Re: [PATCH v4 06/13] riscv: Improve zacas fully-ordered cmpxchg()
From: Alexandre Ghiti <hidden>
Date: 2024-08-01 06:15:58
Also in:
linux-arch, linux-doc, linux-riscv, lkml
Hi Drew, On Wed, Jul 31, 2024 at 11:59 AM Andrew Jones [off-list ref] wrote:
On Wed, Jul 31, 2024 at 09:23:58AM GMT, Alexandre Ghiti wrote:quoted
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)We won't get release semantics if the exchange fails. Does that matter?quoted
Suggested-by: Andrea Parri <redacted> Signed-off-by: Alexandre Ghiti <redacted> --- arch/riscv/include/asm/cmpxchg.h | 72 +++++++++++++++++++------------- 1 file changed, 44 insertions(+), 28 deletions(-)diff --git a/arch/riscv/include/asm/cmpxchg.h b/arch/riscv/include/asm/cmpxchg.h index ebcd4a30ae60..391730367213 100644 --- a/arch/riscv/include/asm/cmpxchg.h +++ b/arch/riscv/include/asm/cmpxchg.h@@ -107,8 +107,10 @@ * store NEW in MEM. Return the initial value in MEM. Success is * 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, \ + cas_prepend, cas_append, \ + r, p, o, n) \ ({ \ if (IS_ENABLED(CONFIG_RISCV_ISA_ZABHA) && \ IS_ENABLED(CONFIG_RISCV_ISA_ZACAS) && \@@ -117,9 +119,9 @@ r = o; \ \ __asm__ __volatile__ ( \ - prepend \ + cas_prepend \ " amocas" cas_sfx " %0, %z2, %1\n" \ - append \ + cas_append \ : "+&r" (r), "+A" (*(p)) \ : "rJ" (n) \ : "memory"); \@@ -134,7 +136,7 @@ ulong __rc; \ \ __asm__ __volatile__ ( \ - prepend \ + sc_prepend \ "0: lr.w %0, %2\n" \ " and %1, %0, %z5\n" \ " bne %1, %z3, 1f\n" \@@ -142,7 +144,7 @@ " 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), \@@ -153,16 +155,19 @@ } \ }) -#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, \ + cas_prepend, cas_append, \ + r, p, co, o, n) \ ({ \ if (IS_ENABLED(CONFIG_RISCV_ISA_ZACAS) && \ riscv_has_extension_unlikely(RISCV_ISA_EXT_ZACAS)) { \ r = o; \ \ __asm__ __volatile__ ( \ - prepend \ - " amocas" sc_cas_sfx " %0, %z2, %1\n" \ - append \ + cas_prepend \ + " amocas" cas_sfx " %0, %z2, %1\n" \ + cas_append \ : "+&r" (r), "+A" (*(p)) \ : "rJ" (n) \ : "memory"); \@@ -170,12 +175,12 @@ register unsigned int __rc; \ \ __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" \nit: If patch3 hadn't renamed sc_sfx to sc_cas_sfx then we wouldn't need to rename it again now.
You're right, if you don't mind I'll leave it as is though as it makes the previous patch more consistent.
quoted
" bnez %1, 0b\n" \ - append \ + sc_append \ "1:\n" \ : "=&r" (r), "=&r" (__rc), "+A" (*(p)) \ : "rJ" (co o), "rJ" (n) \@@ -183,7 +188,9 @@ } \ }) -#define _arch_cmpxchg(ptr, old, new, sc_cas_sfx, prepend, append) \ +#define _arch_cmpxchg(ptr, old, new, sc_sfx, cas_sfx, \ + sc_prepend, sc_append, \ + cas_prepend, cas_append) \ ({ \ __typeof__(ptr) __ptr = (ptr); \ __typeof__(*(__ptr)) __old = (old); \@@ -192,22 +199,28 @@ \ switch (sizeof(*__ptr)) { \ case 1: \ - __arch_cmpxchg_masked(sc_cas_sfx, ".b" sc_cas_sfx, \ - prepend, append, \ - __ret, __ptr, __old, __new); \ + __arch_cmpxchg_masked(sc_sfx, ".b" cas_sfx, \ + sc_prepend, sc_append, \ + cas_prepend, cas_append, \ + __ret, __ptr, __old, __new); \ break; \ case 2: \ - __arch_cmpxchg_masked(sc_cas_sfx, ".h" sc_cas_sfx, \ - prepend, append, \ - __ret, __ptr, __old, __new); \ + __arch_cmpxchg_masked(sc_sfx, ".h" cas_sfx, \ + sc_prepend, sc_append, \ + cas_prepend, cas_append, \ + __ret, __ptr, __old, __new); \ break; \ case 4: \ - __arch_cmpxchg(".w", ".w" sc_cas_sfx, prepend, append, \ - __ret, __ptr, (long), __old, __new); \ + __arch_cmpxchg(".w", ".w" sc_sfx, ".w" cas_sfx, \ + sc_prepend, sc_append, \ + cas_prepend, cas_append, \ + __ret, __ptr, (long), __old, __new); \ break; \ case 8: \ - __arch_cmpxchg(".d", ".d" sc_cas_sfx, prepend, append, \ - __ret, __ptr, /**/, __old, __new); \ + __arch_cmpxchg(".d", ".d" sc_sfx, ".d" cas_sfx, \ + sc_prepend, sc_append, \ + cas_prepend, cas_append, \ + __ret, __ptr, /**/, __old, __new); \ break; \ default: \ BUILD_BUG(); \@@ -216,16 +229,19 @@ }) #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), "", "", \ + "", RISCV_ACQUIRE_BARRIER, "", RISCV_ACQUIRE_BARRIER) #define arch_cmpxchg_release(ptr, o, n) \ - _arch_cmpxchg((ptr), (o), (n), "", RISCV_RELEASE_BARRIER, "") + _arch_cmpxchg((ptr), (o), (n), "", "", \ + RISCV_RELEASE_BARRIER, "", 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, "", "")These aren't the easiest things to read, but I can't think of a way to improve it other than maybe some macro annotations. E.g. #define SC_SFX(x) x #define CAS_SFX(x) x #define SC_PREPEND(x) x #define SC_APPEND(x) x #define CAS_PREPEND(x) x #define CAS_APPEND(x) x #define arch_cmpxchg(ptr, o, n) \ _arch_cmpxchg(ptr, o, n, \ SC_SFX(".rl"), CAS_SFX(".aqrl"), \ SC_PREPEND(""), SC_APPEND(RISCV_FULL_BARRIER), \ CAS_PREPEND(""), CAS_APPEND(""))
That's a very good idea, it's been hard to review even for me :) I could add comments too, but I like your solution, so unless I find something better in the next 30min, I'll implement that. Thanks, Alex
quoted
#define arch_cmpxchg_local(ptr, o, n) \ arch_cmpxchg_relaxed((ptr), (o), (n)) -- 2.39.2Thanks, drew