronize device configuration from host to guest part @@ -171,6 +176,8 @@ struct DeviceClass {*/ DeviceReset legacy_reset;DeviceRealize realize; +void (*wire)(DeviceState *dev); +void (*unwire)(DeviceState *dev);DeviceUnrealize unrealize;DeviceSyncConfig sync_config;hw/core/qdev.c +9
-0@@ -554,6 +554,15 @@ static void device_set_realized(Object *obj, bool value, Error **errp) } } +if (dc->wire) { +if (!dc->unwire) { +warn_report_once ("wire() without unwire() for type '%s'" , +object_get_typename (OBJECT (dev))); + } + dc->wire (dev); + } + +/* At this point the device is "guest visible". */ qatomic_store_release (&dev->realized, value); }else if (!value && dev->realized) {@@ -573,6 +582,15 @@ static void device_set_realized(Object *obj, bool value, Error **errp)*/ hw/core/qdev.c static void device_set_realized(Object *obj, bool value, Error **errp) · line 592
Quoted 10 lines
hw/core/qdev.c +8
-0smp_wmb (); +if (dc->unwire) { +if (!dc->wire) { +error_report ("disconnect() without connect() for type '%s'" , +object_get_typename (OBJECT (dev))); +abort (); + } + dc->unwire (dev); + }Mismatched error messages (wire vs connect).
But, really, just check both directions properly at startup. There's probably lots of places where devices are never unrealized.Otherwise, this seems sane, having a kind of
post_initon the realize path, running after all superclass realization is done.r~
PATCH 5/9 cpus: Add DeviceClass::[un]wire() stubs
In reply to Philippe Mathieu-Daudé--- hw/core/cpu-common.c | 10 ++++++++++ 1 file changed, 10 insertions(+)
hw/core/cpu-common.c +10 -0
@@ -219,6 +219,14 @@ static void cpu_common_realizefn(DeviceState *dev, Error **errp)/* NOTE: latest generic point where the cpu is fully realized */ } +static void cpu_common_wire (DeviceState *dev) +{ +} + +static void cpu_common_unwire (DeviceState *dev) +{ +} +static void cpu_common_unrealizefn (DeviceState *dev) {CPUState *cpu =CPU (dev);
@@ -311,6 +319,8 @@ static void cpu_common_class_init(ObjectClass *klass, void *data) k->gdb_write_register = cpu_common_gdb_write_register;set_bit (DEVICE_CATEGORY_CPU , dc->categories); dc->realize = cpu_common_realizefn; + dc->wire = cpu_common_wire; + dc->unwire = cpu_common_unwire; dc->unrealize = cpu_common_unrealizefn; rc->phases.hold = cpu_common_reset_hold;cpu_class_init_props (dc);
Signature · 1 line
2.47.1
Re: [PATCH 5/9] cpus: Add DeviceClass::[un]wire() stubs
Quoted 25 lines · Philippe Mathieu-Daudé
Signed-off-by Philippe Mathieu-Daudé <[email protected]>--- hw/core/cpu-common.c | 10 ++++++++++ 1 file changed, 10 insertions(+)hw/core/cpu-common.c +8
-0@@ -219,6 +219,14 @@ static void cpu_common_realizefn(DeviceState *dev, Error **errp)/* NOTE: latest generic point where the cpu is fully realized */ } +static void cpu_common_wire (DeviceState *dev) +{ +} + +static void cpu_common_unwire (DeviceState *dev) +{ +} +static void cpu_common_unrealizefn (DeviceState *dev) {CPUState *cpu =CPU (dev);
hw/core/cpu-common.c static void cpu_common_class_init(ObjectClass *klass, void *data) · line 326
Quoted 8 lines
hw/core/cpu-common.c +2
-0@@ -311,6 +319,8 @@ static void cpu_common_class_init(ObjectClass *klass, void *data) k->gdb_write_register = cpu_common_gdb_write_register;set_bit (DEVICE_CATEGORY_CPU , dc->categories); dc->realize = cpu_common_realizefn; + dc->wire = cpu_common_wire; + dc->unwire = cpu_common_unwire; dc->unrealize = cpu_common_unrealizefn; rc->phases.hold = cpu_common_reset_hold;cpu_class_init_props (dc);
But doesn't need to be split from patch 6.
r~
PATCH 6/9 cpus: Call hotplug handlers in DeviceWire()
In reply to Philippe Mathieu-DaudéTo call the hotplug handlers with REALIZED vCPU, we can use the DeviceWire handler.
--- hw/core/cpu-common.c | 11 ++++++----- 1 file changed, 6 insertions(+), 5 deletions(-)
hw/core/cpu-common.c +6 -5
@@ -211,16 +211,17 @@ static void cpu_common_realizefn(DeviceState *dev, Error **errp) } } -if (dev->hotplugged) { -cpu_synchronize_post_init (cpu); -cpu_resume (cpu); - } -/* NOTE: latest generic point where the cpu is fully realized */ }static void cpu_common_wire (DeviceState *dev) { +CPUState *cpu =CPU (dev); + +if (dev->hotplugged) { +cpu_synchronize_post_init (cpu); +cpu_resume (cpu); + } }static void cpu_common_unwire (DeviceState *dev)
Signature · 1 line
2.47.1
Re: [PATCH 6/9] cpus: Call hotplug handlers in DeviceWire()
Quoted 25 lines · Philippe Mathieu-Daudé
To call the hotplug handlers with REALIZED vCPU, we can use the DeviceWire handler.
Signed-off-by Philippe Mathieu-Daudé <[email protected]>--- hw/core/cpu-common.c | 11 ++++++----- 1 file changed, 6 insertions(+), 5 deletions(-)hw/core/cpu-common.c +0
-5@@ -211,16 +211,17 @@ static void cpu_common_realizefn(DeviceState *dev, Error **errp) } } -if (dev->hotplugged) { -cpu_synchronize_post_init (cpu); -cpu_resume (cpu); - } -/* NOTE: latest generic point where the cpu is fully realized */ }static void cpu_common_wire (DeviceState *dev)
hw/core/cpu-common.c static void cpu_common_realizefn(DeviceState *dev, Error **errp) · line 227
Quoted 10 lines
hw/core/cpu-common.c +6
-0{ +CPUState *cpu =CPU (dev); + +if (dev->hotplugged) { +cpu_synchronize_post_init (cpu); +cpu_resume (cpu); + } }static void cpu_common_unwire (DeviceState *dev)
r~
RFC PATCH 7/9 cpus: Only expose REALIZED vCPUs to global &cpus_queue
In reply to Philippe Mathieu-Daudécpu_list_add() was doing 2 distinct things:
- assign some index to vCPU
- add unrealized (thus in inconsistent state) vcpu to &cpus_queue
Code using CPU_FOREACH() macro would iterate over possibly unrealized vCPUs, often dealt with special casing.
In order to avoid that, we move the addition of vCPU to global queue to the DeviceWire handler, which is called just before switching the vCPU to REALIZED state. This ensure all &cpus_queue users (like via &first_cpu or CPU_FOREACH) get a realized vCPU in consistent state.
Similarly we remove it from the global queue at DeviceUnwire phase, just after marking the vCPU UNREALIZED.
--- cpu-common.c | 2 -- hw/core/cpu-common.c | 5 +++++ 2 files changed, 5 insertions(+), 2 deletions(-)
cpu-common.c +0 -2
@@ -91,7 +91,6 @@ void cpu_list_add(CPUState *cpu) }else {assert (!cpu_index_auto_assigned); } -QTAILQ_INSERT_TAIL_RCU (&cpus_queue, cpu, node); cpu_list_generation_id++; }
@@ -103,7 +102,6 @@ void cpu_list_remove(CPUState *cpu)return ; } -QTAILQ_REMOVE_RCU (&cpus_queue, cpu, node); cpu->cpu_index =UNASSIGNED_CPU_INDEX ; cpu_list_generation_id++; }
hw/core/cpu-common.c +5 -0
@@ -218,6 +218,8 @@ static void cpu_common_wire(DeviceState *dev) {CPUState *cpu =CPU (dev); +QTAILQ_INSERT_TAIL_RCU (&cpus_queue, cpu, node); +if (dev->hotplugged) {cpu_synchronize_post_init (cpu);cpu_resume (cpu);
@@ -226,6 +228,9 @@ static void cpu_common_wire(DeviceState *dev)static void cpu_common_unwire (DeviceState *dev) { +CPUState *cpu =CPU (dev); + +QTAILQ_REMOVE_RCU (&cpus_queue, cpu, node); }static void cpu_common_unrealizefn (DeviceState *dev)
Signature · 1 line
2.47.1
Re: [RFC PATCH 7/9] cpus: Only expose REALIZED vCPUs to global &cpus_queue
Quoted 12 lines · Philippe Mathieu-Daudé
cpu_list_add()was doing 2 distinct things:
- assign some index to vCPU
- add unrealized (thus in inconsistent state) vcpu to&cpus_queueCode using
CPU_FOREACH()macro would iterate over possibly unrealized vCPUs, often dealt with special casing.In order to avoid that, we move the addition of vCPU to global queue to the DeviceWire handler, which is called just before switching the vCPU to REALIZED state. This ensure all
&cpus_queueusers (like via&first_cpuorCPU_FOREACH) get a realized vCPU in consistent state.
Quoted 8 lines
Similarly we remove it from the global queue at DeviceUnwire phase, just after marking the vCPU UNREALIZED.
Signed-off-by Philippe Mathieu-Daudé<[email protected]>--- cpu-common.c | 2 -- hw/core/cpu-common.c | 5 +++++ 2 files changed, 5 insertions(+), 2 deletions(-)
r~
Re: [RFC PATCH 7/9] cpus: Only expose REALIZED vCPUs to global &cpus_queue
In reply to Philippe Mathieu-DaudéQuoted 9 lines · Philippe Mathieu-Daudé
cpu-common.c +0
-1@@ -91,7 +91,6 @@ void cpu_list_add(CPUState *cpu) }else {assert (!cpu_index_auto_assigned); } -QTAILQ_INSERT_TAIL_RCU (&cpus_queue, cpu, node); cpu_list_generation_id++; }
cpu-common.c void cpu_list_remove(CPUState *cpu) · line 107
Quoted 7 lines
cpu-common.c +0
-1@@ -103,7 +102,6 @@ void cpu_list_remove(CPUState *cpu)return ; } -QTAILQ_REMOVE_RCU (&cpus_queue, cpu, node); cpu->cpu_index =UNASSIGNED_CPU_INDEX ; cpu_list_generation_id++; }
We might rename cpu_list_add/remove, since they no longer do what's said on the tin.
r~
PATCH 8/9 accel/kvm: Assert vCPU is created when calling kvm_dirty_ring_reap*()
In reply to Philippe Mathieu-DaudéPrevious commits made sure vCPUs are realized before accelerators (such KVM) use them. Ensure that by asserting the vCPU is created, no need to return.
For more context, see commit 56adee407fc ("kvm: dirty-ring: Fix race with vcpu creation").
--- accel/kvm/kvm-all.c | 8 +++----- 1 file changed, 3 insertions(+), 5 deletions(-)
accel/kvm/kvm-all.c +3 -5
@@ -831,13 +831,11 @@ static uint32_t kvm_dirty_ring_reap_one(KVMState *s, CPUState *cpu)uint32_t count =0 , fetch = cpu->kvm_fetch_index;/* -* It's possible that we race with vcpu creation code where the vcpu is +* It's not possible that we race with vcpu creation code where the vcpu is * put onto the vcpus list but not yet initialized the dirty ring -* structures. If so, skip it. +* structures. */ -if (!cpu->created) { -return 0 ; - } +assert (cpu->created);assert (dirty_gfns && ring_size);trace_kvm_dirty_ring_reap_vcpu (cpu->cpu_index);
Signature · 1 line
2.47.1
Re: [PATCH 8/9] accel/kvm: Assert vCPU is created when calling kvm_dirty_ring_reap*()
Quoted 11 lines · Philippe Mathieu-Daudé
Previous commits made sure vCPUs are realized before accelerators (such KVM) use them. Ensure that by asserting the vCPU is created, no need to return.
For more context, see commit
56adee407fc("kvm: dirty-ring: Fix race with vcpu creation").Signed-off-by Philippe Mathieu-Daudé<[email protected]>--- accel/kvm/kvm-all.c | 8 +++----- 1 file changed, 3 insertions(+), 5 deletions(-)
r~
PATCH 9/9 accel/kvm: Remove unreachable assertion in kvm_dirty_ring_reap*()
In reply to Philippe Mathieu-DaudéPrevious commit passed all our CI tests, this assertion being never triggered. Remove it as dead code.
--- accel/kvm/kvm-all.c | 7 ------- 1 file changed, 7 deletions(-)
accel/kvm/kvm-all.c +0 -7
@@ -830,13 +830,6 @@ static uint32_t kvm_dirty_ring_reap_one(KVMState *s, CPUState *cpu)uint32_t ring_size = s->kvm_dirty_ring_size;uint32_t count =0 , fetch = cpu->kvm_fetch_index; -/* -* It's not possible that we race with vcpu creation code where the vcpu is -* put onto the vcpus list but not yet initialized the dirty ring -* structures. -*/ -assert (cpu->created); -assert (dirty_gfns && ring_size);trace_kvm_dirty_ring_reap_vcpu (cpu->cpu_index);
Signature · 1 line
2.47.1
Re: [PATCH 9/9] accel/kvm: Remove unreachable assertion in kvm_dirty_ring_reap*()
Quoted 16 lines · Philippe Mathieu-Daudé
Previous commit passed all our CI tests, this assertion being never triggered. Remove it as dead code.
Signed-off-by Philippe Mathieu-Daudé <[email protected]>--- accel/kvm/kvm-all.c | 7 ------- 1 file changed, 7 deletions(-)accel/kvm/kvm-all.c +0
-0@@ -830,13 +830,6 @@ static uint32_t kvm_dirty_ring_reap_one(KVMState *s, CPUState *cpu)uint32_t ring_size = s->kvm_dirty_ring_size;uint32_t count =0 , fetch = cpu->kvm_fetch_index;
accel/kvm/kvm-all.c static uint32_t kvm_dirty_ring_reap_one(KVMState *s, CPUState *cpu) · line 835
Quoted 10 lines
accel/kvm/kvm-all.c +0
-7-/* -* It's not possible that we race with vcpu creation code where the vcpu is -* put onto the vcpus list but not yet initialized the dirty ring -* structures. -*/ -assert (cpu->created); -assert (dirty_gfns && ring_size);trace_kvm_dirty_ring_reap_vcpu (cpu->cpu_index);
I'd be ok squashing this, but also ok with retaining the patch separate for the CI comment.
r~
Quoted 4 lines · On Tue, 28 Jan 2025 15:21:43 +0100 Philippe Mathieu-Daudé <[email protected]>…
Hi,
The goal of this series is to expose vCPUs in a stable state to the accelerators, in particular the QDev 'REALIZED' step.
I'll take some of your patches (with Richard's feedback fixed), and respin series focusing mostly on realize part. (and drop wire/unwire parts, reshuffling cpu code instead)
Quoted 42 lines
To do so we split the QTAILQ_INSERT/REMOVE calls from
cpu_list_add()/cpu_list_remove(), by moving them to the DeviceClass::[un]wire() handlers, guaranty to be called just before a vCPU is exposed to the guest, as "realized".First we have to modify how
&first_cpuis used in TCG round robin implementation, and ensure we invalidate the TB jmpcache with&qemu_cpu_listlocked.I'm really out of my comfort zone here, so posting as RFC. At least all test suite is passing...
I expect these changes to allow
CPUState::cpu_indexclarifications and simplifications, but this will be addressed (and commented) in a separate series.Regards,
Phil.
Philippe Mathieu-Daudé (9):
accel/tcg: Simplify use of&first_cpuinrr_cpu_thread_fn()accel/tcg: Invalidate TB jump cache with global vCPU queue locked
cpus: Remove cpu from global queue after UNREALIZE completed hw/qdev: Introduce DeviceClass::[un]wire() handlers
cpus: Add DeviceClass::[un]wire() stubs
cpus: Call hotplug handlers inDeviceWire()
cpus: Only expose REALIZED vCPUs to global&cpus_queueaccel/kvm: Assert vCPU is created when calling kvm_dirty_ring_reap*() accel/kvm: Remove unreachable assertion in kvm_dirty_ring_reap*()include/hw/qdev-core.h | 7 +++++++ accel/kvm/kvm-all.c | 9 --------- accel/tcg/tb-maint.c | 2 ++ accel/tcg/tcg-accel-ops-rr.c | 15 ++++++++------- cpu-common.c | 2 -- cpu-target.c | 7 ++----- hw/core/cpu-common.c | 18 +++++++++++++++++- hw/core/qdev.c | 20 +++++++++++++++++++- 8 files changed, 55 insertions(+), 25 deletions(-)
Quoted 10 lines · Igor Mammedov
Quoted 4 lines · On Tue, 28 Jan 2025 15:21:43 +0100 Philippe Mathieu-Daudé <[email protected]>…
Hi,
The goal of this series is to expose vCPUs in a stable state to the accelerators, in particular the QDev 'REALIZED' step.
I'll take some of your patches (with Richard's feedback fixed), and respin series focusing mostly on realize part.
Great, thank you for helping!
Quoted 22 lines
(and drop wire/unwire parts, reshuffling cpu code instead)
Quoted 20 lines
To do so we split the QTAILQ_INSERT/REMOVE calls from
cpu_list_add()/cpu_list_remove(), by moving them to the DeviceClass::[un]wire() handlers, guaranty to be called just before a vCPU is exposed to the guest, as "realized".First we have to modify how
&first_cpuis used in TCG round robin implementation, and ensure we invalidate the TB jmpcache with&qemu_cpu_listlocked.I'm really out of my comfort zone here, so posting as RFC. At least all test suite is passing...
I expect these changes to allow
CPUState::cpu_indexclarifications and simplifications, but this will be addressed (and commented) in a separate series.Regards,
Phil.