| Message ID | 20240529090132.59434-4-roger.pau@citrix.com (mailing list archive) |
|---|---|
| State | Superseded |
| Headers | show |
| Series | x86/irq: fixes for CPU hot{,un}plug | expand |
On 29.05.2024 11:01, Roger Pau Monne wrote: > Due to the current rwlock logic, if the CPU calling get_cpu_maps() does so from > a cpu_hotplug_{begin,done}() region the function will still return success, > because a CPU taking the rwlock in read mode after having taken it in write > mode is allowed. Such behavior however defeats the purpose of get_cpu_maps(), > as it should always return false when called with a CPU hot{,un}plug operation > is in progress. I'm not sure I can agree with this. The CPU doing said operation ought to be aware of what it is itself doing. And all other CPUs will get back false from get_cpu_maps(). > Otherwise the logic in send_IPI_mask() for example is wrong, > as it could decide to use the shorthand even when a CPU operation is in > progress. It's also not becoming clear what's wrong there: As long as a CPU isn't offline enough to not be in cpu_online_map anymore, it may well need to still be the target of IPIs, and targeting it with a shorthand then is still fine. In any event this would again affect only the CPU leading the CPU operation, which should clearly know at which point(s) it is okay to send IPIs. Are we actually sending any IPIs from within CPU-online or CPU-offline paths? Together with the earlier paragraph the critical window would be between the CPU being taken off of cpu_online_map and the CPU actually going "dead" (i.e. on x86: its LAPIC becoming unresponsive to other than INIT/SIPI). And even then the question would be what bad, if any, would happen to that CPU if an IPI was still targeted at it by way of using the shorthand. I'm pretty sure it runs with IRQs off at that time, so no ordinary IRQ could be delivered. > Adjust the logic in get_cpu_maps() to return false when the CPUs lock is > already hold in write mode by the current CPU, as read_trylock() would > otherwise return true. > > Fixes: 868a01021c6f ('rwlock: allow recursive read locking when already locked in write mode') I'm puzzled by this as well: Prior to that and the change referenced by its Fixes: tag, recursive spin locks were used. For the purposes here that's the same as permitting read locking even when the write lock is already held by the local CPU. Jan
On Wed, May 29, 2024 at 03:35:04PM +0200, Jan Beulich wrote: > On 29.05.2024 11:01, Roger Pau Monne wrote: > > Due to the current rwlock logic, if the CPU calling get_cpu_maps() does so from > > a cpu_hotplug_{begin,done}() region the function will still return success, > > because a CPU taking the rwlock in read mode after having taken it in write > > mode is allowed. Such behavior however defeats the purpose of get_cpu_maps(), > > as it should always return false when called with a CPU hot{,un}plug operation > > is in progress. > > I'm not sure I can agree with this. The CPU doing said operation ought to be > aware of what it is itself doing. And all other CPUs will get back false from > get_cpu_maps(). Well, the CPU is aware in the context of cpu_{up,down}(), but not in the interrupts that might be handled while that operation is in progress, see below for a concrete example. > > Otherwise the logic in send_IPI_mask() for example is wrong, > > as it could decide to use the shorthand even when a CPU operation is in > > progress. > > It's also not becoming clear what's wrong there: As long as a CPU isn't > offline enough to not be in cpu_online_map anymore, it may well need to still > be the target of IPIs, and targeting it with a shorthand then is still fine. The issue is in the online path: there's a window where the CPU is online (and the lapic active), but cpu_online_map hasn't been updated yet. A specific example would be time_calibration() being executed on the CPU that is running cpu_up(). That could result in a shorthand IPI being used, but the mask in r.cpu_calibration_map not containing the CPU that's being brought up online because it's not yet added to cpu_online_map. Then the number of CPUs actually running time_calibration_rendezvous_fn won't match the weight of the cpumask in r.cpu_calibration_map. > In any event this would again affect only the CPU leading the CPU operation, > which should clearly know at which point(s) it is okay to send IPIs. Are we > actually sending any IPIs from within CPU-online or CPU-offline paths? Yes, I've seen the time rendezvous happening while in the middle of a hotplug operation, and the CPU coordinating the rendezvous being the one doing the CPU hotplug operation, so get_cpu_maps() returning true. > Together with the earlier paragraph the critical window would be between the > CPU being taken off of cpu_online_map and the CPU actually going "dead" (i.e. > on x86: its LAPIC becoming unresponsive to other than INIT/SIPI). And even > then the question would be what bad, if any, would happen to that CPU if an > IPI was still targeted at it by way of using the shorthand. I'm pretty sure > it runs with IRQs off at that time, so no ordinary IRQ could be delivered. > > > Adjust the logic in get_cpu_maps() to return false when the CPUs lock is > > already hold in write mode by the current CPU, as read_trylock() would > > otherwise return true. > > > > Fixes: 868a01021c6f ('rwlock: allow recursive read locking when already locked in write mode') > > I'm puzzled by this as well: Prior to that and the change referenced by its > Fixes: tag, recursive spin locks were used. For the purposes here that's the > same as permitting read locking even when the write lock is already held by > the local CPU. I see, so the Fixes should be: x86/smp: use APIC ALLBUT destination shorthand when possible Instead, which is the commit that started using get_cpu_maps() in send_IPI_mask(). Thanks, Roger.
On 29.05.2024 17:03, Roger Pau Monné wrote: > On Wed, May 29, 2024 at 03:35:04PM +0200, Jan Beulich wrote: >> On 29.05.2024 11:01, Roger Pau Monne wrote: >>> Due to the current rwlock logic, if the CPU calling get_cpu_maps() does so from >>> a cpu_hotplug_{begin,done}() region the function will still return success, >>> because a CPU taking the rwlock in read mode after having taken it in write >>> mode is allowed. Such behavior however defeats the purpose of get_cpu_maps(), >>> as it should always return false when called with a CPU hot{,un}plug operation >>> is in progress. >> >> I'm not sure I can agree with this. The CPU doing said operation ought to be >> aware of what it is itself doing. And all other CPUs will get back false from >> get_cpu_maps(). > > Well, the CPU is aware in the context of cpu_{up,down}(), but not in > the interrupts that might be handled while that operation is in > progress, see below for a concrete example. > >>> Otherwise the logic in send_IPI_mask() for example is wrong, >>> as it could decide to use the shorthand even when a CPU operation is in >>> progress. >> >> It's also not becoming clear what's wrong there: As long as a CPU isn't >> offline enough to not be in cpu_online_map anymore, it may well need to still >> be the target of IPIs, and targeting it with a shorthand then is still fine. > > The issue is in the online path: there's a window where the CPU is > online (and the lapic active), but cpu_online_map hasn't been updated > yet. A specific example would be time_calibration() being executed on > the CPU that is running cpu_up(). That could result in a shorthand > IPI being used, but the mask in r.cpu_calibration_map not containing > the CPU that's being brought up online because it's not yet added to > cpu_online_map. Then the number of CPUs actually running > time_calibration_rendezvous_fn won't match the weight of the cpumask > in r.cpu_calibration_map. I see, but maybe only partly. Prior to the CPU having its bit set in cpu_online_map, can it really take interrupts already? Shouldn't it be running with IRQs off until later, thus preventing it from making it into the rendezvous function in the first place? But yes, I can see how the IRQ (IPI) then being delivered later (once IRQs are enabled) might cause problems, too. Plus, with how the rendezvous function is invoked (via on_selected_cpus() with the mask copied from cpu_online_map), the first check in smp_call_function_interrupt() ought to prevent the function from being called on the CPU being onlined. A problem would arise though if the IPI arrived later and call_data was already (partly or fully) overwritten with the next request. >> In any event this would again affect only the CPU leading the CPU operation, >> which should clearly know at which point(s) it is okay to send IPIs. Are we >> actually sending any IPIs from within CPU-online or CPU-offline paths? > > Yes, I've seen the time rendezvous happening while in the middle of a > hotplug operation, and the CPU coordinating the rendezvous being the > one doing the CPU hotplug operation, so get_cpu_maps() returning true. Right, yet together with ... >> Together with the earlier paragraph the critical window would be between the >> CPU being taken off of cpu_online_map and the CPU actually going "dead" (i.e. >> on x86: its LAPIC becoming unresponsive to other than INIT/SIPI). And even >> then the question would be what bad, if any, would happen to that CPU if an >> IPI was still targeted at it by way of using the shorthand. I'm pretty sure >> it runs with IRQs off at that time, so no ordinary IRQ could be delivered. >> >>> Adjust the logic in get_cpu_maps() to return false when the CPUs lock is >>> already hold in write mode by the current CPU, as read_trylock() would >>> otherwise return true. >>> >>> Fixes: 868a01021c6f ('rwlock: allow recursive read locking when already locked in write mode') >> >> I'm puzzled by this as well: Prior to that and the change referenced by its >> Fixes: tag, recursive spin locks were used. For the purposes here that's the >> same as permitting read locking even when the write lock is already held by >> the local CPU. > > I see, so the Fixes should be: > > x86/smp: use APIC ALLBUT destination shorthand when possible > > Instead, which is the commit that started using get_cpu_maps() in > send_IPI_mask(). ... this I then wonder whether it's really only the condition in send_IPI_mask() which needs further amending, rather than fiddling with get_cpu_maps(). Jan
On Wed, May 29, 2024 at 05:49:48PM +0200, Jan Beulich wrote: > On 29.05.2024 17:03, Roger Pau Monné wrote: > > On Wed, May 29, 2024 at 03:35:04PM +0200, Jan Beulich wrote: > >> On 29.05.2024 11:01, Roger Pau Monne wrote: > >>> Due to the current rwlock logic, if the CPU calling get_cpu_maps() does so from > >>> a cpu_hotplug_{begin,done}() region the function will still return success, > >>> because a CPU taking the rwlock in read mode after having taken it in write > >>> mode is allowed. Such behavior however defeats the purpose of get_cpu_maps(), > >>> as it should always return false when called with a CPU hot{,un}plug operation > >>> is in progress. > >> > >> I'm not sure I can agree with this. The CPU doing said operation ought to be > >> aware of what it is itself doing. And all other CPUs will get back false from > >> get_cpu_maps(). > > > > Well, the CPU is aware in the context of cpu_{up,down}(), but not in > > the interrupts that might be handled while that operation is in > > progress, see below for a concrete example. > > > >>> Otherwise the logic in send_IPI_mask() for example is wrong, > >>> as it could decide to use the shorthand even when a CPU operation is in > >>> progress. > >> > >> It's also not becoming clear what's wrong there: As long as a CPU isn't > >> offline enough to not be in cpu_online_map anymore, it may well need to still > >> be the target of IPIs, and targeting it with a shorthand then is still fine. > > > > The issue is in the online path: there's a window where the CPU is > > online (and the lapic active), but cpu_online_map hasn't been updated > > yet. A specific example would be time_calibration() being executed on > > the CPU that is running cpu_up(). That could result in a shorthand > > IPI being used, but the mask in r.cpu_calibration_map not containing > > the CPU that's being brought up online because it's not yet added to > > cpu_online_map. Then the number of CPUs actually running > > time_calibration_rendezvous_fn won't match the weight of the cpumask > > in r.cpu_calibration_map. > > I see, but maybe only partly. Prior to the CPU having its bit set in > cpu_online_map, can it really take interrupts already? Shouldn't it be > running with IRQs off until later, thus preventing it from making it > into the rendezvous function in the first place? But yes, I can see > how the IRQ (IPI) then being delivered later (once IRQs are enabled) > might cause problems, too. The interrupt will get set in IRR and handled when interrupts are enabled. > > Plus, with how the rendezvous function is invoked (via > on_selected_cpus() with the mask copied from cpu_online_map), the > first check in smp_call_function_interrupt() ought to prevent the > function from being called on the CPU being onlined. A problem would > arise though if the IPI arrived later and call_data was already > (partly or fully) overwritten with the next request. Yeah, there's a small window where the fields in call_data are out of sync. > >> In any event this would again affect only the CPU leading the CPU operation, > >> which should clearly know at which point(s) it is okay to send IPIs. Are we > >> actually sending any IPIs from within CPU-online or CPU-offline paths? > > > > Yes, I've seen the time rendezvous happening while in the middle of a > > hotplug operation, and the CPU coordinating the rendezvous being the > > one doing the CPU hotplug operation, so get_cpu_maps() returning true. > > Right, yet together with ... > > >> Together with the earlier paragraph the critical window would be between the > >> CPU being taken off of cpu_online_map and the CPU actually going "dead" (i.e. > >> on x86: its LAPIC becoming unresponsive to other than INIT/SIPI). And even > >> then the question would be what bad, if any, would happen to that CPU if an > >> IPI was still targeted at it by way of using the shorthand. I'm pretty sure > >> it runs with IRQs off at that time, so no ordinary IRQ could be delivered. > >> > >>> Adjust the logic in get_cpu_maps() to return false when the CPUs lock is > >>> already hold in write mode by the current CPU, as read_trylock() would > >>> otherwise return true. > >>> > >>> Fixes: 868a01021c6f ('rwlock: allow recursive read locking when already locked in write mode') > >> > >> I'm puzzled by this as well: Prior to that and the change referenced by its > >> Fixes: tag, recursive spin locks were used. For the purposes here that's the > >> same as permitting read locking even when the write lock is already held by > >> the local CPU. > > > > I see, so the Fixes should be: > > > > x86/smp: use APIC ALLBUT destination shorthand when possible > > > > Instead, which is the commit that started using get_cpu_maps() in > > send_IPI_mask(). > > ... this I then wonder whether it's really only the condition in > send_IPI_mask() which needs further amending, rather than fiddling with > get_cpu_maps(). That the other option, but I have impression it's more fragile to adjust the condition in send_IPI_mask() rather than fiddle with get_cpu_maps(). However if that's the preference I can adjust. Thanks, Roger.
On 29.05.2024 18:14, Roger Pau Monné wrote: > On Wed, May 29, 2024 at 05:49:48PM +0200, Jan Beulich wrote: >> On 29.05.2024 17:03, Roger Pau Monné wrote: >>> On Wed, May 29, 2024 at 03:35:04PM +0200, Jan Beulich wrote: >>>> On 29.05.2024 11:01, Roger Pau Monne wrote: >>>>> Due to the current rwlock logic, if the CPU calling get_cpu_maps() does so from >>>>> a cpu_hotplug_{begin,done}() region the function will still return success, >>>>> because a CPU taking the rwlock in read mode after having taken it in write >>>>> mode is allowed. Such behavior however defeats the purpose of get_cpu_maps(), >>>>> as it should always return false when called with a CPU hot{,un}plug operation >>>>> is in progress. >>>> >>>> I'm not sure I can agree with this. The CPU doing said operation ought to be >>>> aware of what it is itself doing. And all other CPUs will get back false from >>>> get_cpu_maps(). >>> >>> Well, the CPU is aware in the context of cpu_{up,down}(), but not in >>> the interrupts that might be handled while that operation is in >>> progress, see below for a concrete example. >>> >>>>> Otherwise the logic in send_IPI_mask() for example is wrong, >>>>> as it could decide to use the shorthand even when a CPU operation is in >>>>> progress. >>>> >>>> It's also not becoming clear what's wrong there: As long as a CPU isn't >>>> offline enough to not be in cpu_online_map anymore, it may well need to still >>>> be the target of IPIs, and targeting it with a shorthand then is still fine. >>> >>> The issue is in the online path: there's a window where the CPU is >>> online (and the lapic active), but cpu_online_map hasn't been updated >>> yet. A specific example would be time_calibration() being executed on >>> the CPU that is running cpu_up(). That could result in a shorthand >>> IPI being used, but the mask in r.cpu_calibration_map not containing >>> the CPU that's being brought up online because it's not yet added to >>> cpu_online_map. Then the number of CPUs actually running >>> time_calibration_rendezvous_fn won't match the weight of the cpumask >>> in r.cpu_calibration_map. >> >> I see, but maybe only partly. Prior to the CPU having its bit set in >> cpu_online_map, can it really take interrupts already? Shouldn't it be >> running with IRQs off until later, thus preventing it from making it >> into the rendezvous function in the first place? But yes, I can see >> how the IRQ (IPI) then being delivered later (once IRQs are enabled) >> might cause problems, too. > > The interrupt will get set in IRR and handled when interrupts are > enabled. > >> >> Plus, with how the rendezvous function is invoked (via >> on_selected_cpus() with the mask copied from cpu_online_map), the >> first check in smp_call_function_interrupt() ought to prevent the >> function from being called on the CPU being onlined. A problem would >> arise though if the IPI arrived later and call_data was already >> (partly or fully) overwritten with the next request. > > Yeah, there's a small window where the fields in call_data are out of > sync. > >>>> In any event this would again affect only the CPU leading the CPU operation, >>>> which should clearly know at which point(s) it is okay to send IPIs. Are we >>>> actually sending any IPIs from within CPU-online or CPU-offline paths? >>> >>> Yes, I've seen the time rendezvous happening while in the middle of a >>> hotplug operation, and the CPU coordinating the rendezvous being the >>> one doing the CPU hotplug operation, so get_cpu_maps() returning true. >> >> Right, yet together with ... >> >>>> Together with the earlier paragraph the critical window would be between the >>>> CPU being taken off of cpu_online_map and the CPU actually going "dead" (i.e. >>>> on x86: its LAPIC becoming unresponsive to other than INIT/SIPI). And even >>>> then the question would be what bad, if any, would happen to that CPU if an >>>> IPI was still targeted at it by way of using the shorthand. I'm pretty sure >>>> it runs with IRQs off at that time, so no ordinary IRQ could be delivered. >>>> >>>>> Adjust the logic in get_cpu_maps() to return false when the CPUs lock is >>>>> already hold in write mode by the current CPU, as read_trylock() would >>>>> otherwise return true. >>>>> >>>>> Fixes: 868a01021c6f ('rwlock: allow recursive read locking when already locked in write mode') >>>> >>>> I'm puzzled by this as well: Prior to that and the change referenced by its >>>> Fixes: tag, recursive spin locks were used. For the purposes here that's the >>>> same as permitting read locking even when the write lock is already held by >>>> the local CPU. >>> >>> I see, so the Fixes should be: >>> >>> x86/smp: use APIC ALLBUT destination shorthand when possible >>> >>> Instead, which is the commit that started using get_cpu_maps() in >>> send_IPI_mask(). >> >> ... this I then wonder whether it's really only the condition in >> send_IPI_mask() which needs further amending, rather than fiddling with >> get_cpu_maps(). > > That the other option, but I have impression it's more fragile to > adjust the condition in send_IPI_mask() rather than fiddle with > get_cpu_maps(). > > However if that's the preference I can adjust. I guess we need other REST input here then. The two of us clearly disagree on what use of get_cpu_maps() is meant to guarantee. And I deem fiddling with common code here more risky (and more intrusive - the other change would be a single-line code change afaict, plus extending the related comment). Jan
On Fri, May 31, 2024 at 09:02:20AM +0200, Jan Beulich wrote: > On 29.05.2024 18:14, Roger Pau Monné wrote: > > On Wed, May 29, 2024 at 05:49:48PM +0200, Jan Beulich wrote: > >> On 29.05.2024 17:03, Roger Pau Monné wrote: > >>> On Wed, May 29, 2024 at 03:35:04PM +0200, Jan Beulich wrote: > >>>> On 29.05.2024 11:01, Roger Pau Monne wrote: > >>>>> Due to the current rwlock logic, if the CPU calling get_cpu_maps() does so from > >>>>> a cpu_hotplug_{begin,done}() region the function will still return success, > >>>>> because a CPU taking the rwlock in read mode after having taken it in write > >>>>> mode is allowed. Such behavior however defeats the purpose of get_cpu_maps(), > >>>>> as it should always return false when called with a CPU hot{,un}plug operation > >>>>> is in progress. > >>>> > >>>> I'm not sure I can agree with this. The CPU doing said operation ought to be > >>>> aware of what it is itself doing. And all other CPUs will get back false from > >>>> get_cpu_maps(). > >>> > >>> Well, the CPU is aware in the context of cpu_{up,down}(), but not in > >>> the interrupts that might be handled while that operation is in > >>> progress, see below for a concrete example. > >>> > >>>>> Otherwise the logic in send_IPI_mask() for example is wrong, > >>>>> as it could decide to use the shorthand even when a CPU operation is in > >>>>> progress. > >>>> > >>>> It's also not becoming clear what's wrong there: As long as a CPU isn't > >>>> offline enough to not be in cpu_online_map anymore, it may well need to still > >>>> be the target of IPIs, and targeting it with a shorthand then is still fine. > >>> > >>> The issue is in the online path: there's a window where the CPU is > >>> online (and the lapic active), but cpu_online_map hasn't been updated > >>> yet. A specific example would be time_calibration() being executed on > >>> the CPU that is running cpu_up(). That could result in a shorthand > >>> IPI being used, but the mask in r.cpu_calibration_map not containing > >>> the CPU that's being brought up online because it's not yet added to > >>> cpu_online_map. Then the number of CPUs actually running > >>> time_calibration_rendezvous_fn won't match the weight of the cpumask > >>> in r.cpu_calibration_map. > >> > >> I see, but maybe only partly. Prior to the CPU having its bit set in > >> cpu_online_map, can it really take interrupts already? Shouldn't it be > >> running with IRQs off until later, thus preventing it from making it > >> into the rendezvous function in the first place? But yes, I can see > >> how the IRQ (IPI) then being delivered later (once IRQs are enabled) > >> might cause problems, too. > > > > The interrupt will get set in IRR and handled when interrupts are > > enabled. > > > >> > >> Plus, with how the rendezvous function is invoked (via > >> on_selected_cpus() with the mask copied from cpu_online_map), the > >> first check in smp_call_function_interrupt() ought to prevent the > >> function from being called on the CPU being onlined. A problem would > >> arise though if the IPI arrived later and call_data was already > >> (partly or fully) overwritten with the next request. > > > > Yeah, there's a small window where the fields in call_data are out of > > sync. > > > >>>> In any event this would again affect only the CPU leading the CPU operation, > >>>> which should clearly know at which point(s) it is okay to send IPIs. Are we > >>>> actually sending any IPIs from within CPU-online or CPU-offline paths? > >>> > >>> Yes, I've seen the time rendezvous happening while in the middle of a > >>> hotplug operation, and the CPU coordinating the rendezvous being the > >>> one doing the CPU hotplug operation, so get_cpu_maps() returning true. > >> > >> Right, yet together with ... > >> > >>>> Together with the earlier paragraph the critical window would be between the > >>>> CPU being taken off of cpu_online_map and the CPU actually going "dead" (i.e. > >>>> on x86: its LAPIC becoming unresponsive to other than INIT/SIPI). And even > >>>> then the question would be what bad, if any, would happen to that CPU if an > >>>> IPI was still targeted at it by way of using the shorthand. I'm pretty sure > >>>> it runs with IRQs off at that time, so no ordinary IRQ could be delivered. > >>>> > >>>>> Adjust the logic in get_cpu_maps() to return false when the CPUs lock is > >>>>> already hold in write mode by the current CPU, as read_trylock() would > >>>>> otherwise return true. > >>>>> > >>>>> Fixes: 868a01021c6f ('rwlock: allow recursive read locking when already locked in write mode') > >>>> > >>>> I'm puzzled by this as well: Prior to that and the change referenced by its > >>>> Fixes: tag, recursive spin locks were used. For the purposes here that's the > >>>> same as permitting read locking even when the write lock is already held by > >>>> the local CPU. > >>> > >>> I see, so the Fixes should be: > >>> > >>> x86/smp: use APIC ALLBUT destination shorthand when possible > >>> > >>> Instead, which is the commit that started using get_cpu_maps() in > >>> send_IPI_mask(). > >> > >> ... this I then wonder whether it's really only the condition in > >> send_IPI_mask() which needs further amending, rather than fiddling with > >> get_cpu_maps(). > > > > That the other option, but I have impression it's more fragile to > > adjust the condition in send_IPI_mask() rather than fiddle with > > get_cpu_maps(). > > > > However if that's the preference I can adjust. > > I guess we need other REST input here then. The two of us clearly disagree on > what use of get_cpu_maps() is meant to guarantee. And I deem fiddling with > common code here more risky (and more intrusive - the other change would be > a single-line code change afaict, plus extending the related comment). How do you envision that other change to be done? Adding an extra variable and toggling it in cpu_hotplug_{begin,done}() to signal whether a CPU hotplug is in progress? If I go this route I would like to add a comment to get_cpu_maps() in order to notice this IMO weird property of succeeding when calling from a CPU that's performing a hotplug operation. Thanks, Roger.
On 31.05.2024 09:31, Roger Pau Monné wrote: > On Fri, May 31, 2024 at 09:02:20AM +0200, Jan Beulich wrote: >> On 29.05.2024 18:14, Roger Pau Monné wrote: >>> On Wed, May 29, 2024 at 05:49:48PM +0200, Jan Beulich wrote: >>>> On 29.05.2024 17:03, Roger Pau Monné wrote: >>>>> On Wed, May 29, 2024 at 03:35:04PM +0200, Jan Beulich wrote: >>>>>> On 29.05.2024 11:01, Roger Pau Monne wrote: >>>>>>> Due to the current rwlock logic, if the CPU calling get_cpu_maps() does so from >>>>>>> a cpu_hotplug_{begin,done}() region the function will still return success, >>>>>>> because a CPU taking the rwlock in read mode after having taken it in write >>>>>>> mode is allowed. Such behavior however defeats the purpose of get_cpu_maps(), >>>>>>> as it should always return false when called with a CPU hot{,un}plug operation >>>>>>> is in progress. >>>>>> >>>>>> I'm not sure I can agree with this. The CPU doing said operation ought to be >>>>>> aware of what it is itself doing. And all other CPUs will get back false from >>>>>> get_cpu_maps(). >>>>> >>>>> Well, the CPU is aware in the context of cpu_{up,down}(), but not in >>>>> the interrupts that might be handled while that operation is in >>>>> progress, see below for a concrete example. >>>>> >>>>>>> Otherwise the logic in send_IPI_mask() for example is wrong, >>>>>>> as it could decide to use the shorthand even when a CPU operation is in >>>>>>> progress. >>>>>> >>>>>> It's also not becoming clear what's wrong there: As long as a CPU isn't >>>>>> offline enough to not be in cpu_online_map anymore, it may well need to still >>>>>> be the target of IPIs, and targeting it with a shorthand then is still fine. >>>>> >>>>> The issue is in the online path: there's a window where the CPU is >>>>> online (and the lapic active), but cpu_online_map hasn't been updated >>>>> yet. A specific example would be time_calibration() being executed on >>>>> the CPU that is running cpu_up(). That could result in a shorthand >>>>> IPI being used, but the mask in r.cpu_calibration_map not containing >>>>> the CPU that's being brought up online because it's not yet added to >>>>> cpu_online_map. Then the number of CPUs actually running >>>>> time_calibration_rendezvous_fn won't match the weight of the cpumask >>>>> in r.cpu_calibration_map. >>>> >>>> I see, but maybe only partly. Prior to the CPU having its bit set in >>>> cpu_online_map, can it really take interrupts already? Shouldn't it be >>>> running with IRQs off until later, thus preventing it from making it >>>> into the rendezvous function in the first place? But yes, I can see >>>> how the IRQ (IPI) then being delivered later (once IRQs are enabled) >>>> might cause problems, too. >>> >>> The interrupt will get set in IRR and handled when interrupts are >>> enabled. >>> >>>> >>>> Plus, with how the rendezvous function is invoked (via >>>> on_selected_cpus() with the mask copied from cpu_online_map), the >>>> first check in smp_call_function_interrupt() ought to prevent the >>>> function from being called on the CPU being onlined. A problem would >>>> arise though if the IPI arrived later and call_data was already >>>> (partly or fully) overwritten with the next request. >>> >>> Yeah, there's a small window where the fields in call_data are out of >>> sync. >>> >>>>>> In any event this would again affect only the CPU leading the CPU operation, >>>>>> which should clearly know at which point(s) it is okay to send IPIs. Are we >>>>>> actually sending any IPIs from within CPU-online or CPU-offline paths? >>>>> >>>>> Yes, I've seen the time rendezvous happening while in the middle of a >>>>> hotplug operation, and the CPU coordinating the rendezvous being the >>>>> one doing the CPU hotplug operation, so get_cpu_maps() returning true. >>>> >>>> Right, yet together with ... >>>> >>>>>> Together with the earlier paragraph the critical window would be between the >>>>>> CPU being taken off of cpu_online_map and the CPU actually going "dead" (i.e. >>>>>> on x86: its LAPIC becoming unresponsive to other than INIT/SIPI). And even >>>>>> then the question would be what bad, if any, would happen to that CPU if an >>>>>> IPI was still targeted at it by way of using the shorthand. I'm pretty sure >>>>>> it runs with IRQs off at that time, so no ordinary IRQ could be delivered. >>>>>> >>>>>>> Adjust the logic in get_cpu_maps() to return false when the CPUs lock is >>>>>>> already hold in write mode by the current CPU, as read_trylock() would >>>>>>> otherwise return true. >>>>>>> >>>>>>> Fixes: 868a01021c6f ('rwlock: allow recursive read locking when already locked in write mode') >>>>>> >>>>>> I'm puzzled by this as well: Prior to that and the change referenced by its >>>>>> Fixes: tag, recursive spin locks were used. For the purposes here that's the >>>>>> same as permitting read locking even when the write lock is already held by >>>>>> the local CPU. >>>>> >>>>> I see, so the Fixes should be: >>>>> >>>>> x86/smp: use APIC ALLBUT destination shorthand when possible >>>>> >>>>> Instead, which is the commit that started using get_cpu_maps() in >>>>> send_IPI_mask(). >>>> >>>> ... this I then wonder whether it's really only the condition in >>>> send_IPI_mask() which needs further amending, rather than fiddling with >>>> get_cpu_maps(). >>> >>> That the other option, but I have impression it's more fragile to >>> adjust the condition in send_IPI_mask() rather than fiddle with >>> get_cpu_maps(). >>> >>> However if that's the preference I can adjust. >> >> I guess we need other REST input here then. The two of us clearly disagree on >> what use of get_cpu_maps() is meant to guarantee. And I deem fiddling with >> common code here more risky (and more intrusive - the other change would be >> a single-line code change afaict, plus extending the related comment). > > How do you envision that other change to be done? Adding an extra > variable and toggling it in cpu_hotplug_{begin,done}() to signal > whether a CPU hotplug is in progress? I was thinking of an is-write-locked-by-me check on cpu_add_remove_lock. Jan
On Fri, May 31, 2024 at 10:33:58AM +0200, Jan Beulich wrote: > On 31.05.2024 09:31, Roger Pau Monné wrote: > > On Fri, May 31, 2024 at 09:02:20AM +0200, Jan Beulich wrote: > >> On 29.05.2024 18:14, Roger Pau Monné wrote: > >>> On Wed, May 29, 2024 at 05:49:48PM +0200, Jan Beulich wrote: > >>>> On 29.05.2024 17:03, Roger Pau Monné wrote: > >>>>> On Wed, May 29, 2024 at 03:35:04PM +0200, Jan Beulich wrote: > >>>>>> On 29.05.2024 11:01, Roger Pau Monne wrote: > >>>>>>> Due to the current rwlock logic, if the CPU calling get_cpu_maps() does so from > >>>>>>> a cpu_hotplug_{begin,done}() region the function will still return success, > >>>>>>> because a CPU taking the rwlock in read mode after having taken it in write > >>>>>>> mode is allowed. Such behavior however defeats the purpose of get_cpu_maps(), > >>>>>>> as it should always return false when called with a CPU hot{,un}plug operation > >>>>>>> is in progress. > >>>>>> > >>>>>> I'm not sure I can agree with this. The CPU doing said operation ought to be > >>>>>> aware of what it is itself doing. And all other CPUs will get back false from > >>>>>> get_cpu_maps(). > >>>>> > >>>>> Well, the CPU is aware in the context of cpu_{up,down}(), but not in > >>>>> the interrupts that might be handled while that operation is in > >>>>> progress, see below for a concrete example. > >>>>> > >>>>>>> Otherwise the logic in send_IPI_mask() for example is wrong, > >>>>>>> as it could decide to use the shorthand even when a CPU operation is in > >>>>>>> progress. > >>>>>> > >>>>>> It's also not becoming clear what's wrong there: As long as a CPU isn't > >>>>>> offline enough to not be in cpu_online_map anymore, it may well need to still > >>>>>> be the target of IPIs, and targeting it with a shorthand then is still fine. > >>>>> > >>>>> The issue is in the online path: there's a window where the CPU is > >>>>> online (and the lapic active), but cpu_online_map hasn't been updated > >>>>> yet. A specific example would be time_calibration() being executed on > >>>>> the CPU that is running cpu_up(). That could result in a shorthand > >>>>> IPI being used, but the mask in r.cpu_calibration_map not containing > >>>>> the CPU that's being brought up online because it's not yet added to > >>>>> cpu_online_map. Then the number of CPUs actually running > >>>>> time_calibration_rendezvous_fn won't match the weight of the cpumask > >>>>> in r.cpu_calibration_map. > >>>> > >>>> I see, but maybe only partly. Prior to the CPU having its bit set in > >>>> cpu_online_map, can it really take interrupts already? Shouldn't it be > >>>> running with IRQs off until later, thus preventing it from making it > >>>> into the rendezvous function in the first place? But yes, I can see > >>>> how the IRQ (IPI) then being delivered later (once IRQs are enabled) > >>>> might cause problems, too. > >>> > >>> The interrupt will get set in IRR and handled when interrupts are > >>> enabled. > >>> > >>>> > >>>> Plus, with how the rendezvous function is invoked (via > >>>> on_selected_cpus() with the mask copied from cpu_online_map), the > >>>> first check in smp_call_function_interrupt() ought to prevent the > >>>> function from being called on the CPU being onlined. A problem would > >>>> arise though if the IPI arrived later and call_data was already > >>>> (partly or fully) overwritten with the next request. > >>> > >>> Yeah, there's a small window where the fields in call_data are out of > >>> sync. > >>> > >>>>>> In any event this would again affect only the CPU leading the CPU operation, > >>>>>> which should clearly know at which point(s) it is okay to send IPIs. Are we > >>>>>> actually sending any IPIs from within CPU-online or CPU-offline paths? > >>>>> > >>>>> Yes, I've seen the time rendezvous happening while in the middle of a > >>>>> hotplug operation, and the CPU coordinating the rendezvous being the > >>>>> one doing the CPU hotplug operation, so get_cpu_maps() returning true. > >>>> > >>>> Right, yet together with ... > >>>> > >>>>>> Together with the earlier paragraph the critical window would be between the > >>>>>> CPU being taken off of cpu_online_map and the CPU actually going "dead" (i.e. > >>>>>> on x86: its LAPIC becoming unresponsive to other than INIT/SIPI). And even > >>>>>> then the question would be what bad, if any, would happen to that CPU if an > >>>>>> IPI was still targeted at it by way of using the shorthand. I'm pretty sure > >>>>>> it runs with IRQs off at that time, so no ordinary IRQ could be delivered. > >>>>>> > >>>>>>> Adjust the logic in get_cpu_maps() to return false when the CPUs lock is > >>>>>>> already hold in write mode by the current CPU, as read_trylock() would > >>>>>>> otherwise return true. > >>>>>>> > >>>>>>> Fixes: 868a01021c6f ('rwlock: allow recursive read locking when already locked in write mode') > >>>>>> > >>>>>> I'm puzzled by this as well: Prior to that and the change referenced by its > >>>>>> Fixes: tag, recursive spin locks were used. For the purposes here that's the > >>>>>> same as permitting read locking even when the write lock is already held by > >>>>>> the local CPU. > >>>>> > >>>>> I see, so the Fixes should be: > >>>>> > >>>>> x86/smp: use APIC ALLBUT destination shorthand when possible > >>>>> > >>>>> Instead, which is the commit that started using get_cpu_maps() in > >>>>> send_IPI_mask(). > >>>> > >>>> ... this I then wonder whether it's really only the condition in > >>>> send_IPI_mask() which needs further amending, rather than fiddling with > >>>> get_cpu_maps(). > >>> > >>> That the other option, but I have impression it's more fragile to > >>> adjust the condition in send_IPI_mask() rather than fiddle with > >>> get_cpu_maps(). > >>> > >>> However if that's the preference I can adjust. > >> > >> I guess we need other REST input here then. The two of us clearly disagree on > >> what use of get_cpu_maps() is meant to guarantee. And I deem fiddling with > >> common code here more risky (and more intrusive - the other change would be > >> a single-line code change afaict, plus extending the related comment). > > > > How do you envision that other change to be done? Adding an extra > > variable and toggling it in cpu_hotplug_{begin,done}() to signal > > whether a CPU hotplug is in progress? > > I was thinking of an is-write-locked-by-me check on cpu_add_remove_lock. Oh, so basically open-coding what I proposed here as get_cpu_maps() in send_IPI_mask(). Unless anyone else expresses interest in my current proposal I would switch to that. Thanks, Roger.
diff --git a/xen/common/cpu.c b/xen/common/cpu.c index 6173220e771b..d76f80fe2e99 100644 --- a/xen/common/cpu.c +++ b/xen/common/cpu.c @@ -49,7 +49,8 @@ static DEFINE_RWLOCK(cpu_add_remove_lock); bool get_cpu_maps(void) { - return read_trylock(&cpu_add_remove_lock); + return !rw_is_write_locked_by_me(&cpu_add_remove_lock) && + read_trylock(&cpu_add_remove_lock); } void put_cpu_maps(void) diff --git a/xen/include/xen/rwlock.h b/xen/include/xen/rwlock.h index a2e98cad343e..4e7802821859 100644 --- a/xen/include/xen/rwlock.h +++ b/xen/include/xen/rwlock.h @@ -316,6 +316,8 @@ static always_inline void write_lock_irq(rwlock_t *l) #define rw_is_locked(l) _rw_is_locked(l) #define rw_is_write_locked(l) _rw_is_write_locked(l) +#define rw_is_write_locked_by_me(l) \ + lock_evaluate_nospec(_is_write_locked_by_me(atomic_read(&(l)->cnts))) typedef struct percpu_rwlock percpu_rwlock_t;
Due to the current rwlock logic, if the CPU calling get_cpu_maps() does so from a cpu_hotplug_{begin,done}() region the function will still return success, because a CPU taking the rwlock in read mode after having taken it in write mode is allowed. Such behavior however defeats the purpose of get_cpu_maps(), as it should always return false when called with a CPU hot{,un}plug operation is in progress. Otherwise the logic in send_IPI_mask() for example is wrong, as it could decide to use the shorthand even when a CPU operation is in progress. Adjust the logic in get_cpu_maps() to return false when the CPUs lock is already hold in write mode by the current CPU, as read_trylock() would otherwise return true. Fixes: 868a01021c6f ('rwlock: allow recursive read locking when already locked in write mode') Signed-off-by: Roger Pau Monné <roger.pau@citrix.com> --- xen/common/cpu.c | 3 ++- xen/include/xen/rwlock.h | 2 ++ 2 files changed, 4 insertions(+), 1 deletion(-)