Timed may_goto
Subsystems (6)
Kumar Kartikeya Dwivedi
maintainerv2 cover Timed may_goto
This series replaces the current implementation of cond_break, which uses the may_goto instruction, and counts 8 million iterations per stack frame, with an implementation based on sampling time locally on the CPU.
This is done to permit a longer time for a given loop per-program invocation. The accounting is still done per-stack frame, but the count is used to instead amortize the cost of the logic to sample and check the time spent since the start.
This is needed for expressing more complicated algorithms (spin locks, waiting loops, etc.) in BPF programs without false positive expiration of the loop. For instance, the plan is to make use of this for implementing spin locks for BPF arena [0].
For the loop as follows:
for (int i = 0;; i++) {}
Testing on a bare-metal Sapphire Rapids Intel server yields the following table (taking an average of 25 runs).
+-----------------------------+--------------+--------------+------------------+ | Loop type | Iterations | Time (ms) | Time/iter (ns) | +-----------------------------|--------------+--------------+------------------+ | may_goto | 8388608 | 3 | 0.36 | | timed_may_goto (count=65535)| 589674932 | 250 | 0.42 | | bpf_for | 8388608 | 10 | 1.19 | +-----------------------------+--------------+--------------+------------------+
Here, count is used to amortize the time sampling and checking logic.
Obviously, this is the limit of an empty loop. Given the complexity of the loop body, the time spent in the loop can be longer. Cancellations will address the task of imposing an upper bound on program runtime.
For now, the implementation only supports x86.
[0]: https://lore.kernel.org/bpf/[email protected]
Changelog:
----------
v1 -> v2
v1: https://lore.kernel.org/bpf/[email protected]
* Address comments from Alexei
* Use kernel comment style for new code.
* Remove p->count == 0 check in bpf_check_timed_may_goto.
* Add comments on AX as argument/retval calling convention.
* Add comments describing how the counting logic works.
* Use BPF_EMIT_CALL instead of open-coding instruction encoding.
* Change if ax != 1 goto pc+X condition to if ax != 0 goto pc+X.
Kumar Kartikeya Dwivedi (2):
bpf: Add verifier support for timed may_goto bpf, x86: Add x86 JIT support for timed may_goto
arch/x86/net/Makefile | 2 +- arch/x86/net/bpf_jit_comp.c | 5 ++ arch/x86/net/bpf_timed_may_goto.S | 52 ++++++++++++++ include/linux/bpf.h | 1 + include/linux/filter.h | 8 +++ kernel/bpf/core.c | 32 +++++++++ kernel/bpf/verifier.c | 70 ++++++++++++++++--- .../bpf/progs/verifier_bpf_fastcall.c | 58 ++++++++++++--- .../selftests/bpf/progs/verifier_may_goto_1.c | 34 ++++++++- 9 files changed, 241 insertions(+), 21 deletions(-) create mode 100644 arch/x86/net/bpf_timed_may_goto.S
base-commit: 7586e2169c77a444d235a98ac858272d3dcec16e
Signature · 1 line
2.43.5
Kumar Kartikeya Dwivedi
maintainerPATCH v2 1/2 bpf: Add verifier support for timed may_goto
Implement support in the verifier for replacing may_goto implementation from a counter-based approach to one which samples time on the local CPU to have a bigger loop bound.
We implement it by maintaining 16-bytes per-stack frame, and using 8 bytes for maintaining the count for amortizing time sampling, and 8 bytes for the starting timestamp. To minimize overhead, we need to avoid spilling and filling of registers around this sequence, so we push this cost into the time sampling function 'arch_bpf_timed_may_goto'. This is a JIT-specific wrapper around bpf_check_timed_may_goto which returns us the count to store into the stack through BPF_REG_AX. All caller-saved registers (r0-r5) are guaranteed to remain untouched.
The loop can be broken by returning count as 0, otherwise we dispatch into the function when the count drops to 0, and the runtime chooses to refresh it (by returning count as BPF_MAX_TIMED_LOOPS) or returning 0 and aborting the loop on next iteration.
Since the check for 0 is done right after loading the count from the stack, all subsequent cond_break sequences should immediately break as well, of the same loop or subsequent loops in the program.
We pass in the stack_depth of the count (and thus the timestamp, by adding 8 to it) to the arch_bpf_timed_may_goto call so that it can be passed in to bpf_check_timed_may_goto as an argument after r1 is saved, by adding the offset to r10/fp. This adjustment will be arch specific, and the next patch will introduce support for x86.
Note that depending on loop complexity, time spent in the loop can be more than the current limit (250 ms), but imposing an upper bound on program runtime is an orthogonal problem which will be addressed when program cancellations are supported.
The current time afforded by cond_break may not be enough for cases where BPF programs want to implement locking algorithms inline, and use cond_break as a promise to the verifier that they will eventually terminate.
Below are some benchmarking numbers on the time taken per-iteration for an empty loop that counts the number of iterations until cond_break fires. For comparison, we compare it against bpf_for/bpf_repeat which is another way to achieve the same number of spins (BPF_MAX_LOOPS). The hardware used for benchmarking was a Sapphire Rapids Intel server with performance governor enabled, mitigations were enabled.
+-----------------------------+--------------+--------------+------------------+ | Loop type | Iterations | Time (ms) | Time/iter (ns) | +-----------------------------|--------------+--------------+------------------+ | may_goto | 8388608 | 3 | 0.36 | | timed_may_goto (count=65535)| 589674932 | 250 | 0.42 | | bpf_for | 8388608 | 10 | 1.19 | +-----------------------------+--------------+--------------+------------------+
This gives a good approximation at low overhead while staying close to the current implementation.
--- include/linux/bpf.h | 1 + include/linux/filter.h | 8 +++++ kernel/bpf/core.c | 32 +++++++++++++++++++ kernel/bpf/verifier.c | 70 +++++++++++++++++++++++++++++++++++++----- 4 files changed, 103 insertions(+), 8 deletions(-)
include/linux/bpf.h +1 -0
@@ -1987,6 +1987,7 @@ struct bpf_array {*/ enum {BPF_MAX_LOOPS =8 *1024 *1024 , +BPF_MAX_TIMED_LOOPS =0xffff , };#define BPF_F_ACCESS_MASK (BPF_F_RDONLY | \
include/linux/filter.h +8 -0
@@ -669,6 +669,11 @@ struct bpf_prog_stats {struct u64_stats_sync syncp; }__aligned(2 * sizeof(u64)) ; +struct bpf_timed_may_goto { +u64 count; +u64 timestamp; +}; +struct sk_filter {refcount_t refcnt;struct rcu_head rcu;
@@ -1130,8 +1135,11 @@ bool bpf_jit_supports_ptr_xchg(void);bool bpf_jit_supports_arena (void );bool bpf_jit_supports_insn (struct bpf_insn *insn,bool in_arena);bool bpf_jit_supports_private_stack (void ); +bool bpf_jit_supports_timed_may_goto (void );u64 bpf_arch_uaddress_limit (void );void arch_bpf_stack_walk (bool (*consume_fn)(void *cookie,u64 ip,u64 sp,u64 bp),void *cookie); +u64 arch_bpf_timed_may_goto (void ); +u64 bpf_check_timed_may_goto (struct bpf_timed_may_goto *);bool bpf_helper_changes_pkt_data (enum bpf_func_id func_id);static inline bool bpf_dump_raw_ok (const struct cred *cred)
kernel/bpf/core.c +32 -0
@@ -3069,6 +3069,38 @@ void __weak arch_bpf_stack_walk(bool (*consume_fn)(void *cookie, u64 ip, u64 sp, { } +bool __weak bpf_jit_supports_timed_may_goto (void ) +{ +return false; +} + +u64 __weak arch_bpf_timed_may_goto (void ) +{ +return 0 ; +} + +u64 bpf_check_timed_may_goto (struct bpf_timed_may_goto *p) +{ +u64 time =ktime_get_mono_fast_ns (); + +/* +* Populate the timestamp for this stack frame, and refresh count. +*/ +if (!p->timestamp) { + p->timestamp = time; +return BPF_MAX_TIMED_LOOPS ; + } +/* +* Check if we've exhausted our time slice, and zero count. +*/ +if (time - p->timestamp >= (NSEC_PER_SEC /4 )) +return 0 ; +/* +* Refresh the count for the stack frame. +*/ +return BPF_MAX_TIMED_LOOPS ; +} +/* for configs without MMU or 32-bit */ __weak const struct bpf_map_ops arena_map_ops;__weak u64 bpf_arena_get_user_vm_start (struct bpf_arena *arena)
kernel/bpf/verifier.c +62 -8
99 lines not on this page · raw message
Signature · 1 line
2.43.5
Re: [PATCH bpf-next v2 1/2] bpf: Add verifier support for timed may_goto
Quoted 10 lines · Kumar Kartikeya Dwivedi
kernel/bpf/core.c +9
-0+u64 bpf_check_timed_may_goto (struct bpf_timed_may_goto *p) +{ +u64 time =ktime_get_mono_fast_ns (); + +/* +* Populate the timestamp for this stack frame, and refresh count. +*/ +if (!p->timestamp) { + p->timestamp = time;
kernel/bpf/core.c void __weak arch_bpf_stack_walk(bool (*consume_fn)(void *cookie, u64 ip, u64 sp, · line 3100
Quoted 10 lines
kernel/bpf/core.c +10
-0+return BPF_MAX_TIMED_LOOPS ; + } +/* +* Check if we've exhausted our time slice, and zero count. +*/ +if (time - p->timestamp >= (NSEC_PER_SEC /4 )) +return 0 ; +/* +* Refresh the count for the stack frame. +*/
I converted the comments back to single line comments.
Quoted 75 lines
kernel/bpf/core.c +3
-0+return BPF_MAX_TIMED_LOOPS ; +} +/* for configs without MMU or 32-bit */ __weak const struct bpf_map_ops arena_map_ops;__weak u64 bpf_arena_get_user_vm_start (struct bpf_arena *arena)kernel/bpf/verifier.c +47
-2@@ -21572,7 +21572,50 @@ static int do_misc_fixups(struct bpf_verifier_env *env) goto next_insn; } -if (is_may_goto_insn (insn)) { +if (is_may_goto_insn (insn) &&bpf_jit_supports_timed_may_goto ()) { +int stack_off_cnt = -stack_depth -16 ; + +/* +* Two 8 byte slots, depth-16 stores the count, and +* depth-8 stores the start timestamp of the loop. +* +* The starting value of count is BPF_MAX_TIMED_LOOPS +* (0xffff). Every iteration loads it and subs it by 1, +* until the value becomes 0 in AX (thus, 1 in stack), +* after which we call arch_bpf_timed_may_goto, which +* either sets AX to 0xffff to keep looping, or to 0 +* upon timeout. AX is then stored into the stack. In +* the next iteration, we either see 0 and break out, or +* continue iterating until the next time value is 0 +* after subtraction, rinse and repeat. +*/ + stack_depth_extra =16 ; + insn_buf[0 ] =BPF_LDX_MEM (BPF_DW ,BPF_REG_AX ,BPF_REG_10 , stack_off_cnt); +if (insn->off >=0 ) + insn_buf[1 ] =BPF_JMP_IMM (BPF_JEQ ,BPF_REG_AX ,0 , insn->off +5 ); +else + insn_buf[1 ] =BPF_JMP_IMM (BPF_JEQ ,BPF_REG_AX ,0 , insn->off -1 ); + insn_buf[2 ] =BPF_ALU64_IMM (BPF_SUB ,BPF_REG_AX ,1 ); + insn_buf[3 ] =BPF_JMP_IMM (BPF_JNE ,BPF_REG_AX ,0 ,2 ); +/* +* AX is used as an argument to pass in stack_off_cnt +* (to add to r10/fp), and also as the return value of +* the call to arch_bpf_timed_may_goto. +*/ + insn_buf[4 ] =BPF_MOV64_IMM (BPF_REG_AX , stack_off_cnt); + insn_buf[5 ] =BPF_EMIT_CALL (arch_bpf_timed_may_goto); + insn_buf[6 ] =BPF_STX_MEM (BPF_DW ,BPF_REG_10 ,BPF_REG_AX , stack_off_cnt); + cnt =7 ; + + new_prog =bpf_patch_insn_data (env, i + delta, insn_buf, cnt); +if (!new_prog) +return -ENOMEM ; + + delta += cnt -1 ; + env->prog = prog = new_prog; + insn = new_prog->insnsi + i + delta; + goto next_insn; + }else if (is_may_goto_insn (insn)) {int stack_off = -stack_depth -8 ; stack_depth_extra =8 ;@@ -22113,23 +22156,34 @@ static int do_misc_fixups(struct bpf_verifier_env *env) env->prog->aux->stack_depth = subprogs[0 ].stack_depth;for (i =0 ; i < env->subprog_cnt; i++) { +int delta =bpf_jit_supports_timed_may_goto () ?2 :1 ;int subprog_start = subprogs[i].start;int stack_slots = subprogs[i].stack_extra /8 ; +int slots = delta, cnt =0 ;if (!stack_slots)continue ; -if (stack_slots >1 ) { +/*
kernel/bpf/verifier.c static int do_misc_fixups(struct bpf_verifier_env *env) · line 22128
Quoted 10 lines
kernel/bpf/verifier.c +3
-3+* We need two slots in case timed may_goto is supported. +*/ +if (stack_slots > slots) {verbose (env,"verifier bug: stack_slots supports may_goto only\n" );return -EFAULT ; } -/* Add ST insn to subprog prologue to init extra stack */ - insn_buf[0 ] =BPF_ST_MEM (BPF_DW ,BPF_REG_FP , - -subprogs[i].stack_depth,BPF_MAX_LOOPS );
and here addedstack_depth = subprogs[i].stack_depth;
to reduce copy paste in below lines...
kernel/bpf/verifier.c static int do_misc_fixups(struct bpf_verifier_env *env) · line 22182
Quoted 9 lines
kernel/bpf/verifier.c +9
-0+if (bpf_jit_supports_timed_may_goto ()) { + insn_buf[cnt++] =BPF_ST_MEM (BPF_DW ,BPF_REG_FP , -subprogs[i].stack_depth, +BPF_MAX_TIMED_LOOPS ); + insn_buf[cnt++] =BPF_ST_MEM (BPF_DW ,BPF_REG_FP , -subprogs[i].stack_depth +8 ,0 ); + }else { +/* Add ST insn to subprog prologue to init extra stack */ + insn_buf[cnt++] =BPF_ST_MEM (BPF_DW ,BPF_REG_FP , -subprogs[i].stack_depth, +BPF_MAX_LOOPS ); + }
and applied.
Kumar Kartikeya Dwivedi
maintainerPATCH v2 2/2 bpf, x86: Add x86 JIT support for timed may_goto
In reply to Kumar Kartikeya DwivediImplement the arch_bpf_timed_may_goto function using inline assembly to have control over which registers are spilled, and use our special protocol of using BPF_REG_AX as an argument into the function, and as the return value when going back.
Emit call depth accounting for the call made from this stub, and ensure we don't have naked returns (when rethunk mitigations are enabled) by falling back to the RET macro (instead of retq). After popping all saved registers, the return address into the BPF program should be on top of the stack.
Since the JIT support is now enabled, ensure selftests which are checking the produced may_goto sequences do not break by adjusting them. Make sure we still test the old may_goto sequence on other architectures, while testing the new sequence on x86_64.
--- arch/x86/net/Makefile | 2 +- arch/x86/net/bpf_jit_comp.c | 5 ++ arch/x86/net/bpf_timed_may_goto.S | 52 +++++++++++++++++ .../bpf/progs/verifier_bpf_fastcall.c | 58 +++++++++++++++---- .../selftests/bpf/progs/verifier_may_goto_1.c | 34 ++++++++++- 5 files changed, 138 insertions(+), 13 deletions(-) create mode 100644 arch/x86/net/bpf_timed_may_goto.S
arch/x86/net/Makefile +1 -1
@@ -6,5 +6,5 @@ ifeq ($(CONFIG_X86_32),y) obj-$(CONFIG_BPF_JIT) += bpf_jit_comp32.o else - obj-$(CONFIG_BPF_JIT) += bpf_jit_comp.o + obj-$(CONFIG_BPF_JIT) += bpf_jit_comp.o bpf_timed_may_goto.o endif
arch/x86/net/bpf_jit_comp.c +5 -0
@@ -3791,3 +3791,8 @@ u64 bpf_arch_uaddress_limit(void) {return 0 ; } + +bool bpf_jit_supports_timed_may_goto (void ) +{ +return true; +}
arch/x86/net/bpf_timed_may_goto.S new file +52 -0
52 lines not on this page · raw message
tools/testing/selftests/bpf/progs/verifier_bpf_fastcall.c +48 -10
71 lines not on this page · raw message
tools/testing/selftests/bpf/progs/verifier_may_goto_1.c +32 -2
@@ -69,8 +69,38 @@ __naked void may_goto_batch_1(void) } SEC("raw_tp") -__description("may_goto batch with offsets 2/0") +__description("may_goto batch with offsets 2/0 - x86_64") __arch_x86_64 +__xlated("0: *(u64 *)(r10 -16) = 65535") +__xlated("1: *(u64 *)(r10 -8) = 0") +__xlated("2: r11 = *(u64 *)(r10 -16)") +__xlated("3: if r11 == 0x0 goto pc+6") +__xlated("4: r11 -= 1") +__xlated("5: if r11 != 0x0 goto pc+2") +__xlated("6: r11 = -16") +__xlated("7: call unknown") +__xlated("8: *(u64 *)(r10 -16) = r11") +__xlated("9: r0 = 1") +__xlated("10: r0 = 2") +__xlated("11: exit") +__success +__naked void may_goto_batch_2_x86_64 (void ) +{ + asmvolatile ( +".8byte %[may_goto1];" +".8byte %[may_goto3];" +"r0 = 1;" +"r0 = 2;" +"exit;" + : + : __imm_insn(may_goto1, BPF_RAW_INSN(BPF_JMP | BPF_JCOND, 0, 0, 2 /* offset */, 0)), +__imm_insn (may_goto3,BPF_RAW_INSN (BPF_JMP |BPF_JCOND ,0 ,0 ,0 /* offset */ ,0 )) + : __clobber_all); +} + +SEC("raw_tp") +__description("may_goto batch with offsets 2/0 - arm64") +__arch_arm64 __xlated("0: *(u64 *)(r10 -8) = 8388608") __xlated("1: r11 = *(u64 *)(r10 -8)")__xlated ("2: if r11 == 0x0 goto pc+3" )
@@ -80,7 +110,7 @@ __xlated("5: r0 = 1") __xlated("6: r0 = 2") __xlated("7: exit") __success -__naked void may_goto_batch_2 (void ) +__naked void may_goto_batch_2_arm64 (void ) { asmvolatile ( ".8byte %[may_goto1];"
Signature · 1 line
2.43.5
This series was applied to bpf/bpf-next.git (master) by Alexei Starovoitov <[email protected]>:
Hello:
This series was applied to bpf/bpf-next.git (master) by Alexei Starovoitov <[email protected]>:
Quoted 10 lines · On Mon, 3 Mar 2025 16:32:37 -0800 you wrote:
This series replaces the current implementation of
cond_break, which uses themay_gotoinstruction, and counts 8 million iterations per stack frame, with an implementation based on sampling time locally on the CPU.This is done to permit a longer time for a given loop per-program invocation. The accounting is still done per-stack frame, but the count is used to instead amortize the cost of the logic to sample and check the time spent since the start.
[...]
Here is the summary with links:
- [bpf-next,v2,1/2] bpf: Add verifier support for timed may_goto
https://git.kernel.org/bpf/bpf-next/c/13a664f46e34 - [bpf-next,v2,2/2] bpf, x86: Add x86 JIT support for timed may_goto https://git.kernel.org/bpf/bpf-next/c/2cb0a5215274
You are awesome, thank you!
Signature · 2 lines
Deet-doot-dot, I am a bot. https://korg.docs.kernel.org/patchwork/pwbot.html