diff mbox series

[bpf-next,3/5] selftests/bpf: fix BPF_CORE_READ_BITFIELD() macro

Message ID 20210423233058.3386115-4-andrii@kernel.org (mailing list archive)
State Superseded
Delegated to: BPF
Headers show
Series CO-RE relocation selftests fixes | expand

Checks

Context Check Description
netdev/cover_letter success Link
netdev/fixes_present success Link
netdev/patch_count success Link
netdev/tree_selection success Clearly marked for bpf-next
netdev/subject_prefix success Link
netdev/cc_maintainers warning 9 maintainers not CCed: yhs@fb.com kpsingh@kernel.org kafai@fb.com clang-built-linux@googlegroups.com ast@kernel.org nathan@kernel.org john.fastabend@gmail.com songliubraving@fb.com ndesaulniers@google.com
netdev/source_inline success Was 0 now: 0
netdev/verify_signedoff success Link
netdev/module_param success Was 0 now: 0
netdev/build_32bit success Errors and warnings before: 0 this patch: 0
netdev/kdoc success Errors and warnings before: 0 this patch: 0
netdev/verify_fixes success Link
netdev/checkpatch fail ERROR: trailing statements should be on next line WARNING: Avoid unnecessary line continuations
netdev/build_allmodconfig_warn success Errors and warnings before: 0 this patch: 0
netdev/header_inline success Link

Commit Message

Andrii Nakryiko April 23, 2021, 11:30 p.m. UTC
Fix BPF_CORE_READ_BITFIELD() macro used for reading CO-RE-relocatable
bitfields. Missing breaks in a switch caused 8-byte reads always. This can
confuse libbpf because it does strict checks that memory load size corresponds
to the original size of the field, which in this case quite often would be
wrong.

After fixing that, we run into another problem, which quite subtle, so worth
documenting here. The issue is in Clang optimization and CO-RE relocation
interactions. Without that asm volatile construct (also known as
barrier_var()), Clang will re-order BYTE_OFFSET and BYTE_SIZE relocations and
will apply BYTE_OFFSET 4 times for each switch case arm. This will result in
the same error from libbpf about mismatch of memory load size and original
field size. I.e., if we were reading u32, we'd still have *(u8 *), *(u16 *),
*(u32 *), and *(u64 *) memory loads, three of which will fail. Using
barrier_var() forces Clang to apply BYTE_OFFSET relocation first (and once) to
calculate p, after which value of p is used without relocation in each of
switch case arms, doing appropiately-sized memory load.

Here's the list of relevant relocations and pieces of generated BPF code
before and after this patch for test_core_reloc_bitfields_direct selftests.

BEFORE
=====
 #45: core_reloc: insn #160 --> [5] + 0:5: byte_sz --> struct core_reloc_bitfields.u32
 #46: core_reloc: insn #167 --> [5] + 0:5: byte_off --> struct core_reloc_bitfields.u32
 #47: core_reloc: insn #174 --> [5] + 0:5: byte_off --> struct core_reloc_bitfields.u32
 #48: core_reloc: insn #178 --> [5] + 0:5: byte_off --> struct core_reloc_bitfields.u32
 #49: core_reloc: insn #182 --> [5] + 0:5: byte_off --> struct core_reloc_bitfields.u32

     157:       18 02 00 00 00 00 00 00 00 00 00 00 00 00 00 00 r2 = 0 ll
     159:       7b 12 20 01 00 00 00 00 *(u64 *)(r2 + 288) = r1
     160:       b7 02 00 00 04 00 00 00 r2 = 4
; BYTE_SIZE relocation here                 ^^^
     161:       66 02 07 00 03 00 00 00 if w2 s> 3 goto +7 <LBB0_63>
     162:       16 02 0d 00 01 00 00 00 if w2 == 1 goto +13 <LBB0_65>
     163:       16 02 01 00 02 00 00 00 if w2 == 2 goto +1 <LBB0_66>
     164:       05 00 12 00 00 00 00 00 goto +18 <LBB0_69>

0000000000000528 <LBB0_66>:
     165:       18 01 00 00 00 00 00 00 00 00 00 00 00 00 00 00 r1 = 0 ll
     167:       69 11 08 00 00 00 00 00 r1 = *(u16 *)(r1 + 8)
; BYTE_OFFSET relo here w/ WRONG size        ^^^^^^^^^^^^^^^^
     168:       05 00 0e 00 00 00 00 00 goto +14 <LBB0_69>

0000000000000548 <LBB0_63>:
     169:       16 02 0a 00 04 00 00 00 if w2 == 4 goto +10 <LBB0_67>
     170:       16 02 01 00 08 00 00 00 if w2 == 8 goto +1 <LBB0_68>
     171:       05 00 0b 00 00 00 00 00 goto +11 <LBB0_69>

0000000000000560 <LBB0_68>:
     172:       18 01 00 00 00 00 00 00 00 00 00 00 00 00 00 00 r1 = 0 ll
     174:       79 11 08 00 00 00 00 00 r1 = *(u64 *)(r1 + 8)
; BYTE_OFFSET relo here w/ WRONG size        ^^^^^^^^^^^^^^^^
     175:       05 00 07 00 00 00 00 00 goto +7 <LBB0_69>

0000000000000580 <LBB0_65>:
     176:       18 01 00 00 00 00 00 00 00 00 00 00 00 00 00 00 r1 = 0 ll
     178:       71 11 08 00 00 00 00 00 r1 = *(u8 *)(r1 + 8)
; BYTE_OFFSET relo here w/ WRONG size        ^^^^^^^^^^^^^^^^
     179:       05 00 03 00 00 00 00 00 goto +3 <LBB0_69>

00000000000005a0 <LBB0_67>:
     180:       18 01 00 00 00 00 00 00 00 00 00 00 00 00 00 00 r1 = 0 ll
     182:       61 11 08 00 00 00 00 00 r1 = *(u32 *)(r1 + 8)
; BYTE_OFFSET relo here w/ RIGHT size        ^^^^^^^^^^^^^^^^

00000000000005b8 <LBB0_69>:
     183:       67 01 00 00 20 00 00 00 r1 <<= 32
     184:       b7 02 00 00 00 00 00 00 r2 = 0
     185:       16 02 02 00 00 00 00 00 if w2 == 0 goto +2 <LBB0_71>
     186:       c7 01 00 00 20 00 00 00 r1 s>>= 32
     187:       05 00 01 00 00 00 00 00 goto +1 <LBB0_72>

00000000000005e0 <LBB0_71>:
     188:       77 01 00 00 20 00 00 00 r1 >>= 32

AFTER
=====

 #30: core_reloc: insn #132 --> [5] + 0:5: byte_off --> struct core_reloc_bitfields.u32
 #31: core_reloc: insn #134 --> [5] + 0:5: byte_sz --> struct core_reloc_bitfields.u32

     129:       18 02 00 00 00 00 00 00 00 00 00 00 00 00 00 00 r2 = 0 ll
     131:       7b 12 20 01 00 00 00 00 *(u64 *)(r2 + 288) = r1
     132:       b7 01 00 00 08 00 00 00 r1 = 8
; BYTE_OFFSET relo here                     ^^^
; no size check for non-memory dereferencing instructions
     133:       0f 12 00 00 00 00 00 00 r2 += r1
     134:       b7 03 00 00 04 00 00 00 r3 = 4
; BYTE_SIZE relocation here                 ^^^
     135:       66 03 05 00 03 00 00 00 if w3 s> 3 goto +5 <LBB0_63>
     136:       16 03 09 00 01 00 00 00 if w3 == 1 goto +9 <LBB0_65>
     137:       16 03 01 00 02 00 00 00 if w3 == 2 goto +1 <LBB0_66>
     138:       05 00 0a 00 00 00 00 00 goto +10 <LBB0_69>

0000000000000458 <LBB0_66>:
     139:       69 21 00 00 00 00 00 00 r1 = *(u16 *)(r2 + 0)
; NO CO-RE relocation here                   ^^^^^^^^^^^^^^^^
     140:       05 00 08 00 00 00 00 00 goto +8 <LBB0_69>

0000000000000468 <LBB0_63>:
     141:       16 03 06 00 04 00 00 00 if w3 == 4 goto +6 <LBB0_67>
     142:       16 03 01 00 08 00 00 00 if w3 == 8 goto +1 <LBB0_68>
     143:       05 00 05 00 00 00 00 00 goto +5 <LBB0_69>

0000000000000480 <LBB0_68>:
     144:       79 21 00 00 00 00 00 00 r1 = *(u64 *)(r2 + 0)
; NO CO-RE relocation here                   ^^^^^^^^^^^^^^^^
     145:       05 00 03 00 00 00 00 00 goto +3 <LBB0_69>

0000000000000490 <LBB0_65>:
     146:       71 21 00 00 00 00 00 00 r1 = *(u8 *)(r2 + 0)
; NO CO-RE relocation here                   ^^^^^^^^^^^^^^^^
     147:       05 00 01 00 00 00 00 00 goto +1 <LBB0_69>

00000000000004a0 <LBB0_67>:
     148:       61 21 00 00 00 00 00 00 r1 = *(u32 *)(r2 + 0)
; NO CO-RE relocation here                   ^^^^^^^^^^^^^^^^

00000000000004a8 <LBB0_69>:
     149:       67 01 00 00 20 00 00 00 r1 <<= 32
     150:       b7 02 00 00 00 00 00 00 r2 = 0
     151:       16 02 02 00 00 00 00 00 if w2 == 0 goto +2 <LBB0_71>
     152:       c7 01 00 00 20 00 00 00 r1 s>>= 32
     153:       05 00 01 00 00 00 00 00 goto +1 <LBB0_72>

00000000000004d0 <LBB0_71>:
     154:       77 01 00 00 20 00 00 00 r1 >>= 323

Fixes: ee26dade0e3b ("libbpf: Add support for relocatable bitfields")
Signed-off-by: Andrii Nakryiko <andrii@kernel.org>
---
 tools/lib/bpf/bpf_core_read.h | 16 ++++++++++++----
 1 file changed, 12 insertions(+), 4 deletions(-)

Comments

Lorenz Bauer April 26, 2021, 8:10 a.m. UTC | #1
On Sat, 24 Apr 2021 at 00:36, Andrii Nakryiko <andrii@kernel.org> wrote:
>
> Fix BPF_CORE_READ_BITFIELD() macro used for reading CO-RE-relocatable
> bitfields. Missing breaks in a switch caused 8-byte reads always. This can
> confuse libbpf because it does strict checks that memory load size corresponds
> to the original size of the field, which in this case quite often would be
> wrong.

Acked-by: Lorenz Bauer <lmb@cloudflare.com>
diff mbox series

Patch

diff --git a/tools/lib/bpf/bpf_core_read.h b/tools/lib/bpf/bpf_core_read.h
index 53b3e199fb25..09ebe3db5f2f 100644
--- a/tools/lib/bpf/bpf_core_read.h
+++ b/tools/lib/bpf/bpf_core_read.h
@@ -88,11 +88,19 @@  enum bpf_enum_value_kind {
 	const void *p = (const void *)s + __CORE_RELO(s, field, BYTE_OFFSET); \
 	unsigned long long val;						      \
 									      \
+	/* This is a so-called barrier_var() operation that makes specified   \
+	 * variable "a black box" for optimizing compiler.		      \
+	 * It forces compiler to perform BYTE_OFFSET relocation on p and use  \
+	 * its calculated value in the switch below, instead of applying      \
+	 * the same relocation 4 times for each individual memory load.       \
+	 */								      \
+	asm volatile("" : "=r"(p) : "0"(p));				      \
+									      \
 	switch (__CORE_RELO(s, field, BYTE_SIZE)) {			      \
-	case 1: val = *(const unsigned char *)p;			      \
-	case 2: val = *(const unsigned short *)p;			      \
-	case 4: val = *(const unsigned int *)p;				      \
-	case 8: val = *(const unsigned long long *)p;			      \
+	case 1: val = *(const unsigned char *)p; break;			      \
+	case 2: val = *(const unsigned short *)p; break;		      \
+	case 4: val = *(const unsigned int *)p; break;			      \
+	case 8: val = *(const unsigned long long *)p; break;		      \
 	}								      \
 	val <<= __CORE_RELO(s, field, LSHIFT_U64);			      \
 	if (__CORE_RELO(s, field, SIGNED))				      \