xref: /freebsd/sys/dev/ntb/ntb_transport.c (revision c22994e3ad2d05f45f5e9fc78140a5d6b3d7e567)
1 /*-
2  * Copyright (c) 2016-2017 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 = 256*1024*1024;
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 vbase / phys_addr / tx_size */
181 	caddr_t		vbase;
182 	size_t		tx_size;
183 	/* Rx buff is virt_addr / dma_addr / rx_size */
184 	bus_dma_tag_t	dma_tag;
185 	bus_dmamap_t	dma_map;
186 	caddr_t		virt_addr;
187 	bus_addr_t	dma_addr;
188 	size_t		rx_size;
189 	/* rx_size increased to size alignment requirements of the hardware. */
190 	size_t		buff_size;
191 };
192 
193 struct ntb_transport_child {
194 	device_t	dev;
195 	int		consumer;
196 	int		qpoff;
197 	int		qpcnt;
198 	struct ntb_transport_child *next;
199 };
200 
201 struct ntb_transport_ctx {
202 	device_t		 dev;
203 	struct ntb_transport_child *child;
204 	struct ntb_transport_mw	*mw_vec;
205 	struct ntb_transport_qp	*qp_vec;
206 	unsigned		mw_count;
207 	unsigned		qp_count;
208 	uint64_t		qp_bitmap;
209 	volatile bool		link_is_up;
210 	enum ntb_speed		link_speed;
211 	enum ntb_width		link_width;
212 	struct callout		link_work;
213 	struct callout		link_watchdog;
214 	struct task		link_cleanup;
215 };
216 
217 enum {
218 	NTBT_DESC_DONE_FLAG = 1 << 0,
219 	NTBT_LINK_DOWN_FLAG = 1 << 1,
220 };
221 
222 struct ntb_payload_header {
223 	ntb_q_idx_t ver;
224 	uint32_t len;
225 	uint32_t flags;
226 };
227 
228 enum {
229 	/*
230 	 * The order of this enum is part of the remote protocol.  Do not
231 	 * reorder without bumping protocol version (and it's probably best
232 	 * to keep the protocol in lock-step with the Linux NTB driver.
233 	 */
234 	NTBT_VERSION = 0,
235 	NTBT_QP_LINKS,
236 	NTBT_NUM_QPS,
237 	NTBT_NUM_MWS,
238 	/*
239 	 * N.B.: transport_link_work assumes MW1 enums = MW0 + 2.
240 	 */
241 	NTBT_MW0_SZ_HIGH,
242 	NTBT_MW0_SZ_LOW,
243 	NTBT_MW1_SZ_HIGH,
244 	NTBT_MW1_SZ_LOW,
245 
246 	/*
247 	 * Some NTB-using hardware have a watchdog to work around NTB hangs; if
248 	 * a register or doorbell isn't written every few seconds, the link is
249 	 * torn down.  Write an otherwise unused register every few seconds to
250 	 * work around this watchdog.
251 	 */
252 	NTBT_WATCHDOG_SPAD = 15
253 };
254 
255 #define QP_TO_MW(nt, qp)	((qp) % nt->mw_count)
256 #define NTB_QP_DEF_NUM_ENTRIES	100
257 #define NTB_LINK_DOWN_TIMEOUT	100
258 
259 static int ntb_transport_probe(device_t dev);
260 static int ntb_transport_attach(device_t dev);
261 static int ntb_transport_detach(device_t dev);
262 static void ntb_transport_init_queue(struct ntb_transport_ctx *nt,
263     unsigned int qp_num);
264 static int ntb_process_tx(struct ntb_transport_qp *qp,
265     struct ntb_queue_entry *entry);
266 static void ntb_transport_rxc_db(void *arg, int pending);
267 static int ntb_process_rxc(struct ntb_transport_qp *qp);
268 static void ntb_memcpy_rx(struct ntb_transport_qp *qp,
269     struct ntb_queue_entry *entry, void *offset);
270 static inline void ntb_rx_copy_callback(struct ntb_transport_qp *qp,
271     void *data);
272 static void ntb_complete_rxc(struct ntb_transport_qp *qp);
273 static void ntb_transport_doorbell_callback(void *data, uint32_t vector);
274 static void ntb_transport_event_callback(void *data);
275 static void ntb_transport_link_work(void *arg);
276 static int ntb_set_mw(struct ntb_transport_ctx *, int num_mw, size_t size);
277 static void ntb_free_mw(struct ntb_transport_ctx *nt, int num_mw);
278 static int ntb_transport_setup_qp_mw(struct ntb_transport_ctx *nt,
279     unsigned int qp_num);
280 static void ntb_qp_link_work(void *arg);
281 static void ntb_transport_link_cleanup(struct ntb_transport_ctx *nt);
282 static void ntb_transport_link_cleanup_work(void *, int);
283 static void ntb_qp_link_down(struct ntb_transport_qp *qp);
284 static void ntb_qp_link_down_reset(struct ntb_transport_qp *qp);
285 static void ntb_qp_link_cleanup(struct ntb_transport_qp *qp);
286 static void ntb_send_link_down(struct ntb_transport_qp *qp);
287 static void ntb_list_add(struct mtx *lock, struct ntb_queue_entry *entry,
288     struct ntb_queue_list *list);
289 static struct ntb_queue_entry *ntb_list_rm(struct mtx *lock,
290     struct ntb_queue_list *list);
291 static struct ntb_queue_entry *ntb_list_mv(struct mtx *lock,
292     struct ntb_queue_list *from, struct ntb_queue_list *to);
293 static void xeon_link_watchdog_hb(void *);
294 
295 static const struct ntb_ctx_ops ntb_transport_ops = {
296 	.link_event = ntb_transport_event_callback,
297 	.db_event = ntb_transport_doorbell_callback,
298 };
299 
300 MALLOC_DEFINE(M_NTB_T, "ntb_transport", "ntb transport driver");
301 
302 static inline void
303 iowrite32(uint32_t val, void *addr)
304 {
305 
306 	bus_space_write_4(X86_BUS_SPACE_MEM, 0/* HACK */, (uintptr_t)addr,
307 	    val);
308 }
309 
310 /* Transport Init and teardown */
311 
312 static void
313 xeon_link_watchdog_hb(void *arg)
314 {
315 	struct ntb_transport_ctx *nt;
316 
317 	nt = arg;
318 	ntb_spad_write(nt->dev, NTBT_WATCHDOG_SPAD, 0);
319 	callout_reset(&nt->link_watchdog, 1 * hz, xeon_link_watchdog_hb, nt);
320 }
321 
322 static int
323 ntb_transport_probe(device_t dev)
324 {
325 
326 	device_set_desc(dev, "NTB Transport");
327 	return (0);
328 }
329 
330 static int
331 ntb_transport_attach(device_t dev)
332 {
333 	struct ntb_transport_ctx *nt = device_get_softc(dev);
334 	struct ntb_transport_child **cpp = &nt->child;
335 	struct ntb_transport_child *nc;
336 	struct ntb_transport_mw *mw;
337 	uint64_t db_bitmap;
338 	int rc, i, db_count, spad_count, qp, qpu, qpo, qpt;
339 	char cfg[128] = "";
340 	char buf[32];
341 	char *n, *np, *c, *name;
342 
343 	nt->dev = dev;
344 	nt->mw_count = ntb_mw_count(dev);
345 	spad_count = ntb_spad_count(dev);
346 	db_bitmap = ntb_db_valid_mask(dev);
347 	db_count = flsll(db_bitmap);
348 	KASSERT(db_bitmap == (1 << db_count) - 1,
349 	    ("Doorbells are not sequential (%jx).\n", db_bitmap));
350 
351 	if (nt->mw_count == 0) {
352 		device_printf(dev, "At least 1 memory window required.\n");
353 		return (ENXIO);
354 	}
355 	if (spad_count < 6) {
356 		device_printf(dev, "At least 6 scratchpads required.\n");
357 		return (ENXIO);
358 	}
359 	if (spad_count < 4 + 2 * nt->mw_count) {
360 		nt->mw_count = (spad_count - 4) / 2;
361 		device_printf(dev, "Scratchpads enough only for %d "
362 		    "memory windows.\n", nt->mw_count);
363 	}
364 	if (db_bitmap == 0) {
365 		device_printf(dev, "At least one doorbell required.\n");
366 		return (ENXIO);
367 	}
368 
369 	nt->mw_vec = malloc(nt->mw_count * sizeof(*nt->mw_vec), M_NTB_T,
370 	    M_WAITOK | M_ZERO);
371 	for (i = 0; i < nt->mw_count; i++) {
372 		mw = &nt->mw_vec[i];
373 
374 		rc = ntb_mw_get_range(dev, i, &mw->phys_addr, &mw->vbase,
375 		    &mw->phys_size, &mw->xlat_align, &mw->xlat_align_size,
376 		    &mw->addr_limit);
377 		if (rc != 0)
378 			goto err;
379 
380 		mw->tx_size = mw->phys_size;
381 		if (max_mw_size != 0 && mw->tx_size > max_mw_size) {
382 			device_printf(dev, "Memory window %d limited from "
383 			    "%ju to %ju\n", i, (uintmax_t)mw->phys_size,
384 			    max_mw_size);
385 			mw->tx_size = max_mw_size;
386 		}
387 
388 		mw->rx_size = 0;
389 		mw->buff_size = 0;
390 		mw->virt_addr = NULL;
391 		mw->dma_addr = 0;
392 
393 		rc = ntb_mw_set_wc(dev, i, VM_MEMATTR_WRITE_COMBINING);
394 		if (rc)
395 			ntb_printf(0, "Unable to set mw%d caching\n", i);
396 
397 		/*
398 		 * Try to preallocate receive memory early, since there may
399 		 * be not enough contiguous memory later.  It is quite likely
400 		 * that NTB windows are symmetric and this allocation remain,
401 		 * but even if not, we will just reallocate it later.
402 		 */
403 		ntb_set_mw(nt, i, mw->tx_size);
404 	}
405 
406 	qpu = 0;
407 	qpo = imin(db_count, nt->mw_count);
408 	qpt = db_count;
409 
410 	snprintf(buf, sizeof(buf), "hint.%s.%d.config", device_get_name(dev),
411 	    device_get_unit(dev));
412 	TUNABLE_STR_FETCH(buf, cfg, sizeof(cfg));
413 	n = cfg;
414 	i = 0;
415 	while ((c = strsep(&n, ",")) != NULL) {
416 		np = c;
417 		name = strsep(&np, ":");
418 		if (name != NULL && name[0] == 0)
419 			name = NULL;
420 		qp = (np && np[0] != 0) ? strtol(np, NULL, 10) : qpo - qpu;
421 		if (qp <= 0)
422 			qp = 1;
423 
424 		if (qp > qpt - qpu) {
425 			device_printf(dev, "Not enough resources for config\n");
426 			break;
427 		}
428 
429 		nc = malloc(sizeof(*nc), M_DEVBUF, M_WAITOK | M_ZERO);
430 		nc->consumer = i;
431 		nc->qpoff = qpu;
432 		nc->qpcnt = qp;
433 		nc->dev = device_add_child(dev, name, -1);
434 		if (nc->dev == NULL) {
435 			device_printf(dev, "Can not add child.\n");
436 			break;
437 		}
438 		device_set_ivars(nc->dev, nc);
439 		*cpp = nc;
440 		cpp = &nc->next;
441 
442 		if (bootverbose) {
443 			device_printf(dev, "%d \"%s\": queues %d",
444 			    i, name, qpu);
445 			if (qp > 1)
446 				printf("-%d", qpu + qp - 1);
447 			printf("\n");
448 		}
449 
450 		qpu += qp;
451 		i++;
452 	}
453 	nt->qp_count = qpu;
454 
455 	nt->qp_vec = malloc(nt->qp_count * sizeof(*nt->qp_vec), M_NTB_T,
456 	    M_WAITOK | M_ZERO);
457 
458 	for (i = 0; i < nt->qp_count; i++)
459 		ntb_transport_init_queue(nt, i);
460 
461 	callout_init(&nt->link_work, 0);
462 	callout_init(&nt->link_watchdog, 0);
463 	TASK_INIT(&nt->link_cleanup, 0, ntb_transport_link_cleanup_work, nt);
464 	nt->link_is_up = false;
465 
466 	rc = ntb_set_ctx(dev, nt, &ntb_transport_ops);
467 	if (rc != 0)
468 		goto err;
469 
470 	ntb_link_enable(dev, NTB_SPEED_AUTO, NTB_WIDTH_AUTO);
471 
472 	for (i = 0; i < nt->mw_count; i++) {
473 		mw = &nt->mw_vec[i];
474 		rc = ntb_mw_set_trans(nt->dev, i, mw->dma_addr, mw->buff_size);
475 		if (rc != 0)
476 			ntb_printf(0, "load time mw%d xlat fails, rc %d\n", i, rc);
477 	}
478 
479 	if (enable_xeon_watchdog != 0)
480 		callout_reset(&nt->link_watchdog, 0, xeon_link_watchdog_hb, nt);
481 
482 	bus_generic_attach(dev);
483 	return (0);
484 
485 err:
486 	free(nt->qp_vec, M_NTB_T);
487 	free(nt->mw_vec, M_NTB_T);
488 	return (rc);
489 }
490 
491 static int
492 ntb_transport_detach(device_t dev)
493 {
494 	struct ntb_transport_ctx *nt = device_get_softc(dev);
495 	struct ntb_transport_child **cpp = &nt->child;
496 	struct ntb_transport_child *nc;
497 	int error = 0, i;
498 
499 	while ((nc = *cpp) != NULL) {
500 		*cpp = (*cpp)->next;
501 		error = device_delete_child(dev, nc->dev);
502 		if (error)
503 			break;
504 		free(nc, M_DEVBUF);
505 	}
506 	KASSERT(nt->qp_bitmap == 0,
507 	    ("Some queues not freed on detach (%jx)", nt->qp_bitmap));
508 
509 	ntb_transport_link_cleanup(nt);
510 	taskqueue_drain(taskqueue_swi, &nt->link_cleanup);
511 	callout_drain(&nt->link_work);
512 	callout_drain(&nt->link_watchdog);
513 
514 	ntb_link_disable(dev);
515 	ntb_clear_ctx(dev);
516 
517 	for (i = 0; i < nt->mw_count; i++)
518 		ntb_free_mw(nt, i);
519 
520 	free(nt->qp_vec, M_NTB_T);
521 	free(nt->mw_vec, M_NTB_T);
522 	return (0);
523 }
524 
525 static int
526 ntb_transport_print_child(device_t dev, device_t child)
527 {
528 	struct ntb_transport_child *nc = device_get_ivars(child);
529 	int retval;
530 
531 	retval = bus_print_child_header(dev, child);
532 	if (nc->qpcnt > 0) {
533 		printf(" queue %d", nc->qpoff);
534 		if (nc->qpcnt > 1)
535 			printf("-%d", nc->qpoff + nc->qpcnt - 1);
536 	}
537 	retval += printf(" at consumer %d", nc->consumer);
538 	retval += bus_print_child_domain(dev, child);
539 	retval += bus_print_child_footer(dev, child);
540 
541 	return (retval);
542 }
543 
544 static int
545 ntb_transport_child_location_str(device_t dev, device_t child, char *buf,
546     size_t buflen)
547 {
548 	struct ntb_transport_child *nc = device_get_ivars(child);
549 
550 	snprintf(buf, buflen, "consumer=%d", nc->consumer);
551 	return (0);
552 }
553 
554 int
555 ntb_transport_queue_count(device_t dev)
556 {
557 	struct ntb_transport_child *nc = device_get_ivars(dev);
558 
559 	return (nc->qpcnt);
560 }
561 
562 static void
563 ntb_transport_init_queue(struct ntb_transport_ctx *nt, unsigned int qp_num)
564 {
565 	struct ntb_transport_mw *mw;
566 	struct ntb_transport_qp *qp;
567 	vm_paddr_t mw_base;
568 	uint64_t qp_offset;
569 	size_t tx_size;
570 	unsigned num_qps_mw, mw_num, mw_count;
571 
572 	mw_count = nt->mw_count;
573 	mw_num = QP_TO_MW(nt, qp_num);
574 	mw = &nt->mw_vec[mw_num];
575 
576 	qp = &nt->qp_vec[qp_num];
577 	qp->qp_num = qp_num;
578 	qp->transport = nt;
579 	qp->dev = nt->dev;
580 	qp->client_ready = false;
581 	qp->event_handler = NULL;
582 	ntb_qp_link_down_reset(qp);
583 
584 	if (mw_num < nt->qp_count % mw_count)
585 		num_qps_mw = nt->qp_count / mw_count + 1;
586 	else
587 		num_qps_mw = nt->qp_count / mw_count;
588 
589 	mw_base = mw->phys_addr;
590 
591 	tx_size = mw->tx_size / num_qps_mw;
592 	qp_offset = tx_size * (qp_num / mw_count);
593 
594 	qp->tx_mw = mw->vbase + qp_offset;
595 	KASSERT(qp->tx_mw != NULL, ("uh oh?"));
596 
597 	/* XXX Assumes that a vm_paddr_t is equivalent to bus_addr_t */
598 	qp->tx_mw_phys = mw_base + qp_offset;
599 	KASSERT(qp->tx_mw_phys != 0, ("uh oh?"));
600 
601 	tx_size -= sizeof(struct ntb_rx_info);
602 	qp->rx_info = (void *)(qp->tx_mw + tx_size);
603 
604 	/* Due to house-keeping, there must be at least 2 buffs */
605 	qp->tx_max_frame = qmin(transport_mtu, tx_size / 2);
606 	qp->tx_max_entry = tx_size / qp->tx_max_frame;
607 
608 	callout_init(&qp->link_work, 0);
609 	callout_init(&qp->rx_full, 1);
610 
611 	mtx_init(&qp->ntb_rx_q_lock, "ntb rx q", NULL, MTX_SPIN);
612 	mtx_init(&qp->ntb_tx_free_q_lock, "ntb tx free q", NULL, MTX_SPIN);
613 	mtx_init(&qp->tx_lock, "ntb transport tx", NULL, MTX_DEF);
614 	TASK_INIT(&qp->rxc_db_work, 0, ntb_transport_rxc_db, qp);
615 	qp->rxc_tq = taskqueue_create("ntbt_rx", M_WAITOK,
616 	    taskqueue_thread_enqueue, &qp->rxc_tq);
617 	taskqueue_start_threads(&qp->rxc_tq, 1, PI_NET, "%s rx%d",
618 	    device_get_nameunit(nt->dev), qp_num);
619 
620 	STAILQ_INIT(&qp->rx_post_q);
621 	STAILQ_INIT(&qp->rx_pend_q);
622 	STAILQ_INIT(&qp->tx_free_q);
623 }
624 
625 void
626 ntb_transport_free_queue(struct ntb_transport_qp *qp)
627 {
628 	struct ntb_transport_ctx *nt = qp->transport;
629 	struct ntb_queue_entry *entry;
630 
631 	callout_drain(&qp->link_work);
632 
633 	ntb_db_set_mask(qp->dev, 1ull << qp->qp_num);
634 	taskqueue_drain_all(qp->rxc_tq);
635 	taskqueue_free(qp->rxc_tq);
636 
637 	qp->cb_data = NULL;
638 	qp->rx_handler = NULL;
639 	qp->tx_handler = NULL;
640 	qp->event_handler = NULL;
641 
642 	while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_pend_q)))
643 		free(entry, M_NTB_T);
644 
645 	while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_post_q)))
646 		free(entry, M_NTB_T);
647 
648 	while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
649 		free(entry, M_NTB_T);
650 
651 	nt->qp_bitmap &= ~(1 << qp->qp_num);
652 }
653 
654 /**
655  * ntb_transport_create_queue - Create a new NTB transport layer queue
656  * @rx_handler: receive callback function
657  * @tx_handler: transmit callback function
658  * @event_handler: event callback function
659  *
660  * Create a new NTB transport layer queue and provide the queue with a callback
661  * routine for both transmit and receive.  The receive callback routine will be
662  * used to pass up data when the transport has received it on the queue.   The
663  * transmit callback routine will be called when the transport has completed the
664  * transmission of the data on the queue and the data is ready to be freed.
665  *
666  * RETURNS: pointer to newly created ntb_queue, NULL on error.
667  */
668 struct ntb_transport_qp *
669 ntb_transport_create_queue(device_t dev, int q,
670     const struct ntb_queue_handlers *handlers, void *data)
671 {
672 	struct ntb_transport_child *nc = device_get_ivars(dev);
673 	struct ntb_transport_ctx *nt = device_get_softc(device_get_parent(dev));
674 	struct ntb_queue_entry *entry;
675 	struct ntb_transport_qp *qp;
676 	int i;
677 
678 	if (q < 0 || q >= nc->qpcnt)
679 		return (NULL);
680 
681 	qp = &nt->qp_vec[nc->qpoff + q];
682 	nt->qp_bitmap |= (1 << qp->qp_num);
683 	qp->cb_data = data;
684 	qp->rx_handler = handlers->rx_handler;
685 	qp->tx_handler = handlers->tx_handler;
686 	qp->event_handler = handlers->event_handler;
687 
688 	for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) {
689 		entry = malloc(sizeof(*entry), M_NTB_T, M_WAITOK | M_ZERO);
690 		entry->cb_data = data;
691 		entry->buf = NULL;
692 		entry->len = transport_mtu;
693 		entry->qp = qp;
694 		ntb_list_add(&qp->ntb_rx_q_lock, entry, &qp->rx_pend_q);
695 	}
696 
697 	for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) {
698 		entry = malloc(sizeof(*entry), M_NTB_T, M_WAITOK | M_ZERO);
699 		entry->qp = qp;
700 		ntb_list_add(&qp->ntb_tx_free_q_lock, entry, &qp->tx_free_q);
701 	}
702 
703 	ntb_db_clear(dev, 1ull << qp->qp_num);
704 	return (qp);
705 }
706 
707 /**
708  * ntb_transport_link_up - Notify NTB transport of client readiness to use queue
709  * @qp: NTB transport layer queue to be enabled
710  *
711  * Notify NTB transport layer of client readiness to use queue
712  */
713 void
714 ntb_transport_link_up(struct ntb_transport_qp *qp)
715 {
716 	struct ntb_transport_ctx *nt = qp->transport;
717 
718 	qp->client_ready = true;
719 
720 	ntb_printf(2, "qp %d client ready\n", qp->qp_num);
721 
722 	if (nt->link_is_up)
723 		callout_reset(&qp->link_work, 0, ntb_qp_link_work, qp);
724 }
725 
726 
727 
728 /* Transport Tx */
729 
730 /**
731  * ntb_transport_tx_enqueue - Enqueue a new NTB queue entry
732  * @qp: NTB transport layer queue the entry is to be enqueued on
733  * @cb: per buffer pointer for callback function to use
734  * @data: pointer to data buffer that will be sent
735  * @len: length of the data buffer
736  *
737  * Enqueue a new transmit buffer onto the transport queue from which a NTB
738  * payload will be transmitted.  This assumes that a lock is being held to
739  * serialize access to the qp.
740  *
741  * RETURNS: An appropriate ERRNO error value on error, or zero for success.
742  */
743 int
744 ntb_transport_tx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
745     unsigned int len)
746 {
747 	struct ntb_queue_entry *entry;
748 	int rc;
749 
750 	if (!qp->link_is_up || len == 0) {
751 		CTR0(KTR_NTB, "TX: link not up");
752 		return (EINVAL);
753 	}
754 
755 	entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
756 	if (entry == NULL) {
757 		CTR0(KTR_NTB, "TX: could not get entry from tx_free_q");
758 		qp->tx_err_no_buf++;
759 		return (EBUSY);
760 	}
761 	CTR1(KTR_NTB, "TX: got entry %p from tx_free_q", entry);
762 
763 	entry->cb_data = cb;
764 	entry->buf = data;
765 	entry->len = len;
766 	entry->flags = 0;
767 
768 	mtx_lock(&qp->tx_lock);
769 	rc = ntb_process_tx(qp, entry);
770 	mtx_unlock(&qp->tx_lock);
771 	if (rc != 0) {
772 		ntb_list_add(&qp->ntb_tx_free_q_lock, entry, &qp->tx_free_q);
773 		CTR1(KTR_NTB,
774 		    "TX: process_tx failed. Returning entry %p to tx_free_q",
775 		    entry);
776 	}
777 	return (rc);
778 }
779 
780 static void
781 ntb_tx_copy_callback(void *data)
782 {
783 	struct ntb_queue_entry *entry = data;
784 	struct ntb_transport_qp *qp = entry->qp;
785 	struct ntb_payload_header *hdr = entry->x_hdr;
786 
787 	iowrite32(entry->flags | NTBT_DESC_DONE_FLAG, &hdr->flags);
788 	CTR1(KTR_NTB, "TX: hdr %p set DESC_DONE", hdr);
789 
790 	ntb_peer_db_set(qp->dev, 1ull << qp->qp_num);
791 
792 	/*
793 	 * The entry length can only be zero if the packet is intended to be a
794 	 * "link down" or similar.  Since no payload is being sent in these
795 	 * cases, there is nothing to add to the completion queue.
796 	 */
797 	if (entry->len > 0) {
798 		qp->tx_bytes += entry->len;
799 
800 		if (qp->tx_handler)
801 			qp->tx_handler(qp, qp->cb_data, entry->buf,
802 			    entry->len);
803 		else
804 			m_freem(entry->buf);
805 		entry->buf = NULL;
806 	}
807 
808 	CTR3(KTR_NTB,
809 	    "TX: entry %p sent. hdr->ver = %u, hdr->flags = 0x%x, Returning "
810 	    "to tx_free_q", entry, hdr->ver, hdr->flags);
811 	ntb_list_add(&qp->ntb_tx_free_q_lock, entry, &qp->tx_free_q);
812 }
813 
814 static void
815 ntb_memcpy_tx(struct ntb_queue_entry *entry, void *offset)
816 {
817 
818 	CTR2(KTR_NTB, "TX: copying %d bytes to offset %p", entry->len, offset);
819 	if (entry->buf != NULL) {
820 		m_copydata((struct mbuf *)entry->buf, 0, entry->len, offset);
821 
822 		/*
823 		 * Ensure that the data is fully copied before setting the
824 		 * flags
825 		 */
826 		wmb();
827 	}
828 
829 	ntb_tx_copy_callback(entry);
830 }
831 
832 static void
833 ntb_async_tx(struct ntb_transport_qp *qp, struct ntb_queue_entry *entry)
834 {
835 	struct ntb_payload_header *hdr;
836 	void *offset;
837 
838 	offset = qp->tx_mw + qp->tx_max_frame * qp->tx_index;
839 	hdr = (struct ntb_payload_header *)((char *)offset + qp->tx_max_frame -
840 	    sizeof(struct ntb_payload_header));
841 	entry->x_hdr = hdr;
842 
843 	iowrite32(entry->len, &hdr->len);
844 	iowrite32(qp->tx_pkts, &hdr->ver);
845 
846 	ntb_memcpy_tx(entry, offset);
847 }
848 
849 static int
850 ntb_process_tx(struct ntb_transport_qp *qp, struct ntb_queue_entry *entry)
851 {
852 
853 	CTR3(KTR_NTB,
854 	    "TX: process_tx: tx_pkts=%lu, tx_index=%u, remote entry=%u",
855 	    qp->tx_pkts, qp->tx_index, qp->remote_rx_info->entry);
856 	if (qp->tx_index == qp->remote_rx_info->entry) {
857 		CTR0(KTR_NTB, "TX: ring full");
858 		qp->tx_ring_full++;
859 		return (EAGAIN);
860 	}
861 
862 	if (entry->len > qp->tx_max_frame - sizeof(struct ntb_payload_header)) {
863 		if (qp->tx_handler != NULL)
864 			qp->tx_handler(qp, qp->cb_data, entry->buf,
865 			    EIO);
866 		else
867 			m_freem(entry->buf);
868 
869 		entry->buf = NULL;
870 		ntb_list_add(&qp->ntb_tx_free_q_lock, entry, &qp->tx_free_q);
871 		CTR1(KTR_NTB,
872 		    "TX: frame too big. returning entry %p to tx_free_q",
873 		    entry);
874 		return (0);
875 	}
876 	CTR2(KTR_NTB, "TX: copying entry %p to index %u", entry, qp->tx_index);
877 	ntb_async_tx(qp, entry);
878 
879 	qp->tx_index++;
880 	qp->tx_index %= qp->tx_max_entry;
881 
882 	qp->tx_pkts++;
883 
884 	return (0);
885 }
886 
887 /* Transport Rx */
888 static void
889 ntb_transport_rxc_db(void *arg, int pending __unused)
890 {
891 	struct ntb_transport_qp *qp = arg;
892 	uint64_t qp_mask = 1ull << qp->qp_num;
893 	int rc;
894 
895 	CTR0(KTR_NTB, "RX: transport_rx");
896 again:
897 	while ((rc = ntb_process_rxc(qp)) == 0)
898 		;
899 	CTR1(KTR_NTB, "RX: process_rxc returned %d", rc);
900 
901 	if ((ntb_db_read(qp->dev) & qp_mask) != 0) {
902 		/* If db is set, clear it and check queue once more. */
903 		ntb_db_clear(qp->dev, qp_mask);
904 		goto again;
905 	}
906 	if (qp->link_is_up)
907 		ntb_db_clear_mask(qp->dev, qp_mask);
908 }
909 
910 static int
911 ntb_process_rxc(struct ntb_transport_qp *qp)
912 {
913 	struct ntb_payload_header *hdr;
914 	struct ntb_queue_entry *entry;
915 	caddr_t offset;
916 
917 	offset = qp->rx_buff + qp->rx_max_frame * qp->rx_index;
918 	hdr = (void *)(offset + qp->rx_max_frame -
919 	    sizeof(struct ntb_payload_header));
920 
921 	CTR1(KTR_NTB, "RX: process_rxc rx_index = %u", qp->rx_index);
922 	if ((hdr->flags & NTBT_DESC_DONE_FLAG) == 0) {
923 		CTR0(KTR_NTB, "RX: hdr not done");
924 		qp->rx_ring_empty++;
925 		return (EAGAIN);
926 	}
927 
928 	if ((hdr->flags & NTBT_LINK_DOWN_FLAG) != 0) {
929 		CTR0(KTR_NTB, "RX: link down");
930 		ntb_qp_link_down(qp);
931 		hdr->flags = 0;
932 		return (EAGAIN);
933 	}
934 
935 	if (hdr->ver != (uint32_t)qp->rx_pkts) {
936 		CTR2(KTR_NTB,"RX: ver != rx_pkts (%x != %lx). "
937 		    "Returning entry to rx_pend_q", hdr->ver, qp->rx_pkts);
938 		qp->rx_err_ver++;
939 		return (EIO);
940 	}
941 
942 	entry = ntb_list_mv(&qp->ntb_rx_q_lock, &qp->rx_pend_q, &qp->rx_post_q);
943 	if (entry == NULL) {
944 		qp->rx_err_no_buf++;
945 		CTR0(KTR_NTB, "RX: No entries in rx_pend_q");
946 		return (EAGAIN);
947 	}
948 	callout_stop(&qp->rx_full);
949 	CTR1(KTR_NTB, "RX: rx entry %p from rx_pend_q", entry);
950 
951 	entry->x_hdr = hdr;
952 	entry->index = qp->rx_index;
953 
954 	if (hdr->len > entry->len) {
955 		CTR2(KTR_NTB, "RX: len too long. Wanted %ju got %ju",
956 		    (uintmax_t)hdr->len, (uintmax_t)entry->len);
957 		qp->rx_err_oflow++;
958 
959 		entry->len = -EIO;
960 		entry->flags |= NTBT_DESC_DONE_FLAG;
961 
962 		ntb_complete_rxc(qp);
963 	} else {
964 		qp->rx_bytes += hdr->len;
965 		qp->rx_pkts++;
966 
967 		CTR1(KTR_NTB, "RX: received %ld rx_pkts", qp->rx_pkts);
968 
969 		entry->len = hdr->len;
970 
971 		ntb_memcpy_rx(qp, entry, offset);
972 	}
973 
974 	qp->rx_index++;
975 	qp->rx_index %= qp->rx_max_entry;
976 	return (0);
977 }
978 
979 static void
980 ntb_memcpy_rx(struct ntb_transport_qp *qp, struct ntb_queue_entry *entry,
981     void *offset)
982 {
983 	struct ifnet *ifp = entry->cb_data;
984 	unsigned int len = entry->len;
985 
986 	CTR2(KTR_NTB, "RX: copying %d bytes from offset %p", len, offset);
987 
988 	entry->buf = (void *)m_devget(offset, len, 0, ifp, NULL);
989 	if (entry->buf == NULL)
990 		entry->len = -ENOMEM;
991 
992 	/* Ensure that the data is globally visible before clearing the flag */
993 	wmb();
994 
995 	CTR2(KTR_NTB, "RX: copied entry %p to mbuf %p.", entry, entry->buf);
996 	ntb_rx_copy_callback(qp, entry);
997 }
998 
999 static inline void
1000 ntb_rx_copy_callback(struct ntb_transport_qp *qp, void *data)
1001 {
1002 	struct ntb_queue_entry *entry;
1003 
1004 	entry = data;
1005 	entry->flags |= NTBT_DESC_DONE_FLAG;
1006 	ntb_complete_rxc(qp);
1007 }
1008 
1009 static void
1010 ntb_complete_rxc(struct ntb_transport_qp *qp)
1011 {
1012 	struct ntb_queue_entry *entry;
1013 	struct mbuf *m;
1014 	unsigned len;
1015 
1016 	CTR0(KTR_NTB, "RX: rx_completion_task");
1017 
1018 	mtx_lock_spin(&qp->ntb_rx_q_lock);
1019 
1020 	while (!STAILQ_EMPTY(&qp->rx_post_q)) {
1021 		entry = STAILQ_FIRST(&qp->rx_post_q);
1022 		if ((entry->flags & NTBT_DESC_DONE_FLAG) == 0)
1023 			break;
1024 
1025 		entry->x_hdr->flags = 0;
1026 		iowrite32(entry->index, &qp->rx_info->entry);
1027 
1028 		STAILQ_REMOVE_HEAD(&qp->rx_post_q, entry);
1029 
1030 		len = entry->len;
1031 		m = entry->buf;
1032 
1033 		/*
1034 		 * Re-initialize queue_entry for reuse; rx_handler takes
1035 		 * ownership of the mbuf.
1036 		 */
1037 		entry->buf = NULL;
1038 		entry->len = transport_mtu;
1039 		entry->cb_data = qp->cb_data;
1040 
1041 		STAILQ_INSERT_TAIL(&qp->rx_pend_q, entry, entry);
1042 
1043 		mtx_unlock_spin(&qp->ntb_rx_q_lock);
1044 
1045 		CTR2(KTR_NTB, "RX: completing entry %p, mbuf %p", entry, m);
1046 		if (qp->rx_handler != NULL && qp->client_ready)
1047 			qp->rx_handler(qp, qp->cb_data, m, len);
1048 		else
1049 			m_freem(m);
1050 
1051 		mtx_lock_spin(&qp->ntb_rx_q_lock);
1052 	}
1053 
1054 	mtx_unlock_spin(&qp->ntb_rx_q_lock);
1055 }
1056 
1057 static void
1058 ntb_transport_doorbell_callback(void *data, uint32_t vector)
1059 {
1060 	struct ntb_transport_ctx *nt = data;
1061 	struct ntb_transport_qp *qp;
1062 	uint64_t vec_mask;
1063 	unsigned qp_num;
1064 
1065 	vec_mask = ntb_db_vector_mask(nt->dev, vector);
1066 	vec_mask &= nt->qp_bitmap;
1067 	if ((vec_mask & (vec_mask - 1)) != 0)
1068 		vec_mask &= ntb_db_read(nt->dev);
1069 	if (vec_mask != 0) {
1070 		ntb_db_set_mask(nt->dev, vec_mask);
1071 		ntb_db_clear(nt->dev, vec_mask);
1072 	}
1073 	while (vec_mask != 0) {
1074 		qp_num = ffsll(vec_mask) - 1;
1075 
1076 		qp = &nt->qp_vec[qp_num];
1077 		if (qp->link_is_up)
1078 			taskqueue_enqueue(qp->rxc_tq, &qp->rxc_db_work);
1079 
1080 		vec_mask &= ~(1ull << qp_num);
1081 	}
1082 }
1083 
1084 /* Link Event handler */
1085 static void
1086 ntb_transport_event_callback(void *data)
1087 {
1088 	struct ntb_transport_ctx *nt = data;
1089 
1090 	if (ntb_link_is_up(nt->dev, &nt->link_speed, &nt->link_width)) {
1091 		ntb_printf(1, "HW link up\n");
1092 		callout_reset(&nt->link_work, 0, ntb_transport_link_work, nt);
1093 	} else {
1094 		ntb_printf(1, "HW link down\n");
1095 		taskqueue_enqueue(taskqueue_swi, &nt->link_cleanup);
1096 	}
1097 }
1098 
1099 /* Link bring up */
1100 static void
1101 ntb_transport_link_work(void *arg)
1102 {
1103 	struct ntb_transport_ctx *nt = arg;
1104 	struct ntb_transport_mw *mw;
1105 	device_t dev = nt->dev;
1106 	struct ntb_transport_qp *qp;
1107 	uint64_t val64, size;
1108 	uint32_t val;
1109 	unsigned i;
1110 	int rc;
1111 
1112 	/* send the local info, in the opposite order of the way we read it */
1113 	for (i = 0; i < nt->mw_count; i++) {
1114 		size = nt->mw_vec[i].tx_size;
1115 		ntb_peer_spad_write(dev, NTBT_MW0_SZ_HIGH + (i * 2),
1116 		    size >> 32);
1117 		ntb_peer_spad_write(dev, NTBT_MW0_SZ_LOW + (i * 2), size);
1118 	}
1119 	ntb_peer_spad_write(dev, NTBT_NUM_MWS, nt->mw_count);
1120 	ntb_peer_spad_write(dev, NTBT_NUM_QPS, nt->qp_count);
1121 	ntb_peer_spad_write(dev, NTBT_QP_LINKS, 0);
1122 	ntb_peer_spad_write(dev, NTBT_VERSION, NTB_TRANSPORT_VERSION);
1123 
1124 	/* Query the remote side for its info */
1125 	val = 0;
1126 	ntb_spad_read(dev, NTBT_VERSION, &val);
1127 	if (val != NTB_TRANSPORT_VERSION)
1128 		goto out;
1129 
1130 	ntb_spad_read(dev, NTBT_NUM_QPS, &val);
1131 	if (val != nt->qp_count)
1132 		goto out;
1133 
1134 	ntb_spad_read(dev, NTBT_NUM_MWS, &val);
1135 	if (val != nt->mw_count)
1136 		goto out;
1137 
1138 	for (i = 0; i < nt->mw_count; i++) {
1139 		ntb_spad_read(dev, NTBT_MW0_SZ_HIGH + (i * 2), &val);
1140 		val64 = (uint64_t)val << 32;
1141 
1142 		ntb_spad_read(dev, NTBT_MW0_SZ_LOW + (i * 2), &val);
1143 		val64 |= val;
1144 
1145 		mw = &nt->mw_vec[i];
1146 		mw->rx_size = val64;
1147 		val64 = roundup(val64, mw->xlat_align_size);
1148 		if (mw->buff_size != val64) {
1149 
1150 			rc = ntb_set_mw(nt, i, val64);
1151 			if (rc != 0) {
1152 				ntb_printf(0, "link up set mw%d fails, rc %d\n",
1153 				    i, rc);
1154 				goto free_mws;
1155 			}
1156 
1157 			/* Notify HW the memory location of the receive buffer */
1158 			rc = ntb_mw_set_trans(nt->dev, i, mw->dma_addr,
1159 			    mw->buff_size);
1160 			if (rc != 0) {
1161 				ntb_printf(0, "link up mw%d xlat fails, rc %d\n",
1162 				     i, rc);
1163 				goto free_mws;
1164 			}
1165 		}
1166 	}
1167 
1168 	nt->link_is_up = true;
1169 	ntb_printf(1, "transport link up\n");
1170 
1171 	for (i = 0; i < nt->qp_count; i++) {
1172 		qp = &nt->qp_vec[i];
1173 
1174 		ntb_transport_setup_qp_mw(nt, i);
1175 
1176 		if (qp->client_ready)
1177 			callout_reset(&qp->link_work, 0, ntb_qp_link_work, qp);
1178 	}
1179 
1180 	return;
1181 
1182 free_mws:
1183 	for (i = 0; i < nt->mw_count; i++)
1184 		ntb_free_mw(nt, i);
1185 out:
1186 	if (ntb_link_is_up(dev, &nt->link_speed, &nt->link_width))
1187 		callout_reset(&nt->link_work,
1188 		    NTB_LINK_DOWN_TIMEOUT * hz / 1000, ntb_transport_link_work, nt);
1189 }
1190 
1191 struct ntb_load_cb_args {
1192 	bus_addr_t addr;
1193 	int error;
1194 };
1195 
1196 static void
1197 ntb_load_cb(void *xsc, bus_dma_segment_t *segs, int nsegs, int error)
1198 {
1199 	struct ntb_load_cb_args *cba = (struct ntb_load_cb_args *)xsc;
1200 
1201 	if (!(cba->error = error))
1202 		cba->addr = segs[0].ds_addr;
1203 }
1204 
1205 static int
1206 ntb_set_mw(struct ntb_transport_ctx *nt, int num_mw, size_t size)
1207 {
1208 	struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
1209 	struct ntb_load_cb_args cba;
1210 	size_t buff_size;
1211 
1212 	if (size == 0)
1213 		return (EINVAL);
1214 
1215 	buff_size = roundup(size, mw->xlat_align_size);
1216 
1217 	/* No need to re-setup */
1218 	if (mw->buff_size == buff_size)
1219 		return (0);
1220 
1221 	if (mw->buff_size != 0)
1222 		ntb_free_mw(nt, num_mw);
1223 
1224 	/* Alloc memory for receiving data.  Must be aligned */
1225 	mw->buff_size = buff_size;
1226 
1227 	if (bus_dma_tag_create(bus_get_dma_tag(nt->dev), mw->xlat_align, 0,
1228 	    mw->addr_limit, BUS_SPACE_MAXADDR,
1229 	    NULL, NULL, mw->buff_size, 1, mw->buff_size,
1230 	    0, NULL, NULL, &mw->dma_tag)) {
1231 		ntb_printf(0, "Unable to create MW tag of size %zu\n",
1232 		    mw->buff_size);
1233 		mw->buff_size = 0;
1234 		return (ENOMEM);
1235 	}
1236 	if (bus_dmamem_alloc(mw->dma_tag, (void **)&mw->virt_addr,
1237 	    BUS_DMA_WAITOK | BUS_DMA_ZERO, &mw->dma_map)) {
1238 		bus_dma_tag_destroy(mw->dma_tag);
1239 		ntb_printf(0, "Unable to allocate MW buffer of size %zu\n",
1240 		    mw->buff_size);
1241 		mw->buff_size = 0;
1242 		return (ENOMEM);
1243 	}
1244 	if (bus_dmamap_load(mw->dma_tag, mw->dma_map, mw->virt_addr,
1245 	    mw->buff_size, ntb_load_cb, &cba, BUS_DMA_NOWAIT) || cba.error) {
1246 		bus_dmamem_free(mw->dma_tag, mw->virt_addr, mw->dma_map);
1247 		bus_dma_tag_destroy(mw->dma_tag);
1248 		ntb_printf(0, "Unable to load MW buffer of size %zu\n",
1249 		    mw->buff_size);
1250 		mw->buff_size = 0;
1251 		return (ENOMEM);
1252 	}
1253 	mw->dma_addr = cba.addr;
1254 
1255 	return (0);
1256 }
1257 
1258 static void
1259 ntb_free_mw(struct ntb_transport_ctx *nt, int num_mw)
1260 {
1261 	struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
1262 
1263 	if (mw->virt_addr == NULL)
1264 		return;
1265 
1266 	ntb_mw_clear_trans(nt->dev, num_mw);
1267 	bus_dmamap_unload(mw->dma_tag, mw->dma_map);
1268 	bus_dmamem_free(mw->dma_tag, mw->virt_addr, mw->dma_map);
1269 	bus_dma_tag_destroy(mw->dma_tag);
1270 	mw->buff_size = 0;
1271 	mw->virt_addr = NULL;
1272 }
1273 
1274 static int
1275 ntb_transport_setup_qp_mw(struct ntb_transport_ctx *nt, unsigned int qp_num)
1276 {
1277 	struct ntb_transport_qp *qp = &nt->qp_vec[qp_num];
1278 	struct ntb_transport_mw *mw;
1279 	void *offset;
1280 	ntb_q_idx_t i;
1281 	size_t rx_size;
1282 	unsigned num_qps_mw, mw_num, mw_count;
1283 
1284 	mw_count = nt->mw_count;
1285 	mw_num = QP_TO_MW(nt, qp_num);
1286 	mw = &nt->mw_vec[mw_num];
1287 
1288 	if (mw->virt_addr == NULL)
1289 		return (ENOMEM);
1290 
1291 	if (mw_num < nt->qp_count % mw_count)
1292 		num_qps_mw = nt->qp_count / mw_count + 1;
1293 	else
1294 		num_qps_mw = nt->qp_count / mw_count;
1295 
1296 	rx_size = mw->rx_size / num_qps_mw;
1297 	qp->rx_buff = mw->virt_addr + rx_size * (qp_num / mw_count);
1298 	rx_size -= sizeof(struct ntb_rx_info);
1299 
1300 	qp->remote_rx_info = (void*)(qp->rx_buff + rx_size);
1301 
1302 	/* Due to house-keeping, there must be at least 2 buffs */
1303 	qp->rx_max_frame = qmin(transport_mtu, rx_size / 2);
1304 	qp->rx_max_entry = rx_size / qp->rx_max_frame;
1305 	qp->rx_index = 0;
1306 
1307 	qp->remote_rx_info->entry = qp->rx_max_entry - 1;
1308 
1309 	/* Set up the hdr offsets with 0s */
1310 	for (i = 0; i < qp->rx_max_entry; i++) {
1311 		offset = (void *)(qp->rx_buff + qp->rx_max_frame * (i + 1) -
1312 		    sizeof(struct ntb_payload_header));
1313 		memset(offset, 0, sizeof(struct ntb_payload_header));
1314 	}
1315 
1316 	qp->rx_pkts = 0;
1317 	qp->tx_pkts = 0;
1318 	qp->tx_index = 0;
1319 
1320 	return (0);
1321 }
1322 
1323 static void
1324 ntb_qp_link_work(void *arg)
1325 {
1326 	struct ntb_transport_qp *qp = arg;
1327 	device_t dev = qp->dev;
1328 	struct ntb_transport_ctx *nt = qp->transport;
1329 	int i;
1330 	uint32_t val;
1331 
1332 	/* Report queues that are up on our side */
1333 	for (i = 0, val = 0; i < nt->qp_count; i++) {
1334 		if (nt->qp_vec[i].client_ready)
1335 			val |= (1 << i);
1336 	}
1337 	ntb_peer_spad_write(dev, NTBT_QP_LINKS, val);
1338 
1339 	/* See if the remote side is up */
1340 	ntb_spad_read(dev, NTBT_QP_LINKS, &val);
1341 	if ((val & (1ull << qp->qp_num)) != 0) {
1342 		ntb_printf(2, "qp %d link up\n", qp->qp_num);
1343 		qp->link_is_up = true;
1344 
1345 		if (qp->event_handler != NULL)
1346 			qp->event_handler(qp->cb_data, NTB_LINK_UP);
1347 
1348 		ntb_db_clear_mask(dev, 1ull << qp->qp_num);
1349 	} else if (nt->link_is_up)
1350 		callout_reset(&qp->link_work,
1351 		    NTB_LINK_DOWN_TIMEOUT * hz / 1000, ntb_qp_link_work, qp);
1352 }
1353 
1354 /* Link down event*/
1355 static void
1356 ntb_transport_link_cleanup(struct ntb_transport_ctx *nt)
1357 {
1358 	struct ntb_transport_qp *qp;
1359 	int i;
1360 
1361 	callout_drain(&nt->link_work);
1362 	nt->link_is_up = 0;
1363 
1364 	/* Pass along the info to any clients */
1365 	for (i = 0; i < nt->qp_count; i++) {
1366 		if ((nt->qp_bitmap & (1 << i)) != 0) {
1367 			qp = &nt->qp_vec[i];
1368 			ntb_qp_link_cleanup(qp);
1369 			callout_drain(&qp->link_work);
1370 		}
1371 	}
1372 
1373 	/*
1374 	 * The scratchpad registers keep the values if the remote side
1375 	 * goes down, blast them now to give them a sane value the next
1376 	 * time they are accessed
1377 	 */
1378 	ntb_spad_clear(nt->dev);
1379 }
1380 
1381 static void
1382 ntb_transport_link_cleanup_work(void *arg, int pending __unused)
1383 {
1384 
1385 	ntb_transport_link_cleanup(arg);
1386 }
1387 
1388 static void
1389 ntb_qp_link_down(struct ntb_transport_qp *qp)
1390 {
1391 
1392 	ntb_qp_link_cleanup(qp);
1393 }
1394 
1395 static void
1396 ntb_qp_link_down_reset(struct ntb_transport_qp *qp)
1397 {
1398 
1399 	qp->link_is_up = false;
1400 	ntb_db_set_mask(qp->dev, 1ull << qp->qp_num);
1401 
1402 	qp->tx_index = qp->rx_index = 0;
1403 	qp->tx_bytes = qp->rx_bytes = 0;
1404 	qp->tx_pkts = qp->rx_pkts = 0;
1405 
1406 	qp->rx_ring_empty = 0;
1407 	qp->tx_ring_full = 0;
1408 
1409 	qp->rx_err_no_buf = qp->tx_err_no_buf = 0;
1410 	qp->rx_err_oflow = qp->rx_err_ver = 0;
1411 }
1412 
1413 static void
1414 ntb_qp_link_cleanup(struct ntb_transport_qp *qp)
1415 {
1416 
1417 	callout_drain(&qp->link_work);
1418 	ntb_qp_link_down_reset(qp);
1419 
1420 	if (qp->event_handler != NULL)
1421 		qp->event_handler(qp->cb_data, NTB_LINK_DOWN);
1422 }
1423 
1424 /* Link commanded down */
1425 /**
1426  * ntb_transport_link_down - Notify NTB transport to no longer enqueue data
1427  * @qp: NTB transport layer queue to be disabled
1428  *
1429  * Notify NTB transport layer of client's desire to no longer receive data on
1430  * transport queue specified.  It is the client's responsibility to ensure all
1431  * entries on queue are purged or otherwise handled appropriately.
1432  */
1433 void
1434 ntb_transport_link_down(struct ntb_transport_qp *qp)
1435 {
1436 	struct ntb_transport_ctx *nt = qp->transport;
1437 	int i;
1438 	uint32_t val;
1439 
1440 	qp->client_ready = false;
1441 	for (i = 0, val = 0; i < nt->qp_count; i++) {
1442 		if (nt->qp_vec[i].client_ready)
1443 			val |= (1 << i);
1444 	}
1445 	ntb_peer_spad_write(qp->dev, NTBT_QP_LINKS, val);
1446 
1447 	if (qp->link_is_up)
1448 		ntb_send_link_down(qp);
1449 	else
1450 		callout_drain(&qp->link_work);
1451 }
1452 
1453 /**
1454  * ntb_transport_link_query - Query transport link state
1455  * @qp: NTB transport layer queue to be queried
1456  *
1457  * Query connectivity to the remote system of the NTB transport queue
1458  *
1459  * RETURNS: true for link up or false for link down
1460  */
1461 bool
1462 ntb_transport_link_query(struct ntb_transport_qp *qp)
1463 {
1464 
1465 	return (qp->link_is_up);
1466 }
1467 
1468 /**
1469  * ntb_transport_link_speed - Query transport link speed
1470  * @qp: NTB transport layer queue to be queried
1471  *
1472  * Query connection speed to the remote system of the NTB transport queue
1473  *
1474  * RETURNS: link speed in bits per second
1475  */
1476 uint64_t
1477 ntb_transport_link_speed(struct ntb_transport_qp *qp)
1478 {
1479 	struct ntb_transport_ctx *nt = qp->transport;
1480 	uint64_t rate;
1481 
1482 	if (!nt->link_is_up)
1483 		return (0);
1484 	switch (nt->link_speed) {
1485 	case NTB_SPEED_GEN1:
1486 		rate = 2500000000 * 8 / 10;
1487 		break;
1488 	case NTB_SPEED_GEN2:
1489 		rate = 5000000000 * 8 / 10;
1490 		break;
1491 	case NTB_SPEED_GEN3:
1492 		rate = 8000000000 * 128 / 130;
1493 		break;
1494 	case NTB_SPEED_GEN4:
1495 		rate = 16000000000 * 128 / 130;
1496 		break;
1497 	default:
1498 		return (0);
1499 	}
1500 	if (nt->link_width <= 0)
1501 		return (0);
1502 	return (rate * nt->link_width);
1503 }
1504 
1505 static void
1506 ntb_send_link_down(struct ntb_transport_qp *qp)
1507 {
1508 	struct ntb_queue_entry *entry;
1509 	int i, rc;
1510 
1511 	if (!qp->link_is_up)
1512 		return;
1513 
1514 	for (i = 0; i < NTB_LINK_DOWN_TIMEOUT; i++) {
1515 		entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
1516 		if (entry != NULL)
1517 			break;
1518 		pause("NTB Wait for link down", hz / 10);
1519 	}
1520 
1521 	if (entry == NULL)
1522 		return;
1523 
1524 	entry->cb_data = NULL;
1525 	entry->buf = NULL;
1526 	entry->len = 0;
1527 	entry->flags = NTBT_LINK_DOWN_FLAG;
1528 
1529 	mtx_lock(&qp->tx_lock);
1530 	rc = ntb_process_tx(qp, entry);
1531 	mtx_unlock(&qp->tx_lock);
1532 	if (rc != 0)
1533 		printf("ntb: Failed to send link down\n");
1534 
1535 	ntb_qp_link_down_reset(qp);
1536 }
1537 
1538 
1539 /* List Management */
1540 
1541 static void
1542 ntb_list_add(struct mtx *lock, struct ntb_queue_entry *entry,
1543     struct ntb_queue_list *list)
1544 {
1545 
1546 	mtx_lock_spin(lock);
1547 	STAILQ_INSERT_TAIL(list, entry, entry);
1548 	mtx_unlock_spin(lock);
1549 }
1550 
1551 static struct ntb_queue_entry *
1552 ntb_list_rm(struct mtx *lock, struct ntb_queue_list *list)
1553 {
1554 	struct ntb_queue_entry *entry;
1555 
1556 	mtx_lock_spin(lock);
1557 	if (STAILQ_EMPTY(list)) {
1558 		entry = NULL;
1559 		goto out;
1560 	}
1561 	entry = STAILQ_FIRST(list);
1562 	STAILQ_REMOVE_HEAD(list, entry);
1563 out:
1564 	mtx_unlock_spin(lock);
1565 
1566 	return (entry);
1567 }
1568 
1569 static struct ntb_queue_entry *
1570 ntb_list_mv(struct mtx *lock, struct ntb_queue_list *from,
1571     struct ntb_queue_list *to)
1572 {
1573 	struct ntb_queue_entry *entry;
1574 
1575 	mtx_lock_spin(lock);
1576 	if (STAILQ_EMPTY(from)) {
1577 		entry = NULL;
1578 		goto out;
1579 	}
1580 	entry = STAILQ_FIRST(from);
1581 	STAILQ_REMOVE_HEAD(from, entry);
1582 	STAILQ_INSERT_TAIL(to, entry, entry);
1583 
1584 out:
1585 	mtx_unlock_spin(lock);
1586 	return (entry);
1587 }
1588 
1589 /**
1590  * ntb_transport_qp_num - Query the qp number
1591  * @qp: NTB transport layer queue to be queried
1592  *
1593  * Query qp number of the NTB transport queue
1594  *
1595  * RETURNS: a zero based number specifying the qp number
1596  */
1597 unsigned char ntb_transport_qp_num(struct ntb_transport_qp *qp)
1598 {
1599 
1600 	return (qp->qp_num);
1601 }
1602 
1603 /**
1604  * ntb_transport_max_size - Query the max payload size of a qp
1605  * @qp: NTB transport layer queue to be queried
1606  *
1607  * Query the maximum payload size permissible on the given qp
1608  *
1609  * RETURNS: the max payload size of a qp
1610  */
1611 unsigned int
1612 ntb_transport_max_size(struct ntb_transport_qp *qp)
1613 {
1614 
1615 	return (qp->tx_max_frame - sizeof(struct ntb_payload_header));
1616 }
1617 
1618 unsigned int
1619 ntb_transport_tx_free_entry(struct ntb_transport_qp *qp)
1620 {
1621 	unsigned int head = qp->tx_index;
1622 	unsigned int tail = qp->remote_rx_info->entry;
1623 
1624 	return (tail >= head ? tail - head : qp->tx_max_entry + tail - head);
1625 }
1626 
1627 static device_method_t ntb_transport_methods[] = {
1628 	/* Device interface */
1629 	DEVMETHOD(device_probe,     ntb_transport_probe),
1630 	DEVMETHOD(device_attach,    ntb_transport_attach),
1631 	DEVMETHOD(device_detach,    ntb_transport_detach),
1632 	/* Bus interface */
1633 	DEVMETHOD(bus_child_location_str, ntb_transport_child_location_str),
1634 	DEVMETHOD(bus_print_child,  ntb_transport_print_child),
1635 	DEVMETHOD_END
1636 };
1637 
1638 devclass_t ntb_transport_devclass;
1639 static DEFINE_CLASS_0(ntb_transport, ntb_transport_driver,
1640     ntb_transport_methods, sizeof(struct ntb_transport_ctx));
1641 DRIVER_MODULE(ntb_transport, ntb_hw, ntb_transport_driver,
1642     ntb_transport_devclass, NULL, NULL);
1643 MODULE_DEPEND(ntb_transport, ntb, 1, 1, 1);
1644 MODULE_VERSION(ntb_transport, 1);
1645