xen: credit1: fix tickling when it happens from a remote pCPU
PATCH 1/1 xen: credit1: fix tickling when it happens from a remote pCPU
especially if that is also from a different cpupool than the processor of the vCPU that triggered the tickling.
In fact, it is possible that we get as far as calling vcpu_unblock()--> vcpu_wake()-->csched_vcpu_wake()-->__runq_tickle() for the vCPU 'vc', but all while running on a pCPU that is different from 'vc->processor'.
For instance, this can happen when an HVM domain runs in a cpupool, with a different scheduler than the default one, and issues IOREQs to Dom0, running in Pool-0 with the default scheduler. In fact, right in this case, the following crash can be observed:
(XEN) ----[ Xen-4.7-unstable x86_64 debug=y Tainted: C ]----
(XEN) CPU: 7
(XEN) RIP: e008:[<ffff82d0801230de>] __runq_tickle+0x18f/0x430
(XEN) RFLAGS: 0000000000010086 CONTEXT: hypervisor (d1v0)
(XEN) rax: 0000000000000001 rbx: ffff8303184fee00 rcx: 0000000000000000
(XEN) ... ... ...
(XEN) Xen stack trace from rsp=ffff83031fa57a08:
(XEN) ffff82d0801fe664 ffff82d08033c820 0000000100000002 0000000a00000001
(XEN) 0000000000006831 0000000000000000 0000000000000000 0000000000000000
(XEN) ... ... ...
(XEN) Xen call trace:
(XEN) [<ffff82d0801230de>] __runq_tickle+0x18f/0x430 (XEN) [<ffff82d08012348a>] csched_vcpu_wake+0x10b/0x110 (XEN) [<ffff82d08012b421>] vcpu_wake+0x20a/0x3ce (XEN) [<ffff82d08012b91c>] vcpu_unblock+0x4b/0x4e (XEN) [<ffff82d080167bd0>] vcpu_kick+0x17/0x61 (XEN) [<ffff82d080167c46>] vcpu_mark_events_pending+0x2c/0x2f (XEN) [<ffff82d08010ac35>] evtchn_fifo_set_pending+0x381/0x3f6 (XEN) [<ffff82d08010a0f6>] notify_via_xen_event_channel+0xc9/0xd6 (XEN) [<ffff82d0801c29ed>] hvm_send_ioreq+0x3e9/0x441 (XEN) [<ffff82d0801bba7d>] hvmemul_do_io+0x23f/0x2d2 (XEN) [<ffff82d0801bbb43>] hvmemul_do_io_buffer+0x33/0x64 (XEN) [<ffff82d0801bc92b>] hvmemul_do_pio_buffer+0x35/0x37 (XEN) [<ffff82d0801cc49f>] handle_pio+0x58/0x14c (XEN) [<ffff82d0801eabcb>] vmx_vmexit_handler+0x16b3/0x1bea (XEN) [<ffff82d0801efd21>] vmx_asm_vmexit_handler+0x41/0xc0
In this case, pCPU 7 is not in Pool-0, while the (Dom0's) vCPU being woken is. pCPU's 7 pool has a different scheduler than credit, but it is, however, right from pCPU 7 that we are waking the Dom0's vCPUs. Therefore, the current code tries to access csched_balance_mask for pCPU 7, but that is not defined, and hence the Oops.
(Note that, in case the two pools run the same scheduler we see no Oops, but things are still conceptually wrong.)
Cure things by providing a second macro allowing to fetch the scratch mask of a specific pCPU (instead than always using smp_processor_id()), and use that one in __runq_tickle(), with such pCPU equal to the processor of the vCPU being woken.
---
Cc: George Dunlap <[email protected]>
Cc: Juergen Gross <[email protected]>
--- xen/common/sched_credit.c | 15 +++++++++------ 1 file changed, 9 insertions(+), 6 deletions(-)
xen/common/sched_credit.c +9 -6
@@ -168,14 +168,16 @@ struct csched_pcpu { };/* -* Convenience macro for accessing the per-PCPU cpumask we need for +* Convenience macros for accessing the per-PCPU cpumask we need for * implementing the two steps (soft and hard affinity) balancing logic. * It is stored in csched_pcpu so that serialization is not an issue, -* as there is a csched_pcpu for each PCPU and we always hold the -* runqueue spin-lock when using this. +* as there is a csched_pcpu for each PCPU, and we always hold the +* runqueue lock for the proper PCPU when using these. */ #define csched_balance_mask (CSCHED_PCPU(smp_processor_id())->balance_mask) +#define csched_balance_mask_cpu (c) (CSCHED_PCPU(c)->balance_mask) +/* * Virtual CPU */
@@ -412,9 +414,10 @@ __runq_tickle(unsigned int cpu, struct csched_vcpu *new)/* Are there idlers suitable for new (for this balance step)? */ csched_balance_cpumask (new->vcpu, balance_step, - csched_balance_mask); -cpumask_and (csched_balance_mask, csched_balance_mask, &idle_mask); - new_idlers_empty =cpumask_empty (csched_balance_mask); +csched_balance_mask_cpu (cpu)); +cpumask_and (csched_balance_mask_cpu (cpu), +csched_balance_mask_cpu (cpu), &idle_mask); + new_idlers_empty =cpumask_empty (csched_balance_mask_cpu (cpu));/* * Let's not be too harsh! If there aren't idlers suitable
Quoted 48 lines · Dario Faggioli
especially if that is also from a different cpupool than the processor of the vCPU that triggered the tickling.
In fact, it is possible that we get as far as calling vcpu_unblock()--> vcpu_wake()-->csched_vcpu_wake()-->__runq_tickle() for the vCPU 'vc', but all while running on a pCPU that is different from '
vc->processor'.For instance, this can happen when an HVM domain runs in a cpupool, with a different scheduler than the default one, and issues IOREQs to Dom0, running in Pool-0 with the default scheduler. In fact, right in this case, the following crash can be observed:
(XEN) ----[ Xen-4.7-unstable
x86_64debug=y Tainted: C ]----
(XEN) CPU: 7
(XEN) RIP: e008:[<ffff82d0801230de>]__runq_tickle+0x18f/0x430
(XEN) RFLAGS: 0000000000010086 CONTEXT: hypervisor (d1v0)
(XEN) rax: 0000000000000001 rbx:ffff8303184fee00rcx: 0000000000000000
(XEN) ... ... ...
(XEN) Xen stack trace from rsp=ffff83031fa57a08:
(XEN)ffff82d0801fe664ffff82d08033c82000000001000000020000000a00000001
(XEN) 0000000000006831 0000000000000000 0000000000000000 0000000000000000
(XEN) ... ... ...
(XEN) Xen call trace:(XEN) [<ffff82d0801230de>] __runq_tickle+0x18f/0x430 (XEN) [<ffff82d08012348a>] csched_vcpu_wake+0x10b/0x110 (XEN) [<ffff82d08012b421>] vcpu_wake+0x20a/0x3ce (XEN) [<ffff82d08012b91c>] vcpu_unblock+0x4b/0x4e (XEN) [<ffff82d080167bd0>] vcpu_kick+0x17/0x61 (XEN) [<ffff82d080167c46>] vcpu_mark_events_pending+0x2c/0x2f (XEN) [<ffff82d08010ac35>] evtchn_fifo_set_pending+0x381/0x3f6 (XEN) [<ffff82d08010a0f6>] notify_via_xen_event_channel+0xc9/0xd6 (XEN) [<ffff82d0801c29ed>] hvm_send_ioreq+0x3e9/0x441 (XEN) [<ffff82d0801bba7d>] hvmemul_do_io+0x23f/0x2d2 (XEN) [<ffff82d0801bbb43>] hvmemul_do_io_buffer+0x33/0x64 (XEN) [<ffff82d0801bc92b>] hvmemul_do_pio_buffer+0x35/0x37 (XEN) [<ffff82d0801cc49f>] handle_pio+0x58/0x14c (XEN) [<ffff82d0801eabcb>] vmx_vmexit_handler+0x16b3/0x1bea (XEN) [<ffff82d0801efd21>] vmx_asm_vmexit_handler+0x41/0xc0In this case, pCPU 7 is not in Pool-0, while the (Dom0's) vCPU being woken is. pCPU's 7 pool has a different scheduler than credit, but it is, however, right from pCPU 7 that we are waking the Dom0's vCPUs. Therefore, the current code tries to access
csched_balance_maskfor pCPU 7, but that is not defined, and hence the Oops.(Note that, in case the two pools run the same scheduler we see no Oops, but things are still conceptually wrong.)
Quoted 6 lines
Cure things by providing a second macro allowing to fetch the scratch mask of a specific pCPU (instead than always using
smp_processor_id()), and use that one in__runq_tickle(), with such pCPU equal to the processor of the vCPU being woken.Signed-off-by Dario Faggioli <[email protected]>
regardless whether you address my suggestion below or not.
Quoted 15 lines
---
Cc: George Dunlap <[email protected]>
Cc: Juergen Gross <[email protected]>--- xen/common/sched_credit.c | 15 +++++++++------ 1 file changed, 9 insertions(+), 6 deletions(-)xen/common/sched_credit.c +0
-0@@ -168,14 +168,16 @@ struct csched_pcpu { };/*
xen/common/sched_credit.c struct csched_pcpu { · line 177
Quoted 10 lines
xen/common/sched_credit.c +3
-3-* Convenience macro for accessing the per-PCPU cpumask we need for +* Convenience macros for accessing the per-PCPU cpumask we need for * implementing the two steps (soft and hard affinity) balancing logic. * It is stored in csched_pcpu so that serialization is not an issue, -* as there is a csched_pcpu for each PCPU and we always hold the -* runqueue spin-lock when using this. +* as there is a csched_pcpu for each PCPU, and we always hold the +* runqueue lock for the proper PCPU when using these. */ #define csched_balance_mask (CSCHED_PCPU(smp_processor_id())->balance_mask)
Wouldn't it make sense to get rid of that macro? After your patch it is used only in csched_runq_steal() which is called with cpu being always smp_processor_id(). You'd eliminate a possible source of future error and avoid multiple evaluation of smp_processor_id().
Juergen
Quoted 22 lines
xen/common/sched_credit.c +6
-3+#define csched_balance_mask_cpu (c) (CSCHED_PCPU(c)->balance_mask) +/* * Virtual CPU */ @@ -412,9 +414,10 @@ __runq_tickle(unsigned int cpu, struct csched_vcpu *new)/* Are there idlers suitable for new (for this balance step)? */ csched_balance_cpumask (new->vcpu, balance_step, - csched_balance_mask); -cpumask_and (csched_balance_mask, csched_balance_mask, &idle_mask); - new_idlers_empty =cpumask_empty (csched_balance_mask); +csched_balance_mask_cpu (cpu)); +cpumask_and (csched_balance_mask_cpu (cpu), +csched_balance_mask_cpu (cpu), &idle_mask); + new_idlers_empty =cpumask_empty (csched_balance_mask_cpu (cpu));/* * Let's not be too harsh! If there aren't idlers suitable
Quoted 9 lines
Quoted 1 line
Quoted 1 line · On 24.09.15 at 06:31, <[email protected]> wrote:
xen/common/sched_credit.c +1
-1@@ -168,14 +168,16 @@ struct csched_pcpu { };/* -* Convenience macro for accessing the per-PCPU cpumask we need for +* Convenience macros for accessing the per-PCPU cpumask we need for
xen/common/sched_credit.c struct csched_pcpu { · line 179
Quoted 10 lines
xen/common/sched_credit.c +3
-2* implementing the two steps (soft and hard affinity) balancing logic. * It is stored in csched_pcpu so that serialization is not an issue, -* as there is a csched_pcpu for each PCPU and we always hold the -* runqueue spin-lock when using this. +* as there is a csched_pcpu for each PCPU, and we always hold the +* runqueue lock for the proper PCPU when using these. */ #define csched_balance_mask (CSCHED_PCPU(smp_processor_id())->balance_mask) +#define csched_balance_mask_cpu (c) (CSCHED_PCPU(c)->balance_mask)
csched_runq_steal() gets called with peer_cpu's runqueue lock held afaics, but uses smp_processor_id()'s balance_mask. I.e. it looks to me that what Jürgen suggested as an option is actually a requirement.
Jan
_______________________________________________
Xen-devel mailing list
[email protected]
http://lists.xen.org/xen-devel
Quoted 4 lines · Jan Beulich
Quoted 3 lines
Quoted 1 line
Quoted 1 line · On 24.09.15 at 06:31, <[email protected]> wrote:
xen/common/sched_credit.c +0
-0
Quoted 9 lines
Quoted 4 lines
#define
csched_balance_mask(CSCHED_PCPU(smp_processor_id())
->balance_mask)+#define
csched_balance_mask_cpu(c)(CSCHED_PCPU(c)->balance_mask)
csched_runq_steal()gets called with peer_cpu's runqueue lock held afaics, but uses smp_processor_id()'sbalance_mask. I.e. it looks to me that what Jürgen suggested as an option is actually a requirement.
And I'm very much agreeable on taking his suggestion, because I actually like it.
Correctness should not be an issue, though. In fact, here is the story about csched_runq_steal():
schedule()
cpu = smp_processor_id()
lock = pcpu_schedule_lock_irq(cpu);
sched = this_cpu(scheduler);
next_slice = sched->do_schedule(sched, ...);
|
--> csched_schedule()
cpu = smp_processor_id();
snext = csched_load_balance(..., cpu, ...);
peer_cpu = xxx;
lock = pcpu_schedule_trylock(peer_cpu);
speer = csched_runq_steal(peer_cpu, cpu, ...);
csched_balance_cpumask(..., csched_balance_mask);
pcpu_schedule_unlock(lock, peer_cpu); pcpu_schedule_unlock_irq(lock, cpu);
So, we can safely use smp_processor_id()'s scratch space, as we own its runqueue lock too.
In any case, thanks a lot for having a look.
Dario
Signature · 4 lines
<<This happens because I choose it to happen!>> (Raistlin Majere) ----------------------------------------------------------------- Dario Faggioli, Ph.D, http://about.me/dario.faggioli Senior Software Engineer, Citrix Systems R&D Ltd., Cambridge (UK)
_______________________________________________
Xen-devel mailing list
[email protected]
http://lists.xen.org/xen-devel
Quoted 14 lines · Juergen Gross
Quoted 1 line · Dario Faggioli
Signed-off-by Dario Faggioli <[email protected]>Reviewed-by Juergen Gross <[email protected]>regardless whether you address my suggestion below or not.
xen/common/sched_credit.c struct csched_pcpu { · line 170
Quoted 6 lines
xen/common/sched_credit.c +0
-1@@ -168,14 +168,16 @@ struct csched_pcpu {#define csched_balance_mask (CSCHED_PCPU(smp_processor_id()) ->balance_mask)
Quoted 7 lines
Wouldn't it make sense to get rid of that macro? After your patch it is
used only incsched_runq_steal()which is called with cpu being alwayssmp_processor_id(). You'd eliminate a possible source of future error and avoid multiple evaluation ofsmp_processor_id().
Right. As said to Jan, I like this. I'll respin the patch with only the new macro.
Thanks and Regards,
Dario
Signature · 4 lines
<<This happens because I choose it to happen!>> (Raistlin Majere) ----------------------------------------------------------------- Dario Faggioli, Ph.D, http://about.me/dario.faggioli Senior Software Engineer, Citrix Systems R&D Ltd., Cambridge (UK)
_______________________________________________
Xen-devel mailing list
[email protected]
http://lists.xen.org/xen-devel