int gpuvm_pre_init(int flags) {
if(flags != GPUVM_THREADS_BEFORE_INIT &&
flags != GPUVM_THREADS_AFTER_INIT) {
fprintf(stderr, "gpuvm_pre_init: invalid flags\n");
return GPUVM_EARG;
}
if(flags == GPUVM_THREADS_BEFORE_INIT) {
// case before OpenCL initialization
if(pre_runtime_threads_g) {
free(pre_runtime_threads_g);
pre_runtime_threads_g = 0;
pre_runtime_nthreads_g = 0;
}
int pre_runtime_nthreads = get_threads(&pre_runtime_threads_g);
if(pre_runtime_nthreads < 0)
return -1;
pre_runtime_nthreads_g = pre_runtime_nthreads;
//fprintf(stderr, "%d threads before OpenCL runtime init\n",
// pre_runtime_nthreads_g);
return 0;
} else if(flags == GPUVM_THREADS_AFTER_INIT) {
// case after OpenCL initialization
if(!pre_runtime_threads_g) {
fprintf(stderr, "gpuvm_pre_init: list of threads must be recorded "
"first\n");
return GPUVM_ESTATE;
}
thread_t *after_runtime_threads, *diff_threads;
// get new threads
int after_runtime_nthreads = get_threads(&after_runtime_threads);
if(after_runtime_nthreads < 0)
return after_runtime_nthreads;
//fprintf(stderr, "%d threads after OpenCL runtime init\n",
// after_runtime_nthreads);
// get difference
int diff_nthreads = threads_diff
(&diff_threads, after_runtime_threads, after_runtime_nthreads,
pre_runtime_threads_g, pre_runtime_nthreads_g);
if(diff_nthreads < 0) {
free(after_runtime_threads);
return diff_nthreads;
}
// save difference as "immune threads"
if(diff_nthreads > MAX_NTHREADS) {
fprintf(stderr, "gpuvm_pre_init: too many immune threads\n");
return -1;
}
immune_nthreads_g = diff_nthreads;
//fprintf(stderr, "immune_nthreads_g = %d\n", immune_nthreads_g);
memcpy(immune_threads_g, diff_threads, immune_nthreads_g * sizeof(thread_t));
free(after_runtime_threads);
free(diff_threads);
return 0;
} // if(flags == ...)
}
int backtrace_ptrace(int pid, int *tids, int *index, int nr_tids)
{
#if !defined (NO_LIBUNWIND_PTRACE)
int i, count, rc = 0;
int *threads = NULL;
count = get_threads(pid, &threads);
if (!count || threads == NULL)
return -1;
if (tids != NULL) {
if (adjust_threads(threads, count, tids, index, nr_tids) < 0)
return -1;
free(threads);
count = nr_tids;
threads = tids;
}
if (attach_process(pid) < 0)
return -1;
for (i = 0; i < count; ++i) {
void *upt_info;
printf("-------------------- thread %d (%d) --------------------\n",
(index != NULL ? index[i] : i+1), threads[i]);
if (threads[i] != pid && attach_thread(threads[i]) < 0) {
rc = -1;
break;
}
upt_info = _UPT_create(threads[i]);
if (backtrace_thread(&_UPT_accessors, upt_info) < 0)
rc = -1;
_UPT_destroy(upt_info);
if (threads[i] != pid && detach_thread(threads[i]))
rc = -1;
if (rc < 0)
break;
}
free(threads);
if (detach_process(pid) < 0)
return -1;
return rc;
#else
return -1;
#endif /* NO_LIBUNWIND_PTRACE */
}
void mp_manager::check_threads(section* pSection)
{
volatile thread* pBound;
InterlockedExchangePointer(&pBound , pSection->m_pbound_thread);
if (pBound )
{
if (true )
{
volatile KickProc pKick;
InterlockedExchangePointer(&pKick, pBound->m_pKickProc);
if (pKick && pBound)
{
(*pKick)((thread*) pBound ,(section*) pBound->m_pexclusive_section);
}
else
{
SetEvent( pBound->m_handle);
}
}
}
else
for (int i = 0 ; i < get_threads().size() ; i ++ )
{
thread* next = get_threads()[i];
//volatile KickProc pKick;
volatile section* pSection;
//InterlockedExchangePointer(&pKick, next->m_pKickProc);
InterlockedExchangePointer(&pSection, next->m_pexclusive_section);
if (pSection) continue;
if (InterlockedCompareExchange(&next->m_sleeping,0,0) )
{
//printf("Kicking thread %d" , next->m_id);
SetEvent( next->m_handle);
}
break;
}
}
int get_physical_cores()
{
int regs[4];
boost::simd::config::x86::cpuid(regs, 0x00000000);
if(boost::simd::config::x86::get_vendor() == boost::simd::config::x86::intel)
{
return (get_logical_cores()/get_threads());
}
else if(boost::simd::config::x86::get_vendor() == boost::simd::config::x86::amd)
{
boost::simd::config::x86::cpuid(regs, 0x80000008);
return (get_byte(regs[2], 0) + 1);
}
else return -1;
}
int main(int argc, char *argv[])
{
DWORD dwPid = 7668;//GetCurrentProcessId();
HANDLE hProcess = OpenProcess(PROCESS_ALL_ACCESS, FALSE, dwPid);
for (auto region : get_regions(hProcess))
print_range((ULONG_PTR)region.BaseAddress, (ULONG_PTR)region.BaseAddress + region.RegionSize);
printf("\nDLLs:\n");
for (auto module : get_modules(hProcess))
print_range((ULONG_PTR)module.lpBaseOfDll, (ULONG_PTR)module.lpBaseOfDll + (ULONG_PTR)module.lpBaseOfDll);
printf("\nThreads:\n");
for (auto thread : get_threads(dwPid))
print_range((ULONG_PTR)thread.Eip, (ULONG_PTR)thread.Esp);
getchar();
return 0;
}
/*
* save process' memory maps, stack contents, thread identifiers and registers
*/
struct snapshot *get_snapshot(int pid, int *tids, int *index, int nr_tids)
{
struct snapshot *res;
int attached_tid = 0;
int i, n_frames;
long page, label, rc;
struct mem_data_chunk **stacks_cover = NULL;
if ((page = sysconf(_SC_PAGESIZE)) < 0) {
perror("get pagesize");
return NULL;
}
--page;
res = calloc(1, sizeof(struct snapshot));
/*
* create memory_map structure corresponding to process' maps
*/
res->map = create_maps(pid);
if (res->map == NULL)
goto get_snapshot_fail;
/*
* get process' threads
*/
res->num_threads = get_threads(pid, &res->tids);
if (res->num_threads < 0 || res->tids == NULL)
goto get_snapshot_fail;
/*
* user-provided list of threads
*/
if (tids != NULL) {
if (adjust_threads(res->tids, res->num_threads, tids, index, nr_tids) < 0)
goto get_snapshot_fail;
free(res->tids);
res->num_threads = nr_tids;
res->tids = tids;
}
res->cur_thr = 0;
res->regs = malloc(sizeof(res->regs[0])*res->num_threads);
if (res->regs == NULL) {
perror("malloc");
goto get_snapshot_fail;
}
/* FREEZE PROCESS */
if (attach_process(pid) < 0)
goto get_snapshot_fail;
for (i = 0; i < res->num_threads; ++i) {
struct iovec iov;
/*
* we have already attached to main thread. call attach_thread()
* for other ones
*/
attached_tid = res->tids[i];
if (res->tids[i] != pid && attach_thread(res->tids[i]) < 0)
goto get_snapshot_fail_attached;
/*
* save thread's registers
*/
iov.iov_len = sizeof(res->regs[0]);
iov.iov_base = &res->regs[i];
rc = ptrace(PTRACE_GETREGSET, res->tids[i], NT_PRSTATUS, &iov);
if (rc < 0) {
perror("PTRACE_GETREGSET");
goto get_snapshot_fail_attached;
}
/*
* save label on memory region. it will indicate that memory contents
* upper than this point (%rsp) will needed to unwind stacks
*/
label = SP_REG(&res->regs[i]) & ~page;
rc = mem_map_add_label(res->map, (void *)label, res->num_threads);
if (rc < 0) {
fprintf(stderr, "failed to add label 0x%lx [rsp 0x%llx thread %d]\n",
label, (long long unsigned int)SP_REG(&res->regs[i]), res->tids[i]);
goto get_snapshot_fail_attached;
}
/*
* detach from thread. it will still be frozen due to SIGSTOP
*/
if (res->tids[i] != pid && detach_thread(res->tids[i]) < 0)
goto get_snapshot_fail_attached;
}
/*
* arrange data chunks to copy memory contents. in most cases the chunks
* will start from %rsp pointing somewhere in thread's stack
* to the end of the stack region
*/
stacks_cover = malloc(sizeof(struct mem_data_chunk*) * res->num_threads);
n_frames = mem_map_build_label_cover(res->map, stack_size,
stacks_cover, page + 1);
if (stacks_cover == NULL) {
fprintf(stderr, "error: stacks cover == NULL, n_frames=%d\n", n_frames);
goto get_snapshot_fail_attached;
}
/*
* copy memory contents
*/
rc = copy_memory(pid, stacks_cover, n_frames);
if (rc < 0)
goto get_snapshot_fail_attached;
/* UNFREEZE PROCESS */
if (detach_process(pid) < 0)
goto get_snapshot_fail;
if (opt_verbose) {
for (i = 0; i < n_frames; ++i) {
struct mem_data_chunk *chunk = stacks_cover[i];
printf("chunk #%d: 0x%lx-0x%lx length: %ldK\n",
i, (size_t)chunk->start,
(size_t)chunk->start + chunk->length,
chunk->length >> 10);
}
}
free(stacks_cover);
return res;
get_snapshot_fail_attached:
if (attached_tid)
detach_thread(attached_tid);
detach_process(pid);
get_snapshot_fail:
if (opt_verbose) {
fprintf(stderr, "maps of %d:\n", pid);
print_proc_maps(pid);
}
free(stacks_cover);
snapshot_destroy(res);
return NULL;
}
int backtrace_ptrace(int pid, int *tids, int *index, int nr_tids)
{
#if !defined (NO_LIBUNWIND_PTRACE)
int i, count, rc = 0;
int *threads = NULL;
count = get_threads(pid, &threads);
if (!count || threads == NULL)
return -1;
if (tids != NULL) {
if (adjust_threads(threads, count, tids, index, nr_tids) < 0)
return -1;
free(threads);
count = nr_tids;
threads = tids;
}
if (attach_process(pid) < 0)
return -1;
for (i = 0; i < count; ++i) {
void *upt_info;
int x;
char comm[16];
char end_pad[25] = "------------------------";
x = get_thread_comm(threads[i], comm, sizeof(comm));
if (x > 0 && x <= sizeof(end_pad))
{
end_pad[sizeof(end_pad) - x] = '\0';
printf("-------------- thread %d (%d) (%s) %s\n", (index != NULL ? index[i] : i + 1), threads[i], comm, end_pad);
}
if (threads[i] != pid && attach_thread(threads[i]) < 0) {
rc = -1;
break;
}
upt_info = _UPT_create(threads[i]);
if (backtrace_thread(&_UPT_accessors, upt_info) < 0)
rc = -1;
_UPT_destroy(upt_info);
if (threads[i] != pid && detach_thread(threads[i]))
rc = -1;
if (rc < 0)
break;
}
free(threads);
if (detach_process(pid) < 0)
return -1;
return rc;
#else
return -1;
#endif /* NO_LIBUNWIND_PTRACE */
}
/*
* save process' memory maps, stack contents, thread identifiers and registers
*/
struct snapshot *get_snapshot(int pid, int *tids, int *index, int nr_tids)
{
struct snapshot *res;
int i, attached_tid, n_frames;
long page, label, rc;
struct mem_data_chunk **stacks_cover = NULL;
int v_major, v_minor;
int use_process_vm_readv = 0;
if ((page = sysconf(_SC_PAGESIZE)) < 0) {
perror("get pagesize");
return NULL;
}
--page;
res = calloc(1, sizeof(struct snapshot));
/*
* create memory_map structure corresponding to process' maps
*/
res->map = create_maps(pid);
if (res->map == NULL)
goto get_snapshot_fail;
/*
* get process' threads
*/
res->num_threads = get_threads(pid, &res->tids);
if (res->num_threads < 0 || res->tids == NULL)
goto get_snapshot_fail;
/*
* user-provided list of threads
*/
if (tids != NULL) {
if (adjust_threads(res->tids, res->num_threads, tids, index, nr_tids) < 0)
goto get_snapshot_fail;
free(res->tids);
res->num_threads = nr_tids;
res->tids = tids;
}
res->cur_thr = 0;
res->regs = malloc(sizeof(struct user_regs_struct)*res->num_threads);
if (res->regs == NULL) {
perror("malloc");
goto get_snapshot_fail;
}
/*
* decide how to copy memory contents of the process. on newer kernels
* proc_vm_readv() is used by default. on older kernels or when the option
* --proc-mem is specified read the file /proc/<pid>/mem
*/
if (!opt_proc_mem) {
if (get_kernel_version(&v_major, &v_minor) < 0)
goto get_snapshot_fail;
if (((v_major << 16) | v_minor) >= 0x30002)
use_process_vm_readv = 1;
} else {
use_process_vm_readv = 0;
}
/* FREEZE PROCESS */
if (attach_process(pid) < 0)
goto get_snapshot_fail;
for (i = 0; i < res->num_threads; ++i) {
/*
* we have already attached to main thread. call attach_thread()
* for other ones
*/
attached_tid = res->tids[i];
if (res->tids[i] != pid && attach_thread(res->tids[i]) < 0)
goto get_snapshot_fail_attached;
/*
* save thread's registers
*/
rc = ptrace(PTRACE_GETREGS, res->tids[i], NULL, &res->regs[i]);
if (rc < 0) {
perror("PTRACE_GETREGS");
goto get_snapshot_fail_attached;
}
/*
* save label on memory region. it will indicate that memory contents
* upper than this point (%rsp) will needed to unwind stacks
*/
label = res->regs[i].rsp & ~page;
rc = mem_map_add_label(res->map, (void *)label, res->num_threads);
if (rc < 0) {
fprintf(stderr, "failed to add label 0x%lx [rsp 0x%lx thread %d]\n",
label, res->regs[i].rsp, res->tids[i]);
goto get_snapshot_fail_attached;
}
/*
* detach from thread. it will still be frozen due to SIGSTOP
*/
if (res->tids[i] != pid && detach_thread(res->tids[i]) < 0)
goto get_snapshot_fail_attached;
}
/*
* arrange data chunks to copy memory contents. in most cases the chunks
* will start from %rsp pointing somewhere in thread's stack
* to the end of the stack region
*/
stacks_cover = malloc(sizeof(struct mem_data_chunk*) * res->num_threads);
n_frames = mem_map_build_label_cover(res->map, stack_size,
stacks_cover, page + 1);
if (stacks_cover == NULL) {
fprintf(stderr, "error: stacks cover == NULL, n_frames=%d\n", n_frames);
goto get_snapshot_fail_attached;
}
/*
* copy memory contents
*/
rc = use_process_vm_readv ?
copy_memory_process_vm_readv(pid, stacks_cover, n_frames) :
copy_memory_proc_mem(pid, stacks_cover, n_frames);
if (rc < 0)
goto get_snapshot_fail_attached;
/* UNFREEZE PROCESS */
if (detach_process(pid) < 0)
goto get_snapshot_fail;
if (opt_verbose) {
for (i = 0; i < n_frames; ++i) {
struct mem_data_chunk *chunk = stacks_cover[i];
printf("chunk #%d: 0x%lx-0x%lx length: %ldK\n",
i, (size_t)chunk->start,
(size_t)chunk->start + chunk->length,
chunk->length >> 10);
}
}
free(stacks_cover);
return res;
get_snapshot_fail_attached:
if (attached_tid)
detach_thread(attached_tid);
detach_process(pid);
get_snapshot_fail:
if (opt_verbose) {
fprintf(stderr, "maps of %d:\n", pid);
print_proc_maps(pid);
}
free(stacks_cover);
snapshot_destroy(res);
return NULL;
}
