linux/arch/sparc/kernel/kprobes.c

594 lines
16 KiB
C
Raw Normal View History

/* arch/sparc64/kernel/kprobes.c
*
* Copyright (C) 2004 David S. Miller <davem@davemloft.net>
*/
#include <linux/kernel.h>
#include <linux/kprobes.h>
#include <linux/module.h>
#include <linux/kdebug.h>
#include <asm/signal.h>
#include <asm/cacheflush.h>
#include <asm/uaccess.h>
/* We do not have hardware single-stepping on sparc64.
* So we implement software single-stepping with breakpoint
* traps. The top-level scheme is similar to that used
* in the x86 kprobes implementation.
*
* In the kprobe->ainsn.insn[] array we store the original
* instruction at index zero and a break instruction at
* index one.
*
* When we hit a kprobe we:
* - Run the pre-handler
* - Remember "regs->tnpc" and interrupt level stored in
* "regs->tstate" so we can restore them later
* - Disable PIL interrupts
* - Set regs->tpc to point to kprobe->ainsn.insn[0]
* - Set regs->tnpc to point to kprobe->ainsn.insn[1]
* - Mark that we are actively in a kprobe
*
* At this point we wait for the second breakpoint at
* kprobe->ainsn.insn[1] to hit. When it does we:
* - Run the post-handler
* - Set regs->tpc to "remembered" regs->tnpc stored above,
* restore the PIL interrupt level in "regs->tstate" as well
* - Make any adjustments necessary to regs->tnpc in order
* to handle relative branches correctly. See below.
* - Mark that we are no longer actively in a kprobe.
*/
DEFINE_PER_CPU(struct kprobe *, current_kprobe) = NULL;
DEFINE_PER_CPU(struct kprobe_ctlblk, kprobe_ctlblk);
struct kretprobe_blackpoint kretprobe_blacklist[] = {{NULL, NULL}};
int __kprobes arch_prepare_kprobe(struct kprobe *p)
{
p->ainsn.insn[0] = *p->addr;
[SPARC64]: Fix several kprobes bugs. - relbranch_fixup(), for non-branches, would end up setting regs->tnpc incorrectly, in fact it would set it equal to regs->tpc which would cause that instruction to execute twice Also, if this is not a PC-relative branch, we should just leave regs->tnpc as-is. This covers cases like 'jmpl' which branch to absolute values. - To be absolutely %100 safe, we need to flush the instruction cache for all assignments to kprobe->ainsn.insn[], including cases like add_aggr_kprobe() - prev_kprobe's status field needs to be 'unsigned long' to match the type of the value it is saving - jprobes were totally broken: = jprobe_return() can run in the stack frame of the jprobe handler, or in an even deeper stack frame, thus we'll be in the wrong register window than the one from the original probe state. So unwind using 'restore' instructions, if necessary, right before we do the jprobe_return() breakpoint trap. = There is no reason to save/restore the register window saved at %sp at jprobe trigger time. Those registers cannot be modified by the jprobe handler. Also, this code was saving and restoring "sizeof (struct sparc_stackf)" bytes. Depending upon the caller, this could clobber unrelated stack frame pieces if there is only a basic 128-byte register window stored on the stack, without the argument save area. So just saving and restoring struct pt_regs is sufficient. = Kill the "jprobe_saved_esp", totally unused. Also, delete "jprobe_saved_regs_location", with the stack frame unwind now done explicitly by jprobe_return(), this check is superfluous. Signed-off-by: David S. Miller <davem@davemloft.net>
2006-12-10 10:42:03 +00:00
flushi(&p->ainsn.insn[0]);
p->ainsn.insn[1] = BREAKPOINT_INSTRUCTION_2;
[SPARC64]: Fix several kprobes bugs. - relbranch_fixup(), for non-branches, would end up setting regs->tnpc incorrectly, in fact it would set it equal to regs->tpc which would cause that instruction to execute twice Also, if this is not a PC-relative branch, we should just leave regs->tnpc as-is. This covers cases like 'jmpl' which branch to absolute values. - To be absolutely %100 safe, we need to flush the instruction cache for all assignments to kprobe->ainsn.insn[], including cases like add_aggr_kprobe() - prev_kprobe's status field needs to be 'unsigned long' to match the type of the value it is saving - jprobes were totally broken: = jprobe_return() can run in the stack frame of the jprobe handler, or in an even deeper stack frame, thus we'll be in the wrong register window than the one from the original probe state. So unwind using 'restore' instructions, if necessary, right before we do the jprobe_return() breakpoint trap. = There is no reason to save/restore the register window saved at %sp at jprobe trigger time. Those registers cannot be modified by the jprobe handler. Also, this code was saving and restoring "sizeof (struct sparc_stackf)" bytes. Depending upon the caller, this could clobber unrelated stack frame pieces if there is only a basic 128-byte register window stored on the stack, without the argument save area. So just saving and restoring struct pt_regs is sufficient. = Kill the "jprobe_saved_esp", totally unused. Also, delete "jprobe_saved_regs_location", with the stack frame unwind now done explicitly by jprobe_return(), this check is superfluous. Signed-off-by: David S. Miller <davem@davemloft.net>
2006-12-10 10:42:03 +00:00
flushi(&p->ainsn.insn[1]);
[PATCH] Move kprobe [dis]arming into arch specific code The architecture independent code of the current kprobes implementation is arming and disarming kprobes at registration time. The problem is that the code is assuming that arming and disarming is a just done by a simple write of some magic value to an address. This is problematic for ia64 where our instructions look more like structures, and we can not insert break points by just doing something like: *p->addr = BREAKPOINT_INSTRUCTION; The following patch to 2.6.12-rc4-mm2 adds two new architecture dependent functions: * void arch_arm_kprobe(struct kprobe *p) * void arch_disarm_kprobe(struct kprobe *p) and then adds the new functions for each of the architectures that already implement kprobes (spar64/ppc64/i386/x86_64). I thought arch_[dis]arm_kprobe was the most descriptive of what was really happening, but each of the architectures already had a disarm_kprobe() function that was really a "disarm and do some other clean-up items as needed when you stumble across a recursive kprobe." So... I took the liberty of changing the code that was calling disarm_kprobe() to call arch_disarm_kprobe(), and then do the cleanup in the block of code dealing with the recursive kprobe case. So far this patch as been tested on i386, x86_64, and ppc64, but still needs to be tested in sparc64. Signed-off-by: Rusty Lynch <rusty.lynch@intel.com> Signed-off-by: Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-23 07:09:25 +00:00
p->opcode = *p->addr;
return 0;
[PATCH] Move kprobe [dis]arming into arch specific code The architecture independent code of the current kprobes implementation is arming and disarming kprobes at registration time. The problem is that the code is assuming that arming and disarming is a just done by a simple write of some magic value to an address. This is problematic for ia64 where our instructions look more like structures, and we can not insert break points by just doing something like: *p->addr = BREAKPOINT_INSTRUCTION; The following patch to 2.6.12-rc4-mm2 adds two new architecture dependent functions: * void arch_arm_kprobe(struct kprobe *p) * void arch_disarm_kprobe(struct kprobe *p) and then adds the new functions for each of the architectures that already implement kprobes (spar64/ppc64/i386/x86_64). I thought arch_[dis]arm_kprobe was the most descriptive of what was really happening, but each of the architectures already had a disarm_kprobe() function that was really a "disarm and do some other clean-up items as needed when you stumble across a recursive kprobe." So... I took the liberty of changing the code that was calling disarm_kprobe() to call arch_disarm_kprobe(), and then do the cleanup in the block of code dealing with the recursive kprobe case. So far this patch as been tested on i386, x86_64, and ppc64, but still needs to be tested in sparc64. Signed-off-by: Rusty Lynch <rusty.lynch@intel.com> Signed-off-by: Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-23 07:09:25 +00:00
}
void __kprobes arch_arm_kprobe(struct kprobe *p)
[PATCH] Move kprobe [dis]arming into arch specific code The architecture independent code of the current kprobes implementation is arming and disarming kprobes at registration time. The problem is that the code is assuming that arming and disarming is a just done by a simple write of some magic value to an address. This is problematic for ia64 where our instructions look more like structures, and we can not insert break points by just doing something like: *p->addr = BREAKPOINT_INSTRUCTION; The following patch to 2.6.12-rc4-mm2 adds two new architecture dependent functions: * void arch_arm_kprobe(struct kprobe *p) * void arch_disarm_kprobe(struct kprobe *p) and then adds the new functions for each of the architectures that already implement kprobes (spar64/ppc64/i386/x86_64). I thought arch_[dis]arm_kprobe was the most descriptive of what was really happening, but each of the architectures already had a disarm_kprobe() function that was really a "disarm and do some other clean-up items as needed when you stumble across a recursive kprobe." So... I took the liberty of changing the code that was calling disarm_kprobe() to call arch_disarm_kprobe(), and then do the cleanup in the block of code dealing with the recursive kprobe case. So far this patch as been tested on i386, x86_64, and ppc64, but still needs to be tested in sparc64. Signed-off-by: Rusty Lynch <rusty.lynch@intel.com> Signed-off-by: Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-23 07:09:25 +00:00
{
*p->addr = BREAKPOINT_INSTRUCTION;
flushi(p->addr);
}
void __kprobes arch_disarm_kprobe(struct kprobe *p)
[PATCH] Move kprobe [dis]arming into arch specific code The architecture independent code of the current kprobes implementation is arming and disarming kprobes at registration time. The problem is that the code is assuming that arming and disarming is a just done by a simple write of some magic value to an address. This is problematic for ia64 where our instructions look more like structures, and we can not insert break points by just doing something like: *p->addr = BREAKPOINT_INSTRUCTION; The following patch to 2.6.12-rc4-mm2 adds two new architecture dependent functions: * void arch_arm_kprobe(struct kprobe *p) * void arch_disarm_kprobe(struct kprobe *p) and then adds the new functions for each of the architectures that already implement kprobes (spar64/ppc64/i386/x86_64). I thought arch_[dis]arm_kprobe was the most descriptive of what was really happening, but each of the architectures already had a disarm_kprobe() function that was really a "disarm and do some other clean-up items as needed when you stumble across a recursive kprobe." So... I took the liberty of changing the code that was calling disarm_kprobe() to call arch_disarm_kprobe(), and then do the cleanup in the block of code dealing with the recursive kprobe case. So far this patch as been tested on i386, x86_64, and ppc64, but still needs to be tested in sparc64. Signed-off-by: Rusty Lynch <rusty.lynch@intel.com> Signed-off-by: Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-23 07:09:25 +00:00
{
*p->addr = p->opcode;
flushi(p->addr);
}
static void __kprobes save_previous_kprobe(struct kprobe_ctlblk *kcb)
{
kcb->prev_kprobe.kp = kprobe_running();
kcb->prev_kprobe.status = kcb->kprobe_status;
kcb->prev_kprobe.orig_tnpc = kcb->kprobe_orig_tnpc;
kcb->prev_kprobe.orig_tstate_pil = kcb->kprobe_orig_tstate_pil;
}
static void __kprobes restore_previous_kprobe(struct kprobe_ctlblk *kcb)
{
__get_cpu_var(current_kprobe) = kcb->prev_kprobe.kp;
kcb->kprobe_status = kcb->prev_kprobe.status;
kcb->kprobe_orig_tnpc = kcb->prev_kprobe.orig_tnpc;
kcb->kprobe_orig_tstate_pil = kcb->prev_kprobe.orig_tstate_pil;
}
static void __kprobes set_current_kprobe(struct kprobe *p, struct pt_regs *regs,
struct kprobe_ctlblk *kcb)
{
__get_cpu_var(current_kprobe) = p;
kcb->kprobe_orig_tnpc = regs->tnpc;
kcb->kprobe_orig_tstate_pil = (regs->tstate & TSTATE_PIL);
}
static void __kprobes prepare_singlestep(struct kprobe *p, struct pt_regs *regs,
struct kprobe_ctlblk *kcb)
{
regs->tstate |= TSTATE_PIL;
/*single step inline, if it a breakpoint instruction*/
if (p->opcode == BREAKPOINT_INSTRUCTION) {
regs->tpc = (unsigned long) p->addr;
regs->tnpc = kcb->kprobe_orig_tnpc;
} else {
regs->tpc = (unsigned long) &p->ainsn.insn[0];
regs->tnpc = (unsigned long) &p->ainsn.insn[1];
}
}
static int __kprobes kprobe_handler(struct pt_regs *regs)
{
struct kprobe *p;
void *addr = (void *) regs->tpc;
int ret = 0;
struct kprobe_ctlblk *kcb;
/*
* We don't want to be preempted for the entire
* duration of kprobe processing
*/
preempt_disable();
kcb = get_kprobe_ctlblk();
if (kprobe_running()) {
p = get_kprobe(addr);
if (p) {
if (kcb->kprobe_status == KPROBE_HIT_SS) {
regs->tstate = ((regs->tstate & ~TSTATE_PIL) |
kcb->kprobe_orig_tstate_pil);
goto no_kprobe;
}
/* We have reentered the kprobe_handler(), since
* another probe was hit while within the handler.
* We here save the original kprobes variables and
* just single step on the instruction of the new probe
* without calling any user handlers.
*/
save_previous_kprobe(kcb);
set_current_kprobe(p, regs, kcb);
kprobes_inc_nmissed_count(p);
kcb->kprobe_status = KPROBE_REENTER;
prepare_singlestep(p, regs, kcb);
return 1;
} else {
if (*(u32 *)addr != BREAKPOINT_INSTRUCTION) {
/* The breakpoint instruction was removed by
* another cpu right after we hit, no further
* handling of this interrupt is appropriate
*/
ret = 1;
goto no_kprobe;
}
p = __get_cpu_var(current_kprobe);
if (p->break_handler && p->break_handler(p, regs))
goto ss_probe;
}
goto no_kprobe;
}
p = get_kprobe(addr);
if (!p) {
if (*(u32 *)addr != BREAKPOINT_INSTRUCTION) {
/*
* The breakpoint instruction was removed right
* after we hit it. Another cpu has removed
* either a probepoint or a debugger breakpoint
* at this address. In either case, no further
* handling of this interrupt is appropriate.
*/
ret = 1;
}
/* Not one of ours: let kernel handle it */
goto no_kprobe;
}
set_current_kprobe(p, regs, kcb);
kcb->kprobe_status = KPROBE_HIT_ACTIVE;
if (p->pre_handler && p->pre_handler(p, regs))
return 1;
ss_probe:
prepare_singlestep(p, regs, kcb);
kcb->kprobe_status = KPROBE_HIT_SS;
return 1;
no_kprobe:
preempt_enable_no_resched();
return ret;
}
/* If INSN is a relative control transfer instruction,
* return the corrected branch destination value.
*
[SPARC64]: Fix several kprobes bugs. - relbranch_fixup(), for non-branches, would end up setting regs->tnpc incorrectly, in fact it would set it equal to regs->tpc which would cause that instruction to execute twice Also, if this is not a PC-relative branch, we should just leave regs->tnpc as-is. This covers cases like 'jmpl' which branch to absolute values. - To be absolutely %100 safe, we need to flush the instruction cache for all assignments to kprobe->ainsn.insn[], including cases like add_aggr_kprobe() - prev_kprobe's status field needs to be 'unsigned long' to match the type of the value it is saving - jprobes were totally broken: = jprobe_return() can run in the stack frame of the jprobe handler, or in an even deeper stack frame, thus we'll be in the wrong register window than the one from the original probe state. So unwind using 'restore' instructions, if necessary, right before we do the jprobe_return() breakpoint trap. = There is no reason to save/restore the register window saved at %sp at jprobe trigger time. Those registers cannot be modified by the jprobe handler. Also, this code was saving and restoring "sizeof (struct sparc_stackf)" bytes. Depending upon the caller, this could clobber unrelated stack frame pieces if there is only a basic 128-byte register window stored on the stack, without the argument save area. So just saving and restoring struct pt_regs is sufficient. = Kill the "jprobe_saved_esp", totally unused. Also, delete "jprobe_saved_regs_location", with the stack frame unwind now done explicitly by jprobe_return(), this check is superfluous. Signed-off-by: David S. Miller <davem@davemloft.net>
2006-12-10 10:42:03 +00:00
* regs->tpc and regs->tnpc still hold the values of the
* program counters at the time of trap due to the execution
* of the BREAKPOINT_INSTRUCTION_2 at p->ainsn.insn[1]
*
*/
[SPARC64]: Fix several kprobes bugs. - relbranch_fixup(), for non-branches, would end up setting regs->tnpc incorrectly, in fact it would set it equal to regs->tpc which would cause that instruction to execute twice Also, if this is not a PC-relative branch, we should just leave regs->tnpc as-is. This covers cases like 'jmpl' which branch to absolute values. - To be absolutely %100 safe, we need to flush the instruction cache for all assignments to kprobe->ainsn.insn[], including cases like add_aggr_kprobe() - prev_kprobe's status field needs to be 'unsigned long' to match the type of the value it is saving - jprobes were totally broken: = jprobe_return() can run in the stack frame of the jprobe handler, or in an even deeper stack frame, thus we'll be in the wrong register window than the one from the original probe state. So unwind using 'restore' instructions, if necessary, right before we do the jprobe_return() breakpoint trap. = There is no reason to save/restore the register window saved at %sp at jprobe trigger time. Those registers cannot be modified by the jprobe handler. Also, this code was saving and restoring "sizeof (struct sparc_stackf)" bytes. Depending upon the caller, this could clobber unrelated stack frame pieces if there is only a basic 128-byte register window stored on the stack, without the argument save area. So just saving and restoring struct pt_regs is sufficient. = Kill the "jprobe_saved_esp", totally unused. Also, delete "jprobe_saved_regs_location", with the stack frame unwind now done explicitly by jprobe_return(), this check is superfluous. Signed-off-by: David S. Miller <davem@davemloft.net>
2006-12-10 10:42:03 +00:00
static unsigned long __kprobes relbranch_fixup(u32 insn, struct kprobe *p,
struct pt_regs *regs)
{
[SPARC64]: Fix several kprobes bugs. - relbranch_fixup(), for non-branches, would end up setting regs->tnpc incorrectly, in fact it would set it equal to regs->tpc which would cause that instruction to execute twice Also, if this is not a PC-relative branch, we should just leave regs->tnpc as-is. This covers cases like 'jmpl' which branch to absolute values. - To be absolutely %100 safe, we need to flush the instruction cache for all assignments to kprobe->ainsn.insn[], including cases like add_aggr_kprobe() - prev_kprobe's status field needs to be 'unsigned long' to match the type of the value it is saving - jprobes were totally broken: = jprobe_return() can run in the stack frame of the jprobe handler, or in an even deeper stack frame, thus we'll be in the wrong register window than the one from the original probe state. So unwind using 'restore' instructions, if necessary, right before we do the jprobe_return() breakpoint trap. = There is no reason to save/restore the register window saved at %sp at jprobe trigger time. Those registers cannot be modified by the jprobe handler. Also, this code was saving and restoring "sizeof (struct sparc_stackf)" bytes. Depending upon the caller, this could clobber unrelated stack frame pieces if there is only a basic 128-byte register window stored on the stack, without the argument save area. So just saving and restoring struct pt_regs is sufficient. = Kill the "jprobe_saved_esp", totally unused. Also, delete "jprobe_saved_regs_location", with the stack frame unwind now done explicitly by jprobe_return(), this check is superfluous. Signed-off-by: David S. Miller <davem@davemloft.net>
2006-12-10 10:42:03 +00:00
unsigned long real_pc = (unsigned long) p->addr;
/* Branch not taken, no mods necessary. */
[SPARC64]: Fix several kprobes bugs. - relbranch_fixup(), for non-branches, would end up setting regs->tnpc incorrectly, in fact it would set it equal to regs->tpc which would cause that instruction to execute twice Also, if this is not a PC-relative branch, we should just leave regs->tnpc as-is. This covers cases like 'jmpl' which branch to absolute values. - To be absolutely %100 safe, we need to flush the instruction cache for all assignments to kprobe->ainsn.insn[], including cases like add_aggr_kprobe() - prev_kprobe's status field needs to be 'unsigned long' to match the type of the value it is saving - jprobes were totally broken: = jprobe_return() can run in the stack frame of the jprobe handler, or in an even deeper stack frame, thus we'll be in the wrong register window than the one from the original probe state. So unwind using 'restore' instructions, if necessary, right before we do the jprobe_return() breakpoint trap. = There is no reason to save/restore the register window saved at %sp at jprobe trigger time. Those registers cannot be modified by the jprobe handler. Also, this code was saving and restoring "sizeof (struct sparc_stackf)" bytes. Depending upon the caller, this could clobber unrelated stack frame pieces if there is only a basic 128-byte register window stored on the stack, without the argument save area. So just saving and restoring struct pt_regs is sufficient. = Kill the "jprobe_saved_esp", totally unused. Also, delete "jprobe_saved_regs_location", with the stack frame unwind now done explicitly by jprobe_return(), this check is superfluous. Signed-off-by: David S. Miller <davem@davemloft.net>
2006-12-10 10:42:03 +00:00
if (regs->tnpc == regs->tpc + 0x4UL)
return real_pc + 0x8UL;
/* The three cases are call, branch w/prediction,
* and traditional branch.
*/
if ((insn & 0xc0000000) == 0x40000000 ||
(insn & 0xc1c00000) == 0x00400000 ||
(insn & 0xc1c00000) == 0x00800000) {
[SPARC64]: Fix several kprobes bugs. - relbranch_fixup(), for non-branches, would end up setting regs->tnpc incorrectly, in fact it would set it equal to regs->tpc which would cause that instruction to execute twice Also, if this is not a PC-relative branch, we should just leave regs->tnpc as-is. This covers cases like 'jmpl' which branch to absolute values. - To be absolutely %100 safe, we need to flush the instruction cache for all assignments to kprobe->ainsn.insn[], including cases like add_aggr_kprobe() - prev_kprobe's status field needs to be 'unsigned long' to match the type of the value it is saving - jprobes were totally broken: = jprobe_return() can run in the stack frame of the jprobe handler, or in an even deeper stack frame, thus we'll be in the wrong register window than the one from the original probe state. So unwind using 'restore' instructions, if necessary, right before we do the jprobe_return() breakpoint trap. = There is no reason to save/restore the register window saved at %sp at jprobe trigger time. Those registers cannot be modified by the jprobe handler. Also, this code was saving and restoring "sizeof (struct sparc_stackf)" bytes. Depending upon the caller, this could clobber unrelated stack frame pieces if there is only a basic 128-byte register window stored on the stack, without the argument save area. So just saving and restoring struct pt_regs is sufficient. = Kill the "jprobe_saved_esp", totally unused. Also, delete "jprobe_saved_regs_location", with the stack frame unwind now done explicitly by jprobe_return(), this check is superfluous. Signed-off-by: David S. Miller <davem@davemloft.net>
2006-12-10 10:42:03 +00:00
unsigned long ainsn_addr;
ainsn_addr = (unsigned long) &p->ainsn.insn[0];
/* The instruction did all the work for us
* already, just apply the offset to the correct
* instruction location.
*/
[SPARC64]: Fix several kprobes bugs. - relbranch_fixup(), for non-branches, would end up setting regs->tnpc incorrectly, in fact it would set it equal to regs->tpc which would cause that instruction to execute twice Also, if this is not a PC-relative branch, we should just leave regs->tnpc as-is. This covers cases like 'jmpl' which branch to absolute values. - To be absolutely %100 safe, we need to flush the instruction cache for all assignments to kprobe->ainsn.insn[], including cases like add_aggr_kprobe() - prev_kprobe's status field needs to be 'unsigned long' to match the type of the value it is saving - jprobes were totally broken: = jprobe_return() can run in the stack frame of the jprobe handler, or in an even deeper stack frame, thus we'll be in the wrong register window than the one from the original probe state. So unwind using 'restore' instructions, if necessary, right before we do the jprobe_return() breakpoint trap. = There is no reason to save/restore the register window saved at %sp at jprobe trigger time. Those registers cannot be modified by the jprobe handler. Also, this code was saving and restoring "sizeof (struct sparc_stackf)" bytes. Depending upon the caller, this could clobber unrelated stack frame pieces if there is only a basic 128-byte register window stored on the stack, without the argument save area. So just saving and restoring struct pt_regs is sufficient. = Kill the "jprobe_saved_esp", totally unused. Also, delete "jprobe_saved_regs_location", with the stack frame unwind now done explicitly by jprobe_return(), this check is superfluous. Signed-off-by: David S. Miller <davem@davemloft.net>
2006-12-10 10:42:03 +00:00
return (real_pc + (regs->tnpc - ainsn_addr));
}
[SPARC64]: Fix several kprobes bugs. - relbranch_fixup(), for non-branches, would end up setting regs->tnpc incorrectly, in fact it would set it equal to regs->tpc which would cause that instruction to execute twice Also, if this is not a PC-relative branch, we should just leave regs->tnpc as-is. This covers cases like 'jmpl' which branch to absolute values. - To be absolutely %100 safe, we need to flush the instruction cache for all assignments to kprobe->ainsn.insn[], including cases like add_aggr_kprobe() - prev_kprobe's status field needs to be 'unsigned long' to match the type of the value it is saving - jprobes were totally broken: = jprobe_return() can run in the stack frame of the jprobe handler, or in an even deeper stack frame, thus we'll be in the wrong register window than the one from the original probe state. So unwind using 'restore' instructions, if necessary, right before we do the jprobe_return() breakpoint trap. = There is no reason to save/restore the register window saved at %sp at jprobe trigger time. Those registers cannot be modified by the jprobe handler. Also, this code was saving and restoring "sizeof (struct sparc_stackf)" bytes. Depending upon the caller, this could clobber unrelated stack frame pieces if there is only a basic 128-byte register window stored on the stack, without the argument save area. So just saving and restoring struct pt_regs is sufficient. = Kill the "jprobe_saved_esp", totally unused. Also, delete "jprobe_saved_regs_location", with the stack frame unwind now done explicitly by jprobe_return(), this check is superfluous. Signed-off-by: David S. Miller <davem@davemloft.net>
2006-12-10 10:42:03 +00:00
/* It is jmpl or some other absolute PC modification instruction,
* leave NPC as-is.
*/
return regs->tnpc;
}
/* If INSN is an instruction which writes it's PC location
* into a destination register, fix that up.
*/
static void __kprobes retpc_fixup(struct pt_regs *regs, u32 insn,
unsigned long real_pc)
{
unsigned long *slot = NULL;
[SPARC64]: Fix several kprobes bugs. - relbranch_fixup(), for non-branches, would end up setting regs->tnpc incorrectly, in fact it would set it equal to regs->tpc which would cause that instruction to execute twice Also, if this is not a PC-relative branch, we should just leave regs->tnpc as-is. This covers cases like 'jmpl' which branch to absolute values. - To be absolutely %100 safe, we need to flush the instruction cache for all assignments to kprobe->ainsn.insn[], including cases like add_aggr_kprobe() - prev_kprobe's status field needs to be 'unsigned long' to match the type of the value it is saving - jprobes were totally broken: = jprobe_return() can run in the stack frame of the jprobe handler, or in an even deeper stack frame, thus we'll be in the wrong register window than the one from the original probe state. So unwind using 'restore' instructions, if necessary, right before we do the jprobe_return() breakpoint trap. = There is no reason to save/restore the register window saved at %sp at jprobe trigger time. Those registers cannot be modified by the jprobe handler. Also, this code was saving and restoring "sizeof (struct sparc_stackf)" bytes. Depending upon the caller, this could clobber unrelated stack frame pieces if there is only a basic 128-byte register window stored on the stack, without the argument save area. So just saving and restoring struct pt_regs is sufficient. = Kill the "jprobe_saved_esp", totally unused. Also, delete "jprobe_saved_regs_location", with the stack frame unwind now done explicitly by jprobe_return(), this check is superfluous. Signed-off-by: David S. Miller <davem@davemloft.net>
2006-12-10 10:42:03 +00:00
/* Simplest case is 'call', which always uses %o7 */
if ((insn & 0xc0000000) == 0x40000000) {
slot = &regs->u_regs[UREG_I7];
}
[SPARC64]: Fix several kprobes bugs. - relbranch_fixup(), for non-branches, would end up setting regs->tnpc incorrectly, in fact it would set it equal to regs->tpc which would cause that instruction to execute twice Also, if this is not a PC-relative branch, we should just leave regs->tnpc as-is. This covers cases like 'jmpl' which branch to absolute values. - To be absolutely %100 safe, we need to flush the instruction cache for all assignments to kprobe->ainsn.insn[], including cases like add_aggr_kprobe() - prev_kprobe's status field needs to be 'unsigned long' to match the type of the value it is saving - jprobes were totally broken: = jprobe_return() can run in the stack frame of the jprobe handler, or in an even deeper stack frame, thus we'll be in the wrong register window than the one from the original probe state. So unwind using 'restore' instructions, if necessary, right before we do the jprobe_return() breakpoint trap. = There is no reason to save/restore the register window saved at %sp at jprobe trigger time. Those registers cannot be modified by the jprobe handler. Also, this code was saving and restoring "sizeof (struct sparc_stackf)" bytes. Depending upon the caller, this could clobber unrelated stack frame pieces if there is only a basic 128-byte register window stored on the stack, without the argument save area. So just saving and restoring struct pt_regs is sufficient. = Kill the "jprobe_saved_esp", totally unused. Also, delete "jprobe_saved_regs_location", with the stack frame unwind now done explicitly by jprobe_return(), this check is superfluous. Signed-off-by: David S. Miller <davem@davemloft.net>
2006-12-10 10:42:03 +00:00
/* 'jmpl' encodes the register inside of the opcode */
if ((insn & 0xc1f80000) == 0x81c00000) {
unsigned long rd = ((insn >> 25) & 0x1f);
if (rd <= 15) {
slot = &regs->u_regs[rd];
} else {
/* Hard case, it goes onto the stack. */
flushw_all();
rd -= 16;
slot = (unsigned long *)
(regs->u_regs[UREG_FP] + STACK_BIAS);
slot += rd;
}
}
if (slot != NULL)
*slot = real_pc;
}
/*
* Called after single-stepping. p->addr is the address of the
[SPARC64]: Fix several kprobes bugs. - relbranch_fixup(), for non-branches, would end up setting regs->tnpc incorrectly, in fact it would set it equal to regs->tpc which would cause that instruction to execute twice Also, if this is not a PC-relative branch, we should just leave regs->tnpc as-is. This covers cases like 'jmpl' which branch to absolute values. - To be absolutely %100 safe, we need to flush the instruction cache for all assignments to kprobe->ainsn.insn[], including cases like add_aggr_kprobe() - prev_kprobe's status field needs to be 'unsigned long' to match the type of the value it is saving - jprobes were totally broken: = jprobe_return() can run in the stack frame of the jprobe handler, or in an even deeper stack frame, thus we'll be in the wrong register window than the one from the original probe state. So unwind using 'restore' instructions, if necessary, right before we do the jprobe_return() breakpoint trap. = There is no reason to save/restore the register window saved at %sp at jprobe trigger time. Those registers cannot be modified by the jprobe handler. Also, this code was saving and restoring "sizeof (struct sparc_stackf)" bytes. Depending upon the caller, this could clobber unrelated stack frame pieces if there is only a basic 128-byte register window stored on the stack, without the argument save area. So just saving and restoring struct pt_regs is sufficient. = Kill the "jprobe_saved_esp", totally unused. Also, delete "jprobe_saved_regs_location", with the stack frame unwind now done explicitly by jprobe_return(), this check is superfluous. Signed-off-by: David S. Miller <davem@davemloft.net>
2006-12-10 10:42:03 +00:00
* instruction which has been replaced by the breakpoint
* instruction. To avoid the SMP problems that can occur when we
* temporarily put back the original opcode to single-step, we
* single-stepped a copy of the instruction. The address of this
[SPARC64]: Fix several kprobes bugs. - relbranch_fixup(), for non-branches, would end up setting regs->tnpc incorrectly, in fact it would set it equal to regs->tpc which would cause that instruction to execute twice Also, if this is not a PC-relative branch, we should just leave regs->tnpc as-is. This covers cases like 'jmpl' which branch to absolute values. - To be absolutely %100 safe, we need to flush the instruction cache for all assignments to kprobe->ainsn.insn[], including cases like add_aggr_kprobe() - prev_kprobe's status field needs to be 'unsigned long' to match the type of the value it is saving - jprobes were totally broken: = jprobe_return() can run in the stack frame of the jprobe handler, or in an even deeper stack frame, thus we'll be in the wrong register window than the one from the original probe state. So unwind using 'restore' instructions, if necessary, right before we do the jprobe_return() breakpoint trap. = There is no reason to save/restore the register window saved at %sp at jprobe trigger time. Those registers cannot be modified by the jprobe handler. Also, this code was saving and restoring "sizeof (struct sparc_stackf)" bytes. Depending upon the caller, this could clobber unrelated stack frame pieces if there is only a basic 128-byte register window stored on the stack, without the argument save area. So just saving and restoring struct pt_regs is sufficient. = Kill the "jprobe_saved_esp", totally unused. Also, delete "jprobe_saved_regs_location", with the stack frame unwind now done explicitly by jprobe_return(), this check is superfluous. Signed-off-by: David S. Miller <davem@davemloft.net>
2006-12-10 10:42:03 +00:00
* copy is &p->ainsn.insn[0].
*
* This function prepares to return from the post-single-step
* breakpoint trap.
*/
static void __kprobes resume_execution(struct kprobe *p,
struct pt_regs *regs, struct kprobe_ctlblk *kcb)
{
u32 insn = p->ainsn.insn[0];
[SPARC64]: Fix several kprobes bugs. - relbranch_fixup(), for non-branches, would end up setting regs->tnpc incorrectly, in fact it would set it equal to regs->tpc which would cause that instruction to execute twice Also, if this is not a PC-relative branch, we should just leave regs->tnpc as-is. This covers cases like 'jmpl' which branch to absolute values. - To be absolutely %100 safe, we need to flush the instruction cache for all assignments to kprobe->ainsn.insn[], including cases like add_aggr_kprobe() - prev_kprobe's status field needs to be 'unsigned long' to match the type of the value it is saving - jprobes were totally broken: = jprobe_return() can run in the stack frame of the jprobe handler, or in an even deeper stack frame, thus we'll be in the wrong register window than the one from the original probe state. So unwind using 'restore' instructions, if necessary, right before we do the jprobe_return() breakpoint trap. = There is no reason to save/restore the register window saved at %sp at jprobe trigger time. Those registers cannot be modified by the jprobe handler. Also, this code was saving and restoring "sizeof (struct sparc_stackf)" bytes. Depending upon the caller, this could clobber unrelated stack frame pieces if there is only a basic 128-byte register window stored on the stack, without the argument save area. So just saving and restoring struct pt_regs is sufficient. = Kill the "jprobe_saved_esp", totally unused. Also, delete "jprobe_saved_regs_location", with the stack frame unwind now done explicitly by jprobe_return(), this check is superfluous. Signed-off-by: David S. Miller <davem@davemloft.net>
2006-12-10 10:42:03 +00:00
regs->tnpc = relbranch_fixup(insn, p, regs);
/* This assignment must occur after relbranch_fixup() */
regs->tpc = kcb->kprobe_orig_tnpc;
[SPARC64]: Fix several kprobes bugs. - relbranch_fixup(), for non-branches, would end up setting regs->tnpc incorrectly, in fact it would set it equal to regs->tpc which would cause that instruction to execute twice Also, if this is not a PC-relative branch, we should just leave regs->tnpc as-is. This covers cases like 'jmpl' which branch to absolute values. - To be absolutely %100 safe, we need to flush the instruction cache for all assignments to kprobe->ainsn.insn[], including cases like add_aggr_kprobe() - prev_kprobe's status field needs to be 'unsigned long' to match the type of the value it is saving - jprobes were totally broken: = jprobe_return() can run in the stack frame of the jprobe handler, or in an even deeper stack frame, thus we'll be in the wrong register window than the one from the original probe state. So unwind using 'restore' instructions, if necessary, right before we do the jprobe_return() breakpoint trap. = There is no reason to save/restore the register window saved at %sp at jprobe trigger time. Those registers cannot be modified by the jprobe handler. Also, this code was saving and restoring "sizeof (struct sparc_stackf)" bytes. Depending upon the caller, this could clobber unrelated stack frame pieces if there is only a basic 128-byte register window stored on the stack, without the argument save area. So just saving and restoring struct pt_regs is sufficient. = Kill the "jprobe_saved_esp", totally unused. Also, delete "jprobe_saved_regs_location", with the stack frame unwind now done explicitly by jprobe_return(), this check is superfluous. Signed-off-by: David S. Miller <davem@davemloft.net>
2006-12-10 10:42:03 +00:00
retpc_fixup(regs, insn, (unsigned long) p->addr);
regs->tstate = ((regs->tstate & ~TSTATE_PIL) |
kcb->kprobe_orig_tstate_pil);
}
static int __kprobes post_kprobe_handler(struct pt_regs *regs)
{
struct kprobe *cur = kprobe_running();
struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
if (!cur)
return 0;
if ((kcb->kprobe_status != KPROBE_REENTER) && cur->post_handler) {
kcb->kprobe_status = KPROBE_HIT_SSDONE;
cur->post_handler(cur, regs, 0);
}
resume_execution(cur, regs, kcb);
/*Restore back the original saved kprobes variables and continue. */
if (kcb->kprobe_status == KPROBE_REENTER) {
restore_previous_kprobe(kcb);
goto out;
}
reset_current_kprobe();
out:
preempt_enable_no_resched();
return 1;
}
int __kprobes kprobe_fault_handler(struct pt_regs *regs, int trapnr)
{
struct kprobe *cur = kprobe_running();
struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
const struct exception_table_entry *entry;
switch(kcb->kprobe_status) {
case KPROBE_HIT_SS:
case KPROBE_REENTER:
/*
* We are here because the instruction being single
* stepped caused a page fault. We reset the current
* kprobe and the tpc points back to the probe address
* and allow the page fault handler to continue as a
* normal page fault.
*/
regs->tpc = (unsigned long)cur->addr;
regs->tnpc = kcb->kprobe_orig_tnpc;
regs->tstate = ((regs->tstate & ~TSTATE_PIL) |
kcb->kprobe_orig_tstate_pil);
if (kcb->kprobe_status == KPROBE_REENTER)
restore_previous_kprobe(kcb);
else
reset_current_kprobe();
preempt_enable_no_resched();
break;
case KPROBE_HIT_ACTIVE:
case KPROBE_HIT_SSDONE:
/*
* We increment the nmissed count for accounting,
* we can also use npre/npostfault count for accouting
* these specific fault cases.
*/
kprobes_inc_nmissed_count(cur);
/*
* We come here because instructions in the pre/post
* handler caused the page_fault, this could happen
* if handler tries to access user space by
* copy_from_user(), get_user() etc. Let the
* user-specified handler try to fix it first.
*/
if (cur->fault_handler && cur->fault_handler(cur, regs, trapnr))
return 1;
/*
* In case the user-specified fault handler returned
* zero, try to fix up.
*/
entry = search_exception_tables(regs->tpc);
if (entry) {
regs->tpc = entry->fixup;
regs->tnpc = regs->tpc + 4;
return 1;
}
/*
* fixup_exception() could not handle it,
* Let do_page_fault() fix it.
*/
break;
default:
break;
}
return 0;
}
/*
* Wrapper routine to for handling exceptions.
*/
int __kprobes kprobe_exceptions_notify(struct notifier_block *self,
unsigned long val, void *data)
{
struct die_args *args = (struct die_args *)data;
2005-11-07 09:00:07 +00:00
int ret = NOTIFY_DONE;
if (args->regs && user_mode(args->regs))
return ret;
switch (val) {
case DIE_DEBUG:
if (kprobe_handler(args->regs))
2005-11-07 09:00:07 +00:00
ret = NOTIFY_STOP;
break;
case DIE_DEBUG_2:
if (post_kprobe_handler(args->regs))
2005-11-07 09:00:07 +00:00
ret = NOTIFY_STOP;
break;
default:
break;
}
2005-11-07 09:00:07 +00:00
return ret;
}
asmlinkage void __kprobes kprobe_trap(unsigned long trap_level,
struct pt_regs *regs)
{
BUG_ON(trap_level != 0x170 && trap_level != 0x171);
if (user_mode(regs)) {
local_irq_enable();
bad_trap(regs, trap_level);
return;
}
/* trap_level == 0x170 --> ta 0x70
* trap_level == 0x171 --> ta 0x71
*/
if (notify_die((trap_level == 0x170) ? DIE_DEBUG : DIE_DEBUG_2,
(trap_level == 0x170) ? "debug" : "debug_2",
regs, 0, trap_level, SIGTRAP) != NOTIFY_STOP)
bad_trap(regs, trap_level);
}
/* Jprobes support. */
int __kprobes setjmp_pre_handler(struct kprobe *p, struct pt_regs *regs)
{
struct jprobe *jp = container_of(p, struct jprobe, kp);
struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
memcpy(&(kcb->jprobe_saved_regs), regs, sizeof(*regs));
regs->tpc = (unsigned long) jp->entry;
regs->tnpc = ((unsigned long) jp->entry) + 0x4UL;
regs->tstate |= TSTATE_PIL;
return 1;
}
void __kprobes jprobe_return(void)
{
[SPARC64]: Fix several kprobes bugs. - relbranch_fixup(), for non-branches, would end up setting regs->tnpc incorrectly, in fact it would set it equal to regs->tpc which would cause that instruction to execute twice Also, if this is not a PC-relative branch, we should just leave regs->tnpc as-is. This covers cases like 'jmpl' which branch to absolute values. - To be absolutely %100 safe, we need to flush the instruction cache for all assignments to kprobe->ainsn.insn[], including cases like add_aggr_kprobe() - prev_kprobe's status field needs to be 'unsigned long' to match the type of the value it is saving - jprobes were totally broken: = jprobe_return() can run in the stack frame of the jprobe handler, or in an even deeper stack frame, thus we'll be in the wrong register window than the one from the original probe state. So unwind using 'restore' instructions, if necessary, right before we do the jprobe_return() breakpoint trap. = There is no reason to save/restore the register window saved at %sp at jprobe trigger time. Those registers cannot be modified by the jprobe handler. Also, this code was saving and restoring "sizeof (struct sparc_stackf)" bytes. Depending upon the caller, this could clobber unrelated stack frame pieces if there is only a basic 128-byte register window stored on the stack, without the argument save area. So just saving and restoring struct pt_regs is sufficient. = Kill the "jprobe_saved_esp", totally unused. Also, delete "jprobe_saved_regs_location", with the stack frame unwind now done explicitly by jprobe_return(), this check is superfluous. Signed-off-by: David S. Miller <davem@davemloft.net>
2006-12-10 10:42:03 +00:00
struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
register unsigned long orig_fp asm("g1");
orig_fp = kcb->jprobe_saved_regs.u_regs[UREG_FP];
__asm__ __volatile__("\n"
"1: cmp %%sp, %0\n\t"
"blu,a,pt %%xcc, 1b\n\t"
" restore\n\t"
".globl jprobe_return_trap_instruction\n"
"jprobe_return_trap_instruction:\n\t"
[SPARC64]: Fix several kprobes bugs. - relbranch_fixup(), for non-branches, would end up setting regs->tnpc incorrectly, in fact it would set it equal to regs->tpc which would cause that instruction to execute twice Also, if this is not a PC-relative branch, we should just leave regs->tnpc as-is. This covers cases like 'jmpl' which branch to absolute values. - To be absolutely %100 safe, we need to flush the instruction cache for all assignments to kprobe->ainsn.insn[], including cases like add_aggr_kprobe() - prev_kprobe's status field needs to be 'unsigned long' to match the type of the value it is saving - jprobes were totally broken: = jprobe_return() can run in the stack frame of the jprobe handler, or in an even deeper stack frame, thus we'll be in the wrong register window than the one from the original probe state. So unwind using 'restore' instructions, if necessary, right before we do the jprobe_return() breakpoint trap. = There is no reason to save/restore the register window saved at %sp at jprobe trigger time. Those registers cannot be modified by the jprobe handler. Also, this code was saving and restoring "sizeof (struct sparc_stackf)" bytes. Depending upon the caller, this could clobber unrelated stack frame pieces if there is only a basic 128-byte register window stored on the stack, without the argument save area. So just saving and restoring struct pt_regs is sufficient. = Kill the "jprobe_saved_esp", totally unused. Also, delete "jprobe_saved_regs_location", with the stack frame unwind now done explicitly by jprobe_return(), this check is superfluous. Signed-off-by: David S. Miller <davem@davemloft.net>
2006-12-10 10:42:03 +00:00
"ta 0x70"
: /* no outputs */
: "r" (orig_fp));
}
extern void jprobe_return_trap_instruction(void);
int __kprobes longjmp_break_handler(struct kprobe *p, struct pt_regs *regs)
{
u32 *addr = (u32 *) regs->tpc;
struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
if (addr == (u32 *) jprobe_return_trap_instruction) {
memcpy(regs, &(kcb->jprobe_saved_regs), sizeof(*regs));
preempt_enable_no_resched();
return 1;
}
return 0;
}
kprobes: improve kretprobe scalability with hashed locking Currently list of kretprobe instances are stored in kretprobe object (as used_instances,free_instances) and in kretprobe hash table. We have one global kretprobe lock to serialise the access to these lists. This causes only one kretprobe handler to execute at a time. Hence affects system performance, particularly on SMP systems and when return probe is set on lot of functions (like on all systemcalls). Solution proposed here gives fine-grain locks that performs better on SMP system compared to present kretprobe implementation. Solution: 1) Instead of having one global lock to protect kretprobe instances present in kretprobe object and kretprobe hash table. We will have two locks, one lock for protecting kretprobe hash table and another lock for kretporbe object. 2) We hold lock present in kretprobe object while we modify kretprobe instance in kretprobe object and we hold per-hash-list lock while modifying kretprobe instances present in that hash list. To prevent deadlock, we never grab a per-hash-list lock while holding a kretprobe lock. 3) We can remove used_instances from struct kretprobe, as we can track used instances of kretprobe instances using kretprobe hash table. Time duration for kernel compilation ("make -j 8") on a 8-way ppc64 system with return probes set on all systemcalls looks like this. cacheline non-cacheline Un-patched kernel aligned patch aligned patch =============================================================================== real 9m46.784s 9m54.412s 10m2.450s user 40m5.715s 40m7.142s 40m4.273s sys 2m57.754s 2m58.583s 3m17.430s =========================================================== Time duration for kernel compilation ("make -j 8) on the same system, when kernel is not probed. ========================= real 9m26.389s user 40m8.775s sys 2m7.283s ========================= Signed-off-by: Srinivasa DS <srinivasa@in.ibm.com> Signed-off-by: Jim Keniston <jkenisto@us.ibm.com> Acked-by: Ananth N Mavinakayanahalli <ananth@in.ibm.com> Cc: Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com> Cc: David S. Miller <davem@davemloft.net> Cc: Masami Hiramatsu <mhiramat@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-07-25 08:46:04 +00:00
/* The value stored in the return address register is actually 2
* instructions before where the callee will return to.
* Sequences usually look something like this
*
* call some_function <--- return register points here
* nop <--- call delay slot
* whatever <--- where callee returns to
*
* To keep trampoline_probe_handler logic simpler, we normalize the
* value kept in ri->ret_addr so we don't need to keep adjusting it
* back and forth.
*/
void __kprobes arch_prepare_kretprobe(struct kretprobe_instance *ri,
struct pt_regs *regs)
{
ri->ret_addr = (kprobe_opcode_t *)(regs->u_regs[UREG_RETPC] + 8);
/* Replace the return addr with trampoline addr */
regs->u_regs[UREG_RETPC] =
((unsigned long)kretprobe_trampoline) - 8;
}
/*
* Called when the probe at kretprobe trampoline is hit
*/
int __kprobes trampoline_probe_handler(struct kprobe *p, struct pt_regs *regs)
{
struct kretprobe_instance *ri = NULL;
struct hlist_head *head, empty_rp;
struct hlist_node *node, *tmp;
unsigned long flags, orig_ret_address = 0;
unsigned long trampoline_address =(unsigned long)&kretprobe_trampoline;
INIT_HLIST_HEAD(&empty_rp);
kprobes: improve kretprobe scalability with hashed locking Currently list of kretprobe instances are stored in kretprobe object (as used_instances,free_instances) and in kretprobe hash table. We have one global kretprobe lock to serialise the access to these lists. This causes only one kretprobe handler to execute at a time. Hence affects system performance, particularly on SMP systems and when return probe is set on lot of functions (like on all systemcalls). Solution proposed here gives fine-grain locks that performs better on SMP system compared to present kretprobe implementation. Solution: 1) Instead of having one global lock to protect kretprobe instances present in kretprobe object and kretprobe hash table. We will have two locks, one lock for protecting kretprobe hash table and another lock for kretporbe object. 2) We hold lock present in kretprobe object while we modify kretprobe instance in kretprobe object and we hold per-hash-list lock while modifying kretprobe instances present in that hash list. To prevent deadlock, we never grab a per-hash-list lock while holding a kretprobe lock. 3) We can remove used_instances from struct kretprobe, as we can track used instances of kretprobe instances using kretprobe hash table. Time duration for kernel compilation ("make -j 8") on a 8-way ppc64 system with return probes set on all systemcalls looks like this. cacheline non-cacheline Un-patched kernel aligned patch aligned patch =============================================================================== real 9m46.784s 9m54.412s 10m2.450s user 40m5.715s 40m7.142s 40m4.273s sys 2m57.754s 2m58.583s 3m17.430s =========================================================== Time duration for kernel compilation ("make -j 8) on the same system, when kernel is not probed. ========================= real 9m26.389s user 40m8.775s sys 2m7.283s ========================= Signed-off-by: Srinivasa DS <srinivasa@in.ibm.com> Signed-off-by: Jim Keniston <jkenisto@us.ibm.com> Acked-by: Ananth N Mavinakayanahalli <ananth@in.ibm.com> Cc: Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com> Cc: David S. Miller <davem@davemloft.net> Cc: Masami Hiramatsu <mhiramat@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-07-25 08:46:04 +00:00
kretprobe_hash_lock(current, &head, &flags);
/*
* It is possible to have multiple instances associated with a given
* task either because an multiple functions in the call path
* have a return probe installed on them, and/or more than one return
* return probe was registered for a target function.
*
* We can handle this because:
* - instances are always inserted at the head of the list
* - when multiple return probes are registered for the same
* function, the first instance's ret_addr will point to the
* real return address, and all the rest will point to
* kretprobe_trampoline
*/
hlist_for_each_entry_safe(ri, node, tmp, head, hlist) {
if (ri->task != current)
/* another task is sharing our hash bucket */
continue;
if (ri->rp && ri->rp->handler)
ri->rp->handler(ri, regs);
orig_ret_address = (unsigned long)ri->ret_addr;
recycle_rp_inst(ri, &empty_rp);
if (orig_ret_address != trampoline_address)
/*
* This is the real return address. Any other
* instances associated with this task are for
* other calls deeper on the call stack
*/
break;
}
kretprobe_assert(ri, orig_ret_address, trampoline_address);
regs->tpc = orig_ret_address;
regs->tnpc = orig_ret_address + 4;
reset_current_kprobe();
kprobes: improve kretprobe scalability with hashed locking Currently list of kretprobe instances are stored in kretprobe object (as used_instances,free_instances) and in kretprobe hash table. We have one global kretprobe lock to serialise the access to these lists. This causes only one kretprobe handler to execute at a time. Hence affects system performance, particularly on SMP systems and when return probe is set on lot of functions (like on all systemcalls). Solution proposed here gives fine-grain locks that performs better on SMP system compared to present kretprobe implementation. Solution: 1) Instead of having one global lock to protect kretprobe instances present in kretprobe object and kretprobe hash table. We will have two locks, one lock for protecting kretprobe hash table and another lock for kretporbe object. 2) We hold lock present in kretprobe object while we modify kretprobe instance in kretprobe object and we hold per-hash-list lock while modifying kretprobe instances present in that hash list. To prevent deadlock, we never grab a per-hash-list lock while holding a kretprobe lock. 3) We can remove used_instances from struct kretprobe, as we can track used instances of kretprobe instances using kretprobe hash table. Time duration for kernel compilation ("make -j 8") on a 8-way ppc64 system with return probes set on all systemcalls looks like this. cacheline non-cacheline Un-patched kernel aligned patch aligned patch =============================================================================== real 9m46.784s 9m54.412s 10m2.450s user 40m5.715s 40m7.142s 40m4.273s sys 2m57.754s 2m58.583s 3m17.430s =========================================================== Time duration for kernel compilation ("make -j 8) on the same system, when kernel is not probed. ========================= real 9m26.389s user 40m8.775s sys 2m7.283s ========================= Signed-off-by: Srinivasa DS <srinivasa@in.ibm.com> Signed-off-by: Jim Keniston <jkenisto@us.ibm.com> Acked-by: Ananth N Mavinakayanahalli <ananth@in.ibm.com> Cc: Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com> Cc: David S. Miller <davem@davemloft.net> Cc: Masami Hiramatsu <mhiramat@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-07-25 08:46:04 +00:00
kretprobe_hash_unlock(current, &flags);
preempt_enable_no_resched();
hlist_for_each_entry_safe(ri, node, tmp, &empty_rp, hlist) {
hlist_del(&ri->hlist);
kfree(ri);
}
/*
* By returning a non-zero value, we are telling
* kprobe_handler() that we don't want the post_handler
* to run (and have re-enabled preemption)
*/
return 1;
}
void kretprobe_trampoline_holder(void)
{
asm volatile(".global kretprobe_trampoline\n"
"kretprobe_trampoline:\n"
"\tnop\n"
"\tnop\n");
}
static struct kprobe trampoline_p = {
.addr = (kprobe_opcode_t *) &kretprobe_trampoline,
.pre_handler = trampoline_probe_handler
};
int __init arch_init_kprobes(void)
{
return register_kprobe(&trampoline_p);
}
int __kprobes arch_trampoline_kprobe(struct kprobe *p)
{
if (p->addr == (kprobe_opcode_t *)&kretprobe_trampoline)
return 1;
return 0;
}