ftrace學(xué)習(xí)筆記
[導(dǎo)讀]目錄1.前言2.ARM64棧幀結(jié)構(gòu)3.編譯階段3.1未開啟ftrace時的blk_update_request3.2開啟ftrace時的blk_update_request4.鏈接階段4.1未開啟ftrace時的blk_update_request4.2開啟ftrace時的blk...
目錄
總結(jié)一下：通過對aarch64代碼反匯編的分析，可以得出：1.?????每個函數(shù)在入口處首先會分配棧空間，且一次分配，確定棧頂，之后sp將不再變化；2.?????每個函數(shù)的棧頂部存放的是caller的棧頂指針，即fun1的棧頂存放的是main棧頂指針;3.?????對于最后一級callee函數(shù)，由于x29保存了上一級caller的棧頂sp指針，因此不在需要入棧保存，如示例中fun2執(zhí)行時，此時x29指向fun1的棧頂sp下面我們將根據(jù)是否開啟ftrace配置，并區(qū)分編譯階段、鏈接階段和運行階段，分別查看鉤子函數(shù)的替換及構(gòu)建情況。
bl 0xffff80001002c370
- 1. 前言
- 2. ARM64棧幀結(jié)構(gòu)
- 3. 編譯階段
- 3.1 未開啟ftrace時的blk_update_request
- 3.2 開啟ftrace時的blk_update_request
- 4. 鏈接階段
- 4.1 未開啟ftrace時的blk_update_request
- 4.2 開啟ftrace時的blk_update_request
- 5. 運行階段
- 5.1 ftrace_init執(zhí)行后的blk_update_request
- 5.2 設(shè)定trace函數(shù)blk_update_request
- 6. 鉤子函數(shù)的替換過程
- 7.總結(jié)
- 參考文檔
1. 前言
本文主要是根據(jù)閱碼場 《Linux內(nèi)核tracers的實現(xiàn)原理與應(yīng)用》視頻課程,我自己在aarch64上的實踐。通過觀察鉤子函數(shù)的創(chuàng)建過程以及替換過程，理解trace的原理。本文同樣以blk_update_request函數(shù)為例進(jìn)行說明。kernel版本：5.10平臺：arm642.ARM64棧幀結(jié)構(gòu)
在開始介紹arm64架構(gòu)下的ftrace之前，先來簡要說明一下arm64棧幀的相關(guān)知識。arm64有31個通用寄存器r0-r30，其中r0-r7用于Parameter/result 寄存器; r29為Frame Pointer寄存器，r30為Link寄存器，指向上級函數(shù)的返回地址；SP為棧指針。將以如下代碼為例，說明它的棧幀結(jié)構(gòu)：/* * ARCH: armv8 * GCC版本：aarch64-linux-gnu-gcc (Linaro GCC 5.4-2017.01) 5.4.1 20161213 */intfun2(int c,int d){?? return0;}?intfun1(int a,int b){?? int c = 1;?? int d = 2;?????fun2(c, d);?? return0;}?intmain(int argc,char **argv){?? int a = 0;?? int b = 1;?? fun1(a,b);}aarch64-linux-gnu-objdump -d a.out 反匯編后的結(jié)果為：0000000000400530 : ? /* 更新sp到fun2的棧底 */? 400530:?????? d10043ff??????? sub???? sp, sp, #0x10? 400534:?????? b9000fe0??????? str???? w0, [sp,#12]? 400538:?????? b9000be1??????? str???? w1, [sp,#8]? 40053c:?????? 52800000??????? mov???? w0, #0x0??????????????????????? // #0? 400540:?????? 910043ff??????? add???? sp, sp, #0x10? 400544:?????? d65f03c0??????? ret?0000000000400548 : ? /* 分配48字節(jié)?？臻g，先更新sp=sp-48, 再入棧x29, x30, 此時sp指向棧頂 */? 400548:?????? a9bd7bfd??????? stp???? x29, x30, [sp,#-48]!? /* x29、sp指向棧頂*/? 40054c:?????? 910003fd??????? mov???? x29, sp? /* 入棧fun1參數(shù)0 */? 400550:?????? b9001fa0??????? str???? w0, [x29,#28]? /* 入棧fun1參數(shù)1 */? 400554:?????? b9001ba1??????? str???? w1, [x29,#24]? /* 入棧fun1局部變量c */? 400558:?????? 52800020??????? mov???? w0, #0x1??????????????????????? // #1? 40055c:?????? b9002fa0??????? str???? w0, [x29,#44]? /* 入棧fun1局部變量d */? 400560:?????? 52800040??????? mov???? w0, #0x2??????????????????????? // #2? 400564:?????? b9002ba0??????? str???? w0, [x29,#40]? 400568:?????? b9402ba1??????? ldr???? w1, [x29,#40]? 40056c:?????? b9402fa0??????? ldr???? w0, [x29,#44]? /* 跳轉(zhuǎn)到fun2 */? 400570:?????? 97fffff0??????? bl????? 400530 ? 400574:?????? 52800000??????? mov???? w0, #0x0??????????????????????? // #0? 400578:?????? a8c37bfd??????? ldp???? x29, x30, [sp],#48? 40057c:?????? d65f03c0??????? ret?0000000000400580 : ? /* 分配48字節(jié)?？臻g，先更新sp=sp-48, 再入棧x29, x30, 此時sp指向棧頂*/? 400580:?????? a9bd7bfd??????? stp???? x29, x30, [sp,#-48]!? /* x29、sp指向棧頂*/? 400584:?????? 910003fd??????? mov???? x29, sp? /* 入棧main參數(shù)0 */? 400588:?????? b9001fa0??????? str???? w0, [x29,#28]? /* 入棧main參數(shù)1 */? 40058c:?????? f9000ba1??????? str???? x1, [x29,#16]? /* 入棧變量a */? 400590:?????? b9002fbf??????? str???? wzr, [x29,#44]? 400594:?????? 52800020??????? mov???? w0, #0x1??????????????????????? // #1? /* 入棧變量b */? 400598:?????? b9002ba0??????? str???? w0, [x29,#40]? 40059c:?????? b9402ba1??????? ldr???? w1, [x29,#40]? 4005a0:?????? b9402fa0??????? ldr???? w0, [x29,#44]? /* 跳轉(zhuǎn)到fun1 */? 4005a4:?????? 97ffffe9??????? bl????? 400548 ? 4005a8:?????? 52800000??????? mov???? w0, #0x0??????????????????????? // #0? 4005ac:?????? a8c37bfd??????? ldp???? x29, x30, [sp],#48? 4005b0:?????? d65f03c0??????? ret? 4005b4:?????? 00000000??????? .inst?? 0x00000000 ; undefined對應(yīng)棧幀結(jié)構(gòu)為：總結(jié)一下：通過對aarch64代碼反匯編的分析，可以得出：1.?????每個函數(shù)在入口處首先會分配棧空間，且一次分配，確定棧頂，之后sp將不再變化；2.?????每個函數(shù)的棧頂部存放的是caller的棧頂指針，即fun1的棧頂存放的是main棧頂指針;3.?????對于最后一級callee函數(shù)，由于x29保存了上一級caller的棧頂sp指針，因此不在需要入棧保存，如示例中fun2執(zhí)行時，此時x29指向fun1的棧頂sp下面我們將根據(jù)是否開啟ftrace配置，并區(qū)分編譯階段、鏈接階段和運行階段，分別查看鉤子函數(shù)的替換及構(gòu)建情況。
3. 編譯階段
3.1 未開啟ftrace時的blk_update_request
00000000000012ac : ??? 12ac:?????? d10183ff??????? sub??? ?sp, sp, #0x60??? 12b0:?????? a9017bfd??????? stp???? x29, x30, [sp,#16]??? 12b4:?????? 910043fd??????? add???? x29, sp, #0x10??? 12b8:?????? a90253f3??????? stp???? x19, x20, [sp,#32]??? 12bc:?????? a9035bf5??????? stp???? x21, x22, [sp,#48]??? 12c0:?????? a90463f7??????? stp???? x23, x24, [sp,#64]??? 12c4:?????? f9002bf9??????? str???? x25, [sp,#80]??? 12c8:?????? aa0003f6??????? mov???? x22, x0??? 12cc:?????? 53001c38??????? uxtb??? w24, w1??? 12d0:?????? 2a0203f5??????? mov???? w21, w2??? 12d4:?????? 2a1803e0??????? mov???? w0, w24??? 12d8:?????? 94000000??????? bl????? 12c ??? ...在未使能內(nèi)核配置項CONFIG_FTRACE時，反匯編blk_update_request函數(shù)可以看出，不包含鉤子函數(shù)。3.2 開啟ftrace時的blk_update_request
0000000000003f10 : ??? 3f10:?????? d10183ff??????? sub???? sp, sp, #0x60??? 3f14:?????? a9017bfd??????? stp???? x29, x30, [sp,#16]??? 3f18:?????? 910043fd??????? add???? x29, sp, #0x10??? 3f1c:?????? a90253f3??????? stp???? x19, x20, [sp,#32]??? 3f20:?????? a9035bf5??????? stp???? x21, x22, [sp,#48]??? 3f24:?????? a90463f7??????? stp???? x23, x24, [sp,#64]??? 3f28:?????? f9002bf9??????? str???? x25, [sp,#80]??? 3f2c:?????? aa0003f6??????? mov???? x22, x0??? 3f30:?????? 53001c38??????? uxtb??? w24, w1??? 3f34:?????? 2a0203f5??????? mov???? w21, w2??? 3f38:?????? aa1e03e0??????? mov???? x0, x30??? 3f3c:?????? 94000000??????? bl????? 0 <_mcount>??? ...在使能內(nèi)核配置項CONFIG_FTRACE時，可以看到blk_update_request函數(shù)增加了如下部分：??? 3f3c:?????? 94000000??????? bl????? 0 <_mcount>那么 bl 0 <_mcount> 是由誰在何時插入的呢? 答案是編譯器在編譯時插入，編譯選項-pg -mrecord-mcoun會在編譯時在每個可trace函數(shù)插入bl 0 <_mcount>，并將所有可trace的函數(shù)放到一個__mcount_loc的section中。通過查看blk-core.o的可重定位段，可以看到有大量的地址需要定位到_mcount函數(shù)，其中3f3c地址正是位于blk_update_request，它會在鏈接階段被重定位到_mcount函數(shù)的地址。ubuntu@VM-0-9-ubuntu:~/qemu/kernel/linux/block$ aarch64-linux-gnu-objdump -r blk-core.o | grep _mcount0000000000000014 R_AARCH64_CALL26? _mcount000000000000005c R_AARCH64_CALL26? _mcount00000000000000ac R_AARCH64_CALL26? _mcount0000000000000108 R_AARCH64_CALL26? _mcount0000000000000164 R_AARCH64_CALL26? _mcount00000000000001bc R_AARCH64_CALL26? _mcount0000000000000214 R_AARCH64_CALL26? _mcount...0000000000003f3c R_AARCH64_CALL26? _mcount...我們還可以看到，blk-core.o有一個.rela__mcount_loc的可重定位段，里面存放了所有需要可trace函數(shù)中需要重定位到函數(shù)_mcount的地址。ubuntu@VM-0-9-ubuntu:~/qemu/kernel/linux/block$ aarch64-linux-gnu-objdump -r blk-core.o...RELOCATION RECORDS FOR [__mcount_loc]:OFFSET????? ?????TYPE????????????? VALUE0000000000000000 R_AARCH64_ABS64?? .text 0x00000000000000140000000000000008 R_AARCH64_ABS64?? .text 0x000000000000005c0000000000000010 R_AARCH64_ABS64?? .text 0x00000000000000ac0000000000000018 R_AARCH64_ABS64?? .text 0x0000000000000108...00000000000001b8 R_AARCH64_ABS64?? .text 0x0000000000003f3c...4. 鏈接階段
4.1 未開啟ftrace時的blk_update_request
未使能內(nèi)核配置項CONFIG_FTRACE時，鏈接階段與編譯階段一樣，反匯編blk_update_request函數(shù)可以看出，不包含鉤子函數(shù)4.2 開啟ftrace時的blk_update_request
ffff8000104e43c8 : ffff8000104e43c8:?????? d10183ff??????? sub???? sp, sp, #0x60ffff8000104e43cc:?????? a9017bfd??????? stp???? x29, x30, [sp,#16]ffff8000104e43d0:?????? 910043fd??????? add???? x29, sp, #0x10ffff8000104e43d4:?????? a90253f3??????? stp???? x19, x20, [sp,#32]ffff8000104e43d8:?????? a9035bf5??????? stp???? x21, x22, [sp,#48]ffff8000104e43dc:?????? a90463f7??????? stp???? x23, x24, [sp,#64]ffff8000104e43e0:?????? f9002bf9??????? str???? x25, [sp,#80]ffff8000104e43e4:?????? aa0003f6??????? mov???? x22, x0ffff8000104e43e8:?????? 53001c38??????? uxtb??? w24, w1ffff8000104e43ec:?????? 2a0203f5??????? mov???? w21, w2ffff8000104e43f0:?????? aa1e03e0??????? mov???? x0, x30ffff8000104e43f4:?????? 97ed1fde??????? bl????? ffff80001002c36c <_mcount>ffff8000104e43f8:?????? 2a1803e0??????? mov???? w0, w24ffff8000104e43fc:?????? 97fff432??????? bl????? ffff8000104e14c4 ...在鏈接階段，使能內(nèi)核配置項CONFIG_FTRACE時，可以看到編譯階段的如下代碼??? 3f3c:?????? 94000000??????? bl????? 0 <_mcount>在鏈接階段已經(jīng)被替換為：ffff8000104e43f4:?????? 97ed1fde??????? bl????? ffff80001002c36c <_mcount>其中_mcount函數(shù)反匯編為：ffff80001002c36c <_mcount>:ffff80001002c36c:?????? d65f03c0??????? ret5. 運行階段
5.1ftrace_init執(zhí)行后的blk_update_request
(gdb) x/20i blk_update_request?? 0xffff8000104e43c8 :???? sub???? sp, sp, #0x60 ?? 0xffff8000104e43cc :?? stp???? x29, x30, [sp,#16] ?? 0xffff8000104e43d0 :?? add???? x29, sp, #0x10 ?? 0xffff8000104e43d4 :? stp???? x19, x20, [sp,#32] ?? 0xffff8000104e43d8 :? stp???? x21, x22, [sp,#48] ?? 0xffff8000104e43dc :? stp???? x23, x24, [sp,#64] ?? 0xffff8000104e43e0 :? str???? x25, [sp,#80] ??0xffff8000104e43e4 :? mov???? x22, x0 ?? 0xffff8000104e43e8 :? uxtb??? w24, w1 ?? 0xffff8000104e43ec :? mov???? w21, w2 ?? 0xffff8000104e43f0 :? mov???? x0, x30 ??0xffff8000104e43f4 :? nop ?? 0xffff8000104e43f8 :? mov???? w0, w24 ?? 0xffff8000104e43fc :? bl????? 0xffff8000104e14c4 內(nèi)核在start_kernel執(zhí)行時，會調(diào)用ftrace_init，它會將所有可trace函數(shù)中的_mcount進(jìn)行替換，如上可以看出鏈接階段的 bl ffff80001002c36c <_mcount> 已經(jīng)被替換為nop指令5.2 設(shè)定trace函數(shù)blk_update_request
執(zhí)行如下命令來trace函數(shù)blk_update_requestubuntu@VM-0-9-ubuntu:~$echo blk_update_request > /sys/kernel/debug/tracing/set_ftrace_filterubuntu@VM-0-9-ubuntu:~$echo function > /sys/kernel/debug/tracing/current_tracer我們再來查看blk_update_request反匯編代碼(gdb) x/20i blk_update_request?? 0xffff8000104e43c8 :???? sub???? sp, sp, #0x60 ?? 0xffff8000104e43cc :?? stp???? x29, x30, [sp,#16] ?? 0xffff8000104e43d0 :?? add???? x29, sp, #0x10 ?? 0xffff8000104e43d4 :? stp???? x19, x20, [sp,#32] ?? 0xffff8000104e43d8 :? stp???? x21, x22, [sp,#48] ?? 0xffff8000104e43dc :? stp???? x23, x24, [sp,#64] ?? 0xffff8000104e43e0 :? str???? x25, [sp,#80] ?? 0xffff8000104e43e4 :? mov???? x22, x0 ?? 0xffff8000104e43e8 :? uxtb??? w24, w1 ?? 0xffff8000104e43ec :? mov???? w21, w2 ?? 0xffff8000104e43f0 :? mov???? x0, x30 ?? 0xffff8000104e43f4 :? bl????? 0xffff80001002c370 ?? 0xffff8000104e43f8 :? mov???? w0, w24 ?? 0xffff8000104e43fc :? bl????? 0xffff8000104e14c4 可以看到之前在blk_update_request的nop指令被替換成bl 0xffff80001002c370
(gdb) disassemble ftrace_callerDump of assembler code for function ftrace_caller:?? 0xffff80001002c374 < 0>:???? stp???? x29, x30, [sp,#-16]!?? 0xffff80001002c378 < 4>:???? mov???? x29, sp?? // x30是blk_update_request的lr，-4是當(dāng)前執(zhí)行函數(shù)的入口地址，也就是ftrace_caller的ip?? // 它將作為參數(shù)0傳遞給ftrace_ops_no_ops?? 0xffff80001002c37c < 8>:???? sub???? x0, x30, #0x4?? // 參考前面arm64棧幀結(jié)構(gòu)，x29指向上一級函數(shù)blk_update_request棧頂?? //[x29]指向blk_mq_end_request函數(shù)的棧頂?? //[[x29] 8]為blk_mq_end_request的ip(實際是ip的下條指令)?? 0xffff80001002c380 < 12>:??? ldr???? x1, [x29]?? 0xffff80001002c384 < 16>:??? ldr???? x1, [x1,#8]?? 0xffff80001002c388 < 20>:??? bl????? 0xffff800010188ffc ?? 0xffff80001002c38c < 24>:??? nop?? 0xffff80001002c390 < 28>:??? ldp???? x29, x30, [sp],#16?? 0xffff80001002c394 < 32>:??? retEnd of assembler dump.可以看到ftrace_caller會調(diào)用ftrace_ops_no_ops，我們在ftrace_ops_no_ops源碼中看到它會遍歷ftrace_ops_list鏈表，并執(zhí)行這個鏈表上的回調(diào)函數(shù)，這里看下ftrace_ops_list上都鏈接了哪些func(gdb) p *ftrace_ops_list$4 = {? func = 0xffff8000101a0b1c , //ftrace_ops_list鏈表唯一func? next = 0xffff800011c5a438 , //說明ftrace_ops_list鏈表只有一個func? flags = 8273,??private = 0xffff800011cf94e8 , ??saved_func = 0xffff8000101a0b1c , ??local_hash = {??? notrace_hash = 0xffff800010cf7118 , ????filter_hash = 0xffff00000720af80,????regex_lock = {????? owner = {??????? counter = 0????? },......從ftrace_ops_list鏈表中可以看到只有一個function_trace_call函數(shù)組成，因此可以說ftrace_caller最終會調(diào)用到function_trace_call。通過前面的分析，我們一步步找到了blk_update_request的鉤子函數(shù)function_trace_call，其函數(shù)原型如下，其中參數(shù)ip指向ftrace_caller，參數(shù)parent_ip指向blk_mq_end_request：staticvoidfunction_trace_call(unsignedlong ip, unsignedlong parent_ip,???????????????????????????????????????????????????????????? struct ftrace_ops *op, struct pt_regs *pt_regs)下一節(jié)我們將追蹤鉤子函數(shù)的構(gòu)造以及替換過程。6. 鉤子函數(shù)的替換過程
前面我們看到blk_update_request的nop指令被替換成bl ftrace_caller，那么此處的ftrace_caller是在哪里定義的呢？我們可以看到arch/arm64/kernel/entry-ftrace.S有如下的定義：/* * void ftrace_caller(unsigned long return_address) * @return_address: return address to instrumented function * * This function is a counterpart of _mcount() in 'static' ftrace, and * makes calls to: *???? - tracer function to probe instrumented function's entry, *???? - ftrace_graph_caller to set up an exit hook */SYM_FUNC_START(ftrace_caller)??????? mcount_enter???????? mcount_get_pc0? x0????????????? //???? function's pc??????? mcount_get_lr?? x1????????????? //???? function's lr?SYM_INNER_LABEL(ftrace_call, SYM_L_GLOBAL)????? // tracer(pc, lr);??????? nop???????????????????????????? // This will be replaced with "bl xxx"??????????????????????????????????????? // where xxx can be any kind of tracer.?#ifdef CONFIG_FUNCTION_GRAPH_TRACERSYM_INNER_LABEL(ftrace_graph_call, SYM_L_GLOBAL) // ftrace_graph_caller();??????? nop???????????????????????????? // If enabled, this will be replaced??????????????????????????????????????? // "b ftrace_graph_caller"#endif???????? mcount_exitSYM_FUNC_END(ftrace_caller)通過 gdb可以看到ftrace_caller的反匯編代碼如下：(gdb) disassemble ftrace_callerDump of assembler code for function ftrace_caller:?? 0xffff80001002c370 < 0>:???? stp???? x29, x30, [sp,#-16]!?? 0xffff80001002c374 < 4>:???? mov???? x29, sp?? 0xffff80001002c378 < 8>:???? sub???? x0, x30, #0x4?? 0xffff80001002c37c < 12>:??? ldr???? x1, [x29]?? 0xffff80001002c380 < 16>:??? ldr???? x1, [x1,#8]?? 0xffff80001002c384 < 20>:??? nop????????????????? /*ftrace_call*/?? 0xffff80001002c388 < 24>:??? nop????????????????? /*ftrace_graph_call,暫不討論*/?? 0xffff80001002c38c < 28>:??? ldp???? x29, x30, [sp],#16?? 0xffff80001002c390 < 32>:??? retEnd of assembler dump.當(dāng)執(zhí)行echo blk_update_request >set_ftrace_filter時相當(dāng)于使能了blk_update_request的鉤子替換標(biāo)志，當(dāng)執(zhí)行echo function >current_tracer時會檢查這個標(biāo)志，并執(zhí)行替換，它會產(chǎn)生如下的調(diào)用鏈：/sys/kernel/debug/tracing # echo function > current_tracer[?? 45.632002] CPU: 0 PID: 111 Comm: sh Not tainted 5.10.0-dirty #35[?? 45.632457] Hardware name: linux,dummy-virt (DT)[?? 45.632697] Call trace:[?? 45.632981]? dump_backtrace 0x0/0x1f8[?? 45.633169]? show_stack 0x2c/0x7c[?? 45.634039]? ftrace_modify_all_code 0x38/0x118[?? 45.634269]? arch_ftrace_update_code 0x10/0x18[?? 45.634495]? ftrace_run_update_code 0x2c/0x48[?? 45.634727]? ftrace_startup_enable 0x40/0x4c[?? 45.634943]? ftrace_startup 0xec/0x11c[?? 45.635137]? register_ftrace_function 0x68/0x84[?? 45.635369]? function_trace_init 0xa0/0xc4[?? 45.635574]? tracer_init 0x28/0x34[?? 45.635768]? tracing_set_tracer 0x11c/0x17c[?? 45.635982]? tracing_set_trace_write 0x124/0x170[?? 45.636224]? vfs_write 0x16c/0x368[?? 45.636409]? ksys_write 0x74/0x10c[?? 45.636594]? __arm64_sys_write 0x28/0x34[?? 45.636923]? el0_svc_common 0xf0/0x174[?? 45.637138]? do_el0_svc 0x84/0x90[?? 45.637330]? el0_svc 0x1c/0x28[?? 45.637510]? el0_sync_handler 0x3c/0xac[?? 45.637721]? el0_sync 0x140/0x180進(jìn)一步查看ftrace_modify_all_code的代碼，我們可以看到如下的調(diào)用流程：ftrace_modify_all_code(command)? \--ftrace_update_ftrace_func(ftrace_ops_list_func)?????? |--pc = (unsignedlong) 
下載文檔
