Merge patch series "bpf, riscv: use BPF prog pack allocator in BPF JIT"

palmer-dabbelt · palmer-dabbelt · commit 77eea559bae9 · 2023-09-08T11:25:25.000-07:00
Puranjay Mohan <puranjay12@gmail.com> says: Here is some data to prove the V2 fixes the problem: Without this series: root@rv-selftester:~/src/kselftest/bpf# time ./test_tag test_tag: OK (40945 tests) real 7m47.562s user 0m24.145s sys 6m37.064s With this series applied: root@rv-selftester:~/src/selftest/bpf# time ./test_tag test_tag: OK (40945 tests) real 7m29.472s user 0m25.865s sys 6m18.401s BPF programs currently consume a page each on RISCV. For systems with many BPF programs, this adds significant pressure to instruction TLB. High iTLB pressure usually causes slow down for the whole system. Song Liu introduced the BPF prog pack allocator[1] to mitigate the above issue. It packs multiple BPF programs into a single huge page. It is currently only enabled for the x86_64 BPF JIT. I enabled this allocator on the ARM64 BPF JIT[2]. It is being reviewed now. This patch series enables the BPF prog pack allocator for the RISCV BPF JIT. ====================================================== Performance Analysis of prog pack allocator on RISCV64 ====================================================== Test setup: =========== Host machine: Debian GNU/Linux 11 (bullseye) Qemu Version: QEMU emulator version 8.0.3 (Debian 1:8.0.3+dfsg-1) u-boot-qemu Version: 2023.07+dfsg-1 opensbi Version: 1.3-1 To test the performance of the BPF prog pack allocator on RV, a stresser tool[4] linked below was built. This tool loads 8 BPF programs on the system and triggers 5 of them in an infinite loop by doing system calls. The runner script starts 20 instances of the above which loads 8*20=160 BPF programs on the system, 5*20=100 of which are being constantly triggered. The script is passed a command which would be run in the above environment. The script was run with following perf command: ./run.sh "perf stat -a \ -e iTLB-load-misses \ -e dTLB-load-misses \ -e dTLB-store-misses \ -e instructions \ --timeout 60000" The output of the above command is discussed below before and after enabling the BPF prog pack allocator. The tests were run on qemu-system-riscv64 with 8 cpus, 16G memory. The rootfs was created using Bjorn's riscv-cross-builder[5] docker container linked below. Results ======= Before enabling prog pack allocator: ------------------------------------ Performance counter stats for 'system wide': 4939048 iTLB-load-misses 5468689 dTLB-load-misses 465234 dTLB-store-misses 1441082097998 instructions 60.045791200 seconds time elapsed After enabling prog pack allocator: ----------------------------------- Performance counter stats for 'system wide': 3430035 iTLB-load-misses 5008745 dTLB-load-misses 409944 dTLB-store-misses 1441535637988 instructions 60.046296600 seconds time elapsed Improvements in metrics ======================= It was expected that the iTLB-load-misses would decrease as now a single huge page is used to keep all the BPF programs compared to a single page for each program earlier. -------------------------------------------- The improvement in iTLB-load-misses: -30.5 % -------------------------------------------- I repeated this expriment more than 100 times in different setups and the improvement was always greater than 30%. This patch series is boot tested on the Starfive VisionFive 2 board[6]. The performance analysis was not done on the board because it doesn't expose iTLB-load-misses, etc. The stresser program was run on the board to test the loading and unloading of BPF programs [1] https://lore.kernel.org/bpf/20220204185742.271030-1-song@kernel.org/ [2] https://lore.kernel.org/all/20230626085811.3192402-1-puranjay12@gmail.com/ [3] https://lore.kernel.org/all/20230626085811.3192402-2-puranjay12@gmail.com/ [4] https://github.com/puranjaymohan/BPF-Allocator-Bench [5] https://github.com/bjoto/riscv-cross-builder [6] https://www.starfivetech.com/en/site/boards * b4-shazam-merge: bpf, riscv: use prog pack allocator in the BPF JIT riscv: implement a memset like function for text riscv: extend patch_text_nosync() for multiple pages bpf: make bpf_prog_pack allocator portable Link: https://lore.kernel.org/r/20230831131229.497941-1-puranjay12@gmail.com Signed-off-by: Palmer Dabbelt <palmer@rivosinc.com>
diff --git a/arch/riscv/include/asm/patch.h b/arch/riscv/include/asm/patch.h
@@ -7,6 +7,7 @@
 #define _ASM_RISCV_PATCH_H
 
 int patch_text_nosync(void *addr, const void *insns, size_t len);
+int patch_text_set_nosync(void *addr, u8 c, size_t len);
 int patch_text(void *addr, u32 *insns, int ninsns);
 
 extern int riscv_patch_in_stop_machine;
diff --git a/arch/riscv/kernel/patch.c b/arch/riscv/kernel/patch.c
@@ -6,6 +6,7 @@
 #include <linux/spinlock.h>
 #include <linux/mm.h>
 #include <linux/memory.h>
+#include <linux/string.h>
 #include <linux/uaccess.h>
 #include <linux/stop_machine.h>
 #include <asm/kprobes.h>
@@ -53,12 +54,51 @@ static void patch_unmap(int fixmap)
 }
 NOKPROBE_SYMBOL(patch_unmap);
 
-static int patch_insn_write(void *addr, const void *insn, size_t len)
+static int __patch_insn_set(void *addr, u8 c, size_t len)
+{
+	void *waddr = addr;
+	bool across_pages = (((uintptr_t)addr & ~PAGE_MASK) + len) > PAGE_SIZE;
+
+	/*
+	 * Only two pages can be mapped at a time for writing.
+	 */
+	if (len + offset_in_page(addr) > 2 * PAGE_SIZE)
+		return -EINVAL;
+	/*
+	 * Before reaching here, it was expected to lock the text_mutex
+	 * already, so we don't need to give another lock here and could
+	 * ensure that it was safe between each cores.
+	 */
+	lockdep_assert_held(&text_mutex);
+
+	if (across_pages)
+		patch_map(addr + PAGE_SIZE, FIX_TEXT_POKE1);
+
+	waddr = patch_map(addr, FIX_TEXT_POKE0);
+
+	memset(waddr, c, len);
+
+	patch_unmap(FIX_TEXT_POKE0);
+
+	if (across_pages)
+		patch_unmap(FIX_TEXT_POKE1);
+
+	return 0;
+}
+NOKPROBE_SYMBOL(__patch_insn_set);
+
+static int __patch_insn_write(void *addr, const void *insn, size_t len)
 {
 	void *waddr = addr;
 	bool across_pages = (((uintptr_t) addr & ~PAGE_MASK) + len) > PAGE_SIZE;
 	int ret;
 
+	/*
+	 * Only two pages can be mapped at a time for writing.
+	 */
+	if (len + offset_in_page(addr) > 2 * PAGE_SIZE)
+		return -EINVAL;
+
 	/*
 	 * Before reaching here, it was expected to lock the text_mutex
 	 * already, so we don't need to give another lock here and could
@@ -74,7 +114,7 @@ static int patch_insn_write(void *addr, const void *insn, size_t len)
 		lockdep_assert_held(&text_mutex);
 
 	if (across_pages)
-		patch_map(addr + len, FIX_TEXT_POKE1);
+		patch_map(addr + PAGE_SIZE, FIX_TEXT_POKE1);
 
 	waddr = patch_map(addr, FIX_TEXT_POKE0);
 
@@ -87,15 +127,79 @@ static int patch_insn_write(void *addr, const void *insn, size_t len)
 
 	return ret;
 }
-NOKPROBE_SYMBOL(patch_insn_write);
+NOKPROBE_SYMBOL(__patch_insn_write);
 #else
-static int patch_insn_write(void *addr, const void *insn, size_t len)
+static int __patch_insn_set(void *addr, u8 c, size_t len)
+{
+	memset(addr, c, len);
+
+	return 0;
+}
+NOKPROBE_SYMBOL(__patch_insn_set);
+
+static int __patch_insn_write(void *addr, const void *insn, size_t len)
 {
 	return copy_to_kernel_nofault(addr, insn, len);
 }
-NOKPROBE_SYMBOL(patch_insn_write);
+NOKPROBE_SYMBOL(__patch_insn_write);
 #endif /* CONFIG_MMU */
 
+static int patch_insn_set(void *addr, u8 c, size_t len)
+{
+	size_t patched = 0;
+	size_t size;
+	int ret = 0;
+
+	/*
+	 * __patch_insn_set() can only work on 2 pages at a time so call it in a
+	 * loop with len <= 2 * PAGE_SIZE.
+	 */
+	while (patched < len && !ret) {
+		size = min_t(size_t, PAGE_SIZE * 2 - offset_in_page(addr + patched), len - patched);
+		ret = __patch_insn_set(addr + patched, c, size);
+
+		patched += size;
+	}
+
+	return ret;
+}
+NOKPROBE_SYMBOL(patch_insn_set);
+
+int patch_text_set_nosync(void *addr, u8 c, size_t len)
+{
+	u32 *tp = addr;
+	int ret;
+
+	ret = patch_insn_set(tp, c, len);
+
+	if (!ret)
+		flush_icache_range((uintptr_t)tp, (uintptr_t)tp + len);
+
+	return ret;
+}
+NOKPROBE_SYMBOL(patch_text_set_nosync);
+
+static int patch_insn_write(void *addr, const void *insn, size_t len)
+{
+	size_t patched = 0;
+	size_t size;
+	int ret = 0;
+
+	/*
+	 * Copy the instructions to the destination address, two pages at a time
+	 * because __patch_insn_write() can only handle len <= 2 * PAGE_SIZE.
+	 */
+	while (patched < len && !ret) {
+		size = min_t(size_t, PAGE_SIZE * 2 - offset_in_page(addr + patched), len - patched);
+		ret = __patch_insn_write(addr + patched, insn + patched, size);
+
+		patched += size;
+	}
+
+	return ret;
+}
+NOKPROBE_SYMBOL(patch_insn_write);
+
 int patch_text_nosync(void *addr, const void *insns, size_t len)
 {
 	u32 *tp = addr;
diff --git a/arch/riscv/net/bpf_jit.h b/arch/riscv/net/bpf_jit.h
@@ -68,6 +68,7 @@ static inline bool is_creg(u8 reg)
 struct rv_jit_context {
 	struct bpf_prog *prog;
 	u16 *insns;		/* RV insns */
+	u16 *ro_insns;
 	int ninsns;
 	int prologue_len;
 	int epilogue_offset;
@@ -85,7 +86,9 @@ static inline int ninsns_rvoff(int ninsns)
 
 struct rv_jit_data {
 	struct bpf_binary_header *header;
+	struct bpf_binary_header *ro_header;
 	u8 *image;
+	u8 *ro_image;
 	struct rv_jit_context ctx;
 };
 
diff --git a/arch/riscv/net/bpf_jit_comp64.c b/arch/riscv/net/bpf_jit_comp64.c
@@ -144,7 +144,11 @@ static bool in_auipc_jalr_range(s64 val)
 /* Emit fixed-length instructions for address */
 static int emit_addr(u8 rd, u64 addr, bool extra_pass, struct rv_jit_context *ctx)
 {
-	u64 ip = (u64)(ctx->insns + ctx->ninsns);
+	/*
+	 * Use the ro_insns(RX) to calculate the offset as the BPF program will
+	 * finally run from this memory region.
+	 */
+	u64 ip = (u64)(ctx->ro_insns + ctx->ninsns);
 	s64 off = addr - ip;
 	s64 upper = (off + (1 << 11)) >> 12;
 	s64 lower = off & 0xfff;
@@ -464,8 +468,12 @@ static int emit_call(u64 addr, bool fixed_addr, struct rv_jit_context *ctx)
 	s64 off = 0;
 	u64 ip;
 
-	if (addr && ctx->insns) {
-		ip = (u64)(long)(ctx->insns + ctx->ninsns);
+	if (addr && ctx->insns && ctx->ro_insns) {
+		/*
+		 * Use the ro_insns(RX) to calculate the offset as the BPF
+		 * program will finally run from this memory region.
+		 */
+		ip = (u64)(long)(ctx->ro_insns + ctx->ninsns);
 		off = addr - ip;
 	}
 
@@ -578,9 +586,10 @@ static int add_exception_handler(const struct bpf_insn *insn,
 {
 	struct exception_table_entry *ex;
 	unsigned long pc;
-	off_t offset;
+	off_t ins_offset;
+	off_t fixup_offset;
 
-	if (!ctx->insns || !ctx->prog->aux->extable ||
+	if (!ctx->insns || !ctx->ro_insns || !ctx->prog->aux->extable ||
 	    (BPF_MODE(insn->code) != BPF_PROBE_MEM && BPF_MODE(insn->code) != BPF_PROBE_MEMSX))
 		return 0;
 
@@ -594,12 +603,17 @@ static int add_exception_handler(const struct bpf_insn *insn,
 		return -EINVAL;
 
 	ex = &ctx->prog->aux->extable[ctx->nexentries];
-	pc = (unsigned long)&ctx->insns[ctx->ninsns - insn_len];
+	pc = (unsigned long)&ctx->ro_insns[ctx->ninsns - insn_len];
 
-	offset = pc - (long)&ex->insn;
-	if (WARN_ON_ONCE(offset >= 0 || offset < INT_MIN))
+	/*
+	 * This is the relative offset of the instruction that may fault from
+	 * the exception table itself. This will be written to the exception
+	 * table and if this instruction faults, the destination register will
+	 * be set to '0' and the execution will jump to the next instruction.
+	 */
+	ins_offset = pc - (long)&ex->insn;
+	if (WARN_ON_ONCE(ins_offset >= 0 || ins_offset < INT_MIN))
 		return -ERANGE;
-	ex->insn = offset;
 
 	/*
 	 * Since the extable follows the program, the fixup offset is always
@@ -608,12 +622,25 @@ static int add_exception_handler(const struct bpf_insn *insn,
 	 * bits. We don't need to worry about buildtime or runtime sort
 	 * modifying the upper bits because the table is already sorted, and
 	 * isn't part of the main exception table.
+	 *
+	 * The fixup_offset is set to the next instruction from the instruction
+	 * that may fault. The execution will jump to this after handling the
+	 * fault.
 	 */
-	offset = (long)&ex->fixup - (pc + insn_len * sizeof(u16));
-	if (!FIELD_FIT(BPF_FIXUP_OFFSET_MASK, offset))
+	fixup_offset = (long)&ex->fixup - (pc + insn_len * sizeof(u16));
+	if (!FIELD_FIT(BPF_FIXUP_OFFSET_MASK, fixup_offset))
 		return -ERANGE;
 
-	ex->fixup = FIELD_PREP(BPF_FIXUP_OFFSET_MASK, offset) |
+	/*
+	 * The offsets above have been calculated using the RO buffer but we
+	 * need to use the R/W buffer for writes.
+	 * switch ex to rw buffer for writing.
+	 */
+	ex = (void *)ctx->insns + ((void *)ex - (void *)ctx->ro_insns);
+
+	ex->insn = ins_offset;
+
+	ex->fixup = FIELD_PREP(BPF_FIXUP_OFFSET_MASK, fixup_offset) |
 		FIELD_PREP(BPF_FIXUP_REG_MASK, dst_reg);
 	ex->type = EX_TYPE_BPF;
 
@@ -1007,6 +1034,7 @@ int arch_prepare_bpf_trampoline(struct bpf_tramp_image *im, void *image,
 
 	ctx.ninsns = 0;
 	ctx.insns = NULL;
+	ctx.ro_insns = NULL;
 	ret = __arch_prepare_bpf_trampoline(im, m, tlinks, func_addr, flags, &ctx);
 	if (ret < 0)
 		return ret;
@@ -1015,7 +1043,15 @@ int arch_prepare_bpf_trampoline(struct bpf_tramp_image *im, void *image,
 		return -EFBIG;
 
 	ctx.ninsns = 0;
+	/*
+	 * The bpf_int_jit_compile() uses a RW buffer (ctx.insns) to write the
+	 * JITed instructions and later copies it to a RX region (ctx.ro_insns).
+	 * It also uses ctx.ro_insns to calculate offsets for jumps etc. As the
+	 * trampoline image uses the same memory area for writing and execution,
+	 * both ctx.insns and ctx.ro_insns can be set to image.
+	 */
 	ctx.insns = image;
+	ctx.ro_insns = image;
 	ret = __arch_prepare_bpf_trampoline(im, m, tlinks, func_addr, flags, &ctx);
 	if (ret < 0)
 		return ret;
diff --git a/arch/riscv/net/bpf_jit_core.c b/arch/riscv/net/bpf_jit_core.c
diff --git a/kernel/bpf/core.c b/kernel/bpf/core.c
