vfio: fix DMA mapping with default container
[dpdk.git] / lib / librte_distributor / rte_distributor.c
index d505983..7df97df 100644 (file)
@@ -8,17 +8,18 @@
 #include <rte_mbuf.h>
 #include <rte_memory.h>
 #include <rte_cycles.h>
-#include <rte_compat.h>
+#include <rte_function_versioning.h>
 #include <rte_memzone.h>
 #include <rte_errno.h>
 #include <rte_string_fns.h>
 #include <rte_eal_memconfig.h>
 #include <rte_pause.h>
+#include <rte_tailq.h>
 
-#include "rte_distributor_private.h"
 #include "rte_distributor.h"
 #include "rte_distributor_v20.h"
 #include "rte_distributor_v1705.h"
+#include "distributor_private.h"
 
 TAILQ_HEAD(rte_dist_burst_list, rte_distributor);
 
@@ -48,8 +49,11 @@ rte_distributor_request_pkt_v1705(struct rte_distributor *d,
        }
 
        retptr64 = &(buf->retptr64[0]);
-       /* Spin while handshake bits are set (scheduler clears it) */
-       while (unlikely(*retptr64 & RTE_DISTRIB_GET_BUF)) {
+       /* Spin while handshake bits are set (scheduler clears it).
+        * Sync with worker on GET_BUF flag.
+        */
+       while (unlikely(__atomic_load_n(retptr64, __ATOMIC_ACQUIRE)
+                       & RTE_DISTRIB_GET_BUF)) {
                rte_pause();
                uint64_t t = rte_rdtsc()+100;
 
@@ -74,8 +78,10 @@ rte_distributor_request_pkt_v1705(struct rte_distributor *d,
        /*
         * Finally, set the GET_BUF  to signal to distributor that cache
         * line is ready for processing
+        * Sync with distributor to release retptrs
         */
-       *retptr64 |= RTE_DISTRIB_GET_BUF;
+       __atomic_store_n(retptr64, *retptr64 | RTE_DISTRIB_GET_BUF,
+                       __ATOMIC_RELEASE);
 }
 BIND_DEFAULT_SYMBOL(rte_distributor_request_pkt, _v1705, 17.05);
 MAP_STATIC_SYMBOL(void rte_distributor_request_pkt(struct rte_distributor *d,
@@ -97,8 +103,11 @@ rte_distributor_poll_pkt_v1705(struct rte_distributor *d,
                return (pkts[0]) ? 1 : 0;
        }
 
-       /* If bit is set, return */
-       if (buf->bufptr64[0] & RTE_DISTRIB_GET_BUF)
+       /* If bit is set, return
+        * Sync with distributor to acquire bufptrs
+        */
+       if (__atomic_load_n(&(buf->bufptr64[0]), __ATOMIC_ACQUIRE)
+               & RTE_DISTRIB_GET_BUF)
                return -1;
 
        /* since bufptr64 is signed, this should be an arithmetic shift */
@@ -113,8 +122,10 @@ rte_distributor_poll_pkt_v1705(struct rte_distributor *d,
         * so now we've got the contents of the cacheline into an  array of
         * mbuf pointers, so toggle the bit so scheduler can start working
         * on the next cacheline while we're working.
+        * Sync with distributor on GET_BUF flag. Release bufptrs.
         */
-       buf->bufptr64[0] |= RTE_DISTRIB_GET_BUF;
+       __atomic_store_n(&(buf->bufptr64[0]),
+               buf->bufptr64[0] | RTE_DISTRIB_GET_BUF, __ATOMIC_RELEASE);
 
        return count;
 }
@@ -173,6 +184,8 @@ rte_distributor_return_pkt_v1705(struct rte_distributor *d,
                        return -EINVAL;
        }
 
+       /* Sync with distributor to acquire retptrs */
+       __atomic_thread_fence(__ATOMIC_ACQUIRE);
        for (i = 0; i < RTE_DIST_BURST_SIZE; i++)
                /* Switch off the return bit first */
                buf->retptr64[i] &= ~RTE_DISTRIB_RETURN_BUF;
@@ -181,8 +194,11 @@ rte_distributor_return_pkt_v1705(struct rte_distributor *d,
                buf->retptr64[i] = (((int64_t)(uintptr_t)oldpkt[i]) <<
                        RTE_DISTRIB_FLAG_BITS) | RTE_DISTRIB_RETURN_BUF;
 
-       /* set the GET_BUF but even if we got no returns */
-       buf->retptr64[0] |= RTE_DISTRIB_GET_BUF;
+       /* set the GET_BUF but even if we got no returns.
+        * Sync with distributor on GET_BUF flag. Release retptrs.
+        */
+       __atomic_store_n(&(buf->retptr64[0]),
+               buf->retptr64[0] | RTE_DISTRIB_GET_BUF, __ATOMIC_RELEASE);
 
        return 0;
 }
@@ -272,7 +288,9 @@ handle_returns(struct rte_distributor *d, unsigned int wkr)
        unsigned int count = 0;
        unsigned int i;
 
-       if (buf->retptr64[0] & RTE_DISTRIB_GET_BUF) {
+       /* Sync on GET_BUF flag. Acquire retptrs. */
+       if (__atomic_load_n(&(buf->retptr64[0]), __ATOMIC_ACQUIRE)
+               & RTE_DISTRIB_GET_BUF) {
                for (i = 0; i < RTE_DIST_BURST_SIZE; i++) {
                        if (buf->retptr64[i] & RTE_DISTRIB_RETURN_BUF) {
                                oldbuf = ((uintptr_t)(buf->retptr64[i] >>
@@ -285,8 +303,10 @@ handle_returns(struct rte_distributor *d, unsigned int wkr)
                }
                d->returns.start = ret_start;
                d->returns.count = ret_count;
-               /* Clear for the worker to populate with more returns */
-               buf->retptr64[0] = 0;
+               /* Clear for the worker to populate with more returns.
+                * Sync with distributor on GET_BUF flag. Release retptrs.
+                */
+               __atomic_store_n(&(buf->retptr64[0]), 0, __ATOMIC_RELEASE);
        }
        return count;
 }
@@ -306,7 +326,9 @@ release(struct rte_distributor *d, unsigned int wkr)
        struct rte_distributor_buffer *buf = &(d->bufs[wkr]);
        unsigned int i;
 
-       while (!(d->bufs[wkr].bufptr64[0] & RTE_DISTRIB_GET_BUF))
+       /* Sync with worker on GET_BUF flag */
+       while (!(__atomic_load_n(&(d->bufs[wkr].bufptr64[0]), __ATOMIC_ACQUIRE)
+               & RTE_DISTRIB_GET_BUF))
                rte_pause();
 
        handle_returns(d, wkr);
@@ -326,8 +348,11 @@ release(struct rte_distributor *d, unsigned int wkr)
 
        d->backlog[wkr].count = 0;
 
-       /* Clear the GET bit */
-       buf->bufptr64[0] &= ~RTE_DISTRIB_GET_BUF;
+       /* Clear the GET bit.
+        * Sync with worker on GET_BUF flag. Release bufptrs.
+        */
+       __atomic_store_n(&(buf->bufptr64[0]),
+               buf->bufptr64[0] & ~RTE_DISTRIB_GET_BUF, __ATOMIC_RELEASE);
        return  buf->count;
 
 }
@@ -354,7 +379,9 @@ rte_distributor_process_v1705(struct rte_distributor *d,
        if (unlikely(num_mbufs == 0)) {
                /* Flush out all non-full cache-lines to workers. */
                for (wid = 0 ; wid < d->num_workers; wid++) {
-                       if (d->bufs[wid].bufptr64[0] & RTE_DISTRIB_GET_BUF) {
+                       /* Sync with worker on GET_BUF flag. */
+                       if (__atomic_load_n(&(d->bufs[wid].bufptr64[0]),
+                               __ATOMIC_ACQUIRE) & RTE_DISTRIB_GET_BUF) {
                                release(d, wid);
                                handle_returns(d, wid);
                        }
@@ -366,7 +393,9 @@ rte_distributor_process_v1705(struct rte_distributor *d,
                uint16_t matches[RTE_DIST_BURST_SIZE];
                unsigned int pkts;
 
-               if (d->bufs[wkr].bufptr64[0] & RTE_DISTRIB_GET_BUF)
+               /* Sync with worker on GET_BUF flag. */
+               if (__atomic_load_n(&(d->bufs[wkr].bufptr64[0]),
+                       __ATOMIC_ACQUIRE) & RTE_DISTRIB_GET_BUF)
                        d->bufs[wkr].count = 0;
 
                if ((num_mbufs - next_idx) < RTE_DIST_BURST_SIZE)
@@ -464,7 +493,9 @@ rte_distributor_process_v1705(struct rte_distributor *d,
 
        /* Flush out all non-full cache-lines to workers. */
        for (wid = 0 ; wid < d->num_workers; wid++)
-               if ((d->bufs[wid].bufptr64[0] & RTE_DISTRIB_GET_BUF))
+               /* Sync with worker on GET_BUF flag. */
+               if ((__atomic_load_n(&(d->bufs[wid].bufptr64[0]),
+                       __ATOMIC_ACQUIRE) & RTE_DISTRIB_GET_BUF))
                        release(d, wid);
 
        return num_mbufs;
@@ -541,6 +572,9 @@ rte_distributor_flush_v1705(struct rte_distributor *d)
        while (total_outstanding(d) > 0)
                rte_distributor_process(d, NULL, 0);
 
+       /* wait 10ms to allow all worker drain the pkts */
+       rte_delay_us(10000);
+
        /*
         * Send empty burst to all workers to allow them to exit
         * gracefully, should they need to.
@@ -570,7 +604,9 @@ rte_distributor_clear_returns_v1705(struct rte_distributor *d)
 
        /* throw away returns, so workers can exit */
        for (wkr = 0; wkr < d->num_workers; wkr++)
-               d->bufs[wkr].retptr64[0] = 0;
+               /* Sync with worker. Release retptrs. */
+               __atomic_store_n(&(d->bufs[wkr].retptr64[0]), 0,
+                               __ATOMIC_RELEASE);
 }
 BIND_DEFAULT_SYMBOL(rte_distributor_clear_returns, _v1705, 17.05);
 MAP_STATIC_SYMBOL(void rte_distributor_clear_returns(struct rte_distributor *d),
@@ -595,6 +631,12 @@ rte_distributor_create_v1705(const char *name,
        RTE_BUILD_BUG_ON((sizeof(*d) & RTE_CACHE_LINE_MASK) != 0);
        RTE_BUILD_BUG_ON((RTE_DISTRIB_MAX_WORKERS & 7) != 0);
 
+       if (name == NULL || num_workers >=
+               (unsigned int)RTE_MIN(RTE_DISTRIB_MAX_WORKERS, RTE_MAX_LCORE)) {
+               rte_errno = EINVAL;
+               return NULL;
+       }
+
        if (alg_type == RTE_DIST_ALG_SINGLE) {
                d = malloc(sizeof(struct rte_distributor));
                if (d == NULL) {
@@ -612,11 +654,6 @@ rte_distributor_create_v1705(const char *name,
                return d;
        }
 
-       if (name == NULL || num_workers >= RTE_DISTRIB_MAX_WORKERS) {
-               rte_errno = EINVAL;
-               return NULL;
-       }
-
        snprintf(mz_name, sizeof(mz_name), RTE_DISTRIB_PREFIX"%s", name);
        mz = rte_memzone_reserve(mz_name, sizeof(*d), socket_id, NO_FLAGS);
        if (mz == NULL) {
@@ -625,7 +662,7 @@ rte_distributor_create_v1705(const char *name,
        }
 
        d = mz->addr;
-       snprintf(d->name, sizeof(d->name), "%s", name);
+       strlcpy(d->name, name, sizeof(d->name));
        d->num_workers = num_workers;
        d->alg_type = alg_type;
 
@@ -645,9 +682,9 @@ rte_distributor_create_v1705(const char *name,
                                          rte_dist_burst_list);
 
 
-       rte_rwlock_write_lock(RTE_EAL_TAILQ_RWLOCK);
+       rte_mcfg_tailq_write_lock();
        TAILQ_INSERT_TAIL(dist_burst_list, d, next);
-       rte_rwlock_write_unlock(RTE_EAL_TAILQ_RWLOCK);
+       rte_mcfg_tailq_write_unlock();
 
        return d;
 }