bpf, arm64: use BPF prog pack allocator in BPF JIT
Subsystems (4)
v4 cover bpf, arm64: use BPF prog pack allocator in BPF JIT
BPF programs currently consume a page each on ARM64. For systems with many BPF programs, this adds significant pressure to instruction TLB. High iTLB pressure usually causes slow down for the whole system.
Song Liu introduced the BPF prog pack allocator[1] to mitigate the above issue. It packs multiple BPF programs into a single huge page. It is currently only enabled for the x86_64 BPF JIT.
This patch series enables the BPF prog pack allocator for the ARM64 BPF JIT.
====================================================
Performance Analysis of prog pack allocator on ARM64
====================================================
To test the performance of the BPF prog pack allocator on ARM64, a stresser tool[2] was built. This tool loads 8 BPF programs on the system and triggers 5 of them in an infinite loop by doing system calls.
The runner script starts 20 instances of the above which loads 8*20=160 BPF programs on the system, 5*20=100 of which are being constantly triggered.
In the above environment we try to build Python-3.8.4 and try to find different iTLB metrics for the compilation done by gcc-12.2.0.
The source code[3] is configured with the following command:
./configure --enable-optimizations --with-ensurepip=install
Then the runner script is executed with the following command:
./run.sh "perf stat -e ITLB_WALK,L1I_TLB,INST_RETIRED,iTLB-load-misses -a make -j32"
This builds Python while 160 BPF programs are loaded and 100 are being constantly triggered and measures iTLB related metrics.
The output of the above command is discussed below before and after enabling the BPF prog pack allocator.
The tests were run on qemu-system-aarch64 with 32 cpus, 4G memory, -machine virt, -cpu host, and -enable-kvm.
Results
-------
Before enabling prog pack allocator:
------------------------------------
Performance counter stats for 'system wide':
333278635 ITLB_WALK
6762692976558 L1I_TLB
25359571423901 INST_RETIRED
15824054789 iTLB-load-misses189.029769053 seconds time elapsed
After enabling prog pack allocator:
-----------------------------------
Performance counter stats for 'system wide':
190333544 ITLB_WALK
6712712386528 L1I_TLB
25278233304411 INST_RETIRED
5716757866 iTLB-load-misses185.392650561 seconds time elapsed
Improvements in metrics
-----------------------
Compilation time ---> 1.92% faster
iTLB-load-misses/Sec (Less is better) ---> 63.16% decrease
ITLB_WALK/1000 INST_RETIRED (Less is better) ---> 42.71% decrease
ITLB_Walk/L1I_TLB (Less is better) ---> 42.47% decrease
[1] https://lore.kernel.org/bpf/[email protected]/
[2] https://github.com/puranjaymohan/BPF-Allocator-Bench
[3] https://www.python.org/ftp/python/3.8.4/Python-3.8.4.tgz
Chanes in V3 => V4: Changes only in 3rd patch
1. Fix the I-cache maintenance: Clean the data cache and invalidate the i-Cache only *after* the instructions have been copied to the ROX region.
Chanes in V2 => V3: Changes only in 3rd patch
1. Set prog = orig_prog; in the failure path of bpf_jit_binary_pack_finalize() call.
2. Add comments explaining the usage of the offsets in the exception table.
Changes in v1 => v2:
1. Make the naming consistent in the 3rd patch:
ro_image and image
ro_header and header
ro_image_ptr and image_ptr
2. Use names dst/src in place of addr/opcode in second patch.
3. Add Acked-by: Song Liu <[email protected]> in 1st and 2nd patch.
Puranjay Mohan (3):
bpf: make bpf_prog_pack allocator portable
arm64: patching: Add aarch64_insn_copy()
bpf, arm64: use bpf_jit_binary_pack_alloc
arch/arm64/include/asm/patching.h | 1 + arch/arm64/kernel/patching.c | 39 ++++++++ arch/arm64/net/bpf_jit_comp.c | 145 +++++++++++++++++++++++++----- kernel/bpf/core.c | 8 +- 4 files changed, 165 insertions(+), 28 deletions(-)
Signature · 1 line
2.40.1
PATCH v4 1/3 bpf: make bpf_prog_pack allocator portable
The bpf_prog_pack allocator currently uses module_alloc() and module_memfree() to allocate and free memory. This is not portable because different architectures use different methods for allocating memory for BPF programs. Like ARM64 uses vmalloc()/vfree().
Use bpf_jit_alloc_exec() and bpf_jit_free_exec() for memory management in bpf_prog_pack allocator. Other architectures can override these with their implementation and will be able to use bpf_prog_pack directly.
---
Changes in V3 => v4
- No changes
Changes in V2 => V3:
- No changes
Changes in v1 => v2:
- No code changes.
- Added Acked-by: Song Liu <[email protected]>
kernel/bpf/core.c | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-)
kernel/bpf/core.c +4 -4
@@ -860,7 +860,7 @@ static struct bpf_prog_pack *alloc_new_pack(bpf_jit_fill_hole_t bpf_fill_ill_ins GFP_KERNEL);if (!pack)return NULL ; - pack->ptr =module_alloc (BPF_PROG_PACK_SIZE ); + pack->ptr =bpf_jit_alloc_exec (BPF_PROG_PACK_SIZE );if (!pack->ptr) {kfree (pack);return NULL ;
@@ -884,7 +884,7 @@ void *bpf_prog_pack_alloc(u32 size, bpf_jit_fill_hole_t bpf_fill_ill_insns)mutex_lock (&pack_mutex);if (size >BPF_PROG_PACK_SIZE ) { size =round_up (size,PAGE_SIZE ); - ptr =module_alloc (size); + ptr =bpf_jit_alloc_exec (size);if (ptr) {bpf_fill_ill_insns (ptr, size);set_vm_flush_reset_perms (ptr);
@@ -922,7 +922,7 @@ void bpf_prog_pack_free(struct bpf_binary_header *hdr)mutex_lock (&pack_mutex);if (hdr->size >BPF_PROG_PACK_SIZE ) { -module_memfree (hdr); +bpf_jit_free_exec (hdr); goto out; }
@@ -946,7 +946,7 @@ void bpf_prog_pack_free(struct bpf_binary_header *hdr)if (bitmap_find_next_zero_area (pack->bitmap,BPF_PROG_CHUNK_COUNT ,0 ,BPF_PROG_CHUNK_COUNT ,0 ) ==0 ) {list_del (&pack->list); -module_memfree (pack->ptr); +bpf_jit_free_exec (pack->ptr);kfree (pack); } out:
Signature · 1 line
2.40.1
PATCH v4 2/3 arm64: patching: Add aarch64_insn_copy()
In reply to Puranjay MohanThis will be used by BPF JIT compiler to dump JITed binary to a RX huge page, and thus allow multiple BPF programs sharing the a huge (2MB) page.
The bpf_prog_pack allocator that implements the above feature allocates a RX/RW buffer pair. The JITed code is written to the RW buffer and then this function will be used to copy the code from RW to RX buffer.
---
Changes in v3 => v4
- No changes
Changes in v2 => v3
- No changes
Changes in v1 => v2:
- Renamed addr/opcode to dst/src.
- Added Acked-by: Song Liu <[email protected]>
arch/arm64/include/asm/patching.h | 1 + arch/arm64/kernel/patching.c | 39 +++++++++++++++++++++++++++++++ 2 files changed, 40 insertions(+)
arch/arm64/include/asm/patching.h +1 -0
@@ -8,6 +8,7 @@ int aarch64_insn_read(void *addr, u32 *insnp);int aarch64_insn_write (void *addr,u32 insn);int aarch64_insn_write_literal_u64 (void *addr,u64 val); +void *aarch64_insn_copy (void *dst,const void *src,size_t len);int aarch64_insn_patch_text_nosync (void *addr,u32 insn);int aarch64_insn_patch_text (void *addrs[],u32 insns[],int cnt);
arch/arm64/kernel/patching.c +39 -0
@@ -105,6 +105,45 @@ noinstr int aarch64_insn_write_literal_u64(void *addr, u64 val)return ret; } +/** +* aarch64_insn_copy - Copy instructions into (an unused part of) RX memory +* @dst: address to modify +* @src: source of the copy +* @len: length to copy +* +* Useful for JITs to dump new code blocks into unused regions of RX memory. +*/ +noinstr void *aarch64_insn_copy(void *dst, const void *src, size_t len) +{ +unsigned long flags; +size_t patched =0 ; +size_t size; +void *waddr; +void *ptr; +int ret; + +raw_spin_lock_irqsave (&patch_lock, flags); + +while (patched < len) { + ptr = dst + patched; + size =min_t (size_t,PAGE_SIZE -offset_in_page (ptr), + len - patched); + + waddr =patch_map (ptr,FIX_TEXT_POKE0 ); + ret =copy_to_kernel_nofault (waddr, src + patched, size); +patch_unmap (FIX_TEXT_POKE0 ); + +if (ret <0 ) { +raw_spin_unlock_irqrestore (&patch_lock, flags); +return NULL ; + } + patched += size; + } +raw_spin_unlock_irqrestore (&patch_lock, flags); + +return dst; +} +int __kprobes aarch64_insn_patch_text_nosync (void *addr,u32 insn) {u32 *tp = addr;
Signature · 1 line
2.40.1
PATCH v4 3/3 bpf, arm64: use bpf_jit_binary_pack_alloc
In reply to Puranjay MohanUse bpf_jit_binary_pack_alloc for memory management of JIT binaries in ARM64 BPF JIT. The bpf_jit_binary_pack_alloc creates a pair of RW and RX buffers. The JIT writes the program into the RW buffer. When the JIT is done, the program is copied to the final RX buffer
with bpf_jit_binary_pack_finalize.
Implement bpf_arch_text_copy() and bpf_arch_text_invalidate() for ARM64 JIT as these functions are required by bpf_jit_binary_pack allocator.
---
Changes in v3 => v4:
- Fix the cache maintenance by cleaning the data cache and invalidating the i-Cache only *after* copying the instructions to the ROX region.
Changes in v2 => v3:
- Set prog = orig_prog; in the failure path of bpf_jit_binary_pack_finalize() call.
- Add comments explaining the usage of the offsets in the exception table.
Changes in v1 => v2:
- Made the naming of ro_ prefix consistent.
Now image/header/image_ptr are read/write and
ro_image/ro_header/ro_image_ptr are read-only.
arch/arm64/net/bpf_jit_comp.c | 145 ++++++++++++++++++++++++++++------ 1 file changed, 121 insertions(+), 24 deletions(-)
arch/arm64/net/bpf_jit_comp.c +121 -24
@@ -76,6 +76,7 @@ struct jit_ctx {int *offset;int exentry_idx;__le32 *image; +__le32 *ro_image;u32 stack_size;int fpb_offset; };
@@ -205,6 +206,20 @@ static void jit_fill_hole(void *area, unsigned int size) *ptr++ =cpu_to_le32 (AARCH64_BREAK_FAULT ); } +int bpf_arch_text_invalidate (void *dst,size_t len) +{ +__le32 *ptr; +int ret; + +for (ptr = dst; len >=sizeof (u32); len -=sizeof (u32)) { + ret =aarch64_insn_patch_text_nosync (ptr++,AARCH64_BREAK_FAULT ); +if (ret) +return ret; + } + +return 0 ; +} +static inline int epilogue_offset (const struct jit_ctx *ctx) {int to = ctx->epilogue_offset;
@@ -701,7 +716,8 @@ static int add_exception_handler(const struct bpf_insn *insn, struct jit_ctx *ctx, int dst_reg) { -off_t offset; +off_t ins_offset; +off_t fixup_offset;unsigned long pc;struct exception_table_entry *ex;
@@ -717,12 +733,17 @@ static int add_exception_handler(const struct bpf_insn *insn,return -EINVAL ; ex = &ctx->prog->aux->extable[ctx->exentry_idx]; - pc = (unsigned long )&ctx->image[ctx->idx -1 ]; + pc = (unsigned long )&ctx->ro_image[ctx->idx -1 ]; - offset = pc - (long )&ex->insn; -if (WARN_ON_ONCE (offset >=0 || offset <INT_MIN )) +/* +* This is the relative offset of the instruction that may fault from +* the exception table itself. This will be written to the exception +* table and if this instruction faults, the destination register will +* be set to '0' and the execution will jump to the next instruction. +*/ + ins_offset = pc - (long )&ex->insn; +if (WARN_ON_ONCE (ins_offset >=0 || ins_offset <INT_MIN ))return -ERANGE ; - ex->insn = offset;/* * Since the extable follows the program, the fixup offset is always
@@ -731,12 +752,25 @@ static int add_exception_handler(const struct bpf_insn *insn,* bits. We don't need to worry about buildtime or runtime sort * modifying the upper bits because the table is already sorted, and * isn't part of the main exception table. +* +* The fixup_offset is set to the next instruction from the instruction +* that may fault. The execution will jump to this after handling the +* fault. */ - offset = (long )&ex->fixup - (pc +AARCH64_INSN_SIZE ); -if (!FIELD_FIT (BPF_FIXUP_OFFSET_MASK , offset)) + fixup_offset = (long )&ex->fixup - (pc +AARCH64_INSN_SIZE ); +if (!FIELD_FIT (BPF_FIXUP_OFFSET_MASK , fixup_offset))return -ERANGE ; - ex->fixup =FIELD_PREP (BPF_FIXUP_OFFSET_MASK , offset) | +/* +* The offsets above have been calculated using the RO buffer but we +* need to use the R/W buffer for writes. +* switch ex to rw buffer for writing. +*/ + ex = (void *)ctx->image + ((void *)ex - (void *)ctx->ro_image); + + ex->insn = ins_offset; + + ex->fixup =FIELD_PREP (BPF_FIXUP_OFFSET_MASK , fixup_offset) |FIELD_PREP (BPF_FIXUP_REG_MASK , dst_reg); ex->type =EX_TYPE_BPF ;
@@ -1446,7 +1480,8 @@ static inline void bpf_flush_icache(void *start, void *end)struct arm64_jit_data {struct bpf_binary_header *header; -u8 *image; +u8 *ro_image; +struct bpf_binary_header *ro_header;struct jit_ctx ctx; };
@@ -1455,12 +1490,14 @@ struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)int image_size, prog_size, extable_size, extable_align, extable_offset;struct bpf_prog *tmp, *orig_prog = prog;struct bpf_binary_header *header; +struct bpf_binary_header *ro_header;struct arm64_jit_data *jit_data;bool was_classic =bpf_prog_was_classic (prog);bool tmp_blinded = false;bool extra_pass = false;struct jit_ctx ctx;u8 *image_ptr; +u8 *ro_image_ptr;if (!prog->jit_requested)return orig_prog;
@@ -1487,8 +1524,11 @@ struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog) }if (jit_data->ctx.offset) { ctx = jit_data->ctx; - image_ptr = jit_data->image; + ro_image_ptr = jit_data->ro_image; + ro_header = jit_data->ro_header; header = jit_data->header; + image_ptr = (void *)header + ((void *)ro_image_ptr + - (void *)ro_header); extra_pass = true; prog_size =sizeof (u32) * ctx.idx; goto skip_init_ctx;
@@ -1533,18 +1573,27 @@ struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)/* also allocate space for plt target */ extable_offset =round_up (prog_size +PLT_TARGET_SIZE , extable_align); image_size = extable_offset + extable_size; - header =bpf_jit_binary_alloc (image_size, &image_ptr, -sizeof (u32), jit_fill_hole); -if (header ==NULL ) { + ro_header =bpf_jit_binary_pack_alloc (image_size, &ro_image_ptr, +sizeof (u32), &header, &image_ptr, + jit_fill_hole); +if (!ro_header) { prog = orig_prog; goto out_off; }/* 2. Now, the actual pass. */ +/* +* Use the image(RW) for writing the JITed instructions. But also save +* the ro_image(RX) for calculating the offsets in the image. The RW +* image will be later copied to the RX image from where the program +* will run. The bpf_jit_binary_pack_finalize() will do this copy in the +* final step. +*/ ctx.image = (__le32 *)image_ptr; + ctx.ro_image = (__le32 *)ro_image_ptr;if (extable_size) - prog->aux->extable = (void *)image_ptr + extable_offset; + prog->aux->extable = (void *)ro_image_ptr + extable_offset; skip_init_ctx: ctx.idx =0 ; ctx.exentry_idx =0 ;
@@ -1552,9 +1601,8 @@ struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)build_prologue (&ctx, was_classic);if (build_body (&ctx, extra_pass)) { -bpf_jit_binary_free (header); prog = orig_prog; - goto out_off; + goto out_free_hdr; }build_epilogue (&ctx);
@@ -1562,34 +1610,44 @@ struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)/* 3. Extra pass to validate JITed code. */ if (validate_ctx (&ctx)) { -bpf_jit_binary_free (header); prog = orig_prog; - goto out_off; + goto out_free_hdr; }/* And we're done. */ if (bpf_jit_enable >1 )bpf_jit_dump (prog->len, prog_size,2 , ctx.image); -bpf_flush_icache (header, ctx.image + ctx.idx); -if (!prog->is_func || extra_pass) {if (extra_pass && ctx.idx != jit_data->ctx.idx) {pr_err_once ("multi-func JIT bug %d != %d\n" , ctx.idx, jit_data->ctx.idx); -bpf_jit_binary_free (header); prog->bpf_func =NULL ; prog->jited =0 ; prog->jited_len =0 ; + goto out_free_hdr; + } +if (WARN_ON (bpf_jit_binary_pack_finalize (prog, ro_header, + header))) { +/* ro_header has been freed */ + ro_header =NULL ; + prog = orig_prog; goto out_off; } -bpf_jit_binary_lock_ro (header); +/* +* The instructions have now been copied to the ROX region from +* where they will execute. Now the data cache has to be cleaned to +* the PoU and the I-cache has to be invalidated for the VAs. +*/ +bpf_flush_icache (ro_header, ctx.ro_image + ctx.idx); }else { jit_data->ctx = ctx; - jit_data->image = image_ptr; + jit_data->ro_image = ro_image_ptr; jit_data->header = header; + jit_data->ro_header = ro_header; } - prog->bpf_func = (void *)ctx.image; + + prog->bpf_func = (void *)ctx.ro_image; prog->jited =1 ; prog->jited_len = prog_size;
@@ -1610,6 +1668,14 @@ struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)bpf_jit_prog_release_other (prog, prog == orig_prog ? tmp : orig_prog);return prog; + +out_free_hdr: +if (header) { +bpf_arch_text_copy (&ro_header->size, &header->size, +sizeof (header->size)); +bpf_jit_binary_pack_free (ro_header, header); + } + goto out_off; }bool bpf_jit_supports_kfunc_call (void )
@@ -1617,6 +1683,13 @@ bool bpf_jit_supports_kfunc_call(void)return true; } +void *bpf_arch_text_copy (void *dst,void *src,size_t len) +{ +if (!aarch64_insn_copy (dst, src, len)) +return ERR_PTR (-EINVAL ); +return dst; +} +u64 bpf_jit_alloc_exec_limit (void ) {return VMALLOC_END -VMALLOC_START ;
@@ -2221,3 +2294,27 @@ int bpf_arch_text_poke(void *ip, enum bpf_text_poke_type poke_type,return ret; } + +void bpf_jit_free (struct bpf_prog *prog) +{ +if (prog->jited) { +struct arm64_jit_data *jit_data = prog->aux->jit_data; +struct bpf_binary_header *hdr; + +/* +* If we fail the final pass of JIT (from jit_subprogs), +* the program may not be finalized yet. Call finalize here +* before freeing it. +*/ +if (jit_data) { +bpf_jit_binary_pack_finalize (prog, jit_data->ro_header, + jit_data->header); +kfree (jit_data); + } + hdr =bpf_jit_binary_pack_hdr (prog); +bpf_jit_binary_pack_free (hdr,NULL ); +WARN_ON_ONCE (!bpf_prog_kallsyms_verify_off (prog)); + } + +bpf_prog_unlock_free (prog); +}
Signature · 1 line
2.40.1
Re: [PATCH bpf-next v4 3/3] bpf, arm64: use bpf_jit_binary_pack_alloc
Hi Everyone,
[+cc Björn]
Quoted 12 lines · Puranjay Mohan
Use
bpf_jit_binary_pack_allocfor memory management of JIT binaries in ARM64 BPF JIT. Thebpf_jit_binary_pack_alloccreates a pair of RW and RX buffers. The JIT writes the program into the RW buffer. When the JIT is done, the program is copied to the final RX buffer
withbpf_jit_binary_pack_finalize.Implement
bpf_arch_text_copy()andbpf_arch_text_invalidate()for ARM64 JIT as these functions are required bybpf_jit_binary_packallocator.Signed-off-by Puranjay Mohan <[email protected]>Acked-by Song Liu <[email protected]>
[...]
Quoted 4 lines
arch/arm64/net/bpf_jit_comp.c +4
-0+int bpf_arch_text_invalidate (void *dst,size_t len) +{ +__le32 *ptr; +int ret;
arch/arm64/net/bpf_jit_comp.c static void jit_fill_hole(void *area, unsigned int size) · line 222
Quoted 10 lines
arch/arm64/net/bpf_jit_comp.c +10
-0+ +for (ptr = dst; len >=sizeof (u32); len -=sizeof (u32)) { + ret =aarch64_insn_patch_text_nosync (ptr++,AARCH64_BREAK_FAULT ); +if (ret) +return ret; + } + +return 0 ; +} +
While testing the same patch for riscv bpf jit, Björn found that ./test_tag is taking a lot of
time to complete. He found that bpf_arch_text_invalidate() is calling patch_text_nosync() in RISCV
and aarch64_insn_patch_text_nosync() here in ARM64. Both of the implementations call these functions
in a loop for each word. The problem is that every call to patch_text_nosync()/aarch64_insn_patch_text_nosync()
would clean the cache. This will make this function(bpf_arch_text_invalidate) really slow.
I did some testing using the vmtest.sh script on ARM64 with and without the patches, here are the results:
With Prog Pack patches applied:
=========================
root@(none):/# time ./root/bpf/test_tag test_tag: OK (40945 tests)
real 3m2.001s user 0m1.644s sys 2m40.132s
Without Prog Pack:
===============
root@(none):/# time ./root/bpf/test_tag test_tag: OK (40945 tests)
real 0m26.809s user 0m1.591s sys 0m24.106s
As you can see the current implementation of bpf_arch_text_invalidate() is really slow. I need to implement a new function: aarch64_insn_set_text_nosync() and use it in bpf_arch_text_invalidate()
rather than calling aarch64_insn_patch_text_nosync() in a loop.
In the longer run, it would be really helpful if we have a standard text_poke API like x86 in every architecture.
Thanks,
Puranjay
Daniel Borkmann
maintainerRe: [PATCH bpf-next v4 3/3] bpf, arm64: use bpf_jit_binary_pack_alloc
Quoted 69 lines · Puranjay Mohan
Hi Everyone,
[+cc Björn]
Quoted 12 lines · Puranjay Mohan
Use
bpf_jit_binary_pack_allocfor memory management of JIT binaries in ARM64 BPF JIT. Thebpf_jit_binary_pack_alloccreates a pair of RW and RX buffers. The JIT writes the program into the RW buffer. When the JIT is done, the program is copied to the final RX buffer
withbpf_jit_binary_pack_finalize.Implement
bpf_arch_text_copy()andbpf_arch_text_invalidate()for ARM64 JIT as these functions are required bybpf_jit_binary_packallocator.Signed-off-by Puranjay Mohan <[email protected]>Acked-by Song Liu <[email protected]>[...]
Quoted 4 lines
arch/arm64/net/bpf_jit_comp.c +4
-0+int bpf_arch_text_invalidate (void *dst,size_t len) +{ +__le32 *ptr; +int ret;arch/arm64/net/bpf_jit_comp.c static void jit_fill_hole(void *area, unsigned int size) · line 222
Quoted 10 lines
arch/arm64/net/bpf_jit_comp.c +10
-0+ +for (ptr = dst; len >=sizeof (u32); len -=sizeof (u32)) { + ret =aarch64_insn_patch_text_nosync (ptr++,AARCH64_BREAK_FAULT ); +if (ret) +return ret; + } + +return 0 ; +} +While testing the same patch for riscv bpf jit, Björn found that ./test_tag is taking a lot of
time to complete. He found thatbpf_arch_text_invalidate()is callingpatch_text_nosync()in RISCV
andaarch64_insn_patch_text_nosync()here in ARM64. Both of the implementations call these functions
in a loop for each word. The problem is that every call to patch_text_nosync()/aarch64_insn_patch_text_nosync()
would clean the cache. This will make thisfunction(bpf_arch_text_invalidate)really slow.I did some testing using the
vmtest.shscript on ARM64 with and without the patches, here are the results:With Prog Pack patches applied:
=========================root@(none):/# time ./root/bpf/test_tag test_tag: OK (40945 tests)real 3m2.001s user 0m1.644s sys 2m40.132sWithout Prog Pack:
===============root@(none):/# time ./root/bpf/test_tag test_tag: OK (40945 tests)real 0m26.809s user 0m1.591s sys 0m24.106s
Quoted 5 lines
As you can see the current implementation of
bpf_arch_text_invalidate()is really slow. I need to implement a new function:aarch64_insn_set_text_nosync()and use it inbpf_arch_text_invalidate()
rather than callingaarch64_insn_patch_text_nosync()in a loop.
Ok, thanks for looking into this. I'm going to toss the current version from patchwork in that case and will wait for a v5 from your side.
Quoted 2 lines
In the longer run, it would be really helpful if we have a standard
text_pokeAPI like x86 in every architecture.
Likely independent of this series, but in case you see potential to consolidate generic pieces, it might be worth a look after the arm64/riscv bits landed.
Thanks for working on this, Puranjay!
Quoted 101 lines · Puranjay Mohan
BPF programs currently consume a page each on ARM64. For systems with many BPF programs, this adds significant pressure to instruction TLB. High iTLB pressure usually causes slow down for the whole system.
Song Liu introduced the BPF prog pack allocator[1] to mitigate the above issue. It packs multiple BPF programs into a single huge page. It is currently only enabled for the
x86_64BPF JIT.This patch series enables the BPF prog pack allocator for the ARM64 BPF JIT.
====================================================
Performance Analysis of prog pack allocator on ARM64
====================================================To test the performance of the BPF prog pack allocator on ARM64, a stresser tool[2] was built. This tool loads 8 BPF programs on the system and triggers 5 of them in an infinite loop by doing system calls.
The runner script starts 20 instances of the above which loads 8*20=160 BPF programs on the system, 5*20=100 of which are being constantly triggered.
In the above environment we try to build Python-3.8.4 and try to find different iTLB metrics for the compilation done by gcc-12.2.0.
The source code[3] is configured with the following command:
./configure --enable-optimizations --with-ensurepip=installThen the runner script is executed with the following command:
./run.sh "perf stat -e ITLB_WALK,L1I_TLB,INST_RETIRED,iTLB-load-misses -a make -j32"This builds Python while 160 BPF programs are loaded and 100 are being constantly triggered and measures iTLB related metrics.
The output of the above command is discussed below before and after enabling the BPF prog pack allocator.
The tests were run on qemu-system-aarch64 with 32 cpus, 4G memory, -machine virt, -cpu host, and -enable-kvm.
Results
-------Before enabling prog pack allocator:
------------------------------------Performance counter stats for 'system wide':
333278635 ITLB_WALK 6762692976558 L1I_TLB 25359571423901 INST_RETIRED 15824054789 iTLB-load-misses189.029769053 seconds time elapsedAfter enabling prog pack allocator:
-----------------------------------Performance counter stats for 'system wide':
190333544 ITLB_WALK 6712712386528 L1I_TLB 25278233304411 INST_RETIRED 5716757866 iTLB-load-misses185.392650561 seconds time elapsedImprovements in metrics
-----------------------Compilation time ---> 1.92% faster
iTLB-load-misses/Sec (Less is better) ---> 63.16% decrease
ITLB_WALK/1000INST_RETIRED(Less is better) ---> 42.71% decrease
ITLB_Walk/L1I_TLB (Less is better) ---> 42.47% decrease[1] https://lore.kernel.org/bpf/[email protected]/
[2] https://github.com/puranjaymohan/BPF-Allocator-Bench
[3] https://www.python.org/ftp/python/3.8.4/Python-3.8.4.tgzChanes in V3 => V4: Changes only in 3rd patch
1. Fix the I-cache maintenance: Clean the data cache and invalidate the i-Cache only *after* the instructions have been copied to the ROX region.Chanes in V2 => V3: Changes only in 3rd patch
1. Set prog =orig_prog; in the failure path ofbpf_jit_binary_pack_finalize()call.
2. Add comments explaining the usage of the offsets in the exception table.Changes in v1 => v2:
1. Make the naming consistent in the 3rd patch:
ro_imageand image
ro_headerand header
ro_image_ptrandimage_ptr
2. Use names dst/src in place of addr/opcode in second patch.
3. Add Acked-by: Song Liu <[email protected]> in 1st and 2nd patch.Puranjay Mohan (3):
bpf: makebpf_prog_packallocator portable
arm64: patching: Addaarch64_insn_copy()
bpf, arm64: usebpf_jit_binary_pack_alloc
Quoted 10 lines
arch/arm64/include/asm/patching.h | 1 + arch/arm64/kernel/patching.c | 39 ++++++++ arch/arm64/net/bpf_jit_comp.c | 145 +++++++++++++++++++++++++----- kernel/bpf/core.c | 8 +- 4 files changed, 165 insertions(+), 28 deletions(-)--
2.40.1
FWIW
Thanks for this Puranjay!
Daniel Borkmann
maintainerHi Mark,
Quoted 74 lines · Puranjay Mohan
BPF programs currently consume a page each on ARM64. For systems with many BPF programs, this adds significant pressure to instruction TLB. High iTLB pressure usually causes slow down for the whole system.
Song Liu introduced the BPF prog pack allocator[1] to mitigate the above issue. It packs multiple BPF programs into a single huge page. It is currently only enabled for the
x86_64BPF JIT.This patch series enables the BPF prog pack allocator for the ARM64 BPF JIT.
====================================================
Performance Analysis of prog pack allocator on ARM64
====================================================To test the performance of the BPF prog pack allocator on ARM64, a stresser tool[2] was built. This tool loads 8 BPF programs on the system and triggers 5 of them in an infinite loop by doing system calls.
The runner script starts 20 instances of the above which loads 8*20=160 BPF programs on the system, 5*20=100 of which are being constantly triggered.
In the above environment we try to build Python-3.8.4 and try to find different iTLB metrics for the compilation done by gcc-12.2.0.
The source code[3] is configured with the following command:
./configure --enable-optimizations --with-ensurepip=installThen the runner script is executed with the following command:
./run.sh "perf stat -e ITLB_WALK,L1I_TLB,INST_RETIRED,iTLB-load-misses -a make -j32"This builds Python while 160 BPF programs are loaded and 100 are being constantly triggered and measures iTLB related metrics.
The output of the above command is discussed below before and after enabling the BPF prog pack allocator.
The tests were run on qemu-system-aarch64 with 32 cpus, 4G memory, -machine virt, -cpu host, and -enable-kvm.
Results
-------Before enabling prog pack allocator:
------------------------------------Performance counter stats for 'system wide':
333278635 ITLB_WALK 6762692976558 L1I_TLB 25359571423901 INST_RETIRED 15824054789 iTLB-load-misses189.029769053 seconds time elapsedAfter enabling prog pack allocator:
-----------------------------------Performance counter stats for 'system wide':
190333544 ITLB_WALK 6712712386528 L1I_TLB 25278233304411 INST_RETIRED 5716757866 iTLB-load-misses185.392650561 seconds time elapsedImprovements in metrics
-----------------------Compilation time ---> 1.92% faster
iTLB-load-misses/Sec (Less is better) ---> 63.16% decrease
ITLB_WALK/1000INST_RETIRED(Less is better) ---> 42.71% decrease
ITLB_Walk/L1I_TLB (Less is better) ---> 42.47% decrease
Quoted 7 lines
[1] https://lore.kernel.org/bpf/[email protected]/
[2] https://github.com/puranjaymohan/BPF-Allocator-Bench
[3] https://www.python.org/ftp/python/3.8.4/Python-3.8.4.tgzChanes in V3 => V4: Changes only in 3rd patch
1. Fix the I-cache maintenance: Clean the data cache and invalidate the i-Cache only *after* the instructions have been copied to the ROX region.
If you get a chance to take another look at the v4 changes from Puranjay and in case they look good to you reply with an Ack, that would be great.
Thanks,
Daniel
Mark Rutland
reviewerQuoted 1 line · Daniel Borkmann
Hi Mark,
Hi Daniel,
Quoted 5 lines · Puranjay Mohan
Quoted 4 lines
BPF programs currently consume a page each on ARM64. For systems with many BPF programs, this adds significant pressure to instruction TLB. High iTLB pressure usually causes slow down for the whole system.
Quoted 8 lines
Quoted 5 lines
Song Liu introduced the BPF prog pack allocator[1] to mitigate the above issue. It packs multiple BPF programs into a single huge page. It is currently only enabled for the
x86_64BPF JIT.This patch series enables the BPF prog pack allocator for the ARM64 BPF JIT.
If you get a chance to take another look at the v4 changes from Puranjay and in case they look good to you reply with an Ack, that would be great.
Sure -- this is on my queue of things to look at; it might just take me a few days to get the time to give this a proper look.
Thanks,
Mark.
Daniel Borkmann
maintainerOn 7/3/23 7:15 PM, Mark Rutland wrote:
[...]
Quoted 9 lines
Quoted 9 lines · Puranjay Mohan
Quoted 8 lines
BPF programs currently consume a page each on ARM64. For systems with many BPF programs, this adds significant pressure to instruction TLB. High iTLB pressure usually causes slow down for the whole system.
Song Liu introduced the BPF prog pack allocator[1] to mitigate the above issue. It packs multiple BPF programs into a single huge page. It is currently only enabled for the
x86_64BPF JIT.
Quoted 7 lines
Quoted 4 lines
Quoted 1 line
This patch series enables the BPF prog pack allocator for the ARM64 BPF JIT.
If you get a chance to take another look at the v4 changes from Puranjay and in case they look good to you reply with an Ack, that would be great.
Sure -- this is on my queue of things to look at; it might just take me a few days to get the time to give this a proper look.
Awesome, thanks Mark!
Hi Mark,
Quoted 17 lines · Mark Rutland
Quoted 1 line · Daniel Borkmann
Hi Mark,
Hi Daniel,
Quoted 10 lines · Puranjay Mohan
Quoted 9 lines
BPF programs currently consume a page each on ARM64. For systems with many BPF programs, this adds significant pressure to instruction TLB. High iTLB pressure usually causes slow down for the whole system.
Song Liu introduced the BPF prog pack allocator[1] to mitigate the above issue. It packs multiple BPF programs into a single huge page. It is currently only enabled for the
x86_64BPF JIT.This patch series enables the BPF prog pack allocator for the ARM64 BPF JIT.
Quoted 8 lines
Quoted 2 lines
If you get a chance to take another look at the v4 changes from Puranjay and in case they look good to you reply with an Ack, that would be great.
Sure -- this is on my queue of things to look at; it might just take me a few days to get the time to give this a proper look.
Thanks,
Mark.
I am eagerly looking forward to your feedback on this series.
Thanks,
Puranjay
Hi Mark,
I am really looking forward to your feedback on this series.
Quoted 27 lines · Puranjay Mohan
Hi Mark,
Quoted 23 lines · Mark Rutland
Quoted 1 line · Daniel Borkmann
Hi Mark,
Hi Daniel,
Quoted 5 lines · Puranjay Mohan
Quoted 4 lines
BPF programs currently consume a page each on ARM64. For systems with many BPF programs, this adds significant pressure to instruction TLB. High iTLB pressure usually causes slow down for the whole system.
Quoted 8 lines
Quoted 5 lines
Song Liu introduced the BPF prog pack allocator[1] to mitigate the above issue. It packs multiple BPF programs into a single huge page. It is currently only enabled for the
x86_64BPF JIT.This patch series enables the BPF prog pack allocator for the ARM64 BPF JIT.
If you get a chance to take another look at the v4 changes from Puranjay and in case they look good to you reply with an Ack, that would be great.
Sure -- this is on my queue of things to look at; it might just take me a few days to get the time to give this a proper look.
Quoted 7 lines
Quoted 2 lines
Thanks,
Mark.I am eagerly looking forward to your feedback on this series.
Thanks,
Puranjay
Thanks,
Puranjay
Quoted 36 lines · Puranjay Mohan
Hi Mark,
I am really looking forward to your feedback on this series.Quoted 31 lines · Puranjay Mohan
Hi Mark,
Quoted 17 lines · Mark Rutland
Quoted 1 line · Daniel Borkmann
Hi Mark,
Hi Daniel,
Quoted 10 lines · Puranjay Mohan
Quoted 9 lines
BPF programs currently consume a page each on ARM64. For systems with many BPF programs, this adds significant pressure to instruction TLB. High iTLB pressure usually causes slow down for the whole system.
Song Liu introduced the BPF prog pack allocator[1] to mitigate the above issue. It packs multiple BPF programs into a single huge page. It is currently only enabled for the
x86_64BPF JIT.This patch series enables the BPF prog pack allocator for the ARM64 BPF JIT.
Quoted 8 lines
Quoted 2 lines
If you get a chance to take another look at the v4 changes from Puranjay and in case they look good to you reply with an Ack, that would be great.
Sure -- this is on my queue of things to look at; it might just take me a few days to get the time to give this a proper look.
Thanks,
Mark.I am eagerly looking forward to your feedback on this series.
Mark, Catalin, Florent, KP,
This patch set was submitted on June 26 !
It's not acceptable to delay review for so long.
Please review asap.
Mark Rutland
reviewerHi Alexei,
Quoted 33 lines · Alexei Starovoitov
Quoted 32 lines · Puranjay Mohan
Hi Mark,
I am really looking forward to your feedback on this series.Quoted 27 lines · Puranjay Mohan
Hi Mark,
Quoted 23 lines · Mark Rutland
Quoted 1 line · Daniel Borkmann
Hi Mark,
Hi Daniel,
Quoted 5 lines · Puranjay Mohan
Quoted 4 lines
BPF programs currently consume a page each on ARM64. For systems with many BPF programs, this adds significant pressure to instruction TLB. High iTLB pressure usually causes slow down for the whole system.
Quoted 8 lines
Quoted 5 lines
Song Liu introduced the BPF prog pack allocator[1] to mitigate the above issue. It packs multiple BPF programs into a single huge page. It is currently only enabled for the
x86_64BPF JIT.This patch series enables the BPF prog pack allocator for the ARM64 BPF JIT.
If you get a chance to take another look at the v4 changes from Puranjay and in case they look good to you reply with an Ack, that would be great.
Sure -- this is on my queue of things to look at; it might just take me a few days to get the time to give this a proper look.
Quoted 8 lines
Quoted 4 lines
Quoted 4 lines
Quoted 2 lines
Thanks,
Mark.I am eagerly looking forward to your feedback on this series.
Mark, Catalin, Florent, KP,
This patch set was submitted on June 26 !
I appreciate this was sent a while ago, but I have been stuck on some urgent bug-fixing for the last few weeks, and my review bandwidth is therfore very limited.
Given Puranjay had previously told me he was doing this as a side project for fun, and given no-one had told me this was urgent, I assumed that this wasn't a major blocker and could wait.
I should have sent a holding reply to that effect; sorry.
The series addresses my original concern. However, in looking at it I think there may me a wider potential isssue w.r.t. the way instruction memory gets reused, because as writtten today the architecture doesn't seem to have a guarantee on when instruction fetches are completed and therefore when it's safe to modify instruction memory. Usually we're saved by TLB maintenance, which this series avoids by design.
I unfortunately haven't had the time to dig into that, poke our architects, etc.
So how urgent is this?
Thanks,
Mark.
Quoted 63 lines · Mark Rutland
Hi Alexei,
Quoted 33 lines · Alexei Starovoitov
Quoted 32 lines · Puranjay Mohan
Hi Mark,
I am really looking forward to your feedback on this series.Quoted 27 lines · Puranjay Mohan
Hi Mark,
Quoted 23 lines · Mark Rutland
Quoted 1 line · Daniel Borkmann
Hi Mark,
Hi Daniel,
Quoted 5 lines · Puranjay Mohan
Quoted 4 lines
BPF programs currently consume a page each on ARM64. For systems with many BPF programs, this adds significant pressure to instruction TLB. High iTLB pressure usually causes slow down for the whole system.
Quoted 8 lines
Quoted 5 lines
Song Liu introduced the BPF prog pack allocator[1] to mitigate the above issue. It packs multiple BPF programs into a single huge page. It is currently only enabled for the
x86_64BPF JIT.This patch series enables the BPF prog pack allocator for the ARM64 BPF JIT.
If you get a chance to take another look at the v4 changes from Puranjay and in case they look good to you reply with an Ack, that would be great.
Sure -- this is on my queue of things to look at; it might just take me a few days to get the time to give this a proper look.
Quoted 8 lines
Quoted 4 lines
Quoted 4 lines
Quoted 2 lines
Thanks,
Mark.I am eagerly looking forward to your feedback on this series.
Mark, Catalin, Florent, KP,
This patch set was submitted on June 26 !
I appreciate this was sent a while ago, but I have been stuck on some urgent bug-fixing for the last few weeks, and my review bandwidth is therfore very limited.
Given Puranjay had previously told me he was doing this as a side project for fun, and given no-one had told me this was urgent, I assumed that this wasn't a major blocker and could wait.
I should have sent a holding reply to that effect; sorry.
The series addresses my original concern. However, in looking at it I think there may me a wider potential isssue w.r.t. the way instruction memory gets reused, because as writtten today the architecture doesn't seem to have a guarantee on when instruction fetches are completed and therefore when it's safe to modify instruction memory. Usually we're saved by TLB maintenance, which this series avoids by design.
Quoted 4 lines
I unfortunately haven't had the time to dig into that, poke our architects, etc.
So how urgent is this?
The performance wins are substantial.
We'd like to realize them sooner than later.
Quoted 7 lines · Alexei Starovoitov
Quoted 5 lines · Mark Rutland
Hi Alexei,
Quoted 1 line · Alexei Starovoitov
Mark, Catalin, Florent, KP,
Maybe you've missed my Acked-by for the series Alexei ?
https://lore.kernel.org/all/CABRcYmLAzhG=o2wcBNBtFP34Aj3+eYsEMtMREDT7SqNzBc9-qw@mail.gmail.com/
Quoted 10 lines
Quoted 10 lines
Quoted 1 line
This patch set was submitted on June 26 !
I appreciate this was sent a while ago, but I have been stuck on some urgent bug-fixing for the last few weeks, and my review bandwidth is therfore very limited.
Given Puranjay had previously told me he was doing this as a side project for fun, and given no-one had told me this was urgent, I assumed that this wasn't a major blocker and could wait.
Quoted 8 lines
Quoted 8 lines
I should have sent a holding reply to that effect; sorry.
The series addresses my original concern. However, in looking at it I think there may me a wider potential isssue w.r.t. the way instruction memory gets reused, because as writtten today the architecture doesn't seem to have a guarantee on when instruction fetches are completed and therefore when it's safe to modify instruction memory. Usually we're saved by TLB maintenance, which this series avoids by design.
But I must say that this sits firmly outside of my knowledge of the arm architectural details and I would totally miss this sort of nuance so this is best handled by arm64 maintainers :)
Quoted 7 lines
Quoted 4 lines
I unfortunately haven't had the time to dig into that, poke our architects, etc.
So how urgent is this?
The performance wins are substantial.
We'd like to realize them sooner than later.
I've worked with Mark before, I know for a fact that he is dragged in all directions. Until we figure out a way to clone him we should try to not burn him out too often... :)
Hi Mark,
Quoted 46 lines · Mark Rutland
Hi Alexei,
Quoted 33 lines · Alexei Starovoitov
Quoted 32 lines · Puranjay Mohan
Hi Mark,
I am really looking forward to your feedback on this series.Quoted 27 lines · Puranjay Mohan
Hi Mark,
Quoted 23 lines · Mark Rutland
Quoted 1 line · Daniel Borkmann
Hi Mark,
Hi Daniel,
Quoted 5 lines · Puranjay Mohan
Quoted 4 lines
BPF programs currently consume a page each on ARM64. For systems with many BPF programs, this adds significant pressure to instruction TLB. High iTLB pressure usually causes slow down for the whole system.
Quoted 8 lines
Quoted 5 lines
Song Liu introduced the BPF prog pack allocator[1] to mitigate the above issue. It packs multiple BPF programs into a single huge page. It is currently only enabled for the
x86_64BPF JIT.This patch series enables the BPF prog pack allocator for the ARM64 BPF JIT.
If you get a chance to take another look at the v4 changes from Puranjay and in case they look good to you reply with an Ack, that would be great.
Sure -- this is on my queue of things to look at; it might just take me a few days to get the time to give this a proper look.
Quoted 8 lines
Quoted 4 lines
Quoted 4 lines
Quoted 2 lines
Thanks,
Mark.I am eagerly looking forward to your feedback on this series.
Mark, Catalin, Florent, KP,
This patch set was submitted on June 26 !
Quoted 7 lines
I appreciate this was sent a while ago, but I have been stuck on some urgent bug-fixing for the last few weeks, and my review bandwidth is therfore very limited.
Given Puranjay had previously told me he was doing this as a side project for fun, and given no-one had told me this was urgent, I assumed that this wasn't a major blocker and could wait.
Yes, I am just doing it as a side project for fun. It is not a major blocker.
Quoted 10 lines
I should have sent a holding reply to that effect; sorry.
The series addresses my original concern. However, in looking at it I think there may me a wider potential isssue w.r.t. the way instruction memory gets reused, because as writtten today the architecture doesn't seem to have a guarantee on when instruction fetches are completed and therefore when it's safe to modify instruction memory. Usually we're saved by TLB maintenance, which this series avoids by design.
Quoted 4 lines
I unfortunately haven't had the time to dig into that, poke our architects, etc.
So how urgent is this?
This is not urgent as this is not a blocker for anything.
I just wanted to know if there was something pending from my side.
Please review it whenever you have spare time. Thanks for helping me debug the issue with the cache maintenance.
Thanks,
Puranjay
Mark Rutland
reviewerOn Thu, Aug 03, 2023 at 12:13:00PM +0100, Mark Rutland wrote:
[...]
Quoted 6 lines
However, in looking at it I think
there may me a wider potential isssue w.r.t. the way instruction memory gets reused, because as writtten today the architecture doesn't seem to have a guarantee on when instruction fetches are completed and therefore when it's safe to modify instruction memory. Usually we're saved by TLB maintenance, which this series avoids by design.
Just to confirm on this point specifically, per discussions with our architects, the (architectural) execution of an instruction ensures that there are no outstanding fetches for prior instructions. IIUC that will be clarified the next release of the ARM ARM.
So as long as we're certain all threads have left the old code (e.g. via a flag, RCU tasks rude synchronization, whatever) before we overwrite slots in the shared buffer, we should be good.
We will need to be very careful with the maintenance when installing new code into the shared buffer (e.g. we will require an IPI to all other CPUs), but that should be relatively simple.
I'll go review the latest patches with that in mind.
Thanks,
Mark.