arm64: kprobes: Use BRK instead of single-step when executing instructions out-of-line
PATCH 1/1 arm64: kprobes: Use BRK instead of single-step when executing instructions out-of-line
Commit 36dadef23fcc ("kprobes: Init kprobes in early_initcall") enabled using kprobes from early_initcall. Unfortunately at this point the hardware debug infrastructure is not operational. The OS lock may still be locked, and the hardware watchpoints may have unknown values when kprobe enables debug monitors to single-step instructions.
Rather than using hardware single-step, append a BRK instruction after the instruction to be executed out-of-line.
---
An alternative to fix [1]. I haven't done uprobes yet since they don't suffer the same problem, but could add it to the bottom of my list. Lightly tested with kprobe smoke test and the BPF selftests. Interestingly on Seattle when running the "overhead" BPF selftest, that triggers a kprobe a bunch of times, I see a 20-30% improvement with this patch. I'm guessing it's the cost of touching the debug sysregs?
[1] https://lore.kernel.org/linux-arm-kernel/[email protected]/
--- arch/arm64/include/asm/brk-imm.h | 2 + arch/arm64/include/asm/debug-monitors.h | 1 + arch/arm64/include/asm/kprobes.h | 2 +- arch/arm64/kernel/probes/kprobes.c | 56 +++++++++---------------- 4 files changed, 24 insertions(+), 37 deletions(-)
arch/arm64/include/asm/brk-imm.h +2 -0
@@ -10,6 +10,7 @@* #imm16 values used for BRK instruction generation * 0x004: for installing kprobes * 0x005: for installing uprobes +* 0x006: for kprobe software single-step * Allowed values for kgdb are 0x400 - 0x7ff * 0x100: for triggering a fault on purpose (reserved) * 0x400: for dynamic BRK instruction
@@ -19,6 +20,7 @@*/ #define KPROBES_BRK_IMM 0x004#define UPROBES_BRK_IMM 0x005 +#define KPROBES_BRK_SS_IMM 0x006#define FAULT_BRK_IMM 0x100#define KGDB_DYN_DBG_BRK_IMM 0x400#define KGDB_COMPILED_DBG_BRK_IMM 0x401
arch/arm64/include/asm/debug-monitors.h +1 -0
@@ -53,6 +53,7 @@/* kprobes BRK opcodes with ESR encoding */ #define BRK64_OPCODE_KPROBES (AARCH64_BREAK_MON | (KPROBES_BRK_IMM << 5)) +#define BRK64_OPCODE_KPROBES_SS (AARCH64_BREAK_MON | (KPROBES_BRK_SS_IMM << 5))/* uprobes BRK opcodes with ESR encoding */ #define BRK64_OPCODE_UPROBES (AARCH64_BREAK_MON | (UPROBES_BRK_IMM << 5))
arch/arm64/include/asm/kprobes.h +1 -1
@@ -16,7 +16,7 @@#include <linux/percpu.h> #define __ARCH_WANT_KPROBES_INSN_SLOT -#define MAX_INSN_SIZE 1 +#define MAX_INSN_SIZE 2#define flush_insn_slot (p) do { } while (0)#define kretprobe_blacklist_size 0
arch/arm64/kernel/probes/kprobes.c +20 -36
144 lines not on this page · raw message
Signature · 7 lines
2.29.1 _______________________________________________ linux-arm-kernel mailing list [email protected] http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
Quoted 8 lines · On Thu, 29 Oct 2020 18:34:42 +0100 Jean-Philippe Brucker…
Commit
36dadef23fcc("kprobes: Init kprobes inearly_initcall") enabled using kprobes fromearly_initcall. Unfortunately at this point the hardware debug infrastructure is not operational. The OS lock may still be locked, and the hardware watchpoints may have unknown values when kprobe enables debug monitors to single-step instructions.Rather than using hardware single-step, append a BRK instruction after the instruction to be executed out-of-line.
Ok, this looks good to me. One comment is that we can remove ss_ctx too. ss_ctx has ss_pending bit and match_addr, those are redundant because those can be replaced by KPROBE_HIT_SS and &cur_kprobe->ainsn.api.insn[1] respectively. But that should be done in another patch. To fix the bug, I think this is good.
Quoted 11 lines
Fixes 36dadef23fcc ("kprobes: Init kprobes in early_initcall")Suggested-by Will Deacon <[email protected]>Signed-off-by Jean-Philippe Brucker <[email protected]>---
An alternative to fix [1]. I haven't done uprobes yet since they don't suffer the same problem, but could add it to the bottom of my list. Lightly tested with kprobe smoke test and the BPF selftests. Interestingly on Seattle when running the "overhead" BPF selftest, that triggers a kprobe a bunch of times, I see a 20-30% improvement with this patch. I'm guessing it's the cost of touching the debug sysregs?
Interesting :) I think that is the cost of the debug exception itself. I guess there might be a cost to route the debug logic.
Thank you,
Quoted 214 lines
[1] https://lore.kernel.org/linux-arm-kernel/[email protected]/
--- arch/arm64/include/asm/brk-imm.h | 2 + arch/arm64/include/asm/debug-monitors.h | 1 + arch/arm64/include/asm/kprobes.h | 2 +- arch/arm64/kernel/probes/kprobes.c | 56 +++++++++---------------- 4 files changed, 24 insertions(+), 37 deletions(-)arch/arm64/include/asm/brk-imm.h +2
-0@@ -10,6 +10,7 @@* #imm16 values used for BRK instruction generation * 0x004: for installing kprobes * 0x005: for installing uprobes +* 0x006: for kprobe software single-step * Allowed values for kgdb are 0x400 - 0x7ff * 0x100: for triggering a fault on purpose (reserved) * 0x400: for dynamic BRK instruction @@ -19,6 +20,7 @@*/ #define KPROBES_BRK_IMM 0x004#define UPROBES_BRK_IMM 0x005 +#define KPROBES_BRK_SS_IMM 0x006#define FAULT_BRK_IMM 0x100#define KGDB_DYN_DBG_BRK_IMM 0x400#define KGDB_COMPILED_DBG_BRK_IMM 0x401arch/arm64/include/asm/debug-monitors.h +1
-0@@ -53,6 +53,7 @@/* kprobes BRK opcodes with ESR encoding */ #define BRK64_OPCODE_KPROBES (AARCH64_BREAK_MON | (KPROBES_BRK_IMM << 5)) +#define BRK64_OPCODE_KPROBES_SS (AARCH64_BREAK_MON | (KPROBES_BRK_SS_IMM << 5))/* uprobes BRK opcodes with ESR encoding */ #define BRK64_OPCODE_UPROBES (AARCH64_BREAK_MON | (UPROBES_BRK_IMM << 5))arch/arm64/include/asm/kprobes.h +1
-1@@ -16,7 +16,7 @@#include <linux/percpu.h> #define __ARCH_WANT_KPROBES_INSN_SLOT -#define MAX_INSN_SIZE 1 +#define MAX_INSN_SIZE 2#define flush_insn_slot (p) do { } while (0)#define kretprobe_blacklist_size 0arch/arm64/kernel/probes/kprobes.c +20
-36@@ -36,21 +36,25 @@ DEFINE_PER_CPU(struct kprobe_ctlblk, kprobe_ctlblk);static void __kprobes post_kprobe_handler (struct kprobe_ctlblk *,struct pt_regs *); -static int __kprobes patch_text (kprobe_opcode_t *addr,u32 opcode) +static int __kprobes patch_text (kprobe_opcode_t *addr,u32 opc1,u32 opc2) { -void *addrs[1 ]; -u32 insns[1 ]; +void *addrs[2 ]; +u32 insns[2 ]; addrs[0 ] = addr; - insns[0 ] = opcode; + insns[0 ] = opc1; +if (opc2) { + addrs[1 ] = addr +1 ; + insns[1 ] = opc2; + } -return aarch64_insn_patch_text (addrs, insns,1 ); +return aarch64_insn_patch_text (addrs, insns, opc2 ?2 :1 ); }static void __kprobes arch_prepare_ss_slot (struct kprobe *p) {/* prepare insn slot */ -patch_text (p->ainsn.api.insn, p->opcode); +patch_text (p->ainsn.api.insn, p->opcode,BRK64_OPCODE_KPROBES_SS ); flush_icache_range((uintptr_t) (p->ainsn.api.insn), (uintptr_t) (p->ainsn.api.insn) +@@ -128,13 +132,13 @@ void *alloc_insn_page(void)/* arm kprobe: install breakpoint in text */ void __kprobes arch_arm_kprobe (struct kprobe *p) { -patch_text (p->addr,BRK64_OPCODE_KPROBES ); +patch_text (p->addr,BRK64_OPCODE_KPROBES ,0 ); }/* disarm kprobe: remove breakpoint from text */ void __kprobes arch_disarm_kprobe (struct kprobe *p) { -patch_text (p->addr, p->opcode); +patch_text (p->addr, p->opcode,0 ); }void __kprobes arch_remove_kprobe (struct kprobe *p)@@ -163,20 +167,14 @@ static void __kprobes set_current_kprobe(struct kprobe *p) }/* -* Interrupts need to be disabled before single-step mode is set, and not -* reenabled until after single-step mode ends. -* Without disabling interrupt on local CPU, there is a chance of -* interrupt occurrence in the period of exception return and start of -* out-of-line single-step, that result in wrongly single stepping -* into the interrupt handler. +* Keep interrupts disabled while executing the instruction out-of-line. Since +* the kprobe state is per-CPU, we can't afford to be migrated. */ static void __kprobes kprobes_save_local_irqflag (struct kprobe_ctlblk *kcb,struct pt_regs *regs) { kcb->saved_irqflag = regs->pstate &DAIF_MASK ; regs->pstate |=PSR_I_BIT ; -/* Unmask PSTATE.D for enabling software step exceptions. */ - regs->pstate &= ~PSR_D_BIT ; } static void __kprobes kprobes_restore_local_irqflag(struct kprobe_ctlblk *kcb,@@ -219,10 +217,7 @@ static void __kprobes setup_singlestep(struct kprobe *p, slot = (unsigned long )p->ainsn.api.insn;set_ss_context (kcb, slot);/* mark pending ss */ - -/* IRQs and single stepping do not mix well. */ kprobes_save_local_irqflag (kcb, regs); -kernel_enable_single_step (regs);instruction_pointer_set (regs, slot); }else {/* insn simulation */ @@ -273,12 +268,8 @@ post_kprobe_handler(struct kprobe_ctlblk *kcb, struct pt_regs *regs) }/* call post handler */ kcb->kprobe_status =KPROBE_HIT_SSDONE ; -if (cur->post_handler) { -/* post_handler can hit breakpoint and single step -* again, so we enable D-flag for recursive exception. -*/ +if (cur->post_handler) cur->post_handler (cur, regs,0 ); - }reset_current_kprobe (); }@@ -302,8 +293,6 @@ int __kprobes kprobe_fault_handler(struct pt_regs *regs, unsigned int fsr)if (!instruction_pointer (regs))BUG (); -kernel_disable_single_step (); -if (kcb->kprobe_status ==KPROBE_REENTER )restore_previous_kprobe (kcb);else @@ -365,10 +354,6 @@ static void __kprobes kprobe_handler(struct pt_regs *regs)* pre-handler and it returned non-zero, it will * modify the execution path and no need to single * stepping. Let's just reset current kprobe and exit. -* -* pre_handler can hit a breakpoint and can step thru -* before return, keep PSTATE D-flag enabled until -* pre_handler return back. */ if (!p->pre_handler || !p->pre_handler (p, regs)) {setup_singlestep (p, regs, kcb,0 );@@ -399,7 +384,7 @@ kprobe_ss_hit(struct kprobe_ctlblk *kcb, unsigned long addr) }static int __kprobes -kprobe_single_step_handler (struct pt_regs *regs,unsigned int esr) +kprobe_breakpoint_ss_handler (struct pt_regs *regs,unsigned int esr) {struct kprobe_ctlblk *kcb =get_kprobe_ctlblk ();int retval;@@ -409,16 +394,15 @@ kprobe_single_step_handler(struct pt_regs *regs, unsigned int esr)if (retval ==DBG_HOOK_HANDLED ) {kprobes_restore_local_irqflag (kcb, regs); -kernel_disable_single_step (); -post_kprobe_handler (kcb, regs); }return retval; } -static struct step_hook kprobes_step_hook = { - .fn = kprobe_single_step_handler, +static struct break_hook kprobes_break_ss_hook = { + .imm =KPROBES_BRK_SS_IMM , + .fn = kprobe_breakpoint_ss_handler, }; static int __kprobes@@ -486,7 +470,7 @@ int __kprobes arch_trampoline_kprobe(struct kprobe *p)int __init arch_init_kprobes (void ) {register_kernel_break_hook (&kprobes_break_hook); -register_kernel_step_hook (&kprobes_step_hook); +register_kernel_break_hook (&kprobes_break_ss_hook);return 0 ; }Signature · 1 line
2.29.1
Signature · 6 lines
Masami Hiramatsu <[email protected]> _______________________________________________ linux-arm-kernel mailing list [email protected] http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
PATCH 1/1 arm64: kprobes: Remove redundant kprobe_step_ctx
In reply to Jean-Philippe BruckerThe kprobe_step_ctx (kcb->ss_ctx) has ss_pending and match_addr, but those are redundant because those can be replaced by KPROBE_HIT_SS and &cur_kprobe->ainsn.api.insn[1] respectively.
To simplify the code, remove the kprobe_step_ctx.
--- arch/arm64/include/asm/kprobes.h | 7 ----- arch/arm64/kernel/probes/kprobes.c | 53 ++++++++---------------------------- 2 files changed, 12 insertions(+), 48 deletions(-)
arch/arm64/include/asm/kprobes.h +0 -7
@@ -28,18 +28,11 @@ struct prev_kprobe {unsigned int status; }; -/* Single step context for kprobe */ -struct kprobe_step_ctx { -unsigned long ss_pending; -unsigned long match_addr; -}; -/* per-cpu kprobe control block */ struct kprobe_ctlblk {unsigned int kprobe_status;unsigned long saved_irqflag;struct prev_kprobe prev_kprobe; -struct kprobe_step_ctx ss_ctx; };void arch_remove_kprobe (struct kprobe *);
arch/arm64/kernel/probes/kprobes.c +12 -41
98 lines not on this page · raw message
_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
Re: [PATCH] arm64: kprobes: Remove redundant kprobe_step_ctx
Quoted 6 lines · Masami Hiramatsu
The
kprobe_step_ctx(kcb->ss_ctx) hasss_pendingandmatch_addr, but those are redundant because those can be replaced byKPROBE_HIT_SSand &cur_kprobe->ainsn.api.insn[1] respectively.
To simplify the code, remove thekprobe_step_ctx.Signed-off-by Masami Hiramatsu <[email protected]>
Quoted 127 lines
--- arch/arm64/include/asm/kprobes.h | 7 ----- arch/arm64/kernel/probes/kprobes.c | 53 ++++++++---------------------------- 2 files changed, 12 insertions(+), 48 deletions(-)arch/arm64/include/asm/kprobes.h +0
-7@@ -28,18 +28,11 @@ struct prev_kprobe {unsigned int status; }; -/* Single step context for kprobe */ -struct kprobe_step_ctx { -unsigned long ss_pending; -unsigned long match_addr; -}; -/* per-cpu kprobe control block */ struct kprobe_ctlblk {unsigned int kprobe_status;unsigned long saved_irqflag;struct prev_kprobe prev_kprobe; -struct kprobe_step_ctx ss_ctx; };void arch_remove_kprobe (struct kprobe *);arch/arm64/kernel/probes/kprobes.c +8
-40@@ -34,7 +34,7 @@ DEFINE_PER_CPU(struct kprobe *, current_kprobe) = NULL; DEFINE_PER_CPU(struct kprobe_ctlblk, kprobe_ctlblk);static void __kprobes -post_kprobe_handler (struct kprobe_ctlblk *,struct pt_regs *); +post_kprobe_handler (struct kprobe *,struct kprobe_ctlblk *,struct pt_regs *);static int __kprobes patch_text (kprobe_opcode_t *addr,u32 opc1,u32 opc2) {@@ -81,7 +81,7 @@ static void __kprobes arch_simulate_insn(struct kprobe *p, struct pt_regs *regs) p->ainsn.api.handler ((u32 )p->opcode, (long )p->addr, regs);/* single step simulated, now go for post processing */ -post_kprobe_handler (kcb, regs); +post_kprobe_handler (p, kcb, regs); }int __kprobes arch_prepare_kprobe (struct kprobe *p)@@ -184,19 +184,6 @@ static void __kprobes kprobes_restore_local_irqflag(struct kprobe_ctlblk *kcb, regs->pstate |= kcb->saved_irqflag; } -static void __kprobes -set_ss_context (struct kprobe_ctlblk *kcb,unsigned long addr) -{ - kcb->ss_ctx.ss_pending = true; - kcb->ss_ctx.match_addr = addr +sizeof (kprobe_opcode_t); -} - -static void __kprobes clear_ss_context (struct kprobe_ctlblk *kcb) -{ - kcb->ss_ctx.ss_pending = false; - kcb->ss_ctx.match_addr =0 ; -} -static void __kprobes setup_singlestep (struct kprobe *p,struct pt_regs *regs,struct kprobe_ctlblk *kcb,int reenter)@@ -216,7 +203,6 @@ static void __kprobes setup_singlestep(struct kprobe *p,/* prepare for single stepping */ slot = (unsigned long )p->ainsn.api.insn; -set_ss_context (kcb, slot);/* mark pending ss */ kprobes_save_local_irqflag (kcb, regs);instruction_pointer_set (regs, slot); }else {@@ -250,13 +236,8 @@ static int __kprobes reenter_kprobe(struct kprobe *p, }static void __kprobes -post_kprobe_handler (struct kprobe_ctlblk *kcb,struct pt_regs *regs) +post_kprobe_handler (struct kprobe *cur,struct kprobe_ctlblk *kcb,struct pt_regs *regs) { -struct kprobe *cur =kprobe_running (); - -if (!cur) -return ; -/* return addr restore if non-branching insn */ if (cur->ainsn.api.restore !=0 )instruction_pointer_set (regs, cur->ainsn.api.restore);@@ -371,33 +352,23 @@ static void __kprobes kprobe_handler(struct pt_regs *regs)*/ } -static int __kprobes -kprobe_ss_hit (struct kprobe_ctlblk *kcb,unsigned long addr) -{ -if ((kcb->ss_ctx.ss_pending) - && (kcb->ss_ctx.match_addr == addr)) { -clear_ss_context (kcb);/* clear pending ss */ -return DBG_HOOK_HANDLED ; - } -/* not ours, kprobes should ignore it */ -return DBG_HOOK_ERROR ; -} -static int __kprobes kprobe_breakpoint_ss_handler (struct pt_regs *regs,unsigned int esr) {struct kprobe_ctlblk *kcb =get_kprobe_ctlblk (); -int retval; - -/* return error if this is not our step */ - retval =kprobe_ss_hit (kcb,instruction_pointer (regs)); +unsigned long addr =instruction_pointer (regs); +struct kprobe *cur =kprobe_running (); -if (retval ==DBG_HOOK_HANDLED ) { +if (cur && (kcb->kprobe_status ==KPROBE_HIT_SS ) + && ((unsigned long )&cur->ainsn.api.insn[1 ] == addr)) {kprobes_restore_local_irqflag (kcb, regs); -post_kprobe_handler (kcb, regs); +post_kprobe_handler (cur, kcb, regs);
arch/arm64/kernel/probes/kprobes.c static void __kprobes kprobe_handler(struct pt_regs *regs) · line 374
Quoted 10 lines
arch/arm64/kernel/probes/kprobes.c +4
-1+ +return DBG_HOOK_HANDLED ; } -return retval; +/* not ours, kprobes should ignore it */ +return DBG_HOOK_ERROR ; }static struct break_hook kprobes_break_ss_hook = {
_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
Will Deacon
maintainerQuoted 21 lines · Jean-Philippe Brucker
Commit
36dadef23fcc("kprobes: Init kprobes inearly_initcall") enabled using kprobes fromearly_initcall. Unfortunately at this point the hardware debug infrastructure is not operational. The OS lock may still be locked, and the hardware watchpoints may have unknown values when kprobe enables debug monitors to single-step instructions.Rather than using hardware single-step, append a BRK instruction after the instruction to be executed out-of-line.
Fixes 36dadef23fcc ("kprobes: Init kprobes in early_initcall")Suggested-by Will Deacon <[email protected]>Signed-off-by Jean-Philippe Brucker <[email protected]>---
An alternative to fix [1]. I haven't done uprobes yet since they don't suffer the same problem, but could add it to the bottom of my list. Lightly tested with kprobe smoke test and the BPF selftests. Interestingly on Seattle when running the "overhead" BPF selftest, that triggers a kprobe a bunch of times, I see a 20-30% improvement with this patch. I'm guessing it's the cost of touching the debug sysregs?
Quoted 7 lines
[1] https://lore.kernel.org/linux-arm-kernel/[email protected]/
--- arch/arm64/include/asm/brk-imm.h | 2 + arch/arm64/include/asm/debug-monitors.h | 1 + arch/arm64/include/asm/kprobes.h | 2 +- arch/arm64/kernel/probes/kprobes.c | 56 +++++++++---------------- 4 files changed, 24 insertions(+), 37 deletions(-)
Thanks! I'm about to send fixes for -rc2, but I'd like this to get a little bit of time in linux-next, so I aim to queue it next week for -rc3. Just wanted to make sure you don't think I'm ignored you.
Cheers,
Will
_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
Quoted 30 lines · Will Deacon
Quoted 21 lines · Jean-Philippe Brucker
Commit
36dadef23fcc("kprobes: Init kprobes inearly_initcall") enabled using kprobes fromearly_initcall. Unfortunately at this point the hardware debug infrastructure is not operational. The OS lock may still be locked, and the hardware watchpoints may have unknown values when kprobe enables debug monitors to single-step instructions.Rather than using hardware single-step, append a BRK instruction after the instruction to be executed out-of-line.
Fixes 36dadef23fcc ("kprobes: Init kprobes in early_initcall")Suggested-by Will Deacon <[email protected]>Signed-off-by Jean-Philippe Brucker <[email protected]>---
An alternative to fix [1]. I haven't done uprobes yet since they don't suffer the same problem, but could add it to the bottom of my list. Lightly tested with kprobe smoke test and the BPF selftests. Interestingly on Seattle when running the "overhead" BPF selftest, that triggers a kprobe a bunch of times, I see a 20-30% improvement with this patch. I'm guessing it's the cost of touching the debug sysregs?
Quoted 7 lines
[1] https://lore.kernel.org/linux-arm-kernel/[email protected]/
--- arch/arm64/include/asm/brk-imm.h | 2 + arch/arm64/include/asm/debug-monitors.h | 1 + arch/arm64/include/asm/kprobes.h | 2 +- arch/arm64/kernel/probes/kprobes.c | 56 +++++++++---------------- 4 files changed, 24 insertions(+), 37 deletions(-)
Quoted 3 lines
Thanks! I'm about to send fixes for -rc2, but I'd like this to get a little bit of time in linux-next, so I aim to queue it next week for -rc3. Just wanted to make sure you don't think I'm ignored you.
No worries, I'm more confortable having this sit in linux-next
Thanks,
Jean
_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
Will Deacon
maintainerQuoted 113 lines · Jean-Philippe Brucker
Commit
36dadef23fcc("kprobes: Init kprobes inearly_initcall") enabled using kprobes fromearly_initcall. Unfortunately at this point the hardware debug infrastructure is not operational. The OS lock may still be locked, and the hardware watchpoints may have unknown values when kprobe enables debug monitors to single-step instructions.Rather than using hardware single-step, append a BRK instruction after the instruction to be executed out-of-line.
Fixes 36dadef23fcc ("kprobes: Init kprobes in early_initcall")Suggested-by Will Deacon <[email protected]>Signed-off-by Jean-Philippe Brucker <[email protected]>---
An alternative to fix [1]. I haven't done uprobes yet since they don't suffer the same problem, but could add it to the bottom of my list. Lightly tested with kprobe smoke test and the BPF selftests. Interestingly on Seattle when running the "overhead" BPF selftest, that triggers a kprobe a bunch of times, I see a 20-30% improvement with this patch. I'm guessing it's the cost of touching the debug sysregs?
[1] https://lore.kernel.org/linux-arm-kernel/[email protected]/
--- arch/arm64/include/asm/brk-imm.h | 2 + arch/arm64/include/asm/debug-monitors.h | 1 + arch/arm64/include/asm/kprobes.h | 2 +- arch/arm64/kernel/probes/kprobes.c | 56 +++++++++---------------- 4 files changed, 24 insertions(+), 37 deletions(-)arch/arm64/include/asm/brk-imm.h +2
-0@@ -10,6 +10,7 @@* #imm16 values used for BRK instruction generation * 0x004: for installing kprobes * 0x005: for installing uprobes +* 0x006: for kprobe software single-step * Allowed values for kgdb are 0x400 - 0x7ff * 0x100: for triggering a fault on purpose (reserved) * 0x400: for dynamic BRK instruction @@ -19,6 +20,7 @@*/ #define KPROBES_BRK_IMM 0x004#define UPROBES_BRK_IMM 0x005 +#define KPROBES_BRK_SS_IMM 0x006#define FAULT_BRK_IMM 0x100#define KGDB_DYN_DBG_BRK_IMM 0x400#define KGDB_COMPILED_DBG_BRK_IMM 0x401arch/arm64/include/asm/debug-monitors.h +1
-0@@ -53,6 +53,7 @@/* kprobes BRK opcodes with ESR encoding */ #define BRK64_OPCODE_KPROBES (AARCH64_BREAK_MON | (KPROBES_BRK_IMM << 5)) +#define BRK64_OPCODE_KPROBES_SS (AARCH64_BREAK_MON | (KPROBES_BRK_SS_IMM << 5))/* uprobes BRK opcodes with ESR encoding */ #define BRK64_OPCODE_UPROBES (AARCH64_BREAK_MON | (UPROBES_BRK_IMM << 5))arch/arm64/include/asm/kprobes.h +1
-1@@ -16,7 +16,7 @@#include <linux/percpu.h> #define __ARCH_WANT_KPROBES_INSN_SLOT -#define MAX_INSN_SIZE 1 +#define MAX_INSN_SIZE 2#define flush_insn_slot (p) do { } while (0)#define kretprobe_blacklist_size 0arch/arm64/kernel/probes/kprobes.c +10
-6@@ -36,21 +36,25 @@ DEFINE_PER_CPU(struct kprobe_ctlblk, kprobe_ctlblk);static void __kprobes post_kprobe_handler (struct kprobe_ctlblk *,struct pt_regs *); -static int __kprobes patch_text (kprobe_opcode_t *addr,u32 opcode) +static int __kprobes patch_text (kprobe_opcode_t *addr,u32 opc1,u32 opc2) { -void *addrs[1 ]; -u32 insns[1 ]; +void *addrs[2 ]; +u32 insns[2 ]; addrs[0 ] = addr; - insns[0 ] = opcode; + insns[0 ] = opc1; +if (opc2) { + addrs[1 ] = addr +1 ; + insns[1 ] = opc2; + } -return aarch64_insn_patch_text (addrs, insns,1 ); +return aarch64_insn_patch_text (addrs, insns, opc2 ?2 :1 ); }static void __kprobes arch_prepare_ss_slot (struct kprobe *p) {/* prepare insn slot */ -patch_text (p->ainsn.api.insn, p->opcode); +patch_text (p->ainsn.api.insn, p->opcode,BRK64_OPCODE_KPROBES_SS ); flush_icache_range((uintptr_t) (p->ainsn.api.insn), (uintptr_t) (p->ainsn.api.insn) +@@ -128,13 +132,13 @@ void *alloc_insn_page(void)/* arm kprobe: install breakpoint in text */ void __kprobes arch_arm_kprobe (struct kprobe *p)
arch/arm64/kernel/probes/kprobes.c void *alloc_insn_page(void) · line 141
Quoted 10 lines
arch/arm64/kernel/probes/kprobes.c +2
-2{ -patch_text (p->addr,BRK64_OPCODE_KPROBES ); +patch_text (p->addr,BRK64_OPCODE_KPROBES ,0 ); }/* disarm kprobe: remove breakpoint from text */ void __kprobes arch_disarm_kprobe (struct kprobe *p) { -patch_text (p->addr, p->opcode); +patch_text (p->addr, p->opcode,0 );
I don't think patch_text() is offering an awful lot to these callers. Why don't we just call aarch64_insn_patch_text() directly, which I think will make the code easier to read too?
Quoted 11 lines
arch/arm64/kernel/probes/kprobes.c +0
-6void __kprobes arch_remove_kprobe (struct kprobe *p)@@ -163,20 +167,14 @@ static void __kprobes set_current_kprobe(struct kprobe *p) }/* -* Interrupts need to be disabled before single-step mode is set, and not -* reenabled until after single-step mode ends. -* Without disabling interrupt on local CPU, there is a chance of -* interrupt occurrence in the period of exception return and start of -* out-of-line single-step, that result in wrongly single stepping -* into the interrupt handler.
arch/arm64/kernel/probes/kprobes.c static void __kprobes set_current_kprobe(struct kprobe *p) · line 179
Quoted 10 lines
arch/arm64/kernel/probes/kprobes.c +2
-2+* Keep interrupts disabled while executing the instruction out-of-line. Since +* the kprobe state is per-CPU, we can't afford to be migrated. */ static void __kprobes kprobes_save_local_irqflag (struct kprobe_ctlblk *kcb,struct pt_regs *regs) { kcb->saved_irqflag = regs->pstate &DAIF_MASK ; regs->pstate |=PSR_I_BIT ; -/* Unmask PSTATE.D for enabling software step exceptions. */ - regs->pstate &= ~PSR_D_BIT ;
Can we set the D bit now? Would be one less thing to worry about.
Cheers,
Will
_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
Quoted 119 lines · On Mon, 2 Nov 2020 11:41:52 +0000 Will Deacon <[email protected]> wrote:
Quoted 118 lines · Jean-Philippe Brucker
Commit
36dadef23fcc("kprobes: Init kprobes inearly_initcall") enabled using kprobes fromearly_initcall. Unfortunately at this point the hardware debug infrastructure is not operational. The OS lock may still be locked, and the hardware watchpoints may have unknown values when kprobe enables debug monitors to single-step instructions.Rather than using hardware single-step, append a BRK instruction after the instruction to be executed out-of-line.
Fixes 36dadef23fcc ("kprobes: Init kprobes in early_initcall")Suggested-by Will Deacon <[email protected]>Signed-off-by Jean-Philippe Brucker <[email protected]>---
An alternative to fix [1]. I haven't done uprobes yet since they don't suffer the same problem, but could add it to the bottom of my list. Lightly tested with kprobe smoke test and the BPF selftests. Interestingly on Seattle when running the "overhead" BPF selftest, that triggers a kprobe a bunch of times, I see a 20-30% improvement with this patch. I'm guessing it's the cost of touching the debug sysregs?
[1] https://lore.kernel.org/linux-arm-kernel/[email protected]/
--- arch/arm64/include/asm/brk-imm.h | 2 + arch/arm64/include/asm/debug-monitors.h | 1 + arch/arm64/include/asm/kprobes.h | 2 +- arch/arm64/kernel/probes/kprobes.c | 56 +++++++++---------------- 4 files changed, 24 insertions(+), 37 deletions(-)arch/arm64/include/asm/brk-imm.h +2
-0@@ -10,6 +10,7 @@* #imm16 values used for BRK instruction generation * 0x004: for installing kprobes * 0x005: for installing uprobes +* 0x006: for kprobe software single-step * Allowed values for kgdb are 0x400 - 0x7ff * 0x100: for triggering a fault on purpose (reserved) * 0x400: for dynamic BRK instruction @@ -19,6 +20,7 @@*/ #define KPROBES_BRK_IMM 0x004#define UPROBES_BRK_IMM 0x005 +#define KPROBES_BRK_SS_IMM 0x006#define FAULT_BRK_IMM 0x100#define KGDB_DYN_DBG_BRK_IMM 0x400#define KGDB_COMPILED_DBG_BRK_IMM 0x401arch/arm64/include/asm/debug-monitors.h +1
-0@@ -53,6 +53,7 @@/* kprobes BRK opcodes with ESR encoding */ #define BRK64_OPCODE_KPROBES (AARCH64_BREAK_MON | (KPROBES_BRK_IMM << 5)) +#define BRK64_OPCODE_KPROBES_SS (AARCH64_BREAK_MON | (KPROBES_BRK_SS_IMM << 5))/* uprobes BRK opcodes with ESR encoding */ #define BRK64_OPCODE_UPROBES (AARCH64_BREAK_MON | (UPROBES_BRK_IMM << 5))arch/arm64/include/asm/kprobes.h +1
-1@@ -16,7 +16,7 @@#include <linux/percpu.h> #define __ARCH_WANT_KPROBES_INSN_SLOT -#define MAX_INSN_SIZE 1 +#define MAX_INSN_SIZE 2#define flush_insn_slot (p) do { } while (0)#define kretprobe_blacklist_size 0arch/arm64/kernel/probes/kprobes.c +11
-7@@ -36,21 +36,25 @@ DEFINE_PER_CPU(struct kprobe_ctlblk, kprobe_ctlblk);static void __kprobes post_kprobe_handler (struct kprobe_ctlblk *,struct pt_regs *); -static int __kprobes patch_text (kprobe_opcode_t *addr,u32 opcode) +static int __kprobes patch_text (kprobe_opcode_t *addr,u32 opc1,u32 opc2) { -void *addrs[1 ]; -u32 insns[1 ]; +void *addrs[2 ]; +u32 insns[2 ]; addrs[0 ] = addr; - insns[0 ] = opcode; + insns[0 ] = opc1; +if (opc2) { + addrs[1 ] = addr +1 ; + insns[1 ] = opc2; + } -return aarch64_insn_patch_text (addrs, insns,1 ); +return aarch64_insn_patch_text (addrs, insns, opc2 ?2 :1 ); }static void __kprobes arch_prepare_ss_slot (struct kprobe *p) {/* prepare insn slot */ -patch_text (p->ainsn.api.insn, p->opcode); +patch_text (p->ainsn.api.insn, p->opcode,BRK64_OPCODE_KPROBES_SS ); flush_icache_range((uintptr_t) (p->ainsn.api.insn), (uintptr_t) (p->ainsn.api.insn) +@@ -128,13 +132,13 @@ void *alloc_insn_page(void)/* arm kprobe: install breakpoint in text */ void __kprobes arch_arm_kprobe (struct kprobe *p) { -patch_text (p->addr,BRK64_OPCODE_KPROBES ); +patch_text (p->addr,BRK64_OPCODE_KPROBES ,0 ); }
Quoted 9 lines
arch/arm64/kernel/probes/kprobes.c void *alloc_insn_page(void) · line 141
Quoted 5 lines
arch/arm64/kernel/probes/kprobes.c +1
-1/* disarm kprobe: remove breakpoint from text */ void __kprobes arch_disarm_kprobe (struct kprobe *p) { -patch_text (p->addr, p->opcode); +patch_text (p->addr, p->opcode,0 );I don't think
patch_text()is offering an awful lot to these callers. Why don't we just callaarch64_insn_patch_text()directly, which I think will make the code easier to read too?
Agreed. I guess it's just because histrically used.
Quoted 15 lines
Quoted 15 lines
arch/arm64/kernel/probes/kprobes.c +2
-6void __kprobes arch_remove_kprobe (struct kprobe *p)@@ -163,20 +167,14 @@ static void __kprobes set_current_kprobe(struct kprobe *p) }/* -* Interrupts need to be disabled before single-step mode is set, and not -* reenabled until after single-step mode ends. -* Without disabling interrupt on local CPU, there is a chance of -* interrupt occurrence in the period of exception return and start of -* out-of-line single-step, that result in wrongly single stepping -* into the interrupt handler. +* Keep interrupts disabled while executing the instruction out-of-line. Since +* the kprobe state is per-CPU, we can't afford to be migrated. */ static void __kprobes kprobes_save_local_irqflag(struct kprobe_ctlblk *kcb,
Quoted 8 lines
arch/arm64/kernel/probes/kprobes.c static void __kprobes set_current_kprobe(struct kprobe *p) · line 179
Quoted 6 lines
arch/arm64/kernel/probes/kprobes.c +0
-2struct pt_regs *regs) { kcb->saved_irqflag = regs->pstate &DAIF_MASK ; regs->pstate |=PSR_I_BIT ; -/* Unmask PSTATE.D for enabling software step exceptions. */ - regs->pstate &= ~PSR_D_BIT ;Can we set the D bit now? Would be one less thing to worry about.
Hmm, I think we just ignore D bit above. If someone (kgdb?) is set D flag then there may be no reason to prohibit it on the trampoline. (Of course they must handle this case)
Thank you,
Signature · 6 lines
Masami Hiramatsu <[email protected]> _______________________________________________ linux-arm-kernel mailing list [email protected] http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
Quoted 8 lines · Masami Hiramatsu
Quoted 6 lines
arch/arm64/kernel/probes/kprobes.c void *alloc_insn_page(void) · line 141
Quoted 2 lines
arch/arm64/kernel/probes/kprobes.c +1
-1-patch_text (p->addr, p->opcode); +patch_text (p->addr, p->opcode,0 );I don't think
patch_text()is offering an awful lot to these callers. Why don't we just callaarch64_insn_patch_text()directly, which I think will make the code easier to read too?Agreed. I guess it's just because histrically used.
Yes it's easy to remove, sending a v2
Quoted 25 lines
Quoted 16 lines
Quoted 16 lines
void
__kprobesarch_remove_kprobe(struct kprobe *p)arch/arm64/kernel/probes/kprobes.c +2
-6@@ -163,20 +167,14 @@ static void __kprobes set_current_kprobe(struct kprobe *p) }/* -* Interrupts need to be disabled before single-step mode is set, and not -* reenabled until after single-step mode ends. -* Without disabling interrupt on local CPU, there is a chance of -* interrupt occurrence in the period of exception return and start of -* out-of-line single-step, that result in wrongly single stepping -* into the interrupt handler. +* Keep interrupts disabled while executing the instruction out-of-line. Since +* the kprobe state is per-CPU, we can't afford to be migrated. */ static void __kprobes kprobes_save_local_irqflag(struct kprobe_ctlblk *kcb,Quoted 8 lines
arch/arm64/kernel/probes/kprobes.c static void __kprobes set_current_kprobe(struct kprobe *p) · line 179
Quoted 6 lines
arch/arm64/kernel/probes/kprobes.c +0
-2struct pt_regs *regs) { kcb->saved_irqflag = regs->pstate &DAIF_MASK ; regs->pstate |=PSR_I_BIT ; -/* Unmask PSTATE.D for enabling software step exceptions. */ - regs->pstate &= ~PSR_D_BIT ;Can we set the D bit now? Would be one less thing to worry about.
Quoted 3 lines
Hmm, I think we just ignore D bit above. If someone (kgdb?) is set D flag then there may be no reason to prohibit it on the trampoline. (Of course they must handle this case)
Masking debug just for the trampoline would simplify reasoning about nesting exceptions. Although it prevents from debugging kprobes using KGDB, it seems reasonable to me because I don't think we can provide guarantees that KGDB works reliably with kprobes. For example it doesn't seem like a watchpoint would fire if the instruction performing the access is simulated by kprobe.
Thanks,
Jean
_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
Will Deacon
maintainerQuoted 41 lines · Jean-Philippe Brucker
Quoted 8 lines · Masami Hiramatsu
Quoted 6 lines
arch/arm64/kernel/probes/kprobes.c void *alloc_insn_page(void) · line 141
Quoted 2 lines
arch/arm64/kernel/probes/kprobes.c +1
-1-patch_text (p->addr, p->opcode); +patch_text (p->addr, p->opcode,0 );I don't think
patch_text()is offering an awful lot to these callers. Why don't we just callaarch64_insn_patch_text()directly, which I think will make the code easier to read too?Agreed. I guess it's just because histrically used.
Yes it's easy to remove, sending a v2
Quoted 25 lines
Quoted 16 lines
Quoted 16 lines
void
__kprobesarch_remove_kprobe(struct kprobe *p)arch/arm64/kernel/probes/kprobes.c +2
-6@@ -163,20 +167,14 @@ static void __kprobes set_current_kprobe(struct kprobe *p) }/* -* Interrupts need to be disabled before single-step mode is set, and not -* reenabled until after single-step mode ends. -* Without disabling interrupt on local CPU, there is a chance of -* interrupt occurrence in the period of exception return and start of -* out-of-line single-step, that result in wrongly single stepping -* into the interrupt handler. +* Keep interrupts disabled while executing the instruction out-of-line. Since +* the kprobe state is per-CPU, we can't afford to be migrated. */ static void __kprobes kprobes_save_local_irqflag(struct kprobe_ctlblk *kcb,Quoted 8 lines
arch/arm64/kernel/probes/kprobes.c static void __kprobes set_current_kprobe(struct kprobe *p) · line 179
Quoted 6 lines
arch/arm64/kernel/probes/kprobes.c +0
-2struct pt_regs *regs) { kcb->saved_irqflag = regs->pstate &DAIF_MASK ; regs->pstate |=PSR_I_BIT ; -/* Unmask PSTATE.D for enabling software step exceptions. */ - regs->pstate &= ~PSR_D_BIT ;Can we set the D bit now? Would be one less thing to worry about.
Quoted 3 lines
Hmm, I think we just ignore D bit above. If someone (kgdb?) is set D flag then there may be no reason to prohibit it on the trampoline. (Of course they must handle this case)
Quoted 6 lines
Masking debug just for the trampoline would simplify reasoning about nesting exceptions. Although it prevents from debugging kprobes using KGDB, it seems reasonable to me because I don't think we can provide guarantees that KGDB works reliably with kprobes. For example it doesn't seem like a watchpoint would fire if the instruction performing the access is simulated by kprobe.
Yes, let's just set all of DAIF during the trampoline. Also, while I've got you, If you get a chance, I'd appreciate any feedback on my proposal for reworking our debug exception handling altogether:
https://lore.kernel.org/r/20200626095551.GA9312@willie-the-truck
Will
_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
Quoted 5 lines · Will Deacon
Yes, let's just set all of DAIF during the trampoline. Also, while I've got you, If you get a chance, I'd appreciate any feedback on my proposal for reworking our debug exception handling altogether:
https://lore.kernel.org/r/20200626095551.GA9312@willie-the-truck
Well, I stared at this for a while... It looks fine to me, but I don't have a full picture of the trap infrastructure (not sure whether you were asking me). I could help with testing if you get around to reworking it.
Quoted 6 lines
On taking an interrupt from EL1, stash MDSCR_EL1.SS in a pcpu variable and clear the register bit if it was set. Then unmask only D and leave I set. On return from the exception, set D and restore MDSCR_EL1.SS. If we decide to reschedule, unmask D (i.e. we only step into interrupts if we need a reschedule. Alternatively, we could skip the reschedule if we were stepping.)
Any specific reason to treat reschedule differently, or just to keep things simple? I'm asking because that could be a problem with the current code: when taking an interrupt while stepping EL1, we keep MDSCR_EL1.SS set and unmask D before calling the IRQ handler. The step exception might only be taken after the next context synchronization event (on QEMU it happens at an isb in the timer handler). If the IRQ handler doesn't happen to do any context synchronization and we reschedule, I guess the step exception could happen after the next eret?
Thanks,
Jean
_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
Quoted 18 lines · On Tue, 17 Nov 2020 18:28:12 +0100 Jean-Philippe Brucker…
Quoted 5 lines · Will Deacon
Yes, let's just set all of DAIF during the trampoline. Also, while I've got you, If you get a chance, I'd appreciate any feedback on my proposal for reworking our debug exception handling altogether:
https://lore.kernel.org/r/20200626095551.GA9312@willie-the-truck
Well, I stared at this for a while... It looks fine to me, but I don't have a full picture of the trap infrastructure (not sure whether you were asking me). I could help with testing if you get around to reworking it.
Quoted 6 lines
On taking an interrupt from EL1, stash MDSCR_EL1.SS in a pcpu variable and clear the register bit if it was set. Then unmask only D and leave I set. On return from the exception, set D and restore MDSCR_EL1.SS. If we decide to reschedule, unmask D (i.e. we only step into interrupts if we need a reschedule. Alternatively, we could skip the reschedule if we were stepping.)
Quoted 8 lines
Any specific reason to treat reschedule differently, or just to keep things simple? I'm asking because that could be a problem with the current code: when taking an interrupt while stepping EL1, we keep MDSCR_EL1.SS set and unmask D before calling the IRQ handler. The step exception might only be taken after the next context synchronization event (on QEMU it happens at an isb in the timer handler). If the IRQ handler doesn't happen to do any context synchronization and we reschedule, I guess the step exception could happen after the next eret?
Will, would you have any comment on this?
I guess if SS on EL1 enabled with kprobes, this may happen (SS means SS exception happens)
[orig-code(SS)]->[1st BRK](D set)->[copied insn(!SS)]->[2nd BRK](D unmask)->[orig-code(SS)]
What if we unmask D flag (current code), this may happen
[orig-code(SS)]->[1st BRK](D unmask)->[copied insn(SS)]->[2nd BRK](D unmask)->[orig-code(SS)]
But as you pointed, this may make things complex. (maybe debugger also needs to care about this case, because the single-stepped address is far from the original code, they may not be able to find correct source line.)
So keeping D flag set in this phase is good to me, until we solve the debugger-side support.
Thank you,
Signature · 6 lines
Masami Hiramatsu <[email protected]> _______________________________________________ linux-arm-kernel mailing list [email protected] http://lists.infradead.org/mailman/listinfo/linux-arm-kernel