* [PATCH bpf-next 1/2] bpf: allow wide (u64) aligned stores for some fields of bpf_sock_addr
@ 2019-06-28 23:10 Stanislav Fomichev
2019-06-28 23:10 ` [PATCH bpf-next 2/2] selftests/bpf: add verifier tests for wide stores Stanislav Fomichev
2019-06-30 5:52 ` [PATCH bpf-next 1/2] bpf: allow wide (u64) aligned stores for some fields of bpf_sock_addr Yonghong Song
0 siblings, 2 replies; 11+ messages in thread
From: Stanislav Fomichev @ 2019-06-28 23:10 UTC (permalink / raw)
To: netdev, bpf
Cc: davem, ast, daniel, Stanislav Fomichev, Andrii Nakryiko,
Yonghong Song, kernel test robot
Since commit cd17d7770578 ("bpf/tools: sync bpf.h") clang decided
that it can do a single u64 store into user_ip6[2] instead of two
separate u32 ones:
# 17: (18) r2 = 0x100000000000000
# ; ctx->user_ip6[2] = bpf_htonl(DST_REWRITE_IP6_2);
# 19: (7b) *(u64 *)(r1 +16) = r2
# invalid bpf_context access off=16 size=8
From the compiler point of view it does look like a correct thing
to do, so let's support it on the kernel side.
Credit to Andrii Nakryiko for a proper implementation of
bpf_ctx_wide_store_ok.
Cc: Andrii Nakryiko <andriin@fb.com>
Cc: Yonghong Song <yhs@fb.com>
Fixes: cd17d7770578 ("bpf/tools: sync bpf.h")
Reported-by: kernel test robot <rong.a.chen@intel.com>
Signed-off-by: Stanislav Fomichev <sdf@google.com>
---
include/linux/filter.h | 6 ++++++
net/core/filter.c | 22 ++++++++++++++--------
2 files changed, 20 insertions(+), 8 deletions(-)
diff --git a/include/linux/filter.h b/include/linux/filter.h
index 340f7d648974..3901007e36f1 100644
--- a/include/linux/filter.h
+++ b/include/linux/filter.h
@@ -746,6 +746,12 @@ bpf_ctx_narrow_access_ok(u32 off, u32 size, u32 size_default)
return size <= size_default && (size & (size - 1)) == 0;
}
+#define bpf_ctx_wide_store_ok(off, size, type, field) \
+ (size == sizeof(__u64) && \
+ off >= offsetof(type, field) && \
+ off + sizeof(__u64) <= offsetofend(type, field) && \
+ off % sizeof(__u64) == 0)
+
#define bpf_classic_proglen(fprog) (fprog->len * sizeof(fprog->filter[0]))
static inline void bpf_prog_lock_ro(struct bpf_prog *fp)
diff --git a/net/core/filter.c b/net/core/filter.c
index dc8534be12fc..5d33f2146dab 100644
--- a/net/core/filter.c
+++ b/net/core/filter.c
@@ -6849,6 +6849,16 @@ static bool sock_addr_is_valid_access(int off, int size,
if (!bpf_ctx_narrow_access_ok(off, size, size_default))
return false;
} else {
+ if (bpf_ctx_wide_store_ok(off, size,
+ struct bpf_sock_addr,
+ user_ip6))
+ return true;
+
+ if (bpf_ctx_wide_store_ok(off, size,
+ struct bpf_sock_addr,
+ msg_src_ip6))
+ return true;
+
if (size != size_default)
return false;
}
@@ -7689,9 +7699,6 @@ static u32 xdp_convert_ctx_access(enum bpf_access_type type,
/* SOCK_ADDR_STORE_NESTED_FIELD_OFF() has semantic similar to
* SOCK_ADDR_LOAD_NESTED_FIELD_SIZE_OFF() but for store operation.
*
- * It doesn't support SIZE argument though since narrow stores are not
- * supported for now.
- *
* In addition it uses Temporary Field TF (member of struct S) as the 3rd
* "register" since two registers available in convert_ctx_access are not
* enough: we can't override neither SRC, since it contains value to store, nor
@@ -7699,7 +7706,7 @@ static u32 xdp_convert_ctx_access(enum bpf_access_type type,
* instructions. But we need a temporary place to save pointer to nested
* structure whose field we want to store to.
*/
-#define SOCK_ADDR_STORE_NESTED_FIELD_OFF(S, NS, F, NF, OFF, TF) \
+#define SOCK_ADDR_STORE_NESTED_FIELD_OFF(S, NS, F, NF, SIZE, OFF, TF) \
do { \
int tmp_reg = BPF_REG_9; \
if (si->src_reg == tmp_reg || si->dst_reg == tmp_reg) \
@@ -7710,8 +7717,7 @@ static u32 xdp_convert_ctx_access(enum bpf_access_type type,
offsetof(S, TF)); \
*insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(S, F), tmp_reg, \
si->dst_reg, offsetof(S, F)); \
- *insn++ = BPF_STX_MEM( \
- BPF_FIELD_SIZEOF(NS, NF), tmp_reg, si->src_reg, \
+ *insn++ = BPF_STX_MEM(SIZE, tmp_reg, si->src_reg, \
bpf_target_off(NS, NF, FIELD_SIZEOF(NS, NF), \
target_size) \
+ OFF); \
@@ -7723,8 +7729,8 @@ static u32 xdp_convert_ctx_access(enum bpf_access_type type,
TF) \
do { \
if (type == BPF_WRITE) { \
- SOCK_ADDR_STORE_NESTED_FIELD_OFF(S, NS, F, NF, OFF, \
- TF); \
+ SOCK_ADDR_STORE_NESTED_FIELD_OFF(S, NS, F, NF, SIZE, \
+ OFF, TF); \
} else { \
SOCK_ADDR_LOAD_NESTED_FIELD_SIZE_OFF( \
S, NS, F, NF, SIZE, OFF); \
--
2.22.0.410.gd8fdbe21b5-goog
^ permalink raw reply related [flat|nested] 11+ messages in thread
* [PATCH bpf-next 2/2] selftests/bpf: add verifier tests for wide stores
2019-06-28 23:10 [PATCH bpf-next 1/2] bpf: allow wide (u64) aligned stores for some fields of bpf_sock_addr Stanislav Fomichev
@ 2019-06-28 23:10 ` Stanislav Fomichev
2019-06-30 6:01 ` Yonghong Song
2019-06-30 5:52 ` [PATCH bpf-next 1/2] bpf: allow wide (u64) aligned stores for some fields of bpf_sock_addr Yonghong Song
1 sibling, 1 reply; 11+ messages in thread
From: Stanislav Fomichev @ 2019-06-28 23:10 UTC (permalink / raw)
To: netdev, bpf
Cc: davem, ast, daniel, Stanislav Fomichev, Andrii Nakryiko, Yonghong Song
Make sure that wide stores are allowed at proper (aligned) addresses.
Note that user_ip6 is naturally aligned on 8-byte boundary, so
correct addresses are user_ip6[0] and user_ip6[2]. msg_src_ip6 is,
however, aligned on a 4-byte bondary, so only msg_src_ip6[1]
can be wide-stored.
Cc: Andrii Nakryiko <andriin@fb.com>
Cc: Yonghong Song <yhs@fb.com>
Signed-off-by: Stanislav Fomichev <sdf@google.com>
---
tools/testing/selftests/bpf/test_verifier.c | 17 ++++++--
.../selftests/bpf/verifier/wide_store.c | 40 +++++++++++++++++++
2 files changed, 54 insertions(+), 3 deletions(-)
create mode 100644 tools/testing/selftests/bpf/verifier/wide_store.c
diff --git a/tools/testing/selftests/bpf/test_verifier.c b/tools/testing/selftests/bpf/test_verifier.c
index c5514daf8865..b0773291012a 100644
--- a/tools/testing/selftests/bpf/test_verifier.c
+++ b/tools/testing/selftests/bpf/test_verifier.c
@@ -105,6 +105,7 @@ struct bpf_test {
__u64 data64[TEST_DATA_LEN / 8];
};
} retvals[MAX_TEST_RUNS];
+ enum bpf_attach_type expected_attach_type;
};
/* Note we want this to be 64 bit aligned so that the end of our array is
@@ -850,6 +851,7 @@ static void do_test_single(struct bpf_test *test, bool unpriv,
int fd_prog, expected_ret, alignment_prevented_execution;
int prog_len, prog_type = test->prog_type;
struct bpf_insn *prog = test->insns;
+ struct bpf_load_program_attr attr;
int run_errs, run_successes;
int map_fds[MAX_NR_MAPS];
const char *expected_err;
@@ -881,8 +883,17 @@ static void do_test_single(struct bpf_test *test, bool unpriv,
pflags |= BPF_F_STRICT_ALIGNMENT;
if (test->flags & F_NEEDS_EFFICIENT_UNALIGNED_ACCESS)
pflags |= BPF_F_ANY_ALIGNMENT;
- fd_prog = bpf_verify_program(prog_type, prog, prog_len, pflags,
- "GPL", 0, bpf_vlog, sizeof(bpf_vlog), 4);
+
+ memset(&attr, 0, sizeof(attr));
+ attr.prog_type = prog_type;
+ attr.expected_attach_type = test->expected_attach_type;
+ attr.insns = prog;
+ attr.insns_cnt = prog_len;
+ attr.license = "GPL";
+ attr.log_level = 4;
+ attr.prog_flags = pflags;
+
+ fd_prog = bpf_load_program_xattr(&attr, bpf_vlog, sizeof(bpf_vlog));
if (fd_prog < 0 && !bpf_probe_prog_type(prog_type, 0)) {
printf("SKIP (unsupported program type %d)\n", prog_type);
skips++;
@@ -912,7 +923,7 @@ static void do_test_single(struct bpf_test *test, bool unpriv,
printf("FAIL\nUnexpected success to load!\n");
goto fail_log;
}
- if (!strstr(bpf_vlog, expected_err)) {
+ if (!expected_err || !strstr(bpf_vlog, expected_err)) {
printf("FAIL\nUnexpected error message!\n\tEXP: %s\n\tRES: %s\n",
expected_err, bpf_vlog);
goto fail_log;
diff --git a/tools/testing/selftests/bpf/verifier/wide_store.c b/tools/testing/selftests/bpf/verifier/wide_store.c
new file mode 100644
index 000000000000..c6385f45b114
--- /dev/null
+++ b/tools/testing/selftests/bpf/verifier/wide_store.c
@@ -0,0 +1,40 @@
+#define BPF_SOCK_ADDR(field, off, res, err) \
+{ \
+ "wide store to bpf_sock_addr." #field "[" #off "]", \
+ .insns = { \
+ BPF_MOV64_IMM(BPF_REG_0, 1), \
+ BPF_STX_MEM(BPF_DW, BPF_REG_1, BPF_REG_0, \
+ offsetof(struct bpf_sock_addr, field[off])), \
+ BPF_EXIT_INSN(), \
+ }, \
+ .result = res, \
+ .prog_type = BPF_PROG_TYPE_CGROUP_SOCK_ADDR, \
+ .expected_attach_type = BPF_CGROUP_UDP6_SENDMSG, \
+ .errstr = err, \
+}
+
+/* user_ip6[0] is u64 aligned */
+BPF_SOCK_ADDR(user_ip6, 0, ACCEPT,
+ NULL),
+BPF_SOCK_ADDR(user_ip6, 1, REJECT,
+ "invalid bpf_context access off=12 size=8"),
+BPF_SOCK_ADDR(user_ip6, 2, ACCEPT,
+ NULL),
+BPF_SOCK_ADDR(user_ip6, 3, REJECT,
+ "invalid bpf_context access off=20 size=8"),
+BPF_SOCK_ADDR(user_ip6, 4, REJECT,
+ "invalid bpf_context access off=24 size=8"),
+
+/* msg_src_ip6[0] is _not_ u64 aligned */
+BPF_SOCK_ADDR(msg_src_ip6, 0, REJECT,
+ "invalid bpf_context access off=44 size=8"),
+BPF_SOCK_ADDR(msg_src_ip6, 1, ACCEPT,
+ NULL),
+BPF_SOCK_ADDR(msg_src_ip6, 2, REJECT,
+ "invalid bpf_context access off=52 size=8"),
+BPF_SOCK_ADDR(msg_src_ip6, 3, REJECT,
+ "invalid bpf_context access off=56 size=8"),
+BPF_SOCK_ADDR(msg_src_ip6, 4, REJECT,
+ "invalid bpf_context access off=60 size=8"),
+
+#undef BPF_SOCK_ADDR
--
2.22.0.410.gd8fdbe21b5-goog
^ permalink raw reply related [flat|nested] 11+ messages in thread
* Re: [PATCH bpf-next 1/2] bpf: allow wide (u64) aligned stores for some fields of bpf_sock_addr
2019-06-28 23:10 [PATCH bpf-next 1/2] bpf: allow wide (u64) aligned stores for some fields of bpf_sock_addr Stanislav Fomichev
2019-06-28 23:10 ` [PATCH bpf-next 2/2] selftests/bpf: add verifier tests for wide stores Stanislav Fomichev
@ 2019-06-30 5:52 ` Yonghong Song
2019-07-01 15:36 ` Andrii Nakryiko
2019-07-01 16:01 ` Stanislav Fomichev
1 sibling, 2 replies; 11+ messages in thread
From: Yonghong Song @ 2019-06-30 5:52 UTC (permalink / raw)
To: Stanislav Fomichev, netdev, bpf
Cc: davem, ast, daniel, Andrii Nakryiko, kernel test robot
On 6/28/19 4:10 PM, Stanislav Fomichev wrote:
> Since commit cd17d7770578 ("bpf/tools: sync bpf.h") clang decided
> that it can do a single u64 store into user_ip6[2] instead of two
> separate u32 ones:
>
> # 17: (18) r2 = 0x100000000000000
> # ; ctx->user_ip6[2] = bpf_htonl(DST_REWRITE_IP6_2);
> # 19: (7b) *(u64 *)(r1 +16) = r2
> # invalid bpf_context access off=16 size=8
>
> From the compiler point of view it does look like a correct thing
> to do, so let's support it on the kernel side.
>
> Credit to Andrii Nakryiko for a proper implementation of
> bpf_ctx_wide_store_ok.
>
> Cc: Andrii Nakryiko <andriin@fb.com>
> Cc: Yonghong Song <yhs@fb.com>
> Fixes: cd17d7770578 ("bpf/tools: sync bpf.h")
> Reported-by: kernel test robot <rong.a.chen@intel.com>
> Signed-off-by: Stanislav Fomichev <sdf@google.com>
The change looks good to me with the following nits:
1. could you add a cover letter for the patch set?
typically if the number of patches is more than one,
it would be a good practice with a cover letter.
See bpf_devel_QA.rst .
2. with this change, the comments in uapi bpf.h
are not accurate any more.
__u32 user_ip6[4]; /* Allows 1,2,4-byte read an 4-byte write.
* Stored in network byte order.
*/
__u32 msg_src_ip6[4]; /* Allows 1,2,4-byte read an 4-byte write.
* Stored in network byte order.
*/
now for stores, aligned 8-byte write is permitted.
could you update this as well?
From the typical usage pattern, I did not see a need
for 8-tye read of user_ip6 and msg_src_ip6 yet. So let
us just deal with write for now.
With the above two nits,
Acked-by: Yonghong Song <yhs@fb.com>
> ---
> include/linux/filter.h | 6 ++++++
> net/core/filter.c | 22 ++++++++++++++--------
> 2 files changed, 20 insertions(+), 8 deletions(-)
>
> diff --git a/include/linux/filter.h b/include/linux/filter.h
> index 340f7d648974..3901007e36f1 100644
> --- a/include/linux/filter.h
> +++ b/include/linux/filter.h
> @@ -746,6 +746,12 @@ bpf_ctx_narrow_access_ok(u32 off, u32 size, u32 size_default)
> return size <= size_default && (size & (size - 1)) == 0;
> }
>
> +#define bpf_ctx_wide_store_ok(off, size, type, field) \
> + (size == sizeof(__u64) && \
> + off >= offsetof(type, field) && \
> + off + sizeof(__u64) <= offsetofend(type, field) && \
> + off % sizeof(__u64) == 0)
> +
> #define bpf_classic_proglen(fprog) (fprog->len * sizeof(fprog->filter[0]))
>
> static inline void bpf_prog_lock_ro(struct bpf_prog *fp)
> diff --git a/net/core/filter.c b/net/core/filter.c
> index dc8534be12fc..5d33f2146dab 100644
> --- a/net/core/filter.c
> +++ b/net/core/filter.c
> @@ -6849,6 +6849,16 @@ static bool sock_addr_is_valid_access(int off, int size,
> if (!bpf_ctx_narrow_access_ok(off, size, size_default))
> return false;
> } else {
> + if (bpf_ctx_wide_store_ok(off, size,
> + struct bpf_sock_addr,
> + user_ip6))
> + return true;
> +
> + if (bpf_ctx_wide_store_ok(off, size,
> + struct bpf_sock_addr,
> + msg_src_ip6))
> + return true;
> +
> if (size != size_default)
> return false;
> }
> @@ -7689,9 +7699,6 @@ static u32 xdp_convert_ctx_access(enum bpf_access_type type,
> /* SOCK_ADDR_STORE_NESTED_FIELD_OFF() has semantic similar to
> * SOCK_ADDR_LOAD_NESTED_FIELD_SIZE_OFF() but for store operation.
> *
> - * It doesn't support SIZE argument though since narrow stores are not
> - * supported for now.
> - *
> * In addition it uses Temporary Field TF (member of struct S) as the 3rd
> * "register" since two registers available in convert_ctx_access are not
> * enough: we can't override neither SRC, since it contains value to store, nor
> @@ -7699,7 +7706,7 @@ static u32 xdp_convert_ctx_access(enum bpf_access_type type,
> * instructions. But we need a temporary place to save pointer to nested
> * structure whose field we want to store to.
> */
> -#define SOCK_ADDR_STORE_NESTED_FIELD_OFF(S, NS, F, NF, OFF, TF) \
> +#define SOCK_ADDR_STORE_NESTED_FIELD_OFF(S, NS, F, NF, SIZE, OFF, TF) \
> do { \
> int tmp_reg = BPF_REG_9; \
> if (si->src_reg == tmp_reg || si->dst_reg == tmp_reg) \
> @@ -7710,8 +7717,7 @@ static u32 xdp_convert_ctx_access(enum bpf_access_type type,
> offsetof(S, TF)); \
> *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(S, F), tmp_reg, \
> si->dst_reg, offsetof(S, F)); \
> - *insn++ = BPF_STX_MEM( \
> - BPF_FIELD_SIZEOF(NS, NF), tmp_reg, si->src_reg, \
> + *insn++ = BPF_STX_MEM(SIZE, tmp_reg, si->src_reg, \
> bpf_target_off(NS, NF, FIELD_SIZEOF(NS, NF), \
> target_size) \
> + OFF); \
> @@ -7723,8 +7729,8 @@ static u32 xdp_convert_ctx_access(enum bpf_access_type type,
> TF) \
> do { \
> if (type == BPF_WRITE) { \
> - SOCK_ADDR_STORE_NESTED_FIELD_OFF(S, NS, F, NF, OFF, \
> - TF); \
> + SOCK_ADDR_STORE_NESTED_FIELD_OFF(S, NS, F, NF, SIZE, \
> + OFF, TF); \
> } else { \
> SOCK_ADDR_LOAD_NESTED_FIELD_SIZE_OFF( \
> S, NS, F, NF, SIZE, OFF); \
>
^ permalink raw reply [flat|nested] 11+ messages in thread
* Re: [PATCH bpf-next 2/2] selftests/bpf: add verifier tests for wide stores
2019-06-28 23:10 ` [PATCH bpf-next 2/2] selftests/bpf: add verifier tests for wide stores Stanislav Fomichev
@ 2019-06-30 6:01 ` Yonghong Song
2019-07-01 15:44 ` Andrii Nakryiko
2019-07-01 16:00 ` Stanislav Fomichev
0 siblings, 2 replies; 11+ messages in thread
From: Yonghong Song @ 2019-06-30 6:01 UTC (permalink / raw)
To: Stanislav Fomichev, netdev, bpf; +Cc: davem, ast, daniel, Andrii Nakryiko
On 6/28/19 4:10 PM, Stanislav Fomichev wrote:
> Make sure that wide stores are allowed at proper (aligned) addresses.
> Note that user_ip6 is naturally aligned on 8-byte boundary, so
> correct addresses are user_ip6[0] and user_ip6[2]. msg_src_ip6 is,
> however, aligned on a 4-byte bondary, so only msg_src_ip6[1]
> can be wide-stored.
>
> Cc: Andrii Nakryiko <andriin@fb.com>
> Cc: Yonghong Song <yhs@fb.com>
> Signed-off-by: Stanislav Fomichev <sdf@google.com>
> ---
> tools/testing/selftests/bpf/test_verifier.c | 17 ++++++--
> .../selftests/bpf/verifier/wide_store.c | 40 +++++++++++++++++++
> 2 files changed, 54 insertions(+), 3 deletions(-)
> create mode 100644 tools/testing/selftests/bpf/verifier/wide_store.c
>
> diff --git a/tools/testing/selftests/bpf/test_verifier.c b/tools/testing/selftests/bpf/test_verifier.c
> index c5514daf8865..b0773291012a 100644
> --- a/tools/testing/selftests/bpf/test_verifier.c
> +++ b/tools/testing/selftests/bpf/test_verifier.c
> @@ -105,6 +105,7 @@ struct bpf_test {
> __u64 data64[TEST_DATA_LEN / 8];
> };
> } retvals[MAX_TEST_RUNS];
> + enum bpf_attach_type expected_attach_type;
> };
>
> /* Note we want this to be 64 bit aligned so that the end of our array is
> @@ -850,6 +851,7 @@ static void do_test_single(struct bpf_test *test, bool unpriv,
> int fd_prog, expected_ret, alignment_prevented_execution;
> int prog_len, prog_type = test->prog_type;
> struct bpf_insn *prog = test->insns;
> + struct bpf_load_program_attr attr;
> int run_errs, run_successes;
> int map_fds[MAX_NR_MAPS];
> const char *expected_err;
> @@ -881,8 +883,17 @@ static void do_test_single(struct bpf_test *test, bool unpriv,
> pflags |= BPF_F_STRICT_ALIGNMENT;
> if (test->flags & F_NEEDS_EFFICIENT_UNALIGNED_ACCESS)
> pflags |= BPF_F_ANY_ALIGNMENT;
> - fd_prog = bpf_verify_program(prog_type, prog, prog_len, pflags,
> - "GPL", 0, bpf_vlog, sizeof(bpf_vlog), 4);
> +
> + memset(&attr, 0, sizeof(attr));
> + attr.prog_type = prog_type;
> + attr.expected_attach_type = test->expected_attach_type;
> + attr.insns = prog;
> + attr.insns_cnt = prog_len;
> + attr.license = "GPL";
> + attr.log_level = 4;
> + attr.prog_flags = pflags;
> +
> + fd_prog = bpf_load_program_xattr(&attr, bpf_vlog, sizeof(bpf_vlog));
> if (fd_prog < 0 && !bpf_probe_prog_type(prog_type, 0)) {
> printf("SKIP (unsupported program type %d)\n", prog_type);
> skips++;
> @@ -912,7 +923,7 @@ static void do_test_single(struct bpf_test *test, bool unpriv,
> printf("FAIL\nUnexpected success to load!\n");
> goto fail_log;
> }
> - if (!strstr(bpf_vlog, expected_err)) {
> + if (!expected_err || !strstr(bpf_vlog, expected_err)) {
> printf("FAIL\nUnexpected error message!\n\tEXP: %s\n\tRES: %s\n",
> expected_err, bpf_vlog);
> goto fail_log;
> diff --git a/tools/testing/selftests/bpf/verifier/wide_store.c b/tools/testing/selftests/bpf/verifier/wide_store.c
> new file mode 100644
> index 000000000000..c6385f45b114
> --- /dev/null
> +++ b/tools/testing/selftests/bpf/verifier/wide_store.c
> @@ -0,0 +1,40 @@
> +#define BPF_SOCK_ADDR(field, off, res, err) \
> +{ \
> + "wide store to bpf_sock_addr." #field "[" #off "]", \
> + .insns = { \
> + BPF_MOV64_IMM(BPF_REG_0, 1), \
> + BPF_STX_MEM(BPF_DW, BPF_REG_1, BPF_REG_0, \
> + offsetof(struct bpf_sock_addr, field[off])), \
> + BPF_EXIT_INSN(), \
> + }, \
> + .result = res, \
> + .prog_type = BPF_PROG_TYPE_CGROUP_SOCK_ADDR, \
> + .expected_attach_type = BPF_CGROUP_UDP6_SENDMSG, \
> + .errstr = err, \
> +}
> +
> +/* user_ip6[0] is u64 aligned */
> +BPF_SOCK_ADDR(user_ip6, 0, ACCEPT,
> + NULL),
> +BPF_SOCK_ADDR(user_ip6, 1, REJECT,
> + "invalid bpf_context access off=12 size=8"),
> +BPF_SOCK_ADDR(user_ip6, 2, ACCEPT,
> + NULL),
> +BPF_SOCK_ADDR(user_ip6, 3, REJECT,
> + "invalid bpf_context access off=20 size=8"),
> +BPF_SOCK_ADDR(user_ip6, 4, REJECT,
> + "invalid bpf_context access off=24 size=8"),
With offset 4, we have
#968/p wide store to bpf_sock_addr.user_ip6[4] OK
This test case can be removed. user code typically
won't write bpf_sock_addr.user_ip6[4], and compiler
typically will give a warning since it is out of
array bound. Any particular reason you want to
include this one?
> +
> +/* msg_src_ip6[0] is _not_ u64 aligned */
> +BPF_SOCK_ADDR(msg_src_ip6, 0, REJECT,
> + "invalid bpf_context access off=44 size=8"),
> +BPF_SOCK_ADDR(msg_src_ip6, 1, ACCEPT,
> + NULL),
> +BPF_SOCK_ADDR(msg_src_ip6, 2, REJECT,
> + "invalid bpf_context access off=52 size=8"),
> +BPF_SOCK_ADDR(msg_src_ip6, 3, REJECT,
> + "invalid bpf_context access off=56 size=8"),
> +BPF_SOCK_ADDR(msg_src_ip6, 4, REJECT,
> + "invalid bpf_context access off=60 size=8"),
The same as above, offset=4 case can be removed?
> +
> +#undef BPF_SOCK_ADDR
>
^ permalink raw reply [flat|nested] 11+ messages in thread
* Re: [PATCH bpf-next 1/2] bpf: allow wide (u64) aligned stores for some fields of bpf_sock_addr
2019-06-30 5:52 ` [PATCH bpf-next 1/2] bpf: allow wide (u64) aligned stores for some fields of bpf_sock_addr Yonghong Song
@ 2019-07-01 15:36 ` Andrii Nakryiko
2019-07-01 16:04 ` Stanislav Fomichev
2019-07-01 16:01 ` Stanislav Fomichev
1 sibling, 1 reply; 11+ messages in thread
From: Andrii Nakryiko @ 2019-07-01 15:36 UTC (permalink / raw)
To: Yonghong Song
Cc: Stanislav Fomichev, netdev, bpf, davem, ast, daniel,
Andrii Nakryiko, kernel test robot
On Sat, Jun 29, 2019 at 10:53 PM Yonghong Song <yhs@fb.com> wrote:
>
>
>
> On 6/28/19 4:10 PM, Stanislav Fomichev wrote:
> > Since commit cd17d7770578 ("bpf/tools: sync bpf.h") clang decided
> > that it can do a single u64 store into user_ip6[2] instead of two
> > separate u32 ones:
> >
> > # 17: (18) r2 = 0x100000000000000
> > # ; ctx->user_ip6[2] = bpf_htonl(DST_REWRITE_IP6_2);
> > # 19: (7b) *(u64 *)(r1 +16) = r2
> > # invalid bpf_context access off=16 size=8
> >
> > From the compiler point of view it does look like a correct thing
> > to do, so let's support it on the kernel side.
> >
> > Credit to Andrii Nakryiko for a proper implementation of
> > bpf_ctx_wide_store_ok.
> >
> > Cc: Andrii Nakryiko <andriin@fb.com>
> > Cc: Yonghong Song <yhs@fb.com>
> > Fixes: cd17d7770578 ("bpf/tools: sync bpf.h")
> > Reported-by: kernel test robot <rong.a.chen@intel.com>
> > Signed-off-by: Stanislav Fomichev <sdf@google.com>
>
> The change looks good to me with the following nits:
> 1. could you add a cover letter for the patch set?
> typically if the number of patches is more than one,
> it would be a good practice with a cover letter.
> See bpf_devel_QA.rst .
> 2. with this change, the comments in uapi bpf.h
> are not accurate any more.
> __u32 user_ip6[4]; /* Allows 1,2,4-byte read an 4-byte write.
> * Stored in network byte order.
>
> */
> __u32 msg_src_ip6[4]; /* Allows 1,2,4-byte read an 4-byte write.
> * Stored in network byte order.
> */
> now for stores, aligned 8-byte write is permitted.
> could you update this as well?
>
> From the typical usage pattern, I did not see a need
> for 8-tye read of user_ip6 and msg_src_ip6 yet. So let
> us just deal with write for now.
But I guess it's still possible for clang to optimize two consecutive
4-byte reads into single 8-byte read in some circumstances? If that's
the case, maybe it's a good idea to have corresponding read checks as
well?
But overall this looks good to me:
Acked-by: Andrii Nakryiko <andriin@fb.com>
>
> With the above two nits,
> Acked-by: Yonghong Song <yhs@fb.com>
>
> > ---
> > include/linux/filter.h | 6 ++++++
> > net/core/filter.c | 22 ++++++++++++++--------
> > 2 files changed, 20 insertions(+), 8 deletions(-)
> >
> > diff --git a/include/linux/filter.h b/include/linux/filter.h
> > index 340f7d648974..3901007e36f1 100644
> > --- a/include/linux/filter.h
> > +++ b/include/linux/filter.h
> > @@ -746,6 +746,12 @@ bpf_ctx_narrow_access_ok(u32 off, u32 size, u32 size_default)
> > return size <= size_default && (size & (size - 1)) == 0;
> > }
> >
> > +#define bpf_ctx_wide_store_ok(off, size, type, field) \
> > + (size == sizeof(__u64) && \
> > + off >= offsetof(type, field) && \
> > + off + sizeof(__u64) <= offsetofend(type, field) && \
> > + off % sizeof(__u64) == 0)
> > +
> > #define bpf_classic_proglen(fprog) (fprog->len * sizeof(fprog->filter[0]))
> >
> > static inline void bpf_prog_lock_ro(struct bpf_prog *fp)
> > diff --git a/net/core/filter.c b/net/core/filter.c
> > index dc8534be12fc..5d33f2146dab 100644
> > --- a/net/core/filter.c
> > +++ b/net/core/filter.c
> > @@ -6849,6 +6849,16 @@ static bool sock_addr_is_valid_access(int off, int size,
> > if (!bpf_ctx_narrow_access_ok(off, size, size_default))
> > return false;
> > } else {
> > + if (bpf_ctx_wide_store_ok(off, size,
> > + struct bpf_sock_addr,
> > + user_ip6))
> > + return true;
> > +
> > + if (bpf_ctx_wide_store_ok(off, size,
> > + struct bpf_sock_addr,
> > + msg_src_ip6))
> > + return true;
> > +
> > if (size != size_default)
> > return false;
> > }
> > @@ -7689,9 +7699,6 @@ static u32 xdp_convert_ctx_access(enum bpf_access_type type,
> > /* SOCK_ADDR_STORE_NESTED_FIELD_OFF() has semantic similar to
> > * SOCK_ADDR_LOAD_NESTED_FIELD_SIZE_OFF() but for store operation.
> > *
> > - * It doesn't support SIZE argument though since narrow stores are not
> > - * supported for now.
> > - *
> > * In addition it uses Temporary Field TF (member of struct S) as the 3rd
> > * "register" since two registers available in convert_ctx_access are not
> > * enough: we can't override neither SRC, since it contains value to store, nor
> > @@ -7699,7 +7706,7 @@ static u32 xdp_convert_ctx_access(enum bpf_access_type type,
> > * instructions. But we need a temporary place to save pointer to nested
> > * structure whose field we want to store to.
> > */
> > -#define SOCK_ADDR_STORE_NESTED_FIELD_OFF(S, NS, F, NF, OFF, TF) \
> > +#define SOCK_ADDR_STORE_NESTED_FIELD_OFF(S, NS, F, NF, SIZE, OFF, TF) \
> > do { \
> > int tmp_reg = BPF_REG_9; \
> > if (si->src_reg == tmp_reg || si->dst_reg == tmp_reg) \
> > @@ -7710,8 +7717,7 @@ static u32 xdp_convert_ctx_access(enum bpf_access_type type,
> > offsetof(S, TF)); \
> > *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(S, F), tmp_reg, \
> > si->dst_reg, offsetof(S, F)); \
> > - *insn++ = BPF_STX_MEM( \
> > - BPF_FIELD_SIZEOF(NS, NF), tmp_reg, si->src_reg, \
> > + *insn++ = BPF_STX_MEM(SIZE, tmp_reg, si->src_reg, \
> > bpf_target_off(NS, NF, FIELD_SIZEOF(NS, NF), \
> > target_size) \
> > + OFF); \
> > @@ -7723,8 +7729,8 @@ static u32 xdp_convert_ctx_access(enum bpf_access_type type,
> > TF) \
> > do { \
> > if (type == BPF_WRITE) { \
> > - SOCK_ADDR_STORE_NESTED_FIELD_OFF(S, NS, F, NF, OFF, \
> > - TF); \
> > + SOCK_ADDR_STORE_NESTED_FIELD_OFF(S, NS, F, NF, SIZE, \
> > + OFF, TF); \
> > } else { \
> > SOCK_ADDR_LOAD_NESTED_FIELD_SIZE_OFF( \
> > S, NS, F, NF, SIZE, OFF); \
> >
^ permalink raw reply [flat|nested] 11+ messages in thread
* Re: [PATCH bpf-next 2/2] selftests/bpf: add verifier tests for wide stores
2019-06-30 6:01 ` Yonghong Song
@ 2019-07-01 15:44 ` Andrii Nakryiko
2019-07-01 16:00 ` Stanislav Fomichev
1 sibling, 0 replies; 11+ messages in thread
From: Andrii Nakryiko @ 2019-07-01 15:44 UTC (permalink / raw)
To: Yonghong Song
Cc: Stanislav Fomichev, netdev, bpf, davem, ast, daniel, Andrii Nakryiko
On Sat, Jun 29, 2019 at 11:02 PM Yonghong Song <yhs@fb.com> wrote:
>
>
>
> On 6/28/19 4:10 PM, Stanislav Fomichev wrote:
> > Make sure that wide stores are allowed at proper (aligned) addresses.
> > Note that user_ip6 is naturally aligned on 8-byte boundary, so
> > correct addresses are user_ip6[0] and user_ip6[2]. msg_src_ip6 is,
> > however, aligned on a 4-byte bondary, so only msg_src_ip6[1]
> > can be wide-stored.
> >
> > Cc: Andrii Nakryiko <andriin@fb.com>
> > Cc: Yonghong Song <yhs@fb.com>
> > Signed-off-by: Stanislav Fomichev <sdf@google.com>
> > ---
> > tools/testing/selftests/bpf/test_verifier.c | 17 ++++++--
> > .../selftests/bpf/verifier/wide_store.c | 40 +++++++++++++++++++
> > 2 files changed, 54 insertions(+), 3 deletions(-)
> > create mode 100644 tools/testing/selftests/bpf/verifier/wide_store.c
> >
> > diff --git a/tools/testing/selftests/bpf/test_verifier.c b/tools/testing/selftests/bpf/test_verifier.c
> > index c5514daf8865..b0773291012a 100644
> > --- a/tools/testing/selftests/bpf/test_verifier.c
> > +++ b/tools/testing/selftests/bpf/test_verifier.c
> > @@ -105,6 +105,7 @@ struct bpf_test {
> > __u64 data64[TEST_DATA_LEN / 8];
> > };
> > } retvals[MAX_TEST_RUNS];
> > + enum bpf_attach_type expected_attach_type;
> > };
> >
> > /* Note we want this to be 64 bit aligned so that the end of our array is
> > @@ -850,6 +851,7 @@ static void do_test_single(struct bpf_test *test, bool unpriv,
> > int fd_prog, expected_ret, alignment_prevented_execution;
> > int prog_len, prog_type = test->prog_type;
> > struct bpf_insn *prog = test->insns;
> > + struct bpf_load_program_attr attr;
> > int run_errs, run_successes;
> > int map_fds[MAX_NR_MAPS];
> > const char *expected_err;
> > @@ -881,8 +883,17 @@ static void do_test_single(struct bpf_test *test, bool unpriv,
> > pflags |= BPF_F_STRICT_ALIGNMENT;
> > if (test->flags & F_NEEDS_EFFICIENT_UNALIGNED_ACCESS)
> > pflags |= BPF_F_ANY_ALIGNMENT;
> > - fd_prog = bpf_verify_program(prog_type, prog, prog_len, pflags,
> > - "GPL", 0, bpf_vlog, sizeof(bpf_vlog), 4);
> > +
> > + memset(&attr, 0, sizeof(attr));
> > + attr.prog_type = prog_type;
> > + attr.expected_attach_type = test->expected_attach_type;
> > + attr.insns = prog;
> > + attr.insns_cnt = prog_len;
> > + attr.license = "GPL";
> > + attr.log_level = 4;
> > + attr.prog_flags = pflags;
> > +
> > + fd_prog = bpf_load_program_xattr(&attr, bpf_vlog, sizeof(bpf_vlog));
> > if (fd_prog < 0 && !bpf_probe_prog_type(prog_type, 0)) {
> > printf("SKIP (unsupported program type %d)\n", prog_type);
> > skips++;
> > @@ -912,7 +923,7 @@ static void do_test_single(struct bpf_test *test, bool unpriv,
> > printf("FAIL\nUnexpected success to load!\n");
> > goto fail_log;
> > }
> > - if (!strstr(bpf_vlog, expected_err)) {
> > + if (!expected_err || !strstr(bpf_vlog, expected_err)) {
> > printf("FAIL\nUnexpected error message!\n\tEXP: %s\n\tRES: %s\n",
> > expected_err, bpf_vlog);
> > goto fail_log;
> > diff --git a/tools/testing/selftests/bpf/verifier/wide_store.c b/tools/testing/selftests/bpf/verifier/wide_store.c
> > new file mode 100644
> > index 000000000000..c6385f45b114
> > --- /dev/null
> > +++ b/tools/testing/selftests/bpf/verifier/wide_store.c
> > @@ -0,0 +1,40 @@
> > +#define BPF_SOCK_ADDR(field, off, res, err) \
> > +{ \
> > + "wide store to bpf_sock_addr." #field "[" #off "]", \
> > + .insns = { \
> > + BPF_MOV64_IMM(BPF_REG_0, 1), \
> > + BPF_STX_MEM(BPF_DW, BPF_REG_1, BPF_REG_0, \
> > + offsetof(struct bpf_sock_addr, field[off])), \
> > + BPF_EXIT_INSN(), \
> > + }, \
> > + .result = res, \
> > + .prog_type = BPF_PROG_TYPE_CGROUP_SOCK_ADDR, \
> > + .expected_attach_type = BPF_CGROUP_UDP6_SENDMSG, \
> > + .errstr = err, \
> > +}
> > +
> > +/* user_ip6[0] is u64 aligned */
> > +BPF_SOCK_ADDR(user_ip6, 0, ACCEPT,
> > + NULL),
> > +BPF_SOCK_ADDR(user_ip6, 1, REJECT,
> > + "invalid bpf_context access off=12 size=8"),
> > +BPF_SOCK_ADDR(user_ip6, 2, ACCEPT,
> > + NULL),
> > +BPF_SOCK_ADDR(user_ip6, 3, REJECT,
> > + "invalid bpf_context access off=20 size=8"),
> > +BPF_SOCK_ADDR(user_ip6, 4, REJECT,
> > + "invalid bpf_context access off=24 size=8"),
>
> With offset 4, we have
> #968/p wide store to bpf_sock_addr.user_ip6[4] OK
>
> This test case can be removed. user code typically
> won't write bpf_sock_addr.user_ip6[4], and compiler
> typically will give a warning since it is out of
> array bound. Any particular reason you want to
> include this one?
I agree, user_ip6[4] is essentially 8-byte write to user_port field.
>
>
> > +
> > +/* msg_src_ip6[0] is _not_ u64 aligned */
> > +BPF_SOCK_ADDR(msg_src_ip6, 0, REJECT,
> > + "invalid bpf_context access off=44 size=8"),
> > +BPF_SOCK_ADDR(msg_src_ip6, 1, ACCEPT,
> > + NULL),
> > +BPF_SOCK_ADDR(msg_src_ip6, 2, REJECT,
> > + "invalid bpf_context access off=52 size=8"),
> > +BPF_SOCK_ADDR(msg_src_ip6, 3, REJECT,
> > + "invalid bpf_context access off=56 size=8"),
> > +BPF_SOCK_ADDR(msg_src_ip6, 4, REJECT,
> > + "invalid bpf_context access off=60 size=8"),
>
> The same as above, offset=4 case can be removed?
And this one is a write into a struct hole, which should be rejected
even without wide-store check, right?
>
> > +
> > +#undef BPF_SOCK_ADDR
> >
^ permalink raw reply [flat|nested] 11+ messages in thread
* Re: [PATCH bpf-next 2/2] selftests/bpf: add verifier tests for wide stores
2019-06-30 6:01 ` Yonghong Song
2019-07-01 15:44 ` Andrii Nakryiko
@ 2019-07-01 16:00 ` Stanislav Fomichev
1 sibling, 0 replies; 11+ messages in thread
From: Stanislav Fomichev @ 2019-07-01 16:00 UTC (permalink / raw)
To: Yonghong Song
Cc: Stanislav Fomichev, netdev, bpf, davem, ast, daniel, Andrii Nakryiko
On 06/30, Yonghong Song wrote:
>
>
> On 6/28/19 4:10 PM, Stanislav Fomichev wrote:
> > Make sure that wide stores are allowed at proper (aligned) addresses.
> > Note that user_ip6 is naturally aligned on 8-byte boundary, so
> > correct addresses are user_ip6[0] and user_ip6[2]. msg_src_ip6 is,
> > however, aligned on a 4-byte bondary, so only msg_src_ip6[1]
> > can be wide-stored.
> >
> > Cc: Andrii Nakryiko <andriin@fb.com>
> > Cc: Yonghong Song <yhs@fb.com>
> > Signed-off-by: Stanislav Fomichev <sdf@google.com>
> > ---
> > tools/testing/selftests/bpf/test_verifier.c | 17 ++++++--
> > .../selftests/bpf/verifier/wide_store.c | 40 +++++++++++++++++++
> > 2 files changed, 54 insertions(+), 3 deletions(-)
> > create mode 100644 tools/testing/selftests/bpf/verifier/wide_store.c
> >
> > diff --git a/tools/testing/selftests/bpf/test_verifier.c b/tools/testing/selftests/bpf/test_verifier.c
> > index c5514daf8865..b0773291012a 100644
> > --- a/tools/testing/selftests/bpf/test_verifier.c
> > +++ b/tools/testing/selftests/bpf/test_verifier.c
> > @@ -105,6 +105,7 @@ struct bpf_test {
> > __u64 data64[TEST_DATA_LEN / 8];
> > };
> > } retvals[MAX_TEST_RUNS];
> > + enum bpf_attach_type expected_attach_type;
> > };
> >
> > /* Note we want this to be 64 bit aligned so that the end of our array is
> > @@ -850,6 +851,7 @@ static void do_test_single(struct bpf_test *test, bool unpriv,
> > int fd_prog, expected_ret, alignment_prevented_execution;
> > int prog_len, prog_type = test->prog_type;
> > struct bpf_insn *prog = test->insns;
> > + struct bpf_load_program_attr attr;
> > int run_errs, run_successes;
> > int map_fds[MAX_NR_MAPS];
> > const char *expected_err;
> > @@ -881,8 +883,17 @@ static void do_test_single(struct bpf_test *test, bool unpriv,
> > pflags |= BPF_F_STRICT_ALIGNMENT;
> > if (test->flags & F_NEEDS_EFFICIENT_UNALIGNED_ACCESS)
> > pflags |= BPF_F_ANY_ALIGNMENT;
> > - fd_prog = bpf_verify_program(prog_type, prog, prog_len, pflags,
> > - "GPL", 0, bpf_vlog, sizeof(bpf_vlog), 4);
> > +
> > + memset(&attr, 0, sizeof(attr));
> > + attr.prog_type = prog_type;
> > + attr.expected_attach_type = test->expected_attach_type;
> > + attr.insns = prog;
> > + attr.insns_cnt = prog_len;
> > + attr.license = "GPL";
> > + attr.log_level = 4;
> > + attr.prog_flags = pflags;
> > +
> > + fd_prog = bpf_load_program_xattr(&attr, bpf_vlog, sizeof(bpf_vlog));
> > if (fd_prog < 0 && !bpf_probe_prog_type(prog_type, 0)) {
> > printf("SKIP (unsupported program type %d)\n", prog_type);
> > skips++;
> > @@ -912,7 +923,7 @@ static void do_test_single(struct bpf_test *test, bool unpriv,
> > printf("FAIL\nUnexpected success to load!\n");
> > goto fail_log;
> > }
> > - if (!strstr(bpf_vlog, expected_err)) {
> > + if (!expected_err || !strstr(bpf_vlog, expected_err)) {
> > printf("FAIL\nUnexpected error message!\n\tEXP: %s\n\tRES: %s\n",
> > expected_err, bpf_vlog);
> > goto fail_log;
> > diff --git a/tools/testing/selftests/bpf/verifier/wide_store.c b/tools/testing/selftests/bpf/verifier/wide_store.c
> > new file mode 100644
> > index 000000000000..c6385f45b114
> > --- /dev/null
> > +++ b/tools/testing/selftests/bpf/verifier/wide_store.c
> > @@ -0,0 +1,40 @@
> > +#define BPF_SOCK_ADDR(field, off, res, err) \
> > +{ \
> > + "wide store to bpf_sock_addr." #field "[" #off "]", \
> > + .insns = { \
> > + BPF_MOV64_IMM(BPF_REG_0, 1), \
> > + BPF_STX_MEM(BPF_DW, BPF_REG_1, BPF_REG_0, \
> > + offsetof(struct bpf_sock_addr, field[off])), \
> > + BPF_EXIT_INSN(), \
> > + }, \
> > + .result = res, \
> > + .prog_type = BPF_PROG_TYPE_CGROUP_SOCK_ADDR, \
> > + .expected_attach_type = BPF_CGROUP_UDP6_SENDMSG, \
> > + .errstr = err, \
> > +}
> > +
> > +/* user_ip6[0] is u64 aligned */
> > +BPF_SOCK_ADDR(user_ip6, 0, ACCEPT,
> > + NULL),
> > +BPF_SOCK_ADDR(user_ip6, 1, REJECT,
> > + "invalid bpf_context access off=12 size=8"),
> > +BPF_SOCK_ADDR(user_ip6, 2, ACCEPT,
> > + NULL),
> > +BPF_SOCK_ADDR(user_ip6, 3, REJECT,
> > + "invalid bpf_context access off=20 size=8"),
> > +BPF_SOCK_ADDR(user_ip6, 4, REJECT,
> > + "invalid bpf_context access off=24 size=8"),
>
> With offset 4, we have
> #968/p wide store to bpf_sock_addr.user_ip6[4] OK
>
> This test case can be removed. user code typically
> won't write bpf_sock_addr.user_ip6[4], and compiler
> typically will give a warning since it is out of
> array bound. Any particular reason you want to
> include this one?
Agreed on both, I'm being overly cautious here. They should
be caught by the outer switch and be rejected because of
other reasons.
> > +
> > +/* msg_src_ip6[0] is _not_ u64 aligned */
> > +BPF_SOCK_ADDR(msg_src_ip6, 0, REJECT,
> > + "invalid bpf_context access off=44 size=8"),
> > +BPF_SOCK_ADDR(msg_src_ip6, 1, ACCEPT,
> > + NULL),
> > +BPF_SOCK_ADDR(msg_src_ip6, 2, REJECT,
> > + "invalid bpf_context access off=52 size=8"),
> > +BPF_SOCK_ADDR(msg_src_ip6, 3, REJECT,
> > + "invalid bpf_context access off=56 size=8"),
> > +BPF_SOCK_ADDR(msg_src_ip6, 4, REJECT,
> > + "invalid bpf_context access off=60 size=8"),
>
> The same as above, offset=4 case can be removed?
>
> > +
> > +#undef BPF_SOCK_ADDR
> >
^ permalink raw reply [flat|nested] 11+ messages in thread
* Re: [PATCH bpf-next 1/2] bpf: allow wide (u64) aligned stores for some fields of bpf_sock_addr
2019-06-30 5:52 ` [PATCH bpf-next 1/2] bpf: allow wide (u64) aligned stores for some fields of bpf_sock_addr Yonghong Song
2019-07-01 15:36 ` Andrii Nakryiko
@ 2019-07-01 16:01 ` Stanislav Fomichev
1 sibling, 0 replies; 11+ messages in thread
From: Stanislav Fomichev @ 2019-07-01 16:01 UTC (permalink / raw)
To: Yonghong Song
Cc: Stanislav Fomichev, netdev, bpf, davem, ast, daniel,
Andrii Nakryiko, kernel test robot
On 06/30, Yonghong Song wrote:
>
>
> On 6/28/19 4:10 PM, Stanislav Fomichev wrote:
> > Since commit cd17d7770578 ("bpf/tools: sync bpf.h") clang decided
> > that it can do a single u64 store into user_ip6[2] instead of two
> > separate u32 ones:
> >
> > # 17: (18) r2 = 0x100000000000000
> > # ; ctx->user_ip6[2] = bpf_htonl(DST_REWRITE_IP6_2);
> > # 19: (7b) *(u64 *)(r1 +16) = r2
> > # invalid bpf_context access off=16 size=8
> >
> > From the compiler point of view it does look like a correct thing
> > to do, so let's support it on the kernel side.
> >
> > Credit to Andrii Nakryiko for a proper implementation of
> > bpf_ctx_wide_store_ok.
> >
> > Cc: Andrii Nakryiko <andriin@fb.com>
> > Cc: Yonghong Song <yhs@fb.com>
> > Fixes: cd17d7770578 ("bpf/tools: sync bpf.h")
> > Reported-by: kernel test robot <rong.a.chen@intel.com>
> > Signed-off-by: Stanislav Fomichev <sdf@google.com>
>
> The change looks good to me with the following nits:
> 1. could you add a cover letter for the patch set?
> typically if the number of patches is more than one,
> it would be a good practice with a cover letter.
> See bpf_devel_QA.rst .
> 2. with this change, the comments in uapi bpf.h
> are not accurate any more.
> __u32 user_ip6[4]; /* Allows 1,2,4-byte read an 4-byte write.
> * Stored in network byte order.
>
> */
> __u32 msg_src_ip6[4]; /* Allows 1,2,4-byte read an 4-byte write.
> * Stored in network byte order.
> */
> now for stores, aligned 8-byte write is permitted.
> could you update this as well?
>
> From the typical usage pattern, I did not see a need
> for 8-tye read of user_ip6 and msg_src_ip6 yet. So let
> us just deal with write for now.
>
> With the above two nits,
> Acked-by: Yonghong Song <yhs@fb.com>
Thank you for a review, will follow up with a v2 shortly with both
things addressed!
> > ---
> > include/linux/filter.h | 6 ++++++
> > net/core/filter.c | 22 ++++++++++++++--------
> > 2 files changed, 20 insertions(+), 8 deletions(-)
> >
> > diff --git a/include/linux/filter.h b/include/linux/filter.h
> > index 340f7d648974..3901007e36f1 100644
> > --- a/include/linux/filter.h
> > +++ b/include/linux/filter.h
> > @@ -746,6 +746,12 @@ bpf_ctx_narrow_access_ok(u32 off, u32 size, u32 size_default)
> > return size <= size_default && (size & (size - 1)) == 0;
> > }
> >
> > +#define bpf_ctx_wide_store_ok(off, size, type, field) \
> > + (size == sizeof(__u64) && \
> > + off >= offsetof(type, field) && \
> > + off + sizeof(__u64) <= offsetofend(type, field) && \
> > + off % sizeof(__u64) == 0)
> > +
> > #define bpf_classic_proglen(fprog) (fprog->len * sizeof(fprog->filter[0]))
> >
> > static inline void bpf_prog_lock_ro(struct bpf_prog *fp)
> > diff --git a/net/core/filter.c b/net/core/filter.c
> > index dc8534be12fc..5d33f2146dab 100644
> > --- a/net/core/filter.c
> > +++ b/net/core/filter.c
> > @@ -6849,6 +6849,16 @@ static bool sock_addr_is_valid_access(int off, int size,
> > if (!bpf_ctx_narrow_access_ok(off, size, size_default))
> > return false;
> > } else {
> > + if (bpf_ctx_wide_store_ok(off, size,
> > + struct bpf_sock_addr,
> > + user_ip6))
> > + return true;
> > +
> > + if (bpf_ctx_wide_store_ok(off, size,
> > + struct bpf_sock_addr,
> > + msg_src_ip6))
> > + return true;
> > +
> > if (size != size_default)
> > return false;
> > }
> > @@ -7689,9 +7699,6 @@ static u32 xdp_convert_ctx_access(enum bpf_access_type type,
> > /* SOCK_ADDR_STORE_NESTED_FIELD_OFF() has semantic similar to
> > * SOCK_ADDR_LOAD_NESTED_FIELD_SIZE_OFF() but for store operation.
> > *
> > - * It doesn't support SIZE argument though since narrow stores are not
> > - * supported for now.
> > - *
> > * In addition it uses Temporary Field TF (member of struct S) as the 3rd
> > * "register" since two registers available in convert_ctx_access are not
> > * enough: we can't override neither SRC, since it contains value to store, nor
> > @@ -7699,7 +7706,7 @@ static u32 xdp_convert_ctx_access(enum bpf_access_type type,
> > * instructions. But we need a temporary place to save pointer to nested
> > * structure whose field we want to store to.
> > */
> > -#define SOCK_ADDR_STORE_NESTED_FIELD_OFF(S, NS, F, NF, OFF, TF) \
> > +#define SOCK_ADDR_STORE_NESTED_FIELD_OFF(S, NS, F, NF, SIZE, OFF, TF) \
> > do { \
> > int tmp_reg = BPF_REG_9; \
> > if (si->src_reg == tmp_reg || si->dst_reg == tmp_reg) \
> > @@ -7710,8 +7717,7 @@ static u32 xdp_convert_ctx_access(enum bpf_access_type type,
> > offsetof(S, TF)); \
> > *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(S, F), tmp_reg, \
> > si->dst_reg, offsetof(S, F)); \
> > - *insn++ = BPF_STX_MEM( \
> > - BPF_FIELD_SIZEOF(NS, NF), tmp_reg, si->src_reg, \
> > + *insn++ = BPF_STX_MEM(SIZE, tmp_reg, si->src_reg, \
> > bpf_target_off(NS, NF, FIELD_SIZEOF(NS, NF), \
> > target_size) \
> > + OFF); \
> > @@ -7723,8 +7729,8 @@ static u32 xdp_convert_ctx_access(enum bpf_access_type type,
> > TF) \
> > do { \
> > if (type == BPF_WRITE) { \
> > - SOCK_ADDR_STORE_NESTED_FIELD_OFF(S, NS, F, NF, OFF, \
> > - TF); \
> > + SOCK_ADDR_STORE_NESTED_FIELD_OFF(S, NS, F, NF, SIZE, \
> > + OFF, TF); \
> > } else { \
> > SOCK_ADDR_LOAD_NESTED_FIELD_SIZE_OFF( \
> > S, NS, F, NF, SIZE, OFF); \
> >
^ permalink raw reply [flat|nested] 11+ messages in thread
* Re: [PATCH bpf-next 1/2] bpf: allow wide (u64) aligned stores for some fields of bpf_sock_addr
2019-07-01 15:36 ` Andrii Nakryiko
@ 2019-07-01 16:04 ` Stanislav Fomichev
2019-07-01 17:40 ` Yonghong Song
0 siblings, 1 reply; 11+ messages in thread
From: Stanislav Fomichev @ 2019-07-01 16:04 UTC (permalink / raw)
To: Andrii Nakryiko
Cc: Yonghong Song, Stanislav Fomichev, netdev, bpf, davem, ast,
daniel, Andrii Nakryiko, kernel test robot
On 07/01, Andrii Nakryiko wrote:
> On Sat, Jun 29, 2019 at 10:53 PM Yonghong Song <yhs@fb.com> wrote:
> >
> >
> >
> > On 6/28/19 4:10 PM, Stanislav Fomichev wrote:
> > > Since commit cd17d7770578 ("bpf/tools: sync bpf.h") clang decided
> > > that it can do a single u64 store into user_ip6[2] instead of two
> > > separate u32 ones:
> > >
> > > # 17: (18) r2 = 0x100000000000000
> > > # ; ctx->user_ip6[2] = bpf_htonl(DST_REWRITE_IP6_2);
> > > # 19: (7b) *(u64 *)(r1 +16) = r2
> > > # invalid bpf_context access off=16 size=8
> > >
> > > From the compiler point of view it does look like a correct thing
> > > to do, so let's support it on the kernel side.
> > >
> > > Credit to Andrii Nakryiko for a proper implementation of
> > > bpf_ctx_wide_store_ok.
> > >
> > > Cc: Andrii Nakryiko <andriin@fb.com>
> > > Cc: Yonghong Song <yhs@fb.com>
> > > Fixes: cd17d7770578 ("bpf/tools: sync bpf.h")
> > > Reported-by: kernel test robot <rong.a.chen@intel.com>
> > > Signed-off-by: Stanislav Fomichev <sdf@google.com>
> >
> > The change looks good to me with the following nits:
> > 1. could you add a cover letter for the patch set?
> > typically if the number of patches is more than one,
> > it would be a good practice with a cover letter.
> > See bpf_devel_QA.rst .
> > 2. with this change, the comments in uapi bpf.h
> > are not accurate any more.
> > __u32 user_ip6[4]; /* Allows 1,2,4-byte read an 4-byte write.
> > * Stored in network byte order.
> >
> > */
> > __u32 msg_src_ip6[4]; /* Allows 1,2,4-byte read an 4-byte write.
> > * Stored in network byte order.
> > */
> > now for stores, aligned 8-byte write is permitted.
> > could you update this as well?
> >
> > From the typical usage pattern, I did not see a need
> > for 8-tye read of user_ip6 and msg_src_ip6 yet. So let
> > us just deal with write for now.
>
> But I guess it's still possible for clang to optimize two consecutive
> 4-byte reads into single 8-byte read in some circumstances? If that's
> the case, maybe it's a good idea to have corresponding read checks as
> well?
I guess clang can do those kinds of optimizations. I can put it on my
todo and address later (or when we actually see it out in the wild).
> But overall this looks good to me:
>
> Acked-by: Andrii Nakryiko <andriin@fb.com>
Thanks for a review!
> >
> > With the above two nits,
> > Acked-by: Yonghong Song <yhs@fb.com>
> >
> > > ---
> > > include/linux/filter.h | 6 ++++++
> > > net/core/filter.c | 22 ++++++++++++++--------
> > > 2 files changed, 20 insertions(+), 8 deletions(-)
> > >
> > > diff --git a/include/linux/filter.h b/include/linux/filter.h
> > > index 340f7d648974..3901007e36f1 100644
> > > --- a/include/linux/filter.h
> > > +++ b/include/linux/filter.h
> > > @@ -746,6 +746,12 @@ bpf_ctx_narrow_access_ok(u32 off, u32 size, u32 size_default)
> > > return size <= size_default && (size & (size - 1)) == 0;
> > > }
> > >
> > > +#define bpf_ctx_wide_store_ok(off, size, type, field) \
> > > + (size == sizeof(__u64) && \
> > > + off >= offsetof(type, field) && \
> > > + off + sizeof(__u64) <= offsetofend(type, field) && \
> > > + off % sizeof(__u64) == 0)
> > > +
> > > #define bpf_classic_proglen(fprog) (fprog->len * sizeof(fprog->filter[0]))
> > >
> > > static inline void bpf_prog_lock_ro(struct bpf_prog *fp)
> > > diff --git a/net/core/filter.c b/net/core/filter.c
> > > index dc8534be12fc..5d33f2146dab 100644
> > > --- a/net/core/filter.c
> > > +++ b/net/core/filter.c
> > > @@ -6849,6 +6849,16 @@ static bool sock_addr_is_valid_access(int off, int size,
> > > if (!bpf_ctx_narrow_access_ok(off, size, size_default))
> > > return false;
> > > } else {
> > > + if (bpf_ctx_wide_store_ok(off, size,
> > > + struct bpf_sock_addr,
> > > + user_ip6))
> > > + return true;
> > > +
> > > + if (bpf_ctx_wide_store_ok(off, size,
> > > + struct bpf_sock_addr,
> > > + msg_src_ip6))
> > > + return true;
> > > +
> > > if (size != size_default)
> > > return false;
> > > }
> > > @@ -7689,9 +7699,6 @@ static u32 xdp_convert_ctx_access(enum bpf_access_type type,
> > > /* SOCK_ADDR_STORE_NESTED_FIELD_OFF() has semantic similar to
> > > * SOCK_ADDR_LOAD_NESTED_FIELD_SIZE_OFF() but for store operation.
> > > *
> > > - * It doesn't support SIZE argument though since narrow stores are not
> > > - * supported for now.
> > > - *
> > > * In addition it uses Temporary Field TF (member of struct S) as the 3rd
> > > * "register" since two registers available in convert_ctx_access are not
> > > * enough: we can't override neither SRC, since it contains value to store, nor
> > > @@ -7699,7 +7706,7 @@ static u32 xdp_convert_ctx_access(enum bpf_access_type type,
> > > * instructions. But we need a temporary place to save pointer to nested
> > > * structure whose field we want to store to.
> > > */
> > > -#define SOCK_ADDR_STORE_NESTED_FIELD_OFF(S, NS, F, NF, OFF, TF) \
> > > +#define SOCK_ADDR_STORE_NESTED_FIELD_OFF(S, NS, F, NF, SIZE, OFF, TF) \
> > > do { \
> > > int tmp_reg = BPF_REG_9; \
> > > if (si->src_reg == tmp_reg || si->dst_reg == tmp_reg) \
> > > @@ -7710,8 +7717,7 @@ static u32 xdp_convert_ctx_access(enum bpf_access_type type,
> > > offsetof(S, TF)); \
> > > *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(S, F), tmp_reg, \
> > > si->dst_reg, offsetof(S, F)); \
> > > - *insn++ = BPF_STX_MEM( \
> > > - BPF_FIELD_SIZEOF(NS, NF), tmp_reg, si->src_reg, \
> > > + *insn++ = BPF_STX_MEM(SIZE, tmp_reg, si->src_reg, \
> > > bpf_target_off(NS, NF, FIELD_SIZEOF(NS, NF), \
> > > target_size) \
> > > + OFF); \
> > > @@ -7723,8 +7729,8 @@ static u32 xdp_convert_ctx_access(enum bpf_access_type type,
> > > TF) \
> > > do { \
> > > if (type == BPF_WRITE) { \
> > > - SOCK_ADDR_STORE_NESTED_FIELD_OFF(S, NS, F, NF, OFF, \
> > > - TF); \
> > > + SOCK_ADDR_STORE_NESTED_FIELD_OFF(S, NS, F, NF, SIZE, \
> > > + OFF, TF); \
> > > } else { \
> > > SOCK_ADDR_LOAD_NESTED_FIELD_SIZE_OFF( \
> > > S, NS, F, NF, SIZE, OFF); \
> > >
^ permalink raw reply [flat|nested] 11+ messages in thread
* Re: [PATCH bpf-next 1/2] bpf: allow wide (u64) aligned stores for some fields of bpf_sock_addr
2019-07-01 16:04 ` Stanislav Fomichev
@ 2019-07-01 17:40 ` Yonghong Song
2019-07-01 18:38 ` Stanislav Fomichev
0 siblings, 1 reply; 11+ messages in thread
From: Yonghong Song @ 2019-07-01 17:40 UTC (permalink / raw)
To: Stanislav Fomichev, Andrii Nakryiko
Cc: Stanislav Fomichev, netdev, bpf, davem, ast, daniel,
Andrii Nakryiko, kernel test robot
On 7/1/19 9:04 AM, Stanislav Fomichev wrote:
> On 07/01, Andrii Nakryiko wrote:
>> On Sat, Jun 29, 2019 at 10:53 PM Yonghong Song <yhs@fb.com> wrote:
>>>
>>>
>>>
>>> On 6/28/19 4:10 PM, Stanislav Fomichev wrote:
>>>> Since commit cd17d7770578 ("bpf/tools: sync bpf.h") clang decided
>>>> that it can do a single u64 store into user_ip6[2] instead of two
>>>> separate u32 ones:
>>>>
>>>> # 17: (18) r2 = 0x100000000000000
>>>> # ; ctx->user_ip6[2] = bpf_htonl(DST_REWRITE_IP6_2);
>>>> # 19: (7b) *(u64 *)(r1 +16) = r2
>>>> # invalid bpf_context access off=16 size=8
>>>>
>>>> From the compiler point of view it does look like a correct thing
>>>> to do, so let's support it on the kernel side.
>>>>
>>>> Credit to Andrii Nakryiko for a proper implementation of
>>>> bpf_ctx_wide_store_ok.
>>>>
>>>> Cc: Andrii Nakryiko <andriin@fb.com>
>>>> Cc: Yonghong Song <yhs@fb.com>
>>>> Fixes: cd17d7770578 ("bpf/tools: sync bpf.h")
>>>> Reported-by: kernel test robot <rong.a.chen@intel.com>
>>>> Signed-off-by: Stanislav Fomichev <sdf@google.com>
>>>
>>> The change looks good to me with the following nits:
>>> 1. could you add a cover letter for the patch set?
>>> typically if the number of patches is more than one,
>>> it would be a good practice with a cover letter.
>>> See bpf_devel_QA.rst .
>>> 2. with this change, the comments in uapi bpf.h
>>> are not accurate any more.
>>> __u32 user_ip6[4]; /* Allows 1,2,4-byte read an 4-byte write.
>>> * Stored in network byte order.
>>>
>>> */
>>> __u32 msg_src_ip6[4]; /* Allows 1,2,4-byte read an 4-byte write.
>>> * Stored in network byte order.
>>> */
>>> now for stores, aligned 8-byte write is permitted.
>>> could you update this as well?
>>>
>>> From the typical usage pattern, I did not see a need
>>> for 8-tye read of user_ip6 and msg_src_ip6 yet. So let
>>> us just deal with write for now.
>>
>> But I guess it's still possible for clang to optimize two consecutive
>> 4-byte reads into single 8-byte read in some circumstances? If that's
>> the case, maybe it's a good idea to have corresponding read checks as
>> well?
> I guess clang can do those kinds of optimizations. I can put it on my
> todo and address later (or when we actually see it out in the wild).
Okay, I find a Facebook internal app. does trying to read the 4 bytes
and compare to a predefined loopback address. We may need to handle
read cases as well. But this can be a followup after actual tryout.
>
>> But overall this looks good to me:
>>
>> Acked-by: Andrii Nakryiko <andriin@fb.com>
> Thanks for a review!
>
>>>
>>> With the above two nits,
>>> Acked-by: Yonghong Song <yhs@fb.com>
>>>
>>>> ---
>>>> include/linux/filter.h | 6 ++++++
>>>> net/core/filter.c | 22 ++++++++++++++--------
>>>> 2 files changed, 20 insertions(+), 8 deletions(-)
>>>>
>>>> diff --git a/include/linux/filter.h b/include/linux/filter.h
>>>> index 340f7d648974..3901007e36f1 100644
>>>> --- a/include/linux/filter.h
>>>> +++ b/include/linux/filter.h
>>>> @@ -746,6 +746,12 @@ bpf_ctx_narrow_access_ok(u32 off, u32 size, u32 size_default)
>>>> return size <= size_default && (size & (size - 1)) == 0;
>>>> }
>>>>
>>>> +#define bpf_ctx_wide_store_ok(off, size, type, field) \
>>>> + (size == sizeof(__u64) && \
>>>> + off >= offsetof(type, field) && \
>>>> + off + sizeof(__u64) <= offsetofend(type, field) && \
>>>> + off % sizeof(__u64) == 0)
>>>> +
>>>> #define bpf_classic_proglen(fprog) (fprog->len * sizeof(fprog->filter[0]))
>>>>
>>>> static inline void bpf_prog_lock_ro(struct bpf_prog *fp)
>>>> diff --git a/net/core/filter.c b/net/core/filter.c
>>>> index dc8534be12fc..5d33f2146dab 100644
>>>> --- a/net/core/filter.c
>>>> +++ b/net/core/filter.c
>>>> @@ -6849,6 +6849,16 @@ static bool sock_addr_is_valid_access(int off, int size,
>>>> if (!bpf_ctx_narrow_access_ok(off, size, size_default))
>>>> return false;
>>>> } else {
>>>> + if (bpf_ctx_wide_store_ok(off, size,
>>>> + struct bpf_sock_addr,
>>>> + user_ip6))
>>>> + return true;
>>>> +
>>>> + if (bpf_ctx_wide_store_ok(off, size,
>>>> + struct bpf_sock_addr,
>>>> + msg_src_ip6))
>>>> + return true;
>>>> +
>>>> if (size != size_default)
>>>> return false;
>>>> }
>>>> @@ -7689,9 +7699,6 @@ static u32 xdp_convert_ctx_access(enum bpf_access_type type,
>>>> /* SOCK_ADDR_STORE_NESTED_FIELD_OFF() has semantic similar to
>>>> * SOCK_ADDR_LOAD_NESTED_FIELD_SIZE_OFF() but for store operation.
>>>> *
>>>> - * It doesn't support SIZE argument though since narrow stores are not
>>>> - * supported for now.
>>>> - *
>>>> * In addition it uses Temporary Field TF (member of struct S) as the 3rd
>>>> * "register" since two registers available in convert_ctx_access are not
>>>> * enough: we can't override neither SRC, since it contains value to store, nor
>>>> @@ -7699,7 +7706,7 @@ static u32 xdp_convert_ctx_access(enum bpf_access_type type,
>>>> * instructions. But we need a temporary place to save pointer to nested
>>>> * structure whose field we want to store to.
>>>> */
>>>> -#define SOCK_ADDR_STORE_NESTED_FIELD_OFF(S, NS, F, NF, OFF, TF) \
>>>> +#define SOCK_ADDR_STORE_NESTED_FIELD_OFF(S, NS, F, NF, SIZE, OFF, TF) \
>>>> do { \
>>>> int tmp_reg = BPF_REG_9; \
>>>> if (si->src_reg == tmp_reg || si->dst_reg == tmp_reg) \
>>>> @@ -7710,8 +7717,7 @@ static u32 xdp_convert_ctx_access(enum bpf_access_type type,
>>>> offsetof(S, TF)); \
>>>> *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(S, F), tmp_reg, \
>>>> si->dst_reg, offsetof(S, F)); \
>>>> - *insn++ = BPF_STX_MEM( \
>>>> - BPF_FIELD_SIZEOF(NS, NF), tmp_reg, si->src_reg, \
>>>> + *insn++ = BPF_STX_MEM(SIZE, tmp_reg, si->src_reg, \
>>>> bpf_target_off(NS, NF, FIELD_SIZEOF(NS, NF), \
>>>> target_size) \
>>>> + OFF); \
>>>> @@ -7723,8 +7729,8 @@ static u32 xdp_convert_ctx_access(enum bpf_access_type type,
>>>> TF) \
>>>> do { \
>>>> if (type == BPF_WRITE) { \
>>>> - SOCK_ADDR_STORE_NESTED_FIELD_OFF(S, NS, F, NF, OFF, \
>>>> - TF); \
>>>> + SOCK_ADDR_STORE_NESTED_FIELD_OFF(S, NS, F, NF, SIZE, \
>>>> + OFF, TF); \
>>>> } else { \
>>>> SOCK_ADDR_LOAD_NESTED_FIELD_SIZE_OFF( \
>>>> S, NS, F, NF, SIZE, OFF); \
>>>>
^ permalink raw reply [flat|nested] 11+ messages in thread
* Re: [PATCH bpf-next 1/2] bpf: allow wide (u64) aligned stores for some fields of bpf_sock_addr
2019-07-01 17:40 ` Yonghong Song
@ 2019-07-01 18:38 ` Stanislav Fomichev
0 siblings, 0 replies; 11+ messages in thread
From: Stanislav Fomichev @ 2019-07-01 18:38 UTC (permalink / raw)
To: Yonghong Song
Cc: Andrii Nakryiko, Stanislav Fomichev, netdev, bpf, davem, ast,
daniel, Andrii Nakryiko, kernel test robot
On 07/01, Yonghong Song wrote:
>
>
> On 7/1/19 9:04 AM, Stanislav Fomichev wrote:
> > On 07/01, Andrii Nakryiko wrote:
> >> On Sat, Jun 29, 2019 at 10:53 PM Yonghong Song <yhs@fb.com> wrote:
> >>>
> >>>
> >>>
> >>> On 6/28/19 4:10 PM, Stanislav Fomichev wrote:
> >>>> Since commit cd17d7770578 ("bpf/tools: sync bpf.h") clang decided
> >>>> that it can do a single u64 store into user_ip6[2] instead of two
> >>>> separate u32 ones:
> >>>>
> >>>> # 17: (18) r2 = 0x100000000000000
> >>>> # ; ctx->user_ip6[2] = bpf_htonl(DST_REWRITE_IP6_2);
> >>>> # 19: (7b) *(u64 *)(r1 +16) = r2
> >>>> # invalid bpf_context access off=16 size=8
> >>>>
> >>>> From the compiler point of view it does look like a correct thing
> >>>> to do, so let's support it on the kernel side.
> >>>>
> >>>> Credit to Andrii Nakryiko for a proper implementation of
> >>>> bpf_ctx_wide_store_ok.
> >>>>
> >>>> Cc: Andrii Nakryiko <andriin@fb.com>
> >>>> Cc: Yonghong Song <yhs@fb.com>
> >>>> Fixes: cd17d7770578 ("bpf/tools: sync bpf.h")
> >>>> Reported-by: kernel test robot <rong.a.chen@intel.com>
> >>>> Signed-off-by: Stanislav Fomichev <sdf@google.com>
> >>>
> >>> The change looks good to me with the following nits:
> >>> 1. could you add a cover letter for the patch set?
> >>> typically if the number of patches is more than one,
> >>> it would be a good practice with a cover letter.
> >>> See bpf_devel_QA.rst .
> >>> 2. with this change, the comments in uapi bpf.h
> >>> are not accurate any more.
> >>> __u32 user_ip6[4]; /* Allows 1,2,4-byte read an 4-byte write.
> >>> * Stored in network byte order.
> >>>
> >>> */
> >>> __u32 msg_src_ip6[4]; /* Allows 1,2,4-byte read an 4-byte write.
> >>> * Stored in network byte order.
> >>> */
> >>> now for stores, aligned 8-byte write is permitted.
> >>> could you update this as well?
> >>>
> >>> From the typical usage pattern, I did not see a need
> >>> for 8-tye read of user_ip6 and msg_src_ip6 yet. So let
> >>> us just deal with write for now.
> >>
> >> But I guess it's still possible for clang to optimize two consecutive
> >> 4-byte reads into single 8-byte read in some circumstances? If that's
> >> the case, maybe it's a good idea to have corresponding read checks as
> >> well?
> > I guess clang can do those kinds of optimizations. I can put it on my
> > todo and address later (or when we actually see it out in the wild).
>
> Okay, I find a Facebook internal app. does trying to read the 4 bytes
> and compare to a predefined loopback address. We may need to handle
> read cases as well. But this can be a followup after actual tryout.
Sounds good, will follow up on that.
> >
> >> But overall this looks good to me:
> >>
> >> Acked-by: Andrii Nakryiko <andriin@fb.com>
> > Thanks for a review!
> >
> >>>
> >>> With the above two nits,
> >>> Acked-by: Yonghong Song <yhs@fb.com>
> >>>
> >>>> ---
> >>>> include/linux/filter.h | 6 ++++++
> >>>> net/core/filter.c | 22 ++++++++++++++--------
> >>>> 2 files changed, 20 insertions(+), 8 deletions(-)
> >>>>
> >>>> diff --git a/include/linux/filter.h b/include/linux/filter.h
> >>>> index 340f7d648974..3901007e36f1 100644
> >>>> --- a/include/linux/filter.h
> >>>> +++ b/include/linux/filter.h
> >>>> @@ -746,6 +746,12 @@ bpf_ctx_narrow_access_ok(u32 off, u32 size, u32 size_default)
> >>>> return size <= size_default && (size & (size - 1)) == 0;
> >>>> }
> >>>>
> >>>> +#define bpf_ctx_wide_store_ok(off, size, type, field) \
> >>>> + (size == sizeof(__u64) && \
> >>>> + off >= offsetof(type, field) && \
> >>>> + off + sizeof(__u64) <= offsetofend(type, field) && \
> >>>> + off % sizeof(__u64) == 0)
> >>>> +
> >>>> #define bpf_classic_proglen(fprog) (fprog->len * sizeof(fprog->filter[0]))
> >>>>
> >>>> static inline void bpf_prog_lock_ro(struct bpf_prog *fp)
> >>>> diff --git a/net/core/filter.c b/net/core/filter.c
> >>>> index dc8534be12fc..5d33f2146dab 100644
> >>>> --- a/net/core/filter.c
> >>>> +++ b/net/core/filter.c
> >>>> @@ -6849,6 +6849,16 @@ static bool sock_addr_is_valid_access(int off, int size,
> >>>> if (!bpf_ctx_narrow_access_ok(off, size, size_default))
> >>>> return false;
> >>>> } else {
> >>>> + if (bpf_ctx_wide_store_ok(off, size,
> >>>> + struct bpf_sock_addr,
> >>>> + user_ip6))
> >>>> + return true;
> >>>> +
> >>>> + if (bpf_ctx_wide_store_ok(off, size,
> >>>> + struct bpf_sock_addr,
> >>>> + msg_src_ip6))
> >>>> + return true;
> >>>> +
> >>>> if (size != size_default)
> >>>> return false;
> >>>> }
> >>>> @@ -7689,9 +7699,6 @@ static u32 xdp_convert_ctx_access(enum bpf_access_type type,
> >>>> /* SOCK_ADDR_STORE_NESTED_FIELD_OFF() has semantic similar to
> >>>> * SOCK_ADDR_LOAD_NESTED_FIELD_SIZE_OFF() but for store operation.
> >>>> *
> >>>> - * It doesn't support SIZE argument though since narrow stores are not
> >>>> - * supported for now.
> >>>> - *
> >>>> * In addition it uses Temporary Field TF (member of struct S) as the 3rd
> >>>> * "register" since two registers available in convert_ctx_access are not
> >>>> * enough: we can't override neither SRC, since it contains value to store, nor
> >>>> @@ -7699,7 +7706,7 @@ static u32 xdp_convert_ctx_access(enum bpf_access_type type,
> >>>> * instructions. But we need a temporary place to save pointer to nested
> >>>> * structure whose field we want to store to.
> >>>> */
> >>>> -#define SOCK_ADDR_STORE_NESTED_FIELD_OFF(S, NS, F, NF, OFF, TF) \
> >>>> +#define SOCK_ADDR_STORE_NESTED_FIELD_OFF(S, NS, F, NF, SIZE, OFF, TF) \
> >>>> do { \
> >>>> int tmp_reg = BPF_REG_9; \
> >>>> if (si->src_reg == tmp_reg || si->dst_reg == tmp_reg) \
> >>>> @@ -7710,8 +7717,7 @@ static u32 xdp_convert_ctx_access(enum bpf_access_type type,
> >>>> offsetof(S, TF)); \
> >>>> *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(S, F), tmp_reg, \
> >>>> si->dst_reg, offsetof(S, F)); \
> >>>> - *insn++ = BPF_STX_MEM( \
> >>>> - BPF_FIELD_SIZEOF(NS, NF), tmp_reg, si->src_reg, \
> >>>> + *insn++ = BPF_STX_MEM(SIZE, tmp_reg, si->src_reg, \
> >>>> bpf_target_off(NS, NF, FIELD_SIZEOF(NS, NF), \
> >>>> target_size) \
> >>>> + OFF); \
> >>>> @@ -7723,8 +7729,8 @@ static u32 xdp_convert_ctx_access(enum bpf_access_type type,
> >>>> TF) \
> >>>> do { \
> >>>> if (type == BPF_WRITE) { \
> >>>> - SOCK_ADDR_STORE_NESTED_FIELD_OFF(S, NS, F, NF, OFF, \
> >>>> - TF); \
> >>>> + SOCK_ADDR_STORE_NESTED_FIELD_OFF(S, NS, F, NF, SIZE, \
> >>>> + OFF, TF); \
> >>>> } else { \
> >>>> SOCK_ADDR_LOAD_NESTED_FIELD_SIZE_OFF( \
> >>>> S, NS, F, NF, SIZE, OFF); \
> >>>>
^ permalink raw reply [flat|nested] 11+ messages in thread
end of thread, other threads:[~2019-07-01 18:39 UTC | newest]
Thread overview: 11+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2019-06-28 23:10 [PATCH bpf-next 1/2] bpf: allow wide (u64) aligned stores for some fields of bpf_sock_addr Stanislav Fomichev
2019-06-28 23:10 ` [PATCH bpf-next 2/2] selftests/bpf: add verifier tests for wide stores Stanislav Fomichev
2019-06-30 6:01 ` Yonghong Song
2019-07-01 15:44 ` Andrii Nakryiko
2019-07-01 16:00 ` Stanislav Fomichev
2019-06-30 5:52 ` [PATCH bpf-next 1/2] bpf: allow wide (u64) aligned stores for some fields of bpf_sock_addr Yonghong Song
2019-07-01 15:36 ` Andrii Nakryiko
2019-07-01 16:04 ` Stanislav Fomichev
2019-07-01 17:40 ` Yonghong Song
2019-07-01 18:38 ` Stanislav Fomichev
2019-07-01 16:01 ` Stanislav Fomichev
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).