freebsd-dev/sys/dev/hwpmc/hwpmc_x86.c
Randall Stewart d95b3509e1 Update the hwpmc driver to have the new type HASWELL_XEON. Also
go back through HASWELL, IVY_BRIDGE, IVY_BRIDGE_XEON and SANDY_BRIDGE
to straighten out all the missing PMCs. We also add a new pmc tool
pmcstudy, this allows one to run the various formulas from
the documents "Using Intel Vtune Amplifier XE on XXX Generation platforms" for
IB/SB and Haswell. The tool also allows one to postulate your own
formulas with any of the various PMC's. At some point I will enahance
this to work with Brendan Gregg's flame-graphs so we can flamegraph
various PMC interactions. Note the manual page also needs some
work (lots of work) but gnn has committed to help me with that ;-)
Reviewed by: gnn
MFC after:1 month
Sponsored by:	Netflix Inc.
2015-01-14 12:46:58 +00:00

279 lines
7.7 KiB
C

/*-
* Copyright (c) 2005,2008 Joseph Koshy
* Copyright (c) 2007 The FreeBSD Foundation
* All rights reserved.
*
* Portions of this software were developed by A. Joseph Koshy under
* sponsorship from the FreeBSD Foundation and Google, Inc.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* 1. Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* 2. Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
*
* THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
* ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
* ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
* OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
* HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
* LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
* OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
* SUCH DAMAGE.
*/
#include <sys/cdefs.h>
__FBSDID("$FreeBSD$");
#include <sys/param.h>
#include <sys/bus.h>
#include <sys/pmc.h>
#include <sys/proc.h>
#include <sys/systm.h>
#include <machine/cpu.h>
#include <machine/cputypes.h>
#include <machine/intr_machdep.h>
#if (__FreeBSD_version >= 1100000)
#include <x86/apicvar.h>
#else
#include <machine/apicvar.h>
#endif
#include <machine/pmc_mdep.h>
#include <machine/md_var.h>
#include <vm/vm.h>
#include <vm/vm_param.h>
#include <vm/pmap.h>
#include "hwpmc_soft.h"
/*
* Attempt to walk a user call stack using a too-simple algorithm.
* In the general case we need unwind information associated with
* the executable to be able to walk the user stack.
*
* We are handed a trap frame laid down at the time the PMC interrupt
* was taken. If the application is using frame pointers, the saved
* PC value could be:
* a. at the beginning of a function before the stack frame is laid
* down,
* b. just before a 'ret', after the stack frame has been taken off,
* c. somewhere else in the function with a valid stack frame being
* present,
*
* If the application is not using frame pointers, this algorithm will
* fail to yield an interesting call chain.
*
* TODO: figure out a way to use unwind information.
*/
int
pmc_save_user_callchain(uintptr_t *cc, int nframes, struct trapframe *tf)
{
int n;
uint32_t instr;
uintptr_t fp, oldfp, pc, r, sp;
KASSERT(TRAPF_USERMODE(tf), ("[x86,%d] Not a user trap frame tf=%p",
__LINE__, (void *) tf));
pc = PMC_TRAPFRAME_TO_PC(tf);
oldfp = fp = PMC_TRAPFRAME_TO_FP(tf);
sp = PMC_TRAPFRAME_TO_USER_SP(tf);
*cc++ = pc; n = 1;
r = fp + sizeof(uintptr_t); /* points to return address */
if (!PMC_IN_USERSPACE(pc))
return (n);
if (copyin((void *) pc, &instr, sizeof(instr)) != 0)
return (n);
if (PMC_AT_FUNCTION_PROLOGUE_PUSH_BP(instr) ||
PMC_AT_FUNCTION_EPILOGUE_RET(instr)) { /* ret */
if (copyin((void *) sp, &pc, sizeof(pc)) != 0)
return (n);
} else if (PMC_AT_FUNCTION_PROLOGUE_MOV_SP_BP(instr)) {
sp += sizeof(uintptr_t);
if (copyin((void *) sp, &pc, sizeof(pc)) != 0)
return (n);
} else if (copyin((void *) r, &pc, sizeof(pc)) != 0 ||
copyin((void *) fp, &fp, sizeof(fp)) != 0)
return (n);
for (; n < nframes;) {
if (pc == 0 || !PMC_IN_USERSPACE(pc))
break;
*cc++ = pc; n++;
if (fp < oldfp)
break;
r = fp + sizeof(uintptr_t); /* address of return address */
oldfp = fp;
if (copyin((void *) r, &pc, sizeof(pc)) != 0 ||
copyin((void *) fp, &fp, sizeof(fp)) != 0)
break;
}
return (n);
}
/*
* Walking the kernel call stack.
*
* We are handed the trap frame laid down at the time the PMC
* interrupt was taken. The saved PC could be:
* a. in the lowlevel trap handler, meaning that there isn't a C stack
* to traverse,
* b. at the beginning of a function before the stack frame is laid
* down,
* c. just before a 'ret', after the stack frame has been taken off,
* d. somewhere else in a function with a valid stack frame being
* present.
*
* In case (d), the previous frame pointer is at [%ebp]/[%rbp] and
* the return address is at [%ebp+4]/[%rbp+8].
*
* For cases (b) and (c), the return address is at [%esp]/[%rsp] and
* the frame pointer doesn't need to be changed when going up one
* level in the stack.
*
* For case (a), we check if the PC lies in low-level trap handling
* code, and if so we terminate our trace.
*/
int
pmc_save_kernel_callchain(uintptr_t *cc, int nframes, struct trapframe *tf)
{
int n;
uint32_t instr;
uintptr_t fp, pc, r, sp, stackstart, stackend;
struct thread *td;
KASSERT(TRAPF_USERMODE(tf) == 0,("[x86,%d] not a kernel backtrace",
__LINE__));
td = curthread;
pc = PMC_TRAPFRAME_TO_PC(tf);
fp = PMC_TRAPFRAME_TO_FP(tf);
sp = PMC_TRAPFRAME_TO_KERNEL_SP(tf);
*cc++ = pc;
r = fp + sizeof(uintptr_t); /* points to return address */
if (nframes <= 1)
return (1);
stackstart = (uintptr_t) td->td_kstack;
stackend = (uintptr_t) td->td_kstack + td->td_kstack_pages * PAGE_SIZE;
if (PMC_IN_TRAP_HANDLER(pc) ||
!PMC_IN_KERNEL(pc) ||
!PMC_IN_KERNEL_STACK(r, stackstart, stackend) ||
!PMC_IN_KERNEL_STACK(sp, stackstart, stackend) ||
!PMC_IN_KERNEL_STACK(fp, stackstart, stackend))
return (1);
instr = *(uint32_t *) pc;
/*
* Determine whether the interrupted function was in the
* processing of either laying down its stack frame or taking
* it off.
*
* If we haven't started laying down a stack frame, or are
* just about to return, then our caller's address is at
* *sp, and we don't have a frame to unwind.
*/
if (PMC_AT_FUNCTION_PROLOGUE_PUSH_BP(instr) ||
PMC_AT_FUNCTION_EPILOGUE_RET(instr))
pc = *(uintptr_t *) sp;
else if (PMC_AT_FUNCTION_PROLOGUE_MOV_SP_BP(instr)) {
/*
* The code was midway through laying down a frame.
* At this point sp[0] has a frame back pointer,
* and the caller's address is therefore at sp[1].
*/
sp += sizeof(uintptr_t);
if (!PMC_IN_KERNEL_STACK(sp, stackstart, stackend))
return (1);
pc = *(uintptr_t *) sp;
} else {
/*
* Not in the function prologue or epilogue.
*/
pc = *(uintptr_t *) r;
fp = *(uintptr_t *) fp;
}
for (n = 1; n < nframes; n++) {
*cc++ = pc;
if (PMC_IN_TRAP_HANDLER(pc))
break;
r = fp + sizeof(uintptr_t);
if (!PMC_IN_KERNEL_STACK(fp, stackstart, stackend) ||
!PMC_IN_KERNEL_STACK(r, stackstart, stackend))
break;
pc = *(uintptr_t *) r;
fp = *(uintptr_t *) fp;
}
return (n);
}
/*
* Machine dependent initialization for x86 class platforms.
*/
struct pmc_mdep *
pmc_md_initialize()
{
int i;
struct pmc_mdep *md;
/* determine the CPU kind */
if (cpu_vendor_id == CPU_VENDOR_AMD)
md = pmc_amd_initialize();
else if (cpu_vendor_id == CPU_VENDOR_INTEL)
md = pmc_intel_initialize();
else
return (NULL);
/* disallow sampling if we do not have an LAPIC */
if (md != NULL && !lapic_enable_pmc())
for (i = 0; i < md->pmd_nclass; i++) {
if (i == PMC_CLASS_INDEX_SOFT)
continue;
md->pmd_classdep[i].pcd_caps &= ~PMC_CAP_INTERRUPT;
}
return (md);
}
void
pmc_md_finalize(struct pmc_mdep *md)
{
lapic_disable_pmc();
if (cpu_vendor_id == CPU_VENDOR_AMD)
pmc_amd_finalize(md);
else if (cpu_vendor_id == CPU_VENDOR_INTEL)
pmc_intel_finalize(md);
else
KASSERT(0, ("[x86,%d] Unknown vendor", __LINE__));
}