hw: make TCO watchdog actually work by default for Q35
cover hw: make TCO watchdog actually work by default for Q35
The TCO watchdog is unconditionally integrated into the Q35 machine type by default, but at the same time is unconditionally disabled from firing by a host config option that overrides guest OS attempts to enable it. People have to know to set a magic -global to make it non-broken
IOW we're exposing a broken watchdog by default to all Q35 machines, but which to the guest OS & its apps looks fully functional :-(
This behaviour was set in response to feedback from Michael:
https://lists.gnu.org/archive/html/qemu-devel/2015-06/msg07128.html
"I think sample high is a safer default."
but as explained in the commit message in the last patch, I think the watchdog defaults were already safe without that pin strap setting. The guest OS needs to take explicit action to clear the guest visible 'no reboot' flag, and so we don't need a second guest hidden 'no reboot' flag to override that choice IMHO. Am I missing something ?
NB, I'm toggling this for 7.2 machine type since that's the current git latest machine. Since this has already been "broken" for 7 years though, I am ambivalent about whether we try todo this for 7.2, vs just wait until the 8.0 machine types arrive.
Daniel P. Berrangé (4):
hw/acpi: add trace events for TCO watchdog register access hw/isa: add trace events for ICH9 LPC chip config access hw/watchdog: add trace events for watchdog action handling hw/isa: enable TCO watchdog reboot pin strap by default
hw/acpi/tco.c | 41 +++++++++++++++++++++++++++------------- hw/acpi/trace-events | 2 ++ hw/i386/pc.c | 4 +++- hw/isa/lpc_ich9.c | 5 ++++- hw/isa/trace-events | 4 ++++ hw/watchdog/trace-events | 4 ++++ hw/watchdog/watchdog.c | 4 ++++ tests/qtest/tco-test.c | 2 +- 8 files changed, 50 insertions(+), 16 deletions(-)
Signature · 1 line
2.37.3
PATCH 1/4 hw/acpi: add trace events for TCO watchdog register access
These tracepoints aid in understanding and debugging the guest drivers for the TCO watchdog.
--- hw/acpi/tco.c | 41 ++++++++++++++++++++++++++++------------- hw/acpi/trace-events | 2 ++ 2 files changed, 30 insertions(+), 13 deletions(-)
hw/acpi/tco.c +28 -13
@@ -86,6 +86,7 @@ static inline int can_start_tco_timer(TCOIORegs *tr)static uint32_t tco_ioport_readw (TCOIORegs *tr,uint32_t addr) {uint16_t rld; +uint32_t ret =0 ;switch (addr) {case TCO_RLD :
@@ -96,35 +97,49 @@ static uint32_t tco_ioport_readw(TCOIORegs *tr, uint32_t addr) }else { rld = tr->tco.rld; } -return rld; + ret = rld; +break ;case TCO_DAT_IN : -return tr->tco.din; + ret = tr->tco.din; +break ;case TCO_DAT_OUT : -return tr->tco.dout; + ret = tr->tco.dout; +break ;case TCO1_STS : -return tr->tco.sts1; + ret = tr->tco.sts1; +break ;case TCO2_STS : -return tr->tco.sts2; + ret = tr->tco.sts2; +break ;case TCO1_CNT : -return tr->tco.cnt1; + ret = tr->tco.cnt1; +break ;case TCO2_CNT : -return tr->tco.cnt2; + ret = tr->tco.cnt2; +break ;case TCO_MESSAGE1 : -return tr->tco.msg1; + ret = tr->tco.msg1; +break ;case TCO_MESSAGE2 : -return tr->tco.msg2; + ret = tr->tco.msg2; +break ;case TCO_WDCNT : -return tr->tco.wdcnt; + ret = tr->tco.wdcnt; +break ;case TCO_TMR : -return tr->tco.tmr; + ret = tr->tco.tmr; +break ;case SW_IRQ_GEN : -return tr->sw_irq_gen; + ret = tr->sw_irq_gen; +break ; } -return 0 ; +trace_tco_io_read (addr, ret); +return ret; }static void tco_ioport_writew (TCOIORegs *tr,uint32_t addr,uint32_t val) { +trace_tco_io_write (addr, val);switch (addr) {case TCO_RLD : tr->timeouts_no =0 ;
hw/acpi/trace-events +2 -0
@@ -55,6 +55,8 @@ piix4_gpe_writeb(uint64_t addr, unsigned width, uint64_t val) "addr: 0x%" PRIx64 # tco.c tco_timer_reload(int ticks, int msec) "ticks=%d (%d ms)" tco_timer_expired(int timeouts_no, bool strap, bool no_reboot) "timeouts_no=%d no_reboot=%d/%d" +tco_io_write(uint64_t addr, uint32_t val) "addr=0x%" PRIx64 " val=0x%" PRIx32 +tco_io_read(uint64_t addr, uint32_t val) "addr=0x%" PRIx64 " val=0x%" PRIx32 # erst.c acpi_erst_reg_write(uint64_t addr, uint64_t val, unsigned size) "addr: 0x%04" PRIx64 " <== 0x%016" PRIx64 " (size: %u)"
Signature · 1 line
2.37.3
Re: [PATCH 1/4] hw/acpi: add trace events for TCO watchdog register access
Quoted 89 lines · Daniel P. Berrangé
These tracepoints aid in understanding and debugging the guest drivers for the TCO watchdog.
Signed-off-by Daniel P. Berrangé <[email protected]>--- hw/acpi/tco.c | 41 ++++++++++++++++++++++++++++------------- hw/acpi/trace-events | 2 ++ 2 files changed, 30 insertions(+), 13 deletions(-)hw/acpi/tco.c +28
-13@@ -86,6 +86,7 @@ static inline int can_start_tco_timer(TCOIORegs *tr)static uint32_t tco_ioport_readw (TCOIORegs *tr,uint32_t addr) {uint16_t rld; +uint32_t ret =0 ;switch (addr) {case TCO_RLD :@@ -96,35 +97,49 @@ static uint32_t tco_ioport_readw(TCOIORegs *tr, uint32_t addr) }else { rld = tr->tco.rld; } -return rld; + ret = rld; +break ;case TCO_DAT_IN : -return tr->tco.din; + ret = tr->tco.din; +break ;case TCO_DAT_OUT : -return tr->tco.dout; + ret = tr->tco.dout; +break ;case TCO1_STS : -return tr->tco.sts1; + ret = tr->tco.sts1; +break ;case TCO2_STS : -return tr->tco.sts2; + ret = tr->tco.sts2; +break ;case TCO1_CNT : -return tr->tco.cnt1; + ret = tr->tco.cnt1; +break ;case TCO2_CNT : -return tr->tco.cnt2; + ret = tr->tco.cnt2; +break ;case TCO_MESSAGE1 : -return tr->tco.msg1; + ret = tr->tco.msg1; +break ;case TCO_MESSAGE2 : -return tr->tco.msg2; + ret = tr->tco.msg2; +break ;case TCO_WDCNT : -return tr->tco.wdcnt; + ret = tr->tco.wdcnt; +break ;case TCO_TMR : -return tr->tco.tmr; + ret = tr->tco.tmr; +break ;case SW_IRQ_GEN : -return tr->sw_irq_gen; + ret = tr->sw_irq_gen; +break ; } -return 0 ; +trace_tco_io_read (addr, ret); +return ret; }static void tco_ioport_writew (TCOIORegs *tr,uint32_t addr,uint32_t val) { +trace_tco_io_write (addr, val);switch (addr) {case TCO_RLD : tr->timeouts_no =0 ;hw/acpi/trace-events +0
-0
hw/acpi/trace-events piix4_gpe_writeb(uint64_t addr, unsigned width, uint64_t val) "addr: 0x%" PRIx64 · line 62
Quoted 8 lines
hw/acpi/trace-events +2
-0@@ -55,6 +55,8 @@ piix4_gpe_writeb(uint64_t addr, unsigned width, uint64_t val) "addr: 0x%" PRIx64 # tco.c tco_timer_reload(int ticks, int msec) "ticks=%d (%d ms)" tco_timer_expired(int timeouts_no, bool strap, bool no_reboot) "timeouts_no=%d no_reboot=%d/%d" +tco_io_write(uint64_t addr, uint32_t val) "addr=0x%" PRIx64 " val=0x%" PRIx32 +tco_io_read(uint64_t addr, uint32_t val) "addr=0x%" PRIx64 " val=0x%" PRIx32 # erst.c acpi_erst_reg_write(uint64_t addr, uint64_t val, unsigned size) "addr: 0x%04" PRIx64 " <== 0x%016" PRIx64 " (size: %u)"
Rich.
Signature · 4 lines
Richard Jones, Virtualization Group, Red Hat http://people.redhat.com/~rjones Read my programming and virtualization blog: http://rwmj.wordpress.com virt-builder quickly builds VMs from scratch http://libguestfs.org/virt-builder.1.html
PATCH 2/4 hw/isa: add trace events for ICH9 LPC chip config access
In reply to Daniel P. BerrangéThese tracepoints aid in understanding and debugging the guest drivers for the TCO watchdog.
--- hw/isa/lpc_ich9.c | 3 +++ hw/isa/trace-events | 4 ++++ 2 files changed, 7 insertions(+)
hw/isa/lpc_ich9.c +3 -0
@@ -51,6 +51,7 @@#include "hw/nvram/fw_cfg.h" #include "qemu/cutils.h" #include "hw/acpi/acpi_aml_interface.h" +#include "trace.h" /*****************************************************************************/ /* ICH9 LPC PCI to ISA bridge */
@@ -161,6 +162,7 @@ static void ich9_cc_write(void *opaque, hwaddr addr, {ICH9LPCState *lpc = (ICH9LPCState *)opaque; +trace_ich9_cc_write (addr, val, len);ich9_cc_addr_len (&addr, &len);memcpy (lpc->chip_config + addr, &val, len);pci_bus_fire_intx_routing_notifier (pci_get_bus (&lpc->d));
@@ -176,6 +178,7 @@ static uint64_t ich9_cc_read(void *opaque, hwaddr addr,uint32_t val =0 ;ich9_cc_addr_len (&addr, &len);memcpy (&val, lpc->chip_config + addr, len); +trace_ich9_cc_read (addr, val, len);return val; }
hw/isa/trace-events +4 -0
@@ -21,3 +21,7 @@ via_pm_io_read(uint32_t addr, uint32_t val, int len) "addr 0x%x val 0x%x len 0x% via_pm_io_write(uint32_t addr, uint32_t val, int len) "addr 0x%x val 0x%x len 0x%x" via_superio_read(uint8_t addr, uint8_t val) "addr 0x%x val 0x%x" via_superio_write(uint8_t addr, uint32_t val) "addr 0x%x val 0x%x" + +# lpc_ich9.c +ich9_cc_write(uint64_t addr, uint64_t val, unsigned len) "addr=0x%"PRIx64 " val=0x%"PRIx64 " len=%u" +ich9_cc_read(uint64_t addr, uint64_t val, unsigned len) "addr=0x%"PRIx64 " val=0x%"PRIx64 " len=%u"
Signature · 1 line
2.37.3
Re: [PATCH 2/4] hw/isa: add trace events for ICH9 LPC chip config access
Quoted 42 lines · Daniel P. Berrangé
These tracepoints aid in understanding and debugging the guest drivers for the TCO watchdog.
Signed-off-by Daniel P. Berrangé <[email protected]>--- hw/isa/lpc_ich9.c | 3 +++ hw/isa/trace-events | 4 ++++ 2 files changed, 7 insertions(+)hw/isa/lpc_ich9.c +3
-0@@ -51,6 +51,7 @@#include "hw/nvram/fw_cfg.h" #include "qemu/cutils.h" #include "hw/acpi/acpi_aml_interface.h" +#include "trace.h" /*****************************************************************************/ /* ICH9 LPC PCI to ISA bridge */ @@ -161,6 +162,7 @@ static void ich9_cc_write(void *opaque, hwaddr addr, {ICH9LPCState *lpc = (ICH9LPCState *)opaque; +trace_ich9_cc_write (addr, val, len);ich9_cc_addr_len (&addr, &len);memcpy (lpc->chip_config + addr, &val, len);pci_bus_fire_intx_routing_notifier (pci_get_bus (&lpc->d));@@ -176,6 +178,7 @@ static uint64_t ich9_cc_read(void *opaque, hwaddr addr,uint32_t val =0 ;ich9_cc_addr_len (&addr, &len);memcpy (&val, lpc->chip_config + addr, len); +trace_ich9_cc_read (addr, val, len);return val; }hw/isa/trace-events +0
-0
hw/isa/trace-events via_pm_io_read(uint32_t addr, uint32_t val, int len) "addr 0x%x val 0x%x len 0x% · line 27
Quoted 7 lines
hw/isa/trace-events +4
-0@@ -21,3 +21,7 @@ via_pm_io_read(uint32_t addr, uint32_t val, int len) "addr 0x%x val 0x%x len 0x% via_pm_io_write(uint32_t addr, uint32_t val, int len) "addr 0x%x val 0x%x len 0x%x" via_superio_read(uint8_t addr, uint8_t val) "addr 0x%x val 0x%x" via_superio_write(uint8_t addr, uint32_t val) "addr 0x%x val 0x%x" + +# lpc_ich9.c +ich9_cc_write(uint64_t addr, uint64_t val, unsigned len) "addr=0x%"PRIx64 " val=0x%"PRIx64 " len=%u" +ich9_cc_read(uint64_t addr, uint64_t val, unsigned len) "addr=0x%"PRIx64 " val=0x%"PRIx64 " len=%u"
I can't help thinking that the trace-events file ought to be generated from the source code ...
Rich.
Signature · 5 lines
Richard Jones, Virtualization Group, Red Hat http://people.redhat.com/~rjones Read my programming and virtualization blog: http://rwmj.wordpress.com virt-top is 'top' for virtual machines. Tiny program with many powerful monitoring features, net stats, disk stats, logging, etc. http://people.redhat.com/~rjones/virt-top
PATCH 3/4 hw/watchdog: add trace events for watchdog action handling
In reply to Daniel P. BerrangéThe tracepoints aid in debugging the triggering of watchdog devices.
--- hw/watchdog/trace-events | 4 ++++ hw/watchdog/watchdog.c | 4 ++++ 2 files changed, 8 insertions(+)
hw/watchdog/trace-events +4 -0
@@ -16,3 +16,7 @@ spapr_watchdog_stop(uint64_t num, uint64_t ret) "num=%" PRIu64 " ret=%" PRId64 spapr_watchdog_query(uint64_t caps) "caps=0x%" PRIx64 spapr_watchdog_query_lpm(uint64_t caps) "caps=0x%" PRIx64 spapr_watchdog_expired(uint64_t num, unsigned action) "num=%" PRIu64 " action=%u" + +# watchdog.c +watchdog_perform_action(unsigned int action) "action=%d" +watchdog_set_action(unsigned int action) "action=%d"
hw/watchdog/watchdog.c +4 -0
@@ -30,6 +30,7 @@#include "sysemu/watchdog.h" #include "hw/nmi.h" #include "qemu/help_option.h" +#include "trace.h" static WatchdogAction watchdog_action =WATCHDOG_ACTION_RESET ;
@@ -43,6 +44,8 @@ WatchdogAction get_watchdog_action(void)*/ void watchdog_perform_action (void ) { +trace_watchdog_perform_action (watchdog_action); +switch (watchdog_action) {case WATCHDOG_ACTION_RESET :/* same as 'system_reset' in monitor */ qapi_event_send_watchdog (WATCHDOG_ACTION_RESET );
@@ -89,4 +92,5 @@ void watchdog_perform_action(void)void qmp_watchdog_set_action (WatchdogAction action,Error **errp) { watchdog_action = action; +trace_watchdog_set_action (watchdog_action); }
Signature · 1 line
2.37.3
Re: [PATCH 3/4] hw/watchdog: add trace events for watchdog action handling
Quoted 42 lines · Daniel P. Berrangé
The tracepoints aid in debugging the triggering of watchdog devices.
Signed-off-by Daniel P. Berrangé <[email protected]>--- hw/watchdog/trace-events | 4 ++++ hw/watchdog/watchdog.c | 4 ++++ 2 files changed, 8 insertions(+)hw/watchdog/trace-events +4
-0@@ -16,3 +16,7 @@ spapr_watchdog_stop(uint64_t num, uint64_t ret) "num=%" PRIu64 " ret=%" PRId64 spapr_watchdog_query(uint64_t caps) "caps=0x%" PRIx64 spapr_watchdog_query_lpm(uint64_t caps) "caps=0x%" PRIx64 spapr_watchdog_expired(uint64_t num, unsigned action) "num=%" PRIu64 " action=%u" + +# watchdog.c +watchdog_perform_action(unsigned int action) "action=%d" +watchdog_set_action(unsigned int action) "action=%d"hw/watchdog/watchdog.c +3
-0@@ -30,6 +30,7 @@#include "sysemu/watchdog.h" #include "hw/nmi.h" #include "qemu/help_option.h" +#include "trace.h" static WatchdogAction watchdog_action =WATCHDOG_ACTION_RESET ;@@ -43,6 +44,8 @@ WatchdogAction get_watchdog_action(void)*/ void watchdog_perform_action (void ) { +trace_watchdog_perform_action (watchdog_action); +switch (watchdog_action) {case WATCHDOG_ACTION_RESET :/* same as 'system_reset' in monitor */ qapi_event_send_watchdog (WATCHDOG_ACTION_RESET );
hw/watchdog/watchdog.c void watchdog_perform_action(void) · line 96
Quoted 5 lines
hw/watchdog/watchdog.c +1
-0@@ -89,4 +92,5 @@ void watchdog_perform_action(void)void qmp_watchdog_set_action (WatchdogAction action,Error **errp) { watchdog_action = action; +trace_watchdog_set_action (watchdog_action); }
Rich.
Signature · 4 lines
Richard Jones, Virtualization Group, Red Hat http://people.redhat.com/~rjones Read my programming and virtualization blog: http://rwmj.wordpress.com virt-builder quickly builds VMs from scratch http://libguestfs.org/virt-builder.1.html
Re: [PATCH 3/4] hw/watchdog: add trace events for watchdog action handling
In reply to Daniel P. BerrangéQuoted 13 lines · Daniel P. Berrangé
The tracepoints aid in debugging the triggering of watchdog devices.
Signed-off-by Daniel P. Berrangé <[email protected]>--- hw/watchdog/trace-events | 4 ++++ hw/watchdog/watchdog.c | 4 ++++ 2 files changed, 8 insertions(+)hw/watchdog/trace-events +0
-0
hw/watchdog/trace-events spapr_watchdog_stop(uint64_t num, uint64_t ret) "num=%" PRIu64 " ret=%" PRId64 · line 22
Quoted 7 lines
hw/watchdog/trace-events +4
-0@@ -16,3 +16,7 @@ spapr_watchdog_stop(uint64_t num, uint64_t ret) "num=%" PRIu64 " ret=%" PRId64 spapr_watchdog_query(uint64_t caps) "caps=0x%" PRIx64 spapr_watchdog_query_lpm(uint64_t caps) "caps=0x%" PRIx64 spapr_watchdog_expired(uint64_t num, unsigned action) "num=%" PRIu64 " action=%u" + +# watchdog.c +watchdog_perform_action(unsigned int action) "action=%d" +watchdog_set_action(unsigned int action) "action=%d"
"%u", otherwise:
PATCH 4/4 hw/isa: enable TCO watchdog reboot pin strap by default
In reply to Daniel P. BerrangéThe TCO watchdog implementation default behaviour from POV of the guest OS relies on the initial values for two I/O ports:
* TCO1_CNT == 0x0
Since bit 11 (TCO Timer Halt) is clear, the watchdog state is considered to be initially running
* GCS == 0x20
Since bit 5 (No Reboot) is set, the watchdog will not trigger when the timer expires
This is a safe default, because the No Reboot bit will prevent the watchdog from triggering if the guest OS is unaware of its existance, or is slow in configuring it. When a Linux guest initializes the TCO watchdog, it will attempt to clear the "No Reboot" flag, and read the value back. If the clear was honoured, the driver will treat this as an indicator that the watchdog is functional and create the geust watchdog device.
QEMU implements a second "no reboot" flag, however, via pin straps which overrides the behaviour of the guest controlled "no reboot" flag:
commit 5add35bec1e249bb5345a47008c8f298d4760be4
Author: Paulo Alcantara <[email protected]>
Date: Sun Jun 28 14:58:58 2015 -0300
ich9: implement strap SPKR pin logic
This second 'noreboot' pin was defaulted to high, which also inhibits triggering of the requested watchdog actions, unless QEMU is launched with the magic flag "-global ICH9-LPC.noreboot=false".
This is a bad default as we are exposing a watchdog to every guest OS using the q35 machine type, but preventing it from actually doing what it is designed to do. What is worse is that the guest OS and its apps have no way to know that the watchdog is never going to fire, due to this second 'noreboot' pin.
If a guest OS had no watchdog device at all, then apps whose operation and/or data integrity relies on a watchdog can refuse to launch, and alert the administrator of the problematic deployment. With Q35 machines unconditionally exposing a watchdog though, apps will think their deployment is correct but in fact have no protection at all.
This patch flips the default of the second 'no reboot' flag, so that configured watchdog actions will be honoured out of the box for the 7.2 Q35 machine type onwards, if the guest enables use of the watchdog.
--- hw/i386/pc.c | 4 +++- hw/isa/lpc_ich9.c | 2 +- tests/qtest/tco-test.c | 2 +- 3 files changed, 5 insertions(+), 3 deletions(-)
hw/i386/pc.c +3 -1
@@ -108,7 +108,9 @@ { "qemu64-" TYPE_X86_CPU, "model-id", "QEMU Virtual CPU version " v, },\ { "athlon-" TYPE_X86_CPU, "model-id", "QEMU Virtual CPU version " v, }, -GlobalProperty pc_compat_7_1[] = {}; +GlobalProperty pc_compat_7_1[] = { + { "ICH9-LPC", "noreboot", "true" }, +};const size_t pc_compat_7_1_len =G_N_ELEMENTS (pc_compat_7_1);GlobalProperty pc_compat_7_0[] = {};
hw/isa/lpc_ich9.c +1 -1
@@ -789,7 +789,7 @@ static const VMStateDescription vmstate_ich9_lpc = { };static Property ich9_lpc_properties[] = { -DEFINE_PROP_BOOL ("noreboot" , ICH9LPCState, pin_strap.spkr_hi, true), +DEFINE_PROP_BOOL ("noreboot" , ICH9LPCState, pin_strap.spkr_hi, false),DEFINE_PROP_BOOL ("smm-compat" , ICH9LPCState, pm.smm_compat, false),DEFINE_PROP_BIT64 ("x-smi-broadcast" , ICH9LPCState, smi_host_features,ICH9_LPC_SMI_F_BROADCAST_BIT , true),
tests/qtest/tco-test.c +1 -1
@@ -60,7 +60,7 @@ static void test_init(TestData *d)QTestState *qs; qs =qtest_initf ("-machine q35 %s %s" , - d->noreboot ?"" :"-global ICH9-LPC.noreboot=false" , + d->noreboot ?"-global ICH9-LPC.noreboot=true" :"" , !d->args ?"" : d->args);qtest_irq_intercept_in (qs,"ioapic" );
Signature · 1 line
2.37.3
Re: [PATCH 4/4] hw/isa: enable TCO watchdog reboot pin strap by default
Quoted 13 lines · Daniel P. Berrangé
The TCO watchdog implementation default behaviour from POV of the guest OS relies on the initial values for two I/O ports:
*
TCO1_CNT==0x0Since bit 11 (TCO Timer Halt) is clear, the watchdog state is considered to be initially running
* GCS ==
0x20Since bit 5 (No Reboot) is set, the watchdog will not trigger when the timer expires
Quoted 6 lines
This is a safe default, because the No Reboot bit will prevent the watchdog from triggering if the guest OS is unaware of its existance, or is slow in configuring it. When a Linux guest initializes the TCO watchdog, it will attempt to clear the "No Reboot" flag, and read the value back. If the clear was honoured, the driver will treat this as an indicator that the watchdog is functional and create the geust
Typo: "guest"
Quoted 28 lines
watchdog device.
QEMU implements a second "no reboot" flag, however, via pin straps which overrides the behaviour of the guest controlled "no reboot" flag:
commit
5add35bec1e249bb5345a47008c8f298d4760be4
Author: Paulo Alcantara <[email protected]>
Date: Sun Jun 28 14:58:58 2015 -0300ich9: implement strap SPKR pin logic
This second 'noreboot' pin was defaulted to high, which also inhibits triggering of the requested watchdog actions, unless QEMU is launched with the magic flag "-global ICH9-LPC.noreboot=false".
This is a bad default as we are exposing a watchdog to every guest OS using the q35 machine type, but preventing it from actually doing what it is designed to do. What is worse is that the guest OS and its apps have no way to know that the watchdog is never going to fire, due to this second 'noreboot' pin.
If a guest OS had no watchdog device at all, then apps whose operation and/or data integrity relies on a watchdog can refuse to launch, and alert the administrator of the problematic deployment. With Q35 machines unconditionally exposing a watchdog though, apps will think their deployment is correct but in fact have no protection at all.
Quoted 5 lines
This patch flips the default of the second 'no reboot' flag, so that configured watchdog actions will be honoured out of the box for the 7.2 Q35 machine type onwards, if the guest enables use of the watchdog.
Signed-off-by Daniel P. Berrangé <[email protected]>
Add Fixes: or some other reference to the BZs? We have a few!
https://bugzilla.redhat.com/show_bug.cgi?id=2136889
https://bugzilla.redhat.com/show_bug.cgi?id=2080207
https://bugzilla.redhat.com/show_bug.cgi?id=2137346 (libvirt)
Quoted 38 lines
hw/i386/pc.c | 4 +++- hw/isa/lpc_ich9.c | 2 +- tests/qtest/tco-test.c | 2 +- 3 files changed, 5 insertions(+), 3 deletions(-)hw/i386/pc.c +3
-1@@ -108,7 +108,9 @@ { "qemu64-" TYPE_X86_CPU, "model-id", "QEMU Virtual CPU version " v, },\ { "athlon-" TYPE_X86_CPU, "model-id", "QEMU Virtual CPU version " v, }, -GlobalProperty pc_compat_7_1[] = {}; +GlobalProperty pc_compat_7_1[] = { + { "ICH9-LPC", "noreboot", "true" }, +};const size_t pc_compat_7_1_len =G_N_ELEMENTS (pc_compat_7_1);GlobalProperty pc_compat_7_0[] = {};hw/isa/lpc_ich9.c +1
-1@@ -789,7 +789,7 @@ static const VMStateDescription vmstate_ich9_lpc = { };static Property ich9_lpc_properties[] = { -DEFINE_PROP_BOOL ("noreboot" , ICH9LPCState, pin_strap.spkr_hi, true), +DEFINE_PROP_BOOL ("noreboot" , ICH9LPCState, pin_strap.spkr_hi, false),DEFINE_PROP_BOOL ("smm-compat" , ICH9LPCState, pm.smm_compat, false),DEFINE_PROP_BIT64 ("x-smi-broadcast" , ICH9LPCState, smi_host_features,ICH9_LPC_SMI_F_BROADCAST_BIT , true),tests/qtest/tco-test.c +0
-0
tests/qtest/tco-test.c static void test_init(TestData *d) · line 65
Quoted 7 lines
tests/qtest/tco-test.c +1
-1@@ -60,7 +60,7 @@ static void test_init(TestData *d)QTestState *qs; qs =qtest_initf ("-machine q35 %s %s" , - d->noreboot ?"" :"-global ICH9-LPC.noreboot=false" , + d->noreboot ?"-global ICH9-LPC.noreboot=true" :"" , !d->args ?"" : d->args);qtest_irq_intercept_in (qs,"ioapic" );
Rich.
Signature · 5 lines
Richard Jones, Virtualization Group, Red Hat http://people.redhat.com/~rjones Read my programming and virtualization blog: http://rwmj.wordpress.com Fedora Windows cross-compiler. Compile Windows programs, test, and build Windows installers. Over 100 libraries supported. http://fedoraproject.org/wiki/MinGW
Quoted 5 lines · Daniel P. Berrangé
The TCO watchdog is unconditionally integrated into the Q35 machine type by default, but at the same time is unconditionally disabled from firing by a host config option that overrides guest OS attempts to enable it. People have to know to set a magic -global to make it non-broken
Incidentally I found that originally the TCO watchdog was not unconditionally enabled. Its exposure to the guest could be turned on/off using
-global ICH9-LPC.enable_tco=bool
This was implemented for machine type compat, but it also gave apps a way to disable the watchdog functionality. Unfortunately that ability was discarded in this series:
https://lore.kernel.org/all/[email protected]/
but the 'enable_tco' property still exists in QOM, but silently ignored.
Seems we should either fix the impl of 'enable_tco', or remove the QOM property entirely, so we don't pretend it can be toggled anymore.
With regards,
Daniel
Signature · 3 lines
|: https://berrange.com -o- https://www.flickr.com/photos/dberrange :| |: https://libvirt.org -o- https://fstop138.berrange.com :| |: https://entangle-photo.org -o- https://www.instagram.com/dberrange :|
Quoted 19 lines · Daniel P. Berrangé
Quoted 5 lines · Daniel P. Berrangé
The TCO watchdog is unconditionally integrated into the Q35 machine type by default, but at the same time is unconditionally disabled from firing by a host config option that overrides guest OS attempts to enable it. People have to know to set a magic -global to make it non-broken
Incidentally I found that originally the TCO watchdog was not unconditionally enabled. Its exposure to the guest could be turned on/off using
-global ICH9-LPC.enable_tco=bool
This was implemented for machine type compat, but it also gave apps a way to disable the watchdog functionality. Unfortunately that ability was discarded in this series:
Quoted 8 lines
but the '
enable_tco' property still exists in QOM, but silently ignored.Seems we should either fix the impl of '
enable_tco', or remove the QOM property entirely, so we don't pretend it can be toggled anymore.With regards,
Daniel
i am inclined to say you are right and the fix is to fix the impl.
Quoted 4 lines
Signature · 3 lines
|: https://berrange.com -o- https://www.flickr.com/photos/dberrange :| |: https://libvirt.org -o- https://fstop138.berrange.com :| |: https://entangle-photo.org -o- https://www.instagram.com/dberrange :|
Quoted 23 lines · On Mon, 31 Oct 2022 11:48:58 -0400 "Michael S. Tsirkin" <[email protected]> wrote:
Quoted 22 lines · Daniel P. Berrangé
Quoted 5 lines · Daniel P. Berrangé
The TCO watchdog is unconditionally integrated into the Q35 machine type by default, but at the same time is unconditionally disabled from firing by a host config option that overrides guest OS attempts to enable it. People have to know to set a magic -global to make it non-broken
Incidentally I found that originally the TCO watchdog was not unconditionally enabled. Its exposure to the guest could be turned on/off using
-global ICH9-LPC.enable_tco=bool
This was implemented for machine type compat, but it also gave apps a way to disable the watchdog functionality. Unfortunately that ability was discarded in this series:
https://lore.kernel.org/all/[email protected]/
but the '
enable_tco' property still exists in QOM, but silently ignored.
Quoted 7 lines
Quoted 5 lines
Seems we should either fix the impl of '
enable_tco', or remove the QOM property entirely, so we don't pretend it can be toggled anymore.With regards,
Danieli am inclined to say you are right and the fix is to fix the impl.
Is there need for users to disable whatchdog at all? It was always present since then and no one complained, so perhaps we should ditch property instead fixing it to keep it simple.
Quoted 6 lines
Quoted 5 lines
Signature · 4 lines
-- |: https://berrange.com -o- https://www.flickr.com/photos/dberrange :| |: https://libvirt.org -o- https://fstop138.berrange.com :| |: https://entangle-photo.org -o- https://www.instagram.com/dberrange :|
Quoted 32 lines · Igor Mammedov
Quoted 29 lines · On Mon, 31 Oct 2022 11:48:58 -0400 "Michael S. Tsirkin" <[email protected]> wrote:
Quoted 27 lines · Daniel P. Berrangé
Quoted 5 lines · Daniel P. Berrangé
The TCO watchdog is unconditionally integrated into the Q35 machine type by default, but at the same time is unconditionally disabled from firing by a host config option that overrides guest OS attempts to enable it. People have to know to set a magic -global to make it non-broken
Incidentally I found that originally the TCO watchdog was not unconditionally enabled. Its exposure to the guest could be turned on/off using
-global ICH9-LPC.enable_tco=bool
This was implemented for machine type compat, but it also gave apps a way to disable the watchdog functionality. Unfortunately that ability was discarded in this series:
https://lore.kernel.org/all/[email protected]/
but the '
enable_tco' property still exists in QOM, but silently ignored.Seems we should either fix the impl of '
enable_tco', or remove the QOM property entirely, so we don't pretend it can be toggled anymore.With regards,
Daniel
Quoted 6 lines
Quoted 1 line
i am inclined to say you are right and the fix is to fix the impl.
Is there need for users to disable whatchdog at all? It was always present since then and no one complained, so perhaps we should ditch property instead fixing it to keep it simple.
Thinking about it more, I think we should NOT fix the 'enable_tco' property, because there will be no way for a mgmt appp to tell if they're using a fixed or broken QEMU. So if they use 'enable_tco' on a broken QEMU and then live migrate, they'll get an guest ABI change. If we did want to support disabling it, then we should have a brand new property that apps can probe for.
In the absence of a request to disable watchdog, I'd say we just delete 'enable_tco' right now. If someone wants it in future, we can add it with a new name.
With regards,
Daniel
Signature · 3 lines
|: https://berrange.com -o- https://www.flickr.com/photos/dberrange :| |: https://libvirt.org -o- https://fstop138.berrange.com :| |: https://entangle-photo.org -o- https://www.instagram.com/dberrange :|
Quoted 35 lines · Daniel P. Berrangé
Quoted 34 lines · Igor Mammedov
Quoted 30 lines · On Mon, 31 Oct 2022 11:48:58 -0400 "Michael S. Tsirkin" <[email protected]> wrote:
Quoted 19 lines · Daniel P. Berrangé
Quoted 5 lines · Daniel P. Berrangé
The TCO watchdog is unconditionally integrated into the Q35 machine type by default, but at the same time is unconditionally disabled from firing by a host config option that overrides guest OS attempts to enable it. People have to know to set a magic -global to make it non-broken
Incidentally I found that originally the TCO watchdog was not unconditionally enabled. Its exposure to the guest could be turned on/off using
-global ICH9-LPC.enable_tco=bool
This was implemented for machine type compat, but it also gave apps a way to disable the watchdog functionality. Unfortunately that ability was discarded in this series:
Quoted 8 lines
but the '
enable_tco' property still exists in QOM, but silently ignored.Seems we should either fix the impl of '
enable_tco', or remove the QOM property entirely, so we don't pretend it can be toggled anymore.With regards,
Danieli am inclined to say you are right and the fix is to fix the impl.
Quoted 8 lines
Quoted 4 lines
Is there need for users to disable whatchdog at all? It was always present since then and no one complained, so perhaps we should ditch property instead fixing it to keep it simple.
Thinking about it more, I think we should NOT fix the '
enable_tco' property, because there will be no way for a mgmt appp to tell if they're using a fixed or broken QEMU.
This is always the case for any bug. We don't as a rule add properties for this, it is distro's responsibility to pick bugfixes for features users care about.
What makes this bug different?
Quoted 11 lines
So if they use '
enable_tco' on a broken QEMU and then
live migrate, they'll get an guest ABI change. If we did want to support disabling it, then we should have a brand new property that apps can probe for.In the absence of a request to disable watchdog, I'd say we just delete '
enable_tco' right now. If someone wants it in future, we can add it with a new name.With regards,
Daniel
I am really stressed about watchdog firing and resetting VMs that previously were working fine. Adding this without a safety mechanism to quickly disable in case of problems in the field is not wise imho.
Quoted 4 lines
Signature · 3 lines
|: https://berrange.com -o- https://www.flickr.com/photos/dberrange :| |: https://libvirt.org -o- https://fstop138.berrange.com :| |: https://entangle-photo.org -o- https://www.instagram.com/dberrange :|
Quoted 45 lines · Michael S. Tsirkin
Quoted 35 lines · Daniel P. Berrangé
Quoted 34 lines · Igor Mammedov
Quoted 30 lines · On Mon, 31 Oct 2022 11:48:58 -0400 "Michael S. Tsirkin" <[email protected]> wrote:
Quoted 19 lines · Daniel P. Berrangé
Quoted 5 lines · Daniel P. Berrangé
The TCO watchdog is unconditionally integrated into the Q35 machine type by default, but at the same time is unconditionally disabled from firing by a host config option that overrides guest OS attempts to enable it. People have to know to set a magic -global to make it non-broken
Incidentally I found that originally the TCO watchdog was not unconditionally enabled. Its exposure to the guest could be turned on/off using
-global ICH9-LPC.enable_tco=bool
This was implemented for machine type compat, but it also gave apps a way to disable the watchdog functionality. Unfortunately that ability was discarded in this series:
Quoted 8 lines
but the '
enable_tco' property still exists in QOM, but silently ignored.Seems we should either fix the impl of '
enable_tco', or remove the QOM property entirely, so we don't pretend it can be toggled anymore.With regards,
Danieli am inclined to say you are right and the fix is to fix the impl.
Quoted 8 lines
Quoted 4 lines
Is there need for users to disable whatchdog at all? It was always present since then and no one complained, so perhaps we should ditch property instead fixing it to keep it simple.
Thinking about it more, I think we should NOT fix the '
enable_tco' property, because there will be no way for a mgmt appp to tell if they're using a fixed or broken QEMU.
Quoted 5 lines
This is always the case for any bug. We don't as a rule add properties for this, it is distro's responsibility to pick bugfixes for features users care about.
What makes this bug different?
It is impossible to safely use the enable_tco property because it will break ABI across live migrations between old and new QEMU. If it was a recent bug I'd be ok with just fixing it. Since this property has been broken for 6 years though, IMHO it just needs to be deleted entirely.
Quoted 5 lines
Quoted 5 lines
So if they use '
enable_tco' on a broken QEMU and then
live migrate, they'll get an guest ABI change. If we did want to support disabling it, then we should have a brand new property that apps can probe for.
Quoted 7 lines
Quoted 3 lines
In the absence of a request to disable watchdog, I'd say we just delete '
enable_tco' right now. If someone wants it in future, we can add it with a new name.I am really stressed about watchdog firing and resetting VMs that previously were working fine. Adding this without a safety mechanism to quickly disable in case of problems in the field is not wise imho.
We would only toggle the ICH9-LPC.noreboot flag for new machine types, so wouldn't affect existing VMs, and if anyone has a problem they can explicitly set ICH9-LPC.noreboot back to true, or pick the older machine type again.
With regards,
Daniel
Signature · 3 lines
|: https://berrange.com -o- https://www.flickr.com/photos/dberrange :| |: https://libvirt.org -o- https://fstop138.berrange.com :| |: https://entangle-photo.org -o- https://www.instagram.com/dberrange :|
Quoted 21 lines · Daniel P. Berrangé
The TCO watchdog is unconditionally integrated into the Q35 machine type by default, but at the same time is unconditionally disabled from firing by a host config option that overrides guest OS attempts to enable it. People have to know to set a magic -global to make it non-broken
IOW we're exposing a broken watchdog by default to all Q35 machines, but which to the guest OS & its apps looks fully functional :-(
This behaviour was set in response to feedback from Michael:
https://lists.gnu.org/archive/html/qemu-devel/2015-06/msg07128.html
"I think sample high is a safer default."
but as explained in the commit message in the last patch, I think the watchdog defaults were already safe without that pin strap setting. The guest OS needs to take explicit action to clear the guest visible 'no reboot' flag, and so we don't need a second guest hidden 'no reboot' flag to override that choice IMHO. Am I missing something ?
Quoted 4 lines
NB, I'm toggling this for 7.2 machine type since that's the current git latest machine. Since this has already been "broken" for 7 years though, I am ambivalent about whether we try todo this for 7.2, vs just wait until the 8.0 machine types arrive.
So I expect v2 with minor issues fixed and hopefully a fixed property (we can debate removing it down the road if we want).
Quoted 18 lines
Daniel P. Berrangé (4):
hw/acpi: add trace events for TCO watchdog register access hw/isa: add trace events for ICH9 LPC chip config access hw/watchdog: add trace events for watchdog action handling hw/isa: enable TCO watchdog reboot pin strap by defaulthw/acpi/tco.c | 41 +++++++++++++++++++++++++++------------- hw/acpi/trace-events | 2 ++ hw/i386/pc.c | 4 +++- hw/isa/lpc_ich9.c | 5 ++++- hw/isa/trace-events | 4 ++++ hw/watchdog/trace-events | 4 ++++ hw/watchdog/watchdog.c | 4 ++++ tests/qtest/tco-test.c | 2 +- 8 files changed, 50 insertions(+), 16 deletions(-)Signature · 1 line
2.37.3