dsp_mem_region_t region;
dsp_component_id_t compid;
+ unsigned int stale_caches:1;
+ unsigned int req_sent:1;
} f;
+static noinline void dsp_fault(void)
+{
+ dsp_msg_t msg;
+
+ f.dsp_cache_inv_virt(worker, sizeof(*worker));
+ printf("dsp crash/fault/corruption:\n");
+ printf("state rdy/reap/done: %u %u %u\n",
+ worker->i_ready, worker->i_reaped, worker->i_done);
+ printf("active/boot: %u %u\n",
+ worker->active, worker->boot_cnt);
+
+ if (f.req_sent) {
+ f.dsp_rpc_recv(&msg);
+ f.req_sent = 0;
+ }
+ f.dsp_logbuf_print();
+ spu_config.iUseThread = 0;
+}
+
static void thread_work_start(void)
{
struct region_mem *mem;
// to start the DSP, dsp_rpc_send() must be used,
// but before that, previous request must be finished
- if (worker->req_sent) {
+ if (f.req_sent) {
if (worker->boot_cnt == worker->last_boot_cnt) {
// hopefully still booting
//printf("booting?\n");
if (ret != 0) {
fprintf(stderr, "dsp_rpc_recv failed: %d\n", ret);
f.dsp_logbuf_print();
- worker->req_sent = 0;
+ f.req_sent = 0;
spu_config.iUseThread = 0;
return;
}
f.dsp_cache_inv_virt(&worker->i_done, 64);
worker->last_boot_cnt = worker->boot_cnt;
+ worker->ram_dirty = spu.bMemDirty;
+ spu.bMemDirty = 0;
mem = (void *)f.region.virt_addr;
- memcpy(&mem->spu_config, &spu_config, sizeof(mem->spu_config));
+ memcpy(&mem->in.spu_config, &spu_config, sizeof(mem->in.spu_config));
DSP_MSG_INIT(&msg, f.compid, CCMD_DOIT, f.region.phys_addr, 0);
ret = f.dsp_rpc_send(&msg);
spu_config.iUseThread = 0;
return;
}
- worker->req_sent = 1;
+ f.req_sent = 1;
+
+#if 0
+ f.dsp_rpc_recv(&msg);
+ f.req_sent = 0;
+#endif
}
static int thread_get_i_done(void)
int limit = 1000;
int ns_to;
- ns_to = work->ns_to;
- f.dsp_cache_inv_virt(work->RVB, sizeof(work->RVB[0]) * 2 * ns_to);
- f.dsp_cache_inv_virt(work->SSumLR, sizeof(work->SSumLR[0]) * 2 * ns_to);
- __builtin_prefetch(work->RVB);
- __builtin_prefetch(work->SSumLR);
+ if ((unsigned int)(worker->i_done - worker->i_reaped) > WORK_MAXCNT) {
+ dsp_fault();
+ return;
+ }
while (worker->i_done == worker->i_reaped && limit-- > 0) {
- if (!worker->active) {
+ if (!f.req_sent) {
+ printf("dsp: req not sent?\n");
+ break;
+ }
+
+ if (worker->boot_cnt != worker->last_boot_cnt && !worker->active) {
printf("dsp: broken sync\n");
worker->last_boot_cnt = ~0;
break;
f.dsp_cache_inv_virt(&worker->i_done, 64);
}
+ ns_to = work->ns_to;
+ f.dsp_cache_inv_virt(work->SSumLR, sizeof(work->SSumLR[0]) * 2 * ns_to);
+ preload(work->SSumLR);
+ preload(work->SSumLR + 64/4);
+
+ f.stale_caches = 1; // SB, spuMem
+
if (limit == 0)
printf("dsp: wait timeout\n");
if (worker->i_reaped != worker->i_done - 1)
return;
- if (worker->req_sent && (force || worker->i_done == worker->i_ready)) {
+ if (f.req_sent && (force || worker->i_done == worker->i_ready)) {
dsp_msg_t msg;
int ret;
f.dsp_logbuf_print();
spu_config.iUseThread = 0;
}
- worker->req_sent = 0;
+ f.req_sent = 0;
}
+}
- if (force) {
+static void thread_sync_caches(void)
+{
+ if (f.stale_caches) {
f.dsp_cache_inv_virt(spu.SB, sizeof(spu.SB[0]) * SB_SIZE * 24);
f.dsp_cache_inv_virt(spu.spuMemC + 0x800, 0x800);
+ if (spu.rvb->StartAddr) {
+ int left = 0x40000 - spu.rvb->StartAddr;
+ f.dsp_cache_inv_virt(spu.spuMem + spu.rvb->StartAddr, left * 2);
+ }
+ f.stale_caches = 0;
}
}
}
mem = (void *)f.region.virt_addr;
- memcpy(&mem->spu_config, &spu_config, sizeof(mem->spu_config));
+ memcpy(&mem->in.spu_config, &spu_config, sizeof(mem->in.spu_config));
DSP_MSG_INIT(&init_msg, f.compid, CCMD_INIT, f.region.phys_addr, 0);
ret = f.dsp_rpc(&init_msg, &msg_in);
mem->sizeof_region_mem, sizeof(*mem));
goto fail_init;
}
- if (mem->offsetof_s_chan1 != offsetof(typeof(*mem), s_chan[1])) {
+ if (mem->offsetof_s_chan1 != offsetof(typeof(*mem), in.s_chan[1])) {
fprintf(stderr, "error: size mismatch 2: %d vs %zd\n",
- mem->offsetof_s_chan1, offsetof(typeof(*mem), s_chan[1]));
+ mem->offsetof_s_chan1, offsetof(typeof(*mem), in.s_chan[1]));
goto fail_init;
}
if (mem->offsetof_spos_3_20 != offsetof(typeof(*mem), worker.i[3].ch[20])) {
free(spu.SB);
spu.SB = mem->SB;
free(spu.s_chan);
- spu.s_chan = mem->s_chan;
+ spu.s_chan = mem->in.s_chan;
+ free(spu.rvb);
+ spu.rvb = &mem->in.rvb;
worker = &mem->worker;
printf("spu: C64x DSP ready (id=%d).\n", (int)f.compid);
if (worker == NULL)
return;
- if (worker->req_sent)
+ if (f.req_sent) {
f.dsp_rpc_recv(&msg);
+ f.req_sent = 0;
+ }
f.dsp_logbuf_print();
f.dsp_shm_free(f.region);
spu.spuMemC = NULL;
spu.SB = NULL;
spu.s_chan = NULL;
+ spu.rvb = NULL;
worker = NULL;
}
+/* debug: "access" shared mem from gdb */
+#if 0
+struct region_mem *dbg_dsp_mem;
+
+void dbg_dsp_mem_update(void)
+{
+ struct region_mem *mem;
+
+ if (dbg_dsp_mem == NULL)
+ dbg_dsp_mem = malloc(sizeof(*dbg_dsp_mem));
+ if (dbg_dsp_mem == NULL)
+ return;
+
+ mem = (void *)f.region.virt_addr;
+ f.dsp_cache_inv_virt(mem, sizeof(*mem));
+ memcpy(dbg_dsp_mem, mem, sizeof(*dbg_dsp_mem));
+}
+#endif
+
// vim:shiftwidth=1:expandtab