Re: [Qemu-devel] [RFC PATCH] spapr: support time base offset migration
On 13.09.2013, at 00:20, David Gibson wrote: > On Mon, Sep 09, 2013 at 08:06:53AM +0200, Alexander Graf wrote: >> >> >> Am 09.09.2013 um 07:58 schrieb Alexey Kardashevskiy : >> >>> On 09/09/2013 03:50 PM, Alexander Graf wrote: Am 09.09.2013 um 04:40 schrieb Alexey Kardashevskiy : > On 09/06/2013 01:11 AM, Alexander Graf wrote: >> >> On 05.09.2013, at 16:26, Benjamin Herrenschmidt wrote: >> >>> On Thu, 2013-09-05 at 16:14 +0200, Andreas Färber wrote: >>> Are you thinking of POWER8 having a different frequency than POWER8 in compat mode? Because migration from one -cpu to another is not supported elsewhere. Even if we want to migrate from one POWER7 revision to another, we should let the destination use the revision of the source (guest ABI!), via property if need be. Anything else will lead to confusion as to what is supported and what is not. That -cpu host is the default for convenience shouldn't relieve admins/libvirt to think about sensible setups like they have to on x86. >>> >>> Besides POWER8 uses 512Mhz too :-) It's been architected so it's >>> unlikely to change from now on. >> >> Ok, ditch the tb frequency then. We can always default to 512Mhz when >> it's absent. > > > QEMU uses what kvmppc_get_tbfreq() returns. And ppc_tb_t carries this > value. It does not use 512MHz automatically but migration should always > assume 512MHz :-/ > > If we remove it, I would like to add assert(tb_freq!=512MHz) into > timebase_pre_save() and timebase_post_load(), is that ok? Good point. This would break TCG migration, right? >>> >>> >>> In TCG we do not need any timebase offset at all, the time should migrate >>> as is, no? :) >> >> I hope so, but we need to make sure we don't assert() in that case :). > > Hrm.. that doesn't sound right. Whether you have KVM or TCG, what you > need in the migration stream is enough data points that you can deduce > the guest timebase from the host real time. On KVM that translates > into a diff between guest and host timebase, but on TCG you'll need to > implement that differently I'd claim that for TCG we don't care about accuracy of the time base across live migration, so I'd be happy to simply ignore the adjustments we'd do with KVM there. Otherwise things would get heavily complicated. > >>> >>> It could break something like power5 migration under PR KVM, do not know... >> >> Well, today we don't support full migration with PR KVM yet - IIRC we don't >> have access to all state. But that may change in the future. >> >> I think it's ok to restrict live migration to machines with the same >> tb frequency when kvm is enabled. Whether you implement it through a >> hardcoded 512Mhz or through a timebase value that gets live migrated >> and then compared is up to you - both ways work for me :). > > I don't think it's possible to allow KVM migration between hosts of > different tb frequency. But I think encoding the frequency in the > stream and verifying it is a better option. It might be useful for > migration on 970, and it might be useful for TCG migration and it > might be useful for TCG<->KVM migration. (For a TCG migration > destination you *can* potentially set the tb frequency according to > what you're told). Yup, IIUC that's what Alexey's gut feeling told him too (and I concur) and is what he's doing ;). Alex
Re: [Qemu-devel] [RFC PATCH] spapr: support time base offset migration
On Mon, Sep 09, 2013 at 08:06:53AM +0200, Alexander Graf wrote: > > > Am 09.09.2013 um 07:58 schrieb Alexey Kardashevskiy : > > > On 09/09/2013 03:50 PM, Alexander Graf wrote: > >> > >> > >> Am 09.09.2013 um 04:40 schrieb Alexey Kardashevskiy : > >> > >>> On 09/06/2013 01:11 AM, Alexander Graf wrote: > > On 05.09.2013, at 16:26, Benjamin Herrenschmidt wrote: > > > On Thu, 2013-09-05 at 16:14 +0200, Andreas Färber wrote: > > > >> Are you thinking of POWER8 having a different frequency than POWER8 in > >> compat mode? Because migration from one -cpu to another is not > >> supported > >> elsewhere. > >> > >> Even if we want to migrate from one POWER7 revision to another, we > >> should let the destination use the revision of the source (guest ABI!), > >> via property if need be. Anything else will lead to confusion as to > >> what > >> is supported and what is not. That -cpu host is the default for > >> convenience shouldn't relieve admins/libvirt to think about sensible > >> setups like they have to on x86. > > > > Besides POWER8 uses 512Mhz too :-) It's been architected so it's > > unlikely to change from now on. > > Ok, ditch the tb frequency then. We can always default to 512Mhz when > it's absent. > >>> > >>> > >>> QEMU uses what kvmppc_get_tbfreq() returns. And ppc_tb_t carries this > >>> value. It does not use 512MHz automatically but migration should always > >>> assume 512MHz :-/ > >>> > >>> If we remove it, I would like to add assert(tb_freq!=512MHz) into > >>> timebase_pre_save() and timebase_post_load(), is that ok? > >> > >> Good point. This would break TCG migration, right? > > > > > > In TCG we do not need any timebase offset at all, the time should migrate > > as is, no? :) > > I hope so, but we need to make sure we don't assert() in that case :). Hrm.. that doesn't sound right. Whether you have KVM or TCG, what you need in the migration stream is enough data points that you can deduce the guest timebase from the host real time. On KVM that translates into a diff between guest and host timebase, but on TCG you'll need to implement that differently > > > > It could break something like power5 migration under PR KVM, do not know... > > Well, today we don't support full migration with PR KVM yet - IIRC we don't > have access to all state. But that may change in the future. > > I think it's ok to restrict live migration to machines with the same > tb frequency when kvm is enabled. Whether you implement it through a > hardcoded 512Mhz or through a timebase value that gets live migrated > and then compared is up to you - both ways work for me :). I don't think it's possible to allow KVM migration between hosts of different tb frequency. But I think encoding the frequency in the stream and verifying it is a better option. It might be useful for migration on 970, and it might be useful for TCG migration and it might be useful for TCG<->KVM migration. (For a TCG migration destination you *can* potentially set the tb frequency according to what you're told). -- David Gibson| I'll have my music baroque, and my code david AT gibson.dropbear.id.au | minimalist, thank you. NOT _the_ _other_ | _way_ _around_! http://www.ozlabs.org/~dgibson pgpNzv36uk_j6.pgp Description: PGP signature
Re: [Qemu-devel] [RFC PATCH] spapr: support time base offset migration
On 09.09.2013, at 11:38, Benjamin Herrenschmidt wrote: > On Mon, 2013-09-09 at 11:32 +0200, Alexander Graf wrote: >> On 09.09.2013, at 11:29, Benjamin Herrenschmidt wrote: >> >>> On Mon, 2013-09-09 at 08:06 +0200, Alexander Graf wrote: I think it's ok to restrict live migration to machines with the same tb frequency when kvm is enabled. Whether you implement it through a hardcoded 512Mhz or through a timebase value that gets live migrated and then compared is up to you - both ways work for me :). >>> >>> The latter might be handy if we want to support migration on 970, though >>> we don't have the TBU40 stuff there so adjusting the TB in the host >>> kernel would be ... problematic. >> >> Well, we could save/restore TB when we enter/exit the guest, no? > > Hard to do without introducing drift... We could probably quantify the drift through DEC. But I'm personally not too eager to worry about live migration on 970 :). Alex
Re: [Qemu-devel] [RFC PATCH] spapr: support time base offset migration
On Mon, 2013-09-09 at 11:32 +0200, Alexander Graf wrote: > On 09.09.2013, at 11:29, Benjamin Herrenschmidt wrote: > > > On Mon, 2013-09-09 at 08:06 +0200, Alexander Graf wrote: > >> I think it's ok to restrict live migration to machines with the same > >> tb frequency when kvm is enabled. Whether you implement it through a > >> hardcoded 512Mhz or through a timebase value that gets live migrated > >> and then compared is up to you - both ways work for me :). > > > > The latter might be handy if we want to support migration on 970, though > > we don't have the TBU40 stuff there so adjusting the TB in the host > > kernel would be ... problematic. > > Well, we could save/restore TB when we enter/exit the guest, no? Hard to do without introducing drift... > But I really only want to have something that doesn't block the door for > someone who wants to enable things. > > > Alex
Re: [Qemu-devel] [RFC PATCH] spapr: support time base offset migration
On Mon, 2013-09-09 at 08:06 +0200, Alexander Graf wrote: > I think it's ok to restrict live migration to machines with the same > tb frequency when kvm is enabled. Whether you implement it through a > hardcoded 512Mhz or through a timebase value that gets live migrated > and then compared is up to you - both ways work for me :). The latter might be handy if we want to support migration on 970, though we don't have the TBU40 stuff there so adjusting the TB in the host kernel would be ... problematic. Cheers, Ben.
Re: [Qemu-devel] [RFC PATCH] spapr: support time base offset migration
On 09.09.2013, at 11:29, Benjamin Herrenschmidt wrote: > On Mon, 2013-09-09 at 08:06 +0200, Alexander Graf wrote: >> I think it's ok to restrict live migration to machines with the same >> tb frequency when kvm is enabled. Whether you implement it through a >> hardcoded 512Mhz or through a timebase value that gets live migrated >> and then compared is up to you - both ways work for me :). > > The latter might be handy if we want to support migration on 970, though > we don't have the TBU40 stuff there so adjusting the TB in the host > kernel would be ... problematic. Well, we could save/restore TB when we enter/exit the guest, no? But I really only want to have something that doesn't block the door for someone who wants to enable things. Alex
Re: [Qemu-devel] [RFC PATCH] spapr: support time base offset migration
Am 09.09.2013 um 07:58 schrieb Alexey Kardashevskiy : > On 09/09/2013 03:50 PM, Alexander Graf wrote: >> >> >> Am 09.09.2013 um 04:40 schrieb Alexey Kardashevskiy : >> >>> On 09/06/2013 01:11 AM, Alexander Graf wrote: On 05.09.2013, at 16:26, Benjamin Herrenschmidt wrote: > On Thu, 2013-09-05 at 16:14 +0200, Andreas Färber wrote: > >> Are you thinking of POWER8 having a different frequency than POWER8 in >> compat mode? Because migration from one -cpu to another is not supported >> elsewhere. >> >> Even if we want to migrate from one POWER7 revision to another, we >> should let the destination use the revision of the source (guest ABI!), >> via property if need be. Anything else will lead to confusion as to what >> is supported and what is not. That -cpu host is the default for >> convenience shouldn't relieve admins/libvirt to think about sensible >> setups like they have to on x86. > > Besides POWER8 uses 512Mhz too :-) It's been architected so it's > unlikely to change from now on. Ok, ditch the tb frequency then. We can always default to 512Mhz when it's absent. >>> >>> >>> QEMU uses what kvmppc_get_tbfreq() returns. And ppc_tb_t carries this >>> value. It does not use 512MHz automatically but migration should always >>> assume 512MHz :-/ >>> >>> If we remove it, I would like to add assert(tb_freq!=512MHz) into >>> timebase_pre_save() and timebase_post_load(), is that ok? >> >> Good point. This would break TCG migration, right? > > > In TCG we do not need any timebase offset at all, the time should migrate > as is, no? :) I hope so, but we need to make sure we don't assert() in that case :). > > It could break something like power5 migration under PR KVM, do not know... Well, today we don't support full migration with PR KVM yet - IIRC we don't have access to all state. But that may change in the future. I think it's ok to restrict live migration to machines with the same tb frequency when kvm is enabled. Whether you implement it through a hardcoded 512Mhz or through a timebase value that gets live migrated and then compared is up to you - both ways work for me :). Alex > > > -- > Alexey
Re: [Qemu-devel] [RFC PATCH] spapr: support time base offset migration
On 09/09/2013 03:50 PM, Alexander Graf wrote: > > > Am 09.09.2013 um 04:40 schrieb Alexey Kardashevskiy : > >> On 09/06/2013 01:11 AM, Alexander Graf wrote: >>> >>> On 05.09.2013, at 16:26, Benjamin Herrenschmidt wrote: >>> On Thu, 2013-09-05 at 16:14 +0200, Andreas Färber wrote: > Are you thinking of POWER8 having a different frequency than POWER8 in > compat mode? Because migration from one -cpu to another is not supported > elsewhere. > > Even if we want to migrate from one POWER7 revision to another, we > should let the destination use the revision of the source (guest ABI!), > via property if need be. Anything else will lead to confusion as to what > is supported and what is not. That -cpu host is the default for > convenience shouldn't relieve admins/libvirt to think about sensible > setups like they have to on x86. Besides POWER8 uses 512Mhz too :-) It's been architected so it's unlikely to change from now on. >>> >>> Ok, ditch the tb frequency then. We can always default to 512Mhz when it's >>> absent. >> >> >> QEMU uses what kvmppc_get_tbfreq() returns. And ppc_tb_t carries this >> value. It does not use 512MHz automatically but migration should always >> assume 512MHz :-/ >> >> If we remove it, I would like to add assert(tb_freq!=512MHz) into >> timebase_pre_save() and timebase_post_load(), is that ok? > > Good point. This would break TCG migration, right? In TCG we do not need any timebase offset at all, the time should migrate as is, no? :) It could break something like power5 migration under PR KVM, do not know... -- Alexey
Re: [Qemu-devel] [RFC PATCH] spapr: support time base offset migration
Am 09.09.2013 um 04:40 schrieb Alexey Kardashevskiy : > On 09/06/2013 01:11 AM, Alexander Graf wrote: >> >> On 05.09.2013, at 16:26, Benjamin Herrenschmidt wrote: >> >>> On Thu, 2013-09-05 at 16:14 +0200, Andreas Färber wrote: >>> Are you thinking of POWER8 having a different frequency than POWER8 in compat mode? Because migration from one -cpu to another is not supported elsewhere. Even if we want to migrate from one POWER7 revision to another, we should let the destination use the revision of the source (guest ABI!), via property if need be. Anything else will lead to confusion as to what is supported and what is not. That -cpu host is the default for convenience shouldn't relieve admins/libvirt to think about sensible setups like they have to on x86. >>> >>> Besides POWER8 uses 512Mhz too :-) It's been architected so it's >>> unlikely to change from now on. >> >> Ok, ditch the tb frequency then. We can always default to 512Mhz when it's >> absent. > > > QEMU uses what kvmppc_get_tbfreq() returns. And ppc_tb_t carries this > value. It does not use 512MHz automatically but migration should always > assume 512MHz :-/ > > If we remove it, I would like to add assert(tb_freq!=512MHz) into > timebase_pre_save() and timebase_post_load(), is that ok? Good point. This would break TCG migration, right? Alex > > > -- > Alexey
Re: [Qemu-devel] [RFC PATCH] spapr: support time base offset migration
On 09/06/2013 01:11 AM, Alexander Graf wrote: > > On 05.09.2013, at 16:26, Benjamin Herrenschmidt wrote: > >> On Thu, 2013-09-05 at 16:14 +0200, Andreas Färber wrote: >> >>> Are you thinking of POWER8 having a different frequency than POWER8 in >>> compat mode? Because migration from one -cpu to another is not supported >>> elsewhere. >>> >>> Even if we want to migrate from one POWER7 revision to another, we >>> should let the destination use the revision of the source (guest ABI!), >>> via property if need be. Anything else will lead to confusion as to what >>> is supported and what is not. That -cpu host is the default for >>> convenience shouldn't relieve admins/libvirt to think about sensible >>> setups like they have to on x86. >> >> Besides POWER8 uses 512Mhz too :-) It's been architected so it's >> unlikely to change from now on. > > Ok, ditch the tb frequency then. We can always default to 512Mhz when it's > absent. QEMU uses what kvmppc_get_tbfreq() returns. And ppc_tb_t carries this value. It does not use 512MHz automatically but migration should always assume 512MHz :-/ If we remove it, I would like to add assert(tb_freq!=512MHz) into timebase_pre_save() and timebase_post_load(), is that ok? -- Alexey
Re: [Qemu-devel] [RFC PATCH] spapr: support time base offset migration
On Thu, Sep 05, 2013 at 02:54:46PM +1000, Alexey Kardashevskiy wrote:
> On 09/05/2013 02:30 PM, David Gibson wrote:
> > On Tue, Sep 03, 2013 at 05:31:42PM +1000, Alexey Kardashevskiy wrote:
> >> This allows guests to have a different timebase origin from the host.
> >>
> >> This is needed for migration, where a guest can migrate from one host
> >> to another and the two hosts might have a different timebase origin.
> >> However, the timebase seen by the guest must not go backwards, and
> >> should go forwards only by a small amount corresponding to the time
> >> taken for the migration.
> >>
> >> This is only supported for recent POWER hardware which has the TBU40
> >> (timebase upper 40 bits) register. That includes POWER6, 7, 8 but not
> >> 970.
> >>
> >> This adds kvm_access_one_reg() to access a special register which is not
> >> in env->spr.
> >>
> >> The feature must be present in the host kernel.
> >>
> >> Signed-off-by: Alexey Kardashevskiy
> >> ---
> >>
> >> This is an RFC but not a final patch. Can break something but I just do
> >> not see what.
> >>
> >> ---
> >> hw/ppc/ppc.c | 49
> >> +
> >> include/hw/ppc/ppc.h | 4
> >> target-ppc/kvm.c | 23 +++
> >> target-ppc/machine.c | 44
> >> trace-events | 3 +++
> >> 5 files changed, 123 insertions(+)
> >>
> >> diff --git a/hw/ppc/ppc.c b/hw/ppc/ppc.c
> >> index 1e3cab3..7d08c9a 100644
> >> --- a/hw/ppc/ppc.c
> >> +++ b/hw/ppc/ppc.c
> >> @@ -31,6 +31,7 @@
> >> #include "hw/loader.h"
> >> #include "sysemu/kvm.h"
> >> #include "kvm_ppc.h"
> >> +#include "trace.h"
> >>
> >> //#define PPC_DEBUG_IRQ
> >> #define PPC_DEBUG_TB
> >> @@ -796,6 +797,54 @@ static void cpu_ppc_set_tb_clk (void *opaque,
> >> uint32_t freq)
> >> cpu_ppc_store_purr(cpu, 0xULL);
> >> }
> >>
> >> +/*
> >> + * Calculate timebase on the destination side of migration
> >
> >> + * We calculate new timebase offset as shown below:
> >> + * 1) Gtb2 = Gtb1 + max(tod2 - tod1, 0)
> >> + *Gtb2 = tb2 + off2
> >> + *Gtb1 = tb1 + off1
> >> + * 2) tb2 + off2 = tb1 + off1 + max(tod2 - tod1, 0)
> >> + * 3) off2 = tb1 - tb2 + off1 + max(tod2 - tod1, 0)
> >> + *
> >> + * where:
> >> + * Gtb2 - destination guest timebase
> >> + * tb2 - destination host timebase
> >> + * off2 - destination timebase offset
> >> + * tod2 - destination time of the day
> >> + * Gtb1 - source guest timebase
> >> + * tb1 - source host timebase
> >> + * off1 - source timebase offset
> >> + * tod1 - source time of the day
> >> + *
> >> + * The result we want is in @off2
> >> + *
> >> + * Two conditions must be met for @off2:
> >> + * 1) off2 must be multiple of 2^24 ticks as it will be set via TBU40 SPR
> >> + * 2) Gtb2 >= Gtb1
> >
> > What about the TCG case, where there is not host timebase, only a a
> > host system clock?
>
>
> cpu_get_real_ticks() returns ticks, this is what the patch cares about.
> What is the difference between KVM and TCG here?
Hrm. It may just be a question of variable naming and comments. But
my point is that "host tb" is not a concept that always exists, so it
shouldn't appear in anything external like a migration stream.
> >> + */
> >> +void cpu_ppc_adjust_tb_offset(ppc_tb_t *tb_env)
> >> +{
> >> +uint64_t tb2, tod2, off2;
> >> +int ratio = tb_env->tb_freq / 100;
> >> +struct timeval tv;
> >> +
> >> +tb2 = cpu_get_real_ticks();
> >> +gettimeofday(&tv, NULL);
> >> +tod2 = tv.tv_sec * 100 + tv.tv_usec;
> >> +
> >> +off2 = tb_env->timebase - tb2 + tb_env->tb_offset;
> >> +if (tod2 > tb_env->time_of_the_day) {
> >> +off2 += (tod2 - tb_env->time_of_the_day) * ratio;
> >> +}
> >> +off2 = ROUND_UP(off2, 1 << 24);
> >> +
> >> +trace_ppc_tb_adjust(tb_env->tb_offset, off2,
> >> +(int64_t)off2 - tb_env->tb_offset);
> >> +
> >> +tb_env->tb_offset = off2;
> >> +}
> >> +
> >> /* Set up (once) timebase frequency (in Hz) */
> >> clk_setup_cb cpu_ppc_tb_init (CPUPPCState *env, uint32_t freq)
> >> {
> >> diff --git a/include/hw/ppc/ppc.h b/include/hw/ppc/ppc.h
> >> index 132ab97..235871c 100644
> >> --- a/include/hw/ppc/ppc.h
> >> +++ b/include/hw/ppc/ppc.h
> >> @@ -32,6 +32,9 @@ struct ppc_tb_t {
> >> uint64_t purr_start;
> >> void *opaque;
> >> uint32_t flags;
> >> +/* Cached values for live migration purposes */
> >> +uint64_t timebase;
> >> +uint64_t time_of_the_day;
> >
> > How is the time of day encoded here?
>
>
> Microseconds. I'll put a comment here, I just thought it is quite obvious
> as gettimeofday() returns microseconds.
>
>
> >> };
> >>
> >> /* PPC Timers flags */
> >> @@ -46,6 +49,7 @@ struct ppc_tb_t {
> >> */
> >>
> >> uint64_t cpu_ppc_get_tb(ppc_tb_t *tb_env, uint64_t vmclk, int64_t
> >> tb_offset);
> >> +void cpu_ppc_adjust_tb_offset
Re: [Qemu-devel] [RFC PATCH] spapr: support time base offset migration
On 05.09.2013, at 16:26, Benjamin Herrenschmidt wrote: > On Thu, 2013-09-05 at 16:14 +0200, Andreas Färber wrote: > >> Are you thinking of POWER8 having a different frequency than POWER8 in >> compat mode? Because migration from one -cpu to another is not supported >> elsewhere. >> >> Even if we want to migrate from one POWER7 revision to another, we >> should let the destination use the revision of the source (guest ABI!), >> via property if need be. Anything else will lead to confusion as to what >> is supported and what is not. That -cpu host is the default for >> convenience shouldn't relieve admins/libvirt to think about sensible >> setups like they have to on x86. > > Besides POWER8 uses 512Mhz too :-) It's been architected so it's > unlikely to change from now on. Ok, ditch the tb frequency then. We can always default to 512Mhz when it's absent. Alex
Re: [Qemu-devel] [RFC PATCH] spapr: support time base offset migration
Am 05.09.2013 15:39, schrieb Alexander Graf: > > On 05.09.2013, at 15:36, Benjamin Herrenschmidt wrote: > >> On Thu, 2013-09-05 at 14:37 +0200, Alexander Graf wrote: >> >>> Hrm, I think I'm starting to understand what this is about. So what we want >>> is >>> >>> - timebase in guest >>> - timebase frequency in guest >>> - wall clock time in host >>> >>> That way the receiving end can then take the timebase and add (new_timebase >>> - old_timebase) * tb_freq to the guest's time base. >>> >>> Which gets me to the next question. Can we modify the tb frequency in >>> guests? >> >> No. It's architected at 512Mhz however since P7 I think. Not sure how we >> did before, it's possible that P6 was the same (at least it's sourced >> from more/less the same chip TOD facility). > > I think we should transmit the tb frequency as well to at least allow us to > adjust if a later chip derives here. Are you thinking of POWER8 having a different frequency than POWER8 in compat mode? Because migration from one -cpu to another is not supported elsewhere. Even if we want to migrate from one POWER7 revision to another, we should let the destination use the revision of the source (guest ABI!), via property if need be. Anything else will lead to confusion as to what is supported and what is not. That -cpu host is the default for convenience shouldn't relieve admins/libvirt to think about sensible setups like they have to on x86. Andreas > > But yes, without frequency adjustment I see where you're coming from. We > still only need the timebase the guest sees, not the offset. But we also need > the host wall clock to allow for adjustments :). > > > Alex > -- SUSE LINUX Products GmbH, Maxfeldstr. 5, 90409 Nürnberg, Germany GF: Jeff Hawn, Jennifer Guild, Felix Imendörffer; HRB 16746 AG Nürnberg
Re: [Qemu-devel] [RFC PATCH] spapr: support time base offset migration
On Thu, 2013-09-05 at 16:14 +0200, Andreas Färber wrote: > Are you thinking of POWER8 having a different frequency than POWER8 in > compat mode? Because migration from one -cpu to another is not supported > elsewhere. > > Even if we want to migrate from one POWER7 revision to another, we > should let the destination use the revision of the source (guest ABI!), > via property if need be. Anything else will lead to confusion as to what > is supported and what is not. That -cpu host is the default for > convenience shouldn't relieve admins/libvirt to think about sensible > setups like they have to on x86. Besides POWER8 uses 512Mhz too :-) It's been architected so it's unlikely to change from now on. Cheers, Ben. > Andreas > > > > > But yes, without frequency adjustment I see where you're coming from. We > > still only need the timebase the guest sees, not the offset. But we also > > need the host wall clock to allow for adjustments :). > > > > > > Alex > > > >
Re: [Qemu-devel] [RFC PATCH] spapr: support time base offset migration
On 05.09.2013, at 15:36, Benjamin Herrenschmidt wrote: > On Thu, 2013-09-05 at 14:37 +0200, Alexander Graf wrote: > >> Hrm, I think I'm starting to understand what this is about. So what we want >> is >> >> - timebase in guest >> - timebase frequency in guest >> - wall clock time in host >> >> That way the receiving end can then take the timebase and add (new_timebase >> - old_timebase) * tb_freq to the guest's time base. >> >> Which gets me to the next question. Can we modify the tb frequency in guests? > > No. It's architected at 512Mhz however since P7 I think. Not sure how we > did before, it's possible that P6 was the same (at least it's sourced > from more/less the same chip TOD facility). I think we should transmit the tb frequency as well to at least allow us to adjust if a later chip derives here. But yes, without frequency adjustment I see where you're coming from. We still only need the timebase the guest sees, not the offset. But we also need the host wall clock to allow for adjustments :). Alex
Re: [Qemu-devel] [RFC PATCH] spapr: support time base offset migration
On Thu, 2013-09-05 at 14:37 +0200, Alexander Graf wrote: > Hrm, I think I'm starting to understand what this is about. So what we want is > > - timebase in guest > - timebase frequency in guest > - wall clock time in host > > That way the receiving end can then take the timebase and add (new_timebase - > old_timebase) * tb_freq to the guest's time base. > > Which gets me to the next question. Can we modify the tb frequency in guests? No. It's architected at 512Mhz however since P7 I think. Not sure how we did before, it's possible that P6 was the same (at least it's sourced from more/less the same chip TOD facility). Cheers, Ben.
Re: [Qemu-devel] [RFC PATCH] spapr: support time base offset migration
On Thu, 2013-09-05 at 11:58 +0200, Alexander Graf wrote: > > Yes. I do not really understand the problem here (and I am not > playing > > dump). Do you suggest sending just the guest timebase and do not > send the > > host timebase and the offset (one number instead of two)? I can do > that, > > makes sense, no problem, thanks for the idea. > > Yup, pretty much :). The receiving end should have no business in > knowing how far off the guest and the host timebase were skewed on the > sending end :). Well, yes and no ... we'd like to account for the migration latency on the timebase or the guest view of real time will get skewed since it uses the TB to maintain it's clock. So assuming both hosts are reasonably synchronized with NTP, we want a correlation guest TB / TOD in order to properly make the adjustment on the target. This may not be what Alexey implemented but I think that's what Paulus and I asked for :-) Cheers, Ben.
Re: [Qemu-devel] [RFC PATCH] spapr: support time base offset migration
On 05.09.2013, at 13:44, Benjamin Herrenschmidt wrote: > On Thu, 2013-09-05 at 11:58 +0200, Alexander Graf wrote: > >>> Yes. I do not really understand the problem here (and I am not >> playing >>> dump). Do you suggest sending just the guest timebase and do not >> send the >>> host timebase and the offset (one number instead of two)? I can do >> that, >>> makes sense, no problem, thanks for the idea. >> >> Yup, pretty much :). The receiving end should have no business in >> knowing how far off the guest and the host timebase were skewed on the >> sending end :). > > Well, yes and no ... we'd like to account for the migration latency on > the timebase or the guest view of real time will get skewed since it > uses the TB to maintain it's clock. > > So assuming both hosts are reasonably synchronized with NTP, we want a > correlation guest TB / TOD in order to properly make the adjustment on > the target. Hrm, I think I'm starting to understand what this is about. So what we want is - timebase in guest - timebase frequency in guest - wall clock time in host That way the receiving end can then take the timebase and add (new_timebase - old_timebase) * tb_freq to the guest's time base. Which gets me to the next question. Can we modify the tb frequency in guests? Alex
Re: [Qemu-devel] [RFC PATCH] spapr: support time base offset migration
On 09/05/2013 09:42 PM, Benjamin Herrenschmidt wrote: > On Thu, 2013-09-05 at 19:48 +1000, Alexey Kardashevskiy wrote: I do not have pure guest timebase in QEMU and I need it on the destination. But I have host timebase + offset to calculate it. And tb_offset is already in ppc_tb_t. It looked logical to me to send the existing field and add only the missing part. >>> >>> I still don't understand. You want the guest visible timebase in the >>> migration stream, no? >> >> >> Yes. I do not really understand the problem here (and I am not playing >> dump). Do you suggest sending just the guest timebase and do not send the >> host timebase and the offset (one number instead of two)? I can do that, >> makes sense, no problem, thanks for the idea. > > No, we want to send the guest TB and the corresponding "time of day" so that > the target can adjust the TB based on how long the migration took. Nobody is discussing time_of_the_day, it is there and transmitted. The only question is whether to send host_timebase and tb_offset (2 values) or just guest timebase (1 value). -- Alexey
Re: [Qemu-devel] [RFC PATCH] spapr: support time base offset migration
On Thu, 2013-09-05 at 19:48 +1000, Alexey Kardashevskiy wrote: > >> I do not have pure guest timebase in QEMU and I need it on the destination. > >> But I have host timebase + offset to calculate it. And tb_offset is already > >> in ppc_tb_t. It looked logical to me to send the existing field and add > >> only the missing part. > > > > I still don't understand. You want the guest visible timebase in the > > migration stream, no? > > > Yes. I do not really understand the problem here (and I am not playing > dump). Do you suggest sending just the guest timebase and do not send the > host timebase and the offset (one number instead of two)? I can do that, > makes sense, no problem, thanks for the idea. No, we want to send the guest TB and the corresponding "time of day" so that the target can adjust the TB based on how long the migration took. Cheers, Ben.
Re: [Qemu-devel] [RFC PATCH] spapr: support time base offset migration
On 05.09.2013, at 11:48, Alexey Kardashevskiy wrote:
> On 09/05/2013 07:16 PM, Alexander Graf wrote:
>>
>> On 05.09.2013, at 06:54, Alexey Kardashevskiy wrote:
>>
>>> On 09/05/2013 02:30 PM, David Gibson wrote:
[...]
>>>
> #endif /* TARGET_PPC64 */
>}
>
> @@ -1082,6 +1102,9 @@ int kvm_arch_get_registers(CPUState *cs)
>DPRINTF("Warning: Unable to get VPA information from
> KVM\n");
>}
>}
> +
> +kvm_access_one_reg(cs, 0, KVM_REG_PPC_TB_OFFSET,
> + &env->tb_env->tb_offset);
> #endif
>}
>
> diff --git a/target-ppc/machine.c b/target-ppc/machine.c
> index 12e1512..d1ffc7f 100644
> --- a/target-ppc/machine.c
> +++ b/target-ppc/machine.c
> @@ -1,5 +1,6 @@
> #include "hw/hw.h"
> #include "hw/boards.h"
> +#include "hw/ppc/ppc.h"
> #include "sysemu/kvm.h"
> #include "helper_regs.h"
>
> @@ -459,6 +460,45 @@ static const VMStateDescription vmstate_tlbmas = {
>}
> };
>
> +static void timebase_pre_save(void *opaque)
> +{
> +ppc_tb_t *tb_env = opaque;
> +struct timeval tv;
> +
> +gettimeofday(&tv, NULL);
> +tb_env->time_of_the_day = tv.tv_sec * 100 + tv.tv_usec;
> +tb_env->timebase = cpu_get_real_ticks();
> +}
> +
> +static int timebase_post_load(void *opaque, int version_id)
> +{
> +ppc_tb_t *tb_env = opaque;
> +
> +if (!tb_env) {
> +printf("NO TB!\n");
> +return -1;
> +}
> +cpu_ppc_adjust_tb_offset(tb_env);
> +
> +return 0;
> +}
> +
> +static const VMStateDescription vmstate_timebase = {
> +.name = "cpu/timebase",
> +.version_id = 1,
> +.minimum_version_id = 1,
> +.minimum_version_id_old = 1,
> +.pre_save = timebase_pre_save,
> +.post_load = timebase_post_load,
> +.fields = (VMStateField []) {
> +VMSTATE_UINT64(timebase, ppc_tb_t),
> +VMSTATE_INT64(tb_offset, ppc_tb_t),
tb_offset is inherently a local concept, since it depends on the host
timebase. So how can it belong in the migration stream?
>>>
>>>
>>> I do not have pure guest timebase in QEMU and I need it on the destination.
>>> But I have host timebase + offset to calculate it. And tb_offset is already
>>> in ppc_tb_t. It looked logical to me to send the existing field and add
>>> only the missing part.
>>
>> I still don't understand. You want the guest visible timebase in the
>> migration stream, no?
>
>
> Yes. I do not really understand the problem here (and I am not playing
> dump). Do you suggest sending just the guest timebase and do not send the
> host timebase and the offset (one number instead of two)? I can do that,
> makes sense, no problem, thanks for the idea.
Yup, pretty much :). The receiving end should have no business in knowing how
far off the guest and the host timebase were skewed on the sending end :).
Alex
Re: [Qemu-devel] [RFC PATCH] spapr: support time base offset migration
On 09/05/2013 07:16 PM, Alexander Graf wrote:
>
> On 05.09.2013, at 06:54, Alexey Kardashevskiy wrote:
>
>> On 09/05/2013 02:30 PM, David Gibson wrote:
>>> On Tue, Sep 03, 2013 at 05:31:42PM +1000, Alexey Kardashevskiy wrote:
This allows guests to have a different timebase origin from the host.
This is needed for migration, where a guest can migrate from one host
to another and the two hosts might have a different timebase origin.
However, the timebase seen by the guest must not go backwards, and
should go forwards only by a small amount corresponding to the time
taken for the migration.
This is only supported for recent POWER hardware which has the TBU40
(timebase upper 40 bits) register. That includes POWER6, 7, 8 but not
970.
This adds kvm_access_one_reg() to access a special register which is not
in env->spr.
The feature must be present in the host kernel.
Signed-off-by: Alexey Kardashevskiy
---
This is an RFC but not a final patch. Can break something but I just do
not see what.
---
hw/ppc/ppc.c | 49 +
include/hw/ppc/ppc.h | 4
target-ppc/kvm.c | 23 +++
target-ppc/machine.c | 44
trace-events | 3 +++
5 files changed, 123 insertions(+)
diff --git a/hw/ppc/ppc.c b/hw/ppc/ppc.c
index 1e3cab3..7d08c9a 100644
--- a/hw/ppc/ppc.c
+++ b/hw/ppc/ppc.c
@@ -31,6 +31,7 @@
#include "hw/loader.h"
#include "sysemu/kvm.h"
#include "kvm_ppc.h"
+#include "trace.h"
//#define PPC_DEBUG_IRQ
#define PPC_DEBUG_TB
@@ -796,6 +797,54 @@ static void cpu_ppc_set_tb_clk (void *opaque,
uint32_t freq)
cpu_ppc_store_purr(cpu, 0xULL);
}
+/*
+ * Calculate timebase on the destination side of migration
>>>
+ * We calculate new timebase offset as shown below:
+ * 1) Gtb2 = Gtb1 + max(tod2 - tod1, 0)
+ *Gtb2 = tb2 + off2
+ *Gtb1 = tb1 + off1
+ * 2) tb2 + off2 = tb1 + off1 + max(tod2 - tod1, 0)
+ * 3) off2 = tb1 - tb2 + off1 + max(tod2 - tod1, 0)
+ *
+ * where:
+ * Gtb2 - destination guest timebase
+ * tb2 - destination host timebase
+ * off2 - destination timebase offset
+ * tod2 - destination time of the day
+ * Gtb1 - source guest timebase
+ * tb1 - source host timebase
+ * off1 - source timebase offset
+ * tod1 - source time of the day
+ *
+ * The result we want is in @off2
+ *
+ * Two conditions must be met for @off2:
+ * 1) off2 must be multiple of 2^24 ticks as it will be set via TBU40 SPR
+ * 2) Gtb2 >= Gtb1
>>>
>>> What about the TCG case, where there is not host timebase, only a a
>>> host system clock?
>>
>>
>> cpu_get_real_ticks() returns ticks, this is what the patch cares about.
>> What is the difference between KVM and TCG here?
>>
>>
+ */
+void cpu_ppc_adjust_tb_offset(ppc_tb_t *tb_env)
+{
+uint64_t tb2, tod2, off2;
+int ratio = tb_env->tb_freq / 100;
+struct timeval tv;
+
+tb2 = cpu_get_real_ticks();
+gettimeofday(&tv, NULL);
+tod2 = tv.tv_sec * 100 + tv.tv_usec;
+
+off2 = tb_env->timebase - tb2 + tb_env->tb_offset;
+if (tod2 > tb_env->time_of_the_day) {
+off2 += (tod2 - tb_env->time_of_the_day) * ratio;
+}
+off2 = ROUND_UP(off2, 1 << 24);
+
+trace_ppc_tb_adjust(tb_env->tb_offset, off2,
+(int64_t)off2 - tb_env->tb_offset);
+
+tb_env->tb_offset = off2;
+}
+
/* Set up (once) timebase frequency (in Hz) */
clk_setup_cb cpu_ppc_tb_init (CPUPPCState *env, uint32_t freq)
{
diff --git a/include/hw/ppc/ppc.h b/include/hw/ppc/ppc.h
index 132ab97..235871c 100644
--- a/include/hw/ppc/ppc.h
+++ b/include/hw/ppc/ppc.h
@@ -32,6 +32,9 @@ struct ppc_tb_t {
uint64_t purr_start;
void *opaque;
uint32_t flags;
+/* Cached values for live migration purposes */
+uint64_t timebase;
+uint64_t time_of_the_day;
>>>
>>> How is the time of day encoded here?
>>
>>
>> Microseconds. I'll put a comment here, I just thought it is quite obvious
>> as gettimeofday() returns microseconds.
>>
>>
};
/* PPC Timers flags */
@@ -46,6 +49,7 @@ struct ppc_tb_t {
*/
uint64_t cpu_ppc_get_tb(ppc_tb_t *tb_env, uint64_t vmclk, int64_t
tb_offset);
+void cpu_ppc_adjust_tb_offset(ppc_tb_t *tb_env);
clk_setup_cb cpu_ppc_tb_init (CPUPPCState *env, uint32_t freq);
/* Embedded PowerPC DCR management */
typedef uint32_t (*dcr_read_cb)(void *opaque, int dcrn);
>>>
Re: [Qemu-devel] [RFC PATCH] spapr: support time base offset migration
On 05.09.2013, at 06:54, Alexey Kardashevskiy wrote:
> On 09/05/2013 02:30 PM, David Gibson wrote:
>> On Tue, Sep 03, 2013 at 05:31:42PM +1000, Alexey Kardashevskiy wrote:
>>> This allows guests to have a different timebase origin from the host.
>>>
>>> This is needed for migration, where a guest can migrate from one host
>>> to another and the two hosts might have a different timebase origin.
>>> However, the timebase seen by the guest must not go backwards, and
>>> should go forwards only by a small amount corresponding to the time
>>> taken for the migration.
>>>
>>> This is only supported for recent POWER hardware which has the TBU40
>>> (timebase upper 40 bits) register. That includes POWER6, 7, 8 but not
>>> 970.
>>>
>>> This adds kvm_access_one_reg() to access a special register which is not
>>> in env->spr.
>>>
>>> The feature must be present in the host kernel.
>>>
>>> Signed-off-by: Alexey Kardashevskiy
>>> ---
>>>
>>> This is an RFC but not a final patch. Can break something but I just do not
>>> see what.
>>>
>>> ---
>>> hw/ppc/ppc.c | 49 +
>>> include/hw/ppc/ppc.h | 4
>>> target-ppc/kvm.c | 23 +++
>>> target-ppc/machine.c | 44
>>> trace-events | 3 +++
>>> 5 files changed, 123 insertions(+)
>>>
>>> diff --git a/hw/ppc/ppc.c b/hw/ppc/ppc.c
>>> index 1e3cab3..7d08c9a 100644
>>> --- a/hw/ppc/ppc.c
>>> +++ b/hw/ppc/ppc.c
>>> @@ -31,6 +31,7 @@
>>> #include "hw/loader.h"
>>> #include "sysemu/kvm.h"
>>> #include "kvm_ppc.h"
>>> +#include "trace.h"
>>>
>>> //#define PPC_DEBUG_IRQ
>>> #define PPC_DEBUG_TB
>>> @@ -796,6 +797,54 @@ static void cpu_ppc_set_tb_clk (void *opaque, uint32_t
>>> freq)
>>> cpu_ppc_store_purr(cpu, 0xULL);
>>> }
>>>
>>> +/*
>>> + * Calculate timebase on the destination side of migration
>>
>>> + * We calculate new timebase offset as shown below:
>>> + * 1) Gtb2 = Gtb1 + max(tod2 - tod1, 0)
>>> + *Gtb2 = tb2 + off2
>>> + *Gtb1 = tb1 + off1
>>> + * 2) tb2 + off2 = tb1 + off1 + max(tod2 - tod1, 0)
>>> + * 3) off2 = tb1 - tb2 + off1 + max(tod2 - tod1, 0)
>>> + *
>>> + * where:
>>> + * Gtb2 - destination guest timebase
>>> + * tb2 - destination host timebase
>>> + * off2 - destination timebase offset
>>> + * tod2 - destination time of the day
>>> + * Gtb1 - source guest timebase
>>> + * tb1 - source host timebase
>>> + * off1 - source timebase offset
>>> + * tod1 - source time of the day
>>> + *
>>> + * The result we want is in @off2
>>> + *
>>> + * Two conditions must be met for @off2:
>>> + * 1) off2 must be multiple of 2^24 ticks as it will be set via TBU40 SPR
>>> + * 2) Gtb2 >= Gtb1
>>
>> What about the TCG case, where there is not host timebase, only a a
>> host system clock?
>
>
> cpu_get_real_ticks() returns ticks, this is what the patch cares about.
> What is the difference between KVM and TCG here?
>
>
>>> + */
>>> +void cpu_ppc_adjust_tb_offset(ppc_tb_t *tb_env)
>>> +{
>>> +uint64_t tb2, tod2, off2;
>>> +int ratio = tb_env->tb_freq / 100;
>>> +struct timeval tv;
>>> +
>>> +tb2 = cpu_get_real_ticks();
>>> +gettimeofday(&tv, NULL);
>>> +tod2 = tv.tv_sec * 100 + tv.tv_usec;
>>> +
>>> +off2 = tb_env->timebase - tb2 + tb_env->tb_offset;
>>> +if (tod2 > tb_env->time_of_the_day) {
>>> +off2 += (tod2 - tb_env->time_of_the_day) * ratio;
>>> +}
>>> +off2 = ROUND_UP(off2, 1 << 24);
>>> +
>>> +trace_ppc_tb_adjust(tb_env->tb_offset, off2,
>>> +(int64_t)off2 - tb_env->tb_offset);
>>> +
>>> +tb_env->tb_offset = off2;
>>> +}
>>> +
>>> /* Set up (once) timebase frequency (in Hz) */
>>> clk_setup_cb cpu_ppc_tb_init (CPUPPCState *env, uint32_t freq)
>>> {
>>> diff --git a/include/hw/ppc/ppc.h b/include/hw/ppc/ppc.h
>>> index 132ab97..235871c 100644
>>> --- a/include/hw/ppc/ppc.h
>>> +++ b/include/hw/ppc/ppc.h
>>> @@ -32,6 +32,9 @@ struct ppc_tb_t {
>>> uint64_t purr_start;
>>> void *opaque;
>>> uint32_t flags;
>>> +/* Cached values for live migration purposes */
>>> +uint64_t timebase;
>>> +uint64_t time_of_the_day;
>>
>> How is the time of day encoded here?
>
>
> Microseconds. I'll put a comment here, I just thought it is quite obvious
> as gettimeofday() returns microseconds.
>
>
>>> };
>>>
>>> /* PPC Timers flags */
>>> @@ -46,6 +49,7 @@ struct ppc_tb_t {
>>>*/
>>>
>>> uint64_t cpu_ppc_get_tb(ppc_tb_t *tb_env, uint64_t vmclk, int64_t
>>> tb_offset);
>>> +void cpu_ppc_adjust_tb_offset(ppc_tb_t *tb_env);
>>> clk_setup_cb cpu_ppc_tb_init (CPUPPCState *env, uint32_t freq);
>>> /* Embedded PowerPC DCR management */
>>> typedef uint32_t (*dcr_read_cb)(void *opaque, int dcrn);
>>> diff --git a/target-ppc/kvm.c b/target-ppc/kvm.c
>>> index 7af9e3d..93df955 100644
>>> --- a/target-ppc/kvm.c
>>> +++ b/target-ppc/kvm.c
>>> @@ -35,6 +35,7 @@
>>> #inclu
Re: [Qemu-devel] [RFC PATCH] spapr: support time base offset migration
On 09/05/2013 02:30 PM, David Gibson wrote:
> On Tue, Sep 03, 2013 at 05:31:42PM +1000, Alexey Kardashevskiy wrote:
>> This allows guests to have a different timebase origin from the host.
>>
>> This is needed for migration, where a guest can migrate from one host
>> to another and the two hosts might have a different timebase origin.
>> However, the timebase seen by the guest must not go backwards, and
>> should go forwards only by a small amount corresponding to the time
>> taken for the migration.
>>
>> This is only supported for recent POWER hardware which has the TBU40
>> (timebase upper 40 bits) register. That includes POWER6, 7, 8 but not
>> 970.
>>
>> This adds kvm_access_one_reg() to access a special register which is not
>> in env->spr.
>>
>> The feature must be present in the host kernel.
>>
>> Signed-off-by: Alexey Kardashevskiy
>> ---
>>
>> This is an RFC but not a final patch. Can break something but I just do not
>> see what.
>>
>> ---
>> hw/ppc/ppc.c | 49 +
>> include/hw/ppc/ppc.h | 4
>> target-ppc/kvm.c | 23 +++
>> target-ppc/machine.c | 44
>> trace-events | 3 +++
>> 5 files changed, 123 insertions(+)
>>
>> diff --git a/hw/ppc/ppc.c b/hw/ppc/ppc.c
>> index 1e3cab3..7d08c9a 100644
>> --- a/hw/ppc/ppc.c
>> +++ b/hw/ppc/ppc.c
>> @@ -31,6 +31,7 @@
>> #include "hw/loader.h"
>> #include "sysemu/kvm.h"
>> #include "kvm_ppc.h"
>> +#include "trace.h"
>>
>> //#define PPC_DEBUG_IRQ
>> #define PPC_DEBUG_TB
>> @@ -796,6 +797,54 @@ static void cpu_ppc_set_tb_clk (void *opaque, uint32_t
>> freq)
>> cpu_ppc_store_purr(cpu, 0xULL);
>> }
>>
>> +/*
>> + * Calculate timebase on the destination side of migration
>
>> + * We calculate new timebase offset as shown below:
>> + * 1) Gtb2 = Gtb1 + max(tod2 - tod1, 0)
>> + *Gtb2 = tb2 + off2
>> + *Gtb1 = tb1 + off1
>> + * 2) tb2 + off2 = tb1 + off1 + max(tod2 - tod1, 0)
>> + * 3) off2 = tb1 - tb2 + off1 + max(tod2 - tod1, 0)
>> + *
>> + * where:
>> + * Gtb2 - destination guest timebase
>> + * tb2 - destination host timebase
>> + * off2 - destination timebase offset
>> + * tod2 - destination time of the day
>> + * Gtb1 - source guest timebase
>> + * tb1 - source host timebase
>> + * off1 - source timebase offset
>> + * tod1 - source time of the day
>> + *
>> + * The result we want is in @off2
>> + *
>> + * Two conditions must be met for @off2:
>> + * 1) off2 must be multiple of 2^24 ticks as it will be set via TBU40 SPR
>> + * 2) Gtb2 >= Gtb1
>
> What about the TCG case, where there is not host timebase, only a a
> host system clock?
cpu_get_real_ticks() returns ticks, this is what the patch cares about.
What is the difference between KVM and TCG here?
>> + */
>> +void cpu_ppc_adjust_tb_offset(ppc_tb_t *tb_env)
>> +{
>> +uint64_t tb2, tod2, off2;
>> +int ratio = tb_env->tb_freq / 100;
>> +struct timeval tv;
>> +
>> +tb2 = cpu_get_real_ticks();
>> +gettimeofday(&tv, NULL);
>> +tod2 = tv.tv_sec * 100 + tv.tv_usec;
>> +
>> +off2 = tb_env->timebase - tb2 + tb_env->tb_offset;
>> +if (tod2 > tb_env->time_of_the_day) {
>> +off2 += (tod2 - tb_env->time_of_the_day) * ratio;
>> +}
>> +off2 = ROUND_UP(off2, 1 << 24);
>> +
>> +trace_ppc_tb_adjust(tb_env->tb_offset, off2,
>> +(int64_t)off2 - tb_env->tb_offset);
>> +
>> +tb_env->tb_offset = off2;
>> +}
>> +
>> /* Set up (once) timebase frequency (in Hz) */
>> clk_setup_cb cpu_ppc_tb_init (CPUPPCState *env, uint32_t freq)
>> {
>> diff --git a/include/hw/ppc/ppc.h b/include/hw/ppc/ppc.h
>> index 132ab97..235871c 100644
>> --- a/include/hw/ppc/ppc.h
>> +++ b/include/hw/ppc/ppc.h
>> @@ -32,6 +32,9 @@ struct ppc_tb_t {
>> uint64_t purr_start;
>> void *opaque;
>> uint32_t flags;
>> +/* Cached values for live migration purposes */
>> +uint64_t timebase;
>> +uint64_t time_of_the_day;
>
> How is the time of day encoded here?
Microseconds. I'll put a comment here, I just thought it is quite obvious
as gettimeofday() returns microseconds.
>> };
>>
>> /* PPC Timers flags */
>> @@ -46,6 +49,7 @@ struct ppc_tb_t {
>> */
>>
>> uint64_t cpu_ppc_get_tb(ppc_tb_t *tb_env, uint64_t vmclk, int64_t
>> tb_offset);
>> +void cpu_ppc_adjust_tb_offset(ppc_tb_t *tb_env);
>> clk_setup_cb cpu_ppc_tb_init (CPUPPCState *env, uint32_t freq);
>> /* Embedded PowerPC DCR management */
>> typedef uint32_t (*dcr_read_cb)(void *opaque, int dcrn);
>> diff --git a/target-ppc/kvm.c b/target-ppc/kvm.c
>> index 7af9e3d..93df955 100644
>> --- a/target-ppc/kvm.c
>> +++ b/target-ppc/kvm.c
>> @@ -35,6 +35,7 @@
>> #include "hw/sysbus.h"
>> #include "hw/ppc/spapr.h"
>> #include "hw/ppc/spapr_vio.h"
>> +#include "hw/ppc/ppc.h"
>> #include "sysemu/watchdog.h"
>>
>> //#define DEBUG_KVM
>> @@ -761,6 +762,22
Re: [Qemu-devel] [RFC PATCH] spapr: support time base offset migration
On Tue, Sep 03, 2013 at 05:31:42PM +1000, Alexey Kardashevskiy wrote:
> This allows guests to have a different timebase origin from the host.
>
> This is needed for migration, where a guest can migrate from one host
> to another and the two hosts might have a different timebase origin.
> However, the timebase seen by the guest must not go backwards, and
> should go forwards only by a small amount corresponding to the time
> taken for the migration.
>
> This is only supported for recent POWER hardware which has the TBU40
> (timebase upper 40 bits) register. That includes POWER6, 7, 8 but not
> 970.
>
> This adds kvm_access_one_reg() to access a special register which is not
> in env->spr.
>
> The feature must be present in the host kernel.
>
> Signed-off-by: Alexey Kardashevskiy
> ---
>
> This is an RFC but not a final patch. Can break something but I just do not
> see what.
>
> ---
> hw/ppc/ppc.c | 49 +
> include/hw/ppc/ppc.h | 4
> target-ppc/kvm.c | 23 +++
> target-ppc/machine.c | 44
> trace-events | 3 +++
> 5 files changed, 123 insertions(+)
>
> diff --git a/hw/ppc/ppc.c b/hw/ppc/ppc.c
> index 1e3cab3..7d08c9a 100644
> --- a/hw/ppc/ppc.c
> +++ b/hw/ppc/ppc.c
> @@ -31,6 +31,7 @@
> #include "hw/loader.h"
> #include "sysemu/kvm.h"
> #include "kvm_ppc.h"
> +#include "trace.h"
>
> //#define PPC_DEBUG_IRQ
> #define PPC_DEBUG_TB
> @@ -796,6 +797,54 @@ static void cpu_ppc_set_tb_clk (void *opaque, uint32_t
> freq)
> cpu_ppc_store_purr(cpu, 0xULL);
> }
>
> +/*
> + * Calculate timebase on the destination side of migration
> + * We calculate new timebase offset as shown below:
> + * 1) Gtb2 = Gtb1 + max(tod2 - tod1, 0)
> + *Gtb2 = tb2 + off2
> + *Gtb1 = tb1 + off1
> + * 2) tb2 + off2 = tb1 + off1 + max(tod2 - tod1, 0)
> + * 3) off2 = tb1 - tb2 + off1 + max(tod2 - tod1, 0)
> + *
> + * where:
> + * Gtb2 - destination guest timebase
> + * tb2 - destination host timebase
> + * off2 - destination timebase offset
> + * tod2 - destination time of the day
> + * Gtb1 - source guest timebase
> + * tb1 - source host timebase
> + * off1 - source timebase offset
> + * tod1 - source time of the day
> + *
> + * The result we want is in @off2
> + *
> + * Two conditions must be met for @off2:
> + * 1) off2 must be multiple of 2^24 ticks as it will be set via TBU40 SPR
> + * 2) Gtb2 >= Gtb1
What about the TCG case, where there is not host timebase, only a a
host system clock?
> + */
> +void cpu_ppc_adjust_tb_offset(ppc_tb_t *tb_env)
> +{
> +uint64_t tb2, tod2, off2;
> +int ratio = tb_env->tb_freq / 100;
> +struct timeval tv;
> +
> +tb2 = cpu_get_real_ticks();
> +gettimeofday(&tv, NULL);
> +tod2 = tv.tv_sec * 100 + tv.tv_usec;
> +
> +off2 = tb_env->timebase - tb2 + tb_env->tb_offset;
> +if (tod2 > tb_env->time_of_the_day) {
> +off2 += (tod2 - tb_env->time_of_the_day) * ratio;
> +}
> +off2 = ROUND_UP(off2, 1 << 24);
> +
> +trace_ppc_tb_adjust(tb_env->tb_offset, off2,
> +(int64_t)off2 - tb_env->tb_offset);
> +
> +tb_env->tb_offset = off2;
> +}
> +
> /* Set up (once) timebase frequency (in Hz) */
> clk_setup_cb cpu_ppc_tb_init (CPUPPCState *env, uint32_t freq)
> {
> diff --git a/include/hw/ppc/ppc.h b/include/hw/ppc/ppc.h
> index 132ab97..235871c 100644
> --- a/include/hw/ppc/ppc.h
> +++ b/include/hw/ppc/ppc.h
> @@ -32,6 +32,9 @@ struct ppc_tb_t {
> uint64_t purr_start;
> void *opaque;
> uint32_t flags;
> +/* Cached values for live migration purposes */
> +uint64_t timebase;
> +uint64_t time_of_the_day;
How is the time of day encoded here?
> };
>
> /* PPC Timers flags */
> @@ -46,6 +49,7 @@ struct ppc_tb_t {
> */
>
> uint64_t cpu_ppc_get_tb(ppc_tb_t *tb_env, uint64_t vmclk, int64_t tb_offset);
> +void cpu_ppc_adjust_tb_offset(ppc_tb_t *tb_env);
> clk_setup_cb cpu_ppc_tb_init (CPUPPCState *env, uint32_t freq);
> /* Embedded PowerPC DCR management */
> typedef uint32_t (*dcr_read_cb)(void *opaque, int dcrn);
> diff --git a/target-ppc/kvm.c b/target-ppc/kvm.c
> index 7af9e3d..93df955 100644
> --- a/target-ppc/kvm.c
> +++ b/target-ppc/kvm.c
> @@ -35,6 +35,7 @@
> #include "hw/sysbus.h"
> #include "hw/ppc/spapr.h"
> #include "hw/ppc/spapr_vio.h"
> +#include "hw/ppc/ppc.h"
> #include "sysemu/watchdog.h"
>
> //#define DEBUG_KVM
> @@ -761,6 +762,22 @@ static int kvm_put_vpa(CPUState *cs)
> }
> #endif /* TARGET_PPC64 */
>
> +static int kvm_access_one_reg(CPUState *cs, bool set, __u64 id,
> void *addr)
I think it would be nicer to have seperate set_one_reg and get_one_reg
functions, rather than using a direction parameter.
> +{
> +struct kvm_one_reg reg = {
> +.id = id,
> +.addr = (uintptr_t)addr,
> +};
> +int ret = kvm_vcpu_ioctl(cs, se
Re: [Qemu-devel] [RFC PATCH] spapr: support time base offset migration
On 04.09.2013, at 03:13, Alexey Kardashevskiy wrote:
> On 09/03/2013 07:22 PM, Andreas Färber wrote:
>> Am 03.09.2013 11:07, schrieb Alexey Kardashevskiy:
>>> On 09/03/2013 06:42 PM, Andreas Färber wrote:
Am 03.09.2013 09:31, schrieb Alexey Kardashevskiy:
> diff --git a/target-ppc/machine.c b/target-ppc/machine.c
> index 12e1512..d1ffc7f 100644
> --- a/target-ppc/machine.c
> +++ b/target-ppc/machine.c
>> [...]
> +static const VMStateDescription vmstate_timebase = {
> +.name = "cpu/timebase",
> +.version_id = 1,
> +.minimum_version_id = 1,
> +.minimum_version_id_old = 1,
> +.pre_save = timebase_pre_save,
> +.post_load = timebase_post_load,
> +.fields = (VMStateField []) {
> +VMSTATE_UINT64(timebase, ppc_tb_t),
> +VMSTATE_INT64(tb_offset, ppc_tb_t),
> +VMSTATE_UINT64(time_of_the_day, ppc_tb_t),
> +VMSTATE_UINT32_EQUAL(tb_freq, ppc_tb_t),
> +VMSTATE_END_OF_LIST()
> +},
> +};
> +
> const VMStateDescription vmstate_ppc_cpu = {
> .name = "cpu",
> .version_id = 5,
> @@ -498,6 +538,10 @@ const VMStateDescription vmstate_ppc_cpu = {
> VMSTATE_UINT64_EQUAL(env.insns_flags, PowerPCCPU),
> VMSTATE_UINT64_EQUAL(env.insns_flags2, PowerPCCPU),
> VMSTATE_UINT32_EQUAL(env.nb_BATs, PowerPCCPU),
> +
> +/* Time offset */
> +VMSTATE_STRUCT_POINTER(env.tb_env, PowerPCCPU,
> + vmstate_timebase, ppc_tb_t *),
> VMSTATE_END_OF_LIST()
> },
> .subsections = (VMStateSubsection []) {
Breaks the migration format. ;) You need to bump version_id and use a
macro that accepts the version the field was added in as argument.
>>>
>>> Risking of being called ignorant, I'll still ask - is the patch below what
>>> you mean? I could not find VMSTATE_STRUCT_POINTER_V and I do not believe it
>>> is not there already.
>>
>> Usually the way we do it is to have VMSTATE_STRUCT_POINTER() call
>> VMSTATE_STRUCT_POINTER_V() and thus VMSTATE_STRUCT_POINTER_TEST() call a
>> new VMSTATE_STRUCT_POINTER_TEST_V(), to avoid code duplication of the
>> core array entry. CC'ing Juan. Please do the VMState preparation in a
>> separate patch.
>>
>> ppc usage looks fine.
>>
>>> btw why would it break? Just asking. Is it because the source may send what
>>> the destination cannot handle? Named fields should stop the migration the
>>> same way as version mismatch would have done.
>>
>> Nope, field names do not get transmitted, only the section names.
>> (Otherwise random code refactorings could break the migration format.)
>>
>>> Or the source won't sent what the destination expects and we do not want
>>> this destination guest to continue?
>>
>> There's an incoming stream of data from either live migration or a file,
>> and QEMU must decide whether it can read and how to interpret the raw
>> bytestream. It shouldn't just read random bytes into a new field when
>> they were written from a different field.
>>
>>> Once I was told that migration between different versions of QEMU is not
>>> supported - so what is the point of .version_id field at all?
>>
>> Not sure who told such a thing and in what context. On x86 we try to
>> avoid version_id bumps by adding subsections to allow migration in both
>> ways (including from newer to older) but for ppc, arm and all others we
>> do require that new fields are marked as such. Whether migration is
>> officially supported is a different matter from the VMState wire format.
>
>
> Why is the approach different for x86 and ppc here? I can convert
> VMSTATE_STRUCT_POINTER to a subsection, why should not I? Or ppc is not
> mature enough and therefore there is no need to keep compatibility? Thanks.
Just bump the version.
Alex
Re: [Qemu-devel] [RFC PATCH] spapr: support time base offset migration
On 09/03/2013 07:22 PM, Andreas Färber wrote:
> Am 03.09.2013 11:07, schrieb Alexey Kardashevskiy:
>> On 09/03/2013 06:42 PM, Andreas Färber wrote:
>>> Am 03.09.2013 09:31, schrieb Alexey Kardashevskiy:
diff --git a/target-ppc/machine.c b/target-ppc/machine.c
index 12e1512..d1ffc7f 100644
--- a/target-ppc/machine.c
+++ b/target-ppc/machine.c
> [...]
+static const VMStateDescription vmstate_timebase = {
+.name = "cpu/timebase",
+.version_id = 1,
+.minimum_version_id = 1,
+.minimum_version_id_old = 1,
+.pre_save = timebase_pre_save,
+.post_load = timebase_post_load,
+.fields = (VMStateField []) {
+VMSTATE_UINT64(timebase, ppc_tb_t),
+VMSTATE_INT64(tb_offset, ppc_tb_t),
+VMSTATE_UINT64(time_of_the_day, ppc_tb_t),
+VMSTATE_UINT32_EQUAL(tb_freq, ppc_tb_t),
+VMSTATE_END_OF_LIST()
+},
+};
+
const VMStateDescription vmstate_ppc_cpu = {
.name = "cpu",
.version_id = 5,
@@ -498,6 +538,10 @@ const VMStateDescription vmstate_ppc_cpu = {
VMSTATE_UINT64_EQUAL(env.insns_flags, PowerPCCPU),
VMSTATE_UINT64_EQUAL(env.insns_flags2, PowerPCCPU),
VMSTATE_UINT32_EQUAL(env.nb_BATs, PowerPCCPU),
+
+/* Time offset */
+VMSTATE_STRUCT_POINTER(env.tb_env, PowerPCCPU,
+ vmstate_timebase, ppc_tb_t *),
VMSTATE_END_OF_LIST()
},
.subsections = (VMStateSubsection []) {
>>>
>>> Breaks the migration format. ;) You need to bump version_id and use a
>>> macro that accepts the version the field was added in as argument.
>>
>> Risking of being called ignorant, I'll still ask - is the patch below what
>> you mean? I could not find VMSTATE_STRUCT_POINTER_V and I do not believe it
>> is not there already.
>
> Usually the way we do it is to have VMSTATE_STRUCT_POINTER() call
> VMSTATE_STRUCT_POINTER_V() and thus VMSTATE_STRUCT_POINTER_TEST() call a
> new VMSTATE_STRUCT_POINTER_TEST_V(), to avoid code duplication of the
> core array entry. CC'ing Juan. Please do the VMState preparation in a
> separate patch.
>
> ppc usage looks fine.
>
>> btw why would it break? Just asking. Is it because the source may send what
>> the destination cannot handle? Named fields should stop the migration the
>> same way as version mismatch would have done.
>
> Nope, field names do not get transmitted, only the section names.
> (Otherwise random code refactorings could break the migration format.)
>
>> Or the source won't sent what the destination expects and we do not want
>> this destination guest to continue?
>
> There's an incoming stream of data from either live migration or a file,
> and QEMU must decide whether it can read and how to interpret the raw
> bytestream. It shouldn't just read random bytes into a new field when
> they were written from a different field.
>
>> Once I was told that migration between different versions of QEMU is not
>> supported - so what is the point of .version_id field at all?
>
> Not sure who told such a thing and in what context. On x86 we try to
> avoid version_id bumps by adding subsections to allow migration in both
> ways (including from newer to older) but for ppc, arm and all others we
> do require that new fields are marked as such. Whether migration is
> officially supported is a different matter from the VMState wire format.
Why is the approach different for x86 and ppc here? I can convert
VMSTATE_STRUCT_POINTER to a subsection, why should not I? Or ppc is not
mature enough and therefore there is no need to keep compatibility? Thanks.
>
> Regards,
> Andreas
>
>
>> alexey@ka1:~/pcipassthru/qemu$ git diff
>> diff --git a/include/migration/vmstate.h b/include/migration/vmstate.h
>> index 1c31b5d..7b624bf 100644
>> --- a/include/migration/vmstate.h
>> +++ b/include/migration/vmstate.h
>> @@ -499,6 +499,15 @@ extern const VMStateInfo vmstate_info_bitmap;
>> #define VMSTATE_STRUCT_POINTER(_field, _state, _vmsd, _type) \
>> VMSTATE_STRUCT_POINTER_TEST(_field, _state, NULL, _vmsd, _type)
>>
>> +#define VMSTATE_STRUCT_POINTER_V(_field, _state, _vmsd, _type, _version) {
>> \
>> +.name = (stringify(_field)), \
>> +.version_id = (_version),\
>> +.vmsd = &(_vmsd),\
>> +.size = sizeof(_type), \
>> +.flags= VMS_STRUCT|VMS_POINTER, \
>> +.offset = vmstate_offset_value(_state, _field, _type), \
>> +}
>> +
>> #define VMSTATE_STRUCT_ARRAY(_field, _state, _num, _version, _vmsd, _type) \
>> VMSTATE_STRUCT_ARRAY_TEST(_field, _state, _num, NULL, _version, \
>> _vmsd, _type)
>> diff --git a/target-ppc/machine.c b/target-ppc/ma
Re: [Qemu-devel] [RFC PATCH] spapr: support time base offset migration
Am 03.09.2013 11:07, schrieb Alexey Kardashevskiy:
> On 09/03/2013 06:42 PM, Andreas Färber wrote:
>> Am 03.09.2013 09:31, schrieb Alexey Kardashevskiy:
>>> diff --git a/target-ppc/machine.c b/target-ppc/machine.c
>>> index 12e1512..d1ffc7f 100644
>>> --- a/target-ppc/machine.c
>>> +++ b/target-ppc/machine.c
[...]
>>> +static const VMStateDescription vmstate_timebase = {
>>> +.name = "cpu/timebase",
>>> +.version_id = 1,
>>> +.minimum_version_id = 1,
>>> +.minimum_version_id_old = 1,
>>> +.pre_save = timebase_pre_save,
>>> +.post_load = timebase_post_load,
>>> +.fields = (VMStateField []) {
>>> +VMSTATE_UINT64(timebase, ppc_tb_t),
>>> +VMSTATE_INT64(tb_offset, ppc_tb_t),
>>> +VMSTATE_UINT64(time_of_the_day, ppc_tb_t),
>>> +VMSTATE_UINT32_EQUAL(tb_freq, ppc_tb_t),
>>> +VMSTATE_END_OF_LIST()
>>> +},
>>> +};
>>> +
>>> const VMStateDescription vmstate_ppc_cpu = {
>>> .name = "cpu",
>>> .version_id = 5,
>>> @@ -498,6 +538,10 @@ const VMStateDescription vmstate_ppc_cpu = {
>>> VMSTATE_UINT64_EQUAL(env.insns_flags, PowerPCCPU),
>>> VMSTATE_UINT64_EQUAL(env.insns_flags2, PowerPCCPU),
>>> VMSTATE_UINT32_EQUAL(env.nb_BATs, PowerPCCPU),
>>> +
>>> +/* Time offset */
>>> +VMSTATE_STRUCT_POINTER(env.tb_env, PowerPCCPU,
>>> + vmstate_timebase, ppc_tb_t *),
>>> VMSTATE_END_OF_LIST()
>>> },
>>> .subsections = (VMStateSubsection []) {
>>
>> Breaks the migration format. ;) You need to bump version_id and use a
>> macro that accepts the version the field was added in as argument.
>
> Risking of being called ignorant, I'll still ask - is the patch below what
> you mean? I could not find VMSTATE_STRUCT_POINTER_V and I do not believe it
> is not there already.
Usually the way we do it is to have VMSTATE_STRUCT_POINTER() call
VMSTATE_STRUCT_POINTER_V() and thus VMSTATE_STRUCT_POINTER_TEST() call a
new VMSTATE_STRUCT_POINTER_TEST_V(), to avoid code duplication of the
core array entry. CC'ing Juan. Please do the VMState preparation in a
separate patch.
ppc usage looks fine.
> btw why would it break? Just asking. Is it because the source may send what
> the destination cannot handle? Named fields should stop the migration the
> same way as version mismatch would have done.
Nope, field names do not get transmitted, only the section names.
(Otherwise random code refactorings could break the migration format.)
> Or the source won't sent what the destination expects and we do not want
> this destination guest to continue?
There's an incoming stream of data from either live migration or a file,
and QEMU must decide whether it can read and how to interpret the raw
bytestream. It shouldn't just read random bytes into a new field when
they were written from a different field.
> Once I was told that migration between different versions of QEMU is not
> supported - so what is the point of .version_id field at all?
Not sure who told such a thing and in what context. On x86 we try to
avoid version_id bumps by adding subsections to allow migration in both
ways (including from newer to older) but for ppc, arm and all others we
do require that new fields are marked as such. Whether migration is
officially supported is a different matter from the VMState wire format.
Regards,
Andreas
> alexey@ka1:~/pcipassthru/qemu$ git diff
> diff --git a/include/migration/vmstate.h b/include/migration/vmstate.h
> index 1c31b5d..7b624bf 100644
> --- a/include/migration/vmstate.h
> +++ b/include/migration/vmstate.h
> @@ -499,6 +499,15 @@ extern const VMStateInfo vmstate_info_bitmap;
> #define VMSTATE_STRUCT_POINTER(_field, _state, _vmsd, _type) \
> VMSTATE_STRUCT_POINTER_TEST(_field, _state, NULL, _vmsd, _type)
>
> +#define VMSTATE_STRUCT_POINTER_V(_field, _state, _vmsd, _type, _version) { \
> +.name = (stringify(_field)), \
> +.version_id = (_version),\
> +.vmsd = &(_vmsd),\
> +.size = sizeof(_type), \
> +.flags= VMS_STRUCT|VMS_POINTER, \
> +.offset = vmstate_offset_value(_state, _field, _type), \
> +}
> +
> #define VMSTATE_STRUCT_ARRAY(_field, _state, _num, _version, _vmsd, _type) \
> VMSTATE_STRUCT_ARRAY_TEST(_field, _state, _num, NULL, _version, \
> _vmsd, _type)
> diff --git a/target-ppc/machine.c b/target-ppc/machine.c
> index b4f447c..f79f38e 100644
> --- a/target-ppc/machine.c
> +++ b/target-ppc/machine.c
> @@ -501,7 +501,7 @@ static const VMStateDescription vmstate_timebase = {
>
> const VMStateDescription vmstate_ppc_cpu = {
> .name = "cpu",
> -.version_id = 5,
> +.version_id = 6,
> .minimum_version_id = 5,
> .minimum_version_id_old = 4,
> .load_state_old = cpu_load_old,
Re: [Qemu-devel] [RFC PATCH] spapr: support time base offset migration
On 09/03/2013 06:42 PM, Andreas Färber wrote:
> Am 03.09.2013 09:31, schrieb Alexey Kardashevskiy:
>> This allows guests to have a different timebase origin from the host.
>>
>> This is needed for migration, where a guest can migrate from one host
>> to another and the two hosts might have a different timebase origin.
>> However, the timebase seen by the guest must not go backwards, and
>> should go forwards only by a small amount corresponding to the time
>> taken for the migration.
>>
>> This is only supported for recent POWER hardware which has the TBU40
>> (timebase upper 40 bits) register. That includes POWER6, 7, 8 but not
>> 970.
>>
>> This adds kvm_access_one_reg() to access a special register which is not
>> in env->spr.
>>
>> The feature must be present in the host kernel.
>>
>> Signed-off-by: Alexey Kardashevskiy
>> ---
>>
>> This is an RFC but not a final patch. Can break something but I just do not
>> see what.
>>
>> ---
>> hw/ppc/ppc.c | 49 +
>> include/hw/ppc/ppc.h | 4
>> target-ppc/kvm.c | 23 +++
>> target-ppc/machine.c | 44
>> trace-events | 3 +++
>> 5 files changed, 123 insertions(+)
>>
>> diff --git a/hw/ppc/ppc.c b/hw/ppc/ppc.c
>> index 1e3cab3..7d08c9a 100644
>> --- a/hw/ppc/ppc.c
>> +++ b/hw/ppc/ppc.c
>> @@ -31,6 +31,7 @@
>> #include "hw/loader.h"
>> #include "sysemu/kvm.h"
>> #include "kvm_ppc.h"
>> +#include "trace.h"
>>
>> //#define PPC_DEBUG_IRQ
>> #define PPC_DEBUG_TB
>> @@ -796,6 +797,54 @@ static void cpu_ppc_set_tb_clk (void *opaque, uint32_t
>> freq)
>> cpu_ppc_store_purr(cpu, 0xULL);
>> }
>>
>> +/*
>> + * Calculate timebase on the destination side of migration
>> + *
>> + * We calculate new timebase offset as shown below:
>> + * 1) Gtb2 = Gtb1 + max(tod2 - tod1, 0)
>> + *Gtb2 = tb2 + off2
>> + *Gtb1 = tb1 + off1
>> + * 2) tb2 + off2 = tb1 + off1 + max(tod2 - tod1, 0)
>> + * 3) off2 = tb1 - tb2 + off1 + max(tod2 - tod1, 0)
>> + *
>> + * where:
>> + * Gtb2 - destination guest timebase
>> + * tb2 - destination host timebase
>> + * off2 - destination timebase offset
>> + * tod2 - destination time of the day
>> + * Gtb1 - source guest timebase
>> + * tb1 - source host timebase
>> + * off1 - source timebase offset
>> + * tod1 - source time of the day
>> + *
>> + * The result we want is in @off2
>> + *
>> + * Two conditions must be met for @off2:
>> + * 1) off2 must be multiple of 2^24 ticks as it will be set via TBU40 SPR
>> + * 2) Gtb2 >= Gtb1
>> + */
>> +void cpu_ppc_adjust_tb_offset(ppc_tb_t *tb_env)
>> +{
>> +uint64_t tb2, tod2, off2;
>> +int ratio = tb_env->tb_freq / 100;
>> +struct timeval tv;
>> +
>> +tb2 = cpu_get_real_ticks();
>> +gettimeofday(&tv, NULL);
>> +tod2 = tv.tv_sec * 100 + tv.tv_usec;
>> +
>> +off2 = tb_env->timebase - tb2 + tb_env->tb_offset;
>> +if (tod2 > tb_env->time_of_the_day) {
>> +off2 += (tod2 - tb_env->time_of_the_day) * ratio;
>> +}
>> +off2 = ROUND_UP(off2, 1 << 24);
>> +
>> +trace_ppc_tb_adjust(tb_env->tb_offset, off2,
>> +(int64_t)off2 - tb_env->tb_offset);
>> +
>> +tb_env->tb_offset = off2;
>> +}
>> +
>> /* Set up (once) timebase frequency (in Hz) */
>> clk_setup_cb cpu_ppc_tb_init (CPUPPCState *env, uint32_t freq)
>> {
>> diff --git a/include/hw/ppc/ppc.h b/include/hw/ppc/ppc.h
>> index 132ab97..235871c 100644
>> --- a/include/hw/ppc/ppc.h
>> +++ b/include/hw/ppc/ppc.h
>> @@ -32,6 +32,9 @@ struct ppc_tb_t {
>> uint64_t purr_start;
>> void *opaque;
>> uint32_t flags;
>> +/* Cached values for live migration purposes */
>> +uint64_t timebase;
>> +uint64_t time_of_the_day;
>> };
>>
>> /* PPC Timers flags */
>> @@ -46,6 +49,7 @@ struct ppc_tb_t {
>> */
>>
>> uint64_t cpu_ppc_get_tb(ppc_tb_t *tb_env, uint64_t vmclk, int64_t
>> tb_offset);
>> +void cpu_ppc_adjust_tb_offset(ppc_tb_t *tb_env);
>> clk_setup_cb cpu_ppc_tb_init (CPUPPCState *env, uint32_t freq);
>> /* Embedded PowerPC DCR management */
>> typedef uint32_t (*dcr_read_cb)(void *opaque, int dcrn);
>> diff --git a/target-ppc/kvm.c b/target-ppc/kvm.c
>> index 7af9e3d..93df955 100644
>> --- a/target-ppc/kvm.c
>> +++ b/target-ppc/kvm.c
>> @@ -35,6 +35,7 @@
>> #include "hw/sysbus.h"
>> #include "hw/ppc/spapr.h"
>> #include "hw/ppc/spapr_vio.h"
>> +#include "hw/ppc/ppc.h"
>> #include "sysemu/watchdog.h"
>>
>> //#define DEBUG_KVM
>> @@ -761,6 +762,22 @@ static int kvm_put_vpa(CPUState *cs)
>> }
>> #endif /* TARGET_PPC64 */
>>
>> +static int kvm_access_one_reg(CPUState *cs, bool set, __u64 id, void *addr)
>> +{
>> +struct kvm_one_reg reg = {
>> +.id = id,
>> +.addr = (uintptr_t)addr,
>> +};
>> +int ret = kvm_vcpu_ioctl(cs, set ? KVM_SET_ONE_REG : KVM_GET_ONE_REG,
>> ®);
>> +
>> +if (ret) {
>> +
Re: [Qemu-devel] [RFC PATCH] spapr: support time base offset migration
Am 03.09.2013 09:31, schrieb Alexey Kardashevskiy:
> This allows guests to have a different timebase origin from the host.
>
> This is needed for migration, where a guest can migrate from one host
> to another and the two hosts might have a different timebase origin.
> However, the timebase seen by the guest must not go backwards, and
> should go forwards only by a small amount corresponding to the time
> taken for the migration.
>
> This is only supported for recent POWER hardware which has the TBU40
> (timebase upper 40 bits) register. That includes POWER6, 7, 8 but not
> 970.
>
> This adds kvm_access_one_reg() to access a special register which is not
> in env->spr.
>
> The feature must be present in the host kernel.
>
> Signed-off-by: Alexey Kardashevskiy
> ---
>
> This is an RFC but not a final patch. Can break something but I just do not
> see what.
>
> ---
> hw/ppc/ppc.c | 49 +
> include/hw/ppc/ppc.h | 4
> target-ppc/kvm.c | 23 +++
> target-ppc/machine.c | 44
> trace-events | 3 +++
> 5 files changed, 123 insertions(+)
>
> diff --git a/hw/ppc/ppc.c b/hw/ppc/ppc.c
> index 1e3cab3..7d08c9a 100644
> --- a/hw/ppc/ppc.c
> +++ b/hw/ppc/ppc.c
> @@ -31,6 +31,7 @@
> #include "hw/loader.h"
> #include "sysemu/kvm.h"
> #include "kvm_ppc.h"
> +#include "trace.h"
>
> //#define PPC_DEBUG_IRQ
> #define PPC_DEBUG_TB
> @@ -796,6 +797,54 @@ static void cpu_ppc_set_tb_clk (void *opaque, uint32_t
> freq)
> cpu_ppc_store_purr(cpu, 0xULL);
> }
>
> +/*
> + * Calculate timebase on the destination side of migration
> + *
> + * We calculate new timebase offset as shown below:
> + * 1) Gtb2 = Gtb1 + max(tod2 - tod1, 0)
> + *Gtb2 = tb2 + off2
> + *Gtb1 = tb1 + off1
> + * 2) tb2 + off2 = tb1 + off1 + max(tod2 - tod1, 0)
> + * 3) off2 = tb1 - tb2 + off1 + max(tod2 - tod1, 0)
> + *
> + * where:
> + * Gtb2 - destination guest timebase
> + * tb2 - destination host timebase
> + * off2 - destination timebase offset
> + * tod2 - destination time of the day
> + * Gtb1 - source guest timebase
> + * tb1 - source host timebase
> + * off1 - source timebase offset
> + * tod1 - source time of the day
> + *
> + * The result we want is in @off2
> + *
> + * Two conditions must be met for @off2:
> + * 1) off2 must be multiple of 2^24 ticks as it will be set via TBU40 SPR
> + * 2) Gtb2 >= Gtb1
> + */
> +void cpu_ppc_adjust_tb_offset(ppc_tb_t *tb_env)
> +{
> +uint64_t tb2, tod2, off2;
> +int ratio = tb_env->tb_freq / 100;
> +struct timeval tv;
> +
> +tb2 = cpu_get_real_ticks();
> +gettimeofday(&tv, NULL);
> +tod2 = tv.tv_sec * 100 + tv.tv_usec;
> +
> +off2 = tb_env->timebase - tb2 + tb_env->tb_offset;
> +if (tod2 > tb_env->time_of_the_day) {
> +off2 += (tod2 - tb_env->time_of_the_day) * ratio;
> +}
> +off2 = ROUND_UP(off2, 1 << 24);
> +
> +trace_ppc_tb_adjust(tb_env->tb_offset, off2,
> +(int64_t)off2 - tb_env->tb_offset);
> +
> +tb_env->tb_offset = off2;
> +}
> +
> /* Set up (once) timebase frequency (in Hz) */
> clk_setup_cb cpu_ppc_tb_init (CPUPPCState *env, uint32_t freq)
> {
> diff --git a/include/hw/ppc/ppc.h b/include/hw/ppc/ppc.h
> index 132ab97..235871c 100644
> --- a/include/hw/ppc/ppc.h
> +++ b/include/hw/ppc/ppc.h
> @@ -32,6 +32,9 @@ struct ppc_tb_t {
> uint64_t purr_start;
> void *opaque;
> uint32_t flags;
> +/* Cached values for live migration purposes */
> +uint64_t timebase;
> +uint64_t time_of_the_day;
> };
>
> /* PPC Timers flags */
> @@ -46,6 +49,7 @@ struct ppc_tb_t {
> */
>
> uint64_t cpu_ppc_get_tb(ppc_tb_t *tb_env, uint64_t vmclk, int64_t tb_offset);
> +void cpu_ppc_adjust_tb_offset(ppc_tb_t *tb_env);
> clk_setup_cb cpu_ppc_tb_init (CPUPPCState *env, uint32_t freq);
> /* Embedded PowerPC DCR management */
> typedef uint32_t (*dcr_read_cb)(void *opaque, int dcrn);
> diff --git a/target-ppc/kvm.c b/target-ppc/kvm.c
> index 7af9e3d..93df955 100644
> --- a/target-ppc/kvm.c
> +++ b/target-ppc/kvm.c
> @@ -35,6 +35,7 @@
> #include "hw/sysbus.h"
> #include "hw/ppc/spapr.h"
> #include "hw/ppc/spapr_vio.h"
> +#include "hw/ppc/ppc.h"
> #include "sysemu/watchdog.h"
>
> //#define DEBUG_KVM
> @@ -761,6 +762,22 @@ static int kvm_put_vpa(CPUState *cs)
> }
> #endif /* TARGET_PPC64 */
>
> +static int kvm_access_one_reg(CPUState *cs, bool set, __u64 id, void *addr)
> +{
> +struct kvm_one_reg reg = {
> +.id = id,
> +.addr = (uintptr_t)addr,
> +};
> +int ret = kvm_vcpu_ioctl(cs, set ? KVM_SET_ONE_REG : KVM_GET_ONE_REG,
> ®);
> +
> +if (ret) {
> +DPRINTF("Unable to %s time base offset to KVM: %s\n",
> +set ? "set" : "get", strerror(errno));
> +}
> +
> +return ret;
> +}
> +
> int kvm_arch_put_registers(CPUStat
