Message ID | 20190523105440.27045-2-rkagan@virtuozzo.com (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
Series | establish nesting rule of BQL vs cpu-exclusive | expand |
Roman Kagan <rkagan@virtuozzo.com> writes: > It was introduced in commit b129972c8b41e15b0521895a46fd9c752b68a5e, > with the following motivation: I can't find this commit in my tree. > > Because start_exclusive uses CPU_FOREACH, merge exclusive_lock with > qemu_cpu_list_lock: together with a call to exclusive_idle (via > cpu_exec_start/end) in cpu_list_add, this protects exclusive work > against concurrent CPU addition and removal. > > However, it seems to be redundant, because the cpu-exclusive > infrastructure provides suffificent protection against the newly added > CPU starting execution while the cpu-exclusive work is running, and the > aforementioned traversing of the cpu list is protected by > qemu_cpu_list_lock. > > Besides, this appears to be the only place where the cpu-exclusive > section is entered with the BQL taken, which has been found to trigger > AB-BA deadlock as follows: > > vCPU thread main thread > ----------- ----------- > async_safe_run_on_cpu(self, > async_synic_update) > ... [cpu hot-add] > process_queued_cpu_work() > qemu_mutex_unlock_iothread() > [grab BQL] > start_exclusive() cpu_list_add() > async_synic_update() finish_safe_work() > qemu_mutex_lock_iothread() cpu_exec_start() > > So remove it. This paves the way to establishing a strict nesting rule > of never entering the exclusive section with the BQL taken. > > Signed-off-by: Roman Kagan <rkagan@virtuozzo.com> > --- > cpus-common.c | 8 -------- > 1 file changed, 8 deletions(-) > > diff --git a/cpus-common.c b/cpus-common.c > index 3ca58c64e8..023cfebfa3 100644 > --- a/cpus-common.c > +++ b/cpus-common.c > @@ -69,12 +69,6 @@ static int cpu_get_free_index(void) > return cpu_index; > } > > -static void finish_safe_work(CPUState *cpu) > -{ > - cpu_exec_start(cpu); > - cpu_exec_end(cpu); > -} > - This makes sense to me intellectually but I'm worried I've missed the reason for it being introduced. Without finish_safe_work we have to wait for the actual vCPU thread function to acquire and release the BQL and enter it's first cpu_exec_start(). I guess I'd be happier if we had a hotplug test where we could stress test the operation and be sure we've not just moved the deadlock somewhere else. > void cpu_list_add(CPUState *cpu) > { > qemu_mutex_lock(&qemu_cpu_list_lock); > @@ -86,8 +80,6 @@ void cpu_list_add(CPUState *cpu) > } > QTAILQ_INSERT_TAIL_RCU(&cpus, cpu, node); > qemu_mutex_unlock(&qemu_cpu_list_lock); > - > - finish_safe_work(cpu); > } > > void cpu_list_remove(CPUState *cpu) -- Alex Bennée
On Mon, Jun 24, 2019 at 11:58:23AM +0100, Alex Bennée wrote: > Roman Kagan <rkagan@virtuozzo.com> writes: > > > It was introduced in commit b129972c8b41e15b0521895a46fd9c752b68a5e, > > with the following motivation: > > I can't find this commit in my tree. OOPS, that was supposed to be ab129972c8b41e15b0521895a46fd9c752b68a5e, sorry. > > > > > Because start_exclusive uses CPU_FOREACH, merge exclusive_lock with > > qemu_cpu_list_lock: together with a call to exclusive_idle (via > > cpu_exec_start/end) in cpu_list_add, this protects exclusive work > > against concurrent CPU addition and removal. > > > > However, it seems to be redundant, because the cpu-exclusive > > infrastructure provides suffificent protection against the newly added > > CPU starting execution while the cpu-exclusive work is running, and the > > aforementioned traversing of the cpu list is protected by > > qemu_cpu_list_lock. > > > > Besides, this appears to be the only place where the cpu-exclusive > > section is entered with the BQL taken, which has been found to trigger > > AB-BA deadlock as follows: > > > > vCPU thread main thread > > ----------- ----------- > > async_safe_run_on_cpu(self, > > async_synic_update) > > ... [cpu hot-add] > > process_queued_cpu_work() > > qemu_mutex_unlock_iothread() > > [grab BQL] > > start_exclusive() cpu_list_add() > > async_synic_update() finish_safe_work() > > qemu_mutex_lock_iothread() cpu_exec_start() > > > > So remove it. This paves the way to establishing a strict nesting rule > > of never entering the exclusive section with the BQL taken. > > > > Signed-off-by: Roman Kagan <rkagan@virtuozzo.com> > > --- > > cpus-common.c | 8 -------- > > 1 file changed, 8 deletions(-) > > > > diff --git a/cpus-common.c b/cpus-common.c > > index 3ca58c64e8..023cfebfa3 100644 > > --- a/cpus-common.c > > +++ b/cpus-common.c > > @@ -69,12 +69,6 @@ static int cpu_get_free_index(void) > > return cpu_index; > > } > > > > -static void finish_safe_work(CPUState *cpu) > > -{ > > - cpu_exec_start(cpu); > > - cpu_exec_end(cpu); > > -} > > - > > This makes sense to me intellectually but I'm worried I've missed the > reason for it being introduced. Without finish_safe_work we have to wait > for the actual vCPU thread function to acquire and release the BQL and > enter it's first cpu_exec_start(). > > I guess I'd be happier if we had a hotplug test where we could stress > test the operation and be sure we've not just moved the deadlock > somewhere else. Me too. Unfortunately I haven't managed to come up with an idea how to do this test. One of the race participants, the safe work in a vCPU thread, happens in response to an MSR write by the guest. ATM there's no way to do it without an actual guest running. I'll have a look if I can make a vm test for it, using a linux guest and its /dev/cpu/*/msr. Thanks, Roman. > > > void cpu_list_add(CPUState *cpu) > > { > > qemu_mutex_lock(&qemu_cpu_list_lock); > > @@ -86,8 +80,6 @@ void cpu_list_add(CPUState *cpu) > > } > > QTAILQ_INSERT_TAIL_RCU(&cpus, cpu, node); > > qemu_mutex_unlock(&qemu_cpu_list_lock); > > - > > - finish_safe_work(cpu); > > } > > > > void cpu_list_remove(CPUState *cpu) > > > -- > Alex Bennée >
Roman Kagan <rkagan@virtuozzo.com> writes: > On Mon, Jun 24, 2019 at 11:58:23AM +0100, Alex Bennée wrote: >> Roman Kagan <rkagan@virtuozzo.com> writes: >> >> > It was introduced in commit b129972c8b41e15b0521895a46fd9c752b68a5e, >> > with the following motivation: >> >> I can't find this commit in my tree. > > OOPS, that was supposed to be ab129972c8b41e15b0521895a46fd9c752b68a5e, > sorry. > >> >> > >> > Because start_exclusive uses CPU_FOREACH, merge exclusive_lock with >> > qemu_cpu_list_lock: together with a call to exclusive_idle (via >> > cpu_exec_start/end) in cpu_list_add, this protects exclusive work >> > against concurrent CPU addition and removal. >> > >> > However, it seems to be redundant, because the cpu-exclusive >> > infrastructure provides suffificent protection against the newly added >> > CPU starting execution while the cpu-exclusive work is running, and the >> > aforementioned traversing of the cpu list is protected by >> > qemu_cpu_list_lock. >> > >> > Besides, this appears to be the only place where the cpu-exclusive >> > section is entered with the BQL taken, which has been found to trigger >> > AB-BA deadlock as follows: >> > >> > vCPU thread main thread >> > ----------- ----------- >> > async_safe_run_on_cpu(self, >> > async_synic_update) >> > ... [cpu hot-add] >> > process_queued_cpu_work() >> > qemu_mutex_unlock_iothread() >> > [grab BQL] >> > start_exclusive() cpu_list_add() >> > async_synic_update() finish_safe_work() >> > qemu_mutex_lock_iothread() cpu_exec_start() >> > >> > So remove it. This paves the way to establishing a strict nesting rule >> > of never entering the exclusive section with the BQL taken. >> > >> > Signed-off-by: Roman Kagan <rkagan@virtuozzo.com> >> > --- >> > cpus-common.c | 8 -------- >> > 1 file changed, 8 deletions(-) >> > >> > diff --git a/cpus-common.c b/cpus-common.c >> > index 3ca58c64e8..023cfebfa3 100644 >> > --- a/cpus-common.c >> > +++ b/cpus-common.c >> > @@ -69,12 +69,6 @@ static int cpu_get_free_index(void) >> > return cpu_index; >> > } >> > >> > -static void finish_safe_work(CPUState *cpu) >> > -{ >> > - cpu_exec_start(cpu); >> > - cpu_exec_end(cpu); >> > -} >> > - >> >> This makes sense to me intellectually but I'm worried I've missed the >> reason for it being introduced. Without finish_safe_work we have to wait >> for the actual vCPU thread function to acquire and release the BQL and >> enter it's first cpu_exec_start(). >> >> I guess I'd be happier if we had a hotplug test where we could stress >> test the operation and be sure we've not just moved the deadlock >> somewhere else. > > Me too. Unfortunately I haven't managed to come up with an idea how to > do this test. One of the race participants, the safe work in a vCPU > thread, happens in response to an MSR write by the guest. ATM there's > no way to do it without an actual guest running. I'll have a look if I > can make a vm test for it, using a linux guest and its /dev/cpu/*/msr. Depending on how much machinery is required to trigger this we could add a system mode test. However there isn't much point if it requires duplicating the entire guest hotplug stack. It maybe easier to trigger on ARM - the PCSI sequence isn't overly complicated to deal with but I don't know what the impact of MSIs is. -- Alex Bennée
diff --git a/cpus-common.c b/cpus-common.c index 3ca58c64e8..023cfebfa3 100644 --- a/cpus-common.c +++ b/cpus-common.c @@ -69,12 +69,6 @@ static int cpu_get_free_index(void) return cpu_index; } -static void finish_safe_work(CPUState *cpu) -{ - cpu_exec_start(cpu); - cpu_exec_end(cpu); -} - void cpu_list_add(CPUState *cpu) { qemu_mutex_lock(&qemu_cpu_list_lock); @@ -86,8 +80,6 @@ void cpu_list_add(CPUState *cpu) } QTAILQ_INSERT_TAIL_RCU(&cpus, cpu, node); qemu_mutex_unlock(&qemu_cpu_list_lock); - - finish_safe_work(cpu); } void cpu_list_remove(CPUState *cpu)
It was introduced in commit b129972c8b41e15b0521895a46fd9c752b68a5e, with the following motivation: Because start_exclusive uses CPU_FOREACH, merge exclusive_lock with qemu_cpu_list_lock: together with a call to exclusive_idle (via cpu_exec_start/end) in cpu_list_add, this protects exclusive work against concurrent CPU addition and removal. However, it seems to be redundant, because the cpu-exclusive infrastructure provides suffificent protection against the newly added CPU starting execution while the cpu-exclusive work is running, and the aforementioned traversing of the cpu list is protected by qemu_cpu_list_lock. Besides, this appears to be the only place where the cpu-exclusive section is entered with the BQL taken, which has been found to trigger AB-BA deadlock as follows: vCPU thread main thread ----------- ----------- async_safe_run_on_cpu(self, async_synic_update) ... [cpu hot-add] process_queued_cpu_work() qemu_mutex_unlock_iothread() [grab BQL] start_exclusive() cpu_list_add() async_synic_update() finish_safe_work() qemu_mutex_lock_iothread() cpu_exec_start() So remove it. This paves the way to establishing a strict nesting rule of never entering the exclusive section with the BQL taken. Signed-off-by: Roman Kagan <rkagan@virtuozzo.com> --- cpus-common.c | 8 -------- 1 file changed, 8 deletions(-)