this_cpu_ops.txt 11.2 KB
Newer Older
1
===================
2
this_cpu operations
3 4 5 6
===================

:Author: Christoph Lameter, August 4th, 2014
:Author: Pranith Kumar, Aug 2nd, 2014
7 8

this_cpu operations are a way of optimizing access to per cpu
9 10 11 12
variables associated with the *currently* executing processor. This is
done through the use of segment registers (or a dedicated register where
the cpu permanently stored the beginning of the per cpu	area for a
specific processor).
13

14 15
this_cpu operations add a per cpu variable offset to the processor
specific per cpu base and encode that operation in the instruction
16 17
operating on the per cpu variable.

18
This means that there are no atomicity issues between the calculation of
19
the offset and the operation on the data. Therefore it is not
20
necessary to disable preemption or interrupts to ensure that the
21 22 23 24 25
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
26 27 28
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
29 30 31 32 33 34 35 36
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.

37 38 39 40 41 42 43 44 45
Please note that accesses by remote processors to a per cpu area are
exceptional situations and may impact performance and/or correctness
(remote write operations) of local RMW operations via this_cpu_*.

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
46
preemption and interrupts::
47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67

	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
------------------------------------

68 69 70 71 72 73
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
to relocate a per cpu relative address to the proper per cpu area for
the processor. So the relocation to the per cpu base is encoded in the
instruction via a segment register prefix.

74
For example::
75 76 77 78 79 80

	DEFINE_PER_CPU(int, x);
	int z;

	z = this_cpu_read(x);

81
results in a single instruction::
82 83 84 85

	mov ax, gs:[x]

instead of a sequence of calculation of the address and then a fetch
86
from that address which occurs with the per cpu operations. Before
87 88 89 90
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.

91
Consider the following this_cpu operation::
92 93 94

	this_cpu_inc(x)

95
The above results in the following single instruction (no lock prefix!)::
96 97 98 99

	inc gs:[x]

instead of the following operations required if there is no segment
100
register::
101 102 103 104 105 106 107 108 109

	int *y;
	int cpu;

	cpu = get_cpu();
	y = per_cpu_ptr(&x, cpu);
	(*y)++;
	put_cpu();

110
Note that these operations can only be used on per cpu data that is
111 112
reserved for a specific processor. Without disabling preemption in the
surrounding code this_cpu_inc() will only guarantee that one of the
113
per cpu counters is correctly incremented. However, there is no
114 115 116 117 118 119 120 121 122
guarantee that the OS will not move the process directly before or
after the this_cpu instruction is executed. In general this means that
the value of the individual counters for each processor are
meaningless. The sum of all the per cpu counters is the only value
that is of interest.

Per cpu variables are used for performance reasons. Bouncing cache
lines can be avoided if multiple processors concurrently go through
the same code paths.  Since each processor has its own per cpu
123
variables no concurrent cache line updates take place. The price that
124
has to be paid for this optimization is the need to add up the per cpu
125
counters when the value of a counter is needed.
126 127


128 129 130 131
Special operations
------------------

::
132 133 134 135 136 137 138

	y = this_cpu_ptr(&x)

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
139 140
available. Instead, the offset of the local per cpu area is simply
added to the per cpu offset.
141

142 143 144 145 146
Note that this operation is usually used in a code segment when
preemption has been disabled. The pointer is then used to
access local per cpu data in a critical section. When preemption
is re-enabled this pointer is usually no longer useful since it may
no longer point to per cpu data of the current processor.
147 148 149 150 151


Per cpu variables and offsets
-----------------------------

152
Per cpu variables have *offsets* to the beginning of the per cpu
153 154
area. They do not have addresses although they look like that in the
code. Offsets cannot be directly dereferenced. The offset must be
155
added to a base pointer of a per cpu area of a processor in order to
156 157 158 159 160 161
form a valid address.

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.

162 163
::

164
	DEFINE_PER_CPU(int, x);
165

166 167
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.
168

169 170
::

171
	int __percpu *p = &x;
172

173 174 175
&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.
176 177 178 179 180


Operations on a field of a per cpu structure
--------------------------------------------

181
Let's say we have a percpu structure::
182 183 184 185 186 187 188 189

	struct s {
		int n,m;
	};

	DEFINE_PER_CPU(struct s, p);


190
Operations on these fields are straightforward::
191 192 193 194 195 196

	this_cpu_inc(p.m)

	z = this_cpu_cmpxchg(p.m, 0, 1);


197
If we have an offset to struct s::
198 199 200

	struct s __percpu *ps = &p;

201
	this_cpu_dec(ps->m);
202 203 204 205 206

	z = this_cpu_inc_return(ps->n);


The calculation of the pointer may require the use of this_cpu_ptr()
207
if we do not make use of this_cpu ops later to manipulate fields::
208 209 210 211 212 213 214 215 216 217 218

	struct s *pp;

	pp = this_cpu_ptr(&p);

	pp->m--;

	z = pp->n++;


Variants of this_cpu ops
219
------------------------
220

221
this_cpu ops are interrupt safe. Some architectures do not support
222 223
these per cpu local operations. In that case the operation must be
replaced by code that disables interrupts, then does the operations
224
that are guaranteed to be atomic and then re-enable interrupts. Doing
225 226
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
227 228 229 230 231 232 233 234
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. If any interrupts
still occur while an operation is in progress and if the interrupt too
modifies the variable, then RMW actions can not be guaranteed to be
235
safe::
236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254

	__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)


Will increment x and will not fall-back to code that disables
255 256 257 258 259 260 261 262 263
interrupts on platforms that cannot accomplish atomicity through
address relocation and a Read-Modify-Write operation in the same
instruction.


&this_cpu_ptr(pp)->n vs this_cpu_ptr(&pp->n)
--------------------------------------------

The first operation takes the offset and forms an address and then
264 265
adds the offset of the n field. This may result in two add
instructions emitted by the compiler.
266 267 268 269 270 271 272

The second one first adds the two offsets and then does the
relocation.  IMHO the second form looks cleaner and has an easier time
with (). The second form also is consistent with the way
this_cpu_read() and friends are used.


273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291
Remote access to per cpu data
------------------------------

Per cpu data structures are designed to be used by 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.

There are special cases where you might need to access per cpu data
structures remotely. It is usually safe to do a remote read access
and that is frequently done to summarize counters. Remote write access
something which could be problematic because this_cpu ops do not
have lock semantics. A remote write may interfere with a this_cpu
RMW operation.

Remote write accesses to percpu data structures are highly discouraged
unless absolutely necessary. Please consider using an IPI to wake up
the remote CPU and perform the update to its per cpu area.

To access per-cpu data structure remotely, typically the per_cpu_ptr()
292
function is used::
293 294 295 296 297 298 299 300 301


	DEFINE_PER_CPU(struct data, datap);

	struct data *p = per_cpu_ptr(&datap, cpu);

This makes it explicit that we are getting ready to access a percpu
area remotely.

302
You can also do the following to convert the datap offset to an address::
303 304 305 306 307 308 309 310 311 312 313 314 315 316 317

	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.

Remote access are typically only for reading the status of another cpus
per cpu data. Write accesses can cause unique problems due to the
relaxed synchronization requirements for this_cpu operations.

One example that illustrates some concerns with write operations is
the following scenario that occurs because two per cpu variables
share a cache-line but the relaxed synchronization is applied to
only one process updating the cache-line.

318
Consider the following example::
319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339


	struct test {
		atomic_t a;
		int b;
	};

	DEFINE_PER_CPU(struct test, onecacheline);

There is some concern about what would happen if the field 'a' is updated
remotely from one processor and the local processor would use this_cpu ops
to update field b. Care should be taken that such simultaneous accesses to
data within the same cache line are avoided. Also costly synchronization
may be necessary. IPIs are generally recommended in such scenarios instead
of a remote write to the per cpu area of another processor.

Even in cases where the remote writes are rare, please bear in
mind that a remote write will evict the cache line from the processor
that most likely will access it. If the processor wakes up and finds a
missing local cache line of a per cpu area, its performance and hence
the wake up times will be affected.