[PATCH 22/27] perf tools: Add prologue for BPF programs for fetching arguments

[Wang Nan]

This patch generates prologue for a BPF program which fetch arguments
for it. With this patch, the program can have arguments as follow:

 SEC("lock_page=__lock_page page->flags")
 int lock_page(struct pt_regs *ctx, int err, unsigned long flags)
 {
	 return 1;
 }

This patch passes at most 3 arguments from r3, r4 and r5. r1 is still
the ctx pointer. r2 is used to indicate the successfulness of
dereferencing.

This patch uses r6 to hold ctx (struct pt_regs) and r7 to hold stack
pointer for result. Result of each arguments first store on stack:

 low address
 BPF_REG_FP - 24  ARG3
 BPF_REG_FP - 16  ARG2
 BPF_REG_FP - 8   ARG1
 BPF_REG_FP
 high address

Then loaded into r3, r4 and r5.

The output prologue for offn(...off2(off1(reg)))) should be:

     r6 <- r1			// save ctx into a callee saved register
     r7 <- fp
     r7 <- r7 - stack_offset	// pointer to result slot
     /* load r3 with the offset in pt_regs of 'reg' */
     (r7) <- r3			// make slot valid
     r3 <- r3 + off1		// prepare to read unsafe pointer
     r2 <- 8
     r1 <- r7			// result put onto stack
     call probe_read		// read unsafe pointer
     jnei r0, 0, err		// error checking
     r3 <- (r7)			// read result
     r3 <- r3 + off2		// prepare to read unsafe pointer
     r2 <- 8
     r1 <- r7
     call probe_read
     jnei r0, 0, err
     ...
     /* load r2, r3, r4 from stack */
     goto success
err:
     r2 <- 1
     /* load r3, r4, r5 with 0 */
     goto usercode
success:
     r2 <- 0
usercode:
     r1 <- r6	// restore ctx
     // original user code

If all of arguments reside in register (dereferencing is not
required), gen_prologue_fastpath() will be used to create
fast prologue:

     r3 <- (r1 + offset of reg1)
     r4 <- (r1 + offset of reg2)
     r5 <- (r1 + offset of reg3)
     r2 <- 0

P.S.

eBPF calling convention is defined as:

* r0		- return value from in-kernel function, and exit value
                  for eBPF program
* r1 - r5	- arguments from eBPF program to in-kernel function
* r6 - r9	- callee saved registers that in-kernel function will
                  preserve
* r10		- read-only frame pointer to access stack

Signed-off-by: He Kuang <hekuang@xxxxxxxxxx>
Signed-off-by: Wang Nan <wangnan0@xxxxxxxxxx>
Cc: Alexei Starovoitov <ast@xxxxxxxxxxxx>
Cc: Brendan Gregg <brendan.d.gregg@xxxxxxxxx>
Cc: Daniel Borkmann <daniel@xxxxxxxxxxxxx>
Cc: David Ahern <dsahern@xxxxxxxxx>
Cc: He Kuang <hekuang@xxxxxxxxxx>
Cc: Jiri Olsa <jolsa@xxxxxxxxxx>
Cc: Kaixu Xia <xiakaixu@xxxxxxxxxx>
Cc: Masami Hiramatsu <masami.hiramatsu.pt@xxxxxxxxxxx>
Cc: Namhyung Kim <namhyung@xxxxxxxxxx>
Cc: Paul Mackerras <paulus@xxxxxxxxx>
Cc: Peter Zijlstra <a.p.zijlstra@xxxxxxxxx>
Cc: Zefan Li <lizefan@xxxxxxxxxx>
Cc: pi3orama@xxxxxxx
Cc: Arnaldo Carvalho de Melo <acme@xxxxxxxxxx>
Link: http://lkml.kernel.org/n/1436445342-1402-35-git-send-email-wangnan0@xxxxxxxxxx
---
 tools/perf/util/Build          |   1 +
 tools/perf/util/bpf-prologue.c | 442 +++++++++++++++++++++++++++++++++++++++++
 tools/perf/util/bpf-prologue.h |  34 ++++
 3 files changed, 477 insertions(+)
 create mode 100644 tools/perf/util/bpf-prologue.c
 create mode 100644 tools/perf/util/bpf-prologue.h

diff --git a/tools/perf/util/Build b/tools/perf/util/Build
index 309b2d2..fc2bd80 100644
--- a/tools/perf/util/Build
+++ b/tools/perf/util/Build
@@ -81,6 +81,7 @@ libperf-$(CONFIG_AUXTRACE) += auxtrace.o
 libperf-y += parse-branch-options.o
 
 libperf-$(CONFIG_LIBBPF) += bpf-loader.o
+libperf-$(CONFIG_BPF_PROLOGUE) += bpf-prologue.o
 libperf-$(CONFIG_LIBELF) += symbol-elf.o
 libperf-$(CONFIG_LIBELF) += probe-file.o
 libperf-$(CONFIG_LIBELF) += probe-event.o
diff --git a/tools/perf/util/bpf-prologue.c b/tools/perf/util/bpf-prologue.c
new file mode 100644
index 0000000..2a5f4c7
--- /dev/null
+++ b/tools/perf/util/bpf-prologue.c
@@ -0,0 +1,442 @@
+/*
+ * bpf-prologue.c
+ *
+ * Copyright (C) 2015 He Kuang <hekuang@xxxxxxxxxx>
+ * Copyright (C) 2015 Huawei Inc.
+ */
+
+#include <bpf/libbpf.h>
+#include "perf.h"
+#include "debug.h"
+#include "bpf-prologue.h"
+#include "probe-finder.h"
+#include <dwarf-regs.h>
+#include <linux/filter.h>
+
+#define BPF_REG_SIZE		8
+
+#define JMP_TO_ERROR_CODE	-1
+#define JMP_TO_SUCCESS_CODE	-2
+#define JMP_TO_USER_CODE	-3
+
+struct bpf_insn_pos {
+	struct bpf_insn *begin;
+	struct bpf_insn *end;
+	struct bpf_insn *pos;
+};
+
+static inline int
+pos_get_cnt(struct bpf_insn_pos *pos)
+{
+	return pos->pos - pos->begin;
+}
+
+static int
+append_insn(struct bpf_insn new_insn, struct bpf_insn_pos *pos)
+{
+	if (!pos->pos)
+		return -ERANGE;
+
+	if (pos->pos + 1 >= pos->end) {
+		pr_err("bpf prologue: prologue too long\n");
+		pos->pos = NULL;
+		return -ERANGE;
+	}
+
+	*(pos->pos)++ = new_insn;
+	return 0;
+}
+
+static int
+check_pos(struct bpf_insn_pos *pos)
+{
+	if (!pos->pos || pos->pos >= pos->end)
+		return -ERANGE;
+	return 0;
+}
+
+/* Give it a shorter name */
+#define ins(i, p) append_insn((i), (p))
+
+/*
+ * Give a register name (in 'reg'), generate instruction to
+ * load register into an eBPF register rd:
+ *   'ldd target_reg, offset(ctx_reg)', where:
+ * ctx_reg is pre initialized to pointer of 'struct pt_regs'.
+ */
+static int
+gen_ldx_reg_from_ctx(struct bpf_insn_pos *pos, int ctx_reg,
+		     const char *reg, int target_reg)
+{
+	int offset;
+
+	if (arch_get_reg_info(reg, &offset)) {
+		pr_err("bpf: prologue: failed to get register %s\n",
+		       reg);
+		return -1;
+	}
+	ins(BPF_LDX_MEM(BPF_DW, target_reg, ctx_reg, offset), pos);
+
+	if (check_pos(pos))
+		return -ERANGE;
+	return 0;
+}
+
+/*
+ * Generate a BPF_FUNC_probe_read function call.
+ *
+ * src_base_addr_reg is a register holding base address,
+ * dst_addr_reg is a register holding dest address (on stack),
+ * result is:
+ *
+ *  *[dst_addr_reg] = *([src_base_addr_reg] + offset)
+ *
+ * Arguments of BPF_FUNC_probe_read:
+ *     ARG1: ptr to stack (dest)
+ *     ARG2: size (8)
+ *     ARG3: unsafe ptr (src)
+ */
+static int
+gen_read_mem(struct bpf_insn_pos *pos,
+	     int src_base_addr_reg,
+	     int dst_addr_reg,
+	     long offset)
+{
+	/* mov arg3, src_base_addr_reg */
+	if (src_base_addr_reg != BPF_REG_ARG3)
+		ins(BPF_MOV64_REG(BPF_REG_ARG3, src_base_addr_reg), pos);
+	/* add arg3, #offset */
+	if (offset)
+		ins(BPF_ALU64_IMM(BPF_ADD, BPF_REG_ARG3, offset), pos);
+
+	/* mov arg2, #reg_size */
+	ins(BPF_ALU64_IMM(BPF_MOV, BPF_REG_ARG2, BPF_REG_SIZE), pos);
+
+	/* mov arg1, dst_addr_reg */
+	if (dst_addr_reg != BPF_REG_ARG1)
+		ins(BPF_MOV64_REG(BPF_REG_ARG1, dst_addr_reg), pos);
+
+	/* Call probe_read  */
+	ins(BPF_EMIT_CALL(BPF_FUNC_probe_read), pos);
+	/*
+	 * Error processing: if read fail, goto error code,
+	 * will be relocated. Target should be the start of
+	 * error processing code.
+	 */
+	ins(BPF_JMP_IMM(BPF_JNE, BPF_REG_0, 0, JMP_TO_ERROR_CODE),
+	    pos);
+
+	if (check_pos(pos))
+		return -ERANGE;
+	return 0;
+}
+
+/*
+ * Each arg should be bare register. Fetch and save them into argument
+ * registers (r3 - r5).
+ *
+ * BPF_REG_1 should have been initialized with pointer to
+ * 'struct pt_regs'.
+ */
+static int
+gen_prologue_fastpath(struct bpf_insn_pos *pos,
+		      struct probe_trace_arg *args, int nargs)
+{
+	int i;
+
+	for (i = 0; i < nargs; i++)
+		if (gen_ldx_reg_from_ctx(pos, BPF_REG_1, args[i].value,
+					 BPF_PROLOGUE_START_ARG_REG + i))
+			goto errout;
+
+	if (check_pos(pos))
+		goto errout;
+	return 0;
+errout:
+	return -1;
+}
+
+/*
+ * Slow path:
+ *   At least one argument has the form of 'offset($rx)'.
+ *
+ * Following code first stores them into stack, then loads all of then
+ * to r2 - r5.
+ * Before final loading, the final result should be:
+ *
+ * low address
+ * BPF_REG_FP - 24  ARG3
+ * BPF_REG_FP - 16  ARG2
+ * BPF_REG_FP - 8   ARG1
+ * BPF_REG_FP
+ * high address
+ *
+ * For each argument (described as: offn(...off2(off1(reg)))),
+ * generates following code:
+ *
+ *  r7 <- fp
+ *  r7 <- r7 - stack_offset  // Ideal code should initialize r7 using
+ *                           // fp before generating args. However,
+ *                           // eBPF won't regard r7 as stack pointer
+ *                           // if it is generated by minus 8 from
+ *                           // another stack pointer except fp.
+ *                           // This is why we have to set r7
+ *                           // to fp for each variable.
+ *  r3 <- value of 'reg'-> generated using gen_ldx_reg_from_ctx()
+ *  (r7) <- r3       // skip following instructions for bare reg
+ *  r3 <- r3 + off1  . // skip if off1 == 0
+ *  r2 <- 8           \
+ *  r1 <- r7           |-> generated by gen_read_mem()
+ *  call probe_read    /
+ *  jnei r0, 0, err  ./
+ *  r3 <- (r7)
+ *  r3 <- r3 + off2  . // skip if off2 == 0
+ *  r2 <- 8           \  // r2 may be broken by probe_read, so set again
+ *  r1 <- r7           |-> generated by gen_read_mem()
+ *  call probe_read    /
+ *  jnei r0, 0, err  ./
+ *  ...
+ */
+static int
+gen_prologue_slowpath(struct bpf_insn_pos *pos,
+		      struct probe_trace_arg *args, int nargs)
+{
+	int i;
+
+	for (i = 0; i < nargs; i++) {
+		struct probe_trace_arg *arg = &args[i];
+		const char *reg = arg->value;
+		struct probe_trace_arg_ref *ref = NULL;
+		int stack_offset = (i + 1) * -8;
+
+		pr_debug("prologue: fetch arg %d, base reg is %s\n",
+			 i, reg);
+
+		/* value of base register is stored into ARG3 */
+		if (gen_ldx_reg_from_ctx(pos, BPF_REG_CTX, reg,
+					 BPF_REG_ARG3)) {
+			pr_err("prologue: failed to get offset of register %s\n",
+			       reg);
+			goto errout;
+		}
+
+		/* Make r7 the stack pointer. */
+		ins(BPF_MOV64_REG(BPF_REG_7, BPF_REG_FP), pos);
+		/* r7 += -8 */
+		ins(BPF_ALU64_IMM(BPF_ADD, BPF_REG_7, stack_offset), pos);
+		/*
+		 * Store r3 (base register) onto stack
+		 * Ensure fp[offset] is set.
+		 * fp is the only valid base register when storing
+		 * into stack. We are not allowed to use r7 as base
+		 * register here.
+		 */
+		ins(BPF_STX_MEM(BPF_DW, BPF_REG_FP, BPF_REG_ARG3,
+				stack_offset), pos);
+
+		ref = arg->ref;
+		while (ref) {
+			pr_debug("prologue: arg %d: offset %ld\n",
+				 i, ref->offset);
+			if (gen_read_mem(pos, BPF_REG_3, BPF_REG_7,
+					 ref->offset)) {
+				pr_err("prologue: failed to generate probe_read function call\n");
+				goto errout;
+			}
+
+			ref = ref->next;
+			/*
+			 * Load previous result into ARG3. Use
+			 * BPF_REG_FP instead of r7 because verifier
+			 * allows FP based addressing only.
+			 */
+			if (ref)
+				ins(BPF_LDX_MEM(BPF_DW, BPF_REG_ARG3,
+						BPF_REG_FP, stack_offset), pos);
+		}
+	}
+
+	/* Final pass: read to registers */
+	for (i = 0; i < nargs; i++)
+		ins(BPF_LDX_MEM(BPF_DW, BPF_PROLOGUE_START_ARG_REG + i,
+				BPF_REG_FP, -BPF_REG_SIZE * (i + 1)), pos);
+
+	ins(BPF_JMP_IMM(BPF_JA, BPF_REG_0, 0, JMP_TO_SUCCESS_CODE), pos);
+
+	if (check_pos(pos))
+		goto errout;
+	return 0;
+errout:
+	return -1;
+}
+
+static int
+prologue_relocate(struct bpf_insn_pos *pos, struct bpf_insn *error_code,
+	    struct bpf_insn *success_code, struct bpf_insn *user_code)
+{
+	struct bpf_insn *insn;
+
+	if (check_pos(pos))
+		return -ERANGE;
+
+	for (insn = pos->begin; insn < pos->pos; insn++) {
+		u8 class = BPF_CLASS(insn->code);
+		u8 opcode;
+
+		if (class != BPF_JMP)
+			continue;
+		opcode = BPF_OP(insn->code);
+		if (opcode == BPF_CALL)
+			continue;
+
+		switch (insn->off) {
+		case JMP_TO_ERROR_CODE:
+			insn->off = error_code - (insn + 1);
+			break;
+		case JMP_TO_SUCCESS_CODE:
+			insn->off = success_code - (insn + 1);
+			break;
+		case JMP_TO_USER_CODE:
+			insn->off = user_code - (insn + 1);
+			break;
+		default:
+			pr_err("bpf prologue: internal error: relocation failed\n");
+			return -1;
+		}
+	}
+	return 0;
+}
+
+int bpf__gen_prologue(struct probe_trace_arg *args, int nargs,
+		      struct bpf_insn *new_prog, size_t *new_cnt,
+		      size_t cnt_space)
+{
+	struct bpf_insn *success_code = NULL;
+	struct bpf_insn *error_code = NULL;
+	struct bpf_insn *user_code = NULL;
+	struct bpf_insn_pos pos;
+	bool fastpath = true;
+	int i;
+
+	if (!new_prog || !new_cnt)
+		return -EINVAL;
+
+	pos.begin = new_prog;
+	pos.end = new_prog + cnt_space;
+	pos.pos = new_prog;
+
+	if (!nargs) {
+		ins(BPF_ALU64_IMM(BPF_MOV, BPF_PROLOGUE_FETCH_RESULT_REG, 0),
+		    &pos);
+
+		if (check_pos(&pos))
+			goto errout;
+
+		*new_cnt = pos_get_cnt(&pos);
+		return 0;
+	}
+
+	if (nargs > BPF_PROLOGUE_MAX_ARGS)
+		nargs = BPF_PROLOGUE_MAX_ARGS;
+	if (cnt_space > BPF_MAXINSNS)
+		cnt_space = BPF_MAXINSNS;
+
+	/* First pass: validation */
+	for (i = 0; i < nargs; i++) {
+		struct probe_trace_arg_ref *ref = args[i].ref;
+
+		if (args[i].value[0] == '@') {
+			/* TODO: fetch global variable */
+			pr_err("bpf: prologue: global %s%+ld not support\n",
+				args[i].value, ref ? ref->offset : 0);
+			return -ENOTSUP;
+		}
+
+		while (ref) {
+			/* fastpath is true if all args has ref == NULL */
+			fastpath = false;
+
+			/*
+			 * Instruction encodes immediate value using
+			 * s32, ref->offset is long. On systems which
+			 * can't fill long in s32, refuse to process if
+			 * ref->offset too large (or small).
+			 */
+#ifdef __LP64__
+#define OFFSET_MAX	((1LL << 31) - 1)
+#define OFFSET_MIN	((1LL << 31) * -1)
+			if (ref->offset > OFFSET_MAX ||
+					ref->offset < OFFSET_MIN) {
+				pr_err("bpf: prologue: offset out of bound: %ld\n",
+				       ref->offset);
+				return -E2BIG;
+			}
+#endif
+			ref = ref->next;
+		}
+	}
+	pr_debug("prologue: pass validation\n");
+
+	if (fastpath) {
+		/* If all variables are registers... */
+		pr_debug("prologue: fast path\n");
+		if (gen_prologue_fastpath(&pos, args, nargs))
+			goto errout;
+	} else {
+		pr_debug("prologue: slow path\n");
+
+		/* Initialization: move ctx to a callee saved register. */
+		ins(BPF_MOV64_REG(BPF_REG_CTX, BPF_REG_ARG1), &pos);
+
+		if (gen_prologue_slowpath(&pos, args, nargs))
+			goto errout;
+		/*
+		 * start of ERROR_CODE (only slow pass needs error code)
+		 *   mov r2 <- 1
+		 *   goto usercode
+		 */
+		error_code = pos.pos;
+		ins(BPF_ALU64_IMM(BPF_MOV, BPF_PROLOGUE_FETCH_RESULT_REG, 1),
+		    &pos);
+
+		for (i = 0; i < nargs; i++)
+			ins(BPF_ALU64_IMM(BPF_MOV,
+					  BPF_PROLOGUE_START_ARG_REG + i,
+					  0),
+			    &pos);
+		ins(BPF_JMP_IMM(BPF_JA, BPF_REG_0, 0, JMP_TO_USER_CODE),
+				&pos);
+	}
+
+	/*
+	 * start of SUCCESS_CODE:
+	 *   mov r2 <- 0
+	 *   goto usercode  // skip
+	 */
+	success_code = pos.pos;
+	ins(BPF_ALU64_IMM(BPF_MOV, BPF_PROLOGUE_FETCH_RESULT_REG, 0), &pos);
+
+	/*
+	 * start of USER_CODE:
+	 *   Restore ctx to r1
+	 */
+	user_code = pos.pos;
+	if (!fastpath) {
+		/*
+		 * Only slow path needs restoring of ctx. In fast path,
+		 * register are loaded directly from r1.
+		 */
+		ins(BPF_MOV64_REG(BPF_REG_ARG1, BPF_REG_CTX), &pos);
+		if (prologue_relocate(&pos, error_code, success_code,
+				      user_code))
+			goto errout;
+	}
+
+	if (check_pos(&pos))
+		goto errout;
+
+	*new_cnt = pos_get_cnt(&pos);
+	return 0;
+errout:
+	return -ERANGE;
+}
diff --git a/tools/perf/util/bpf-prologue.h b/tools/perf/util/bpf-prologue.h
new file mode 100644
index 0000000..f1e4c5d
--- /dev/null
+++ b/tools/perf/util/bpf-prologue.h
@@ -0,0 +1,34 @@
+/*
+ * Copyright (C) 2015, He Kuang <hekuang@xxxxxxxxxx>
+ * Copyright (C) 2015, Huawei Inc.
+ */
+#ifndef __BPF_PROLOGUE_H
+#define __BPF_PROLOGUE_H
+
+#include <linux/compiler.h>
+#include <linux/filter.h>
+#include "probe-event.h"
+
+#define BPF_PROLOGUE_MAX_ARGS 3
+#define BPF_PROLOGUE_START_ARG_REG BPF_REG_3
+#define BPF_PROLOGUE_FETCH_RESULT_REG BPF_REG_2
+
+#ifdef HAVE_BPF_PROLOGUE
+int bpf__gen_prologue(struct probe_trace_arg *args, int nargs,
+		      struct bpf_insn *new_prog, size_t *new_cnt,
+		      size_t cnt_space);
+#else
+static inline int
+bpf__gen_prologue(struct probe_trace_arg *args __maybe_unused,
+		  int nargs __maybe_unused,
+		  struct bpf_insn *new_prog __maybe_unused,
+		  size_t *new_cnt,
+		  size_t cnt_space __maybe_unused)
+{
+	if (!new_cnt)
+		return -EINVAL;
+	*new_cnt = 0;
+	return 0;
+}
+#endif
+#endif /* __BPF_PROLOGUE_H */
-- 
1.8.3.4

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/