xref: /openbmc/linux/drivers/ntb/ntb_transport.c (revision 3213486f)
1 /*
2  * This file is provided under a dual BSD/GPLv2 license.  When using or
3  *   redistributing this file, you may do so under either license.
4  *
5  *   GPL LICENSE SUMMARY
6  *
7  *   Copyright(c) 2012 Intel Corporation. All rights reserved.
8  *   Copyright (C) 2015 EMC Corporation. All Rights Reserved.
9  *
10  *   This program is free software; you can redistribute it and/or modify
11  *   it under the terms of version 2 of the GNU General Public License as
12  *   published by the Free Software Foundation.
13  *
14  *   BSD LICENSE
15  *
16  *   Copyright(c) 2012 Intel Corporation. All rights reserved.
17  *   Copyright (C) 2015 EMC Corporation. All Rights Reserved.
18  *
19  *   Redistribution and use in source and binary forms, with or without
20  *   modification, are permitted provided that the following conditions
21  *   are met:
22  *
23  *     * Redistributions of source code must retain the above copyright
24  *       notice, this list of conditions and the following disclaimer.
25  *     * Redistributions in binary form must reproduce the above copy
26  *       notice, this list of conditions and the following disclaimer in
27  *       the documentation and/or other materials provided with the
28  *       distribution.
29  *     * Neither the name of Intel Corporation nor the names of its
30  *       contributors may be used to endorse or promote products derived
31  *       from this software without specific prior written permission.
32  *
33  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
34  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
35  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
36  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
37  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
38  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
39  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
40  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
41  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
42  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
43  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
44  *
45  * PCIe NTB Transport Linux driver
46  *
47  * Contact Information:
48  * Jon Mason <jon.mason@intel.com>
49  */
50 #include <linux/debugfs.h>
51 #include <linux/delay.h>
52 #include <linux/dmaengine.h>
53 #include <linux/dma-mapping.h>
54 #include <linux/errno.h>
55 #include <linux/export.h>
56 #include <linux/interrupt.h>
57 #include <linux/module.h>
58 #include <linux/pci.h>
59 #include <linux/slab.h>
60 #include <linux/types.h>
61 #include <linux/uaccess.h>
62 #include "linux/ntb.h"
63 #include "linux/ntb_transport.h"
64 
65 #define NTB_TRANSPORT_VERSION	4
66 #define NTB_TRANSPORT_VER	"4"
67 #define NTB_TRANSPORT_NAME	"ntb_transport"
68 #define NTB_TRANSPORT_DESC	"Software Queue-Pair Transport over NTB"
69 #define NTB_TRANSPORT_MIN_SPADS (MW0_SZ_HIGH + 2)
70 
71 MODULE_DESCRIPTION(NTB_TRANSPORT_DESC);
72 MODULE_VERSION(NTB_TRANSPORT_VER);
73 MODULE_LICENSE("Dual BSD/GPL");
74 MODULE_AUTHOR("Intel Corporation");
75 
76 static unsigned long max_mw_size;
77 module_param(max_mw_size, ulong, 0644);
78 MODULE_PARM_DESC(max_mw_size, "Limit size of large memory windows");
79 
80 static unsigned int transport_mtu = 0x10000;
81 module_param(transport_mtu, uint, 0644);
82 MODULE_PARM_DESC(transport_mtu, "Maximum size of NTB transport packets");
83 
84 static unsigned char max_num_clients;
85 module_param(max_num_clients, byte, 0644);
86 MODULE_PARM_DESC(max_num_clients, "Maximum number of NTB transport clients");
87 
88 static unsigned int copy_bytes = 1024;
89 module_param(copy_bytes, uint, 0644);
90 MODULE_PARM_DESC(copy_bytes, "Threshold under which NTB will use the CPU to copy instead of DMA");
91 
92 static bool use_dma;
93 module_param(use_dma, bool, 0644);
94 MODULE_PARM_DESC(use_dma, "Use DMA engine to perform large data copy");
95 
96 static struct dentry *nt_debugfs_dir;
97 
98 /* Only two-ports NTB devices are supported */
99 #define PIDX		NTB_DEF_PEER_IDX
100 
101 struct ntb_queue_entry {
102 	/* ntb_queue list reference */
103 	struct list_head entry;
104 	/* pointers to data to be transferred */
105 	void *cb_data;
106 	void *buf;
107 	unsigned int len;
108 	unsigned int flags;
109 	int retries;
110 	int errors;
111 	unsigned int tx_index;
112 	unsigned int rx_index;
113 
114 	struct ntb_transport_qp *qp;
115 	union {
116 		struct ntb_payload_header __iomem *tx_hdr;
117 		struct ntb_payload_header *rx_hdr;
118 	};
119 };
120 
121 struct ntb_rx_info {
122 	unsigned int entry;
123 };
124 
125 struct ntb_transport_qp {
126 	struct ntb_transport_ctx *transport;
127 	struct ntb_dev *ndev;
128 	void *cb_data;
129 	struct dma_chan *tx_dma_chan;
130 	struct dma_chan *rx_dma_chan;
131 
132 	bool client_ready;
133 	bool link_is_up;
134 	bool active;
135 
136 	u8 qp_num;	/* Only 64 QP's are allowed.  0-63 */
137 	u64 qp_bit;
138 
139 	struct ntb_rx_info __iomem *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 list_head tx_free_q;
145 	spinlock_t ntb_tx_free_q_lock;
146 	void __iomem *tx_mw;
147 	phys_addr_t tx_mw_phys;
148 	size_t tx_mw_size;
149 	dma_addr_t tx_mw_dma_addr;
150 	unsigned int tx_index;
151 	unsigned int tx_max_entry;
152 	unsigned int tx_max_frame;
153 
154 	void (*rx_handler)(struct ntb_transport_qp *qp, void *qp_data,
155 			   void *data, int len);
156 	struct list_head rx_post_q;
157 	struct list_head rx_pend_q;
158 	struct list_head rx_free_q;
159 	/* ntb_rx_q_lock: synchronize access to rx_XXXX_q */
160 	spinlock_t ntb_rx_q_lock;
161 	void *rx_buff;
162 	unsigned int rx_index;
163 	unsigned int rx_max_entry;
164 	unsigned int rx_max_frame;
165 	unsigned int rx_alloc_entry;
166 	dma_cookie_t last_cookie;
167 	struct tasklet_struct rxc_db_work;
168 
169 	void (*event_handler)(void *data, int status);
170 	struct delayed_work link_work;
171 	struct work_struct link_cleanup;
172 
173 	struct dentry *debugfs_dir;
174 	struct dentry *debugfs_stats;
175 
176 	/* Stats */
177 	u64 rx_bytes;
178 	u64 rx_pkts;
179 	u64 rx_ring_empty;
180 	u64 rx_err_no_buf;
181 	u64 rx_err_oflow;
182 	u64 rx_err_ver;
183 	u64 rx_memcpy;
184 	u64 rx_async;
185 	u64 tx_bytes;
186 	u64 tx_pkts;
187 	u64 tx_ring_full;
188 	u64 tx_err_no_buf;
189 	u64 tx_memcpy;
190 	u64 tx_async;
191 };
192 
193 struct ntb_transport_mw {
194 	phys_addr_t phys_addr;
195 	resource_size_t phys_size;
196 	void __iomem *vbase;
197 	size_t xlat_size;
198 	size_t buff_size;
199 	size_t alloc_size;
200 	void *alloc_addr;
201 	void *virt_addr;
202 	dma_addr_t dma_addr;
203 };
204 
205 struct ntb_transport_client_dev {
206 	struct list_head entry;
207 	struct ntb_transport_ctx *nt;
208 	struct device dev;
209 };
210 
211 struct ntb_transport_ctx {
212 	struct list_head entry;
213 	struct list_head client_devs;
214 
215 	struct ntb_dev *ndev;
216 
217 	struct ntb_transport_mw *mw_vec;
218 	struct ntb_transport_qp *qp_vec;
219 	unsigned int mw_count;
220 	unsigned int qp_count;
221 	u64 qp_bitmap;
222 	u64 qp_bitmap_free;
223 
224 	bool link_is_up;
225 	struct delayed_work link_work;
226 	struct work_struct link_cleanup;
227 
228 	struct dentry *debugfs_node_dir;
229 };
230 
231 enum {
232 	DESC_DONE_FLAG = BIT(0),
233 	LINK_DOWN_FLAG = BIT(1),
234 };
235 
236 struct ntb_payload_header {
237 	unsigned int ver;
238 	unsigned int len;
239 	unsigned int flags;
240 };
241 
242 enum {
243 	VERSION = 0,
244 	QP_LINKS,
245 	NUM_QPS,
246 	NUM_MWS,
247 	MW0_SZ_HIGH,
248 	MW0_SZ_LOW,
249 };
250 
251 #define dev_client_dev(__dev) \
252 	container_of((__dev), struct ntb_transport_client_dev, dev)
253 
254 #define drv_client(__drv) \
255 	container_of((__drv), struct ntb_transport_client, driver)
256 
257 #define QP_TO_MW(nt, qp)	((qp) % nt->mw_count)
258 #define NTB_QP_DEF_NUM_ENTRIES	100
259 #define NTB_LINK_DOWN_TIMEOUT	10
260 
261 static void ntb_transport_rxc_db(unsigned long data);
262 static const struct ntb_ctx_ops ntb_transport_ops;
263 static struct ntb_client ntb_transport_client;
264 static int ntb_async_tx_submit(struct ntb_transport_qp *qp,
265 			       struct ntb_queue_entry *entry);
266 static void ntb_memcpy_tx(struct ntb_queue_entry *entry, void __iomem *offset);
267 static int ntb_async_rx_submit(struct ntb_queue_entry *entry, void *offset);
268 static void ntb_memcpy_rx(struct ntb_queue_entry *entry, void *offset);
269 
270 
271 static int ntb_transport_bus_match(struct device *dev,
272 				   struct device_driver *drv)
273 {
274 	return !strncmp(dev_name(dev), drv->name, strlen(drv->name));
275 }
276 
277 static int ntb_transport_bus_probe(struct device *dev)
278 {
279 	const struct ntb_transport_client *client;
280 	int rc = -EINVAL;
281 
282 	get_device(dev);
283 
284 	client = drv_client(dev->driver);
285 	rc = client->probe(dev);
286 	if (rc)
287 		put_device(dev);
288 
289 	return rc;
290 }
291 
292 static int ntb_transport_bus_remove(struct device *dev)
293 {
294 	const struct ntb_transport_client *client;
295 
296 	client = drv_client(dev->driver);
297 	client->remove(dev);
298 
299 	put_device(dev);
300 
301 	return 0;
302 }
303 
304 static struct bus_type ntb_transport_bus = {
305 	.name = "ntb_transport",
306 	.match = ntb_transport_bus_match,
307 	.probe = ntb_transport_bus_probe,
308 	.remove = ntb_transport_bus_remove,
309 };
310 
311 static LIST_HEAD(ntb_transport_list);
312 
313 static int ntb_bus_init(struct ntb_transport_ctx *nt)
314 {
315 	list_add_tail(&nt->entry, &ntb_transport_list);
316 	return 0;
317 }
318 
319 static void ntb_bus_remove(struct ntb_transport_ctx *nt)
320 {
321 	struct ntb_transport_client_dev *client_dev, *cd;
322 
323 	list_for_each_entry_safe(client_dev, cd, &nt->client_devs, entry) {
324 		dev_err(client_dev->dev.parent, "%s still attached to bus, removing\n",
325 			dev_name(&client_dev->dev));
326 		list_del(&client_dev->entry);
327 		device_unregister(&client_dev->dev);
328 	}
329 
330 	list_del(&nt->entry);
331 }
332 
333 static void ntb_transport_client_release(struct device *dev)
334 {
335 	struct ntb_transport_client_dev *client_dev;
336 
337 	client_dev = dev_client_dev(dev);
338 	kfree(client_dev);
339 }
340 
341 /**
342  * ntb_transport_unregister_client_dev - Unregister NTB client device
343  * @device_name: Name of NTB client device
344  *
345  * Unregister an NTB client device with the NTB transport layer
346  */
347 void ntb_transport_unregister_client_dev(char *device_name)
348 {
349 	struct ntb_transport_client_dev *client, *cd;
350 	struct ntb_transport_ctx *nt;
351 
352 	list_for_each_entry(nt, &ntb_transport_list, entry)
353 		list_for_each_entry_safe(client, cd, &nt->client_devs, entry)
354 			if (!strncmp(dev_name(&client->dev), device_name,
355 				     strlen(device_name))) {
356 				list_del(&client->entry);
357 				device_unregister(&client->dev);
358 			}
359 }
360 EXPORT_SYMBOL_GPL(ntb_transport_unregister_client_dev);
361 
362 /**
363  * ntb_transport_register_client_dev - Register NTB client device
364  * @device_name: Name of NTB client device
365  *
366  * Register an NTB client device with the NTB transport layer
367  */
368 int ntb_transport_register_client_dev(char *device_name)
369 {
370 	struct ntb_transport_client_dev *client_dev;
371 	struct ntb_transport_ctx *nt;
372 	int node;
373 	int rc, i = 0;
374 
375 	if (list_empty(&ntb_transport_list))
376 		return -ENODEV;
377 
378 	list_for_each_entry(nt, &ntb_transport_list, entry) {
379 		struct device *dev;
380 
381 		node = dev_to_node(&nt->ndev->dev);
382 
383 		client_dev = kzalloc_node(sizeof(*client_dev),
384 					  GFP_KERNEL, node);
385 		if (!client_dev) {
386 			rc = -ENOMEM;
387 			goto err;
388 		}
389 
390 		dev = &client_dev->dev;
391 
392 		/* setup and register client devices */
393 		dev_set_name(dev, "%s%d", device_name, i);
394 		dev->bus = &ntb_transport_bus;
395 		dev->release = ntb_transport_client_release;
396 		dev->parent = &nt->ndev->dev;
397 
398 		rc = device_register(dev);
399 		if (rc) {
400 			kfree(client_dev);
401 			goto err;
402 		}
403 
404 		list_add_tail(&client_dev->entry, &nt->client_devs);
405 		i++;
406 	}
407 
408 	return 0;
409 
410 err:
411 	ntb_transport_unregister_client_dev(device_name);
412 
413 	return rc;
414 }
415 EXPORT_SYMBOL_GPL(ntb_transport_register_client_dev);
416 
417 /**
418  * ntb_transport_register_client - Register NTB client driver
419  * @drv: NTB client driver to be registered
420  *
421  * Register an NTB client driver with the NTB transport layer
422  *
423  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
424  */
425 int ntb_transport_register_client(struct ntb_transport_client *drv)
426 {
427 	drv->driver.bus = &ntb_transport_bus;
428 
429 	if (list_empty(&ntb_transport_list))
430 		return -ENODEV;
431 
432 	return driver_register(&drv->driver);
433 }
434 EXPORT_SYMBOL_GPL(ntb_transport_register_client);
435 
436 /**
437  * ntb_transport_unregister_client - Unregister NTB client driver
438  * @drv: NTB client driver to be unregistered
439  *
440  * Unregister an NTB client driver with the NTB transport layer
441  *
442  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
443  */
444 void ntb_transport_unregister_client(struct ntb_transport_client *drv)
445 {
446 	driver_unregister(&drv->driver);
447 }
448 EXPORT_SYMBOL_GPL(ntb_transport_unregister_client);
449 
450 static ssize_t debugfs_read(struct file *filp, char __user *ubuf, size_t count,
451 			    loff_t *offp)
452 {
453 	struct ntb_transport_qp *qp;
454 	char *buf;
455 	ssize_t ret, out_offset, out_count;
456 
457 	qp = filp->private_data;
458 
459 	if (!qp || !qp->link_is_up)
460 		return 0;
461 
462 	out_count = 1000;
463 
464 	buf = kmalloc(out_count, GFP_KERNEL);
465 	if (!buf)
466 		return -ENOMEM;
467 
468 	out_offset = 0;
469 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
470 			       "\nNTB QP stats:\n\n");
471 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
472 			       "rx_bytes - \t%llu\n", qp->rx_bytes);
473 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
474 			       "rx_pkts - \t%llu\n", qp->rx_pkts);
475 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
476 			       "rx_memcpy - \t%llu\n", qp->rx_memcpy);
477 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
478 			       "rx_async - \t%llu\n", qp->rx_async);
479 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
480 			       "rx_ring_empty - %llu\n", qp->rx_ring_empty);
481 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
482 			       "rx_err_no_buf - %llu\n", qp->rx_err_no_buf);
483 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
484 			       "rx_err_oflow - \t%llu\n", qp->rx_err_oflow);
485 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
486 			       "rx_err_ver - \t%llu\n", qp->rx_err_ver);
487 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
488 			       "rx_buff - \t0x%p\n", qp->rx_buff);
489 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
490 			       "rx_index - \t%u\n", qp->rx_index);
491 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
492 			       "rx_max_entry - \t%u\n", qp->rx_max_entry);
493 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
494 			       "rx_alloc_entry - \t%u\n\n", qp->rx_alloc_entry);
495 
496 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
497 			       "tx_bytes - \t%llu\n", qp->tx_bytes);
498 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
499 			       "tx_pkts - \t%llu\n", qp->tx_pkts);
500 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
501 			       "tx_memcpy - \t%llu\n", qp->tx_memcpy);
502 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
503 			       "tx_async - \t%llu\n", qp->tx_async);
504 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
505 			       "tx_ring_full - \t%llu\n", qp->tx_ring_full);
506 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
507 			       "tx_err_no_buf - %llu\n", qp->tx_err_no_buf);
508 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
509 			       "tx_mw - \t0x%p\n", qp->tx_mw);
510 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
511 			       "tx_index (H) - \t%u\n", qp->tx_index);
512 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
513 			       "RRI (T) - \t%u\n",
514 			       qp->remote_rx_info->entry);
515 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
516 			       "tx_max_entry - \t%u\n", qp->tx_max_entry);
517 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
518 			       "free tx - \t%u\n",
519 			       ntb_transport_tx_free_entry(qp));
520 
521 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
522 			       "\n");
523 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
524 			       "Using TX DMA - \t%s\n",
525 			       qp->tx_dma_chan ? "Yes" : "No");
526 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
527 			       "Using RX DMA - \t%s\n",
528 			       qp->rx_dma_chan ? "Yes" : "No");
529 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
530 			       "QP Link - \t%s\n",
531 			       qp->link_is_up ? "Up" : "Down");
532 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
533 			       "\n");
534 
535 	if (out_offset > out_count)
536 		out_offset = out_count;
537 
538 	ret = simple_read_from_buffer(ubuf, count, offp, buf, out_offset);
539 	kfree(buf);
540 	return ret;
541 }
542 
543 static const struct file_operations ntb_qp_debugfs_stats = {
544 	.owner = THIS_MODULE,
545 	.open = simple_open,
546 	.read = debugfs_read,
547 };
548 
549 static void ntb_list_add(spinlock_t *lock, struct list_head *entry,
550 			 struct list_head *list)
551 {
552 	unsigned long flags;
553 
554 	spin_lock_irqsave(lock, flags);
555 	list_add_tail(entry, list);
556 	spin_unlock_irqrestore(lock, flags);
557 }
558 
559 static struct ntb_queue_entry *ntb_list_rm(spinlock_t *lock,
560 					   struct list_head *list)
561 {
562 	struct ntb_queue_entry *entry;
563 	unsigned long flags;
564 
565 	spin_lock_irqsave(lock, flags);
566 	if (list_empty(list)) {
567 		entry = NULL;
568 		goto out;
569 	}
570 	entry = list_first_entry(list, struct ntb_queue_entry, entry);
571 	list_del(&entry->entry);
572 
573 out:
574 	spin_unlock_irqrestore(lock, flags);
575 
576 	return entry;
577 }
578 
579 static struct ntb_queue_entry *ntb_list_mv(spinlock_t *lock,
580 					   struct list_head *list,
581 					   struct list_head *to_list)
582 {
583 	struct ntb_queue_entry *entry;
584 	unsigned long flags;
585 
586 	spin_lock_irqsave(lock, flags);
587 
588 	if (list_empty(list)) {
589 		entry = NULL;
590 	} else {
591 		entry = list_first_entry(list, struct ntb_queue_entry, entry);
592 		list_move_tail(&entry->entry, to_list);
593 	}
594 
595 	spin_unlock_irqrestore(lock, flags);
596 
597 	return entry;
598 }
599 
600 static int ntb_transport_setup_qp_mw(struct ntb_transport_ctx *nt,
601 				     unsigned int qp_num)
602 {
603 	struct ntb_transport_qp *qp = &nt->qp_vec[qp_num];
604 	struct ntb_transport_mw *mw;
605 	struct ntb_dev *ndev = nt->ndev;
606 	struct ntb_queue_entry *entry;
607 	unsigned int rx_size, num_qps_mw;
608 	unsigned int mw_num, mw_count, qp_count;
609 	unsigned int i;
610 	int node;
611 
612 	mw_count = nt->mw_count;
613 	qp_count = nt->qp_count;
614 
615 	mw_num = QP_TO_MW(nt, qp_num);
616 	mw = &nt->mw_vec[mw_num];
617 
618 	if (!mw->virt_addr)
619 		return -ENOMEM;
620 
621 	if (mw_num < qp_count % mw_count)
622 		num_qps_mw = qp_count / mw_count + 1;
623 	else
624 		num_qps_mw = qp_count / mw_count;
625 
626 	rx_size = (unsigned int)mw->xlat_size / num_qps_mw;
627 	qp->rx_buff = mw->virt_addr + rx_size * (qp_num / mw_count);
628 	rx_size -= sizeof(struct ntb_rx_info);
629 
630 	qp->remote_rx_info = qp->rx_buff + rx_size;
631 
632 	/* Due to housekeeping, there must be atleast 2 buffs */
633 	qp->rx_max_frame = min(transport_mtu, rx_size / 2);
634 	qp->rx_max_entry = rx_size / qp->rx_max_frame;
635 	qp->rx_index = 0;
636 
637 	/*
638 	 * Checking to see if we have more entries than the default.
639 	 * We should add additional entries if that is the case so we
640 	 * can be in sync with the transport frames.
641 	 */
642 	node = dev_to_node(&ndev->dev);
643 	for (i = qp->rx_alloc_entry; i < qp->rx_max_entry; i++) {
644 		entry = kzalloc_node(sizeof(*entry), GFP_KERNEL, node);
645 		if (!entry)
646 			return -ENOMEM;
647 
648 		entry->qp = qp;
649 		ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry,
650 			     &qp->rx_free_q);
651 		qp->rx_alloc_entry++;
652 	}
653 
654 	qp->remote_rx_info->entry = qp->rx_max_entry - 1;
655 
656 	/* setup the hdr offsets with 0's */
657 	for (i = 0; i < qp->rx_max_entry; i++) {
658 		void *offset = (qp->rx_buff + qp->rx_max_frame * (i + 1) -
659 				sizeof(struct ntb_payload_header));
660 		memset(offset, 0, sizeof(struct ntb_payload_header));
661 	}
662 
663 	qp->rx_pkts = 0;
664 	qp->tx_pkts = 0;
665 	qp->tx_index = 0;
666 
667 	return 0;
668 }
669 
670 static void ntb_free_mw(struct ntb_transport_ctx *nt, int num_mw)
671 {
672 	struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
673 	struct pci_dev *pdev = nt->ndev->pdev;
674 
675 	if (!mw->virt_addr)
676 		return;
677 
678 	ntb_mw_clear_trans(nt->ndev, PIDX, num_mw);
679 	dma_free_coherent(&pdev->dev, mw->alloc_size,
680 			  mw->alloc_addr, mw->dma_addr);
681 	mw->xlat_size = 0;
682 	mw->buff_size = 0;
683 	mw->alloc_size = 0;
684 	mw->alloc_addr = NULL;
685 	mw->virt_addr = NULL;
686 }
687 
688 static int ntb_alloc_mw_buffer(struct ntb_transport_mw *mw,
689 			       struct device *dma_dev, size_t align)
690 {
691 	dma_addr_t dma_addr;
692 	void *alloc_addr, *virt_addr;
693 	int rc;
694 
695 	alloc_addr = dma_alloc_coherent(dma_dev, mw->alloc_size,
696 					&dma_addr, GFP_KERNEL);
697 	if (!alloc_addr) {
698 		dev_err(dma_dev, "Unable to alloc MW buff of size %zu\n",
699 			mw->alloc_size);
700 		return -ENOMEM;
701 	}
702 	virt_addr = alloc_addr;
703 
704 	/*
705 	 * we must ensure that the memory address allocated is BAR size
706 	 * aligned in order for the XLAT register to take the value. This
707 	 * is a requirement of the hardware. It is recommended to setup CMA
708 	 * for BAR sizes equal or greater than 4MB.
709 	 */
710 	if (!IS_ALIGNED(dma_addr, align)) {
711 		if (mw->alloc_size > mw->buff_size) {
712 			virt_addr = PTR_ALIGN(alloc_addr, align);
713 			dma_addr = ALIGN(dma_addr, align);
714 		} else {
715 			rc = -ENOMEM;
716 			goto err;
717 		}
718 	}
719 
720 	mw->alloc_addr = alloc_addr;
721 	mw->virt_addr = virt_addr;
722 	mw->dma_addr = dma_addr;
723 
724 	return 0;
725 
726 err:
727 	dma_free_coherent(dma_dev, mw->alloc_size, alloc_addr, dma_addr);
728 
729 	return rc;
730 }
731 
732 static int ntb_set_mw(struct ntb_transport_ctx *nt, int num_mw,
733 		      resource_size_t size)
734 {
735 	struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
736 	struct pci_dev *pdev = nt->ndev->pdev;
737 	size_t xlat_size, buff_size;
738 	resource_size_t xlat_align;
739 	resource_size_t xlat_align_size;
740 	int rc;
741 
742 	if (!size)
743 		return -EINVAL;
744 
745 	rc = ntb_mw_get_align(nt->ndev, PIDX, num_mw, &xlat_align,
746 			      &xlat_align_size, NULL);
747 	if (rc)
748 		return rc;
749 
750 	xlat_size = round_up(size, xlat_align_size);
751 	buff_size = round_up(size, xlat_align);
752 
753 	/* No need to re-setup */
754 	if (mw->xlat_size == xlat_size)
755 		return 0;
756 
757 	if (mw->buff_size)
758 		ntb_free_mw(nt, num_mw);
759 
760 	/* Alloc memory for receiving data.  Must be aligned */
761 	mw->xlat_size = xlat_size;
762 	mw->buff_size = buff_size;
763 	mw->alloc_size = buff_size;
764 
765 	rc = ntb_alloc_mw_buffer(mw, &pdev->dev, xlat_align);
766 	if (rc) {
767 		mw->alloc_size *= 2;
768 		rc = ntb_alloc_mw_buffer(mw, &pdev->dev, xlat_align);
769 		if (rc) {
770 			dev_err(&pdev->dev,
771 				"Unable to alloc aligned MW buff\n");
772 			mw->xlat_size = 0;
773 			mw->buff_size = 0;
774 			mw->alloc_size = 0;
775 			return rc;
776 		}
777 	}
778 
779 	/* Notify HW the memory location of the receive buffer */
780 	rc = ntb_mw_set_trans(nt->ndev, PIDX, num_mw, mw->dma_addr,
781 			      mw->xlat_size);
782 	if (rc) {
783 		dev_err(&pdev->dev, "Unable to set mw%d translation", num_mw);
784 		ntb_free_mw(nt, num_mw);
785 		return -EIO;
786 	}
787 
788 	return 0;
789 }
790 
791 static void ntb_qp_link_down_reset(struct ntb_transport_qp *qp)
792 {
793 	qp->link_is_up = false;
794 	qp->active = false;
795 
796 	qp->tx_index = 0;
797 	qp->rx_index = 0;
798 	qp->rx_bytes = 0;
799 	qp->rx_pkts = 0;
800 	qp->rx_ring_empty = 0;
801 	qp->rx_err_no_buf = 0;
802 	qp->rx_err_oflow = 0;
803 	qp->rx_err_ver = 0;
804 	qp->rx_memcpy = 0;
805 	qp->rx_async = 0;
806 	qp->tx_bytes = 0;
807 	qp->tx_pkts = 0;
808 	qp->tx_ring_full = 0;
809 	qp->tx_err_no_buf = 0;
810 	qp->tx_memcpy = 0;
811 	qp->tx_async = 0;
812 }
813 
814 static void ntb_qp_link_cleanup(struct ntb_transport_qp *qp)
815 {
816 	struct ntb_transport_ctx *nt = qp->transport;
817 	struct pci_dev *pdev = nt->ndev->pdev;
818 
819 	dev_info(&pdev->dev, "qp %d: Link Cleanup\n", qp->qp_num);
820 
821 	cancel_delayed_work_sync(&qp->link_work);
822 	ntb_qp_link_down_reset(qp);
823 
824 	if (qp->event_handler)
825 		qp->event_handler(qp->cb_data, qp->link_is_up);
826 }
827 
828 static void ntb_qp_link_cleanup_work(struct work_struct *work)
829 {
830 	struct ntb_transport_qp *qp = container_of(work,
831 						   struct ntb_transport_qp,
832 						   link_cleanup);
833 	struct ntb_transport_ctx *nt = qp->transport;
834 
835 	ntb_qp_link_cleanup(qp);
836 
837 	if (nt->link_is_up)
838 		schedule_delayed_work(&qp->link_work,
839 				      msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
840 }
841 
842 static void ntb_qp_link_down(struct ntb_transport_qp *qp)
843 {
844 	schedule_work(&qp->link_cleanup);
845 }
846 
847 static void ntb_transport_link_cleanup(struct ntb_transport_ctx *nt)
848 {
849 	struct ntb_transport_qp *qp;
850 	u64 qp_bitmap_alloc;
851 	unsigned int i, count;
852 
853 	qp_bitmap_alloc = nt->qp_bitmap & ~nt->qp_bitmap_free;
854 
855 	/* Pass along the info to any clients */
856 	for (i = 0; i < nt->qp_count; i++)
857 		if (qp_bitmap_alloc & BIT_ULL(i)) {
858 			qp = &nt->qp_vec[i];
859 			ntb_qp_link_cleanup(qp);
860 			cancel_work_sync(&qp->link_cleanup);
861 			cancel_delayed_work_sync(&qp->link_work);
862 		}
863 
864 	if (!nt->link_is_up)
865 		cancel_delayed_work_sync(&nt->link_work);
866 
867 	for (i = 0; i < nt->mw_count; i++)
868 		ntb_free_mw(nt, i);
869 
870 	/* The scratchpad registers keep the values if the remote side
871 	 * goes down, blast them now to give them a sane value the next
872 	 * time they are accessed
873 	 */
874 	count = ntb_spad_count(nt->ndev);
875 	for (i = 0; i < count; i++)
876 		ntb_spad_write(nt->ndev, i, 0);
877 }
878 
879 static void ntb_transport_link_cleanup_work(struct work_struct *work)
880 {
881 	struct ntb_transport_ctx *nt =
882 		container_of(work, struct ntb_transport_ctx, link_cleanup);
883 
884 	ntb_transport_link_cleanup(nt);
885 }
886 
887 static void ntb_transport_event_callback(void *data)
888 {
889 	struct ntb_transport_ctx *nt = data;
890 
891 	if (ntb_link_is_up(nt->ndev, NULL, NULL) == 1)
892 		schedule_delayed_work(&nt->link_work, 0);
893 	else
894 		schedule_work(&nt->link_cleanup);
895 }
896 
897 static void ntb_transport_link_work(struct work_struct *work)
898 {
899 	struct ntb_transport_ctx *nt =
900 		container_of(work, struct ntb_transport_ctx, link_work.work);
901 	struct ntb_dev *ndev = nt->ndev;
902 	struct pci_dev *pdev = ndev->pdev;
903 	resource_size_t size;
904 	u32 val;
905 	int rc = 0, i, spad;
906 
907 	/* send the local info, in the opposite order of the way we read it */
908 	for (i = 0; i < nt->mw_count; i++) {
909 		size = nt->mw_vec[i].phys_size;
910 
911 		if (max_mw_size && size > max_mw_size)
912 			size = max_mw_size;
913 
914 		spad = MW0_SZ_HIGH + (i * 2);
915 		ntb_peer_spad_write(ndev, PIDX, spad, upper_32_bits(size));
916 
917 		spad = MW0_SZ_LOW + (i * 2);
918 		ntb_peer_spad_write(ndev, PIDX, spad, lower_32_bits(size));
919 	}
920 
921 	ntb_peer_spad_write(ndev, PIDX, NUM_MWS, nt->mw_count);
922 
923 	ntb_peer_spad_write(ndev, PIDX, NUM_QPS, nt->qp_count);
924 
925 	ntb_peer_spad_write(ndev, PIDX, VERSION, NTB_TRANSPORT_VERSION);
926 
927 	/* Query the remote side for its info */
928 	val = ntb_spad_read(ndev, VERSION);
929 	dev_dbg(&pdev->dev, "Remote version = %d\n", val);
930 	if (val != NTB_TRANSPORT_VERSION)
931 		goto out;
932 
933 	val = ntb_spad_read(ndev, NUM_QPS);
934 	dev_dbg(&pdev->dev, "Remote max number of qps = %d\n", val);
935 	if (val != nt->qp_count)
936 		goto out;
937 
938 	val = ntb_spad_read(ndev, NUM_MWS);
939 	dev_dbg(&pdev->dev, "Remote number of mws = %d\n", val);
940 	if (val != nt->mw_count)
941 		goto out;
942 
943 	for (i = 0; i < nt->mw_count; i++) {
944 		u64 val64;
945 
946 		val = ntb_spad_read(ndev, MW0_SZ_HIGH + (i * 2));
947 		val64 = (u64)val << 32;
948 
949 		val = ntb_spad_read(ndev, MW0_SZ_LOW + (i * 2));
950 		val64 |= val;
951 
952 		dev_dbg(&pdev->dev, "Remote MW%d size = %#llx\n", i, val64);
953 
954 		rc = ntb_set_mw(nt, i, val64);
955 		if (rc)
956 			goto out1;
957 	}
958 
959 	nt->link_is_up = true;
960 
961 	for (i = 0; i < nt->qp_count; i++) {
962 		struct ntb_transport_qp *qp = &nt->qp_vec[i];
963 
964 		ntb_transport_setup_qp_mw(nt, i);
965 
966 		if (qp->client_ready)
967 			schedule_delayed_work(&qp->link_work, 0);
968 	}
969 
970 	return;
971 
972 out1:
973 	for (i = 0; i < nt->mw_count; i++)
974 		ntb_free_mw(nt, i);
975 
976 	/* if there's an actual failure, we should just bail */
977 	if (rc < 0)
978 		return;
979 
980 out:
981 	if (ntb_link_is_up(ndev, NULL, NULL) == 1)
982 		schedule_delayed_work(&nt->link_work,
983 				      msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
984 }
985 
986 static void ntb_qp_link_work(struct work_struct *work)
987 {
988 	struct ntb_transport_qp *qp = container_of(work,
989 						   struct ntb_transport_qp,
990 						   link_work.work);
991 	struct pci_dev *pdev = qp->ndev->pdev;
992 	struct ntb_transport_ctx *nt = qp->transport;
993 	int val;
994 
995 	WARN_ON(!nt->link_is_up);
996 
997 	val = ntb_spad_read(nt->ndev, QP_LINKS);
998 
999 	ntb_peer_spad_write(nt->ndev, PIDX, QP_LINKS, val | BIT(qp->qp_num));
1000 
1001 	/* query remote spad for qp ready bits */
1002 	dev_dbg_ratelimited(&pdev->dev, "Remote QP link status = %x\n", val);
1003 
1004 	/* See if the remote side is up */
1005 	if (val & BIT(qp->qp_num)) {
1006 		dev_info(&pdev->dev, "qp %d: Link Up\n", qp->qp_num);
1007 		qp->link_is_up = true;
1008 		qp->active = true;
1009 
1010 		if (qp->event_handler)
1011 			qp->event_handler(qp->cb_data, qp->link_is_up);
1012 
1013 		if (qp->active)
1014 			tasklet_schedule(&qp->rxc_db_work);
1015 	} else if (nt->link_is_up)
1016 		schedule_delayed_work(&qp->link_work,
1017 				      msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
1018 }
1019 
1020 static int ntb_transport_init_queue(struct ntb_transport_ctx *nt,
1021 				    unsigned int qp_num)
1022 {
1023 	struct ntb_transport_qp *qp;
1024 	phys_addr_t mw_base;
1025 	resource_size_t mw_size;
1026 	unsigned int num_qps_mw, tx_size;
1027 	unsigned int mw_num, mw_count, qp_count;
1028 	u64 qp_offset;
1029 
1030 	mw_count = nt->mw_count;
1031 	qp_count = nt->qp_count;
1032 
1033 	mw_num = QP_TO_MW(nt, qp_num);
1034 
1035 	qp = &nt->qp_vec[qp_num];
1036 	qp->qp_num = qp_num;
1037 	qp->transport = nt;
1038 	qp->ndev = nt->ndev;
1039 	qp->client_ready = false;
1040 	qp->event_handler = NULL;
1041 	ntb_qp_link_down_reset(qp);
1042 
1043 	if (mw_num < qp_count % mw_count)
1044 		num_qps_mw = qp_count / mw_count + 1;
1045 	else
1046 		num_qps_mw = qp_count / mw_count;
1047 
1048 	mw_base = nt->mw_vec[mw_num].phys_addr;
1049 	mw_size = nt->mw_vec[mw_num].phys_size;
1050 
1051 	if (max_mw_size && mw_size > max_mw_size)
1052 		mw_size = max_mw_size;
1053 
1054 	tx_size = (unsigned int)mw_size / num_qps_mw;
1055 	qp_offset = tx_size * (qp_num / mw_count);
1056 
1057 	qp->tx_mw_size = tx_size;
1058 	qp->tx_mw = nt->mw_vec[mw_num].vbase + qp_offset;
1059 	if (!qp->tx_mw)
1060 		return -EINVAL;
1061 
1062 	qp->tx_mw_phys = mw_base + qp_offset;
1063 	if (!qp->tx_mw_phys)
1064 		return -EINVAL;
1065 
1066 	tx_size -= sizeof(struct ntb_rx_info);
1067 	qp->rx_info = qp->tx_mw + tx_size;
1068 
1069 	/* Due to housekeeping, there must be atleast 2 buffs */
1070 	qp->tx_max_frame = min(transport_mtu, tx_size / 2);
1071 	qp->tx_max_entry = tx_size / qp->tx_max_frame;
1072 
1073 	if (nt->debugfs_node_dir) {
1074 		char debugfs_name[4];
1075 
1076 		snprintf(debugfs_name, 4, "qp%d", qp_num);
1077 		qp->debugfs_dir = debugfs_create_dir(debugfs_name,
1078 						     nt->debugfs_node_dir);
1079 
1080 		qp->debugfs_stats = debugfs_create_file("stats", S_IRUSR,
1081 							qp->debugfs_dir, qp,
1082 							&ntb_qp_debugfs_stats);
1083 	} else {
1084 		qp->debugfs_dir = NULL;
1085 		qp->debugfs_stats = NULL;
1086 	}
1087 
1088 	INIT_DELAYED_WORK(&qp->link_work, ntb_qp_link_work);
1089 	INIT_WORK(&qp->link_cleanup, ntb_qp_link_cleanup_work);
1090 
1091 	spin_lock_init(&qp->ntb_rx_q_lock);
1092 	spin_lock_init(&qp->ntb_tx_free_q_lock);
1093 
1094 	INIT_LIST_HEAD(&qp->rx_post_q);
1095 	INIT_LIST_HEAD(&qp->rx_pend_q);
1096 	INIT_LIST_HEAD(&qp->rx_free_q);
1097 	INIT_LIST_HEAD(&qp->tx_free_q);
1098 
1099 	tasklet_init(&qp->rxc_db_work, ntb_transport_rxc_db,
1100 		     (unsigned long)qp);
1101 
1102 	return 0;
1103 }
1104 
1105 static int ntb_transport_probe(struct ntb_client *self, struct ntb_dev *ndev)
1106 {
1107 	struct ntb_transport_ctx *nt;
1108 	struct ntb_transport_mw *mw;
1109 	unsigned int mw_count, qp_count, spad_count, max_mw_count_for_spads;
1110 	u64 qp_bitmap;
1111 	int node;
1112 	int rc, i;
1113 
1114 	mw_count = ntb_peer_mw_count(ndev);
1115 
1116 	if (!ndev->ops->mw_set_trans) {
1117 		dev_err(&ndev->dev, "Inbound MW based NTB API is required\n");
1118 		return -EINVAL;
1119 	}
1120 
1121 	if (ntb_db_is_unsafe(ndev))
1122 		dev_dbg(&ndev->dev,
1123 			"doorbell is unsafe, proceed anyway...\n");
1124 	if (ntb_spad_is_unsafe(ndev))
1125 		dev_dbg(&ndev->dev,
1126 			"scratchpad is unsafe, proceed anyway...\n");
1127 
1128 	if (ntb_peer_port_count(ndev) != NTB_DEF_PEER_CNT)
1129 		dev_warn(&ndev->dev, "Multi-port NTB devices unsupported\n");
1130 
1131 	node = dev_to_node(&ndev->dev);
1132 
1133 	nt = kzalloc_node(sizeof(*nt), GFP_KERNEL, node);
1134 	if (!nt)
1135 		return -ENOMEM;
1136 
1137 	nt->ndev = ndev;
1138 	spad_count = ntb_spad_count(ndev);
1139 
1140 	/* Limit the MW's based on the availability of scratchpads */
1141 
1142 	if (spad_count < NTB_TRANSPORT_MIN_SPADS) {
1143 		nt->mw_count = 0;
1144 		rc = -EINVAL;
1145 		goto err;
1146 	}
1147 
1148 	max_mw_count_for_spads = (spad_count - MW0_SZ_HIGH) / 2;
1149 	nt->mw_count = min(mw_count, max_mw_count_for_spads);
1150 
1151 	nt->mw_vec = kcalloc_node(mw_count, sizeof(*nt->mw_vec),
1152 				  GFP_KERNEL, node);
1153 	if (!nt->mw_vec) {
1154 		rc = -ENOMEM;
1155 		goto err;
1156 	}
1157 
1158 	for (i = 0; i < mw_count; i++) {
1159 		mw = &nt->mw_vec[i];
1160 
1161 		rc = ntb_peer_mw_get_addr(ndev, i, &mw->phys_addr,
1162 					  &mw->phys_size);
1163 		if (rc)
1164 			goto err1;
1165 
1166 		mw->vbase = ioremap_wc(mw->phys_addr, mw->phys_size);
1167 		if (!mw->vbase) {
1168 			rc = -ENOMEM;
1169 			goto err1;
1170 		}
1171 
1172 		mw->buff_size = 0;
1173 		mw->xlat_size = 0;
1174 		mw->virt_addr = NULL;
1175 		mw->dma_addr = 0;
1176 	}
1177 
1178 	qp_bitmap = ntb_db_valid_mask(ndev);
1179 
1180 	qp_count = ilog2(qp_bitmap);
1181 	if (max_num_clients && max_num_clients < qp_count)
1182 		qp_count = max_num_clients;
1183 	else if (nt->mw_count < qp_count)
1184 		qp_count = nt->mw_count;
1185 
1186 	qp_bitmap &= BIT_ULL(qp_count) - 1;
1187 
1188 	nt->qp_count = qp_count;
1189 	nt->qp_bitmap = qp_bitmap;
1190 	nt->qp_bitmap_free = qp_bitmap;
1191 
1192 	nt->qp_vec = kcalloc_node(qp_count, sizeof(*nt->qp_vec),
1193 				  GFP_KERNEL, node);
1194 	if (!nt->qp_vec) {
1195 		rc = -ENOMEM;
1196 		goto err1;
1197 	}
1198 
1199 	if (nt_debugfs_dir) {
1200 		nt->debugfs_node_dir =
1201 			debugfs_create_dir(pci_name(ndev->pdev),
1202 					   nt_debugfs_dir);
1203 	}
1204 
1205 	for (i = 0; i < qp_count; i++) {
1206 		rc = ntb_transport_init_queue(nt, i);
1207 		if (rc)
1208 			goto err2;
1209 	}
1210 
1211 	INIT_DELAYED_WORK(&nt->link_work, ntb_transport_link_work);
1212 	INIT_WORK(&nt->link_cleanup, ntb_transport_link_cleanup_work);
1213 
1214 	rc = ntb_set_ctx(ndev, nt, &ntb_transport_ops);
1215 	if (rc)
1216 		goto err2;
1217 
1218 	INIT_LIST_HEAD(&nt->client_devs);
1219 	rc = ntb_bus_init(nt);
1220 	if (rc)
1221 		goto err3;
1222 
1223 	nt->link_is_up = false;
1224 	ntb_link_enable(ndev, NTB_SPEED_AUTO, NTB_WIDTH_AUTO);
1225 	ntb_link_event(ndev);
1226 
1227 	return 0;
1228 
1229 err3:
1230 	ntb_clear_ctx(ndev);
1231 err2:
1232 	kfree(nt->qp_vec);
1233 err1:
1234 	while (i--) {
1235 		mw = &nt->mw_vec[i];
1236 		iounmap(mw->vbase);
1237 	}
1238 	kfree(nt->mw_vec);
1239 err:
1240 	kfree(nt);
1241 	return rc;
1242 }
1243 
1244 static void ntb_transport_free(struct ntb_client *self, struct ntb_dev *ndev)
1245 {
1246 	struct ntb_transport_ctx *nt = ndev->ctx;
1247 	struct ntb_transport_qp *qp;
1248 	u64 qp_bitmap_alloc;
1249 	int i;
1250 
1251 	ntb_transport_link_cleanup(nt);
1252 	cancel_work_sync(&nt->link_cleanup);
1253 	cancel_delayed_work_sync(&nt->link_work);
1254 
1255 	qp_bitmap_alloc = nt->qp_bitmap & ~nt->qp_bitmap_free;
1256 
1257 	/* verify that all the qp's are freed */
1258 	for (i = 0; i < nt->qp_count; i++) {
1259 		qp = &nt->qp_vec[i];
1260 		if (qp_bitmap_alloc & BIT_ULL(i))
1261 			ntb_transport_free_queue(qp);
1262 		debugfs_remove_recursive(qp->debugfs_dir);
1263 	}
1264 
1265 	ntb_link_disable(ndev);
1266 	ntb_clear_ctx(ndev);
1267 
1268 	ntb_bus_remove(nt);
1269 
1270 	for (i = nt->mw_count; i--; ) {
1271 		ntb_free_mw(nt, i);
1272 		iounmap(nt->mw_vec[i].vbase);
1273 	}
1274 
1275 	kfree(nt->qp_vec);
1276 	kfree(nt->mw_vec);
1277 	kfree(nt);
1278 }
1279 
1280 static void ntb_complete_rxc(struct ntb_transport_qp *qp)
1281 {
1282 	struct ntb_queue_entry *entry;
1283 	void *cb_data;
1284 	unsigned int len;
1285 	unsigned long irqflags;
1286 
1287 	spin_lock_irqsave(&qp->ntb_rx_q_lock, irqflags);
1288 
1289 	while (!list_empty(&qp->rx_post_q)) {
1290 		entry = list_first_entry(&qp->rx_post_q,
1291 					 struct ntb_queue_entry, entry);
1292 		if (!(entry->flags & DESC_DONE_FLAG))
1293 			break;
1294 
1295 		entry->rx_hdr->flags = 0;
1296 		iowrite32(entry->rx_index, &qp->rx_info->entry);
1297 
1298 		cb_data = entry->cb_data;
1299 		len = entry->len;
1300 
1301 		list_move_tail(&entry->entry, &qp->rx_free_q);
1302 
1303 		spin_unlock_irqrestore(&qp->ntb_rx_q_lock, irqflags);
1304 
1305 		if (qp->rx_handler && qp->client_ready)
1306 			qp->rx_handler(qp, qp->cb_data, cb_data, len);
1307 
1308 		spin_lock_irqsave(&qp->ntb_rx_q_lock, irqflags);
1309 	}
1310 
1311 	spin_unlock_irqrestore(&qp->ntb_rx_q_lock, irqflags);
1312 }
1313 
1314 static void ntb_rx_copy_callback(void *data,
1315 				 const struct dmaengine_result *res)
1316 {
1317 	struct ntb_queue_entry *entry = data;
1318 
1319 	/* we need to check DMA results if we are using DMA */
1320 	if (res) {
1321 		enum dmaengine_tx_result dma_err = res->result;
1322 
1323 		switch (dma_err) {
1324 		case DMA_TRANS_READ_FAILED:
1325 		case DMA_TRANS_WRITE_FAILED:
1326 			entry->errors++;
1327 			/* fall through */
1328 		case DMA_TRANS_ABORTED:
1329 		{
1330 			struct ntb_transport_qp *qp = entry->qp;
1331 			void *offset = qp->rx_buff + qp->rx_max_frame *
1332 					qp->rx_index;
1333 
1334 			ntb_memcpy_rx(entry, offset);
1335 			qp->rx_memcpy++;
1336 			return;
1337 		}
1338 
1339 		case DMA_TRANS_NOERROR:
1340 		default:
1341 			break;
1342 		}
1343 	}
1344 
1345 	entry->flags |= DESC_DONE_FLAG;
1346 
1347 	ntb_complete_rxc(entry->qp);
1348 }
1349 
1350 static void ntb_memcpy_rx(struct ntb_queue_entry *entry, void *offset)
1351 {
1352 	void *buf = entry->buf;
1353 	size_t len = entry->len;
1354 
1355 	memcpy(buf, offset, len);
1356 
1357 	/* Ensure that the data is fully copied out before clearing the flag */
1358 	wmb();
1359 
1360 	ntb_rx_copy_callback(entry, NULL);
1361 }
1362 
1363 static int ntb_async_rx_submit(struct ntb_queue_entry *entry, void *offset)
1364 {
1365 	struct dma_async_tx_descriptor *txd;
1366 	struct ntb_transport_qp *qp = entry->qp;
1367 	struct dma_chan *chan = qp->rx_dma_chan;
1368 	struct dma_device *device;
1369 	size_t pay_off, buff_off, len;
1370 	struct dmaengine_unmap_data *unmap;
1371 	dma_cookie_t cookie;
1372 	void *buf = entry->buf;
1373 
1374 	len = entry->len;
1375 	device = chan->device;
1376 	pay_off = (size_t)offset & ~PAGE_MASK;
1377 	buff_off = (size_t)buf & ~PAGE_MASK;
1378 
1379 	if (!is_dma_copy_aligned(device, pay_off, buff_off, len))
1380 		goto err;
1381 
1382 	unmap = dmaengine_get_unmap_data(device->dev, 2, GFP_NOWAIT);
1383 	if (!unmap)
1384 		goto err;
1385 
1386 	unmap->len = len;
1387 	unmap->addr[0] = dma_map_page(device->dev, virt_to_page(offset),
1388 				      pay_off, len, DMA_TO_DEVICE);
1389 	if (dma_mapping_error(device->dev, unmap->addr[0]))
1390 		goto err_get_unmap;
1391 
1392 	unmap->to_cnt = 1;
1393 
1394 	unmap->addr[1] = dma_map_page(device->dev, virt_to_page(buf),
1395 				      buff_off, len, DMA_FROM_DEVICE);
1396 	if (dma_mapping_error(device->dev, unmap->addr[1]))
1397 		goto err_get_unmap;
1398 
1399 	unmap->from_cnt = 1;
1400 
1401 	txd = device->device_prep_dma_memcpy(chan, unmap->addr[1],
1402 					     unmap->addr[0], len,
1403 					     DMA_PREP_INTERRUPT);
1404 	if (!txd)
1405 		goto err_get_unmap;
1406 
1407 	txd->callback_result = ntb_rx_copy_callback;
1408 	txd->callback_param = entry;
1409 	dma_set_unmap(txd, unmap);
1410 
1411 	cookie = dmaengine_submit(txd);
1412 	if (dma_submit_error(cookie))
1413 		goto err_set_unmap;
1414 
1415 	dmaengine_unmap_put(unmap);
1416 
1417 	qp->last_cookie = cookie;
1418 
1419 	qp->rx_async++;
1420 
1421 	return 0;
1422 
1423 err_set_unmap:
1424 	dmaengine_unmap_put(unmap);
1425 err_get_unmap:
1426 	dmaengine_unmap_put(unmap);
1427 err:
1428 	return -ENXIO;
1429 }
1430 
1431 static void ntb_async_rx(struct ntb_queue_entry *entry, void *offset)
1432 {
1433 	struct ntb_transport_qp *qp = entry->qp;
1434 	struct dma_chan *chan = qp->rx_dma_chan;
1435 	int res;
1436 
1437 	if (!chan)
1438 		goto err;
1439 
1440 	if (entry->len < copy_bytes)
1441 		goto err;
1442 
1443 	res = ntb_async_rx_submit(entry, offset);
1444 	if (res < 0)
1445 		goto err;
1446 
1447 	if (!entry->retries)
1448 		qp->rx_async++;
1449 
1450 	return;
1451 
1452 err:
1453 	ntb_memcpy_rx(entry, offset);
1454 	qp->rx_memcpy++;
1455 }
1456 
1457 static int ntb_process_rxc(struct ntb_transport_qp *qp)
1458 {
1459 	struct ntb_payload_header *hdr;
1460 	struct ntb_queue_entry *entry;
1461 	void *offset;
1462 
1463 	offset = qp->rx_buff + qp->rx_max_frame * qp->rx_index;
1464 	hdr = offset + qp->rx_max_frame - sizeof(struct ntb_payload_header);
1465 
1466 	dev_dbg(&qp->ndev->pdev->dev, "qp %d: RX ver %u len %d flags %x\n",
1467 		qp->qp_num, hdr->ver, hdr->len, hdr->flags);
1468 
1469 	if (!(hdr->flags & DESC_DONE_FLAG)) {
1470 		dev_dbg(&qp->ndev->pdev->dev, "done flag not set\n");
1471 		qp->rx_ring_empty++;
1472 		return -EAGAIN;
1473 	}
1474 
1475 	if (hdr->flags & LINK_DOWN_FLAG) {
1476 		dev_dbg(&qp->ndev->pdev->dev, "link down flag set\n");
1477 		ntb_qp_link_down(qp);
1478 		hdr->flags = 0;
1479 		return -EAGAIN;
1480 	}
1481 
1482 	if (hdr->ver != (u32)qp->rx_pkts) {
1483 		dev_dbg(&qp->ndev->pdev->dev,
1484 			"version mismatch, expected %llu - got %u\n",
1485 			qp->rx_pkts, hdr->ver);
1486 		qp->rx_err_ver++;
1487 		return -EIO;
1488 	}
1489 
1490 	entry = ntb_list_mv(&qp->ntb_rx_q_lock, &qp->rx_pend_q, &qp->rx_post_q);
1491 	if (!entry) {
1492 		dev_dbg(&qp->ndev->pdev->dev, "no receive buffer\n");
1493 		qp->rx_err_no_buf++;
1494 		return -EAGAIN;
1495 	}
1496 
1497 	entry->rx_hdr = hdr;
1498 	entry->rx_index = qp->rx_index;
1499 
1500 	if (hdr->len > entry->len) {
1501 		dev_dbg(&qp->ndev->pdev->dev,
1502 			"receive buffer overflow! Wanted %d got %d\n",
1503 			hdr->len, entry->len);
1504 		qp->rx_err_oflow++;
1505 
1506 		entry->len = -EIO;
1507 		entry->flags |= DESC_DONE_FLAG;
1508 
1509 		ntb_complete_rxc(qp);
1510 	} else {
1511 		dev_dbg(&qp->ndev->pdev->dev,
1512 			"RX OK index %u ver %u size %d into buf size %d\n",
1513 			qp->rx_index, hdr->ver, hdr->len, entry->len);
1514 
1515 		qp->rx_bytes += hdr->len;
1516 		qp->rx_pkts++;
1517 
1518 		entry->len = hdr->len;
1519 
1520 		ntb_async_rx(entry, offset);
1521 	}
1522 
1523 	qp->rx_index++;
1524 	qp->rx_index %= qp->rx_max_entry;
1525 
1526 	return 0;
1527 }
1528 
1529 static void ntb_transport_rxc_db(unsigned long data)
1530 {
1531 	struct ntb_transport_qp *qp = (void *)data;
1532 	int rc, i;
1533 
1534 	dev_dbg(&qp->ndev->pdev->dev, "%s: doorbell %d received\n",
1535 		__func__, qp->qp_num);
1536 
1537 	/* Limit the number of packets processed in a single interrupt to
1538 	 * provide fairness to others
1539 	 */
1540 	for (i = 0; i < qp->rx_max_entry; i++) {
1541 		rc = ntb_process_rxc(qp);
1542 		if (rc)
1543 			break;
1544 	}
1545 
1546 	if (i && qp->rx_dma_chan)
1547 		dma_async_issue_pending(qp->rx_dma_chan);
1548 
1549 	if (i == qp->rx_max_entry) {
1550 		/* there is more work to do */
1551 		if (qp->active)
1552 			tasklet_schedule(&qp->rxc_db_work);
1553 	} else if (ntb_db_read(qp->ndev) & BIT_ULL(qp->qp_num)) {
1554 		/* the doorbell bit is set: clear it */
1555 		ntb_db_clear(qp->ndev, BIT_ULL(qp->qp_num));
1556 		/* ntb_db_read ensures ntb_db_clear write is committed */
1557 		ntb_db_read(qp->ndev);
1558 
1559 		/* an interrupt may have arrived between finishing
1560 		 * ntb_process_rxc and clearing the doorbell bit:
1561 		 * there might be some more work to do.
1562 		 */
1563 		if (qp->active)
1564 			tasklet_schedule(&qp->rxc_db_work);
1565 	}
1566 }
1567 
1568 static void ntb_tx_copy_callback(void *data,
1569 				 const struct dmaengine_result *res)
1570 {
1571 	struct ntb_queue_entry *entry = data;
1572 	struct ntb_transport_qp *qp = entry->qp;
1573 	struct ntb_payload_header __iomem *hdr = entry->tx_hdr;
1574 
1575 	/* we need to check DMA results if we are using DMA */
1576 	if (res) {
1577 		enum dmaengine_tx_result dma_err = res->result;
1578 
1579 		switch (dma_err) {
1580 		case DMA_TRANS_READ_FAILED:
1581 		case DMA_TRANS_WRITE_FAILED:
1582 			entry->errors++;
1583 			/* fall through */
1584 		case DMA_TRANS_ABORTED:
1585 		{
1586 			void __iomem *offset =
1587 				qp->tx_mw + qp->tx_max_frame *
1588 				entry->tx_index;
1589 
1590 			/* resubmit via CPU */
1591 			ntb_memcpy_tx(entry, offset);
1592 			qp->tx_memcpy++;
1593 			return;
1594 		}
1595 
1596 		case DMA_TRANS_NOERROR:
1597 		default:
1598 			break;
1599 		}
1600 	}
1601 
1602 	iowrite32(entry->flags | DESC_DONE_FLAG, &hdr->flags);
1603 
1604 	ntb_peer_db_set(qp->ndev, BIT_ULL(qp->qp_num));
1605 
1606 	/* The entry length can only be zero if the packet is intended to be a
1607 	 * "link down" or similar.  Since no payload is being sent in these
1608 	 * cases, there is nothing to add to the completion queue.
1609 	 */
1610 	if (entry->len > 0) {
1611 		qp->tx_bytes += entry->len;
1612 
1613 		if (qp->tx_handler)
1614 			qp->tx_handler(qp, qp->cb_data, entry->cb_data,
1615 				       entry->len);
1616 	}
1617 
1618 	ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry, &qp->tx_free_q);
1619 }
1620 
1621 static void ntb_memcpy_tx(struct ntb_queue_entry *entry, void __iomem *offset)
1622 {
1623 #ifdef ARCH_HAS_NOCACHE_UACCESS
1624 	/*
1625 	 * Using non-temporal mov to improve performance on non-cached
1626 	 * writes, even though we aren't actually copying from user space.
1627 	 */
1628 	__copy_from_user_inatomic_nocache(offset, entry->buf, entry->len);
1629 #else
1630 	memcpy_toio(offset, entry->buf, entry->len);
1631 #endif
1632 
1633 	/* Ensure that the data is fully copied out before setting the flags */
1634 	wmb();
1635 
1636 	ntb_tx_copy_callback(entry, NULL);
1637 }
1638 
1639 static int ntb_async_tx_submit(struct ntb_transport_qp *qp,
1640 			       struct ntb_queue_entry *entry)
1641 {
1642 	struct dma_async_tx_descriptor *txd;
1643 	struct dma_chan *chan = qp->tx_dma_chan;
1644 	struct dma_device *device;
1645 	size_t len = entry->len;
1646 	void *buf = entry->buf;
1647 	size_t dest_off, buff_off;
1648 	struct dmaengine_unmap_data *unmap;
1649 	dma_addr_t dest;
1650 	dma_cookie_t cookie;
1651 
1652 	device = chan->device;
1653 	dest = qp->tx_mw_dma_addr + qp->tx_max_frame * entry->tx_index;
1654 	buff_off = (size_t)buf & ~PAGE_MASK;
1655 	dest_off = (size_t)dest & ~PAGE_MASK;
1656 
1657 	if (!is_dma_copy_aligned(device, buff_off, dest_off, len))
1658 		goto err;
1659 
1660 	unmap = dmaengine_get_unmap_data(device->dev, 1, GFP_NOWAIT);
1661 	if (!unmap)
1662 		goto err;
1663 
1664 	unmap->len = len;
1665 	unmap->addr[0] = dma_map_page(device->dev, virt_to_page(buf),
1666 				      buff_off, len, DMA_TO_DEVICE);
1667 	if (dma_mapping_error(device->dev, unmap->addr[0]))
1668 		goto err_get_unmap;
1669 
1670 	unmap->to_cnt = 1;
1671 
1672 	txd = device->device_prep_dma_memcpy(chan, dest, unmap->addr[0], len,
1673 					     DMA_PREP_INTERRUPT);
1674 	if (!txd)
1675 		goto err_get_unmap;
1676 
1677 	txd->callback_result = ntb_tx_copy_callback;
1678 	txd->callback_param = entry;
1679 	dma_set_unmap(txd, unmap);
1680 
1681 	cookie = dmaengine_submit(txd);
1682 	if (dma_submit_error(cookie))
1683 		goto err_set_unmap;
1684 
1685 	dmaengine_unmap_put(unmap);
1686 
1687 	dma_async_issue_pending(chan);
1688 
1689 	return 0;
1690 err_set_unmap:
1691 	dmaengine_unmap_put(unmap);
1692 err_get_unmap:
1693 	dmaengine_unmap_put(unmap);
1694 err:
1695 	return -ENXIO;
1696 }
1697 
1698 static void ntb_async_tx(struct ntb_transport_qp *qp,
1699 			 struct ntb_queue_entry *entry)
1700 {
1701 	struct ntb_payload_header __iomem *hdr;
1702 	struct dma_chan *chan = qp->tx_dma_chan;
1703 	void __iomem *offset;
1704 	int res;
1705 
1706 	entry->tx_index = qp->tx_index;
1707 	offset = qp->tx_mw + qp->tx_max_frame * entry->tx_index;
1708 	hdr = offset + qp->tx_max_frame - sizeof(struct ntb_payload_header);
1709 	entry->tx_hdr = hdr;
1710 
1711 	iowrite32(entry->len, &hdr->len);
1712 	iowrite32((u32)qp->tx_pkts, &hdr->ver);
1713 
1714 	if (!chan)
1715 		goto err;
1716 
1717 	if (entry->len < copy_bytes)
1718 		goto err;
1719 
1720 	res = ntb_async_tx_submit(qp, entry);
1721 	if (res < 0)
1722 		goto err;
1723 
1724 	if (!entry->retries)
1725 		qp->tx_async++;
1726 
1727 	return;
1728 
1729 err:
1730 	ntb_memcpy_tx(entry, offset);
1731 	qp->tx_memcpy++;
1732 }
1733 
1734 static int ntb_process_tx(struct ntb_transport_qp *qp,
1735 			  struct ntb_queue_entry *entry)
1736 {
1737 	if (qp->tx_index == qp->remote_rx_info->entry) {
1738 		qp->tx_ring_full++;
1739 		return -EAGAIN;
1740 	}
1741 
1742 	if (entry->len > qp->tx_max_frame - sizeof(struct ntb_payload_header)) {
1743 		if (qp->tx_handler)
1744 			qp->tx_handler(qp, qp->cb_data, NULL, -EIO);
1745 
1746 		ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
1747 			     &qp->tx_free_q);
1748 		return 0;
1749 	}
1750 
1751 	ntb_async_tx(qp, entry);
1752 
1753 	qp->tx_index++;
1754 	qp->tx_index %= qp->tx_max_entry;
1755 
1756 	qp->tx_pkts++;
1757 
1758 	return 0;
1759 }
1760 
1761 static void ntb_send_link_down(struct ntb_transport_qp *qp)
1762 {
1763 	struct pci_dev *pdev = qp->ndev->pdev;
1764 	struct ntb_queue_entry *entry;
1765 	int i, rc;
1766 
1767 	if (!qp->link_is_up)
1768 		return;
1769 
1770 	dev_info(&pdev->dev, "qp %d: Send Link Down\n", qp->qp_num);
1771 
1772 	for (i = 0; i < NTB_LINK_DOWN_TIMEOUT; i++) {
1773 		entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
1774 		if (entry)
1775 			break;
1776 		msleep(100);
1777 	}
1778 
1779 	if (!entry)
1780 		return;
1781 
1782 	entry->cb_data = NULL;
1783 	entry->buf = NULL;
1784 	entry->len = 0;
1785 	entry->flags = LINK_DOWN_FLAG;
1786 
1787 	rc = ntb_process_tx(qp, entry);
1788 	if (rc)
1789 		dev_err(&pdev->dev, "ntb: QP%d unable to send linkdown msg\n",
1790 			qp->qp_num);
1791 
1792 	ntb_qp_link_down_reset(qp);
1793 }
1794 
1795 static bool ntb_dma_filter_fn(struct dma_chan *chan, void *node)
1796 {
1797 	return dev_to_node(&chan->dev->device) == (int)(unsigned long)node;
1798 }
1799 
1800 /**
1801  * ntb_transport_create_queue - Create a new NTB transport layer queue
1802  * @rx_handler: receive callback function
1803  * @tx_handler: transmit callback function
1804  * @event_handler: event callback function
1805  *
1806  * Create a new NTB transport layer queue and provide the queue with a callback
1807  * routine for both transmit and receive.  The receive callback routine will be
1808  * used to pass up data when the transport has received it on the queue.   The
1809  * transmit callback routine will be called when the transport has completed the
1810  * transmission of the data on the queue and the data is ready to be freed.
1811  *
1812  * RETURNS: pointer to newly created ntb_queue, NULL on error.
1813  */
1814 struct ntb_transport_qp *
1815 ntb_transport_create_queue(void *data, struct device *client_dev,
1816 			   const struct ntb_queue_handlers *handlers)
1817 {
1818 	struct ntb_dev *ndev;
1819 	struct pci_dev *pdev;
1820 	struct ntb_transport_ctx *nt;
1821 	struct ntb_queue_entry *entry;
1822 	struct ntb_transport_qp *qp;
1823 	u64 qp_bit;
1824 	unsigned int free_queue;
1825 	dma_cap_mask_t dma_mask;
1826 	int node;
1827 	int i;
1828 
1829 	ndev = dev_ntb(client_dev->parent);
1830 	pdev = ndev->pdev;
1831 	nt = ndev->ctx;
1832 
1833 	node = dev_to_node(&ndev->dev);
1834 
1835 	free_queue = ffs(nt->qp_bitmap_free);
1836 	if (!free_queue)
1837 		goto err;
1838 
1839 	/* decrement free_queue to make it zero based */
1840 	free_queue--;
1841 
1842 	qp = &nt->qp_vec[free_queue];
1843 	qp_bit = BIT_ULL(qp->qp_num);
1844 
1845 	nt->qp_bitmap_free &= ~qp_bit;
1846 
1847 	qp->cb_data = data;
1848 	qp->rx_handler = handlers->rx_handler;
1849 	qp->tx_handler = handlers->tx_handler;
1850 	qp->event_handler = handlers->event_handler;
1851 
1852 	dma_cap_zero(dma_mask);
1853 	dma_cap_set(DMA_MEMCPY, dma_mask);
1854 
1855 	if (use_dma) {
1856 		qp->tx_dma_chan =
1857 			dma_request_channel(dma_mask, ntb_dma_filter_fn,
1858 					    (void *)(unsigned long)node);
1859 		if (!qp->tx_dma_chan)
1860 			dev_info(&pdev->dev, "Unable to allocate TX DMA channel\n");
1861 
1862 		qp->rx_dma_chan =
1863 			dma_request_channel(dma_mask, ntb_dma_filter_fn,
1864 					    (void *)(unsigned long)node);
1865 		if (!qp->rx_dma_chan)
1866 			dev_info(&pdev->dev, "Unable to allocate RX DMA channel\n");
1867 	} else {
1868 		qp->tx_dma_chan = NULL;
1869 		qp->rx_dma_chan = NULL;
1870 	}
1871 
1872 	if (qp->tx_dma_chan) {
1873 		qp->tx_mw_dma_addr =
1874 			dma_map_resource(qp->tx_dma_chan->device->dev,
1875 					 qp->tx_mw_phys, qp->tx_mw_size,
1876 					 DMA_FROM_DEVICE, 0);
1877 		if (dma_mapping_error(qp->tx_dma_chan->device->dev,
1878 				      qp->tx_mw_dma_addr)) {
1879 			qp->tx_mw_dma_addr = 0;
1880 			goto err1;
1881 		}
1882 	}
1883 
1884 	dev_dbg(&pdev->dev, "Using %s memcpy for TX\n",
1885 		qp->tx_dma_chan ? "DMA" : "CPU");
1886 
1887 	dev_dbg(&pdev->dev, "Using %s memcpy for RX\n",
1888 		qp->rx_dma_chan ? "DMA" : "CPU");
1889 
1890 	for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) {
1891 		entry = kzalloc_node(sizeof(*entry), GFP_KERNEL, node);
1892 		if (!entry)
1893 			goto err1;
1894 
1895 		entry->qp = qp;
1896 		ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry,
1897 			     &qp->rx_free_q);
1898 	}
1899 	qp->rx_alloc_entry = NTB_QP_DEF_NUM_ENTRIES;
1900 
1901 	for (i = 0; i < qp->tx_max_entry; i++) {
1902 		entry = kzalloc_node(sizeof(*entry), GFP_KERNEL, node);
1903 		if (!entry)
1904 			goto err2;
1905 
1906 		entry->qp = qp;
1907 		ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
1908 			     &qp->tx_free_q);
1909 	}
1910 
1911 	ntb_db_clear(qp->ndev, qp_bit);
1912 	ntb_db_clear_mask(qp->ndev, qp_bit);
1913 
1914 	dev_info(&pdev->dev, "NTB Transport QP %d created\n", qp->qp_num);
1915 
1916 	return qp;
1917 
1918 err2:
1919 	while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
1920 		kfree(entry);
1921 err1:
1922 	qp->rx_alloc_entry = 0;
1923 	while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q)))
1924 		kfree(entry);
1925 	if (qp->tx_mw_dma_addr)
1926 		dma_unmap_resource(qp->tx_dma_chan->device->dev,
1927 				   qp->tx_mw_dma_addr, qp->tx_mw_size,
1928 				   DMA_FROM_DEVICE, 0);
1929 	if (qp->tx_dma_chan)
1930 		dma_release_channel(qp->tx_dma_chan);
1931 	if (qp->rx_dma_chan)
1932 		dma_release_channel(qp->rx_dma_chan);
1933 	nt->qp_bitmap_free |= qp_bit;
1934 err:
1935 	return NULL;
1936 }
1937 EXPORT_SYMBOL_GPL(ntb_transport_create_queue);
1938 
1939 /**
1940  * ntb_transport_free_queue - Frees NTB transport queue
1941  * @qp: NTB queue to be freed
1942  *
1943  * Frees NTB transport queue
1944  */
1945 void ntb_transport_free_queue(struct ntb_transport_qp *qp)
1946 {
1947 	struct pci_dev *pdev;
1948 	struct ntb_queue_entry *entry;
1949 	u64 qp_bit;
1950 
1951 	if (!qp)
1952 		return;
1953 
1954 	pdev = qp->ndev->pdev;
1955 
1956 	qp->active = false;
1957 
1958 	if (qp->tx_dma_chan) {
1959 		struct dma_chan *chan = qp->tx_dma_chan;
1960 		/* Putting the dma_chan to NULL will force any new traffic to be
1961 		 * processed by the CPU instead of the DAM engine
1962 		 */
1963 		qp->tx_dma_chan = NULL;
1964 
1965 		/* Try to be nice and wait for any queued DMA engine
1966 		 * transactions to process before smashing it with a rock
1967 		 */
1968 		dma_sync_wait(chan, qp->last_cookie);
1969 		dmaengine_terminate_all(chan);
1970 
1971 		dma_unmap_resource(chan->device->dev,
1972 				   qp->tx_mw_dma_addr, qp->tx_mw_size,
1973 				   DMA_FROM_DEVICE, 0);
1974 
1975 		dma_release_channel(chan);
1976 	}
1977 
1978 	if (qp->rx_dma_chan) {
1979 		struct dma_chan *chan = qp->rx_dma_chan;
1980 		/* Putting the dma_chan to NULL will force any new traffic to be
1981 		 * processed by the CPU instead of the DAM engine
1982 		 */
1983 		qp->rx_dma_chan = NULL;
1984 
1985 		/* Try to be nice and wait for any queued DMA engine
1986 		 * transactions to process before smashing it with a rock
1987 		 */
1988 		dma_sync_wait(chan, qp->last_cookie);
1989 		dmaengine_terminate_all(chan);
1990 		dma_release_channel(chan);
1991 	}
1992 
1993 	qp_bit = BIT_ULL(qp->qp_num);
1994 
1995 	ntb_db_set_mask(qp->ndev, qp_bit);
1996 	tasklet_kill(&qp->rxc_db_work);
1997 
1998 	cancel_delayed_work_sync(&qp->link_work);
1999 
2000 	qp->cb_data = NULL;
2001 	qp->rx_handler = NULL;
2002 	qp->tx_handler = NULL;
2003 	qp->event_handler = NULL;
2004 
2005 	while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q)))
2006 		kfree(entry);
2007 
2008 	while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_pend_q))) {
2009 		dev_warn(&pdev->dev, "Freeing item from non-empty rx_pend_q\n");
2010 		kfree(entry);
2011 	}
2012 
2013 	while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_post_q))) {
2014 		dev_warn(&pdev->dev, "Freeing item from non-empty rx_post_q\n");
2015 		kfree(entry);
2016 	}
2017 
2018 	while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
2019 		kfree(entry);
2020 
2021 	qp->transport->qp_bitmap_free |= qp_bit;
2022 
2023 	dev_info(&pdev->dev, "NTB Transport QP %d freed\n", qp->qp_num);
2024 }
2025 EXPORT_SYMBOL_GPL(ntb_transport_free_queue);
2026 
2027 /**
2028  * ntb_transport_rx_remove - Dequeues enqueued rx packet
2029  * @qp: NTB queue to be freed
2030  * @len: pointer to variable to write enqueued buffers length
2031  *
2032  * Dequeues unused buffers from receive queue.  Should only be used during
2033  * shutdown of qp.
2034  *
2035  * RETURNS: NULL error value on error, or void* for success.
2036  */
2037 void *ntb_transport_rx_remove(struct ntb_transport_qp *qp, unsigned int *len)
2038 {
2039 	struct ntb_queue_entry *entry;
2040 	void *buf;
2041 
2042 	if (!qp || qp->client_ready)
2043 		return NULL;
2044 
2045 	entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_pend_q);
2046 	if (!entry)
2047 		return NULL;
2048 
2049 	buf = entry->cb_data;
2050 	*len = entry->len;
2051 
2052 	ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry, &qp->rx_free_q);
2053 
2054 	return buf;
2055 }
2056 EXPORT_SYMBOL_GPL(ntb_transport_rx_remove);
2057 
2058 /**
2059  * ntb_transport_rx_enqueue - Enqueue a new NTB queue entry
2060  * @qp: NTB transport layer queue the entry is to be enqueued on
2061  * @cb: per buffer pointer for callback function to use
2062  * @data: pointer to data buffer that incoming packets will be copied into
2063  * @len: length of the data buffer
2064  *
2065  * Enqueue a new receive buffer onto the transport queue into which a NTB
2066  * payload can be received into.
2067  *
2068  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
2069  */
2070 int ntb_transport_rx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
2071 			     unsigned int len)
2072 {
2073 	struct ntb_queue_entry *entry;
2074 
2075 	if (!qp)
2076 		return -EINVAL;
2077 
2078 	entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q);
2079 	if (!entry)
2080 		return -ENOMEM;
2081 
2082 	entry->cb_data = cb;
2083 	entry->buf = data;
2084 	entry->len = len;
2085 	entry->flags = 0;
2086 	entry->retries = 0;
2087 	entry->errors = 0;
2088 	entry->rx_index = 0;
2089 
2090 	ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry, &qp->rx_pend_q);
2091 
2092 	if (qp->active)
2093 		tasklet_schedule(&qp->rxc_db_work);
2094 
2095 	return 0;
2096 }
2097 EXPORT_SYMBOL_GPL(ntb_transport_rx_enqueue);
2098 
2099 /**
2100  * ntb_transport_tx_enqueue - Enqueue a new NTB queue entry
2101  * @qp: NTB transport layer queue the entry is to be enqueued on
2102  * @cb: per buffer pointer for callback function to use
2103  * @data: pointer to data buffer that will be sent
2104  * @len: length of the data buffer
2105  *
2106  * Enqueue a new transmit buffer onto the transport queue from which a NTB
2107  * payload will be transmitted.  This assumes that a lock is being held to
2108  * serialize access to the qp.
2109  *
2110  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
2111  */
2112 int ntb_transport_tx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
2113 			     unsigned int len)
2114 {
2115 	struct ntb_queue_entry *entry;
2116 	int rc;
2117 
2118 	if (!qp || !qp->link_is_up || !len)
2119 		return -EINVAL;
2120 
2121 	entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
2122 	if (!entry) {
2123 		qp->tx_err_no_buf++;
2124 		return -EBUSY;
2125 	}
2126 
2127 	entry->cb_data = cb;
2128 	entry->buf = data;
2129 	entry->len = len;
2130 	entry->flags = 0;
2131 	entry->errors = 0;
2132 	entry->retries = 0;
2133 	entry->tx_index = 0;
2134 
2135 	rc = ntb_process_tx(qp, entry);
2136 	if (rc)
2137 		ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
2138 			     &qp->tx_free_q);
2139 
2140 	return rc;
2141 }
2142 EXPORT_SYMBOL_GPL(ntb_transport_tx_enqueue);
2143 
2144 /**
2145  * ntb_transport_link_up - Notify NTB transport of client readiness to use queue
2146  * @qp: NTB transport layer queue to be enabled
2147  *
2148  * Notify NTB transport layer of client readiness to use queue
2149  */
2150 void ntb_transport_link_up(struct ntb_transport_qp *qp)
2151 {
2152 	if (!qp)
2153 		return;
2154 
2155 	qp->client_ready = true;
2156 
2157 	if (qp->transport->link_is_up)
2158 		schedule_delayed_work(&qp->link_work, 0);
2159 }
2160 EXPORT_SYMBOL_GPL(ntb_transport_link_up);
2161 
2162 /**
2163  * ntb_transport_link_down - Notify NTB transport to no longer enqueue data
2164  * @qp: NTB transport layer queue to be disabled
2165  *
2166  * Notify NTB transport layer of client's desire to no longer receive data on
2167  * transport queue specified.  It is the client's responsibility to ensure all
2168  * entries on queue are purged or otherwise handled appropriately.
2169  */
2170 void ntb_transport_link_down(struct ntb_transport_qp *qp)
2171 {
2172 	int val;
2173 
2174 	if (!qp)
2175 		return;
2176 
2177 	qp->client_ready = false;
2178 
2179 	val = ntb_spad_read(qp->ndev, QP_LINKS);
2180 
2181 	ntb_peer_spad_write(qp->ndev, PIDX, QP_LINKS, val & ~BIT(qp->qp_num));
2182 
2183 	if (qp->link_is_up)
2184 		ntb_send_link_down(qp);
2185 	else
2186 		cancel_delayed_work_sync(&qp->link_work);
2187 }
2188 EXPORT_SYMBOL_GPL(ntb_transport_link_down);
2189 
2190 /**
2191  * ntb_transport_link_query - Query transport link state
2192  * @qp: NTB transport layer queue to be queried
2193  *
2194  * Query connectivity to the remote system of the NTB transport queue
2195  *
2196  * RETURNS: true for link up or false for link down
2197  */
2198 bool ntb_transport_link_query(struct ntb_transport_qp *qp)
2199 {
2200 	if (!qp)
2201 		return false;
2202 
2203 	return qp->link_is_up;
2204 }
2205 EXPORT_SYMBOL_GPL(ntb_transport_link_query);
2206 
2207 /**
2208  * ntb_transport_qp_num - Query the qp number
2209  * @qp: NTB transport layer queue to be queried
2210  *
2211  * Query qp number of the NTB transport queue
2212  *
2213  * RETURNS: a zero based number specifying the qp number
2214  */
2215 unsigned char ntb_transport_qp_num(struct ntb_transport_qp *qp)
2216 {
2217 	if (!qp)
2218 		return 0;
2219 
2220 	return qp->qp_num;
2221 }
2222 EXPORT_SYMBOL_GPL(ntb_transport_qp_num);
2223 
2224 /**
2225  * ntb_transport_max_size - Query the max payload size of a qp
2226  * @qp: NTB transport layer queue to be queried
2227  *
2228  * Query the maximum payload size permissible on the given qp
2229  *
2230  * RETURNS: the max payload size of a qp
2231  */
2232 unsigned int ntb_transport_max_size(struct ntb_transport_qp *qp)
2233 {
2234 	unsigned int max_size;
2235 	unsigned int copy_align;
2236 	struct dma_chan *rx_chan, *tx_chan;
2237 
2238 	if (!qp)
2239 		return 0;
2240 
2241 	rx_chan = qp->rx_dma_chan;
2242 	tx_chan = qp->tx_dma_chan;
2243 
2244 	copy_align = max(rx_chan ? rx_chan->device->copy_align : 0,
2245 			 tx_chan ? tx_chan->device->copy_align : 0);
2246 
2247 	/* If DMA engine usage is possible, try to find the max size for that */
2248 	max_size = qp->tx_max_frame - sizeof(struct ntb_payload_header);
2249 	max_size = round_down(max_size, 1 << copy_align);
2250 
2251 	return max_size;
2252 }
2253 EXPORT_SYMBOL_GPL(ntb_transport_max_size);
2254 
2255 unsigned int ntb_transport_tx_free_entry(struct ntb_transport_qp *qp)
2256 {
2257 	unsigned int head = qp->tx_index;
2258 	unsigned int tail = qp->remote_rx_info->entry;
2259 
2260 	return tail > head ? tail - head : qp->tx_max_entry + tail - head;
2261 }
2262 EXPORT_SYMBOL_GPL(ntb_transport_tx_free_entry);
2263 
2264 static void ntb_transport_doorbell_callback(void *data, int vector)
2265 {
2266 	struct ntb_transport_ctx *nt = data;
2267 	struct ntb_transport_qp *qp;
2268 	u64 db_bits;
2269 	unsigned int qp_num;
2270 
2271 	db_bits = (nt->qp_bitmap & ~nt->qp_bitmap_free &
2272 		   ntb_db_vector_mask(nt->ndev, vector));
2273 
2274 	while (db_bits) {
2275 		qp_num = __ffs(db_bits);
2276 		qp = &nt->qp_vec[qp_num];
2277 
2278 		if (qp->active)
2279 			tasklet_schedule(&qp->rxc_db_work);
2280 
2281 		db_bits &= ~BIT_ULL(qp_num);
2282 	}
2283 }
2284 
2285 static const struct ntb_ctx_ops ntb_transport_ops = {
2286 	.link_event = ntb_transport_event_callback,
2287 	.db_event = ntb_transport_doorbell_callback,
2288 };
2289 
2290 static struct ntb_client ntb_transport_client = {
2291 	.ops = {
2292 		.probe = ntb_transport_probe,
2293 		.remove = ntb_transport_free,
2294 	},
2295 };
2296 
2297 static int __init ntb_transport_init(void)
2298 {
2299 	int rc;
2300 
2301 	pr_info("%s, version %s\n", NTB_TRANSPORT_DESC, NTB_TRANSPORT_VER);
2302 
2303 	if (debugfs_initialized())
2304 		nt_debugfs_dir = debugfs_create_dir(KBUILD_MODNAME, NULL);
2305 
2306 	rc = bus_register(&ntb_transport_bus);
2307 	if (rc)
2308 		goto err_bus;
2309 
2310 	rc = ntb_register_client(&ntb_transport_client);
2311 	if (rc)
2312 		goto err_client;
2313 
2314 	return 0;
2315 
2316 err_client:
2317 	bus_unregister(&ntb_transport_bus);
2318 err_bus:
2319 	debugfs_remove_recursive(nt_debugfs_dir);
2320 	return rc;
2321 }
2322 module_init(ntb_transport_init);
2323 
2324 static void __exit ntb_transport_exit(void)
2325 {
2326 	ntb_unregister_client(&ntb_transport_client);
2327 	bus_unregister(&ntb_transport_bus);
2328 	debugfs_remove_recursive(nt_debugfs_dir);
2329 }
2330 module_exit(ntb_transport_exit);
2331