[Timer学习](zz)Linux HPET Support

Linux HPET Support

IA-PC HPET (High Precision Event Timer) is a specification which was jointly developed by Intel and Microsoft in the early part of this decade.. The latest version is dated October 2004. It’s stated purpose is to

initially supplement and eventually replace the legacy 8254 Programmable Interval Timer and the Real Time Clock Periodic Interrupt generation functions that are currently used as the ‘de-facto’ timer hardware for IA-PCs.

 

The HPET architecture defines a set of timers that can be used by the operating system. A timer block is a combination of a single counter and up to 32 comparators and match registers. The comparator compares the contents of the match register against the value of a free running monotonic up-counter. When the output of the up-counter equals the value in the match register an interrupt is generated. Each of the comparators can output an interrupt. A maximum of 8 timer blocks are supported for a total of 256 timers. Each timer block can have different clocking attributes. Specific implementations may include only a subset of these timers. A minimum of three timers is required.

The specification contains the following block diagram of the HPET architecture.

Some of the timers may be enabled to generate a periodic interrupt. If a timer is set to be periodic, its period is added to the match register each time a match occurs, thus computing the next time for this timer to generate an interrupt.. An up-counter is usually 64 bits wide but 32-bit implementations are permitted by the specification and 64-bit up-counters can also be driven in 32-bit mode. Up-counters run at a minimum of 10 MHz. which is much faster than the older RTC (Real Time Clock) and can thus produce periodic interrupts at a much higher resolution. The registers associated with these timers are mapped to memory space.

The BIOS uses ACPI ( Advanced Configuration and Power Interface) functionality to inform the operating system of the location of the HPET memory-mapped register space. Here is an example of a disassembled ACPI HPET table from an Intel DX48BT2 (AKA BoneTrail) motherboard.

$ cat /sys/firmware/acpi/tables/HPET > /var/tmp/hpet.out
$ iasl -d /var/tmp/hpet.out
$ cat /var/tmp/hpet.dsl
/*
 * Intel ACPI Component Architecture
 * AML Disassembler version 20090123
 *
 * Disassembly of /var/tmp/hpet.out, Sun Jul  5 19:34:47 2009
 *
 * ACPI Data Table [HPET]
 *
 * Format: [HexOffset DecimalOffset ByteLength]  FieldName : FieldValue
 */

[000h 000  4]                    Signature : "HPET"    /* High Precision Event Timer table */
[004h 004  4]                 Table Length : 00000038
[008h 008  1]                     Revision : 01
[009h 009  1]                     Checksum : CE
[00Ah 010  6]                       Oem ID : "INTEL "
[010h 016  8]                 Oem Table ID : "DX48BT2 "
[018h 024  4]                 Oem Revision : 0000076E
[01Ch 028  4]              Asl Compiler ID : "MSFT"
[020h 032  4]        Asl Compiler Revision : 01000013

[024h 036  4]            Hardware Block ID : 8086A301

[028h 040 12]         Timer Block Register : 
[028h 040  1]                     Space ID : 00 (SystemMemory)
[029h 041  1]                    Bit Width : 00
[02Ah 042  1]                   Bit Offset : 00
[02Bh 043  1]                 Access Width : 00
[02Ch 044  8]                      Address : 00000000FED00000

[034h 052  1]              Sequence Number : 00
[035h 053  2]          Minimum Clock Ticks : 0001
[037h 055  1]        Flags (decoded below) : 00
                              Page Protect : 0
                           4K Page Protect : 0
                          64K Page Protect : 0
Raw Table Data

  0000: 48 50 45 54 38 00 00 00 01 CE 49 4E 54 45 4C 20  HPET8.....INTEL
  0010: 44 58 34 38 42 54 32 20 6E 07 00 00 4D 53 46 54  DX48BT2 n...MSFT
  0020: 13 00 00 01 01 A3 86 80 00 00 00 00 00 00 D0 FE  ................
  0030: 00 00 00 00 00 01 00 00                          ........
$

See page 30 of the HPET v1.0a specification for a detailed breakdown of the individual bits in the Event Time Block (called Hardware Block by the AML disassember). Note that only one Event Timer Block need be described in the HPET table in order to bootstrap an operating system. This is the case here. For non-legacy platforms, the Event Timer Block described in the HPET is the one that provides functionality to replace the 8254/RTC Periodic Interrupt Logic.

Other Event Time Blocks are described in the ACPI namespace. Here is the relevant section from the disassembled ACPI DSDT table.

            Device (HPET)
            {
                Name (_HID, EisaId ("PNP0103"))
                Name (_CRS, ResourceTemplate ()
                {
                    Memory32Fixed (ReadOnly,
                        0xFED00000,         // Address Base
                        0x00004000,         // Address Length
                        )
                })
                Method (_STA, 0, NotSerialized)
                {
                    If (HPEE)
                    {
                        Return (0x0F)
                    }
                    Else
                    {
                        Return (Zero)
                    }
                }
            }

Note the assigned PNPID (PNP0103) for the HPET. Because no _UID is specified it means that there are no other HPET timer blocks.

Here is a list of the HPET-related messages outputted when this particular motherboard is booted up under Fedora 11.

$ dmesg | grep -i HPET
ACPI: HPET CFBF2000, 0038 (r1 INTEL  DX48BT2       76E MSFT  1000013)
ACPI: HPET id: 0x8086a301 base: 0xfed00000
hpet clockevent registered
HPET: 4 timers in total, 0 timers will be used for per-cpu timer
hpet0: at MMIO 0xfed00000, IRQs 2, 8, 0, 0
hpet0: 4 comparators, 64-bit 14.318180 MHz counter
rtc0: alarms up to one month, 114 bytes nvram, hpet irqs
$

The first line is outputted when the ACPI HPET table is read. The second line is outputted when the ACPI HPET table is mapped into memory by …/arch/x86/kernel/acpi/boot.c. The next line is outputted when the HPET legacy interrupts are started and HPET is registered as the global clock. The following line is outputted when the kernel checks to ensure that at least one timer is reserved for userspace (/dev/hpet.) The next two lines of output comes from the HPET device driver (…/drivers/char/hpet.c.) It shows that 2 timers have allocated interrupts and two do not..

Here is the relevant part of the output from /proc/time_list as it relates to HPET:

Tick Device: mode:     1
Broadcast device
Clock Event Device: hpet
 max_delta_ns:   149983005959
 min_delta_ns:   5000
 mult:           61496114
 shift:          32
 mode:           3
 next_event:     9223372036854775807 nsecs
 set_next_event: hpet_legacy_next_event
 set_mode:       hpet_legacy_set_mode
 event_handler:  tick_handle_oneshot_broadcast
tick_broadcast_mask: 00000000
tick_broadcast_oneshot_mask: 00000000

Here is the output from /proc/sys/dev/hpet and /proc/driver/rtc:

$ cat /proc/sys/dev/hpet/max-user-freq
64
$ cat /proc/driver/rtc
rtc_time	: 06:34:31
rtc_date	: 2009-07-06
alrm_time	: **:24:40
alrm_date	: ****-**-**
alarm_IRQ	: no
alrm_pending	: no
24hr		: yes
periodic_IRQ	: no
update_IRQ	: no
HPET_emulated	: yes
DST_enable	: no
periodic_freq	: 1024
batt_status	: okay

 

The HPET driver (/dev/hpet) has a similar API to the Real Time Clock driver. It is a character device which can support any number of HPET devices. The kernel API has three interfaces exported from the driver:

hpet_register( struct hpet_task *tp, int periodic )
hpet_unregister( struct hpet_task *tp )
hpet_control( struct hpet_task *tp, unsigned int cmd, unsigned long arg )

The userspace interface to HPET is defined in the header /usr/include/linux/hpet.h. The current set of supported operations is:

#define HPET_IE_ON      _IO('h', 0x01)                        /* interrupt on */
#define HPET_IE_OFF     _IO('h', 0x02)                        /* interrupt off */
#define HPET_INFO       _IOR('h', 0x03, struct hpet_info)     /* get information */
#define HPET_EPI        _IO('h', 0x04)                        /* enable periodic */
#define HPET_DPI        _IO('h', 0x05)                        /* disable periodic */
#define HPET_IRQFREQ    _IOW('h', 0x6, unsigned long)         /* set frequency */

The following example shows how to use the published interface to access a HPET and call a simple periodic signal handler hpet_alarm between 2 and 99 times a second.

#include <stdio.h>
#include <stdlib.h;>
#include <fcntl.h>
#include <time.h>
#include <sys/types.h>
#include <sys/wait.h>
#include <signal.h>
#include <fcntl.h>
#include <sys/time.h>
#include <linux/hpet.h>
#include <stdint.h>
#include <sys/ioctl.h>
#include <signal.h>

static uint16_t hpet_sigio_count;
static uint64_t secs;

static void
hpet_alarm(int val)
{
   struct timespec t;
   clock_gettime(CLOCK_REALTIME, &t);

   if (!secs)  secs = t.tv_sec;

   fprintf(stderr, "hpet_alarm called. iteration: %2d  secs: %ld  nsecs: %ld \n",
                    hpet_sigio_count, (t.tv_sec - secs) , t.tv_sec * 100000 + t.tv_nsec );

   hpet_sigio_count++;
}

int
main(int argc, const char **argv)
{
    struct sigaction old, new;
    struct hpet_info info;
    int              frequency;
    int              iterations;
    int              retval = 0;
    int              fd;
    int              r, i, value;

    if (argc != 3) {
        fprintf(stderr, "Usage: %s frequency(1-64) iterations(10-99)\n", argv[0]);
        return -1;
    }

    frequency = atoi(argv[1]);
    iterations = atoi(argv[2]);

    if (frequency > 64 || frequency < 1 ) {
        fprintf(stderr, "ERROR: Invalid value for frequency\n");
        return -1;
    }

    if (iterations < 10 || iterations > 99 ) {
        fprintf(stderr, "ERROR: Invalid value for iterations\n");
        return -1;
    }

    hpet_sigio_count = 0;

    sigemptyset(&new.sa_mask);
    new.sa_flags = 0;
    new.sa_handler = hpet_alarm;

    sigaction(SIGIO, NULL, &old);
    sigaction(SIGIO, &new, NULL);

    fd = open("/dev/hpet", O_RDONLY);
    if (fd < 0) {
        fprintf(stderr, "ERROR: Failed to open /dev/hpet\n");
        return -1;
    }

    if ((fcntl(fd, F_SETOWN, getpid()) == 1) ||
        ((value = fcntl(fd, F_GETFL)) == 1) ||
        (fcntl(fd, F_SETFL, value | O_ASYNC) == 1)) {
        fprintf(stderr, "ERROR: fcntl failed\n");
        retval = 1;
        goto fail;
    }

    if (ioctl(fd, HPET_IRQFREQ, frequency) < 0) {
        fprintf(stderr, "ERROR: Could not set /dev/hpet to have a %2dHz timer\n", frequency);
        retval = 2;
        goto fail;
    }

    if (ioctl(fd, HPET_INFO, &info) < 0) {
        fprintf(stderr, "ERROR: failed to get info\n");
        retval = 3;
        goto fail;
    }

    fprintf(stdout, "\nhi_ireqfreq: 0x%lx  hi_flags: %0x%lx  hi_hpet: 0x%x  hi_timer: 0x%x\n\n",
            info.hi_ireqfreq,  info.hi_flags, info.hi_hpet, info.hi_timer);

    r = ioctl(fd, HPET_EPI, 0);
    if (info.hi_flags && (r < 0)) {
        fprintf(stderr, "ERROR:  HPET_EPI failed\n");
        retval = 4;
        goto fail;
    }

    if (ioctl(fd, HPET_IE_ON, 0) < 0) {
        fprintf(stderr, "ERROR: HPET_IE_ON failed\n");
        retval = 5;
        goto fail;
    }

    /* wait for specified number of signal interrupts */
    for (i = 0; i < iterations; i++) {
        (void) pause();
    }

    if (ioctl(fd, HPET_IE_OFF, 0) < 0) {
        fprintf(stderr, "ERROR: HPET_IE_OFF failed\n");
        retval = 6;
    }

fail:
    sigaction(SIGIO, &old, NULL);

    if (fd > 0)
        close(fd);

    return retval;
}

Here is the output from this example when it is invoked with a frequency of 32 and an iteration count of 64.

$ sudo ./hpet_example 32 64

hi_ireqfreq: 0x20  hi_flags: 00  hi_hpet: 0x2  hi_timer: 0x4a1cb9c8

hpet_alarm called. iteration:  0  secs: 0  nsecs: 124683205055050
hpet_alarm called. iteration:  1  secs: 0  nsecs: 124683236313149
hpet_alarm called. iteration:  2  secs: 0  nsecs: 124683267566342
hpet_alarm called. iteration:  3  secs: 0  nsecs: 124683298821905
hpet_alarm called. iteration:  4  secs: 0  nsecs: 124683330077493
hpet_alarm called. iteration:  5  secs: 0  nsecs: 124683361341893
hpet_alarm called. iteration:  6  secs: 0  nsecs: 124683392590764
hpet_alarm called. iteration:  7  secs: 0  nsecs: 124683423849157
hpet_alarm called. iteration:  8  secs: 0  nsecs: 124683455101917
hpet_alarm called. iteration:  9  secs: 0  nsecs: 124683486357683
hpet_alarm called. iteration: 10  secs: 0  nsecs: 124683517617931
hpet_alarm called. iteration: 11  secs: 0  nsecs: 124683548872198
hpet_alarm called. iteration: 12  secs: 1  nsecs: 124682580229541
hpet_alarm called. iteration: 13  secs: 1  nsecs: 124682611481235
hpet_alarm called. iteration: 14  secs: 1  nsecs: 124682642740016
hpet_alarm called. iteration: 15  secs: 1  nsecs: 124682673992697
hpet_alarm called. iteration: 16  secs: 1  nsecs: 124682705247479
hpet_alarm called. iteration: 17  secs: 1  nsecs: 124682736504664
hpet_alarm called. iteration: 18  secs: 1  nsecs: 124682767758840
hpet_alarm called. iteration: 19  secs: 1  nsecs: 124682799014280
hpet_alarm called. iteration: 20  secs: 1  nsecs: 124682830270129
hpet_alarm called. iteration: 21  secs: 1  nsecs: 124682861530334
hpet_alarm called. iteration: 22  secs: 1  nsecs: 124682892784577
hpet_alarm called. iteration: 23  secs: 1  nsecs: 124682924038220
hpet_alarm called. iteration: 24  secs: 1  nsecs: 124682955294110
hpet_alarm called. iteration: 25  secs: 1  nsecs: 124682986550572
hpet_alarm called. iteration: 26  secs: 1  nsecs: 124683017805756
hpet_alarm called. iteration: 27  secs: 1  nsecs: 124683049061117
hpet_alarm called. iteration: 28  secs: 1  nsecs: 124683080318331
hpet_alarm called. iteration: 29  secs: 1  nsecs: 124683111576954
hpet_alarm called. iteration: 30  secs: 1  nsecs: 124683142828988
hpet_alarm called. iteration: 31  secs: 1  nsecs: 124683174083954
hpet_alarm called. iteration: 32  secs: 1  nsecs: 124683205337967
hpet_alarm called. iteration: 33  secs: 1  nsecs: 124683236593144
hpet_alarm called. iteration: 34  secs: 1  nsecs: 124683267851530
hpet_alarm called. iteration: 35  secs: 1  nsecs: 124683299104054
hpet_alarm called. iteration: 36  secs: 1  nsecs: 124683330358748
hpet_alarm called. iteration: 37  secs: 1  nsecs: 124683361617445
hpet_alarm called. iteration: 38  secs: 1  nsecs: 124683392870249
hpet_alarm called. iteration: 39  secs: 1  nsecs: 124683424124489
hpet_alarm called. iteration: 40  secs: 1  nsecs: 124683455379717
hpet_alarm called. iteration: 41  secs: 1  nsecs: 124683486634424
hpet_alarm called. iteration: 42  secs: 1  nsecs: 124683517889149
hpet_alarm called. iteration: 43  secs: 1  nsecs: 124683549144315
hpet_alarm called. iteration: 44  secs: 2  nsecs: 124682580500695
hpet_alarm called. iteration: 45  secs: 2  nsecs: 124682611761325
hpet_alarm called. iteration: 46  secs: 2  nsecs: 124682643011863
hpet_alarm called. iteration: 47  secs: 2  nsecs: 124682674265864
hpet_alarm called. iteration: 48  secs: 2  nsecs: 124682705521034
hpet_alarm called. iteration: 49  secs: 2  nsecs: 124682736776049
hpet_alarm called. iteration: 50  secs: 2  nsecs: 124682768030654
hpet_alarm called. iteration: 51  secs: 2  nsecs: 124682799285398
hpet_alarm called. iteration: 52  secs: 2  nsecs: 124682830544701
hpet_alarm called. iteration: 53  secs: 2  nsecs: 124682861797319
hpet_alarm called. iteration: 54  secs: 2  nsecs: 124682893051578
hpet_alarm called. iteration: 55  secs: 2  nsecs: 124682924306748
hpet_alarm called. iteration: 56  secs: 2  nsecs: 124682955562132
hpet_alarm called. iteration: 57  secs: 2  nsecs: 124682986823545
hpet_alarm called. iteration: 58  secs: 2  nsecs: 124683018073636
hpet_alarm called. iteration: 59  secs: 2  nsecs: 124683049327560
hpet_alarm called. iteration: 60  secs: 2  nsecs: 124683080586707
hpet_alarm called. iteration: 61  secs: 2  nsecs: 124683111841132
hpet_alarm called. iteration: 62  secs: 2  nsecs: 124683143095147
hpet_alarm called. iteration: 63  secs: 2  nsecs: 124683174349985
hpet_alarm called. iteration: 64  secs: 2  nsecs: 124683205607103
$

Well, I think that I have provided you with enough information so that you should now be able to go away and experiment with the HPET interface yourself.

By the way, not all VMware products support HPET. Currently ESX does not provide a virtual HPET to guest operating systems and in some cases it may be necessary to disable HPET altogether because of timer drift in virtual machines. See VMware TimeKeeping for more information.

P.S. I tested the the above example on an Intel DX48BT2 motherboard running a 2.6.29.5-191 kernel.