event/dlb2: optimize dequeue operation
authorTimothy McDaniel <timothy.mcdaniel@intel.com>
Sat, 1 May 2021 19:07:44 +0000 (14:07 -0500)
committerJerin Jacob <jerinj@marvell.com>
Mon, 3 May 2021 12:21:23 +0000 (14:21 +0200)
Convert code to use x86 vector instructions, thereby significantly
improving dequeue performance.

Signed-off-by: Timothy McDaniel <timothy.mcdaniel@intel.com>
Signed-off-by: Harry van Haaren <harry.van.haaren@intel.com>
drivers/event/dlb2/dlb2.c
drivers/event/dlb2/dlb2_priv.h

index 818b1c3..3570678 100644 (file)
@@ -375,6 +375,26 @@ set_default_depth_thresh(const char *key __rte_unused,
        return 0;
 }
 
+static int
+set_vector_opts_disab(const char *key __rte_unused,
+       const char *value,
+       void *opaque)
+{
+       bool *dlb2_vector_opts_disabled = opaque;
+
+       if (value == NULL || opaque == NULL) {
+               DLB2_LOG_ERR("NULL pointer\n");
+               return -EINVAL;
+       }
+
+       if ((*value == 'y') || (*value == 'Y'))
+               *dlb2_vector_opts_disabled = true;
+       else
+               *dlb2_vector_opts_disabled = false;
+
+       return 0;
+}
+
 static int
 set_qid_depth_thresh(const char *key __rte_unused,
                     const char *value,
@@ -1240,6 +1260,37 @@ dlb2_event_enqueue_forward_burst_delayed(void *event_port,
                                         const struct rte_event events[],
                                         uint16_t num);
 
+/* Generate the required bitmask for rotate-style expected QE gen bits.
+ * This requires a pattern of 1's and zeros, starting with expected as
+ * 1 bits, so when hardware writes 0's they're "new". This requires the
+ * ring size to be powers of 2 to wrap correctly.
+ */
+static void
+dlb2_hw_cq_bitmask_init(struct dlb2_port *qm_port, uint32_t cq_depth)
+{
+       uint64_t cq_build_mask = 0;
+       uint32_t i;
+
+       if (cq_depth > 64)
+               return; /* need to fall back to scalar code */
+
+       /*
+        * all 1's in first u64, all zeros in second is correct bit pattern to
+        * start. Special casing == 64 easier than adapting complex loop logic.
+        */
+       if (cq_depth == 64) {
+               qm_port->cq_rolling_mask = 0;
+               qm_port->cq_rolling_mask_2 = -1;
+               return;
+       }
+
+       for (i = 0; i < 64; i += (cq_depth * 2))
+               cq_build_mask |= ((1ULL << cq_depth) - 1) << (i + cq_depth);
+
+       qm_port->cq_rolling_mask = cq_build_mask;
+       qm_port->cq_rolling_mask_2 = cq_build_mask;
+}
+
 static int
 dlb2_hw_create_ldb_port(struct dlb2_eventdev *dlb2,
                        struct dlb2_eventdev_port *ev_port,
@@ -1357,6 +1408,8 @@ dlb2_hw_create_ldb_port(struct dlb2_eventdev *dlb2,
        /* starting value of gen bit - it toggles at wrap time */
        qm_port->gen_bit = 1;
 
+       dlb2_hw_cq_bitmask_init(qm_port, qm_port->cq_depth);
+
        qm_port->int_armed = false;
 
        /* Save off for later use in info and lookup APIs. */
@@ -1408,6 +1461,18 @@ dlb2_hw_create_ldb_port(struct dlb2_eventdev *dlb2,
                             dequeue_depth,
                             qm_port->credits);
        }
+
+       qm_port->use_scalar = false;
+
+#if (!defined RTE_ARCH_X86_64)
+       qm_port->use_scalar = true;
+#else
+       if ((qm_port->cq_depth > 64) ||
+           (!rte_is_power_of_2(qm_port->cq_depth)) ||
+           (dlb2->vector_opts_disabled == true))
+               qm_port->use_scalar = true;
+#endif
+
        rte_spinlock_unlock(&handle->resource_lock);
 
        return 0;
@@ -1553,6 +1618,7 @@ dlb2_hw_create_dir_port(struct dlb2_eventdev *dlb2,
        qm_port->gen_bit_shift = __builtin_popcount(qm_port->cq_depth_mask);
        /* starting value of gen bit - it toggles at wrap time */
        qm_port->gen_bit = 1;
+       dlb2_hw_cq_bitmask_init(qm_port, qm_port->cq_depth);
 
        qm_port->int_armed = false;
 
@@ -1593,6 +1659,16 @@ dlb2_hw_create_dir_port(struct dlb2_eventdev *dlb2,
                             dequeue_depth,
                             credit_high_watermark);
        }
+
+#if (!defined RTE_ARCH_X86_64)
+       qm_port->use_scalar = true;
+#else
+       if ((qm_port->cq_depth > 64) ||
+           (!rte_is_power_of_2(qm_port->cq_depth)) ||
+           (dlb2->vector_opts_disabled == true))
+               qm_port->use_scalar = true;
+#endif
+
        rte_spinlock_unlock(&handle->resource_lock);
 
        return 0;
@@ -2987,10 +3063,11 @@ dlb2_event_release(struct dlb2_eventdev *dlb2,
                int j = 0;
 
                /* Zero-out QEs */
-               qm_port->qe4[0].cmd_byte = 0;
-               qm_port->qe4[1].cmd_byte = 0;
-               qm_port->qe4[2].cmd_byte = 0;
-               qm_port->qe4[3].cmd_byte = 0;
+               _mm_storeu_si128((void *)&qm_port->qe4[0], _mm_setzero_si128());
+               _mm_storeu_si128((void *)&qm_port->qe4[1], _mm_setzero_si128());
+               _mm_storeu_si128((void *)&qm_port->qe4[2], _mm_setzero_si128());
+               _mm_storeu_si128((void *)&qm_port->qe4[3], _mm_setzero_si128());
+
 
                for (; j < DLB2_NUM_QES_PER_CACHE_LINE && (i + j) < n; j++) {
                        int16_t thresh = qm_port->token_pop_thresh;
@@ -3020,7 +3097,7 @@ dlb2_event_release(struct dlb2_eventdev *dlb2,
 
 sw_credit_update:
        /* each release returns one credit */
-       if (!ev_port->outstanding_releases) {
+       if (unlikely(!ev_port->outstanding_releases)) {
                DLB2_LOG_ERR("%s: Outstanding releases underflowed.\n",
                             __func__);
                return;
@@ -3137,7 +3214,7 @@ dlb2_dequeue_wait(struct dlb2_eventdev *dlb2,
        return 0;
 }
 
-static inline int
+static __rte_noinline int
 dlb2_process_dequeue_qes(struct dlb2_eventdev_port *ev_port,
                         struct dlb2_port *qm_port,
                         struct rte_event *events,
@@ -3406,8 +3483,7 @@ dlb2_recv_qe_sparse(struct dlb2_port *qm_port, struct dlb2_dequeue_qe *qe)
 
        cq_addr = dlb2_port[qm_port->id][PORT_TYPE(qm_port)].cq_base;
 
-       idx = qm_port->cq_idx;
-
+       idx = qm_port->cq_idx_unmasked & qm_port->cq_depth_mask;
        /* Load the next 4 QEs */
        addr[0] = (uintptr_t)&cq_addr[idx];
        addr[1] = (uintptr_t)&cq_addr[(idx +  4) & qm_port->cq_depth_mask];
@@ -3452,6 +3528,272 @@ dlb2_recv_qe_sparse(struct dlb2_port *qm_port, struct dlb2_dequeue_qe *qe)
        return __builtin_popcount(gen_bits);
 }
 
+static inline void
+_process_deq_qes_vec_impl(struct dlb2_port *qm_port,
+                         struct rte_event *events,
+                         __m128i v_qe_3,
+                         __m128i v_qe_2,
+                         __m128i v_qe_1,
+                         __m128i v_qe_0,
+                         __m128i v_qe_meta,
+                         __m128i v_qe_status,
+                         uint32_t valid_events)
+{
+       /* Look up the event QIDs, using the hardware QIDs to index the
+        * port's QID mapping.
+        *
+        * Each v_qe_[0-4] is just a 16-byte load of the whole QE. It is
+        * passed along in registers as the QE data is required later.
+        *
+        * v_qe_meta is an u32 unpack of all 4x QEs. A.k.a, it contains one
+        * 32-bit slice of each QE, so makes up a full SSE register. This
+        * allows parallel processing of 4x QEs in a single register.
+        */
+
+       __m128i v_qid_done = {0};
+       int hw_qid0 = _mm_extract_epi8(v_qe_meta, 2);
+       int hw_qid1 = _mm_extract_epi8(v_qe_meta, 6);
+       int hw_qid2 = _mm_extract_epi8(v_qe_meta, 10);
+       int hw_qid3 = _mm_extract_epi8(v_qe_meta, 14);
+
+       int ev_qid0 = qm_port->qid_mappings[hw_qid0];
+       int ev_qid1 = qm_port->qid_mappings[hw_qid1];
+       int ev_qid2 = qm_port->qid_mappings[hw_qid2];
+       int ev_qid3 = qm_port->qid_mappings[hw_qid3];
+
+       v_qid_done = _mm_insert_epi8(v_qid_done, ev_qid0, 2);
+       v_qid_done = _mm_insert_epi8(v_qid_done, ev_qid1, 6);
+       v_qid_done = _mm_insert_epi8(v_qid_done, ev_qid2, 10);
+       v_qid_done = _mm_insert_epi8(v_qid_done, ev_qid3, 14);
+
+       /* Schedule field remapping using byte shuffle
+        * - Full byte containing sched field handled here (op, rsvd are zero)
+        * - Note sanitizing the register requires two masking ANDs:
+        *   1) to strip prio/msg_type from byte for correct shuffle lookup
+        *   2) to strip any non-sched-field lanes from any results to OR later
+        * - Final byte result is >> 10 to another byte-lane inside the u32.
+        *   This makes the final combination OR easier to make the rte_event.
+        */
+       __m128i v_sched_done;
+       __m128i v_sched_bits;
+       {
+               static const uint8_t sched_type_map[16] = {
+                       [DLB2_SCHED_ATOMIC] = RTE_SCHED_TYPE_ATOMIC,
+                       [DLB2_SCHED_UNORDERED] = RTE_SCHED_TYPE_PARALLEL,
+                       [DLB2_SCHED_ORDERED] = RTE_SCHED_TYPE_ORDERED,
+                       [DLB2_SCHED_DIRECTED] = RTE_SCHED_TYPE_ATOMIC,
+               };
+               static const uint8_t sched_and_mask[16] = {
+                       0x00, 0x00, 0x00, 0x03,
+                       0x00, 0x00, 0x00, 0x03,
+                       0x00, 0x00, 0x00, 0x03,
+                       0x00, 0x00, 0x00, 0x03,
+               };
+               const __m128i v_sched_map = _mm_loadu_si128(
+                                            (const __m128i *)sched_type_map);
+               __m128i v_sched_mask = _mm_loadu_si128(
+                                            (const __m128i *)&sched_and_mask);
+               v_sched_bits = _mm_and_si128(v_qe_meta, v_sched_mask);
+               __m128i v_sched_remapped = _mm_shuffle_epi8(v_sched_map,
+                                                           v_sched_bits);
+               __m128i v_preshift = _mm_and_si128(v_sched_remapped,
+                                                  v_sched_mask);
+               v_sched_done = _mm_srli_epi32(v_preshift, 10);
+       }
+
+       /* Priority handling
+        * - QE provides 3 bits of priority
+        * - Shift << 3 to move to MSBs for byte-prio in rte_event
+        * - Mask bits to avoid pollution, leaving only 3 prio MSBs in reg
+        */
+       __m128i v_prio_done;
+       {
+               static const uint8_t prio_mask[16] = {
+                       0x00, 0x00, 0x00, 0x07 << 5,
+                       0x00, 0x00, 0x00, 0x07 << 5,
+                       0x00, 0x00, 0x00, 0x07 << 5,
+                       0x00, 0x00, 0x00, 0x07 << 5,
+               };
+               __m128i v_prio_mask  = _mm_loadu_si128(
+                                               (const __m128i *)prio_mask);
+               __m128i v_prio_shifted = _mm_slli_epi32(v_qe_meta, 3);
+               v_prio_done = _mm_and_si128(v_prio_shifted, v_prio_mask);
+       }
+
+       /* Event Sub/Type handling:
+        * we want to keep the lower 12 bits of each QE. Shift up by 20 bits
+        * to get the sub/ev type data into rte_event location, clearing the
+        * lower 20 bits in the process.
+        */
+       __m128i v_types_done;
+       {
+               static const uint8_t event_mask[16] = {
+                       0x0f, 0x00, 0x00, 0x00,
+                       0x0f, 0x00, 0x00, 0x00,
+                       0x0f, 0x00, 0x00, 0x00,
+                       0x0f, 0x00, 0x00, 0x00,
+               };
+               static const uint8_t sub_event_mask[16] = {
+                       0xff, 0x00, 0x00, 0x00,
+                       0xff, 0x00, 0x00, 0x00,
+                       0xff, 0x00, 0x00, 0x00,
+                       0xff, 0x00, 0x00, 0x00,
+               };
+               static const uint8_t flow_mask[16] = {
+                       0xff, 0xff, 0x00, 0x00,
+                       0xff, 0xff, 0x00, 0x00,
+                       0xff, 0xff, 0x00, 0x00,
+                       0xff, 0xff, 0x00, 0x00,
+               };
+               __m128i v_event_mask  = _mm_loadu_si128(
+                                       (const __m128i *)event_mask);
+               __m128i v_sub_event_mask  = _mm_loadu_si128(
+                                       (const __m128i *)sub_event_mask);
+               __m128i v_flow_mask  = _mm_loadu_si128(
+                                      (const __m128i *)flow_mask);
+               __m128i v_sub = _mm_srli_epi32(v_qe_meta, 8);
+               v_sub = _mm_and_si128(v_sub, v_sub_event_mask);
+               __m128i v_type = _mm_and_si128(v_qe_meta, v_event_mask);
+               v_type = _mm_slli_epi32(v_type, 8);
+               v_types_done = _mm_or_si128(v_type, v_sub);
+               v_types_done = _mm_slli_epi32(v_types_done, 20);
+               __m128i v_flow = _mm_and_si128(v_qe_status, v_flow_mask);
+               v_types_done = _mm_or_si128(v_types_done, v_flow);
+       }
+
+       /* Combine QID, Sched and Prio fields, then Shift >> 8 bits to align
+        * with the rte_event, allowing unpacks to move/blend with payload.
+        */
+       __m128i v_q_s_p_done;
+       {
+               __m128i v_qid_sched = _mm_or_si128(v_qid_done, v_sched_done);
+               __m128i v_q_s_prio = _mm_or_si128(v_qid_sched, v_prio_done);
+               v_q_s_p_done = _mm_srli_epi32(v_q_s_prio, 8);
+       }
+
+       __m128i v_unpk_ev_23, v_unpk_ev_01, v_ev_2, v_ev_3, v_ev_0, v_ev_1;
+
+       /* Unpack evs into u64 metadata, then indiv events */
+       v_unpk_ev_23 = _mm_unpackhi_epi32(v_types_done, v_q_s_p_done);
+       v_unpk_ev_01 = _mm_unpacklo_epi32(v_types_done, v_q_s_p_done);
+
+       switch (valid_events) {
+       case 4:
+               v_ev_3 = _mm_blend_epi16(v_unpk_ev_23, v_qe_3, 0x0F);
+               v_ev_3 = _mm_alignr_epi8(v_ev_3, v_ev_3, 8);
+               _mm_storeu_si128((__m128i *)&events[3], v_ev_3);
+               /* fallthrough */
+       case 3:
+               v_ev_2 = _mm_unpacklo_epi64(v_unpk_ev_23, v_qe_2);
+               _mm_storeu_si128((__m128i *)&events[2], v_ev_2);
+               /* fallthrough */
+       case 2:
+               v_ev_1 = _mm_blend_epi16(v_unpk_ev_01, v_qe_1, 0x0F);
+               v_ev_1 = _mm_alignr_epi8(v_ev_1, v_ev_1, 8);
+               _mm_storeu_si128((__m128i *)&events[1], v_ev_1);
+               /* fallthrough */
+       case 1:
+               v_ev_0 = _mm_unpacklo_epi64(v_unpk_ev_01, v_qe_0);
+               _mm_storeu_si128((__m128i *)&events[0], v_ev_0);
+       }
+}
+
+static __rte_always_inline int
+dlb2_recv_qe_sparse_vec(struct dlb2_port *qm_port, void *events,
+                       uint32_t max_events)
+{
+       /* Using unmasked idx for perf, and masking manually */
+       uint16_t idx = qm_port->cq_idx_unmasked;
+       volatile struct dlb2_dequeue_qe *cq_addr;
+
+       cq_addr = dlb2_port[qm_port->id][PORT_TYPE(qm_port)].cq_base;
+
+       uintptr_t qe_ptr_3 = (uintptr_t)&cq_addr[(idx + 12) &
+                                                qm_port->cq_depth_mask];
+       uintptr_t qe_ptr_2 = (uintptr_t)&cq_addr[(idx +  8) &
+                                                qm_port->cq_depth_mask];
+       uintptr_t qe_ptr_1 = (uintptr_t)&cq_addr[(idx +  4) &
+                                                qm_port->cq_depth_mask];
+       uintptr_t qe_ptr_0 = (uintptr_t)&cq_addr[(idx +  0) &
+                                                qm_port->cq_depth_mask];
+
+       /* Load QEs from CQ: use compiler barriers to avoid load reordering */
+       __m128i v_qe_3 = _mm_loadu_si128((const __m128i *)qe_ptr_3);
+       rte_compiler_barrier();
+       __m128i v_qe_2 = _mm_loadu_si128((const __m128i *)qe_ptr_2);
+       rte_compiler_barrier();
+       __m128i v_qe_1 = _mm_loadu_si128((const __m128i *)qe_ptr_1);
+       rte_compiler_barrier();
+       __m128i v_qe_0 = _mm_loadu_si128((const __m128i *)qe_ptr_0);
+
+       /* Generate the pkt_shuffle mask;
+        * - Avoids load in otherwise load-heavy section of code
+        * - Moves bytes 3,7,11,15 (gen bit bytes) to LSB bytes in XMM
+        */
+       const uint32_t stat_shuf_bytes = (15 << 24) | (11 << 16) | (7 << 8) | 3;
+       __m128i v_zeros = _mm_setzero_si128();
+       __m128i v_ffff = _mm_cmpeq_epi8(v_zeros, v_zeros);
+       __m128i v_stat_shuf_mask = _mm_insert_epi32(v_ffff, stat_shuf_bytes, 0);
+
+       /* Extract u32 components required from the QE
+        * - QE[64 to 95 ] for metadata (qid, sched, prio, event type, ...)
+        * - QE[96 to 127] for status (cq gen bit, error)
+        *
+        * Note that stage 1 of the unpacking is re-used for both u32 extracts
+        */
+       __m128i v_qe_02 = _mm_unpackhi_epi32(v_qe_0, v_qe_2);
+       __m128i v_qe_13 = _mm_unpackhi_epi32(v_qe_1, v_qe_3);
+       __m128i v_qe_status = _mm_unpackhi_epi32(v_qe_02, v_qe_13);
+       __m128i v_qe_meta   = _mm_unpacklo_epi32(v_qe_02, v_qe_13);
+
+       /* Status byte (gen_bit, error) handling:
+        * - Shuffle to lanes 0,1,2,3, clear all others
+        * - Shift right by 7 for gen bit to MSB, movemask to scalar
+        * - Shift right by 2 for error bit to MSB, movemask to scalar
+        */
+       __m128i v_qe_shuffled = _mm_shuffle_epi8(v_qe_status, v_stat_shuf_mask);
+       __m128i v_qes_shift_gen_bit = _mm_slli_epi32(v_qe_shuffled, 7);
+       int32_t qe_gen_bits = _mm_movemask_epi8(v_qes_shift_gen_bit) & 0xf;
+
+       /* Expected vs Reality of QE Gen bits
+        * - cq_rolling_mask provides expected bits
+        * - QE loads, unpacks/shuffle and movemask provides reality
+        * - XOR of the two gives bitmask of new packets
+        * - POPCNT to get the number of new events
+        */
+       uint64_t rolling = qm_port->cq_rolling_mask & 0xF;
+       uint64_t qe_xor_bits = (qe_gen_bits ^ rolling);
+       uint32_t count_new = __builtin_popcount(qe_xor_bits);
+       count_new = RTE_MIN(count_new, max_events);
+       if (!count_new)
+               return 0;
+
+       /* emulate a 128 bit rotate using 2x 64-bit numbers and bit-shifts */
+
+       uint64_t m_rshift = qm_port->cq_rolling_mask >> count_new;
+       uint64_t m_lshift = qm_port->cq_rolling_mask << (64 - count_new);
+       uint64_t m2_rshift = qm_port->cq_rolling_mask_2 >> count_new;
+       uint64_t m2_lshift = qm_port->cq_rolling_mask_2 << (64 - count_new);
+
+       /* shifted out of m2 into MSB of m */
+       qm_port->cq_rolling_mask = (m_rshift | m2_lshift);
+
+       /* shifted out of m "looped back" into MSB of m2 */
+       qm_port->cq_rolling_mask_2 = (m2_rshift | m_lshift);
+
+       /* Prefetch the next QEs - should run as IPC instead of cycles */
+       rte_prefetch0(&cq_addr[(idx + 16) & qm_port->cq_depth_mask]);
+       rte_prefetch0(&cq_addr[(idx + 20) & qm_port->cq_depth_mask]);
+       rte_prefetch0(&cq_addr[(idx + 24) & qm_port->cq_depth_mask]);
+       rte_prefetch0(&cq_addr[(idx + 28) & qm_port->cq_depth_mask]);
+
+       /* Convert QEs from XMM regs to events and store events directly */
+       _process_deq_qes_vec_impl(qm_port, events, v_qe_3, v_qe_2, v_qe_1,
+                                 v_qe_0, v_qe_meta, v_qe_status, count_new);
+
+       return count_new;
+}
+
 static inline void
 dlb2_inc_cq_idx(struct dlb2_port *qm_port, int cnt)
 {
@@ -3469,25 +3811,15 @@ dlb2_hw_dequeue_sparse(struct dlb2_eventdev *dlb2,
                       uint16_t max_num,
                       uint64_t dequeue_timeout_ticks)
 {
-       uint64_t timeout;
        uint64_t start_ticks = 0ULL;
        struct dlb2_port *qm_port;
        int num = 0;
+       bool use_scalar;
+       uint64_t timeout;
 
        qm_port = &ev_port->qm_port;
+       use_scalar = qm_port->use_scalar;
 
-       /* We have a special implementation for waiting. Wait can be:
-        * 1) no waiting at all
-        * 2) busy poll only
-        * 3) wait for interrupt. If wakeup and poll time
-        * has expired, then return to caller
-        * 4) umonitor/umwait repeatedly up to poll time
-        */
-
-       /* If configured for per dequeue wait, then use wait value provided
-        * to this API. Otherwise we must use the global
-        * value from eventdev config time.
-        */
        if (!dlb2->global_dequeue_wait)
                timeout = dequeue_timeout_ticks;
        else
@@ -3495,35 +3827,41 @@ dlb2_hw_dequeue_sparse(struct dlb2_eventdev *dlb2,
 
        start_ticks = rte_get_timer_cycles();
 
+       use_scalar = use_scalar || (max_num & 0x3);
+
        while (num < max_num) {
                struct dlb2_dequeue_qe qes[DLB2_NUM_QES_PER_CACHE_LINE];
                int num_avail;
-
-               /* Copy up to 4 QEs from the current cache line into qes */
-               num_avail = dlb2_recv_qe_sparse(qm_port, qes);
-
-               /* But don't process more than the user requested */
-               num_avail = RTE_MIN(num_avail, max_num - num);
-
-               dlb2_inc_cq_idx(qm_port, num_avail << 2);
-
-               if (num_avail == DLB2_NUM_QES_PER_CACHE_LINE)
-                       num += dlb2_process_dequeue_four_qes(ev_port,
-                                                             qm_port,
-                                                             &events[num],
-                                                             &qes[0]);
-               else if (num_avail)
-                       num += dlb2_process_dequeue_qes(ev_port,
-                                                        qm_port,
-                                                        &events[num],
-                                                        &qes[0],
-                                                        num_avail);
-               else if ((timeout == 0) || (num > 0))
-                       /* Not waiting in any form, or 1+ events received? */
-                       break;
-               else if (dlb2_dequeue_wait(dlb2, ev_port, qm_port,
-                                          timeout, start_ticks))
-                       break;
+               if (use_scalar) {
+                       num_avail = dlb2_recv_qe_sparse(qm_port, qes);
+                       num_avail = RTE_MIN(num_avail, max_num - num);
+                       dlb2_inc_cq_idx(qm_port, num_avail << 2);
+                       if (num_avail == DLB2_NUM_QES_PER_CACHE_LINE)
+                               num += dlb2_process_dequeue_four_qes(ev_port,
+                                                                 qm_port,
+                                                                 &events[num],
+                                                                 &qes[0]);
+                       else if (num_avail)
+                               num += dlb2_process_dequeue_qes(ev_port,
+                                                               qm_port,
+                                                               &events[num],
+                                                               &qes[0],
+                                                               num_avail);
+               } else { /* !use_scalar */
+                       num_avail = dlb2_recv_qe_sparse_vec(qm_port,
+                                                           &events[num],
+                                                           max_num - num);
+                       num += num_avail;
+                       dlb2_inc_cq_idx(qm_port, num_avail << 2);
+                       DLB2_INC_STAT(ev_port->stats.traffic.rx_ok, num_avail);
+               }
+               if (!num_avail) {
+                       if (num > 0)
+                               break;
+                       else if (dlb2_dequeue_wait(dlb2, ev_port, qm_port,
+                                                  timeout, start_ticks))
+                               break;
+               }
        }
 
        qm_port->owed_tokens += num;
@@ -4083,6 +4421,7 @@ dlb2_primary_eventdev_probe(struct rte_eventdev *dev,
        dlb2->poll_interval = dlb2_args->poll_interval;
        dlb2->sw_credit_quanta = dlb2_args->sw_credit_quanta;
        dlb2->default_depth_thresh = dlb2_args->default_depth_thresh;
+       dlb2->vector_opts_disabled = dlb2_args->vector_opts_disabled;
 
        err = dlb2_iface_open(&dlb2->qm_instance, name);
        if (err < 0) {
@@ -4186,6 +4525,7 @@ dlb2_parse_params(const char *params,
                                             DLB2_POLL_INTERVAL_ARG,
                                             DLB2_SW_CREDIT_QUANTA_ARG,
                                             DLB2_DEPTH_THRESH_ARG,
+                                            DLB2_VECTOR_OPTS_DISAB_ARG,
                                             NULL };
 
        if (params != NULL && params[0] != '\0') {
@@ -4299,6 +4639,17 @@ dlb2_parse_params(const char *params,
                                return ret;
                        }
 
+                       ret = rte_kvargs_process(kvlist,
+                                       DLB2_VECTOR_OPTS_DISAB_ARG,
+                                       set_vector_opts_disab,
+                                       &dlb2_args->vector_opts_disabled);
+                       if (ret != 0) {
+                               DLB2_LOG_ERR("%s: Error parsing vector opts disabled",
+                                            name);
+                               rte_kvargs_free(kvlist);
+                               return ret;
+                       }
+
                        rte_kvargs_free(kvlist);
                }
        }
index cf120c9..3140764 100644 (file)
@@ -38,6 +38,7 @@
 #define DLB2_POLL_INTERVAL_ARG "poll_interval"
 #define DLB2_SW_CREDIT_QUANTA_ARG "sw_credit_quanta"
 #define DLB2_DEPTH_THRESH_ARG "default_depth_thresh"
+#define DLB2_VECTOR_OPTS_DISAB_ARG "vector_opts_disable"
 
 /* Begin HW related defines and structs */
 
@@ -205,9 +206,9 @@ enum dlb2_enqueue_type {
 /* hw-specific format - do not change */
 
 struct dlb2_event_type {
-       uint8_t major:4;
-       uint8_t unused:4;
-       uint8_t sub;
+       uint16_t major:4;
+       uint16_t unused:4;
+       uint16_t sub:8;
 };
 
 union dlb2_opaque_data {
@@ -351,6 +352,12 @@ struct dlb2_port {
        uint16_t cq_idx_unmasked;
        uint16_t cq_depth_mask;
        uint16_t gen_bit_shift;
+       uint64_t cq_rolling_mask; /*
+                                  * rotate to always have right expected
+                                  * gen bits
+                                  */
+       uint64_t cq_rolling_mask_2;
+       void *cq_addr_cached; /* avoid multiple refs */
        enum dlb2_port_state state;
        enum dlb2_configuration_state config_state;
        int num_mapped_qids;
@@ -360,6 +367,7 @@ struct dlb2_port {
        struct dlb2_cq_pop_qe *consume_qe;
        struct dlb2_eventdev *dlb2; /* back ptr */
        struct dlb2_eventdev_port *ev_port; /* back ptr */
+       bool use_scalar; /* force usage of scalar code */
 };
 
 /* Per-process per-port mmio and memory pointers */
@@ -513,9 +521,9 @@ struct dlb2_queue {
        uint32_t num_qid_inflights; /* User config */
        uint32_t num_atm_inflights; /* User config */
        enum dlb2_configuration_state config_state;
-       int sched_type; /* LB queue only */
-       uint32_t id;
-       bool is_directed;
+       int  sched_type; /* LB queue only */
+       uint8_t id;
+       bool     is_directed;
 };
 
 struct dlb2_eventdev_queue {
@@ -558,6 +566,7 @@ struct dlb2_eventdev {
        uint32_t new_event_limit;
        int max_num_events_override;
        int num_dir_credits_override;
+       bool vector_opts_disabled;
        volatile enum dlb2_run_state run_state;
        uint16_t num_dir_queues; /* total num of evdev dir queues requested */
        union {
@@ -617,6 +626,7 @@ struct dlb2_devargs {
        int poll_interval;
        int sw_credit_quanta;
        int default_depth_thresh;
+       bool vector_opts_disabled;
 };
 
 /* End Eventdev related defines and structs */