| Message ID | 20231211122322.15815-2-julien@xen.org (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | xen/x86: Optimize timer_irq_works() | expand |
On 11.12.2023 13:23, Julien Grall wrote: > --- a/docs/misc/xen-command-line.pandoc > +++ b/docs/misc/xen-command-line.pandoc > @@ -2535,6 +2535,17 @@ pages) must also be specified via the tbuf_size parameter. > ### tickle_one_idle_cpu > > `= <boolean>` > > +### pit-irq-works (x86) > +> `=<boolean>` > + > +> Default: `false` > + > +Disables the code which tests for broken timer IRQ sources. Enabling > +this option will reduce boot time on HW where the timer works properly. > + > +If the system is unstable when enabling the option, then it means you > +may have a broken HW and therefore the testing cannot be be skipped. > + > ### timer_slop > > `= <integer>` With the rename this now needs to move up to retain sorting. > --- a/xen/arch/x86/io_apic.c > +++ b/xen/arch/x86/io_apic.c > @@ -57,6 +57,14 @@ bool __initdata ioapic_ack_forced; > int __read_mostly nr_ioapic_entries[MAX_IO_APICS]; > int __read_mostly nr_ioapics; > > +/* > + * The logic to check if the timer is working is expensive. So allow > + * the admin to bypass it if they know their platform doesn't have > + * a buggy timer. > + */ > +static bool __initdata pit_irq_works; > +boolean_param("pit-irq-works", pit_irq_works); > + > /* > * Rough estimation of how many shared IRQs there are, can > * be changed anytime. > @@ -1502,6 +1510,9 @@ static int __init timer_irq_works(void) > { > unsigned long t1, flags; > > + if ( pit_irq_works ) > + return 1; When the check is placed here, what exactly use of the option means is system dependent. I consider this somewhat risky, so I'd prefer if the check was put on the "normal" path in check_timer(). That way it'll affect only the one case which we can generally consider "known good", but not the cases where the virtual wire setups are being probed. I.e. if (pin1 != -1) { /* * Ok, does IRQ0 through the IOAPIC work? */ unmask_IO_APIC_irq(irq_to_desc(0)); if (pit_irq_works || timer_irq_works()) { local_irq_restore(flags); return; } Plus this way changes to the various fallback paths can also be done without needing to consider users who might be making use of the new option. Jan
Hi, On 14/12/2023 10:10, Jan Beulich wrote: > On 11.12.2023 13:23, Julien Grall wrote: >> --- a/docs/misc/xen-command-line.pandoc >> +++ b/docs/misc/xen-command-line.pandoc >> @@ -2535,6 +2535,17 @@ pages) must also be specified via the tbuf_size parameter. >> ### tickle_one_idle_cpu >> > `= <boolean>` >> >> +### pit-irq-works (x86) >> +> `=<boolean>` >> + >> +> Default: `false` >> + >> +Disables the code which tests for broken timer IRQ sources. Enabling >> +this option will reduce boot time on HW where the timer works properly. >> + >> +If the system is unstable when enabling the option, then it means you >> +may have a broken HW and therefore the testing cannot be be skipped. >> + >> ### timer_slop >> > `= <integer>` > > With the rename this now needs to move up to retain sorting. Ok. > >> --- a/xen/arch/x86/io_apic.c >> +++ b/xen/arch/x86/io_apic.c >> @@ -57,6 +57,14 @@ bool __initdata ioapic_ack_forced; >> int __read_mostly nr_ioapic_entries[MAX_IO_APICS]; >> int __read_mostly nr_ioapics; >> >> +/* >> + * The logic to check if the timer is working is expensive. So allow >> + * the admin to bypass it if they know their platform doesn't have >> + * a buggy timer. >> + */ >> +static bool __initdata pit_irq_works; >> +boolean_param("pit-irq-works", pit_irq_works); >> + >> /* >> * Rough estimation of how many shared IRQs there are, can >> * be changed anytime. >> @@ -1502,6 +1510,9 @@ static int __init timer_irq_works(void) >> { >> unsigned long t1, flags; >> >> + if ( pit_irq_works ) >> + return 1; > > When the check is placed here, what exactly use of the option means is > system dependent. I consider this somewhat risky, so I'd prefer if the > check was put on the "normal" path in check_timer(). That way it'll > affect only the one case which we can generally consider "known good", > but not the cases where the virtual wire setups are being probed. I.e. I am not against restricting when we allow skipping the timer check. But in that case, I wonder why Linux is doing it differently? After all, this code is heavily borrowed from Linux. So shouldn't we follow what they are doing? Cheers,
On 14.12.2023 11:14, Julien Grall wrote: > On 14/12/2023 10:10, Jan Beulich wrote: >> On 11.12.2023 13:23, Julien Grall wrote: >>> --- a/xen/arch/x86/io_apic.c >>> +++ b/xen/arch/x86/io_apic.c >>> @@ -57,6 +57,14 @@ bool __initdata ioapic_ack_forced; >>> int __read_mostly nr_ioapic_entries[MAX_IO_APICS]; >>> int __read_mostly nr_ioapics; >>> >>> +/* >>> + * The logic to check if the timer is working is expensive. So allow >>> + * the admin to bypass it if they know their platform doesn't have >>> + * a buggy timer. >>> + */ >>> +static bool __initdata pit_irq_works; >>> +boolean_param("pit-irq-works", pit_irq_works); >>> + >>> /* >>> * Rough estimation of how many shared IRQs there are, can >>> * be changed anytime. >>> @@ -1502,6 +1510,9 @@ static int __init timer_irq_works(void) >>> { >>> unsigned long t1, flags; >>> >>> + if ( pit_irq_works ) >>> + return 1; >> >> When the check is placed here, what exactly use of the option means is >> system dependent. I consider this somewhat risky, so I'd prefer if the >> check was put on the "normal" path in check_timer(). That way it'll >> affect only the one case which we can generally consider "known good", >> but not the cases where the virtual wire setups are being probed. I.e. > > I am not against restricting when we allow skipping the timer check. But > in that case, I wonder why Linux is doing it differently? Sadly Linux'es git history doesn't go back far enough (begins only at past 2.6.11), so I can't (easily) find the patch (and description) for the x86-64 change. The later i386 change is justified mainly by paravirt needs, so isn't applicable here. I wouldn't therefore exclude that my point above wasn't even taken into consideration. Furthermore their command line option is "no_timer_check", which to me firmly says "don't check" without regard to whether the source (PIT) is actually okay. That's different with the option name you (imo validly) chose. Jan
Hi Jan, On 14/12/2023 10:35, Jan Beulich wrote: > On 14.12.2023 11:14, Julien Grall wrote: >> On 14/12/2023 10:10, Jan Beulich wrote: >>> On 11.12.2023 13:23, Julien Grall wrote: >>>> --- a/xen/arch/x86/io_apic.c >>>> +++ b/xen/arch/x86/io_apic.c >>>> @@ -57,6 +57,14 @@ bool __initdata ioapic_ack_forced; >>>> int __read_mostly nr_ioapic_entries[MAX_IO_APICS]; >>>> int __read_mostly nr_ioapics; >>>> >>>> +/* >>>> + * The logic to check if the timer is working is expensive. So allow >>>> + * the admin to bypass it if they know their platform doesn't have >>>> + * a buggy timer. >>>> + */ >>>> +static bool __initdata pit_irq_works; >>>> +boolean_param("pit-irq-works", pit_irq_works); >>>> + >>>> /* >>>> * Rough estimation of how many shared IRQs there are, can >>>> * be changed anytime. >>>> @@ -1502,6 +1510,9 @@ static int __init timer_irq_works(void) >>>> { >>>> unsigned long t1, flags; >>>> >>>> + if ( pit_irq_works ) >>>> + return 1; >>> >>> When the check is placed here, what exactly use of the option means is >>> system dependent. I consider this somewhat risky, so I'd prefer if the >>> check was put on the "normal" path in check_timer(). That way it'll >>> affect only the one case which we can generally consider "known good", >>> but not the cases where the virtual wire setups are being probed. I.e. By "known good", do you mean the following: diff --git a/xen/arch/x86/io_apic.c b/xen/arch/x86/io_apic.c index c89fbed8d675..c39d39ee951a 100644 --- a/xen/arch/x86/io_apic.c +++ b/xen/arch/x86/io_apic.c @@ -1960,7 +1959,8 @@ static void __init check_timer(void) * Ok, does IRQ0 through the IOAPIC work? */ unmask_IO_APIC_irq(irq_to_desc(0)); - if (timer_irq_works()) { + if (pit_irq_works || timer_irq_works()) { + printk("====== pirq_irq_works %d =====\n", pit_irq_works); local_irq_restore(flags); return; } >> >> I am not against restricting when we allow skipping the timer check. But >> in that case, I wonder why Linux is doing it differently? > > Sadly Linux'es git history doesn't go back far enough (begins only at past > 2.6.11), so I can't (easily) find the patch (and description) for the x86-64 > change. The later i386 change is justified mainly by paravirt needs, so > isn't applicable here. I wouldn't therefore exclude that my point above > wasn't even taken into consideration. Furthermore their command line option > is "no_timer_check", which to me firmly says "don't check" without regard to > whether the source (PIT) is actually okay. That's different with the option > name you (imo validly) chose. Just to note that the name was suggested by Roger. I have to admit that I didn't check if this made sense for the existing placement. Anyway, I tested the change on the HW where I wanted to skip the timer check. And I can confirm this is still skipping the timer check. So I will send a new version with the diff above and some updated comments. Cheers,
On 02.01.2024 20:09, Julien Grall wrote: > Hi Jan, > > On 14/12/2023 10:35, Jan Beulich wrote: >> On 14.12.2023 11:14, Julien Grall wrote: >>> On 14/12/2023 10:10, Jan Beulich wrote: >>>> On 11.12.2023 13:23, Julien Grall wrote: >>>>> --- a/xen/arch/x86/io_apic.c >>>>> +++ b/xen/arch/x86/io_apic.c >>>>> @@ -57,6 +57,14 @@ bool __initdata ioapic_ack_forced; >>>>> int __read_mostly nr_ioapic_entries[MAX_IO_APICS]; >>>>> int __read_mostly nr_ioapics; >>>>> >>>>> +/* >>>>> + * The logic to check if the timer is working is expensive. So allow >>>>> + * the admin to bypass it if they know their platform doesn't have >>>>> + * a buggy timer. >>>>> + */ >>>>> +static bool __initdata pit_irq_works; >>>>> +boolean_param("pit-irq-works", pit_irq_works); >>>>> + >>>>> /* >>>>> * Rough estimation of how many shared IRQs there are, can >>>>> * be changed anytime. >>>>> @@ -1502,6 +1510,9 @@ static int __init timer_irq_works(void) >>>>> { >>>>> unsigned long t1, flags; >>>>> >>>>> + if ( pit_irq_works ) >>>>> + return 1; >>>> >>>> When the check is placed here, what exactly use of the option means is >>>> system dependent. I consider this somewhat risky, so I'd prefer if the >>>> check was put on the "normal" path in check_timer(). That way it'll >>>> affect only the one case which we can generally consider "known good", >>>> but not the cases where the virtual wire setups are being probed. I.e. > > By "known good", do you mean the following: > > diff --git a/xen/arch/x86/io_apic.c b/xen/arch/x86/io_apic.c > index c89fbed8d675..c39d39ee951a 100644 > --- a/xen/arch/x86/io_apic.c > +++ b/xen/arch/x86/io_apic.c > @@ -1960,7 +1959,8 @@ static void __init check_timer(void) > * Ok, does IRQ0 through the IOAPIC work? > */ > unmask_IO_APIC_irq(irq_to_desc(0)); > - if (timer_irq_works()) { > + if (pit_irq_works || timer_irq_works()) { > + printk("====== pirq_irq_works %d =====\n", pit_irq_works); > local_irq_restore(flags); > return; > } Yes. >>> I am not against restricting when we allow skipping the timer check. But >>> in that case, I wonder why Linux is doing it differently? >> >> Sadly Linux'es git history doesn't go back far enough (begins only at past >> 2.6.11), so I can't (easily) find the patch (and description) for the x86-64 >> change. The later i386 change is justified mainly by paravirt needs, so >> isn't applicable here. I wouldn't therefore exclude that my point above >> wasn't even taken into consideration. Furthermore their command line option >> is "no_timer_check", which to me firmly says "don't check" without regard to >> whether the source (PIT) is actually okay. That's different with the option >> name you (imo validly) chose. > > Just to note that the name was suggested by Roger. I have to admit that > I didn't check if this made sense for the existing placement. Roger, thoughts? Jan > Anyway, I tested the change on the HW where I wanted to skip the timer > check. And I can confirm this is still skipping the timer check. > > So I will send a new version with the diff above and some updated comments. > > Cheers, >
On Thu, Jan 04, 2024 at 09:54:30AM +0100, Jan Beulich wrote: > On 02.01.2024 20:09, Julien Grall wrote: > > Hi Jan, > > > > On 14/12/2023 10:35, Jan Beulich wrote: > >> On 14.12.2023 11:14, Julien Grall wrote: > >>> On 14/12/2023 10:10, Jan Beulich wrote: > >>>> On 11.12.2023 13:23, Julien Grall wrote: > >>>>> --- a/xen/arch/x86/io_apic.c > >>>>> +++ b/xen/arch/x86/io_apic.c > >>>>> @@ -57,6 +57,14 @@ bool __initdata ioapic_ack_forced; > >>>>> int __read_mostly nr_ioapic_entries[MAX_IO_APICS]; > >>>>> int __read_mostly nr_ioapics; > >>>>> > >>>>> +/* > >>>>> + * The logic to check if the timer is working is expensive. So allow > >>>>> + * the admin to bypass it if they know their platform doesn't have > >>>>> + * a buggy timer. > >>>>> + */ > >>>>> +static bool __initdata pit_irq_works; > >>>>> +boolean_param("pit-irq-works", pit_irq_works); > >>>>> + > >>>>> /* > >>>>> * Rough estimation of how many shared IRQs there are, can > >>>>> * be changed anytime. > >>>>> @@ -1502,6 +1510,9 @@ static int __init timer_irq_works(void) > >>>>> { > >>>>> unsigned long t1, flags; > >>>>> > >>>>> + if ( pit_irq_works ) > >>>>> + return 1; > >>>> > >>>> When the check is placed here, what exactly use of the option means is > >>>> system dependent. I consider this somewhat risky, so I'd prefer if the > >>>> check was put on the "normal" path in check_timer(). That way it'll > >>>> affect only the one case which we can generally consider "known good", > >>>> but not the cases where the virtual wire setups are being probed. I.e. > > > > By "known good", do you mean the following: > > > > diff --git a/xen/arch/x86/io_apic.c b/xen/arch/x86/io_apic.c > > index c89fbed8d675..c39d39ee951a 100644 > > --- a/xen/arch/x86/io_apic.c > > +++ b/xen/arch/x86/io_apic.c > > @@ -1960,7 +1959,8 @@ static void __init check_timer(void) > > * Ok, does IRQ0 through the IOAPIC work? > > */ > > unmask_IO_APIC_irq(irq_to_desc(0)); > > - if (timer_irq_works()) { > > + if (pit_irq_works || timer_irq_works()) { > > + printk("====== pirq_irq_works %d =====\n", pit_irq_works); > > local_irq_restore(flags); > > return; > > } > > Yes. > > >>> I am not against restricting when we allow skipping the timer check. But > >>> in that case, I wonder why Linux is doing it differently? > >> > >> Sadly Linux'es git history doesn't go back far enough (begins only at past > >> 2.6.11), so I can't (easily) find the patch (and description) for the x86-64 > >> change. The later i386 change is justified mainly by paravirt needs, so > >> isn't applicable here. I wouldn't therefore exclude that my point above > >> wasn't even taken into consideration. Furthermore their command line option > >> is "no_timer_check", which to me firmly says "don't check" without regard to > >> whether the source (PIT) is actually okay. That's different with the option > >> name you (imo validly) chose. > > > > Just to note that the name was suggested by Roger. I have to admit that > > I didn't check if this made sense for the existing placement. > > Roger, thoughts? Right, with the usage of HPET in legacy replacement mode we are no longer exclusively testing the PIT, so might make sense to use a more generic name, timer-irq-works or some such. Thanks, Roger.
diff --git a/docs/misc/xen-command-line.pandoc b/docs/misc/xen-command-line.pandoc index 8e65f8bd18bf..c382b061b302 100644 --- a/docs/misc/xen-command-line.pandoc +++ b/docs/misc/xen-command-line.pandoc @@ -2535,6 +2535,17 @@ pages) must also be specified via the tbuf_size parameter. ### tickle_one_idle_cpu > `= <boolean>` +### pit-irq-works (x86) +> `=<boolean>` + +> Default: `false` + +Disables the code which tests for broken timer IRQ sources. Enabling +this option will reduce boot time on HW where the timer works properly. + +If the system is unstable when enabling the option, then it means you +may have a broken HW and therefore the testing cannot be be skipped. + ### timer_slop > `= <integer>` diff --git a/xen/arch/x86/io_apic.c b/xen/arch/x86/io_apic.c index d11c880544e6..238b6c1c2837 100644 --- a/xen/arch/x86/io_apic.c +++ b/xen/arch/x86/io_apic.c @@ -57,6 +57,14 @@ bool __initdata ioapic_ack_forced; int __read_mostly nr_ioapic_entries[MAX_IO_APICS]; int __read_mostly nr_ioapics; +/* + * The logic to check if the timer is working is expensive. So allow + * the admin to bypass it if they know their platform doesn't have + * a buggy timer. + */ +static bool __initdata pit_irq_works; +boolean_param("pit-irq-works", pit_irq_works); + /* * Rough estimation of how many shared IRQs there are, can * be changed anytime. @@ -1502,6 +1510,9 @@ static int __init timer_irq_works(void) { unsigned long t1, flags; + if ( pit_irq_works ) + return 1; + t1 = ACCESS_ONCE(pit0_ticks); local_save_flags(flags);