xref: /openbmc/linux/drivers/ntb/ntb_transport.c (revision a36954f5)
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 struct ntb_queue_entry {
99 	/* ntb_queue list reference */
100 	struct list_head entry;
101 	/* pointers to data to be transferred */
102 	void *cb_data;
103 	void *buf;
104 	unsigned int len;
105 	unsigned int flags;
106 	int retries;
107 	int errors;
108 	unsigned int tx_index;
109 	unsigned int rx_index;
110 
111 	struct ntb_transport_qp *qp;
112 	union {
113 		struct ntb_payload_header __iomem *tx_hdr;
114 		struct ntb_payload_header *rx_hdr;
115 	};
116 };
117 
118 struct ntb_rx_info {
119 	unsigned int entry;
120 };
121 
122 struct ntb_transport_qp {
123 	struct ntb_transport_ctx *transport;
124 	struct ntb_dev *ndev;
125 	void *cb_data;
126 	struct dma_chan *tx_dma_chan;
127 	struct dma_chan *rx_dma_chan;
128 
129 	bool client_ready;
130 	bool link_is_up;
131 	bool active;
132 
133 	u8 qp_num;	/* Only 64 QP's are allowed.  0-63 */
134 	u64 qp_bit;
135 
136 	struct ntb_rx_info __iomem *rx_info;
137 	struct ntb_rx_info *remote_rx_info;
138 
139 	void (*tx_handler)(struct ntb_transport_qp *qp, void *qp_data,
140 			   void *data, int len);
141 	struct list_head tx_free_q;
142 	spinlock_t ntb_tx_free_q_lock;
143 	void __iomem *tx_mw;
144 	dma_addr_t tx_mw_phys;
145 	unsigned int tx_index;
146 	unsigned int tx_max_entry;
147 	unsigned int tx_max_frame;
148 
149 	void (*rx_handler)(struct ntb_transport_qp *qp, void *qp_data,
150 			   void *data, int len);
151 	struct list_head rx_post_q;
152 	struct list_head rx_pend_q;
153 	struct list_head rx_free_q;
154 	/* ntb_rx_q_lock: synchronize access to rx_XXXX_q */
155 	spinlock_t ntb_rx_q_lock;
156 	void *rx_buff;
157 	unsigned int rx_index;
158 	unsigned int rx_max_entry;
159 	unsigned int rx_max_frame;
160 	unsigned int rx_alloc_entry;
161 	dma_cookie_t last_cookie;
162 	struct tasklet_struct rxc_db_work;
163 
164 	void (*event_handler)(void *data, int status);
165 	struct delayed_work link_work;
166 	struct work_struct link_cleanup;
167 
168 	struct dentry *debugfs_dir;
169 	struct dentry *debugfs_stats;
170 
171 	/* Stats */
172 	u64 rx_bytes;
173 	u64 rx_pkts;
174 	u64 rx_ring_empty;
175 	u64 rx_err_no_buf;
176 	u64 rx_err_oflow;
177 	u64 rx_err_ver;
178 	u64 rx_memcpy;
179 	u64 rx_async;
180 	u64 dma_rx_prep_err;
181 	u64 tx_bytes;
182 	u64 tx_pkts;
183 	u64 tx_ring_full;
184 	u64 tx_err_no_buf;
185 	u64 tx_memcpy;
186 	u64 tx_async;
187 	u64 dma_tx_prep_err;
188 };
189 
190 struct ntb_transport_mw {
191 	phys_addr_t phys_addr;
192 	resource_size_t phys_size;
193 	resource_size_t xlat_align;
194 	resource_size_t xlat_align_size;
195 	void __iomem *vbase;
196 	size_t xlat_size;
197 	size_t buff_size;
198 	void *virt_addr;
199 	dma_addr_t dma_addr;
200 };
201 
202 struct ntb_transport_client_dev {
203 	struct list_head entry;
204 	struct ntb_transport_ctx *nt;
205 	struct device dev;
206 };
207 
208 struct ntb_transport_ctx {
209 	struct list_head entry;
210 	struct list_head client_devs;
211 
212 	struct ntb_dev *ndev;
213 
214 	struct ntb_transport_mw *mw_vec;
215 	struct ntb_transport_qp *qp_vec;
216 	unsigned int mw_count;
217 	unsigned int qp_count;
218 	u64 qp_bitmap;
219 	u64 qp_bitmap_free;
220 
221 	bool link_is_up;
222 	struct delayed_work link_work;
223 	struct work_struct link_cleanup;
224 
225 	struct dentry *debugfs_node_dir;
226 };
227 
228 enum {
229 	DESC_DONE_FLAG = BIT(0),
230 	LINK_DOWN_FLAG = BIT(1),
231 };
232 
233 struct ntb_payload_header {
234 	unsigned int ver;
235 	unsigned int len;
236 	unsigned int flags;
237 };
238 
239 enum {
240 	VERSION = 0,
241 	QP_LINKS,
242 	NUM_QPS,
243 	NUM_MWS,
244 	MW0_SZ_HIGH,
245 	MW0_SZ_LOW,
246 };
247 
248 #define dev_client_dev(__dev) \
249 	container_of((__dev), struct ntb_transport_client_dev, dev)
250 
251 #define drv_client(__drv) \
252 	container_of((__drv), struct ntb_transport_client, driver)
253 
254 #define QP_TO_MW(nt, qp)	((qp) % nt->mw_count)
255 #define NTB_QP_DEF_NUM_ENTRIES	100
256 #define NTB_LINK_DOWN_TIMEOUT	10
257 #define DMA_RETRIES		20
258 #define DMA_OUT_RESOURCE_TO	msecs_to_jiffies(50)
259 
260 static void ntb_transport_rxc_db(unsigned long data);
261 static const struct ntb_ctx_ops ntb_transport_ops;
262 static struct ntb_client ntb_transport_client;
263 static int ntb_async_tx_submit(struct ntb_transport_qp *qp,
264 			       struct ntb_queue_entry *entry);
265 static void ntb_memcpy_tx(struct ntb_queue_entry *entry, void __iomem *offset);
266 static int ntb_async_rx_submit(struct ntb_queue_entry *entry, void *offset);
267 static void ntb_memcpy_rx(struct ntb_queue_entry *entry, void *offset);
268 
269 
270 static int ntb_transport_bus_match(struct device *dev,
271 				   struct device_driver *drv)
272 {
273 	return !strncmp(dev_name(dev), drv->name, strlen(drv->name));
274 }
275 
276 static int ntb_transport_bus_probe(struct device *dev)
277 {
278 	const struct ntb_transport_client *client;
279 	int rc = -EINVAL;
280 
281 	get_device(dev);
282 
283 	client = drv_client(dev->driver);
284 	rc = client->probe(dev);
285 	if (rc)
286 		put_device(dev);
287 
288 	return rc;
289 }
290 
291 static int ntb_transport_bus_remove(struct device *dev)
292 {
293 	const struct ntb_transport_client *client;
294 
295 	client = drv_client(dev->driver);
296 	client->remove(dev);
297 
298 	put_device(dev);
299 
300 	return 0;
301 }
302 
303 static struct bus_type ntb_transport_bus = {
304 	.name = "ntb_transport",
305 	.match = ntb_transport_bus_match,
306 	.probe = ntb_transport_bus_probe,
307 	.remove = ntb_transport_bus_remove,
308 };
309 
310 static LIST_HEAD(ntb_transport_list);
311 
312 static int ntb_bus_init(struct ntb_transport_ctx *nt)
313 {
314 	list_add_tail(&nt->entry, &ntb_transport_list);
315 	return 0;
316 }
317 
318 static void ntb_bus_remove(struct ntb_transport_ctx *nt)
319 {
320 	struct ntb_transport_client_dev *client_dev, *cd;
321 
322 	list_for_each_entry_safe(client_dev, cd, &nt->client_devs, entry) {
323 		dev_err(client_dev->dev.parent, "%s still attached to bus, removing\n",
324 			dev_name(&client_dev->dev));
325 		list_del(&client_dev->entry);
326 		device_unregister(&client_dev->dev);
327 	}
328 
329 	list_del(&nt->entry);
330 }
331 
332 static void ntb_transport_client_release(struct device *dev)
333 {
334 	struct ntb_transport_client_dev *client_dev;
335 
336 	client_dev = dev_client_dev(dev);
337 	kfree(client_dev);
338 }
339 
340 /**
341  * ntb_transport_unregister_client_dev - Unregister NTB client device
342  * @device_name: Name of NTB client device
343  *
344  * Unregister an NTB client device with the NTB transport layer
345  */
346 void ntb_transport_unregister_client_dev(char *device_name)
347 {
348 	struct ntb_transport_client_dev *client, *cd;
349 	struct ntb_transport_ctx *nt;
350 
351 	list_for_each_entry(nt, &ntb_transport_list, entry)
352 		list_for_each_entry_safe(client, cd, &nt->client_devs, entry)
353 			if (!strncmp(dev_name(&client->dev), device_name,
354 				     strlen(device_name))) {
355 				list_del(&client->entry);
356 				device_unregister(&client->dev);
357 			}
358 }
359 EXPORT_SYMBOL_GPL(ntb_transport_unregister_client_dev);
360 
361 /**
362  * ntb_transport_register_client_dev - Register NTB client device
363  * @device_name: Name of NTB client device
364  *
365  * Register an NTB client device with the NTB transport layer
366  */
367 int ntb_transport_register_client_dev(char *device_name)
368 {
369 	struct ntb_transport_client_dev *client_dev;
370 	struct ntb_transport_ctx *nt;
371 	int node;
372 	int rc, i = 0;
373 
374 	if (list_empty(&ntb_transport_list))
375 		return -ENODEV;
376 
377 	list_for_each_entry(nt, &ntb_transport_list, entry) {
378 		struct device *dev;
379 
380 		node = dev_to_node(&nt->ndev->dev);
381 
382 		client_dev = kzalloc_node(sizeof(*client_dev),
383 					  GFP_KERNEL, node);
384 		if (!client_dev) {
385 			rc = -ENOMEM;
386 			goto err;
387 		}
388 
389 		dev = &client_dev->dev;
390 
391 		/* setup and register client devices */
392 		dev_set_name(dev, "%s%d", device_name, i);
393 		dev->bus = &ntb_transport_bus;
394 		dev->release = ntb_transport_client_release;
395 		dev->parent = &nt->ndev->dev;
396 
397 		rc = device_register(dev);
398 		if (rc) {
399 			kfree(client_dev);
400 			goto err;
401 		}
402 
403 		list_add_tail(&client_dev->entry, &nt->client_devs);
404 		i++;
405 	}
406 
407 	return 0;
408 
409 err:
410 	ntb_transport_unregister_client_dev(device_name);
411 
412 	return rc;
413 }
414 EXPORT_SYMBOL_GPL(ntb_transport_register_client_dev);
415 
416 /**
417  * ntb_transport_register_client - Register NTB client driver
418  * @drv: NTB client driver to be registered
419  *
420  * Register an NTB client driver with the NTB transport layer
421  *
422  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
423  */
424 int ntb_transport_register_client(struct ntb_transport_client *drv)
425 {
426 	drv->driver.bus = &ntb_transport_bus;
427 
428 	if (list_empty(&ntb_transport_list))
429 		return -ENODEV;
430 
431 	return driver_register(&drv->driver);
432 }
433 EXPORT_SYMBOL_GPL(ntb_transport_register_client);
434 
435 /**
436  * ntb_transport_unregister_client - Unregister NTB client driver
437  * @drv: NTB client driver to be unregistered
438  *
439  * Unregister an NTB client driver with the NTB transport layer
440  *
441  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
442  */
443 void ntb_transport_unregister_client(struct ntb_transport_client *drv)
444 {
445 	driver_unregister(&drv->driver);
446 }
447 EXPORT_SYMBOL_GPL(ntb_transport_unregister_client);
448 
449 static ssize_t debugfs_read(struct file *filp, char __user *ubuf, size_t count,
450 			    loff_t *offp)
451 {
452 	struct ntb_transport_qp *qp;
453 	char *buf;
454 	ssize_t ret, out_offset, out_count;
455 
456 	qp = filp->private_data;
457 
458 	if (!qp || !qp->link_is_up)
459 		return 0;
460 
461 	out_count = 1000;
462 
463 	buf = kmalloc(out_count, GFP_KERNEL);
464 	if (!buf)
465 		return -ENOMEM;
466 
467 	out_offset = 0;
468 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
469 			       "\nNTB QP stats:\n\n");
470 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
471 			       "rx_bytes - \t%llu\n", qp->rx_bytes);
472 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
473 			       "rx_pkts - \t%llu\n", qp->rx_pkts);
474 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
475 			       "rx_memcpy - \t%llu\n", qp->rx_memcpy);
476 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
477 			       "rx_async - \t%llu\n", qp->rx_async);
478 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
479 			       "rx_ring_empty - %llu\n", qp->rx_ring_empty);
480 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
481 			       "rx_err_no_buf - %llu\n", qp->rx_err_no_buf);
482 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
483 			       "rx_err_oflow - \t%llu\n", qp->rx_err_oflow);
484 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
485 			       "rx_err_ver - \t%llu\n", qp->rx_err_ver);
486 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
487 			       "rx_buff - \t0x%p\n", qp->rx_buff);
488 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
489 			       "rx_index - \t%u\n", qp->rx_index);
490 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
491 			       "rx_max_entry - \t%u\n", qp->rx_max_entry);
492 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
493 			       "rx_alloc_entry - \t%u\n\n", qp->rx_alloc_entry);
494 
495 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
496 			       "tx_bytes - \t%llu\n", qp->tx_bytes);
497 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
498 			       "tx_pkts - \t%llu\n", qp->tx_pkts);
499 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
500 			       "tx_memcpy - \t%llu\n", qp->tx_memcpy);
501 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
502 			       "tx_async - \t%llu\n", qp->tx_async);
503 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
504 			       "tx_ring_full - \t%llu\n", qp->tx_ring_full);
505 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
506 			       "tx_err_no_buf - %llu\n", qp->tx_err_no_buf);
507 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
508 			       "tx_mw - \t0x%p\n", qp->tx_mw);
509 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
510 			       "tx_index (H) - \t%u\n", qp->tx_index);
511 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
512 			       "RRI (T) - \t%u\n",
513 			       qp->remote_rx_info->entry);
514 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
515 			       "tx_max_entry - \t%u\n", qp->tx_max_entry);
516 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
517 			       "free tx - \t%u\n",
518 			       ntb_transport_tx_free_entry(qp));
519 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
520 			       "DMA tx prep err - \t%llu\n",
521 			       qp->dma_tx_prep_err);
522 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
523 			       "DMA rx prep err - \t%llu\n",
524 			       qp->dma_rx_prep_err);
525 
526 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
527 			       "\n");
528 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
529 			       "Using TX DMA - \t%s\n",
530 			       qp->tx_dma_chan ? "Yes" : "No");
531 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
532 			       "Using RX DMA - \t%s\n",
533 			       qp->rx_dma_chan ? "Yes" : "No");
534 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
535 			       "QP Link - \t%s\n",
536 			       qp->link_is_up ? "Up" : "Down");
537 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
538 			       "\n");
539 
540 	if (out_offset > out_count)
541 		out_offset = out_count;
542 
543 	ret = simple_read_from_buffer(ubuf, count, offp, buf, out_offset);
544 	kfree(buf);
545 	return ret;
546 }
547 
548 static const struct file_operations ntb_qp_debugfs_stats = {
549 	.owner = THIS_MODULE,
550 	.open = simple_open,
551 	.read = debugfs_read,
552 };
553 
554 static void ntb_list_add(spinlock_t *lock, struct list_head *entry,
555 			 struct list_head *list)
556 {
557 	unsigned long flags;
558 
559 	spin_lock_irqsave(lock, flags);
560 	list_add_tail(entry, list);
561 	spin_unlock_irqrestore(lock, flags);
562 }
563 
564 static struct ntb_queue_entry *ntb_list_rm(spinlock_t *lock,
565 					   struct list_head *list)
566 {
567 	struct ntb_queue_entry *entry;
568 	unsigned long flags;
569 
570 	spin_lock_irqsave(lock, flags);
571 	if (list_empty(list)) {
572 		entry = NULL;
573 		goto out;
574 	}
575 	entry = list_first_entry(list, struct ntb_queue_entry, entry);
576 	list_del(&entry->entry);
577 
578 out:
579 	spin_unlock_irqrestore(lock, flags);
580 
581 	return entry;
582 }
583 
584 static struct ntb_queue_entry *ntb_list_mv(spinlock_t *lock,
585 					   struct list_head *list,
586 					   struct list_head *to_list)
587 {
588 	struct ntb_queue_entry *entry;
589 	unsigned long flags;
590 
591 	spin_lock_irqsave(lock, flags);
592 
593 	if (list_empty(list)) {
594 		entry = NULL;
595 	} else {
596 		entry = list_first_entry(list, struct ntb_queue_entry, entry);
597 		list_move_tail(&entry->entry, to_list);
598 	}
599 
600 	spin_unlock_irqrestore(lock, flags);
601 
602 	return entry;
603 }
604 
605 static int ntb_transport_setup_qp_mw(struct ntb_transport_ctx *nt,
606 				     unsigned int qp_num)
607 {
608 	struct ntb_transport_qp *qp = &nt->qp_vec[qp_num];
609 	struct ntb_transport_mw *mw;
610 	struct ntb_dev *ndev = nt->ndev;
611 	struct ntb_queue_entry *entry;
612 	unsigned int rx_size, num_qps_mw;
613 	unsigned int mw_num, mw_count, qp_count;
614 	unsigned int i;
615 	int node;
616 
617 	mw_count = nt->mw_count;
618 	qp_count = nt->qp_count;
619 
620 	mw_num = QP_TO_MW(nt, qp_num);
621 	mw = &nt->mw_vec[mw_num];
622 
623 	if (!mw->virt_addr)
624 		return -ENOMEM;
625 
626 	if (qp_count % mw_count && mw_num + 1 < qp_count / mw_count)
627 		num_qps_mw = qp_count / mw_count + 1;
628 	else
629 		num_qps_mw = qp_count / mw_count;
630 
631 	rx_size = (unsigned int)mw->xlat_size / num_qps_mw;
632 	qp->rx_buff = mw->virt_addr + rx_size * (qp_num / mw_count);
633 	rx_size -= sizeof(struct ntb_rx_info);
634 
635 	qp->remote_rx_info = qp->rx_buff + rx_size;
636 
637 	/* Due to housekeeping, there must be atleast 2 buffs */
638 	qp->rx_max_frame = min(transport_mtu, rx_size / 2);
639 	qp->rx_max_entry = rx_size / qp->rx_max_frame;
640 	qp->rx_index = 0;
641 
642 	/*
643 	 * Checking to see if we have more entries than the default.
644 	 * We should add additional entries if that is the case so we
645 	 * can be in sync with the transport frames.
646 	 */
647 	node = dev_to_node(&ndev->dev);
648 	for (i = qp->rx_alloc_entry; i < qp->rx_max_entry; i++) {
649 		entry = kzalloc_node(sizeof(*entry), GFP_ATOMIC, node);
650 		if (!entry)
651 			return -ENOMEM;
652 
653 		entry->qp = qp;
654 		ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry,
655 			     &qp->rx_free_q);
656 		qp->rx_alloc_entry++;
657 	}
658 
659 	qp->remote_rx_info->entry = qp->rx_max_entry - 1;
660 
661 	/* setup the hdr offsets with 0's */
662 	for (i = 0; i < qp->rx_max_entry; i++) {
663 		void *offset = (qp->rx_buff + qp->rx_max_frame * (i + 1) -
664 				sizeof(struct ntb_payload_header));
665 		memset(offset, 0, sizeof(struct ntb_payload_header));
666 	}
667 
668 	qp->rx_pkts = 0;
669 	qp->tx_pkts = 0;
670 	qp->tx_index = 0;
671 
672 	return 0;
673 }
674 
675 static void ntb_free_mw(struct ntb_transport_ctx *nt, int num_mw)
676 {
677 	struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
678 	struct pci_dev *pdev = nt->ndev->pdev;
679 
680 	if (!mw->virt_addr)
681 		return;
682 
683 	ntb_mw_clear_trans(nt->ndev, num_mw);
684 	dma_free_coherent(&pdev->dev, mw->buff_size,
685 			  mw->virt_addr, mw->dma_addr);
686 	mw->xlat_size = 0;
687 	mw->buff_size = 0;
688 	mw->virt_addr = NULL;
689 }
690 
691 static int ntb_set_mw(struct ntb_transport_ctx *nt, int num_mw,
692 		      resource_size_t size)
693 {
694 	struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
695 	struct pci_dev *pdev = nt->ndev->pdev;
696 	size_t xlat_size, buff_size;
697 	int rc;
698 
699 	if (!size)
700 		return -EINVAL;
701 
702 	xlat_size = round_up(size, mw->xlat_align_size);
703 	buff_size = round_up(size, mw->xlat_align);
704 
705 	/* No need to re-setup */
706 	if (mw->xlat_size == xlat_size)
707 		return 0;
708 
709 	if (mw->buff_size)
710 		ntb_free_mw(nt, num_mw);
711 
712 	/* Alloc memory for receiving data.  Must be aligned */
713 	mw->xlat_size = xlat_size;
714 	mw->buff_size = buff_size;
715 
716 	mw->virt_addr = dma_alloc_coherent(&pdev->dev, buff_size,
717 					   &mw->dma_addr, GFP_KERNEL);
718 	if (!mw->virt_addr) {
719 		mw->xlat_size = 0;
720 		mw->buff_size = 0;
721 		dev_err(&pdev->dev, "Unable to alloc MW buff of size %zu\n",
722 			buff_size);
723 		return -ENOMEM;
724 	}
725 
726 	/*
727 	 * we must ensure that the memory address allocated is BAR size
728 	 * aligned in order for the XLAT register to take the value. This
729 	 * is a requirement of the hardware. It is recommended to setup CMA
730 	 * for BAR sizes equal or greater than 4MB.
731 	 */
732 	if (!IS_ALIGNED(mw->dma_addr, mw->xlat_align)) {
733 		dev_err(&pdev->dev, "DMA memory %pad is not aligned\n",
734 			&mw->dma_addr);
735 		ntb_free_mw(nt, num_mw);
736 		return -ENOMEM;
737 	}
738 
739 	/* Notify HW the memory location of the receive buffer */
740 	rc = ntb_mw_set_trans(nt->ndev, num_mw, mw->dma_addr, mw->xlat_size);
741 	if (rc) {
742 		dev_err(&pdev->dev, "Unable to set mw%d translation", num_mw);
743 		ntb_free_mw(nt, num_mw);
744 		return -EIO;
745 	}
746 
747 	return 0;
748 }
749 
750 static void ntb_qp_link_down_reset(struct ntb_transport_qp *qp)
751 {
752 	qp->link_is_up = false;
753 	qp->active = false;
754 
755 	qp->tx_index = 0;
756 	qp->rx_index = 0;
757 	qp->rx_bytes = 0;
758 	qp->rx_pkts = 0;
759 	qp->rx_ring_empty = 0;
760 	qp->rx_err_no_buf = 0;
761 	qp->rx_err_oflow = 0;
762 	qp->rx_err_ver = 0;
763 	qp->rx_memcpy = 0;
764 	qp->rx_async = 0;
765 	qp->tx_bytes = 0;
766 	qp->tx_pkts = 0;
767 	qp->tx_ring_full = 0;
768 	qp->tx_err_no_buf = 0;
769 	qp->tx_memcpy = 0;
770 	qp->tx_async = 0;
771 	qp->dma_tx_prep_err = 0;
772 	qp->dma_rx_prep_err = 0;
773 }
774 
775 static void ntb_qp_link_cleanup(struct ntb_transport_qp *qp)
776 {
777 	struct ntb_transport_ctx *nt = qp->transport;
778 	struct pci_dev *pdev = nt->ndev->pdev;
779 
780 	dev_info(&pdev->dev, "qp %d: Link Cleanup\n", qp->qp_num);
781 
782 	cancel_delayed_work_sync(&qp->link_work);
783 	ntb_qp_link_down_reset(qp);
784 
785 	if (qp->event_handler)
786 		qp->event_handler(qp->cb_data, qp->link_is_up);
787 }
788 
789 static void ntb_qp_link_cleanup_work(struct work_struct *work)
790 {
791 	struct ntb_transport_qp *qp = container_of(work,
792 						   struct ntb_transport_qp,
793 						   link_cleanup);
794 	struct ntb_transport_ctx *nt = qp->transport;
795 
796 	ntb_qp_link_cleanup(qp);
797 
798 	if (nt->link_is_up)
799 		schedule_delayed_work(&qp->link_work,
800 				      msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
801 }
802 
803 static void ntb_qp_link_down(struct ntb_transport_qp *qp)
804 {
805 	schedule_work(&qp->link_cleanup);
806 }
807 
808 static void ntb_transport_link_cleanup(struct ntb_transport_ctx *nt)
809 {
810 	struct ntb_transport_qp *qp;
811 	u64 qp_bitmap_alloc;
812 	unsigned int i, count;
813 
814 	qp_bitmap_alloc = nt->qp_bitmap & ~nt->qp_bitmap_free;
815 
816 	/* Pass along the info to any clients */
817 	for (i = 0; i < nt->qp_count; i++)
818 		if (qp_bitmap_alloc & BIT_ULL(i)) {
819 			qp = &nt->qp_vec[i];
820 			ntb_qp_link_cleanup(qp);
821 			cancel_work_sync(&qp->link_cleanup);
822 			cancel_delayed_work_sync(&qp->link_work);
823 		}
824 
825 	if (!nt->link_is_up)
826 		cancel_delayed_work_sync(&nt->link_work);
827 
828 	/* The scratchpad registers keep the values if the remote side
829 	 * goes down, blast them now to give them a sane value the next
830 	 * time they are accessed
831 	 */
832 	count = ntb_spad_count(nt->ndev);
833 	for (i = 0; i < count; i++)
834 		ntb_spad_write(nt->ndev, i, 0);
835 }
836 
837 static void ntb_transport_link_cleanup_work(struct work_struct *work)
838 {
839 	struct ntb_transport_ctx *nt =
840 		container_of(work, struct ntb_transport_ctx, link_cleanup);
841 
842 	ntb_transport_link_cleanup(nt);
843 }
844 
845 static void ntb_transport_event_callback(void *data)
846 {
847 	struct ntb_transport_ctx *nt = data;
848 
849 	if (ntb_link_is_up(nt->ndev, NULL, NULL) == 1)
850 		schedule_delayed_work(&nt->link_work, 0);
851 	else
852 		schedule_work(&nt->link_cleanup);
853 }
854 
855 static void ntb_transport_link_work(struct work_struct *work)
856 {
857 	struct ntb_transport_ctx *nt =
858 		container_of(work, struct ntb_transport_ctx, link_work.work);
859 	struct ntb_dev *ndev = nt->ndev;
860 	struct pci_dev *pdev = ndev->pdev;
861 	resource_size_t size;
862 	u32 val;
863 	int rc = 0, i, spad;
864 
865 	/* send the local info, in the opposite order of the way we read it */
866 	for (i = 0; i < nt->mw_count; i++) {
867 		size = nt->mw_vec[i].phys_size;
868 
869 		if (max_mw_size && size > max_mw_size)
870 			size = max_mw_size;
871 
872 		spad = MW0_SZ_HIGH + (i * 2);
873 		ntb_peer_spad_write(ndev, spad, upper_32_bits(size));
874 
875 		spad = MW0_SZ_LOW + (i * 2);
876 		ntb_peer_spad_write(ndev, spad, lower_32_bits(size));
877 	}
878 
879 	ntb_peer_spad_write(ndev, NUM_MWS, nt->mw_count);
880 
881 	ntb_peer_spad_write(ndev, NUM_QPS, nt->qp_count);
882 
883 	ntb_peer_spad_write(ndev, VERSION, NTB_TRANSPORT_VERSION);
884 
885 	/* Query the remote side for its info */
886 	val = ntb_spad_read(ndev, VERSION);
887 	dev_dbg(&pdev->dev, "Remote version = %d\n", val);
888 	if (val != NTB_TRANSPORT_VERSION)
889 		goto out;
890 
891 	val = ntb_spad_read(ndev, NUM_QPS);
892 	dev_dbg(&pdev->dev, "Remote max number of qps = %d\n", val);
893 	if (val != nt->qp_count)
894 		goto out;
895 
896 	val = ntb_spad_read(ndev, NUM_MWS);
897 	dev_dbg(&pdev->dev, "Remote number of mws = %d\n", val);
898 	if (val != nt->mw_count)
899 		goto out;
900 
901 	for (i = 0; i < nt->mw_count; i++) {
902 		u64 val64;
903 
904 		val = ntb_spad_read(ndev, MW0_SZ_HIGH + (i * 2));
905 		val64 = (u64)val << 32;
906 
907 		val = ntb_spad_read(ndev, MW0_SZ_LOW + (i * 2));
908 		val64 |= val;
909 
910 		dev_dbg(&pdev->dev, "Remote MW%d size = %#llx\n", i, val64);
911 
912 		rc = ntb_set_mw(nt, i, val64);
913 		if (rc)
914 			goto out1;
915 	}
916 
917 	nt->link_is_up = true;
918 
919 	for (i = 0; i < nt->qp_count; i++) {
920 		struct ntb_transport_qp *qp = &nt->qp_vec[i];
921 
922 		ntb_transport_setup_qp_mw(nt, i);
923 
924 		if (qp->client_ready)
925 			schedule_delayed_work(&qp->link_work, 0);
926 	}
927 
928 	return;
929 
930 out1:
931 	for (i = 0; i < nt->mw_count; i++)
932 		ntb_free_mw(nt, i);
933 
934 	/* if there's an actual failure, we should just bail */
935 	if (rc < 0) {
936 		ntb_link_disable(ndev);
937 		return;
938 	}
939 
940 out:
941 	if (ntb_link_is_up(ndev, NULL, NULL) == 1)
942 		schedule_delayed_work(&nt->link_work,
943 				      msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
944 }
945 
946 static void ntb_qp_link_work(struct work_struct *work)
947 {
948 	struct ntb_transport_qp *qp = container_of(work,
949 						   struct ntb_transport_qp,
950 						   link_work.work);
951 	struct pci_dev *pdev = qp->ndev->pdev;
952 	struct ntb_transport_ctx *nt = qp->transport;
953 	int val;
954 
955 	WARN_ON(!nt->link_is_up);
956 
957 	val = ntb_spad_read(nt->ndev, QP_LINKS);
958 
959 	ntb_peer_spad_write(nt->ndev, QP_LINKS, val | BIT(qp->qp_num));
960 
961 	/* query remote spad for qp ready bits */
962 	dev_dbg_ratelimited(&pdev->dev, "Remote QP link status = %x\n", val);
963 
964 	/* See if the remote side is up */
965 	if (val & BIT(qp->qp_num)) {
966 		dev_info(&pdev->dev, "qp %d: Link Up\n", qp->qp_num);
967 		qp->link_is_up = true;
968 		qp->active = true;
969 
970 		if (qp->event_handler)
971 			qp->event_handler(qp->cb_data, qp->link_is_up);
972 
973 		if (qp->active)
974 			tasklet_schedule(&qp->rxc_db_work);
975 	} else if (nt->link_is_up)
976 		schedule_delayed_work(&qp->link_work,
977 				      msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
978 }
979 
980 static int ntb_transport_init_queue(struct ntb_transport_ctx *nt,
981 				    unsigned int qp_num)
982 {
983 	struct ntb_transport_qp *qp;
984 	phys_addr_t mw_base;
985 	resource_size_t mw_size;
986 	unsigned int num_qps_mw, tx_size;
987 	unsigned int mw_num, mw_count, qp_count;
988 	u64 qp_offset;
989 
990 	mw_count = nt->mw_count;
991 	qp_count = nt->qp_count;
992 
993 	mw_num = QP_TO_MW(nt, qp_num);
994 
995 	qp = &nt->qp_vec[qp_num];
996 	qp->qp_num = qp_num;
997 	qp->transport = nt;
998 	qp->ndev = nt->ndev;
999 	qp->client_ready = false;
1000 	qp->event_handler = NULL;
1001 	ntb_qp_link_down_reset(qp);
1002 
1003 	if (qp_count % mw_count && mw_num + 1 < qp_count / mw_count)
1004 		num_qps_mw = qp_count / mw_count + 1;
1005 	else
1006 		num_qps_mw = qp_count / mw_count;
1007 
1008 	mw_base = nt->mw_vec[mw_num].phys_addr;
1009 	mw_size = nt->mw_vec[mw_num].phys_size;
1010 
1011 	tx_size = (unsigned int)mw_size / num_qps_mw;
1012 	qp_offset = tx_size * (qp_num / mw_count);
1013 
1014 	qp->tx_mw = nt->mw_vec[mw_num].vbase + qp_offset;
1015 	if (!qp->tx_mw)
1016 		return -EINVAL;
1017 
1018 	qp->tx_mw_phys = mw_base + qp_offset;
1019 	if (!qp->tx_mw_phys)
1020 		return -EINVAL;
1021 
1022 	tx_size -= sizeof(struct ntb_rx_info);
1023 	qp->rx_info = qp->tx_mw + tx_size;
1024 
1025 	/* Due to housekeeping, there must be atleast 2 buffs */
1026 	qp->tx_max_frame = min(transport_mtu, tx_size / 2);
1027 	qp->tx_max_entry = tx_size / qp->tx_max_frame;
1028 
1029 	if (nt->debugfs_node_dir) {
1030 		char debugfs_name[4];
1031 
1032 		snprintf(debugfs_name, 4, "qp%d", qp_num);
1033 		qp->debugfs_dir = debugfs_create_dir(debugfs_name,
1034 						     nt->debugfs_node_dir);
1035 
1036 		qp->debugfs_stats = debugfs_create_file("stats", S_IRUSR,
1037 							qp->debugfs_dir, qp,
1038 							&ntb_qp_debugfs_stats);
1039 	} else {
1040 		qp->debugfs_dir = NULL;
1041 		qp->debugfs_stats = NULL;
1042 	}
1043 
1044 	INIT_DELAYED_WORK(&qp->link_work, ntb_qp_link_work);
1045 	INIT_WORK(&qp->link_cleanup, ntb_qp_link_cleanup_work);
1046 
1047 	spin_lock_init(&qp->ntb_rx_q_lock);
1048 	spin_lock_init(&qp->ntb_tx_free_q_lock);
1049 
1050 	INIT_LIST_HEAD(&qp->rx_post_q);
1051 	INIT_LIST_HEAD(&qp->rx_pend_q);
1052 	INIT_LIST_HEAD(&qp->rx_free_q);
1053 	INIT_LIST_HEAD(&qp->tx_free_q);
1054 
1055 	tasklet_init(&qp->rxc_db_work, ntb_transport_rxc_db,
1056 		     (unsigned long)qp);
1057 
1058 	return 0;
1059 }
1060 
1061 static int ntb_transport_probe(struct ntb_client *self, struct ntb_dev *ndev)
1062 {
1063 	struct ntb_transport_ctx *nt;
1064 	struct ntb_transport_mw *mw;
1065 	unsigned int mw_count, qp_count, spad_count, max_mw_count_for_spads;
1066 	u64 qp_bitmap;
1067 	int node;
1068 	int rc, i;
1069 
1070 	mw_count = ntb_mw_count(ndev);
1071 
1072 	if (ntb_db_is_unsafe(ndev))
1073 		dev_dbg(&ndev->dev,
1074 			"doorbell is unsafe, proceed anyway...\n");
1075 	if (ntb_spad_is_unsafe(ndev))
1076 		dev_dbg(&ndev->dev,
1077 			"scratchpad is unsafe, proceed anyway...\n");
1078 
1079 	node = dev_to_node(&ndev->dev);
1080 
1081 	nt = kzalloc_node(sizeof(*nt), GFP_KERNEL, node);
1082 	if (!nt)
1083 		return -ENOMEM;
1084 
1085 	nt->ndev = ndev;
1086 	spad_count = ntb_spad_count(ndev);
1087 
1088 	/* Limit the MW's based on the availability of scratchpads */
1089 
1090 	if (spad_count < NTB_TRANSPORT_MIN_SPADS) {
1091 		nt->mw_count = 0;
1092 		rc = -EINVAL;
1093 		goto err;
1094 	}
1095 
1096 	max_mw_count_for_spads = (spad_count - MW0_SZ_HIGH) / 2;
1097 	nt->mw_count = min(mw_count, max_mw_count_for_spads);
1098 
1099 	nt->mw_vec = kzalloc_node(mw_count * sizeof(*nt->mw_vec),
1100 				  GFP_KERNEL, node);
1101 	if (!nt->mw_vec) {
1102 		rc = -ENOMEM;
1103 		goto err;
1104 	}
1105 
1106 	for (i = 0; i < mw_count; i++) {
1107 		mw = &nt->mw_vec[i];
1108 
1109 		rc = ntb_mw_get_range(ndev, i, &mw->phys_addr, &mw->phys_size,
1110 				      &mw->xlat_align, &mw->xlat_align_size);
1111 		if (rc)
1112 			goto err1;
1113 
1114 		mw->vbase = ioremap_wc(mw->phys_addr, mw->phys_size);
1115 		if (!mw->vbase) {
1116 			rc = -ENOMEM;
1117 			goto err1;
1118 		}
1119 
1120 		mw->buff_size = 0;
1121 		mw->xlat_size = 0;
1122 		mw->virt_addr = NULL;
1123 		mw->dma_addr = 0;
1124 	}
1125 
1126 	qp_bitmap = ntb_db_valid_mask(ndev);
1127 
1128 	qp_count = ilog2(qp_bitmap);
1129 	if (max_num_clients && max_num_clients < qp_count)
1130 		qp_count = max_num_clients;
1131 	else if (mw_count < qp_count)
1132 		qp_count = mw_count;
1133 
1134 	qp_bitmap &= BIT_ULL(qp_count) - 1;
1135 
1136 	nt->qp_count = qp_count;
1137 	nt->qp_bitmap = qp_bitmap;
1138 	nt->qp_bitmap_free = qp_bitmap;
1139 
1140 	nt->qp_vec = kzalloc_node(qp_count * sizeof(*nt->qp_vec),
1141 				  GFP_KERNEL, node);
1142 	if (!nt->qp_vec) {
1143 		rc = -ENOMEM;
1144 		goto err1;
1145 	}
1146 
1147 	if (nt_debugfs_dir) {
1148 		nt->debugfs_node_dir =
1149 			debugfs_create_dir(pci_name(ndev->pdev),
1150 					   nt_debugfs_dir);
1151 	}
1152 
1153 	for (i = 0; i < qp_count; i++) {
1154 		rc = ntb_transport_init_queue(nt, i);
1155 		if (rc)
1156 			goto err2;
1157 	}
1158 
1159 	INIT_DELAYED_WORK(&nt->link_work, ntb_transport_link_work);
1160 	INIT_WORK(&nt->link_cleanup, ntb_transport_link_cleanup_work);
1161 
1162 	rc = ntb_set_ctx(ndev, nt, &ntb_transport_ops);
1163 	if (rc)
1164 		goto err2;
1165 
1166 	INIT_LIST_HEAD(&nt->client_devs);
1167 	rc = ntb_bus_init(nt);
1168 	if (rc)
1169 		goto err3;
1170 
1171 	nt->link_is_up = false;
1172 	ntb_link_enable(ndev, NTB_SPEED_AUTO, NTB_WIDTH_AUTO);
1173 	ntb_link_event(ndev);
1174 
1175 	return 0;
1176 
1177 err3:
1178 	ntb_clear_ctx(ndev);
1179 err2:
1180 	kfree(nt->qp_vec);
1181 err1:
1182 	while (i--) {
1183 		mw = &nt->mw_vec[i];
1184 		iounmap(mw->vbase);
1185 	}
1186 	kfree(nt->mw_vec);
1187 err:
1188 	kfree(nt);
1189 	return rc;
1190 }
1191 
1192 static void ntb_transport_free(struct ntb_client *self, struct ntb_dev *ndev)
1193 {
1194 	struct ntb_transport_ctx *nt = ndev->ctx;
1195 	struct ntb_transport_qp *qp;
1196 	u64 qp_bitmap_alloc;
1197 	int i;
1198 
1199 	ntb_transport_link_cleanup(nt);
1200 	cancel_work_sync(&nt->link_cleanup);
1201 	cancel_delayed_work_sync(&nt->link_work);
1202 
1203 	qp_bitmap_alloc = nt->qp_bitmap & ~nt->qp_bitmap_free;
1204 
1205 	/* verify that all the qp's are freed */
1206 	for (i = 0; i < nt->qp_count; i++) {
1207 		qp = &nt->qp_vec[i];
1208 		if (qp_bitmap_alloc & BIT_ULL(i))
1209 			ntb_transport_free_queue(qp);
1210 		debugfs_remove_recursive(qp->debugfs_dir);
1211 	}
1212 
1213 	ntb_link_disable(ndev);
1214 	ntb_clear_ctx(ndev);
1215 
1216 	ntb_bus_remove(nt);
1217 
1218 	for (i = nt->mw_count; i--; ) {
1219 		ntb_free_mw(nt, i);
1220 		iounmap(nt->mw_vec[i].vbase);
1221 	}
1222 
1223 	kfree(nt->qp_vec);
1224 	kfree(nt->mw_vec);
1225 	kfree(nt);
1226 }
1227 
1228 static void ntb_complete_rxc(struct ntb_transport_qp *qp)
1229 {
1230 	struct ntb_queue_entry *entry;
1231 	void *cb_data;
1232 	unsigned int len;
1233 	unsigned long irqflags;
1234 
1235 	spin_lock_irqsave(&qp->ntb_rx_q_lock, irqflags);
1236 
1237 	while (!list_empty(&qp->rx_post_q)) {
1238 		entry = list_first_entry(&qp->rx_post_q,
1239 					 struct ntb_queue_entry, entry);
1240 		if (!(entry->flags & DESC_DONE_FLAG))
1241 			break;
1242 
1243 		entry->rx_hdr->flags = 0;
1244 		iowrite32(entry->rx_index, &qp->rx_info->entry);
1245 
1246 		cb_data = entry->cb_data;
1247 		len = entry->len;
1248 
1249 		list_move_tail(&entry->entry, &qp->rx_free_q);
1250 
1251 		spin_unlock_irqrestore(&qp->ntb_rx_q_lock, irqflags);
1252 
1253 		if (qp->rx_handler && qp->client_ready)
1254 			qp->rx_handler(qp, qp->cb_data, cb_data, len);
1255 
1256 		spin_lock_irqsave(&qp->ntb_rx_q_lock, irqflags);
1257 	}
1258 
1259 	spin_unlock_irqrestore(&qp->ntb_rx_q_lock, irqflags);
1260 }
1261 
1262 static void ntb_rx_copy_callback(void *data,
1263 				 const struct dmaengine_result *res)
1264 {
1265 	struct ntb_queue_entry *entry = data;
1266 
1267 	/* we need to check DMA results if we are using DMA */
1268 	if (res) {
1269 		enum dmaengine_tx_result dma_err = res->result;
1270 
1271 		switch (dma_err) {
1272 		case DMA_TRANS_READ_FAILED:
1273 		case DMA_TRANS_WRITE_FAILED:
1274 			entry->errors++;
1275 		case DMA_TRANS_ABORTED:
1276 		{
1277 			struct ntb_transport_qp *qp = entry->qp;
1278 			void *offset = qp->rx_buff + qp->rx_max_frame *
1279 					qp->rx_index;
1280 
1281 			ntb_memcpy_rx(entry, offset);
1282 			qp->rx_memcpy++;
1283 			return;
1284 		}
1285 
1286 		case DMA_TRANS_NOERROR:
1287 		default:
1288 			break;
1289 		}
1290 	}
1291 
1292 	entry->flags |= DESC_DONE_FLAG;
1293 
1294 	ntb_complete_rxc(entry->qp);
1295 }
1296 
1297 static void ntb_memcpy_rx(struct ntb_queue_entry *entry, void *offset)
1298 {
1299 	void *buf = entry->buf;
1300 	size_t len = entry->len;
1301 
1302 	memcpy(buf, offset, len);
1303 
1304 	/* Ensure that the data is fully copied out before clearing the flag */
1305 	wmb();
1306 
1307 	ntb_rx_copy_callback(entry, NULL);
1308 }
1309 
1310 static int ntb_async_rx_submit(struct ntb_queue_entry *entry, void *offset)
1311 {
1312 	struct dma_async_tx_descriptor *txd;
1313 	struct ntb_transport_qp *qp = entry->qp;
1314 	struct dma_chan *chan = qp->rx_dma_chan;
1315 	struct dma_device *device;
1316 	size_t pay_off, buff_off, len;
1317 	struct dmaengine_unmap_data *unmap;
1318 	dma_cookie_t cookie;
1319 	void *buf = entry->buf;
1320 	int retries = 0;
1321 
1322 	len = entry->len;
1323 	device = chan->device;
1324 	pay_off = (size_t)offset & ~PAGE_MASK;
1325 	buff_off = (size_t)buf & ~PAGE_MASK;
1326 
1327 	if (!is_dma_copy_aligned(device, pay_off, buff_off, len))
1328 		goto err;
1329 
1330 	unmap = dmaengine_get_unmap_data(device->dev, 2, GFP_NOWAIT);
1331 	if (!unmap)
1332 		goto err;
1333 
1334 	unmap->len = len;
1335 	unmap->addr[0] = dma_map_page(device->dev, virt_to_page(offset),
1336 				      pay_off, len, DMA_TO_DEVICE);
1337 	if (dma_mapping_error(device->dev, unmap->addr[0]))
1338 		goto err_get_unmap;
1339 
1340 	unmap->to_cnt = 1;
1341 
1342 	unmap->addr[1] = dma_map_page(device->dev, virt_to_page(buf),
1343 				      buff_off, len, DMA_FROM_DEVICE);
1344 	if (dma_mapping_error(device->dev, unmap->addr[1]))
1345 		goto err_get_unmap;
1346 
1347 	unmap->from_cnt = 1;
1348 
1349 	for (retries = 0; retries < DMA_RETRIES; retries++) {
1350 		txd = device->device_prep_dma_memcpy(chan,
1351 						     unmap->addr[1],
1352 						     unmap->addr[0], len,
1353 						     DMA_PREP_INTERRUPT);
1354 		if (txd)
1355 			break;
1356 
1357 		set_current_state(TASK_INTERRUPTIBLE);
1358 		schedule_timeout(DMA_OUT_RESOURCE_TO);
1359 	}
1360 
1361 	if (!txd) {
1362 		qp->dma_rx_prep_err++;
1363 		goto err_get_unmap;
1364 	}
1365 
1366 	txd->callback_result = ntb_rx_copy_callback;
1367 	txd->callback_param = entry;
1368 	dma_set_unmap(txd, unmap);
1369 
1370 	cookie = dmaengine_submit(txd);
1371 	if (dma_submit_error(cookie))
1372 		goto err_set_unmap;
1373 
1374 	dmaengine_unmap_put(unmap);
1375 
1376 	qp->last_cookie = cookie;
1377 
1378 	qp->rx_async++;
1379 
1380 	return 0;
1381 
1382 err_set_unmap:
1383 	dmaengine_unmap_put(unmap);
1384 err_get_unmap:
1385 	dmaengine_unmap_put(unmap);
1386 err:
1387 	return -ENXIO;
1388 }
1389 
1390 static void ntb_async_rx(struct ntb_queue_entry *entry, void *offset)
1391 {
1392 	struct ntb_transport_qp *qp = entry->qp;
1393 	struct dma_chan *chan = qp->rx_dma_chan;
1394 	int res;
1395 
1396 	if (!chan)
1397 		goto err;
1398 
1399 	if (entry->len < copy_bytes)
1400 		goto err;
1401 
1402 	res = ntb_async_rx_submit(entry, offset);
1403 	if (res < 0)
1404 		goto err;
1405 
1406 	if (!entry->retries)
1407 		qp->rx_async++;
1408 
1409 	return;
1410 
1411 err:
1412 	ntb_memcpy_rx(entry, offset);
1413 	qp->rx_memcpy++;
1414 }
1415 
1416 static int ntb_process_rxc(struct ntb_transport_qp *qp)
1417 {
1418 	struct ntb_payload_header *hdr;
1419 	struct ntb_queue_entry *entry;
1420 	void *offset;
1421 
1422 	offset = qp->rx_buff + qp->rx_max_frame * qp->rx_index;
1423 	hdr = offset + qp->rx_max_frame - sizeof(struct ntb_payload_header);
1424 
1425 	dev_dbg(&qp->ndev->pdev->dev, "qp %d: RX ver %u len %d flags %x\n",
1426 		qp->qp_num, hdr->ver, hdr->len, hdr->flags);
1427 
1428 	if (!(hdr->flags & DESC_DONE_FLAG)) {
1429 		dev_dbg(&qp->ndev->pdev->dev, "done flag not set\n");
1430 		qp->rx_ring_empty++;
1431 		return -EAGAIN;
1432 	}
1433 
1434 	if (hdr->flags & LINK_DOWN_FLAG) {
1435 		dev_dbg(&qp->ndev->pdev->dev, "link down flag set\n");
1436 		ntb_qp_link_down(qp);
1437 		hdr->flags = 0;
1438 		return -EAGAIN;
1439 	}
1440 
1441 	if (hdr->ver != (u32)qp->rx_pkts) {
1442 		dev_dbg(&qp->ndev->pdev->dev,
1443 			"version mismatch, expected %llu - got %u\n",
1444 			qp->rx_pkts, hdr->ver);
1445 		qp->rx_err_ver++;
1446 		return -EIO;
1447 	}
1448 
1449 	entry = ntb_list_mv(&qp->ntb_rx_q_lock, &qp->rx_pend_q, &qp->rx_post_q);
1450 	if (!entry) {
1451 		dev_dbg(&qp->ndev->pdev->dev, "no receive buffer\n");
1452 		qp->rx_err_no_buf++;
1453 		return -EAGAIN;
1454 	}
1455 
1456 	entry->rx_hdr = hdr;
1457 	entry->rx_index = qp->rx_index;
1458 
1459 	if (hdr->len > entry->len) {
1460 		dev_dbg(&qp->ndev->pdev->dev,
1461 			"receive buffer overflow! Wanted %d got %d\n",
1462 			hdr->len, entry->len);
1463 		qp->rx_err_oflow++;
1464 
1465 		entry->len = -EIO;
1466 		entry->flags |= DESC_DONE_FLAG;
1467 
1468 		ntb_complete_rxc(qp);
1469 	} else {
1470 		dev_dbg(&qp->ndev->pdev->dev,
1471 			"RX OK index %u ver %u size %d into buf size %d\n",
1472 			qp->rx_index, hdr->ver, hdr->len, entry->len);
1473 
1474 		qp->rx_bytes += hdr->len;
1475 		qp->rx_pkts++;
1476 
1477 		entry->len = hdr->len;
1478 
1479 		ntb_async_rx(entry, offset);
1480 	}
1481 
1482 	qp->rx_index++;
1483 	qp->rx_index %= qp->rx_max_entry;
1484 
1485 	return 0;
1486 }
1487 
1488 static void ntb_transport_rxc_db(unsigned long data)
1489 {
1490 	struct ntb_transport_qp *qp = (void *)data;
1491 	int rc, i;
1492 
1493 	dev_dbg(&qp->ndev->pdev->dev, "%s: doorbell %d received\n",
1494 		__func__, qp->qp_num);
1495 
1496 	/* Limit the number of packets processed in a single interrupt to
1497 	 * provide fairness to others
1498 	 */
1499 	for (i = 0; i < qp->rx_max_entry; i++) {
1500 		rc = ntb_process_rxc(qp);
1501 		if (rc)
1502 			break;
1503 	}
1504 
1505 	if (i && qp->rx_dma_chan)
1506 		dma_async_issue_pending(qp->rx_dma_chan);
1507 
1508 	if (i == qp->rx_max_entry) {
1509 		/* there is more work to do */
1510 		if (qp->active)
1511 			tasklet_schedule(&qp->rxc_db_work);
1512 	} else if (ntb_db_read(qp->ndev) & BIT_ULL(qp->qp_num)) {
1513 		/* the doorbell bit is set: clear it */
1514 		ntb_db_clear(qp->ndev, BIT_ULL(qp->qp_num));
1515 		/* ntb_db_read ensures ntb_db_clear write is committed */
1516 		ntb_db_read(qp->ndev);
1517 
1518 		/* an interrupt may have arrived between finishing
1519 		 * ntb_process_rxc and clearing the doorbell bit:
1520 		 * there might be some more work to do.
1521 		 */
1522 		if (qp->active)
1523 			tasklet_schedule(&qp->rxc_db_work);
1524 	}
1525 }
1526 
1527 static void ntb_tx_copy_callback(void *data,
1528 				 const struct dmaengine_result *res)
1529 {
1530 	struct ntb_queue_entry *entry = data;
1531 	struct ntb_transport_qp *qp = entry->qp;
1532 	struct ntb_payload_header __iomem *hdr = entry->tx_hdr;
1533 
1534 	/* we need to check DMA results if we are using DMA */
1535 	if (res) {
1536 		enum dmaengine_tx_result dma_err = res->result;
1537 
1538 		switch (dma_err) {
1539 		case DMA_TRANS_READ_FAILED:
1540 		case DMA_TRANS_WRITE_FAILED:
1541 			entry->errors++;
1542 		case DMA_TRANS_ABORTED:
1543 		{
1544 			void __iomem *offset =
1545 				qp->tx_mw + qp->tx_max_frame *
1546 				entry->tx_index;
1547 
1548 			/* resubmit via CPU */
1549 			ntb_memcpy_tx(entry, offset);
1550 			qp->tx_memcpy++;
1551 			return;
1552 		}
1553 
1554 		case DMA_TRANS_NOERROR:
1555 		default:
1556 			break;
1557 		}
1558 	}
1559 
1560 	iowrite32(entry->flags | DESC_DONE_FLAG, &hdr->flags);
1561 
1562 	ntb_peer_db_set(qp->ndev, BIT_ULL(qp->qp_num));
1563 
1564 	/* The entry length can only be zero if the packet is intended to be a
1565 	 * "link down" or similar.  Since no payload is being sent in these
1566 	 * cases, there is nothing to add to the completion queue.
1567 	 */
1568 	if (entry->len > 0) {
1569 		qp->tx_bytes += entry->len;
1570 
1571 		if (qp->tx_handler)
1572 			qp->tx_handler(qp, qp->cb_data, entry->cb_data,
1573 				       entry->len);
1574 	}
1575 
1576 	ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry, &qp->tx_free_q);
1577 }
1578 
1579 static void ntb_memcpy_tx(struct ntb_queue_entry *entry, void __iomem *offset)
1580 {
1581 #ifdef ARCH_HAS_NOCACHE_UACCESS
1582 	/*
1583 	 * Using non-temporal mov to improve performance on non-cached
1584 	 * writes, even though we aren't actually copying from user space.
1585 	 */
1586 	__copy_from_user_inatomic_nocache(offset, entry->buf, entry->len);
1587 #else
1588 	memcpy_toio(offset, entry->buf, entry->len);
1589 #endif
1590 
1591 	/* Ensure that the data is fully copied out before setting the flags */
1592 	wmb();
1593 
1594 	ntb_tx_copy_callback(entry, NULL);
1595 }
1596 
1597 static int ntb_async_tx_submit(struct ntb_transport_qp *qp,
1598 			       struct ntb_queue_entry *entry)
1599 {
1600 	struct dma_async_tx_descriptor *txd;
1601 	struct dma_chan *chan = qp->tx_dma_chan;
1602 	struct dma_device *device;
1603 	size_t len = entry->len;
1604 	void *buf = entry->buf;
1605 	size_t dest_off, buff_off;
1606 	struct dmaengine_unmap_data *unmap;
1607 	dma_addr_t dest;
1608 	dma_cookie_t cookie;
1609 	int retries = 0;
1610 
1611 	device = chan->device;
1612 	dest = qp->tx_mw_phys + qp->tx_max_frame * entry->tx_index;
1613 	buff_off = (size_t)buf & ~PAGE_MASK;
1614 	dest_off = (size_t)dest & ~PAGE_MASK;
1615 
1616 	if (!is_dma_copy_aligned(device, buff_off, dest_off, len))
1617 		goto err;
1618 
1619 	unmap = dmaengine_get_unmap_data(device->dev, 1, GFP_NOWAIT);
1620 	if (!unmap)
1621 		goto err;
1622 
1623 	unmap->len = len;
1624 	unmap->addr[0] = dma_map_page(device->dev, virt_to_page(buf),
1625 				      buff_off, len, DMA_TO_DEVICE);
1626 	if (dma_mapping_error(device->dev, unmap->addr[0]))
1627 		goto err_get_unmap;
1628 
1629 	unmap->to_cnt = 1;
1630 
1631 	for (retries = 0; retries < DMA_RETRIES; retries++) {
1632 		txd = device->device_prep_dma_memcpy(chan, dest,
1633 						     unmap->addr[0], len,
1634 						     DMA_PREP_INTERRUPT);
1635 		if (txd)
1636 			break;
1637 
1638 		set_current_state(TASK_INTERRUPTIBLE);
1639 		schedule_timeout(DMA_OUT_RESOURCE_TO);
1640 	}
1641 
1642 	if (!txd) {
1643 		qp->dma_tx_prep_err++;
1644 		goto err_get_unmap;
1645 	}
1646 
1647 	txd->callback_result = ntb_tx_copy_callback;
1648 	txd->callback_param = entry;
1649 	dma_set_unmap(txd, unmap);
1650 
1651 	cookie = dmaengine_submit(txd);
1652 	if (dma_submit_error(cookie))
1653 		goto err_set_unmap;
1654 
1655 	dmaengine_unmap_put(unmap);
1656 
1657 	dma_async_issue_pending(chan);
1658 
1659 	return 0;
1660 err_set_unmap:
1661 	dmaengine_unmap_put(unmap);
1662 err_get_unmap:
1663 	dmaengine_unmap_put(unmap);
1664 err:
1665 	return -ENXIO;
1666 }
1667 
1668 static void ntb_async_tx(struct ntb_transport_qp *qp,
1669 			 struct ntb_queue_entry *entry)
1670 {
1671 	struct ntb_payload_header __iomem *hdr;
1672 	struct dma_chan *chan = qp->tx_dma_chan;
1673 	void __iomem *offset;
1674 	int res;
1675 
1676 	entry->tx_index = qp->tx_index;
1677 	offset = qp->tx_mw + qp->tx_max_frame * entry->tx_index;
1678 	hdr = offset + qp->tx_max_frame - sizeof(struct ntb_payload_header);
1679 	entry->tx_hdr = hdr;
1680 
1681 	iowrite32(entry->len, &hdr->len);
1682 	iowrite32((u32)qp->tx_pkts, &hdr->ver);
1683 
1684 	if (!chan)
1685 		goto err;
1686 
1687 	if (entry->len < copy_bytes)
1688 		goto err;
1689 
1690 	res = ntb_async_tx_submit(qp, entry);
1691 	if (res < 0)
1692 		goto err;
1693 
1694 	if (!entry->retries)
1695 		qp->tx_async++;
1696 
1697 	return;
1698 
1699 err:
1700 	ntb_memcpy_tx(entry, offset);
1701 	qp->tx_memcpy++;
1702 }
1703 
1704 static int ntb_process_tx(struct ntb_transport_qp *qp,
1705 			  struct ntb_queue_entry *entry)
1706 {
1707 	if (qp->tx_index == qp->remote_rx_info->entry) {
1708 		qp->tx_ring_full++;
1709 		return -EAGAIN;
1710 	}
1711 
1712 	if (entry->len > qp->tx_max_frame - sizeof(struct ntb_payload_header)) {
1713 		if (qp->tx_handler)
1714 			qp->tx_handler(qp, qp->cb_data, NULL, -EIO);
1715 
1716 		ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
1717 			     &qp->tx_free_q);
1718 		return 0;
1719 	}
1720 
1721 	ntb_async_tx(qp, entry);
1722 
1723 	qp->tx_index++;
1724 	qp->tx_index %= qp->tx_max_entry;
1725 
1726 	qp->tx_pkts++;
1727 
1728 	return 0;
1729 }
1730 
1731 static void ntb_send_link_down(struct ntb_transport_qp *qp)
1732 {
1733 	struct pci_dev *pdev = qp->ndev->pdev;
1734 	struct ntb_queue_entry *entry;
1735 	int i, rc;
1736 
1737 	if (!qp->link_is_up)
1738 		return;
1739 
1740 	dev_info(&pdev->dev, "qp %d: Send Link Down\n", qp->qp_num);
1741 
1742 	for (i = 0; i < NTB_LINK_DOWN_TIMEOUT; i++) {
1743 		entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
1744 		if (entry)
1745 			break;
1746 		msleep(100);
1747 	}
1748 
1749 	if (!entry)
1750 		return;
1751 
1752 	entry->cb_data = NULL;
1753 	entry->buf = NULL;
1754 	entry->len = 0;
1755 	entry->flags = LINK_DOWN_FLAG;
1756 
1757 	rc = ntb_process_tx(qp, entry);
1758 	if (rc)
1759 		dev_err(&pdev->dev, "ntb: QP%d unable to send linkdown msg\n",
1760 			qp->qp_num);
1761 
1762 	ntb_qp_link_down_reset(qp);
1763 }
1764 
1765 static bool ntb_dma_filter_fn(struct dma_chan *chan, void *node)
1766 {
1767 	return dev_to_node(&chan->dev->device) == (int)(unsigned long)node;
1768 }
1769 
1770 /**
1771  * ntb_transport_create_queue - Create a new NTB transport layer queue
1772  * @rx_handler: receive callback function
1773  * @tx_handler: transmit callback function
1774  * @event_handler: event callback function
1775  *
1776  * Create a new NTB transport layer queue and provide the queue with a callback
1777  * routine for both transmit and receive.  The receive callback routine will be
1778  * used to pass up data when the transport has received it on the queue.   The
1779  * transmit callback routine will be called when the transport has completed the
1780  * transmission of the data on the queue and the data is ready to be freed.
1781  *
1782  * RETURNS: pointer to newly created ntb_queue, NULL on error.
1783  */
1784 struct ntb_transport_qp *
1785 ntb_transport_create_queue(void *data, struct device *client_dev,
1786 			   const struct ntb_queue_handlers *handlers)
1787 {
1788 	struct ntb_dev *ndev;
1789 	struct pci_dev *pdev;
1790 	struct ntb_transport_ctx *nt;
1791 	struct ntb_queue_entry *entry;
1792 	struct ntb_transport_qp *qp;
1793 	u64 qp_bit;
1794 	unsigned int free_queue;
1795 	dma_cap_mask_t dma_mask;
1796 	int node;
1797 	int i;
1798 
1799 	ndev = dev_ntb(client_dev->parent);
1800 	pdev = ndev->pdev;
1801 	nt = ndev->ctx;
1802 
1803 	node = dev_to_node(&ndev->dev);
1804 
1805 	free_queue = ffs(nt->qp_bitmap_free);
1806 	if (!free_queue)
1807 		goto err;
1808 
1809 	/* decrement free_queue to make it zero based */
1810 	free_queue--;
1811 
1812 	qp = &nt->qp_vec[free_queue];
1813 	qp_bit = BIT_ULL(qp->qp_num);
1814 
1815 	nt->qp_bitmap_free &= ~qp_bit;
1816 
1817 	qp->cb_data = data;
1818 	qp->rx_handler = handlers->rx_handler;
1819 	qp->tx_handler = handlers->tx_handler;
1820 	qp->event_handler = handlers->event_handler;
1821 
1822 	dma_cap_zero(dma_mask);
1823 	dma_cap_set(DMA_MEMCPY, dma_mask);
1824 
1825 	if (use_dma) {
1826 		qp->tx_dma_chan =
1827 			dma_request_channel(dma_mask, ntb_dma_filter_fn,
1828 					    (void *)(unsigned long)node);
1829 		if (!qp->tx_dma_chan)
1830 			dev_info(&pdev->dev, "Unable to allocate TX DMA channel\n");
1831 
1832 		qp->rx_dma_chan =
1833 			dma_request_channel(dma_mask, ntb_dma_filter_fn,
1834 					    (void *)(unsigned long)node);
1835 		if (!qp->rx_dma_chan)
1836 			dev_info(&pdev->dev, "Unable to allocate RX DMA channel\n");
1837 	} else {
1838 		qp->tx_dma_chan = NULL;
1839 		qp->rx_dma_chan = NULL;
1840 	}
1841 
1842 	dev_dbg(&pdev->dev, "Using %s memcpy for TX\n",
1843 		qp->tx_dma_chan ? "DMA" : "CPU");
1844 
1845 	dev_dbg(&pdev->dev, "Using %s memcpy for RX\n",
1846 		qp->rx_dma_chan ? "DMA" : "CPU");
1847 
1848 	for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) {
1849 		entry = kzalloc_node(sizeof(*entry), GFP_ATOMIC, node);
1850 		if (!entry)
1851 			goto err1;
1852 
1853 		entry->qp = qp;
1854 		ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry,
1855 			     &qp->rx_free_q);
1856 	}
1857 	qp->rx_alloc_entry = NTB_QP_DEF_NUM_ENTRIES;
1858 
1859 	for (i = 0; i < qp->tx_max_entry; i++) {
1860 		entry = kzalloc_node(sizeof(*entry), GFP_ATOMIC, node);
1861 		if (!entry)
1862 			goto err2;
1863 
1864 		entry->qp = qp;
1865 		ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
1866 			     &qp->tx_free_q);
1867 	}
1868 
1869 	ntb_db_clear(qp->ndev, qp_bit);
1870 	ntb_db_clear_mask(qp->ndev, qp_bit);
1871 
1872 	dev_info(&pdev->dev, "NTB Transport QP %d created\n", qp->qp_num);
1873 
1874 	return qp;
1875 
1876 err2:
1877 	while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
1878 		kfree(entry);
1879 err1:
1880 	qp->rx_alloc_entry = 0;
1881 	while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q)))
1882 		kfree(entry);
1883 	if (qp->tx_dma_chan)
1884 		dma_release_channel(qp->tx_dma_chan);
1885 	if (qp->rx_dma_chan)
1886 		dma_release_channel(qp->rx_dma_chan);
1887 	nt->qp_bitmap_free |= qp_bit;
1888 err:
1889 	return NULL;
1890 }
1891 EXPORT_SYMBOL_GPL(ntb_transport_create_queue);
1892 
1893 /**
1894  * ntb_transport_free_queue - Frees NTB transport queue
1895  * @qp: NTB queue to be freed
1896  *
1897  * Frees NTB transport queue
1898  */
1899 void ntb_transport_free_queue(struct ntb_transport_qp *qp)
1900 {
1901 	struct pci_dev *pdev;
1902 	struct ntb_queue_entry *entry;
1903 	u64 qp_bit;
1904 
1905 	if (!qp)
1906 		return;
1907 
1908 	pdev = qp->ndev->pdev;
1909 
1910 	qp->active = false;
1911 
1912 	if (qp->tx_dma_chan) {
1913 		struct dma_chan *chan = qp->tx_dma_chan;
1914 		/* Putting the dma_chan to NULL will force any new traffic to be
1915 		 * processed by the CPU instead of the DAM engine
1916 		 */
1917 		qp->tx_dma_chan = NULL;
1918 
1919 		/* Try to be nice and wait for any queued DMA engine
1920 		 * transactions to process before smashing it with a rock
1921 		 */
1922 		dma_sync_wait(chan, qp->last_cookie);
1923 		dmaengine_terminate_all(chan);
1924 		dma_release_channel(chan);
1925 	}
1926 
1927 	if (qp->rx_dma_chan) {
1928 		struct dma_chan *chan = qp->rx_dma_chan;
1929 		/* Putting the dma_chan to NULL will force any new traffic to be
1930 		 * processed by the CPU instead of the DAM engine
1931 		 */
1932 		qp->rx_dma_chan = NULL;
1933 
1934 		/* Try to be nice and wait for any queued DMA engine
1935 		 * transactions to process before smashing it with a rock
1936 		 */
1937 		dma_sync_wait(chan, qp->last_cookie);
1938 		dmaengine_terminate_all(chan);
1939 		dma_release_channel(chan);
1940 	}
1941 
1942 	qp_bit = BIT_ULL(qp->qp_num);
1943 
1944 	ntb_db_set_mask(qp->ndev, qp_bit);
1945 	tasklet_kill(&qp->rxc_db_work);
1946 
1947 	cancel_delayed_work_sync(&qp->link_work);
1948 
1949 	qp->cb_data = NULL;
1950 	qp->rx_handler = NULL;
1951 	qp->tx_handler = NULL;
1952 	qp->event_handler = NULL;
1953 
1954 	while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q)))
1955 		kfree(entry);
1956 
1957 	while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_pend_q))) {
1958 		dev_warn(&pdev->dev, "Freeing item from non-empty rx_pend_q\n");
1959 		kfree(entry);
1960 	}
1961 
1962 	while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_post_q))) {
1963 		dev_warn(&pdev->dev, "Freeing item from non-empty rx_post_q\n");
1964 		kfree(entry);
1965 	}
1966 
1967 	while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
1968 		kfree(entry);
1969 
1970 	qp->transport->qp_bitmap_free |= qp_bit;
1971 
1972 	dev_info(&pdev->dev, "NTB Transport QP %d freed\n", qp->qp_num);
1973 }
1974 EXPORT_SYMBOL_GPL(ntb_transport_free_queue);
1975 
1976 /**
1977  * ntb_transport_rx_remove - Dequeues enqueued rx packet
1978  * @qp: NTB queue to be freed
1979  * @len: pointer to variable to write enqueued buffers length
1980  *
1981  * Dequeues unused buffers from receive queue.  Should only be used during
1982  * shutdown of qp.
1983  *
1984  * RETURNS: NULL error value on error, or void* for success.
1985  */
1986 void *ntb_transport_rx_remove(struct ntb_transport_qp *qp, unsigned int *len)
1987 {
1988 	struct ntb_queue_entry *entry;
1989 	void *buf;
1990 
1991 	if (!qp || qp->client_ready)
1992 		return NULL;
1993 
1994 	entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_pend_q);
1995 	if (!entry)
1996 		return NULL;
1997 
1998 	buf = entry->cb_data;
1999 	*len = entry->len;
2000 
2001 	ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry, &qp->rx_free_q);
2002 
2003 	return buf;
2004 }
2005 EXPORT_SYMBOL_GPL(ntb_transport_rx_remove);
2006 
2007 /**
2008  * ntb_transport_rx_enqueue - Enqueue a new NTB queue entry
2009  * @qp: NTB transport layer queue the entry is to be enqueued on
2010  * @cb: per buffer pointer for callback function to use
2011  * @data: pointer to data buffer that incoming packets will be copied into
2012  * @len: length of the data buffer
2013  *
2014  * Enqueue a new receive buffer onto the transport queue into which a NTB
2015  * payload can be received into.
2016  *
2017  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
2018  */
2019 int ntb_transport_rx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
2020 			     unsigned int len)
2021 {
2022 	struct ntb_queue_entry *entry;
2023 
2024 	if (!qp)
2025 		return -EINVAL;
2026 
2027 	entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q);
2028 	if (!entry)
2029 		return -ENOMEM;
2030 
2031 	entry->cb_data = cb;
2032 	entry->buf = data;
2033 	entry->len = len;
2034 	entry->flags = 0;
2035 	entry->retries = 0;
2036 	entry->errors = 0;
2037 	entry->rx_index = 0;
2038 
2039 	ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry, &qp->rx_pend_q);
2040 
2041 	if (qp->active)
2042 		tasklet_schedule(&qp->rxc_db_work);
2043 
2044 	return 0;
2045 }
2046 EXPORT_SYMBOL_GPL(ntb_transport_rx_enqueue);
2047 
2048 /**
2049  * ntb_transport_tx_enqueue - Enqueue a new NTB queue entry
2050  * @qp: NTB transport layer queue the entry is to be enqueued on
2051  * @cb: per buffer pointer for callback function to use
2052  * @data: pointer to data buffer that will be sent
2053  * @len: length of the data buffer
2054  *
2055  * Enqueue a new transmit buffer onto the transport queue from which a NTB
2056  * payload will be transmitted.  This assumes that a lock is being held to
2057  * serialize access to the qp.
2058  *
2059  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
2060  */
2061 int ntb_transport_tx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
2062 			     unsigned int len)
2063 {
2064 	struct ntb_queue_entry *entry;
2065 	int rc;
2066 
2067 	if (!qp || !qp->link_is_up || !len)
2068 		return -EINVAL;
2069 
2070 	entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
2071 	if (!entry) {
2072 		qp->tx_err_no_buf++;
2073 		return -EBUSY;
2074 	}
2075 
2076 	entry->cb_data = cb;
2077 	entry->buf = data;
2078 	entry->len = len;
2079 	entry->flags = 0;
2080 	entry->errors = 0;
2081 	entry->retries = 0;
2082 	entry->tx_index = 0;
2083 
2084 	rc = ntb_process_tx(qp, entry);
2085 	if (rc)
2086 		ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
2087 			     &qp->tx_free_q);
2088 
2089 	return rc;
2090 }
2091 EXPORT_SYMBOL_GPL(ntb_transport_tx_enqueue);
2092 
2093 /**
2094  * ntb_transport_link_up - Notify NTB transport of client readiness to use queue
2095  * @qp: NTB transport layer queue to be enabled
2096  *
2097  * Notify NTB transport layer of client readiness to use queue
2098  */
2099 void ntb_transport_link_up(struct ntb_transport_qp *qp)
2100 {
2101 	if (!qp)
2102 		return;
2103 
2104 	qp->client_ready = true;
2105 
2106 	if (qp->transport->link_is_up)
2107 		schedule_delayed_work(&qp->link_work, 0);
2108 }
2109 EXPORT_SYMBOL_GPL(ntb_transport_link_up);
2110 
2111 /**
2112  * ntb_transport_link_down - Notify NTB transport to no longer enqueue data
2113  * @qp: NTB transport layer queue to be disabled
2114  *
2115  * Notify NTB transport layer of client's desire to no longer receive data on
2116  * transport queue specified.  It is the client's responsibility to ensure all
2117  * entries on queue are purged or otherwise handled appropriately.
2118  */
2119 void ntb_transport_link_down(struct ntb_transport_qp *qp)
2120 {
2121 	int val;
2122 
2123 	if (!qp)
2124 		return;
2125 
2126 	qp->client_ready = false;
2127 
2128 	val = ntb_spad_read(qp->ndev, QP_LINKS);
2129 
2130 	ntb_peer_spad_write(qp->ndev, QP_LINKS,
2131 			    val & ~BIT(qp->qp_num));
2132 
2133 	if (qp->link_is_up)
2134 		ntb_send_link_down(qp);
2135 	else
2136 		cancel_delayed_work_sync(&qp->link_work);
2137 }
2138 EXPORT_SYMBOL_GPL(ntb_transport_link_down);
2139 
2140 /**
2141  * ntb_transport_link_query - Query transport link state
2142  * @qp: NTB transport layer queue to be queried
2143  *
2144  * Query connectivity to the remote system of the NTB transport queue
2145  *
2146  * RETURNS: true for link up or false for link down
2147  */
2148 bool ntb_transport_link_query(struct ntb_transport_qp *qp)
2149 {
2150 	if (!qp)
2151 		return false;
2152 
2153 	return qp->link_is_up;
2154 }
2155 EXPORT_SYMBOL_GPL(ntb_transport_link_query);
2156 
2157 /**
2158  * ntb_transport_qp_num - Query the qp number
2159  * @qp: NTB transport layer queue to be queried
2160  *
2161  * Query qp number of the NTB transport queue
2162  *
2163  * RETURNS: a zero based number specifying the qp number
2164  */
2165 unsigned char ntb_transport_qp_num(struct ntb_transport_qp *qp)
2166 {
2167 	if (!qp)
2168 		return 0;
2169 
2170 	return qp->qp_num;
2171 }
2172 EXPORT_SYMBOL_GPL(ntb_transport_qp_num);
2173 
2174 /**
2175  * ntb_transport_max_size - Query the max payload size of a qp
2176  * @qp: NTB transport layer queue to be queried
2177  *
2178  * Query the maximum payload size permissible on the given qp
2179  *
2180  * RETURNS: the max payload size of a qp
2181  */
2182 unsigned int ntb_transport_max_size(struct ntb_transport_qp *qp)
2183 {
2184 	unsigned int max_size;
2185 	unsigned int copy_align;
2186 	struct dma_chan *rx_chan, *tx_chan;
2187 
2188 	if (!qp)
2189 		return 0;
2190 
2191 	rx_chan = qp->rx_dma_chan;
2192 	tx_chan = qp->tx_dma_chan;
2193 
2194 	copy_align = max(rx_chan ? rx_chan->device->copy_align : 0,
2195 			 tx_chan ? tx_chan->device->copy_align : 0);
2196 
2197 	/* If DMA engine usage is possible, try to find the max size for that */
2198 	max_size = qp->tx_max_frame - sizeof(struct ntb_payload_header);
2199 	max_size = round_down(max_size, 1 << copy_align);
2200 
2201 	return max_size;
2202 }
2203 EXPORT_SYMBOL_GPL(ntb_transport_max_size);
2204 
2205 unsigned int ntb_transport_tx_free_entry(struct ntb_transport_qp *qp)
2206 {
2207 	unsigned int head = qp->tx_index;
2208 	unsigned int tail = qp->remote_rx_info->entry;
2209 
2210 	return tail > head ? tail - head : qp->tx_max_entry + tail - head;
2211 }
2212 EXPORT_SYMBOL_GPL(ntb_transport_tx_free_entry);
2213 
2214 static void ntb_transport_doorbell_callback(void *data, int vector)
2215 {
2216 	struct ntb_transport_ctx *nt = data;
2217 	struct ntb_transport_qp *qp;
2218 	u64 db_bits;
2219 	unsigned int qp_num;
2220 
2221 	db_bits = (nt->qp_bitmap & ~nt->qp_bitmap_free &
2222 		   ntb_db_vector_mask(nt->ndev, vector));
2223 
2224 	while (db_bits) {
2225 		qp_num = __ffs(db_bits);
2226 		qp = &nt->qp_vec[qp_num];
2227 
2228 		if (qp->active)
2229 			tasklet_schedule(&qp->rxc_db_work);
2230 
2231 		db_bits &= ~BIT_ULL(qp_num);
2232 	}
2233 }
2234 
2235 static const struct ntb_ctx_ops ntb_transport_ops = {
2236 	.link_event = ntb_transport_event_callback,
2237 	.db_event = ntb_transport_doorbell_callback,
2238 };
2239 
2240 static struct ntb_client ntb_transport_client = {
2241 	.ops = {
2242 		.probe = ntb_transport_probe,
2243 		.remove = ntb_transport_free,
2244 	},
2245 };
2246 
2247 static int __init ntb_transport_init(void)
2248 {
2249 	int rc;
2250 
2251 	pr_info("%s, version %s\n", NTB_TRANSPORT_DESC, NTB_TRANSPORT_VER);
2252 
2253 	if (debugfs_initialized())
2254 		nt_debugfs_dir = debugfs_create_dir(KBUILD_MODNAME, NULL);
2255 
2256 	rc = bus_register(&ntb_transport_bus);
2257 	if (rc)
2258 		goto err_bus;
2259 
2260 	rc = ntb_register_client(&ntb_transport_client);
2261 	if (rc)
2262 		goto err_client;
2263 
2264 	return 0;
2265 
2266 err_client:
2267 	bus_unregister(&ntb_transport_bus);
2268 err_bus:
2269 	debugfs_remove_recursive(nt_debugfs_dir);
2270 	return rc;
2271 }
2272 module_init(ntb_transport_init);
2273 
2274 static void __exit ntb_transport_exit(void)
2275 {
2276 	ntb_unregister_client(&ntb_transport_client);
2277 	bus_unregister(&ntb_transport_bus);
2278 	debugfs_remove_recursive(nt_debugfs_dir);
2279 }
2280 module_exit(ntb_transport_exit);
2281