powerpc/64/ftrace: mprofile-kernel patch out mflr
Nicholas Piggin
reviewerRFC PATCH 1/1 powerpc/64/ftrace: mprofile-kernel patch out mflr
The new mprofile-kernel mcount sequence is
mflr r0
bl _mcount
Dynamic ftrace patches the branch instruction with a noop, but leaves the mflr. mflr is executed by the branch unit that can only execute one per cycle on POWER9 and shared with branches, so it would be nice to avoid it where possible.
This patch is a hacky proof of concept to nop out the mflr. Can we do this or are there races or other issues with it?
--- arch/powerpc/kernel/trace/ftrace.c | 77 +++++++++++++++++++++++++++++- 1 file changed, 75 insertions(+), 2 deletions(-)
arch/powerpc/kernel/trace/ftrace.c +75 -2
@@ -172,6 +172,19 @@ __ftrace_make_nop(struct module *mod,pr_err ("Unexpected instruction %08x around bl _mcount\n" , op); return -EINVAL; } + +if (patch_instruction ((unsigned int *)ip, pop)) { +pr_err ("Patching NOP failed.\n" ); +return -EPERM ; + } + +if (op ==PPC_INST_MFLR ) { +if (patch_instruction ((unsigned int *)(ip -4 ), pop)) { +pr_err ("Patching NOP failed.\n" ); +return -EPERM ; + } + } +#else /* * Our original call site looks like:
@@ -202,13 +215,14 @@ __ftrace_make_nop(struct module *mod,pr_err ("Expected %08x found %08x\n" ,PPC_INST_LD_TOC , op); return -EINVAL; } -#endif /* CONFIG_MPROFILE_KERNEL */ if (patch_instruction ((unsigned int *)ip, pop)) {pr_err ("Patching NOP failed.\n" );return -EPERM ; } +#endif /* CONFIG_MPROFILE_KERNEL */ +return 0 ; }
@@ -421,6 +435,20 @@ static int __ftrace_make_nop_kernel(struct dyn_ftrace *rec, unsigned long addr)return -EPERM ; } +#ifdef CONFIG_MPROFILE_KERNEL +if (probe_kernel_read (&op, (void *)(ip -4 ),4 )) { +pr_err ("Fetching instruction at %lx failed.\n" , ip -4 ); +return -EFAULT ; + } + +if (op ==PPC_INST_MFLR ) { +if (patch_instruction ((unsigned int *)(ip -4 ),PPC_INST_NOP )) { +pr_err ("Patching NOP failed.\n" ); +return -EPERM ; + } + } +#endif +return 0 ; }
@@ -437,9 +465,20 @@ int ftrace_make_nop(struct module *mod,*/ if (test_24bit_addr (ip, addr)) {/* within range */ +int rc; + old =ftrace_call_replace (ip, addr,1 ); new =PPC_INST_NOP ; -return ftrace_modify_code (ip, old, new); + rc =ftrace_modify_code (ip, old, new); +if (rc) +return rc; +#ifdef CONFIG_MPROFILE_KERNEL + old =PPC_INST_MFLR ; + new =PPC_INST_NOP ; +ftrace_modify_code (ip -4 , old, new); +/* old mprofile kernel will error because no mflr */ +#endif +return rc; }else if (core_kernel_text (ip))return __ftrace_make_nop_kernel (rec, addr);
@@ -562,6 +601,20 @@ __ftrace_make_call(struct dyn_ftrace *rec, unsigned long addr)return -EINVAL ; } +#ifdef CONFIG_MPROFILE_KERNEL +if (probe_kernel_read (op, (ip -4 ),4 )) { +pr_err ("Fetching instruction at %lx failed.\n" , (unsigned long )(ip -4 )); +return -EFAULT ; + } + +if (op[0 ] ==PPC_INST_NOP ) { +if (patch_instruction ((ip -4 ),PPC_INST_MFLR )) { +pr_err ("Patching mflr failed.\n" ); +return -EINVAL ; + } + } +#endif +if (patch_branch (ip, tramp,BRANCH_SET_LINK )) {pr_err ("REL24 out of range!\n" );return -EINVAL ;
@@ -650,6 +703,20 @@ static int __ftrace_make_call_kernel(struct dyn_ftrace *rec, unsigned long addr)return -EINVAL ; } +#ifdef CONFIG_MPROFILE_KERNEL +if (probe_kernel_read (&op, (ip -4 ),4 )) { +pr_err ("Fetching instruction at %lx failed.\n" , (unsigned long )(ip -4 )); +return -EFAULT ; + } + +if (op ==PPC_INST_NOP ) { +if (patch_instruction ((ip -4 ),PPC_INST_MFLR )) { +pr_err ("Patching mflr failed.\n" ); +return -EINVAL ; + } + } +#endif +if (patch_branch (ip, tramp,BRANCH_SET_LINK )) {pr_err ("Error patching branch to ftrace tramp!\n" );return -EINVAL ;
@@ -670,6 +737,12 @@ int ftrace_make_call(struct dyn_ftrace *rec, unsigned long addr)*/ if (test_24bit_addr (ip, addr)) {/* within range */ +#ifdef CONFIG_MPROFILE_KERNEL + old =PPC_INST_NOP ; + new =PPC_INST_MFLR ; +ftrace_modify_code (ip -4 , old, new); +/* old mprofile-kernel sequence will error because no nop */ +#endif old =PPC_INST_NOP ; new =ftrace_call_replace (ip, addr,1 );return ftrace_modify_code (ip, old, new);
Signature · 1 line
2.20.1
Hi Nick,
Quoted 5 lines · Nicholas Piggin
The new mprofile-kernel mcount sequence is
mflr r0
bl_mcount
Quoted 7 lines
Dynamic ftrace patches the branch instruction with a noop, but leaves the mflr. mflr is executed by the branch unit that can only execute one per cycle on POWER9 and shared with branches, so it would be nice to avoid it where possible.
This patch is a hacky proof of concept to nop out the mflr. Can we do this or are there races or other issues with it?
Thanks for implementing this. I don't see a problem with this.
- Naveen
Michael Ellerman
reviewerQuoted 5 lines · Nicholas Piggin
The new mprofile-kernel mcount sequence is
mflr r0
bl_mcount
Quoted 7 lines
Dynamic ftrace patches the branch instruction with a noop, but leaves the mflr. mflr is executed by the branch unit that can only execute one per cycle on POWER9 and shared with branches, so it would be nice to avoid it where possible.
This patch is a hacky proof of concept to nop out the mflr. Can we do this or are there races or other issues with it?
There's a race, isn't there?
We have a function foo which currently has tracing disabled, so the mflr and bl are nop'ed out.
CPU 0 CPU 1
==================================
bl foo
nop (ie. not mflr)
-> interrupt
something else enable tracing for foo
... patch mflr and branch
<- rfi
bl _mcount
So we end up in _mcount() but with r0 not populated.
Or else what am I missing?
cheers
Quoted 140 lines
arch/powerpc/kernel/trace/ftrace.c +75
-2@@ -172,6 +172,19 @@ __ftrace_make_nop(struct module *mod,pr_err ("Unexpected instruction %08x around bl _mcount\n" , op); return -EINVAL; } + +if (patch_instruction ((unsigned int *)ip, pop)) { +pr_err ("Patching NOP failed.\n" ); +return -EPERM ; + } + +if (op ==PPC_INST_MFLR ) { +if (patch_instruction ((unsigned int *)(ip -4 ), pop)) { +pr_err ("Patching NOP failed.\n" ); +return -EPERM ; + } + } +#else /* * Our original call site looks like: @@ -202,13 +215,14 @@ __ftrace_make_nop(struct module *mod,pr_err ("Expected %08x found %08x\n" ,PPC_INST_LD_TOC , op); return -EINVAL; } -#endif /* CONFIG_MPROFILE_KERNEL */ if (patch_instruction ((unsigned int *)ip, pop)) {pr_err ("Patching NOP failed.\n" );return -EPERM ; } +#endif /* CONFIG_MPROFILE_KERNEL */ +return 0 ; }@@ -421,6 +435,20 @@ static int __ftrace_make_nop_kernel(struct dyn_ftrace *rec, unsigned long addr)return -EPERM ; } +#ifdef CONFIG_MPROFILE_KERNEL +if (probe_kernel_read (&op, (void *)(ip -4 ),4 )) { +pr_err ("Fetching instruction at %lx failed.\n" , ip -4 ); +return -EFAULT ; + } + +if (op ==PPC_INST_MFLR ) { +if (patch_instruction ((unsigned int *)(ip -4 ),PPC_INST_NOP )) { +pr_err ("Patching NOP failed.\n" ); +return -EPERM ; + } + } +#endif +return 0 ; }@@ -437,9 +465,20 @@ int ftrace_make_nop(struct module *mod,*/ if (test_24bit_addr (ip, addr)) {/* within range */ +int rc; + old =ftrace_call_replace (ip, addr,1 ); new =PPC_INST_NOP ; -return ftrace_modify_code (ip, old, new); + rc =ftrace_modify_code (ip, old, new); +if (rc) +return rc; +#ifdef CONFIG_MPROFILE_KERNEL + old =PPC_INST_MFLR ; + new =PPC_INST_NOP ; +ftrace_modify_code (ip -4 , old, new); +/* old mprofile kernel will error because no mflr */ +#endif +return rc; }else if (core_kernel_text (ip))return __ftrace_make_nop_kernel (rec, addr);@@ -562,6 +601,20 @@ __ftrace_make_call(struct dyn_ftrace *rec, unsigned long addr)return -EINVAL ; } +#ifdef CONFIG_MPROFILE_KERNEL +if (probe_kernel_read (op, (ip -4 ),4 )) { +pr_err ("Fetching instruction at %lx failed.\n" , (unsigned long )(ip -4 )); +return -EFAULT ; + } + +if (op[0 ] ==PPC_INST_NOP ) { +if (patch_instruction ((ip -4 ),PPC_INST_MFLR )) { +pr_err ("Patching mflr failed.\n" ); +return -EINVAL ; + } + } +#endif +if (patch_branch (ip, tramp,BRANCH_SET_LINK )) {pr_err ("REL24 out of range!\n" );return -EINVAL ;@@ -650,6 +703,20 @@ static int __ftrace_make_call_kernel(struct dyn_ftrace *rec, unsigned long addr)return -EINVAL ; } +#ifdef CONFIG_MPROFILE_KERNEL +if (probe_kernel_read (&op, (ip -4 ),4 )) { +pr_err ("Fetching instruction at %lx failed.\n" , (unsigned long )(ip -4 )); +return -EFAULT ; + } + +if (op ==PPC_INST_NOP ) { +if (patch_instruction ((ip -4 ),PPC_INST_MFLR )) { +pr_err ("Patching mflr failed.\n" ); +return -EINVAL ; + } + } +#endif +if (patch_branch (ip, tramp,BRANCH_SET_LINK )) {pr_err ("Error patching branch to ftrace tramp!\n" );return -EINVAL ;@@ -670,6 +737,12 @@ int ftrace_make_call(struct dyn_ftrace *rec, unsigned long addr)*/ if (test_24bit_addr (ip, addr)) {/* within range */ +#ifdef CONFIG_MPROFILE_KERNEL + old =PPC_INST_NOP ; + new =PPC_INST_MFLR ; +ftrace_modify_code (ip -4 , old, new); +/* old mprofile-kernel sequence will error because no nop */ +#endif old =PPC_INST_NOP ; new =ftrace_call_replace (ip, addr,1 );return ftrace_modify_code (ip, old, new);Signature · 1 line
2.20.1
Quoted 29 lines · Michael Ellerman
Quoted 5 lines · Nicholas Piggin
The new mprofile-kernel mcount sequence is
mflr r0
bl_mcountQuoted 7 lines
Dynamic ftrace patches the branch instruction with a noop, but leaves the mflr. mflr is executed by the branch unit that can only execute one per cycle on POWER9 and shared with branches, so it would be nice to avoid it where possible.
This patch is a hacky proof of concept to nop out the mflr. Can we do this or are there races or other issues with it?
There's a race, isn't there?
We have a function foo which currently has tracing disabled, so the mflr and bl are nop'ed out.
CPU 0 CPU 1
==================================
bl foo
nop (ie. not mflr)
-> interrupt
something else enable tracing for foo
... patch mflr and branch
<- rfi
bl_mcount
Quoted 1 line
So we end up in
_mcount()but with r0 not populated.
Good catch! Looks like we need to patch the mflr with a "b +8" similar to what we do in __ftrace_make_nop().
- Naveen
Michael Ellerman
reviewerQuoted 30 lines · Naveen N. Rao
Quoted 29 lines · Michael Ellerman
Quoted 5 lines · Nicholas Piggin
The new mprofile-kernel mcount sequence is
mflr r0
bl_mcountQuoted 7 lines
Dynamic ftrace patches the branch instruction with a noop, but leaves the mflr. mflr is executed by the branch unit that can only execute one per cycle on POWER9 and shared with branches, so it would be nice to avoid it where possible.
This patch is a hacky proof of concept to nop out the mflr. Can we do this or are there races or other issues with it?
There's a race, isn't there?
We have a function foo which currently has tracing disabled, so the mflr and bl are nop'ed out.
CPU 0 CPU 1
==================================
bl foo
nop (ie. not mflr)
-> interrupt
something else enable tracing for foo
... patch mflr and branch
<- rfi
bl_mcount
Quoted 4 lines
Quoted 1 line
So we end up in
_mcount()but with r0 not populated.Good catch! Looks like we need to patch the mflr with a "b +8" similar to what we do in
__ftrace_make_nop().
Would that actually make it any faster though? Nick?
cheers
Quoted 31 lines · Michael Ellerman
Quoted 30 lines · Naveen N. Rao
Quoted 29 lines · Michael Ellerman
Quoted 5 lines · Nicholas Piggin
The new mprofile-kernel mcount sequence is
mflr r0
bl_mcountQuoted 7 lines
Dynamic ftrace patches the branch instruction with a noop, but leaves the mflr. mflr is executed by the branch unit that can only execute one per cycle on POWER9 and shared with branches, so it would be nice to avoid it where possible.
This patch is a hacky proof of concept to nop out the mflr. Can we do this or are there races or other issues with it?
There's a race, isn't there?
We have a function foo which currently has tracing disabled, so the mflr and bl are nop'ed out.
CPU 0 CPU 1
==================================
bl foo
nop (ie. not mflr)
-> interrupt
something else enable tracing for foo
... patch mflr and branch
<- rfi
bl_mcount
Quoted 5 lines
Quoted 3 lines
Quoted 1 line
So we end up in
_mcount()but with r0 not populated.Good catch! Looks like we need to patch the mflr with a "b +8" similar to what we do in
__ftrace_make_nop().Would that actually make it any faster though? Nick?
Ok, how about doing this as a 2-step process?
1. patch 'mflr r0' with a 'b +8'
synchronize_rcu_tasks()
2. convert 'b +8' to a 'nop'
- Naveen
Michael Ellerman
reviewerQuoted 37 lines · Naveen N. Rao
Quoted 36 lines · Michael Ellerman
Quoted 34 lines · Naveen N. Rao
Quoted 29 lines · Michael Ellerman
Quoted 5 lines · Nicholas Piggin
The new mprofile-kernel mcount sequence is
mflr r0
bl_mcountQuoted 7 lines
Dynamic ftrace patches the branch instruction with a noop, but leaves the mflr. mflr is executed by the branch unit that can only execute one per cycle on POWER9 and shared with branches, so it would be nice to avoid it where possible.
This patch is a hacky proof of concept to nop out the mflr. Can we do this or are there races or other issues with it?
There's a race, isn't there?
We have a function foo which currently has tracing disabled, so the mflr and bl are nop'ed out.
CPU 0 CPU 1
==================================
bl foo
nop (ie. not mflr)
-> interrupt
something else enable tracing for foo
... patch mflr and branch
<- rfi
bl_mcountQuoted 1 line
So we end up in
_mcount()but with r0 not populated.Good catch! Looks like we need to patch the mflr with a "b +8" similar to what we do in
__ftrace_make_nop().
Quoted 6 lines
Quoted 1 line
Would that actually make it any faster though? Nick?
Ok, how about doing this as a 2-step process?
1. patch 'mflr r0' with a 'b +8'
synchronize_rcu_tasks()
2. convert 'b +8' to a 'nop'
I think that would work, if I understand synchronize_rcu_tasks().
I worry that it will make the enable/disable expensive. But could be worth trying.
cheers
Nicholas Piggin
reviewerNaveen N. Rao's on May 14, 2019 6:32 pm:
Quoted 37 lines
Quoted 36 lines · Michael Ellerman
Quoted 34 lines · Naveen N. Rao
Quoted 29 lines · Michael Ellerman
Quoted 5 lines · Nicholas Piggin
The new mprofile-kernel mcount sequence is
mflr r0
bl_mcountQuoted 7 lines
Dynamic ftrace patches the branch instruction with a noop, but leaves the mflr. mflr is executed by the branch unit that can only execute one per cycle on POWER9 and shared with branches, so it would be nice to avoid it where possible.
This patch is a hacky proof of concept to nop out the mflr. Can we do this or are there races or other issues with it?
There's a race, isn't there?
We have a function foo which currently has tracing disabled, so the mflr and bl are nop'ed out.
CPU 0 CPU 1
==================================
bl foo
nop (ie. not mflr)
-> interrupt
something else enable tracing for foo
... patch mflr and branch
<- rfi
bl_mcountQuoted 1 line
So we end up in
_mcount()but with r0 not populated.Good catch! Looks like we need to patch the mflr with a "b +8" similar to what we do in
__ftrace_make_nop().
Quoted 6 lines
Quoted 1 line
Would that actually make it any faster though? Nick?
Ok, how about doing this as a 2-step process?
1. patch 'mflr r0' with a 'b +8'
synchronize_rcu_tasks()
2. convert 'b +8' to a 'nop'
Good idea. Well the mflr r0 is harmless, so you can leave that in. You just need to ensure it's not removed before the bl is. So nop the bl _mcount, then synchronize_rcu_tasks(), then nop the mflr?
Thanks,
Nick
Quoted 39 lines · Nicholas Piggin
Naveen N. Rao's on May 14, 2019 6:32 pm:
Quoted 38 lines
Quoted 36 lines · Michael Ellerman
Quoted 30 lines · Naveen N. Rao
Quoted 29 lines · Michael Ellerman
Quoted 5 lines · Nicholas Piggin
The new mprofile-kernel mcount sequence is
mflr r0
bl_mcountQuoted 7 lines
Dynamic ftrace patches the branch instruction with a noop, but leaves the mflr. mflr is executed by the branch unit that can only execute one per cycle on POWER9 and shared with branches, so it would be nice to avoid it where possible.
This patch is a hacky proof of concept to nop out the mflr. Can we do this or are there races or other issues with it?
There's a race, isn't there?
We have a function foo which currently has tracing disabled, so the mflr and bl are nop'ed out.
CPU 0 CPU 1
==================================
bl foo
nop (ie. not mflr)
-> interrupt
something else enable tracing for foo
... patch mflr and branch
<- rfi
bl_mcountQuoted 3 lines
Quoted 1 line
So we end up in
_mcount()but with r0 not populated.Good catch! Looks like we need to patch the mflr with a "b +8" similar to what we do in
__ftrace_make_nop().Would that actually make it any faster though? Nick?
Quoted 8 lines
Quoted 4 lines
Ok, how about doing this as a 2-step process?
1. patch 'mflr r0' with a 'b +8'
synchronize_rcu_tasks()
2. convert 'b +8' to a 'nop'Good idea. Well the mflr r0 is harmless, so you can leave that in. You just need to ensure it's not removed before the bl is. So nop the bl
_mcount, thensynchronize_rcu_tasks(), then nop the mflr?
The problem actually seems to be when we try to patch in the branch to _mcount(), rather than when we are patching in the nop instructions (i.e., the race is when we try to enable the function tracer, rather than while disabling it).
When we disable ftrace, we only need to ensure we patch out the branch to _mcount() before patching out the preceding 'mflr r0'. I don't think we need a synchronize_rcu_tasks() in that case.
While enabling ftrace, we will first need to patch the preceding 'mflr r0' (which would now be a 'nop') with 'b +8', then use synchronize_rcu_tasks() and finally patch in 'bl _mcount()' followed by 'mflr r0'.
I think that's what you meant, just that my reference to patching 'mflr r0' with a 'b +8' should have called out that the mflr would have been nop'ed out.
- Naveen
Nicholas Piggin
reviewerNaveen N. Rao's on May 17, 2019 4:22 am:
Quoted 50 lines
Quoted 40 lines · Nicholas Piggin
Naveen N. Rao's on May 14, 2019 6:32 pm:
Quoted 39 lines
Quoted 37 lines · Michael Ellerman
Quoted 30 lines · Naveen N. Rao
Quoted 29 lines · Michael Ellerman
Quoted 5 lines · Nicholas Piggin
The new mprofile-kernel mcount sequence is
mflr r0
bl_mcountQuoted 7 lines
Dynamic ftrace patches the branch instruction with a noop, but leaves the mflr. mflr is executed by the branch unit that can only execute one per cycle on POWER9 and shared with branches, so it would be nice to avoid it where possible.
This patch is a hacky proof of concept to nop out the mflr. Can we do this or are there races or other issues with it?
There's a race, isn't there?
We have a function foo which currently has tracing disabled, so the mflr and bl are nop'ed out.
CPU 0 CPU 1
==================================
bl foo
nop (ie. not mflr)
-> interrupt
something else enable tracing for foo
... patch mflr and branch
<- rfi
bl_mcountQuoted 4 lines
Quoted 1 line
So we end up in
_mcount()but with r0 not populated.Good catch! Looks like we need to patch the mflr with a "b +8" similar to what we do in
__ftrace_make_nop().Would that actually make it any faster though? Nick?
Quoted 8 lines
Quoted 4 lines
Ok, how about doing this as a 2-step process?
1. patch 'mflr r0' with a 'b +8'
synchronize_rcu_tasks()
2. convert 'b +8' to a 'nop'Good idea. Well the mflr r0 is harmless, so you can leave that in. You just need to ensure it's not removed before the bl is. So nop the bl
_mcount, thensynchronize_rcu_tasks(), then nop the mflr?
Quoted 8 lines
The problem actually seems to be when we try to patch in the branch to
_mcount(), rather than when we are patching in the nop instructions (i.e., the race is when we try to enable the function tracer, rather than while disabling it).When we disable ftrace, we only need to ensure we patch out the branch to
_mcount()before patching out the preceding 'mflr r0'. I don't think we need asynchronize_rcu_tasks()in that case.
That's probably right.
Quoted 8 lines
While enabling ftrace, we will first need to patch the preceding 'mflr r0' (which would now be a 'nop') with 'b +8', then use
synchronize_rcu_tasks()and finally patch in 'bl_mcount()' followed by 'mflr r0'.I think that's what you meant, just that my reference to patching 'mflr r0' with a 'b +8' should have called out that the mflr would have been nop'ed out.
I meant that we don't need the b +8 anywhere, because the mflr r0 is harmless. Enabling ftrace just needs to patch in 'mflr r0', and then synchronize_rcu_tasks(), and then patch in 'bl _mcount'. I think?
Thanks,
Nick
Quoted 9 lines · Nicholas Piggin
Naveen N. Rao's on May 17, 2019 4:22 am:
Quoted 3 lines
While enabling ftrace, we will first need to patch the preceding 'mflr r0' (which would now be a 'nop') with 'b +8', then use
synchronize_rcu_tasks()and finally patch in 'bl_mcount()' followed by 'mflr r0'.I think that's what you meant, just that my reference to patching 'mflr r0' with a 'b +8' should have called out that the mflr would have been nop'ed out.
I meant that we don't need the b +8 anywhere, because the mflr r0 is harmless. Enabling ftrace just needs to patch in 'mflr r0', and then
synchronize_rcu_tasks(), and then patch in 'bl_mcount'. I think?
Ah, that's a good point! That should be enough and it simplifies this further.
Thanks,
Naveen