hsa_status_t HSA_API
hsa_amd_profiling_get_dispatch_time(hsa_agent_t agent_handle,
hsa_signal_t hsa_signal,
hsa_amd_profiling_dispatch_time_t* time) {
IS_OPEN();
IS_BAD_PTR(time);
core::Agent* agent = core::Agent::Convert(agent_handle);
IS_VALID(agent);
core::Signal* signal = core::Signal::Convert(hsa_signal);
IS_VALID(signal);
if (agent->device_type() != core::Agent::kAmdGpuDevice) {
return HSA_STATUS_ERROR_INVALID_AGENT;
}
amd::GpuAgentInt* gpu_agent = static_cast<amd::GpuAgentInt*>(agent);
gpu_agent->TranslateTime(signal, *time);
return HSA_STATUS_SUCCESS;
}
hsa_status_t HSA_API hsa_amd_queue_cu_set_mask(const hsa_queue_t* queue,
uint32_t num_cu_mask_count,
const uint32_t* cu_mask) {
IS_OPEN();
IS_BAD_PTR(cu_mask);
core::Queue* cmd_queue = core::Queue::Convert(queue);
IS_VALID(cmd_queue);
return cmd_queue->SetCUMasking(num_cu_mask_count, cu_mask);
}
hsa_status_t HSA_API
hsa_amd_signal_async_handler(hsa_signal_t hsa_signal,
hsa_signal_condition_t cond,
hsa_signal_value_t value,
hsa_amd_signal_handler handler, void* arg) {
IS_OPEN();
core::Signal* signal = core::Signal::Convert(hsa_signal);
IS_VALID(signal);
IS_BAD_PTR(handler);
return core::Runtime::runtime_singleton_->SetAsyncSignalHandler(
hsa_signal, cond, value, handler, arg);
}
hsa_status_t HSA_API hsa_ext_get_memory_type(hsa_agent_t agent_handle,
hsa_amd_memory_type_t* type) {
const core::Agent* agent = core::Agent::Convert(agent_handle);
IS_VALID(agent);
IS_BAD_PTR(type);
if (agent->device_type() != core::Agent::kAmdGpuDevice) {
return HSA_STATUS_ERROR_INVALID_AGENT;
}
const amd::GpuAgentInt* gpu_agent =
static_cast<const amd::GpuAgentInt*>(agent);
*type = gpu_agent->memory_type();
return HSA_STATUS_SUCCESS;
}
static int slp_io_start(const struct lu_env *env, const struct cl_io_slice *ios)
{
struct ccc_io *cio = cl2ccc_io(env, ios);
struct cl_io *io = ios->cis_io;
struct cl_object *obj = io->ci_obj;
struct inode *inode = ccc_object_inode(obj);
int err, ret;
loff_t pos;
long cnt;
struct llu_io_group *iogroup;
struct lustre_rw_params p = {0};
int iovidx;
struct intnl_stat *st = llu_i2stat(inode);
struct llu_inode_info *lli = llu_i2info(inode);
struct llu_io_session *session = cl2slp_io(env, ios)->sio_session;
int write = io->ci_type == CIT_WRITE;
int exceed = 0;
CLOBINVRNT(env, obj, ccc_object_invariant(obj));
if (write) {
pos = io->u.ci_wr.wr.crw_pos;
cnt = io->u.ci_wr.wr.crw_count;
} else {
pos = io->u.ci_rd.rd.crw_pos;
cnt = io->u.ci_rd.rd.crw_count;
}
if (io->u.ci_wr.wr_append) {
p.lrp_lock_mode = LCK_PW;
} else {
p.lrp_brw_flags = OBD_BRW_SRVLOCK;
p.lrp_lock_mode = LCK_NL;
}
iogroup = get_io_group(inode, max_io_pages(cnt, cio->cui_nrsegs), &p);
if (IS_ERR(iogroup))
RETURN(PTR_ERR(iogroup));
err = ccc_prep_size(env, obj, io, pos, cnt, &exceed);
if (err != 0 || (write == 0 && exceed != 0))
GOTO(out, err);
CDEBUG(D_INODE,
"%s ino %lu, %lu bytes, offset "LPU64", i_size "LPU64"\n",
write ? "Write" : "Read", (unsigned long)st->st_ino,
cnt, (__u64)pos, (__u64)st->st_size);
if (write && io->u.ci_wr.wr_append)
pos = io->u.ci_wr.wr.crw_pos = st->st_size; /* XXX? Do we need to change io content too here? */
/* XXX What about if one write syscall writes at 2 different offsets? */
for (iovidx = 0; iovidx < cio->cui_nrsegs; iovidx++) {
char *buf = (char *) cio->cui_iov[iovidx].iov_base;
long count = cio->cui_iov[iovidx].iov_len;
if (!count)
continue;
if (cnt < count)
count = cnt;
if (IS_BAD_PTR(buf) || IS_BAD_PTR(buf + count)) {
GOTO(out, err = -EFAULT);
}
if (io->ci_type == CIT_READ) {
if (/* local_lock && */ pos >= st->st_size)
break;
} else if (io->ci_type == CIT_WRITE) {
if (pos >= lli->lli_maxbytes) {
GOTO(out, err = -EFBIG);
}
if (pos + count >= lli->lli_maxbytes)
count = lli->lli_maxbytes - pos;
} else {
LBUG();
}
ret = llu_queue_pio(env, io, iogroup, buf, count, pos);
if (ret < 0) {
GOTO(out, err = ret);
} else {
io->ci_nob += ret;
pos += ret;
cnt -= ret;
if (io->ci_type == CIT_WRITE) {
// obd_adjust_kms(exp, lsm, pos, 0); // XXX
if (pos > st->st_size)
st->st_size = pos;
}
if (!cnt)
break;
}
}
LASSERT(cnt == 0 || io->ci_type == CIT_READ); /* libsysio should guarantee this */
if (!iogroup->lig_rc)
session->lis_rwcount += iogroup->lig_rwcount;
else if (!session->lis_rc)
session->lis_rc = iogroup->lig_rc;
err = 0;
out:
put_io_group(iogroup);
return err;
}