xref: /freebsd/sys/dev/ntb/if_ntb/if_ntb.c (revision a4dc509f723944821bcfcc52005ff87c9a5dee5b)
1 /*-
2  * Copyright (C) 2013 Intel Corporation
3  * Copyright (C) 2015 EMC Corporation
4  * All rights reserved.
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25  * SUCH DAMAGE.
26  */
27 
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
30 
31 #include <sys/param.h>
32 #include <sys/kernel.h>
33 #include <sys/systm.h>
34 #include <sys/bitset.h>
35 #include <sys/bus.h>
36 #include <sys/ktr.h>
37 #include <sys/limits.h>
38 #include <sys/lock.h>
39 #include <sys/malloc.h>
40 #include <sys/module.h>
41 #include <sys/mutex.h>
42 #include <sys/queue.h>
43 #include <sys/socket.h>
44 #include <sys/sockio.h>
45 #include <sys/sysctl.h>
46 #include <sys/taskqueue.h>
47 
48 #include <net/if.h>
49 #include <net/if_media.h>
50 #include <net/if_types.h>
51 #include <net/if_var.h>
52 #include <net/bpf.h>
53 #include <net/ethernet.h>
54 
55 #include <vm/vm.h>
56 #include <vm/pmap.h>
57 
58 #include <machine/bus.h>
59 #include <machine/cpufunc.h>
60 #include <machine/pmap.h>
61 
62 #include "../ntb_hw/ntb_hw.h"
63 
64 /*
65  * The Non-Transparent Bridge (NTB) is a device on some Intel processors that
66  * allows you to connect two systems using a PCI-e link.
67  *
68  * This module contains a protocol for sending and receiving messages, and
69  * exposes that protocol through a simulated ethernet device called ntb.
70  *
71  * NOTE: Much of the code in this module is shared with Linux. Any patches may
72  * be picked up and redistributed in Linux with a dual GPL/BSD license.
73  */
74 
75 #define QP_SETSIZE	64
76 BITSET_DEFINE(_qpset, QP_SETSIZE);
77 #define test_bit(pos, addr)	BIT_ISSET(QP_SETSIZE, (pos), (addr))
78 #define set_bit(pos, addr)	BIT_SET(QP_SETSIZE, (pos), (addr))
79 #define clear_bit(pos, addr)	BIT_CLR(QP_SETSIZE, (pos), (addr))
80 #define ffs_bit(addr)		BIT_FFS(QP_SETSIZE, (addr))
81 
82 #define KTR_NTB KTR_SPARE3
83 
84 #define NTB_TRANSPORT_VERSION	4
85 #define NTB_RX_MAX_PKTS		64
86 #define	NTB_RXQ_SIZE		300
87 
88 enum ntb_link_event {
89 	NTB_LINK_DOWN = 0,
90 	NTB_LINK_UP,
91 };
92 
93 static unsigned int transport_mtu = 0x10000 + ETHER_HDR_LEN + ETHER_CRC_LEN;
94 
95 static uint64_t max_mw_size;
96 SYSCTL_UQUAD(_hw_ntb, OID_AUTO, max_mw_size, CTLFLAG_RDTUN, &max_mw_size, 0,
97     "If enabled (non-zero), limit the size of large memory windows. "
98     "Both sides of the NTB MUST set the same value here.");
99 
100 static unsigned int max_num_clients;
101 SYSCTL_UINT(_hw_ntb, OID_AUTO, max_num_clients, CTLFLAG_RDTUN,
102     &max_num_clients, 0, "Maximum number of NTB transport clients.  "
103     "0 (default) - use all available NTB memory windows; "
104     "positive integer N - Limit to N memory windows.");
105 
106 STAILQ_HEAD(ntb_queue_list, ntb_queue_entry);
107 
108 typedef unsigned ntb_q_idx_t;
109 
110 struct ntb_queue_entry {
111 	/* ntb_queue list reference */
112 	STAILQ_ENTRY(ntb_queue_entry) entry;
113 
114 	/* info on data to be transferred */
115 	void		*cb_data;
116 	void		*buf;
117 	unsigned	len;
118 	unsigned	flags;
119 
120 	struct ntb_transport_qp		*qp;
121 	struct ntb_payload_header	*x_hdr;
122 	ntb_q_idx_t	index;
123 };
124 
125 struct ntb_rx_info {
126 	ntb_q_idx_t	entry;
127 };
128 
129 struct ntb_transport_qp {
130 	struct ntb_transport_ctx	*transport;
131 	struct ntb_softc	*ntb;
132 
133 	void			*cb_data;
134 
135 	bool			client_ready;
136 	bool			link_is_up;
137 	uint8_t			qp_num;	/* Only 64 QPs are allowed.  0-63 */
138 
139 	struct ntb_rx_info	*rx_info;
140 	struct ntb_rx_info	*remote_rx_info;
141 
142 	void (*tx_handler)(struct ntb_transport_qp *qp, void *qp_data,
143 	    void *data, int len);
144 	struct ntb_queue_list	tx_free_q;
145 	struct mtx		ntb_tx_free_q_lock;
146 	void			*tx_mw;
147 	bus_addr_t		tx_mw_phys;
148 	ntb_q_idx_t		tx_index;
149 	ntb_q_idx_t		tx_max_entry;
150 	uint64_t		tx_max_frame;
151 
152 	void (*rx_handler)(struct ntb_transport_qp *qp, void *qp_data,
153 	    void *data, int len);
154 	struct ntb_queue_list	rx_post_q;
155 	struct ntb_queue_list	rx_pend_q;
156 	struct ntb_queue_list	rx_free_q;
157 	/* ntb_rx_q_lock: synchronize access to rx_XXXX_q */
158 	struct mtx		ntb_rx_q_lock;
159 	struct task		rx_completion_task;
160 	struct task		rxc_db_work;
161 	void			*rx_buff;
162 	ntb_q_idx_t		rx_index;
163 	ntb_q_idx_t		rx_max_entry;
164 	uint64_t		rx_max_frame;
165 
166 	void (*event_handler)(void *data, enum ntb_link_event status);
167 	struct callout		link_work;
168 	struct callout		queue_full;
169 	struct callout		rx_full;
170 
171 	uint64_t		last_rx_no_buf;
172 
173 	/* Stats */
174 	uint64_t		rx_bytes;
175 	uint64_t		rx_pkts;
176 	uint64_t		rx_ring_empty;
177 	uint64_t		rx_err_no_buf;
178 	uint64_t		rx_err_oflow;
179 	uint64_t		rx_err_ver;
180 	uint64_t		tx_bytes;
181 	uint64_t		tx_pkts;
182 	uint64_t		tx_ring_full;
183 	uint64_t		tx_err_no_buf;
184 };
185 
186 struct ntb_queue_handlers {
187 	void (*rx_handler)(struct ntb_transport_qp *qp, void *qp_data,
188 	    void *data, int len);
189 	void (*tx_handler)(struct ntb_transport_qp *qp, void *qp_data,
190 	    void *data, int len);
191 	void (*event_handler)(void *data, enum ntb_link_event status);
192 };
193 
194 struct ntb_transport_mw {
195 	vm_paddr_t	phys_addr;
196 	size_t		phys_size;
197 	size_t		xlat_align;
198 	size_t		xlat_align_size;
199 	/* Tx buff is off vbase / phys_addr */
200 	void		*vbase;
201 	size_t		xlat_size;
202 	size_t		buff_size;
203 	/* Rx buff is off virt_addr / dma_addr */
204 	void		*virt_addr;
205 	bus_addr_t	dma_addr;
206 };
207 
208 struct ntb_transport_ctx {
209 	struct ntb_softc	*ntb;
210 	struct ifnet		*ifp;
211 	struct ntb_transport_mw	mw_vec[NTB_MAX_NUM_MW];
212 	struct ntb_transport_qp	*qp_vec;
213 	struct _qpset		qp_bitmap;
214 	struct _qpset		qp_bitmap_free;
215 	unsigned		mw_count;
216 	unsigned		qp_count;
217 	enum ntb_link_event	link_is_up;
218 	struct callout		link_work;
219 	uint64_t		bufsize;
220 	u_char			eaddr[ETHER_ADDR_LEN];
221 	struct mtx		tx_lock;
222 	struct mtx		rx_lock;
223 
224 	/* The hardcoded single queuepair in ntb_setup_interface() */
225 	struct ntb_transport_qp *qp;
226 };
227 
228 static struct ntb_transport_ctx net_softc;
229 
230 enum {
231 	IF_NTB_DESC_DONE_FLAG = 1 << 0,
232 	IF_NTB_LINK_DOWN_FLAG = 1 << 1,
233 };
234 
235 struct ntb_payload_header {
236 	uint64_t ver;
237 	uint64_t len;
238 	uint64_t flags;
239 };
240 
241 enum {
242 	/*
243 	 * The order of this enum is part of the if_ntb remote protocol.  Do
244 	 * not reorder without bumping protocol version (and it's probably best
245 	 * to keep the protocol in lock-step with the Linux NTB driver.
246 	 */
247 	IF_NTB_VERSION = 0,
248 	IF_NTB_QP_LINKS,
249 	IF_NTB_NUM_QPS,
250 	IF_NTB_NUM_MWS,
251 	/*
252 	 * N.B.: transport_link_work assumes MW1 enums = MW0 + 2.
253 	 */
254 	IF_NTB_MW0_SZ_HIGH,
255 	IF_NTB_MW0_SZ_LOW,
256 	IF_NTB_MW1_SZ_HIGH,
257 	IF_NTB_MW1_SZ_LOW,
258 	IF_NTB_MAX_SPAD,
259 };
260 
261 #define QP_TO_MW(nt, qp)	((qp) % nt->mw_count)
262 #define NTB_QP_DEF_NUM_ENTRIES	100
263 #define NTB_LINK_DOWN_TIMEOUT	10
264 
265 static int ntb_handle_module_events(struct module *m, int what, void *arg);
266 static int ntb_setup_interface(void);
267 static int ntb_teardown_interface(void);
268 static void ntb_net_init(void *arg);
269 static int ntb_ioctl(struct ifnet *ifp, u_long command, caddr_t data);
270 static void ntb_start(struct ifnet *ifp);
271 static void ntb_net_tx_handler(struct ntb_transport_qp *qp, void *qp_data,
272     void *data, int len);
273 static void ntb_net_rx_handler(struct ntb_transport_qp *qp, void *qp_data,
274     void *data, int len);
275 static void ntb_net_event_handler(void *data, enum ntb_link_event status);
276 static int ntb_transport_init(struct ntb_softc *ntb);
277 static void ntb_transport_free(struct ntb_transport_ctx *);
278 static void ntb_transport_init_queue(struct ntb_transport_ctx *nt,
279     unsigned int qp_num);
280 static void ntb_transport_free_queue(struct ntb_transport_qp *qp);
281 static struct ntb_transport_qp *ntb_transport_create_queue(void *data,
282     struct ntb_softc *pdev, const struct ntb_queue_handlers *handlers);
283 static void ntb_transport_link_up(struct ntb_transport_qp *qp);
284 static int ntb_transport_tx_enqueue(struct ntb_transport_qp *qp, void *cb,
285     void *data, unsigned int len);
286 static int ntb_process_tx(struct ntb_transport_qp *qp,
287     struct ntb_queue_entry *entry);
288 static void ntb_memcpy_tx(struct ntb_transport_qp *qp,
289     struct ntb_queue_entry *entry, void *offset);
290 static void ntb_qp_full(void *arg);
291 static void ntb_transport_rxc_db(void *arg, int pending);
292 static int ntb_process_rxc(struct ntb_transport_qp *qp);
293 static void ntb_memcpy_rx(struct ntb_transport_qp *qp,
294     struct ntb_queue_entry *entry, void *offset);
295 static inline void ntb_rx_copy_callback(struct ntb_transport_qp *qp,
296     void *data);
297 static void ntb_complete_rxc(void *arg, int pending);
298 static void ntb_transport_doorbell_callback(void *data, uint32_t vector);
299 static void ntb_transport_event_callback(void *data);
300 static void ntb_transport_link_work(void *arg);
301 static int ntb_set_mw(struct ntb_transport_ctx *, int num_mw, size_t size);
302 static void ntb_free_mw(struct ntb_transport_ctx *nt, int num_mw);
303 static int ntb_transport_setup_qp_mw(struct ntb_transport_ctx *nt,
304     unsigned int qp_num);
305 static void ntb_qp_link_work(void *arg);
306 static void ntb_transport_link_cleanup(struct ntb_transport_ctx *nt);
307 static void ntb_qp_link_down(struct ntb_transport_qp *qp);
308 static void ntb_qp_link_down_reset(struct ntb_transport_qp *qp);
309 static void ntb_qp_link_cleanup(struct ntb_transport_qp *qp);
310 static void ntb_transport_link_down(struct ntb_transport_qp *qp);
311 static void ntb_send_link_down(struct ntb_transport_qp *qp);
312 static void ntb_list_add(struct mtx *lock, struct ntb_queue_entry *entry,
313     struct ntb_queue_list *list);
314 static struct ntb_queue_entry *ntb_list_rm(struct mtx *lock,
315     struct ntb_queue_list *list);
316 static struct ntb_queue_entry *ntb_list_mv(struct mtx *lock,
317     struct ntb_queue_list *from, struct ntb_queue_list *to);
318 static void create_random_local_eui48(u_char *eaddr);
319 static unsigned int ntb_transport_max_size(struct ntb_transport_qp *qp);
320 
321 static const struct ntb_ctx_ops ntb_transport_ops = {
322 	.link_event = ntb_transport_event_callback,
323 	.db_event = ntb_transport_doorbell_callback,
324 };
325 
326 MALLOC_DEFINE(M_NTB_IF, "if_ntb", "ntb network driver");
327 
328 /* Module setup and teardown */
329 static int
330 ntb_handle_module_events(struct module *m, int what, void *arg)
331 {
332 	int err = 0;
333 
334 	switch (what) {
335 	case MOD_LOAD:
336 		err = ntb_setup_interface();
337 		break;
338 	case MOD_UNLOAD:
339 		err = ntb_teardown_interface();
340 		break;
341 	default:
342 		err = EOPNOTSUPP;
343 		break;
344 	}
345 	return (err);
346 }
347 
348 static moduledata_t if_ntb_mod = {
349 	"if_ntb",
350 	ntb_handle_module_events,
351 	NULL
352 };
353 
354 DECLARE_MODULE(if_ntb, if_ntb_mod, SI_SUB_KLD, SI_ORDER_ANY);
355 MODULE_DEPEND(if_ntb, ntb_hw, 1, 1, 1);
356 
357 static int
358 ntb_setup_interface(void)
359 {
360 	struct ifnet *ifp;
361 	struct ntb_queue_handlers handlers = { ntb_net_rx_handler,
362 	    ntb_net_tx_handler, ntb_net_event_handler };
363 	int rc;
364 
365 	net_softc.ntb = devclass_get_softc(devclass_find("ntb_hw"), 0);
366 	if (net_softc.ntb == NULL) {
367 		printf("ntb: Cannot find devclass\n");
368 		return (ENXIO);
369 	}
370 
371 	rc = ntb_transport_init(net_softc.ntb);
372 	if (rc != 0) {
373 		printf("ntb: Cannot init transport: %d\n", rc);
374 		return (rc);
375 	}
376 
377 	ifp = net_softc.ifp = if_alloc(IFT_ETHER);
378 	if (ifp == NULL) {
379 		ntb_transport_free(&net_softc);
380 		printf("ntb: Cannot allocate ifnet structure\n");
381 		return (ENOMEM);
382 	}
383 
384 	net_softc.qp = ntb_transport_create_queue(ifp, net_softc.ntb,
385 	    &handlers);
386 	if_initname(ifp, "ntb", 0);
387 	ifp->if_init = ntb_net_init;
388 	ifp->if_softc = &net_softc;
389 	ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX;
390 	ifp->if_ioctl = ntb_ioctl;
391 	ifp->if_start = ntb_start;
392 	IFQ_SET_MAXLEN(&ifp->if_snd, IFQ_MAXLEN);
393 	ifp->if_snd.ifq_drv_maxlen = IFQ_MAXLEN;
394 	IFQ_SET_READY(&ifp->if_snd);
395 	create_random_local_eui48(net_softc.eaddr);
396 	ether_ifattach(ifp, net_softc.eaddr);
397 	ifp->if_capabilities = IFCAP_HWCSUM | IFCAP_JUMBO_MTU;
398 	ifp->if_capenable = ifp->if_capabilities;
399 
400 	ntb_transport_link_up(net_softc.qp);
401 	net_softc.bufsize = ntb_transport_max_size(net_softc.qp) +
402 	    sizeof(struct ether_header);
403 	return (0);
404 }
405 
406 static int
407 ntb_teardown_interface(void)
408 {
409 
410 	if (net_softc.qp != NULL)
411 		ntb_transport_link_down(net_softc.qp);
412 
413 	if (net_softc.ifp != NULL) {
414 		ether_ifdetach(net_softc.ifp);
415 		if_free(net_softc.ifp);
416 	}
417 
418 	if (net_softc.qp != NULL) {
419 		ntb_transport_free_queue(net_softc.qp);
420 		ntb_transport_free(&net_softc);
421 	}
422 
423 	return (0);
424 }
425 
426 /* Network device interface */
427 
428 static void
429 ntb_net_init(void *arg)
430 {
431 	struct ntb_transport_ctx *ntb_softc = arg;
432 	struct ifnet *ifp = ntb_softc->ifp;
433 
434 	ifp->if_drv_flags |= IFF_DRV_RUNNING;
435 	ifp->if_drv_flags &= ~IFF_DRV_OACTIVE;
436 	ifp->if_flags |= IFF_UP;
437 	if_link_state_change(ifp, LINK_STATE_UP);
438 }
439 
440 static int
441 ntb_ioctl(struct ifnet *ifp, u_long command, caddr_t data)
442 {
443 	struct ntb_transport_ctx *nt = ifp->if_softc;
444 	struct ifreq *ifr = (struct ifreq *)data;
445 	int error = 0;
446 
447 	switch (command) {
448 	case SIOCSIFMTU:
449 	    {
450 		if (ifr->ifr_mtu > ntb_transport_max_size(nt->qp) -
451 		    ETHER_HDR_LEN - ETHER_CRC_LEN) {
452 			error = EINVAL;
453 			break;
454 		}
455 
456 		ifp->if_mtu = ifr->ifr_mtu;
457 		break;
458 	    }
459 	default:
460 		error = ether_ioctl(ifp, command, data);
461 		break;
462 	}
463 
464 	return (error);
465 }
466 
467 
468 static void
469 ntb_start(struct ifnet *ifp)
470 {
471 	struct mbuf *m_head;
472 	struct ntb_transport_ctx *nt = ifp->if_softc;
473 	int rc;
474 
475 	mtx_lock(&nt->tx_lock);
476 	ifp->if_drv_flags &= ~IFF_DRV_OACTIVE;
477 	CTR0(KTR_NTB, "TX: ntb_start");
478 	while (!IFQ_DRV_IS_EMPTY(&ifp->if_snd)) {
479 		IFQ_DRV_DEQUEUE(&ifp->if_snd, m_head);
480 		CTR1(KTR_NTB, "TX: start mbuf %p", m_head);
481 		rc = ntb_transport_tx_enqueue(nt->qp, m_head, m_head,
482 			     m_length(m_head, NULL));
483 		if (rc != 0) {
484 			CTR1(KTR_NTB,
485 			    "TX: could not tx mbuf %p. Returning to snd q",
486 			    m_head);
487 			if (rc == EAGAIN) {
488 				ifp->if_drv_flags |= IFF_DRV_OACTIVE;
489 				IFQ_DRV_PREPEND(&ifp->if_snd, m_head);
490 				callout_reset(&nt->qp->queue_full, hz / 1000,
491 				    ntb_qp_full, ifp);
492 			}
493 			break;
494 		}
495 
496 	}
497 	mtx_unlock(&nt->tx_lock);
498 }
499 
500 /* Network Device Callbacks */
501 static void
502 ntb_net_tx_handler(struct ntb_transport_qp *qp, void *qp_data, void *data,
503     int len)
504 {
505 
506 	m_freem(data);
507 	CTR1(KTR_NTB, "TX: tx_handler freeing mbuf %p", data);
508 }
509 
510 static void
511 ntb_net_rx_handler(struct ntb_transport_qp *qp, void *qp_data, void *data,
512     int len)
513 {
514 	struct mbuf *m = data;
515 	struct ifnet *ifp = qp_data;
516 
517 	CTR0(KTR_NTB, "RX: rx handler");
518 	(*ifp->if_input)(ifp, m);
519 }
520 
521 static void
522 ntb_net_event_handler(void *data, enum ntb_link_event status)
523 {
524 	struct ifnet *ifp;
525 
526 	ifp = data;
527 	(void)ifp;
528 
529 	/* XXX The Linux driver munges with the carrier status here. */
530 
531 	switch (status) {
532 	case NTB_LINK_DOWN:
533 		break;
534 	case NTB_LINK_UP:
535 		break;
536 	default:
537 		panic("Bogus ntb_link_event %u\n", status);
538 	}
539 }
540 
541 /* Transport Init and teardown */
542 
543 static int
544 ntb_transport_init(struct ntb_softc *ntb)
545 {
546 	struct ntb_transport_ctx *nt = &net_softc;
547 	struct ntb_transport_mw *mw;
548 	uint64_t qp_bitmap;
549 	int rc;
550 	unsigned i;
551 
552 	nt->mw_count = ntb_mw_count(ntb);
553 	for (i = 0; i < nt->mw_count; i++) {
554 		mw = &nt->mw_vec[i];
555 
556 		rc = ntb_mw_get_range(ntb, i, &mw->phys_addr, &mw->vbase,
557 		    &mw->phys_size, &mw->xlat_align, &mw->xlat_align_size);
558 		if (rc != 0)
559 			goto err;
560 
561 		mw->buff_size = 0;
562 		mw->xlat_size = 0;
563 		mw->virt_addr = 0;
564 		mw->dma_addr = 0;
565 	}
566 
567 	qp_bitmap = ntb_db_valid_mask(ntb);
568 	nt->qp_count = flsll(qp_bitmap);
569 	KASSERT(nt->qp_count != 0, ("bogus db bitmap"));
570 	nt->qp_count -= 1;
571 
572 	if (max_num_clients != 0 && max_num_clients < nt->qp_count)
573 		nt->qp_count = max_num_clients;
574 	else if (nt->mw_count < nt->qp_count)
575 		nt->qp_count = nt->mw_count;
576 	KASSERT(nt->qp_count <= QP_SETSIZE, ("invalid qp_count"));
577 
578 	mtx_init(&nt->tx_lock, "ntb transport tx", NULL, MTX_DEF);
579 	mtx_init(&nt->rx_lock, "ntb transport rx", NULL, MTX_DEF);
580 
581 	nt->qp_vec = malloc(nt->qp_count * sizeof(*nt->qp_vec), M_NTB_IF,
582 	    M_WAITOK | M_ZERO);
583 
584 	for (i = 0; i < nt->qp_count; i++) {
585 		set_bit(i, &nt->qp_bitmap);
586 		set_bit(i, &nt->qp_bitmap_free);
587 		ntb_transport_init_queue(nt, i);
588 	}
589 
590 	callout_init(&nt->link_work, 0);
591 
592 	rc = ntb_set_ctx(ntb, nt, &ntb_transport_ops);
593 	if (rc != 0)
594 		goto err;
595 
596 	nt->link_is_up = false;
597 	ntb_link_enable(ntb, NTB_SPEED_AUTO, NTB_WIDTH_AUTO);
598 	ntb_link_event(ntb);
599 	return (0);
600 
601 err:
602 	free(nt->qp_vec, M_NTB_IF);
603 	nt->qp_vec = NULL;
604 	return (rc);
605 }
606 
607 static void
608 ntb_transport_free(struct ntb_transport_ctx *nt)
609 {
610 	struct ntb_softc *ntb = nt->ntb;
611 	struct _qpset qp_bitmap_alloc;
612 	uint8_t i;
613 
614 	ntb_transport_link_cleanup(nt);
615 
616 	callout_drain(&nt->link_work);
617 
618 	BIT_COPY(QP_SETSIZE, &nt->qp_bitmap, &qp_bitmap_alloc);
619 	BIT_NAND(QP_SETSIZE, &qp_bitmap_alloc, &nt->qp_bitmap_free);
620 
621 	/* Verify that all the QPs are freed */
622 	for (i = 0; i < nt->qp_count; i++)
623 		if (test_bit(i, &qp_bitmap_alloc))
624 			ntb_transport_free_queue(&nt->qp_vec[i]);
625 
626 	ntb_link_disable(ntb);
627 	ntb_clear_ctx(ntb);
628 
629 	for (i = 0; i < nt->mw_count; i++)
630 		ntb_free_mw(nt, i);
631 
632 	free(nt->qp_vec, M_NTB_IF);
633 }
634 
635 static void
636 ntb_transport_init_queue(struct ntb_transport_ctx *nt, unsigned int qp_num)
637 {
638 	struct ntb_transport_mw *mw;
639 	struct ntb_transport_qp *qp;
640 	vm_paddr_t mw_base;
641 	uint64_t mw_size, qp_offset;
642 	size_t tx_size;
643 	unsigned num_qps_mw, mw_num, mw_count;
644 
645 	mw_count = nt->mw_count;
646 	mw_num = QP_TO_MW(nt, qp_num);
647 	mw = &nt->mw_vec[mw_num];
648 
649 	qp = &nt->qp_vec[qp_num];
650 	qp->qp_num = qp_num;
651 	qp->transport = nt;
652 	qp->ntb = nt->ntb;
653 	qp->client_ready = false;
654 	qp->event_handler = NULL;
655 	ntb_qp_link_down_reset(qp);
656 
657 	if (nt->qp_count % mw_count && mw_num + 1 < nt->qp_count / mw_count)
658 		num_qps_mw = nt->qp_count / mw_count + 1;
659 	else
660 		num_qps_mw = nt->qp_count / mw_count;
661 
662 	mw_base = mw->phys_addr;
663 	mw_size = mw->phys_size;
664 
665 	tx_size = mw_size / num_qps_mw;
666 	qp_offset = tx_size * qp_num / mw_count;
667 
668 	qp->tx_mw = (char *)mw->vbase + qp_offset;
669 	KASSERT(qp->tx_mw != NULL, ("uh oh?"));
670 
671 	/* XXX Assumes that a vm_paddr_t is equivalent to bus_addr_t */
672 	qp->tx_mw_phys = mw_base + qp_offset;
673 	KASSERT(qp->tx_mw_phys != 0, ("uh oh?"));
674 
675 	tx_size -= sizeof(struct ntb_rx_info);
676 	qp->rx_info = (void *)((char *)qp->tx_mw + tx_size);
677 
678 	/* Due to house-keeping, there must be at least 2 buffs */
679 	qp->tx_max_frame = qmin(tx_size / 2,
680 	    transport_mtu + sizeof(struct ntb_payload_header));
681 	qp->tx_max_entry = tx_size / qp->tx_max_frame;
682 
683 	callout_init(&qp->link_work, 0);
684 	callout_init(&qp->queue_full, 1);
685 	callout_init(&qp->rx_full, 1);
686 
687 	mtx_init(&qp->ntb_rx_q_lock, "ntb rx q", NULL, MTX_SPIN);
688 	mtx_init(&qp->ntb_tx_free_q_lock, "ntb tx free q", NULL, MTX_SPIN);
689 	TASK_INIT(&qp->rx_completion_task, 0, ntb_complete_rxc, qp);
690 	TASK_INIT(&qp->rxc_db_work, 0, ntb_transport_rxc_db, qp);
691 
692 	STAILQ_INIT(&qp->rx_post_q);
693 	STAILQ_INIT(&qp->rx_pend_q);
694 	STAILQ_INIT(&qp->rx_free_q);
695 	STAILQ_INIT(&qp->tx_free_q);
696 }
697 
698 static void
699 ntb_transport_free_queue(struct ntb_transport_qp *qp)
700 {
701 	struct ntb_queue_entry *entry;
702 
703 	if (qp == NULL)
704 		return;
705 
706 	callout_drain(&qp->link_work);
707 
708 	ntb_db_set_mask(qp->ntb, 1ull << qp->qp_num);
709 	taskqueue_drain(taskqueue_swi, &qp->rxc_db_work);
710 	taskqueue_drain(taskqueue_swi, &qp->rx_completion_task);
711 
712 	qp->cb_data = NULL;
713 	qp->rx_handler = NULL;
714 	qp->tx_handler = NULL;
715 	qp->event_handler = NULL;
716 
717 	while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q)))
718 		free(entry, M_NTB_IF);
719 
720 	while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_pend_q)))
721 		free(entry, M_NTB_IF);
722 
723 	while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_post_q)))
724 		free(entry, M_NTB_IF);
725 
726 	while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
727 		free(entry, M_NTB_IF);
728 
729 	set_bit(qp->qp_num, &qp->transport->qp_bitmap_free);
730 }
731 
732 /**
733  * ntb_transport_create_queue - Create a new NTB transport layer queue
734  * @rx_handler: receive callback function
735  * @tx_handler: transmit callback function
736  * @event_handler: event callback function
737  *
738  * Create a new NTB transport layer queue and provide the queue with a callback
739  * routine for both transmit and receive.  The receive callback routine will be
740  * used to pass up data when the transport has received it on the queue.   The
741  * transmit callback routine will be called when the transport has completed the
742  * transmission of the data on the queue and the data is ready to be freed.
743  *
744  * RETURNS: pointer to newly created ntb_queue, NULL on error.
745  */
746 static struct ntb_transport_qp *
747 ntb_transport_create_queue(void *data, struct ntb_softc *ntb,
748     const struct ntb_queue_handlers *handlers)
749 {
750 	struct ntb_queue_entry *entry;
751 	struct ntb_transport_qp *qp;
752 	struct ntb_transport_ctx *nt;
753 	unsigned int free_queue;
754 	int i;
755 
756 	nt = ntb_get_ctx(ntb, NULL);
757 	KASSERT(nt != NULL, ("bogus"));
758 
759 	free_queue = ffs_bit(&nt->qp_bitmap);
760 	if (free_queue == 0)
761 		return (NULL);
762 
763 	/* decrement free_queue to make it zero based */
764 	free_queue--;
765 
766 	qp = &nt->qp_vec[free_queue];
767 	clear_bit(1ull << qp->qp_num, &nt->qp_bitmap_free);
768 	qp->cb_data = data;
769 	qp->rx_handler = handlers->rx_handler;
770 	qp->tx_handler = handlers->tx_handler;
771 	qp->event_handler = handlers->event_handler;
772 
773 	for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) {
774 		entry = malloc(sizeof(*entry), M_NTB_IF, M_WAITOK | M_ZERO);
775 		entry->cb_data = nt->ifp;
776 		entry->buf = NULL;
777 		entry->len = transport_mtu;
778 		ntb_list_add(&qp->ntb_rx_q_lock, entry, &qp->rx_pend_q);
779 	}
780 
781 	for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) {
782 		entry = malloc(sizeof(*entry), M_NTB_IF, M_WAITOK | M_ZERO);
783 		ntb_list_add(&qp->ntb_tx_free_q_lock, entry, &qp->tx_free_q);
784 	}
785 
786 	ntb_db_clear(ntb, 1ull << qp->qp_num);
787 	ntb_db_clear_mask(ntb, 1ull << qp->qp_num);
788 	return (qp);
789 }
790 
791 /**
792  * ntb_transport_link_up - Notify NTB transport of client readiness to use queue
793  * @qp: NTB transport layer queue to be enabled
794  *
795  * Notify NTB transport layer of client readiness to use queue
796  */
797 static void
798 ntb_transport_link_up(struct ntb_transport_qp *qp)
799 {
800 
801 	if (qp == NULL)
802 		return;
803 
804 	qp->client_ready = true;
805 	if (bootverbose)
806 		if_printf(qp->transport->ifp, "qp client ready\n");
807 
808 	if (qp->transport->link_is_up)
809 		callout_reset(&qp->link_work, 0, ntb_qp_link_work, qp);
810 }
811 
812 
813 
814 /* Transport Tx */
815 
816 /**
817  * ntb_transport_tx_enqueue - Enqueue a new NTB queue entry
818  * @qp: NTB transport layer queue the entry is to be enqueued on
819  * @cb: per buffer pointer for callback function to use
820  * @data: pointer to data buffer that will be sent
821  * @len: length of the data buffer
822  *
823  * Enqueue a new transmit buffer onto the transport queue from which a NTB
824  * payload will be transmitted.  This assumes that a lock is being held to
825  * serialize access to the qp.
826  *
827  * RETURNS: An appropriate ERRNO error value on error, or zero for success.
828  */
829 static int
830 ntb_transport_tx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
831     unsigned int len)
832 {
833 	struct ntb_queue_entry *entry;
834 	int rc;
835 
836 	if (qp == NULL || !qp->link_is_up || len == 0) {
837 		CTR0(KTR_NTB, "TX: link not up");
838 		return (EINVAL);
839 	}
840 
841 	entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
842 	if (entry == NULL) {
843 		CTR0(KTR_NTB, "TX: could not get entry from tx_free_q");
844 		qp->tx_err_no_buf++;
845 		return (EBUSY);
846 	}
847 	CTR1(KTR_NTB, "TX: got entry %p from tx_free_q", entry);
848 
849 	entry->cb_data = cb;
850 	entry->buf = data;
851 	entry->len = len;
852 	entry->flags = 0;
853 
854 	rc = ntb_process_tx(qp, entry);
855 	if (rc != 0) {
856 		ntb_list_add(&qp->ntb_tx_free_q_lock, entry, &qp->tx_free_q);
857 		CTR1(KTR_NTB,
858 		    "TX: process_tx failed. Returning entry %p to tx_free_q",
859 		    entry);
860 	}
861 	return (rc);
862 }
863 
864 static int
865 ntb_process_tx(struct ntb_transport_qp *qp, struct ntb_queue_entry *entry)
866 {
867 	void *offset;
868 
869 	offset = (char *)qp->tx_mw + qp->tx_max_frame * qp->tx_index;
870 	CTR3(KTR_NTB,
871 	    "TX: process_tx: tx_pkts=%u, tx_index=%u, remote entry=%u",
872 	    qp->tx_pkts, qp->tx_index, qp->remote_rx_info->entry);
873 	if (qp->tx_index == qp->remote_rx_info->entry) {
874 		CTR0(KTR_NTB, "TX: ring full");
875 		qp->tx_ring_full++;
876 		return (EAGAIN);
877 	}
878 
879 	if (entry->len > qp->tx_max_frame - sizeof(struct ntb_payload_header)) {
880 		if (qp->tx_handler != NULL)
881 			qp->tx_handler(qp, qp->cb_data, entry->buf,
882 				       EIO);
883 
884 		ntb_list_add(&qp->ntb_tx_free_q_lock, entry, &qp->tx_free_q);
885 		CTR1(KTR_NTB,
886 		    "TX: frame too big. returning entry %p to tx_free_q",
887 		    entry);
888 		return (0);
889 	}
890 	CTR2(KTR_NTB, "TX: copying entry %p to offset %p", entry, offset);
891 	ntb_memcpy_tx(qp, entry, offset);
892 
893 	qp->tx_index++;
894 	qp->tx_index %= qp->tx_max_entry;
895 
896 	qp->tx_pkts++;
897 
898 	return (0);
899 }
900 
901 static void
902 ntb_memcpy_tx(struct ntb_transport_qp *qp, struct ntb_queue_entry *entry,
903     void *offset)
904 {
905 	struct ntb_payload_header *hdr;
906 
907 	/* This piece is from Linux' ntb_async_tx() */
908 	hdr = (struct ntb_payload_header *)((char *)offset + qp->tx_max_frame -
909 	    sizeof(struct ntb_payload_header));
910 	entry->x_hdr = hdr;
911 	hdr->len = entry->len; /* TODO: replace with bus_space_write */
912 	hdr->ver = qp->tx_pkts; /* TODO: replace with bus_space_write */
913 
914 	/* This piece is ntb_memcpy_tx() */
915 	CTR2(KTR_NTB, "TX: copying %d bytes to offset %p", entry->len, offset);
916 	if (entry->buf != NULL) {
917 		m_copydata((struct mbuf *)entry->buf, 0, entry->len, offset);
918 
919 		/*
920 		 * Ensure that the data is fully copied before setting the
921 		 * flags
922 		 */
923 		wmb();
924 	}
925 
926 	/* The rest is ntb_tx_copy_callback() */
927 	/* TODO: replace with bus_space_write */
928 	hdr->flags = entry->flags | IF_NTB_DESC_DONE_FLAG;
929 
930 	ntb_peer_db_set(qp->ntb, 1ull << qp->qp_num);
931 
932 	/*
933 	 * The entry length can only be zero if the packet is intended to be a
934 	 * "link down" or similar.  Since no payload is being sent in these
935 	 * cases, there is nothing to add to the completion queue.
936 	 */
937 	if (entry->len > 0) {
938 		qp->tx_bytes += entry->len;
939 
940 		if (qp->tx_handler)
941 			qp->tx_handler(qp, qp->cb_data, entry->cb_data,
942 				       entry->len);
943 	}
944 
945 	CTR2(KTR_NTB,
946 	    "TX: entry %p sent. hdr->ver = %d, Returning to tx_free_q", entry,
947 	    hdr->ver);
948 	ntb_list_add(&qp->ntb_tx_free_q_lock, entry, &qp->tx_free_q);
949 }
950 
951 static void
952 ntb_qp_full(void *arg)
953 {
954 
955 	CTR0(KTR_NTB, "TX: qp_full callout");
956 	ntb_start(arg);
957 }
958 
959 /* Transport Rx */
960 static void
961 ntb_transport_rxc_db(void *arg, int pending __unused)
962 {
963 	struct ntb_transport_qp *qp = arg;
964 	ntb_q_idx_t i;
965 	int rc;
966 
967 	/*
968 	 * Limit the number of packets processed in a single interrupt to
969 	 * provide fairness to others
970 	 */
971 	CTR0(KTR_NTB, "RX: transport_rx");
972 	mtx_lock(&qp->transport->rx_lock);
973 	for (i = 0; i < qp->rx_max_entry; i++) {
974 		rc = ntb_process_rxc(qp);
975 		if (rc != 0) {
976 			CTR0(KTR_NTB, "RX: process_rxc failed");
977 			break;
978 		}
979 	}
980 	mtx_unlock(&qp->transport->rx_lock);
981 
982 	if (i == qp->rx_max_entry)
983 		taskqueue_enqueue(taskqueue_swi, &qp->rxc_db_work);
984 	else if ((ntb_db_read(qp->ntb) & (1ull << qp->qp_num)) != 0) {
985 		/* If db is set, clear it and read it back to commit clear. */
986 		ntb_db_clear(qp->ntb, 1ull << qp->qp_num);
987 		(void)ntb_db_read(qp->ntb);
988 
989 		/*
990 		 * An interrupt may have arrived between finishing
991 		 * ntb_process_rxc and clearing the doorbell bit: there might
992 		 * be some more work to do.
993 		 */
994 		taskqueue_enqueue(taskqueue_swi, &qp->rxc_db_work);
995 	}
996 }
997 
998 static int
999 ntb_process_rxc(struct ntb_transport_qp *qp)
1000 {
1001 	struct ntb_payload_header *hdr;
1002 	struct ntb_queue_entry *entry;
1003 	void *offset;
1004 
1005 	offset = (void *)
1006 	    ((char *)qp->rx_buff + qp->rx_max_frame * qp->rx_index);
1007 	hdr = (void *)
1008 	    ((char *)offset + qp->rx_max_frame -
1009 		sizeof(struct ntb_payload_header));
1010 
1011 	CTR1(KTR_NTB, "RX: process_rxc rx_index = %u", qp->rx_index);
1012 	if ((hdr->flags & IF_NTB_DESC_DONE_FLAG) == 0) {
1013 		CTR0(KTR_NTB, "RX: hdr not done");
1014 		qp->rx_ring_empty++;
1015 		return (EAGAIN);
1016 	}
1017 
1018 	if ((hdr->flags & IF_NTB_LINK_DOWN_FLAG) != 0) {
1019 		CTR0(KTR_NTB, "RX: link down");
1020 		ntb_qp_link_down(qp);
1021 		hdr->flags = 0;
1022 		return (EAGAIN);
1023 	}
1024 
1025 	if (hdr->ver != (uint32_t)qp->rx_pkts) {
1026 		CTR2(KTR_NTB,"RX: ver != rx_pkts (%x != %lx). "
1027 		    "Returning entry %p to rx_pend_q", hdr->ver, qp->rx_pkts);
1028 		qp->rx_err_ver++;
1029 		return (EIO);
1030 	}
1031 
1032 	entry = ntb_list_mv(&qp->ntb_rx_q_lock, &qp->rx_pend_q, &qp->rx_post_q);
1033 	if (entry == NULL) {
1034 		qp->rx_err_no_buf++;
1035 		CTR0(KTR_NTB, "RX: No entries in rx_pend_q");
1036 		return (EAGAIN);
1037 	}
1038 	callout_stop(&qp->rx_full);
1039 	CTR1(KTR_NTB, "RX: rx entry %p from rx_pend_q", entry);
1040 
1041 	entry->x_hdr = hdr;
1042 	entry->index = qp->rx_index;
1043 
1044 	if (hdr->len > entry->len) {
1045 		CTR2(KTR_NTB, "RX: len too long. Wanted %ju got %ju",
1046 		    (uintmax_t)hdr->len, (uintmax_t)entry->len);
1047 		qp->rx_err_oflow++;
1048 
1049 		entry->len = -EIO;
1050 		entry->flags |= IF_NTB_DESC_DONE_FLAG;
1051 
1052 		taskqueue_enqueue(taskqueue_swi, &qp->rx_completion_task);
1053 	} else {
1054 		qp->rx_bytes += hdr->len;
1055 		qp->rx_pkts++;
1056 
1057 		CTR1(KTR_NTB, "RX: received %ld rx_pkts", qp->rx_pkts);
1058 
1059 		entry->len = hdr->len;
1060 
1061 		ntb_memcpy_rx(qp, entry, offset);
1062 	}
1063 
1064 	qp->rx_index++;
1065 	qp->rx_index %= qp->rx_max_entry;
1066 	return (0);
1067 }
1068 
1069 static void
1070 ntb_memcpy_rx(struct ntb_transport_qp *qp, struct ntb_queue_entry *entry,
1071     void *offset)
1072 {
1073 	struct ifnet *ifp = entry->cb_data;
1074 	unsigned int len = entry->len;
1075 	struct mbuf *m;
1076 
1077 	CTR2(KTR_NTB, "RX: copying %d bytes from offset %p", len, offset);
1078 	m = m_devget(offset, len, 0, ifp, NULL);
1079 	m->m_pkthdr.csum_flags = CSUM_IP_CHECKED | CSUM_IP_VALID;
1080 
1081 	entry->buf = (void *)m;
1082 
1083 	/* Ensure that the data is globally visible before clearing the flag */
1084 	wmb();
1085 
1086 	CTR2(KTR_NTB, "RX: copied entry %p to mbuf %p.", entry, m);
1087 	ntb_rx_copy_callback(qp, entry);
1088 }
1089 
1090 static inline void
1091 ntb_rx_copy_callback(struct ntb_transport_qp *qp, void *data)
1092 {
1093 	struct ntb_queue_entry *entry;
1094 
1095 	entry = data;
1096 	entry->flags |= IF_NTB_DESC_DONE_FLAG;
1097 	taskqueue_enqueue(taskqueue_swi, &qp->rx_completion_task);
1098 }
1099 
1100 static void
1101 ntb_complete_rxc(void *arg, int pending)
1102 {
1103 	struct ntb_transport_qp *qp = arg;
1104 	struct ntb_queue_entry *entry;
1105 	struct mbuf *m;
1106 	unsigned len;
1107 
1108 	CTR0(KTR_NTB, "RX: rx_completion_task");
1109 
1110 	mtx_lock_spin(&qp->ntb_rx_q_lock);
1111 
1112 	while (!STAILQ_EMPTY(&qp->rx_post_q)) {
1113 		entry = STAILQ_FIRST(&qp->rx_post_q);
1114 		if ((entry->flags & IF_NTB_DESC_DONE_FLAG) == 0)
1115 			break;
1116 
1117 		entry->x_hdr->flags = 0;
1118 		/* XXX bus_space_write */
1119 		qp->rx_info->entry = entry->index;
1120 
1121 		len = entry->len;
1122 		m = entry->buf;
1123 
1124 		STAILQ_REMOVE_HEAD(&qp->rx_post_q, entry);
1125 		STAILQ_INSERT_TAIL(&qp->rx_free_q, entry, entry);
1126 
1127 		mtx_unlock_spin(&qp->ntb_rx_q_lock);
1128 
1129 		CTR2(KTR_NTB, "RX: completing entry %p, mbuf %p", entry, m);
1130 		if (qp->rx_handler != NULL && qp->client_ready)
1131 			qp->rx_handler(qp, qp->cb_data, m, len);
1132 
1133 		mtx_lock_spin(&qp->ntb_rx_q_lock);
1134 	}
1135 
1136 	mtx_unlock_spin(&qp->ntb_rx_q_lock);
1137 }
1138 
1139 static void
1140 ntb_transport_doorbell_callback(void *data, uint32_t vector)
1141 {
1142 	struct ntb_transport_ctx *nt = data;
1143 	struct ntb_transport_qp *qp;
1144 	struct _qpset db_bits;
1145 	uint64_t vec_mask;
1146 	unsigned qp_num;
1147 
1148 	BIT_COPY(QP_SETSIZE, &nt->qp_bitmap, &db_bits);
1149 	BIT_NAND(QP_SETSIZE, &db_bits, &nt->qp_bitmap_free);
1150 
1151 	vec_mask = ntb_db_vector_mask(nt->ntb, vector);
1152 	while (vec_mask != 0) {
1153 		qp_num = ffsll(vec_mask) - 1;
1154 
1155 		if (test_bit(qp_num, &db_bits)) {
1156 			qp = &nt->qp_vec[qp_num];
1157 			taskqueue_enqueue(taskqueue_swi, &qp->rxc_db_work);
1158 		}
1159 
1160 		vec_mask &= ~(1ull << qp_num);
1161 	}
1162 }
1163 
1164 /* Link Event handler */
1165 static void
1166 ntb_transport_event_callback(void *data)
1167 {
1168 	struct ntb_transport_ctx *nt = data;
1169 
1170 	if (ntb_link_is_up(nt->ntb, NULL, NULL)) {
1171 		if (bootverbose)
1172 			if_printf(nt->ifp, "HW link up\n");
1173 		callout_reset(&nt->link_work, 0, ntb_transport_link_work, nt);
1174 	} else {
1175 		if (bootverbose)
1176 			if_printf(nt->ifp, "HW link down\n");
1177 		ntb_transport_link_cleanup(nt);
1178 	}
1179 }
1180 
1181 /* Link bring up */
1182 static void
1183 ntb_transport_link_work(void *arg)
1184 {
1185 	struct ntb_transport_ctx *nt = arg;
1186 	struct ntb_softc *ntb = nt->ntb;
1187 	struct ntb_transport_qp *qp;
1188 	uint64_t val64, size;
1189 	uint32_t val;
1190 	unsigned i;
1191 	int rc;
1192 
1193 	/* send the local info, in the opposite order of the way we read it */
1194 	for (i = 0; i < nt->mw_count; i++) {
1195 		size = nt->mw_vec[i].phys_size;
1196 
1197 		if (max_mw_size != 0 && size > max_mw_size)
1198 			size = max_mw_size;
1199 
1200 		ntb_peer_spad_write(ntb, IF_NTB_MW0_SZ_HIGH + (i * 2),
1201 		    size >> 32);
1202 		ntb_peer_spad_write(ntb, IF_NTB_MW0_SZ_LOW + (i * 2), size);
1203 	}
1204 
1205 	ntb_peer_spad_write(ntb, IF_NTB_NUM_MWS, nt->mw_count);
1206 
1207 	ntb_peer_spad_write(ntb, IF_NTB_NUM_QPS, nt->qp_count);
1208 
1209 	ntb_peer_spad_write(ntb, IF_NTB_VERSION, NTB_TRANSPORT_VERSION);
1210 
1211 	/* Query the remote side for its info */
1212 	val = 0;
1213 	ntb_spad_read(ntb, IF_NTB_VERSION, &val);
1214 	if (val != NTB_TRANSPORT_VERSION)
1215 		goto out;
1216 
1217 	ntb_spad_read(ntb, IF_NTB_NUM_QPS, &val);
1218 	if (val != nt->qp_count)
1219 		goto out;
1220 
1221 	ntb_spad_read(ntb, IF_NTB_NUM_MWS, &val);
1222 	if (val != nt->mw_count)
1223 		goto out;
1224 
1225 	for (i = 0; i < nt->mw_count; i++) {
1226 		ntb_spad_read(ntb, IF_NTB_MW0_SZ_HIGH + (i * 2), &val);
1227 		val64 = (uint64_t)val << 32;
1228 
1229 		ntb_spad_read(ntb, IF_NTB_MW0_SZ_LOW + (i * 2), &val);
1230 		val64 |= val;
1231 
1232 		rc = ntb_set_mw(nt, i, val64);
1233 		if (rc != 0)
1234 			goto free_mws;
1235 	}
1236 
1237 	nt->link_is_up = true;
1238 	if (bootverbose)
1239 		if_printf(nt->ifp, "transport link up\n");
1240 
1241 	for (i = 0; i < nt->qp_count; i++) {
1242 		qp = &nt->qp_vec[i];
1243 
1244 		ntb_transport_setup_qp_mw(nt, i);
1245 
1246 		if (qp->client_ready)
1247 			callout_reset(&qp->link_work, 0, ntb_qp_link_work, qp);
1248 	}
1249 
1250 	return;
1251 
1252 free_mws:
1253 	for (i = 0; i < nt->mw_count; i++)
1254 		ntb_free_mw(nt, i);
1255 out:
1256 	if (ntb_link_is_up(ntb, NULL, NULL))
1257 		callout_reset(&nt->link_work,
1258 		    NTB_LINK_DOWN_TIMEOUT * hz / 1000, ntb_transport_link_work, nt);
1259 }
1260 
1261 static int
1262 ntb_set_mw(struct ntb_transport_ctx *nt, int num_mw, size_t size)
1263 {
1264 	struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
1265 	size_t xlat_size, buff_size;
1266 	int rc;
1267 
1268 	if (size == 0)
1269 		return (EINVAL);
1270 
1271 	xlat_size = roundup(size, mw->xlat_align_size);
1272 	buff_size = roundup(size, mw->xlat_align);
1273 
1274 	/* No need to re-setup */
1275 	if (mw->xlat_size == xlat_size)
1276 		return (0);
1277 
1278 	if (mw->buff_size != 0)
1279 		ntb_free_mw(nt, num_mw);
1280 
1281 	/* Alloc memory for receiving data.  Must be aligned */
1282 	mw->xlat_size = xlat_size;
1283 	mw->buff_size = buff_size;
1284 
1285 	mw->virt_addr = contigmalloc(mw->buff_size, M_NTB_IF, M_ZERO, 0,
1286 	    BUS_SPACE_MAXADDR, mw->xlat_align, 0);
1287 	if (mw->virt_addr == NULL) {
1288 		mw->xlat_size = 0;
1289 		mw->buff_size = 0;
1290 		printf("ntb: Unable to allocate MW buffer of size %zu\n",
1291 		    mw->xlat_size);
1292 		return (ENOMEM);
1293 	}
1294 	/* TODO: replace with bus_space_* functions */
1295 	mw->dma_addr = vtophys(mw->virt_addr);
1296 
1297 	/*
1298 	 * Ensure that the allocation from contigmalloc is aligned as
1299 	 * requested.  XXX: This may not be needed -- brought in for parity
1300 	 * with the Linux driver.
1301 	 */
1302 	if (mw->dma_addr % mw->xlat_align != 0) {
1303 		if_printf(nt->ifp,
1304 		    "DMA memory 0x%jx not aligned to BAR size 0x%zx\n",
1305 		    (uintmax_t)mw->dma_addr, size);
1306 		ntb_free_mw(nt, num_mw);
1307 		return (ENOMEM);
1308 	}
1309 
1310 	/* Notify HW the memory location of the receive buffer */
1311 	rc = ntb_mw_set_trans(nt->ntb, num_mw, mw->dma_addr, mw->xlat_size);
1312 	if (rc) {
1313 		if_printf(nt->ifp, "Unable to set mw%d translation", num_mw);
1314 		ntb_free_mw(nt, num_mw);
1315 		return (rc);
1316 	}
1317 
1318 	return (0);
1319 }
1320 
1321 static void
1322 ntb_free_mw(struct ntb_transport_ctx *nt, int num_mw)
1323 {
1324 	struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
1325 
1326 	if (mw->virt_addr == NULL)
1327 		return;
1328 
1329 	ntb_mw_clear_trans(nt->ntb, num_mw);
1330 	contigfree(mw->virt_addr, mw->xlat_size, M_NTB_IF);
1331 	mw->xlat_size = 0;
1332 	mw->buff_size = 0;
1333 	mw->virt_addr = NULL;
1334 }
1335 
1336 static int
1337 ntb_transport_setup_qp_mw(struct ntb_transport_ctx *nt, unsigned int qp_num)
1338 {
1339 	struct ntb_transport_qp *qp = &nt->qp_vec[qp_num];
1340 	struct ntb_transport_mw *mw;
1341 	void *offset;
1342 	ntb_q_idx_t i;
1343 	size_t rx_size;
1344 	unsigned num_qps_mw, mw_num, mw_count;
1345 
1346 	mw_count = nt->mw_count;
1347 	mw_num = QP_TO_MW(nt, qp_num);
1348 	mw = &nt->mw_vec[mw_num];
1349 
1350 	if (mw->virt_addr == NULL)
1351 		return (ENOMEM);
1352 
1353 	if (nt->qp_count % mw_count && mw_num + 1 < nt->qp_count / mw_count)
1354 		num_qps_mw = nt->qp_count / mw_count + 1;
1355 	else
1356 		num_qps_mw = nt->qp_count / mw_count;
1357 
1358 	rx_size = mw->xlat_size / num_qps_mw;
1359 	qp->rx_buff = (char *)mw->virt_addr + rx_size * qp_num / mw_count;
1360 	rx_size -= sizeof(struct ntb_rx_info);
1361 
1362 	qp->remote_rx_info = (void*)((char *)qp->rx_buff + rx_size);
1363 
1364 	/* Due to house-keeping, there must be at least 2 buffs */
1365 	qp->rx_max_frame = qmin(rx_size / 2,
1366 	    transport_mtu + sizeof(struct ntb_payload_header));
1367 	qp->rx_max_entry = rx_size / qp->rx_max_frame;
1368 	qp->rx_index = 0;
1369 
1370 	qp->remote_rx_info->entry = qp->rx_max_entry - 1;
1371 
1372 	/* Set up the hdr offsets with 0s */
1373 	for (i = 0; i < qp->rx_max_entry; i++) {
1374 		offset = (void *)((uint8_t *)qp->rx_buff +
1375 		    qp->rx_max_frame * (i + 1) -
1376 		    sizeof(struct ntb_payload_header));
1377 		memset(offset, 0, sizeof(struct ntb_payload_header));
1378 	}
1379 
1380 	qp->rx_pkts = 0;
1381 	qp->tx_pkts = 0;
1382 	qp->tx_index = 0;
1383 
1384 	return (0);
1385 }
1386 
1387 static void
1388 ntb_qp_link_work(void *arg)
1389 {
1390 	struct ntb_transport_qp *qp = arg;
1391 	struct ntb_softc *ntb = qp->ntb;
1392 	struct ntb_transport_ctx *nt = qp->transport;
1393 	uint32_t val, dummy;
1394 
1395 	ntb_spad_read(ntb, IF_NTB_QP_LINKS, &val);
1396 
1397 	ntb_peer_spad_write(ntb, IF_NTB_QP_LINKS, val | (1ull << qp->qp_num));
1398 
1399 	/* query remote spad for qp ready bits */
1400 	ntb_peer_spad_read(ntb, IF_NTB_QP_LINKS, &dummy);
1401 
1402 	/* See if the remote side is up */
1403 	if ((val & (1ull << qp->qp_num)) != 0) {
1404 		if (bootverbose)
1405 			if_printf(nt->ifp, "qp link up\n");
1406 		qp->link_is_up = true;
1407 
1408 		if (qp->event_handler != NULL)
1409 			qp->event_handler(qp->cb_data, NTB_LINK_UP);
1410 
1411 		taskqueue_enqueue(taskqueue_swi, &qp->rxc_db_work);
1412 	} else if (nt->link_is_up)
1413 		callout_reset(&qp->link_work,
1414 		    NTB_LINK_DOWN_TIMEOUT * hz / 1000, ntb_qp_link_work, qp);
1415 }
1416 
1417 /* Link down event*/
1418 static void
1419 ntb_transport_link_cleanup(struct ntb_transport_ctx *nt)
1420 {
1421 	struct ntb_transport_qp *qp;
1422 	struct _qpset qp_bitmap_alloc;
1423 	unsigned i;
1424 
1425 	BIT_COPY(QP_SETSIZE, &nt->qp_bitmap, &qp_bitmap_alloc);
1426 	BIT_NAND(QP_SETSIZE, &qp_bitmap_alloc, &nt->qp_bitmap_free);
1427 
1428 	/* Pass along the info to any clients */
1429 	for (i = 0; i < nt->qp_count; i++)
1430 		if (test_bit(i, &qp_bitmap_alloc)) {
1431 			qp = &nt->qp_vec[i];
1432 			ntb_qp_link_cleanup(qp);
1433 			callout_drain(&qp->link_work);
1434 		}
1435 
1436 	if (!nt->link_is_up)
1437 		callout_drain(&nt->link_work);
1438 
1439 	/*
1440 	 * The scratchpad registers keep the values if the remote side
1441 	 * goes down, blast them now to give them a sane value the next
1442 	 * time they are accessed
1443 	 */
1444 	for (i = 0; i < IF_NTB_MAX_SPAD; i++)
1445 		ntb_spad_write(nt->ntb, i, 0);
1446 }
1447 
1448 
1449 static void
1450 ntb_qp_link_down(struct ntb_transport_qp *qp)
1451 {
1452 
1453 	ntb_qp_link_cleanup(qp);
1454 }
1455 
1456 static void
1457 ntb_qp_link_down_reset(struct ntb_transport_qp *qp)
1458 {
1459 
1460 	qp->link_is_up = false;
1461 
1462 	qp->tx_index = qp->rx_index = 0;
1463 	qp->tx_bytes = qp->rx_bytes = 0;
1464 	qp->tx_pkts = qp->rx_pkts = 0;
1465 
1466 	qp->rx_ring_empty = 0;
1467 	qp->tx_ring_full = 0;
1468 
1469 	qp->rx_err_no_buf = qp->tx_err_no_buf = 0;
1470 	qp->rx_err_oflow = qp->rx_err_ver = 0;
1471 }
1472 
1473 static void
1474 ntb_qp_link_cleanup(struct ntb_transport_qp *qp)
1475 {
1476 	struct ntb_transport_ctx *nt = qp->transport;
1477 
1478 	callout_drain(&qp->link_work);
1479 	ntb_qp_link_down_reset(qp);
1480 
1481 	if (qp->event_handler != NULL)
1482 		qp->event_handler(qp->cb_data, NTB_LINK_DOWN);
1483 
1484 	if (nt->link_is_up)
1485 		callout_reset(&qp->link_work,
1486 		    NTB_LINK_DOWN_TIMEOUT * hz / 1000, ntb_qp_link_work, qp);
1487 }
1488 
1489 /* Link commanded down */
1490 /**
1491  * ntb_transport_link_down - Notify NTB transport to no longer enqueue data
1492  * @qp: NTB transport layer queue to be disabled
1493  *
1494  * Notify NTB transport layer of client's desire to no longer receive data on
1495  * transport queue specified.  It is the client's responsibility to ensure all
1496  * entries on queue are purged or otherwise handled appropriately.
1497  */
1498 static void
1499 ntb_transport_link_down(struct ntb_transport_qp *qp)
1500 {
1501 	uint32_t val;
1502 
1503 	if (qp == NULL)
1504 		return;
1505 
1506 	qp->client_ready = false;
1507 
1508 	ntb_spad_read(qp->ntb, IF_NTB_QP_LINKS, &val);
1509 
1510 	ntb_peer_spad_write(qp->ntb, IF_NTB_QP_LINKS,
1511 	   val & ~(1 << qp->qp_num));
1512 
1513 	if (qp->link_is_up)
1514 		ntb_send_link_down(qp);
1515 	else
1516 		callout_drain(&qp->link_work);
1517 }
1518 
1519 static void
1520 ntb_send_link_down(struct ntb_transport_qp *qp)
1521 {
1522 	struct ntb_queue_entry *entry;
1523 	int i, rc;
1524 
1525 	if (!qp->link_is_up)
1526 		return;
1527 
1528 	for (i = 0; i < NTB_LINK_DOWN_TIMEOUT; i++) {
1529 		entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
1530 		if (entry != NULL)
1531 			break;
1532 		pause("NTB Wait for link down", hz / 10);
1533 	}
1534 
1535 	if (entry == NULL)
1536 		return;
1537 
1538 	entry->cb_data = NULL;
1539 	entry->buf = NULL;
1540 	entry->len = 0;
1541 	entry->flags = IF_NTB_LINK_DOWN_FLAG;
1542 
1543 	mtx_lock(&qp->transport->tx_lock);
1544 	rc = ntb_process_tx(qp, entry);
1545 	if (rc != 0)
1546 		printf("ntb: Failed to send link down\n");
1547 	mtx_unlock(&qp->transport->tx_lock);
1548 
1549 	ntb_qp_link_down_reset(qp);
1550 }
1551 
1552 
1553 /* List Management */
1554 
1555 static void
1556 ntb_list_add(struct mtx *lock, struct ntb_queue_entry *entry,
1557     struct ntb_queue_list *list)
1558 {
1559 
1560 	mtx_lock_spin(lock);
1561 	STAILQ_INSERT_TAIL(list, entry, entry);
1562 	mtx_unlock_spin(lock);
1563 }
1564 
1565 static struct ntb_queue_entry *
1566 ntb_list_rm(struct mtx *lock, struct ntb_queue_list *list)
1567 {
1568 	struct ntb_queue_entry *entry;
1569 
1570 	mtx_lock_spin(lock);
1571 	if (STAILQ_EMPTY(list)) {
1572 		entry = NULL;
1573 		goto out;
1574 	}
1575 	entry = STAILQ_FIRST(list);
1576 	STAILQ_REMOVE_HEAD(list, entry);
1577 out:
1578 	mtx_unlock_spin(lock);
1579 
1580 	return (entry);
1581 }
1582 
1583 static struct ntb_queue_entry *
1584 ntb_list_mv(struct mtx *lock, struct ntb_queue_list *from,
1585     struct ntb_queue_list *to)
1586 {
1587 	struct ntb_queue_entry *entry;
1588 
1589 	mtx_lock_spin(lock);
1590 	if (STAILQ_EMPTY(from)) {
1591 		entry = NULL;
1592 		goto out;
1593 	}
1594 	entry = STAILQ_FIRST(from);
1595 	STAILQ_REMOVE_HEAD(from, entry);
1596 	STAILQ_INSERT_TAIL(to, entry, entry);
1597 
1598 out:
1599 	mtx_unlock_spin(lock);
1600 	return (entry);
1601 }
1602 
1603 /* Helper functions */
1604 /* TODO: This too should really be part of the kernel */
1605 #define EUI48_MULTICAST			1 << 0
1606 #define EUI48_LOCALLY_ADMINISTERED	1 << 1
1607 static void
1608 create_random_local_eui48(u_char *eaddr)
1609 {
1610 	static uint8_t counter = 0;
1611 	uint32_t seed = ticks;
1612 
1613 	eaddr[0] = EUI48_LOCALLY_ADMINISTERED;
1614 	memcpy(&eaddr[1], &seed, sizeof(uint32_t));
1615 	eaddr[5] = counter++;
1616 }
1617 
1618 /**
1619  * ntb_transport_max_size - Query the max payload size of a qp
1620  * @qp: NTB transport layer queue to be queried
1621  *
1622  * Query the maximum payload size permissible on the given qp
1623  *
1624  * RETURNS: the max payload size of a qp
1625  */
1626 static unsigned int
1627 ntb_transport_max_size(struct ntb_transport_qp *qp)
1628 {
1629 
1630 	if (qp == NULL)
1631 		return (0);
1632 
1633 	return (qp->tx_max_frame - sizeof(struct ntb_payload_header));
1634 }
1635