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

Merge remote-tracking branch 'origin/fhi72-no-polling' into integration_2025_w07 (!3200)

Browse Source 
FHI72: remove polling, fix memory leak

- FHI72 used polling to track the next slot. This is not necessary, a
  simple queue does the job and allows to save one core that was
  previously at 100%
- Remove a workaround to avoid a memory leak (fixes #812); allows to
  also run RFsim/USRP/otehr radios when compiling for FHI72
This commit is contained in:
Robert Schmidt
2025-02-17 15:10:01 +01:00
parent 7d574ffcfb e1d7306a43
commit f0d6d42aec
6 changed files with 113 additions and 103 deletions
Download Patch File Download Diff File Expand all files Collapse all files
common/utils/threadPool/notified_fifo.h
+1
View File
@@ -23,6 +23,7 @@
#define NOTIFIED_FIFO_H
#include "pthread_utils.h"
#include <stdint.h>
#include <malloc.h>
#include <pthread.h>
#include "time_meas.h"
#include <memory.h>
doc/ORAN_FHI7.2_Tutorial.md
+15
View File
@@ -475,6 +475,19 @@ cmake .. -GNinja -DOAI_FHI72=ON -Dxran_LOCATION=$HOME/phy/fhi_lib/lib
ninja nr-softmodem oran_fhlib_5g params_libconfig
```
Note that in tags 2025.w06 and prior, the FHI72 driver used polling to wait for
the next slot. This is inefficient as it burns CPU time, and has been replaced
with a more efficient mechanism. Nevertheless, if you experience problems that
did not occur previously, it is possible to re-enable polling, either with
`build_oai` like this
./build_oai --gNB --ninja -t oran_fhlib_5g --cmake-opt -Dxran_LOCATION=$HOME/phy/fhi_lib/lib --cmake-opt -DOAI_FHI72_USE_POLLING=ON
or with `cmake` like so
cmake .. -GNinja -DOAI_FHI72=ON -Dxran_LOCATION=$HOME/phy/fhi_lib/lib -DOAI_FHI72_USE_POLLING=ON
ninja oran_fhlib_5g
# Configuration
RU and DU configurations have a circular dependency: you have to configure DU MAC address in the RU configuration and the RU MAC address, VLAN and Timing advance parameters in the DU configuration.
@@ -1040,6 +1053,8 @@ In this case, you should reverify that `ptp4l` and `phc2sys` are working, e.g.,
do not do any jumps (during the last hour). While an occasional jump is not
necessarily problematic for the gNB, many such messages mean that the system is
not working, and UEs might not be able to attach or reach good performance.
Also, you can try to compile with polling (see [the build
section](.#build-oai-gnb)) to see if it resolves the problem.
# Operation with multiple RUs
executables/nr-gnb.c
-2
View File
@@ -219,7 +219,6 @@ static void rx_func(processingData_L1_t *info)
gNB->if_inst->NR_UL_indication(&UL_INFO);
stop_meas(&gNB->ul_indication_stats);
#ifndef OAI_FHI72
notifiedFIFO_elt_t *res = newNotifiedFIFO_elt(sizeof(processingData_L1_t), 0, &gNB->L1_rx_out, NULL);
processingData_L1_t *syncMsg = NotifiedFifoData(res);
syncMsg->gNB = gNB;
@@ -228,7 +227,6 @@ static void rx_func(processingData_L1_t *info)
res->key = slot_rx;
LOG_D(NR_PHY, "Signaling completion for %d.%d (mod_slot %d) on L1_rx_out\n", frame_rx, slot_rx, slot_rx % RU_RX_SLOT_DEPTH);
pushNotifiedFIFO(&gNB->L1_rx_out, res);
#endif
}
}
executables/nr-ru.c
+44 -32
View File
@@ -1084,6 +1084,48 @@ void ru_tx_func(void *param)
}//else emulate_rf
}
/* @brief wait for the next RX TTI to be free
*
* Certain radios, e.g., RFsim, can run faster than real-time. This might
* create problems, e.g., if RX and TX get too far from each other. This
* function ensures that a maximum of 4 RX slots are processed at a time (and
* not more than those four are started).
*
* Through the queue L1_rx_out, we are informed about completed RX jobs.
* rx_tti_busy keeps track of individual slots that have been started; this
* function blocks until the current frame/slot is completed, signaled through
* a message.
*
* @param L1_rx_out the queue from which to read completed RX jobs
* @param rx_tti_busy array to mark RX job completion
* @param frame_rx the frame to wait for
* @param slot_rx the slot to wait for
*/
static bool wait_free_rx_tti(notifiedFIFO_t *L1_rx_out, bool rx_tti_busy[RU_RX_SLOT_DEPTH], int frame_rx, int slot_rx)
{
int idx = slot_rx % RU_RX_SLOT_DEPTH;
if (rx_tti_busy[idx]) {
bool not_done = true;
LOG_D(NR_PHY, "%d.%d Waiting to access RX slot %d\n", frame_rx, slot_rx, idx);
// block and wait for frame_rx/slot_rx free from previous slot processing.
// as we can get other slots, we loop on the queue
while (not_done) {
notifiedFIFO_elt_t *res = pullNotifiedFIFO(L1_rx_out);
if (!res)
return false;
processingData_L1_t *info = NotifiedFifoData(res);
LOG_D(NR_PHY, "%d.%d Got access to RX slot %d.%d (%d)\n", frame_rx, slot_rx, info->frame_rx, info->slot_rx, idx);
rx_tti_busy[info->slot_rx % RU_RX_SLOT_DEPTH] = false;
if ((info->slot_rx % RU_RX_SLOT_DEPTH) == idx)
not_done = false;
delNotifiedFIFO_elt(res);
}
}
// set the tti to busy: the caller will process this slot now
rx_tti_busy[idx] = true;
return true;
}
void *ru_thread(void *param)
{
static int ru_thread_status;
@@ -1097,9 +1139,7 @@ void *ru_thread(void *param)
char threadname[40];
int initial_wait = 0;
#ifndef OAI_FHI72
bool rx_tti_busy[RU_RX_SLOT_DEPTH] = {false};
#endif
// set default return value
ru_thread_status = 0;
// set default return value
@@ -1245,39 +1285,12 @@ void *ru_thread(void *param)
if (ru->idx != 0)
proc->frame_tx = (proc->frame_tx + proc->frame_offset) & 1023;
#ifndef OAI_FHI72
// do RX front-end processing (frequency-shift, dft) if needed
int slot_type = nr_slot_select(&ru->config, proc->frame_rx, proc->tti_rx);
if (slot_type == NR_UPLINK_SLOT || slot_type == NR_MIXED_SLOT) {
if (!wait_free_rx_tti(&gNB->L1_rx_out, rx_tti_busy, proc->frame_rx, proc->tti_rx))
break; // nothing to wait for: we have to stop
if (ru->feprx) {
if (rx_tti_busy[proc->tti_rx % RU_RX_SLOT_DEPTH]) {
bool not_done = true;
LOG_D(NR_PHY, "%d.%d Waiting to access RX slot %d\n", proc->frame_rx, proc->tti_rx, proc->tti_rx % RU_RX_SLOT_DEPTH);
// now we block and wait our slot memory zone is freed from previous slot processing
// as we can get other slots ending, we loop on the queue
notifiedFIFO_elt_t *res = NULL;
while (not_done) {
res = pullNotifiedFIFO(&gNB->L1_rx_out);
if (!res)
break;
processingData_L1_t *info = (processingData_L1_t *)NotifiedFifoData(res);
LOG_D(NR_PHY,
"%d.%d Got access to RX slot %d.%d (%d)\n",
proc->frame_rx,
proc->tti_rx,
info->frame_rx,
info->slot_rx,
proc->tti_rx % RU_RX_SLOT_DEPTH);
rx_tti_busy[info->slot_rx % RU_RX_SLOT_DEPTH] = false;
if ((info->slot_rx % RU_RX_SLOT_DEPTH) == (proc->tti_rx % RU_RX_SLOT_DEPTH))
not_done = false;
delNotifiedFIFO_elt(res);
}
if (!res)
break;
}
// set the tti that was generated to busy
rx_tti_busy[proc->tti_rx % RU_RX_SLOT_DEPTH] = true;
ru->feprx(ru,proc->tti_rx);
LOG_D(NR_PHY, "Setting %d.%d (%d) to busy\n", proc->frame_rx, proc->tti_rx, proc->tti_rx % RU_RX_SLOT_DEPTH);
//LOG_M("rxdata.m","rxs",ru->common.rxdata[0],1228800,1,1);
@@ -1315,7 +1328,6 @@ void *ru_thread(void *param)
} // end if (prach_id > 0)
} // end if (ru->feprx)
} // end if (slot_type == NR_UPLINK_SLOT || slot_type == NR_MIXED_SLOT) {
#endif
notifiedFIFO_elt_t *resTx = newNotifiedFIFO_elt(sizeof(processingData_L1tx_t), 0, &gNB->L1_tx_out, NULL);
processingData_L1tx_t *syncMsgTx = NotifiedFifoData(resTx);
radio/fhi_72/CMakeLists.txt
+2 -3
View File
@@ -39,10 +39,9 @@ target_link_libraries(oran_fhlib_5g PRIVATE ${T_LIB})
# TODO: can be removed?
target_include_directories(oran_fhlib_5g PRIVATE ${xran_INCLUDE_DIRS}/../src/)
add_boolean_option(OAI_FHI72_USE_POLLING OFF "Enable polling in FHI72 driver" ON)
set_target_properties(oran_fhlib_5g PROPERTIES LIBRARY_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR})
add_custom_command(TARGET oran_fhlib_5g POST_BUILD
COMMAND ${CMAKE_COMMAND} -E create_symlink liboran_fhlib_5g.so liboai_transpro.so
WORKING_DIRECTORY ${CMAKE_BINARY_DIR})
message(STATUS "Temporary added -DOAI_FHI72")
target_compile_definitions(nr-softmodem PUBLIC OAI_FHI72)
radio/fhi_72/oaioran.c
+51 -66
View File
@@ -37,26 +37,14 @@
#include "oran-config.h" // for g_kbar
#define USE_POLLING 1
#include "common/utils/threadPool/notified_fifo.h"
#if OAI_FHI72_USE_POLLING
#define USE_POLLING
#endif
// Declare variable useful for the send buffer function
volatile uint8_t first_call_set = 0;
volatile uint8_t first_rx_set = 0;
volatile int first_read_set = 0;
// Variable declaration useful for fill IQ samples from file
#define IQ_PLAYBACK_BUFFER_BYTES (XRAN_NUM_OF_SLOT_IN_TDD_LOOP * N_SYM_PER_SLOT * XRAN_MAX_PRBS * N_SC_PER_PRB * 4L)
/*
int rx_tti;
int rx_sym;
volatile uint32_t rx_cb_tti = 0;
volatile uint32_t rx_cb_frame = 0;
volatile uint32_t rx_cb_subframe = 0;
volatile uint32_t rx_cb_slot = 0;
*/
#define GetFrameNum(tti, SFNatSecStart, numSubFramePerSystemFrame, numSlotPerSubFrame) \
((((uint32_t)tti / ((uint32_t)numSubFramePerSystemFrame * (uint32_t)numSlotPerSubFrame)) + SFNatSecStart) & 0x3FF)
#define GetSlotNum(tti, numSlotPerSfn) ((uint32_t)tti % ((uint32_t)numSlotPerSfn))
int xran_is_prach_slot(uint8_t PortId, uint32_t subframe_id, uint32_t slot_id);
#include "common/utils/LOG/log.h"
@@ -64,7 +52,7 @@ int xran_is_prach_slot(uint8_t PortId, uint32_t subframe_id, uint32_t slot_id);
#ifndef USE_POLLING
extern notifiedFIFO_t oran_sync_fifo;
#else
volatile oran_sync_info_t oran_sync_info;
volatile oran_sync_info_t oran_sync_info = {0};
#endif
void oai_xran_fh_rx_callback(void *pCallbackTag, xran_status_t status)
{
@@ -124,47 +112,42 @@ void oai_xran_fh_rx_callback(void *pCallbackTag, xran_status_t status)
}
}
#endif
if (first_call_set) {
if (!first_rx_set) {
LOG_I(NR_PHY, "first_rx is set (num_ports %d), first_read_set %d\n", num_ports, first_read_set);
}
first_rx_set = 1;
if (first_read_set == 1) {
slot2 = slot + (subframe * slots_per_subframe);
rx_RU[ru_id][slot2] = 1;
if (last_frame > 0 && frame > 0
&& ((slot2 > 0 && last_frame != frame) || (slot2 == 0 && last_frame != ((1024 + frame - 1) & 1023))))
LOG_E(PHY, "Jump in frame counter last_frame %d => %d, slot %d\n", last_frame, frame, slot2);
for (int i = 0; i < num_ports; i++) {
if (rx_RU[i][slot2] == 0)
return;
}
for (int i = 0; i < num_ports; i++)
rx_RU[i][slot2] = 0;
// if xran did not call xran_physide_dl_tti callback, it's not ready yet.
// wait till first callback to advance counters, because otherwise users
// would see periodic output with only "0" in stats counters
if (!first_call_set)
return;
slot2 = slot + (subframe * slots_per_subframe);
rx_RU[ru_id][slot2] = 1;
if (last_frame > 0 && frame > 0
&& ((slot2 > 0 && last_frame != frame) || (slot2 == 0 && last_frame != ((1024 + frame - 1) & 1023))))
LOG_E(PHY, "Jump in frame counter last_frame %d => %d, slot %d\n", last_frame, frame, slot2);
for (int i = 0; i < num_ports; i++) {
if (rx_RU[i][slot2] == 0)
return;
}
for (int i = 0; i < num_ports; i++)
rx_RU[i][slot2] = 0;
if (last_slot == -1 || slot2 != last_slot) {
if (last_slot == -1 || slot2 != last_slot) {
#ifndef USE_POLLING
notifiedFIFO_elt_t *req = newNotifiedFIFO_elt(sizeof(oran_sync_info_t), 0, &oran_sync_fifo, NULL);
oran_sync_info_t *info = (oran_sync_info_t *)NotifiedFifoData(req);
info->sl = slot2;
info->f = frame;
LOG_D(PHY, "Push %d.%d.%d (slot %d, subframe %d,last_slot %d)\n", frame, info->sl, slot, ru_id, subframe, last_slot);
notifiedFIFO_elt_t *req = newNotifiedFIFO_elt(sizeof(oran_sync_info_t), 0, &oran_sync_fifo, NULL);
oran_sync_info_t *info = NotifiedFifoData(req);
info->tti = tti;
info->sl = slot2;
info->f = frame;
LOG_D(PHY, "Push %d.%d.%d (slot %d, subframe %d,last_slot %d)\n", frame, info->sl, slot, ru_id, subframe, last_slot);
pushNotifiedFIFO(&oran_sync_fifo, req);
#else
LOG_D(PHY, "Writing %d.%d.%d (slot %d, subframe %d,last_slot %d)\n", frame, slot2, ru_id, slot, subframe, last_slot);
oran_sync_info.tti = tti;
oran_sync_info.sl = slot2;
oran_sync_info.f = frame;
LOG_D(PHY, "Writing %d.%d.%d (slot %d, subframe %d,last_slot %d)\n", frame, slot2, ru_id, slot, subframe, last_slot);
oran_sync_info.tti = tti;
oran_sync_info.sl = slot2;
oran_sync_info.f = frame;
#endif
#ifndef USE_POLLING
pushNotifiedFIFO(&oran_sync_fifo, req);
#else
#endif
} else
LOG_E(PHY, "Cannot Push %d.%d.%d (slot %d, subframe %d,last_slot %d)\n", frame, slot2, ru_id, slot, subframe, last_slot);
last_slot = slot2;
last_frame = frame;
} // first_read_set == 1
} // first_call_set
} else
LOG_E(PHY, "Cannot Push %d.%d.%d (slot %d, subframe %d,last_slot %d)\n", frame, slot2, ru_id, slot, subframe, last_slot);
last_slot = slot2;
last_frame = frame;
} // rx_sym == 7
}
void oai_xran_fh_srs_callback(void *pCallbackTag, xran_status_t status)
@@ -286,8 +269,6 @@ int xran_fh_rx_read_slot(ru_info_t *ru, int *frame, int *slot)
void *ptr = NULL;
int32_t *pos = NULL;
int idx = 0;
static int last_slot = -1;
first_read_set = 1;
static int64_t old_rx_counter[XRAN_PORTS_NUM] = {0};
static int64_t old_tx_counter[XRAN_PORTS_NUM] = {0};
@@ -295,24 +276,28 @@ int xran_fh_rx_read_slot(ru_info_t *ru, int *frame, int *slot)
static int outcnt = 0;
#ifndef USE_POLLING
// pull next even from oran_sync_fifo
notifiedFIFO_elt_t *res = pollNotifiedFIFO(&oran_sync_fifo);
while (res == NULL) {
res = pollNotifiedFIFO(&oran_sync_fifo);
notifiedFIFO_elt_t *res = pullNotifiedFIFO(&oran_sync_fifo);
notifiedFIFO_elt_t *f;
while ((f = pollNotifiedFIFO(&oran_sync_fifo)) != NULL) {
oran_sync_info_t *old_info = NotifiedFifoData(res);
oran_sync_info_t *new_info = NotifiedFifoData(f);
LOG_E(PHY, "Detected double sync message %d.%d => %d.%d\n", old_info->f, old_info->sl, new_info->f, new_info->sl);
delNotifiedFIFO_elt(res);
res = f;
}
oran_sync_info_t *info = (oran_sync_info_t *)NotifiedFifoData(res);
oran_sync_info_t *info = NotifiedFifoData(res);
*slot = info->sl;
*frame = info->f;
delNotifiedFIFO_elt(res);
#else
LOG_D(PHY, "In xran_fh_rx_read_slot, first_rx_set %d\n", first_rx_set);
while (first_rx_set == 0) {
}
*slot = oran_sync_info.sl;
*frame = oran_sync_info.f;
uint32_t tti_in = oran_sync_info.tti;
static int last_slot = -1;
LOG_D(PHY, "oran slot %d, last_slot %d\n", *slot, last_slot);
int cnt = 0;
// while (*slot == last_slot) {