Message ID | 20230822010933.2247671-1-ushankar@purestorage.com (mailing list archive) |
---|---|
State | Handled Elsewhere |
Headers | show |
Series | Revert "ACPI, APEI, use raw spinlock in ERST" | expand |
Hi, Uday, Uday Shankar <ushankar@purestorage.com> writes: > The commit > > commit 3b38bb5f7f06356a89bb2cbf92ad346025e95192 > Author: Huang Ying <ying.huang@intel.com> > Date: Thu Dec 2 10:40:53 2010 +0800 > > ACPI, APEI, use raw spinlock in ERST > > ERST writing may be used in NMI or Machine Check Exception handler. So > it need to use raw spinlock instead of normal spinlock. This patch > fixes it. > > Signed-off-by: Huang Ying <ying.huang@intel.com> > Signed-off-by: Len Brown <len.brown@intel.com> > > turns erst_lock from a regular spinlock into a raw spinlock. This change > does not help address the fundamental problem of locking in NMI handlers > - NMIs can interrupt anything, including the critical section of the > lock the handler wants to acquire. The original commit cannot resolve the nested erst_lock issue, but that's not the intention of the commit too. Let me provide some background. ERST is mainly used to log the hardware error. While, hardware error may be reported via NMI (e.g., ACPI APEI GHES NMI), so we need to call ERST functions in NMI handlers. Where normal spinlock cannot be used because they will be converted to sleepable rt_mutex in RT kernel. Yes. This isn't a perfect solution. But it works in most cases in practice. Or, do you have some better solution? -- Best Regards, Huang, Ying > The change does, however, introduce > an illegal lock nesting pattern - while holding the erst_lock, there is > a code path which calls printk, which may acquire a regular spinlock. > This issue was caught by lockdep, and the report follows. Revert > 3b38bb5f7f06 to address the issue. > > [ BUG: Invalid wait context ] > 6.5.0-iopolicies+ #4 Not tainted > ----------------------------- > systemd/1 is trying to lock: > ffffffff9ec053b8 (&port_lock_key){....}-{3:3}, at: serial8250_console_write+0x325/0x440 > other info that might help us debug this: > context-{5:5} > 7 locks held by systemd/1: > stack backtrace: > CPU: 2 PID: 1 Comm: systemd Not tainted 6.5.0-iopolicies+ #4 > Hardware name: /0JP31P, BIOS 2.7.0 05/23/2018 > Call Trace: > <TASK> > dump_stack_lvl+0x63/0xb0 > __lock_acquire+0x434/0xd30 > lock_acquire+0xcd/0x2b0 > ? serial8250_console_write+0x325/0x440 > ? __lock_acquire+0xab0/0xd30 > _raw_spin_lock_irqsave+0x39/0x60 > ? serial8250_console_write+0x325/0x440 > serial8250_console_write+0x325/0x440 > ? console_flush_all+0x1b5/0x590 > console_flush_all+0x1ec/0x590 > ? console_flush_all+0x184/0x590 > console_unlock+0x7c/0xd0 > vprintk_emit+0x261/0x380 > _printk+0x5c/0x80 > erst_read+0x1f8/0x210 > ? erst_get_record_id_next+0x26f/0x340 > erst_read_record+0x29/0xa0 > erst_reader+0xad/0x2e0 > ? __pfx_read_tsc+0x10/0x10 > ? ktime_get_real_fast_ns+0x47/0xa0 > ? __pfx_pstore_fill_super+0x10/0x10 > pstore_get_backend_records+0xc6/0x2d0 > ? __pfx_pstore_fill_super+0x10/0x10 > pstore_get_records+0x2b/0x50 > pstore_fill_super+0xb7/0xd0 > mount_single+0x90/0xc0 > ? security_capable+0x3c/0x60 > legacy_get_tree+0x34/0x60 > vfs_get_tree+0x2c/0xe0 > path_mount+0x703/0xa90 > do_mount+0x7d/0xa0 > __x64_sys_mount+0xdc/0x100 > do_syscall_64+0x3e/0x90 > entry_SYSCALL_64_after_hwframe+0x6e/0xd8 > RIP: 0033:0x7f3fdafa42be > Code: 48 8b 0d cd 4b 38 00 f7 d8 64 89 01 48 83 c8 ff c3 66 2e 0f 1f 84 00 00 00 00 00 90 f3 0f 1e fa 49 89 ca b8 a5 00 00 00 0f 05 <48> 3d 01 f0 ff ff 73 01 c3 48 8b 0d 9a 4b 38 00 f7 d8 64 89 01 48 > RSP: 002b:00007ffdf2b2be08 EFLAGS: 00000246 ORIG_RAX: 00000000000000a5 > RAX: ffffffffffffffda RBX: 000055b567f49e78 RCX: 00007f3fdafa42be > RDX: 000055b567d0294a RSI: 000055b567d02942 RDI: 000055b567d0294a > RBP: 0000000000000007 R08: 0000000000000000 R09: 0000000000000002 > R10: 000000000000000e R11: 0000000000000246 R12: 0000000000000000 > R13: 0000000000000000 R14: 0000000000000000 R15: 00007ffdf2b2c050 > </TASK> > > Signed-off-by: Uday Shankar <ushankar@purestorage.com> > Reviewed-by: Joern Engel <joern@purestorage.com> > --- > drivers/acpi/apei/erst.c | 26 +++++++++++++------------- > 1 file changed, 13 insertions(+), 13 deletions(-) > > diff --git a/drivers/acpi/apei/erst.c b/drivers/acpi/apei/erst.c > index 247989060e29..b2c806840a11 100644 > --- a/drivers/acpi/apei/erst.c > +++ b/drivers/acpi/apei/erst.c > @@ -78,7 +78,7 @@ static struct erst_erange { > * It is used to provide exclusive accessing for ERST Error Log > * Address Range too. > */ > -static DEFINE_RAW_SPINLOCK(erst_lock); > +static DEFINE_SPINLOCK(erst_lock); > > static inline int erst_errno(int command_status) > { > @@ -413,9 +413,9 @@ ssize_t erst_get_record_count(void) > if (erst_disable) > return -ENODEV; > > - raw_spin_lock_irqsave(&erst_lock, flags); > + spin_lock_irqsave(&erst_lock, flags); > count = __erst_get_record_count(); > - raw_spin_unlock_irqrestore(&erst_lock, flags); > + spin_unlock_irqrestore(&erst_lock, flags); > > return count; > } > @@ -480,9 +480,9 @@ static int __erst_record_id_cache_add_one(void) > > id = prev_id = first_id = APEI_ERST_INVALID_RECORD_ID; > retry: > - raw_spin_lock_irqsave(&erst_lock, flags); > + spin_lock_irqsave(&erst_lock, flags); > rc = __erst_get_next_record_id(&id); > - raw_spin_unlock_irqrestore(&erst_lock, flags); > + spin_unlock_irqrestore(&erst_lock, flags); > if (rc == -ENOENT) > return 0; > if (rc) > @@ -777,17 +777,17 @@ int erst_write(const struct cper_record_header *record) > return -EINVAL; > > if (erst_erange.attr & ERST_RANGE_NVRAM) { > - if (!raw_spin_trylock_irqsave(&erst_lock, flags)) > + if (!spin_trylock_irqsave(&erst_lock, flags)) > return -EBUSY; > rc = __erst_write_to_nvram(record); > - raw_spin_unlock_irqrestore(&erst_lock, flags); > + spin_unlock_irqrestore(&erst_lock, flags); > return rc; > } > > if (record->record_length > erst_erange.size) > return -EINVAL; > > - if (!raw_spin_trylock_irqsave(&erst_lock, flags)) > + if (!spin_trylock_irqsave(&erst_lock, flags)) > return -EBUSY; > memcpy(erst_erange.vaddr, record, record->record_length); > rcd_erange = erst_erange.vaddr; > @@ -795,7 +795,7 @@ int erst_write(const struct cper_record_header *record) > memcpy(&rcd_erange->persistence_information, "ER", 2); > > rc = __erst_write_to_storage(0); > - raw_spin_unlock_irqrestore(&erst_lock, flags); > + spin_unlock_irqrestore(&erst_lock, flags); > > return rc; > } > @@ -849,9 +849,9 @@ ssize_t erst_read(u64 record_id, struct cper_record_header *record, > if (erst_disable) > return -ENODEV; > > - raw_spin_lock_irqsave(&erst_lock, flags); > + spin_lock_irqsave(&erst_lock, flags); > len = __erst_read(record_id, record, buflen); > - raw_spin_unlock_irqrestore(&erst_lock, flags); > + spin_unlock_irqrestore(&erst_lock, flags); > return len; > } > EXPORT_SYMBOL_GPL(erst_read); > @@ -936,12 +936,12 @@ int erst_clear(u64 record_id) > rc = mutex_lock_interruptible(&erst_record_id_cache.lock); > if (rc) > return rc; > - raw_spin_lock_irqsave(&erst_lock, flags); > + spin_lock_irqsave(&erst_lock, flags); > if (erst_erange.attr & ERST_RANGE_NVRAM) > rc = __erst_clear_from_nvram(record_id); > else > rc = __erst_clear_from_storage(record_id); > - raw_spin_unlock_irqrestore(&erst_lock, flags); > + spin_unlock_irqrestore(&erst_lock, flags); > if (rc) > goto out; > entries = erst_record_id_cache.entries;
On Tue, Aug 22, 2023 at 09:56:38AM +0800, Huang, Ying wrote: > > ERST is mainly used to log the hardware error. While, hardware error > may be reported via NMI (e.g., ACPI APEI GHES NMI), so we need to call > ERST functions in NMI handlers. Where normal spinlock cannot be used > because they will be converted to sleepable rt_mutex in RT kernel. Non-sleeping spinlocks cannot be used in NMI context either. raw_spin_lock_irqsave() will prevent regular interrupts, but not NMI. So taking a spinlock inside an NMI can trigger a deadlock. Am I missing something here? Jörn -- All art is but imitation of nature. -- Lucius Annaeus Seneca
On Mon, Aug 21, 2023 at 08:16:05PM -0700, Jörn Engel wrote: > On Tue, Aug 22, 2023 at 09:56:38AM +0800, Huang, Ying wrote: > > > > ERST is mainly used to log the hardware error. While, hardware error > > may be reported via NMI (e.g., ACPI APEI GHES NMI), so we need to call > > ERST functions in NMI handlers. Where normal spinlock cannot be used > > because they will be converted to sleepable rt_mutex in RT kernel. > > Non-sleeping spinlocks cannot be used in NMI context either. > raw_spin_lock_irqsave() will prevent regular interrupts, but not NMI. > So taking a spinlock inside an NMI can trigger a deadlock. > > Am I missing something here? > > Jörn > > -- > All art is but imitation of nature. > -- Lucius Annaeus Seneca Also want to point out that both before and after this commit, we only use trylock from erst_write, which looks like the only function touched by this patch which can be called from NMI context. Trylock should be safe in NMI context regardless of the type of synchronization used (raw_spinlock, spinlock, or otherwise).
Jrn Engel <joern@purestorage.com> writes: > On Tue, Aug 22, 2023 at 09:56:38AM +0800, Huang, Ying wrote: >> >> ERST is mainly used to log the hardware error. While, hardware error >> may be reported via NMI (e.g., ACPI APEI GHES NMI), so we need to call >> ERST functions in NMI handlers. Where normal spinlock cannot be used >> because they will be converted to sleepable rt_mutex in RT kernel. > > Non-sleeping spinlocks cannot be used in NMI context either. > raw_spin_lock_irqsave() will prevent regular interrupts, but not NMI. > So taking a spinlock inside an NMI can trigger a deadlock. Thanks reminding from Uday, trylock can avoid the deadlock. More details in reply to Uday's email. -- Best Regards, Huang, Ying > Am I missing something here? > > Jrn > > -- > All art is but imitation of nature. > -- Lucius Annaeus Seneca
Uday Shankar <ushankar@purestorage.com> writes: > On Mon, Aug 21, 2023 at 08:16:05PM -0700, Jörn Engel wrote: >> On Tue, Aug 22, 2023 at 09:56:38AM +0800, Huang, Ying wrote: >> > >> > ERST is mainly used to log the hardware error. While, hardware error >> > may be reported via NMI (e.g., ACPI APEI GHES NMI), so we need to call >> > ERST functions in NMI handlers. Where normal spinlock cannot be used >> > because they will be converted to sleepable rt_mutex in RT kernel. >> >> Non-sleeping spinlocks cannot be used in NMI context either. >> raw_spin_lock_irqsave() will prevent regular interrupts, but not NMI. >> So taking a spinlock inside an NMI can trigger a deadlock. >> >> Am I missing something here? >> >> Jörn >> >> -- >> All art is but imitation of nature. >> -- Lucius Annaeus Seneca > > Also want to point out that both before and after this commit, we only > use trylock from erst_write, which looks like the only function touched > by this patch which can be called from NMI context. Trylock should be > safe in NMI context regardless of the type of synchronization used > (raw_spinlock, spinlock, or otherwise). Thanks for reminding! That's a good point. Checked the implementation of rt_mutex version of spin_trylock(). One possible code path is, spin_trylock() rt_spin_trylock() __rt_spin_trylock() rt_mutex_slowtrylock() raw_spin_lock_irqsave() IIUC, the deadlock is still possible for rt_mutex. While it seems that the deadlock isn't possible for raw_spinlock. If so, it's still better to use raw_spinlock. -- Best Regards, Huang, Ying
On Wed, Aug 23, 2023 at 10:38:23AM +0800, Huang, Ying wrote: > > Thanks for reminding! That's a good point. > > Checked the implementation of rt_mutex version of spin_trylock(). One > possible code path is, > > spin_trylock() > rt_spin_trylock() > __rt_spin_trylock() > rt_mutex_slowtrylock() > raw_spin_lock_irqsave() You are right. I'm actually quite surprised how we turn a trylock into a spinning lock. Now I know, thank you! Jörn -- "Security vulnerabilities are here to stay." -- Scott Culp, Manager of the Microsoft Security Response Center, 2001
On Tue, Aug 22, 2023 at 07:57:52PM -0700, Jörn Engel wrote: > > You are right. I'm actually quite surprised how we turn a trylock into > a spinning lock. Now I know, thank you! On that subject, why do we turn a trylock into a spinning lock? Thomas, is there a good argument against a patch like the one below? If the question was ever considered, it seems to have happened outside of mainline. At least I cannot find it in git history. Jörn -- If you cannot make a function run 100x faster, maybe you can call it 100x less often. -- Unknown diff --git a/kernel/locking/rtmutex.c b/kernel/locking/rtmutex.c index fcd9ad3f7f2e..5bda84d5879a 100644 --- a/kernel/locking/rtmutex.c +++ b/kernel/locking/rtmutex.c @@ -1255,11 +1255,13 @@ static int __sched rt_mutex_slowtrylock(struct rt_mutex_base *lock) /* * The mutex has currently no owner. Lock the wait lock and try to * acquire the lock. We use irqsave here to support early boot calls. */ - raw_spin_lock_irqsave(&lock->wait_lock, flags); + ret = raw_spin_trylock_irqsave(&lock->wait_lock, flags); + if (!ret) + return 0; ret = __rt_mutex_slowtrylock(lock); raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
diff --git a/drivers/acpi/apei/erst.c b/drivers/acpi/apei/erst.c index 247989060e29..b2c806840a11 100644 --- a/drivers/acpi/apei/erst.c +++ b/drivers/acpi/apei/erst.c @@ -78,7 +78,7 @@ static struct erst_erange { * It is used to provide exclusive accessing for ERST Error Log * Address Range too. */ -static DEFINE_RAW_SPINLOCK(erst_lock); +static DEFINE_SPINLOCK(erst_lock); static inline int erst_errno(int command_status) { @@ -413,9 +413,9 @@ ssize_t erst_get_record_count(void) if (erst_disable) return -ENODEV; - raw_spin_lock_irqsave(&erst_lock, flags); + spin_lock_irqsave(&erst_lock, flags); count = __erst_get_record_count(); - raw_spin_unlock_irqrestore(&erst_lock, flags); + spin_unlock_irqrestore(&erst_lock, flags); return count; } @@ -480,9 +480,9 @@ static int __erst_record_id_cache_add_one(void) id = prev_id = first_id = APEI_ERST_INVALID_RECORD_ID; retry: - raw_spin_lock_irqsave(&erst_lock, flags); + spin_lock_irqsave(&erst_lock, flags); rc = __erst_get_next_record_id(&id); - raw_spin_unlock_irqrestore(&erst_lock, flags); + spin_unlock_irqrestore(&erst_lock, flags); if (rc == -ENOENT) return 0; if (rc) @@ -777,17 +777,17 @@ int erst_write(const struct cper_record_header *record) return -EINVAL; if (erst_erange.attr & ERST_RANGE_NVRAM) { - if (!raw_spin_trylock_irqsave(&erst_lock, flags)) + if (!spin_trylock_irqsave(&erst_lock, flags)) return -EBUSY; rc = __erst_write_to_nvram(record); - raw_spin_unlock_irqrestore(&erst_lock, flags); + spin_unlock_irqrestore(&erst_lock, flags); return rc; } if (record->record_length > erst_erange.size) return -EINVAL; - if (!raw_spin_trylock_irqsave(&erst_lock, flags)) + if (!spin_trylock_irqsave(&erst_lock, flags)) return -EBUSY; memcpy(erst_erange.vaddr, record, record->record_length); rcd_erange = erst_erange.vaddr; @@ -795,7 +795,7 @@ int erst_write(const struct cper_record_header *record) memcpy(&rcd_erange->persistence_information, "ER", 2); rc = __erst_write_to_storage(0); - raw_spin_unlock_irqrestore(&erst_lock, flags); + spin_unlock_irqrestore(&erst_lock, flags); return rc; } @@ -849,9 +849,9 @@ ssize_t erst_read(u64 record_id, struct cper_record_header *record, if (erst_disable) return -ENODEV; - raw_spin_lock_irqsave(&erst_lock, flags); + spin_lock_irqsave(&erst_lock, flags); len = __erst_read(record_id, record, buflen); - raw_spin_unlock_irqrestore(&erst_lock, flags); + spin_unlock_irqrestore(&erst_lock, flags); return len; } EXPORT_SYMBOL_GPL(erst_read); @@ -936,12 +936,12 @@ int erst_clear(u64 record_id) rc = mutex_lock_interruptible(&erst_record_id_cache.lock); if (rc) return rc; - raw_spin_lock_irqsave(&erst_lock, flags); + spin_lock_irqsave(&erst_lock, flags); if (erst_erange.attr & ERST_RANGE_NVRAM) rc = __erst_clear_from_nvram(record_id); else rc = __erst_clear_from_storage(record_id); - raw_spin_unlock_irqrestore(&erst_lock, flags); + spin_unlock_irqrestore(&erst_lock, flags); if (rc) goto out; entries = erst_record_id_cache.entries;