Message ID | 1421807520-12030-1-git-send-email-cyrilbur@gmail.com (mailing list archive) |
---|---|
State | Accepted |
Commit | 3df76a9dcc74d5f012b94ea01ed6e7aaf8362c5a |
Delegated to: | Michael Ellerman |
Headers | show |
On Wed, 2015-01-21 at 13:32 +1100, Cyril Bur wrote: > The need to handle ibm,suspend_me specially from within ppc_rtas has left an > endian bug exposed as rtas_ibm_suspend_me actually performs HCALLs and should > have its params in CPU endian. That needs a much better explanation. Key points: - ppc_rtas() is a syscall, which takes arguments in BE - ibm,suspend-me is not a real RTAS call and is handled specially in there - ibm,suspend-me is actually implemented by an hcall - there is currently a bug on LE, because rtas_ibm_suspend_me() takes the ppc_rtas() args and feeds them directly to the hcall > Have ppc_rtas send the params correctly and also interpret the result > correctly. That's a second bug which you should also mention above. > Removed the convoluted use of the rtas_args struct to pass params to > rtas_ibm_suspend_me in favour of passing what it needs directly. > > Signed-off-by: Cyril Bur <cyrilbur@gmail.com> > --- > This patch has been tested with KVM both LE and BE and on PowerVM both LE and > BE. Under QEMU/KVM the migration happens without touching the these code > pathes. > For PowerVM there is no obvious regression on BE and the LE code path now > provides the correct parameters to the hypervisor Fold that into the changelog, it's worth remembering. cheers
On Wed, 2015-01-21 at 14:33 +1100, Michael Ellerman wrote: > On Wed, 2015-01-21 at 13:32 +1100, Cyril Bur wrote: > > The need to handle ibm,suspend_me specially from within ppc_rtas has left an > > endian bug exposed as rtas_ibm_suspend_me actually performs HCALLs and should > > have its params in CPU endian. > > That needs a much better explanation. > Agreed > Key points: > - ppc_rtas() is a syscall, which takes arguments in BE > - ibm,suspend-me is not a real RTAS call and is handled specially in there > - ibm,suspend-me is actually implemented by an hcall > - there is currently a bug on LE, because rtas_ibm_suspend_me() takes the > ppc_rtas() args and feeds them directly to the hcall > I've tried to write that out neatly and orderly. Here's how that went: RTAS events require arguments be passed in big endian while hypercalls have their arguments passed in registers and the values should therefore be in CPU endian. The ibm,suspend_me 'RTAS' call makes a sequence of hypercalls to setup one true RTAS call. This means that ibm,suspend_me is handled specially in the ppc_rtas syscall. The ppc_rtas syscall has its arguments in big endian and can therefore pass these arguments directly to the rtas call. ibm,suspend_me is handled specially from within ppc_rtas (by calling rtas_ibm_suspend_me) which has left an endian bug on little endian systems due to the requirement of hypercalls. The return value from rtas_ibm_suspend me gets returned in cpu endian, and is left unconverted, also a bug on little endian systems. rtas_ibm_suspend_me does not actually make use of the rtas_args that it is passed. This patch removes the convoluted use of the rtas_args struct to pass params to rtas_ibm_suspend_me in favour of passing what it needs as actual arguments. This patch also ensures the two callers of rtas_ibm_suspend_me pass function parameters in cpu endian and in the case of ppc_rtas, converts the return value. migrate_store (the other caller of rtas_ibm_suspend_me) is from a sysfs file which deals with everything in cpu endian so this function only underwent cleanup. > > Have ppc_rtas send the params correctly and also interpret the result > > correctly. > > That's a second bug which you should also mention above. > > > Removed the convoluted use of the rtas_args struct to pass params to > > rtas_ibm_suspend_me in favour of passing what it needs directly. > > > > Signed-off-by: Cyril Bur <cyrilbur@gmail.com> > > --- > > This patch has been tested with KVM both LE and BE and on PowerVM both LE and > > BE. Under QEMU/KVM the migration happens without touching the these code > > pathes. > > For PowerVM there is no obvious regression on BE and the LE code path now > > provides the correct parameters to the hypervisor > > Fold that into the changelog, it's worth remembering. > > cheers > >
On Thu, 2015-01-22 at 16:37 +1100, Cyril Bur wrote: > On Wed, 2015-01-21 at 14:33 +1100, Michael Ellerman wrote: > > On Wed, 2015-01-21 at 13:32 +1100, Cyril Bur wrote: > > > The need to handle ibm,suspend_me specially from within ppc_rtas has left an > > > endian bug exposed as rtas_ibm_suspend_me actually performs HCALLs and should > > > have its params in CPU endian. > > > > That needs a much better explanation. > I've tried to write that out neatly and orderly. Here's how that went: Yep that's much better. Minor points of style that I like are to use "()" on functions to make it clear, eg. ppc_rtas(), and to put strings in quotes, eg. "ibm,suspend_me". I've fixed it up and incorporated your testing notes, end result is: Author: cyrilbur@gmail.com <cyrilbur@gmail.com> Date: Wed Jan 21 13:32:00 2015 +1100 powerpc/pseries: Fix endian problems with LE migration RTAS events require arguments be passed in big endian while hypercalls have their arguments passed in registers and the values should therefore be in CPU endian. The "ibm,suspend_me" 'RTAS' call makes a sequence of hypercalls to setup one true RTAS call. This means that "ibm,suspend_me" is handled specially in the ppc_rtas() syscall. The ppc_rtas() syscall has its arguments in big endian and can therefore pass these arguments directly to the RTAS call. "ibm,suspend_me" is handled specially from within ppc_rtas() (by calling rtas_ibm_suspend_me()) which has left an endian bug on little endian systems due to the requirement of hypercalls. The return value from rtas_ibm_suspend_me() gets returned in cpu endian, and is left unconverted, also a bug on little endian systems. rtas_ibm_suspend_me() does not actually make use of the rtas_args that it is passed. This patch removes the convoluted use of the rtas_args struct to pass params to rtas_ibm_suspend_me() in favour of passing what it needs as actual arguments. This patch also ensures the two callers of rtas_ibm_suspend_me() pass function parameters in cpu endian and in the case of ppc_rtas(), converts the return value. migrate_store() (the other caller of rtas_ibm_suspend_me()) is from a sysfs file which deals with everything in cpu endian so this function only underwent cleanup. This patch has been tested with KVM both LE and BE and on PowerVM both LE and BE. Under QEMU/KVM the migration happens without touching these code pathes. For PowerVM there is no obvious regression on BE and the LE code path now provides the correct parameters to the hypervisor. Signed-off-by: Cyril Bur <cyrilbur@gmail.com> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au> cheers
diff --git a/arch/powerpc/include/asm/rtas.h b/arch/powerpc/include/asm/rtas.h index b390f55..2e23e92 100644 --- a/arch/powerpc/include/asm/rtas.h +++ b/arch/powerpc/include/asm/rtas.h @@ -327,7 +327,7 @@ extern int rtas_suspend_cpu(struct rtas_suspend_me_data *data); extern int rtas_suspend_last_cpu(struct rtas_suspend_me_data *data); extern int rtas_online_cpus_mask(cpumask_var_t cpus); extern int rtas_offline_cpus_mask(cpumask_var_t cpus); -extern int rtas_ibm_suspend_me(struct rtas_args *); +extern int rtas_ibm_suspend_me(u64 handle, int *vasi_return); struct rtc_time; extern unsigned long rtas_get_boot_time(void); diff --git a/arch/powerpc/kernel/rtas.c b/arch/powerpc/kernel/rtas.c index 4af905e..21c45a2 100644 --- a/arch/powerpc/kernel/rtas.c +++ b/arch/powerpc/kernel/rtas.c @@ -897,7 +897,7 @@ int rtas_offline_cpus_mask(cpumask_var_t cpus) } EXPORT_SYMBOL(rtas_offline_cpus_mask); -int rtas_ibm_suspend_me(struct rtas_args *args) +int rtas_ibm_suspend_me(u64 handle, int *vasi_return) { long state; long rc; @@ -911,8 +911,7 @@ int rtas_ibm_suspend_me(struct rtas_args *args) return -ENOSYS; /* Make sure the state is valid */ - rc = plpar_hcall(H_VASI_STATE, retbuf, - ((u64)args->args[0] << 32) | args->args[1]); + rc = plpar_hcall(H_VASI_STATE, retbuf, handle); state = retbuf[0]; @@ -920,12 +919,12 @@ int rtas_ibm_suspend_me(struct rtas_args *args) printk(KERN_ERR "rtas_ibm_suspend_me: vasi_state returned %ld\n",rc); return rc; } else if (state == H_VASI_ENABLED) { - args->args[args->nargs] = RTAS_NOT_SUSPENDABLE; + *vasi_return = RTAS_NOT_SUSPENDABLE; return 0; } else if (state != H_VASI_SUSPENDING) { printk(KERN_ERR "rtas_ibm_suspend_me: vasi_state returned state %ld\n", state); - args->args[args->nargs] = -1; + *vasi_return = -1; return 0; } @@ -973,7 +972,7 @@ out: return atomic_read(&data.error); } #else /* CONFIG_PPC_PSERIES */ -int rtas_ibm_suspend_me(struct rtas_args *args) +int rtas_ibm_suspend_me(u64 handle, int *vasi_return) { return -ENOSYS; } @@ -1053,7 +1052,16 @@ asmlinkage int ppc_rtas(struct rtas_args __user *uargs) /* Need to handle ibm,suspend_me call specially */ if (token == ibm_suspend_me_token) { - rc = rtas_ibm_suspend_me(&args); + + /* + * rtas_ibm_suspend_me assumes args are in cpu endian, or at least the + * hcall within it requires it. + */ + int vasi_rc = 0; + u64 handle = ((u64)be32_to_cpu(args.args[0]) << 32) + | be32_to_cpu(args.args[1]); + rc = rtas_ibm_suspend_me(handle, &vasi_rc); + args.rets[0] = cpu_to_be32(vasi_rc); if (rc) return rc; goto copy_return; diff --git a/arch/powerpc/platforms/pseries/mobility.c b/arch/powerpc/platforms/pseries/mobility.c index e7cb6d4..90cf3dc 100644 --- a/arch/powerpc/platforms/pseries/mobility.c +++ b/arch/powerpc/platforms/pseries/mobility.c @@ -316,34 +316,24 @@ void post_mobility_fixup(void) static ssize_t migrate_store(struct class *class, struct class_attribute *attr, const char *buf, size_t count) { - struct rtas_args args; u64 streamid; int rc; + int vasi_rc = 0; rc = kstrtou64(buf, 0, &streamid); if (rc) return rc; - memset(&args, 0, sizeof(args)); - args.token = rtas_token("ibm,suspend-me"); - args.nargs = 2; - args.nret = 1; - - args.args[0] = streamid >> 32 ; - args.args[1] = streamid & 0xffffffff; - args.rets = &args.args[args.nargs]; - do { - args.rets[0] = 0; - rc = rtas_ibm_suspend_me(&args); - if (!rc && args.rets[0] == RTAS_NOT_SUSPENDABLE) + rc = rtas_ibm_suspend_me(streamid, &vasi_rc); + if (!rc && vasi_rc == RTAS_NOT_SUSPENDABLE) ssleep(1); - } while (!rc && args.rets[0] == RTAS_NOT_SUSPENDABLE); + } while (!rc && vasi_rc == RTAS_NOT_SUSPENDABLE); if (rc) return rc; - else if (args.rets[0]) - return args.rets[0]; + if (vasi_rc) + return vasi_rc; post_mobility_fixup(); return count;
The need to handle ibm,suspend_me specially from within ppc_rtas has left an endian bug exposed as rtas_ibm_suspend_me actually performs HCALLs and should have its params in CPU endian. Have ppc_rtas send the params correctly and also interpret the result correctly. Removed the convoluted use of the rtas_args struct to pass params to rtas_ibm_suspend_me in favour of passing what it needs directly. Signed-off-by: Cyril Bur <cyrilbur@gmail.com> --- This patch has been tested with KVM both LE and BE and on PowerVM both LE and BE. Under QEMU/KVM the migration happens without touching the these code pathes. For PowerVM there is no obvious regression on BE and the LE code path now provides the correct parameters to the hypervisor --- arch/powerpc/include/asm/rtas.h | 2 +- arch/powerpc/kernel/rtas.c | 22 +++++++++++++++------- arch/powerpc/platforms/pseries/mobility.c | 22 ++++++---------------- 3 files changed, 22 insertions(+), 24 deletions(-)