/*
* icount_start -- begin instruction counting
*
* Inputs:
* life_remaining -- simulate a crash after counting this many
* instructions. pass in 0ull to disable this
* feature.
*
* There is some variability on how far the parent can get before the child
* attaches to it for tracing, especially when the system is loaded down
* (e.g. due to zillions of parallel icount traces happening at once).
* To coordinate this, we use a multi-step procedure where the tracee
* runs until it gets to the function pretrigger() and the tracer() detects
* that and raises a signal (SIGRTMIN+15) in the tracee, who then catches
* it and longjmps out of signal handler, back here, and then calls trigger().
*
* The multi-step coordination sounds complex, but it handles the
* surprisingly common cases where the tracer attached to the tracee
* before it even finished executing the libc fork() wrapper, and the
* other end of the spectrum where the tracee ran way too far ahead
* before the tracer attached and the tracer missed the call to trigger().
*/
void
icount_start(unsigned long life_remaining)
{
if (Tracer_pid) {
/* nested call not allowed */
icount_stop();
FATAL("icount_start called while counting already active");
}
Total = 0ull;
if (pipe(Tracerpipe) < 0)
FATALSYS("pipe");
if (signal(SIGRTMIN+15, handler) == SIG_ERR)
FATALSYS("signal: SIGRTMIN+15");
if ((Tracer_pid = fork()) < 0)
FATALSYS("fork");
else if (Tracer_pid) {
/* parent */
close(Tracerpipe[1]);
if (!setjmp(Trigjmp))
pretrigger();
close(-1); /* harmless syscall for tracer */
trigger();
return;
} else {
/* child */
close(Tracerpipe[0]);
tracer(life_remaining);
}
}
/*
* icount_stop -- stop counting instructions
*/
void
icount_stop(void)
{
int status;
if (read(Tracerpipe[0], &Total, sizeof(Total)) < 0)
FATALSYS("read from pipe");
close(Tracerpipe[0]);
/* wait for child */
if (waitpid(Tracer_pid, &status, 0) < 0)
FATALSYS("waitpid(pid=%d)", Tracer_pid);
Tracer_pid = 0;
}
/*
* pmem_flush_cache -- flush processor cache for the given range
*
* This is the fit version (fault injection test) that uses copy-on-write.
*/
void
pmem_flush_cache_fit(void *addr, size_t len, int flags)
{
uintptr_t uptr;
if (!PM_base)
FATAL("pmem_map hasn't been called");
/*
* even though pwrite() can take any random byte addresses and
* lengths, we simulate cache flushing by writing the full 64B
* chunks that cover the given range.
*/
for (uptr = (uintptr_t)addr & ~(ALIGN - 1);
uptr < (uintptr_t)addr + len; uptr += ALIGN)
if (pwrite(PM_fd, (void *)uptr, ALIGN, uptr - PM_base) < 0)
FATALSYS("pwrite len %d offset %lu", len,
addr - PM_base);
}
/*
* pmem_flush_cache -- flush processor cache for the given range
*
* This is the msync-based version.
*/
void
pmem_flush_cache_msync(void *addr, size_t len, int flags)
{
uintptr_t uptr;
/*
* msync requires len to be a multiple of pagesize, so
* adjust addr and len to represent the full 4k chunks
* covering the given range.
*/
/* increase len by the amount we gain when we round addr down */
len += (uintptr_t)addr & (ALIGN - 1);
/* round addr down to page boundary */
uptr = (uintptr_t)addr & ~(ALIGN - 1);
/* round len up to multiple of page size */
len = (len + (ALIGN - 1)) & ~(ALIGN - 1);
if (msync((void *)uptr, len, MS_SYNC) < 0)
FATALSYS("msync");
}
int
main(int argc, char *argv[])
{
int opt;
int iflag = 0;
unsigned long icount;
const char *path;
struct stat stbuf;
size_t size;
int fd;
char *pmaddr;
Myname = argv[0];
while ((opt = getopt(argc, argv, "FMdi:")) != -1) {
switch (opt) {
case 'F':
pmem_fit_mode();
break;
case 'M':
pmem_msync_mode();
break;
case 'd':
Debug++;
break;
case 'i':
iflag++;
icount = strtoul(optarg, NULL, 10);
break;
default:
USAGE(NULL);
}
}
if (optind >= argc)
USAGE("No path given");
path = argv[optind++];
if (stat(path, &stbuf) < 0) {
/*
* file didn't exist, create it with DEFAULT_SIZE
*/
if ((fd = open(path, O_CREAT|O_RDWR, 0666)) < 0)
FATALSYS("can't create %s", path);
if ((errno = posix_fallocate(fd, 0, DEFAULT_SIZE)) != 0)
FATALSYS("posix_fallocate");
size = DEFAULT_SIZE;
} else {
/*
* file exists, just open it
*/
if ((fd = open(path, O_RDWR)) < 0)
FATALSYS("open %s", path);
size = stbuf.st_size;
}
/*
* map the file into our address space.
*/
if ((pmaddr = pmem_map(fd, size)) == NULL)
FATALSYS("pmem_map");
if (optind < argc) { /* strings supplied as arguments? */
int i;
char *ptr = pmaddr;
if (iflag)
icount_start(icount); /* start instruction count */
for (i = optind; i < argc; i++) {
size_t len = strlen(argv[i]) + 1; /* includes '\0' */
if (len > size)
FATAL("no more room for %d-byte string", len);
/* store to Persistent Memory */
strcpy(ptr, argv[i]);
/* make that change durable */
pmem_persist(ptr, len, 0);
ptr += len;
size -= len;
}
if (iflag) {
icount_stop(); /* end instruction count */
printf("Total instruction count: %lu\n",
icount_total());
}
} else {
char *ptr = pmaddr;
char *sep = "";
/*
* dump out all the strings we find in Persistent Memory
*/
while (ptr < &pmaddr[size]) {
/* load from Persistent Memory */
if (isprint(*ptr)) {
putc(*ptr, stdout);
sep = "\n";
} else if (*ptr == '\0') {
fputs(sep, stdout);
sep = "";
}
ptr++;
}
}
exit(0);
}
/*
* tracer -- internal function used in child process to trace parent
*/
static void
tracer(unsigned long ttl)
{
pid_t ppid = getppid();
int status;
int triggered = 0;
int signaled = 0;
if (ptrace(PTRACE_ATTACH, ppid, 0, 0) < 0)
FATALSYS("PTRACE_ATTACH");
while (1) {
if (waitpid(ppid, &status, 0) < 0)
FATALSYS("waitpid(pid=%d)", ppid);
if (WIFSTOPPED(status)) {
struct user_regs_struct regs;
if (triggered)
Total++;
if (ttl && Total >= ttl) {
if (kill(ppid, SIGKILL) < 0)
FATAL("SIGKILL %d", ppid);
printf("Program terminated after %lu "
"instructions\n",
Total);
fflush(stdout);
_exit(0);
}
if (ptrace(PTRACE_GETREGS, ppid, 0, ®s) < 0)
FATALSYS("PTRACE_GETREGS");
if ((unsigned long)regs.rip >=
(unsigned long)pretrigger &&
(unsigned long)regs.rip <
(unsigned long)trigger) {
if (!signaled) {
if (ptrace(PTRACE_SYSCALL, ppid, 0,
SIGRTMIN+15) < 0)
FATALSYS("PTRACE_SYSCALL");
signaled = 1;
continue;
}
} else if ((unsigned long) regs.rip ==
(unsigned long) trigger) {
triggered = 1;
} else if ((unsigned long) regs.rip ==
(unsigned long) icount_stop) {
if (ptrace(PTRACE_DETACH, ppid, 0, 0) < 0)
FATALSYS("PTRACE_DETACH");
break;
}
if (ptrace(PTRACE_SINGLESTEP, ppid, 0, 0) < 0)
FATALSYS("PTRACE_SINGLESTEP");
} else if (WIFEXITED(status))
FATAL("tracee: exit %d", WEXITSTATUS(status));
else if (WIFSIGNALED(status))
FATAL("tracee: %s", strsignal(WTERMSIG(status)));
else
FATAL("unexpected wait status: 0x%x", status);
}
/*
* our counting is done, send the count back to the tracee
* via the pipe and exit.
*/
if (write(Tracerpipe[1], &Total, sizeof(Total)) < 0)
FATALSYS("write to pipe");
close(Tracerpipe[1]);
_exit(0);
}
/*
* pmemalloc_check -- check the consistency of a pmem pool
*
* Inputs:
* path -- path to the file which contains the memory pool
*
* The current state of the pmem pool is printed. This routine does
* not make any changes to the pmem pool (maps it read-only, in fact).
* It is not necessary to call pmemalloc_init() before calling this.
*/
void pmemalloc_check(const char *path)
{
void *pmp;
int fd;
struct stat stbuf;
struct clump *clp;
struct clump *lastclp;
struct pool_header *hdrp;
size_t clumptotal;
/*
* stats we keep for each type of memory:
* stats[PMEM_STATE_FREE] for free clumps
* stats[PMEM_STATE_RESERVED] for reserved clumps
* stats[PMEM_STATE_ACTIVATING] for activating clumps
* stats[PMEM_STATE_ACTIVE] for active clumps
* stats[PMEM_STATE_FREEING] for freeing clumps
* stats[PMEM_STATE_UNUSED] for overall totals
*/
struct {
size_t largest;
size_t smallest;
size_t bytes;
unsigned count;
} stats[PMEM_STATE_UNUSED + 1] = { 0 };
const char *names[] = {
"Free",
"Reserved",
"Activating",
"Active",
"Freeing",
"TOTAL",
};
int i;
DEBUG("path=%s", path);
if ((fd = open(path, O_RDONLY)) < 0)
FATALSYS("%s", path);
if (fstat(fd, &stbuf) < 0)
FATALSYS("fstat");
DEBUG("file size 0x%lx", stbuf.st_size);
if (stbuf.st_size < PMEM_MIN_POOL_SIZE)
FATAL("size %lu too small (must be at least %lu)",
stbuf.st_size, PMEM_MIN_POOL_SIZE);
if ((pmp = mmap(NULL, stbuf.st_size, PROT_READ, MAP_SHARED,
fd, 0)) == MAP_FAILED)
FATALSYS("mmap");
DEBUG("pmp %lx", pmp);
close(fd);
hdrp = PMEM(pmp, (struct pool_header *)PMEM_HDR_OFFSET);
DEBUG(" hdrp 0x%lx (off 0x%lx)", hdrp, OFF(pmp, hdrp));
if (strcmp(hdrp->signature, PMEM_SIGNATURE))
FATAL("failed signature check");
DEBUG("signature check passed");
clp = PMEM(pmp, (struct clump *)PMEM_CLUMP_OFFSET);
/*
* location of last clump is calculated by rounding the file
* size down to a multiple of 64, and then subtracting off
* another 64 to hold the struct clump. the last clump is
* indicated by a size of zero.
*/
lastclp = PMEM(pmp, (struct clump *)
(stbuf.st_size & ~(PMEM_CHUNK_SIZE - 1)) - PMEM_CHUNK_SIZE);
DEBUG(" clp 0x%lx (off 0x%lx)", clp, OFF(pmp, clp));
DEBUG("lastclp 0x%lx (off 0x%lx)", lastclp, OFF(pmp, lastclp));
clumptotal = (uintptr_t)lastclp - (uintptr_t)clp;
DEBUG("expected clumptotal: %lu", clumptotal);
/*
* check that:
*
* the overhead size (stuff up to CLUMP_OFFSET)
* + clumptotal
* + last clump marker (CHUNK_SIZE)
* + any bytes we rounded off the end
* = file size
*/
if (PMEM_CLUMP_OFFSET + clumptotal +
(stbuf.st_size & (PMEM_CHUNK_SIZE - 1)) + PMEM_CHUNK_SIZE
== stbuf.st_size) {
DEBUG("section sizes correctly add up to file size");
} else {
FATAL("CLUMP_OFFSET %d + clumptotal %lu + rounded %d + "
"CHUNK_SIZE %d = %lu, (not st_size %lu)",
PMEM_CLUMP_OFFSET, clumptotal,
(stbuf.st_size & (PMEM_CHUNK_SIZE - 1)),
PMEM_CHUNK_SIZE,
PMEM_CLUMP_OFFSET + clumptotal +
(stbuf.st_size & (PMEM_CHUNK_SIZE - 1)) +
PMEM_CHUNK_SIZE,
stbuf.st_size);
}
if (clp->size == 0)
FATAL("no clumps found");
while (clp->size) {
size_t sz = clp->size & ~PMEM_STATE_MASK;
int state = clp->size & PMEM_STATE_MASK;
DEBUG("[%u]clump size 0x%lx state %d",
OFF(pmp, clp), sz, state);
DEBUG("on: 0x%lx 0x%lx 0x%lx 0x%lx 0x%lx 0x%lx",
clp->on[0].off, clp->on[0].ptr_,
clp->on[1].off, clp->on[1].ptr_,
clp->on[2].off, clp->on[2].ptr_);
if (sz > stats[PMEM_STATE_UNUSED].largest)
stats[PMEM_STATE_UNUSED].largest = sz;
if (stats[PMEM_STATE_UNUSED].smallest == 0 ||
sz < stats[PMEM_STATE_UNUSED].smallest)
stats[PMEM_STATE_UNUSED].smallest = sz;
stats[PMEM_STATE_UNUSED].bytes += sz;
stats[PMEM_STATE_UNUSED].count++;
switch (state) {
case PMEM_STATE_FREE:
DEBUG("clump state: free");
ASSERTeq(clp->on[0].off, 0);
ASSERTeq(clp->on[1].off, 0);
ASSERTeq(clp->on[2].off, 0);
break;
case PMEM_STATE_RESERVED:
DEBUG("clump state: reserved");
break;
case PMEM_STATE_ACTIVATING:
DEBUG("clump state: activating");
break;
case PMEM_STATE_ACTIVE:
DEBUG("clump state: active");
ASSERTeq(clp->on[0].off, 0);
ASSERTeq(clp->on[1].off, 0);
ASSERTeq(clp->on[2].off, 0);
break;
case PMEM_STATE_FREEING:
DEBUG("clump state: freeing");
break;
default:
FATAL("unknown clump state: %d", state);
}
if (sz > stats[state].largest)
stats[state].largest = sz;
if (stats[state].smallest == 0 ||
sz < stats[state].smallest)
stats[state].smallest = sz;
stats[state].bytes += sz;
stats[state].count++;
clp = (struct clump *)((uintptr_t)clp + sz);
DEBUG("next clp 0x%lx, offset 0x%lx", clp, OFF(pmp, clp));
}
if (clp == lastclp)
DEBUG("all clump space accounted for");
else
FATAL("clump list stopped at %lx instead of %lx", clp, lastclp);
if (munmap(pmp, stbuf.st_size) < 0)
FATALSYS("munmap");
/*
* print the report
*/
printf("Summary of pmem pool:\n");
printf("File size: %lu, %d allocatable bytes in pool\n\n",
stbuf.st_size, clumptotal);
printf(" State Bytes Clumps Largest Smallest\n");
for (i = 0; i < PMEM_STATE_UNUSED + 1; i++) {
printf("%10s %10d %10d %10d %10d\n",
names[i],
stats[i].bytes,
stats[i].count,
stats[i].largest,
stats[i].smallest);
}
}
/*
* main() for the allocating thread.
*/
void *
alloc_main( void *tparam )
{
unsigned long thread_num = (unsigned long) tparam;
int retval;
void **mbx_list = (*mbx_array_ptr)[thread_num];
pthread_t free_thread;
void *rel_mem_ptr_ = NULL;
unsigned int malloc_size = 0;
int mbx;
DEBUG( "Enter alloc thread %d\n", thread_num );
retval = pthread_create( &free_thread,
NULL,
free_main,
(void *) thread_num );
if (retval) {
errno = retval;
FATALSYS( "free thread create in thread %l\n", thread_num );
}
/*
* Wait for the starting gun to send all the threads
* off and running....
*/
pthread_mutex_lock( &start_lock );
while (!b_start_flag) {
pthread_cond_wait( &start_cv, &start_lock );
}
pthread_mutex_unlock( &start_lock );
/*
* Until the parent thread says to stop just keep looping
* through the mailboxes looking for empty ones. When one is
* found, do a random sized alloc and put a pointer in the mailbox
* for the freeing thread to free.
*/
while (!b_all_stop) {
for (mbx=0; mbx<MAILBOXES; ++mbx) {
if (mbx_list[mbx] == NULL) {
malloc_size = (int)random() % max_malloc;
if( (rel_mem_ptr_=pmemalloc_reserve( pmp,
malloc_size )) != NULL ) {
pmemalloc_onactive( pmp, rel_mem_ptr_,
&mbx_list[mbx],
rel_mem_ptr_ );
pmemalloc_activate( pmp, rel_mem_ptr_ );
DEBUG( "malloc %d bytes", malloc_size );
} else {
/*
* Malloc failed.
* Sleep to let the Freeing thread catch up.
*/
DEBUG( "malloc failed for size %d\n", malloc_size );
sleep(0);
}
} /* end if mbx empty */
} /* end for each mbx */
} /* end while not all_stop */
/*
* Time to stop. Wait for the freeing thread to complete.
*/
retval = pthread_join( free_thread, NULL );
if (retval) {
errno = retval;
FATALSYS( "Join with freeing thread %d\n", thread_num );
}
pthread_exit( NULL );
}
int
main(int argc, char *argv[])
{
const char *path;
int opt;
int retval;
unsigned long thrd;
int mbx;
void **sa_ptr;
mailbox_array_t *mbx_offset_;
Myname = argv[0];
while ((opt = getopt(argc, argv, "t:r:s:d")) != -1) {
switch (opt) {
case 't':
if (sscanf(optarg, "%u", &num_threads) == EOF) {
USAGE( "-t option error");
}
if (num_threads > MAX_THREADS) {
fprintf( stderr, "using max threads %d\n", MAX_THREADS );
num_threads = MAX_THREADS;
}
break;
case 'r':
if (sscanf(optarg, "%u", &runtime)==EOF) {
USAGE("-r option error");
}
break;
case 's':
if (sscanf(optarg, "%u", &max_malloc)==EOF)
USAGE("-s option error");
break;
case 'd':
Debug=TRUE;
break;
default:
USAGE(NULL);
}
} /* end while opt */
if (optind >= argc)
USAGE("No path given");
path = argv[optind++];
if (optind < argc)
USAGE(NULL);
/*
* Use the alloc_init lib function to open the pool
* via pmfs, and map it into our address space.
* This returns a regular (absolute) pointer.
*/
if ((pmp = pmemalloc_init(path, POOL_SIZE)) == NULL)
FATALSYS("pmemalloc_init on %s", path);
/*
* Fetch our static info.
* The first word is used to store a relative pointer to
* the mailbox array. The library function converts this
* to an absolute pointer.
*/
sa_ptr = (void**)pmemalloc_static_area(pmp);
/* The static area for a new pmem pool is zero'd */
if (*sa_ptr == NULL) {
/*
* Create and initialize the mailbox array in PM
*/
if ((mbx_offset_=pmemalloc_reserve(pmp,
sizeof(mailbox_array_t))) == NULL )
FATALSYS("pmemalloc mailbox array");
/*
* Place a pointer to this array in the first word of the
* static area on activation
*/
pmemalloc_onactive( pmp, mbx_offset_, (void**)sa_ptr, mbx_offset_ );
pmemalloc_activate( pmp, mbx_offset_ );
/* Set the static, regular pointer to be used in the program */
mbx_array_ptr = PMEM( pmp, mbx_offset_ );
for (thrd=0; thrd<MAX_THREADS; ++thrd) {
for (mbx=0; mbx<MAILBOXES; ++mbx) {
(*mbx_array_ptr)[thrd][mbx] = NULL;
}
}
} else {
/*
* This region already exists from a previous run.
* Free any pmem spaces still in the mailbox.
*/
mbx_array_ptr = PMEM( pmp, (mailbox_array_t*)*sa_ptr );
for (thrd=0; thrd<MAX_THREADS; ++thrd) {
for (mbx=0; mbx<MAILBOXES; ++mbx) {
if ((*mbx_array_ptr)[thrd][mbx] != NULL) {
pmemalloc_onfree( pmp,
(*mbx_array_ptr)[thrd][mbx],
&(*mbx_array_ptr)[thrd][mbx],
NULL );
pmemalloc_free( pmp, (*mbx_array_ptr)[thrd][mbx] );
}
}
}
}
/* Commit the initialized mailbox to persistent media */
pmem_persist( mbx_array_ptr, sizeof(mailbox_array_t), 0 );
DEBUG( "Number of threads = %d", num_threads);
DEBUG( "Runtime: %d seconds", runtime );
DEBUG( "Max alloc size %d bytes", max_malloc );
/*
* Create each allocating thread. Each allocating thread
* will create its corresponding freeing thread.
* Once each each thread is created, signal the start condition
* so they all start running around the same time.
*/
for (thrd=0; thrd<num_threads; ++thrd) {
retval = pthread_create( &alloc_threads[thrd],
NULL,
alloc_main,
(void *) thrd );
if (retval) {
errno = retval;
FATALSYS( "alloc thread create %d\n", thrd );
}
}
/* Give the new threads a chance to start */
sleep(0);
pthread_mutex_lock( &start_lock );
b_start_flag = TRUE;
pthread_cond_broadcast( &start_cv );
pthread_mutex_unlock( &start_lock );
/* Let run for the desired seconds then tell all threads to stop */
sleep( runtime );
b_all_stop = TRUE;
/* Wait for each alloating thread to complete. */
for (thrd=0; thrd<num_threads; ++thrd) {
retval = pthread_join( alloc_threads[thrd], NULL );
if (retval) {
errno = retval;
FATALSYS( "Allocating thread JOIN %d", thrd );
}
}
/* Commit the final mailbox array to persistent media */
pmem_persist( mbx_array_ptr, sizeof(mailbox_array_t), 0 );
DEBUG("Done.");
exit(0);
}
int main(int argc, char *argv[]) {
const char *path;
int opt;
int fflag = 0;
int iflag = 0;
unsigned long icount;
void *pmp;
struct static_info *sp;
struct node *parent_;
struct node *np_;
Myname = argv[0];
while ((opt = getopt(argc, argv, "FMdfi:")) != -1) {
switch (opt) {
case 'F':
pmem_fit_mode();
break;
case 'M':
pmem_msync_mode();
break;
case 'd':
Debug++;
break;
case 'f':
fflag++;
break;
case 'i':
iflag++;
icount = strtoul(optarg, NULL, 10);
break;
default:
USAGE(NULL);
}
}
if (optind >= argc)
USAGE("No path given");
path = argv[optind++];
if ((pmp = pmemalloc_init(path, MY_POOL_SIZE)) == NULL)
FATALSYS("pmemalloc_init on %s", path);
/* fetch our static info */
sp = (struct static_info *) pmemalloc_static_area(pmp);
if (optind < argc) { /* numbers supplied as arguments? */
int i;
if (fflag)
USAGE("unexpected extra arguments given with -f flag");
if (iflag)
icount_start(icount); /* start instruction count */
for (i = optind; i < argc; i++) {
int value = atoi(argv[i]);
if ((np_ = pmemalloc_reserve(pmp, sizeof(*np_))) == NULL)
FATALSYS("pmemalloc_reserve");
/* link it in at the beginning of the list */
PMEM(pmp, np_)->next_ = sp->rootnp_;
PMEM(pmp, np_)->value = value;
pmemalloc_onactive(pmp, np_, (void **) &sp->rootnp_, np_);
pmemalloc_activate(pmp, np_);
}
if (iflag) {
icount_stop(); /* end instruction count */
printf("Total instruction count: %lu\n", icount_total());
}
} else if (fflag) {
/*
* remove first item from list
*/
if (sp->rootnp_ == NULL)
FATAL("the list is empty");
if (iflag)
icount_start(icount); /* start instruction count */
np_ = sp->rootnp_;
pmemalloc_onfree(pmp, np_, (void **) &sp->rootnp_, PMEM(pmp, np_)->next_);
pmemalloc_free(pmp, np_);
if (iflag) {
icount_stop(); /* end instruction count */
printf("Total instruction count: %lu\n", icount_total());
}
} else {
char *sep = "";
/*
* traverse the list, printing the numbers found
*/
np_ = sp->rootnp_;
while (np_) {
printf("%s%d", sep, PMEM(pmp, np_)->value);
sep = " ";
np_ = PMEM(pmp, np_)->next_;
}
printf("\n");
}
DEBUG("Done.");
exit(0);
}