qemu/docs/atomics.txt
<<
>>
Prefs
   1CPUs perform independent memory operations effectively in random order.
   2but this can be a problem for CPU-CPU interaction (including interactions
   3between QEMU and the guest).  Multi-threaded programs use various tools
   4to instruct the compiler and the CPU to restrict the order to something
   5that is consistent with the expectations of the programmer.
   6
   7The most basic tool is locking.  Mutexes, condition variables and
   8semaphores are used in QEMU, and should be the default approach to
   9synchronization.  Anything else is considerably harder, but it's
  10also justified more often than one would like.  The two tools that
  11are provided by qemu/atomic.h are memory barriers and atomic operations.
  12
  13Macros defined by qemu/atomic.h fall in three camps:
  14
  15- compiler barriers: barrier();
  16
  17- weak atomic access and manual memory barriers: atomic_read(),
  18  atomic_set(), smp_rmb(), smp_wmb(), smp_mb(), smp_read_barrier_depends();
  19
  20- sequentially consistent atomic access: everything else.
  21
  22
  23COMPILER MEMORY BARRIER
  24=======================
  25
  26barrier() prevents the compiler from moving the memory accesses either
  27side of it to the other side.  The compiler barrier has no direct effect
  28on the CPU, which may then reorder things however it wishes.
  29
  30barrier() is mostly used within qemu/atomic.h itself.  On some
  31architectures, CPU guarantees are strong enough that blocking compiler
  32optimizations already ensures the correct order of execution.  In this
  33case, qemu/atomic.h will reduce stronger memory barriers to simple
  34compiler barriers.
  35
  36Still, barrier() can be useful when writing code that can be interrupted
  37by signal handlers.
  38
  39
  40SEQUENTIALLY CONSISTENT ATOMIC ACCESS
  41=====================================
  42
  43Most of the operations in the qemu/atomic.h header ensure *sequential
  44consistency*, where "the result of any execution is the same as if the
  45operations of all the processors were executed in some sequential order,
  46and the operations of each individual processor appear in this sequence
  47in the order specified by its program".
  48
  49qemu/atomic.h provides the following set of atomic read-modify-write
  50operations:
  51
  52    void atomic_inc(ptr)
  53    void atomic_dec(ptr)
  54    void atomic_add(ptr, val)
  55    void atomic_sub(ptr, val)
  56    void atomic_and(ptr, val)
  57    void atomic_or(ptr, val)
  58
  59    typeof(*ptr) atomic_fetch_inc(ptr)
  60    typeof(*ptr) atomic_fetch_dec(ptr)
  61    typeof(*ptr) atomic_fetch_add(ptr, val)
  62    typeof(*ptr) atomic_fetch_sub(ptr, val)
  63    typeof(*ptr) atomic_fetch_and(ptr, val)
  64    typeof(*ptr) atomic_fetch_or(ptr, val)
  65    typeof(*ptr) atomic_xchg(ptr, val)
  66    typeof(*ptr) atomic_cmpxchg(ptr, old, new)
  67
  68all of which return the old value of *ptr.  These operations are
  69polymorphic; they operate on any type that is as wide as an int.
  70
  71Sequentially consistent loads and stores can be done using:
  72
  73    atomic_fetch_add(ptr, 0) for loads
  74    atomic_xchg(ptr, val) for stores
  75
  76However, they are quite expensive on some platforms, notably POWER and
  77ARM.  Therefore, qemu/atomic.h provides two primitives with slightly
  78weaker constraints:
  79
  80    typeof(*ptr) atomic_mb_read(ptr)
  81    void         atomic_mb_set(ptr, val)
  82
  83The semantics of these primitives map to Java volatile variables,
  84and are strongly related to memory barriers as used in the Linux
  85kernel (see below).
  86
  87As long as you use atomic_mb_read and atomic_mb_set, accesses cannot
  88be reordered with each other, and it is also not possible to reorder
  89"normal" accesses around them.
  90
  91However, and this is the important difference between
  92atomic_mb_read/atomic_mb_set and sequential consistency, it is important
  93for both threads to access the same volatile variable.  It is not the
  94case that everything visible to thread A when it writes volatile field f
  95becomes visible to thread B after it reads volatile field g. The store
  96and load have to "match" (i.e., be performed on the same volatile
  97field) to achieve the right semantics.
  98
  99
 100These operations operate on any type that is as wide as an int or smaller.
 101
 102
 103WEAK ATOMIC ACCESS AND MANUAL MEMORY BARRIERS
 104=============================================
 105
 106Compared to sequentially consistent atomic access, programming with
 107weaker consistency models can be considerably more complicated.
 108In general, if the algorithm you are writing includes both writes
 109and reads on the same side, it is generally simpler to use sequentially
 110consistent primitives.
 111
 112When using this model, variables are accessed with atomic_read() and
 113atomic_set(), and restrictions to the ordering of accesses is enforced
 114using the smp_rmb(), smp_wmb(), smp_mb() and smp_read_barrier_depends()
 115memory barriers.
 116
 117atomic_read() and atomic_set() prevents the compiler from using
 118optimizations that might otherwise optimize accesses out of existence
 119on the one hand, or that might create unsolicited accesses on the other.
 120In general this should not have any effect, because the same compiler
 121barriers are already implied by memory barriers.  However, it is useful
 122to do so, because it tells readers which variables are shared with
 123other threads, and which are local to the current thread or protected
 124by other, more mundane means.
 125
 126Memory barriers control the order of references to shared memory.
 127They come in four kinds:
 128
 129- smp_rmb() guarantees that all the LOAD operations specified before
 130  the barrier will appear to happen before all the LOAD operations
 131  specified after the barrier with respect to the other components of
 132  the system.
 133
 134  In other words, smp_rmb() puts a partial ordering on loads, but is not
 135  required to have any effect on stores.
 136
 137- smp_wmb() guarantees that all the STORE operations specified before
 138  the barrier will appear to happen before all the STORE operations
 139  specified after the barrier with respect to the other components of
 140  the system.
 141
 142  In other words, smp_wmb() puts a partial ordering on stores, but is not
 143  required to have any effect on loads.
 144
 145- smp_mb() guarantees that all the LOAD and STORE operations specified
 146  before the barrier will appear to happen before all the LOAD and
 147  STORE operations specified after the barrier with respect to the other
 148  components of the system.
 149
 150  smp_mb() puts a partial ordering on both loads and stores.  It is
 151  stronger than both a read and a write memory barrier; it implies both
 152  smp_rmb() and smp_wmb(), but it also prevents STOREs coming before the
 153  barrier from overtaking LOADs coming after the barrier and vice versa.
 154
 155- smp_read_barrier_depends() is a weaker kind of read barrier.  On
 156  most processors, whenever two loads are performed such that the
 157  second depends on the result of the first (e.g., the first load
 158  retrieves the address to which the second load will be directed),
 159  the processor will guarantee that the first LOAD will appear to happen
 160  before the second with respect to the other components of the system.
 161  However, this is not always true---for example, it was not true on
 162  Alpha processors.  Whenever this kind of access happens to shared
 163  memory (that is not protected by a lock), a read barrier is needed,
 164  and smp_read_barrier_depends() can be used instead of smp_rmb().
 165
 166  Note that the first load really has to have a _data_ dependency and not
 167  a control dependency.  If the address for the second load is dependent
 168  on the first load, but the dependency is through a conditional rather
 169  than actually loading the address itself, then it's a _control_
 170  dependency and a full read barrier or better is required.
 171
 172
 173This is the set of barriers that is required *between* two atomic_read()
 174and atomic_set() operations to achieve sequential consistency:
 175
 176                    |               2nd operation             |
 177                    |-----------------------------------------|
 178     1st operation  | (after last) | atomic_read | atomic_set |
 179     ---------------+--------------+-------------+------------|
 180     (before first) |              | none        | smp_wmb()  |
 181     ---------------+--------------+-------------+------------|
 182     atomic_read    | smp_rmb()    | smp_rmb()*  | **         |
 183     ---------------+--------------+-------------+------------|
 184     atomic_set     | none         | smp_mb()*** | smp_wmb()  |
 185     ---------------+--------------+-------------+------------|
 186
 187       * Or smp_read_barrier_depends().
 188
 189      ** This requires a load-store barrier.  How to achieve this varies
 190         depending on the machine, but in practice smp_rmb()+smp_wmb()
 191         should have the desired effect.  For example, on PowerPC the
 192         lwsync instruction is a combined load-load, load-store and
 193         store-store barrier.
 194
 195     *** This requires a store-load barrier.  On most machines, the only
 196         way to achieve this is a full barrier.
 197
 198
 199You can see that the two possible definitions of atomic_mb_read()
 200and atomic_mb_set() are the following:
 201
 202    1) atomic_mb_read(p)   = atomic_read(p); smp_rmb()
 203       atomic_mb_set(p, v) = smp_wmb(); atomic_set(p, v); smp_mb()
 204
 205    2) atomic_mb_read(p)   = smp_mb() atomic_read(p); smp_rmb()
 206       atomic_mb_set(p, v) = smp_wmb(); atomic_set(p, v);
 207
 208Usually the former is used, because smp_mb() is expensive and a program
 209normally has more reads than writes.  Therefore it makes more sense to
 210make atomic_mb_set() the more expensive operation.
 211
 212There are two common cases in which atomic_mb_read and atomic_mb_set
 213generate too many memory barriers, and thus it can be useful to manually
 214place barriers instead:
 215
 216- when a data structure has one thread that is always a writer
 217  and one thread that is always a reader, manual placement of
 218  memory barriers makes the write side faster.  Furthermore,
 219  correctness is easy to check for in this case using the "pairing"
 220  trick that is explained below:
 221
 222     thread 1                                thread 1
 223     -------------------------               ------------------------
 224     (other writes)
 225                                             smp_wmb()
 226     atomic_mb_set(&a, x)                    atomic_set(&a, x)
 227                                             smp_wmb()
 228     atomic_mb_set(&b, y)                    atomic_set(&b, y)
 229
 230                                       =>
 231     thread 2                                thread 2
 232     -------------------------               ------------------------
 233     y = atomic_mb_read(&b)                  y = atomic_read(&b)
 234                                             smp_rmb()
 235     x = atomic_mb_read(&a)                  x = atomic_read(&a)
 236                                             smp_rmb()
 237
 238- sometimes, a thread is accessing many variables that are otherwise
 239  unrelated to each other (for example because, apart from the current
 240  thread, exactly one other thread will read or write each of these
 241  variables).  In this case, it is possible to "hoist" the implicit
 242  barriers provided by atomic_mb_read() and atomic_mb_set() outside
 243  a loop.  For example, the above definition atomic_mb_read() gives
 244  the following transformation:
 245
 246     n = 0;                                  n = 0;
 247     for (i = 0; i < 10; i++)          =>    for (i = 0; i < 10; i++)
 248       n += atomic_mb_read(&a[i]);             n += atomic_read(&a[i]);
 249                                             smp_rmb();
 250
 251  Similarly, atomic_mb_set() can be transformed as follows:
 252  smp_mb():
 253
 254                                             smp_wmb();
 255     for (i = 0; i < 10; i++)          =>    for (i = 0; i < 10; i++)
 256       atomic_mb_set(&a[i], false);            atomic_set(&a[i], false);
 257                                             smp_mb();
 258
 259
 260The two tricks can be combined.  In this case, splitting a loop in
 261two lets you hoist the barriers out of the loops _and_ eliminate the
 262expensive smp_mb():
 263
 264                                             smp_wmb();
 265     for (i = 0; i < 10; i++) {        =>    for (i = 0; i < 10; i++)
 266       atomic_mb_set(&a[i], false);            atomic_set(&a[i], false);
 267       atomic_mb_set(&b[i], false);          smb_wmb();
 268     }                                       for (i = 0; i < 10; i++)
 269                                               atomic_set(&a[i], false);
 270                                             smp_mb();
 271
 272  The other thread can still use atomic_mb_read()/atomic_mb_set()
 273
 274
 275Memory barrier pairing
 276----------------------
 277
 278A useful rule of thumb is that memory barriers should always, or almost
 279always, be paired with another barrier.  In the case of QEMU, however,
 280note that the other barrier may actually be in a driver that runs in
 281the guest!
 282
 283For the purposes of pairing, smp_read_barrier_depends() and smp_rmb()
 284both count as read barriers.  A read barrier shall pair with a write
 285barrier or a full barrier; a write barrier shall pair with a read
 286barrier or a full barrier.  A full barrier can pair with anything.
 287For example:
 288
 289        thread 1             thread 2
 290        ===============      ===============
 291        a = 1;
 292        smp_wmb();
 293        b = 2;               x = b;
 294                             smp_rmb();
 295                             y = a;
 296
 297Note that the "writing" thread is accessing the variables in the
 298opposite order as the "reading" thread.  This is expected: stores
 299before the write barrier will normally match the loads after the
 300read barrier, and vice versa.  The same is true for more than 2
 301access and for data dependency barriers:
 302
 303        thread 1             thread 2
 304        ===============      ===============
 305        b[2] = 1;
 306        smp_wmb();
 307        x->i = 2;
 308        smp_wmb();
 309        a = x;               x = a;
 310                             smp_read_barrier_depends();
 311                             y = x->i;
 312                             smp_read_barrier_depends();
 313                             z = b[y];
 314
 315smp_wmb() also pairs with atomic_mb_read(), and smp_rmb() also pairs
 316with atomic_mb_set().
 317
 318
 319COMPARISON WITH LINUX KERNEL MEMORY BARRIERS
 320============================================
 321
 322Here is a list of differences between Linux kernel atomic operations
 323and memory barriers, and the equivalents in QEMU:
 324
 325- atomic operations in Linux are always on a 32-bit int type and
 326  use a boxed atomic_t type; atomic operations in QEMU are polymorphic
 327  and use normal C types.
 328
 329- atomic_read and atomic_set in Linux give no guarantee at all;
 330  atomic_read and atomic_set in QEMU include a compiler barrier
 331  (similar to the READ_ONCE/WRITE_ONCE macros in Linux).
 332
 333- most atomic read-modify-write operations in Linux return void;
 334  in QEMU, all of them return the old value of the variable.
 335
 336- different atomic read-modify-write operations in Linux imply
 337  a different set of memory barriers; in QEMU, all of them enforce
 338  sequential consistency, which means they imply full memory barriers
 339  before and after the operation.
 340
 341- Linux does not have an equivalent of atomic_mb_read() and
 342  atomic_mb_set().  In particular, note that set_mb() is a little
 343  weaker than atomic_mb_set().
 344
 345
 346SOURCES
 347=======
 348
 349* Documentation/memory-barriers.txt from the Linux kernel
 350
 351* "The JSR-133 Cookbook for Compiler Writers", available at
 352  http://g.oswego.edu/dl/jmm/cookbook.html
 353