/* Macros for atomic functionality for tile. Copyright (C) 2011-2020 Free Software Foundation, Inc. Contributed by Walter Lee (walt@tilera.com) This file is free software; you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation; either version 3, or (at your option) any later version. This file is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details. Under Section 7 of GPL version 3, you are granted additional permissions described in the GCC Runtime Library Exception, version 3.1, as published by the Free Software Foundation. You should have received a copy of the GNU General Public License and a copy of the GCC Runtime Library Exception along with this program; see the files COPYING3 and COPYING.RUNTIME respectively. If not, see . */ /* Provides macros for common atomic functionality. */ #ifndef _ATOMIC_H_ #define _ATOMIC_H_ #ifdef __tilegx__ /* Atomic instruction macros The macros provided by atomic.h simplify access to the TILE-Gx architecture's atomic instructions. The architecture provides a variety of atomic instructions, including "exchange", "compare and exchange", "fetch and ADD", "fetch and AND", "fetch and OR", and "fetch and ADD if greater than or equal to zero". No barrier or fence semantics are implied by any of the atomic instructions for manipulating memory; you must specify the barriers that you wish explicitly, using the provided macros. Any integral 32- or 64-bit value can be used as the argument to these macros, such as "int", "long long", "unsigned long", etc. The pointers must be aligned to 4 or 8 bytes for 32- or 64-bit data. The "exchange" and "compare and exchange" macros may also take pointer values. We use the pseudo-type "VAL" in the documentation to indicate the use of an appropriate type. */ #else /* Atomic instruction macros The macros provided by atomic.h simplify access to the Tile architecture's atomic instructions. Since the architecture supports test-and-set as its only in-silicon atomic operation, many of the operations provided by this header are implemented as fast-path calls to Linux emulation routines. Using the kernel for atomic operations allows userspace to take advantage of the kernel's existing atomic-integer support (managed by a distributed array of locks). The kernel provides proper ordering among simultaneous atomic operations on different cores, and guarantees a process cannot be context-switched part way through an atomic operation. By virtue of sharing the kernel atomic implementation, the userspace atomic operations are compatible with the atomic methods provided by the kernel's futex() syscall API. Note that these operations never cause Linux kernel scheduling, and are in fact invisible to the kernel; they simply act as regular function calls but with an elevated privilege level. Note that the kernel's distributed lock array is hashed by using only VA bits from the atomic value's address (to avoid the performance hit of page table locking and multiple page-table lookups to get the PA) and only the VA bits that are below page granularity (to properly lock simultaneous accesses to the same page mapped at different VAs). As a result, simultaneous atomic operations on values whose addresses are at the same offset on a page will contend in the kernel for the same lock array element. No barrier or fence semantics are implied by any of the atomic instructions for manipulating memory; you must specify the barriers that you wish explicitly, using the provided macros. Any integral 32- or 64-bit value can be used as the argument to these macros, such as "int", "long long", "unsigned long", etc. The pointers must be aligned to 4 or 8 bytes for 32- or 64-bit data. The "exchange" and "compare and exchange" macros may also take pointer values. We use the pseudo-type "VAL" in the documentation to indicate the use of an appropriate type. The 32-bit routines are implemented using a single kernel fast syscall, as is the 64-bit compare-and-exchange. The other 64-bit routines are implemented by looping over the 64-bit compare-and-exchange routine, so may be potentially less efficient. */ #endif #ifdef __tilegx__ #define SPR_CMPEXCH_VALUE 0x2780 #else #define __NR_FAST_cmpxchg -1 #define __NR_FAST_atomic_update -2 #define __NR_FAST_cmpxchg64 -3 #endif /* 32-bit integer compare-and-exchange. */ static __inline __attribute__ ((always_inline)) int arch_atomic_val_compare_and_exchange_4 (volatile int *mem, int oldval, int newval) { #ifdef __tilegx__ __insn_mtspr (SPR_CMPEXCH_VALUE, oldval); return __insn_cmpexch4 (mem, newval); #else int result; __asm__ __volatile__ ("swint1":"=R00" (result), "=m" (*mem):"R10" (__NR_FAST_cmpxchg), "R00" (mem), "R01" (oldval), "R02" (newval), "m" (*mem):"r20", "r21", "r22", "r23", "r24", "r25", "r26", "r27", "r28", "r29", "memory"); return result; #endif } /* 64-bit integer compare-and-exchange. */ static __inline __attribute__ ((always_inline)) long long arch_atomic_val_compare_and_exchange_8 (volatile long long *mem, long long oldval, long long newval) { #ifdef __tilegx__ __insn_mtspr (SPR_CMPEXCH_VALUE, oldval); return __insn_cmpexch (mem, newval); #else unsigned int result_lo, result_hi; unsigned int oldval_lo = oldval & 0xffffffffu, oldval_hi = oldval >> 32; unsigned int newval_lo = newval & 0xffffffffu, newval_hi = newval >> 32; __asm__ __volatile__ ("swint1":"=R00" (result_lo), "=R01" (result_hi), "=m" (*mem):"R10" (__NR_FAST_cmpxchg64), "R00" (mem), "R02" (oldval_lo), "R03" (oldval_hi), "R04" (newval_lo), "R05" (newval_hi), "m" (*mem):"r20", "r21", "r22", "r23", "r24", "r25", "r26", "r27", "r28", "r29", "memory"); return ((long long) result_hi) << 32 | result_lo; #endif } /* This non-existent symbol is called for sizes other than "4" and "8", indicating a bug in the caller. */ extern int __arch_atomic_error_bad_argument_size (void) __attribute__ ((warning ("sizeof atomic argument not 4 or 8"))); #define arch_atomic_val_compare_and_exchange(mem, o, n) \ __extension__ ({ \ (__typeof(*(mem)))(__typeof(*(mem)-*(mem))) \ ((sizeof(*(mem)) == 8) ? \ arch_atomic_val_compare_and_exchange_8( \ (volatile long long*)(mem), (__typeof((o)-(o)))(o), \ (__typeof((n)-(n)))(n)) : \ (sizeof(*(mem)) == 4) ? \ arch_atomic_val_compare_and_exchange_4( \ (volatile int*)(mem), (__typeof((o)-(o)))(o), \ (__typeof((n)-(n)))(n)) : \ __arch_atomic_error_bad_argument_size()); \ }) #define arch_atomic_bool_compare_and_exchange(mem, o, n) \ __extension__ ({ \ __typeof(o) __o = (o); \ __builtin_expect( \ __o == arch_atomic_val_compare_and_exchange((mem), __o, (n)), 1); \ }) /* Loop with compare_and_exchange until we guess the correct value. Normally "expr" will be an expression using __old and __value. */ #define __arch_atomic_update_cmpxchg(mem, value, expr) \ __extension__ ({ \ __typeof(value) __value = (value); \ __typeof(*(mem)) *__mem = (mem), __old = *__mem, __guess; \ do { \ __guess = __old; \ __old = arch_atomic_val_compare_and_exchange(__mem, __old, (expr)); \ } while (__builtin_expect(__old != __guess, 0)); \ __old; \ }) #ifdef __tilegx__ /* Generic atomic op with 8- or 4-byte variant. The _mask, _addend, and _expr arguments are ignored on tilegx. */ #define __arch_atomic_update(mem, value, op, _mask, _addend, _expr) \ __extension__ ({ \ ((__typeof(*(mem))) \ ((sizeof(*(mem)) == 8) ? (__typeof(*(mem)-*(mem)))__insn_##op( \ (volatile void *)(mem), \ (long long)(__typeof((value)-(value)))(value)) : \ (sizeof(*(mem)) == 4) ? (int)__insn_##op##4( \ (volatile void *)(mem), \ (int)(__typeof((value)-(value)))(value)) : \ __arch_atomic_error_bad_argument_size())); \ }) #else /* This uses TILEPro's fast syscall support to atomically compute: int old = *ptr; *ptr = (old & mask) + addend; return old; This primitive can be used for atomic exchange, add, or, and. Only 32-bit support is provided. */ static __inline __attribute__ ((always_inline)) int __arch_atomic_update_4 (volatile int *mem, int mask, int addend) { int result; __asm__ __volatile__ ("swint1":"=R00" (result), "=m" (*mem):"R10" (__NR_FAST_atomic_update), "R00" (mem), "R01" (mask), "R02" (addend), "m" (*mem):"r20", "r21", "r22", "r23", "r24", "r25", "r26", "r27", "r28", "r29", "memory"); return result; } /* Generic atomic op with 8- or 4-byte variant. The _op argument is ignored on tilepro. */ #define __arch_atomic_update(mem, value, _op, mask, addend, expr) \ __extension__ ({ \ (__typeof(*(mem)))(__typeof(*(mem)-*(mem))) \ ((sizeof(*(mem)) == 8) ? \ __arch_atomic_update_cmpxchg((mem), (value), (expr)) : \ (sizeof(*(mem)) == 4) ? \ __arch_atomic_update_4((volatile int*)(mem), \ (__typeof((mask)-(mask)))(mask), \ (__typeof((addend)-(addend)))(addend)) : \ __arch_atomic_error_bad_argument_size()); \ }) #endif /* __tilegx__ */ #define arch_atomic_exchange(mem, newvalue) \ __arch_atomic_update(mem, newvalue, exch, 0, newvalue, __value) #define arch_atomic_add(mem, value) \ __arch_atomic_update(mem, value, fetchadd, -1, value, __old + __value) #define arch_atomic_sub(mem, value) arch_atomic_add((mem), -(value)) #define arch_atomic_increment(mem) arch_atomic_add((mem), 1) #define arch_atomic_decrement(mem) arch_atomic_add((mem), -1) #define arch_atomic_and(mem, mask) \ __arch_atomic_update(mem, mask, fetchand, mask, 0, __old & __value) #define arch_atomic_or(mem, mask) \ __arch_atomic_update(mem, mask, fetchor, ~mask, mask, __old | __value) #define arch_atomic_xor(mem, mask) \ __arch_atomic_update_cmpxchg(mem, mask, __old ^ __value) #define arch_atomic_nand(mem, mask) \ __arch_atomic_update_cmpxchg(mem, mask, ~(__old & __value)) #define arch_atomic_bit_set(mem, bit) \ __extension__ ({ \ __typeof(*(mem)) __mask = (__typeof(*(mem)))1 << (bit); \ __mask & arch_atomic_or((mem), __mask); \ }) #define arch_atomic_bit_clear(mem, bit) \ __extension__ ({ \ __typeof(*(mem)) __mask = (__typeof(*(mem)))1 << (bit); \ __mask & arch_atomic_and((mem), ~__mask); \ }) #ifdef __tilegx__ /* Atomically store a new value to memory. Note that you can freely use types of any size here, unlike the other atomic routines, which require 32- or 64-bit types. This accessor is provided for compatibility with TILEPro, which required an explicit atomic operation for stores that needed to be atomic with respect to other atomic methods in this header. */ #define arch_atomic_write(mem, value) ((void) (*(mem) = (value))) #else #define arch_atomic_write(mem, value) \ do { \ __typeof(mem) __aw_mem = (mem); \ __typeof(value) __aw_val = (value); \ unsigned int *__aw_mem32, __aw_intval, __aw_val32, __aw_off, __aw_mask; \ __aw_intval = (__typeof((value) - (value)))__aw_val; \ switch (sizeof(*__aw_mem)) { \ case 8: \ __arch_atomic_update_cmpxchg(__aw_mem, __aw_val, __value); \ break; \ case 4: \ __arch_atomic_update_4((int *)__aw_mem, 0, __aw_intval); \ break; \ case 2: \ __aw_off = 8 * ((long)__aw_mem & 0x2); \ __aw_mask = 0xffffU << __aw_off; \ __aw_mem32 = (unsigned int *)((long)__aw_mem & ~0x2); \ __aw_val32 = (__aw_intval << __aw_off) & __aw_mask; \ __arch_atomic_update_cmpxchg(__aw_mem32, __aw_val32, \ (__old & ~__aw_mask) | __value); \ break; \ case 1: \ __aw_off = 8 * ((long)__aw_mem & 0x3); \ __aw_mask = 0xffU << __aw_off; \ __aw_mem32 = (unsigned int *)((long)__aw_mem & ~0x3); \ __aw_val32 = (__aw_intval << __aw_off) & __aw_mask; \ __arch_atomic_update_cmpxchg(__aw_mem32, __aw_val32, \ (__old & ~__aw_mask) | __value); \ break; \ } \ } while (0) #endif /* Compiler barrier. This macro prevents loads or stores from being moved by the compiler across the macro. Any loaded value that was loaded before this macro must then be reloaded by the compiler. */ #define arch_atomic_compiler_barrier() __asm__ __volatile__("" ::: "memory") /* Full memory barrier. This macro has the semantics of arch_atomic_compiler_barrer(), but also ensures that previous stores are visible to other cores, and that all previous loaded values have been placed into their target register on this core. */ #define arch_atomic_full_barrier() __insn_mf() /* Read memory barrier. Ensure that all reads by this processor that occurred prior to the read memory barrier have completed, and that no reads that occur after the read memory barrier on this processor are initiated before the barrier. On current TILE chips a read barrier is implemented as a full barrier, but this may not be true in later versions of the architecture. See also arch_atomic_acquire_barrier() for the appropriate idiom to use to ensure no reads are lifted above an atomic lock instruction. */ #define arch_atomic_read_barrier() arch_atomic_full_barrier() /* Write memory barrier. Ensure that all writes by this processor that occurred prior to the write memory barrier have completed, and that no writes that occur after the write memory barrier on this processor are initiated before the barrier. On current TILE chips a write barrier is implemented as a full barrier, but this may not be true in later versions of the architecture. See also arch_atomic_release_barrier() for the appropriate idiom to use to ensure all writes are complete prior to an atomic unlock instruction. */ #define arch_atomic_write_barrier() arch_atomic_full_barrier() /* Lock acquisition barrier. Ensure that no load operations that follow this macro in the program can issue prior to the barrier. Without such a barrier, the compiler can reorder them to issue earlier, or the hardware can issue them speculatively. The latter is not currently done in the Tile microarchitecture, but using this operation improves portability to future implementations. This operation is intended to be used as part of the "acquire" path for locking, that is, when entering a critical section. This should be done after the atomic operation that actually acquires the lock, and in conjunction with a "control dependency" that checks the atomic operation result to see if the lock was in fact acquired. See the arch_atomic_read_barrier() macro for a heavier-weight barrier to use in certain unusual constructs, or arch_atomic_acquire_barrier_value() if no control dependency exists. */ #define arch_atomic_acquire_barrier() arch_atomic_compiler_barrier() /* Lock release barrier. Ensure that no store operations that precede this macro in the program complete subsequent to the barrier. Without such a barrier, the compiler can reorder stores to issue later, or stores can be still outstanding in the memory network. This operation is intended to be used as part of the "release" path for locking, that is, when leaving a critical section. This should be done before the operation (such as a store of zero) that actually releases the lock. */ #define arch_atomic_release_barrier() arch_atomic_write_barrier() /* Barrier until the read of a particular value is complete. This is occasionally useful when constructing certain locking scenarios. For example, you might write a routine that issues an atomic instruction to enter a critical section, then reads one or more values within the critical section without checking to see if the critical section was in fact acquired, and only later checks the atomic instruction result to see if the lock was acquired. If so the routine could properly release the lock and know that the values that were read were valid. In this scenario, it is required to wait for the result of the atomic instruction, even if the value itself is not checked. This guarantees that if the atomic instruction succeeded in taking the lock, the lock was held before any reads in the critical section issued. */ #define arch_atomic_acquire_barrier_value(val) \ __asm__ __volatile__("move %0, %0" :: "r"(val)) /* Access the given variable in memory exactly once. In some contexts, an algorithm may need to force access to memory, since otherwise the compiler may think it can optimize away a memory load or store; for example, in a loop when polling memory to see if another cpu has updated it yet. Generally this is only required for certain very carefully hand-tuned algorithms; using it unnecessarily may result in performance losses. A related use of this macro is to ensure that the compiler does not rematerialize the value of "x" by reloading it from memory unexpectedly; the "volatile" marking will prevent the compiler from being able to rematerialize. This is helpful if an algorithm needs to read a variable without locking, but needs it to have the same value if it ends up being used several times within the algorithm. Note that multiple uses of this macro are guaranteed to be ordered, i.e. the compiler will not reorder stores or loads that are wrapped in arch_atomic_access_once(). */ #define arch_atomic_access_once(x) (*(volatile __typeof(x) *)&(x)) #endif /* !_ATOMIC_H_ */