Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
kernel
os
linux
riscv: Improve zacas fully-ordered cmpxchg()
The current fully-ordered cmpxchgXX() implementation results in:
amocas.X.rl a5,a4,(s1)
fence rw,rw
This provides enough sync but we can actually use the following better
mapping instead:
amocas.X.aqrl a5,a4,(s1)
Suggested-by: Andrea Parri <andrea@rivosinc.com>
Signed-off-by: Alexandre Ghiti <alexghiti@rivosinc.com>
Reviewed-by: Andrew Jones <ajones@ventanamicro.com>
Reviewed-by: Andrea Parri <parri.andrea@gmail.com>
Link: https://lore.kernel.org/r/20241103145153.105097-7-alexghiti@rivosinc.com
Signed-off-by: Palmer Dabbelt <palmer@rivosinc.com>
authored by
Alexandre Ghiti
and committed by
Palmer Dabbelt
6116e22e
1658ef43
+64
-28
+64
-28
arch/riscv/include/asm/cmpxchg.h
+64
-28
arch/riscv/include/asm/cmpxchg.h
···
107
107
* store NEW in MEM. Return the initial value in MEM. Success is
108
108
* indicated by comparing RETURN with OLD.
109
109
*/
110
-
111
-
#define __arch_cmpxchg_masked(sc_sfx, cas_sfx, prepend, append, r, p, o, n) \
110
+
#define __arch_cmpxchg_masked(sc_sfx, cas_sfx, \
111
+
sc_prepend, sc_append, \
112
+
cas_prepend, cas_append, \
113
+
r, p, o, n) \
112
114
({ \
113
115
if (IS_ENABLED(CONFIG_RISCV_ISA_ZABHA) && \
114
116
IS_ENABLED(CONFIG_RISCV_ISA_ZACAS) && \
···
119
117
r = o; \
120
118
\
121
119
__asm__ __volatile__ ( \
122
-
prepend \
120
+
cas_prepend \
123
121
" amocas" cas_sfx " %0, %z2, %1\n" \
124
-
append \
122
+
cas_append \
125
123
: "+&r" (r), "+A" (*(p)) \
126
124
: "rJ" (n) \
127
125
: "memory"); \
···
136
134
ulong __rc; \
137
135
\
138
136
__asm__ __volatile__ ( \
139
-
prepend \
137
+
sc_prepend \
140
138
"0: lr.w %0, %2\n" \
141
139
" and %1, %0, %z5\n" \
142
140
" bne %1, %z3, 1f\n" \
···
144
142
" or %1, %1, %z4\n" \
145
143
" sc.w" sc_sfx " %1, %1, %2\n" \
146
144
" bnez %1, 0b\n" \
147
-
append \
145
+
sc_append \
148
146
"1:\n" \
149
147
: "=&r" (__retx), "=&r" (__rc), "+A" (*(__ptr32b)) \
150
148
: "rJ" ((long)__oldx), "rJ" (__newx), \
···
155
153
} \
156
154
})
157
155
158
-
#define __arch_cmpxchg(lr_sfx, sc_cas_sfx, prepend, append, r, p, co, o, n) \
156
+
#define __arch_cmpxchg(lr_sfx, sc_sfx, cas_sfx, \
157
+
sc_prepend, sc_append, \
158
+
cas_prepend, cas_append, \
159
+
r, p, co, o, n) \
159
160
({ \
160
161
if (IS_ENABLED(CONFIG_RISCV_ISA_ZACAS) && \
161
162
riscv_has_extension_unlikely(RISCV_ISA_EXT_ZACAS)) { \
162
163
r = o; \
163
164
\
164
165
__asm__ __volatile__ ( \
165
-
prepend \
166
-
" amocas" sc_cas_sfx " %0, %z2, %1\n" \
167
-
append \
166
+
cas_prepend \
167
+
" amocas" cas_sfx " %0, %z2, %1\n" \
168
+
cas_append \
168
169
: "+&r" (r), "+A" (*(p)) \
169
170
: "rJ" (n) \
170
171
: "memory"); \
···
175
170
register unsigned int __rc; \
176
171
\
177
172
__asm__ __volatile__ ( \
178
-
prepend \
173
+
sc_prepend \
179
174
"0: lr" lr_sfx " %0, %2\n" \
180
175
" bne %0, %z3, 1f\n" \
181
-
" sc" sc_cas_sfx " %1, %z4, %2\n" \
176
+
" sc" sc_sfx " %1, %z4, %2\n" \
182
177
" bnez %1, 0b\n" \
183
-
append \
178
+
sc_append \
184
179
"1:\n" \
185
180
: "=&r" (r), "=&r" (__rc), "+A" (*(p)) \
186
181
: "rJ" (co o), "rJ" (n) \
···
188
183
} \
189
184
})
190
185
191
-
#define _arch_cmpxchg(ptr, old, new, sc_cas_sfx, prepend, append) \
186
+
#define _arch_cmpxchg(ptr, old, new, sc_sfx, cas_sfx, \
187
+
sc_prepend, sc_append, \
188
+
cas_prepend, cas_append) \
192
189
({ \
193
190
__typeof__(ptr) __ptr = (ptr); \
194
191
__typeof__(*(__ptr)) __old = (old); \
···
199
192
\
200
193
switch (sizeof(*__ptr)) { \
201
194
case 1: \
202
-
__arch_cmpxchg_masked(sc_cas_sfx, ".b" sc_cas_sfx, \
203
-
prepend, append, \
204
-
__ret, __ptr, __old, __new); \
195
+
__arch_cmpxchg_masked(sc_sfx, ".b" cas_sfx, \
196
+
sc_prepend, sc_append, \
197
+
cas_prepend, cas_append, \
198
+
__ret, __ptr, __old, __new); \
205
199
break; \
206
200
case 2: \
207
-
__arch_cmpxchg_masked(sc_cas_sfx, ".h" sc_cas_sfx, \
208
-
prepend, append, \
209
-
__ret, __ptr, __old, __new); \
201
+
__arch_cmpxchg_masked(sc_sfx, ".h" cas_sfx, \
202
+
sc_prepend, sc_append, \
203
+
cas_prepend, cas_append, \
204
+
__ret, __ptr, __old, __new); \
210
205
break; \
211
206
case 4: \
212
-
__arch_cmpxchg(".w", ".w" sc_cas_sfx, prepend, append, \
213
-
__ret, __ptr, (long), __old, __new); \
207
+
__arch_cmpxchg(".w", ".w" sc_sfx, ".w" cas_sfx, \
208
+
sc_prepend, sc_append, \
209
+
cas_prepend, cas_append, \
210
+
__ret, __ptr, (long), __old, __new); \
214
211
break; \
215
212
case 8: \
216
-
__arch_cmpxchg(".d", ".d" sc_cas_sfx, prepend, append, \
217
-
__ret, __ptr, /**/, __old, __new); \
213
+
__arch_cmpxchg(".d", ".d" sc_sfx, ".d" cas_sfx, \
214
+
sc_prepend, sc_append, \
215
+
cas_prepend, cas_append, \
216
+
__ret, __ptr, /**/, __old, __new); \
218
217
break; \
219
218
default: \
220
219
BUILD_BUG(); \
···
228
215
(__typeof__(*(__ptr)))__ret; \
229
216
})
230
217
218
+
/*
219
+
* These macros are here to improve the readability of the arch_cmpxchg_XXX()
220
+
* macros.
221
+
*/
222
+
#define SC_SFX(x) x
223
+
#define CAS_SFX(x) x
224
+
#define SC_PREPEND(x) x
225
+
#define SC_APPEND(x) x
226
+
#define CAS_PREPEND(x) x
227
+
#define CAS_APPEND(x) x
228
+
231
229
#define arch_cmpxchg_relaxed(ptr, o, n) \
232
-
_arch_cmpxchg((ptr), (o), (n), "", "", "")
230
+
_arch_cmpxchg((ptr), (o), (n), \
231
+
SC_SFX(""), CAS_SFX(""), \
232
+
SC_PREPEND(""), SC_APPEND(""), \
233
+
CAS_PREPEND(""), CAS_APPEND(""))
233
234
234
235
#define arch_cmpxchg_acquire(ptr, o, n) \
235
-
_arch_cmpxchg((ptr), (o), (n), "", "", RISCV_ACQUIRE_BARRIER)
236
+
_arch_cmpxchg((ptr), (o), (n), \
237
+
SC_SFX(""), CAS_SFX(""), \
238
+
SC_PREPEND(""), SC_APPEND(RISCV_ACQUIRE_BARRIER), \
239
+
CAS_PREPEND(""), CAS_APPEND(RISCV_ACQUIRE_BARRIER))
236
240
237
241
#define arch_cmpxchg_release(ptr, o, n) \
238
-
_arch_cmpxchg((ptr), (o), (n), "", RISCV_RELEASE_BARRIER, "")
242
+
_arch_cmpxchg((ptr), (o), (n), \
243
+
SC_SFX(""), CAS_SFX(""), \
244
+
SC_PREPEND(RISCV_RELEASE_BARRIER), SC_APPEND(""), \
245
+
CAS_PREPEND(RISCV_RELEASE_BARRIER), CAS_APPEND(""))
239
246
240
247
#define arch_cmpxchg(ptr, o, n) \
241
-
_arch_cmpxchg((ptr), (o), (n), ".rl", "", " fence rw, rw\n")
248
+
_arch_cmpxchg((ptr), (o), (n), \
249
+
SC_SFX(".rl"), CAS_SFX(".aqrl"), \
250
+
SC_PREPEND(""), SC_APPEND(RISCV_FULL_BARRIER), \
251
+
CAS_PREPEND(""), CAS_APPEND(""))
242
252
243
253
#define arch_cmpxchg_local(ptr, o, n) \
244
254
arch_cmpxchg_relaxed((ptr), (o), (n))