2 * Copyright (c) 2016-2017 Alexander Motin <mav@FreeBSD.org>
3 * Copyright (C) 2013 Intel Corporation
4 * Copyright (C) 2015 EMC Corporation
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
10 * 1. Redistributions of source code must retain the above copyright
11 * notice, this list of conditions and the following disclaimer.
12 * 2. Redistributions in binary form must reproduce the above copyright
13 * notice, this list of conditions and the following disclaimer in the
14 * documentation and/or other materials provided with the distribution.
16 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
17 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
20 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30 * The Non-Transparent Bridge (NTB) is a device that allows you to connect
31 * two or more systems using a PCI-e links, providing remote memory access.
33 * This module contains a transport for sending and receiving messages by
34 * writing to remote memory window(s) provided by underlying NTB device.
36 * NOTE: Much of the code in this module is shared with Linux. Any patches may
37 * be picked up and redistributed in Linux with a dual GPL/BSD license.
40 #include <sys/cdefs.h>
41 __FBSDID("$FreeBSD$");
43 #include <sys/param.h>
44 #include <sys/kernel.h>
45 #include <sys/systm.h>
48 #include <sys/limits.h>
50 #include <sys/malloc.h>
52 #include <sys/module.h>
53 #include <sys/mutex.h>
54 #include <sys/queue.h>
55 #include <sys/sysctl.h>
56 #include <sys/taskqueue.h>
61 #include <machine/bus.h>
64 #include "ntb_transport.h"
66 #define KTR_NTB KTR_SPARE3
68 #define NTB_TRANSPORT_VERSION 4
70 static SYSCTL_NODE(_hw, OID_AUTO, ntb_transport, CTLFLAG_RW, 0, "ntb_transport");
72 static unsigned g_ntb_transport_debug_level;
73 SYSCTL_UINT(_hw_ntb_transport, OID_AUTO, debug_level, CTLFLAG_RWTUN,
74 &g_ntb_transport_debug_level, 0,
75 "ntb_transport log level -- higher is more verbose");
76 #define ntb_printf(lvl, ...) do { \
77 if ((lvl) <= g_ntb_transport_debug_level) { \
78 printf(__VA_ARGS__); \
82 static unsigned transport_mtu = 0x10000;
84 static uint64_t max_mw_size;
85 SYSCTL_UQUAD(_hw_ntb_transport, OID_AUTO, max_mw_size, CTLFLAG_RDTUN, &max_mw_size, 0,
86 "If enabled (non-zero), limit the size of large memory windows. "
87 "Both sides of the NTB MUST set the same value here.");
89 static unsigned enable_xeon_watchdog;
90 SYSCTL_UINT(_hw_ntb_transport, OID_AUTO, enable_xeon_watchdog, CTLFLAG_RDTUN,
91 &enable_xeon_watchdog, 0, "If non-zero, write a register every second to "
92 "keep a watchdog from tearing down the NTB link");
94 STAILQ_HEAD(ntb_queue_list, ntb_queue_entry);
96 typedef uint32_t ntb_q_idx_t;
98 struct ntb_queue_entry {
99 /* ntb_queue list reference */
100 STAILQ_ENTRY(ntb_queue_entry) entry;
102 /* info on data to be transferred */
108 struct ntb_transport_qp *qp;
109 struct ntb_payload_header *x_hdr;
117 struct ntb_transport_qp {
118 struct ntb_transport_ctx *transport;
124 volatile bool link_is_up;
125 uint8_t qp_num; /* Only 64 QPs are allowed. 0-63 */
127 struct ntb_rx_info *rx_info;
128 struct ntb_rx_info *remote_rx_info;
130 void (*tx_handler)(struct ntb_transport_qp *qp, void *qp_data,
131 void *data, int len);
132 struct ntb_queue_list tx_free_q;
133 struct mtx ntb_tx_free_q_lock;
135 bus_addr_t tx_mw_phys;
136 ntb_q_idx_t tx_index;
137 ntb_q_idx_t tx_max_entry;
138 uint64_t tx_max_frame;
140 void (*rx_handler)(struct ntb_transport_qp *qp, void *qp_data,
141 void *data, int len);
142 struct ntb_queue_list rx_post_q;
143 struct ntb_queue_list rx_pend_q;
144 /* ntb_rx_q_lock: synchronize access to rx_XXXX_q */
145 struct mtx ntb_rx_q_lock;
146 struct task rxc_db_work;
147 struct taskqueue *rxc_tq;
149 ntb_q_idx_t rx_index;
150 ntb_q_idx_t rx_max_entry;
151 uint64_t rx_max_frame;
153 void (*event_handler)(void *data, enum ntb_link_event status);
154 struct callout link_work;
155 struct callout rx_full;
157 uint64_t last_rx_no_buf;
162 uint64_t rx_ring_empty;
163 uint64_t rx_err_no_buf;
164 uint64_t rx_err_oflow;
168 uint64_t tx_ring_full;
169 uint64_t tx_err_no_buf;
174 struct ntb_transport_mw {
175 vm_paddr_t phys_addr;
178 size_t xlat_align_size;
179 bus_addr_t addr_limit;
180 /* Tx buff is off vbase / phys_addr */
184 /* Rx buff is off virt_addr / dma_addr */
189 struct ntb_transport_child {
194 struct ntb_transport_child *next;
197 struct ntb_transport_ctx {
199 struct ntb_transport_child *child;
200 struct ntb_transport_mw *mw_vec;
201 struct ntb_transport_qp *qp_vec;
205 volatile bool link_is_up;
206 enum ntb_speed link_speed;
207 enum ntb_width link_width;
208 struct callout link_work;
209 struct callout link_watchdog;
210 struct task link_cleanup;
214 NTBT_DESC_DONE_FLAG = 1 << 0,
215 NTBT_LINK_DOWN_FLAG = 1 << 1,
218 struct ntb_payload_header {
226 * The order of this enum is part of the remote protocol. Do not
227 * reorder without bumping protocol version (and it's probably best
228 * to keep the protocol in lock-step with the Linux NTB driver.
235 * N.B.: transport_link_work assumes MW1 enums = MW0 + 2.
243 * Some NTB-using hardware have a watchdog to work around NTB hangs; if
244 * a register or doorbell isn't written every few seconds, the link is
245 * torn down. Write an otherwise unused register every few seconds to
246 * work around this watchdog.
248 NTBT_WATCHDOG_SPAD = 15
251 #define QP_TO_MW(nt, qp) ((qp) % nt->mw_count)
252 #define NTB_QP_DEF_NUM_ENTRIES 100
253 #define NTB_LINK_DOWN_TIMEOUT 10
255 static int ntb_transport_probe(device_t dev);
256 static int ntb_transport_attach(device_t dev);
257 static int ntb_transport_detach(device_t dev);
258 static void ntb_transport_init_queue(struct ntb_transport_ctx *nt,
259 unsigned int qp_num);
260 static int ntb_process_tx(struct ntb_transport_qp *qp,
261 struct ntb_queue_entry *entry);
262 static void ntb_transport_rxc_db(void *arg, int pending);
263 static int ntb_process_rxc(struct ntb_transport_qp *qp);
264 static void ntb_memcpy_rx(struct ntb_transport_qp *qp,
265 struct ntb_queue_entry *entry, void *offset);
266 static inline void ntb_rx_copy_callback(struct ntb_transport_qp *qp,
268 static void ntb_complete_rxc(struct ntb_transport_qp *qp);
269 static void ntb_transport_doorbell_callback(void *data, uint32_t vector);
270 static void ntb_transport_event_callback(void *data);
271 static void ntb_transport_link_work(void *arg);
272 static int ntb_set_mw(struct ntb_transport_ctx *, int num_mw, size_t size);
273 static void ntb_free_mw(struct ntb_transport_ctx *nt, int num_mw);
274 static int ntb_transport_setup_qp_mw(struct ntb_transport_ctx *nt,
275 unsigned int qp_num);
276 static void ntb_qp_link_work(void *arg);
277 static void ntb_transport_link_cleanup(struct ntb_transport_ctx *nt);
278 static void ntb_transport_link_cleanup_work(void *, int);
279 static void ntb_qp_link_down(struct ntb_transport_qp *qp);
280 static void ntb_qp_link_down_reset(struct ntb_transport_qp *qp);
281 static void ntb_qp_link_cleanup(struct ntb_transport_qp *qp);
282 static void ntb_send_link_down(struct ntb_transport_qp *qp);
283 static void ntb_list_add(struct mtx *lock, struct ntb_queue_entry *entry,
284 struct ntb_queue_list *list);
285 static struct ntb_queue_entry *ntb_list_rm(struct mtx *lock,
286 struct ntb_queue_list *list);
287 static struct ntb_queue_entry *ntb_list_mv(struct mtx *lock,
288 struct ntb_queue_list *from, struct ntb_queue_list *to);
289 static void xeon_link_watchdog_hb(void *);
291 static const struct ntb_ctx_ops ntb_transport_ops = {
292 .link_event = ntb_transport_event_callback,
293 .db_event = ntb_transport_doorbell_callback,
296 MALLOC_DEFINE(M_NTB_T, "ntb_transport", "ntb transport driver");
299 iowrite32(uint32_t val, void *addr)
302 bus_space_write_4(X86_BUS_SPACE_MEM, 0/* HACK */, (uintptr_t)addr,
306 /* Transport Init and teardown */
309 xeon_link_watchdog_hb(void *arg)
311 struct ntb_transport_ctx *nt;
314 ntb_spad_write(nt->dev, NTBT_WATCHDOG_SPAD, 0);
315 callout_reset(&nt->link_watchdog, 1 * hz, xeon_link_watchdog_hb, nt);
319 ntb_transport_probe(device_t dev)
322 device_set_desc(dev, "NTB Transport");
327 ntb_transport_attach(device_t dev)
329 struct ntb_transport_ctx *nt = device_get_softc(dev);
330 struct ntb_transport_child **cpp = &nt->child;
331 struct ntb_transport_child *nc;
332 struct ntb_transport_mw *mw;
334 int rc, i, db_count, spad_count, qp, qpu, qpo, qpt;
337 char *n, *np, *c, *name;
340 nt->mw_count = ntb_mw_count(dev);
341 spad_count = ntb_spad_count(dev);
342 db_bitmap = ntb_db_valid_mask(dev);
343 db_count = flsll(db_bitmap);
344 KASSERT(db_bitmap == (1 << db_count) - 1,
345 ("Doorbells are not sequential (%jx).\n", db_bitmap));
347 if (nt->mw_count == 0) {
348 device_printf(dev, "At least 1 memory window required.\n");
351 if (spad_count < 6) {
352 device_printf(dev, "At least 6 scratchpads required.\n");
355 if (spad_count < 4 + 2 * nt->mw_count) {
356 nt->mw_count = (spad_count - 4) / 2;
357 device_printf(dev, "Scratchpads enough only for %d "
358 "memory windows.\n", nt->mw_count);
360 if (db_bitmap == 0) {
361 device_printf(dev, "At least one doorbell required.\n");
365 nt->mw_vec = malloc(nt->mw_count * sizeof(*nt->mw_vec), M_NTB_T,
367 for (i = 0; i < nt->mw_count; i++) {
370 rc = ntb_mw_get_range(dev, i, &mw->phys_addr, &mw->vbase,
371 &mw->phys_size, &mw->xlat_align, &mw->xlat_align_size,
378 mw->virt_addr = NULL;
381 rc = ntb_mw_set_wc(dev, i, VM_MEMATTR_WRITE_COMBINING);
383 ntb_printf(0, "Unable to set mw%d caching\n", i);
387 qpo = imin(db_count, nt->mw_count);
390 snprintf(buf, sizeof(buf), "hint.%s.%d.config", device_get_name(dev),
391 device_get_unit(dev));
392 TUNABLE_STR_FETCH(buf, cfg, sizeof(cfg));
395 while ((c = strsep(&n, ",")) != NULL) {
397 name = strsep(&np, ":");
398 if (name != NULL && name[0] == 0)
400 qp = (np && np[0] != 0) ? strtol(np, NULL, 10) : qpo - qpu;
404 if (qp > qpt - qpu) {
405 device_printf(dev, "Not enough resources for config\n");
409 nc = malloc(sizeof(*nc), M_DEVBUF, M_WAITOK | M_ZERO);
413 nc->dev = device_add_child(dev, name, -1);
414 if (nc->dev == NULL) {
415 device_printf(dev, "Can not add child.\n");
418 device_set_ivars(nc->dev, nc);
423 device_printf(dev, "%d \"%s\": queues %d",
426 printf("-%d", qpu + qp - 1);
435 nt->qp_vec = malloc(nt->qp_count * sizeof(*nt->qp_vec), M_NTB_T,
438 for (i = 0; i < nt->qp_count; i++)
439 ntb_transport_init_queue(nt, i);
441 callout_init(&nt->link_work, 0);
442 callout_init(&nt->link_watchdog, 0);
443 TASK_INIT(&nt->link_cleanup, 0, ntb_transport_link_cleanup_work, nt);
444 nt->link_is_up = false;
446 rc = ntb_set_ctx(dev, nt, &ntb_transport_ops);
450 ntb_link_enable(dev, NTB_SPEED_AUTO, NTB_WIDTH_AUTO);
452 if (enable_xeon_watchdog != 0)
453 callout_reset(&nt->link_watchdog, 0, xeon_link_watchdog_hb, nt);
455 bus_generic_attach(dev);
459 free(nt->qp_vec, M_NTB_T);
460 free(nt->mw_vec, M_NTB_T);
465 ntb_transport_detach(device_t dev)
467 struct ntb_transport_ctx *nt = device_get_softc(dev);
468 struct ntb_transport_child **cpp = &nt->child;
469 struct ntb_transport_child *nc;
472 while ((nc = *cpp) != NULL) {
474 error = device_delete_child(dev, nc->dev);
479 KASSERT(nt->qp_bitmap == 0,
480 ("Some queues not freed on detach (%jx)", nt->qp_bitmap));
482 ntb_transport_link_cleanup(nt);
483 taskqueue_drain(taskqueue_swi, &nt->link_cleanup);
484 callout_drain(&nt->link_work);
485 callout_drain(&nt->link_watchdog);
487 ntb_link_disable(dev);
490 for (i = 0; i < nt->mw_count; i++)
493 free(nt->qp_vec, M_NTB_T);
494 free(nt->mw_vec, M_NTB_T);
499 ntb_transport_print_child(device_t dev, device_t child)
501 struct ntb_transport_child *nc = device_get_ivars(child);
504 retval = bus_print_child_header(dev, child);
506 printf(" queue %d", nc->qpoff);
508 printf("-%d", nc->qpoff + nc->qpcnt - 1);
510 retval += printf(" at consumer %d", nc->consumer);
511 retval += bus_print_child_domain(dev, child);
512 retval += bus_print_child_footer(dev, child);
518 ntb_transport_child_location_str(device_t dev, device_t child, char *buf,
521 struct ntb_transport_child *nc = device_get_ivars(child);
523 snprintf(buf, buflen, "consumer=%d", nc->consumer);
528 ntb_transport_queue_count(device_t dev)
530 struct ntb_transport_child *nc = device_get_ivars(dev);
536 ntb_transport_init_queue(struct ntb_transport_ctx *nt, unsigned int qp_num)
538 struct ntb_transport_mw *mw;
539 struct ntb_transport_qp *qp;
541 uint64_t mw_size, qp_offset;
543 unsigned num_qps_mw, mw_num, mw_count;
545 mw_count = nt->mw_count;
546 mw_num = QP_TO_MW(nt, qp_num);
547 mw = &nt->mw_vec[mw_num];
549 qp = &nt->qp_vec[qp_num];
553 qp->client_ready = false;
554 qp->event_handler = NULL;
555 ntb_qp_link_down_reset(qp);
557 if (mw_num < nt->qp_count % mw_count)
558 num_qps_mw = nt->qp_count / mw_count + 1;
560 num_qps_mw = nt->qp_count / mw_count;
562 mw_base = mw->phys_addr;
563 mw_size = mw->phys_size;
565 tx_size = mw_size / num_qps_mw;
566 qp_offset = tx_size * (qp_num / mw_count);
568 qp->tx_mw = mw->vbase + qp_offset;
569 KASSERT(qp->tx_mw != NULL, ("uh oh?"));
571 /* XXX Assumes that a vm_paddr_t is equivalent to bus_addr_t */
572 qp->tx_mw_phys = mw_base + qp_offset;
573 KASSERT(qp->tx_mw_phys != 0, ("uh oh?"));
575 tx_size -= sizeof(struct ntb_rx_info);
576 qp->rx_info = (void *)(qp->tx_mw + tx_size);
578 /* Due to house-keeping, there must be at least 2 buffs */
579 qp->tx_max_frame = qmin(transport_mtu, tx_size / 2);
580 qp->tx_max_entry = tx_size / qp->tx_max_frame;
582 callout_init(&qp->link_work, 0);
583 callout_init(&qp->rx_full, 1);
585 mtx_init(&qp->ntb_rx_q_lock, "ntb rx q", NULL, MTX_SPIN);
586 mtx_init(&qp->ntb_tx_free_q_lock, "ntb tx free q", NULL, MTX_SPIN);
587 mtx_init(&qp->tx_lock, "ntb transport tx", NULL, MTX_DEF);
588 TASK_INIT(&qp->rxc_db_work, 0, ntb_transport_rxc_db, qp);
589 qp->rxc_tq = taskqueue_create("ntbt_rx", M_WAITOK,
590 taskqueue_thread_enqueue, &qp->rxc_tq);
591 taskqueue_start_threads(&qp->rxc_tq, 1, PI_NET, "%s rx%d",
592 device_get_nameunit(nt->dev), qp_num);
594 STAILQ_INIT(&qp->rx_post_q);
595 STAILQ_INIT(&qp->rx_pend_q);
596 STAILQ_INIT(&qp->tx_free_q);
600 ntb_transport_free_queue(struct ntb_transport_qp *qp)
602 struct ntb_transport_ctx *nt = qp->transport;
603 struct ntb_queue_entry *entry;
605 callout_drain(&qp->link_work);
607 ntb_db_set_mask(qp->dev, 1ull << qp->qp_num);
608 taskqueue_drain_all(qp->rxc_tq);
609 taskqueue_free(qp->rxc_tq);
612 qp->rx_handler = NULL;
613 qp->tx_handler = NULL;
614 qp->event_handler = NULL;
616 while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_pend_q)))
617 free(entry, M_NTB_T);
619 while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_post_q)))
620 free(entry, M_NTB_T);
622 while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
623 free(entry, M_NTB_T);
625 nt->qp_bitmap &= ~(1 << qp->qp_num);
629 * ntb_transport_create_queue - Create a new NTB transport layer queue
630 * @rx_handler: receive callback function
631 * @tx_handler: transmit callback function
632 * @event_handler: event callback function
634 * Create a new NTB transport layer queue and provide the queue with a callback
635 * routine for both transmit and receive. The receive callback routine will be
636 * used to pass up data when the transport has received it on the queue. The
637 * transmit callback routine will be called when the transport has completed the
638 * transmission of the data on the queue and the data is ready to be freed.
640 * RETURNS: pointer to newly created ntb_queue, NULL on error.
642 struct ntb_transport_qp *
643 ntb_transport_create_queue(device_t dev, int q,
644 const struct ntb_queue_handlers *handlers, void *data)
646 struct ntb_transport_child *nc = device_get_ivars(dev);
647 struct ntb_transport_ctx *nt = device_get_softc(device_get_parent(dev));
648 struct ntb_queue_entry *entry;
649 struct ntb_transport_qp *qp;
652 if (q < 0 || q >= nc->qpcnt)
655 qp = &nt->qp_vec[nc->qpoff + q];
656 nt->qp_bitmap |= (1 << qp->qp_num);
658 qp->rx_handler = handlers->rx_handler;
659 qp->tx_handler = handlers->tx_handler;
660 qp->event_handler = handlers->event_handler;
662 for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) {
663 entry = malloc(sizeof(*entry), M_NTB_T, M_WAITOK | M_ZERO);
664 entry->cb_data = data;
666 entry->len = transport_mtu;
668 ntb_list_add(&qp->ntb_rx_q_lock, entry, &qp->rx_pend_q);
671 for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) {
672 entry = malloc(sizeof(*entry), M_NTB_T, M_WAITOK | M_ZERO);
674 ntb_list_add(&qp->ntb_tx_free_q_lock, entry, &qp->tx_free_q);
677 ntb_db_clear(dev, 1ull << qp->qp_num);
682 * ntb_transport_link_up - Notify NTB transport of client readiness to use queue
683 * @qp: NTB transport layer queue to be enabled
685 * Notify NTB transport layer of client readiness to use queue
688 ntb_transport_link_up(struct ntb_transport_qp *qp)
690 struct ntb_transport_ctx *nt = qp->transport;
692 qp->client_ready = true;
694 ntb_printf(2, "qp %d client ready\n", qp->qp_num);
697 callout_reset(&qp->link_work, 0, ntb_qp_link_work, qp);
705 * ntb_transport_tx_enqueue - Enqueue a new NTB queue entry
706 * @qp: NTB transport layer queue the entry is to be enqueued on
707 * @cb: per buffer pointer for callback function to use
708 * @data: pointer to data buffer that will be sent
709 * @len: length of the data buffer
711 * Enqueue a new transmit buffer onto the transport queue from which a NTB
712 * payload will be transmitted. This assumes that a lock is being held to
713 * serialize access to the qp.
715 * RETURNS: An appropriate ERRNO error value on error, or zero for success.
718 ntb_transport_tx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
721 struct ntb_queue_entry *entry;
724 if (!qp->link_is_up || len == 0) {
725 CTR0(KTR_NTB, "TX: link not up");
729 entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
731 CTR0(KTR_NTB, "TX: could not get entry from tx_free_q");
735 CTR1(KTR_NTB, "TX: got entry %p from tx_free_q", entry);
742 mtx_lock(&qp->tx_lock);
743 rc = ntb_process_tx(qp, entry);
744 mtx_unlock(&qp->tx_lock);
746 ntb_list_add(&qp->ntb_tx_free_q_lock, entry, &qp->tx_free_q);
748 "TX: process_tx failed. Returning entry %p to tx_free_q",
755 ntb_tx_copy_callback(void *data)
757 struct ntb_queue_entry *entry = data;
758 struct ntb_transport_qp *qp = entry->qp;
759 struct ntb_payload_header *hdr = entry->x_hdr;
761 iowrite32(entry->flags | NTBT_DESC_DONE_FLAG, &hdr->flags);
762 CTR1(KTR_NTB, "TX: hdr %p set DESC_DONE", hdr);
764 ntb_peer_db_set(qp->dev, 1ull << qp->qp_num);
767 * The entry length can only be zero if the packet is intended to be a
768 * "link down" or similar. Since no payload is being sent in these
769 * cases, there is nothing to add to the completion queue.
771 if (entry->len > 0) {
772 qp->tx_bytes += entry->len;
775 qp->tx_handler(qp, qp->cb_data, entry->buf,
783 "TX: entry %p sent. hdr->ver = %u, hdr->flags = 0x%x, Returning "
784 "to tx_free_q", entry, hdr->ver, hdr->flags);
785 ntb_list_add(&qp->ntb_tx_free_q_lock, entry, &qp->tx_free_q);
789 ntb_memcpy_tx(struct ntb_queue_entry *entry, void *offset)
792 CTR2(KTR_NTB, "TX: copying %d bytes to offset %p", entry->len, offset);
793 if (entry->buf != NULL) {
794 m_copydata((struct mbuf *)entry->buf, 0, entry->len, offset);
797 * Ensure that the data is fully copied before setting the
803 ntb_tx_copy_callback(entry);
807 ntb_async_tx(struct ntb_transport_qp *qp, struct ntb_queue_entry *entry)
809 struct ntb_payload_header *hdr;
812 offset = qp->tx_mw + qp->tx_max_frame * qp->tx_index;
813 hdr = (struct ntb_payload_header *)((char *)offset + qp->tx_max_frame -
814 sizeof(struct ntb_payload_header));
817 iowrite32(entry->len, &hdr->len);
818 iowrite32(qp->tx_pkts, &hdr->ver);
820 ntb_memcpy_tx(entry, offset);
824 ntb_process_tx(struct ntb_transport_qp *qp, struct ntb_queue_entry *entry)
828 "TX: process_tx: tx_pkts=%lu, tx_index=%u, remote entry=%u",
829 qp->tx_pkts, qp->tx_index, qp->remote_rx_info->entry);
830 if (qp->tx_index == qp->remote_rx_info->entry) {
831 CTR0(KTR_NTB, "TX: ring full");
836 if (entry->len > qp->tx_max_frame - sizeof(struct ntb_payload_header)) {
837 if (qp->tx_handler != NULL)
838 qp->tx_handler(qp, qp->cb_data, entry->buf,
844 ntb_list_add(&qp->ntb_tx_free_q_lock, entry, &qp->tx_free_q);
846 "TX: frame too big. returning entry %p to tx_free_q",
850 CTR2(KTR_NTB, "TX: copying entry %p to index %u", entry, qp->tx_index);
851 ntb_async_tx(qp, entry);
854 qp->tx_index %= qp->tx_max_entry;
863 ntb_transport_rxc_db(void *arg, int pending __unused)
865 struct ntb_transport_qp *qp = arg;
866 uint64_t qp_mask = 1ull << qp->qp_num;
869 CTR0(KTR_NTB, "RX: transport_rx");
871 while ((rc = ntb_process_rxc(qp)) == 0)
873 CTR1(KTR_NTB, "RX: process_rxc returned %d", rc);
875 if ((ntb_db_read(qp->dev) & qp_mask) != 0) {
876 /* If db is set, clear it and check queue once more. */
877 ntb_db_clear(qp->dev, qp_mask);
881 ntb_db_clear_mask(qp->dev, qp_mask);
885 ntb_process_rxc(struct ntb_transport_qp *qp)
887 struct ntb_payload_header *hdr;
888 struct ntb_queue_entry *entry;
891 offset = qp->rx_buff + qp->rx_max_frame * qp->rx_index;
892 hdr = (void *)(offset + qp->rx_max_frame -
893 sizeof(struct ntb_payload_header));
895 CTR1(KTR_NTB, "RX: process_rxc rx_index = %u", qp->rx_index);
896 if ((hdr->flags & NTBT_DESC_DONE_FLAG) == 0) {
897 CTR0(KTR_NTB, "RX: hdr not done");
902 if ((hdr->flags & NTBT_LINK_DOWN_FLAG) != 0) {
903 CTR0(KTR_NTB, "RX: link down");
904 ntb_qp_link_down(qp);
909 if (hdr->ver != (uint32_t)qp->rx_pkts) {
910 CTR2(KTR_NTB,"RX: ver != rx_pkts (%x != %lx). "
911 "Returning entry to rx_pend_q", hdr->ver, qp->rx_pkts);
916 entry = ntb_list_mv(&qp->ntb_rx_q_lock, &qp->rx_pend_q, &qp->rx_post_q);
919 CTR0(KTR_NTB, "RX: No entries in rx_pend_q");
922 callout_stop(&qp->rx_full);
923 CTR1(KTR_NTB, "RX: rx entry %p from rx_pend_q", entry);
926 entry->index = qp->rx_index;
928 if (hdr->len > entry->len) {
929 CTR2(KTR_NTB, "RX: len too long. Wanted %ju got %ju",
930 (uintmax_t)hdr->len, (uintmax_t)entry->len);
934 entry->flags |= NTBT_DESC_DONE_FLAG;
936 ntb_complete_rxc(qp);
938 qp->rx_bytes += hdr->len;
941 CTR1(KTR_NTB, "RX: received %ld rx_pkts", qp->rx_pkts);
943 entry->len = hdr->len;
945 ntb_memcpy_rx(qp, entry, offset);
949 qp->rx_index %= qp->rx_max_entry;
954 ntb_memcpy_rx(struct ntb_transport_qp *qp, struct ntb_queue_entry *entry,
957 struct ifnet *ifp = entry->cb_data;
958 unsigned int len = entry->len;
960 CTR2(KTR_NTB, "RX: copying %d bytes from offset %p", len, offset);
962 entry->buf = (void *)m_devget(offset, len, 0, ifp, NULL);
963 if (entry->buf == NULL)
964 entry->len = -ENOMEM;
966 /* Ensure that the data is globally visible before clearing the flag */
969 CTR2(KTR_NTB, "RX: copied entry %p to mbuf %p.", entry, entry->buf);
970 ntb_rx_copy_callback(qp, entry);
974 ntb_rx_copy_callback(struct ntb_transport_qp *qp, void *data)
976 struct ntb_queue_entry *entry;
979 entry->flags |= NTBT_DESC_DONE_FLAG;
980 ntb_complete_rxc(qp);
984 ntb_complete_rxc(struct ntb_transport_qp *qp)
986 struct ntb_queue_entry *entry;
990 CTR0(KTR_NTB, "RX: rx_completion_task");
992 mtx_lock_spin(&qp->ntb_rx_q_lock);
994 while (!STAILQ_EMPTY(&qp->rx_post_q)) {
995 entry = STAILQ_FIRST(&qp->rx_post_q);
996 if ((entry->flags & NTBT_DESC_DONE_FLAG) == 0)
999 entry->x_hdr->flags = 0;
1000 iowrite32(entry->index, &qp->rx_info->entry);
1002 STAILQ_REMOVE_HEAD(&qp->rx_post_q, entry);
1008 * Re-initialize queue_entry for reuse; rx_handler takes
1009 * ownership of the mbuf.
1012 entry->len = transport_mtu;
1013 entry->cb_data = qp->cb_data;
1015 STAILQ_INSERT_TAIL(&qp->rx_pend_q, entry, entry);
1017 mtx_unlock_spin(&qp->ntb_rx_q_lock);
1019 CTR2(KTR_NTB, "RX: completing entry %p, mbuf %p", entry, m);
1020 if (qp->rx_handler != NULL && qp->client_ready)
1021 qp->rx_handler(qp, qp->cb_data, m, len);
1025 mtx_lock_spin(&qp->ntb_rx_q_lock);
1028 mtx_unlock_spin(&qp->ntb_rx_q_lock);
1032 ntb_transport_doorbell_callback(void *data, uint32_t vector)
1034 struct ntb_transport_ctx *nt = data;
1035 struct ntb_transport_qp *qp;
1039 vec_mask = ntb_db_vector_mask(nt->dev, vector);
1040 vec_mask &= nt->qp_bitmap;
1041 if ((vec_mask & (vec_mask - 1)) != 0)
1042 vec_mask &= ntb_db_read(nt->dev);
1043 if (vec_mask != 0) {
1044 ntb_db_set_mask(nt->dev, vec_mask);
1045 ntb_db_clear(nt->dev, vec_mask);
1047 while (vec_mask != 0) {
1048 qp_num = ffsll(vec_mask) - 1;
1050 qp = &nt->qp_vec[qp_num];
1052 taskqueue_enqueue(qp->rxc_tq, &qp->rxc_db_work);
1054 vec_mask &= ~(1ull << qp_num);
1058 /* Link Event handler */
1060 ntb_transport_event_callback(void *data)
1062 struct ntb_transport_ctx *nt = data;
1064 if (ntb_link_is_up(nt->dev, &nt->link_speed, &nt->link_width)) {
1065 ntb_printf(1, "HW link up\n");
1066 callout_reset(&nt->link_work, 0, ntb_transport_link_work, nt);
1068 ntb_printf(1, "HW link down\n");
1069 taskqueue_enqueue(taskqueue_swi, &nt->link_cleanup);
1075 ntb_transport_link_work(void *arg)
1077 struct ntb_transport_ctx *nt = arg;
1078 device_t dev = nt->dev;
1079 struct ntb_transport_qp *qp;
1080 uint64_t val64, size;
1085 /* send the local info, in the opposite order of the way we read it */
1086 for (i = 0; i < nt->mw_count; i++) {
1087 size = nt->mw_vec[i].phys_size;
1089 if (max_mw_size != 0 && size > max_mw_size)
1092 ntb_peer_spad_write(dev, NTBT_MW0_SZ_HIGH + (i * 2),
1094 ntb_peer_spad_write(dev, NTBT_MW0_SZ_LOW + (i * 2), size);
1096 ntb_peer_spad_write(dev, NTBT_NUM_MWS, nt->mw_count);
1097 ntb_peer_spad_write(dev, NTBT_NUM_QPS, nt->qp_count);
1098 ntb_peer_spad_write(dev, NTBT_QP_LINKS, 0);
1099 ntb_peer_spad_write(dev, NTBT_VERSION, NTB_TRANSPORT_VERSION);
1101 /* Query the remote side for its info */
1103 ntb_spad_read(dev, NTBT_VERSION, &val);
1104 if (val != NTB_TRANSPORT_VERSION)
1107 ntb_spad_read(dev, NTBT_NUM_QPS, &val);
1108 if (val != nt->qp_count)
1111 ntb_spad_read(dev, NTBT_NUM_MWS, &val);
1112 if (val != nt->mw_count)
1115 for (i = 0; i < nt->mw_count; i++) {
1116 ntb_spad_read(dev, NTBT_MW0_SZ_HIGH + (i * 2), &val);
1117 val64 = (uint64_t)val << 32;
1119 ntb_spad_read(dev, NTBT_MW0_SZ_LOW + (i * 2), &val);
1122 rc = ntb_set_mw(nt, i, val64);
1127 nt->link_is_up = true;
1128 ntb_printf(1, "transport link up\n");
1130 for (i = 0; i < nt->qp_count; i++) {
1131 qp = &nt->qp_vec[i];
1133 ntb_transport_setup_qp_mw(nt, i);
1135 if (qp->client_ready)
1136 callout_reset(&qp->link_work, 0, ntb_qp_link_work, qp);
1142 for (i = 0; i < nt->mw_count; i++)
1145 if (ntb_link_is_up(dev, &nt->link_speed, &nt->link_width))
1146 callout_reset(&nt->link_work,
1147 NTB_LINK_DOWN_TIMEOUT * hz / 1000, ntb_transport_link_work, nt);
1151 ntb_set_mw(struct ntb_transport_ctx *nt, int num_mw, size_t size)
1153 struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
1154 size_t xlat_size, buff_size;
1160 xlat_size = roundup(size, mw->xlat_align_size);
1161 buff_size = xlat_size;
1163 /* No need to re-setup */
1164 if (mw->xlat_size == xlat_size)
1167 if (mw->buff_size != 0)
1168 ntb_free_mw(nt, num_mw);
1170 /* Alloc memory for receiving data. Must be aligned */
1171 mw->xlat_size = xlat_size;
1172 mw->buff_size = buff_size;
1174 mw->virt_addr = contigmalloc(mw->buff_size, M_NTB_T, M_ZERO, 0,
1175 mw->addr_limit, mw->xlat_align, 0);
1176 if (mw->virt_addr == NULL) {
1177 ntb_printf(0, "Unable to allocate MW buffer of size %zu/%zu\n",
1178 mw->buff_size, mw->xlat_size);
1183 /* TODO: replace with bus_space_* functions */
1184 mw->dma_addr = vtophys(mw->virt_addr);
1187 * Ensure that the allocation from contigmalloc is aligned as
1188 * requested. XXX: This may not be needed -- brought in for parity
1189 * with the Linux driver.
1191 if (mw->dma_addr % mw->xlat_align != 0) {
1193 "DMA memory 0x%jx not aligned to BAR size 0x%zx\n",
1194 (uintmax_t)mw->dma_addr, size);
1195 ntb_free_mw(nt, num_mw);
1199 /* Notify HW the memory location of the receive buffer */
1200 rc = ntb_mw_set_trans(nt->dev, num_mw, mw->dma_addr, mw->xlat_size);
1202 ntb_printf(0, "Unable to set mw%d translation\n", num_mw);
1203 ntb_free_mw(nt, num_mw);
1211 ntb_free_mw(struct ntb_transport_ctx *nt, int num_mw)
1213 struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
1215 if (mw->virt_addr == NULL)
1218 ntb_mw_clear_trans(nt->dev, num_mw);
1219 contigfree(mw->virt_addr, mw->xlat_size, M_NTB_T);
1222 mw->virt_addr = NULL;
1226 ntb_transport_setup_qp_mw(struct ntb_transport_ctx *nt, unsigned int qp_num)
1228 struct ntb_transport_qp *qp = &nt->qp_vec[qp_num];
1229 struct ntb_transport_mw *mw;
1233 unsigned num_qps_mw, mw_num, mw_count;
1235 mw_count = nt->mw_count;
1236 mw_num = QP_TO_MW(nt, qp_num);
1237 mw = &nt->mw_vec[mw_num];
1239 if (mw->virt_addr == NULL)
1242 if (mw_num < nt->qp_count % mw_count)
1243 num_qps_mw = nt->qp_count / mw_count + 1;
1245 num_qps_mw = nt->qp_count / mw_count;
1247 rx_size = mw->xlat_size / num_qps_mw;
1248 qp->rx_buff = mw->virt_addr + rx_size * (qp_num / mw_count);
1249 rx_size -= sizeof(struct ntb_rx_info);
1251 qp->remote_rx_info = (void*)(qp->rx_buff + rx_size);
1253 /* Due to house-keeping, there must be at least 2 buffs */
1254 qp->rx_max_frame = qmin(transport_mtu, rx_size / 2);
1255 qp->rx_max_entry = rx_size / qp->rx_max_frame;
1258 qp->remote_rx_info->entry = qp->rx_max_entry - 1;
1260 /* Set up the hdr offsets with 0s */
1261 for (i = 0; i < qp->rx_max_entry; i++) {
1262 offset = (void *)(qp->rx_buff + qp->rx_max_frame * (i + 1) -
1263 sizeof(struct ntb_payload_header));
1264 memset(offset, 0, sizeof(struct ntb_payload_header));
1275 ntb_qp_link_work(void *arg)
1277 struct ntb_transport_qp *qp = arg;
1278 device_t dev = qp->dev;
1279 struct ntb_transport_ctx *nt = qp->transport;
1283 /* Report queues that are up on our side */
1284 for (i = 0, val = 0; i < nt->qp_count; i++) {
1285 if (nt->qp_vec[i].client_ready)
1288 ntb_peer_spad_write(dev, NTBT_QP_LINKS, val);
1290 /* See if the remote side is up */
1291 ntb_spad_read(dev, NTBT_QP_LINKS, &val);
1292 if ((val & (1ull << qp->qp_num)) != 0) {
1293 ntb_printf(2, "qp %d link up\n", qp->qp_num);
1294 qp->link_is_up = true;
1296 if (qp->event_handler != NULL)
1297 qp->event_handler(qp->cb_data, NTB_LINK_UP);
1299 ntb_db_clear_mask(dev, 1ull << qp->qp_num);
1300 } else if (nt->link_is_up)
1301 callout_reset(&qp->link_work,
1302 NTB_LINK_DOWN_TIMEOUT * hz / 1000, ntb_qp_link_work, qp);
1305 /* Link down event*/
1307 ntb_transport_link_cleanup(struct ntb_transport_ctx *nt)
1309 struct ntb_transport_qp *qp;
1312 callout_drain(&nt->link_work);
1315 /* Pass along the info to any clients */
1316 for (i = 0; i < nt->qp_count; i++) {
1317 if ((nt->qp_bitmap & (1 << i)) != 0) {
1318 qp = &nt->qp_vec[i];
1319 ntb_qp_link_cleanup(qp);
1320 callout_drain(&qp->link_work);
1325 * The scratchpad registers keep the values if the remote side
1326 * goes down, blast them now to give them a sane value the next
1327 * time they are accessed
1329 ntb_spad_clear(nt->dev);
1333 ntb_transport_link_cleanup_work(void *arg, int pending __unused)
1336 ntb_transport_link_cleanup(arg);
1340 ntb_qp_link_down(struct ntb_transport_qp *qp)
1343 ntb_qp_link_cleanup(qp);
1347 ntb_qp_link_down_reset(struct ntb_transport_qp *qp)
1350 qp->link_is_up = false;
1351 ntb_db_set_mask(qp->dev, 1ull << qp->qp_num);
1353 qp->tx_index = qp->rx_index = 0;
1354 qp->tx_bytes = qp->rx_bytes = 0;
1355 qp->tx_pkts = qp->rx_pkts = 0;
1357 qp->rx_ring_empty = 0;
1358 qp->tx_ring_full = 0;
1360 qp->rx_err_no_buf = qp->tx_err_no_buf = 0;
1361 qp->rx_err_oflow = qp->rx_err_ver = 0;
1365 ntb_qp_link_cleanup(struct ntb_transport_qp *qp)
1368 callout_drain(&qp->link_work);
1369 ntb_qp_link_down_reset(qp);
1371 if (qp->event_handler != NULL)
1372 qp->event_handler(qp->cb_data, NTB_LINK_DOWN);
1375 /* Link commanded down */
1377 * ntb_transport_link_down - Notify NTB transport to no longer enqueue data
1378 * @qp: NTB transport layer queue to be disabled
1380 * Notify NTB transport layer of client's desire to no longer receive data on
1381 * transport queue specified. It is the client's responsibility to ensure all
1382 * entries on queue are purged or otherwise handled appropriately.
1385 ntb_transport_link_down(struct ntb_transport_qp *qp)
1387 struct ntb_transport_ctx *nt = qp->transport;
1391 qp->client_ready = false;
1392 for (i = 0, val = 0; i < nt->qp_count; i++) {
1393 if (nt->qp_vec[i].client_ready)
1396 ntb_peer_spad_write(qp->dev, NTBT_QP_LINKS, val);
1399 ntb_send_link_down(qp);
1401 callout_drain(&qp->link_work);
1405 * ntb_transport_link_query - Query transport link state
1406 * @qp: NTB transport layer queue to be queried
1408 * Query connectivity to the remote system of the NTB transport queue
1410 * RETURNS: true for link up or false for link down
1413 ntb_transport_link_query(struct ntb_transport_qp *qp)
1416 return (qp->link_is_up);
1420 * ntb_transport_link_speed - Query transport link speed
1421 * @qp: NTB transport layer queue to be queried
1423 * Query connection speed to the remote system of the NTB transport queue
1425 * RETURNS: link speed in bits per second
1428 ntb_transport_link_speed(struct ntb_transport_qp *qp)
1430 struct ntb_transport_ctx *nt = qp->transport;
1433 if (!nt->link_is_up)
1435 switch (nt->link_speed) {
1436 case NTB_SPEED_GEN1:
1437 rate = 2500000000 * 8 / 10;
1439 case NTB_SPEED_GEN2:
1440 rate = 5000000000 * 8 / 10;
1442 case NTB_SPEED_GEN3:
1443 rate = 8000000000 * 128 / 130;
1445 case NTB_SPEED_GEN4:
1446 rate = 16000000000 * 128 / 130;
1451 if (nt->link_width <= 0)
1453 return (rate * nt->link_width);
1457 ntb_send_link_down(struct ntb_transport_qp *qp)
1459 struct ntb_queue_entry *entry;
1462 if (!qp->link_is_up)
1465 for (i = 0; i < NTB_LINK_DOWN_TIMEOUT; i++) {
1466 entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
1469 pause("NTB Wait for link down", hz / 10);
1475 entry->cb_data = NULL;
1478 entry->flags = NTBT_LINK_DOWN_FLAG;
1480 mtx_lock(&qp->tx_lock);
1481 rc = ntb_process_tx(qp, entry);
1482 mtx_unlock(&qp->tx_lock);
1484 printf("ntb: Failed to send link down\n");
1486 ntb_qp_link_down_reset(qp);
1490 /* List Management */
1493 ntb_list_add(struct mtx *lock, struct ntb_queue_entry *entry,
1494 struct ntb_queue_list *list)
1497 mtx_lock_spin(lock);
1498 STAILQ_INSERT_TAIL(list, entry, entry);
1499 mtx_unlock_spin(lock);
1502 static struct ntb_queue_entry *
1503 ntb_list_rm(struct mtx *lock, struct ntb_queue_list *list)
1505 struct ntb_queue_entry *entry;
1507 mtx_lock_spin(lock);
1508 if (STAILQ_EMPTY(list)) {
1512 entry = STAILQ_FIRST(list);
1513 STAILQ_REMOVE_HEAD(list, entry);
1515 mtx_unlock_spin(lock);
1520 static struct ntb_queue_entry *
1521 ntb_list_mv(struct mtx *lock, struct ntb_queue_list *from,
1522 struct ntb_queue_list *to)
1524 struct ntb_queue_entry *entry;
1526 mtx_lock_spin(lock);
1527 if (STAILQ_EMPTY(from)) {
1531 entry = STAILQ_FIRST(from);
1532 STAILQ_REMOVE_HEAD(from, entry);
1533 STAILQ_INSERT_TAIL(to, entry, entry);
1536 mtx_unlock_spin(lock);
1541 * ntb_transport_qp_num - Query the qp number
1542 * @qp: NTB transport layer queue to be queried
1544 * Query qp number of the NTB transport queue
1546 * RETURNS: a zero based number specifying the qp number
1548 unsigned char ntb_transport_qp_num(struct ntb_transport_qp *qp)
1551 return (qp->qp_num);
1555 * ntb_transport_max_size - Query the max payload size of a qp
1556 * @qp: NTB transport layer queue to be queried
1558 * Query the maximum payload size permissible on the given qp
1560 * RETURNS: the max payload size of a qp
1563 ntb_transport_max_size(struct ntb_transport_qp *qp)
1566 return (qp->tx_max_frame - sizeof(struct ntb_payload_header));
1570 ntb_transport_tx_free_entry(struct ntb_transport_qp *qp)
1572 unsigned int head = qp->tx_index;
1573 unsigned int tail = qp->remote_rx_info->entry;
1575 return (tail >= head ? tail - head : qp->tx_max_entry + tail - head);
1578 static device_method_t ntb_transport_methods[] = {
1579 /* Device interface */
1580 DEVMETHOD(device_probe, ntb_transport_probe),
1581 DEVMETHOD(device_attach, ntb_transport_attach),
1582 DEVMETHOD(device_detach, ntb_transport_detach),
1584 DEVMETHOD(bus_child_location_str, ntb_transport_child_location_str),
1585 DEVMETHOD(bus_print_child, ntb_transport_print_child),
1589 devclass_t ntb_transport_devclass;
1590 static DEFINE_CLASS_0(ntb_transport, ntb_transport_driver,
1591 ntb_transport_methods, sizeof(struct ntb_transport_ctx));
1592 DRIVER_MODULE(ntb_transport, ntb_hw, ntb_transport_driver,
1593 ntb_transport_devclass, NULL, NULL);
1594 MODULE_DEPEND(ntb_transport, ntb, 1, 1, 1);
1595 MODULE_VERSION(ntb_transport, 1);