2 * Copyright (c) 2016 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 {
193 struct ntb_transport_child *next;
196 struct ntb_transport_ctx {
198 struct ntb_transport_child *child;
199 struct ntb_transport_mw *mw_vec;
200 struct ntb_transport_qp *qp_vec;
204 volatile bool link_is_up;
205 enum ntb_speed link_speed;
206 enum ntb_width link_width;
207 struct callout link_work;
208 struct callout link_watchdog;
209 struct task link_cleanup;
213 NTBT_DESC_DONE_FLAG = 1 << 0,
214 NTBT_LINK_DOWN_FLAG = 1 << 1,
217 struct ntb_payload_header {
225 * The order of this enum is part of the remote protocol. Do not
226 * reorder without bumping protocol version (and it's probably best
227 * to keep the protocol in lock-step with the Linux NTB driver.
234 * N.B.: transport_link_work assumes MW1 enums = MW0 + 2.
242 * Some NTB-using hardware have a watchdog to work around NTB hangs; if
243 * a register or doorbell isn't written every few seconds, the link is
244 * torn down. Write an otherwise unused register every few seconds to
245 * work around this watchdog.
247 NTBT_WATCHDOG_SPAD = 15
250 #define QP_TO_MW(nt, qp) ((qp) % nt->mw_count)
251 #define NTB_QP_DEF_NUM_ENTRIES 100
252 #define NTB_LINK_DOWN_TIMEOUT 10
254 static int ntb_transport_probe(device_t dev);
255 static int ntb_transport_attach(device_t dev);
256 static int ntb_transport_detach(device_t dev);
257 static void ntb_transport_init_queue(struct ntb_transport_ctx *nt,
258 unsigned int qp_num);
259 static int ntb_process_tx(struct ntb_transport_qp *qp,
260 struct ntb_queue_entry *entry);
261 static void ntb_transport_rxc_db(void *arg, int pending);
262 static int ntb_process_rxc(struct ntb_transport_qp *qp);
263 static void ntb_memcpy_rx(struct ntb_transport_qp *qp,
264 struct ntb_queue_entry *entry, void *offset);
265 static inline void ntb_rx_copy_callback(struct ntb_transport_qp *qp,
267 static void ntb_complete_rxc(struct ntb_transport_qp *qp);
268 static void ntb_transport_doorbell_callback(void *data, uint32_t vector);
269 static void ntb_transport_event_callback(void *data);
270 static void ntb_transport_link_work(void *arg);
271 static int ntb_set_mw(struct ntb_transport_ctx *, int num_mw, size_t size);
272 static void ntb_free_mw(struct ntb_transport_ctx *nt, int num_mw);
273 static int ntb_transport_setup_qp_mw(struct ntb_transport_ctx *nt,
274 unsigned int qp_num);
275 static void ntb_qp_link_work(void *arg);
276 static void ntb_transport_link_cleanup(struct ntb_transport_ctx *nt);
277 static void ntb_transport_link_cleanup_work(void *, int);
278 static void ntb_qp_link_down(struct ntb_transport_qp *qp);
279 static void ntb_qp_link_down_reset(struct ntb_transport_qp *qp);
280 static void ntb_qp_link_cleanup(struct ntb_transport_qp *qp);
281 static void ntb_send_link_down(struct ntb_transport_qp *qp);
282 static void ntb_list_add(struct mtx *lock, struct ntb_queue_entry *entry,
283 struct ntb_queue_list *list);
284 static struct ntb_queue_entry *ntb_list_rm(struct mtx *lock,
285 struct ntb_queue_list *list);
286 static struct ntb_queue_entry *ntb_list_mv(struct mtx *lock,
287 struct ntb_queue_list *from, struct ntb_queue_list *to);
288 static void xeon_link_watchdog_hb(void *);
290 static const struct ntb_ctx_ops ntb_transport_ops = {
291 .link_event = ntb_transport_event_callback,
292 .db_event = ntb_transport_doorbell_callback,
295 MALLOC_DEFINE(M_NTB_T, "ntb_transport", "ntb transport driver");
298 iowrite32(uint32_t val, void *addr)
301 bus_space_write_4(X86_BUS_SPACE_MEM, 0/* HACK */, (uintptr_t)addr,
305 /* Transport Init and teardown */
308 xeon_link_watchdog_hb(void *arg)
310 struct ntb_transport_ctx *nt;
313 ntb_spad_write(nt->dev, NTBT_WATCHDOG_SPAD, 0);
314 callout_reset(&nt->link_watchdog, 1 * hz, xeon_link_watchdog_hb, nt);
318 ntb_transport_probe(device_t dev)
321 device_set_desc(dev, "NTB Transport");
326 ntb_transport_attach(device_t dev)
328 struct ntb_transport_ctx *nt = device_get_softc(dev);
329 struct ntb_transport_child **cpp = &nt->child;
330 struct ntb_transport_child *nc;
331 struct ntb_transport_mw *mw;
333 int rc, i, db_count, spad_count, qp, qpu, qpo, qpt;
336 char *n, *np, *c, *name;
339 nt->mw_count = ntb_mw_count(dev);
340 spad_count = ntb_spad_count(dev);
341 db_bitmap = ntb_db_valid_mask(dev);
342 db_count = flsll(db_bitmap);
343 KASSERT(db_bitmap == (1 << db_count) - 1,
344 ("Doorbells are not sequential (%jx).\n", db_bitmap));
346 device_printf(dev, "%d memory windows, %d scratchpads, "
347 "%d doorbells\n", nt->mw_count, spad_count, db_count);
349 if (nt->mw_count == 0) {
350 device_printf(dev, "At least 1 memory window required.\n");
353 if (spad_count < 6) {
354 device_printf(dev, "At least 6 scratchpads required.\n");
357 if (spad_count < 4 + 2 * nt->mw_count) {
358 nt->mw_count = (spad_count - 4) / 2;
359 device_printf(dev, "Scratchpads enough only for %d "
360 "memory windows.\n", nt->mw_count);
362 if (db_bitmap == 0) {
363 device_printf(dev, "At least one doorbell required.\n");
367 nt->mw_vec = malloc(nt->mw_count * sizeof(*nt->mw_vec), M_NTB_T,
369 for (i = 0; i < nt->mw_count; i++) {
372 rc = ntb_mw_get_range(dev, i, &mw->phys_addr, &mw->vbase,
373 &mw->phys_size, &mw->xlat_align, &mw->xlat_align_size,
380 mw->virt_addr = NULL;
383 rc = ntb_mw_set_wc(dev, i, VM_MEMATTR_WRITE_COMBINING);
385 ntb_printf(0, "Unable to set mw%d caching\n", i);
389 qpo = imin(db_count, nt->mw_count);
392 snprintf(buf, sizeof(buf), "hint.%s.%d.config", device_get_name(dev),
393 device_get_unit(dev));
394 TUNABLE_STR_FETCH(buf, cfg, sizeof(cfg));
397 while ((c = strsep(&n, ",")) != NULL) {
399 name = strsep(&np, ":");
400 if (name != NULL && name[0] == 0)
402 qp = (np && np[0] != 0) ? strtol(np, NULL, 10) : qpo - qpu;
406 if (qp > qpt - qpu) {
407 device_printf(dev, "Not enough resources for config\n");
411 nc = malloc(sizeof(*nc), M_DEVBUF, M_WAITOK | M_ZERO);
414 nc->dev = device_add_child(dev, name, -1);
415 if (nc->dev == NULL) {
416 device_printf(dev, "Can not add child.\n");
419 device_set_ivars(nc->dev, nc);
424 device_printf(dev, "%d \"%s\": queues %d",
427 printf("-%d", qpu + qp - 1);
436 nt->qp_vec = malloc(nt->qp_count * sizeof(*nt->qp_vec), M_NTB_T,
439 for (i = 0; i < nt->qp_count; i++)
440 ntb_transport_init_queue(nt, i);
442 callout_init(&nt->link_work, 0);
443 callout_init(&nt->link_watchdog, 0);
444 TASK_INIT(&nt->link_cleanup, 0, ntb_transport_link_cleanup_work, nt);
446 rc = ntb_set_ctx(dev, nt, &ntb_transport_ops);
450 nt->link_is_up = false;
451 ntb_link_enable(dev, NTB_SPEED_AUTO, NTB_WIDTH_AUTO);
453 if (enable_xeon_watchdog != 0)
454 callout_reset(&nt->link_watchdog, 0, xeon_link_watchdog_hb, nt);
456 bus_generic_attach(dev);
460 free(nt->qp_vec, M_NTB_T);
461 free(nt->mw_vec, M_NTB_T);
466 ntb_transport_detach(device_t dev)
468 struct ntb_transport_ctx *nt = device_get_softc(dev);
469 struct ntb_transport_child **cpp = &nt->child;
470 struct ntb_transport_child *nc;
473 while ((nc = *cpp) != NULL) {
475 error = device_delete_child(dev, nc->dev);
480 KASSERT(nt->qp_bitmap == 0,
481 ("Some queues not freed on detach (%jx)", nt->qp_bitmap));
483 ntb_transport_link_cleanup(nt);
484 taskqueue_drain(taskqueue_swi, &nt->link_cleanup);
485 callout_drain(&nt->link_work);
486 callout_drain(&nt->link_watchdog);
488 ntb_link_disable(dev);
491 for (i = 0; i < nt->mw_count; i++)
494 free(nt->qp_vec, M_NTB_T);
495 free(nt->mw_vec, M_NTB_T);
500 ntb_transport_queue_count(device_t dev)
502 struct ntb_transport_child *nc = device_get_ivars(dev);
508 ntb_transport_init_queue(struct ntb_transport_ctx *nt, unsigned int qp_num)
510 struct ntb_transport_mw *mw;
511 struct ntb_transport_qp *qp;
513 uint64_t mw_size, qp_offset;
515 unsigned num_qps_mw, mw_num, mw_count;
517 mw_count = nt->mw_count;
518 mw_num = QP_TO_MW(nt, qp_num);
519 mw = &nt->mw_vec[mw_num];
521 qp = &nt->qp_vec[qp_num];
525 qp->client_ready = false;
526 qp->event_handler = NULL;
527 ntb_qp_link_down_reset(qp);
529 if (mw_num < nt->qp_count % mw_count)
530 num_qps_mw = nt->qp_count / mw_count + 1;
532 num_qps_mw = nt->qp_count / mw_count;
534 mw_base = mw->phys_addr;
535 mw_size = mw->phys_size;
537 tx_size = mw_size / num_qps_mw;
538 qp_offset = tx_size * (qp_num / mw_count);
540 qp->tx_mw = mw->vbase + qp_offset;
541 KASSERT(qp->tx_mw != NULL, ("uh oh?"));
543 /* XXX Assumes that a vm_paddr_t is equivalent to bus_addr_t */
544 qp->tx_mw_phys = mw_base + qp_offset;
545 KASSERT(qp->tx_mw_phys != 0, ("uh oh?"));
547 tx_size -= sizeof(struct ntb_rx_info);
548 qp->rx_info = (void *)(qp->tx_mw + tx_size);
550 /* Due to house-keeping, there must be at least 2 buffs */
551 qp->tx_max_frame = qmin(transport_mtu, tx_size / 2);
552 qp->tx_max_entry = tx_size / qp->tx_max_frame;
554 callout_init(&qp->link_work, 0);
555 callout_init(&qp->rx_full, 1);
557 mtx_init(&qp->ntb_rx_q_lock, "ntb rx q", NULL, MTX_SPIN);
558 mtx_init(&qp->ntb_tx_free_q_lock, "ntb tx free q", NULL, MTX_SPIN);
559 mtx_init(&qp->tx_lock, "ntb transport tx", NULL, MTX_DEF);
560 TASK_INIT(&qp->rxc_db_work, 0, ntb_transport_rxc_db, qp);
561 qp->rxc_tq = taskqueue_create("ntbt_rx", M_WAITOK,
562 taskqueue_thread_enqueue, &qp->rxc_tq);
563 taskqueue_start_threads(&qp->rxc_tq, 1, PI_NET, "%s rx%d",
564 device_get_nameunit(nt->dev), qp_num);
566 STAILQ_INIT(&qp->rx_post_q);
567 STAILQ_INIT(&qp->rx_pend_q);
568 STAILQ_INIT(&qp->tx_free_q);
572 ntb_transport_free_queue(struct ntb_transport_qp *qp)
574 struct ntb_transport_ctx *nt = qp->transport;
575 struct ntb_queue_entry *entry;
577 callout_drain(&qp->link_work);
579 ntb_db_set_mask(qp->dev, 1ull << qp->qp_num);
580 taskqueue_drain_all(qp->rxc_tq);
581 taskqueue_free(qp->rxc_tq);
584 qp->rx_handler = NULL;
585 qp->tx_handler = NULL;
586 qp->event_handler = NULL;
588 while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_pend_q)))
589 free(entry, M_NTB_T);
591 while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_post_q)))
592 free(entry, M_NTB_T);
594 while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
595 free(entry, M_NTB_T);
597 nt->qp_bitmap &= ~(1 << qp->qp_num);
601 * ntb_transport_create_queue - Create a new NTB transport layer queue
602 * @rx_handler: receive callback function
603 * @tx_handler: transmit callback function
604 * @event_handler: event callback function
606 * Create a new NTB transport layer queue and provide the queue with a callback
607 * routine for both transmit and receive. The receive callback routine will be
608 * used to pass up data when the transport has received it on the queue. The
609 * transmit callback routine will be called when the transport has completed the
610 * transmission of the data on the queue and the data is ready to be freed.
612 * RETURNS: pointer to newly created ntb_queue, NULL on error.
614 struct ntb_transport_qp *
615 ntb_transport_create_queue(device_t dev, int q,
616 const struct ntb_queue_handlers *handlers, void *data)
618 struct ntb_transport_child *nc = device_get_ivars(dev);
619 struct ntb_transport_ctx *nt = device_get_softc(device_get_parent(dev));
620 struct ntb_queue_entry *entry;
621 struct ntb_transport_qp *qp;
624 if (q < 0 || q >= nc->qpcnt)
627 qp = &nt->qp_vec[nc->qpoff + q];
628 nt->qp_bitmap |= (1 << qp->qp_num);
630 qp->rx_handler = handlers->rx_handler;
631 qp->tx_handler = handlers->tx_handler;
632 qp->event_handler = handlers->event_handler;
634 for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) {
635 entry = malloc(sizeof(*entry), M_NTB_T, M_WAITOK | M_ZERO);
636 entry->cb_data = data;
638 entry->len = transport_mtu;
640 ntb_list_add(&qp->ntb_rx_q_lock, entry, &qp->rx_pend_q);
643 for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) {
644 entry = malloc(sizeof(*entry), M_NTB_T, M_WAITOK | M_ZERO);
646 ntb_list_add(&qp->ntb_tx_free_q_lock, entry, &qp->tx_free_q);
649 ntb_db_clear(dev, 1ull << qp->qp_num);
654 * ntb_transport_link_up - Notify NTB transport of client readiness to use queue
655 * @qp: NTB transport layer queue to be enabled
657 * Notify NTB transport layer of client readiness to use queue
660 ntb_transport_link_up(struct ntb_transport_qp *qp)
662 struct ntb_transport_ctx *nt = qp->transport;
664 qp->client_ready = true;
666 ntb_printf(2, "qp %d client ready\n", qp->qp_num);
669 callout_reset(&qp->link_work, 0, ntb_qp_link_work, qp);
677 * ntb_transport_tx_enqueue - Enqueue a new NTB queue entry
678 * @qp: NTB transport layer queue the entry is to be enqueued on
679 * @cb: per buffer pointer for callback function to use
680 * @data: pointer to data buffer that will be sent
681 * @len: length of the data buffer
683 * Enqueue a new transmit buffer onto the transport queue from which a NTB
684 * payload will be transmitted. This assumes that a lock is being held to
685 * serialize access to the qp.
687 * RETURNS: An appropriate ERRNO error value on error, or zero for success.
690 ntb_transport_tx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
693 struct ntb_queue_entry *entry;
696 if (!qp->link_is_up || len == 0) {
697 CTR0(KTR_NTB, "TX: link not up");
701 entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
703 CTR0(KTR_NTB, "TX: could not get entry from tx_free_q");
707 CTR1(KTR_NTB, "TX: got entry %p from tx_free_q", entry);
714 mtx_lock(&qp->tx_lock);
715 rc = ntb_process_tx(qp, entry);
716 mtx_unlock(&qp->tx_lock);
718 ntb_list_add(&qp->ntb_tx_free_q_lock, entry, &qp->tx_free_q);
720 "TX: process_tx failed. Returning entry %p to tx_free_q",
727 ntb_tx_copy_callback(void *data)
729 struct ntb_queue_entry *entry = data;
730 struct ntb_transport_qp *qp = entry->qp;
731 struct ntb_payload_header *hdr = entry->x_hdr;
733 iowrite32(entry->flags | NTBT_DESC_DONE_FLAG, &hdr->flags);
734 CTR1(KTR_NTB, "TX: hdr %p set DESC_DONE", hdr);
736 ntb_peer_db_set(qp->dev, 1ull << qp->qp_num);
739 * The entry length can only be zero if the packet is intended to be a
740 * "link down" or similar. Since no payload is being sent in these
741 * cases, there is nothing to add to the completion queue.
743 if (entry->len > 0) {
744 qp->tx_bytes += entry->len;
747 qp->tx_handler(qp, qp->cb_data, entry->buf,
755 "TX: entry %p sent. hdr->ver = %u, hdr->flags = 0x%x, Returning "
756 "to tx_free_q", entry, hdr->ver, hdr->flags);
757 ntb_list_add(&qp->ntb_tx_free_q_lock, entry, &qp->tx_free_q);
761 ntb_memcpy_tx(struct ntb_queue_entry *entry, void *offset)
764 CTR2(KTR_NTB, "TX: copying %d bytes to offset %p", entry->len, offset);
765 if (entry->buf != NULL) {
766 m_copydata((struct mbuf *)entry->buf, 0, entry->len, offset);
769 * Ensure that the data is fully copied before setting the
775 ntb_tx_copy_callback(entry);
779 ntb_async_tx(struct ntb_transport_qp *qp, struct ntb_queue_entry *entry)
781 struct ntb_payload_header *hdr;
784 offset = qp->tx_mw + qp->tx_max_frame * qp->tx_index;
785 hdr = (struct ntb_payload_header *)((char *)offset + qp->tx_max_frame -
786 sizeof(struct ntb_payload_header));
789 iowrite32(entry->len, &hdr->len);
790 iowrite32(qp->tx_pkts, &hdr->ver);
792 ntb_memcpy_tx(entry, offset);
796 ntb_process_tx(struct ntb_transport_qp *qp, struct ntb_queue_entry *entry)
800 "TX: process_tx: tx_pkts=%lu, tx_index=%u, remote entry=%u",
801 qp->tx_pkts, qp->tx_index, qp->remote_rx_info->entry);
802 if (qp->tx_index == qp->remote_rx_info->entry) {
803 CTR0(KTR_NTB, "TX: ring full");
808 if (entry->len > qp->tx_max_frame - sizeof(struct ntb_payload_header)) {
809 if (qp->tx_handler != NULL)
810 qp->tx_handler(qp, qp->cb_data, entry->buf,
816 ntb_list_add(&qp->ntb_tx_free_q_lock, entry, &qp->tx_free_q);
818 "TX: frame too big. returning entry %p to tx_free_q",
822 CTR2(KTR_NTB, "TX: copying entry %p to index %u", entry, qp->tx_index);
823 ntb_async_tx(qp, entry);
826 qp->tx_index %= qp->tx_max_entry;
835 ntb_transport_rxc_db(void *arg, int pending __unused)
837 struct ntb_transport_qp *qp = arg;
840 CTR0(KTR_NTB, "RX: transport_rx");
842 while ((rc = ntb_process_rxc(qp)) == 0)
844 CTR1(KTR_NTB, "RX: process_rxc returned %d", rc);
846 if ((ntb_db_read(qp->dev) & (1ull << qp->qp_num)) != 0) {
847 /* If db is set, clear it and check queue once more. */
848 ntb_db_clear(qp->dev, 1ull << qp->qp_num);
854 ntb_process_rxc(struct ntb_transport_qp *qp)
856 struct ntb_payload_header *hdr;
857 struct ntb_queue_entry *entry;
860 offset = qp->rx_buff + qp->rx_max_frame * qp->rx_index;
861 hdr = (void *)(offset + qp->rx_max_frame -
862 sizeof(struct ntb_payload_header));
864 CTR1(KTR_NTB, "RX: process_rxc rx_index = %u", qp->rx_index);
865 if ((hdr->flags & NTBT_DESC_DONE_FLAG) == 0) {
866 CTR0(KTR_NTB, "RX: hdr not done");
871 if ((hdr->flags & NTBT_LINK_DOWN_FLAG) != 0) {
872 CTR0(KTR_NTB, "RX: link down");
873 ntb_qp_link_down(qp);
878 if (hdr->ver != (uint32_t)qp->rx_pkts) {
879 CTR2(KTR_NTB,"RX: ver != rx_pkts (%x != %lx). "
880 "Returning entry to rx_pend_q", hdr->ver, qp->rx_pkts);
885 entry = ntb_list_mv(&qp->ntb_rx_q_lock, &qp->rx_pend_q, &qp->rx_post_q);
888 CTR0(KTR_NTB, "RX: No entries in rx_pend_q");
891 callout_stop(&qp->rx_full);
892 CTR1(KTR_NTB, "RX: rx entry %p from rx_pend_q", entry);
895 entry->index = qp->rx_index;
897 if (hdr->len > entry->len) {
898 CTR2(KTR_NTB, "RX: len too long. Wanted %ju got %ju",
899 (uintmax_t)hdr->len, (uintmax_t)entry->len);
903 entry->flags |= NTBT_DESC_DONE_FLAG;
905 ntb_complete_rxc(qp);
907 qp->rx_bytes += hdr->len;
910 CTR1(KTR_NTB, "RX: received %ld rx_pkts", qp->rx_pkts);
912 entry->len = hdr->len;
914 ntb_memcpy_rx(qp, entry, offset);
918 qp->rx_index %= qp->rx_max_entry;
923 ntb_memcpy_rx(struct ntb_transport_qp *qp, struct ntb_queue_entry *entry,
926 struct ifnet *ifp = entry->cb_data;
927 unsigned int len = entry->len;
929 CTR2(KTR_NTB, "RX: copying %d bytes from offset %p", len, offset);
931 entry->buf = (void *)m_devget(offset, len, 0, ifp, NULL);
932 if (entry->buf == NULL)
933 entry->len = -ENOMEM;
935 /* Ensure that the data is globally visible before clearing the flag */
938 CTR2(KTR_NTB, "RX: copied entry %p to mbuf %p.", entry, entry->buf);
939 ntb_rx_copy_callback(qp, entry);
943 ntb_rx_copy_callback(struct ntb_transport_qp *qp, void *data)
945 struct ntb_queue_entry *entry;
948 entry->flags |= NTBT_DESC_DONE_FLAG;
949 ntb_complete_rxc(qp);
953 ntb_complete_rxc(struct ntb_transport_qp *qp)
955 struct ntb_queue_entry *entry;
959 CTR0(KTR_NTB, "RX: rx_completion_task");
961 mtx_lock_spin(&qp->ntb_rx_q_lock);
963 while (!STAILQ_EMPTY(&qp->rx_post_q)) {
964 entry = STAILQ_FIRST(&qp->rx_post_q);
965 if ((entry->flags & NTBT_DESC_DONE_FLAG) == 0)
968 entry->x_hdr->flags = 0;
969 iowrite32(entry->index, &qp->rx_info->entry);
971 STAILQ_REMOVE_HEAD(&qp->rx_post_q, entry);
977 * Re-initialize queue_entry for reuse; rx_handler takes
978 * ownership of the mbuf.
981 entry->len = transport_mtu;
982 entry->cb_data = qp->cb_data;
984 STAILQ_INSERT_TAIL(&qp->rx_pend_q, entry, entry);
986 mtx_unlock_spin(&qp->ntb_rx_q_lock);
988 CTR2(KTR_NTB, "RX: completing entry %p, mbuf %p", entry, m);
989 if (qp->rx_handler != NULL && qp->client_ready)
990 qp->rx_handler(qp, qp->cb_data, m, len);
994 mtx_lock_spin(&qp->ntb_rx_q_lock);
997 mtx_unlock_spin(&qp->ntb_rx_q_lock);
1001 ntb_transport_doorbell_callback(void *data, uint32_t vector)
1003 struct ntb_transport_ctx *nt = data;
1004 struct ntb_transport_qp *qp;
1008 vec_mask = ntb_db_vector_mask(nt->dev, vector);
1009 vec_mask &= nt->qp_bitmap;
1010 if ((vec_mask & (vec_mask - 1)) != 0)
1011 vec_mask &= ntb_db_read(nt->dev);
1012 while (vec_mask != 0) {
1013 qp_num = ffsll(vec_mask) - 1;
1015 qp = &nt->qp_vec[qp_num];
1017 taskqueue_enqueue(qp->rxc_tq, &qp->rxc_db_work);
1019 vec_mask &= ~(1ull << qp_num);
1023 /* Link Event handler */
1025 ntb_transport_event_callback(void *data)
1027 struct ntb_transport_ctx *nt = data;
1029 if (ntb_link_is_up(nt->dev, &nt->link_speed, &nt->link_width)) {
1030 ntb_printf(1, "HW link up\n");
1031 callout_reset(&nt->link_work, 0, ntb_transport_link_work, nt);
1033 ntb_printf(1, "HW link down\n");
1034 taskqueue_enqueue(taskqueue_swi, &nt->link_cleanup);
1040 ntb_transport_link_work(void *arg)
1042 struct ntb_transport_ctx *nt = arg;
1043 device_t dev = nt->dev;
1044 struct ntb_transport_qp *qp;
1045 uint64_t val64, size;
1050 /* send the local info, in the opposite order of the way we read it */
1051 for (i = 0; i < nt->mw_count; i++) {
1052 size = nt->mw_vec[i].phys_size;
1054 if (max_mw_size != 0 && size > max_mw_size)
1057 ntb_peer_spad_write(dev, NTBT_MW0_SZ_HIGH + (i * 2),
1059 ntb_peer_spad_write(dev, NTBT_MW0_SZ_LOW + (i * 2), size);
1061 ntb_peer_spad_write(dev, NTBT_NUM_MWS, nt->mw_count);
1062 ntb_peer_spad_write(dev, NTBT_NUM_QPS, nt->qp_count);
1063 ntb_peer_spad_write(dev, NTBT_QP_LINKS, 0);
1064 ntb_peer_spad_write(dev, NTBT_VERSION, NTB_TRANSPORT_VERSION);
1066 /* Query the remote side for its info */
1068 ntb_spad_read(dev, NTBT_VERSION, &val);
1069 if (val != NTB_TRANSPORT_VERSION)
1072 ntb_spad_read(dev, NTBT_NUM_QPS, &val);
1073 if (val != nt->qp_count)
1076 ntb_spad_read(dev, NTBT_NUM_MWS, &val);
1077 if (val != nt->mw_count)
1080 for (i = 0; i < nt->mw_count; i++) {
1081 ntb_spad_read(dev, NTBT_MW0_SZ_HIGH + (i * 2), &val);
1082 val64 = (uint64_t)val << 32;
1084 ntb_spad_read(dev, NTBT_MW0_SZ_LOW + (i * 2), &val);
1087 rc = ntb_set_mw(nt, i, val64);
1092 nt->link_is_up = true;
1093 ntb_printf(1, "transport link up\n");
1095 for (i = 0; i < nt->qp_count; i++) {
1096 qp = &nt->qp_vec[i];
1098 ntb_transport_setup_qp_mw(nt, i);
1100 if (qp->client_ready)
1101 callout_reset(&qp->link_work, 0, ntb_qp_link_work, qp);
1107 for (i = 0; i < nt->mw_count; i++)
1110 if (ntb_link_is_up(dev, &nt->link_speed, &nt->link_width))
1111 callout_reset(&nt->link_work,
1112 NTB_LINK_DOWN_TIMEOUT * hz / 1000, ntb_transport_link_work, nt);
1116 ntb_set_mw(struct ntb_transport_ctx *nt, int num_mw, size_t size)
1118 struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
1119 size_t xlat_size, buff_size;
1125 xlat_size = roundup(size, mw->xlat_align_size);
1126 buff_size = xlat_size;
1128 /* No need to re-setup */
1129 if (mw->xlat_size == xlat_size)
1132 if (mw->buff_size != 0)
1133 ntb_free_mw(nt, num_mw);
1135 /* Alloc memory for receiving data. Must be aligned */
1136 mw->xlat_size = xlat_size;
1137 mw->buff_size = buff_size;
1139 mw->virt_addr = contigmalloc(mw->buff_size, M_NTB_T, M_ZERO, 0,
1140 mw->addr_limit, mw->xlat_align, 0);
1141 if (mw->virt_addr == NULL) {
1142 ntb_printf(0, "Unable to allocate MW buffer of size %zu/%zu\n",
1143 mw->buff_size, mw->xlat_size);
1148 /* TODO: replace with bus_space_* functions */
1149 mw->dma_addr = vtophys(mw->virt_addr);
1152 * Ensure that the allocation from contigmalloc is aligned as
1153 * requested. XXX: This may not be needed -- brought in for parity
1154 * with the Linux driver.
1156 if (mw->dma_addr % mw->xlat_align != 0) {
1158 "DMA memory 0x%jx not aligned to BAR size 0x%zx\n",
1159 (uintmax_t)mw->dma_addr, size);
1160 ntb_free_mw(nt, num_mw);
1164 /* Notify HW the memory location of the receive buffer */
1165 rc = ntb_mw_set_trans(nt->dev, num_mw, mw->dma_addr, mw->xlat_size);
1167 ntb_printf(0, "Unable to set mw%d translation\n", num_mw);
1168 ntb_free_mw(nt, num_mw);
1176 ntb_free_mw(struct ntb_transport_ctx *nt, int num_mw)
1178 struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
1180 if (mw->virt_addr == NULL)
1183 ntb_mw_clear_trans(nt->dev, num_mw);
1184 contigfree(mw->virt_addr, mw->xlat_size, M_NTB_T);
1187 mw->virt_addr = NULL;
1191 ntb_transport_setup_qp_mw(struct ntb_transport_ctx *nt, unsigned int qp_num)
1193 struct ntb_transport_qp *qp = &nt->qp_vec[qp_num];
1194 struct ntb_transport_mw *mw;
1198 unsigned num_qps_mw, mw_num, mw_count;
1200 mw_count = nt->mw_count;
1201 mw_num = QP_TO_MW(nt, qp_num);
1202 mw = &nt->mw_vec[mw_num];
1204 if (mw->virt_addr == NULL)
1207 if (mw_num < nt->qp_count % mw_count)
1208 num_qps_mw = nt->qp_count / mw_count + 1;
1210 num_qps_mw = nt->qp_count / mw_count;
1212 rx_size = mw->xlat_size / num_qps_mw;
1213 qp->rx_buff = mw->virt_addr + rx_size * (qp_num / mw_count);
1214 rx_size -= sizeof(struct ntb_rx_info);
1216 qp->remote_rx_info = (void*)(qp->rx_buff + rx_size);
1218 /* Due to house-keeping, there must be at least 2 buffs */
1219 qp->rx_max_frame = qmin(transport_mtu, rx_size / 2);
1220 qp->rx_max_entry = rx_size / qp->rx_max_frame;
1223 qp->remote_rx_info->entry = qp->rx_max_entry - 1;
1225 /* Set up the hdr offsets with 0s */
1226 for (i = 0; i < qp->rx_max_entry; i++) {
1227 offset = (void *)(qp->rx_buff + qp->rx_max_frame * (i + 1) -
1228 sizeof(struct ntb_payload_header));
1229 memset(offset, 0, sizeof(struct ntb_payload_header));
1240 ntb_qp_link_work(void *arg)
1242 struct ntb_transport_qp *qp = arg;
1243 device_t dev = qp->dev;
1244 struct ntb_transport_ctx *nt = qp->transport;
1248 /* Report queues that are up on our side */
1249 for (i = 0, val = 0; i < nt->qp_count; i++) {
1250 if (nt->qp_vec[i].client_ready)
1253 ntb_peer_spad_write(dev, NTBT_QP_LINKS, val);
1255 /* See if the remote side is up */
1256 ntb_spad_read(dev, NTBT_QP_LINKS, &val);
1257 if ((val & (1ull << qp->qp_num)) != 0) {
1258 ntb_printf(2, "qp %d link up\n", qp->qp_num);
1259 qp->link_is_up = true;
1261 if (qp->event_handler != NULL)
1262 qp->event_handler(qp->cb_data, NTB_LINK_UP);
1264 ntb_db_clear_mask(dev, 1ull << qp->qp_num);
1265 } else if (nt->link_is_up)
1266 callout_reset(&qp->link_work,
1267 NTB_LINK_DOWN_TIMEOUT * hz / 1000, ntb_qp_link_work, qp);
1270 /* Link down event*/
1272 ntb_transport_link_cleanup(struct ntb_transport_ctx *nt)
1274 struct ntb_transport_qp *qp;
1277 /* Pass along the info to any clients */
1278 for (i = 0; i < nt->qp_count; i++) {
1279 if ((nt->qp_bitmap & (1 << i)) != 0) {
1280 qp = &nt->qp_vec[i];
1281 ntb_qp_link_cleanup(qp);
1282 callout_drain(&qp->link_work);
1286 if (!nt->link_is_up)
1287 callout_drain(&nt->link_work);
1290 * The scratchpad registers keep the values if the remote side
1291 * goes down, blast them now to give them a sane value the next
1292 * time they are accessed
1294 ntb_spad_clear(nt->dev);
1298 ntb_transport_link_cleanup_work(void *arg, int pending __unused)
1301 ntb_transport_link_cleanup(arg);
1305 ntb_qp_link_down(struct ntb_transport_qp *qp)
1308 ntb_qp_link_cleanup(qp);
1312 ntb_qp_link_down_reset(struct ntb_transport_qp *qp)
1315 qp->link_is_up = false;
1316 ntb_db_set_mask(qp->dev, 1ull << qp->qp_num);
1318 qp->tx_index = qp->rx_index = 0;
1319 qp->tx_bytes = qp->rx_bytes = 0;
1320 qp->tx_pkts = qp->rx_pkts = 0;
1322 qp->rx_ring_empty = 0;
1323 qp->tx_ring_full = 0;
1325 qp->rx_err_no_buf = qp->tx_err_no_buf = 0;
1326 qp->rx_err_oflow = qp->rx_err_ver = 0;
1330 ntb_qp_link_cleanup(struct ntb_transport_qp *qp)
1333 callout_drain(&qp->link_work);
1334 ntb_qp_link_down_reset(qp);
1336 if (qp->event_handler != NULL)
1337 qp->event_handler(qp->cb_data, NTB_LINK_DOWN);
1340 /* Link commanded down */
1342 * ntb_transport_link_down - Notify NTB transport to no longer enqueue data
1343 * @qp: NTB transport layer queue to be disabled
1345 * Notify NTB transport layer of client's desire to no longer receive data on
1346 * transport queue specified. It is the client's responsibility to ensure all
1347 * entries on queue are purged or otherwise handled appropriately.
1350 ntb_transport_link_down(struct ntb_transport_qp *qp)
1352 struct ntb_transport_ctx *nt = qp->transport;
1356 qp->client_ready = false;
1357 for (i = 0, val = 0; i < nt->qp_count; i++) {
1358 if (nt->qp_vec[i].client_ready)
1361 ntb_peer_spad_write(qp->dev, NTBT_QP_LINKS, val);
1364 ntb_send_link_down(qp);
1366 callout_drain(&qp->link_work);
1370 * ntb_transport_link_query - Query transport link state
1371 * @qp: NTB transport layer queue to be queried
1373 * Query connectivity to the remote system of the NTB transport queue
1375 * RETURNS: true for link up or false for link down
1378 ntb_transport_link_query(struct ntb_transport_qp *qp)
1381 return (qp->link_is_up);
1385 * ntb_transport_link_speed - Query transport link speed
1386 * @qp: NTB transport layer queue to be queried
1388 * Query connection speed to the remote system of the NTB transport queue
1390 * RETURNS: link speed in bits per second
1393 ntb_transport_link_speed(struct ntb_transport_qp *qp)
1395 struct ntb_transport_ctx *nt = qp->transport;
1398 if (!nt->link_is_up)
1400 switch (nt->link_speed) {
1401 case NTB_SPEED_GEN1:
1402 rate = 2500000000 * 8 / 10;
1404 case NTB_SPEED_GEN2:
1405 rate = 5000000000 * 8 / 10;
1407 case NTB_SPEED_GEN3:
1408 rate = 8000000000 * 128 / 130;
1410 case NTB_SPEED_GEN4:
1411 rate = 16000000000 * 128 / 130;
1416 if (nt->link_width <= 0)
1418 return (rate * nt->link_width);
1422 ntb_send_link_down(struct ntb_transport_qp *qp)
1424 struct ntb_queue_entry *entry;
1427 if (!qp->link_is_up)
1430 for (i = 0; i < NTB_LINK_DOWN_TIMEOUT; i++) {
1431 entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
1434 pause("NTB Wait for link down", hz / 10);
1440 entry->cb_data = NULL;
1443 entry->flags = NTBT_LINK_DOWN_FLAG;
1445 mtx_lock(&qp->tx_lock);
1446 rc = ntb_process_tx(qp, entry);
1447 mtx_unlock(&qp->tx_lock);
1449 printf("ntb: Failed to send link down\n");
1451 ntb_qp_link_down_reset(qp);
1455 /* List Management */
1458 ntb_list_add(struct mtx *lock, struct ntb_queue_entry *entry,
1459 struct ntb_queue_list *list)
1462 mtx_lock_spin(lock);
1463 STAILQ_INSERT_TAIL(list, entry, entry);
1464 mtx_unlock_spin(lock);
1467 static struct ntb_queue_entry *
1468 ntb_list_rm(struct mtx *lock, struct ntb_queue_list *list)
1470 struct ntb_queue_entry *entry;
1472 mtx_lock_spin(lock);
1473 if (STAILQ_EMPTY(list)) {
1477 entry = STAILQ_FIRST(list);
1478 STAILQ_REMOVE_HEAD(list, entry);
1480 mtx_unlock_spin(lock);
1485 static struct ntb_queue_entry *
1486 ntb_list_mv(struct mtx *lock, struct ntb_queue_list *from,
1487 struct ntb_queue_list *to)
1489 struct ntb_queue_entry *entry;
1491 mtx_lock_spin(lock);
1492 if (STAILQ_EMPTY(from)) {
1496 entry = STAILQ_FIRST(from);
1497 STAILQ_REMOVE_HEAD(from, entry);
1498 STAILQ_INSERT_TAIL(to, entry, entry);
1501 mtx_unlock_spin(lock);
1506 * ntb_transport_qp_num - Query the qp number
1507 * @qp: NTB transport layer queue to be queried
1509 * Query qp number of the NTB transport queue
1511 * RETURNS: a zero based number specifying the qp number
1513 unsigned char ntb_transport_qp_num(struct ntb_transport_qp *qp)
1516 return (qp->qp_num);
1520 * ntb_transport_max_size - Query the max payload size of a qp
1521 * @qp: NTB transport layer queue to be queried
1523 * Query the maximum payload size permissible on the given qp
1525 * RETURNS: the max payload size of a qp
1528 ntb_transport_max_size(struct ntb_transport_qp *qp)
1531 return (qp->tx_max_frame - sizeof(struct ntb_payload_header));
1535 ntb_transport_tx_free_entry(struct ntb_transport_qp *qp)
1537 unsigned int head = qp->tx_index;
1538 unsigned int tail = qp->remote_rx_info->entry;
1540 return (tail >= head ? tail - head : qp->tx_max_entry + tail - head);
1543 static device_method_t ntb_transport_methods[] = {
1544 /* Device interface */
1545 DEVMETHOD(device_probe, ntb_transport_probe),
1546 DEVMETHOD(device_attach, ntb_transport_attach),
1547 DEVMETHOD(device_detach, ntb_transport_detach),
1551 devclass_t ntb_transport_devclass;
1552 static DEFINE_CLASS_0(ntb_transport, ntb_transport_driver,
1553 ntb_transport_methods, sizeof(struct ntb_transport_ctx));
1554 DRIVER_MODULE(ntb_transport, ntb_hw, ntb_transport_driver,
1555 ntb_transport_devclass, NULL, NULL);
1556 MODULE_DEPEND(ntb_transport, ntb, 1, 1, 1);
1557 MODULE_VERSION(ntb_transport, 1);