Re: [PATCH] LoongArch: Support dbar with different hints

From: WANG Xuerui
Date: Tue May 16 2023 - 09:24:53 EST

On 2023/5/16 20:45, Huacai Chen wrote:
Traditionally, LoongArch uses "dbar 0" (full completion barrier) for
everything. But the full completion barrier is a performance killer, so
Loongson-3A6000 and newer processors introduce different hints:

"have made finer granularity hints available"


Bit4: ordering or completion (0: completion, 1: ordering)
Bit3: barrier for previous read (0: true, 1: false)
Bit2: barrier for previous write (0: true, 1: false)
Bit1: barrier for succeeding read (0: true, 1: false)
Bit0: barrier for succedding write (0: true, 1: false)

"succeeding"


Hint 0x700: barrier for "read after read" from the same address, which
is needed by LL-SC loops.

"needed by LL-SC loops on older models"?


This patch enable various hints for different memory barries, it brings
performance improvements for Loongson-3A6000 series, and doesn't impact
Loongson-3A5000 series because they treat all hints as "dbar 0".

"This patch makes use of the various new hints for different kinds of memory barriers. It brings performance improvements on Loongson-3A6000 series, while not affecting the existing models because all variants are treated as 'dbar 0' there."


Signed-off-by: Jun Yi <yijun@xxxxxxxxxxx>
Signed-off-by: Huacai Chen <chenhuacai@xxxxxxxxxxx>
---
arch/loongarch/include/asm/barrier.h | 130 ++++++++++++---------------
arch/loongarch/include/asm/io.h | 2 +-
arch/loongarch/kernel/smp.c | 2 +-
arch/loongarch/mm/tlbex.S | 6 +-
4 files changed, 60 insertions(+), 80 deletions(-)

diff --git a/arch/loongarch/include/asm/barrier.h b/arch/loongarch/include/asm/barrier.h
index cda977675854..0286ae7e3636 100644
--- a/arch/loongarch/include/asm/barrier.h
+++ b/arch/loongarch/include/asm/barrier.h
@@ -5,27 +5,56 @@
#ifndef __ASM_BARRIER_H
#define __ASM_BARRIER_H
-#define __sync() __asm__ __volatile__("dbar 0" : : : "memory")
+/*
+ * Hint types:

"Hint encoding" might be more appropriate?

+ *
+ * Bit4: ordering or completion (0: completion, 1: ordering)
+ * Bit3: barrier for previous read (0: true, 1: false)
+ * Bit2: barrier for previous write (0: true, 1: false)
+ * Bit1: barrier for succeeding read (0: true, 1: false)
+ * Bit0: barrier for succedding write (0: true, 1: false)

"succeeding"

+ *
+ * Hint 0x700: barrier for "read after read" from the same address
+ */
+
+#define DBAR(hint) __asm__ __volatile__("dbar %0 " : : "I"(hint) : "memory")

Why not __builtin_loongarch_dbar (should be usable with both GCC and Clang) or __dbar (in <larchintrin.h>)?

+
+#define crwrw 0b00000
+#define cr_r_ 0b00101
+#define c_w_w 0b01010
-#define fast_wmb() __sync()
-#define fast_rmb() __sync()
-#define fast_mb() __sync()
-#define fast_iob() __sync()
-#define wbflush() __sync()
+#define orwrw 0b10000
+#define or_r_ 0b10101
+#define o_w_w 0b11010
-#define wmb() fast_wmb()
-#define rmb() fast_rmb()
-#define mb() fast_mb()
-#define iob() fast_iob()
+#define orw_w 0b10010
+#define or_rw 0b10100
-#define __smp_mb() __asm__ __volatile__("dbar 0" : : : "memory")
-#define __smp_rmb() __asm__ __volatile__("dbar 0" : : : "memory")
-#define __smp_wmb() __asm__ __volatile__("dbar 0" : : : "memory")
+#define c_sync() DBAR(crwrw)
+#define c_rsync() DBAR(cr_r_)
+#define c_wsync() DBAR(c_w_w)
+
+#define o_sync() DBAR(orwrw)
+#define o_rsync() DBAR(or_r_)
+#define o_wsync() DBAR(o_w_w)
+
+#define ldacq_mb() DBAR(or_rw)
+#define strel_mb() DBAR(orw_w)
+
+#define mb() c_sync()
+#define rmb() c_rsync()
+#define wmb() c_wsync()
+#define iob() c_sync()
+#define wbflush() c_sync()
+
+#define __smp_mb() o_sync()
+#define __smp_rmb() o_rsync()
+#define __smp_wmb() o_wsync()
#ifdef CONFIG_SMP
-#define __WEAK_LLSC_MB " dbar 0 \n"
+#define __WEAK_LLSC_MB " dbar 0x700 \n"
#else
-#define __WEAK_LLSC_MB " \n"
+#define __WEAK_LLSC_MB " \n"
#endif
#define __smp_mb__before_atomic() barrier()
@@ -59,68 +88,19 @@ static inline unsigned long array_index_mask_nospec(unsigned long index,
return mask;
}
-#define __smp_load_acquire(p) \
-({ \
- union { typeof(*p) __val; char __c[1]; } __u; \
- unsigned long __tmp = 0; \
- compiletime_assert_atomic_type(*p); \
- switch (sizeof(*p)) { \
- case 1: \
- *(__u8 *)__u.__c = *(volatile __u8 *)p; \
- __smp_mb(); \
- break; \
- case 2: \
- *(__u16 *)__u.__c = *(volatile __u16 *)p; \
- __smp_mb(); \
- break; \
- case 4: \
- __asm__ __volatile__( \
- "amor_db.w %[val], %[tmp], %[mem] \n" \
- : [val] "=&r" (*(__u32 *)__u.__c) \
- : [mem] "ZB" (*(u32 *) p), [tmp] "r" (__tmp) \
- : "memory"); \
- break; \
- case 8: \
- __asm__ __volatile__( \
- "amor_db.d %[val], %[tmp], %[mem] \n" \
- : [val] "=&r" (*(__u64 *)__u.__c) \
- : [mem] "ZB" (*(u64 *) p), [tmp] "r" (__tmp) \
- : "memory"); \
- break; \
- } \
- (typeof(*p))__u.__val; \
+#define __smp_load_acquire(p) \
+({ \
+ typeof(*p) ___p1 = READ_ONCE(*p); \
+ compiletime_assert_atomic_type(*p); \
+ ldacq_mb(); \
+ ___p1; \
})
-#define __smp_store_release(p, v) \
-do { \
- union { typeof(*p) __val; char __c[1]; } __u = \
- { .__val = (__force typeof(*p)) (v) }; \
- unsigned long __tmp; \
- compiletime_assert_atomic_type(*p); \
- switch (sizeof(*p)) { \
- case 1: \
- __smp_mb(); \
- *(volatile __u8 *)p = *(__u8 *)__u.__c; \
- break; \
- case 2: \
- __smp_mb(); \
- *(volatile __u16 *)p = *(__u16 *)__u.__c; \
- break; \
- case 4: \
- __asm__ __volatile__( \
- "amswap_db.w %[tmp], %[val], %[mem] \n" \
- : [mem] "+ZB" (*(u32 *)p), [tmp] "=&r" (__tmp) \
- : [val] "r" (*(__u32 *)__u.__c) \
- : ); \
- break; \
- case 8: \
- __asm__ __volatile__( \
- "amswap_db.d %[tmp], %[val], %[mem] \n" \
- : [mem] "+ZB" (*(u64 *)p), [tmp] "=&r" (__tmp) \
- : [val] "r" (*(__u64 *)__u.__c) \
- : ); \
- break; \
- } \
+#define __smp_store_release(p, v) \
+do { \
+ compiletime_assert_atomic_type(*p); \
+ strel_mb(); \
+ WRITE_ONCE(*p, v); \
} while (0)
#define __smp_store_mb(p, v) \
diff --git a/arch/loongarch/include/asm/io.h b/arch/loongarch/include/asm/io.h
index 545e2708fbf7..1c9410220040 100644
--- a/arch/loongarch/include/asm/io.h
+++ b/arch/loongarch/include/asm/io.h
@@ -62,7 +62,7 @@ extern pgprot_t pgprot_wc;
#define ioremap_cache(offset, size) \
ioremap_prot((offset), (size), pgprot_val(PAGE_KERNEL))
-#define mmiowb() asm volatile ("dbar 0" ::: "memory")
+#define mmiowb() wmb()
/*
* String version of I/O memory access operations.
diff --git a/arch/loongarch/kernel/smp.c b/arch/loongarch/kernel/smp.c
index ed167e244cda..8daa97148c8e 100644
--- a/arch/loongarch/kernel/smp.c
+++ b/arch/loongarch/kernel/smp.c
@@ -118,7 +118,7 @@ static u32 ipi_read_clear(int cpu)
action = iocsr_read32(LOONGARCH_IOCSR_IPI_STATUS);
/* Clear the ipi register to clear the interrupt */
iocsr_write32(action, LOONGARCH_IOCSR_IPI_CLEAR);
- smp_mb();
+ wbflush();
return action;
}
diff --git a/arch/loongarch/mm/tlbex.S b/arch/loongarch/mm/tlbex.S
index 244e2f5aeee5..240ced55586e 100644
--- a/arch/loongarch/mm/tlbex.S
+++ b/arch/loongarch/mm/tlbex.S
@@ -184,7 +184,7 @@ tlb_huge_update_load:
ertn
nopage_tlb_load:
- dbar 0
+ dbar 0x700
csrrd ra, EXCEPTION_KS2
la_abs t0, tlb_do_page_fault_0
jr t0
@@ -333,7 +333,7 @@ tlb_huge_update_store:
ertn
nopage_tlb_store:
- dbar 0
+ dbar 0x700
csrrd ra, EXCEPTION_KS2
la_abs t0, tlb_do_page_fault_1
jr t0
@@ -480,7 +480,7 @@ tlb_huge_update_modify:
ertn
nopage_tlb_modify:
- dbar 0
+ dbar 0x700
csrrd ra, EXCEPTION_KS2
la_abs t0, tlb_do_page_fault_1
jr t0

Due to not having the manuals, I cannot review any further, but I can at least see that semantics on older models should be reasonably conserved, and the new bits seem reasonable. Nice cleanups anyway, and thanks to all of you for listening to community voices!

--
WANG "xen0n" Xuerui

Linux/LoongArch mailing list: https://lore.kernel.org/loongarch/