maki/openairinterface5g
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Merge remote-tracking branch 'origin/nr-ue-threading-improvements' into integration_2024_w48 (!3091)

Browse Source 
Update NR UE threading model

This updates the NR UE threading model by preventing calling threadpool
from threadpool and fixes deadlock issues related to processSlotTX. By
making processSlotTX run to completion instead of starting and waiting
for other threads to finish it prevents locking thread pool cores,
therefore prevening deadlocks.

There is a slight difference here compared to the previous version where
I do not use the UL actor as I believe it is not necessary,
processSlotTX satisfies the run-to-completion requirement now. If in the
future any UL procedure would be parallelized using the thread pool it
should either be done in a run-to-completion model or the processSlotTX
function should be taken off the thread pool.

The documentation has been updated.
This commit is contained in:
Robert Schmidt
2024-11-29 15:51:56 +01:00
parent a61b5f303d 97269a9d2d
commit ca3739e806
25 changed files with 631 additions and 111 deletions
Download Patch File Download Diff File Expand all files Collapse all files
executables/nr-ue.c
+19 -60
View File
@@ -100,6 +100,11 @@
static void *NRUE_phy_stub_standalone_pnf_task(void *arg);
static void start_process_slot_tx(void* arg) {
notifiedFIFO_elt_t *newTx = arg;
pushTpool(&(get_nrUE_params()->Tpool), newTx);
}
static size_t dump_L1_UE_meas_stats(PHY_VARS_NR_UE *ue, char *output, size_t max_len)
{
const char *begin = output;
@@ -488,56 +493,10 @@ void processSlotTX(void *arg)
nr_phy_data_tx_t phy_data = {0};
bool sl_tx_action = false;
// Force sequential execution, even if we launch in // for all slots
// at least ULstatus variable is a pure race condition that is quickly detected by assert() in the code because one thread sets it
// to active, so the other thread try to steal&run the ul work
if (rxtxD->stream_status == STREAM_STATUS_SYNCED) {
notifiedFIFO_elt_t *res = pullNotifiedFIFO(UE->tx_resume_ind_fifo + proc->nr_slot_tx);
delNotifiedFIFO_elt(res);
}
if (UE->if_inst)
UE->if_inst->slot_indication(UE->Mod_id);
LOG_D(PHY,
"SlotTx %d.%d => slot type %d, wait: %d \n",
proc->frame_tx,
proc->nr_slot_tx,
proc->tx_slot_type,
rxtxD->tx_wait_for_dlsch);
if (proc->tx_slot_type == NR_UPLINK_SLOT || proc->tx_slot_type == NR_MIXED_SLOT) {
if (rxtxD->tx_wait_for_dlsch)
LOG_D(PHY, "enter wait for tx, slot %d, nb events to wait %d; ", proc->nr_slot_tx, rxtxD->tx_wait_for_dlsch);
// wait for rx slots to send indication (if any) that DLSCH decoding is finished
for(int i=0; i < rxtxD->tx_wait_for_dlsch; i++) {
notifiedFIFO_elt_t *res = pullNotifiedFIFO(UE->tx_resume_ind_fifo + proc->nr_slot_tx);
delNotifiedFIFO_elt(res);
}
LOG_D(PHY, "completed wait for tx, slot %d\n", proc->nr_slot_tx);
/*
This herafter code is costing some perfomance for a check that should be useless
But, we face today several bugs arround the matching between events in UE->tx_resume_ind_fifo[slot]
and the corresponding tx_wait_for_dlsch[slot]
The algorithm is we accumlate the actions that should end before processing a tx slot in tx_wait_for_dlsch[slot]
later, other threads push events in UE->tx_resume_ind_fifo[slot]
so, the tx encoding starts only when related actions are done (mainly DLSCH ACK/NACK to encode PUCCH)
if there is a bug that misses to send a event in UE->tx_resume_ind_fifo[slot], the process hangs, we detect the issue
if there is a bug that makes a extra event in UE->tx_resume_ind_fifo[slot], and if we drop the hereafter check
the system runs with random race conditions, very hard to debug
Likely we should later remove completly UE->tx_resume_ind_fifo with notifications,
instead,
we may run in place the processSlotTX() when the conditions are met (when a decreasing tx_wait_for_dlsch[slot] will become 0)
It will remove the condition signals (for a thread safe semaphore or counter) and make the system simpler
*/
notifiedFIFO_elt_t *res = pollNotifiedFIFO(UE->tx_resume_ind_fifo + proc->nr_slot_tx);
if (res)
LOG_E(NR_PHY,
"Internal error: extra event on Tx waiting queue for slot %d, event comes from rx slot %d\n",
proc->nr_slot_tx,
*(int *)NotifiedFifoData(res));
if (UE->sl_mode == 2 && proc->tx_slot_type == NR_SIDELINK_SLOT) {
// trigger L2 to run ue_sidelink_scheduler thru IF module
if (UE->if_inst != NULL && UE->if_inst->sl_indication != NULL) {
@@ -586,13 +545,10 @@ void processSlotTX(void *arg)
}
}
notifiedFIFO_elt_t *newElt = newNotifiedFIFO_elt(sizeof(int), 0, NULL, NULL);
int *msgData = (int *)NotifiedFifoData(newElt);
int slots_per_frame = (UE->sl_mode == 2) ? UE->SL_UE_PHY_PARAMS.sl_frame_params.slots_per_frame
: UE->frame_parms.slots_per_frame;
int newslot = (proc->nr_slot_tx + 1) % slots_per_frame;
*msgData = newslot;
pushNotifiedFIFO(UE->tx_resume_ind_fifo + newslot, newElt);
int next_slot = (proc->nr_slot_tx + 1) % slots_per_frame;
dynamic_barrier_join(&UE->process_slot_tx_barriers[next_slot]);
RU_write(rxtxD, sl_tx_action);
TracyCZoneEnd(ctx);
}
@@ -849,7 +805,7 @@ void *UE_thread(void *arg)
int num_ind_fifo = nb_slot_frame;
for(int i = 0; i < num_ind_fifo; i++) {
initNotifiedFIFO(UE->tx_resume_ind_fifo + i);
dynamic_barrier_init(&UE->process_slot_tx_barriers[i]);
}
int shiftForNextFrame = 0;
int intialSyncOffset = 0;
@@ -924,7 +880,7 @@ void *UE_thread(void *arg)
}
syncMsg->UE = UE;
memset(&syncMsg->proc, 0, sizeof(syncMsg->proc));
pushTpool(&(get_nrUE_params()->Tpool), Msg);
pushNotifiedFIFO(&UE->sync_actor.fifo, Msg);
trashed_frames = 0;
syncRunning = true;
continue;
@@ -964,9 +920,7 @@ void *UE_thread(void *arg)
// We have resynchronized, maybe after RF loss so we need to purge any existing context
memset(tx_wait_for_dlsch, 0, sizeof(tx_wait_for_dlsch));
for (int i = 0; i < num_ind_fifo; i++) {
notifiedFIFO_elt_t *res;
while ((res = pollNotifiedFIFO(UE->tx_resume_ind_fifo + i)))
delNotifiedFIFO_elt(res);
dynamic_barrier_reset(&UE->process_slot_tx_barriers[i]);
}
continue;
}
@@ -1055,7 +1009,7 @@ void *UE_thread(void *arg)
int ret = UE_dl_preprocessing(UE, &curMsgRx->proc, tx_wait_for_dlsch, &curMsgRx->phy_data, &stats_printed);
if (ret != INT_MAX)
shiftForNextFrame = ret;
pushTpool(&(get_nrUE_params()->Tpool), newRx);
pushNotifiedFIFO(&UE->dl_actors[curMsg.proc.nr_slot_rx % NUM_DL_ACTORS].fifo, newRx);
// Start TX slot processing here. It runs in parallel with RX slot processing
// in current code, DURATION_RX_TO_TX constant is the limit to get UL data to encode from a RX slot
@@ -1065,11 +1019,16 @@ void *UE_thread(void *arg)
curMsgTx->writeBlockSize = writeBlockSize;
curMsgTx->proc.timestamp_tx = writeTimestamp;
curMsgTx->UE = UE;
curMsgTx->tx_wait_for_dlsch = tx_wait_for_dlsch[curMsgTx->proc.nr_slot_tx];
curMsgTx->stream_status = stream_status;
int sync_to_previous_thread = stream_status == STREAM_STATUS_SYNCED ? 1 : 0;
int slot = curMsgTx->proc.nr_slot_tx;
dynamic_barrier_update(&UE->process_slot_tx_barriers[slot],
tx_wait_for_dlsch[slot] + sync_to_previous_thread,
start_process_slot_tx,
newTx);
stream_status = STREAM_STATUS_SYNCED;
tx_wait_for_dlsch[curMsgTx->proc.nr_slot_tx] = 0;
pushTpool(&(get_nrUE_params()->Tpool), newTx);
tx_wait_for_dlsch[slot] = 0;
}
return NULL;
executables/nr-uesoftmodem.c
+5
View File
@@ -88,6 +88,7 @@ unsigned short config_frames[4] = {2,9,11,13};
#include <openair1/PHY/MODULATION/nr_modulation.h>
#include "openair2/GNB_APP/gnb_paramdef.h"
#include "pdcp.h"
#include "actor.h"
extern const char *duplex_mode[];
THREAD_STRUCT thread_struct;
@@ -495,6 +496,10 @@ int main(int argc, char **argv)
}
UE[CC_id]->sl_mode = get_softmodem_params()->sl_mode;
init_actor(&UE[CC_id]->sync_actor, "SYNC_", -1);
for (int i = 0; i < NUM_DL_ACTORS; i++) {
init_actor(&UE[CC_id]->dl_actors[i], "DL_", -1);
}
init_nr_ue_vars(UE[CC_id], inst);
if (UE[CC_id]->sl_mode) {