"Nicholas Piggin" <npig...@gmail.com> writes: > On Wed Sep 14, 2022 at 3:39 AM AEST, Leonardo Brás wrote: >> On Mon, 2022-09-12 at 14:58 -0500, Nathan Lynch wrote: >> > Leonardo Brás <leobra...@gmail.com> writes: >> > > On Fri, 2022-09-09 at 09:04 -0500, Nathan Lynch wrote: >> > > > Leonardo Brás <leobra...@gmail.com> writes: >> > > > > On Wed, 2022-09-07 at 17:01 -0500, Nathan Lynch wrote: >> > > > > > At the time this was submitted by Leonardo, I confirmed -- or >> > > > > > thought >> > > > > > I had confirmed -- with PowerVM partition firmware development that >> > > > > > the following RTAS functions: >> > > > > > >> > > > > > - ibm,get-xive >> > > > > > - ibm,int-off >> > > > > > - ibm,int-on >> > > > > > - ibm,set-xive >> > > > > > >> > > > > > were safe to call on multiple CPUs simultaneously, not only with >> > > > > > respect to themselves as indicated by PAPR, but with arbitrary >> > > > > > other >> > > > > > RTAS calls: >> > > > > > >> > > > > > https://lore.kernel.org/linuxppc-dev/875zcy2v8o....@linux.ibm.com/ >> > > > > > >> > > > > > Recent discussion with firmware development makes it clear that >> > > > > > this >> > > > > > is not true, and that the code in commit b664db8e3f97 >> > > > > > ("powerpc/rtas: >> > > > > > Implement reentrant rtas call") is unsafe, likely explaining >> > > > > > several >> > > > > > strange bugs we've seen in internal testing involving DLPAR and >> > > > > > LPM. These scenarios use ibm,configure-connector, whose internal >> > > > > > state >> > > > > > can be corrupted by the concurrent use of the "reentrant" >> > > > > > functions, >> > > > > > leading to symptoms like endless busy statuses from RTAS. >> > > > > >> > > > > Oh, does not it means PowerVM is not compliant to the PAPR specs? >> > > > >> > > > No, it means the premise of commit b664db8e3f97 ("powerpc/rtas: >> > > > Implement reentrant rtas call") change is incorrect. The "reentrant" >> > > > property described in the spec applies only to the individual RTAS >> > > > functions. The OS can invoke (for example) ibm,set-xive on multiple >> > > > CPUs >> > > > simultaneously, but it must adhere to the more general requirement to >> > > > serialize with other RTAS functions. >> > > > >> > > >> > > I see. Thanks for explaining that part! >> > > I agree: reentrant calls that way don't look as useful on Linux than I >> > > previously thought. >> > > >> > > OTOH, I think that instead of reverting the change, we could make use of >> > > the >> > > correct information and fix the current implementation. (This could help >> > > when we >> > > do the same rtas call in multiple cpus) >> > >> > Hmm I'm happy to be mistaken here, but I doubt we ever really need to do >> > that. I'm not seeing the need. >> > >> > > I have an idea of a patch to fix this. >> > > Do you think it would be ok if I sent that, to prospect being an >> > > alternative to >> > > this reversion? >> > >> > It is my preference, and I believe it is more common, to revert to the >> > well-understood prior state, imperfect as it may be. The revert can be >> > backported to -stable and distros while development and review of >> > another approach proceeds. >> >> Ok then, as long as you are aware of the kdump bug, I'm good. >> >> FWIW: >> Reviewed-by: Leonardo Bras <leobra...@gmail.com> > > A shame. I guess a reader/writer lock would not be much help because > the crash is probably more likely to hit longer running rtas calls? > > Alternative is just cheat and do this...? > > Thanks, > Nick > > diff --git a/arch/powerpc/kernel/rtas.c b/arch/powerpc/kernel/rtas.c > index 693133972294..89728714a06e 100644 > --- a/arch/powerpc/kernel/rtas.c > +++ b/arch/powerpc/kernel/rtas.c > @@ -26,6 +26,7 @@ > #include <linux/syscalls.h> > #include <linux/of.h> > #include <linux/of_fdt.h> > +#include <linux/panic.h> > > #include <asm/interrupt.h> > #include <asm/rtas.h> > @@ -97,6 +98,19 @@ static unsigned long lock_rtas(void) > { > unsigned long flags; > > + if (atomic_read(&panic_cpu) == raw_smp_processor_id()) { > + /* > + * Crash in progress on this CPU. Other CPUs should be > + * stopped by now, so skip the lock in case it was being > + * held, and is now needed for crashing e.g., kexec > + * (machine_kexec_mask_interrupts) requires rtas calls. > + * > + * It's possible this could have caused rtas state > breakage > + * but the alternative is deadlock. > + */ > + return 0; > + } > + > local_irq_save(flags); > preempt_disable(); > arch_spin_lock(&rtas.lock); > @@ -105,6 +119,9 @@ static unsigned long lock_rtas(void) > > static void unlock_rtas(unsigned long flags) > { > + if (atomic_read(&panic_cpu) == raw_smp_processor_id()) > + return; > + > arch_spin_unlock(&rtas.lock); > local_irq_restore(flags); > preempt_enable();
Looks correct. I wonder - would it be worth making the panic path use a separate "emergency" rtas_args buffer as well? If a CPU is actually "stuck" in RTAS at panic time, then leaving rtas.args untouched might make the ibm,int-off, ibm,set-xive, ibm,os-term, and any other RTAS calls we incur on the panic path more likely to succeed. Building on yours, something like (sorry, it's ugly): diff --git a/arch/powerpc/kernel/rtas.c b/arch/powerpc/kernel/rtas.c index 693133972294..4865d26e7391 100644 --- a/arch/powerpc/kernel/rtas.c +++ b/arch/powerpc/kernel/rtas.c @@ -73,6 +73,8 @@ struct rtas_t rtas = { }; EXPORT_SYMBOL(rtas); +static struct rtas_args emergency_rtas_args; + DEFINE_SPINLOCK(rtas_data_buf_lock); EXPORT_SYMBOL(rtas_data_buf_lock); @@ -93,20 +95,24 @@ EXPORT_SYMBOL(rtas_flash_term_hook); * such as having the timebase stopped which would lockup with * normal locks and spinlock debugging enabled */ -static unsigned long lock_rtas(void) +static struct rtas_args *lock_rtas(unsigned long *flags) { - unsigned long flags; + if (atomic_read(&panic_cpu) == raw_smp_processor_id()) + return &emergency_rtas_args; - local_irq_save(flags); + local_irq_save(*flags); preempt_disable(); arch_spin_lock(&rtas.lock); - return flags; + return &rtas.args; } -static void unlock_rtas(unsigned long flags) +static void unlock_rtas(struct rtas_args *args, unsigned long *flags) { + if (atomic_read(&panic_cpu) == raw_smp_processor_id()) + return; + arch_spin_unlock(&rtas.lock); - local_irq_restore(flags); + local_irq_restore(*flags); preempt_enable(); } @@ -117,14 +123,15 @@ static void unlock_rtas(unsigned long flags) */ static void call_rtas_display_status(unsigned char c) { - unsigned long s; + struct rtas_args *args; + unsigned long flags; if (!rtas.base) return; - s = lock_rtas(); - rtas_call_unlocked(&rtas.args, 10, 1, 1, NULL, c); - unlock_rtas(s); + args = lock_rtas(&flags); + rtas_call_unlocked(args, 10, 1, 1, NULL, c); + unlock_rtas(args, &flags); } static void call_rtas_display_status_delay(char c) @@ -468,7 +475,7 @@ int rtas_call(int token, int nargs, int nret, int *outputs, ...) { va_list list; int i; - unsigned long s; + unsigned long flags; struct rtas_args *rtas_args; char *buff_copy = NULL; int ret; @@ -481,10 +488,7 @@ int rtas_call(int token, int nargs, int nret, int *outputs, ...) return -1; } - s = lock_rtas(); - - /* We use the global rtas args buffer */ - rtas_args = &rtas.args; + rtas_args = lock_rtas(&flags); va_start(list, outputs); va_rtas_call_unlocked(rtas_args, token, nargs, nret, list); @@ -500,7 +504,7 @@ int rtas_call(int token, int nargs, int nret, int *outputs, ...) outputs[i] = be32_to_cpu(rtas_args->rets[i+1]); ret = (nret > 0)? be32_to_cpu(rtas_args->rets[0]): 0; - unlock_rtas(s); + unlock_rtas(rtas_args, &flags); if (buff_copy) { log_error(buff_copy, ERR_TYPE_RTAS_LOG, 0); @@ -1190,6 +1194,7 @@ static void __init rtas_syscall_filter_init(void) /* We assume to be passed big endian arguments */ SYSCALL_DEFINE1(rtas, struct rtas_args __user *, uargs) { + struct rtas_args *argsp; struct rtas_args args; unsigned long flags; char *buff_copy, *errbuf = NULL; @@ -1249,18 +1254,18 @@ SYSCALL_DEFINE1(rtas, struct rtas_args __user *, uargs) buff_copy = get_errorlog_buffer(); - flags = lock_rtas(); + argsp = lock_rtas(&flags); - rtas.args = args; - do_enter_rtas(__pa(&rtas.args)); - args = rtas.args; + *argsp = args; + do_enter_rtas(__pa(argsp)); + args = *argsp; /* A -1 return code indicates that the last command couldn't be completed due to a hardware error. */ if (be32_to_cpu(args.rets[0]) == -1) errbuf = __fetch_rtas_last_error(buff_copy); - unlock_rtas(flags); + unlock_rtas(argsp, &flags); if (buff_copy) { if (errbuf)