Merge branch 'bpf-romap-known-scalars'

Andrii Nakryiko says:

====================
With BPF maps supporting direct map access (currently, array_map w/ single
element, used for global data) that are read-only both from system call and
BPF side, it's possible for BPF verifier to track its contents as known
constants.

Now it's possible for user-space control app to pre-initialize read-only map
(e.g., for .rodata section) with user-provided flags and parameters and rely
on BPF verifier to detect and eliminate dead code resulting from specific
combination of input parameters.

v1->v2:
- BPF_F_RDONLY means nothing, stick to just map->frozen (Daniel);
- stick to passing just offset into map_direct_value_addr (Martin).
====================
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>

kernel/bpf/verifier.c

@@ -2739,6 +2739,41 @@ static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
 	reg->smax_value = reg->umax_value;
 }
 
+static bool bpf_map_is_rdonly(const struct bpf_map *map)
+{
+	return (map->map_flags & BPF_F_RDONLY_PROG) && map->frozen;
+}
+
+static int bpf_map_direct_read(struct bpf_map *map, int off, int size, u64 *val)
+{
+	void *ptr;
+	u64 addr;
+	int err;
+
+	err = map->ops->map_direct_value_addr(map, &addr, off);
+	if (err)
+		return err;
+	ptr = (void *)addr + off;
+
+	switch (size) {
+	case sizeof(u8):
+		*val = (u64)*(u8 *)ptr;
+		break;
+	case sizeof(u16):
+		*val = (u64)*(u16 *)ptr;
+		break;
+	case sizeof(u32):
+		*val = (u64)*(u32 *)ptr;
+		break;
+	case sizeof(u64):
+		*val = *(u64 *)ptr;
+		break;
+	default:
+		return -EINVAL;
+	}
+	return 0;
+}
+
 /* check whether memory at (regno + off) is accessible for t = (read | write)
  * if t==write, value_regno is a register which value is stored into memory
  * if t==read, value_regno is a register which will receive the value from memory
@@ -2776,9 +2811,27 @@ static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regn
 		if (err)
 			return err;
 		err = check_map_access(env, regno, off, size, false);
-		if (!err && t == BPF_READ && value_regno >= 0)
-			mark_reg_unknown(env, regs, value_regno);
+		if (!err && t == BPF_READ && value_regno >= 0) {
+			struct bpf_map *map = reg->map_ptr;
+
+			/* if map is read-only, track its contents as scalars */
+			if (tnum_is_const(reg->var_off) &&
+			    bpf_map_is_rdonly(map) &&
+			    map->ops->map_direct_value_addr) {
+				int map_off = off + reg->var_off.value;
+				u64 val = 0;
 
+				err = bpf_map_direct_read(map, map_off, size,
+							  &val);
+				if (err)
+					return err;
+
+				regs[value_regno].type = SCALAR_VALUE;
+				__mark_reg_known(&regs[value_regno], val);
+			} else {
+				mark_reg_unknown(env, regs, value_regno);
+			}
+		}
 	} else if (reg->type == PTR_TO_CTX) {
 		enum bpf_reg_type reg_type = SCALAR_VALUE;
 

tools/testing/selftests/bpf/prog_tests/rdonly_maps.c

+// SPDX-License-Identifier: GPL-2.0
+#include <test_progs.h>
+
+struct bss {
+	unsigned did_run;
+	unsigned iters;
+	unsigned sum;
+};
+
+struct rdonly_map_subtest {
+	const char *subtest_name;
+	const char *prog_name;
+	unsigned exp_iters;
+	unsigned exp_sum;
+};
+
+void test_rdonly_maps(void)
+{
+	const char *prog_name_skip_loop = "raw_tracepoint/sys_enter:skip_loop";
+	const char *prog_name_part_loop = "raw_tracepoint/sys_enter:part_loop";
+	const char *prog_name_full_loop = "raw_tracepoint/sys_enter:full_loop";
+	const char *file = "test_rdonly_maps.o";
+	struct rdonly_map_subtest subtests[] = {
+		{ "skip loop", prog_name_skip_loop, 0, 0 },
+		{ "part loop", prog_name_part_loop, 3, 2 + 3 + 4 },
+		{ "full loop", prog_name_full_loop, 4, 2 + 3 + 4 + 5 },
+	};
+	int i, err, zero = 0, duration = 0;
+	struct bpf_link *link = NULL;
+	struct bpf_program *prog;
+	struct bpf_map *bss_map;
+	struct bpf_object *obj;
+	struct bss bss;
+
+	obj = bpf_object__open_file(file, NULL);
+	if (CHECK(IS_ERR(obj), "obj_open", "err %ld\n", PTR_ERR(obj)))
+		return;
+
+	bpf_object__for_each_program(prog, obj) {
+		bpf_program__set_raw_tracepoint(prog);
+	}
+
+	err = bpf_object__load(obj);
+	if (CHECK(err, "obj_load", "err %d errno %d\n", err, errno))
+		goto cleanup;
+
+	bss_map = bpf_object__find_map_by_name(obj, "test_rdo.bss");
+	if (CHECK(!bss_map, "find_bss_map", "failed\n"))
+		goto cleanup;
+
+	for (i = 0; i < ARRAY_SIZE(subtests); i++) {
+		const struct rdonly_map_subtest *t = &subtests[i];
+
+		if (!test__start_subtest(t->subtest_name))
+			continue;
+
+		prog = bpf_object__find_program_by_title(obj, t->prog_name);
+		if (CHECK(!prog, "find_prog", "prog '%s' not found\n",
+			  t->prog_name))
+			goto cleanup;
+
+		memset(&bss, 0, sizeof(bss));
+		err = bpf_map_update_elem(bpf_map__fd(bss_map), &zero, &bss, 0);
+		if (CHECK(err, "set_bss", "failed to set bss data: %d\n", err))
+			goto cleanup;
+
+		link = bpf_program__attach_raw_tracepoint(prog, "sys_enter");
+		if (CHECK(IS_ERR(link), "attach_prog", "prog '%s', err %ld\n",
+			  t->prog_name, PTR_ERR(link))) {
+			link = NULL;
+			goto cleanup;
+		}
+
+		/* trigger probe */
+		usleep(1);
+
+		bpf_link__destroy(link);
+		link = NULL;
+
+		err = bpf_map_lookup_elem(bpf_map__fd(bss_map), &zero, &bss);
+		if (CHECK(err, "get_bss", "failed to get bss data: %d\n", err))
+			goto cleanup;
+		if (CHECK(bss.did_run == 0, "check_run",
+			  "prog '%s' didn't run?\n", t->prog_name))
+			goto cleanup;
+		if (CHECK(bss.iters != t->exp_iters, "check_iters",
+			  "prog '%s' iters: %d, expected: %d\n",
+			  t->prog_name, bss.iters, t->exp_iters))
+			goto cleanup;
+		if (CHECK(bss.sum != t->exp_sum, "check_sum",
+			  "prog '%s' sum: %d, expected: %d\n",
+			  t->prog_name, bss.sum, t->exp_sum))
+			goto cleanup;
+	}
+
+cleanup:
+	bpf_link__destroy(link);
+	bpf_object__close(obj);
+}

tools/testing/selftests/bpf/progs/test_rdonly_maps.c

+// SPDX-License-Identifier: GPL-2.0
+// Copyright (c) 2019 Facebook
+
+#include <linux/ptrace.h>
+#include <linux/bpf.h>
+#include "bpf_helpers.h"
+
+static volatile const struct {
+	unsigned a[4];
+	/*
+	 * if the struct's size is multiple of 16, compiler will put it into
+	 * .rodata.cst16 section, which is not recognized by libbpf; work
+	 * around this by ensuring we don't have 16-aligned struct
+	 */
+	char _y;
+} rdonly_values = { .a = {2, 3, 4, 5} };
+
+static volatile struct {
+	unsigned did_run;
+	unsigned iters;
+	unsigned sum;
+} res;
+
+SEC("raw_tracepoint/sys_enter:skip_loop")
+int skip_loop(struct pt_regs *ctx)
+{
+	/* prevent compiler to optimize everything out */
+	unsigned * volatile p = (void *)&rdonly_values.a;
+	unsigned iters = 0, sum = 0;
+
+	/* we should never enter this loop */
+	while (*p & 1) {
+		iters++;
+		sum += *p;
+		p++;
+	}
+	res.did_run = 1;
+	res.iters = iters;
+	res.sum = sum;
+	return 0;
+}
+
+SEC("raw_tracepoint/sys_enter:part_loop")
+int part_loop(struct pt_regs *ctx)
+{
+	/* prevent compiler to optimize everything out */
+	unsigned * volatile p = (void *)&rdonly_values.a;
+	unsigned iters = 0, sum = 0;
+
+	/* validate verifier can derive loop termination */
+	while (*p < 5) {
+		iters++;
+		sum += *p;
+		p++;
+	}
+	res.did_run = 1;
+	res.iters = iters;
+	res.sum = sum;
+	return 0;
+}
+
+SEC("raw_tracepoint/sys_enter:full_loop")
+int full_loop(struct pt_regs *ctx)
+{
+	/* prevent compiler to optimize everything out */
+	unsigned * volatile p = (void *)&rdonly_values.a;
+	int i = sizeof(rdonly_values.a) / sizeof(rdonly_values.a[0]);
+	unsigned iters = 0, sum = 0;
+
+	/* validate verifier can allow full loop as well */
+	while (i > 0 ) {
+		iters++;
+		sum += *p;
+		p++;
+		i--;
+	}
+	res.did_run = 1;
+	res.iters = iters;
+	res.sum = sum;
+	return 0;
+}
+
+char _license[] SEC("license") = "GPL";