Skip to contentSkip to conversation

doc: Add remote CPU access details and others to this_cpu_ops.txt

PATCH Superseded Pranith Kumar 2 participants 12 messages Jul 16, 2014 – Jul 18, 2014 2 messages not in the archive Documentation

v2 of 3 Superseded Latest is v3 →

lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 doc: Add remote CPU access details and others to this_cpu_ops.txt

Add more details from a recent kernel newbies mailing list discussion here:
http://www.spinics.net/lists/newbies/msg52747.html

Also list the operations available and add details about safe accesses.

v2:
* updated with comments from Christoph Lameter

CC Christoph Lameter <[email protected]>
---
 Documentation/this_cpu_ops.txt | 129 ++++++++++++++++++++++++++++++++++-------
 scripts/bin2c                  | Bin 0 -> 8706 bytes
 2 files changed, 108 insertions(+), 21 deletions(-)
 create mode 100755 scripts/bin2c
Documentation/this_cpu_ops.txt +108 -21
index 1a4ce7e..b1faa9d 100644
@@ -13,22 +13,51 @@ operating on the per cpu variable.
 
 This means there are no atomicity issues between the calculation of
 the offset and the operation on the data. Therefore it is not
-necessary to disable preempt or interrupts to ensure that the
+necessary to disable preemption or interrupts to ensure that the
 processor is not changed between the calculation of the address and
 the operation on the data.
 
 Read-modify-write operations are of particular interest. Frequently
 processors have special lower latency instructions that can operate
-without the typical synchronization overhead but still provide some
-sort of relaxed atomicity guarantee. The x86 for example can execute
-RMV (Read Modify Write) instructions like inc/dec/cmpxchg without the
+without the typical synchronization overhead, but still provide some
+sort of relaxed atomicity guarantees. The x86, for example, can execute
+RMW (Read Modify Write) instructions like inc/dec/cmpxchg without the
 lock prefix and the associated latency penalty.
 
 Access to the variable without the lock prefix is not synchronized but
 synchronization is not necessary since we are dealing with per cpu
 data specific to the currently executing processor. Only the current
 processor should be accessing that variable and therefore there are no
-concurrency issues with other processors in the system.
+concurrency issues with other processors in the system. Please see the 
+section "Remote access to per-CPU data" if you need remote access.
+
+The main use of the this_cpu operations has been to optimize counter
+operations. 
+
+The following this_cpu() operations with implied preemption protection
+are defined. These operations can be used without worrying about
+preemption and interrupts.
+	
+	this_cpu_add()
+	this_cpu_read(pcp)
+	this_cpu_write(pcp, val)	
+	this_cpu_add(pcp, val)		
+	this_cpu_and(pcp, val)		
+	this_cpu_or(pcp, val)		
+	this_cpu_add_return(pcp, val)	
+	this_cpu_xchg(pcp, nval)	
+	this_cpu_cmpxchg(pcp, oval, nval) 
+	this_cpu_cmpxchg_double(pcp1, pcp2, oval1, oval2, nval1, nval2) 
+	this_cpu_sub(pcp, val)		
+	this_cpu_inc(pcp)		
+	this_cpu_dec(pcp)		
+	this_cpu_sub_return(pcp, val)	
+	this_cpu_inc_return(pcp)	
+	this_cpu_dec_return(pcp)	
+
+
+Inner working of this_cpu operations
+------------------------------------
 
 On x86 the fs: or the gs: segment registers contain the base of the
 per cpu area. It is then possible to simply use the segment override
@@ -53,12 +82,11 @@ this_cpu_ops such sequence also required preempt disable/enable to
 prevent the kernel from moving the thread to a different processor
 while the calculation is performed.
 
-The main use of the this_cpu operations has been to optimize counter
-operations.
+Consider the following this_cpu operation
 
 	this_cpu_inc(x)
 
-results in the following single instruction (no lock prefix!)
+The above results in the following single instruction (no lock prefix!)
 
 	inc gs:[x]
 
@@ -100,11 +128,10 @@ Takes the offset of a per cpu variable (&x !) and returns the address
 of the per cpu variable that belongs to the currently executing
 processor.  this_cpu_ptr avoids multiple steps that the common
 get_cpu/put_cpu sequence requires. No processor number is
-available. Instead the offset of the local per cpu area is simply
+available. Instead, the offset of the local per cpu area is simply
 added to the percpu offset.
 
 
-
 Per cpu variables and offsets
 -----------------------------
 
@@ -118,15 +145,16 @@ Therefore the use of x or &x outside of the context of per cpu
 operations is invalid and will generally be treated like a NULL
 pointer dereference.
 
-In the context of per cpu operations
+	DEFINE_PER_CPU(int, x);
 
-	x is a per cpu variable. Most this_cpu operations take a cpu
-	variable.
+In the context of per cpu operations the above implies that x is a per
+cpu variable. Most this_cpu operations take a cpu variable.
 
-	&x is the *offset* a per cpu variable. this_cpu_ptr() takes
-	the offset of a per cpu variable which makes this look a bit
-	strange.
+	int __percpu *p = &x;
 
+&x and hence p is the *offset* of a per cpu variable. this_cpu_ptr() 
+takes the offset of a per cpu variable which makes this look a bit
+strange.
 
 
 Operations on a field of a per cpu structure
@@ -172,16 +200,40 @@ if we do not make use of this_cpu ops later to manipulate fields:
 Variants of this_cpu ops
 -------------------------
 
-this_cpu ops are interrupt safe. Some architecture do not support
+this_cpu ops are interrupt safe. Some architectures do not support
 these per cpu local operations. In that case the operation must be
 replaced by code that disables interrupts, then does the operations
 that are guaranteed to be atomic and then reenable interrupts. Doing
 so is expensive. If there are other reasons why the scheduler cannot
 change the processor we are executing on then there is no reason to
-disable interrupts. For that purpose the __this_cpu operations are
-provided. For example.
+disable interrupts. For that purpose the following __this_cpu operations
+are provided. 
+
+These operations have no guarantee against concurrent interrupts or
+preemption. If a per cpu variable is not used in an interrupt context
+and the scheduler cannot preempt then they are safe.
+
+Note that interrupts may still occur while an operation is
+in progress and if the interrupt too modifies the variable, then RMW
+actions may not be atomic.
+
+	__this_cpu_add()
+	__this_cpu_read(pcp)
+	__this_cpu_write(pcp, val)	
+	__this_cpu_add(pcp, val)		
+	__this_cpu_and(pcp, val)		
+	__this_cpu_or(pcp, val)		
+	__this_cpu_add_return(pcp, val)	
+	__this_cpu_xchg(pcp, nval)	
+	__this_cpu_cmpxchg(pcp, oval, nval) 
+	__this_cpu_cmpxchg_double(pcp1, pcp2, oval1, oval2, nval1, nval2) 
+	__this_cpu_sub(pcp, val)		
+	__this_cpu_inc(pcp)		
+	__this_cpu_dec(pcp)		
+	__this_cpu_sub_return(pcp, val)	
+	__this_cpu_inc_return(pcp)	
+	__this_cpu_dec_return(pcp)	
 
-	__this_cpu_inc(x);
 
 Will increment x and will not fallback to code that disables
 interrupts on platforms that cannot accomplish atomicity through
@@ -189,7 +241,6 @@ address relocation and a Read-Modify-Write operation in the same
 instruction.
 
 
-
 &this_cpu_ptr(pp)->n vs this_cpu_ptr(&pp->n)
 --------------------------------------------
 
@@ -202,4 +253,40 @@ with (). The second form also is consistent with the way
 this_cpu_read() and friends are used.
 
 
+Remote access to per-CPU data
+------------------------------
+
+Per cpu data structures are for the use of one cpu exclusively. If you
+use the variables as intended, this_cpu_ops() are guaranteed to be
+"atomic" as no other CPU has access to these data structures. But there
+are special cases where you might need to access per cpu data structures
+remotely. One such case is to perform maintenance tasks of idle CPUs
+without having to wake them up using IPIs for power saving purposes.
+It is usually safe to do a remote read access and that is frequently
+done to summarize counters. Remote write access is the one which is
+problematic. This is not recommended unless absolutely necessary.
+
+To access per-cpu data structure remotely, you need to convert the 
+per-cpu offset to a pointer using this_cpu_ptr().
+
+	DEFINE_PER_CPU(struct data, datap);
+	struct data *p = per_cpu(&datap, cpu);
+
+p can now be passed to another CPU to be updated remotely.
+
+This makes it explicit that we are getting ready to access a foreing
+percpu area.
+
+You can also do the following to convert the datap offset to an address
+
+	struct data *p = this_cpu_ptr(&datap);
+
+but, passing of pointers calculated via this_cpu_ptr to other cpus is
+unusual and should be avoided.
+
+Such remote accesses to per CPU data are not guaranteed to be atomic
+anymore. You will have to use atomic_t and rely on the standard atomic
+operations for these remote accesses to be atomic.
+
 Christoph Lameter, April 3rd, 2013
+Pranith Kumar, July 16th, 2014
scripts/bin2c new file binary +0 -0
new file mode 100755
index 0000000000000000000000000000000000000000..58f44ba9ae94ca23445e244fb83132e779ffe4ad
GIT binary patch
Signature · 1 line
1.9.1
#1 lore raw

Documentation/this_cpu_ops.txt with (). The second form also is consistent with the way · line 289

Quoted 10 lines · Pranith Kumar
Documentation/this_cpu_ops.txt +10 -0
+You can also do the following to convert the datap offset to an address
+
+	struct data *p = this_cpu_ptr(&datap);
+
+but, passing of pointers calculated via this_cpu_ptr to other cpus is
+unusual and should be avoided.
+
+Such remote accesses to per CPU data are not guaranteed to be atomic
+anymore. You will have to use atomic_t and rely on the standard atomic
+operations for these remote accesses to be atomic.

The use of atomic_t implies a remote write operation to a percpu area.

atomic_t needs to be avoided. If data needs to be modified from multiple cpus then it usually does not belong into a percpu area.

A message that is not in the archive (sent off-list or not yet received): [email protected]

#3 lore raw
Quoted 11 lines · Pranith Kumar
Quoted 5 lines

The use of atomic_t implies a remote write operation to a percpu area.

atomic_t needs to be avoided. If data needs to be modified from multiple cpus then it usually does not belong into a percpu area.

Yes, I think I made it pretty clear that remote accesses need to be avoided unless absolutely necessary. But, there will be scenarios where mostly local data will need to be have remote accesses. In such scenarios, isn't using atomic_t better? FYI, that is how RCU code currently works. It uses atomic_t in per cpu areas to ensure atomicity for remote accesses.

The RCU code has .... ummmm... some issues with percpu usage and should not be taken as a good example. If you look at the RCU code it looks horrible with numerous barriers around remote percpu read/wrirte accesses and one wonders if that code is actually ok.

Quoted 2 lines

If data needs to be modified from multiple cpus only very rarely, doesn't it make sense to use per-cpu areas?

I would suggest that this should not occur. You can always "modify" remote percpu areas by generating an IPI on that cpu to make that processor update its own per cpu data.

#4 lore raw
Quoted 14 lines · Christoph Lameter
Quoted 11 lines · Pranith Kumar
Quoted 5 lines

The use of atomic_t implies a remote write operation to a percpu area.

atomic_t needs to be avoided. If data needs to be modified from multiple cpus then it usually does not belong into a percpu area.

Yes, I think I made it pretty clear that remote accesses need to be avoided unless absolutely necessary. But, there will be scenarios where mostly local data will need to be have remote accesses. In such scenarios, isn't using atomic_t better? FYI, that is how RCU code currently works. It uses atomic_t in per cpu areas to ensure atomicity for remote accesses.

Quoted 4 lines

The RCU code has .... ummmm... some issues with percpu usage and should not be taken as a good example. If you look at the RCU code it looks horrible with numerous barriers around remote percpu read/wrirte accesses and one wonders if that code is actually ok.

Well, it is running in all our kernels with not many reported issues, isn't it ;) And yes, that is one of the extra-ordinary situations where we use per-cpu data. Once you've extracted a pointer to the per-cpu area -and- ensure that concurrent accesses do not happen(or happen with enough guarantees using barriers), what is the case against remote accesses? I am asking from a correctness and a performance point of view, not style/aesthetics.

Quoted 8 lines
Quoted 3 lines

If data needs to be modified from multiple cpus only very rarely, doesn't it make sense to use per-cpu areas?

I would suggest that this should not occur. You can always "modify" remote percpu areas by generating an IPI on that cpu to make that processor update its own per cpu data.

The case against doing that is not to wake up CPUs which are in idle/sleep states. I think mentioning it here that remote accesses are strongly discouraged with a reasonable explanation of the implications should be enough. There might always be rare situations where remote accesses might be necessary.

--
Pranith

#5 lore raw
Quoted 11 lines · Pranith Kumar
Quoted 4 lines

The RCU code has .... ummmm... some issues with percpu usage and should not be taken as a good example. If you look at the RCU code it looks horrible with numerous barriers around remote percpu read/wrirte accesses and one wonders if that code is actually ok.

Well, it is running in all our kernels with not many reported issues, isn't it ;) And yes, that is one of the extra-ordinary situations where we use per-cpu data. Once you've extracted a pointer to the per-cpu area -and- ensure that concurrent accesses do not happen(or happen with enough guarantees using barriers), what is the case against remote accesses? I am asking from a correctness and a performance point of view, not style/aesthetics.

Could be working but I do not want it to be mentioned in the documentation given the problems it causes. IPI is preferable.

Quoted 8 lines
Quoted 7 lines
Quoted 2 lines

If data needs to be modified from multiple cpus only very rarely, doesn't it make sense to use per-cpu areas?

I would suggest that this should not occur. You can always "modify" remote percpu areas by generating an IPI on that cpu to make that processor update its own per cpu data.

Quoted 4 lines

The case against doing that is not to wake up CPUs which are in idle/sleep states. I think mentioning it here that remote accesses are strongly discouraged with a reasonable explanation of the implications should be enough. There might always be rare situations where remote accesses might be necessary.

Remote percpu updates are extremely rare events. If the cpu is idle/asleep then usually no updates are needed because no activity is occurring on that cpu.

#6 lore raw
Quoted 14 lines · Christoph Lameter
Quoted 11 lines · Pranith Kumar
Quoted 4 lines

The RCU code has .... ummmm... some issues with percpu usage and should not be taken as a good example. If you look at the RCU code it looks horrible with numerous barriers around remote percpu read/wrirte accesses and one wonders if that code is actually ok.

Well, it is running in all our kernels with not many reported issues, isn't it ;) And yes, that is one of the extra-ordinary situations where we use per-cpu data. Once you've extracted a pointer to the per-cpu area -and- ensure that concurrent accesses do not happen(or happen with enough guarantees using barriers), what is the case against remote accesses? I am asking from a correctness and a performance point of view, not style/aesthetics.

Quoted 2 lines

Could be working but I do not want it to be mentioned in the documentation given the problems it causes. IPI is preferable.

I can mention that IPI is preferable. What is that you don't want mentioned? atomic_t?

Quoted 9 lines
Quoted 9 lines
Quoted 8 lines
Quoted 3 lines

If data needs to be modified from multiple cpus only very rarely, doesn't it make sense to use per-cpu areas?

I would suggest that this should not occur. You can always "modify" remote percpu areas by generating an IPI on that cpu to make that processor update its own per cpu data.

Quoted 9 lines
Quoted 4 lines

The case against doing that is not to wake up CPUs which are in idle/sleep states. I think mentioning it here that remote accesses are strongly discouraged with a reasonable explanation of the implications should be enough. There might always be rare situations where remote accesses might be necessary.

Remote percpu updates are extremely rare events. If the cpu is idle/asleep then usually no updates are needed because no activity is occurring on that cpu.

Yes, -usually- that is the case. But we are talking about the extreme rare event where we need to update some remote CPU`s per-cpu data without waking it up from sleep/idle. How do you suggest we handle this? I don't think suggesting not to use per-cpu areas because of this is a good idea, since we lose a lot of performance in the most common cases.

Thoughts?

--
Pranith

#7 lore raw
Quoted 1 line · Pranith Kumar

I can mention that IPI is preferable. What is that you don't want mentioned? atomic_t?

Definitely not as an example. atomic_t in per cpu areas is self contradicting. The per cpu area is exclusively for that processor whereas an atomic_t is supposed to be accessed from multiple processors.

Quoted 10 lines
Quoted 4 lines

Remote percpu updates are extremely rare events. If the cpu is idle/asleep then usually no updates are needed because no activity is occurring on that cpu.

Yes, -usually- that is the case. But we are talking about the extreme rare event where we need to update some remote CPU`s per-cpu data without waking it up from sleep/idle. How do you suggest we handle this? I don't think suggesting not to use per-cpu areas because of this is a good idea, since we lose a lot of performance in the most common cases.

If you modify a percpu area then that is usually done because that cpu needs to take some action. An IPI is fine.

Otherwise yes I would suggest not use a percpu area but a separate data structure for synchronization.

#8 lore raw
Quoted 19 lines · Christoph Lameter
Quoted 1 line · Pranith Kumar

I can mention that IPI is preferable. What is that you don't want mentioned? atomic_t?

Definitely not as an example. atomic_t in per cpu areas is self contradicting. The per cpu area is exclusively for that processor whereas an atomic_t is supposed to be accessed from multiple processors.

Quoted 10 lines
Quoted 4 lines

Remote percpu updates are extremely rare events. If the cpu is idle/asleep then usually no updates are needed because no activity is occurring on that cpu.

Yes, -usually- that is the case. But we are talking about the extreme rare event where we need to update some remote CPU`s per-cpu data without waking it up from sleep/idle. How do you suggest we handle this? I don't think suggesting not to use per-cpu areas because of this is a good idea, since we lose a lot of performance in the most common cases.

Quoted 6 lines

If you modify a percpu area then that is usually done because that cpu needs to take some action. An IPI is fine.

Otherwise yes I would suggest not use a percpu area but a separate data structure for synchronization.

Yes, I will add this information to the doc. Thanks!

--
Pranith

#9 lore raw
In reply to Pranith Kumar

Regarding atomic_t in per cpu areas: I am uncomfortable especially because both locked and unlocked RMW write operations could be acting on values in the same cacheline. I am concerned that the unlocked operation could have an unpredictable result.

f.e. the following per cpu data structure

struct test {
	atomic_t a;
	int b;
} onecacheline;

Local cpu does

	this_cpu_inc(onecacheline.b);

If this is racing with a remote cpus:

	atomic_inc(percpu(&a, cpu))

then we have on x86 a increment operation with locked semantics racing with an unlocked one on the same cacheline.

#10 lore raw
Quoted 16 lines · Christoph Lameter

Regarding atomic_t in per cpu areas: I am uncomfortable especially because both locked and unlocked RMW write operations could be acting on values in the same cacheline. I am concerned that the unlocked operation could have an unpredictable result.

f.e. the following per cpu data structure

struct test {
	atomic_t a;
	int b;
} onecacheline;

Local cpu does

Quoted 9 lines
	this_cpu_inc(onecacheline.b);

If this is racing with a remote cpus:

	atomic_inc(percpu(&a, cpu))

then we have on x86 a increment operation with locked semantics racing with an unlocked one on the same cacheline.

OK, I will add this as a warning in the documentation. Thanks!

--
Pranith

#11 lore raw
Quoted 3 lines · Pranith Kumar
Quoted 2 lines

then we have on x86 a increment operation with locked semantics racing with an unlocked one on the same cacheline.

OK, I will add this as a warning in the documentation. Thanks!

Note that I have not been able to get beyond a bad feeling. Looked up various sources of information about how the lock prefix works. Still I am not sure if this works or not.

A message that is not in the archive (sent off-list or not yet received): [email protected]