static char *tp_from_combined_area(const struct tls_info *info,
void *combined_area, size_t tdb_size) {
size_t tls_size = info->tdata_size + info->tbss_size;
ptrdiff_t tdboff = __nacl_tp_tdb_offset(tdb_size);
if (tdboff < 0) {
/*
* The combined area is big enough to hold the TDB and then be aligned
* up to the $tp alignment requirement. If the whole area is aligned
* to the $tp requirement, then aligning the beginning of the area
* would give us the beginning unchanged, which is not what we need.
* Instead, align from the putative end of the TDB, to decide where
* $tp--the true end of the TDB--should actually lie.
*/
return aligned_addr((char *) combined_area + tdb_size, info->tls_alignment);
} else {
/*
* The linker increases the size of the TLS block up to its alignment
* requirement, and that total is subtracted from the $tp address to
* access the TLS area. To keep that final address properly aligned,
* we need to align up from the allocated space and then add the
* aligned size.
*/
tls_size = aligned_size(tls_size, info->tls_alignment);
return aligned_addr((char *) combined_area, info->tls_alignment) + tls_size;
}
}
static inline nc_thread_descriptor_t *nc_get_tdb(void) {
/*
* Fetch the thread-specific data pointer. This is usually just
* a wrapper around __libnacl_irt_tls.tls_get() but we don't use
* that here so that the IRT build can override the definition.
*/
return (void *) ((char *) __nacl_read_tp() + __nacl_tp_tdb_offset(TDB_SIZE));
}
void *__nacl_tls_initialize_memory(void *combined_area, size_t tdb_size) {
const struct tls_info *info = get_tls_info();
size_t tls_size = info->tdata_size + info->tbss_size;
char *combined_area_end =
(char *) combined_area + __nacl_tls_combined_size(tdb_size);
void *tp = tp_from_combined_area(info, combined_area, tdb_size);
char *start = tp;
if (__nacl_tp_tls_offset(0) > 0) {
/*
* From $tp, we skip the header size and then must round up from
* there to the required alignment (which is what the linker will
* will do when calculating TPOFF relocations at link time). The
* end result is that the offset from $tp matches the one chosen
* by the linker exactly and that the final address is aligned to
* info->tls_alignment (since $tp was already aligned to at least
* that much).
*/
start += aligned_size(__nacl_tp_tls_offset(tls_size), info->tls_alignment);
} else {
/*
* We'll subtract the aligned size of the TLS block from $tp, which
* must itself already be adequately aligned.
*/
start += __nacl_tp_tls_offset(aligned_size(tls_size, info->tls_alignment));
}
/* Sanity check. (But avoid pulling in assert() here.) */
if (start + info->tdata_size + info->tbss_size > combined_area_end)
simple_abort();
memcpy(start, info->tdata_start, info->tdata_size);
memset(start + info->tdata_size, 0, info->tbss_size);
if (__nacl_tp_tdb_offset(tdb_size) == 0) {
/*
* On x86 (but not on ARM), the TDB sits directly at $tp and the
* first word there must hold the $tp pointer itself.
*/
void *tdb = (char *) tp + __nacl_tp_tdb_offset(tdb_size);
*(void **) tdb = tdb;
}
return tp;
}
void *__nacl_tls_data_bss_initialize_from_template(void *combined_area,
size_t tdb_size) {
if (__nacl_tp_tdb_offset(tdb_size) != 0) {
/*
* This needs more work for ARM.
* For now abort via null pointer dereference.
*/
while (1) *(volatile int *) 0;
} else {
void *tdb = aligned_addr(((char *) combined_area) + tls_size +
SAFETY_PADDING , tls_align);
return _dl_allocate_tls(tdb);
}
}
static struct nc_combined_tdb *get_irt_tdb(void *thread_ptr) {
struct nc_combined_tdb *tdb = (void *) ((uintptr_t) thread_ptr +
__nacl_tp_tdb_offset(sizeof(*tdb)));
return tdb;
}
int pthread_create(pthread_t *thread_id,
const pthread_attr_t *attr,
void *(*start_routine)(void *),
void *arg) {
int retval = EAGAIN;
void *esp;
/* Declare the variables outside of the while scope. */
nc_thread_memory_block_t *stack_node = NULL;
char *thread_stack = NULL;
nc_thread_descriptor_t *new_tdb = NULL;
nc_basic_thread_data_t *new_basic_data = NULL;
nc_thread_memory_block_t *tls_node = NULL;
size_t stacksize = PTHREAD_STACK_DEFAULT;
void *new_tp;
/* TODO(gregoryd) - right now a single lock is used, try to optimize? */
pthread_mutex_lock(&__nc_thread_management_lock);
do {
/* Allocate the combined TLS + TDB block---see tls.h for explanation. */
tls_node = nc_allocate_memory_block_mu(TLS_AND_TDB_MEMORY,
__nacl_tls_combined_size(TDB_SIZE));
if (NULL == tls_node)
break;
new_tp = __nacl_tls_initialize_memory(nc_memory_block_to_payload(tls_node),
TDB_SIZE);
new_tdb = (nc_thread_descriptor_t *)
((char *) new_tp + __nacl_tp_tdb_offset(TDB_SIZE));
/*
* TODO(gregoryd): consider creating a pool of basic_data structs,
* similar to stack and TLS+TDB (probably when adding the support for
* variable stack size).
*/
new_basic_data = malloc(sizeof(*new_basic_data));
if (NULL == new_basic_data) {
/*
* The tdb should be zero intialized.
* This just re-emphasizes this requirement.
*/
new_tdb->basic_data = NULL;
break;
}
nc_tdb_init(new_tdb, new_basic_data);
new_tdb->tls_node = tls_node;
/*
* All the required members of the tdb must be initialized before
* the thread is started and actually before the global lock is released,
* since another thread can call pthread_join() or pthread_detach().
*/
new_tdb->start_func = start_routine;
new_tdb->state = arg;
if (attr != NULL) {
new_tdb->joinable = attr->joinable;
stacksize = attr->stacksize;
}
/* Allocate the stack for the thread. */
stack_node = nc_allocate_memory_block_mu(THREAD_STACK_MEMORY, stacksize);
if (NULL == stack_node) {
retval = EAGAIN;
break;
}
thread_stack = align((uint32_t) nc_memory_block_to_payload(stack_node),
kStackAlignment);
new_tdb->stack_node = stack_node;
retval = 0;
} while (0);
if (0 != retval) {
pthread_mutex_unlock(&__nc_thread_management_lock);
goto ret; /* error */
}
/*
* Speculatively increase the thread count. If thread creation
* fails, we will decrease it back. This way the thread count will
* never be lower than the actual number of threads, but can briefly
* be higher than that.
*/
++__nc_running_threads_counter;
/*
* Save the new thread id. This can not be done after the syscall,
* because the child thread could have already finished by that
* time. If thread creation fails, it will be overriden with -1
* later.
*/
*thread_id = new_basic_data;
pthread_mutex_unlock(&__nc_thread_management_lock);
/*
* Calculate the top-of-stack location. The very first location is a
* zero address of architecture-dependent width, needed to satisfy the
* normal ABI alignment requirements for the stack. (On some machines
* this is the dummy return address of the thread-start function.)
*
* Both thread_stack and stacksize are multiples of 16.
*/
esp = (void *) (thread_stack + stacksize - kStackPadBelowAlign);
memset(esp, 0, kStackPadBelowAlign);
/* Start the thread. */
retval = irt_thread.thread_create(
FUN_TO_VOID_PTR(nc_thread_starter), esp, new_tp);
if (0 != retval) {
pthread_mutex_lock(&__nc_thread_management_lock);
/* TODO(gregoryd) : replace with atomic decrement? */
--__nc_running_threads_counter;
pthread_mutex_unlock(&__nc_thread_management_lock);
goto ret;
}
assert(0 == retval);
ret:
if (0 != retval) {
/* Failed to create a thread. */
pthread_mutex_lock(&__nc_thread_management_lock);
nc_release_tls_node(tls_node, new_tdb);
if (new_basic_data) {
nc_release_basic_data_mu(new_basic_data);
}
if (stack_node) {
stack_node->is_used = 0;
nc_free_memory_block_mu(THREAD_STACK_MEMORY, stack_node);
}
pthread_mutex_unlock(&__nc_thread_management_lock);
*thread_id = NACL_PTHREAD_ILLEGAL_THREAD_ID;
}
return retval;
}