Change gcore in order to get rid of the procfs accesses and use FreeBSD's

specific sysctls and ptrace interfaces.
This change switches a bit gcore POLA that is summarized here:
- now gcore can recognize threads within the process and handle dumps
  on thread-scope
- the process to be analyzed will be stopped during its gcore run
- gcore may not work with processes which are actively being analyzed
  by gdb or truss
- the ptrace interface may cause syscalls to return EINTR, thus
  interferring with signals handling within the process

Side note: <janitor task> the interface can be further lifted in order to
get rid of the very last procfs interfaces remnants and made more
suitable for copying with sysctl/ptrace interface </janitor task>.

Obtained from:	Sandvine Incorporated
Reviewed by:	emaste, rwatson
Sponsored by:	Sandvine Incorporated
MFC:		1 month
This commit is contained in:
Attilio Rao 2009-11-25 15:23:14 +00:00
parent 7a7043c787
commit 2e7ecbfbc8
Notes: svn2git 2020-12-20 02:59:44 +00:00
svn path=/head/; revision=199805
4 changed files with 182 additions and 195 deletions

View File

@ -1,6 +1,7 @@
# @(#)Makefile 8.1 (Berkeley) 6/6/93
# $FreeBSD$
LDADD+= -lutil
PROG= gcore
SRCS= elfcore.c gcore.c

View File

@ -1,4 +1,5 @@
/*-
* Copyright (c) 2007 Sandvine Incorporated
* Copyright (c) 1998 John D. Polstra
* All rights reserved.
*
@ -29,8 +30,12 @@ __FBSDID("$FreeBSD$");
#include <sys/param.h>
#include <sys/procfs.h>
#include <sys/ptrace.h>
#include <sys/queue.h>
#include <sys/linker_set.h>
#include <sys/sysctl.h>
#include <sys/user.h>
#include <sys/wait.h>
#include <machine/elf.h>
#include <vm/vm_param.h>
#include <vm/vm.h>
@ -44,6 +49,7 @@ __FBSDID("$FreeBSD$");
#include <stdlib.h>
#include <string.h>
#include <unistd.h>
#include <libutil.h>
#include "extern.h"
@ -69,16 +75,15 @@ static void cb_put_phdr(vm_map_entry_t, void *);
static void cb_size_segment(vm_map_entry_t, void *);
static void each_writable_segment(vm_map_entry_t, segment_callback,
void *closure);
static void elf_corehdr(int fd, pid_t, vm_map_entry_t, int numsegs,
void *hdr, size_t hdrsize);
static void elf_puthdr(vm_map_entry_t, void *, size_t *,
const prstatus_t *, const prfpregset_t *, const prpsinfo_t *, int numsegs);
static void elf_detach(void); /* atexit() handler. */
static void elf_puthdr(pid_t, vm_map_entry_t, void *, size_t *, int numsegs);
static void elf_putnote(void *dst, size_t *off, const char *name, int type,
const void *desc, size_t descsz);
static void freemap(vm_map_entry_t);
static void readhdrinfo(pid_t, prstatus_t *, prfpregset_t *, prpsinfo_t *);
static vm_map_entry_t readmap(pid_t);
static pid_t g_pid; /* Pid being dumped, global for elf_detach */
static int
elf_ident(int efd, pid_t pid __unused, char *binfile __unused)
{
@ -93,6 +98,14 @@ elf_ident(int efd, pid_t pid __unused, char *binfile __unused)
return (0);
}
static void
elf_detach(void)
{
if (g_pid != 0)
ptrace(PT_DETACH, g_pid, (caddr_t)1, 0);
}
/*
* Write an ELF coredump for the given pid to the given fd.
*/
@ -103,11 +116,20 @@ elf_coredump(int efd __unused, int fd, pid_t pid)
struct sseg_closure seginfo;
void *hdr;
size_t hdrsize;
char memname[64];
int memfd;
Elf_Phdr *php;
int i;
/* Attach to process to dump. */
g_pid = pid;
if (atexit(elf_detach) != 0)
err(1, "atexit");
errno = 0;
ptrace(PT_ATTACH, pid, NULL, 0);
if (errno)
err(1, "PT_ATTACH");
if (waitpid(pid, NULL, 0) == -1)
err(1, "waitpid");
/* Get the program's memory map. */
map = readmap(pid);
@ -122,28 +144,31 @@ elf_coredump(int efd __unused, int fd, pid_t pid)
* size is calculated.
*/
hdrsize = 0;
elf_puthdr(map, (void *)NULL, &hdrsize,
(const prstatus_t *)NULL, (const prfpregset_t *)NULL,
(const prpsinfo_t *)NULL, seginfo.count);
elf_puthdr(pid, map, NULL, &hdrsize, seginfo.count);
/*
* Allocate memory for building the header, fill it up,
* and write it out.
*/
if ((hdr = malloc(hdrsize)) == NULL)
if ((hdr = calloc(1, hdrsize)) == NULL)
errx(1, "out of memory");
elf_corehdr(fd, pid, map, seginfo.count, hdr, hdrsize);
/* Fill in the header. */
hdrsize = 0;
elf_puthdr(pid, map, hdr, &hdrsize, seginfo.count);
/* Write it to the core file. */
if (write(fd, hdr, hdrsize) == -1)
err(1, "write");
/* Write the contents of all of the writable segments. */
snprintf(memname, sizeof memname, "/proc/%d/mem", pid);
if ((memfd = open(memname, O_RDONLY)) == -1)
err(1, "cannot open %s", memname);
php = (Elf_Phdr *)((char *)hdr + sizeof(Elf_Ehdr)) + 1;
for (i = 0; i < seginfo.count; i++) {
struct ptrace_io_desc iorequest;
uintmax_t nleft = php->p_filesz;
lseek(memfd, (off_t)php->p_vaddr, SEEK_SET);
iorequest.piod_op = PIOD_READ_D;
iorequest.piod_offs = (caddr_t)php->p_vaddr;
while (nleft > 0) {
char buf[8*1024];
size_t nwant;
@ -153,12 +178,12 @@ elf_coredump(int efd __unused, int fd, pid_t pid)
nwant = sizeof buf;
else
nwant = nleft;
ngot = read(memfd, buf, nwant);
if (ngot == -1)
err(1, "read from %s", memname);
iorequest.piod_addr = buf;
iorequest.piod_len = nwant;
ptrace(PT_IO, pid, (caddr_t)&iorequest, 0);
ngot = iorequest.piod_len;
if ((size_t)ngot < nwant)
errx(1, "short read from %s:"
" wanted %zu, got %zd", memname,
errx(1, "short read wanted %d, got %d",
nwant, ngot);
ngot = write(fd, buf, nwant);
if (ngot == -1)
@ -166,10 +191,10 @@ elf_coredump(int efd __unused, int fd, pid_t pid)
if ((size_t)ngot != nwant)
errx(1, "short write");
nleft -= nwant;
iorequest.piod_offs += ngot;
}
php++;
}
close(memfd);
free(hdr);
freemap(map);
}
@ -231,30 +256,25 @@ each_writable_segment(vm_map_entry_t map, segment_callback func, void *closure)
(*func)(entry, closure);
}
/*
* Write the core file header to the file, including padding up to
* the page boundary.
*/
static void
elf_corehdr(int fd, pid_t pid, vm_map_entry_t map, int numsegs, void *hdr,
size_t hdrsize)
elf_getstatus(pid_t pid, prpsinfo_t *psinfo)
{
size_t off;
prstatus_t status;
prfpregset_t fpregset;
prpsinfo_t psinfo;
struct kinfo_proc kobj;
int name[4];
size_t len;
/* Gather the information for the header. */
readhdrinfo(pid, &status, &fpregset, &psinfo);
name[0] = CTL_KERN;
name[1] = KERN_PROC;
name[2] = KERN_PROC_PID;
name[3] = pid;
/* Fill in the header. */
memset(hdr, 0, hdrsize);
off = 0;
elf_puthdr(map, hdr, &off, &status, &fpregset, &psinfo, numsegs);
/* Write it to the core file. */
if (write(fd, hdr, hdrsize) == -1)
err(1, "write");
len = sizeof(kobj);
if (sysctl(name, 4, &kobj, &len, NULL, 0) == -1)
err(1, "error accessing kern.proc.pid.%u sysctl", pid);
if (kobj.ki_pid != pid)
err(1, "error accessing kern.proc.pid.%u sysctl datas", pid);
strncpy(psinfo->pr_fname, kobj.ki_comm, MAXCOMLEN);
strncpy(psinfo->pr_psargs, psinfo->pr_fname, PRARGSZ);
}
/*
@ -262,13 +282,24 @@ elf_corehdr(int fd, pid_t pid, vm_map_entry_t map, int numsegs, void *hdr,
* be NULL, in which case the header is sized but not actually generated.
*/
static void
elf_puthdr(vm_map_entry_t map, void *dst, size_t *off, const prstatus_t *status,
const prfpregset_t *fpregset, const prpsinfo_t *psinfo, int numsegs)
elf_puthdr(pid_t pid, vm_map_entry_t map, void *dst, size_t *off, int numsegs)
{
struct {
prstatus_t status;
prfpregset_t fpregset;
prpsinfo_t psinfo;
} *tempdata;
size_t ehoff;
size_t phoff;
size_t noteoff;
size_t notesz;
size_t threads;
lwpid_t *tids;
int i;
prstatus_t *status;
prfpregset_t *fpregset;
prpsinfo_t *psinfo;
ehoff = *off;
*off += sizeof(Elf_Ehdr);
@ -277,14 +308,68 @@ elf_puthdr(vm_map_entry_t map, void *dst, size_t *off, const prstatus_t *status,
*off += (numsegs + 1) * sizeof(Elf_Phdr);
noteoff = *off;
elf_putnote(dst, off, "FreeBSD", NT_PRSTATUS, status,
sizeof *status);
elf_putnote(dst, off, "FreeBSD", NT_FPREGSET, fpregset,
sizeof *fpregset);
if (dst != NULL) {
if ((tempdata = calloc(1, sizeof(*tempdata))) == NULL)
errx(1, "out of memory");
status = &tempdata->status;
fpregset = &tempdata->fpregset;
psinfo = &tempdata->psinfo;
} else {
tempdata = NULL;
status = NULL;
fpregset = NULL;
psinfo = NULL;
}
errno = 0;
threads = ptrace(PT_GETNUMLWPS, pid, NULL, 0);
if (errno)
err(1, "PT_GETNUMLWPS");
if (dst != NULL) {
psinfo->pr_version = PRPSINFO_VERSION;
psinfo->pr_psinfosz = sizeof(prpsinfo_t);
elf_getstatus(pid, psinfo);
}
elf_putnote(dst, off, "FreeBSD", NT_PRPSINFO, psinfo,
sizeof *psinfo);
if (dst != NULL) {
tids = malloc(threads * sizeof(*tids));
if (tids == NULL)
errx(1, "out of memory");
errno = 0;
ptrace(PT_GETLWPLIST, pid, (void *)tids, threads);
if (errno)
err(1, "PT_GETLWPLIST");
}
for (i = 0; i < threads; ++i) {
if (dst != NULL) {
status->pr_version = PRSTATUS_VERSION;
status->pr_statussz = sizeof(prstatus_t);
status->pr_gregsetsz = sizeof(gregset_t);
status->pr_fpregsetsz = sizeof(fpregset_t);
status->pr_osreldate = __FreeBSD_version;
status->pr_pid = tids[i];
ptrace(PT_GETREGS, tids[i], (void *)&status->pr_reg, 0);
ptrace(PT_GETFPREGS, tids[i], (void *)fpregset, 0);
}
elf_putnote(dst, off, "FreeBSD", NT_PRSTATUS, status,
sizeof *status);
elf_putnote(dst, off, "FreeBSD", NT_FPREGSET, fpregset,
sizeof *fpregset);
}
notesz = *off - noteoff;
if (dst != NULL) {
free(tids);
free(tempdata);
}
/* Align up to a page boundary for the program segments. */
*off = round_page(*off);
@ -381,70 +466,7 @@ freemap(vm_map_entry_t map)
}
/*
* Read the process information necessary to fill in the core file's header.
*/
static void
readhdrinfo(pid_t pid, prstatus_t *status, prfpregset_t *fpregset,
prpsinfo_t *psinfo)
{
char name[64];
char line[256];
int fd;
int i;
int n;
memset(status, 0, sizeof *status);
status->pr_version = PRSTATUS_VERSION;
status->pr_statussz = sizeof(prstatus_t);
status->pr_gregsetsz = sizeof(gregset_t);
status->pr_fpregsetsz = sizeof(fpregset_t);
status->pr_osreldate = __FreeBSD_version;
status->pr_pid = pid;
memset(fpregset, 0, sizeof *fpregset);
memset(psinfo, 0, sizeof *psinfo);
psinfo->pr_version = PRPSINFO_VERSION;
psinfo->pr_psinfosz = sizeof(prpsinfo_t);
/* Read the general registers. */
snprintf(name, sizeof name, "/proc/%d/regs", pid);
if ((fd = open(name, O_RDONLY)) == -1)
err(1, "cannot open %s", name);
if ((n = read(fd, &status->pr_reg, sizeof status->pr_reg)) == -1)
err(1, "read error from %s", name);
if ((size_t)n < sizeof(status->pr_reg))
errx(1, "short read from %s: wanted %zu, got %d", name,
sizeof status->pr_reg, n);
close(fd);
/* Read the floating point registers. */
snprintf(name, sizeof name, "/proc/%d/fpregs", pid);
if ((fd = open(name, O_RDONLY)) == -1)
err(1, "cannot open %s", name);
if ((n = read(fd, fpregset, sizeof *fpregset)) == -1)
err(1, "read error from %s", name);
if ((size_t)n < sizeof(*fpregset))
errx(1, "short read from %s: wanted %zu, got %d", name,
sizeof *fpregset, n);
close(fd);
/* Read and parse the process status. */
snprintf(name, sizeof name, "/proc/%d/status", pid);
if ((fd = open(name, O_RDONLY)) == -1)
err(1, "cannot open %s", name);
if ((n = read(fd, line, sizeof line - 1)) == -1)
err(1, "read error from %s", name);
if (n > MAXCOMLEN)
n = MAXCOMLEN;
for (i = 0; i < n && line[i] != ' '; i++)
psinfo->pr_fname[i] = line[i];
strncpy(psinfo->pr_psargs, psinfo->pr_fname, PRARGSZ);
close(fd);
}
/*
* Read the process's memory map using procfs, and return a list of
* Read the process's memory map using kinfo_getvmmap(), and return a list of
* VM map entries. Only the non-device read/writable segments are
* returned. The map entries in the list aren't fully filled in; only
* the items we need are present.
@ -452,83 +474,44 @@ readhdrinfo(pid_t pid, prstatus_t *status, prfpregset_t *fpregset,
static vm_map_entry_t
readmap(pid_t pid)
{
char mapname[64];
int mapfd;
ssize_t mapsize;
size_t bufsize;
char *mapbuf;
int pos;
vm_map_entry_t map;
vm_map_entry_t *linkp;
vm_map_entry_t ent, *linkp, map;
struct kinfo_vmentry *vmentl, *kve;
int i, nitems;
snprintf(mapname, sizeof mapname, "/proc/%d/map", pid);
if ((mapfd = open(mapname, O_RDONLY)) == -1)
err(1, "cannot open %s", mapname);
vmentl = kinfo_getvmmap(pid, &nitems);
if (vmentl == NULL)
err(1, "cannot retrieve mappings for %u process", pid);
/*
* Procfs requires (for consistency) that the entire memory map
* be read with a single read() call. Start with a reasonably sized
* buffer, and double it until it is big enough.
*/
bufsize = 8 * 1024;
mapbuf = NULL;
for ( ; ; ) {
if ((mapbuf = realloc(mapbuf, bufsize + 1)) == NULL)
errx(1, "out of memory");
mapsize = read(mapfd, mapbuf, bufsize);
if (mapsize != -1 || errno != EFBIG)
break;
bufsize *= 2;
/* This lseek shouldn't be necessary, but it is. */
lseek(mapfd, (off_t)0, SEEK_SET);
}
if (mapsize == -1)
err(1, "read error from %s", mapname);
if (mapsize == 0)
errx(1, "empty map file %s", mapname);
mapbuf[mapsize] = 0;
close(mapfd);
pos = 0;
map = NULL;
linkp = &map;
while (pos < mapsize) {
vm_map_entry_t ent;
u_long start;
u_long end;
char prot[4];
char type[16];
int n;
int len;
for (i = 0; i < nitems; i++) {
kve = &vmentl[i];
len = 0;
n = sscanf(mapbuf + pos, "%lx %lx %*d %*d %*x %3[-rwx]"
" %*d %*d %*x %*s %*s %16s %*s%*[\n]%n",
&start, &end, prot, type, &len);
if (n != 4 || len == 0)
errx(1, "ill-formed line in %s starting at character %d", mapname, pos + 1);
pos += len;
/* Ignore segments of the wrong kind, and unwritable ones */
if (strncmp(prot, "rw", 2) != 0 ||
(strcmp(type, "default") != 0 &&
strcmp(type, "vnode") != 0 &&
strcmp(type, "swap") != 0))
/*
* Ignore segments of the wrong kind and ones which are not
* readable and writable.
*/
if ((kve->kve_protection & KVME_PROT_WRITE) == 0 ||
(kve->kve_protection & KVME_PROT_READ) == 0 ||
(kve->kve_type != KVME_TYPE_DEFAULT &&
kve->kve_type != KVME_TYPE_VNODE &&
kve->kve_type != KVME_TYPE_SWAP))
continue;
if ((ent = (vm_map_entry_t)calloc(1, sizeof *ent)) == NULL)
ent = calloc(1, sizeof(*ent));
if (ent == NULL)
errx(1, "out of memory");
ent->start = start;
ent->end = end;
ent->start = (vm_offset_t)kve->kve_start;
ent->end = (vm_offset_t)kve->kve_end;
ent->protection = VM_PROT_READ | VM_PROT_WRITE;
if (prot[2] == 'x')
ent->protection |= VM_PROT_EXECUTE;
if ((kve->kve_protection & KVME_PROT_EXEC) != 0)
ent->protection |= VM_PROT_EXECUTE;
*linkp = ent;
linkp = &ent->next;
}
free(mapbuf);
return map;
free(vmentl);
return (map);
}
struct dumpers elfdump = { elf_ident, elf_coredump };

View File

@ -32,7 +32,7 @@
.\" @(#)gcore.1 8.2 (Berkeley) 4/18/94
.\" $FreeBSD$
.\"
.Dd April 18, 1994
.Dd November 18, 2009
.Dt GCORE 1
.Os
.Sh NAME
@ -55,11 +55,6 @@ By default, the core is written to the file
The process identifier,
.Ar pid ,
must be given on the command line.
If no executable image is
specified,
.Nm
will use
.Dq Pa /proc/<pid>/file .
.Pp
The following options are available:
.Bl -tag -width indent
@ -80,8 +75,6 @@ The same effect can be achieved manually with
.Bl -tag -width /var/log/messages -compact
.It Pa core.<pid>
the core image
.It Pa /proc/<pid>/file
the executable image
.El
.Sh HISTORY
A
@ -89,12 +82,15 @@ A
utility appeared in
.Bx 4.2 .
.Sh BUGS
Context switches or paging activity that occur while
Because of the
.Xr ptrace 2
usage
.Nm
is running may cause the program to become confused.
For best results, use
.Fl s
to temporarily stop the target process.
may not work with processes which are actively investigated with
.Xr truss 1
or
.Xr gdb 1 .
Additionally, interruptable sleeps may exit with EINTR.
.Pp
The
.Nm

View File

@ -61,6 +61,7 @@ __FBSDID("$FreeBSD$");
#include <sys/time.h>
#include <sys/stat.h>
#include <sys/linker_set.h>
#include <sys/sysctl.h>
#include <err.h>
#include <fcntl.h>
@ -83,10 +84,11 @@ SET_DECLARE(dumpset, struct dumpers);
int
main(int argc, char *argv[])
{
int ch, efd, fd, sflag;
int ch, efd, fd, name[4], sflag;
char *binfile, *corefile;
char fname[MAXPATHLEN];
char passpath[MAXPATHLEN], fname[MAXPATHLEN];
struct dumpers **d, *dumper;
size_t len;
sflag = 0;
corefile = NULL;
@ -109,9 +111,14 @@ main(int argc, char *argv[])
switch (argc) {
case 1:
pid = atoi(argv[0]);
asprintf(&binfile, "/proc/%d/file", pid);
if (binfile == NULL)
errx(1, "allocation failure");
name[0] = CTL_KERN;
name[1] = KERN_PROC;
name[2] = KERN_PROC_PATHNAME;
name[3] = pid;
len = sizeof(passpath);
if (sysctl(name, 4, passpath, &len, NULL, 0) == -1)
errx(1, "kern.proc.pathname failure");
binfile = passpath;
break;
case 2:
pid = atoi(argv[1]);