x86, bitops, variable_test_bit should return 1 not -1 on a match
Subsystems (2)
PATCH 1/1 x86, bitops, variable_test_bit should return 1 not -1 on a match
This issue was noticed while debugging a CPU hotplug issue. On x86 with (NR_CPUS > 1) the cpu_online() define is cpumask_test_cpu(). cpumask_test_cpu() should return 1 if the cpu is set in cpumask and 0 otherwise.
However, cpumask_test_cpu() returns -1 if the cpu in the cpumask is set and 0 otherwise. This happens because cpumask_test_cpu() calls test_bit() which is a define that will call variable_test_bit().
variable_test_bit() calls the assembler instruction sbb (Subtract with Borrow, " Subtracts the source from the destination, and subtracts 1 extra if the Carry Flag is set. Results are returned in "dest".)
A bit match results in -1 being returned from variable_test_bit() if a match occurs, not 1 as the function is supposed to. This can be easily resolved by adding a "!!" to force 0 or 1 as a return.
It looks like the code never does, for example, (test_bit() == 1) so this change should not have any impact.
--- arch/x86/include/asm/bitops.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
arch/x86/include/asm/bitops.h +1 -1
@@ -320,7 +320,7 @@ static inline int variable_test_bit(long nr, volatile const unsigned long *addr) : "=r" (oldbit) : "m" (*(unsigned long *)addr), "Ir" (nr)); -return oldbit; +return !!oldbit; }#if 0 /* Fool kernel-doc since it doesn't do macros yet */
Signature · 1 line
1.7.9.3
Quoted 35 lines · Prarit Bhargava
This issue was noticed while debugging a CPU hotplug issue. On x86 with (
NR_CPUS> 1) thecpu_online()define iscpumask_test_cpu().cpumask_test_cpu()should return 1 if the cpu is set in cpumask and 0 otherwise.However,
cpumask_test_cpu()returns -1 if the cpu in the cpumask is set and 0 otherwise. This happens becausecpumask_test_cpu()callstest_bit()which is a define that will callvariable_test_bit().
variable_test_bit()calls the assembler instruction sbb (Subtract with Borrow, " Subtracts the source from the destination, and subtracts 1 extra if the Carry Flag is set. Results are returned in "dest".)A bit match results in -1 being returned from
variable_test_bit()if a match occurs, not 1 as the function is supposed to. This can be easily resolved by adding a "!!" to force 0 or 1 as a return.It looks like the code never does, for example, (
test_bit()== 1) so this change should not have any impact.Cc Thomas Gleixner <[email protected]>Cc Ingo Molnar <[email protected]>Cc "H. Peter Anvin" <[email protected]>Signed-off-by Prarit Bhargava <[email protected]>--- arch/x86/include/asm/bitops.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)arch/x86/include/asm/bitops.h +0
-0
arch/x86/include/asm/bitops.h static inline int variable_test_bit(long nr, volatile const unsigned long *addr) · line 326
Quoted 8 lines
arch/x86/include/asm/bitops.h +1
-1@@ -320,7 +320,7 @@ static inline int variable_test_bit(long nr, volatile const unsigned long *addr) : "=r" (oldbit) : "m" (*(unsigned long *)addr), "Ir" (nr)); -return oldbit; +return !!oldbit; }#if 0 /* Fool kernel-doc since it doesn't do macros yet */
Ok, I think this is a good fix to improve the robustness of this primitive, unless someone objects.
I tried to find the CPU hotplug code that broke with cpu_online() returning -1 but failed - all current mainline usage sites seem to be testing for nonzero in one way or another. Could you please point it out?
Thanks,
Ingo
H. Peter Anvin
reviewerWrong fix, though. Instead we should change it to use the set instruction, which would also make it easier to use the CC_SET/CC_OUT proposed macros to use assembly out in the future.
The downside with set is that it only sets a single byte, the upside is that it always outputs 0 or 1, and apparently if the output variable is your bool gcc can use that for optimization.
Quoted 66 lines · Ingo Molnar
Quoted 11 lines · Prarit Bhargava
This issue was noticed while debugging a CPU hotplug issue. On x86 with (
NR_CPUS> 1) thecpu_online()define iscpumask_test_cpu().cpumask_test_cpu()should return 1 if the cpu is set in cpumask and 0 otherwise.However,
cpumask_test_cpu()returns -1 if the cpu in the cpumask is set and 0 otherwise. This happens becausecpumask_test_cpu()callstest_bit()which is a define that will callvariable_test_bit().
variable_test_bit()calls the assembler instruction sbb (Subtract with Borrow, " Subtracts the source from the destination, andsubtracts 1
Quoted 3 lines
extra if the Carry Flag is set. Results are returned in "dest".)
A bit match results in -1 being returned from
variable_test_bit()ifa
Quoted 1 line
match occurs, not 1 as the function is supposed to. This can be
easily
Quoted 3 lines
resolved by adding a "!!" to force 0 or 1 as a return.
It looks like the code never does, for example, (
test_bit()== 1) sothis
Quoted 8 lines
change should not have any impact.
Cc Thomas Gleixner <[email protected]>Cc Ingo Molnar <[email protected]>Cc "H. Peter Anvin" <[email protected]>Signed-off-by Prarit Bhargava <[email protected]>Review
Quoted 5 lines
--- arch/x86/include/asm/bitops.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)(quoted diff) +0
-0
b/arch/x86/include/asm/bitops.harch/x86/include/asm/bitops.h static inline int variable_test_bit(long nr,
Quoted 4 lines
index cfe3b95..a87a5fb 100644
arch/x86/include/asm/bitops.h +0
-0@@ -320,7 +320,7 @@ static inline int variable_test_bit(long nr,volatile const unsigned long *addr)
arch/x86/include/asm/bitops.h static inline int variable_test_bit(long nr, volatile const unsigned long *addr) · line 326
Quoted 8 lines
arch/x86/include/asm/bitops.h +1
-1: "=r" (oldbit) : "m" (*(unsigned long *)addr), "Ir" (nr)); -return oldbit; +return !!oldbit; }#if 0 /* Fool kernel-doc since it doesn't do macros yet */ Ok, I think this is a good fix to improve the robustness of this primitive, unless
someone objects.I tried to find the CPU hotplug code that broke with
cpu_online()returning -1 but
failed - all current mainline usage sites seem to be testing for nonzero in one
way or another. Could you please point it out?Thanks,
Ingo
Signature · 1 line
Sent from my Android device with K-9 Mail. Please excuse my brevity.
Quoted 4 lines · H. Peter Anvin
Wrong fix, though. Instead we should change it to use the set instruction, which would also make it easier to use the CC_SET/CC_OUT proposed macros to use assembly out in the future.
The downside with set is that it only sets a single byte, the upside is that it always outputs 0 or 1, and apparently if the output variable is your bool gcc can use that for optimization.
hpa -- can you send a pointer to that discussion?
Thanks,
P.
PATCH v2 1/1 x86, bitops, variable_test_bit should return 1 not -1 on a match
In reply to H. Peter AnvinThis issue was noticed while debugging a CPU hotplug issue. On x86 with (NR_CPUS > 1) the cpu_online() define is cpumask_test_cpu(). cpumask_test_cpu() should return 1 if the cpu is set in cpumask and 0 otherwise.
However, cpumask_test_cpu() returns -1 if the cpu in the cpumask is set and 0 otherwise. This happens because cpumask_test_cpu() calls test_bit() which is a define that will call variable_test_bit().
variable_test_bit() calls the assembler instruction sbb (Subtract with Borrow, " Subtracts the source from the destination, and subtracts 1 extra if the Carry Flag is set. Results are returned in "dest".)
A bit match results in -1 being returned from variable_test_bit() if a match occurs, not 1 as the function is supposed to.
It looks like the code never does, for example, (test_bit() == 1) so this change should not have any impact.
[v2]: hpa: Use setc, (Set if Carry, "Sets the byte in the operand to 1 if the Carry Flag is set, otherwise sets the operand to 0.") instead of !!
--- arch/x86/include/asm/bitops.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)
arch/x86/include/asm/bitops.h +2 -2
@@ -313,10 +313,10 @@ static __always_inline int constant_test_bit(long nr, const volatile unsigned lostatic inline int variable_test_bit (long nr,volatile const unsigned long *addr) { -int oldbit; +u8 oldbit; asmvolatile ("bt %2,%1\n\t" -"sbb %0,%0" +"setc %0" :"=r" (oldbit) :"m" (*(unsigned long *)addr),"Ir" (nr));
Signature · 1 line
1.7.9.3
re-ping on this. Just making sure this wasn't dropped on the floor.
P.
Quoted 50 lines · Prarit Bhargava
This issue was noticed while debugging a CPU hotplug issue. On x86 with (
NR_CPUS> 1) thecpu_online()define iscpumask_test_cpu().cpumask_test_cpu()should return 1 if the cpu is set in cpumask and 0 otherwise.However,
cpumask_test_cpu()returns -1 if the cpu in the cpumask is set and 0 otherwise. This happens becausecpumask_test_cpu()callstest_bit()which is a define that will callvariable_test_bit().
variable_test_bit()calls the assembler instruction sbb (Subtract with Borrow, " Subtracts the source from the destination, and subtracts 1 extra if the Carry Flag is set. Results are returned in "dest".)A bit match results in -1 being returned from
variable_test_bit()if a match occurs, not 1 as the function is supposed to.It looks like the code never does, for example, (
test_bit()== 1) so this change should not have any impact.[v2]: hpa: Use setc, (Set if Carry, "Sets the byte in the operand to 1 if the Carry Flag is set, otherwise sets the operand to 0.") instead of !!
Cc Thomas Gleixner <[email protected]>Cc Ingo Molnar <[email protected]>Cc "H. Peter Anvin" <[email protected]>Signed-off-by Prarit Bhargava <[email protected]>--- arch/x86/include/asm/bitops.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)arch/x86/include/asm/bitops.h +2
-2@@ -313,10 +313,10 @@ static __always_inline int constant_test_bit(long nr, const volatile unsigned lostatic inline int variable_test_bit (long nr,volatile const unsigned long *addr) { -int oldbit; +u8 oldbit; asmvolatile ("bt %2,%1\n\t" -"sbb %0,%0" +"setc %0" :"=r" (oldbit) :"m" (*(unsigned long *)addr),"Ir" (nr));
Quoted 1 line · Prarit Bhargava
re-ping on this. Just making sure this wasn't dropped on the floor.
So I didn't apply it back when I saw your patch because I didn't see where you addressed/analyzed the second paragraph of hpa's review:
"The downside with set is that it only sets a single byte, the upside is that it
always outputs 0 or 1, and apparently if the output variable is your bool gcc can use that for optimization."
That's a valid concern and should be addressed.
Thanks,
Ingo
Quoted 4 lines · H. Peter Anvin
Wrong fix, though. Instead we should change it to use the set instruction, which would also make it easier to use the CC_SET/CC_OUT proposed macros to use assembly out in the future.
The downside with set is that it only sets a single byte, the upside is that it always outputs 0 or 1, and apparently if the output variable is your bool gcc can use that for optimization.
hpa, I didn't realize your comment was suggesting a change. I've done a google search on "gcc bool optimization" (and various incantations of that) and didn't find anything that lead me in the direction of making a change. Could you elaborate on what the issue is?
Thanks,
P.
Quoted 45 lines · Ingo Molnar
Quoted 42 lines · Prarit Bhargava
This issue was noticed while debugging a CPU hotplug issue. On x86 with (
NR_CPUS> 1) thecpu_online()define iscpumask_test_cpu().cpumask_test_cpu()should return 1 if the cpu is set in cpumask and 0 otherwise.However,
cpumask_test_cpu()returns -1 if the cpu in the cpumask is set and 0 otherwise. This happens becausecpumask_test_cpu()callstest_bit()which is a define that will callvariable_test_bit().
variable_test_bit()calls the assembler instruction sbb (Subtract with Borrow, " Subtracts the source from the destination, and subtracts 1 extra if the Carry Flag is set. Results are returned in "dest".)A bit match results in -1 being returned from
variable_test_bit()if a match occurs, not 1 as the function is supposed to. This can be easily resolved by adding a "!!" to force 0 or 1 as a return.It looks like the code never does, for example, (
test_bit()== 1) so this change should not have any impact.Cc Thomas Gleixner <[email protected]>Cc Ingo Molnar <[email protected]>Cc "H. Peter Anvin" <[email protected]>Signed-off-by Prarit Bhargava <[email protected]>--- arch/x86/include/asm/bitops.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)arch/x86/include/asm/bitops.h +1
-1@@ -320,7 +320,7 @@ static inline int variable_test_bit(long nr, volatile const unsigned long *addr) : "=r" (oldbit) : "m" (*(unsigned long *)addr), "Ir" (nr)); -return oldbit; +return !!oldbit; }
Quoted 8 lines
arch/x86/include/asm/bitops.h static inline int variable_test_bit(long nr, volatile const unsigned long *addr) · line 326
Quoted 1 line
arch/x86/include/asm/bitops.h +0
-0#if 0 /* Fool kernel-doc since it doesn't do macros yet */ Ok, I think this is a good fix to improve the robustness of this primitive, unless someone objects.
I tried to find the CPU hotplug code that broke with
cpu_online()returning -1 but failed - all current mainline usage sites seem to be testing for nonzero in one way or another. Could you please point it out?
I'm sorry Ingo, I think my description may have confused you. I was debugging a cpu hotplug issue[1] and did
printk("cpu %d cpu online status %d\n", cpu, cpu_online(cpu));
as a debug printk. This printed out
cpu 3 cpu online status -1
which was really confusing. That lead me down the rabbit hole of looking at the sbb assembler instruction in variable_test_bit() to figure out why I was seeing -1.
P.
[1] The bug looks like it has to do with the system's firmware, not cpu hotplug.
Quoted 60 lines · Prarit Bhargava
Quoted 45 lines · Ingo Molnar
Quoted 42 lines · Prarit Bhargava
This issue was noticed while debugging a CPU hotplug issue. On x86 with (
NR_CPUS> 1) thecpu_online()define iscpumask_test_cpu().cpumask_test_cpu()should return 1 if the cpu is set in cpumask and 0 otherwise.However,
cpumask_test_cpu()returns -1 if the cpu in the cpumask is set and 0 otherwise. This happens becausecpumask_test_cpu()callstest_bit()which is a define that will callvariable_test_bit().
variable_test_bit()calls the assembler instruction sbb (Subtract with Borrow, " Subtracts the source from the destination, and subtracts 1 extra if the Carry Flag is set. Results are returned in "dest".)A bit match results in -1 being returned from
variable_test_bit()if a match occurs, not 1 as the function is supposed to. This can be easily resolved by adding a "!!" to force 0 or 1 as a return.It looks like the code never does, for example, (
test_bit()== 1) so this change should not have any impact.Cc Thomas Gleixner <[email protected]>Cc Ingo Molnar <[email protected]>Cc "H. Peter Anvin" <[email protected]>Signed-off-by Prarit Bhargava <[email protected]>--- arch/x86/include/asm/bitops.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)arch/x86/include/asm/bitops.h +1
-1@@ -320,7 +320,7 @@ static inline int variable_test_bit(long nr, volatile const unsigned long *addr) : "=r" (oldbit) : "m" (*(unsigned long *)addr), "Ir" (nr)); -return oldbit; +return !!oldbit; }Quoted 8 lines
arch/x86/include/asm/bitops.h static inline int variable_test_bit(long nr, volatile const unsigned long *addr) · line 326
Quoted 1 line
arch/x86/include/asm/bitops.h +0
-0#if 0 /* Fool kernel-doc since it doesn't do macros yet */ Ok, I think this is a good fix to improve the robustness of this primitive, unless someone objects.
I tried to find the CPU hotplug code that broke with
cpu_online()returning -1 but failed - all current mainline usage sites seem to be testing for nonzero in one way or another. Could you please point it out?I'm sorry Ingo, I think my description may have confused you. I was debugging a cpu hotplug issue[1] and did
Quoted 8 lines
printk("cpu %d cpu online status %d\n", cpu, cpu_online(cpu));as a debug printk. This printed out
cpu 3 cpu online status -1
which was really confusing. That lead me down the rabbit hole of looking at the sbb assembler instruction in
variable_test_bit()to figure out why I was seeing -1.
Ok, fair enough!
Still worth fixing IMHO.
Thanks,
Ingo