1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2014-2018 Chelsio Communications.
14 #include <netinet/in.h>
16 #include <rte_byteorder.h>
17 #include <rte_common.h>
18 #include <rte_cycles.h>
19 #include <rte_interrupts.h>
21 #include <rte_debug.h>
23 #include <rte_branch_prediction.h>
24 #include <rte_memory.h>
25 #include <rte_memzone.h>
26 #include <rte_tailq.h>
28 #include <rte_alarm.h>
29 #include <rte_ether.h>
30 #include <ethdev_driver.h>
31 #include <rte_malloc.h>
32 #include <rte_random.h>
35 #include "base/common.h"
36 #include "base/t4_regs.h"
37 #include "base/t4_msg.h"
40 static inline void ship_tx_pkt_coalesce_wr(struct adapter *adap,
41 struct sge_eth_txq *txq);
44 * Max number of Rx buffers we replenish at a time.
46 #define MAX_RX_REFILL 64U
48 #define NOMEM_TMR_IDX (SGE_NTIMERS - 1)
51 * Max Tx descriptor space we allow for an Ethernet packet to be inlined
54 #define MAX_IMM_TX_PKT_LEN 256
57 * Max size of a WR sent through a control Tx queue.
59 #define MAX_CTRL_WR_LEN SGE_MAX_WR_LEN
62 * Bits 0..3 of rx_sw_desc.dma_addr have special meaning. The hardware uses
63 * these to specify the buffer size as an index into the SGE Free List Buffer
64 * Size register array. We also use bit 4, when the buffer has been unmapped
65 * for DMA, but this is of course never sent to the hardware and is only used
66 * to prevent double unmappings. All of the above requires that the Free List
67 * Buffers which we allocate have the bottom 5 bits free (0) -- i.e. are
68 * 32-byte or or a power of 2 greater in alignment. Since the SGE's minimal
69 * Free List Buffer alignment is 32 bytes, this works out for us ...
72 RX_BUF_FLAGS = 0x1f, /* bottom five bits are special */
73 RX_BUF_SIZE = 0x0f, /* bottom three bits are for buf sizes */
74 RX_UNMAPPED_BUF = 0x10, /* buffer is not mapped */
78 * txq_avail - return the number of available slots in a Tx queue
81 * Returns the number of descriptors in a Tx queue available to write new
84 static inline unsigned int txq_avail(const struct sge_txq *q)
86 return q->size - 1 - q->in_use;
89 static int map_mbuf(struct rte_mbuf *mbuf, dma_addr_t *addr)
91 struct rte_mbuf *m = mbuf;
93 for (; m; m = m->next, addr++) {
94 *addr = m->buf_iova + rte_pktmbuf_headroom(m);
105 * free_tx_desc - reclaims Tx descriptors and their buffers
106 * @q: the Tx queue to reclaim descriptors from
107 * @n: the number of descriptors to reclaim
109 * Reclaims Tx descriptors from an SGE Tx queue and frees the associated
110 * Tx buffers. Called with the Tx queue lock held.
112 static void free_tx_desc(struct sge_txq *q, unsigned int n)
114 struct tx_sw_desc *d;
115 unsigned int cidx = 0;
119 if (d->mbuf) { /* an SGL is present */
120 rte_pktmbuf_free(d->mbuf);
123 if (d->coalesce.idx) {
126 for (i = 0; i < d->coalesce.idx; i++) {
127 rte_pktmbuf_free(d->coalesce.mbuf[i]);
128 d->coalesce.mbuf[i] = NULL;
133 if (++cidx == q->size) {
137 RTE_MBUF_PREFETCH_TO_FREE(&q->sdesc->mbuf->pool);
141 static void reclaim_tx_desc(struct sge_txq *q, unsigned int n)
143 struct tx_sw_desc *d;
144 unsigned int cidx = q->cidx;
148 if (d->mbuf) { /* an SGL is present */
149 rte_pktmbuf_free(d->mbuf);
153 if (++cidx == q->size) {
162 * fl_cap - return the capacity of a free-buffer list
165 * Returns the capacity of a free-buffer list. The capacity is less than
166 * the size because one descriptor needs to be left unpopulated, otherwise
167 * HW will think the FL is empty.
169 static inline unsigned int fl_cap(const struct sge_fl *fl)
171 return fl->size - 8; /* 1 descriptor = 8 buffers */
175 * fl_starving - return whether a Free List is starving.
176 * @adapter: pointer to the adapter
179 * Tests specified Free List to see whether the number of buffers
180 * available to the hardware has fallen below our "starvation"
183 static inline bool fl_starving(const struct adapter *adapter,
184 const struct sge_fl *fl)
186 const struct sge *s = &adapter->sge;
188 return fl->avail - fl->pend_cred <= s->fl_starve_thres;
191 static inline unsigned int get_buf_size(struct adapter *adapter,
192 const struct rx_sw_desc *d)
194 return adapter->sge.fl_buffer_size[d->dma_addr & RX_BUF_SIZE];
198 * free_rx_bufs - free the Rx buffers on an SGE free list
199 * @q: the SGE free list to free buffers from
200 * @n: how many buffers to free
202 * Release the next @n buffers on an SGE free-buffer Rx queue. The
203 * buffers must be made inaccessible to HW before calling this function.
205 static void free_rx_bufs(struct sge_fl *q, int n)
207 unsigned int cidx = q->cidx;
208 struct rx_sw_desc *d;
213 rte_pktmbuf_free(d->buf);
217 if (++cidx == q->size) {
227 * unmap_rx_buf - unmap the current Rx buffer on an SGE free list
228 * @q: the SGE free list
230 * Unmap the current buffer on an SGE free-buffer Rx queue. The
231 * buffer must be made inaccessible to HW before calling this function.
233 * This is similar to @free_rx_bufs above but does not free the buffer.
234 * Do note that the FL still loses any further access to the buffer.
236 static void unmap_rx_buf(struct sge_fl *q)
238 if (++q->cidx == q->size)
243 static inline void ring_fl_db(struct adapter *adap, struct sge_fl *q)
245 if (q->pend_cred >= 64) {
246 u32 val = adap->params.arch.sge_fl_db;
248 if (is_t4(adap->params.chip))
249 val |= V_PIDX(q->pend_cred / 8);
251 val |= V_PIDX_T5(q->pend_cred / 8);
254 * Make sure all memory writes to the Free List queue are
255 * committed before we tell the hardware about them.
260 * If we don't have access to the new User Doorbell (T5+), use
261 * the old doorbell mechanism; otherwise use the new BAR2
264 if (unlikely(!q->bar2_addr)) {
265 u32 reg = is_pf4(adap) ? MYPF_REG(A_SGE_PF_KDOORBELL) :
269 t4_write_reg_relaxed(adap, reg,
270 val | V_QID(q->cntxt_id));
272 writel_relaxed(val | V_QID(q->bar2_qid),
273 (void *)((uintptr_t)q->bar2_addr +
277 * This Write memory Barrier will force the write to
278 * the User Doorbell area to be flushed.
286 static inline void set_rx_sw_desc(struct rx_sw_desc *sd, void *buf,
290 sd->dma_addr = mapping; /* includes size low bits */
294 * refill_fl_usembufs - refill an SGE Rx buffer ring with mbufs
296 * @q: the ring to refill
297 * @n: the number of new buffers to allocate
299 * (Re)populate an SGE free-buffer queue with up to @n new packet buffers,
300 * allocated with the supplied gfp flags. The caller must assure that
301 * @n does not exceed the queue's capacity. If afterwards the queue is
302 * found critically low mark it as starving in the bitmap of starving FLs.
304 * Returns the number of buffers allocated.
306 static unsigned int refill_fl_usembufs(struct adapter *adap, struct sge_fl *q,
309 struct sge_eth_rxq *rxq = container_of(q, struct sge_eth_rxq, fl);
310 struct rx_sw_desc *sd = &q->sdesc[q->pidx];
311 __be64 *d = &q->desc[q->pidx];
312 struct rte_mbuf *buf_bulk[n];
313 unsigned int cred = q->avail;
316 ret = rte_mempool_get_bulk(rxq->rspq.mb_pool, (void *)buf_bulk, n);
317 if (unlikely(ret != 0)) {
318 dev_debug(adap, "%s: failed to allocated fl entries in bulk ..\n",
321 rxq->rspq.eth_dev->data->rx_mbuf_alloc_failed++;
325 for (i = 0; i < n; i++) {
326 struct rte_mbuf *mbuf = buf_bulk[i];
330 dev_debug(adap, "%s: mbuf alloc failed\n", __func__);
332 rxq->rspq.eth_dev->data->rx_mbuf_alloc_failed++;
336 rte_mbuf_refcnt_set(mbuf, 1);
337 mbuf->data_off = RTE_PKTMBUF_HEADROOM;
340 mbuf->port = rxq->rspq.port_id;
342 mapping = (dma_addr_t)(mbuf->buf_iova + mbuf->data_off);
343 mapping |= q->fl_buf_size_idx;
344 *d++ = cpu_to_be64(mapping);
345 set_rx_sw_desc(sd, mbuf, mapping);
349 if (++q->pidx == q->size) {
356 out: cred = q->avail - cred;
357 q->pend_cred += cred;
360 if (unlikely(fl_starving(adap, q))) {
362 * Make sure data has been written to free list
372 * refill_fl - refill an SGE Rx buffer ring with mbufs
374 * @q: the ring to refill
375 * @n: the number of new buffers to allocate
377 * (Re)populate an SGE free-buffer queue with up to @n new packet buffers,
378 * allocated with the supplied gfp flags. The caller must assure that
379 * @n does not exceed the queue's capacity. Returns the number of buffers
382 static unsigned int refill_fl(struct adapter *adap, struct sge_fl *q, int n)
384 return refill_fl_usembufs(adap, q, n);
387 static inline void __refill_fl(struct adapter *adap, struct sge_fl *fl)
389 refill_fl(adap, fl, min(MAX_RX_REFILL, fl_cap(fl) - fl->avail));
393 * Return the number of reclaimable descriptors in a Tx queue.
395 static inline int reclaimable(const struct sge_txq *q)
397 int hw_cidx = ntohs(q->stat->cidx);
401 return hw_cidx + q->size;
406 * reclaim_completed_tx - reclaims completed Tx descriptors
407 * @q: the Tx queue to reclaim completed descriptors from
409 * Reclaims Tx descriptors that the SGE has indicated it has processed.
411 void reclaim_completed_tx(struct sge_txq *q)
413 unsigned int avail = reclaimable(q);
416 /* reclaim as much as possible */
417 reclaim_tx_desc(q, avail);
419 avail = reclaimable(q);
424 * sgl_len - calculates the size of an SGL of the given capacity
425 * @n: the number of SGL entries
427 * Calculates the number of flits needed for a scatter/gather list that
428 * can hold the given number of entries.
430 static inline unsigned int sgl_len(unsigned int n)
433 * A Direct Scatter Gather List uses 32-bit lengths and 64-bit PCI DMA
434 * addresses. The DSGL Work Request starts off with a 32-bit DSGL
435 * ULPTX header, then Length0, then Address0, then, for 1 <= i <= N,
436 * repeated sequences of { Length[i], Length[i+1], Address[i],
437 * Address[i+1] } (this ensures that all addresses are on 64-bit
438 * boundaries). If N is even, then Length[N+1] should be set to 0 and
439 * Address[N+1] is omitted.
441 * The following calculation incorporates all of the above. It's
442 * somewhat hard to follow but, briefly: the "+2" accounts for the
443 * first two flits which include the DSGL header, Length0 and
444 * Address0; the "(3*(n-1))/2" covers the main body of list entries (3
445 * flits for every pair of the remaining N) +1 if (n-1) is odd; and
446 * finally the "+((n-1)&1)" adds the one remaining flit needed if
450 return (3 * n) / 2 + (n & 1) + 2;
454 * flits_to_desc - returns the num of Tx descriptors for the given flits
455 * @n: the number of flits
457 * Returns the number of Tx descriptors needed for the supplied number
460 static inline unsigned int flits_to_desc(unsigned int n)
462 return DIV_ROUND_UP(n, 8);
466 * is_eth_imm - can an Ethernet packet be sent as immediate data?
469 * Returns whether an Ethernet packet is small enough to fit as
470 * immediate data. Return value corresponds to the headroom required.
472 static inline int is_eth_imm(const struct rte_mbuf *m)
474 unsigned int hdrlen = (m->ol_flags & RTE_MBUF_F_TX_TCP_SEG) ?
475 sizeof(struct cpl_tx_pkt_lso_core) : 0;
477 hdrlen += sizeof(struct cpl_tx_pkt);
478 if (m->pkt_len <= MAX_IMM_TX_PKT_LEN - hdrlen)
485 * calc_tx_flits - calculate the number of flits for a packet Tx WR
487 * @adap: adapter structure pointer
489 * Returns the number of flits needed for a Tx WR for the given Ethernet
490 * packet, including the needed WR and CPL headers.
492 static inline unsigned int calc_tx_flits(const struct rte_mbuf *m,
493 struct adapter *adap)
495 size_t wr_size = is_pf4(adap) ? sizeof(struct fw_eth_tx_pkt_wr) :
496 sizeof(struct fw_eth_tx_pkt_vm_wr);
501 * If the mbuf is small enough, we can pump it out as a work request
502 * with only immediate data. In that case we just have to have the
503 * TX Packet header plus the mbuf data in the Work Request.
506 hdrlen = is_eth_imm(m);
508 return DIV_ROUND_UP(m->pkt_len + hdrlen, sizeof(__be64));
511 * Otherwise, we're going to have to construct a Scatter gather list
512 * of the mbuf body and fragments. We also include the flits necessary
513 * for the TX Packet Work Request and CPL. We always have a firmware
514 * Write Header (incorporated as part of the cpl_tx_pkt_lso and
515 * cpl_tx_pkt structures), followed by either a TX Packet Write CPL
516 * message or, if we're doing a Large Send Offload, an LSO CPL message
517 * with an embedded TX Packet Write CPL message.
519 flits = sgl_len(m->nb_segs);
521 flits += (wr_size + sizeof(struct cpl_tx_pkt_lso_core) +
522 sizeof(struct cpl_tx_pkt_core)) / sizeof(__be64);
525 sizeof(struct cpl_tx_pkt_core)) / sizeof(__be64);
530 * write_sgl - populate a scatter/gather list for a packet
532 * @q: the Tx queue we are writing into
533 * @sgl: starting location for writing the SGL
534 * @end: points right after the end of the SGL
535 * @start: start offset into mbuf main-body data to include in the SGL
536 * @addr: address of mapped region
538 * Generates a scatter/gather list for the buffers that make up a packet.
539 * The caller must provide adequate space for the SGL that will be written.
540 * The SGL includes all of the packet's page fragments and the data in its
541 * main body except for the first @start bytes. @sgl must be 16-byte
542 * aligned and within a Tx descriptor with available space. @end points
543 * write after the end of the SGL but does not account for any potential
544 * wrap around, i.e., @end > @sgl.
546 static void write_sgl(struct rte_mbuf *mbuf, struct sge_txq *q,
547 struct ulptx_sgl *sgl, u64 *end, unsigned int start,
548 const dma_addr_t *addr)
551 struct ulptx_sge_pair *to;
552 struct rte_mbuf *m = mbuf;
553 unsigned int nfrags = m->nb_segs;
554 struct ulptx_sge_pair buf[nfrags / 2];
556 len = m->data_len - start;
557 sgl->len0 = htonl(len);
558 sgl->addr0 = rte_cpu_to_be_64(addr[0]);
560 sgl->cmd_nsge = htonl(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
561 V_ULPTX_NSGE(nfrags));
562 if (likely(--nfrags == 0))
565 * Most of the complexity below deals with the possibility we hit the
566 * end of the queue in the middle of writing the SGL. For this case
567 * only we create the SGL in a temporary buffer and then copy it.
569 to = (u8 *)end > (u8 *)q->stat ? buf : sgl->sge;
571 for (i = 0; nfrags >= 2; nfrags -= 2, to++) {
573 to->len[0] = rte_cpu_to_be_32(m->data_len);
574 to->addr[0] = rte_cpu_to_be_64(addr[++i]);
576 to->len[1] = rte_cpu_to_be_32(m->data_len);
577 to->addr[1] = rte_cpu_to_be_64(addr[++i]);
581 to->len[0] = rte_cpu_to_be_32(m->data_len);
582 to->len[1] = rte_cpu_to_be_32(0);
583 to->addr[0] = rte_cpu_to_be_64(addr[i + 1]);
585 if (unlikely((u8 *)end > (u8 *)q->stat)) {
586 unsigned int part0 = RTE_PTR_DIFF((u8 *)q->stat,
591 memcpy(sgl->sge, buf, part0);
592 part1 = RTE_PTR_DIFF((u8 *)end, (u8 *)q->stat);
593 rte_memcpy(q->desc, RTE_PTR_ADD((u8 *)buf, part0), part1);
594 end = RTE_PTR_ADD((void *)q->desc, part1);
596 if ((uintptr_t)end & 8) /* 0-pad to multiple of 16 */
600 #define IDXDIFF(head, tail, wrap) \
601 ((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
603 #define Q_IDXDIFF(q, idx) IDXDIFF((q)->pidx, (q)->idx, (q)->size)
604 #define R_IDXDIFF(q, idx) IDXDIFF((q)->cidx, (q)->idx, (q)->size)
606 #define PIDXDIFF(head, tail, wrap) \
607 ((tail) >= (head) ? (tail) - (head) : (wrap) - (head) + (tail))
608 #define P_IDXDIFF(q, idx) PIDXDIFF((q)->cidx, idx, (q)->size)
611 * ring_tx_db - ring a Tx queue's doorbell
614 * @n: number of new descriptors to give to HW
616 * Ring the doorbell for a Tx queue.
618 static inline void ring_tx_db(struct adapter *adap, struct sge_txq *q)
620 int n = Q_IDXDIFF(q, dbidx);
623 * Make sure that all writes to the TX Descriptors are committed
624 * before we tell the hardware about them.
629 * If we don't have access to the new User Doorbell (T5+), use the old
630 * doorbell mechanism; otherwise use the new BAR2 mechanism.
632 if (unlikely(!q->bar2_addr)) {
636 * For T4 we need to participate in the Doorbell Recovery
640 t4_write_reg(adap, MYPF_REG(A_SGE_PF_KDOORBELL),
641 V_QID(q->cntxt_id) | val);
644 q->db_pidx = q->pidx;
646 u32 val = V_PIDX_T5(n);
649 * T4 and later chips share the same PIDX field offset within
650 * the doorbell, but T5 and later shrank the field in order to
651 * gain a bit for Doorbell Priority. The field was absurdly
652 * large in the first place (14 bits) so we just use the T5
653 * and later limits and warn if a Queue ID is too large.
655 WARN_ON(val & F_DBPRIO);
657 writel(val | V_QID(q->bar2_qid),
658 (void *)((uintptr_t)q->bar2_addr + SGE_UDB_KDOORBELL));
661 * This Write Memory Barrier will force the write to the User
662 * Doorbell area to be flushed. This is needed to prevent
663 * writes on different CPUs for the same queue from hitting
664 * the adapter out of order. This is required when some Work
665 * Requests take the Write Combine Gather Buffer path (user
666 * doorbell area offset [SGE_UDB_WCDOORBELL..+63]) and some
667 * take the traditional path where we simply increment the
668 * PIDX (User Doorbell area SGE_UDB_KDOORBELL) and have the
669 * hardware DMA read the actual Work Request.
677 * Figure out what HW csum a packet wants and return the appropriate control
680 static u64 hwcsum(enum chip_type chip, const struct rte_mbuf *m)
684 if (m->ol_flags & RTE_MBUF_F_TX_IP_CKSUM) {
685 switch (m->ol_flags & RTE_MBUF_F_TX_L4_MASK) {
686 case RTE_MBUF_F_TX_TCP_CKSUM:
687 csum_type = TX_CSUM_TCPIP;
689 case RTE_MBUF_F_TX_UDP_CKSUM:
690 csum_type = TX_CSUM_UDPIP;
699 if (likely(csum_type >= TX_CSUM_TCPIP)) {
700 u64 hdr_len = V_TXPKT_IPHDR_LEN(m->l3_len);
701 int eth_hdr_len = m->l2_len;
703 if (CHELSIO_CHIP_VERSION(chip) <= CHELSIO_T5)
704 hdr_len |= V_TXPKT_ETHHDR_LEN(eth_hdr_len);
706 hdr_len |= V_T6_TXPKT_ETHHDR_LEN(eth_hdr_len);
707 return V_TXPKT_CSUM_TYPE(csum_type) | hdr_len;
711 * unknown protocol, disable HW csum
712 * and hope a bad packet is detected
714 return F_TXPKT_L4CSUM_DIS;
717 static inline void txq_advance(struct sge_txq *q, unsigned int n)
721 if (q->pidx >= q->size)
725 #define MAX_COALESCE_LEN 64000
727 static inline bool wraps_around(struct sge_txq *q, int ndesc)
729 return (q->pidx + ndesc) > q->size ? true : false;
732 static void tx_timer_cb(void *data)
734 struct adapter *adap = (struct adapter *)data;
735 struct sge_eth_txq *txq = &adap->sge.ethtxq[0];
737 unsigned int coal_idx;
739 /* monitor any pending tx */
740 for (i = 0; i < adap->sge.max_ethqsets; i++, txq++) {
741 if (t4_os_trylock(&txq->txq_lock)) {
742 coal_idx = txq->q.coalesce.idx;
744 if (coal_idx == txq->q.last_coal_idx &&
745 txq->q.pidx == txq->q.last_pidx) {
746 ship_tx_pkt_coalesce_wr(adap, txq);
748 txq->q.last_coal_idx = coal_idx;
749 txq->q.last_pidx = txq->q.pidx;
752 t4_os_unlock(&txq->txq_lock);
755 rte_eal_alarm_set(50, tx_timer_cb, (void *)adap);
759 * ship_tx_pkt_coalesce_wr - finalizes and ships a coalesce WR
760 * @ adap: adapter structure
763 * writes the different fields of the pkts WR and sends it.
765 static inline void ship_tx_pkt_coalesce_wr(struct adapter *adap,
766 struct sge_eth_txq *txq)
768 struct fw_eth_tx_pkts_vm_wr *vmwr;
769 const size_t fw_hdr_copy_len = (sizeof(vmwr->ethmacdst) +
770 sizeof(vmwr->ethmacsrc) +
771 sizeof(vmwr->ethtype) +
772 sizeof(vmwr->vlantci));
773 struct fw_eth_tx_pkts_wr *wr;
774 struct sge_txq *q = &txq->q;
778 /* fill the pkts WR header */
779 wr = (void *)&q->desc[q->pidx];
780 vmwr = (void *)&q->desc[q->pidx];
782 wr_mid = V_FW_WR_LEN16(DIV_ROUND_UP(q->coalesce.flits, 2));
783 ndesc = flits_to_desc(q->coalesce.flits);
784 wr->equiq_to_len16 = htonl(wr_mid);
785 wr->plen = cpu_to_be16(q->coalesce.len);
786 wr->npkt = q->coalesce.idx;
789 wr->type = q->coalesce.type;
790 if (likely(wr->type != 0))
791 wr->op_pkd = htonl(V_FW_WR_OP(FW_ETH_TX_PKTS2_WR));
793 wr->op_pkd = htonl(V_FW_WR_OP(FW_ETH_TX_PKTS_WR));
795 wr->op_pkd = htonl(V_FW_WR_OP(FW_ETH_TX_PKTS_VM_WR));
797 memcpy((void *)vmwr->ethmacdst, (void *)q->coalesce.ethmacdst,
801 /* zero out coalesce structure members */
802 memset((void *)&q->coalesce, 0, sizeof(struct eth_coalesce));
804 txq_advance(q, ndesc);
805 txq->stats.coal_wr++;
806 txq->stats.coal_pkts += wr->npkt;
808 if (Q_IDXDIFF(q, equeidx) >= q->size / 2) {
809 q->equeidx = q->pidx;
810 wr_mid |= F_FW_WR_EQUEQ;
811 wr->equiq_to_len16 = htonl(wr_mid);
817 * should_tx_packet_coalesce - decides whether to coalesce an mbuf or not
818 * @txq: tx queue where the mbuf is sent
819 * @mbuf: mbuf to be sent
820 * @nflits: return value for number of flits needed
821 * @adap: adapter structure
823 * This function decides if a packet should be coalesced or not.
825 static inline int should_tx_packet_coalesce(struct sge_eth_txq *txq,
826 struct rte_mbuf *mbuf,
827 unsigned int *nflits,
828 struct adapter *adap)
830 struct fw_eth_tx_pkts_vm_wr *wr;
831 const size_t fw_hdr_copy_len = (sizeof(wr->ethmacdst) +
832 sizeof(wr->ethmacsrc) +
833 sizeof(wr->ethtype) +
834 sizeof(wr->vlantci));
835 struct sge_txq *q = &txq->q;
836 unsigned int flits, ndesc;
837 unsigned char type = 0;
838 int credits, wr_size;
840 /* use coal WR type 1 when no frags are present */
841 type = (mbuf->nb_segs == 1) ? 1 : 0;
846 if (q->coalesce.idx && memcmp((void *)q->coalesce.ethmacdst,
847 rte_pktmbuf_mtod(mbuf, void *),
849 ship_tx_pkt_coalesce_wr(adap, txq);
852 if (unlikely(type != q->coalesce.type && q->coalesce.idx))
853 ship_tx_pkt_coalesce_wr(adap, txq);
855 /* calculate the number of flits required for coalescing this packet
856 * without the 2 flits of the WR header. These are added further down
857 * if we are just starting in new PKTS WR. sgl_len doesn't account for
858 * the possible 16 bytes alignment ULP TX commands so we do it here.
860 flits = (sgl_len(mbuf->nb_segs) + 1) & ~1U;
862 flits += (sizeof(struct ulp_txpkt) +
863 sizeof(struct ulptx_idata)) / sizeof(__be64);
864 flits += sizeof(struct cpl_tx_pkt_core) / sizeof(__be64);
867 /* If coalescing is on, the mbuf is added to a pkts WR */
868 if (q->coalesce.idx) {
869 ndesc = DIV_ROUND_UP(q->coalesce.flits + flits, 8);
870 credits = txq_avail(q) - ndesc;
872 if (unlikely(wraps_around(q, ndesc)))
875 /* If we are wrapping or this is last mbuf then, send the
876 * already coalesced mbufs and let the non-coalesce pass
879 if (unlikely(credits < 0)) {
880 ship_tx_pkt_coalesce_wr(adap, txq);
884 /* If the max coalesce len or the max WR len is reached
885 * ship the WR and keep coalescing on.
887 if (unlikely((q->coalesce.len + mbuf->pkt_len >
889 (q->coalesce.flits + flits >
891 ship_tx_pkt_coalesce_wr(adap, txq);
898 /* start a new pkts WR, the WR header is not filled below */
899 wr_size = is_pf4(adap) ? sizeof(struct fw_eth_tx_pkts_wr) :
900 sizeof(struct fw_eth_tx_pkts_vm_wr);
901 flits += wr_size / sizeof(__be64);
902 ndesc = flits_to_desc(q->coalesce.flits + flits);
903 credits = txq_avail(q) - ndesc;
905 if (unlikely(wraps_around(q, ndesc)))
908 if (unlikely(credits < 0))
911 q->coalesce.flits += wr_size / sizeof(__be64);
912 q->coalesce.type = type;
913 q->coalesce.ptr = (unsigned char *)&q->desc[q->pidx] +
914 q->coalesce.flits * sizeof(__be64);
916 memcpy((void *)q->coalesce.ethmacdst,
917 rte_pktmbuf_mtod(mbuf, void *), fw_hdr_copy_len);
922 * tx_do_packet_coalesce - add an mbuf to a coalesce WR
923 * @txq: sge_eth_txq used send the mbuf
924 * @mbuf: mbuf to be sent
925 * @flits: flits needed for this mbuf
926 * @adap: adapter structure
927 * @pi: port_info structure
928 * @addr: mapped address of the mbuf
930 * Adds an mbuf to be sent as part of a coalesce WR by filling a
931 * ulp_tx_pkt command, ulp_tx_sc_imm command, cpl message and
932 * ulp_tx_sc_dsgl command.
934 static inline int tx_do_packet_coalesce(struct sge_eth_txq *txq,
935 struct rte_mbuf *mbuf,
936 int flits, struct adapter *adap,
937 const struct port_info *pi,
938 dma_addr_t *addr, uint16_t nb_pkts)
941 struct sge_txq *q = &txq->q;
942 struct ulp_txpkt *mc;
943 struct ulptx_idata *sc_imm;
944 struct cpl_tx_pkt_core *cpl;
945 struct tx_sw_desc *sd;
946 unsigned int idx = q->coalesce.idx, len = mbuf->pkt_len;
948 if (q->coalesce.type == 0) {
949 mc = (struct ulp_txpkt *)q->coalesce.ptr;
950 mc->cmd_dest = htonl(V_ULPTX_CMD(4) | V_ULP_TXPKT_DEST(0) |
951 V_ULP_TXPKT_FID(adap->sge.fw_evtq.cntxt_id) |
953 mc->len = htonl(DIV_ROUND_UP(flits, 2));
954 sc_imm = (struct ulptx_idata *)(mc + 1);
955 sc_imm->cmd_more = htonl(V_ULPTX_CMD(ULP_TX_SC_IMM) |
957 sc_imm->len = htonl(sizeof(*cpl));
958 end = (u64 *)mc + flits;
959 cpl = (struct cpl_tx_pkt_core *)(sc_imm + 1);
961 end = (u64 *)q->coalesce.ptr + flits;
962 cpl = (struct cpl_tx_pkt_core *)q->coalesce.ptr;
965 /* update coalesce structure for this txq */
966 q->coalesce.flits += flits;
967 q->coalesce.ptr += flits * sizeof(__be64);
968 q->coalesce.len += mbuf->pkt_len;
970 /* fill the cpl message, same as in t4_eth_xmit, this should be kept
971 * similar to t4_eth_xmit
973 if (mbuf->ol_flags & RTE_MBUF_F_TX_IP_CKSUM) {
974 cntrl = hwcsum(adap->params.chip, mbuf) |
978 cntrl = F_TXPKT_L4CSUM_DIS | F_TXPKT_IPCSUM_DIS;
981 if (mbuf->ol_flags & RTE_MBUF_F_TX_VLAN) {
982 txq->stats.vlan_ins++;
983 cntrl |= F_TXPKT_VLAN_VLD | V_TXPKT_VLAN(mbuf->vlan_tci);
986 cpl->ctrl0 = htonl(V_TXPKT_OPCODE(CPL_TX_PKT_XT));
988 cpl->ctrl0 |= htonl(V_TXPKT_INTF(pi->tx_chan) |
989 V_TXPKT_PF(adap->pf));
991 cpl->ctrl0 |= htonl(V_TXPKT_INTF(pi->port_id));
992 cpl->pack = htons(0);
993 cpl->len = htons(len);
994 cpl->ctrl1 = cpu_to_be64(cntrl);
995 write_sgl(mbuf, q, (struct ulptx_sgl *)(cpl + 1), end, 0, addr);
997 txq->stats.tx_bytes += len;
999 sd = &q->sdesc[q->pidx + (idx >> 1)];
1001 if (sd->coalesce.idx) {
1004 for (i = 0; i < sd->coalesce.idx; i++) {
1005 rte_pktmbuf_free(sd->coalesce.mbuf[i]);
1006 sd->coalesce.mbuf[i] = NULL;
1011 /* store pointers to the mbuf and the sgl used in free_tx_desc.
1012 * each tx desc can hold two pointers corresponding to the value
1013 * of ETH_COALESCE_PKT_PER_DESC
1015 sd->coalesce.mbuf[idx & 1] = mbuf;
1016 sd->coalesce.sgl[idx & 1] = (struct ulptx_sgl *)(cpl + 1);
1017 sd->coalesce.idx = (idx & 1) + 1;
1019 /* Send the coalesced work request, only if max reached. However,
1020 * if lower latency is preferred over throughput, then don't wait
1021 * for coalescing the next Tx burst and send the packets now.
1024 if (q->coalesce.idx == adap->params.max_tx_coalesce_num ||
1025 (adap->devargs.tx_mode_latency && q->coalesce.idx >= nb_pkts))
1026 ship_tx_pkt_coalesce_wr(adap, txq);
1032 * t4_eth_xmit - add a packet to an Ethernet Tx queue
1033 * @txq: the egress queue
1036 * Add a packet to an SGE Ethernet Tx queue. Runs with softirqs disabled.
1038 int t4_eth_xmit(struct sge_eth_txq *txq, struct rte_mbuf *mbuf,
1041 const struct port_info *pi;
1042 struct cpl_tx_pkt_lso_core *lso;
1043 struct adapter *adap;
1044 struct rte_mbuf *m = mbuf;
1045 struct fw_eth_tx_pkt_wr *wr;
1046 struct fw_eth_tx_pkt_vm_wr *vmwr;
1047 struct cpl_tx_pkt_core *cpl;
1048 struct tx_sw_desc *d;
1049 dma_addr_t addr[m->nb_segs];
1050 unsigned int flits, ndesc, cflits;
1051 int l3hdr_len, l4hdr_len, eth_xtra_len;
1053 int should_coal, credits;
1059 /* Reject xmit if queue is stopped */
1060 if (unlikely(txq->flags & EQ_STOPPED))
1064 * The chip min packet length is 10 octets but play safe and reject
1065 * anything shorter than an Ethernet header.
1067 if (unlikely(m->pkt_len < RTE_ETHER_HDR_LEN)) {
1069 rte_pktmbuf_free(m);
1073 max_pkt_len = txq->data->mtu + RTE_ETHER_HDR_LEN + RTE_ETHER_CRC_LEN;
1074 if ((!(m->ol_flags & RTE_MBUF_F_TX_TCP_SEG)) &&
1075 (unlikely(m->pkt_len > max_pkt_len)))
1078 pi = txq->data->dev_private;
1081 cntrl = F_TXPKT_L4CSUM_DIS | F_TXPKT_IPCSUM_DIS;
1082 /* align the end of coalesce WR to a 512 byte boundary */
1083 txq->q.coalesce.max = (8 - (txq->q.pidx & 7)) * 8;
1085 if ((m->ol_flags & RTE_MBUF_F_TX_TCP_SEG) == 0) {
1086 should_coal = should_tx_packet_coalesce(txq, mbuf, &cflits, adap);
1087 if (should_coal > 0) {
1088 if (unlikely(map_mbuf(mbuf, addr) < 0)) {
1089 dev_warn(adap, "%s: mapping err for coalesce\n",
1091 txq->stats.mapping_err++;
1094 return tx_do_packet_coalesce(txq, mbuf, cflits, adap,
1096 } else if (should_coal < 0) {
1101 if (txq->q.coalesce.idx)
1102 ship_tx_pkt_coalesce_wr(adap, txq);
1104 flits = calc_tx_flits(m, adap);
1105 ndesc = flits_to_desc(flits);
1106 credits = txq_avail(&txq->q) - ndesc;
1108 if (unlikely(credits < 0)) {
1109 dev_debug(adap, "%s: Tx ring %u full; credits = %d\n",
1110 __func__, txq->q.cntxt_id, credits);
1114 if (unlikely(map_mbuf(m, addr) < 0)) {
1115 txq->stats.mapping_err++;
1119 wr_mid = V_FW_WR_LEN16(DIV_ROUND_UP(flits, 2));
1120 if (Q_IDXDIFF(&txq->q, equeidx) >= 64) {
1121 txq->q.equeidx = txq->q.pidx;
1122 wr_mid |= F_FW_WR_EQUEQ;
1125 wr = (void *)&txq->q.desc[txq->q.pidx];
1126 vmwr = (void *)&txq->q.desc[txq->q.pidx];
1127 wr->equiq_to_len16 = htonl(wr_mid);
1129 wr->r3 = rte_cpu_to_be_64(0);
1130 end = (u64 *)wr + flits;
1132 const size_t fw_hdr_copy_len = (sizeof(vmwr->ethmacdst) +
1133 sizeof(vmwr->ethmacsrc) +
1134 sizeof(vmwr->ethtype) +
1135 sizeof(vmwr->vlantci));
1137 vmwr->r3[0] = rte_cpu_to_be_32(0);
1138 vmwr->r3[1] = rte_cpu_to_be_32(0);
1139 memcpy((void *)vmwr->ethmacdst, rte_pktmbuf_mtod(m, void *),
1141 end = (u64 *)vmwr + flits;
1146 /* Coalescing skipped and we send through normal path */
1147 if (!(m->ol_flags & RTE_MBUF_F_TX_TCP_SEG)) {
1148 wr->op_immdlen = htonl(V_FW_WR_OP(is_pf4(adap) ?
1150 FW_ETH_TX_PKT_VM_WR) |
1151 V_FW_WR_IMMDLEN(len));
1153 cpl = (void *)(wr + 1);
1155 cpl = (void *)(vmwr + 1);
1156 if (m->ol_flags & RTE_MBUF_F_TX_IP_CKSUM) {
1157 cntrl = hwcsum(adap->params.chip, m) |
1159 txq->stats.tx_cso++;
1163 lso = (void *)(wr + 1);
1165 lso = (void *)(vmwr + 1);
1166 v6 = (m->ol_flags & RTE_MBUF_F_TX_IPV6) != 0;
1167 l3hdr_len = m->l3_len;
1168 l4hdr_len = m->l4_len;
1169 eth_xtra_len = m->l2_len - RTE_ETHER_HDR_LEN;
1170 len += sizeof(*lso);
1171 wr->op_immdlen = htonl(V_FW_WR_OP(is_pf4(adap) ?
1173 FW_ETH_TX_PKT_VM_WR) |
1174 V_FW_WR_IMMDLEN(len));
1175 lso->lso_ctrl = htonl(V_LSO_OPCODE(CPL_TX_PKT_LSO) |
1176 F_LSO_FIRST_SLICE | F_LSO_LAST_SLICE |
1178 V_LSO_ETHHDR_LEN(eth_xtra_len / 4) |
1179 V_LSO_IPHDR_LEN(l3hdr_len / 4) |
1180 V_LSO_TCPHDR_LEN(l4hdr_len / 4));
1181 lso->ipid_ofst = htons(0);
1182 lso->mss = htons(m->tso_segsz);
1183 lso->seqno_offset = htonl(0);
1184 if (is_t4(adap->params.chip))
1185 lso->len = htonl(m->pkt_len);
1187 lso->len = htonl(V_LSO_T5_XFER_SIZE(m->pkt_len));
1188 cpl = (void *)(lso + 1);
1190 if (CHELSIO_CHIP_VERSION(adap->params.chip) <= CHELSIO_T5)
1191 cntrl = V_TXPKT_ETHHDR_LEN(eth_xtra_len);
1193 cntrl = V_T6_TXPKT_ETHHDR_LEN(eth_xtra_len);
1195 cntrl |= V_TXPKT_CSUM_TYPE(v6 ? TX_CSUM_TCPIP6 :
1197 V_TXPKT_IPHDR_LEN(l3hdr_len);
1199 txq->stats.tx_cso += m->tso_segsz;
1202 if (m->ol_flags & RTE_MBUF_F_TX_VLAN) {
1203 txq->stats.vlan_ins++;
1204 cntrl |= F_TXPKT_VLAN_VLD | V_TXPKT_VLAN(m->vlan_tci);
1207 cpl->ctrl0 = htonl(V_TXPKT_OPCODE(CPL_TX_PKT_XT));
1209 cpl->ctrl0 |= htonl(V_TXPKT_INTF(pi->tx_chan) |
1210 V_TXPKT_PF(adap->pf));
1212 cpl->ctrl0 |= htonl(V_TXPKT_INTF(pi->port_id) |
1215 cpl->pack = htons(0);
1216 cpl->len = htons(m->pkt_len);
1217 cpl->ctrl1 = cpu_to_be64(cntrl);
1220 txq->stats.tx_bytes += m->pkt_len;
1221 last_desc = txq->q.pidx + ndesc - 1;
1222 if (last_desc >= (int)txq->q.size)
1223 last_desc -= txq->q.size;
1225 d = &txq->q.sdesc[last_desc];
1226 if (d->coalesce.idx) {
1229 for (i = 0; i < d->coalesce.idx; i++) {
1230 rte_pktmbuf_free(d->coalesce.mbuf[i]);
1231 d->coalesce.mbuf[i] = NULL;
1233 d->coalesce.idx = 0;
1235 write_sgl(m, &txq->q, (struct ulptx_sgl *)(cpl + 1), end, 0,
1237 txq->q.sdesc[last_desc].mbuf = m;
1238 txq->q.sdesc[last_desc].sgl = (struct ulptx_sgl *)(cpl + 1);
1239 txq_advance(&txq->q, ndesc);
1240 ring_tx_db(adap, &txq->q);
1245 * reclaim_completed_tx_imm - reclaim completed control-queue Tx descs
1246 * @q: the SGE control Tx queue
1248 * This is a variant of reclaim_completed_tx() that is used for Tx queues
1249 * that send only immediate data (presently just the control queues) and
1250 * thus do not have any mbufs to release.
1252 static inline void reclaim_completed_tx_imm(struct sge_txq *q)
1254 int hw_cidx = ntohs(q->stat->cidx);
1255 int reclaim = hw_cidx - q->cidx;
1260 q->in_use -= reclaim;
1265 * is_imm - check whether a packet can be sent as immediate data
1268 * Returns true if a packet can be sent as a WR with immediate data.
1270 static inline int is_imm(const struct rte_mbuf *mbuf)
1272 return mbuf->pkt_len <= MAX_CTRL_WR_LEN;
1276 * inline_tx_mbuf: inline a packet's data into TX descriptors
1277 * @q: the TX queue where the packet will be inlined
1278 * @from: pointer to data portion of packet
1279 * @to: pointer after cpl where data has to be inlined
1280 * @len: length of data to inline
1282 * Inline a packet's contents directly to TX descriptors, starting at
1283 * the given position within the TX DMA ring.
1284 * Most of the complexity of this operation is dealing with wrap arounds
1285 * in the middle of the packet we want to inline.
1287 static void inline_tx_mbuf(const struct sge_txq *q, caddr_t from, caddr_t *to,
1290 int left = RTE_PTR_DIFF(q->stat, *to);
1292 if (likely((uintptr_t)*to + len <= (uintptr_t)q->stat)) {
1293 rte_memcpy(*to, from, len);
1294 *to = RTE_PTR_ADD(*to, len);
1296 rte_memcpy(*to, from, left);
1297 from = RTE_PTR_ADD(from, left);
1299 rte_memcpy((void *)q->desc, from, left);
1300 *to = RTE_PTR_ADD((void *)q->desc, left);
1305 * ctrl_xmit - send a packet through an SGE control Tx queue
1306 * @q: the control queue
1309 * Send a packet through an SGE control Tx queue. Packets sent through
1310 * a control queue must fit entirely as immediate data.
1312 static int ctrl_xmit(struct sge_ctrl_txq *q, struct rte_mbuf *mbuf)
1315 struct fw_wr_hdr *wr;
1318 if (unlikely(!is_imm(mbuf))) {
1320 rte_pktmbuf_free(mbuf);
1324 reclaim_completed_tx_imm(&q->q);
1325 ndesc = DIV_ROUND_UP(mbuf->pkt_len, sizeof(struct tx_desc));
1326 t4_os_lock(&q->ctrlq_lock);
1328 q->full = txq_avail(&q->q) < ndesc ? 1 : 0;
1329 if (unlikely(q->full)) {
1330 t4_os_unlock(&q->ctrlq_lock);
1334 wr = (struct fw_wr_hdr *)&q->q.desc[q->q.pidx];
1336 inline_tx_mbuf(&q->q, rte_pktmbuf_mtod(mbuf, caddr_t),
1337 &dst, mbuf->data_len);
1339 txq_advance(&q->q, ndesc);
1340 if (unlikely(txq_avail(&q->q) < 64))
1341 wr->lo |= htonl(F_FW_WR_EQUEQ);
1345 ring_tx_db(q->adapter, &q->q);
1346 t4_os_unlock(&q->ctrlq_lock);
1348 rte_pktmbuf_free(mbuf);
1353 * t4_mgmt_tx - send a management message
1354 * @q: the control queue
1355 * @mbuf: the packet containing the management message
1357 * Send a management message through control queue.
1359 int t4_mgmt_tx(struct sge_ctrl_txq *q, struct rte_mbuf *mbuf)
1361 return ctrl_xmit(q, mbuf);
1365 * alloc_ring - allocate resources for an SGE descriptor ring
1366 * @dev: the port associated with the queue
1367 * @z_name: memzone's name
1368 * @queue_id: queue index
1369 * @socket_id: preferred socket id for memory allocations
1370 * @nelem: the number of descriptors
1371 * @elem_size: the size of each descriptor
1372 * @stat_size: extra space in HW ring for status information
1373 * @sw_size: the size of the SW state associated with each ring element
1374 * @phys: the physical address of the allocated ring
1375 * @metadata: address of the array holding the SW state for the ring
1377 * Allocates resources for an SGE descriptor ring, such as Tx queues,
1378 * free buffer lists, or response queues. Each SGE ring requires
1379 * space for its HW descriptors plus, optionally, space for the SW state
1380 * associated with each HW entry (the metadata). The function returns
1381 * three values: the virtual address for the HW ring (the return value
1382 * of the function), the bus address of the HW ring, and the address
1385 static void *alloc_ring(struct rte_eth_dev *dev, const char *z_name,
1386 uint16_t queue_id, int socket_id, size_t nelem,
1387 size_t elem_size, size_t stat_size, size_t sw_size,
1388 dma_addr_t *phys, void *metadata)
1390 size_t len = CXGBE_MAX_RING_DESC_SIZE * elem_size + stat_size;
1391 char z_name_sw[RTE_MEMZONE_NAMESIZE];
1392 const struct rte_memzone *tz;
1395 snprintf(z_name_sw, sizeof(z_name_sw), "eth_p%d_q%d_%s_sw_ring",
1396 dev->data->port_id, queue_id, z_name);
1398 dev_debug(adapter, "%s: nelem = %zu; elem_size = %zu; sw_size = %zu; "
1399 "stat_size = %zu; queue_id = %u; socket_id = %d; z_name = %s;"
1400 " z_name_sw = %s\n", __func__, nelem, elem_size, sw_size,
1401 stat_size, queue_id, socket_id, z_name, z_name_sw);
1404 * Allocate TX/RX ring hardware descriptors. A memzone large enough to
1405 * handle the maximum ring size is allocated in order to allow for
1406 * resizing in later calls to the queue setup function.
1408 tz = rte_eth_dma_zone_reserve(dev, z_name, queue_id, len, 4096,
1413 memset(tz->addr, 0, len);
1415 s = rte_zmalloc_socket(z_name_sw, nelem * sw_size,
1416 RTE_CACHE_LINE_SIZE, socket_id);
1419 dev_err(adapter, "%s: failed to get sw_ring memory\n",
1425 *(void **)metadata = s;
1427 *phys = (uint64_t)tz->iova;
1431 #define CXGB4_MSG_AN ((void *)1)
1434 * rspq_next - advance to the next entry in a response queue
1437 * Updates the state of a response queue to advance it to the next entry.
1439 static inline void rspq_next(struct sge_rspq *q)
1441 q->cur_desc = (const __be64 *)((const char *)q->cur_desc + q->iqe_len);
1442 if (unlikely(++q->cidx == q->size)) {
1445 q->cur_desc = q->desc;
1449 static inline void cxgbe_set_mbuf_info(struct rte_mbuf *pkt, uint32_t ptype,
1452 pkt->packet_type |= ptype;
1453 pkt->ol_flags |= ol_flags;
1456 static inline void cxgbe_fill_mbuf_info(struct adapter *adap,
1457 const struct cpl_rx_pkt *cpl,
1458 struct rte_mbuf *pkt)
1463 if (adap->params.tp.rx_pkt_encap)
1464 err_vec = G_T6_COMPR_RXERR_VEC(ntohs(cpl->err_vec));
1466 err_vec = ntohs(cpl->err_vec);
1468 csum_ok = cpl->csum_calc && !err_vec;
1471 cxgbe_set_mbuf_info(pkt, RTE_PTYPE_L2_ETHER_VLAN,
1472 RTE_MBUF_F_RX_VLAN | RTE_MBUF_F_RX_VLAN_STRIPPED);
1474 cxgbe_set_mbuf_info(pkt, RTE_PTYPE_L2_ETHER, 0);
1476 if (cpl->l2info & htonl(F_RXF_IP))
1477 cxgbe_set_mbuf_info(pkt, RTE_PTYPE_L3_IPV4,
1478 csum_ok ? RTE_MBUF_F_RX_IP_CKSUM_GOOD :
1479 RTE_MBUF_F_RX_IP_CKSUM_BAD);
1480 else if (cpl->l2info & htonl(F_RXF_IP6))
1481 cxgbe_set_mbuf_info(pkt, RTE_PTYPE_L3_IPV6,
1482 csum_ok ? RTE_MBUF_F_RX_IP_CKSUM_GOOD :
1483 RTE_MBUF_F_RX_IP_CKSUM_BAD);
1485 if (cpl->l2info & htonl(F_RXF_TCP))
1486 cxgbe_set_mbuf_info(pkt, RTE_PTYPE_L4_TCP,
1487 csum_ok ? RTE_MBUF_F_RX_L4_CKSUM_GOOD :
1488 RTE_MBUF_F_RX_L4_CKSUM_BAD);
1489 else if (cpl->l2info & htonl(F_RXF_UDP))
1490 cxgbe_set_mbuf_info(pkt, RTE_PTYPE_L4_UDP,
1491 csum_ok ? RTE_MBUF_F_RX_L4_CKSUM_GOOD :
1492 RTE_MBUF_F_RX_L4_CKSUM_BAD);
1496 * process_responses - process responses from an SGE response queue
1497 * @q: the ingress queue to process
1498 * @budget: how many responses can be processed in this round
1499 * @rx_pkts: mbuf to put the pkts
1501 * Process responses from an SGE response queue up to the supplied budget.
1502 * Responses include received packets as well as control messages from FW
1505 * Additionally choose the interrupt holdoff time for the next interrupt
1506 * on this queue. If the system is under memory shortage use a fairly
1507 * long delay to help recovery.
1509 static int process_responses(struct sge_rspq *q, int budget,
1510 struct rte_mbuf **rx_pkts)
1512 int ret = 0, rsp_type;
1513 int budget_left = budget;
1514 const struct rsp_ctrl *rc;
1515 struct sge_eth_rxq *rxq = container_of(q, struct sge_eth_rxq, rspq);
1517 while (likely(budget_left)) {
1518 if (q->cidx == ntohs(q->stat->pidx))
1521 rc = (const struct rsp_ctrl *)
1522 ((const char *)q->cur_desc + (q->iqe_len - sizeof(*rc)));
1525 * Ensure response has been read
1528 rsp_type = G_RSPD_TYPE(rc->u.type_gen);
1530 if (likely(rsp_type == X_RSPD_TYPE_FLBUF)) {
1531 struct sge *s = &q->adapter->sge;
1532 unsigned int stat_pidx;
1535 stat_pidx = ntohs(q->stat->pidx);
1536 stat_pidx_diff = P_IDXDIFF(q, stat_pidx);
1537 while (stat_pidx_diff && budget_left) {
1538 const struct rx_sw_desc *rsd =
1539 &rxq->fl.sdesc[rxq->fl.cidx];
1540 const struct rss_header *rss_hdr =
1541 (const void *)q->cur_desc;
1542 const struct cpl_rx_pkt *cpl =
1543 (const void *)&q->cur_desc[1];
1544 struct rte_mbuf *pkt, *npkt;
1547 rc = (const struct rsp_ctrl *)
1548 ((const char *)q->cur_desc +
1549 (q->iqe_len - sizeof(*rc)));
1551 rsp_type = G_RSPD_TYPE(rc->u.type_gen);
1552 if (unlikely(rsp_type != X_RSPD_TYPE_FLBUF))
1555 len = ntohl(rc->pldbuflen_qid);
1556 BUG_ON(!(len & F_RSPD_NEWBUF));
1559 len = G_RSPD_LEN(len);
1562 /* Chain mbufs into len if necessary */
1564 struct rte_mbuf *new_pkt = rsd->buf;
1566 bufsz = min(get_buf_size(q->adapter,
1568 new_pkt->data_len = bufsz;
1569 unmap_rx_buf(&rxq->fl);
1571 npkt->next = new_pkt;
1574 rsd = &rxq->fl.sdesc[rxq->fl.cidx];
1579 cxgbe_fill_mbuf_info(q->adapter, cpl, pkt);
1581 if (!rss_hdr->filter_tid &&
1582 rss_hdr->hash_type) {
1583 pkt->ol_flags |= RTE_MBUF_F_RX_RSS_HASH;
1585 ntohl(rss_hdr->hash_val);
1589 pkt->vlan_tci = ntohs(cpl->vlan);
1591 rte_pktmbuf_adj(pkt, s->pktshift);
1593 rxq->stats.rx_bytes += pkt->pkt_len;
1594 rx_pkts[budget - budget_left] = pkt;
1601 } else if (likely(rsp_type == X_RSPD_TYPE_CPL)) {
1602 ret = q->handler(q, q->cur_desc, NULL);
1604 ret = q->handler(q, (const __be64 *)rc, CXGB4_MSG_AN);
1607 if (unlikely(ret)) {
1608 /* couldn't process descriptor, back off for recovery */
1609 q->next_intr_params = V_QINTR_TIMER_IDX(NOMEM_TMR_IDX);
1618 * If this is a Response Queue with an associated Free List and
1619 * there's room for another chunk of new Free List buffer pointers,
1620 * refill the Free List.
1623 if (q->offset >= 0 && fl_cap(&rxq->fl) - rxq->fl.avail >= 64)
1624 __refill_fl(q->adapter, &rxq->fl);
1626 return budget - budget_left;
1629 int cxgbe_poll(struct sge_rspq *q, struct rte_mbuf **rx_pkts,
1630 unsigned int budget, unsigned int *work_done)
1632 struct sge_eth_rxq *rxq = container_of(q, struct sge_eth_rxq, rspq);
1633 unsigned int cidx_inc;
1634 unsigned int params;
1637 if (unlikely(rxq->flags & IQ_STOPPED)) {
1642 *work_done = process_responses(q, budget, rx_pkts);
1645 cidx_inc = R_IDXDIFF(q, gts_idx);
1647 if (q->offset >= 0 && fl_cap(&rxq->fl) - rxq->fl.avail >= 64)
1648 __refill_fl(q->adapter, &rxq->fl);
1650 params = q->intr_params;
1651 q->next_intr_params = params;
1652 val = V_CIDXINC(cidx_inc) | V_SEINTARM(params);
1654 if (unlikely(!q->bar2_addr)) {
1655 u32 reg = is_pf4(q->adapter) ? MYPF_REG(A_SGE_PF_GTS) :
1656 T4VF_SGE_BASE_ADDR +
1659 t4_write_reg(q->adapter, reg,
1660 val | V_INGRESSQID((u32)q->cntxt_id));
1662 writel(val | V_INGRESSQID(q->bar2_qid),
1663 (void *)((uintptr_t)q->bar2_addr + SGE_UDB_GTS));
1664 /* This Write memory Barrier will force the
1665 * write to the User Doorbell area to be
1670 q->gts_idx = q->cidx;
1676 * bar2_address - return the BAR2 address for an SGE Queue's Registers
1677 * @adapter: the adapter
1678 * @qid: the SGE Queue ID
1679 * @qtype: the SGE Queue Type (Egress or Ingress)
1680 * @pbar2_qid: BAR2 Queue ID or 0 for Queue ID inferred SGE Queues
1682 * Returns the BAR2 address for the SGE Queue Registers associated with
1683 * @qid. If BAR2 SGE Registers aren't available, returns NULL. Also
1684 * returns the BAR2 Queue ID to be used with writes to the BAR2 SGE
1685 * Queue Registers. If the BAR2 Queue ID is 0, then "Inferred Queue ID"
1686 * Registers are supported (e.g. the Write Combining Doorbell Buffer).
1688 static void __iomem *bar2_address(struct adapter *adapter, unsigned int qid,
1689 enum t4_bar2_qtype qtype,
1690 unsigned int *pbar2_qid)
1695 ret = t4_bar2_sge_qregs(adapter, qid, qtype, &bar2_qoffset, pbar2_qid);
1699 return adapter->bar2 + bar2_qoffset;
1702 int t4_sge_eth_rxq_start(struct adapter *adap, struct sge_eth_rxq *rxq)
1704 unsigned int fl_id = rxq->fl.size ? rxq->fl.cntxt_id : 0xffff;
1706 rxq->flags &= ~IQ_STOPPED;
1707 return t4_iq_start_stop(adap, adap->mbox, true, adap->pf, 0,
1708 rxq->rspq.cntxt_id, fl_id, 0xffff);
1711 int t4_sge_eth_rxq_stop(struct adapter *adap, struct sge_eth_rxq *rxq)
1713 unsigned int fl_id = rxq->fl.size ? rxq->fl.cntxt_id : 0xffff;
1715 rxq->flags |= IQ_STOPPED;
1716 return t4_iq_start_stop(adap, adap->mbox, false, adap->pf, 0,
1717 rxq->rspq.cntxt_id, fl_id, 0xffff);
1720 static int cxgbe_sge_fl_buf_size_index(const struct adapter *adap,
1721 struct rte_mempool *mp)
1723 int fl_buf_size, size, delta, min_delta = INT_MAX;
1724 unsigned int i, match = UINT_MAX;
1725 const struct sge *s = &adap->sge;
1727 size = rte_pktmbuf_data_room_size(mp) - RTE_PKTMBUF_HEADROOM;
1728 for (i = 0; i < RTE_DIM(s->fl_buffer_size); i++) {
1729 fl_buf_size = s->fl_buffer_size[i];
1730 if (fl_buf_size > size || fl_buf_size == 0)
1733 delta = size - fl_buf_size;
1736 if (delta < min_delta) {
1742 if (match == UINT_MAX) {
1744 "Could not find valid buffer size for mbuf data room: %d\n",
1753 * @intr_idx: MSI/MSI-X vector if >=0, -(absolute qid + 1) if < 0
1754 * @cong: < 0 -> no congestion feedback, >= 0 -> congestion channel map
1756 int t4_sge_alloc_rxq(struct adapter *adap, struct sge_rspq *iq, bool fwevtq,
1757 struct rte_eth_dev *eth_dev, int intr_idx,
1758 struct sge_fl *fl, rspq_handler_t hnd, int cong,
1759 struct rte_mempool *mp, int queue_id, int socket_id)
1761 struct port_info *pi = eth_dev->data->dev_private;
1762 u8 pciechan, fl_buf_size_idx = 0;
1763 struct sge *s = &adap->sge;
1764 unsigned int nb_refill;
1769 ret = cxgbe_sge_fl_buf_size_index(adap, mp);
1773 fl_buf_size_idx = ret;
1776 /* Size needs to be multiple of 16, including status entry. */
1777 iq->size = cxgbe_roundup(iq->size, 16);
1779 iq->desc = alloc_ring(eth_dev, fwevtq ? "fwq_ring" : "rx_ring",
1780 queue_id, socket_id, iq->size, iq->iqe_len,
1781 0, 0, &iq->phys_addr, NULL);
1785 memset(&c, 0, sizeof(c));
1786 c.op_to_vfn = htonl(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST |
1787 F_FW_CMD_WRITE | F_FW_CMD_EXEC);
1790 pciechan = pi->tx_chan;
1791 c.op_to_vfn |= htonl(V_FW_IQ_CMD_PFN(adap->pf) |
1792 V_FW_IQ_CMD_VFN(0));
1794 c.iqns_to_fl0congen =
1795 htonl(F_FW_IQ_CMD_IQFLINTCONGEN |
1796 V_FW_IQ_CMD_IQTYPE(cong ?
1798 FW_IQ_IQTYPE_OFLD) |
1801 pciechan = pi->port_id;
1804 c.alloc_to_len16 = htonl(F_FW_IQ_CMD_ALLOC | F_FW_IQ_CMD_IQSTART |
1806 c.type_to_iqandstindex =
1807 htonl(V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) |
1808 V_FW_IQ_CMD_IQASYNCH(fwevtq) |
1809 V_FW_IQ_CMD_VIID(pi->viid) |
1810 V_FW_IQ_CMD_IQANDST(intr_idx < 0) |
1811 V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_STATUS_PAGE) |
1812 V_FW_IQ_CMD_IQANDSTINDEX(intr_idx >= 0 ? intr_idx :
1814 c.iqdroprss_to_iqesize =
1815 htons(V_FW_IQ_CMD_IQPCIECH(pciechan) |
1816 F_FW_IQ_CMD_IQGTSMODE |
1817 V_FW_IQ_CMD_IQINTCNTTHRESH(iq->pktcnt_idx) |
1818 V_FW_IQ_CMD_IQESIZE(ilog2(iq->iqe_len) - 4));
1819 c.iqsize = htons(iq->size);
1820 c.iqaddr = cpu_to_be64(iq->phys_addr);
1823 unsigned int chip_ver = CHELSIO_CHIP_VERSION(adap->params.chip);
1826 * Allocate the ring for the hardware free list (with space
1827 * for its status page) along with the associated software
1828 * descriptor ring. The free list size needs to be a multiple
1829 * of the Egress Queue Unit and at least 2 Egress Units larger
1830 * than the SGE's Egress Congestion Threshold
1831 * (fl_starve_thres - 1).
1833 if (fl->size < s->fl_starve_thres - 1 + 2 * 8)
1834 fl->size = s->fl_starve_thres - 1 + 2 * 8;
1835 fl->size = cxgbe_roundup(fl->size, 8);
1837 fl->desc = alloc_ring(eth_dev, "fl_ring", queue_id, socket_id,
1838 fl->size, sizeof(__be64), s->stat_len,
1839 sizeof(struct rx_sw_desc),
1840 &fl->addr, &fl->sdesc);
1846 flsz = fl->size / 8 + s->stat_len / sizeof(struct tx_desc);
1847 c.iqns_to_fl0congen |=
1848 htonl(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE) |
1849 F_FW_IQ_CMD_FL0FETCHRO | F_FW_IQ_CMD_FL0DATARO);
1850 if (is_pf4(adap) && cong >= 0)
1851 c.iqns_to_fl0congen |=
1852 htonl(V_FW_IQ_CMD_FL0CNGCHMAP(cong) |
1853 F_FW_IQ_CMD_FL0CONGCIF |
1854 F_FW_IQ_CMD_FL0CONGEN);
1856 /* In T6, for egress queue type FL there is internal overhead
1857 * of 16B for header going into FLM module.
1858 * Hence maximum allowed burst size will be 448 bytes.
1860 c.fl0dcaen_to_fl0cidxfthresh =
1861 htons(V_FW_IQ_CMD_FL0FBMIN(chip_ver <= CHELSIO_T5 ?
1862 X_FETCHBURSTMIN_128B :
1863 X_FETCHBURSTMIN_64B) |
1864 V_FW_IQ_CMD_FL0FBMAX(chip_ver <= CHELSIO_T5 ?
1865 X_FETCHBURSTMAX_512B :
1866 X_FETCHBURSTMAX_256B));
1867 c.fl0size = htons(flsz);
1868 c.fl0addr = cpu_to_be64(fl->addr);
1872 ret = t4_wr_mbox(adap, adap->mbox, &c, sizeof(c), &c);
1874 ret = t4vf_wr_mbox(adap, &c, sizeof(c), &c);
1878 iq->cur_desc = iq->desc;
1882 iq->next_intr_params = iq->intr_params;
1883 iq->cntxt_id = ntohs(c.iqid);
1884 iq->abs_id = ntohs(c.physiqid);
1885 iq->bar2_addr = bar2_address(adap, iq->cntxt_id, T4_BAR2_QTYPE_INGRESS,
1887 iq->size--; /* subtract status entry */
1888 iq->stat = (void *)&iq->desc[iq->size * 8];
1889 iq->eth_dev = eth_dev;
1891 iq->port_id = eth_dev->data->port_id;
1894 /* set offset to -1 to distinguish ingress queues without FL */
1895 iq->offset = fl ? 0 : -1;
1898 fl->cntxt_id = ntohs(c.fl0id);
1903 fl->alloc_failed = 0;
1904 fl->fl_buf_size_idx = fl_buf_size_idx;
1907 * Note, we must initialize the BAR2 Free List User Doorbell
1908 * information before refilling the Free List!
1910 fl->bar2_addr = bar2_address(adap, fl->cntxt_id,
1911 T4_BAR2_QTYPE_EGRESS,
1914 nb_refill = refill_fl(adap, fl, fl_cap(fl));
1915 if (nb_refill != fl_cap(fl)) {
1917 dev_err(adap, "%s: mbuf alloc failed with error: %d\n",
1924 * For T5 and later we attempt to set up the Congestion Manager values
1925 * of the new RX Ethernet Queue. This should really be handled by
1926 * firmware because it's more complex than any host driver wants to
1927 * get involved with and it's different per chip and this is almost
1928 * certainly wrong. Formware would be wrong as well, but it would be
1929 * a lot easier to fix in one place ... For now we do something very
1930 * simple (and hopefully less wrong).
1932 if (is_pf4(adap) && !is_t4(adap->params.chip) && cong >= 0) {
1933 u8 cng_ch_bits_log = adap->params.arch.cng_ch_bits_log;
1934 u32 param, val, ch_map = 0;
1937 param = (V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
1938 V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
1939 V_FW_PARAMS_PARAM_YZ(iq->cntxt_id));
1941 val = V_CONMCTXT_CNGTPMODE(X_CONMCTXT_CNGTPMODE_QUEUE);
1943 val = V_CONMCTXT_CNGTPMODE(
1944 X_CONMCTXT_CNGTPMODE_CHANNEL);
1945 for (i = 0; i < 4; i++) {
1946 if (cong & (1 << i))
1947 ch_map |= 1 << (i << cng_ch_bits_log);
1949 val |= V_CONMCTXT_CNGCHMAP(ch_map);
1951 ret = t4_set_params(adap, adap->mbox, adap->pf, 0, 1,
1954 dev_warn(adap->pdev_dev, "Failed to set Congestion Manager Context for Ingress Queue %d: %d\n",
1955 iq->cntxt_id, -ret);
1961 t4_iq_free(adap, adap->mbox, adap->pf, 0, FW_IQ_TYPE_FL_INT_CAP,
1962 iq->cntxt_id, fl->cntxt_id, 0xffff);
1969 if (fl && fl->desc) {
1970 rte_free(fl->sdesc);
1978 static void init_txq(struct adapter *adap, struct sge_txq *q, unsigned int id,
1979 unsigned int abs_id)
1983 q->bar2_addr = bar2_address(adap, q->cntxt_id, T4_BAR2_QTYPE_EGRESS,
1990 q->coalesce.idx = 0;
1991 q->coalesce.len = 0;
1992 q->coalesce.flits = 0;
1993 q->last_coal_idx = 0;
1995 q->stat = (void *)&q->desc[q->size];
1998 int t4_sge_eth_txq_start(struct sge_eth_txq *txq)
2001 * TODO: For flow-control, queue may be stopped waiting to reclaim
2003 * Ensure queue is in EQ_STOPPED state before starting it.
2005 if (!(txq->flags & EQ_STOPPED))
2008 txq->flags &= ~EQ_STOPPED;
2013 int t4_sge_eth_txq_stop(struct sge_eth_txq *txq)
2015 txq->flags |= EQ_STOPPED;
2020 int t4_sge_alloc_eth_txq(struct adapter *adap, struct sge_eth_txq *txq,
2021 struct rte_eth_dev *eth_dev, uint16_t queue_id,
2022 unsigned int iqid, int socket_id)
2025 struct fw_eq_eth_cmd c;
2026 struct sge *s = &adap->sge;
2027 struct port_info *pi = eth_dev->data->dev_private;
2030 /* Add status entries */
2031 nentries = txq->q.size + s->stat_len / sizeof(struct tx_desc);
2033 txq->q.desc = alloc_ring(eth_dev, "tx_ring", queue_id, socket_id,
2034 txq->q.size, sizeof(struct tx_desc),
2035 s->stat_len, sizeof(struct tx_sw_desc),
2036 &txq->q.phys_addr, &txq->q.sdesc);
2040 memset(&c, 0, sizeof(c));
2041 c.op_to_vfn = htonl(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST |
2042 F_FW_CMD_WRITE | F_FW_CMD_EXEC);
2044 pciechan = pi->tx_chan;
2045 c.op_to_vfn |= htonl(V_FW_EQ_ETH_CMD_PFN(adap->pf) |
2046 V_FW_EQ_ETH_CMD_VFN(0));
2048 pciechan = pi->port_id;
2051 c.alloc_to_len16 = htonl(F_FW_EQ_ETH_CMD_ALLOC |
2052 F_FW_EQ_ETH_CMD_EQSTART | (sizeof(c) / 16));
2053 c.autoequiqe_to_viid = htonl(F_FW_EQ_ETH_CMD_AUTOEQUEQE |
2054 V_FW_EQ_ETH_CMD_VIID(pi->viid));
2055 c.fetchszm_to_iqid =
2056 htonl(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
2057 V_FW_EQ_ETH_CMD_PCIECHN(pciechan) |
2058 F_FW_EQ_ETH_CMD_FETCHRO | V_FW_EQ_ETH_CMD_IQID(iqid));
2060 htonl(V_FW_EQ_ETH_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
2061 V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
2062 V_FW_EQ_ETH_CMD_EQSIZE(nentries));
2063 c.eqaddr = rte_cpu_to_be_64(txq->q.phys_addr);
2066 ret = t4_wr_mbox(adap, adap->mbox, &c, sizeof(c), &c);
2068 ret = t4vf_wr_mbox(adap, &c, sizeof(c), &c);
2070 rte_free(txq->q.sdesc);
2071 txq->q.sdesc = NULL;
2076 init_txq(adap, &txq->q, G_FW_EQ_ETH_CMD_EQID(ntohl(c.eqid_pkd)),
2077 G_FW_EQ_ETH_CMD_PHYSEQID(ntohl(c.physeqid_pkd)));
2079 txq->stats.pkts = 0;
2080 txq->stats.tx_cso = 0;
2081 txq->stats.coal_wr = 0;
2082 txq->stats.vlan_ins = 0;
2083 txq->stats.tx_bytes = 0;
2084 txq->stats.coal_pkts = 0;
2085 txq->stats.mapping_err = 0;
2086 txq->flags |= EQ_STOPPED;
2087 txq->eth_dev = eth_dev;
2088 txq->data = eth_dev->data;
2089 t4_os_lock_init(&txq->txq_lock);
2093 int t4_sge_alloc_ctrl_txq(struct adapter *adap, struct sge_ctrl_txq *txq,
2094 struct rte_eth_dev *eth_dev, uint16_t queue_id,
2095 unsigned int iqid, int socket_id)
2098 struct fw_eq_ctrl_cmd c;
2099 struct sge *s = &adap->sge;
2100 struct port_info *pi = eth_dev->data->dev_private;
2102 /* Add status entries */
2103 nentries = txq->q.size + s->stat_len / sizeof(struct tx_desc);
2105 txq->q.desc = alloc_ring(eth_dev, "ctrl_tx_ring", queue_id,
2106 socket_id, txq->q.size, sizeof(struct tx_desc),
2107 0, 0, &txq->q.phys_addr, NULL);
2111 memset(&c, 0, sizeof(c));
2112 c.op_to_vfn = htonl(V_FW_CMD_OP(FW_EQ_CTRL_CMD) | F_FW_CMD_REQUEST |
2113 F_FW_CMD_WRITE | F_FW_CMD_EXEC |
2114 V_FW_EQ_CTRL_CMD_PFN(adap->pf) |
2115 V_FW_EQ_CTRL_CMD_VFN(0));
2116 c.alloc_to_len16 = htonl(F_FW_EQ_CTRL_CMD_ALLOC |
2117 F_FW_EQ_CTRL_CMD_EQSTART | (sizeof(c) / 16));
2118 c.cmpliqid_eqid = htonl(V_FW_EQ_CTRL_CMD_CMPLIQID(0));
2119 c.physeqid_pkd = htonl(0);
2120 c.fetchszm_to_iqid =
2121 htonl(V_FW_EQ_CTRL_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
2122 V_FW_EQ_CTRL_CMD_PCIECHN(pi->tx_chan) |
2123 F_FW_EQ_CTRL_CMD_FETCHRO | V_FW_EQ_CTRL_CMD_IQID(iqid));
2125 htonl(V_FW_EQ_CTRL_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
2126 V_FW_EQ_CTRL_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
2127 V_FW_EQ_CTRL_CMD_EQSIZE(nentries));
2128 c.eqaddr = cpu_to_be64(txq->q.phys_addr);
2130 ret = t4_wr_mbox(adap, adap->mbox, &c, sizeof(c), &c);
2136 init_txq(adap, &txq->q, G_FW_EQ_CTRL_CMD_EQID(ntohl(c.cmpliqid_eqid)),
2137 G_FW_EQ_CTRL_CMD_EQID(ntohl(c. physeqid_pkd)));
2138 txq->adapter = adap;
2143 static void free_txq(struct sge_txq *q)
2150 static void free_rspq_fl(struct adapter *adap, struct sge_rspq *rq,
2153 unsigned int fl_id = fl ? fl->cntxt_id : 0xffff;
2155 t4_iq_free(adap, adap->mbox, adap->pf, 0, FW_IQ_TYPE_FL_INT_CAP,
2156 rq->cntxt_id, fl_id, 0xffff);
2162 free_rx_bufs(fl, fl->avail);
2163 rte_free(fl->sdesc);
2171 * Clear all queues of the port
2173 * Note: This function must only be called after rx and tx path
2174 * of the port have been disabled.
2176 void t4_sge_eth_clear_queues(struct port_info *pi)
2178 struct adapter *adap = pi->adapter;
2179 struct sge_eth_rxq *rxq;
2180 struct sge_eth_txq *txq;
2183 rxq = &adap->sge.ethrxq[pi->first_rxqset];
2184 for (i = 0; i < pi->n_rx_qsets; i++, rxq++) {
2186 t4_sge_eth_rxq_stop(adap, rxq);
2189 txq = &adap->sge.ethtxq[pi->first_txqset];
2190 for (i = 0; i < pi->n_tx_qsets; i++, txq++) {
2192 struct sge_txq *q = &txq->q;
2194 t4_sge_eth_txq_stop(txq);
2195 reclaim_completed_tx(q);
2196 free_tx_desc(q, q->size);
2197 q->equeidx = q->pidx;
2202 void t4_sge_eth_rxq_release(struct adapter *adap, struct sge_eth_rxq *rxq)
2204 if (rxq->rspq.desc) {
2205 t4_sge_eth_rxq_stop(adap, rxq);
2206 free_rspq_fl(adap, &rxq->rspq, rxq->fl.size ? &rxq->fl : NULL);
2210 void t4_sge_eth_txq_release(struct adapter *adap, struct sge_eth_txq *txq)
2213 t4_sge_eth_txq_stop(txq);
2214 reclaim_completed_tx(&txq->q);
2215 t4_eth_eq_free(adap, adap->mbox, adap->pf, 0, txq->q.cntxt_id);
2216 free_tx_desc(&txq->q, txq->q.size);
2217 rte_free(txq->q.sdesc);
2222 void t4_sge_eth_release_queues(struct port_info *pi)
2224 struct adapter *adap = pi->adapter;
2225 struct sge_eth_rxq *rxq;
2226 struct sge_eth_txq *txq;
2229 rxq = &adap->sge.ethrxq[pi->first_rxqset];
2230 /* clean up Ethernet Tx/Rx queues */
2231 for (i = 0; i < pi->n_rx_qsets; i++, rxq++) {
2232 /* Free only the queues allocated */
2233 if (rxq->rspq.desc) {
2234 t4_sge_eth_rxq_release(adap, rxq);
2235 rte_eth_dma_zone_free(rxq->rspq.eth_dev, "fl_ring", i);
2236 rte_eth_dma_zone_free(rxq->rspq.eth_dev, "rx_ring", i);
2237 rxq->rspq.eth_dev = NULL;
2241 txq = &adap->sge.ethtxq[pi->first_txqset];
2242 for (i = 0; i < pi->n_tx_qsets; i++, txq++) {
2243 /* Free only the queues allocated */
2245 t4_sge_eth_txq_release(adap, txq);
2246 rte_eth_dma_zone_free(txq->eth_dev, "tx_ring", i);
2247 txq->eth_dev = NULL;
2252 void t4_sge_tx_monitor_start(struct adapter *adap)
2254 rte_eal_alarm_set(50, tx_timer_cb, (void *)adap);
2257 void t4_sge_tx_monitor_stop(struct adapter *adap)
2259 rte_eal_alarm_cancel(tx_timer_cb, (void *)adap);
2263 * t4_free_sge_resources - free SGE resources
2264 * @adap: the adapter
2266 * Frees resources used by the SGE queue sets.
2268 void t4_free_sge_resources(struct adapter *adap)
2272 /* clean up control Tx queues */
2273 for (i = 0; i < ARRAY_SIZE(adap->sge.ctrlq); i++) {
2274 struct sge_ctrl_txq *cq = &adap->sge.ctrlq[i];
2277 reclaim_completed_tx_imm(&cq->q);
2278 t4_ctrl_eq_free(adap, adap->mbox, adap->pf, 0,
2280 rte_eth_dma_zone_free(adap->eth_dev, "ctrl_tx_ring", i);
2281 rte_mempool_free(cq->mb_pool);
2286 /* clean up firmware event queue */
2287 if (adap->sge.fw_evtq.desc) {
2288 free_rspq_fl(adap, &adap->sge.fw_evtq, NULL);
2289 rte_eth_dma_zone_free(adap->eth_dev, "fwq_ring", 0);
2294 * t4_sge_init - initialize SGE
2295 * @adap: the adapter
2297 * Performs SGE initialization needed every time after a chip reset.
2298 * We do not initialize any of the queues here, instead the driver
2299 * top-level must request those individually.
2301 * Called in two different modes:
2303 * 1. Perform actual hardware initialization and record hard-coded
2304 * parameters which were used. This gets used when we're the
2305 * Master PF and the Firmware Configuration File support didn't
2306 * work for some reason.
2308 * 2. We're not the Master PF or initialization was performed with
2309 * a Firmware Configuration File. In this case we need to grab
2310 * any of the SGE operating parameters that we need to have in
2311 * order to do our job and make sure we can live with them ...
2313 static int t4_sge_init_soft(struct adapter *adap)
2315 u32 timer_value_0_and_1, timer_value_2_and_3, timer_value_4_and_5;
2316 struct sge *s = &adap->sge;
2317 u32 ingress_rx_threshold;
2321 * Verify that CPL messages are going to the Ingress Queue for
2322 * process_responses() and that only packet data is going to the
2325 if ((t4_read_reg(adap, A_SGE_CONTROL) & F_RXPKTCPLMODE) !=
2326 V_RXPKTCPLMODE(X_RXPKTCPLMODE_SPLIT)) {
2327 dev_err(adap, "bad SGE CPL MODE\n");
2332 * Read the Host Buffer Register Array indices that we want to
2335 for (i = 0; i < RTE_DIM(s->fl_buffer_size); i++)
2336 s->fl_buffer_size[i] = t4_read_reg(adap,
2337 A_SGE_FL_BUFFER_SIZE0 +
2341 * Retrieve our RX interrupt holdoff timer values and counter
2342 * threshold values from the SGE parameters.
2344 timer_value_0_and_1 = t4_read_reg(adap, A_SGE_TIMER_VALUE_0_AND_1);
2345 timer_value_2_and_3 = t4_read_reg(adap, A_SGE_TIMER_VALUE_2_AND_3);
2346 timer_value_4_and_5 = t4_read_reg(adap, A_SGE_TIMER_VALUE_4_AND_5);
2347 s->timer_val[0] = core_ticks_to_us(adap,
2348 G_TIMERVALUE0(timer_value_0_and_1));
2349 s->timer_val[1] = core_ticks_to_us(adap,
2350 G_TIMERVALUE1(timer_value_0_and_1));
2351 s->timer_val[2] = core_ticks_to_us(adap,
2352 G_TIMERVALUE2(timer_value_2_and_3));
2353 s->timer_val[3] = core_ticks_to_us(adap,
2354 G_TIMERVALUE3(timer_value_2_and_3));
2355 s->timer_val[4] = core_ticks_to_us(adap,
2356 G_TIMERVALUE4(timer_value_4_and_5));
2357 s->timer_val[5] = core_ticks_to_us(adap,
2358 G_TIMERVALUE5(timer_value_4_and_5));
2360 ingress_rx_threshold = t4_read_reg(adap, A_SGE_INGRESS_RX_THRESHOLD);
2361 s->counter_val[0] = G_THRESHOLD_0(ingress_rx_threshold);
2362 s->counter_val[1] = G_THRESHOLD_1(ingress_rx_threshold);
2363 s->counter_val[2] = G_THRESHOLD_2(ingress_rx_threshold);
2364 s->counter_val[3] = G_THRESHOLD_3(ingress_rx_threshold);
2369 int t4_sge_init(struct adapter *adap)
2371 struct sge *s = &adap->sge;
2372 u32 sge_control, sge_conm_ctrl;
2373 int ret, egress_threshold;
2376 * Ingress Padding Boundary and Egress Status Page Size are set up by
2377 * t4_fixup_host_params().
2379 sge_control = t4_read_reg(adap, A_SGE_CONTROL);
2380 s->pktshift = G_PKTSHIFT(sge_control);
2381 s->stat_len = (sge_control & F_EGRSTATUSPAGESIZE) ? 128 : 64;
2382 ret = t4_sge_init_soft(adap);
2384 dev_err(adap, "%s: t4_sge_init_soft failed, error %d\n",
2390 * A FL with <= fl_starve_thres buffers is starving and a periodic
2391 * timer will attempt to refill it. This needs to be larger than the
2392 * SGE's Egress Congestion Threshold. If it isn't, then we can get
2393 * stuck waiting for new packets while the SGE is waiting for us to
2394 * give it more Free List entries. (Note that the SGE's Egress
2395 * Congestion Threshold is in units of 2 Free List pointers.) For T4,
2396 * there was only a single field to control this. For T5 there's the
2397 * original field which now only applies to Unpacked Mode Free List
2398 * buffers and a new field which only applies to Packed Mode Free List
2401 sge_conm_ctrl = t4_read_reg(adap, A_SGE_CONM_CTRL);
2402 if (is_t4(adap->params.chip) || adap->use_unpacked_mode)
2403 egress_threshold = G_EGRTHRESHOLD(sge_conm_ctrl);
2405 egress_threshold = G_EGRTHRESHOLDPACKING(sge_conm_ctrl);
2406 s->fl_starve_thres = 2 * egress_threshold + 1;
2411 int t4vf_sge_init(struct adapter *adap)
2413 struct sge_params *sge_params = &adap->params.sge;
2414 u32 sge_ingress_queues_per_page;
2415 u32 sge_egress_queues_per_page;
2416 u32 sge_ingress_rx_threshold;
2417 u32 sge_timer_value_0_and_1;
2418 u32 sge_timer_value_2_and_3;
2419 u32 sge_timer_value_4_and_5;
2420 u32 sge_congestion_control;
2421 struct sge *s = &adap->sge;
2422 unsigned int s_hps, s_qpp;
2423 u32 sge_host_page_size;
2424 u32 params[7], vals[7];
2429 /* query basic params from fw */
2430 params[0] = (V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_REG) |
2431 V_FW_PARAMS_PARAM_XYZ(A_SGE_CONTROL));
2432 params[1] = (V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_REG) |
2433 V_FW_PARAMS_PARAM_XYZ(A_SGE_HOST_PAGE_SIZE));
2434 params[2] = (V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_REG) |
2435 V_FW_PARAMS_PARAM_XYZ(A_SGE_TIMER_VALUE_0_AND_1));
2436 params[3] = (V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_REG) |
2437 V_FW_PARAMS_PARAM_XYZ(A_SGE_TIMER_VALUE_2_AND_3));
2438 params[4] = (V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_REG) |
2439 V_FW_PARAMS_PARAM_XYZ(A_SGE_TIMER_VALUE_4_AND_5));
2440 v = t4vf_query_params(adap, 7, params, vals);
2441 if (v != FW_SUCCESS)
2444 sge_control = vals[0];
2445 sge_host_page_size = vals[1];
2446 sge_timer_value_0_and_1 = vals[2];
2447 sge_timer_value_2_and_3 = vals[3];
2448 sge_timer_value_4_and_5 = vals[4];
2450 for (i = 0; i < RTE_DIM(s->fl_buffer_size); i++) {
2451 params[0] = (V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_REG) |
2452 V_FW_PARAMS_PARAM_XYZ(A_SGE_FL_BUFFER_SIZE0 +
2454 v = t4vf_query_params(adap, 1, params, vals);
2455 if (v != FW_SUCCESS)
2458 s->fl_buffer_size[i] = vals[0];
2462 * Start by vetting the basic SGE parameters which have been set up by
2463 * the Physical Function Driver.
2466 if ((sge_control & F_RXPKTCPLMODE) !=
2467 V_RXPKTCPLMODE(X_RXPKTCPLMODE_SPLIT)) {
2468 dev_err(adapter->pdev_dev, "bad SGE CPL MODE\n");
2472 params[0] = (V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_REG) |
2473 V_FW_PARAMS_PARAM_XYZ(A_SGE_INGRESS_RX_THRESHOLD));
2474 params[1] = (V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_REG) |
2475 V_FW_PARAMS_PARAM_XYZ(A_SGE_CONM_CTRL));
2476 v = t4vf_query_params(adap, 2, params, vals);
2477 if (v != FW_SUCCESS)
2479 sge_ingress_rx_threshold = vals[0];
2480 sge_congestion_control = vals[1];
2481 params[0] = (V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_REG) |
2482 V_FW_PARAMS_PARAM_XYZ(A_SGE_EGRESS_QUEUES_PER_PAGE_VF));
2483 params[1] = (V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_REG) |
2484 V_FW_PARAMS_PARAM_XYZ(A_SGE_INGRESS_QUEUES_PER_PAGE_VF));
2485 v = t4vf_query_params(adap, 2, params, vals);
2486 if (v != FW_SUCCESS) {
2487 dev_warn(adap, "Unable to get VF SGE Queues/Page; "
2488 "probably old firmware.\n");
2491 sge_egress_queues_per_page = vals[0];
2492 sge_ingress_queues_per_page = vals[1];
2495 * We need the Queues/Page for our VF. This is based on the
2496 * PF from which we're instantiated and is indexed in the
2497 * register we just read.
2499 s_hps = (S_HOSTPAGESIZEPF0 +
2500 (S_HOSTPAGESIZEPF1 - S_HOSTPAGESIZEPF0) * adap->pf);
2502 ((sge_host_page_size >> s_hps) & M_HOSTPAGESIZEPF0);
2504 s_qpp = (S_QUEUESPERPAGEPF0 +
2505 (S_QUEUESPERPAGEPF1 - S_QUEUESPERPAGEPF0) * adap->pf);
2506 sge_params->eq_qpp =
2507 ((sge_egress_queues_per_page >> s_qpp)
2508 & M_QUEUESPERPAGEPF0);
2509 sge_params->iq_qpp =
2510 ((sge_ingress_queues_per_page >> s_qpp)
2511 & M_QUEUESPERPAGEPF0);
2514 * Now translate the queried parameters into our internal forms.
2516 s->stat_len = ((sge_control & F_EGRSTATUSPAGESIZE)
2518 s->pktshift = G_PKTSHIFT(sge_control);
2521 * A FL with <= fl_starve_thres buffers is starving and a periodic
2522 * timer will attempt to refill it. This needs to be larger than the
2523 * SGE's Egress Congestion Threshold. If it isn't, then we can get
2524 * stuck waiting for new packets while the SGE is waiting for us to
2525 * give it more Free List entries. (Note that the SGE's Egress
2526 * Congestion Threshold is in units of 2 Free List pointers.)
2528 switch (CHELSIO_CHIP_VERSION(adap->params.chip)) {
2530 s->fl_starve_thres =
2531 G_EGRTHRESHOLDPACKING(sge_congestion_control);
2535 s->fl_starve_thres =
2536 G_T6_EGRTHRESHOLDPACKING(sge_congestion_control);
2539 s->fl_starve_thres = s->fl_starve_thres * 2 + 1;
2542 * Save RX interrupt holdoff timer values and counter
2543 * threshold values from the SGE parameters.
2545 s->timer_val[0] = core_ticks_to_us(adap,
2546 G_TIMERVALUE0(sge_timer_value_0_and_1));
2547 s->timer_val[1] = core_ticks_to_us(adap,
2548 G_TIMERVALUE1(sge_timer_value_0_and_1));
2549 s->timer_val[2] = core_ticks_to_us(adap,
2550 G_TIMERVALUE2(sge_timer_value_2_and_3));
2551 s->timer_val[3] = core_ticks_to_us(adap,
2552 G_TIMERVALUE3(sge_timer_value_2_and_3));
2553 s->timer_val[4] = core_ticks_to_us(adap,
2554 G_TIMERVALUE4(sge_timer_value_4_and_5));
2555 s->timer_val[5] = core_ticks_to_us(adap,
2556 G_TIMERVALUE5(sge_timer_value_4_and_5));
2557 s->counter_val[0] = G_THRESHOLD_0(sge_ingress_rx_threshold);
2558 s->counter_val[1] = G_THRESHOLD_1(sge_ingress_rx_threshold);
2559 s->counter_val[2] = G_THRESHOLD_2(sge_ingress_rx_threshold);
2560 s->counter_val[3] = G_THRESHOLD_3(sge_ingress_rx_threshold);