[PATCH v2] add acpi pmtimer support
This patch makes seabios use the acpi pmtimer instead of tsc for timekeeping. The pmtimer has a fixed frequency and doesn't need calibration, thus it doesn't suffer from calibration errors due to a loaded host machine. [ v2: add CONFIG_PMTIMER ] Signed-off-by: Gerd Hoffmann <kraxel@redhat.com> --- src/Kconfig | 6 ++++++ src/clock.c | 31 +++++++++++++++++++++++++++++++ src/pciinit.c | 5 +++++ src/util.h | 1 + 4 files changed, 43 insertions(+), 0 deletions(-) diff --git a/src/Kconfig b/src/Kconfig index 6de3e71..b5dd63b 100644 --- a/src/Kconfig +++ b/src/Kconfig @@ -222,6 +222,12 @@ menu "Hardware support" default y help Initialize the Memory Type Range Registers (on emulators). + config PMTIMER + depends on !COREBOOT + bool "Use ACPI timer" + default y + help + Use the ACPI timer instead of the TSC for timekeeping (on qemu). endmenu menu "BIOS interfaces" diff --git a/src/clock.c b/src/clock.c index 69e9f17..b4abf37 100644 --- a/src/clock.c +++ b/src/clock.c @@ -129,11 +129,42 @@ emulate_tsc(void) return ret; } +u16 pmtimer_ioport VAR16VISIBLE; +u32 pmtimer_wraps VARLOW; +u32 pmtimer_last VARLOW; + +void pmtimer_init(u16 ioport, u32 khz) +{ + if (!CONFIG_PMTIMER) + return; + dprintf(1, "Using pmtimer, ioport 0x%x, freq %d kHz\n", ioport, khz); + SET_GLOBAL(pmtimer_ioport, ioport); + SET_GLOBAL(cpu_khz, khz); +} + +static u64 pmtimer_get(void) +{ + u16 ioport = GET_GLOBAL(pmtimer_ioport); + u32 wraps = GET_LOW(pmtimer_wraps); + u32 pmtimer = inl(ioport); + + if (pmtimer < GET_LOW(pmtimer_last)) { + wraps++; + SET_LOW(pmtimer_wraps, wraps); + } + SET_LOW(pmtimer_last, pmtimer); + + dprintf(9, "pmtimer: %u:%u\n", wraps, pmtimer); + return (u64)wraps << 24 | pmtimer; +} + static u64 get_tsc(void) { if (unlikely(GET_GLOBAL(no_tsc))) return emulate_tsc(); + if (CONFIG_PMTIMER && GET_GLOBAL(pmtimer_ioport)) + return pmtimer_get(); return rdtscll(); } diff --git a/src/pciinit.c b/src/pciinit.c index 68f302a..31115ee 100644 --- a/src/pciinit.c +++ b/src/pciinit.c @@ -180,6 +180,9 @@ static const struct pci_device_id pci_class_tbl[] = { PCI_DEVICE_END, }; +/* PM Timer ticks per second (HZ) */ +#define PM_TIMER_FREQUENCY 3579545 + /* PIIX4 Power Management device (for ACPI) */ static void piix4_pm_init(struct pci_device *pci, void *arg) { @@ -191,6 +194,8 @@ static void piix4_pm_init(struct pci_device *pci, void *arg) pci_config_writeb(bdf, 0x80, 0x01); /* enable PM io space */ pci_config_writel(bdf, 0x90, PORT_SMB_BASE | 1); pci_config_writeb(bdf, 0xd2, 0x09); /* enable SMBus io space */ + + pmtimer_init(PORT_ACPI_PM_BASE + 0x08, PM_TIMER_FREQUENCY / 1000); } static const struct pci_device_id pci_device_tbl[] = { diff --git a/src/util.h b/src/util.h index 89e928c..1603a57 100644 --- a/src/util.h +++ b/src/util.h @@ -312,6 +312,7 @@ void lpt_setup(void); // clock.c #define PIT_TICK_RATE 1193180 // Underlying HZ of PIT #define PIT_TICK_INTERVAL 65536 // Default interval for 18.2Hz timer +void pmtimer_init(u16 ioport, u32 khz); int check_tsc(u64 end); void timer_setup(void); void ndelay(u32 count); -- 1.7.1
On Tue, Aug 14, 2012 at 07:29:19AM +0200, Gerd Hoffmann wrote:
This patch makes seabios use the acpi pmtimer instead of tsc for timekeeping. The pmtimer has a fixed frequency and doesn't need calibration, thus it doesn't suffer from calibration errors due to a loaded host machine.
The patch looks okay to me, but is it still needed? (I recall seeing something on the kvm list about a bug fix to the main timer.) [...]
+static u64 pmtimer_get(void) +{ + u16 ioport = GET_GLOBAL(pmtimer_ioport); + u32 wraps = GET_LOW(pmtimer_wraps); + u32 pmtimer = inl(ioport); + + if (pmtimer < GET_LOW(pmtimer_last)) { + wraps++; + SET_LOW(pmtimer_wraps, wraps); + } + SET_LOW(pmtimer_last, pmtimer); + + dprintf(9, "pmtimer: %u:%u\n", wraps, pmtimer); + return (u64)wraps << 24 | pmtimer;
BTW, why is this "<< 24", and if it should be that way, shouldn't the pmtimer be "inl(ioport) & 0xffffff" ? -Kevin
On 09/02/2012 11:42 PM, Kevin O'Connor wrote:
On Tue, Aug 14, 2012 at 07:29:19AM +0200, Gerd Hoffmann wrote:
This patch makes seabios use the acpi pmtimer instead of tsc for timekeeping. The pmtimer has a fixed frequency and doesn't need calibration, thus it doesn't suffer from calibration errors due to a loaded host machine.
The patch looks okay to me, but is it still needed? (I recall seeing something on the kvm list about a bug fix to the main timer.)
Timing will always be fragile in a vm, so I think this can make things more robust. -- error compiling committee.c: too many arguments to function
On 09/02/12 22:42, Kevin O'Connor wrote:
On Tue, Aug 14, 2012 at 07:29:19AM +0200, Gerd Hoffmann wrote:
This patch makes seabios use the acpi pmtimer instead of tsc for timekeeping. The pmtimer has a fixed frequency and doesn't need calibration, thus it doesn't suffer from calibration errors due to a loaded host machine.
The patch looks okay to me, but is it still needed? (I recall seeing something on the kvm list about a bug fix to the main timer.)
It is still a good idea to make timing in a virtual machine more robust.
+ u32 pmtimer = inl(ioport);
+ return (u64)wraps << 24 | pmtimer;
BTW, why is this "<< 24", and if it should be that way, shouldn't the pmtimer be "inl(ioport) & 0xffffff" ?
The pmtimer is defined to be 24 bits wide, so the shift is correct. But, yes, the ioport read should better be masked to be on the safe side. v3 will go out in a minute. cheers, Gerd
On 09/05/12 01:27, Gerd Hoffmann wrote:
On 09/02/12 22:42, Kevin O'Connor wrote:
On Tue, Aug 14, 2012 at 07:29:19AM +0200, Gerd Hoffmann wrote:
This patch makes seabios use the acpi pmtimer instead of tsc for timekeeping. The pmtimer has a fixed frequency and doesn't need calibration, thus it doesn't suffer from calibration errors due to a loaded host machine. The patch looks okay to me, but is it still needed? (I recall seeing something on the kvm list about a bug fix to the main timer.) It is still a good idea to make timing in a virtual machine more robust.
+ u32 pmtimer = inl(ioport); + return (u64)wraps << 24 | pmtimer; BTW, why is this "<< 24", and if it should be that way, shouldn't the pmtimer be "inl(ioport) & 0xffffff" ? The pmtimer is defined to be 24 bits wide, so the shift is correct. This is not true in general. It can be either 24 or 32 bits. What it is depends on ACPI data (acpi_gbl_FADT->tmr_val_ext). However it is valid to only used 24 bits.
*/ /*
But, yes, the ioport read should better be masked to be on the safe side. v3 will go out in a minute.
cheers, Gerd
_______________________________________________ SeaBIOS mailing list SeaBIOS@seabios.org http://www.seabios.org/mailman/listinfo/seabios -Don Slutz
Hi,
+ u32 pmtimer = inl(ioport); + return (u64)wraps << 24 | pmtimer; BTW, why is this "<< 24", and if it should be that way, shouldn't the pmtimer be "inl(ioport) & 0xffffff" ? The pmtimer is defined to be 24 bits wide, so the shift is correct. This is not true in general. It can be either 24 or 32 bits. What it is depends on ACPI data (acpi_gbl_FADT->tmr_val_ext).
The piix4 emulated by qemu has 24 bits.
However it is valid to only used 24 bits.
And we certainly want to mask the ioport read (as suggested by kevin and done in v3 of the patch) so we only pick up the 24 bits we actually use. thanks Gerd
participants (4)
-
Avi Kivity -
Don Slutz -
Gerd Hoffmann -
Kevin O'Connor