BPF: Generate locked insn for __sync_fetch_and_add() with cpu v1/v2 #106494
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
cc @brycekahle |
|
✅ With the latest revision this PR passed the C/C++ code formatter. |
1aa05fd to
7e0c7e1
Compare
4ast
reviewed
Aug 29, 2024
| @@ -1,19 +1,22 @@ | |||
| ; RUN: llc < %s -march=bpfel -verify-machineinstrs -show-mc-encoding | FileCheck --check-prefixes=CHECK,CHECK-V2 %s | |||
| ; RUN: llc < %s -march=bpfel -verify-machineinstrs -show-mc-encoding -mcpu=v3 | FileCheck --check-prefixes=CHECK,CHECK-V3 %s | |||
| ; RUN: llc < %s -march=bpfel -verify-machineinstrs -show-mc-encoding | FileCheck %s | |||
There was a problem hiding this comment.
should we add -mcpu=v1 here and maybe flip a default to -mcpu=v3 ?
So that by default memory ordering is preserved and users that want to see xadd insn for (void)fetch_and_add
have to explicitly request it with -mcpu=v1.
There was a problem hiding this comment.
We can do this. This may cause verification failure or other things from v1 to v3 though. I guess users can either debug it to make it work with v3 or use -mcpu=v1 to restore its old behavior. We can have a separate patch just to flip the default from v1 to v3. I can then post it to bpf mailing list to let people know.
Regarding to selftests, current in llvm/test/CodeGen/BPF directory (not including sub-directory), we only '-mcpu=v1' for llvm-objdump since llvm-objdump has been improved to dump insn based on cpu v4.
Otherwise, -mcpu=v1 is the default and not specified in the test.
[~/work/llvm-project/llvm/test/CodeGen/BPF (main)]$ grep "mcpu=v1" *.ll
objdump_cond_op.ll:; RUN: llc -mtriple=bpfel -filetype=obj -o - %s | llvm-objdump --no-print-imm-hex --mcpu=v1 -d - | FileCheck %s
objdump_static_var.ll:; RUN: llc -mtriple=bpfel -filetype=obj -o - %s | llvm-objdump --no-print-imm-hex --mcpu=v1 -d - | FileCheck --check-prefix=CHECK %s
objdump_static_var.ll:; RUN: llc -mtriple=bpfeb -filetype=obj -o - %s | llvm-objdump --no-print-imm-hex --mcpu=v1 -d - | FileCheck --check-prefix=CHECK %s
[~/work/llvm-project/llvm/test/CodeGen/BPF (main)]$
Later when we want to make cpu v3 as the default, all related tests can be modified at that time.
eddyz87
reviewed
Aug 29, 2024
added 3 commits
August 30, 2024 10:24
…ch_add insn (llvm#101428)" This reverts commit c566769. See discussion in [1]. Currently, with -mcpu=v1/v2, atomic_fetch_add() insn is generated for (void)__sync_fetch_and_add(...). This breaks backward compatibility since there are codes who runs on old system (< 5.12) which does not support atomci_fetch_add(). Now let revert previous comment ([1]) and add additional logic in the next patch to ensure for (void)__sync_fetch_and_add(...), v1/v2: generate locked add insn >= v3: generate atomic_fetch_add insn [1] llvm#101428
In previous commit, atomic_fetch_and_*() operations are converted to atomic_<op>()'s if there are no return values. This is not what we want, we would like to preserve atomic_fetch_and_*() insn so bpf jit can add proper barrier insns. Preserving atomic_fetch_and_*() are okay for most __sync_fetch_and_*() functions, but not for __sync_fetch_and_add() since __sync_fetch_and_add() has been used to generic locked insns in cpu=v1/v2. So after preserving atomic_fetch_and_*() even if return value is not used, XFADDD in BPFInstrInfo.td is adjusted to emit locked add insn for cpu v1/v2 and to emit atomic_fetch_and_add() for cpu >= v3.
The patch is from Eduard. class XADD and class ATOMIC_NOFETCH are almost the same except the pattern matching string. Removing class XADD and defining two Pat's (pattern -> insn) for 32bit and 64bit XADD/XADDD to use ATOMIC_NOFETACH can simplify code a lot.
7e0c7e1 to
1d37bc7
Compare
eddyz87
approved these changes
Aug 30, 2024
|
LLVM Buildbot has detected a new failure on builder Full details are available at: https://lab.llvm.org/buildbot/#/builders/143/builds/1815 Here is the relevant piece of the build log for the reference |
yonghong-song
pushed a commit
to yonghong-song/llvm-project
that referenced
this pull request
Sep 2, 2024
Before llvm20, (void)__sync_fetch_and_add(...) always generates locked xadd insns. In linux kernel upstream discussion [1], it is found that for arm64 architecture, the original semantics of (void)__sync_fetch_and_add(...), i.e., __atomic_fetch_add(...), is preferred in order for jit to emit proper native barrier insns. In llvm commits [2] and [3], (void)__sync_fetch_and_add(...) will generate the following insns: - for cpu v1/v2: locked xadd insns to keep backward compatibility - for cpu v3/v4: __atomic_fetch_add() insns To ensure proper barrier semantics for (void)__sync_fetch_and_add(...), cpu v3/v4 is recommended. This patch enables cpu=v3 as the default cpu version. For users wanting to use cpu v1, -mcpu=v1 needs to be explicitly added to clang/llc command line. [1] https://lore.kernel.org/bpf/ZqqiQQWRnz7H93Hc@google.com/T/#mb68d67bc8f39e35a0c3db52468b9de59b79f021f [2] llvm#101428 [3] llvm#106494
yonghong-song
pushed a commit
to yonghong-song/llvm-project
that referenced
this pull request
Sep 3, 2024
Before llvm20, (void)__sync_fetch_and_add(...) always generates locked xadd insns. In linux kernel upstream discussion [1], it is found that for arm64 architecture, the original semantics of (void)__sync_fetch_and_add(...), i.e., __atomic_fetch_add(...), is preferred in order for jit to emit proper native barrier insns. In llvm commits [2] and [3], (void)__sync_fetch_and_add(...) will generate the following insns: - for cpu v1/v2: locked xadd insns to keep backward compatibility - for cpu v3/v4: __atomic_fetch_add() insns To ensure proper barrier semantics for (void)__sync_fetch_and_add(...), cpu v3/v4 is recommended. This patch enables cpu=v3 as the default cpu version. For users wanting to use cpu v1, -mcpu=v1 needs to be explicitly added to clang/llc command line. [1] https://lore.kernel.org/bpf/ZqqiQQWRnz7H93Hc@google.com/T/#mb68d67bc8f39e35a0c3db52468b9de59b79f021f [2] llvm#101428 [3] llvm#106494
yonghong-song
added a commit
that referenced
this pull request
Sep 3, 2024
Before llvm20, (void)__sync_fetch_and_add(...) always generates locked xadd insns. In linux kernel upstream discussion [1], it is found that for arm64 architecture, the original semantics of (void)__sync_fetch_and_add(...), i.e., __atomic_fetch_add(...), is preferred in order for jit to emit proper native barrier insns. In llvm commits [2] and [3], (void)__sync_fetch_and_add(...) will generate the following insns: - for cpu v1/v2: locked xadd insns to keep backward compatibility - for cpu v3/v4: __atomic_fetch_add() insns To ensure proper barrier semantics for (void)__sync_fetch_and_add(...), cpu v3/v4 is recommended. This patch enables cpu=v3 as the default cpu version. For users wanting to use cpu v1, -mcpu=v1 needs to be explicitly added to clang/llc command line. [1] https://lore.kernel.org/bpf/ZqqiQQWRnz7H93Hc@google.com/T/#mb68d67bc8f39e35a0c3db52468b9de59b79f021f [2] #101428 [3] #106494
intel-lab-lkp
pushed a commit
to intel-lab-lkp/linux
that referenced
this pull request
Sep 9, 2024
llvm change [1] made a change such that __sync_fetch_and_{and,or,xor}()
will generate atomic_fetch_*() insns even if the return value is not used.
This is a deliberate choice to make sure barrier semantics are preserved
from source code to asm insn.
But the change in [1] caused arena_atomics selftest failure.
test_arena_atomics:PASS:arena atomics skeleton open 0 nsec
libbpf: prog 'and': BPF program load failed: Permission denied
libbpf: prog 'and': -- BEGIN PROG LOAD LOG --
arg#0 reference type('UNKNOWN ') size cannot be determined: -22
0: R1=ctx() R10=fp0
; if (pid != (bpf_get_current_pid_tgid() >> 32)) @ arena_atomics.c:87
0: (18) r1 = 0xffffc90000064000 ; R1_w=map_value(map=arena_at.bss,ks=4,vs=4)
2: (61) r6 = *(u32 *)(r1 +0) ; R1_w=map_value(map=arena_at.bss,ks=4,vs=4) R6_w=scalar(smin=0,smax=umax=0xffffffff,v
ar_off=(0x0; 0xffffffff))
3: (85) call bpf_get_current_pid_tgid#14 ; R0_w=scalar()
4: (77) r0 >>= 32 ; R0_w=scalar(smin=0,smax=umax=0xffffffff,var_off=(0x0; 0xffffffff))
5: (5d) if r0 != r6 goto pc+11 ; R0_w=scalar(smin=0,smax=umax=0xffffffff,var_off=(0x0; 0xffffffff)) R6_w=scalar(smin=0,smax=umax=0xffffffff,var_off=(0x0; 0x)
; __sync_fetch_and_and(&and64_value, 0x011ull << 32); @ arena_atomics.c:91
6: (18) r1 = 0x100000000060 ; R1_w=scalar()
8: (bf) r1 = addr_space_cast(r1, 0, 1) ; R1_w=arena
9: (18) r2 = 0x1100000000 ; R2_w=0x1100000000
11: (db) r2 = atomic64_fetch_and((u64 *)(r1 +0), r2)
BPF_ATOMIC stores into R1 arena is not allowed
processed 9 insns (limit 1000000) max_states_per_insn 0 total_states 0 peak_states 0 mark_read 0
-- END PROG LOAD LOG --
libbpf: prog 'and': failed to load: -13
libbpf: failed to load object 'arena_atomics'
libbpf: failed to load BPF skeleton 'arena_atomics': -13
test_arena_atomics:FAIL:arena atomics skeleton load unexpected error: -13 (errno 13)
#3 arena_atomics:FAIL
The reason of the failure is due to [2] where atomic{64,}_fetch_{and,or,xor}() are not
allowed by arena addresses.
Version 2 of the patch fixed the issue by using inline asm ([3]). But further discussion
suggested to find a way from source to generate locked insn which is more user
friendly. So in not-merged llvm patch ([4]), if relax memory ordering is used and
the return value is not used, locked insn could be generated.
So with llvm patch [4] to compile the bpf selftest, the following code
__c11_atomic_fetch_and(&and64_value, 0x011ull << 32, memory_order_relaxed);
is able to generate locked insn, hence fixing the selftest failure.
[1] llvm/llvm-project#106494
[2] d503a04 ("bpf: Add support for certain atomics in bpf_arena to x86 JIT")
[3] https://lore.kernel.org/bpf/20240803025928.4184433-1-yonghong.song@linux.dev/
[4] llvm/llvm-project#107343
Signed-off-by: Yonghong Song <yonghong.song@linux.dev>
kernel-patches-daemon-bpf bot
pushed a commit
to kernel-patches/bpf
that referenced
this pull request
Sep 9, 2024
llvm change [1] made a change such that __sync_fetch_and_{and,or,xor}()
will generate atomic_fetch_*() insns even if the return value is not used.
This is a deliberate choice to make sure barrier semantics are preserved
from source code to asm insn.
But the change in [1] caused arena_atomics selftest failure.
test_arena_atomics:PASS:arena atomics skeleton open 0 nsec
libbpf: prog 'and': BPF program load failed: Permission denied
libbpf: prog 'and': -- BEGIN PROG LOAD LOG --
arg#0 reference type('UNKNOWN ') size cannot be determined: -22
0: R1=ctx() R10=fp0
; if (pid != (bpf_get_current_pid_tgid() >> 32)) @ arena_atomics.c:87
0: (18) r1 = 0xffffc90000064000 ; R1_w=map_value(map=arena_at.bss,ks=4,vs=4)
2: (61) r6 = *(u32 *)(r1 +0) ; R1_w=map_value(map=arena_at.bss,ks=4,vs=4) R6_w=scalar(smin=0,smax=umax=0xffffffff,v
ar_off=(0x0; 0xffffffff))
3: (85) call bpf_get_current_pid_tgid#14 ; R0_w=scalar()
4: (77) r0 >>= 32 ; R0_w=scalar(smin=0,smax=umax=0xffffffff,var_off=(0x0; 0xffffffff))
5: (5d) if r0 != r6 goto pc+11 ; R0_w=scalar(smin=0,smax=umax=0xffffffff,var_off=(0x0; 0xffffffff)) R6_w=scalar(smin=0,smax=umax=0xffffffff,var_off=(0x0; 0x)
; __sync_fetch_and_and(&and64_value, 0x011ull << 32); @ arena_atomics.c:91
6: (18) r1 = 0x100000000060 ; R1_w=scalar()
8: (bf) r1 = addr_space_cast(r1, 0, 1) ; R1_w=arena
9: (18) r2 = 0x1100000000 ; R2_w=0x1100000000
11: (db) r2 = atomic64_fetch_and((u64 *)(r1 +0), r2)
BPF_ATOMIC stores into R1 arena is not allowed
processed 9 insns (limit 1000000) max_states_per_insn 0 total_states 0 peak_states 0 mark_read 0
-- END PROG LOAD LOG --
libbpf: prog 'and': failed to load: -13
libbpf: failed to load object 'arena_atomics'
libbpf: failed to load BPF skeleton 'arena_atomics': -13
test_arena_atomics:FAIL:arena atomics skeleton load unexpected error: -13 (errno 13)
#3 arena_atomics:FAIL
The reason of the failure is due to [2] where atomic{64,}_fetch_{and,or,xor}() are not
allowed by arena addresses.
Version 2 of the patch fixed the issue by using inline asm ([3]). But further discussion
suggested to find a way from source to generate locked insn which is more user
friendly. So in not-merged llvm patch ([4]), if relax memory ordering is used and
the return value is not used, locked insn could be generated.
So with llvm patch [4] to compile the bpf selftest, the following code
__c11_atomic_fetch_and(&and64_value, 0x011ull << 32, memory_order_relaxed);
is able to generate locked insn, hence fixing the selftest failure.
[1] llvm/llvm-project#106494
[2] d503a04 ("bpf: Add support for certain atomics in bpf_arena to x86 JIT")
[3] https://lore.kernel.org/bpf/20240803025928.4184433-1-yonghong.song@linux.dev/
[4] llvm/llvm-project#107343
Signed-off-by: Yonghong Song <yonghong.song@linux.dev>
kernel-patches-daemon-bpf-rc bot
pushed a commit
to kernel-patches/bpf-rc
that referenced
this pull request
Sep 9, 2024
llvm change [1] made a change such that __sync_fetch_and_{and,or,xor}()
will generate atomic_fetch_*() insns even if the return value is not used.
This is a deliberate choice to make sure barrier semantics are preserved
from source code to asm insn.
But the change in [1] caused arena_atomics selftest failure.
test_arena_atomics:PASS:arena atomics skeleton open 0 nsec
libbpf: prog 'and': BPF program load failed: Permission denied
libbpf: prog 'and': -- BEGIN PROG LOAD LOG --
arg#0 reference type('UNKNOWN ') size cannot be determined: -22
0: R1=ctx() R10=fp0
; if (pid != (bpf_get_current_pid_tgid() >> 32)) @ arena_atomics.c:87
0: (18) r1 = 0xffffc90000064000 ; R1_w=map_value(map=arena_at.bss,ks=4,vs=4)
2: (61) r6 = *(u32 *)(r1 +0) ; R1_w=map_value(map=arena_at.bss,ks=4,vs=4) R6_w=scalar(smin=0,smax=umax=0xffffffff,v
ar_off=(0x0; 0xffffffff))
3: (85) call bpf_get_current_pid_tgid#14 ; R0_w=scalar()
4: (77) r0 >>= 32 ; R0_w=scalar(smin=0,smax=umax=0xffffffff,var_off=(0x0; 0xffffffff))
5: (5d) if r0 != r6 goto pc+11 ; R0_w=scalar(smin=0,smax=umax=0xffffffff,var_off=(0x0; 0xffffffff)) R6_w=scalar(smin=0,smax=umax=0xffffffff,var_off=(0x0; 0x)
; __sync_fetch_and_and(&and64_value, 0x011ull << 32); @ arena_atomics.c:91
6: (18) r1 = 0x100000000060 ; R1_w=scalar()
8: (bf) r1 = addr_space_cast(r1, 0, 1) ; R1_w=arena
9: (18) r2 = 0x1100000000 ; R2_w=0x1100000000
11: (db) r2 = atomic64_fetch_and((u64 *)(r1 +0), r2)
BPF_ATOMIC stores into R1 arena is not allowed
processed 9 insns (limit 1000000) max_states_per_insn 0 total_states 0 peak_states 0 mark_read 0
-- END PROG LOAD LOG --
libbpf: prog 'and': failed to load: -13
libbpf: failed to load object 'arena_atomics'
libbpf: failed to load BPF skeleton 'arena_atomics': -13
test_arena_atomics:FAIL:arena atomics skeleton load unexpected error: -13 (errno 13)
#3 arena_atomics:FAIL
The reason of the failure is due to [2] where atomic{64,}_fetch_{and,or,xor}() are not
allowed by arena addresses.
Version 2 of the patch fixed the issue by using inline asm ([3]). But further discussion
suggested to find a way from source to generate locked insn which is more user
friendly. So in not-merged llvm patch ([4]), if relax memory ordering is used and
the return value is not used, locked insn could be generated.
So with llvm patch [4] to compile the bpf selftest, the following code
__c11_atomic_fetch_and(&and64_value, 0x011ull << 32, memory_order_relaxed);
is able to generate locked insn, hence fixing the selftest failure.
[1] llvm/llvm-project#106494
[2] d503a04 ("bpf: Add support for certain atomics in bpf_arena to x86 JIT")
[3] https://lore.kernel.org/bpf/20240803025928.4184433-1-yonghong.song@linux.dev/
[4] llvm/llvm-project#107343
Signed-off-by: Yonghong Song <yonghong.song@linux.dev>
kernel-patches-daemon-bpf bot
pushed a commit
to kernel-patches/bpf
that referenced
this pull request
Sep 9, 2024
llvm change [1] made a change such that __sync_fetch_and_{and,or,xor}()
will generate atomic_fetch_*() insns even if the return value is not used.
This is a deliberate choice to make sure barrier semantics are preserved
from source code to asm insn.
But the change in [1] caused arena_atomics selftest failure.
test_arena_atomics:PASS:arena atomics skeleton open 0 nsec
libbpf: prog 'and': BPF program load failed: Permission denied
libbpf: prog 'and': -- BEGIN PROG LOAD LOG --
arg#0 reference type('UNKNOWN ') size cannot be determined: -22
0: R1=ctx() R10=fp0
; if (pid != (bpf_get_current_pid_tgid() >> 32)) @ arena_atomics.c:87
0: (18) r1 = 0xffffc90000064000 ; R1_w=map_value(map=arena_at.bss,ks=4,vs=4)
2: (61) r6 = *(u32 *)(r1 +0) ; R1_w=map_value(map=arena_at.bss,ks=4,vs=4) R6_w=scalar(smin=0,smax=umax=0xffffffff,v
ar_off=(0x0; 0xffffffff))
3: (85) call bpf_get_current_pid_tgid#14 ; R0_w=scalar()
4: (77) r0 >>= 32 ; R0_w=scalar(smin=0,smax=umax=0xffffffff,var_off=(0x0; 0xffffffff))
5: (5d) if r0 != r6 goto pc+11 ; R0_w=scalar(smin=0,smax=umax=0xffffffff,var_off=(0x0; 0xffffffff)) R6_w=scalar(smin=0,smax=umax=0xffffffff,var_off=(0x0; 0x)
; __sync_fetch_and_and(&and64_value, 0x011ull << 32); @ arena_atomics.c:91
6: (18) r1 = 0x100000000060 ; R1_w=scalar()
8: (bf) r1 = addr_space_cast(r1, 0, 1) ; R1_w=arena
9: (18) r2 = 0x1100000000 ; R2_w=0x1100000000
11: (db) r2 = atomic64_fetch_and((u64 *)(r1 +0), r2)
BPF_ATOMIC stores into R1 arena is not allowed
processed 9 insns (limit 1000000) max_states_per_insn 0 total_states 0 peak_states 0 mark_read 0
-- END PROG LOAD LOG --
libbpf: prog 'and': failed to load: -13
libbpf: failed to load object 'arena_atomics'
libbpf: failed to load BPF skeleton 'arena_atomics': -13
test_arena_atomics:FAIL:arena atomics skeleton load unexpected error: -13 (errno 13)
#3 arena_atomics:FAIL
The reason of the failure is due to [2] where atomic{64,}_fetch_{and,or,xor}() are not
allowed by arena addresses.
Version 2 of the patch fixed the issue by using inline asm ([3]). But further discussion
suggested to find a way from source to generate locked insn which is more user
friendly. So in not-merged llvm patch ([4]), if relax memory ordering is used and
the return value is not used, locked insn could be generated.
So with llvm patch [4] to compile the bpf selftest, the following code
__c11_atomic_fetch_and(&and64_value, 0x011ull << 32, memory_order_relaxed);
is able to generate locked insn, hence fixing the selftest failure.
[1] llvm/llvm-project#106494
[2] d503a04 ("bpf: Add support for certain atomics in bpf_arena to x86 JIT")
[3] https://lore.kernel.org/bpf/20240803025928.4184433-1-yonghong.song@linux.dev/
[4] llvm/llvm-project#107343
Signed-off-by: Yonghong Song <yonghong.song@linux.dev>
kernel-patches-daemon-bpf-rc bot
pushed a commit
to kernel-patches/bpf-rc
that referenced
this pull request
Sep 9, 2024
llvm change [1] made a change such that __sync_fetch_and_{and,or,xor}()
will generate atomic_fetch_*() insns even if the return value is not used.
This is a deliberate choice to make sure barrier semantics are preserved
from source code to asm insn.
But the change in [1] caused arena_atomics selftest failure.
test_arena_atomics:PASS:arena atomics skeleton open 0 nsec
libbpf: prog 'and': BPF program load failed: Permission denied
libbpf: prog 'and': -- BEGIN PROG LOAD LOG --
arg#0 reference type('UNKNOWN ') size cannot be determined: -22
0: R1=ctx() R10=fp0
; if (pid != (bpf_get_current_pid_tgid() >> 32)) @ arena_atomics.c:87
0: (18) r1 = 0xffffc90000064000 ; R1_w=map_value(map=arena_at.bss,ks=4,vs=4)
2: (61) r6 = *(u32 *)(r1 +0) ; R1_w=map_value(map=arena_at.bss,ks=4,vs=4) R6_w=scalar(smin=0,smax=umax=0xffffffff,v
ar_off=(0x0; 0xffffffff))
3: (85) call bpf_get_current_pid_tgid#14 ; R0_w=scalar()
4: (77) r0 >>= 32 ; R0_w=scalar(smin=0,smax=umax=0xffffffff,var_off=(0x0; 0xffffffff))
5: (5d) if r0 != r6 goto pc+11 ; R0_w=scalar(smin=0,smax=umax=0xffffffff,var_off=(0x0; 0xffffffff)) R6_w=scalar(smin=0,smax=umax=0xffffffff,var_off=(0x0; 0x)
; __sync_fetch_and_and(&and64_value, 0x011ull << 32); @ arena_atomics.c:91
6: (18) r1 = 0x100000000060 ; R1_w=scalar()
8: (bf) r1 = addr_space_cast(r1, 0, 1) ; R1_w=arena
9: (18) r2 = 0x1100000000 ; R2_w=0x1100000000
11: (db) r2 = atomic64_fetch_and((u64 *)(r1 +0), r2)
BPF_ATOMIC stores into R1 arena is not allowed
processed 9 insns (limit 1000000) max_states_per_insn 0 total_states 0 peak_states 0 mark_read 0
-- END PROG LOAD LOG --
libbpf: prog 'and': failed to load: -13
libbpf: failed to load object 'arena_atomics'
libbpf: failed to load BPF skeleton 'arena_atomics': -13
test_arena_atomics:FAIL:arena atomics skeleton load unexpected error: -13 (errno 13)
#3 arena_atomics:FAIL
The reason of the failure is due to [2] where atomic{64,}_fetch_{and,or,xor}() are not
allowed by arena addresses.
Version 2 of the patch fixed the issue by using inline asm ([3]). But further discussion
suggested to find a way from source to generate locked insn which is more user
friendly. So in not-merged llvm patch ([4]), if relax memory ordering is used and
the return value is not used, locked insn could be generated.
So with llvm patch [4] to compile the bpf selftest, the following code
__c11_atomic_fetch_and(&and64_value, 0x011ull << 32, memory_order_relaxed);
is able to generate locked insn, hence fixing the selftest failure.
[1] llvm/llvm-project#106494
[2] d503a04 ("bpf: Add support for certain atomics in bpf_arena to x86 JIT")
[3] https://lore.kernel.org/bpf/20240803025928.4184433-1-yonghong.song@linux.dev/
[4] llvm/llvm-project#107343
Signed-off-by: Yonghong Song <yonghong.song@linux.dev>
kernel-patches-daemon-bpf bot
pushed a commit
to kernel-patches/bpf
that referenced
this pull request
Sep 9, 2024
llvm change [1] made a change such that __sync_fetch_and_{and,or,xor}()
will generate atomic_fetch_*() insns even if the return value is not used.
This is a deliberate choice to make sure barrier semantics are preserved
from source code to asm insn.
But the change in [1] caused arena_atomics selftest failure.
test_arena_atomics:PASS:arena atomics skeleton open 0 nsec
libbpf: prog 'and': BPF program load failed: Permission denied
libbpf: prog 'and': -- BEGIN PROG LOAD LOG --
arg#0 reference type('UNKNOWN ') size cannot be determined: -22
0: R1=ctx() R10=fp0
; if (pid != (bpf_get_current_pid_tgid() >> 32)) @ arena_atomics.c:87
0: (18) r1 = 0xffffc90000064000 ; R1_w=map_value(map=arena_at.bss,ks=4,vs=4)
2: (61) r6 = *(u32 *)(r1 +0) ; R1_w=map_value(map=arena_at.bss,ks=4,vs=4) R6_w=scalar(smin=0,smax=umax=0xffffffff,v
ar_off=(0x0; 0xffffffff))
3: (85) call bpf_get_current_pid_tgid#14 ; R0_w=scalar()
4: (77) r0 >>= 32 ; R0_w=scalar(smin=0,smax=umax=0xffffffff,var_off=(0x0; 0xffffffff))
5: (5d) if r0 != r6 goto pc+11 ; R0_w=scalar(smin=0,smax=umax=0xffffffff,var_off=(0x0; 0xffffffff)) R6_w=scalar(smin=0,smax=umax=0xffffffff,var_off=(0x0; 0x)
; __sync_fetch_and_and(&and64_value, 0x011ull << 32); @ arena_atomics.c:91
6: (18) r1 = 0x100000000060 ; R1_w=scalar()
8: (bf) r1 = addr_space_cast(r1, 0, 1) ; R1_w=arena
9: (18) r2 = 0x1100000000 ; R2_w=0x1100000000
11: (db) r2 = atomic64_fetch_and((u64 *)(r1 +0), r2)
BPF_ATOMIC stores into R1 arena is not allowed
processed 9 insns (limit 1000000) max_states_per_insn 0 total_states 0 peak_states 0 mark_read 0
-- END PROG LOAD LOG --
libbpf: prog 'and': failed to load: -13
libbpf: failed to load object 'arena_atomics'
libbpf: failed to load BPF skeleton 'arena_atomics': -13
test_arena_atomics:FAIL:arena atomics skeleton load unexpected error: -13 (errno 13)
#3 arena_atomics:FAIL
The reason of the failure is due to [2] where atomic{64,}_fetch_{and,or,xor}() are not
allowed by arena addresses.
Version 2 of the patch fixed the issue by using inline asm ([3]). But further discussion
suggested to find a way from source to generate locked insn which is more user
friendly. So in not-merged llvm patch ([4]), if relax memory ordering is used and
the return value is not used, locked insn could be generated.
So with llvm patch [4] to compile the bpf selftest, the following code
__c11_atomic_fetch_and(&and64_value, 0x011ull << 32, memory_order_relaxed);
is able to generate locked insn, hence fixing the selftest failure.
[1] llvm/llvm-project#106494
[2] d503a04 ("bpf: Add support for certain atomics in bpf_arena to x86 JIT")
[3] https://lore.kernel.org/bpf/20240803025928.4184433-1-yonghong.song@linux.dev/
[4] llvm/llvm-project#107343
Signed-off-by: Yonghong Song <yonghong.song@linux.dev>
kernel-patches-daemon-bpf-rc bot
pushed a commit
to kernel-patches/bpf-rc
that referenced
this pull request
Sep 9, 2024
llvm change [1] made a change such that __sync_fetch_and_{and,or,xor}()
will generate atomic_fetch_*() insns even if the return value is not used.
This is a deliberate choice to make sure barrier semantics are preserved
from source code to asm insn.
But the change in [1] caused arena_atomics selftest failure.
test_arena_atomics:PASS:arena atomics skeleton open 0 nsec
libbpf: prog 'and': BPF program load failed: Permission denied
libbpf: prog 'and': -- BEGIN PROG LOAD LOG --
arg#0 reference type('UNKNOWN ') size cannot be determined: -22
0: R1=ctx() R10=fp0
; if (pid != (bpf_get_current_pid_tgid() >> 32)) @ arena_atomics.c:87
0: (18) r1 = 0xffffc90000064000 ; R1_w=map_value(map=arena_at.bss,ks=4,vs=4)
2: (61) r6 = *(u32 *)(r1 +0) ; R1_w=map_value(map=arena_at.bss,ks=4,vs=4) R6_w=scalar(smin=0,smax=umax=0xffffffff,v
ar_off=(0x0; 0xffffffff))
3: (85) call bpf_get_current_pid_tgid#14 ; R0_w=scalar()
4: (77) r0 >>= 32 ; R0_w=scalar(smin=0,smax=umax=0xffffffff,var_off=(0x0; 0xffffffff))
5: (5d) if r0 != r6 goto pc+11 ; R0_w=scalar(smin=0,smax=umax=0xffffffff,var_off=(0x0; 0xffffffff)) R6_w=scalar(smin=0,smax=umax=0xffffffff,var_off=(0x0; 0x)
; __sync_fetch_and_and(&and64_value, 0x011ull << 32); @ arena_atomics.c:91
6: (18) r1 = 0x100000000060 ; R1_w=scalar()
8: (bf) r1 = addr_space_cast(r1, 0, 1) ; R1_w=arena
9: (18) r2 = 0x1100000000 ; R2_w=0x1100000000
11: (db) r2 = atomic64_fetch_and((u64 *)(r1 +0), r2)
BPF_ATOMIC stores into R1 arena is not allowed
processed 9 insns (limit 1000000) max_states_per_insn 0 total_states 0 peak_states 0 mark_read 0
-- END PROG LOAD LOG --
libbpf: prog 'and': failed to load: -13
libbpf: failed to load object 'arena_atomics'
libbpf: failed to load BPF skeleton 'arena_atomics': -13
test_arena_atomics:FAIL:arena atomics skeleton load unexpected error: -13 (errno 13)
#3 arena_atomics:FAIL
The reason of the failure is due to [2] where atomic{64,}_fetch_{and,or,xor}() are not
allowed by arena addresses.
Version 2 of the patch fixed the issue by using inline asm ([3]). But further discussion
suggested to find a way from source to generate locked insn which is more user
friendly. So in not-merged llvm patch ([4]), if relax memory ordering is used and
the return value is not used, locked insn could be generated.
So with llvm patch [4] to compile the bpf selftest, the following code
__c11_atomic_fetch_and(&and64_value, 0x011ull << 32, memory_order_relaxed);
is able to generate locked insn, hence fixing the selftest failure.
[1] llvm/llvm-project#106494
[2] d503a04 ("bpf: Add support for certain atomics in bpf_arena to x86 JIT")
[3] https://lore.kernel.org/bpf/20240803025928.4184433-1-yonghong.song@linux.dev/
[4] llvm/llvm-project#107343
Signed-off-by: Yonghong Song <yonghong.song@linux.dev>
kernel-patches-daemon-bpf-rc bot
pushed a commit
to kernel-patches/bpf-rc
that referenced
this pull request
Sep 10, 2024
llvm change [1] made a change such that __sync_fetch_and_{and,or,xor}()
will generate atomic_fetch_*() insns even if the return value is not used.
This is a deliberate choice to make sure barrier semantics are preserved
from source code to asm insn.
But the change in [1] caused arena_atomics selftest failure.
test_arena_atomics:PASS:arena atomics skeleton open 0 nsec
libbpf: prog 'and': BPF program load failed: Permission denied
libbpf: prog 'and': -- BEGIN PROG LOAD LOG --
arg#0 reference type('UNKNOWN ') size cannot be determined: -22
0: R1=ctx() R10=fp0
; if (pid != (bpf_get_current_pid_tgid() >> 32)) @ arena_atomics.c:87
0: (18) r1 = 0xffffc90000064000 ; R1_w=map_value(map=arena_at.bss,ks=4,vs=4)
2: (61) r6 = *(u32 *)(r1 +0) ; R1_w=map_value(map=arena_at.bss,ks=4,vs=4) R6_w=scalar(smin=0,smax=umax=0xffffffff,v
ar_off=(0x0; 0xffffffff))
3: (85) call bpf_get_current_pid_tgid#14 ; R0_w=scalar()
4: (77) r0 >>= 32 ; R0_w=scalar(smin=0,smax=umax=0xffffffff,var_off=(0x0; 0xffffffff))
5: (5d) if r0 != r6 goto pc+11 ; R0_w=scalar(smin=0,smax=umax=0xffffffff,var_off=(0x0; 0xffffffff)) R6_w=scalar(smin=0,smax=umax=0xffffffff,var_off=(0x0; 0x)
; __sync_fetch_and_and(&and64_value, 0x011ull << 32); @ arena_atomics.c:91
6: (18) r1 = 0x100000000060 ; R1_w=scalar()
8: (bf) r1 = addr_space_cast(r1, 0, 1) ; R1_w=arena
9: (18) r2 = 0x1100000000 ; R2_w=0x1100000000
11: (db) r2 = atomic64_fetch_and((u64 *)(r1 +0), r2)
BPF_ATOMIC stores into R1 arena is not allowed
processed 9 insns (limit 1000000) max_states_per_insn 0 total_states 0 peak_states 0 mark_read 0
-- END PROG LOAD LOG --
libbpf: prog 'and': failed to load: -13
libbpf: failed to load object 'arena_atomics'
libbpf: failed to load BPF skeleton 'arena_atomics': -13
test_arena_atomics:FAIL:arena atomics skeleton load unexpected error: -13 (errno 13)
#3 arena_atomics:FAIL
The reason of the failure is due to [2] where atomic{64,}_fetch_{and,or,xor}() are not
allowed by arena addresses.
Version 2 of the patch fixed the issue by using inline asm ([3]). But further discussion
suggested to find a way from source to generate locked insn which is more user
friendly. So in not-merged llvm patch ([4]), if relax memory ordering is used and
the return value is not used, locked insn could be generated.
So with llvm patch [4] to compile the bpf selftest, the following code
__c11_atomic_fetch_and(&and64_value, 0x011ull << 32, memory_order_relaxed);
is able to generate locked insn, hence fixing the selftest failure.
[1] llvm/llvm-project#106494
[2] d503a04 ("bpf: Add support for certain atomics in bpf_arena to x86 JIT")
[3] https://lore.kernel.org/bpf/20240803025928.4184433-1-yonghong.song@linux.dev/
[4] llvm/llvm-project#107343
Signed-off-by: Yonghong Song <yonghong.song@linux.dev>
kernel-patches-daemon-bpf bot
pushed a commit
to kernel-patches/bpf
that referenced
this pull request
Sep 10, 2024
llvm change [1] made a change such that __sync_fetch_and_{and,or,xor}()
will generate atomic_fetch_*() insns even if the return value is not used.
This is a deliberate choice to make sure barrier semantics are preserved
from source code to asm insn.
But the change in [1] caused arena_atomics selftest failure.
test_arena_atomics:PASS:arena atomics skeleton open 0 nsec
libbpf: prog 'and': BPF program load failed: Permission denied
libbpf: prog 'and': -- BEGIN PROG LOAD LOG --
arg#0 reference type('UNKNOWN ') size cannot be determined: -22
0: R1=ctx() R10=fp0
; if (pid != (bpf_get_current_pid_tgid() >> 32)) @ arena_atomics.c:87
0: (18) r1 = 0xffffc90000064000 ; R1_w=map_value(map=arena_at.bss,ks=4,vs=4)
2: (61) r6 = *(u32 *)(r1 +0) ; R1_w=map_value(map=arena_at.bss,ks=4,vs=4) R6_w=scalar(smin=0,smax=umax=0xffffffff,v
ar_off=(0x0; 0xffffffff))
3: (85) call bpf_get_current_pid_tgid#14 ; R0_w=scalar()
4: (77) r0 >>= 32 ; R0_w=scalar(smin=0,smax=umax=0xffffffff,var_off=(0x0; 0xffffffff))
5: (5d) if r0 != r6 goto pc+11 ; R0_w=scalar(smin=0,smax=umax=0xffffffff,var_off=(0x0; 0xffffffff)) R6_w=scalar(smin=0,smax=umax=0xffffffff,var_off=(0x0; 0x)
; __sync_fetch_and_and(&and64_value, 0x011ull << 32); @ arena_atomics.c:91
6: (18) r1 = 0x100000000060 ; R1_w=scalar()
8: (bf) r1 = addr_space_cast(r1, 0, 1) ; R1_w=arena
9: (18) r2 = 0x1100000000 ; R2_w=0x1100000000
11: (db) r2 = atomic64_fetch_and((u64 *)(r1 +0), r2)
BPF_ATOMIC stores into R1 arena is not allowed
processed 9 insns (limit 1000000) max_states_per_insn 0 total_states 0 peak_states 0 mark_read 0
-- END PROG LOAD LOG --
libbpf: prog 'and': failed to load: -13
libbpf: failed to load object 'arena_atomics'
libbpf: failed to load BPF skeleton 'arena_atomics': -13
test_arena_atomics:FAIL:arena atomics skeleton load unexpected error: -13 (errno 13)
#3 arena_atomics:FAIL
The reason of the failure is due to [2] where atomic{64,}_fetch_{and,or,xor}() are not
allowed by arena addresses.
Version 2 of the patch fixed the issue by using inline asm ([3]). But further discussion
suggested to find a way from source to generate locked insn which is more user
friendly. So in not-merged llvm patch ([4]), if relax memory ordering is used and
the return value is not used, locked insn could be generated.
So with llvm patch [4] to compile the bpf selftest, the following code
__c11_atomic_fetch_and(&and64_value, 0x011ull << 32, memory_order_relaxed);
is able to generate locked insn, hence fixing the selftest failure.
[1] llvm/llvm-project#106494
[2] d503a04 ("bpf: Add support for certain atomics in bpf_arena to x86 JIT")
[3] https://lore.kernel.org/bpf/20240803025928.4184433-1-yonghong.song@linux.dev/
[4] llvm/llvm-project#107343
Signed-off-by: Yonghong Song <yonghong.song@linux.dev>
kernel-patches-daemon-bpf bot
pushed a commit
to kernel-patches/bpf
that referenced
this pull request
Sep 10, 2024
llvm change [1] made a change such that __sync_fetch_and_{and,or,xor}()
will generate atomic_fetch_*() insns even if the return value is not used.
This is a deliberate choice to make sure barrier semantics are preserved
from source code to asm insn.
But the change in [1] caused arena_atomics selftest failure.
test_arena_atomics:PASS:arena atomics skeleton open 0 nsec
libbpf: prog 'and': BPF program load failed: Permission denied
libbpf: prog 'and': -- BEGIN PROG LOAD LOG --
arg#0 reference type('UNKNOWN ') size cannot be determined: -22
0: R1=ctx() R10=fp0
; if (pid != (bpf_get_current_pid_tgid() >> 32)) @ arena_atomics.c:87
0: (18) r1 = 0xffffc90000064000 ; R1_w=map_value(map=arena_at.bss,ks=4,vs=4)
2: (61) r6 = *(u32 *)(r1 +0) ; R1_w=map_value(map=arena_at.bss,ks=4,vs=4) R6_w=scalar(smin=0,smax=umax=0xffffffff,v
ar_off=(0x0; 0xffffffff))
3: (85) call bpf_get_current_pid_tgid#14 ; R0_w=scalar()
4: (77) r0 >>= 32 ; R0_w=scalar(smin=0,smax=umax=0xffffffff,var_off=(0x0; 0xffffffff))
5: (5d) if r0 != r6 goto pc+11 ; R0_w=scalar(smin=0,smax=umax=0xffffffff,var_off=(0x0; 0xffffffff)) R6_w=scalar(smin=0,smax=umax=0xffffffff,var_off=(0x0; 0x)
; __sync_fetch_and_and(&and64_value, 0x011ull << 32); @ arena_atomics.c:91
6: (18) r1 = 0x100000000060 ; R1_w=scalar()
8: (bf) r1 = addr_space_cast(r1, 0, 1) ; R1_w=arena
9: (18) r2 = 0x1100000000 ; R2_w=0x1100000000
11: (db) r2 = atomic64_fetch_and((u64 *)(r1 +0), r2)
BPF_ATOMIC stores into R1 arena is not allowed
processed 9 insns (limit 1000000) max_states_per_insn 0 total_states 0 peak_states 0 mark_read 0
-- END PROG LOAD LOG --
libbpf: prog 'and': failed to load: -13
libbpf: failed to load object 'arena_atomics'
libbpf: failed to load BPF skeleton 'arena_atomics': -13
test_arena_atomics:FAIL:arena atomics skeleton load unexpected error: -13 (errno 13)
#3 arena_atomics:FAIL
The reason of the failure is due to [2] where atomic{64,}_fetch_{and,or,xor}() are not
allowed by arena addresses.
Version 2 of the patch fixed the issue by using inline asm ([3]). But further discussion
suggested to find a way from source to generate locked insn which is more user
friendly. So in not-merged llvm patch ([4]), if relax memory ordering is used and
the return value is not used, locked insn could be generated.
So with llvm patch [4] to compile the bpf selftest, the following code
__c11_atomic_fetch_and(&and64_value, 0x011ull << 32, memory_order_relaxed);
is able to generate locked insn, hence fixing the selftest failure.
[1] llvm/llvm-project#106494
[2] d503a04 ("bpf: Add support for certain atomics in bpf_arena to x86 JIT")
[3] https://lore.kernel.org/bpf/20240803025928.4184433-1-yonghong.song@linux.dev/
[4] llvm/llvm-project#107343
Signed-off-by: Yonghong Song <yonghong.song@linux.dev>
kernel-patches-daemon-bpf-rc bot
pushed a commit
to kernel-patches/bpf-rc
that referenced
this pull request
Sep 10, 2024
llvm change [1] made a change such that __sync_fetch_and_{and,or,xor}()
will generate atomic_fetch_*() insns even if the return value is not used.
This is a deliberate choice to make sure barrier semantics are preserved
from source code to asm insn.
But the change in [1] caused arena_atomics selftest failure.
test_arena_atomics:PASS:arena atomics skeleton open 0 nsec
libbpf: prog 'and': BPF program load failed: Permission denied
libbpf: prog 'and': -- BEGIN PROG LOAD LOG --
arg#0 reference type('UNKNOWN ') size cannot be determined: -22
0: R1=ctx() R10=fp0
; if (pid != (bpf_get_current_pid_tgid() >> 32)) @ arena_atomics.c:87
0: (18) r1 = 0xffffc90000064000 ; R1_w=map_value(map=arena_at.bss,ks=4,vs=4)
2: (61) r6 = *(u32 *)(r1 +0) ; R1_w=map_value(map=arena_at.bss,ks=4,vs=4) R6_w=scalar(smin=0,smax=umax=0xffffffff,v
ar_off=(0x0; 0xffffffff))
3: (85) call bpf_get_current_pid_tgid#14 ; R0_w=scalar()
4: (77) r0 >>= 32 ; R0_w=scalar(smin=0,smax=umax=0xffffffff,var_off=(0x0; 0xffffffff))
5: (5d) if r0 != r6 goto pc+11 ; R0_w=scalar(smin=0,smax=umax=0xffffffff,var_off=(0x0; 0xffffffff)) R6_w=scalar(smin=0,smax=umax=0xffffffff,var_off=(0x0; 0x)
; __sync_fetch_and_and(&and64_value, 0x011ull << 32); @ arena_atomics.c:91
6: (18) r1 = 0x100000000060 ; R1_w=scalar()
8: (bf) r1 = addr_space_cast(r1, 0, 1) ; R1_w=arena
9: (18) r2 = 0x1100000000 ; R2_w=0x1100000000
11: (db) r2 = atomic64_fetch_and((u64 *)(r1 +0), r2)
BPF_ATOMIC stores into R1 arena is not allowed
processed 9 insns (limit 1000000) max_states_per_insn 0 total_states 0 peak_states 0 mark_read 0
-- END PROG LOAD LOG --
libbpf: prog 'and': failed to load: -13
libbpf: failed to load object 'arena_atomics'
libbpf: failed to load BPF skeleton 'arena_atomics': -13
test_arena_atomics:FAIL:arena atomics skeleton load unexpected error: -13 (errno 13)
#3 arena_atomics:FAIL
The reason of the failure is due to [2] where atomic{64,}_fetch_{and,or,xor}() are not
allowed by arena addresses.
Version 2 of the patch fixed the issue by using inline asm ([3]). But further discussion
suggested to find a way from source to generate locked insn which is more user
friendly. So in not-merged llvm patch ([4]), if relax memory ordering is used and
the return value is not used, locked insn could be generated.
So with llvm patch [4] to compile the bpf selftest, the following code
__c11_atomic_fetch_and(&and64_value, 0x011ull << 32, memory_order_relaxed);
is able to generate locked insn, hence fixing the selftest failure.
[1] llvm/llvm-project#106494
[2] d503a04 ("bpf: Add support for certain atomics in bpf_arena to x86 JIT")
[3] https://lore.kernel.org/bpf/20240803025928.4184433-1-yonghong.song@linux.dev/
[4] llvm/llvm-project#107343
Signed-off-by: Yonghong Song <yonghong.song@linux.dev>
kernel-patches-daemon-bpf bot
pushed a commit
to kernel-patches/bpf
that referenced
this pull request
Sep 11, 2024
llvm change [1] made a change such that __sync_fetch_and_{and,or,xor}()
will generate atomic_fetch_*() insns even if the return value is not used.
This is a deliberate choice to make sure barrier semantics are preserved
from source code to asm insn.
But the change in [1] caused arena_atomics selftest failure.
test_arena_atomics:PASS:arena atomics skeleton open 0 nsec
libbpf: prog 'and': BPF program load failed: Permission denied
libbpf: prog 'and': -- BEGIN PROG LOAD LOG --
arg#0 reference type('UNKNOWN ') size cannot be determined: -22
0: R1=ctx() R10=fp0
; if (pid != (bpf_get_current_pid_tgid() >> 32)) @ arena_atomics.c:87
0: (18) r1 = 0xffffc90000064000 ; R1_w=map_value(map=arena_at.bss,ks=4,vs=4)
2: (61) r6 = *(u32 *)(r1 +0) ; R1_w=map_value(map=arena_at.bss,ks=4,vs=4) R6_w=scalar(smin=0,smax=umax=0xffffffff,v
ar_off=(0x0; 0xffffffff))
3: (85) call bpf_get_current_pid_tgid#14 ; R0_w=scalar()
4: (77) r0 >>= 32 ; R0_w=scalar(smin=0,smax=umax=0xffffffff,var_off=(0x0; 0xffffffff))
5: (5d) if r0 != r6 goto pc+11 ; R0_w=scalar(smin=0,smax=umax=0xffffffff,var_off=(0x0; 0xffffffff)) R6_w=scalar(smin=0,smax=umax=0xffffffff,var_off=(0x0; 0x)
; __sync_fetch_and_and(&and64_value, 0x011ull << 32); @ arena_atomics.c:91
6: (18) r1 = 0x100000000060 ; R1_w=scalar()
8: (bf) r1 = addr_space_cast(r1, 0, 1) ; R1_w=arena
9: (18) r2 = 0x1100000000 ; R2_w=0x1100000000
11: (db) r2 = atomic64_fetch_and((u64 *)(r1 +0), r2)
BPF_ATOMIC stores into R1 arena is not allowed
processed 9 insns (limit 1000000) max_states_per_insn 0 total_states 0 peak_states 0 mark_read 0
-- END PROG LOAD LOG --
libbpf: prog 'and': failed to load: -13
libbpf: failed to load object 'arena_atomics'
libbpf: failed to load BPF skeleton 'arena_atomics': -13
test_arena_atomics:FAIL:arena atomics skeleton load unexpected error: -13 (errno 13)
#3 arena_atomics:FAIL
The reason of the failure is due to [2] where atomic{64,}_fetch_{and,or,xor}() are not
allowed by arena addresses.
Version 2 of the patch fixed the issue by using inline asm ([3]). But further discussion
suggested to find a way from source to generate locked insn which is more user
friendly. So in not-merged llvm patch ([4]), if relax memory ordering is used and
the return value is not used, locked insn could be generated.
So with llvm patch [4] to compile the bpf selftest, the following code
__c11_atomic_fetch_and(&and64_value, 0x011ull << 32, memory_order_relaxed);
is able to generate locked insn, hence fixing the selftest failure.
[1] llvm/llvm-project#106494
[2] d503a04 ("bpf: Add support for certain atomics in bpf_arena to x86 JIT")
[3] https://lore.kernel.org/bpf/20240803025928.4184433-1-yonghong.song@linux.dev/
[4] llvm/llvm-project#107343
Signed-off-by: Yonghong Song <yonghong.song@linux.dev>
kernel-patches-daemon-bpf-rc bot
pushed a commit
to kernel-patches/bpf-rc
that referenced
this pull request
Sep 11, 2024
llvm change [1] made a change such that __sync_fetch_and_{and,or,xor}()
will generate atomic_fetch_*() insns even if the return value is not used.
This is a deliberate choice to make sure barrier semantics are preserved
from source code to asm insn.
But the change in [1] caused arena_atomics selftest failure.
test_arena_atomics:PASS:arena atomics skeleton open 0 nsec
libbpf: prog 'and': BPF program load failed: Permission denied
libbpf: prog 'and': -- BEGIN PROG LOAD LOG --
arg#0 reference type('UNKNOWN ') size cannot be determined: -22
0: R1=ctx() R10=fp0
; if (pid != (bpf_get_current_pid_tgid() >> 32)) @ arena_atomics.c:87
0: (18) r1 = 0xffffc90000064000 ; R1_w=map_value(map=arena_at.bss,ks=4,vs=4)
2: (61) r6 = *(u32 *)(r1 +0) ; R1_w=map_value(map=arena_at.bss,ks=4,vs=4) R6_w=scalar(smin=0,smax=umax=0xffffffff,v
ar_off=(0x0; 0xffffffff))
3: (85) call bpf_get_current_pid_tgid#14 ; R0_w=scalar()
4: (77) r0 >>= 32 ; R0_w=scalar(smin=0,smax=umax=0xffffffff,var_off=(0x0; 0xffffffff))
5: (5d) if r0 != r6 goto pc+11 ; R0_w=scalar(smin=0,smax=umax=0xffffffff,var_off=(0x0; 0xffffffff)) R6_w=scalar(smin=0,smax=umax=0xffffffff,var_off=(0x0; 0x)
; __sync_fetch_and_and(&and64_value, 0x011ull << 32); @ arena_atomics.c:91
6: (18) r1 = 0x100000000060 ; R1_w=scalar()
8: (bf) r1 = addr_space_cast(r1, 0, 1) ; R1_w=arena
9: (18) r2 = 0x1100000000 ; R2_w=0x1100000000
11: (db) r2 = atomic64_fetch_and((u64 *)(r1 +0), r2)
BPF_ATOMIC stores into R1 arena is not allowed
processed 9 insns (limit 1000000) max_states_per_insn 0 total_states 0 peak_states 0 mark_read 0
-- END PROG LOAD LOG --
libbpf: prog 'and': failed to load: -13
libbpf: failed to load object 'arena_atomics'
libbpf: failed to load BPF skeleton 'arena_atomics': -13
test_arena_atomics:FAIL:arena atomics skeleton load unexpected error: -13 (errno 13)
#3 arena_atomics:FAIL
The reason of the failure is due to [2] where atomic{64,}_fetch_{and,or,xor}() are not
allowed by arena addresses.
Version 2 of the patch fixed the issue by using inline asm ([3]). But further discussion
suggested to find a way from source to generate locked insn which is more user
friendly. So in not-merged llvm patch ([4]), if relax memory ordering is used and
the return value is not used, locked insn could be generated.
So with llvm patch [4] to compile the bpf selftest, the following code
__c11_atomic_fetch_and(&and64_value, 0x011ull << 32, memory_order_relaxed);
is able to generate locked insn, hence fixing the selftest failure.
[1] llvm/llvm-project#106494
[2] d503a04 ("bpf: Add support for certain atomics in bpf_arena to x86 JIT")
[3] https://lore.kernel.org/bpf/20240803025928.4184433-1-yonghong.song@linux.dev/
[4] llvm/llvm-project#107343
Signed-off-by: Yonghong Song <yonghong.song@linux.dev>
Link: https://lore.kernel.org/r/20240909223431.1666305-1-yonghong.song@linux.dev
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
|
Hi, this commit broke some of our code, I am not sure whether this is intentional or not. It seems to me that anything using the return value of e.g. To reproduce: On the parent of this commit (57fe53c) the following test is working: With this commit (06c531e), the same test fails: Like I said, I'm not sure if this is intentional or not - we just switched to inline assembly and call |
|
@th0rex thanks for the reporting. The newer behavior in llvm20 is correct. The old one has some issues and some instructions supposed not support in v1/v2 but somehow we did it and their usage is not intentional. But unfortunately there is no way to do backport. See below for some details. You can see a compiler error will be emitted for cpu v1 if the return value is used. The same will be for v2. For v3, it will be okay since v3 enables 32-bit subregister and we allows new insn atomic_fetch_add. for cpu v1/v2, the atomic_fetch_add is not available. But for 64bit __sync_fetch_and_add(), it seems going through. So there is a discrepancy between 32bit and 64bit. But intention is for <= llvm19, for v1/v2, no return value should be used (using locked insn) and compiler should issue an error when it does use return value. For v3, it is okay to use return value (using atomic_fetch_add insn). llvm20 fixed this issue. So for 64bit insn, using inline asm with atomc_fetch_add() is the right choice if you intends to use cpu v1/v2. |
|
Thanks for the detailed explanation! This makes sense, we'll keep using inline assembly with |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
There are two commits in this pull request:
atomic_fetch_and_*()toatomic_<op>()conversion if return value is not used, but preserving__sync_fetch_and_add()to locked insn with cpu v1/v2.