1/*-------------------------------------------------------------------------
6 * Hardware and compiler dependent functions for manipulating memory
7 * atomically and dealing with cache coherency. Used to implement locking
8 * facilities and lockless algorithms/data structures.
10 * To bring up postgres on a platform/compiler at the very least
11 * implementations for the following operations should be provided:
12 * * pg_compiler_barrier(), pg_write_barrier(), pg_read_barrier()
13 * * pg_atomic_compare_exchange_u32(), pg_atomic_fetch_add_u32()
14 * * pg_atomic_test_set_flag(), pg_atomic_init_flag(), pg_atomic_clear_flag()
15 * * PG_HAVE_8BYTE_SINGLE_COPY_ATOMICITY should be defined if appropriate.
17 * There exist generic, hardware independent, implementations for several
18 * compilers which might be sufficient, although possibly not optimal, for a
19 * new platform. If no such generic implementation is available spinlocks will
20 * be used to implement the 64-bit parts of the API.
22 * Implement _u64 atomics if and only if your platform can use them
23 * efficiently (and obviously correctly).
25 * Use higher level functionality (lwlocks, spinlocks, heavyweight locks)
26 * whenever possible. Writing correct code using these facilities is hard.
28 * For an introduction to using memory barriers within the PostgreSQL backend,
29 * see src/backend/storage/lmgr/README.barrier
31 * Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group
32 * Portions Copyright (c) 1994, Regents of the University of California
34 * src/include/port/atomics.h
36 *-------------------------------------------------------------------------
42#error "atomics.h may not be included from frontend code"
45 #define INSIDE_ATOMICS_H
50 * First a set of architecture specific files is included.
52 * These files can provide the full set of atomics or can do pretty much
53 * nothing if all the compilers commonly used on these platforms provide
56 * Don't add an inline assembly of the actual atomic operations if all the
57 * common implementations of your platform provide intrinsics. Intrinsics are
58 * much easier to understand and potentially support more architectures.
60 * It will often make sense to define memory barrier semantics here, since
61 * e.g. generic compiler intrinsics for x86 memory barriers can't know that
62 * postgres doesn't need x86 read/write barriers do anything more than a
66#if defined(__arm__) || defined(__arm) || defined(__aarch64__)
68#elif defined(__i386__) || defined(__i386) || defined(__x86_64__)
70#elif defined(__ppc__) || defined(__powerpc__) || defined(__ppc64__) || defined(__powerpc64__)
75 * Compiler specific, but architecture independent implementations.
77 * Provide architecture independent implementations of the atomic
78 * facilities. At the very least compiler barriers should be provided, but a
79 * full implementation of
80 * * pg_compiler_barrier(), pg_write_barrier(), pg_read_barrier()
81 * * pg_atomic_compare_exchange_u32(), pg_atomic_fetch_add_u32()
82 * using compiler intrinsics are a good idea.
85 * gcc or compatible, including clang and icc.
87#if defined(__GNUC__) || defined(__INTEL_COMPILER)
89#elif defined(_MSC_VER)
92/* Unknown compiler. */
95/* Fail if we couldn't find implementations of required facilities. */
96#if !defined(PG_HAVE_ATOMIC_U32_SUPPORT)
97#error "could not find an implementation of pg_atomic_uint32"
99#if !defined(pg_compiler_barrier_impl)
100#error "could not find an implementation of pg_compiler_barrier"
102#if !defined(pg_memory_barrier_impl)
103#error "could not find an implementation of pg_memory_barrier_impl"
108 * Provide a spinlock-based implementation of the 64 bit variants, if
114 * Provide additional operations using supported infrastructure. These are
115 * expected to be efficient if the underlying atomic operations are efficient.
121 * pg_compiler_barrier - prevent the compiler from moving code across
123 * A compiler barrier need not (and preferably should not) emit any actual
124 * machine code, but must act as an optimization fence: the compiler must not
125 * reorder loads or stores to main memory around the barrier. However, the
126 * CPU may still reorder loads or stores at runtime, if the architecture's
127 * memory model permits this.
129 #define pg_compiler_barrier() pg_compiler_barrier_impl()
132 * pg_memory_barrier - prevent the CPU from reordering memory access
134 * A memory barrier must act as a compiler barrier, and in addition must
135 * guarantee that all loads and stores issued prior to the barrier are
136 * completed before any loads or stores issued after the barrier. Unless
137 * loads and stores are totally ordered (which is not the case on most
138 * architectures) this requires issuing some sort of memory fencing
141 #define pg_memory_barrier() pg_memory_barrier_impl()
144 * pg_(read|write)_barrier - prevent the CPU from reordering memory access
146 * A read barrier must act as a compiler barrier, and in addition must
147 * guarantee that any loads issued prior to the barrier are completed before
148 * any loads issued after the barrier. Similarly, a write barrier acts
149 * as a compiler barrier, and also orders stores. Read and write barriers
150 * are thus weaker than a full memory barrier, but stronger than a compiler
151 * barrier. In practice, on machines with strong memory ordering, read and
152 * write barriers may require nothing more than a compiler barrier.
154 #define pg_read_barrier() pg_read_barrier_impl()
155 #define pg_write_barrier() pg_write_barrier_impl()
158 * Spinloop delay - Allow CPU to relax in busy loops
160 #define pg_spin_delay() pg_spin_delay_impl()
163 * pg_atomic_init_flag - initialize atomic flag.
165 * No barrier semantics.
170 pg_atomic_init_flag_impl(ptr);
174 * pg_atomic_test_set_flag - TAS()
176 * Returns true if the flag has successfully been set, false otherwise.
178 * Acquire (including read barrier) semantics.
183 return pg_atomic_test_set_flag_impl(ptr);
187 * pg_atomic_unlocked_test_flag - Check if the lock is free
189 * Returns true if the flag currently is not set, false otherwise.
191 * No barrier semantics.
196 return pg_atomic_unlocked_test_flag_impl(ptr);
200 * pg_atomic_clear_flag - release lock set by TAS()
202 * Release (including write barrier) semantics.
207 pg_atomic_clear_flag_impl(ptr);
212 * pg_atomic_init_u32 - initialize atomic variable
214 * Has to be done before any concurrent usage..
216 * No barrier semantics.
227 * pg_atomic_read_u32 - unlocked read from atomic variable.
229 * The read is guaranteed to return a value as it has been written by this or
230 * another process at some point in the past. There's however no cache
231 * coherency interaction guaranteeing the value hasn't since been written to
234 * No barrier semantics.
244 * pg_atomic_read_membarrier_u32 - read with barrier semantics.
246 * This read is guaranteed to return the current value, provided that the value
247 * is only ever updated via operations with barrier semantics, such as
248 * pg_atomic_compare_exchange_u32() and pg_atomic_write_membarrier_u32().
249 * While this may be less performant than pg_atomic_read_u32(), it may be
250 * easier to reason about correctness with this function in less performance-
253 * Full barrier semantics.
260 return pg_atomic_read_membarrier_u32_impl(ptr);
264 * pg_atomic_write_u32 - write to atomic variable.
266 * The write is guaranteed to succeed as a whole, i.e. it's not possible to
267 * observe a partial write for any reader. Note that this correctly interacts
268 * with pg_atomic_compare_exchange_u32, in contrast to
269 * pg_atomic_unlocked_write_u32().
271 * No barrier semantics.
282 * pg_atomic_unlocked_write_u32 - unlocked write to atomic variable.
284 * The write is guaranteed to succeed as a whole, i.e. it's not possible to
285 * observe a partial write for any reader. But note that writing this way is
286 * not guaranteed to correctly interact with read-modify-write operations like
287 * pg_atomic_compare_exchange_u32. This should only be used in cases where
288 * minor performance regressions due to atomics emulation are unacceptable.
290 * No barrier semantics.
301 * pg_atomic_write_membarrier_u32 - write with barrier semantics.
303 * The write is guaranteed to succeed as a whole, i.e., it's not possible to
304 * observe a partial write for any reader. Note that this correctly interacts
305 * with both pg_atomic_compare_exchange_u32() and
306 * pg_atomic_read_membarrier_u32(). While this may be less performant than
307 * pg_atomic_write_u32(), it may be easier to reason about correctness with
308 * this function in less performance-sensitive code.
310 * Full barrier semantics.
317 pg_atomic_write_membarrier_u32_impl(ptr,
val);
321 * pg_atomic_exchange_u32 - exchange newval with current value
323 * Returns the old value of 'ptr' before the swap.
325 * Full barrier semantics.
336 * pg_atomic_compare_exchange_u32 - CAS operation
338 * Atomically compare the current value of ptr with *expected and store newval
339 * iff ptr and *expected have the same value. The current value of *ptr will
340 * always be stored in *expected.
342 * Return true if values have been exchanged, false otherwise.
344 * Full barrier semantics.
357 * pg_atomic_fetch_add_u32 - atomically add to variable
359 * Returns the value of ptr before the arithmetic operation.
361 * Full barrier semantics.
371 * pg_atomic_fetch_sub_u32 - atomically subtract from variable
373 * Returns the value of ptr before the arithmetic operation. Note that sub_
374 * may not be INT_MIN due to platform limitations.
376 * Full barrier semantics.
383 return pg_atomic_fetch_sub_u32_impl(ptr, sub_);
387 * pg_atomic_fetch_and_u32 - atomically bit-and and_ with variable
389 * Returns the value of ptr before the arithmetic operation.
391 * Full barrier semantics.
397 return pg_atomic_fetch_and_u32_impl(ptr, and_);
401 * pg_atomic_fetch_or_u32 - atomically bit-or or_ with variable
403 * Returns the value of ptr before the arithmetic operation.
405 * Full barrier semantics.
411 return pg_atomic_fetch_or_u32_impl(ptr, or_);
415 * pg_atomic_add_fetch_u32 - atomically add to variable
417 * Returns the value of ptr after the arithmetic operation.
419 * Full barrier semantics.
425 return pg_atomic_add_fetch_u32_impl(ptr, add_);
429 * pg_atomic_sub_fetch_u32 - atomically subtract from variable
431 * Returns the value of ptr after the arithmetic operation. Note that sub_ may
432 * not be INT_MIN due to platform limitations.
434 * Full barrier semantics.
441 return pg_atomic_sub_fetch_u32_impl(ptr, sub_);
445 * The 64 bit operations have the same semantics as their 32bit counterparts
446 * if they are available. Check the corresponding 32bit function for
454 * Can't necessarily enforce alignment - and don't need it - when using
455 * the spinlock based fallback implementation. Therefore only assert when
458#ifndef PG_HAVE_ATOMIC_U64_SIMULATION
467#ifndef PG_HAVE_ATOMIC_U64_SIMULATION
476#ifndef PG_HAVE_ATOMIC_U64_SIMULATION
479 return pg_atomic_read_membarrier_u64_impl(ptr);
485#ifndef PG_HAVE_ATOMIC_U64_SIMULATION
494#ifndef PG_HAVE_ATOMIC_U64_SIMULATION
497 pg_atomic_write_membarrier_u64_impl(ptr,
val);
503#ifndef PG_HAVE_ATOMIC_U64_SIMULATION
506 return pg_atomic_exchange_u64_impl(ptr,
newval);
513#ifndef PG_HAVE_ATOMIC_U64_SIMULATION
522#ifndef PG_HAVE_ATOMIC_U64_SIMULATION
531#ifndef PG_HAVE_ATOMIC_U64_SIMULATION
535 return pg_atomic_fetch_sub_u64_impl(ptr, sub_);
541#ifndef PG_HAVE_ATOMIC_U64_SIMULATION
544 return pg_atomic_fetch_and_u64_impl(ptr, and_);
550#ifndef PG_HAVE_ATOMIC_U64_SIMULATION
553 return pg_atomic_fetch_or_u64_impl(ptr, or_);
559#ifndef PG_HAVE_ATOMIC_U64_SIMULATION
562 return pg_atomic_add_fetch_u64_impl(ptr, add_);
568#ifndef PG_HAVE_ATOMIC_U64_SIMULATION
572 return pg_atomic_sub_fetch_u64_impl(ptr, sub_);
576 * Monotonically advance the given variable using only atomic operations until
577 * it's at least the target value. Returns the latest value observed, which
578 * may or may not be the target value.
580 * Full barrier semantics (even when value is unchanged).
587#ifndef PG_HAVE_ATOMIC_U64_SIMULATION
592 if (currval >= target)
598 while (currval < target)
607#undef INSIDE_ATOMICS_H
609#endif /* ATOMICS_H */
static bool pg_atomic_compare_exchange_u32_impl(volatile pg_atomic_uint32 *ptr, uint32 *expected, uint32 newval)
static uint32 pg_atomic_fetch_add_u32_impl(volatile pg_atomic_uint32 *ptr, int32 add_)
uint64 pg_atomic_fetch_add_u64_impl(volatile pg_atomic_uint64 *ptr, int64 add_)
void pg_atomic_init_u64_impl(volatile pg_atomic_uint64 *ptr, uint64 val_)
bool pg_atomic_compare_exchange_u64_impl(volatile pg_atomic_uint64 *ptr, uint64 *expected, uint64 newval)
static uint32 pg_atomic_fetch_and_u32(volatile pg_atomic_uint32 *ptr, uint32 and_)
static bool pg_atomic_compare_exchange_u32(volatile pg_atomic_uint32 *ptr, uint32 *expected, uint32 newval)
static void pg_atomic_write_u64(volatile pg_atomic_uint64 *ptr, uint64 val)
static void pg_atomic_clear_flag(volatile pg_atomic_flag *ptr)
static uint32 pg_atomic_fetch_or_u32(volatile pg_atomic_uint32 *ptr, uint32 or_)
static uint32 pg_atomic_sub_fetch_u32(volatile pg_atomic_uint32 *ptr, int32 sub_)
#define pg_memory_barrier()
static uint32 pg_atomic_read_membarrier_u32(volatile pg_atomic_uint32 *ptr)
static uint32 pg_atomic_fetch_sub_u32(volatile pg_atomic_uint32 *ptr, int32 sub_)
static bool pg_atomic_compare_exchange_u64(volatile pg_atomic_uint64 *ptr, uint64 *expected, uint64 newval)
static void pg_atomic_unlocked_write_u32(volatile pg_atomic_uint32 *ptr, uint32 val)
static void pg_atomic_init_u32(volatile pg_atomic_uint32 *ptr, uint32 val)
static uint64 pg_atomic_read_membarrier_u64(volatile pg_atomic_uint64 *ptr)
static uint32 pg_atomic_fetch_add_u32(volatile pg_atomic_uint32 *ptr, int32 add_)
static uint32 pg_atomic_add_fetch_u32(volatile pg_atomic_uint32 *ptr, int32 add_)
static uint64 pg_atomic_monotonic_advance_u64(volatile pg_atomic_uint64 *ptr, uint64 target)
static uint64 pg_atomic_fetch_add_u64(volatile pg_atomic_uint64 *ptr, int64 add_)
static bool pg_atomic_test_set_flag(volatile pg_atomic_flag *ptr)
static uint64 pg_atomic_sub_fetch_u64(volatile pg_atomic_uint64 *ptr, int64 sub_)
static bool pg_atomic_unlocked_test_flag(volatile pg_atomic_flag *ptr)
static void pg_atomic_write_u32(volatile pg_atomic_uint32 *ptr, uint32 val)
static void pg_atomic_write_membarrier_u32(volatile pg_atomic_uint32 *ptr, uint32 val)
static uint64 pg_atomic_fetch_and_u64(volatile pg_atomic_uint64 *ptr, uint64 and_)
static uint32 pg_atomic_read_u32(volatile pg_atomic_uint32 *ptr)
static uint64 pg_atomic_fetch_or_u64(volatile pg_atomic_uint64 *ptr, uint64 or_)
static uint64 pg_atomic_add_fetch_u64(volatile pg_atomic_uint64 *ptr, int64 add_)
static uint32 pg_atomic_exchange_u32(volatile pg_atomic_uint32 *ptr, uint32 newval)
static void pg_atomic_init_u64(volatile pg_atomic_uint64 *ptr, uint64 val)
static void pg_atomic_write_membarrier_u64(volatile pg_atomic_uint64 *ptr, uint64 val)
static uint64 pg_atomic_read_u64(volatile pg_atomic_uint64 *ptr)
static uint64 pg_atomic_fetch_sub_u64(volatile pg_atomic_uint64 *ptr, int64 sub_)
static uint64 pg_atomic_exchange_u64(volatile pg_atomic_uint64 *ptr, uint64 newval)
static void pg_atomic_init_flag(volatile pg_atomic_flag *ptr)
#define AssertPointerAlignment(ptr, bndr)
static uint32 pg_atomic_exchange_u32_impl(volatile pg_atomic_uint32 *ptr, uint32 newval)
static void pg_atomic_unlocked_write_u32_impl(volatile pg_atomic_uint32 *ptr, uint32 val)
static uint64 pg_atomic_read_u64_impl(volatile pg_atomic_uint64 *ptr)
static void pg_atomic_write_u32_impl(volatile pg_atomic_uint32 *ptr, uint32 val)
static void pg_atomic_write_u64_impl(volatile pg_atomic_uint64 *ptr, uint64 val)
static uint32 pg_atomic_read_u32_impl(volatile pg_atomic_uint32 *ptr)
static void pg_atomic_init_u32_impl(volatile pg_atomic_uint32 *ptr, uint32 val_)
Assert(PointerIsAligned(start, uint64))