static void controller_smart_access_complete (controller *currctlr, ioreq_event *curr)
{
struct ioq *queue = currctlr->devices[curr->devno].queue;
ioreq_event *done = ioreq_copy(curr);
int devno = curr->devno;
int numout;
/* Responds to completion interrupt */
done->type = IO_INTERRUPT_COMPLETE;
currctlr->outbusowned = controller_get_downward_busno(currctlr, done, NULL);
controller_send_event_down_path(currctlr, done, currctlr->ovrhd_disk_complete);
currctlr->outbusowned = -1;
/* Handles request completion, including call-backs into cache */
curr = ioqueue_physical_access_done(queue, curr);
while ((done = curr)) {
curr = curr->next;
/* call back into cache with completion -- let it do request_complete */
controller_smart_wakeup(currctlr, currctlr->cache->cache_disk_access_complete(currctlr->cache, done));
}
/* Initiate another request, if any pending */
numout = ioqueue_get_reqoutstanding(queue);
if ((numout < currctlr->devices[devno].maxoutstanding) && (curr = ioqueue_get_next_request(queue))) {
controller_send_event_down_path(currctlr, curr, currctlr->ovrhd_disk_request);
}
}
static ioreq_event * handle_new_request (iodriver *curriodriver, ioreq_event *curr)
{
struct ioq *queue = curriodriver->devices[(curr->devno)].queue;
ioreq_event *ret = NULL;
/*
fprintf(outputfile, "\n*** handle_new_request:: time %f, devno %d, blkno %d, bcount %d\n\n",
simtime, curr->devno, curr->blkno, curr->bcount);
fprintf(outputfile, "handle_new_request:: calling ioqueue_add_new_request\n");
*/
ioqueue_add_new_request(queue, curr);
if (check_send_out_request(curriodriver, curr->devno)) {
/*
fprintf(outputfile, "handle_new_request:: calling ioqueue_get_next_request\n");
*/
ret = ioqueue_get_next_request(queue);
if (ret != NULL) {
schedule_disk_access(curriodriver, ret);
ret->time = IODRIVER_IMMEDSCHED_TIME * curriodriver->scale;
}
}
return(ret);
}
static void simpledisk_completion_done (ioreq_event *curr)
{
simpledisk_t *currdisk = getsimpledisk (curr->devno);
// fprintf (outputfile, "Entering simpledisk_completion for disk %d: %12.6f\n", currdisk->devno, simtime);
addtoextraq((event *) curr);
if (currdisk->busowned != -1) {
bus_ownership_release(currdisk->busowned);
currdisk->busowned = -1;
}
/* check for and start next queued request, if any */
curr = ioqueue_get_next_request(currdisk->queue);
if (curr != NULL) {
ASSERT (currdisk->media_busy == FALSE);
if (curr->flags & READ) {
currdisk->media_busy = TRUE;
stat_update (&currdisk->stat.acctimestats, currdisk->acctime);
curr->time = simtime + currdisk->acctime;
curr->type = DEVICE_ACCESS_COMPLETE;
addtointq ((event *)curr);
} else {
curr->type = IO_INTERRUPT_ARRIVE;
curr->cause = RECONNECT;
simpledisk_send_event_up_path (curr, currdisk->bus_transaction_latency);
currdisk->reconnect_reason = IO_INTERRUPT_ARRIVE;
}
}
}
static void ssd_check_channel_activity (ssd_t *currdisk)
{
while (1) {
ioreq_event *curr = currdisk->completion_queue;
currdisk->channel_activity = curr;
if (curr != NULL) {
currdisk->completion_queue = curr->next;
if (curr->flags & READ) {
/* transfer data up the line: curr->bcount, which is still set to */
/* original requested value, indicates how many blks to transfer. */
curr->type = DEVICE_DATA_TRANSFER_COMPLETE;
ssd_send_event_up_path(curr, (double) 0.0);
} else {
ssd_request_complete (curr);
}
} else {
curr = ioqueue_get_next_request(currdisk->queue);
currdisk->channel_activity = curr;
if (curr != NULL) {
if (curr->flags & READ) {
ssd_media_access_request(curr);
continue;
} else {
curr->cause = RECONNECT;
curr->type = IO_INTERRUPT_ARRIVE;
currdisk->reconnect_reason = IO_INTERRUPT_ARRIVE;
ssd_send_event_up_path (curr, currdisk->bus_transaction_latency);
}
}
}
break;
}
}
static void ssd_request_arrive (ioreq_event *curr)
{
ssd_t *currdisk;
// fprintf (outputfile, "Entering ssd_request_arrive: %12.6f\n", simtime);
// fprintf (outputfile, "ssd = %d, blkno = %d, bcount = %d, read = %d\n",curr->devno, curr->blkno, curr->bcount, (READ & curr->flags));
currdisk = getssd(curr->devno);
// verify that request is valid.
if ((curr->blkno < 0) || (curr->bcount <= 0) ||
((curr->blkno + curr->bcount) > currdisk->numblocks)) {
fprintf(outputfile3, "Invalid set of blocks requested from ssd - blkno %d, bcount %d, numblocks %d\n", curr->blkno, curr->bcount, currdisk->numblocks);
exit(1);
}
/* create a new request, set it up for initial interrupt */
ioqueue_add_new_request(currdisk->queue, curr);
if (currdisk->channel_activity == NULL) {
curr = ioqueue_get_next_request(currdisk->queue);
currdisk->busowned = ssd_get_busno(curr);
currdisk->channel_activity = curr;
currdisk->reconnect_reason = IO_INTERRUPT_ARRIVE;
if (curr->flags & READ) {
ssd_media_access_request (curr);
ssd_check_channel_activity(currdisk);
} else {
curr->cause = READY_TO_TRANSFER;
curr->type = IO_INTERRUPT_ARRIVE;
ssd_send_event_up_path(curr, currdisk->bus_transaction_latency);
}
}
}
static void simpledisk_request_arrive (ioreq_event *curr)
{
ioreq_event *intrp;
simpledisk_t *currdisk;
#ifdef DEBUG_SIMPLEDISK
fprintf (outputfile, "*** %f: simpledisk_request_arrive - devno %d, blkno %d, bcount %d, flags 0x%x\n", simtime, curr->devno, curr->blkno, curr->bcount, curr->flags );
#endif
currdisk = getsimpledisk(curr->devno);
/* verify that request is valid. */
if ((curr->blkno < 0) || (curr->bcount <= 0) ||
((curr->blkno + curr->bcount) > currdisk->numblocks)) {
fprintf(stderr, "Invalid set of blocks requested from simpledisk - blkno %lld, bcount %d, numblocks %lld\n", curr->blkno, curr->bcount, currdisk->numblocks);
exit(1);
}
/* create a new request, set it up for initial interrupt */
currdisk->busowned = simpledisk_get_busno(curr);
if (ioqueue_get_reqoutstanding (currdisk->queue) == 0) {
ioqueue_add_new_request(currdisk->queue, curr);
curr = ioqueue_get_next_request (currdisk->queue);
intrp = curr;
/* initiate media access if request is a READ */
if (curr->flags & READ) {
ioreq_event *tmp = ioreq_copy (curr);
currdisk->media_busy = TRUE;
stat_update (&currdisk->stat.acctimestats, currdisk->acctime);
tmp->time = simtime + currdisk->acctime;
tmp->type = DEVICE_ACCESS_COMPLETE;
addtointq ((event *)tmp);
}
/* if not disconnecting, then the READY_TO_TRANSFER is like a RECONNECT */
currdisk->reconnect_reason = IO_INTERRUPT_ARRIVE;
if (curr->flags & READ) {
intrp->cause = (currdisk->neverdisconnect) ? READY_TO_TRANSFER : DISCONNECT;
} else {
intrp->cause = READY_TO_TRANSFER;
}
} else {
intrp = ioreq_copy(curr);
ioqueue_add_new_request(currdisk->queue, curr);
intrp->cause = DISCONNECT;
}
intrp->type = IO_INTERRUPT_ARRIVE;
simpledisk_send_event_up_path(intrp, currdisk->bus_transaction_latency);
}
static void controller_smart_issue_access (void *issuefuncparam, ioreq_event *curr)
{
controller *currctlr = issuefuncparam;
struct ioq *queue = currctlr->devices[curr->devno].queue;
int numout = ioqueue_get_reqoutstanding(queue);
/* in case the cache changes to which device the request is sent */
//fprintf (stderr, "busno %x, buspath %x, slotno %x, slotpath %x\n", curr->busno, currctlr->devices[curr->devno].buspath.value, curr->slotno, currctlr->devices[curr->devno].slotpath.value);
curr->busno = currctlr->devices[curr->devno].buspath.value;
curr->slotno = currctlr->devices[curr->devno].slotpath.value;
ioqueue_add_new_request(queue, curr);
if (numout < currctlr->devices[curr->devno].maxoutstanding) {
ioreq_event *sched = ioqueue_get_next_request(queue);
controller_send_event_down_path(currctlr, sched, currctlr->ovrhd_disk_request);
}
}
void iodriver_access_complete (int iodriverno, intr_event *intrp)
{
int i;
int numreqs;
ioreq_event *tmp;
ioreq_event *del;
ioreq_event *req;
int devno;
int skip = 0;
ctlr *ctl = NULL;
time_t now;
if (iodrivers[iodriverno]->type == STANDALONE) {
req = ioreq_copy((ioreq_event *) intrp->infoptr);
}
else {
req = (ioreq_event *) intrp->infoptr;
}
#ifdef DEBUG_IODRIVER
fprintf (outputfile, "*** %f: iodriver_access_complete - devno %d, blkno %d, bcount %d, read %d\n", simtime, req->devno, req->blkno, req->bcount, (req->flags & READ));
fflush(outputfile);
#endif
time( & now );
disksim_exectrace( "Request completion: simtime %f, devno %d, blkno %lld, bcount %d, flags %X, time %s\n", simtime, req->devno, req->blkno, req->bcount, req->flags, asctime( localtime(& now)) );
if (iodrivers[iodriverno]->devices[(req->devno)].queuectlr != -1)
{
int ctlrno = iodrivers[iodriverno]->devices[(req->devno)].queuectlr;
ctl = &iodrivers[iodriverno]->ctlrs[ctlrno];
tmp = ctl->oversized;
numreqs = 1;
while (((numreqs) || (tmp != ctl->oversized))
&& (tmp)
&& (tmp->next)
&& ((tmp->next->devno != req->devno)
|| (tmp->next->opid != req->opid)
|| (req->blkno < tmp->next->blkno)
|| (req->blkno >= (tmp->next->blkno + tmp->next->bcount))))
{
// fprintf (outputfile, "oversized request in list: opid %d, blkno %lld, bcount %d\n", tmp->opid, tmp->blkno, tmp->bcount);
numreqs = 0;
tmp = tmp->next;
}
if ((tmp)
&& (tmp->next->devno == req->devno)
&& (tmp->next->opid == req->opid)
&& (req->blkno >= tmp->next->blkno)
&& (req->blkno < (tmp->next->blkno + tmp->next->bcount)))
{
fprintf (outputfile, "%f, part of oversized request completed: opid %d, blkno %lld, bcount %d, maxreqsize %d\n", simtime, req->opid, req->blkno, req->bcount, ctl->maxreqsize);
if ((req->blkno + ctl->maxreqsize) < (tmp->next->blkno + tmp->next->bcount))
{
fprintf (outputfile, "more to go\n");
req->blkno += ctl->maxreqsize;
req->bcount = min(ctl->maxreqsize,
(tmp->next->blkno + tmp->next->bcount - req->blkno));
goto schedule_next;
}
else {
fprintf (outputfile, "done for real\n");
addtoextraq((event *) req);
req = tmp->next;
tmp->next = tmp->next->next;
if (ctl->oversized == req) {
ctl->oversized = (req != req->next) ? req->next : NULL;
}
req->next = NULL;
}
}
}
devno = req->devno;
req = ioqueue_physical_access_done(iodrivers[iodriverno]->devices[devno].queue, req);
if (ctl) {
ctl->numoutstanding--;
}
// special case for validate:
if (disksim->traceformat == VALIDATE) {
tmp = (ioreq_event *) getfromextraq();
io_validate_do_stats1();
tmp = iotrace_validate_get_ioreq_event(disksim->iotracefile, tmp);
if (tmp) {
io_validate_do_stats2(tmp);
tmp->type = IO_REQUEST_ARRIVE;
addtointq((event *) tmp);
disksim_exectrace("Request issue: simtime %f, devno %d, blkno %lld, time %f\n",
simtime, tmp->devno, tmp->blkno, tmp->time);
}
else {
disksim_simstop();
}
}
else if (disksim->closedios) {
tmp = (ioreq_event *) io_get_next_external_event(disksim->iotracefile);
if (tmp) {
io_using_external_event ((event *)tmp);
tmp->time = simtime + disksim->closedthinktime;
tmp->type = IO_REQUEST_ARRIVE;
addtointq((event *) tmp);
} else {
disksim_simstop();
}
}
while (req) {
tmp = req;
req = req->next;
tmp->next = NULL;
update_iodriver_statistics();
if ((numreqs = logorg_mapcomplete(sysorgs, numsysorgs, tmp)) == COMPLETE) {
/* update up overall I/O system stats for this completed request */
ioreq_event *temp = ioqueue_get_specific_request (OVERALLQUEUE, tmp);
ioreq_event *temp2 = ioqueue_physical_access_done (OVERALLQUEUE, temp);
ASSERT (temp2 != NULL);
addtoextraq((event *)temp);
temp = NULL;
if (iodrivers[iodriverno]->type != STANDALONE) {
iodriver_add_to_intrp_eventlist(intrp,
io_done_notify(tmp),
iodrivers[iodriverno]->scale);
} else {
io_done_notify (tmp);
}
}
else if (numreqs > 0) {
for (i = 0; i < numreqs; i++) {
del = tmp->next;
tmp->next = del->next;
del->next = NULL;
del->type = IO_REQUEST_ARRIVE;
del->flags |= MAPPED;
skip |= (del->devno == devno);
if (iodrivers[iodriverno]->type == STANDALONE)
{
del->time += simtime + 0.0000000001; /* to affect an ordering */
addtointq((event *) del);
}
else {
iodriver_add_to_intrp_eventlist(intrp,
(event *) del,
iodrivers[iodriverno]->scale);
}
}
}
addtoextraq((event *) tmp);
}
if ((iodrivers[iodriverno]->consttime == IODRIVER_TRACED_QUEUE_TIMES)
|| (iodrivers[iodriverno]->consttime == IODRIVER_TRACED_BOTH_TIMES))
{
if (ioqueue_get_number_in_queue(iodrivers[iodriverno]->devices[devno].queue) > 0)
{
iodrivers[iodriverno]->devices[devno].flag = 1;
iodrivers[iodriverno]->devices[devno].lastevent = simtime;
}
return;
}
if (skip) {
return;
}
// fprintf(outputfile, "iodriver_access_complete:: calling ioqueue_get_next_request\n");
req = ioqueue_get_next_request(iodrivers[iodriverno]->devices[devno].queue);
// fprintf (outputfile, "next scheduled: req %p, req->blkno %d, req->flags %x\n", req, ((req) ? req->blkno : 0), ((req) ? req->flags : 0));
schedule_next:
if (req) {
req->type = IO_ACCESS_ARRIVE;
req->next = NULL;
if (ctl) {
ctl->numoutstanding++;
}
if (iodrivers[iodriverno]->type == STANDALONE) {
req->time = simtime;
addtointq((event *) req);
}
else {
iodriver_add_to_intrp_eventlist(intrp,
(event *) req,
iodrivers[iodriverno]->scale);
}
}
}
/*
* collects 1 request from each chip in the gang
*/
static void ssd_collect_req_in_gang
(ssd_t *s, int gang_num, ssd_req ***rd_q, ssd_req ***wr_q, int *rd_total, int *wr_total)
{
int i;
int start;
gang_metadata *g;
g = &s->gang_meta[gang_num];
// start from the first element of the gang
start = gang_num * s->params.elements_per_gang;
i = start;
*rd_total = 0;
*wr_total = 0;
do {
ssd_element *elem;
ioreq_event *req;
int tot_rd_reqs;
int tot_wr_reqs;
int j;
elem = &s->elements[i];
ASSERT(ioqueue_get_reqoutstanding(elem->queue) == 0);
j = i % s->params.elements_per_gang;
// collect the requests
tot_rd_reqs = 0;
tot_wr_reqs = 0;
if ((req = ioqueue_get_next_request(elem->queue)) != NULL) {
int found;
if (req->flags & READ) {
found = ssd_already_present(rd_q[j], tot_rd_reqs, req);
} else {
found = ssd_already_present(wr_q[j], tot_wr_reqs, req);
}
if (!found) {
// this is a valid request
ssd_req *r = malloc(sizeof(ssd_req));
r->blk = req->blkno;
r->count = req->bcount;
r->is_read = req->flags & READ;
r->org_req = req;
r->plane_num = -1; // we don't know to which plane this req will be directed at
if (req->flags & READ) {
rd_q[j][tot_rd_reqs] = r;
tot_rd_reqs ++;
} else {
wr_q[j][tot_wr_reqs] = r;
tot_wr_reqs ++;
}
} else {
// throw this request -- it doesn't make sense
stat_update (&s->stat.acctimestats, 0);
req->time = simtime;
req->ssd_elem_num = i;
req->ssd_gang_num = gang_num;
req->type = DEVICE_ACCESS_COMPLETE;
addtointq ((event *)req);
}
ASSERT((tot_rd_reqs < MAX_REQS) && (tot_wr_reqs < MAX_REQS))
}
*rd_total = *rd_total + tot_rd_reqs;
*wr_total = *wr_total + tot_wr_reqs;
// go to the next element
i = ssd_next_elem_in_gang(s, gang_num, i);
} while (i != start);
static void ssd_activate_elem(ssd_t *currdisk, int elem_num)
{
ioreq_event *req;
ssd_req **read_reqs;
ssd_req **write_reqs;
int i;
int read_total = 0;
int write_total = 0;
double schtime = 0;
int max_reqs;
int tot_reqs_issued;
double max_time_taken = 0;
ssd_element *elem = &currdisk->elements[elem_num];
// if the media is busy, we can't do anything, so return
if (elem->media_busy == TRUE) {
return;
}
ASSERT(ioqueue_get_reqoutstanding(elem->queue) == 0);
// we can invoke cleaning in the background whether there
// is request waiting or not
if (currdisk->params.cleaning_in_background) {
// if cleaning was invoked, wait until
// it is over ...
if (ssd_invoke_element_cleaning(elem_num, currdisk)) {
return;
}
}
ASSERT(elem->metadata.reqs_waiting == ioqueue_get_number_in_queue(elem->queue));
if (elem->metadata.reqs_waiting > 0) {
// invoke cleaning in foreground when there are requests waiting
if (!currdisk->params.cleaning_in_background) {
// if cleaning was invoked, wait until
// it is over ...
if (ssd_invoke_element_cleaning(elem_num, currdisk)) {
return;
}
}
// how many reqs can we issue at once
if (currdisk->params.copy_back == SSD_COPY_BACK_DISABLE) {
max_reqs = 1;
} else {
if (currdisk->params.num_parunits == 1) {
max_reqs = 1;
} else {
max_reqs = MAX_REQS_ELEM_QUEUE;
}
}
// ideally, we should issue one req per plane, overlapping them all.
// in order to simplify the overlapping strategy, let's issue
// requests of the same type together.
read_reqs = (ssd_req **) malloc(max_reqs * sizeof(ssd_req *));
write_reqs = (ssd_req **) malloc(max_reqs * sizeof(ssd_req *));
// collect the requests
while ((req = ioqueue_get_next_request(elem->queue)) != NULL) {
int found = 0;
elem->metadata.reqs_waiting --;
// see if we already have the same request in the list.
// this usually doesn't happen -- but on synthetic traces
// this weird case can occur.
if (req->flags & READ) {
found = ssd_already_present(read_reqs, read_total, req);
} else {
found = ssd_already_present(write_reqs, write_total, req);
}
if (!found) {
// this is a valid request
ssd_req *r = malloc(sizeof(ssd_req));
r->blk = req->blkno;
r->count = req->bcount;
r->is_read = req->flags & READ;
r->org_req = req;
r->plane_num = -1; // we don't know to which plane this req will be directed at
if (req->flags & READ) {
read_reqs[read_total] = r;
read_total ++;
} else {
write_reqs[write_total] = r;
write_total ++;
}
// if we have more reqs than we can handle, quit
if ((read_total >= max_reqs) ||
(write_total >= max_reqs)) {
break;
}
} else {
// throw this request -- it doesn't make sense
stat_update (&currdisk->stat.acctimestats, 0);
req->time = simtime;
req->ssd_elem_num = elem_num;
req->type = DEVICE_ACCESS_COMPLETE;
addtointq ((event *)req);
}
}
if (read_total > 0) {
// first issue all the read requests (it doesn't matter what we
// issue first). i chose read because reads are mostly synchronous.
// find the time taken to serve these requests.
ssd_compute_access_time(currdisk, elem_num, read_reqs, read_total);
// add an event for each request completion
for (i = 0; i < read_total; i ++) {
elem->media_busy = TRUE;
// find the maximum time taken by a request
if (schtime < read_reqs[i]->schtime) {
schtime = read_reqs[i]->schtime;
}
stat_update (&currdisk->stat.acctimestats, read_reqs[i]->acctime);
read_reqs[i]->org_req->time = simtime + read_reqs[i]->schtime;
read_reqs[i]->org_req->ssd_elem_num = elem_num;
read_reqs[i]->org_req->type = DEVICE_ACCESS_COMPLETE;
//printf("R: blk %d elem %d acctime %f simtime %f\n", read_reqs[i]->blk,
// elem_num, read_reqs[i]->acctime, read_reqs[i]->org_req->time);
addtointq ((event *)read_reqs[i]->org_req);
free(read_reqs[i]);
}
}
free(read_reqs);
max_time_taken = schtime;
if (write_total > 0) {
// next issue the write requests
ssd_compute_access_time(currdisk, elem_num, write_reqs, write_total);
// add an event for each request completion.
// note that we can issue the writes only after all the reads above are
// over. so, include the maximum read time when creating the event.
for (i = 0; i < write_total; i ++) {
elem->media_busy = TRUE;
stat_update (&currdisk->stat.acctimestats, write_reqs[i]->acctime);
write_reqs[i]->org_req->time = simtime + schtime + write_reqs[i]->schtime;
//printf("blk %d elem %d acc time %f\n", write_reqs[i]->blk, elem_num, write_reqs[i]->acctime);
if (max_time_taken < (schtime+write_reqs[i]->schtime)) {
max_time_taken = (schtime+write_reqs[i]->schtime);
}
write_reqs[i]->org_req->ssd_elem_num = elem_num;
write_reqs[i]->org_req->type = DEVICE_ACCESS_COMPLETE;
//printf("W: blk %d elem %d acctime %f simtime %f\n", write_reqs[i]->blk,
// elem_num, write_reqs[i]->acctime, write_reqs[i]->org_req->time);
addtointq ((event *)write_reqs[i]->org_req);
free(write_reqs[i]);
}
}
free(write_reqs);
// statistics
tot_reqs_issued = read_total + write_total;
ASSERT(tot_reqs_issued > 0);
currdisk->elements[elem_num].stat.tot_reqs_issued += tot_reqs_issued;
currdisk->elements[elem_num].stat.tot_time_taken += max_time_taken;
}
}