x86/apic: Fix restoring boot irq mode in reboot and kexec/kdump
Subsystems (3)
v3 cover x86/apic: Fix restoring boot irq mode in reboot and kexec/kdump
A regression bug was introduced in below commit.
commit 522e66464467 ("x86/apic: Disable I/O APIC before shutdown of the local APIC")
It caused the action to fail which we try to restore boot irq mode in reboot and kexec/kdump. Details can be seen in patch 0002.
Warning can always be seen during kdump kernel boot on qemu/kvm platform. Our customer even saw casual kdump kernel hang once in ~30 attempts during stress testing of kdump on KVM machine.
This is v3 post, patches are rearranged and changed according to Eric's suggestions.
Baoquan He (5):
x86/apic: Split out restore_boot_irq_mode from disable_IO_APIC x86/apic: Fix restoring boot irq mode in reboot and kexec/kdump x86/apic: Remove useless disable_IO_APIC
x86/apic: Rename variable/function related to x86_io_apic_ops x86/apic: Set up through-local-APIC on boot CPU if 'noapic' specified
arch/x86/include/asm/io_apic.h | 9 +++++---- arch/x86/include/asm/x86_init.h | 8 ++++---- arch/x86/kernel/apic/apic.c | 2 +- arch/x86/kernel/apic/io_apic.c | 16 ++++------------ arch/x86/kernel/crash.c | 3 ++- arch/x86/kernel/machine_kexec_32.c | 7 +++---- arch/x86/kernel/machine_kexec_64.c | 7 +++---- arch/x86/kernel/reboot.c | 3 ++- arch/x86/kernel/x86_init.c | 6 +++--- arch/x86/xen/apic.c | 2 +- drivers/iommu/irq_remapping.c | 4 ++-- 11 files changed, 30 insertions(+), 37 deletions(-)
Signature · 1 line
2.13.6
PATCH v3 1/5 x86/apic: Split out restore_boot_irq_mode from disable_IO_APIC
This is a preparation patch. Split out the code which restores boot irq mode from disable_IO_APIC() and wrap into a new function restore_boot_irq_mode(). No functional change.
--- arch/x86/include/asm/io_apic.h | 1 + arch/x86/kernel/apic/io_apic.c | 5 +++++ 2 files changed, 6 insertions(+)
arch/x86/include/asm/io_apic.h +1 -0
@@ -193,6 +193,7 @@ static inline unsigned int io_apic_read(unsigned int apic, unsigned int reg)extern void setup_IO_APIC (void );extern void enable_IO_APIC (void );extern void disable_IO_APIC (void ); +extern void restore_boot_irq_mode (void );extern int IO_APIC_get_PCI_irq_vector (int bus,int devfn,int pin);extern void print_IO_APICs (void );#else /* !CONFIG_X86_IO_APIC */
arch/x86/kernel/apic/io_apic.c +5 -0
@@ -1448,6 +1448,11 @@ void disable_IO_APIC(void)*/ clear_IO_APIC (); +restore_boot_irq_mode (); +} + +void restore_boot_irq_mode (void ) +{if (!nr_legacy_irqs ())return ;
Signature · 1 line
2.13.6
PATCH v3 2/5 x86/apic: Fix restoring boot irq mode in reboot and kexec/kdump
In reply to Baoquan HeThis is a regression fix.
Before, to fix erratum AVR31, commit 522e66464467 ("x86/apic: Disable I/O APIC before shutdown of the local APIC") moved lapic_shutdown() calling after disable_IO_APIC(). This introdued a regression. The root cause is that disable_IO_APIC() not only clears IO_APIC, also restore boot irq mode by setting LAPIC/APIC/IMCR, lapic_shutdown() after disable_IO_APIC() will disable LAPIC and ruin the possible virtual wire mode setting which the code has been trying to do all along.
The consequence is, in KVM guest kernel always prints warning as below during kexec/kdump kernel boots up. That happened in setup_local_APIC() since 'do { xxx } while (queued && max_loops > 0)' loop does not function well any more if pending irq exists in APIC IRR after LAPIC is disabled. And people even saw casual kdump kernel hang once in ~30 attempts during stress testing of kdump on KVM machine.
[ 0.001000] WARNING: CPU: 0 PID: 0 at arch/x86/kernel/apic/apic.c:1467 setup_local_APIC+0x228/0x330 [ 0.001000] Modules linked in: [ 0.001000] CPU: 0 PID: 0 Comm: swapper/0 Not tainted 4.15.0-rc5+ #3 [ 0.001000] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.10.2-1.fc26 04/01/2014 [ 0.001000] RIP: 0010:setup_local_APIC+0x228/0x330 [ 0.001000] RSP: 0000:ffffffffb6e03eb8 EFLAGS: 00010286 [ 0.001000] RAX: 0000009edb4c4d84 RBX: 0000000000000000 RCX: 00000000b099d800 [ 0.001000] RDX: 0000009e00000000 RSI: 0000000000000000 RDI: 0000000000000810 [ 0.001000] RBP: 0000000000000000 R08: ffffffffffffffff R09: 0000000000000001 [ 0.001000] R10: ffff98ce6a801c00 R11: 0761076d072f0776 R12: 0000000000000001 [ 0.001000] R13: 00000000000000f0 R14: 0000000000004000 R15: ffffffffffffc6ff [ 0.001000] FS: 0000000000000000(0000) GS:ffff98ce6bc00000(0000) knlGS:0000000000000000 [ 0.001000] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 0.001000] CR2: 00000000ffffffff CR3: 0000000022209000 CR4: 00000000000406b0 [ 0.001000] Call Trace: [ 0.001000] apic_bsp_setup+0x56/0x74 [ 0.001000] x86_late_time_init+0x11/0x16 [ 0.001000] start_kernel+0x3c9/0x486 [ 0.001000] secondary_startup_64+0xa5/0xb0 [ 0.001000] Code: 00 85 c9 74 2d 0f 31 c1 e1 0a 48 c1 e2 20 41 89 cf 4c 03 7c 24 08 48 09 d0 49 29 c7 4c 89 3c 24 48 83 3c 24 00 0f 8f 8f fe ff ff <0f> ff e9 10 ff ff ff 48 83 2c 24 01 eb e7 48 83 c4 18 5b 5d 41 [ 0.001000] ---[ end trace b88e71b9a6ebebdd ]--- [ 0.001000] masked ExtINT on CPU#0
To fix this, just break down disable_IO_APIC(), then call clear_IO_APIC() to stop IO_APIC where disable_IO_APIC() was called, and call restore_boot_irq_mode() to restore boot irq mode before reboot or kexec/kdump jump.
--- arch/x86/include/asm/io_apic.h | 1 + arch/x86/kernel/apic/io_apic.c | 2 +- arch/x86/kernel/crash.c | 3 ++- arch/x86/kernel/machine_kexec_32.c | 2 +- arch/x86/kernel/machine_kexec_64.c | 2 +- arch/x86/kernel/reboot.c | 3 ++- 6 files changed, 8 insertions(+), 5 deletions(-)
arch/x86/include/asm/io_apic.h +1 -0
@@ -193,6 +193,7 @@ static inline unsigned int io_apic_read(unsigned int apic, unsigned int reg)extern void setup_IO_APIC (void );extern void enable_IO_APIC (void );extern void disable_IO_APIC (void ); +extern void clear_IO_APIC (void );extern void restore_boot_irq_mode (void );extern int IO_APIC_get_PCI_irq_vector (int bus,int devfn,int pin);extern void print_IO_APICs (void );
arch/x86/kernel/apic/io_apic.c +1 -1
@@ -587,7 +587,7 @@ static void clear_IO_APIC_pin(unsigned int apic, unsigned int pin) mpc_ioapic_id(apic), pin); } -static void clear_IO_APIC (void ) +void clear_IO_APIC (void ) {int apic, pin;
arch/x86/kernel/crash.c +2 -1
@@ -199,9 +199,10 @@ void native_machine_crash_shutdown(struct pt_regs *regs)#ifdef CONFIG_X86_IO_APIC /* Prevent crash_kexec() from deadlocking on ioapic_lock. */ ioapic_zap_locks (); -disable_IO_APIC (); +clear_IO_APIC ();#endif lapic_shutdown (); +restore_boot_irq_mode ();#ifdef CONFIG_HPET_TIMER hpet_disable ();#endif
arch/x86/kernel/machine_kexec_32.c +1 -1
@@ -199,7 +199,7 @@ void machine_kexec(struct kimage *image)* one form or other. kexec jump path also need * one. */ -disable_IO_APIC (); +restore_boot_irq_mode ();#endif }
arch/x86/kernel/machine_kexec_64.c +1 -1
@@ -297,7 +297,7 @@ void machine_kexec(struct kimage *image)* one form or other. kexec jump path also need * one. */ -disable_IO_APIC (); +restore_boot_irq_mode ();#endif }
arch/x86/kernel/reboot.c +2 -1
@@ -666,7 +666,7 @@ void native_machine_shutdown(void)* Even without the erratum, it still makes sense to quiet IO APIC * before disabling Local APIC. */ -disable_IO_APIC (); +clear_IO_APIC ();#endif #ifdef CONFIG_SMP
@@ -680,6 +680,7 @@ void native_machine_shutdown(void)#endif lapic_shutdown (); +restore_boot_irq_mode ();#ifdef CONFIG_HPET_TIMER hpet_disable ();
Signature · 1 line
2.13.6
Quoted 42 lines · Baoquan He
This is a regression fix.
Before, to fix erratum AVR31, commit
522e66464467("x86/apic: Disable I/O APIC before shutdown of the local APIC") movedlapic_shutdown()calling afterdisable_IO_APIC(). This introdued a regression. The root cause is thatdisable_IO_APIC()not only clearsIO_APIC, also restore boot irq mode by setting LAPIC/APIC/IMCR,lapic_shutdown()afterdisable_IO_APIC()will disable LAPIC and ruin the possible virtual wire mode setting which the code has been trying to do all along.The consequence is, in KVM guest kernel always prints warning as below during kexec/kdump kernel boots up. That happened in
setup_local_APIC()since 'do { xxx } while (queued &&max_loops> 0)' loop does not function well any more if pending irq exists in APIC IRR after LAPIC is disabled. And people even saw casual kdump kernel hang once in ~30 attempts during stress testing of kdump on KVM machine.[ 0.001000] WARNING: CPU: 0 PID: 0 at arch/x86/kernel/apic/apic.c:1467 setup_local_APIC+0x228/0x330 [ 0.001000] Modules linked in: [ 0.001000] CPU: 0 PID: 0 Comm: swapper/0 Not tainted 4.15.0-rc5+ #3 [ 0.001000] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.10.2-1.fc26 04/01/2014 [ 0.001000] RIP: 0010:setup_local_APIC+0x228/0x330 [ 0.001000] RSP: 0000:ffffffffb6e03eb8 EFLAGS: 00010286 [ 0.001000] RAX: 0000009edb4c4d84 RBX: 0000000000000000 RCX: 00000000b099d800 [ 0.001000] RDX: 0000009e00000000 RSI: 0000000000000000 RDI: 0000000000000810 [ 0.001000] RBP: 0000000000000000 R08: ffffffffffffffff R09: 0000000000000001 [ 0.001000] R10: ffff98ce6a801c00 R11: 0761076d072f0776 R12: 0000000000000001 [ 0.001000] R13: 00000000000000f0 R14: 0000000000004000 R15: ffffffffffffc6ff [ 0.001000] FS: 0000000000000000(0000) GS:ffff98ce6bc00000(0000) knlGS:0000000000000000 [ 0.001000] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 0.001000] CR2: 00000000ffffffff CR3: 0000000022209000 CR4: 00000000000406b0 [ 0.001000] Call Trace: [ 0.001000] apic_bsp_setup+0x56/0x74 [ 0.001000] x86_late_time_init+0x11/0x16 [ 0.001000] start_kernel+0x3c9/0x486 [ 0.001000] secondary_startup_64+0xa5/0xb0 [ 0.001000] Code: 00 85 c9 74 2d 0f 31 c1 e1 0a 48 c1 e2 20 41 89 cf 4c 03 7c 24 08 48 09 d0 49 29 c7 4c 89 3c 24 48 83 3c 24 00 0f 8f 8f fe ff ff <0f> ff e9 10 ff ff ff 48 83 2c 24 01 eb e7 48 83 c4 18 5b 5d 41 [ 0.001000] ---[ end trace b88e71b9a6ebebdd ]--- [ 0.001000] masked ExtINT on CPU#0
Quoted 4 lines
To fix this, just break down
disable_IO_APIC(), then callclear_IO_APIC()to stopIO_APICwheredisable_IO_APIC()was called, and callrestore_boot_irq_mode()to restore boot irq mode before reboot or kexec/kdump jump.
Two things here.
a) This is missing a fixes tag and a CC stable.
b) What makes your change to the KEXEC_JUMP code path safe? Have the lapic and ioapic already been shut down?
The KEXEC_JUMP changes to machine_kexec_32.c and machine_kexec_64.c either need to be documented in the change long why they are safe so that this change becomes obviously safe and correct.
Otherwise we risk and trivial and obvious looking change causing another regression like changing the order of lapic_shutdown and disable_IOAPIC did.
Eric
Quoted 99 lines
Signed-off-by Baoquan He <[email protected]>--- arch/x86/include/asm/io_apic.h | 1 + arch/x86/kernel/apic/io_apic.c | 2 +- arch/x86/kernel/crash.c | 3 ++- arch/x86/kernel/machine_kexec_32.c | 2 +- arch/x86/kernel/machine_kexec_64.c | 2 +- arch/x86/kernel/reboot.c | 3 ++- 6 files changed, 8 insertions(+), 5 deletions(-)arch/x86/include/asm/io_apic.h +1
-0@@ -193,6 +193,7 @@ static inline unsigned int io_apic_read(unsigned int apic, unsigned int reg)extern void setup_IO_APIC (void );extern void enable_IO_APIC (void );extern void disable_IO_APIC (void ); +extern void clear_IO_APIC (void );extern void restore_boot_irq_mode (void );extern int IO_APIC_get_PCI_irq_vector (int bus,int devfn,int pin);extern void print_IO_APICs (void );arch/x86/kernel/apic/io_apic.c +1
-1@@ -587,7 +587,7 @@ static void clear_IO_APIC_pin(unsigned int apic, unsigned int pin) mpc_ioapic_id(apic), pin); } -static void clear_IO_APIC (void ) +void clear_IO_APIC (void ) {int apic, pin;arch/x86/kernel/crash.c +2
-1@@ -199,9 +199,10 @@ void native_machine_crash_shutdown(struct pt_regs *regs)#ifdef CONFIG_X86_IO_APIC /* Prevent crash_kexec() from deadlocking on ioapic_lock. */ ioapic_zap_locks (); -disable_IO_APIC (); +clear_IO_APIC ();#endif lapic_shutdown (); +restore_boot_irq_mode ();#ifdef CONFIG_HPET_TIMER hpet_disable ();#endif arch/x86/kernel/machine_kexec_32.c +1
-1@@ -199,7 +199,7 @@ void machine_kexec(struct kimage *image)* one form or other. kexec jump path also need * one. */ -disable_IO_APIC (); +restore_boot_irq_mode ();#endif }arch/x86/kernel/machine_kexec_64.c +1
-1@@ -297,7 +297,7 @@ void machine_kexec(struct kimage *image)* one form or other. kexec jump path also need * one. */ -disable_IO_APIC (); +restore_boot_irq_mode ();#endif }arch/x86/kernel/reboot.c +2
-1@@ -666,7 +666,7 @@ void native_machine_shutdown(void)* Even without the erratum, it still makes sense to quiet IO APIC * before disabling Local APIC. */ -disable_IO_APIC (); +clear_IO_APIC ();#endif #ifdef CONFIG_SMP @@ -680,6 +680,7 @@ void native_machine_shutdown(void)#endif lapic_shutdown (); +restore_boot_irq_mode ();#ifdef CONFIG_HPET_TIMER hpet_disable ();
Quoted 54 lines · Eric W. Biederman
Quoted 42 lines · Baoquan He
This is a regression fix.
Before, to fix erratum AVR31, commit
522e66464467("x86/apic: Disable I/O APIC before shutdown of the local APIC") movedlapic_shutdown()calling afterdisable_IO_APIC(). This introdued a regression. The root cause is thatdisable_IO_APIC()not only clearsIO_APIC, also restore boot irq mode by setting LAPIC/APIC/IMCR,lapic_shutdown()afterdisable_IO_APIC()will disable LAPIC and ruin the possible virtual wire mode setting which the code has been trying to do all along.The consequence is, in KVM guest kernel always prints warning as below during kexec/kdump kernel boots up. That happened in
setup_local_APIC()since 'do { xxx } while (queued &&max_loops> 0)' loop does not function well any more if pending irq exists in APIC IRR after LAPIC is disabled. And people even saw casual kdump kernel hang once in ~30 attempts during stress testing of kdump on KVM machine.[ 0.001000] WARNING: CPU: 0 PID: 0 at arch/x86/kernel/apic/apic.c:1467 setup_local_APIC+0x228/0x330 [ 0.001000] Modules linked in: [ 0.001000] CPU: 0 PID: 0 Comm: swapper/0 Not tainted 4.15.0-rc5+ #3 [ 0.001000] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.10.2-1.fc26 04/01/2014 [ 0.001000] RIP: 0010:setup_local_APIC+0x228/0x330 [ 0.001000] RSP: 0000:ffffffffb6e03eb8 EFLAGS: 00010286 [ 0.001000] RAX: 0000009edb4c4d84 RBX: 0000000000000000 RCX: 00000000b099d800 [ 0.001000] RDX: 0000009e00000000 RSI: 0000000000000000 RDI: 0000000000000810 [ 0.001000] RBP: 0000000000000000 R08: ffffffffffffffff R09: 0000000000000001 [ 0.001000] R10: ffff98ce6a801c00 R11: 0761076d072f0776 R12: 0000000000000001 [ 0.001000] R13: 00000000000000f0 R14: 0000000000004000 R15: ffffffffffffc6ff [ 0.001000] FS: 0000000000000000(0000) GS:ffff98ce6bc00000(0000) knlGS:0000000000000000 [ 0.001000] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 0.001000] CR2: 00000000ffffffff CR3: 0000000022209000 CR4: 00000000000406b0 [ 0.001000] Call Trace: [ 0.001000] apic_bsp_setup+0x56/0x74 [ 0.001000] x86_late_time_init+0x11/0x16 [ 0.001000] start_kernel+0x3c9/0x486 [ 0.001000] secondary_startup_64+0xa5/0xb0 [ 0.001000] Code: 00 85 c9 74 2d 0f 31 c1 e1 0a 48 c1 e2 20 41 89 cf 4c 03 7c 24 08 48 09 d0 49 29 c7 4c 89 3c 24 48 83 3c 24 00 0f 8f 8f fe ff ff <0f> ff e9 10 ff ff ff 48 83 2c 24 01 eb e7 48 83 c4 18 5b 5d 41 [ 0.001000] ---[ end trace b88e71b9a6ebebdd ]--- [ 0.001000] masked ExtINT on CPU#0Quoted 4 lines
To fix this, just break down
disable_IO_APIC(), then callclear_IO_APIC()to stopIO_APICwheredisable_IO_APIC()was called, and callrestore_boot_irq_mode()to restore boot irq mode before reboot or kexec/kdump jump.Two things here.
a) This is missing a fixes tag and a CC stable.
b) What makes your change to theKEXEC_JUMPcode path safe? Have the lapic and ioapic already been shut down?
Quoted 7 lines
The
KEXEC_JUMPchanges tomachine_kexec_32.candmachine_kexec_64.ceither need to be documented in the change long why they are safe so that this change becomes obviously safe and correct.Otherwise we risk and trivial and obvious looking change causing another regression like changing the order of
lapic_shutdownanddisable_IOAPICdid.
Makes sense. Will change as you suggested and repost. Thanks a lot.
Thanks
Baoquan
Quoted 97 lines
Quoted 97 lines
--- arch/x86/include/asm/io_apic.h | 1 + arch/x86/kernel/apic/io_apic.c | 2 +- arch/x86/kernel/crash.c | 3 ++- arch/x86/kernel/machine_kexec_32.c | 2 +- arch/x86/kernel/machine_kexec_64.c | 2 +- arch/x86/kernel/reboot.c | 3 ++- 6 files changed, 8 insertions(+), 5 deletions(-)arch/x86/include/asm/io_apic.h +1
-0@@ -193,6 +193,7 @@ static inline unsigned int io_apic_read(unsigned int apic, unsigned int reg)extern void setup_IO_APIC (void );extern void enable_IO_APIC (void );extern void disable_IO_APIC (void ); +extern void clear_IO_APIC (void );extern void restore_boot_irq_mode (void );extern int IO_APIC_get_PCI_irq_vector (int bus,int devfn,int pin);extern void print_IO_APICs (void );arch/x86/kernel/apic/io_apic.c +1
-1@@ -587,7 +587,7 @@ static void clear_IO_APIC_pin(unsigned int apic, unsigned int pin) mpc_ioapic_id(apic), pin); } -static void clear_IO_APIC (void ) +void clear_IO_APIC (void ) {int apic, pin;arch/x86/kernel/crash.c +2
-1@@ -199,9 +199,10 @@ void native_machine_crash_shutdown(struct pt_regs *regs)#ifdef CONFIG_X86_IO_APIC /* Prevent crash_kexec() from deadlocking on ioapic_lock. */ ioapic_zap_locks (); -disable_IO_APIC (); +clear_IO_APIC ();#endif lapic_shutdown (); +restore_boot_irq_mode ();#ifdef CONFIG_HPET_TIMER hpet_disable ();#endif arch/x86/kernel/machine_kexec_32.c +1
-1@@ -199,7 +199,7 @@ void machine_kexec(struct kimage *image)* one form or other. kexec jump path also need * one. */ -disable_IO_APIC (); +restore_boot_irq_mode ();#endif }arch/x86/kernel/machine_kexec_64.c +1
-1@@ -297,7 +297,7 @@ void machine_kexec(struct kimage *image)* one form or other. kexec jump path also need * one. */ -disable_IO_APIC (); +restore_boot_irq_mode ();#endif }arch/x86/kernel/reboot.c +2
-1@@ -666,7 +666,7 @@ void native_machine_shutdown(void)* Even without the erratum, it still makes sense to quiet IO APIC * before disabling Local APIC. */ -disable_IO_APIC (); +clear_IO_APIC ();#endif #ifdef CONFIG_SMP @@ -680,6 +680,7 @@ void native_machine_shutdown(void)#endif lapic_shutdown (); +restore_boot_irq_mode ();#ifdef CONFIG_HPET_TIMER hpet_disable ();
Hi Baoquan,
Quoted 13 lines · Eric W. Biederman
Quoted 11 lines · Baoquan He
This is a regression fix.
Before, to fix erratum AVR31, commit
522e66464467("x86/apic: Disable I/O APIC before shutdown of the local APIC") movedlapic_shutdown()calling afterdisable_IO_APIC(). This introdued a regression. The root cause is thatdisable_IO_APIC()not only clearsIO_APIC, also restore boot irq mode by setting LAPIC/APIC/IMCR,lapic_shutdown()afterdisable_IO_APIC()will disable LAPIC and ruin the possible virtual wire mode setting which the code has been trying to do all along.
Quoted 3 lines
Quoted 3 lines
The consequence is, in KVM guest kernel always prints warning as below during kexec/kdump kernel boots up. That happened in
setup_local_APIC()since 'do { xxx } while (queued &&max_loops> 0)' loop does not function
I am not sure another thing here
AFAIK, according to the order of SMP machine shutdown, other CPUs will be stopped firstly, then the last CPU disable its local apic.
--machine_shutdown
|----...... |----stop_other_cpus() |----local_shutdown()
So, the pending interrupts exist only in BSP and only be ACKed by BSP. is it right?(I validated this by print the value of APIC_IRR/ISR of all CPUs, found only BSP had the non-zero value).
If it is right, We will do not need check the pending interrupt for each cpus.
BTW, the pending interrupt check code is mixed with the local APIC setup code, that looks messy. How about extracting the code which related to the crash interrupt check, put it into a new function and only invoked it when the CPU is BSP?
arch/x86/kernel/apic/apic.c +52 -8
@@ -1408,6 +1408,57 @@ static void lapic_setup_esr(void) oldvalue, value); } + +static void clear_crash_pending_intr (void ) +{ +long long max_loops = cpu_khz ? cpu_khz :1000000 ; +unsigned long long tsc =0 , ntsc; +unsigned int value, queued; +int i, j, acked =0 ; + +if (boot_cpu_has (X86_FEATURE_TSC )) + tsc =rdtsc (); +/* +* After a crash, we no longer service the interrupts and a pending +* interrupt from previous kernel might still have ISR bit set. +* +* Most probably by now CPU has serviced that pending interrupt and +* it might not have done the ack_APIC_irq() because it thought, +* interrupt came from i8259 as ExtInt. LAPIC did not get EOI so it +* does not clear the ISR bit and cpu thinks it has already serivced +* the interrupt. Hence a vector might get locked. It was noticed +* for timer irq (vector 0x31). Issue an extra EOI to clear ISR. +*/ +do { + queued =0 ; +for (i =APIC_ISR_NR -1 ; i >=0 ; i--) + queued |=apic_read (APIC_IRR + i*0x10 ); + +for (i =APIC_ISR_NR -1 ; i >=0 ; i--) { + value =apic_read (APIC_ISR + i*0x10 ); +for (j =31 ; j >=0 ; j--) { +if (value & (1 <<j)) { +ack_APIC_irq (); + acked++; + } + } + } +if (acked >256 ) { +printk (KERN_ERR "LAPIC pending interrupts after %d EOI\n" , + acked); +break ; + } +if (queued) { +if (boot_cpu_has (X86_FEATURE_TSC ) && cpu_khz) { + ntsc =rdtsc (); + max_loops = (cpu_khz <<10 ) - (ntsc - tsc); + }else + max_loops--; + } + }while (queued && max_loops >0 ); +WARN_ON (max_loops <=0 ); +} +/** * setup_local_APIC - setup the local APIC *
@@ -1417,13 +1468,7 @@ static void lapic_setup_esr(void)void setup_local_APIC (void ) {int cpu =smp_processor_id (); -unsigned int value, queued; -int i, j, acked =0 ; -unsigned long long tsc =0 , ntsc; -long long max_loops = cpu_khz ? cpu_khz :1000000 ; - -if (boot_cpu_has (X86_FEATURE_TSC )) - tsc =rdtsc (); - tsc =rdtsc (); +unsigned int value;if (disable_apic) {
disable_ioapic_support();
@@ -1475,45 +1520,8 @@ void setup_local_APIC(void)
value &= ~APIC_TPRI_MASK;
apic_write(APIC_TASKPRI, value);- /*
- * After a crash, we no longer service the interrupts and a pending - * interrupt from previous kernel might still have ISR bit set. - * - * Most probably by now CPU has serviced that pending interrupt and - * it might not have done the ack_APIC_irq() because it thought, - * interrupt came from i8259 as ExtInt. LAPIC did not get EOI so it - * does not clear the ISR bit and cpu thinks it has already serivced - * the interrupt. Hence a vector might get locked. It was noticed - * for timer irq (vector 0x31). Issue an extra EOI to clear ISR. - */
- do {
- queued = 0;
- for (i = APIC_ISR_NR - 1; i >= 0; i--)
- queued |= apic_read(APIC_IRR + i*0x10);
-
- for (i = APIC_ISR_NR - 1; i >= 0; i--) {
- value = apic_read(APIC_ISR + i*0x10);
- for (j = 31; j >= 0; j--) {
- if (value & (1<<j)) {
- ack_APIC_irq();
- acked++;
- }
- }
- }
- if (acked > 256) {
- printk(KERN_ERR "LAPIC pending interrupts after %d EOI\n",
- acked);
- break;
- }
- if (queued) {
- if (boot_cpu_has(X86_FEATURE_TSC) && cpu_khz) {
- ntsc = rdtsc();
- max_loops = (cpu_khz << 10) - (ntsc - tsc);
- } else
- max_loops--;
- }
- } while (queued && max_loops > 0); - WARN_ON(max_loops <= 0); + if (!cpu)
+ clear_crash_pending_intr();
/*
* Now that we are all set up, enable the APIC
Thanks,
dou.
Quoted 151 lines
Quoted 28 lines
well any more if pending irq exists in APIC IRR after LAPIC is disabled. And people even saw casual kdump kernel hang once in ~30 attempts during stress testing of kdump on KVM machine.
[ 0.001000] WARNING: CPU: 0 PID: 0 at arch/x86/kernel/apic/apic.c:1467 setup_local_APIC+0x228/0x330 [ 0.001000] Modules linked in: [ 0.001000] CPU: 0 PID: 0 Comm: swapper/0 Not tainted 4.15.0-rc5+ #3 [ 0.001000] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.10.2-1.fc26 04/01/2014 [ 0.001000] RIP: 0010:setup_local_APIC+0x228/0x330 [ 0.001000] RSP: 0000:ffffffffb6e03eb8 EFLAGS: 00010286 [ 0.001000] RAX: 0000009edb4c4d84 RBX: 0000000000000000 RCX: 00000000b099d800 [ 0.001000] RDX: 0000009e00000000 RSI: 0000000000000000 RDI: 0000000000000810 [ 0.001000] RBP: 0000000000000000 R08: ffffffffffffffff R09: 0000000000000001 [ 0.001000] R10: ffff98ce6a801c00 R11: 0761076d072f0776 R12: 0000000000000001 [ 0.001000] R13: 00000000000000f0 R14: 0000000000004000 R15: ffffffffffffc6ff [ 0.001000] FS: 0000000000000000(0000) GS:ffff98ce6bc00000(0000) knlGS:0000000000000000 [ 0.001000] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 0.001000] CR2: 00000000ffffffff CR3: 0000000022209000 CR4: 00000000000406b0 [ 0.001000] Call Trace: [ 0.001000] apic_bsp_setup+0x56/0x74 [ 0.001000] x86_late_time_init+0x11/0x16 [ 0.001000] start_kernel+0x3c9/0x486 [ 0.001000] secondary_startup_64+0xa5/0xb0 [ 0.001000] Code: 00 85 c9 74 2d 0f 31 c1 e1 0a 48 c1 e2 20 41 89 cf 4c 03 7c 24 08 48 09 d0 49 29 c7 4c 89 3c 24 48 83 3c 24 00 0f 8f 8f fe ff ff <0f> ff e9 10 ff ff ff 48 83 2c 24 01 eb e7 48 83 c4 18 5b 5d 41 [ 0.001000] ---[ end trace b88e71b9a6ebebdd ]--- [ 0.001000] masked ExtINT on CPU#0Quoted 4 lines
To fix this, just break down
disable_IO_APIC(), then callclear_IO_APIC()to stopIO_APICwheredisable_IO_APIC()was called, and callrestore_boot_irq_mode()to restore boot irq mode before reboot or kexec/kdump jump.Two things here.
a) This is missing a fixes tag and a CC stable.
b) What makes your change to theKEXEC_JUMPcode path safe? Have the lapic and ioapic already been shut down?The
KEXEC_JUMPchanges tomachine_kexec_32.candmachine_kexec_64.ceither need to be documented in the change long why they are safe so that this change becomes obviously safe and correct.Otherwise we risk and trivial and obvious looking change causing another regression like changing the order of
lapic_shutdownanddisable_IOAPICdid.Eric
Quoted 99 lines
Signed-off-by Baoquan He <[email protected]>--- arch/x86/include/asm/io_apic.h | 1 + arch/x86/kernel/apic/io_apic.c | 2 +- arch/x86/kernel/crash.c | 3 ++- arch/x86/kernel/machine_kexec_32.c | 2 +- arch/x86/kernel/machine_kexec_64.c | 2 +- arch/x86/kernel/reboot.c | 3 ++- 6 files changed, 8 insertions(+), 5 deletions(-)arch/x86/include/asm/io_apic.h +1
-0@@ -193,6 +193,7 @@ static inline unsigned int io_apic_read(unsigned int apic, unsigned int reg)extern void setup_IO_APIC (void );extern void enable_IO_APIC (void );extern void disable_IO_APIC (void ); +extern void clear_IO_APIC (void );extern void restore_boot_irq_mode (void );extern int IO_APIC_get_PCI_irq_vector (int bus,int devfn,int pin);extern void print_IO_APICs (void );arch/x86/kernel/apic/io_apic.c +1
-1@@ -587,7 +587,7 @@ static void clear_IO_APIC_pin(unsigned int apic, unsigned int pin) mpc_ioapic_id(apic), pin); } -static void clear_IO_APIC (void ) +void clear_IO_APIC (void ) {int apic, pin;arch/x86/kernel/crash.c +2
-1@@ -199,9 +199,10 @@ void native_machine_crash_shutdown(struct pt_regs *regs)#ifdef CONFIG_X86_IO_APIC /* Prevent crash_kexec() from deadlocking on ioapic_lock. */ ioapic_zap_locks (); -disable_IO_APIC (); +clear_IO_APIC ();#endif lapic_shutdown (); +restore_boot_irq_mode ();#ifdef CONFIG_HPET_TIMER hpet_disable ();#endif arch/x86/kernel/machine_kexec_32.c +1
-1@@ -199,7 +199,7 @@ void machine_kexec(struct kimage *image)* one form or other. kexec jump path also need * one. */ -disable_IO_APIC (); +restore_boot_irq_mode ();#endif }arch/x86/kernel/machine_kexec_64.c +1
-1@@ -297,7 +297,7 @@ void machine_kexec(struct kimage *image)* one form or other. kexec jump path also need * one. */ -disable_IO_APIC (); +restore_boot_irq_mode ();#endif }arch/x86/kernel/reboot.c +2
-1@@ -666,7 +666,7 @@ void native_machine_shutdown(void)* Even without the erratum, it still makes sense to quiet IO APIC * before disabling Local APIC. */ -disable_IO_APIC (); +clear_IO_APIC ();#endif #ifdef CONFIG_SMP @@ -680,6 +680,7 @@ void native_machine_shutdown(void)#endif lapic_shutdown (); +restore_boot_irq_mode ();#ifdef CONFIG_HPET_TIMER hpet_disable ();
Quoted 34 lines · Dou Liyang
Hi Baoquan,
Quoted 13 lines · Eric W. Biederman
Quoted 11 lines · Baoquan He
This is a regression fix.
Before, to fix erratum AVR31, commit
522e66464467("x86/apic: Disable I/O APIC before shutdown of the local APIC") movedlapic_shutdown()calling afterdisable_IO_APIC(). This introdued a regression. The root cause is thatdisable_IO_APIC()not only clearsIO_APIC, also restore boot irq mode by setting LAPIC/APIC/IMCR,lapic_shutdown()afterdisable_IO_APIC()will disable LAPIC and ruin the possible virtual wire mode setting which the code has been trying to do all along.Quoted 3 lines
Quoted 3 lines
The consequence is, in KVM guest kernel always prints warning as below during kexec/kdump kernel boots up. That happened in
setup_local_APIC()since 'do { xxx } while (queued &&max_loops> 0)' loop does not functionI am not sure another thing here
AFAIK, according to the order of SMP machine shutdown, other CPUs will be stopped firstly, then the last CPU disable its local apic.
--machine_shutdown
|----...... |----stop_other_cpus() |----local_shutdown()So, the pending interrupts exist only in BSP and only be ACKed by BSP. is it right?(I validated this by print the value of APIC_IRR/ISR of all CPUs, found only BSP had the non-zero value).
Quoted 7 lines
If it is right, We will do not need check the pending interrupt for each cpus.
BTW, the pending interrupt check code is mixed with the local APIC setup code, that looks messy. How about extracting the code which related to the crash interrupt check, put it into a new function and only invoked it when the CPU is BSP?
Yes, looks a good code cleaning up. Virtual wire irq mode only go through the boot cpu's lapic.
In fact, or put in apic_bsp_setup() since it's only related to boot cpu to make it look cleaner if can do.
arch/x86/kernel/apic/apic.c +1 -0
@@ -2380,6 +2380,7 @@ void __init apic_bsp_setup(bool upmode)connect_bsp_APIC ();if (upmode)apic_bsp_up_setup (); +clear_crash_pending_intr ();setup_local_APIC ();enable_IO_APIC ();
Quoted 6 lines
arch/x86/kernel/apic/apic.c +0
-0
arch/x86/kernel/apic/apic.c static void lapic_setup_esr(void) · line 1417
Quoted 10 lines
arch/x86/kernel/apic/apic.c +7
-0@@ -1408,6 +1408,57 @@ static void lapic_setup_esr(void) oldvalue, value); } + +static void clear_crash_pending_intr (void ) +{ +long long max_loops = cpu_khz ? cpu_khz :1000000 ; +unsigned long long tsc =0 , ntsc; +unsigned int value, queued; +int i, j, acked =0 ;
Better do a check here.
if (disable_apic) return;
Quoted 264 lines
arch/x86/kernel/apic/apic.c +44
-0+ +if (boot_cpu_has (X86_FEATURE_TSC )) + tsc =rdtsc (); +/* +* After a crash, we no longer service the interrupts and a pending +* interrupt from previous kernel might still have ISR bit set. +* +* Most probably by now CPU has serviced that pending interrupt and +* it might not have done the ack_APIC_irq() because it thought, +* interrupt came from i8259 as ExtInt. LAPIC did not get EOI so it +* does not clear the ISR bit and cpu thinks it has already serivced +* the interrupt. Hence a vector might get locked. It was noticed +* for timer irq (vector 0x31). Issue an extra EOI to clear ISR. +*/ +do { + queued =0 ; +for (i =APIC_ISR_NR -1 ; i >=0 ; i--) + queued |=apic_read (APIC_IRR + i*0x10 ); + +for (i =APIC_ISR_NR -1 ; i >=0 ; i--) { + value =apic_read (APIC_ISR + i*0x10 ); +for (j =31 ; j >=0 ; j--) { +if (value & (1 <<j)) { +ack_APIC_irq (); + acked++; + } + } + } +if (acked >256 ) { +printk (KERN_ERR "LAPIC pending interrupts after %d EOI\n", + acked); +break ; + } +if (queued) { +if (boot_cpu_has (X86_FEATURE_TSC ) && cpu_khz) { + ntsc =rdtsc (); + max_loops = (cpu_khz <<10 ) - (ntsc - tsc); + }else + max_loops--; + } + }while (queued && max_loops >0 ); +WARN_ON (max_loops <=0 ); +} +/** * setup_local_APIC - setup the local APIC *
arch/x86/kernel/apic/apic.c +3
-47@@ -1417,13 +1468,7 @@ static void lapic_setup_esr(void)void setup_local_APIC (void ) {int cpu =smp_processor_id (); -unsigned int value, queued; -int i, j, acked =0 ; -unsigned long long tsc =0 , ntsc; -long long max_loops = cpu_khz ? cpu_khz :1000000 ; - -if (boot_cpu_has (X86_FEATURE_TSC )) - tsc =rdtsc (); - tsc =rdtsc (); +unsigned int value;if (disable_apic) {disable_ioapic_support ();@@ -1475,45 +1520,8 @@ void setup_local_APIC(void) value &= ~APIC_TPRI_MASK ;apic_write (APIC_TASKPRI , value); -/* -* After a crash, we no longer service the interrupts and a pending -* interrupt from previous kernel might still have ISR bit set. -* -* Most probably by now CPU has serviced that pending interrupt and -* it might not have done the ack_APIC_irq() because it thought, -* interrupt came from i8259 as ExtInt. LAPIC did not get EOI so it -* does not clear the ISR bit and cpu thinks it has already serivced -* the interrupt. Hence a vector might get locked. It was noticed -* for timer irq (vector 0x31). Issue an extra EOI to clear ISR. -*/ -do { - queued =0 ; -for (i =APIC_ISR_NR -1 ; i >=0 ; i--) - queued |=apic_read (APIC_IRR + i*0x10 ); - -for (i =APIC_ISR_NR -1 ; i >=0 ; i--) { - value =apic_read (APIC_ISR + i*0x10 ); -for (j =31 ; j >=0 ; j--) { -if (value & (1 <<j)) { -ack_APIC_irq (); - acked++; - } - } - } -if (acked >256 ) { -printk (KERN_ERR "LAPIC pending interrupts after %d EOI\n", - acked); -break ; - } -if (queued) { -if (boot_cpu_has (X86_FEATURE_TSC ) && cpu_khz) { - ntsc =rdtsc (); - max_loops = (cpu_khz <<10 ) - (ntsc - tsc); - }else - max_loops--; - } - }while (queued && max_loops >0 ); -WARN_ON (max_loops <=0 ); + if (!cpu) + clear_crash_pending_intr();/* * Now that we are all set up, enable the APIC
Thanks,
dou.Quoted 144 lines
Quoted 28 lines
well any more if pending irq exists in APIC IRR after LAPIC is disabled. And people even saw casual kdump kernel hang once in ~30 attempts during stress testing of kdump on KVM machine.
[ 0.001000] WARNING: CPU: 0 PID: 0 at arch/x86/kernel/apic/apic.c:1467 setup_local_APIC+0x228/0x330 [ 0.001000] Modules linked in: [ 0.001000] CPU: 0 PID: 0 Comm: swapper/0 Not tainted 4.15.0-rc5+ #3 [ 0.001000] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.10.2-1.fc26 04/01/2014 [ 0.001000] RIP: 0010:setup_local_APIC+0x228/0x330 [ 0.001000] RSP: 0000:ffffffffb6e03eb8 EFLAGS: 00010286 [ 0.001000] RAX: 0000009edb4c4d84 RBX: 0000000000000000 RCX: 00000000b099d800 [ 0.001000] RDX: 0000009e00000000 RSI: 0000000000000000 RDI: 0000000000000810 [ 0.001000] RBP: 0000000000000000 R08: ffffffffffffffff R09: 0000000000000001 [ 0.001000] R10: ffff98ce6a801c00 R11: 0761076d072f0776 R12: 0000000000000001 [ 0.001000] R13: 00000000000000f0 R14: 0000000000004000 R15: ffffffffffffc6ff [ 0.001000] FS: 0000000000000000(0000) GS:ffff98ce6bc00000(0000) knlGS:0000000000000000 [ 0.001000] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 0.001000] CR2: 00000000ffffffff CR3: 0000000022209000 CR4: 00000000000406b0 [ 0.001000] Call Trace: [ 0.001000] apic_bsp_setup+0x56/0x74 [ 0.001000] x86_late_time_init+0x11/0x16 [ 0.001000] start_kernel+0x3c9/0x486 [ 0.001000] secondary_startup_64+0xa5/0xb0 [ 0.001000] Code: 00 85 c9 74 2d 0f 31 c1 e1 0a 48 c1 e2 20 41 89 cf 4c 03 7c 24 08 48 09 d0 49 29 c7 4c 89 3c 24 48 83 3c 24 00 0f 8f 8f fe ff ff <0f> ff e9 10 ff ff ff 48 83 2c 24 01 eb e7 48 83 c4 18 5b 5d 41 [ 0.001000] ---[ end trace b88e71b9a6ebebdd ]--- [ 0.001000] masked ExtINT on CPU#0Quoted 4 lines
To fix this, just break down
disable_IO_APIC(), then callclear_IO_APIC()to stopIO_APICwheredisable_IO_APIC()was called, and callrestore_boot_irq_mode()to restore boot irq mode before reboot or kexec/kdump jump.Two things here.
a) This is missing a fixes tag and a CC stable.
b) What makes your change to theKEXEC_JUMPcode path safe? Have the lapic and ioapic already been shut down?The
KEXEC_JUMPchanges tomachine_kexec_32.candmachine_kexec_64.ceither need to be documented in the change long why they are safe so that this change becomes obviously safe and correct.Otherwise we risk and trivial and obvious looking change causing another regression like changing the order of
lapic_shutdownanddisable_IOAPICdid.Eric
Quoted 92 lines
Signed-off-by Baoquan He <[email protected]>--- arch/x86/include/asm/io_apic.h | 1 + arch/x86/kernel/apic/io_apic.c | 2 +- arch/x86/kernel/crash.c | 3 ++- arch/x86/kernel/machine_kexec_32.c | 2 +- arch/x86/kernel/machine_kexec_64.c | 2 +- arch/x86/kernel/reboot.c | 3 ++- 6 files changed, 8 insertions(+), 5 deletions(-)arch/x86/include/asm/io_apic.h +1
-0@@ -193,6 +193,7 @@ static inline unsigned int io_apic_read(unsigned int apic, unsigned int reg)extern void setup_IO_APIC (void );extern void enable_IO_APIC (void );extern void disable_IO_APIC (void ); +extern void clear_IO_APIC (void );extern void restore_boot_irq_mode (void );extern int IO_APIC_get_PCI_irq_vector (int bus,int devfn,int pin);extern void print_IO_APICs (void );arch/x86/kernel/apic/io_apic.c +1
-1@@ -587,7 +587,7 @@ static void clear_IO_APIC_pin(unsigned int apic, unsigned int pin) mpc_ioapic_id(apic), pin); } -static void clear_IO_APIC (void ) +void clear_IO_APIC (void ) {int apic, pin;arch/x86/kernel/crash.c +2
-1@@ -199,9 +199,10 @@ void native_machine_crash_shutdown(struct pt_regs *regs)#ifdef CONFIG_X86_IO_APIC /* Prevent crash_kexec() from deadlocking on ioapic_lock. */ ioapic_zap_locks (); -disable_IO_APIC (); +clear_IO_APIC ();#endif lapic_shutdown (); +restore_boot_irq_mode ();#ifdef CONFIG_HPET_TIMER hpet_disable ();#endif arch/x86/kernel/machine_kexec_32.c +1
-1@@ -199,7 +199,7 @@ void machine_kexec(struct kimage *image)* one form or other. kexec jump path also need * one. */ -disable_IO_APIC (); +restore_boot_irq_mode ();#endif }arch/x86/kernel/machine_kexec_64.c +1
-1@@ -297,7 +297,7 @@ void machine_kexec(struct kimage *image)* one form or other. kexec jump path also need * one. */ -disable_IO_APIC (); +restore_boot_irq_mode ();#endif }arch/x86/kernel/reboot.c +2
-1@@ -666,7 +666,7 @@ void native_machine_shutdown(void)* Even without the erratum, it still makes sense to quiet IO APIC * before disabling Local APIC. */ -disable_IO_APIC (); +clear_IO_APIC ();#endif #ifdef CONFIG_SMP @@ -680,6 +680,7 @@ void native_machine_shutdown(void)#endif lapic_shutdown (); +restore_boot_irq_mode ();#ifdef CONFIG_HPET_TIMER hpet_disable ();
Quoted 25 lines · Dou Liyang
Hi Baoquan,
Quoted 13 lines · Eric W. Biederman
Quoted 11 lines · Baoquan He
This is a regression fix.
Before, to fix erratum AVR31, commit
522e66464467("x86/apic: Disable I/O APIC before shutdown of the local APIC") movedlapic_shutdown()calling afterdisable_IO_APIC(). This introdued a regression. The root cause is thatdisable_IO_APIC()not only clearsIO_APIC, also restore boot irq mode by setting LAPIC/APIC/IMCR,lapic_shutdown()afterdisable_IO_APIC()will disable LAPIC and ruin the possible virtual wire mode setting which the code has been trying to do all along.Quoted 3 lines
Quoted 3 lines
The consequence is, in KVM guest kernel always prints warning as below during kexec/kdump kernel boots up. That happened in
setup_local_APIC()since 'do { xxx } while (queued &&max_loops> 0)' loop does not functionI am not sure another thing here
AFAIK, according to the order of SMP machine shutdown, other CPUs will be stopped firstly, then the last CPU disable its local apic.
Quoted 8 lines
--machine_shutdown
|----...... |----stop_other_cpus() |----local_shutdown()So, the pending interrupts exist only in BSP and only be ACKed by BSP. is it right?(I validated this by print the value of APIC_IRR/ISR of all CPUs, found only BSP had the non-zero value).
We don't know. In the kexec on panic case we try and shutdown the other cpus but we have a timeout because we might fail.
Further you have to be careful with the concept of boot cpu. In the normal kexec case shutdown on the boot cpu and leave it running. In the kexec on panic case we shutdown on an arbitrary cpu.
Quoted 2 lines
If it is right, We will do not need check the pending interrupt for each cpus.
It is also cheap if there are no pending interrupts as there is nothing to do.
Quoted 4 lines
BTW, the pending interrupt check code is mixed with the local APIC setup code, that looks messy. How about extracting the code which related to the crash interrupt check, put it into a new function and only invoked it when the CPU is BSP?
Moving it into it's own function makes sense. Let's not taint the concept with ``crash''. We don't know that the only way this will
ever happen is from kexec on panic. We only know it was easy to
trigger the condition from kexec on panic.
There are a lot of cases I can think of that interrupts might fire while interrupts are disabled because the kernel is booting. A normal kexec is also possible.
Throw in MSI interrupts and transitioning from one state to another in non-legacy apic mode and we might be more likely to get some irqs in a pending state.
Your patch makes me nervous as it is not just code motion but a much more substantial change.
As much as I agree that we need to fix the regression in the apic shutdown code that is causing problems. What we really need to do is to completely remove the apic shutdown code from the kexec on panic code path. Over the long term that will provide us with a much more robust path to getting crash dumps and the like.
Eric
Quoted 132 lines
arch/x86/kernel/apic/apic.c +51
-0@@ -1408,6 +1408,57 @@ static void lapic_setup_esr(void) oldvalue, value); } + +static void clear_crash_pending_intr (void ) +{ +long long max_loops = cpu_khz ? cpu_khz :1000000 ; +unsigned long long tsc =0 , ntsc; +unsigned int value, queued; +int i, j, acked =0 ; + +if (boot_cpu_has (X86_FEATURE_TSC )) + tsc =rdtsc (); +/* +* After a crash, we no longer service the interrupts and a pending +* interrupt from previous kernel might still have ISR bit set. +* +* Most probably by now CPU has serviced that pending interrupt and +* it might not have done the ack_APIC_irq() because it thought, +* interrupt came from i8259 as ExtInt. LAPIC did not get EOI so it +* does not clear the ISR bit and cpu thinks it has already serivced +* the interrupt. Hence a vector might get locked. It was noticed +* for timer irq (vector 0x31). Issue an extra EOI to clear ISR. +*/ +do { + queued =0 ; +for (i =APIC_ISR_NR -1 ; i >=0 ; i--) + queued |=apic_read (APIC_IRR + i*0x10 ); + +for (i =APIC_ISR_NR -1 ; i >=0 ; i--) { + value =apic_read (APIC_ISR + i*0x10 ); +for (j =31 ; j >=0 ; j--) { +if (value & (1 <<j)) { +ack_APIC_irq (); + acked++; + } + } + } +if (acked >256 ) { +printk (KERN_ERR "LAPIC pending interrupts after %d EOI\n", + acked); +break ; + } +if (queued) { +if (boot_cpu_has (X86_FEATURE_TSC ) && cpu_khz) { + ntsc =rdtsc (); + max_loops = (cpu_khz <<10 ) - (ntsc - tsc); + }else + max_loops--; + } + }while (queued && max_loops >0 ); +WARN_ON (max_loops <=0 ); +} +/** * setup_local_APIC - setup the local APIC *
arch/x86/kernel/apic/apic.c +3
-47@@ -1417,13 +1468,7 @@ static void lapic_setup_esr(void)void setup_local_APIC (void ) {int cpu =smp_processor_id (); -unsigned int value, queued; -int i, j, acked =0 ; -unsigned long long tsc =0 , ntsc; -long long max_loops = cpu_khz ? cpu_khz :1000000 ; - -if (boot_cpu_has (X86_FEATURE_TSC )) - tsc =rdtsc (); - tsc =rdtsc (); +unsigned int value;if (disable_apic) {disable_ioapic_support ();@@ -1475,45 +1520,8 @@ void setup_local_APIC(void) value &= ~APIC_TPRI_MASK ;apic_write (APIC_TASKPRI , value); -/* -* After a crash, we no longer service the interrupts and a pending -* interrupt from previous kernel might still have ISR bit set. -* -* Most probably by now CPU has serviced that pending interrupt and -* it might not have done the ack_APIC_irq() because it thought, -* interrupt came from i8259 as ExtInt. LAPIC did not get EOI so it -* does not clear the ISR bit and cpu thinks it has already serivced -* the interrupt. Hence a vector might get locked. It was noticed -* for timer irq (vector 0x31). Issue an extra EOI to clear ISR. -*/ -do { - queued =0 ; -for (i =APIC_ISR_NR -1 ; i >=0 ; i--) - queued |=apic_read (APIC_IRR + i*0x10 ); - -for (i =APIC_ISR_NR -1 ; i >=0 ; i--) { - value =apic_read (APIC_ISR + i*0x10 ); -for (j =31 ; j >=0 ; j--) { -if (value & (1 <<j)) { -ack_APIC_irq (); - acked++; - } - } - } -if (acked >256 ) { -printk (KERN_ERR "LAPIC pending interrupts after %d EOI\n", - acked); -break ; - } -if (queued) { -if (boot_cpu_has (X86_FEATURE_TSC ) && cpu_khz) { - ntsc =rdtsc (); - max_loops = (cpu_khz <<10 ) - (ntsc - tsc); - }else - max_loops--; - } - }while (queued && max_loops >0 ); -WARN_ON (max_loops <=0 ); + if (!cpu) + clear_crash_pending_intr();/* * Now that we are all set up, enable the APIC
Thanks,
dou.
Hi Eric,
Quoted 43 lines · Eric W. Biederman
Quoted 25 lines · Dou Liyang
Hi Baoquan,
Quoted 13 lines · Eric W. Biederman
Quoted 11 lines · Baoquan He
This is a regression fix.
Before, to fix erratum AVR31, commit
522e66464467("x86/apic: Disable I/O APIC before shutdown of the local APIC") movedlapic_shutdown()calling afterdisable_IO_APIC(). This introdued a regression. The root cause is thatdisable_IO_APIC()not only clearsIO_APIC, also restore boot irq mode by setting LAPIC/APIC/IMCR,lapic_shutdown()afterdisable_IO_APIC()will disable LAPIC and ruin the possible virtual wire mode setting which the code has been trying to do all along.Quoted 3 lines
Quoted 3 lines
The consequence is, in KVM guest kernel always prints warning as below during kexec/kdump kernel boots up. That happened in
setup_local_APIC()since 'do { xxx } while (queued &&max_loops> 0)' loop does not functionI am not sure another thing here
AFAIK, according to the order of SMP machine shutdown, other CPUs will be stopped firstly, then the last CPU disable its local apic.
Quoted 8 lines
--machine_shutdown
|----...... |----stop_other_cpus() |----local_shutdown()So, the pending interrupts exist only in BSP and only be ACKed by BSP. is it right?(I validated this by print the value of APIC_IRR/ISR of all CPUs, found only BSP had the non-zero value).
We don't know. In the kexec on panic case we try and shutdown the other cpus but we have a timeout because we might fail.
Further you have to be careful with the concept of boot cpu. In the normal kexec case shutdown on the boot cpu and leave it running. In the kexec on panic case we shutdown on an arbitrary cpu.
Quoted 6 lines
Quoted 2 lines
If it is right, We will do not need check the pending interrupt for each cpus.
It is also cheap if there are no pending interrupts as there is nothing to do.
Yes, It's cheap.
But, the local APICs in APs were disabled at that time, not sure if writing to the end-of-interrupt (EOI) register could cause the local APIC to clear the ISR successfully.
If the ISR was not cleared, doing that check is useless.
I can't produce any cases that the lapics in APs have pending interrupts. do you have some suggestions?
Quoted 10 lines
Quoted 4 lines
BTW, the pending interrupt check code is mixed with the local APIC setup code, that looks messy. How about extracting the code which related to the crash interrupt check, put it into a new function and only invoked it when the CPU is BSP?
Moving it into it's own function makes sense. Let's not taint the concept with ``crash''. We don't know that the only way this will
ever happen is from kexec on panic. We only know it was easy to
trigger the condition from kexec on panic.
Yes, I see.
Quoted 8 lines
There are a lot of cases I can think of that interrupts might fire while interrupts are disabled because the kernel is booting. A normal kexec is also possible.
Throw in MSI interrupts and transitioning from one state to another in non-legacy apic mode and we might be more likely to get some irqs in a pending state.
Quoted 8 lines
Your patch makes me nervous as it is not just code motion but a much more substantial change.
As much as I agree that we need to fix the regression in the apic shutdown code that is causing problems. What we really need to do is to completely remove the apic shutdown code from the kexec on panic code path. Over the long term that will provide us with a much more
Wow, indeed, and it is related to many hypervisors on x86, For me, still need more investigations and tests.
Thanks,
dou
Hi Eric,
Quoted 17 lines · Eric W. Biederman
Quoted 14 lines · Baoquan He
This is a regression fix.
Before, to fix erratum AVR31, commit
522e66464467("x86/apic: Disable I/O APIC before shutdown of the local APIC") movedlapic_shutdown()calling afterdisable_IO_APIC(). This introdued a regression. The root cause is thatdisable_IO_APIC()not only clearsIO_APIC, also restore boot irq mode by setting LAPIC/APIC/IMCR,lapic_shutdown()afterdisable_IO_APIC()will disable LAPIC and ruin the possible virtual wire mode setting which the code has been trying to do all along.
To fix this, just break downdisable_IO_APIC(), then callclear_IO_APIC()to stopIO_APICwheredisable_IO_APIC()was called, and callrestore_boot_irq_mode()to restore boot irq mode before reboot or kexec/kdump jump.
Quoted 8 lines
Two things here.
a) This is missing a fixes tag and a CC stable.
b) What makes your change to theKEXEC_JUMPcode path safe? Have the lapic and ioapic already been shut down?The
KEXEC_JUMPchanges tomachine_kexec_32.candmachine_kexec_64.ceither need to be documented in the change long why they are safe so that this change becomes obviously safe and correct.
Re-read the code, I have to admit I didn't check the KEXEC_JUMP code path carefully.
kernel_kexec() {
if (kexec_image->preserve_context) {
...
freeze_processes();
...
disable_nonboot_cpus();
...
else {
...
machine_shutdown();
...
}
machine_kexec(kexec_image);
...
}--machine_shutdown()
--native_machine_shutdown()
--disable_IO_APIC()
--lapic_shutdown()machine_kexec() {
...
if (image->preserve_context) {
disable_IO_APIC();
}
...
}KEXEC_JUMP code path is different than kexec/kdump, it doesn't call lapic_shutdown() before jump. So commit 522e66464467 ("x86/apic: Disable I/O APIC before shutdown of the local APIC") didn't impact it. And here I break down disable_IO_APIC() and change to only call restore_boot_irq_mode() to make a possible danger. I am not an expert on KEXEC_JUMP, and don't know how to test it, so will keep the code implementation consistent as before. For now, I plan to change it as below if you don't object. As you pointed out, I will describe this in patch log.
arch/x86/kernel/machine_kexec_64.c +2 -1
@@ -297,7 +297,7 @@ void machine_kexec(struct kimage *image)* one form or other. kexec jump path also need * one. */ -disable_IO_APIC (); +clear_IO_APIC (); +restore_boot_irq_mode ();#endif }
Quoted 107 lines
Otherwise we risk and trivial and obvious looking change causing another regression like changing the order of
lapic_shutdownanddisable_IOAPICdid.Eric
Quoted 99 lines
Signed-off-by Baoquan He <[email protected]>--- arch/x86/include/asm/io_apic.h | 1 + arch/x86/kernel/apic/io_apic.c | 2 +- arch/x86/kernel/crash.c | 3 ++- arch/x86/kernel/machine_kexec_32.c | 2 +- arch/x86/kernel/machine_kexec_64.c | 2 +- arch/x86/kernel/reboot.c | 3 ++- 6 files changed, 8 insertions(+), 5 deletions(-)arch/x86/include/asm/io_apic.h +1
-0@@ -193,6 +193,7 @@ static inline unsigned int io_apic_read(unsigned int apic, unsigned int reg)extern void setup_IO_APIC (void );extern void enable_IO_APIC (void );extern void disable_IO_APIC (void ); +extern void clear_IO_APIC (void );extern void restore_boot_irq_mode (void );extern int IO_APIC_get_PCI_irq_vector (int bus,int devfn,int pin);extern void print_IO_APICs (void );arch/x86/kernel/apic/io_apic.c +1
-1@@ -587,7 +587,7 @@ static void clear_IO_APIC_pin(unsigned int apic, unsigned int pin) mpc_ioapic_id(apic), pin); } -static void clear_IO_APIC (void ) +void clear_IO_APIC (void ) {int apic, pin;arch/x86/kernel/crash.c +2
-1@@ -199,9 +199,10 @@ void native_machine_crash_shutdown(struct pt_regs *regs)#ifdef CONFIG_X86_IO_APIC /* Prevent crash_kexec() from deadlocking on ioapic_lock. */ ioapic_zap_locks (); -disable_IO_APIC (); +clear_IO_APIC ();#endif lapic_shutdown (); +restore_boot_irq_mode ();#ifdef CONFIG_HPET_TIMER hpet_disable ();#endif arch/x86/kernel/machine_kexec_32.c +1
-1@@ -199,7 +199,7 @@ void machine_kexec(struct kimage *image)* one form or other. kexec jump path also need * one. */ -disable_IO_APIC (); +restore_boot_irq_mode ();#endif }arch/x86/kernel/machine_kexec_64.c +1
-1@@ -297,7 +297,7 @@ void machine_kexec(struct kimage *image)* one form or other. kexec jump path also need * one. */ -disable_IO_APIC (); +restore_boot_irq_mode ();#endif }arch/x86/kernel/reboot.c +2
-1@@ -666,7 +666,7 @@ void native_machine_shutdown(void)* Even without the erratum, it still makes sense to quiet IO APIC * before disabling Local APIC. */ -disable_IO_APIC (); +clear_IO_APIC ();#endif #ifdef CONFIG_SMP @@ -680,6 +680,7 @@ void native_machine_shutdown(void)#endif lapic_shutdown (); +restore_boot_irq_mode ();#ifdef CONFIG_HPET_TIMER hpet_disable ();
Quoted 79 lines · Baoquan He
Hi Eric,
Quoted 17 lines · Eric W. Biederman
Quoted 14 lines · Baoquan He
This is a regression fix.
Before, to fix erratum AVR31, commit
522e66464467("x86/apic: Disable I/O APIC before shutdown of the local APIC") movedlapic_shutdown()calling afterdisable_IO_APIC(). This introdued a regression. The root cause is thatdisable_IO_APIC()not only clearsIO_APIC, also restore boot irq mode by setting LAPIC/APIC/IMCR,lapic_shutdown()afterdisable_IO_APIC()will disable LAPIC and ruin the possible virtual wire mode setting which the code has been trying to do all along.
To fix this, just break downdisable_IO_APIC(), then callclear_IO_APIC()to stopIO_APICwheredisable_IO_APIC()was called, and callrestore_boot_irq_mode()to restore boot irq mode before reboot or kexec/kdump jump.Quoted 8 lines
Two things here.
a) This is missing a fixes tag and a CC stable.
b) What makes your change to theKEXEC_JUMPcode path safe? Have the lapic and ioapic already been shut down?The
KEXEC_JUMPchanges tomachine_kexec_32.candmachine_kexec_64.ceither need to be documented in the change long why they are safe so that this change becomes obviously safe and correct.Re-read the code, I have to admit I didn't check the
KEXEC_JUMPcode path carefully.kernel_kexec() { if (kexec_image->preserve_context) { ... freeze_processes(); ... disable_nonboot_cpus(); ... else { ... machine_shutdown(); ... } machine_kexec(kexec_image); ... }--machine_shutdown()
--native_machine_shutdown() --disable_IO_APIC() --lapic_shutdown()machine_kexec() { ... if (image->preserve_context) { disable_IO_APIC(); } ... }
KEXEC_JUMPcode path is different than kexec/kdump, it doesn't calllapic_shutdown()before jump. So commit522e66464467("x86/apic: Disable I/O APIC before shutdown of the local APIC") didn't impact it. And here I break downdisable_IO_APIC()and change to only callrestore_boot_irq_mode()to make a possible danger. I am not an expert onKEXEC_JUMP, and don't know how to test it, so will keep the code implementation consistent as before. For now, I plan to change it as below if you don't object. As you pointed out, I will describe this in patch log.arch/x86/kernel/machine_kexec_64.c +0
-0
arch/x86/kernel/machine_kexec_64.c void machine_kexec(struct kimage *image) · line 303
Quoted 8 lines
arch/x86/kernel/machine_kexec_64.c +2
-1@@ -297,7 +297,7 @@ void machine_kexec(struct kimage *image)* one form or other. kexec jump path also need * one. */ -disable_IO_APIC (); +clear_IO_APIC (); +restore_boot_irq_mode ();#endif }
Let me give a very concrete suggestion:
Patch 1) Replace "disable_IO_APIC();" with "clear_IO_APIC(); restore_boot_irq_mode();" Patch 2) Move restore_boot_irq_mode(); to fix the regression.
I think that will be a slightly shorter patch sequence than what you are dealing with and one that is slightly easier to read.
We need to sort out KEXEC_JUMP but that is something for another time.
Eric
Quoted 94 lines · Eric W. Biederman
Quoted 79 lines · Baoquan He
Hi Eric,
Quoted 17 lines · Eric W. Biederman
Quoted 14 lines · Baoquan He
This is a regression fix.
Before, to fix erratum AVR31, commit
522e66464467("x86/apic: Disable I/O APIC before shutdown of the local APIC") movedlapic_shutdown()calling afterdisable_IO_APIC(). This introdued a regression. The root cause is thatdisable_IO_APIC()not only clearsIO_APIC, also restore boot irq mode by setting LAPIC/APIC/IMCR,lapic_shutdown()afterdisable_IO_APIC()will disable LAPIC and ruin the possible virtual wire mode setting which the code has been trying to do all along.
To fix this, just break downdisable_IO_APIC(), then callclear_IO_APIC()to stopIO_APICwheredisable_IO_APIC()was called, and callrestore_boot_irq_mode()to restore boot irq mode before reboot or kexec/kdump jump.Quoted 8 lines
Two things here.
a) This is missing a fixes tag and a CC stable.
b) What makes your change to theKEXEC_JUMPcode path safe? Have the lapic and ioapic already been shut down?The
KEXEC_JUMPchanges tomachine_kexec_32.candmachine_kexec_64.ceither need to be documented in the change long why they are safe so that this change becomes obviously safe and correct.Re-read the code, I have to admit I didn't check the
KEXEC_JUMPcode path carefully.kernel_kexec() { if (kexec_image->preserve_context) { ... freeze_processes(); ... disable_nonboot_cpus(); ... else { ... machine_shutdown(); ... } machine_kexec(kexec_image); ... }--machine_shutdown()
--native_machine_shutdown() --disable_IO_APIC() --lapic_shutdown()machine_kexec() { ... if (image->preserve_context) { disable_IO_APIC(); } ... }
KEXEC_JUMPcode path is different than kexec/kdump, it doesn't calllapic_shutdown()before jump. So commit522e66464467("x86/apic: Disable I/O APIC before shutdown of the local APIC") didn't impact it. And here I break downdisable_IO_APIC()and change to only callrestore_boot_irq_mode()to make a possible danger. I am not an expert onKEXEC_JUMP, and don't know how to test it, so will keep the code implementation consistent as before. For now, I plan to change it as below if you don't object. As you pointed out, I will describe this in patch log.arch/x86/kernel/machine_kexec_64.c +0
-0arch/x86/kernel/machine_kexec_64.c void machine_kexec(struct kimage *image) · line 303
Quoted 8 lines
arch/x86/kernel/machine_kexec_64.c +2
-1@@ -297,7 +297,7 @@ void machine_kexec(struct kimage *image)* one form or other. kexec jump path also need * one. */ -disable_IO_APIC (); +clear_IO_APIC (); +restore_boot_irq_mode ();#endif }Let me give a very concrete suggestion:
Patch 1) Replace "disable_IO_APIC();" with "clear_IO_APIC();restore_boot_irq_mode();" Patch 2) Moverestore_boot_irq_mode(); to fix the regression.
Quoted 2 lines
I think that will be a slightly shorter patch sequence than what you are dealing with and one that is slightly easier to read.
Sure, will do.
Thanks for reviewing and suggestion.
Quoted 2 lines
We need to sort out
KEXEC_JUMPbut that is something for another time.
Agree. We ever contacted the author, intel dev, seems they don't maintain it any more. And very few people are interested and report the relevant issues. While recently someone is testing KEXEC_JUMP functionality and updating related doc, will continue tracking it.
Thanks
Baoquan
PATCH v3 3/5 x86/apic: Remove useless disable_IO_APIC
In reply to Baoquan HeNo one uses it anymore.
--- arch/x86/include/asm/io_apic.h | 1 - arch/x86/kernel/apic/io_apic.c | 13 ------------- arch/x86/kernel/machine_kexec_32.c | 5 ++--- arch/x86/kernel/machine_kexec_64.c | 5 ++--- 4 files changed, 4 insertions(+), 20 deletions(-)
arch/x86/include/asm/io_apic.h +0 -1
@@ -192,7 +192,6 @@ static inline unsigned int io_apic_read(unsigned int apic, unsigned int reg)extern void setup_IO_APIC (void );extern void enable_IO_APIC (void ); -extern void disable_IO_APIC (void );extern void clear_IO_APIC (void );extern void restore_boot_irq_mode (void );extern int IO_APIC_get_PCI_irq_vector (int bus,int devfn,int pin);
arch/x86/kernel/apic/io_apic.c +0 -13
@@ -1438,19 +1438,6 @@ void native_disable_io_apic(void)disconnect_bsp_APIC (ioapic_i8259.pin !=-1 ); } -/* -* Not an __init, needed by the reboot code -*/ -void disable_IO_APIC (void ) -{ -/* -* Clear the IO-APIC before rebooting: -*/ -clear_IO_APIC (); - -restore_boot_irq_mode (); -} -void restore_boot_irq_mode (void ) {if (!nr_legacy_irqs ())
arch/x86/kernel/machine_kexec_32.c +2 -3
@@ -195,9 +195,8 @@ void machine_kexec(struct kimage *image)/* * We need to put APICs in legacy mode so that we can * get timer interrupts in second kernel. kexec/kdump -* paths already have calls to disable_IO_APIC() in -* one form or other. kexec jump path also need -* one. +* paths already have calls to restore_boot_irq_mode() +* in one form or other. kexec jump path also need one. */ restore_boot_irq_mode ();#endif
arch/x86/kernel/machine_kexec_64.c +2 -3
@@ -293,9 +293,8 @@ void machine_kexec(struct kimage *image)/* * We need to put APICs in legacy mode so that we can * get timer interrupts in second kernel. kexec/kdump -* paths already have calls to disable_IO_APIC() in -* one form or other. kexec jump path also need -* one. +* paths already have calls to restore_boot_irq_mode() +* in one form or other. kexec jump path also need one. */ restore_boot_irq_mode ();#endif
Signature · 1 line
2.13.6
PATCH v3 4/5 x86/apic: Rename variable/function related to x86_io_apic_ops
In reply to Baoquan HeThe names of x86_io_apic_ops and its two member variables, are misleading. The .read member is to read IO_APIC reg, while .disable which hook native_disable_io_apic/irq_remapping_disable_io_apic is actually used to restore boot irq mode, not disable IO_APIC.
So rename x86_io_apic_ops to x86_apic_ops since it doesn't only handle IO_APIC, also LAPIC.
And also rename its member variables and the related hooks.
--- arch/x86/include/asm/io_apic.h | 6 +++--- arch/x86/include/asm/x86_init.h | 8 ++++---- arch/x86/kernel/apic/io_apic.c | 4 ++-- arch/x86/kernel/x86_init.c | 6 +++--- arch/x86/xen/apic.c | 2 +- drivers/iommu/irq_remapping.c | 4 ++-- 6 files changed, 15 insertions(+), 15 deletions(-)
arch/x86/include/asm/io_apic.h +3 -3
@@ -183,11 +183,11 @@ extern void disable_ioapic_support(void);extern void __init io_apic_init_mappings (void );extern unsigned int native_io_apic_read (unsigned int apic,unsigned int reg); -extern void native_disable_io_apic (void ); +extern void native_restore_boot_irq_mode (void );static inline unsigned int io_apic_read (unsigned int apic,unsigned int reg) { -return x86_io_apic_ops.read (apic, reg); +return x86_apic_ops.io_apic_read (apic, reg); }extern void setup_IO_APIC (void );
@@ -229,7 +229,7 @@ static inline void mp_save_irq(struct mpc_intsrc *m) { }static inline void disable_ioapic_support (void ) { }static inline void io_apic_init_mappings (void ) { }#define native_io_apic_read NULL -#define native_disable_io_apic NULL +#define native_restore_boot_irq_mode NULLstatic inline void setup_IO_APIC (void ) { }static inline void enable_IO_APIC (void ) { }
arch/x86/include/asm/x86_init.h +4 -4
@@ -274,16 +274,16 @@ struct x86_msi_ops {void (*restore_msi_irqs)(struct pci_dev *dev); }; -struct x86_io_apic_ops { -unsigned int (*read) (unsigned int apic,unsigned int reg); -void (*disable)(void ); +struct x86_apic_ops { +unsigned int (*io_apic_read) (unsigned int apic,unsigned int reg); +void (*restore)(void ); };extern struct x86_init_ops x86_init;extern struct x86_cpuinit_ops x86_cpuinit;extern struct x86_platform_ops x86_platform;extern struct x86_msi_ops x86_msi; -extern struct x86_io_apic_ops x86_io_apic_ops; +extern struct x86_apic_ops x86_apic_ops;extern void x86_early_init_platform_quirks (void );extern void x86_init_noop (void );
arch/x86/kernel/apic/io_apic.c +2 -2
@@ -1410,7 +1410,7 @@ void __init enable_IO_APIC(void)clear_IO_APIC (); } -void native_disable_io_apic (void ) +void native_restore_boot_irq_mode (void ) {/* * If the i8259 is routed through an IOAPIC
@@ -1443,7 +1443,7 @@ void restore_boot_irq_mode(void)if (!nr_legacy_irqs ())return ; - x86_io_apic_ops.disable (); + x86_apic_ops.restore (); }#ifdef CONFIG_X86_32
arch/x86/kernel/x86_init.c +3 -3
@@ -146,7 +146,7 @@ void arch_restore_msi_irqs(struct pci_dev *dev) }#endif -struct x86_io_apic_ops x86_io_apic_ops__ro_after_init = { - .read = native_io_apic_read, - .disable = native_disable_io_apic, +struct x86_apic_ops x86_apic_ops__ro_after_init = { + .io_apic_read = native_io_apic_read, + .restore = native_restore_boot_irq_mode, };
arch/x86/xen/apic.c +1 -1
@@ -215,7 +215,7 @@ static void __init xen_apic_check(void) }void __init xen_init_apic (void ) { - x86_io_apic_ops.read = xen_io_apic_read; + x86_apic_ops.io_apic_read = xen_io_apic_read;/* On PV guests the APIC CPUID bit is disabled so none of the * routines end up executing. */ if (!xen_initial_domain ())
drivers/iommu/irq_remapping.c +2 -2
@@ -27,7 +27,7 @@ int disable_irq_post = 0;static int disable_irq_remap;static struct irq_remap_ops *remap_ops; -static void irq_remapping_disable_io_apic (void ) +static void irq_remapping_restore_boot_irq_mode (void ) {/* * With interrupt-remapping, for now we will use virtual wire A
@@ -42,7 +42,7 @@ static void irq_remapping_disable_io_apic(void)static void __init irq_remapping_modify_x86_ops (void ) { - x86_io_apic_ops.disable = irq_remapping_disable_io_apic; + x86_apic_ops.restore = irq_remapping_restore_boot_irq_mode; }static __init int setup_nointremap (char *str)
Signature · 1 line
2.13.6
PATCH v3 5/5 x86/apic: Set up through-local-APIC on boot CPU if 'noapic' specified
In reply to Baoquan HeCurrently kdump kernel becomes very slow if 'noapic' is specified. Normal kernel won't.
Kernel parameter 'noapic' is used to disable IO-APIC in system for testing or special purpose. Here the root cause is that in kdump kernel LAPIC is disabled since commit 522e664644 ("x86/apic: Disable I/O APIC before shutdown of the local APIC"). In this case We need set up through-local-APIC on boot CPU in setup_local_APIC().
While In normal kernel the legacy irq mode is enabled in BIOS. If it is virtual wire mode, the local-APIC has been enabled and set as through-local-APIC.
Though we fix the regression introduced by criminal commit 522e664644, for safety and clarity, better set up through-local-APIC explicitly, but not rely on the default boot irq mode.
Do it now.
--- arch/x86/kernel/apic/apic.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
arch/x86/kernel/apic/apic.c +1 -1
@@ -1570,7 +1570,7 @@ void setup_local_APIC(void)* TODO: set up through-local-APIC from through-I/O-APIC? --macro */ value =apic_read (APIC_LVT0 ) &APIC_LVT_MASKED ; -if (!cpu && (pic_mode || !value)) { +if (!cpu && (pic_mode || !value || skip_ioapic_setup)) { value =APIC_DM_EXTINT ;apic_printk (APIC_VERBOSE ,"enabled ExtINT on CPU#%d\n" , cpu); }else {
Signature · 1 line
2.13.6
Hi all,
One thing confused me.
The disconnect_bsp_APIC() may restore the interrupt delivery mode into virtual wire mode. it uses the vector F as the spurious interrput, But, IMO, using the vector 0xFF(SPURIOUS_APIC_VECTOR) may more suitable and will give us more detail. Why the disconnect_bsp_APIC() use vector F here?
arch/x86/kernel/apic/apic.c +1 -1
@@ -2130,7 +2130,7 @@ void disconnect_bsp_APIC(int virt_wire_setup) value =apic_read (APIC_SPIV ); value &= ~APIC_VECTOR_MASK ; value |=APIC_SPIV_APIC_ENABLED ; - value |=0xf ; + value |=SPURIOUS_APIC_VECTOR ;apic_write (APIC_SPIV , value);if (!virt_wire_setup) {
Thanks,
dou
Quoted 33 lines · Baoquan He
A regression bug was introduced in below commit.
commit522e66464467("x86/apic: Disable I/O APIC before shutdown of the local APIC")It caused the action to fail which we try to restore boot irq mode in reboot and kexec/kdump. Details can be seen in patch 0002.
Warning can always be seen during kdump kernel boot on qemu/kvm platform. Our customer even saw casual kdump kernel hang once in ~30 attempts during stress testing of kdump on KVM machine.
This is v3 post, patches are rearranged and changed according to Eric's suggestions.
Baoquan He (5):
x86/apic: Split outrestore_boot_irq_modefromdisable_IO_APICx86/apic: Fix restoring boot irq mode in reboot and kexec/kdump x86/apic: Remove uselessdisable_IO_APIC
x86/apic: Rename variable/function related tox86_io_apic_opsx86/apic: Set up through-local-APIC on boot CPU if 'noapic' specifiedarch/x86/include/asm/io_apic.h | 9 +++++---- arch/x86/include/asm/x86_init.h | 8 ++++---- arch/x86/kernel/apic/apic.c | 2 +- arch/x86/kernel/apic/io_apic.c | 16 ++++------------ arch/x86/kernel/crash.c | 3 ++- arch/x86/kernel/machine_kexec_32.c | 7 +++---- arch/x86/kernel/machine_kexec_64.c | 7 +++---- arch/x86/kernel/reboot.c | 3 ++- arch/x86/kernel/x86_init.c | 6 +++--- arch/x86/xen/apic.c | 2 +- drivers/iommu/irq_remapping.c | 4 ++-- 11 files changed, 30 insertions(+), 37 deletions(-)
Quoted 9 lines · Dou Liyang
Hi all,
One thing confused me.
The
disconnect_bsp_APIC()may restore the interrupt delivery mode into virtual wire mode. it uses the vector F as the spurious interrput, But, IMO, using the vector 0xFF(SPURIOUS_APIC_VECTOR) may more suitable and will give us more detail. Why thedisconnect_bsp_APIC()use vector F here?
I would say this needs a documentation search before changing this.
This code originates in:208fb93162d5 ("[PATCH] kexec: x86_64: restore apic virtual wire mode on shutdown")
The example in the Multi-Processor Specification v1.4 shows setting up the SPIV to vector 0x0f.
I don't know what is canonical and what will interact best with DOS, and that erra of setup. The vector 0x0f seems an odd choice as it is below 0x20 putting it in the range of vectors that are reserved for processor exceptions.
The constant SPURIOUS_APIC_VECTOR is definitely not something we want to be using at this point as that is a linux specific setting and used when Linux is up and running. So it is completely inapplicable.
This is all about restoring how the apics were configured at boot time so it may be appropriate to copy and store this value, if it was not architectural.
At a practical level at this point I suspect we are ok as the setting of the SPIV this way has not caused any known problems in the last
decade. If someone wants to dig through the archtectural documents
and the real world practice and find a better value and explain the change I would not oppose it.
All I know for certain is using the constant SPURIOUS_APIC_VECTOR is completely inappropriate (as that constant is about how linux uses vectors) and thus the patch below is wrong.
Quoted 6 lines
arch/x86/kernel/apic/apic.c +0
-0
arch/x86/kernel/apic/apic.c void disconnect_bsp_APIC(int virt_wire_setup) · line 2136
Quoted 8 lines
arch/x86/kernel/apic/apic.c +1
-1@@ -2130,7 +2130,7 @@ void disconnect_bsp_APIC(int virt_wire_setup) value =apic_read (APIC_SPIV ); value &= ~APIC_VECTOR_MASK ; value |=APIC_SPIV_APIC_ENABLED ; - value |=0xf ; + value |=SPURIOUS_APIC_VECTOR ;apic_write (APIC_SPIV , value);if (!virt_wire_setup) {
Eric
Hi Eric,
Quoted 12 lines · Eric W. Biederman
Quoted 9 lines · Dou Liyang
Hi all,
One thing confused me.
The
disconnect_bsp_APIC()may restore the interrupt delivery mode into virtual wire mode. it uses the vector F as the spurious interrput, But, IMO, using the vector 0xFF(SPURIOUS_APIC_VECTOR) may more suitable and will give us more detail. Why thedisconnect_bsp_APIC()use vector F here?
Quoted 8 lines
I would say this needs a documentation search before changing this.
This code originates in:
208fb93162d5("[PATCH] kexec:x86_64: restore apic virtual wire mode on shutdown")The example in the Multi-Processor Specification v1.4 shows setting up the SPIV to vector
0x0f.
Thanks for your detailed explanation, I found the example A-1 in this spec.
Quoted 13 lines
I don't know what is canonical and what will interact best with DOS, and that erra of setup. The vector
0x0fseems an odd choice as it is below0x20putting it in the range of vectors that are reserved for processor exceptions.The constant
SPURIOUS_APIC_VECTORis definitely not something we want to be using at this point as that is a linux specific setting and used when Linux is up and running. So it is completely inapplicable.This is all about restoring how the apics were configured at boot time so it may be appropriate to copy and store this value, if it was not architectural.
Quoted 5 lines
At a practical level at this point I suspect we are ok as the setting of the SPIV this way has not caused any known problems in the last
decade. If someone wants to dig through the archtectural documents
and the real world practice and find a better value and explain the change I would not oppose it.
Indeed, I understand.
Thanks,
dou
Quoted 24 lines
All I know for certain is using the constant
SPURIOUS_APIC_VECTORis completely inappropriate (as that constant is about how linux uses vectors) and thus the patch below is wrong.Quoted 6 lines
arch/x86/kernel/apic/apic.c +0
-0arch/x86/kernel/apic/apic.c void disconnect_bsp_APIC(int virt_wire_setup) · line 2136
Quoted 8 lines
arch/x86/kernel/apic/apic.c +1
-1@@ -2130,7 +2130,7 @@ void disconnect_bsp_APIC(int virt_wire_setup) value =apic_read (APIC_SPIV ); value &= ~APIC_VECTOR_MASK ; value |=APIC_SPIV_APIC_ENABLED ; - value |=0xf ; + value |=SPURIOUS_APIC_VECTOR ;apic_write (APIC_SPIV , value);if (!virt_wire_setup) {Eric
Quoted 39 lines · Eric W. Biederman
Quoted 9 lines · Dou Liyang
Hi all,
One thing confused me.
The
disconnect_bsp_APIC()may restore the interrupt delivery mode into virtual wire mode. it uses the vector F as the spurious interrput, But, IMO, using the vector 0xFF(SPURIOUS_APIC_VECTOR) may more suitable and will give us more detail. Why thedisconnect_bsp_APIC()use vector F here?I would say this needs a documentation search before changing this.
This code originates in:
208fb93162d5("[PATCH] kexec:x86_64: restore apic virtual wire mode on shutdown")The example in the Multi-Processor Specification v1.4 shows setting up the SPIV to vector
0x0f.I don't know what is canonical and what will interact best with DOS, and that erra of setup. The vector
0x0fseems an odd choice as it is below0x20putting it in the range of vectors that are reserved for processor exceptions.The constant
SPURIOUS_APIC_VECTORis definitely not something we want to be using at this point as that is a linux specific setting and used when Linux is up and running. So it is completely inapplicable.This is all about restoring how the apics were configured at boot time so it may be appropriate to copy and store this value, if it was not architectural.
At a practical level at this point I suspect we are ok as the setting of the SPIV this way has not caused any known problems in the last
decade. If someone wants to dig through the archtectural documents
and the real world practice and find a better value and explain the change I would not oppose it.
Quoted 3 lines
All I know for certain is using the constant
SPURIOUS_APIC_VECTORis completely inappropriate (as that constant is about how linux uses vectors) and thus the patch below is wrong.
I dig a little deep into doc and code, there are some findings:
1) In Intel® 64 and IA-32 Architectures Software Developer’s Manual and MP-Spec, both mentioned that Spurious-Interrupt Vector Register (SVR) should be 0xxF. For P6 family and Pentium processors, bits 0 through 3 are hardwired to logical ones. Please see 10.9 SPURIOUS INTERRUPT of intel manual vol-3A, part 1.
2) For vector 0xf, we do have a X86_TRAP_SPURIOUS which value is 0xf. In intel manual mentioned as above, Table 6-1. Protected-Mode Exceptions and Interrupts, vector 15 is "(Intel reserved. Do not use.)".
3) I made a debug patch as below to print out the value of SVR at the very beginning of system init, on a bare metal system, its value is 0x10F, means the default value of BIOS setting is 0xF for spurious interrupt vector.
3.1* I tested on qemu-2.9, the printed value is 0x1ff. 3.2* Tested with changing kdump kernel's spurious vector to 0xff, it
works.
3.3* Googled on internet, one guy said he ever checked bare metal
machine and vmware system, values are all 0x10f.
So agree with Eric that we should keep it as is since the value has been there for so long time and no one met issue about it, and no confirmative to support what value we should take, except of the example in MP-Spec.
arch/x86/kernel/traps.c +2 -0
@@ -942,6 +942,8 @@ dotraplinkage void do_iret_error(struct pt_regs *regs, long error_code)void __init trap_init (void ) { +unsigned int v =apic_read (APIC_SPIV ); +pr_info ("... APIC SPIV: %08x\n" , v);/* Init cpu_entry_area before IST entries are set up */ setup_cpu_entry_areas ();
Quoted 17 lines
Quoted 7 lines
arch/x86/kernel/apic/apic.c +0
-0arch/x86/kernel/apic/apic.c void disconnect_bsp_APIC(int virt_wire_setup) · line 2136
Quoted 8 lines
arch/x86/kernel/apic/apic.c +1
-1@@ -2130,7 +2130,7 @@ void disconnect_bsp_APIC(int virt_wire_setup) value =apic_read (APIC_SPIV ); value &= ~APIC_VECTOR_MASK ; value |=APIC_SPIV_APIC_ENABLED ; - value |=0xf ; + value |=SPURIOUS_APIC_VECTOR ;apic_write (APIC_SPIV , value);if (!virt_wire_setup) {Eric