RISC-V: Report vector unaligned access speed hwprobe

Detect if vector misaligned accesses are faster or slower than
equivalent vector byte accesses. This is useful for usermode to know
whether vector byte accesses or vector misaligned accesses have a better
bandwidth for operations like memcpy.

Signed-off-by: Jesse Taube <jesse@rivosinc.com>
Reviewed-by: Charlie Jenkins <charlie@rivosinc.com>
Link: https://lore.kernel.org/r/20241017-jesse_unaligned_vector-v10-5-5b33500160f8@rivosinc.com
Signed-off-by: Palmer Dabbelt <palmer@rivosinc.com>

arch/riscv/Kconfig

@@ -875,6 +875,24 @@ config RISCV_PROBE_VECTOR_UNALIGNED_ACCESS
 	  will dynamically determine the speed of vector unaligned accesses on
 	  the underlying system if they are supported.
 
+config RISCV_SLOW_VECTOR_UNALIGNED_ACCESS
+	bool "Assume the system supports slow vector unaligned memory accesses"
+	depends on NONPORTABLE
+	help
+	  Assume that the system supports slow vector unaligned memory accesses. The
+	  kernel and userspace programs may not be able to run at all on systems
+	  that do not support unaligned memory accesses.
+
+config RISCV_EFFICIENT_VECTOR_UNALIGNED_ACCESS
+	bool "Assume the system supports fast vector unaligned memory accesses"
+	depends on NONPORTABLE
+	help
+	  Assume that the system supports fast vector unaligned memory accesses. When
+	  enabled, this option improves the performance of the kernel on such
+	  systems. However, the kernel and userspace programs will run much more
+	  slowly, or will not be able to run at all, on systems that do not
+	  support efficient unaligned memory accesses.
+
 endchoice
 
 source "arch/riscv/Kconfig.vendor"

arch/riscv/kernel/Makefile

@@ -70,7 +70,8 @@ obj-$(CONFIG_MMU) += vdso.o vdso/
 
 obj-$(CONFIG_RISCV_MISALIGNED)	+= traps_misaligned.o
 obj-$(CONFIG_RISCV_MISALIGNED)	+= unaligned_access_speed.o
-obj-$(CONFIG_RISCV_PROBE_UNALIGNED_ACCESS)	+= copy-unaligned.o
+obj-$(CONFIG_RISCV_PROBE_UNALIGNED_ACCESS)		+= copy-unaligned.o
+obj-$(CONFIG_RISCV_PROBE_VECTOR_UNALIGNED_ACCESS)	+= vec-copy-unaligned.o
 
 obj-$(CONFIG_FPU)		+= fpu.o
 obj-$(CONFIG_FPU)		+= kernel_mode_fpu.o

arch/riscv/kernel/copy-unaligned.h

@@ -10,4 +10,9 @@
 void __riscv_copy_words_unaligned(void *dst, const void *src, size_t size);
 void __riscv_copy_bytes_unaligned(void *dst, const void *src, size_t size);
 
+#ifdef CONFIG_RISCV_PROBE_VECTOR_UNALIGNED_ACCESS
+void __riscv_copy_vec_words_unaligned(void *dst, const void *src, size_t size);
+void __riscv_copy_vec_bytes_unaligned(void *dst, const void *src, size_t size);
+#endif
+
 #endif /* __RISCV_KERNEL_COPY_UNALIGNED_H */

arch/riscv/kernel/sys_hwprobe.c

@@ -228,6 +228,12 @@ static u64 hwprobe_vec_misaligned(const struct cpumask *cpus)
 #else
 static u64 hwprobe_vec_misaligned(const struct cpumask *cpus)
 {
+	if (IS_ENABLED(CONFIG_RISCV_EFFICIENT_VECTOR_UNALIGNED_ACCESS))
+		return RISCV_HWPROBE_MISALIGNED_VECTOR_FAST;
+
+	if (IS_ENABLED(CONFIG_RISCV_SLOW_VECTOR_UNALIGNED_ACCESS))
+		return RISCV_HWPROBE_MISALIGNED_VECTOR_SLOW;
+
 	return RISCV_HWPROBE_MISALIGNED_VECTOR_UNKNOWN;
 }
 #endif

arch/riscv/kernel/unaligned_access_speed.c

@@ -6,11 +6,13 @@
 #include <linux/cpu.h>
 #include <linux/cpumask.h>
 #include <linux/jump_label.h>
+#include <linux/kthread.h>
 #include <linux/mm.h>
 #include <linux/smp.h>
 #include <linux/types.h>
 #include <asm/cpufeature.h>
 #include <asm/hwprobe.h>
+#include <asm/vector.h>
 
 #include "copy-unaligned.h"
 
@@ -268,12 +270,147 @@ static int check_unaligned_access_speed_all_cpus(void)
 }
 #endif
 
+#ifdef CONFIG_RISCV_PROBE_VECTOR_UNALIGNED_ACCESS
+static void check_vector_unaligned_access(struct work_struct *work __always_unused)
+{
+	int cpu = smp_processor_id();
+	u64 start_cycles, end_cycles;
+	u64 word_cycles;
+	u64 byte_cycles;
+	int ratio;
+	unsigned long start_jiffies, now;
+	struct page *page;
+	void *dst;
+	void *src;
+	long speed = RISCV_HWPROBE_MISALIGNED_VECTOR_SLOW;
+
+	if (per_cpu(vector_misaligned_access, cpu) != RISCV_HWPROBE_MISALIGNED_VECTOR_UNKNOWN)
+		return;
+
+	page = alloc_pages(GFP_KERNEL, MISALIGNED_BUFFER_ORDER);
+	if (!page) {
+		pr_warn("Allocation failure, not measuring vector misaligned performance\n");
+		return;
+	}
+
+	/* Make an unaligned destination buffer. */
+	dst = (void *)((unsigned long)page_address(page) | 0x1);
+	/* Unalign src as well, but differently (off by 1 + 2 = 3). */
+	src = dst + (MISALIGNED_BUFFER_SIZE / 2);
+	src += 2;
+	word_cycles = -1ULL;
+
+	/* Do a warmup. */
+	kernel_vector_begin();
+	__riscv_copy_vec_words_unaligned(dst, src, MISALIGNED_COPY_SIZE);
+
+	start_jiffies = jiffies;
+	while ((now = jiffies) == start_jiffies)
+		cpu_relax();
+
+	/*
+	 * For a fixed amount of time, repeatedly try the function, and take
+	 * the best time in cycles as the measurement.
+	 */
+	while (time_before(jiffies, now + (1 << MISALIGNED_ACCESS_JIFFIES_LG2))) {
+		start_cycles = get_cycles64();
+		/* Ensure the CSR read can't reorder WRT to the copy. */
+		mb();
+		__riscv_copy_vec_words_unaligned(dst, src, MISALIGNED_COPY_SIZE);
+		/* Ensure the copy ends before the end time is snapped. */
+		mb();
+		end_cycles = get_cycles64();
+		if ((end_cycles - start_cycles) < word_cycles)
+			word_cycles = end_cycles - start_cycles;
+	}
+
+	byte_cycles = -1ULL;
+	__riscv_copy_vec_bytes_unaligned(dst, src, MISALIGNED_COPY_SIZE);
+	start_jiffies = jiffies;
+	while ((now = jiffies) == start_jiffies)
+		cpu_relax();
+
+	while (time_before(jiffies, now + (1 << MISALIGNED_ACCESS_JIFFIES_LG2))) {
+		start_cycles = get_cycles64();
+		/* Ensure the CSR read can't reorder WRT to the copy. */
+		mb();
+		__riscv_copy_vec_bytes_unaligned(dst, src, MISALIGNED_COPY_SIZE);
+		/* Ensure the copy ends before the end time is snapped. */
+		mb();
+		end_cycles = get_cycles64();
+		if ((end_cycles - start_cycles) < byte_cycles)
+			byte_cycles = end_cycles - start_cycles;
+	}
+
+	kernel_vector_end();
+
+	/* Don't divide by zero. */
+	if (!word_cycles || !byte_cycles) {
+		pr_warn("cpu%d: rdtime lacks granularity needed to measure unaligned vector access speed\n",
+			cpu);
+
+		return;
+	}
+
+	if (word_cycles < byte_cycles)
+		speed = RISCV_HWPROBE_MISALIGNED_VECTOR_FAST;
+
+	ratio = div_u64((byte_cycles * 100), word_cycles);
+	pr_info("cpu%d: Ratio of vector byte access time to vector unaligned word access is %d.%02d, unaligned accesses are %s\n",
+		cpu,
+		ratio / 100,
+		ratio % 100,
+		(speed ==  RISCV_HWPROBE_MISALIGNED_VECTOR_FAST) ? "fast" : "slow");
+
+	per_cpu(vector_misaligned_access, cpu) = speed;
+}
+
+static int riscv_online_cpu_vec(unsigned int cpu)
+{
+	if (!has_vector())
+		return 0;
+
+	if (per_cpu(vector_misaligned_access, cpu) != RISCV_HWPROBE_MISALIGNED_VECTOR_UNSUPPORTED)
+		return 0;
+
+	check_vector_unaligned_access_emulated(NULL);
+	check_vector_unaligned_access(NULL);
+	return 0;
+}
+
+/* Measure unaligned access speed on all CPUs present at boot in parallel. */
+static int vec_check_unaligned_access_speed_all_cpus(void *unused __always_unused)
+{
+	schedule_on_each_cpu(check_vector_unaligned_access);
+
+	/*
+	 * Setup hotplug callbacks for any new CPUs that come online or go
+	 * offline.
+	 */
+	cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "riscv:online",
+				  riscv_online_cpu_vec, NULL);
+
+	return 0;
+}
+#else /* CONFIG_RISCV_PROBE_VECTOR_UNALIGNED_ACCESS */
+static int vec_check_unaligned_access_speed_all_cpus(void *unused __always_unused)
+{
+	return 0;
+}
+#endif
+
 static int check_unaligned_access_all_cpus(void)
 {
-	bool all_cpus_emulated;
+	bool all_cpus_emulated, all_cpus_vec_unsupported;
 
 	all_cpus_emulated = check_unaligned_access_emulated_all_cpus();
-	check_vector_unaligned_access_emulated_all_cpus();
+	all_cpus_vec_unsupported = check_vector_unaligned_access_emulated_all_cpus();
+
+	if (!all_cpus_vec_unsupported &&
+	    IS_ENABLED(CONFIG_RISCV_PROBE_VECTOR_UNALIGNED_ACCESS)) {
+		kthread_run(vec_check_unaligned_access_speed_all_cpus,
+			    NULL, "vec_check_unaligned_access_speed_all_cpus");
+	}
 
 	if (!all_cpus_emulated)
 		return check_unaligned_access_speed_all_cpus();

arch/riscv/kernel/vec-copy-unaligned.S

@@ -0,0 +1,58 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/* Copyright (C) 2024 Rivos Inc. */
+
+#include <linux/args.h>
+#include <linux/linkage.h>
+#include <asm/asm.h>
+
+	.text
+
+#define WORD_EEW 32
+
+#define WORD_SEW CONCATENATE(e, WORD_EEW)
+#define VEC_L CONCATENATE(vle, WORD_EEW).v
+#define VEC_S CONCATENATE(vle, WORD_EEW).v
+
+/* void __riscv_copy_vec_words_unaligned(void *, const void *, size_t) */
+/* Performs a memcpy without aligning buffers, using word loads and stores. */
+/* Note: The size is truncated to a multiple of WORD_EEW */
+SYM_FUNC_START(__riscv_copy_vec_words_unaligned)
+	andi  a4, a2, ~(WORD_EEW-1)
+	beqz  a4, 2f
+	add   a3, a1, a4
+	.option push
+	.option arch, +zve32x
+1:
+	vsetivli t0, 8, WORD_SEW, m8, ta, ma
+	VEC_L v0, (a1)
+	VEC_S v0, (a0)
+	addi  a0, a0, WORD_EEW
+	addi  a1, a1, WORD_EEW
+	bltu  a1, a3, 1b
+
+2:
+	.option pop
+	ret
+SYM_FUNC_END(__riscv_copy_vec_words_unaligned)
+
+/* void __riscv_copy_vec_bytes_unaligned(void *, const void *, size_t) */
+/* Performs a memcpy without aligning buffers, using only byte accesses. */
+/* Note: The size is truncated to a multiple of 8 */
+SYM_FUNC_START(__riscv_copy_vec_bytes_unaligned)
+	andi a4, a2, ~(8-1)
+	beqz a4, 2f
+	add  a3, a1, a4
+	.option push
+	.option arch, +zve32x
+1:
+	vsetivli t0, 8, e8, m8, ta, ma
+	vle8.v v0, (a1)
+	vse8.v v0, (a0)
+	addi a0, a0, 8
+	addi a1, a1, 8
+	bltu a1, a3, 1b
+
+2:
+	.option pop
+	ret
+SYM_FUNC_END(__riscv_copy_vec_bytes_unaligned)