int
fasttrap_pid_probe(struct reg *rp)
{
proc_t *p = curproc;
uintptr_t pc = rp->r_rip - 1;
uintptr_t new_pc = 0;
fasttrap_bucket_t *bucket;
#if defined(sun)
kmutex_t *pid_mtx;
#endif
fasttrap_tracepoint_t *tp, tp_local;
pid_t pid;
dtrace_icookie_t cookie;
uint_t is_enabled = 0;
/*
* It's possible that a user (in a veritable orgy of bad planning)
* could redirect this thread's flow of control before it reached the
* return probe fasttrap. In this case we need to kill the process
* since it's in a unrecoverable state.
*/
if (curthread->t_dtrace_step) {
ASSERT(curthread->t_dtrace_on);
fasttrap_sigtrap(p, curthread, pc);
return (0);
}
/*
* Clear all user tracing flags.
*/
curthread->t_dtrace_ft = 0;
curthread->t_dtrace_pc = 0;
curthread->t_dtrace_npc = 0;
curthread->t_dtrace_scrpc = 0;
curthread->t_dtrace_astpc = 0;
#ifdef __amd64
curthread->t_dtrace_regv = 0;
#endif
#if defined(sun)
/*
* Treat a child created by a call to vfork(2) as if it were its
* parent. We know that there's only one thread of control in such a
* process: this one.
*/
while (p->p_flag & SVFORK) {
p = p->p_parent;
}
#endif
PROC_LOCK(p);
_PHOLD(p);
pid = p->p_pid;
#if defined(sun)
pid_mtx = &cpu_core[CPU->cpu_id].cpuc_pid_lock;
mutex_enter(pid_mtx);
#endif
bucket = &fasttrap_tpoints.fth_table[FASTTRAP_TPOINTS_INDEX(pid, pc)];
/*
* Lookup the tracepoint that the process just hit.
*/
for (tp = bucket->ftb_data; tp != NULL; tp = tp->ftt_next) {
if (pid == tp->ftt_pid && pc == tp->ftt_pc &&
tp->ftt_proc->ftpc_acount != 0)
break;
}
/*
* If we couldn't find a matching tracepoint, either a tracepoint has
* been inserted without using the pid<pid> ioctl interface (see
* fasttrap_ioctl), or somehow we have mislaid this tracepoint.
*/
if (tp == NULL) {
#if defined(sun)
mutex_exit(pid_mtx);
#endif
_PRELE(p);
PROC_UNLOCK(p);
return (-1);
}
/*
* Set the program counter to the address of the traced instruction
* so that it looks right in ustack() output.
*/
rp->r_rip = pc;
if (tp->ftt_ids != NULL) {
fasttrap_id_t *id;
#ifdef __amd64
if (p->p_model == DATAMODEL_LP64) {
for (id = tp->ftt_ids; id != NULL; id = id->fti_next) {
fasttrap_probe_t *probe = id->fti_probe;
if (id->fti_ptype == DTFTP_ENTRY) {
/*
* We note that this was an entry
* probe to help ustack() find the
* first caller.
*/
cookie = dtrace_interrupt_disable();
DTRACE_CPUFLAG_SET(CPU_DTRACE_ENTRY);
dtrace_probe(probe->ftp_id, rp->r_rdi,
rp->r_rsi, rp->r_rdx, rp->r_rcx,
rp->r_r8);
DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_ENTRY);
dtrace_interrupt_enable(cookie);
} else if (id->fti_ptype == DTFTP_IS_ENABLED) {
/*
* Note that in this case, we don't
* call dtrace_probe() since it's only
* an artificial probe meant to change
* the flow of control so that it
* encounters the true probe.
*/
is_enabled = 1;
} else if (probe->ftp_argmap == NULL) {
dtrace_probe(probe->ftp_id, rp->r_rdi,
rp->r_rsi, rp->r_rdx, rp->r_rcx,
rp->r_r8);
} else {
uintptr_t t[5];
fasttrap_usdt_args64(probe, rp,
sizeof (t) / sizeof (t[0]), t);
dtrace_probe(probe->ftp_id, t[0], t[1],
t[2], t[3], t[4]);
}
}
} else {
#else /* __amd64 */
uintptr_t s0, s1, s2, s3, s4, s5;
uint32_t *stack = (uint32_t *)rp->r_esp;
/*
* In 32-bit mode, all arguments are passed on the
* stack. If this is a function entry probe, we need
* to skip the first entry on the stack as it
* represents the return address rather than a
* parameter to the function.
*/
s0 = fasttrap_fuword32_noerr(&stack[0]);
s1 = fasttrap_fuword32_noerr(&stack[1]);
s2 = fasttrap_fuword32_noerr(&stack[2]);
s3 = fasttrap_fuword32_noerr(&stack[3]);
s4 = fasttrap_fuword32_noerr(&stack[4]);
s5 = fasttrap_fuword32_noerr(&stack[5]);
for (id = tp->ftt_ids; id != NULL; id = id->fti_next) {
fasttrap_probe_t *probe = id->fti_probe;
if (id->fti_ptype == DTFTP_ENTRY) {
/*
* We note that this was an entry
* probe to help ustack() find the
* first caller.
*/
cookie = dtrace_interrupt_disable();
DTRACE_CPUFLAG_SET(CPU_DTRACE_ENTRY);
dtrace_probe(probe->ftp_id, s1, s2,
s3, s4, s5);
DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_ENTRY);
dtrace_interrupt_enable(cookie);
} else if (id->fti_ptype == DTFTP_IS_ENABLED) {
/*
* Note that in this case, we don't
* call dtrace_probe() since it's only
* an artificial probe meant to change
* the flow of control so that it
* encounters the true probe.
*/
is_enabled = 1;
} else if (probe->ftp_argmap == NULL) {
dtrace_probe(probe->ftp_id, s0, s1,
s2, s3, s4);
} else {
uint32_t t[5];
fasttrap_usdt_args32(probe, rp,
sizeof (t) / sizeof (t[0]), t);
dtrace_probe(probe->ftp_id, t[0], t[1],
t[2], t[3], t[4]);
}
}
#endif /* __amd64 */
#ifdef __amd64
}
#endif
}
/*
* We're about to do a bunch of work so we cache a local copy of
* the tracepoint to emulate the instruction, and then find the
* tracepoint again later if we need to light up any return probes.
*/
tp_local = *tp;
PROC_UNLOCK(p);
#if defined(sun)
mutex_exit(pid_mtx);
#endif
tp = &tp_local;
/*
* Set the program counter to appear as though the traced instruction
* had completely executed. This ensures that fasttrap_getreg() will
* report the expected value for REG_RIP.
*/
rp->r_rip = pc + tp->ftt_size;
/*
* If there's an is-enabled probe connected to this tracepoint it
* means that there was a 'xorl %eax, %eax' or 'xorq %rax, %rax'
* instruction that was placed there by DTrace when the binary was
* linked. As this probe is, in fact, enabled, we need to stuff 1
* into %eax or %rax. Accordingly, we can bypass all the instruction
* emulation logic since we know the inevitable result. It's possible
* that a user could construct a scenario where the 'is-enabled'
* probe was on some other instruction, but that would be a rather
* exotic way to shoot oneself in the foot.
*/
if (is_enabled) {
rp->r_rax = 1;
new_pc = rp->r_rip;
goto done;
}
/*
* We emulate certain types of instructions to ensure correctness
* (in the case of position dependent instructions) or optimize
* common cases. The rest we have the thread execute back in user-
* land.
*/
switch (tp->ftt_type) {
case FASTTRAP_T_RET:
case FASTTRAP_T_RET16:
{
uintptr_t dst = 0;
uintptr_t addr = 0;
int ret = 0;
/*
* We have to emulate _every_ facet of the behavior of a ret
* instruction including what happens if the load from %esp
* fails; in that case, we send a SIGSEGV.
*/
#ifdef __amd64
if (p->p_model == DATAMODEL_NATIVE) {
ret = dst = fasttrap_fulword((void *)rp->r_rsp);
addr = rp->r_rsp + sizeof (uintptr_t);
} else {
#endif
#ifdef __i386__
uint32_t dst32;
ret = dst32 = fasttrap_fuword32((void *)rp->r_esp);
dst = dst32;
addr = rp->r_esp + sizeof (uint32_t);
#endif
#ifdef __amd64
}
#endif
if (ret == -1) {
fasttrap_sigsegv(p, curthread, rp->r_rsp);
new_pc = pc;
break;
}
if (tp->ftt_type == FASTTRAP_T_RET16)
addr += tp->ftt_dest;
rp->r_rsp = addr;
new_pc = dst;
break;
}
case FASTTRAP_T_JCC:
{
uint_t taken = 0;
switch (tp->ftt_code) {
case FASTTRAP_JO:
taken = (rp->r_rflags & FASTTRAP_EFLAGS_OF) != 0;
break;
case FASTTRAP_JNO:
taken = (rp->r_rflags & FASTTRAP_EFLAGS_OF) == 0;
break;
case FASTTRAP_JB:
taken = (rp->r_rflags & FASTTRAP_EFLAGS_CF) != 0;
break;
case FASTTRAP_JAE:
taken = (rp->r_rflags & FASTTRAP_EFLAGS_CF) == 0;
break;
case FASTTRAP_JE:
taken = (rp->r_rflags & FASTTRAP_EFLAGS_ZF) != 0;
break;
case FASTTRAP_JNE:
taken = (rp->r_rflags & FASTTRAP_EFLAGS_ZF) == 0;
break;
case FASTTRAP_JBE:
taken = (rp->r_rflags & FASTTRAP_EFLAGS_CF) != 0 ||
(rp->r_rflags & FASTTRAP_EFLAGS_ZF) != 0;
break;
case FASTTRAP_JA:
taken = (rp->r_rflags & FASTTRAP_EFLAGS_CF) == 0 &&
(rp->r_rflags & FASTTRAP_EFLAGS_ZF) == 0;
break;
case FASTTRAP_JS:
taken = (rp->r_rflags & FASTTRAP_EFLAGS_SF) != 0;
break;
case FASTTRAP_JNS:
taken = (rp->r_rflags & FASTTRAP_EFLAGS_SF) == 0;
break;
case FASTTRAP_JP:
taken = (rp->r_rflags & FASTTRAP_EFLAGS_PF) != 0;
break;
case FASTTRAP_JNP:
taken = (rp->r_rflags & FASTTRAP_EFLAGS_PF) == 0;
break;
case FASTTRAP_JL:
taken = ((rp->r_rflags & FASTTRAP_EFLAGS_SF) == 0) !=
((rp->r_rflags & FASTTRAP_EFLAGS_OF) == 0);
break;
case FASTTRAP_JGE:
taken = ((rp->r_rflags & FASTTRAP_EFLAGS_SF) == 0) ==
((rp->r_rflags & FASTTRAP_EFLAGS_OF) == 0);
break;
case FASTTRAP_JLE:
taken = (rp->r_rflags & FASTTRAP_EFLAGS_ZF) != 0 ||
((rp->r_rflags & FASTTRAP_EFLAGS_SF) == 0) !=
((rp->r_rflags & FASTTRAP_EFLAGS_OF) == 0);
break;
case FASTTRAP_JG:
taken = (rp->r_rflags & FASTTRAP_EFLAGS_ZF) == 0 &&
((rp->r_rflags & FASTTRAP_EFLAGS_SF) == 0) ==
((rp->r_rflags & FASTTRAP_EFLAGS_OF) == 0);
break;
}
if (taken)
new_pc = tp->ftt_dest;
else
new_pc = pc + tp->ftt_size;
break;
}
case FASTTRAP_T_LOOP:
{
uint_t taken = 0;
#ifdef __amd64
greg_t cx = rp->r_rcx--;
#else
greg_t cx = rp->r_ecx--;
#endif
switch (tp->ftt_code) {
case FASTTRAP_LOOPNZ:
taken = (rp->r_rflags & FASTTRAP_EFLAGS_ZF) == 0 &&
cx != 0;
break;
case FASTTRAP_LOOPZ:
taken = (rp->r_rflags & FASTTRAP_EFLAGS_ZF) != 0 &&
cx != 0;
break;
case FASTTRAP_LOOP:
taken = (cx != 0);
break;
}
if (taken)
new_pc = tp->ftt_dest;
else
new_pc = pc + tp->ftt_size;
break;
}
case FASTTRAP_T_JCXZ:
{
#ifdef __amd64
greg_t cx = rp->r_rcx;
#else
greg_t cx = rp->r_ecx;
#endif
if (cx == 0)
new_pc = tp->ftt_dest;
else
new_pc = pc + tp->ftt_size;
break;
}
case FASTTRAP_T_PUSHL_EBP:
{
int ret = 0;
uintptr_t addr = 0;
#ifdef __amd64
if (p->p_model == DATAMODEL_NATIVE) {
addr = rp->r_rsp - sizeof (uintptr_t);
ret = fasttrap_sulword((void *)addr, &rp->r_rsp);
} else {
#endif
#ifdef __i386__
addr = rp->r_rsp - sizeof (uint32_t);
ret = fasttrap_suword32((void *)addr, &rp->r_rsp);
#endif
#ifdef __amd64
}
#endif
if (ret == -1) {
fasttrap_sigsegv(p, curthread, addr);
new_pc = pc;
break;
}
rp->r_rsp = addr;
new_pc = pc + tp->ftt_size;
break;
}
case FASTTRAP_T_NOP:
new_pc = pc + tp->ftt_size;
break;
case FASTTRAP_T_JMP:
case FASTTRAP_T_CALL:
if (tp->ftt_code == 0) {
new_pc = tp->ftt_dest;
} else {
#ifdef __amd64
uintptr_t value;
#endif
uintptr_t addr = tp->ftt_dest;
if (tp->ftt_base != FASTTRAP_NOREG)
addr += fasttrap_getreg(rp, tp->ftt_base);
if (tp->ftt_index != FASTTRAP_NOREG)
addr += fasttrap_getreg(rp, tp->ftt_index) <<
tp->ftt_scale;
if (tp->ftt_code == 1) {
/*
* If there's a segment prefix for this
* instruction, we'll need to check permissions
* and bounds on the given selector, and adjust
* the address accordingly.
*/
if (tp->ftt_segment != FASTTRAP_SEG_NONE &&
fasttrap_do_seg(tp, rp, &addr) != 0) {
fasttrap_sigsegv(p, curthread, addr);
new_pc = pc;
break;
}
#ifdef __amd64
if (p->p_model == DATAMODEL_NATIVE) {
if ((value = fasttrap_fulword((void *)addr))
== -1) {
fasttrap_sigsegv(p, curthread,
addr);
new_pc = pc;
break;
}
new_pc = value;
} else {
#endif
#ifdef __i386__
uint32_t value32;
addr = (uintptr_t)(uint32_t)addr;
if ((value32 = fasttrap_fuword32((void *)addr))
== -1) {
fasttrap_sigsegv(p, curthread,
addr);
new_pc = pc;
break;
}
new_pc = value32;
#endif
}
#ifdef __amd64
} else {
new_pc = addr;
}
#endif
}
/*
* If this is a call instruction, we need to push the return
* address onto the stack. If this fails, we send the process
* a SIGSEGV and reset the pc to emulate what would happen if
* this instruction weren't traced.
*/
if (tp->ftt_type == FASTTRAP_T_CALL) {
int ret = 0;
uintptr_t addr = 0, pcps;
#ifdef __amd64
if (p->p_model == DATAMODEL_NATIVE) {
addr = rp->r_rsp - sizeof (uintptr_t);
pcps = pc + tp->ftt_size;
ret = fasttrap_sulword((void *)addr, &pcps);
} else {
#endif
#ifdef __i386__
addr = rp->r_rsp - sizeof (uint32_t);
pcps = (uint32_t)(pc + tp->ftt_size);
ret = fasttrap_suword32((void *)addr, &pcps);
#endif
#ifdef __amd64
}
#endif
if (ret == -1) {
fasttrap_sigsegv(p, curthread, addr);
new_pc = pc;
break;
}
rp->r_rsp = addr;
}
break;
case FASTTRAP_T_COMMON:
{
uintptr_t addr;
#if defined(__amd64)
uint8_t scratch[2 * FASTTRAP_MAX_INSTR_SIZE + 22];
#else
uint8_t scratch[2 * FASTTRAP_MAX_INSTR_SIZE + 7];
#endif
uint_t i = 0;
#if defined(sun)
klwp_t *lwp = ttolwp(curthread);
#endif
/*
* Compute the address of the ulwp_t and step over the
* ul_self pointer. The method used to store the user-land
* thread pointer is very different on 32- and 64-bit
* kernels.
*/
#if defined(sun)
#if defined(__amd64)
if (p->p_model == DATAMODEL_LP64) {
addr = lwp->lwp_pcb.pcb_fsbase;
addr += sizeof (void *);
} else {
addr = lwp->lwp_pcb.pcb_gsbase;
addr += sizeof (caddr32_t);
}
#else
addr = USD_GETBASE(&lwp->lwp_pcb.pcb_gsdesc);
addr += sizeof (void *);
#endif
#endif /* sun */
#ifdef __i386__
addr = USD_GETBASE(&curthread->td_pcb->pcb_gsd);
#else
addr = curthread->td_pcb->pcb_gsbase;
#endif
addr += sizeof (void *);
/*
* Generic Instruction Tracing
* ---------------------------
*
* This is the layout of the scratch space in the user-land
* thread structure for our generated instructions.
*
* 32-bit mode bytes
* ------------------------ -----
* a: <original instruction> <= 15
* jmp <pc + tp->ftt_size> 5
* b: <original instruction> <= 15
* int T_DTRACE_RET 2
* -----
* <= 37
*
* 64-bit mode bytes
* ------------------------ -----
* a: <original instruction> <= 15
* jmp 0(%rip) 6
* <pc + tp->ftt_size> 8
* b: <original instruction> <= 15
* int T_DTRACE_RET 2
* -----
* <= 46
*
* The %pc is set to a, and curthread->t_dtrace_astpc is set
* to b. If we encounter a signal on the way out of the
* kernel, trap() will set %pc to curthread->t_dtrace_astpc
* so that we execute the original instruction and re-enter
* the kernel rather than redirecting to the next instruction.
*
* If there are return probes (so we know that we're going to
* need to reenter the kernel after executing the original
* instruction), the scratch space will just contain the
* original instruction followed by an interrupt -- the same
* data as at b.
*
* %rip-relative Addressing
* ------------------------
*
* There's a further complication in 64-bit mode due to %rip-
* relative addressing. While this is clearly a beneficial
* architectural decision for position independent code, it's
* hard not to see it as a personal attack against the pid
* provider since before there was a relatively small set of
* instructions to emulate; with %rip-relative addressing,
* almost every instruction can potentially depend on the
* address at which it's executed. Rather than emulating
* the broad spectrum of instructions that can now be
* position dependent, we emulate jumps and others as in
* 32-bit mode, and take a different tack for instructions
* using %rip-relative addressing.
*
* For every instruction that uses the ModRM byte, the
* in-kernel disassembler reports its location. We use the
* ModRM byte to identify that an instruction uses
* %rip-relative addressing and to see what other registers
* the instruction uses. To emulate those instructions,
* we modify the instruction to be %rax-relative rather than
* %rip-relative (or %rcx-relative if the instruction uses
* %rax; or %r8- or %r9-relative if the REX.B is present so
* we don't have to rewrite the REX prefix). We then load
* the value that %rip would have been into the scratch
* register and generate an instruction to reset the scratch
* register back to its original value. The instruction
* sequence looks like this:
*
* 64-mode %rip-relative bytes
* ------------------------ -----
* a: <modified instruction> <= 15
* movq $<value>, %<scratch> 6
* jmp 0(%rip) 6
* <pc + tp->ftt_size> 8
* b: <modified instruction> <= 15
* int T_DTRACE_RET 2
* -----
* 52
*
* We set curthread->t_dtrace_regv so that upon receiving
* a signal we can reset the value of the scratch register.
*/
ASSERT(tp->ftt_size < FASTTRAP_MAX_INSTR_SIZE);
curthread->t_dtrace_scrpc = addr;
bcopy(tp->ftt_instr, &scratch[i], tp->ftt_size);
i += tp->ftt_size;
#ifdef __amd64
if (tp->ftt_ripmode != 0) {
greg_t *reg = NULL;
ASSERT(p->p_model == DATAMODEL_LP64);
ASSERT(tp->ftt_ripmode &
(FASTTRAP_RIP_1 | FASTTRAP_RIP_2));
/*
* If this was a %rip-relative instruction, we change
* it to be either a %rax- or %rcx-relative
* instruction (depending on whether those registers
* are used as another operand; or %r8- or %r9-
* relative depending on the value of REX.B). We then
* set that register and generate a movq instruction
* to reset the value.
*/
if (tp->ftt_ripmode & FASTTRAP_RIP_X)
scratch[i++] = FASTTRAP_REX(1, 0, 0, 1);
else
scratch[i++] = FASTTRAP_REX(1, 0, 0, 0);
if (tp->ftt_ripmode & FASTTRAP_RIP_1)
scratch[i++] = FASTTRAP_MOV_EAX;
else
scratch[i++] = FASTTRAP_MOV_ECX;
switch (tp->ftt_ripmode) {
case FASTTRAP_RIP_1:
reg = &rp->r_rax;
curthread->t_dtrace_reg = REG_RAX;
break;
case FASTTRAP_RIP_2:
reg = &rp->r_rcx;
curthread->t_dtrace_reg = REG_RCX;
break;
case FASTTRAP_RIP_1 | FASTTRAP_RIP_X:
reg = &rp->r_r8;
curthread->t_dtrace_reg = REG_R8;
break;
case FASTTRAP_RIP_2 | FASTTRAP_RIP_X:
reg = &rp->r_r9;
curthread->t_dtrace_reg = REG_R9;
break;
}
/* LINTED - alignment */
*(uint64_t *)&scratch[i] = *reg;
curthread->t_dtrace_regv = *reg;
*reg = pc + tp->ftt_size;
i += sizeof (uint64_t);
}
#endif
/*
* Generate the branch instruction to what would have
* normally been the subsequent instruction. In 32-bit mode,
* this is just a relative branch; in 64-bit mode this is a
* %rip-relative branch that loads the 64-bit pc value
* immediately after the jmp instruction.
*/
#ifdef __amd64
if (p->p_model == DATAMODEL_LP64) {
scratch[i++] = FASTTRAP_GROUP5_OP;
scratch[i++] = FASTTRAP_MODRM(0, 4, 5);
/* LINTED - alignment */
*(uint32_t *)&scratch[i] = 0;
i += sizeof (uint32_t);
/* LINTED - alignment */
*(uint64_t *)&scratch[i] = pc + tp->ftt_size;
i += sizeof (uint64_t);
} else {
#endif
#ifdef __i386__
/*
* Set up the jmp to the next instruction; note that
* the size of the traced instruction cancels out.
*/
scratch[i++] = FASTTRAP_JMP32;
/* LINTED - alignment */
*(uint32_t *)&scratch[i] = pc - addr - 5;
i += sizeof (uint32_t);
#endif
#ifdef __amd64
}
#endif
curthread->t_dtrace_astpc = addr + i;
bcopy(tp->ftt_instr, &scratch[i], tp->ftt_size);
i += tp->ftt_size;
scratch[i++] = FASTTRAP_INT;
scratch[i++] = T_DTRACE_RET;
ASSERT(i <= sizeof (scratch));
#if defined(sun)
if (fasttrap_copyout(scratch, (char *)addr, i)) {
#else
if (uwrite(curproc, scratch, i, addr)) {
#endif
fasttrap_sigtrap(p, curthread, pc);
new_pc = pc;
break;
}
if (tp->ftt_retids != NULL) {
curthread->t_dtrace_step = 1;
curthread->t_dtrace_ret = 1;
new_pc = curthread->t_dtrace_astpc;
} else {
new_pc = curthread->t_dtrace_scrpc;
}
curthread->t_dtrace_pc = pc;
curthread->t_dtrace_npc = pc + tp->ftt_size;
curthread->t_dtrace_on = 1;
break;
}
default:
panic("fasttrap: mishandled an instruction");
}
done:
/*
* If there were no return probes when we first found the tracepoint,
* we should feel no obligation to honor any return probes that were
* subsequently enabled -- they'll just have to wait until the next
* time around.
*/
if (tp->ftt_retids != NULL) {
/*
* We need to wait until the results of the instruction are
* apparent before invoking any return probes. If this
* instruction was emulated we can just call
* fasttrap_return_common(); if it needs to be executed, we
* need to wait until the user thread returns to the kernel.
*/
if (tp->ftt_type != FASTTRAP_T_COMMON) {
/*
* Set the program counter to the address of the traced
* instruction so that it looks right in ustack()
* output. We had previously set it to the end of the
* instruction to simplify %rip-relative addressing.
*/
rp->r_rip = pc;
fasttrap_return_common(rp, pc, pid, new_pc);
} else {
ASSERT(curthread->t_dtrace_ret != 0);
ASSERT(curthread->t_dtrace_pc == pc);
ASSERT(curthread->t_dtrace_scrpc != 0);
ASSERT(new_pc == curthread->t_dtrace_astpc);
}
}
rp->r_rip = new_pc;
PROC_LOCK(p);
proc_write_regs(curthread, rp);
_PRELE(p);
PROC_UNLOCK(p);
return (0);
}
int
fasttrap_return_probe(struct reg *rp)
{
proc_t *p = curproc;
uintptr_t pc = curthread->t_dtrace_pc;
uintptr_t npc = curthread->t_dtrace_npc;
curthread->t_dtrace_pc = 0;
curthread->t_dtrace_npc = 0;
curthread->t_dtrace_scrpc = 0;
curthread->t_dtrace_astpc = 0;
#if defined(sun)
/*
* Treat a child created by a call to vfork(2) as if it were its
* parent. We know that there's only one thread of control in such a
* process: this one.
*/
while (p->p_flag & SVFORK) {
p = p->p_parent;
}
#endif
/*
* We set rp->r_rip to the address of the traced instruction so
* that it appears to dtrace_probe() that we're on the original
* instruction, and so that the user can't easily detect our
* complex web of lies. dtrace_return_probe() (our caller)
* will correctly set %pc after we return.
*/
rp->r_rip = pc;
fasttrap_return_common(rp, pc, p->p_pid, npc);
return (0);
}
int
kern_ptrace(struct thread *td, int req, pid_t pid, void *addr, int data)
{
struct iovec iov;
struct uio uio;
struct proc *curp, *p, *pp;
struct thread *td2;
struct ptrace_io_desc *piod;
int error, write, tmp;
int proctree_locked = 0;
curp = td->td_proc;
/* Lock proctree before locking the process. */
switch (req) {
case PT_TRACE_ME:
case PT_ATTACH:
case PT_STEP:
case PT_CONTINUE:
case PT_DETACH:
sx_xlock(&proctree_lock);
proctree_locked = 1;
break;
default:
break;
}
write = 0;
if (req == PT_TRACE_ME) {
p = td->td_proc;
PROC_LOCK(p);
} else {
if ((p = pfind(pid)) == NULL) {
if (proctree_locked)
sx_xunlock(&proctree_lock);
return (ESRCH);
}
}
if ((error = p_cansee(td, p)) != 0)
goto fail;
if ((error = p_candebug(td, p)) != 0)
goto fail;
/*
* System processes can't be debugged.
*/
if ((p->p_flag & P_SYSTEM) != 0) {
error = EINVAL;
goto fail;
}
/*
* Permissions check
*/
switch (req) {
case PT_TRACE_ME:
/* Always legal. */
break;
case PT_ATTACH:
/* Self */
if (p->p_pid == td->td_proc->p_pid) {
error = EINVAL;
goto fail;
}
/* Already traced */
if (p->p_flag & P_TRACED) {
error = EBUSY;
goto fail;
}
/* Can't trace an ancestor if you're being traced. */
if (curp->p_flag & P_TRACED) {
for (pp = curp->p_pptr; pp != NULL; pp = pp->p_pptr) {
if (pp == p) {
error = EINVAL;
goto fail;
}
}
}
/* OK */
break;
case PT_READ_I:
case PT_READ_D:
case PT_WRITE_I:
case PT_WRITE_D:
case PT_IO:
case PT_CONTINUE:
case PT_KILL:
case PT_STEP:
case PT_DETACH:
case PT_GETREGS:
case PT_SETREGS:
case PT_GETFPREGS:
case PT_SETFPREGS:
case PT_GETDBREGS:
case PT_SETDBREGS:
/* not being traced... */
if ((p->p_flag & P_TRACED) == 0) {
error = EPERM;
goto fail;
}
/* not being traced by YOU */
if (p->p_pptr != td->td_proc) {
error = EBUSY;
goto fail;
}
/* not currently stopped */
if (!P_SHOULDSTOP(p) || (p->p_flag & P_WAITED) == 0) {
error = EBUSY;
goto fail;
}
/* OK */
break;
default:
error = EINVAL;
goto fail;
}
td2 = FIRST_THREAD_IN_PROC(p);
#ifdef FIX_SSTEP
/*
* Single step fixup ala procfs
*/
FIX_SSTEP(td2); /* XXXKSE */
#endif
/*
* Actually do the requests
*/
td->td_retval[0] = 0;
switch (req) {
case PT_TRACE_ME:
/* set my trace flag and "owner" so it can read/write me */
p->p_flag |= P_TRACED;
p->p_oppid = p->p_pptr->p_pid;
PROC_UNLOCK(p);
sx_xunlock(&proctree_lock);
return (0);
case PT_ATTACH:
/* security check done above */
p->p_flag |= P_TRACED;
p->p_oppid = p->p_pptr->p_pid;
if (p->p_pptr != td->td_proc)
proc_reparent(p, td->td_proc);
data = SIGSTOP;
goto sendsig; /* in PT_CONTINUE below */
case PT_STEP:
case PT_CONTINUE:
case PT_DETACH:
/* XXX data is used even in the PT_STEP case. */
if (req != PT_STEP && (unsigned)data > _SIG_MAXSIG) {
error = EINVAL;
goto fail;
}
_PHOLD(p);
if (req == PT_STEP) {
error = ptrace_single_step(td2);
if (error) {
_PRELE(p);
goto fail;
}
}
if (addr != (void *)1) {
error = ptrace_set_pc(td2, (u_long)(uintfptr_t)addr);
if (error) {
_PRELE(p);
goto fail;
}
}
_PRELE(p);
if (req == PT_DETACH) {
/* reset process parent */
if (p->p_oppid != p->p_pptr->p_pid) {
struct proc *pp;
PROC_UNLOCK(p);
pp = pfind(p->p_oppid);
if (pp == NULL)
pp = initproc;
else
PROC_UNLOCK(pp);
PROC_LOCK(p);
proc_reparent(p, pp);
}
p->p_flag &= ~(P_TRACED | P_WAITED);
p->p_oppid = 0;
/* should we send SIGCHLD? */
}
sendsig:
if (proctree_locked)
sx_xunlock(&proctree_lock);
/* deliver or queue signal */
if (P_SHOULDSTOP(p)) {
p->p_xstat = data;
mtx_lock_spin(&sched_lock);
p->p_flag &= ~(P_STOPPED_TRACE|P_STOPPED_SIG);
thread_unsuspend(p);
setrunnable(td2); /* XXXKSE */
/* Need foreach kse in proc, ... make_kse_queued(). */
mtx_unlock_spin(&sched_lock);
} else if (data)
psignal(p, data);
PROC_UNLOCK(p);
return (0);
case PT_WRITE_I:
case PT_WRITE_D:
write = 1;
/* FALLTHROUGH */
case PT_READ_I:
case PT_READ_D:
PROC_UNLOCK(p);
tmp = 0;
/* write = 0 set above */
iov.iov_base = write ? (caddr_t)&data : (caddr_t)&tmp;
iov.iov_len = sizeof(int);
uio.uio_iov = &iov;
uio.uio_iovcnt = 1;
uio.uio_offset = (off_t)(uintptr_t)addr;
uio.uio_resid = sizeof(int);
uio.uio_segflg = UIO_SYSSPACE; /* i.e.: the uap */
uio.uio_rw = write ? UIO_WRITE : UIO_READ;
uio.uio_td = td;
error = proc_rwmem(p, &uio);
if (uio.uio_resid != 0) {
/*
* XXX proc_rwmem() doesn't currently return ENOSPC,
* so I think write() can bogusly return 0.
* XXX what happens for short writes? We don't want
* to write partial data.
* XXX proc_rwmem() returns EPERM for other invalid
* addresses. Convert this to EINVAL. Does this
* clobber returns of EPERM for other reasons?
*/
if (error == 0 || error == ENOSPC || error == EPERM)
error = EINVAL; /* EOF */
}
if (!write)
td->td_retval[0] = tmp;
return (error);
case PT_IO:
PROC_UNLOCK(p);
piod = addr;
iov.iov_base = piod->piod_addr;
iov.iov_len = piod->piod_len;
uio.uio_iov = &iov;
uio.uio_iovcnt = 1;
uio.uio_offset = (off_t)(uintptr_t)piod->piod_offs;
uio.uio_resid = piod->piod_len;
uio.uio_segflg = UIO_USERSPACE;
uio.uio_td = td;
switch (piod->piod_op) {
case PIOD_READ_D:
case PIOD_READ_I:
uio.uio_rw = UIO_READ;
break;
case PIOD_WRITE_D:
case PIOD_WRITE_I:
uio.uio_rw = UIO_WRITE;
break;
default:
return (EINVAL);
}
error = proc_rwmem(p, &uio);
piod->piod_len -= uio.uio_resid;
return (error);
case PT_KILL:
data = SIGKILL;
goto sendsig; /* in PT_CONTINUE above */
case PT_SETREGS:
_PHOLD(p);
error = proc_write_regs(td2, addr);
_PRELE(p);
PROC_UNLOCK(p);
return (error);
case PT_GETREGS:
_PHOLD(p);
error = proc_read_regs(td2, addr);
_PRELE(p);
PROC_UNLOCK(p);
return (error);
case PT_SETFPREGS:
_PHOLD(p);
error = proc_write_fpregs(td2, addr);
_PRELE(p);
PROC_UNLOCK(p);
return (error);
case PT_GETFPREGS:
_PHOLD(p);
error = proc_read_fpregs(td2, addr);
_PRELE(p);
PROC_UNLOCK(p);
return (error);
case PT_SETDBREGS:
_PHOLD(p);
error = proc_write_dbregs(td2, addr);
_PRELE(p);
PROC_UNLOCK(p);
return (error);
case PT_GETDBREGS:
_PHOLD(p);
error = proc_read_dbregs(td2, addr);
_PRELE(p);
PROC_UNLOCK(p);
return (error);
default:
KASSERT(0, ("unreachable code\n"));
break;
}
KASSERT(0, ("unreachable code\n"));
return (0);
fail:
PROC_UNLOCK(p);
if (proctree_locked)
sx_xunlock(&proctree_lock);
return (error);
}
