xref: /freebsd/sys/dev/ntb/ntb_transport.c (revision cbd30a72ca196976c1c700400ecd424baa1b9c16)
1 /*-
2  * Copyright (c) 2016 Alexander Motin <mav@FreeBSD.org>
3  * Copyright (C) 2013 Intel Corporation
4  * Copyright (C) 2015 EMC Corporation
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
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.
15  *
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
26  * SUCH DAMAGE.
27  */
28 
29 /*
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.
32  *
33  * This module contains a transport for sending and receiving messages by
34  * writing to remote memory window(s) provided by underlying NTB device.
35  *
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.
38  */
39 
40 #include <sys/cdefs.h>
41 __FBSDID("$FreeBSD$");
42 
43 #include <sys/param.h>
44 #include <sys/kernel.h>
45 #include <sys/systm.h>
46 #include <sys/bus.h>
47 #include <sys/ktr.h>
48 #include <sys/limits.h>
49 #include <sys/lock.h>
50 #include <sys/malloc.h>
51 #include <sys/mbuf.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>
57 
58 #include <vm/vm.h>
59 #include <vm/pmap.h>
60 
61 #include <machine/bus.h>
62 
63 #include "ntb.h"
64 #include "ntb_transport.h"
65 
66 #define KTR_NTB KTR_SPARE3
67 
68 #define NTB_TRANSPORT_VERSION	4
69 
70 static SYSCTL_NODE(_hw, OID_AUTO, ntb_transport, CTLFLAG_RW, 0, "ntb_transport");
71 
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__);			\
79 	}						\
80 } while (0)
81 
82 static unsigned transport_mtu = 0x10000;
83 
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.");
88 
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");
93 
94 STAILQ_HEAD(ntb_queue_list, ntb_queue_entry);
95 
96 typedef uint32_t ntb_q_idx_t;
97 
98 struct ntb_queue_entry {
99 	/* ntb_queue list reference */
100 	STAILQ_ENTRY(ntb_queue_entry) entry;
101 
102 	/* info on data to be transferred */
103 	void		*cb_data;
104 	void		*buf;
105 	uint32_t	len;
106 	uint32_t	flags;
107 
108 	struct ntb_transport_qp		*qp;
109 	struct ntb_payload_header	*x_hdr;
110 	ntb_q_idx_t	index;
111 };
112 
113 struct ntb_rx_info {
114 	ntb_q_idx_t	entry;
115 };
116 
117 struct ntb_transport_qp {
118 	struct ntb_transport_ctx	*transport;
119 	device_t		 dev;
120 
121 	void			*cb_data;
122 
123 	bool			client_ready;
124 	volatile bool		link_is_up;
125 	uint8_t			qp_num;	/* Only 64 QPs are allowed.  0-63 */
126 
127 	struct ntb_rx_info	*rx_info;
128 	struct ntb_rx_info	*remote_rx_info;
129 
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;
134 	caddr_t			tx_mw;
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;
139 
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;
148 	caddr_t			rx_buff;
149 	ntb_q_idx_t		rx_index;
150 	ntb_q_idx_t		rx_max_entry;
151 	uint64_t		rx_max_frame;
152 
153 	void (*event_handler)(void *data, enum ntb_link_event status);
154 	struct callout		link_work;
155 	struct callout		rx_full;
156 
157 	uint64_t		last_rx_no_buf;
158 
159 	/* Stats */
160 	uint64_t		rx_bytes;
161 	uint64_t		rx_pkts;
162 	uint64_t		rx_ring_empty;
163 	uint64_t		rx_err_no_buf;
164 	uint64_t		rx_err_oflow;
165 	uint64_t		rx_err_ver;
166 	uint64_t		tx_bytes;
167 	uint64_t		tx_pkts;
168 	uint64_t		tx_ring_full;
169 	uint64_t		tx_err_no_buf;
170 
171 	struct mtx		tx_lock;
172 };
173 
174 struct ntb_transport_mw {
175 	vm_paddr_t	phys_addr;
176 	size_t		phys_size;
177 	size_t		xlat_align;
178 	size_t		xlat_align_size;
179 	bus_addr_t	addr_limit;
180 	/* Tx buff is off vbase / phys_addr */
181 	caddr_t		vbase;
182 	size_t		xlat_size;
183 	size_t		buff_size;
184 	/* Rx buff is off virt_addr / dma_addr */
185 	caddr_t		virt_addr;
186 	bus_addr_t	dma_addr;
187 };
188 
189 struct ntb_transport_child {
190 	device_t	dev;
191 	int		qpoff;
192 	int		qpcnt;
193 	struct ntb_transport_child *next;
194 };
195 
196 struct ntb_transport_ctx {
197 	device_t		 dev;
198 	struct ntb_transport_child *child;
199 	struct ntb_transport_mw	*mw_vec;
200 	struct ntb_transport_qp	*qp_vec;
201 	unsigned		mw_count;
202 	unsigned		qp_count;
203 	uint64_t		qp_bitmap;
204 	volatile bool		link_is_up;
205 	struct callout		link_work;
206 	struct callout		link_watchdog;
207 	struct task		link_cleanup;
208 };
209 
210 enum {
211 	NTBT_DESC_DONE_FLAG = 1 << 0,
212 	NTBT_LINK_DOWN_FLAG = 1 << 1,
213 };
214 
215 struct ntb_payload_header {
216 	ntb_q_idx_t ver;
217 	uint32_t len;
218 	uint32_t flags;
219 };
220 
221 enum {
222 	/*
223 	 * The order of this enum is part of the remote protocol.  Do not
224 	 * reorder without bumping protocol version (and it's probably best
225 	 * to keep the protocol in lock-step with the Linux NTB driver.
226 	 */
227 	NTBT_VERSION = 0,
228 	NTBT_QP_LINKS,
229 	NTBT_NUM_QPS,
230 	NTBT_NUM_MWS,
231 	/*
232 	 * N.B.: transport_link_work assumes MW1 enums = MW0 + 2.
233 	 */
234 	NTBT_MW0_SZ_HIGH,
235 	NTBT_MW0_SZ_LOW,
236 	NTBT_MW1_SZ_HIGH,
237 	NTBT_MW1_SZ_LOW,
238 
239 	/*
240 	 * Some NTB-using hardware have a watchdog to work around NTB hangs; if
241 	 * a register or doorbell isn't written every few seconds, the link is
242 	 * torn down.  Write an otherwise unused register every few seconds to
243 	 * work around this watchdog.
244 	 */
245 	NTBT_WATCHDOG_SPAD = 15
246 };
247 
248 #define QP_TO_MW(nt, qp)	((qp) % nt->mw_count)
249 #define NTB_QP_DEF_NUM_ENTRIES	100
250 #define NTB_LINK_DOWN_TIMEOUT	10
251 
252 static int ntb_transport_probe(device_t dev);
253 static int ntb_transport_attach(device_t dev);
254 static int ntb_transport_detach(device_t dev);
255 static void ntb_transport_init_queue(struct ntb_transport_ctx *nt,
256     unsigned int qp_num);
257 static int ntb_process_tx(struct ntb_transport_qp *qp,
258     struct ntb_queue_entry *entry);
259 static void ntb_transport_rxc_db(void *arg, int pending);
260 static int ntb_process_rxc(struct ntb_transport_qp *qp);
261 static void ntb_memcpy_rx(struct ntb_transport_qp *qp,
262     struct ntb_queue_entry *entry, void *offset);
263 static inline void ntb_rx_copy_callback(struct ntb_transport_qp *qp,
264     void *data);
265 static void ntb_complete_rxc(struct ntb_transport_qp *qp);
266 static void ntb_transport_doorbell_callback(void *data, uint32_t vector);
267 static void ntb_transport_event_callback(void *data);
268 static void ntb_transport_link_work(void *arg);
269 static int ntb_set_mw(struct ntb_transport_ctx *, int num_mw, size_t size);
270 static void ntb_free_mw(struct ntb_transport_ctx *nt, int num_mw);
271 static int ntb_transport_setup_qp_mw(struct ntb_transport_ctx *nt,
272     unsigned int qp_num);
273 static void ntb_qp_link_work(void *arg);
274 static void ntb_transport_link_cleanup(struct ntb_transport_ctx *nt);
275 static void ntb_transport_link_cleanup_work(void *, int);
276 static void ntb_qp_link_down(struct ntb_transport_qp *qp);
277 static void ntb_qp_link_down_reset(struct ntb_transport_qp *qp);
278 static void ntb_qp_link_cleanup(struct ntb_transport_qp *qp);
279 static void ntb_send_link_down(struct ntb_transport_qp *qp);
280 static void ntb_list_add(struct mtx *lock, struct ntb_queue_entry *entry,
281     struct ntb_queue_list *list);
282 static struct ntb_queue_entry *ntb_list_rm(struct mtx *lock,
283     struct ntb_queue_list *list);
284 static struct ntb_queue_entry *ntb_list_mv(struct mtx *lock,
285     struct ntb_queue_list *from, struct ntb_queue_list *to);
286 static void xeon_link_watchdog_hb(void *);
287 
288 static const struct ntb_ctx_ops ntb_transport_ops = {
289 	.link_event = ntb_transport_event_callback,
290 	.db_event = ntb_transport_doorbell_callback,
291 };
292 
293 MALLOC_DEFINE(M_NTB_T, "ntb_transport", "ntb transport driver");
294 
295 static inline void
296 iowrite32(uint32_t val, void *addr)
297 {
298 
299 	bus_space_write_4(X86_BUS_SPACE_MEM, 0/* HACK */, (uintptr_t)addr,
300 	    val);
301 }
302 
303 /* Transport Init and teardown */
304 
305 static void
306 xeon_link_watchdog_hb(void *arg)
307 {
308 	struct ntb_transport_ctx *nt;
309 
310 	nt = arg;
311 	ntb_spad_write(nt->dev, NTBT_WATCHDOG_SPAD, 0);
312 	callout_reset(&nt->link_watchdog, 1 * hz, xeon_link_watchdog_hb, nt);
313 }
314 
315 static int
316 ntb_transport_probe(device_t dev)
317 {
318 
319 	device_set_desc(dev, "NTB Transport");
320 	return (0);
321 }
322 
323 static int
324 ntb_transport_attach(device_t dev)
325 {
326 	struct ntb_transport_ctx *nt = device_get_softc(dev);
327 	struct ntb_transport_child **cpp = &nt->child;
328 	struct ntb_transport_child *nc;
329 	struct ntb_transport_mw *mw;
330 	uint64_t db_bitmap;
331 	int rc, i, db_count, spad_count, qp, qpu, qpo, qpt;
332 	char cfg[128] = "";
333 	char buf[32];
334 	char *n, *np, *c, *name;
335 
336 	nt->dev = dev;
337 	nt->mw_count = ntb_mw_count(dev);
338 	spad_count = ntb_spad_count(dev);
339 	db_bitmap = ntb_db_valid_mask(dev);
340 	db_count = flsll(db_bitmap);
341 	KASSERT(db_bitmap == (1 << db_count) - 1,
342 	    ("Doorbells are not sequential (%jx).\n", db_bitmap));
343 
344 	device_printf(dev, "%d memory windows, %d scratchpads, "
345 	    "%d doorbells\n", nt->mw_count, spad_count, db_count);
346 
347 	if (nt->mw_count == 0) {
348 		device_printf(dev, "At least 1 memory window required.\n");
349 		return (ENXIO);
350 	}
351 	if (spad_count < 6) {
352 		device_printf(dev, "At least 6 scratchpads required.\n");
353 		return (ENXIO);
354 	}
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);
359 	}
360 	if (db_bitmap == 0) {
361 		device_printf(dev, "At least one doorbell required.\n");
362 		return (ENXIO);
363 	}
364 
365 	nt->mw_vec = malloc(nt->mw_count * sizeof(*nt->mw_vec), M_NTB_T,
366 	    M_WAITOK | M_ZERO);
367 	for (i = 0; i < nt->mw_count; i++) {
368 		mw = &nt->mw_vec[i];
369 
370 		rc = ntb_mw_get_range(dev, i, &mw->phys_addr, &mw->vbase,
371 		    &mw->phys_size, &mw->xlat_align, &mw->xlat_align_size,
372 		    &mw->addr_limit);
373 		if (rc != 0)
374 			goto err;
375 
376 		mw->buff_size = 0;
377 		mw->xlat_size = 0;
378 		mw->virt_addr = NULL;
379 		mw->dma_addr = 0;
380 
381 		rc = ntb_mw_set_wc(dev, i, VM_MEMATTR_WRITE_COMBINING);
382 		if (rc)
383 			ntb_printf(0, "Unable to set mw%d caching\n", i);
384 	}
385 
386 	qpu = 0;
387 	qpo = imin(db_count, nt->mw_count);
388 	qpt = db_count;
389 
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));
393 	n = cfg;
394 	i = 0;
395 	while ((c = strsep(&n, ",")) != NULL) {
396 		np = c;
397 		name = strsep(&np, ":");
398 		if (name != NULL && name[0] == 0)
399 			name = NULL;
400 		qp = (np && np[0] != 0) ? strtol(np, NULL, 10) : qpo - qpu;
401 		if (qp <= 0)
402 			qp = 1;
403 
404 		if (qp > qpt - qpu) {
405 			device_printf(dev, "Not enough resources for config\n");
406 			break;
407 		}
408 
409 		nc = malloc(sizeof(*nc), M_DEVBUF, M_WAITOK | M_ZERO);
410 		nc->qpoff = qpu;
411 		nc->qpcnt = qp;
412 		nc->dev = device_add_child(dev, name, -1);
413 		if (nc->dev == NULL) {
414 			device_printf(dev, "Can not add child.\n");
415 			break;
416 		}
417 		device_set_ivars(nc->dev, nc);
418 		*cpp = nc;
419 		cpp = &nc->next;
420 
421 		if (bootverbose) {
422 			device_printf(dev, "%d \"%s\": queues %d",
423 			    i, name, qpu);
424 			if (qp > 1)
425 				printf("-%d", qpu + qp - 1);
426 			printf("\n");
427 		}
428 
429 		qpu += qp;
430 		i++;
431 	}
432 	nt->qp_count = qpu;
433 
434 	nt->qp_vec = malloc(nt->qp_count * sizeof(*nt->qp_vec), M_NTB_T,
435 	    M_WAITOK | M_ZERO);
436 
437 	for (i = 0; i < nt->qp_count; i++)
438 		ntb_transport_init_queue(nt, i);
439 
440 	callout_init(&nt->link_work, 0);
441 	callout_init(&nt->link_watchdog, 0);
442 	TASK_INIT(&nt->link_cleanup, 0, ntb_transport_link_cleanup_work, nt);
443 
444 	rc = ntb_set_ctx(dev, nt, &ntb_transport_ops);
445 	if (rc != 0)
446 		goto err;
447 
448 	nt->link_is_up = false;
449 	ntb_link_enable(dev, NTB_SPEED_AUTO, NTB_WIDTH_AUTO);
450 
451 	if (enable_xeon_watchdog != 0)
452 		callout_reset(&nt->link_watchdog, 0, xeon_link_watchdog_hb, nt);
453 
454 	bus_generic_attach(dev);
455 	return (0);
456 
457 err:
458 	free(nt->qp_vec, M_NTB_T);
459 	free(nt->mw_vec, M_NTB_T);
460 	return (rc);
461 }
462 
463 static int
464 ntb_transport_detach(device_t dev)
465 {
466 	struct ntb_transport_ctx *nt = device_get_softc(dev);
467 	struct ntb_transport_child **cpp = &nt->child;
468 	struct ntb_transport_child *nc;
469 	int error = 0, i;
470 
471 	while ((nc = *cpp) != NULL) {
472 		*cpp = (*cpp)->next;
473 		error = device_delete_child(dev, nc->dev);
474 		if (error)
475 			break;
476 		free(nc, M_DEVBUF);
477 	}
478 	KASSERT(nt->qp_bitmap == 0,
479 	    ("Some queues not freed on detach (%jx)", nt->qp_bitmap));
480 
481 	ntb_transport_link_cleanup(nt);
482 	taskqueue_drain(taskqueue_swi, &nt->link_cleanup);
483 	callout_drain(&nt->link_work);
484 	callout_drain(&nt->link_watchdog);
485 
486 	ntb_link_disable(dev);
487 	ntb_clear_ctx(dev);
488 
489 	for (i = 0; i < nt->mw_count; i++)
490 		ntb_free_mw(nt, i);
491 
492 	free(nt->qp_vec, M_NTB_T);
493 	free(nt->mw_vec, M_NTB_T);
494 	return (0);
495 }
496 
497 int
498 ntb_transport_queue_count(device_t dev)
499 {
500 	struct ntb_transport_child *nc = device_get_ivars(dev);
501 
502 	return (nc->qpcnt);
503 }
504 
505 static void
506 ntb_transport_init_queue(struct ntb_transport_ctx *nt, unsigned int qp_num)
507 {
508 	struct ntb_transport_mw *mw;
509 	struct ntb_transport_qp *qp;
510 	vm_paddr_t mw_base;
511 	uint64_t mw_size, qp_offset;
512 	size_t tx_size;
513 	unsigned num_qps_mw, mw_num, mw_count;
514 
515 	mw_count = nt->mw_count;
516 	mw_num = QP_TO_MW(nt, qp_num);
517 	mw = &nt->mw_vec[mw_num];
518 
519 	qp = &nt->qp_vec[qp_num];
520 	qp->qp_num = qp_num;
521 	qp->transport = nt;
522 	qp->dev = nt->dev;
523 	qp->client_ready = false;
524 	qp->event_handler = NULL;
525 	ntb_qp_link_down_reset(qp);
526 
527 	if (mw_num < nt->qp_count % mw_count)
528 		num_qps_mw = nt->qp_count / mw_count + 1;
529 	else
530 		num_qps_mw = nt->qp_count / mw_count;
531 
532 	mw_base = mw->phys_addr;
533 	mw_size = mw->phys_size;
534 
535 	tx_size = mw_size / num_qps_mw;
536 	qp_offset = tx_size * (qp_num / mw_count);
537 
538 	qp->tx_mw = mw->vbase + qp_offset;
539 	KASSERT(qp->tx_mw != NULL, ("uh oh?"));
540 
541 	/* XXX Assumes that a vm_paddr_t is equivalent to bus_addr_t */
542 	qp->tx_mw_phys = mw_base + qp_offset;
543 	KASSERT(qp->tx_mw_phys != 0, ("uh oh?"));
544 
545 	tx_size -= sizeof(struct ntb_rx_info);
546 	qp->rx_info = (void *)(qp->tx_mw + tx_size);
547 
548 	/* Due to house-keeping, there must be at least 2 buffs */
549 	qp->tx_max_frame = qmin(transport_mtu, tx_size / 2);
550 	qp->tx_max_entry = tx_size / qp->tx_max_frame;
551 
552 	callout_init(&qp->link_work, 0);
553 	callout_init(&qp->rx_full, 1);
554 
555 	mtx_init(&qp->ntb_rx_q_lock, "ntb rx q", NULL, MTX_SPIN);
556 	mtx_init(&qp->ntb_tx_free_q_lock, "ntb tx free q", NULL, MTX_SPIN);
557 	mtx_init(&qp->tx_lock, "ntb transport tx", NULL, MTX_DEF);
558 	TASK_INIT(&qp->rxc_db_work, 0, ntb_transport_rxc_db, qp);
559 	qp->rxc_tq = taskqueue_create("ntbt_rx", M_WAITOK,
560 	    taskqueue_thread_enqueue, &qp->rxc_tq);
561 	taskqueue_start_threads(&qp->rxc_tq, 1, PI_NET, "%s rx%d",
562 	    device_get_nameunit(nt->dev), qp_num);
563 
564 	STAILQ_INIT(&qp->rx_post_q);
565 	STAILQ_INIT(&qp->rx_pend_q);
566 	STAILQ_INIT(&qp->tx_free_q);
567 }
568 
569 void
570 ntb_transport_free_queue(struct ntb_transport_qp *qp)
571 {
572 	struct ntb_transport_ctx *nt = qp->transport;
573 	struct ntb_queue_entry *entry;
574 
575 	callout_drain(&qp->link_work);
576 
577 	ntb_db_set_mask(qp->dev, 1ull << qp->qp_num);
578 	taskqueue_drain_all(qp->rxc_tq);
579 	taskqueue_free(qp->rxc_tq);
580 
581 	qp->cb_data = NULL;
582 	qp->rx_handler = NULL;
583 	qp->tx_handler = NULL;
584 	qp->event_handler = NULL;
585 
586 	while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_pend_q)))
587 		free(entry, M_NTB_T);
588 
589 	while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_post_q)))
590 		free(entry, M_NTB_T);
591 
592 	while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
593 		free(entry, M_NTB_T);
594 
595 	nt->qp_bitmap &= ~(1 << qp->qp_num);
596 }
597 
598 /**
599  * ntb_transport_create_queue - Create a new NTB transport layer queue
600  * @rx_handler: receive callback function
601  * @tx_handler: transmit callback function
602  * @event_handler: event callback function
603  *
604  * Create a new NTB transport layer queue and provide the queue with a callback
605  * routine for both transmit and receive.  The receive callback routine will be
606  * used to pass up data when the transport has received it on the queue.   The
607  * transmit callback routine will be called when the transport has completed the
608  * transmission of the data on the queue and the data is ready to be freed.
609  *
610  * RETURNS: pointer to newly created ntb_queue, NULL on error.
611  */
612 struct ntb_transport_qp *
613 ntb_transport_create_queue(device_t dev, int q,
614     const struct ntb_queue_handlers *handlers, void *data)
615 {
616 	struct ntb_transport_child *nc = device_get_ivars(dev);
617 	struct ntb_transport_ctx *nt = device_get_softc(device_get_parent(dev));
618 	struct ntb_queue_entry *entry;
619 	struct ntb_transport_qp *qp;
620 	int i;
621 
622 	if (q < 0 || q >= nc->qpcnt)
623 		return (NULL);
624 
625 	qp = &nt->qp_vec[nc->qpoff + q];
626 	nt->qp_bitmap |= (1 << qp->qp_num);
627 	qp->cb_data = data;
628 	qp->rx_handler = handlers->rx_handler;
629 	qp->tx_handler = handlers->tx_handler;
630 	qp->event_handler = handlers->event_handler;
631 
632 	for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) {
633 		entry = malloc(sizeof(*entry), M_NTB_T, M_WAITOK | M_ZERO);
634 		entry->cb_data = data;
635 		entry->buf = NULL;
636 		entry->len = transport_mtu;
637 		entry->qp = qp;
638 		ntb_list_add(&qp->ntb_rx_q_lock, entry, &qp->rx_pend_q);
639 	}
640 
641 	for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) {
642 		entry = malloc(sizeof(*entry), M_NTB_T, M_WAITOK | M_ZERO);
643 		entry->qp = qp;
644 		ntb_list_add(&qp->ntb_tx_free_q_lock, entry, &qp->tx_free_q);
645 	}
646 
647 	ntb_db_clear(dev, 1ull << qp->qp_num);
648 	return (qp);
649 }
650 
651 /**
652  * ntb_transport_link_up - Notify NTB transport of client readiness to use queue
653  * @qp: NTB transport layer queue to be enabled
654  *
655  * Notify NTB transport layer of client readiness to use queue
656  */
657 void
658 ntb_transport_link_up(struct ntb_transport_qp *qp)
659 {
660 	struct ntb_transport_ctx *nt = qp->transport;
661 
662 	qp->client_ready = true;
663 
664 	ntb_printf(2, "qp %d client ready\n", qp->qp_num);
665 
666 	if (nt->link_is_up)
667 		callout_reset(&qp->link_work, 0, ntb_qp_link_work, qp);
668 }
669 
670 
671 
672 /* Transport Tx */
673 
674 /**
675  * ntb_transport_tx_enqueue - Enqueue a new NTB queue entry
676  * @qp: NTB transport layer queue the entry is to be enqueued on
677  * @cb: per buffer pointer for callback function to use
678  * @data: pointer to data buffer that will be sent
679  * @len: length of the data buffer
680  *
681  * Enqueue a new transmit buffer onto the transport queue from which a NTB
682  * payload will be transmitted.  This assumes that a lock is being held to
683  * serialize access to the qp.
684  *
685  * RETURNS: An appropriate ERRNO error value on error, or zero for success.
686  */
687 int
688 ntb_transport_tx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
689     unsigned int len)
690 {
691 	struct ntb_queue_entry *entry;
692 	int rc;
693 
694 	if (!qp->link_is_up || len == 0) {
695 		CTR0(KTR_NTB, "TX: link not up");
696 		return (EINVAL);
697 	}
698 
699 	entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
700 	if (entry == NULL) {
701 		CTR0(KTR_NTB, "TX: could not get entry from tx_free_q");
702 		qp->tx_err_no_buf++;
703 		return (EBUSY);
704 	}
705 	CTR1(KTR_NTB, "TX: got entry %p from tx_free_q", entry);
706 
707 	entry->cb_data = cb;
708 	entry->buf = data;
709 	entry->len = len;
710 	entry->flags = 0;
711 
712 	mtx_lock(&qp->tx_lock);
713 	rc = ntb_process_tx(qp, entry);
714 	mtx_unlock(&qp->tx_lock);
715 	if (rc != 0) {
716 		ntb_list_add(&qp->ntb_tx_free_q_lock, entry, &qp->tx_free_q);
717 		CTR1(KTR_NTB,
718 		    "TX: process_tx failed. Returning entry %p to tx_free_q",
719 		    entry);
720 	}
721 	return (rc);
722 }
723 
724 static void
725 ntb_tx_copy_callback(void *data)
726 {
727 	struct ntb_queue_entry *entry = data;
728 	struct ntb_transport_qp *qp = entry->qp;
729 	struct ntb_payload_header *hdr = entry->x_hdr;
730 
731 	iowrite32(entry->flags | NTBT_DESC_DONE_FLAG, &hdr->flags);
732 	CTR1(KTR_NTB, "TX: hdr %p set DESC_DONE", hdr);
733 
734 	ntb_peer_db_set(qp->dev, 1ull << qp->qp_num);
735 
736 	/*
737 	 * The entry length can only be zero if the packet is intended to be a
738 	 * "link down" or similar.  Since no payload is being sent in these
739 	 * cases, there is nothing to add to the completion queue.
740 	 */
741 	if (entry->len > 0) {
742 		qp->tx_bytes += entry->len;
743 
744 		if (qp->tx_handler)
745 			qp->tx_handler(qp, qp->cb_data, entry->buf,
746 			    entry->len);
747 		else
748 			m_freem(entry->buf);
749 		entry->buf = NULL;
750 	}
751 
752 	CTR3(KTR_NTB,
753 	    "TX: entry %p sent. hdr->ver = %u, hdr->flags = 0x%x, Returning "
754 	    "to tx_free_q", entry, hdr->ver, hdr->flags);
755 	ntb_list_add(&qp->ntb_tx_free_q_lock, entry, &qp->tx_free_q);
756 }
757 
758 static void
759 ntb_memcpy_tx(struct ntb_queue_entry *entry, void *offset)
760 {
761 
762 	CTR2(KTR_NTB, "TX: copying %d bytes to offset %p", entry->len, offset);
763 	if (entry->buf != NULL) {
764 		m_copydata((struct mbuf *)entry->buf, 0, entry->len, offset);
765 
766 		/*
767 		 * Ensure that the data is fully copied before setting the
768 		 * flags
769 		 */
770 		wmb();
771 	}
772 
773 	ntb_tx_copy_callback(entry);
774 }
775 
776 static void
777 ntb_async_tx(struct ntb_transport_qp *qp, struct ntb_queue_entry *entry)
778 {
779 	struct ntb_payload_header *hdr;
780 	void *offset;
781 
782 	offset = qp->tx_mw + qp->tx_max_frame * qp->tx_index;
783 	hdr = (struct ntb_payload_header *)((char *)offset + qp->tx_max_frame -
784 	    sizeof(struct ntb_payload_header));
785 	entry->x_hdr = hdr;
786 
787 	iowrite32(entry->len, &hdr->len);
788 	iowrite32(qp->tx_pkts, &hdr->ver);
789 
790 	ntb_memcpy_tx(entry, offset);
791 }
792 
793 static int
794 ntb_process_tx(struct ntb_transport_qp *qp, struct ntb_queue_entry *entry)
795 {
796 
797 	CTR3(KTR_NTB,
798 	    "TX: process_tx: tx_pkts=%lu, tx_index=%u, remote entry=%u",
799 	    qp->tx_pkts, qp->tx_index, qp->remote_rx_info->entry);
800 	if (qp->tx_index == qp->remote_rx_info->entry) {
801 		CTR0(KTR_NTB, "TX: ring full");
802 		qp->tx_ring_full++;
803 		return (EAGAIN);
804 	}
805 
806 	if (entry->len > qp->tx_max_frame - sizeof(struct ntb_payload_header)) {
807 		if (qp->tx_handler != NULL)
808 			qp->tx_handler(qp, qp->cb_data, entry->buf,
809 			    EIO);
810 		else
811 			m_freem(entry->buf);
812 
813 		entry->buf = NULL;
814 		ntb_list_add(&qp->ntb_tx_free_q_lock, entry, &qp->tx_free_q);
815 		CTR1(KTR_NTB,
816 		    "TX: frame too big. returning entry %p to tx_free_q",
817 		    entry);
818 		return (0);
819 	}
820 	CTR2(KTR_NTB, "TX: copying entry %p to index %u", entry, qp->tx_index);
821 	ntb_async_tx(qp, entry);
822 
823 	qp->tx_index++;
824 	qp->tx_index %= qp->tx_max_entry;
825 
826 	qp->tx_pkts++;
827 
828 	return (0);
829 }
830 
831 /* Transport Rx */
832 static void
833 ntb_transport_rxc_db(void *arg, int pending __unused)
834 {
835 	struct ntb_transport_qp *qp = arg;
836 	int rc;
837 
838 	CTR0(KTR_NTB, "RX: transport_rx");
839 again:
840 	while ((rc = ntb_process_rxc(qp)) == 0)
841 		;
842 	CTR1(KTR_NTB, "RX: process_rxc returned %d", rc);
843 
844 	if ((ntb_db_read(qp->dev) & (1ull << qp->qp_num)) != 0) {
845 		/* If db is set, clear it and check queue once more. */
846 		ntb_db_clear(qp->dev, 1ull << qp->qp_num);
847 		goto again;
848 	}
849 }
850 
851 static int
852 ntb_process_rxc(struct ntb_transport_qp *qp)
853 {
854 	struct ntb_payload_header *hdr;
855 	struct ntb_queue_entry *entry;
856 	caddr_t offset;
857 
858 	offset = qp->rx_buff + qp->rx_max_frame * qp->rx_index;
859 	hdr = (void *)(offset + qp->rx_max_frame -
860 	    sizeof(struct ntb_payload_header));
861 
862 	CTR1(KTR_NTB, "RX: process_rxc rx_index = %u", qp->rx_index);
863 	if ((hdr->flags & NTBT_DESC_DONE_FLAG) == 0) {
864 		CTR0(KTR_NTB, "RX: hdr not done");
865 		qp->rx_ring_empty++;
866 		return (EAGAIN);
867 	}
868 
869 	if ((hdr->flags & NTBT_LINK_DOWN_FLAG) != 0) {
870 		CTR0(KTR_NTB, "RX: link down");
871 		ntb_qp_link_down(qp);
872 		hdr->flags = 0;
873 		return (EAGAIN);
874 	}
875 
876 	if (hdr->ver != (uint32_t)qp->rx_pkts) {
877 		CTR2(KTR_NTB,"RX: ver != rx_pkts (%x != %lx). "
878 		    "Returning entry to rx_pend_q", hdr->ver, qp->rx_pkts);
879 		qp->rx_err_ver++;
880 		return (EIO);
881 	}
882 
883 	entry = ntb_list_mv(&qp->ntb_rx_q_lock, &qp->rx_pend_q, &qp->rx_post_q);
884 	if (entry == NULL) {
885 		qp->rx_err_no_buf++;
886 		CTR0(KTR_NTB, "RX: No entries in rx_pend_q");
887 		return (EAGAIN);
888 	}
889 	callout_stop(&qp->rx_full);
890 	CTR1(KTR_NTB, "RX: rx entry %p from rx_pend_q", entry);
891 
892 	entry->x_hdr = hdr;
893 	entry->index = qp->rx_index;
894 
895 	if (hdr->len > entry->len) {
896 		CTR2(KTR_NTB, "RX: len too long. Wanted %ju got %ju",
897 		    (uintmax_t)hdr->len, (uintmax_t)entry->len);
898 		qp->rx_err_oflow++;
899 
900 		entry->len = -EIO;
901 		entry->flags |= NTBT_DESC_DONE_FLAG;
902 
903 		ntb_complete_rxc(qp);
904 	} else {
905 		qp->rx_bytes += hdr->len;
906 		qp->rx_pkts++;
907 
908 		CTR1(KTR_NTB, "RX: received %ld rx_pkts", qp->rx_pkts);
909 
910 		entry->len = hdr->len;
911 
912 		ntb_memcpy_rx(qp, entry, offset);
913 	}
914 
915 	qp->rx_index++;
916 	qp->rx_index %= qp->rx_max_entry;
917 	return (0);
918 }
919 
920 static void
921 ntb_memcpy_rx(struct ntb_transport_qp *qp, struct ntb_queue_entry *entry,
922     void *offset)
923 {
924 	struct ifnet *ifp = entry->cb_data;
925 	unsigned int len = entry->len;
926 
927 	CTR2(KTR_NTB, "RX: copying %d bytes from offset %p", len, offset);
928 
929 	entry->buf = (void *)m_devget(offset, len, 0, ifp, NULL);
930 	if (entry->buf == NULL)
931 		entry->len = -ENOMEM;
932 
933 	/* Ensure that the data is globally visible before clearing the flag */
934 	wmb();
935 
936 	CTR2(KTR_NTB, "RX: copied entry %p to mbuf %p.", entry, entry->buf);
937 	ntb_rx_copy_callback(qp, entry);
938 }
939 
940 static inline void
941 ntb_rx_copy_callback(struct ntb_transport_qp *qp, void *data)
942 {
943 	struct ntb_queue_entry *entry;
944 
945 	entry = data;
946 	entry->flags |= NTBT_DESC_DONE_FLAG;
947 	ntb_complete_rxc(qp);
948 }
949 
950 static void
951 ntb_complete_rxc(struct ntb_transport_qp *qp)
952 {
953 	struct ntb_queue_entry *entry;
954 	struct mbuf *m;
955 	unsigned len;
956 
957 	CTR0(KTR_NTB, "RX: rx_completion_task");
958 
959 	mtx_lock_spin(&qp->ntb_rx_q_lock);
960 
961 	while (!STAILQ_EMPTY(&qp->rx_post_q)) {
962 		entry = STAILQ_FIRST(&qp->rx_post_q);
963 		if ((entry->flags & NTBT_DESC_DONE_FLAG) == 0)
964 			break;
965 
966 		entry->x_hdr->flags = 0;
967 		iowrite32(entry->index, &qp->rx_info->entry);
968 
969 		STAILQ_REMOVE_HEAD(&qp->rx_post_q, entry);
970 
971 		len = entry->len;
972 		m = entry->buf;
973 
974 		/*
975 		 * Re-initialize queue_entry for reuse; rx_handler takes
976 		 * ownership of the mbuf.
977 		 */
978 		entry->buf = NULL;
979 		entry->len = transport_mtu;
980 		entry->cb_data = qp->cb_data;
981 
982 		STAILQ_INSERT_TAIL(&qp->rx_pend_q, entry, entry);
983 
984 		mtx_unlock_spin(&qp->ntb_rx_q_lock);
985 
986 		CTR2(KTR_NTB, "RX: completing entry %p, mbuf %p", entry, m);
987 		if (qp->rx_handler != NULL && qp->client_ready)
988 			qp->rx_handler(qp, qp->cb_data, m, len);
989 		else
990 			m_freem(m);
991 
992 		mtx_lock_spin(&qp->ntb_rx_q_lock);
993 	}
994 
995 	mtx_unlock_spin(&qp->ntb_rx_q_lock);
996 }
997 
998 static void
999 ntb_transport_doorbell_callback(void *data, uint32_t vector)
1000 {
1001 	struct ntb_transport_ctx *nt = data;
1002 	struct ntb_transport_qp *qp;
1003 	uint64_t vec_mask;
1004 	unsigned qp_num;
1005 
1006 	vec_mask = ntb_db_vector_mask(nt->dev, vector);
1007 	vec_mask &= nt->qp_bitmap;
1008 	if ((vec_mask & (vec_mask - 1)) != 0)
1009 		vec_mask &= ntb_db_read(nt->dev);
1010 	while (vec_mask != 0) {
1011 		qp_num = ffsll(vec_mask) - 1;
1012 
1013 		qp = &nt->qp_vec[qp_num];
1014 		if (qp->link_is_up)
1015 			taskqueue_enqueue(qp->rxc_tq, &qp->rxc_db_work);
1016 
1017 		vec_mask &= ~(1ull << qp_num);
1018 	}
1019 }
1020 
1021 /* Link Event handler */
1022 static void
1023 ntb_transport_event_callback(void *data)
1024 {
1025 	struct ntb_transport_ctx *nt = data;
1026 
1027 	if (ntb_link_is_up(nt->dev, NULL, NULL)) {
1028 		ntb_printf(1, "HW link up\n");
1029 		callout_reset(&nt->link_work, 0, ntb_transport_link_work, nt);
1030 	} else {
1031 		ntb_printf(1, "HW link down\n");
1032 		taskqueue_enqueue(taskqueue_swi, &nt->link_cleanup);
1033 	}
1034 }
1035 
1036 /* Link bring up */
1037 static void
1038 ntb_transport_link_work(void *arg)
1039 {
1040 	struct ntb_transport_ctx *nt = arg;
1041 	device_t dev = nt->dev;
1042 	struct ntb_transport_qp *qp;
1043 	uint64_t val64, size;
1044 	uint32_t val;
1045 	unsigned i;
1046 	int rc;
1047 
1048 	/* send the local info, in the opposite order of the way we read it */
1049 	for (i = 0; i < nt->mw_count; i++) {
1050 		size = nt->mw_vec[i].phys_size;
1051 
1052 		if (max_mw_size != 0 && size > max_mw_size)
1053 			size = max_mw_size;
1054 
1055 		ntb_peer_spad_write(dev, NTBT_MW0_SZ_HIGH + (i * 2),
1056 		    size >> 32);
1057 		ntb_peer_spad_write(dev, NTBT_MW0_SZ_LOW + (i * 2), size);
1058 	}
1059 	ntb_peer_spad_write(dev, NTBT_NUM_MWS, nt->mw_count);
1060 	ntb_peer_spad_write(dev, NTBT_NUM_QPS, nt->qp_count);
1061 	ntb_peer_spad_write(dev, NTBT_QP_LINKS, 0);
1062 	ntb_peer_spad_write(dev, NTBT_VERSION, NTB_TRANSPORT_VERSION);
1063 
1064 	/* Query the remote side for its info */
1065 	val = 0;
1066 	ntb_spad_read(dev, NTBT_VERSION, &val);
1067 	if (val != NTB_TRANSPORT_VERSION)
1068 		goto out;
1069 
1070 	ntb_spad_read(dev, NTBT_NUM_QPS, &val);
1071 	if (val != nt->qp_count)
1072 		goto out;
1073 
1074 	ntb_spad_read(dev, NTBT_NUM_MWS, &val);
1075 	if (val != nt->mw_count)
1076 		goto out;
1077 
1078 	for (i = 0; i < nt->mw_count; i++) {
1079 		ntb_spad_read(dev, NTBT_MW0_SZ_HIGH + (i * 2), &val);
1080 		val64 = (uint64_t)val << 32;
1081 
1082 		ntb_spad_read(dev, NTBT_MW0_SZ_LOW + (i * 2), &val);
1083 		val64 |= val;
1084 
1085 		rc = ntb_set_mw(nt, i, val64);
1086 		if (rc != 0)
1087 			goto free_mws;
1088 	}
1089 
1090 	nt->link_is_up = true;
1091 	ntb_printf(1, "transport link up\n");
1092 
1093 	for (i = 0; i < nt->qp_count; i++) {
1094 		qp = &nt->qp_vec[i];
1095 
1096 		ntb_transport_setup_qp_mw(nt, i);
1097 
1098 		if (qp->client_ready)
1099 			callout_reset(&qp->link_work, 0, ntb_qp_link_work, qp);
1100 	}
1101 
1102 	return;
1103 
1104 free_mws:
1105 	for (i = 0; i < nt->mw_count; i++)
1106 		ntb_free_mw(nt, i);
1107 out:
1108 	if (ntb_link_is_up(dev, NULL, NULL))
1109 		callout_reset(&nt->link_work,
1110 		    NTB_LINK_DOWN_TIMEOUT * hz / 1000, ntb_transport_link_work, nt);
1111 }
1112 
1113 static int
1114 ntb_set_mw(struct ntb_transport_ctx *nt, int num_mw, size_t size)
1115 {
1116 	struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
1117 	size_t xlat_size, buff_size;
1118 	int rc;
1119 
1120 	if (size == 0)
1121 		return (EINVAL);
1122 
1123 	xlat_size = roundup(size, mw->xlat_align_size);
1124 	buff_size = xlat_size;
1125 
1126 	/* No need to re-setup */
1127 	if (mw->xlat_size == xlat_size)
1128 		return (0);
1129 
1130 	if (mw->buff_size != 0)
1131 		ntb_free_mw(nt, num_mw);
1132 
1133 	/* Alloc memory for receiving data.  Must be aligned */
1134 	mw->xlat_size = xlat_size;
1135 	mw->buff_size = buff_size;
1136 
1137 	mw->virt_addr = contigmalloc(mw->buff_size, M_NTB_T, M_ZERO, 0,
1138 	    mw->addr_limit, mw->xlat_align, 0);
1139 	if (mw->virt_addr == NULL) {
1140 		ntb_printf(0, "Unable to allocate MW buffer of size %zu/%zu\n",
1141 		    mw->buff_size, mw->xlat_size);
1142 		mw->xlat_size = 0;
1143 		mw->buff_size = 0;
1144 		return (ENOMEM);
1145 	}
1146 	/* TODO: replace with bus_space_* functions */
1147 	mw->dma_addr = vtophys(mw->virt_addr);
1148 
1149 	/*
1150 	 * Ensure that the allocation from contigmalloc is aligned as
1151 	 * requested.  XXX: This may not be needed -- brought in for parity
1152 	 * with the Linux driver.
1153 	 */
1154 	if (mw->dma_addr % mw->xlat_align != 0) {
1155 		ntb_printf(0,
1156 		    "DMA memory 0x%jx not aligned to BAR size 0x%zx\n",
1157 		    (uintmax_t)mw->dma_addr, size);
1158 		ntb_free_mw(nt, num_mw);
1159 		return (ENOMEM);
1160 	}
1161 
1162 	/* Notify HW the memory location of the receive buffer */
1163 	rc = ntb_mw_set_trans(nt->dev, num_mw, mw->dma_addr, mw->xlat_size);
1164 	if (rc) {
1165 		ntb_printf(0, "Unable to set mw%d translation\n", num_mw);
1166 		ntb_free_mw(nt, num_mw);
1167 		return (rc);
1168 	}
1169 
1170 	return (0);
1171 }
1172 
1173 static void
1174 ntb_free_mw(struct ntb_transport_ctx *nt, int num_mw)
1175 {
1176 	struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
1177 
1178 	if (mw->virt_addr == NULL)
1179 		return;
1180 
1181 	ntb_mw_clear_trans(nt->dev, num_mw);
1182 	contigfree(mw->virt_addr, mw->xlat_size, M_NTB_T);
1183 	mw->xlat_size = 0;
1184 	mw->buff_size = 0;
1185 	mw->virt_addr = NULL;
1186 }
1187 
1188 static int
1189 ntb_transport_setup_qp_mw(struct ntb_transport_ctx *nt, unsigned int qp_num)
1190 {
1191 	struct ntb_transport_qp *qp = &nt->qp_vec[qp_num];
1192 	struct ntb_transport_mw *mw;
1193 	void *offset;
1194 	ntb_q_idx_t i;
1195 	size_t rx_size;
1196 	unsigned num_qps_mw, mw_num, mw_count;
1197 
1198 	mw_count = nt->mw_count;
1199 	mw_num = QP_TO_MW(nt, qp_num);
1200 	mw = &nt->mw_vec[mw_num];
1201 
1202 	if (mw->virt_addr == NULL)
1203 		return (ENOMEM);
1204 
1205 	if (mw_num < nt->qp_count % mw_count)
1206 		num_qps_mw = nt->qp_count / mw_count + 1;
1207 	else
1208 		num_qps_mw = nt->qp_count / mw_count;
1209 
1210 	rx_size = mw->xlat_size / num_qps_mw;
1211 	qp->rx_buff = mw->virt_addr + rx_size * (qp_num / mw_count);
1212 	rx_size -= sizeof(struct ntb_rx_info);
1213 
1214 	qp->remote_rx_info = (void*)(qp->rx_buff + rx_size);
1215 
1216 	/* Due to house-keeping, there must be at least 2 buffs */
1217 	qp->rx_max_frame = qmin(transport_mtu, rx_size / 2);
1218 	qp->rx_max_entry = rx_size / qp->rx_max_frame;
1219 	qp->rx_index = 0;
1220 
1221 	qp->remote_rx_info->entry = qp->rx_max_entry - 1;
1222 
1223 	/* Set up the hdr offsets with 0s */
1224 	for (i = 0; i < qp->rx_max_entry; i++) {
1225 		offset = (void *)(qp->rx_buff + qp->rx_max_frame * (i + 1) -
1226 		    sizeof(struct ntb_payload_header));
1227 		memset(offset, 0, sizeof(struct ntb_payload_header));
1228 	}
1229 
1230 	qp->rx_pkts = 0;
1231 	qp->tx_pkts = 0;
1232 	qp->tx_index = 0;
1233 
1234 	return (0);
1235 }
1236 
1237 static void
1238 ntb_qp_link_work(void *arg)
1239 {
1240 	struct ntb_transport_qp *qp = arg;
1241 	device_t dev = qp->dev;
1242 	struct ntb_transport_ctx *nt = qp->transport;
1243 	int i;
1244 	uint32_t val;
1245 
1246 	/* Report queues that are up on our side */
1247 	for (i = 0, val = 0; i < nt->qp_count; i++) {
1248 		if (nt->qp_vec[i].client_ready)
1249 			val |= (1 << i);
1250 	}
1251 	ntb_peer_spad_write(dev, NTBT_QP_LINKS, val);
1252 
1253 	/* See if the remote side is up */
1254 	ntb_spad_read(dev, NTBT_QP_LINKS, &val);
1255 	if ((val & (1ull << qp->qp_num)) != 0) {
1256 		ntb_printf(2, "qp %d link up\n", qp->qp_num);
1257 		qp->link_is_up = true;
1258 
1259 		if (qp->event_handler != NULL)
1260 			qp->event_handler(qp->cb_data, NTB_LINK_UP);
1261 
1262 		ntb_db_clear_mask(dev, 1ull << qp->qp_num);
1263 	} else if (nt->link_is_up)
1264 		callout_reset(&qp->link_work,
1265 		    NTB_LINK_DOWN_TIMEOUT * hz / 1000, ntb_qp_link_work, qp);
1266 }
1267 
1268 /* Link down event*/
1269 static void
1270 ntb_transport_link_cleanup(struct ntb_transport_ctx *nt)
1271 {
1272 	struct ntb_transport_qp *qp;
1273 	int i;
1274 
1275 	/* Pass along the info to any clients */
1276 	for (i = 0; i < nt->qp_count; i++) {
1277 		if ((nt->qp_bitmap & (1 << i)) != 0) {
1278 			qp = &nt->qp_vec[i];
1279 			ntb_qp_link_cleanup(qp);
1280 			callout_drain(&qp->link_work);
1281 		}
1282 	}
1283 
1284 	if (!nt->link_is_up)
1285 		callout_drain(&nt->link_work);
1286 
1287 	/*
1288 	 * The scratchpad registers keep the values if the remote side
1289 	 * goes down, blast them now to give them a sane value the next
1290 	 * time they are accessed
1291 	 */
1292 	ntb_spad_clear(nt->dev);
1293 }
1294 
1295 static void
1296 ntb_transport_link_cleanup_work(void *arg, int pending __unused)
1297 {
1298 
1299 	ntb_transport_link_cleanup(arg);
1300 }
1301 
1302 static void
1303 ntb_qp_link_down(struct ntb_transport_qp *qp)
1304 {
1305 
1306 	ntb_qp_link_cleanup(qp);
1307 }
1308 
1309 static void
1310 ntb_qp_link_down_reset(struct ntb_transport_qp *qp)
1311 {
1312 
1313 	qp->link_is_up = false;
1314 	ntb_db_set_mask(qp->dev, 1ull << qp->qp_num);
1315 
1316 	qp->tx_index = qp->rx_index = 0;
1317 	qp->tx_bytes = qp->rx_bytes = 0;
1318 	qp->tx_pkts = qp->rx_pkts = 0;
1319 
1320 	qp->rx_ring_empty = 0;
1321 	qp->tx_ring_full = 0;
1322 
1323 	qp->rx_err_no_buf = qp->tx_err_no_buf = 0;
1324 	qp->rx_err_oflow = qp->rx_err_ver = 0;
1325 }
1326 
1327 static void
1328 ntb_qp_link_cleanup(struct ntb_transport_qp *qp)
1329 {
1330 
1331 	callout_drain(&qp->link_work);
1332 	ntb_qp_link_down_reset(qp);
1333 
1334 	if (qp->event_handler != NULL)
1335 		qp->event_handler(qp->cb_data, NTB_LINK_DOWN);
1336 }
1337 
1338 /* Link commanded down */
1339 /**
1340  * ntb_transport_link_down - Notify NTB transport to no longer enqueue data
1341  * @qp: NTB transport layer queue to be disabled
1342  *
1343  * Notify NTB transport layer of client's desire to no longer receive data on
1344  * transport queue specified.  It is the client's responsibility to ensure all
1345  * entries on queue are purged or otherwise handled appropriately.
1346  */
1347 void
1348 ntb_transport_link_down(struct ntb_transport_qp *qp)
1349 {
1350 	struct ntb_transport_ctx *nt = qp->transport;
1351 	int i;
1352 	uint32_t val;
1353 
1354 	qp->client_ready = false;
1355 	for (i = 0, val = 0; i < nt->qp_count; i++) {
1356 		if (nt->qp_vec[i].client_ready)
1357 			val |= (1 << i);
1358 	}
1359 	ntb_peer_spad_write(qp->dev, NTBT_QP_LINKS, val);
1360 
1361 	if (qp->link_is_up)
1362 		ntb_send_link_down(qp);
1363 	else
1364 		callout_drain(&qp->link_work);
1365 }
1366 
1367 /**
1368  * ntb_transport_link_query - Query transport link state
1369  * @qp: NTB transport layer queue to be queried
1370  *
1371  * Query connectivity to the remote system of the NTB transport queue
1372  *
1373  * RETURNS: true for link up or false for link down
1374  */
1375 bool
1376 ntb_transport_link_query(struct ntb_transport_qp *qp)
1377 {
1378 
1379 	return (qp->link_is_up);
1380 }
1381 
1382 static void
1383 ntb_send_link_down(struct ntb_transport_qp *qp)
1384 {
1385 	struct ntb_queue_entry *entry;
1386 	int i, rc;
1387 
1388 	if (!qp->link_is_up)
1389 		return;
1390 
1391 	for (i = 0; i < NTB_LINK_DOWN_TIMEOUT; i++) {
1392 		entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
1393 		if (entry != NULL)
1394 			break;
1395 		pause("NTB Wait for link down", hz / 10);
1396 	}
1397 
1398 	if (entry == NULL)
1399 		return;
1400 
1401 	entry->cb_data = NULL;
1402 	entry->buf = NULL;
1403 	entry->len = 0;
1404 	entry->flags = NTBT_LINK_DOWN_FLAG;
1405 
1406 	mtx_lock(&qp->tx_lock);
1407 	rc = ntb_process_tx(qp, entry);
1408 	mtx_unlock(&qp->tx_lock);
1409 	if (rc != 0)
1410 		printf("ntb: Failed to send link down\n");
1411 
1412 	ntb_qp_link_down_reset(qp);
1413 }
1414 
1415 
1416 /* List Management */
1417 
1418 static void
1419 ntb_list_add(struct mtx *lock, struct ntb_queue_entry *entry,
1420     struct ntb_queue_list *list)
1421 {
1422 
1423 	mtx_lock_spin(lock);
1424 	STAILQ_INSERT_TAIL(list, entry, entry);
1425 	mtx_unlock_spin(lock);
1426 }
1427 
1428 static struct ntb_queue_entry *
1429 ntb_list_rm(struct mtx *lock, struct ntb_queue_list *list)
1430 {
1431 	struct ntb_queue_entry *entry;
1432 
1433 	mtx_lock_spin(lock);
1434 	if (STAILQ_EMPTY(list)) {
1435 		entry = NULL;
1436 		goto out;
1437 	}
1438 	entry = STAILQ_FIRST(list);
1439 	STAILQ_REMOVE_HEAD(list, entry);
1440 out:
1441 	mtx_unlock_spin(lock);
1442 
1443 	return (entry);
1444 }
1445 
1446 static struct ntb_queue_entry *
1447 ntb_list_mv(struct mtx *lock, struct ntb_queue_list *from,
1448     struct ntb_queue_list *to)
1449 {
1450 	struct ntb_queue_entry *entry;
1451 
1452 	mtx_lock_spin(lock);
1453 	if (STAILQ_EMPTY(from)) {
1454 		entry = NULL;
1455 		goto out;
1456 	}
1457 	entry = STAILQ_FIRST(from);
1458 	STAILQ_REMOVE_HEAD(from, entry);
1459 	STAILQ_INSERT_TAIL(to, entry, entry);
1460 
1461 out:
1462 	mtx_unlock_spin(lock);
1463 	return (entry);
1464 }
1465 
1466 /**
1467  * ntb_transport_qp_num - Query the qp number
1468  * @qp: NTB transport layer queue to be queried
1469  *
1470  * Query qp number of the NTB transport queue
1471  *
1472  * RETURNS: a zero based number specifying the qp number
1473  */
1474 unsigned char ntb_transport_qp_num(struct ntb_transport_qp *qp)
1475 {
1476 
1477 	return (qp->qp_num);
1478 }
1479 
1480 /**
1481  * ntb_transport_max_size - Query the max payload size of a qp
1482  * @qp: NTB transport layer queue to be queried
1483  *
1484  * Query the maximum payload size permissible on the given qp
1485  *
1486  * RETURNS: the max payload size of a qp
1487  */
1488 unsigned int
1489 ntb_transport_max_size(struct ntb_transport_qp *qp)
1490 {
1491 
1492 	return (qp->tx_max_frame - sizeof(struct ntb_payload_header));
1493 }
1494 
1495 unsigned int
1496 ntb_transport_tx_free_entry(struct ntb_transport_qp *qp)
1497 {
1498 	unsigned int head = qp->tx_index;
1499 	unsigned int tail = qp->remote_rx_info->entry;
1500 
1501 	return (tail >= head ? tail - head : qp->tx_max_entry + tail - head);
1502 }
1503 
1504 static device_method_t ntb_transport_methods[] = {
1505 	/* Device interface */
1506 	DEVMETHOD(device_probe,     ntb_transport_probe),
1507 	DEVMETHOD(device_attach,    ntb_transport_attach),
1508 	DEVMETHOD(device_detach,    ntb_transport_detach),
1509 	DEVMETHOD_END
1510 };
1511 
1512 devclass_t ntb_transport_devclass;
1513 static DEFINE_CLASS_0(ntb_transport, ntb_transport_driver,
1514     ntb_transport_methods, sizeof(struct ntb_transport_ctx));
1515 DRIVER_MODULE(ntb_transport, ntb_hw, ntb_transport_driver,
1516     ntb_transport_devclass, NULL, NULL);
1517 MODULE_DEPEND(ntb_transport, ntb, 1, 1, 1);
1518 MODULE_VERSION(ntb_transport, 1);
1519