Re: [PATCH v2 04/10] riscv: Improve amocas.X use in cmpxchg()
From: Andrea Parri <parri.andrea@gmail.com>
Date: 2024-06-27 13:31:29
Also in:
linux-arch, linux-riscv, lkml
On Wed, Jun 26, 2024 at 03:03:41PM +0200, Alexandre Ghiti wrote:
cmpxchg() uses amocas.X instructions from Zacas and Zabha but still uses the LR/SC acquire/release semantics which require barriers. Let's improve that by using proper amocas acquire/release semantics in order to avoid any of those barriers.
I can't really parse this changelog...
quoted hunk ↗ jump to hunk
Suggested-by: Andrea Parri <redacted> Signed-off-by: Alexandre Ghiti <redacted> --- arch/riscv/include/asm/cmpxchg.h | 60 ++++++++++++++++++-------------- 1 file changed, 33 insertions(+), 27 deletions(-)diff --git a/arch/riscv/include/asm/cmpxchg.h b/arch/riscv/include/asm/cmpxchg.h index b9a3fdcec919..3c65b00a0d36 100644 --- a/arch/riscv/include/asm/cmpxchg.h +++ b/arch/riscv/include/asm/cmpxchg.h@@ -105,7 +105,9 @@ * indicated by comparing RETURN with OLD. */ -#define __arch_cmpxchg_masked(sc_sfx, cas_sfx, prepend, append, r, p, o, n) \ +#define __arch_cmpxchg_masked(sc_sfx, cas_sfx, \ + sc_prepend, sc_append, \ + r, p, o, n) \ ({ \ __label__ no_zacas, zabha, end; \ \@@ -129,7 +131,7 @@ no_zacas:; \ ulong __rc; \ \ __asm__ __volatile__ ( \ - prepend \ + sc_prepend \ "0: lr.w %0, %2\n" \ " and %1, %0, %z5\n" \ " bne %1, %z3, 1f\n" \@@ -137,7 +139,7 @@ no_zacas:; \ " or %1, %1, %z4\n" \ " sc.w" sc_sfx " %1, %1, %2\n" \ " bnez %1, 0b\n" \ - append \ + sc_append \ "1:\n" \ : "=&r" (__retx), "=&r" (__rc), "+A" (*(__ptr32b)) \ : "rJ" ((long)__oldx), "rJ" (__newx), \@@ -150,9 +152,7 @@ no_zacas:; \ zabha: \ if (IS_ENABLED(CONFIG_RISCV_ISA_ZABHA)) { \ __asm__ __volatile__ ( \ - prepend \ " amocas" cas_sfx " %0, %z2, %1\n" \ - append \ : "+&r" (r), "+A" (*(p)) \ : "rJ" (n) \ : "memory"); \@@ -160,7 +160,9 @@ zabha: \ end:; \ }) -#define __arch_cmpxchg(lr_sfx, sc_cas_sfx, prepend, append, r, p, co, o, n) \ +#define __arch_cmpxchg(lr_sfx, sc_sfx, cas_sfx, \ + sc_prepend, sc_append, \ + r, p, co, o, n) \ ({ \ __label__ zacas, end; \ register unsigned int __rc; \@@ -172,12 +174,12 @@ end:; \ } \ \ __asm__ __volatile__ ( \ - prepend \ + sc_prepend \ "0: lr" lr_sfx " %0, %2\n" \ " bne %0, %z3, 1f\n" \ - " sc" sc_cas_sfx " %1, %z4, %2\n" \ + " sc" sc_sfx " %1, %z4, %2\n" \ " bnez %1, 0b\n" \ - append \ + sc_append \ "1:\n" \ : "=&r" (r), "=&r" (__rc), "+A" (*(p)) \ : "rJ" (co o), "rJ" (n) \@@ -187,9 +189,7 @@ end:; \ zacas: \ if (IS_ENABLED(CONFIG_RISCV_ISA_ZACAS)) { \ __asm__ __volatile__ ( \ - prepend \ - " amocas" sc_cas_sfx " %0, %z2, %1\n" \ - append \ + " amocas" cas_sfx " %0, %z2, %1\n" \ : "+&r" (r), "+A" (*(p)) \ : "rJ" (n) \ : "memory"); \@@ -197,7 +197,8 @@ zacas: \ end:; \ }) -#define _arch_cmpxchg(ptr, old, new, sc_sfx, prepend, append) \ +#define _arch_cmpxchg(ptr, old, new, sc_sfx, cas_sfx, \ + sc_prepend, sc_append) \ ({ \ __typeof__(ptr) __ptr = (ptr); \ __typeof__(*(__ptr)) __old = (old); \@@ -206,22 +207,24 @@ end:; \ \ switch (sizeof(*__ptr)) { \ case 1: \ - __arch_cmpxchg_masked(sc_sfx, ".b" sc_sfx, \ - prepend, append, \ - __ret, __ptr, __old, __new); \ + __arch_cmpxchg_masked(sc_sfx, ".b" cas_sfx, \ + sc_prepend, sc_append, \ + __ret, __ptr, __old, __new); \ break; \ case 2: \ - __arch_cmpxchg_masked(sc_sfx, ".h" sc_sfx, \ - prepend, append, \ - __ret, __ptr, __old, __new); \ + __arch_cmpxchg_masked(sc_sfx, ".h" cas_sfx, \ + sc_prepend, sc_append, \ + __ret, __ptr, __old, __new); \ break; \ case 4: \ - __arch_cmpxchg(".w", ".w" sc_sfx, prepend, append, \ - __ret, __ptr, (long), __old, __new); \ + __arch_cmpxchg(".w", ".w" sc_sfx, ".w" cas_sfx, \ + sc_prepend, sc_append, \ + __ret, __ptr, (long), __old, __new); \ break; \ case 8: \ - __arch_cmpxchg(".d", ".d" sc_sfx, prepend, append, \ - __ret, __ptr, /**/, __old, __new); \ + __arch_cmpxchg(".d", ".d" sc_sfx, ".d" cas_sfx, \ + sc_prepend, sc_append, \ + __ret, __ptr, /**/, __old, __new); \ break; \ default: \ BUILD_BUG(); \@@ -230,16 +233,19 @@ end:; \ }) #define arch_cmpxchg_relaxed(ptr, o, n) \ - _arch_cmpxchg((ptr), (o), (n), "", "", "") + _arch_cmpxchg((ptr), (o), (n), "", "", "", "") #define arch_cmpxchg_acquire(ptr, o, n) \ - _arch_cmpxchg((ptr), (o), (n), "", "", RISCV_ACQUIRE_BARRIER) + _arch_cmpxchg((ptr), (o), (n), "", ".aq", \ + "", RISCV_ACQUIRE_BARRIER) #define arch_cmpxchg_release(ptr, o, n) \ - _arch_cmpxchg((ptr), (o), (n), "", RISCV_RELEASE_BARRIER, "") + _arch_cmpxchg((ptr), (o), (n), "", ".rl", \ + RISCV_RELEASE_BARRIER, "") #define arch_cmpxchg(ptr, o, n) \ - _arch_cmpxchg((ptr), (o), (n), ".rl", "", " fence rw, rw\n") + _arch_cmpxchg((ptr), (o), (n), ".rl", ".aqrl", \ + "", RISCV_FULL_BARRIER)
... but this is not what I suggested: my suggestion [1] was about (limited to) the fully-ordered macro arch_cmpxchg(). In fact, I've recently raised some concern about similar changes to the acquire/release macros, cf. [2]. Any particular reasons for doing this? Andrea [1] https://lore.kernel.org/lkml/ZlYff9x12FICHoP0@andrea/ (local) [2] https://lore.kernel.org/lkml/20240505123340.38495-1-puranjay@kernel.org/ (local)