xref: /openbmc/linux/drivers/ntb/ntb_transport.c (revision bc05aa6e)
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 	dma_addr_t tx_mw_phys;
148 	unsigned int tx_index;
149 	unsigned int tx_max_entry;
150 	unsigned int tx_max_frame;
151 
152 	void (*rx_handler)(struct ntb_transport_qp *qp, void *qp_data,
153 			   void *data, int len);
154 	struct list_head rx_post_q;
155 	struct list_head rx_pend_q;
156 	struct list_head rx_free_q;
157 	/* ntb_rx_q_lock: synchronize access to rx_XXXX_q */
158 	spinlock_t ntb_rx_q_lock;
159 	void *rx_buff;
160 	unsigned int rx_index;
161 	unsigned int rx_max_entry;
162 	unsigned int rx_max_frame;
163 	unsigned int rx_alloc_entry;
164 	dma_cookie_t last_cookie;
165 	struct tasklet_struct rxc_db_work;
166 
167 	void (*event_handler)(void *data, int status);
168 	struct delayed_work link_work;
169 	struct work_struct link_cleanup;
170 
171 	struct dentry *debugfs_dir;
172 	struct dentry *debugfs_stats;
173 
174 	/* Stats */
175 	u64 rx_bytes;
176 	u64 rx_pkts;
177 	u64 rx_ring_empty;
178 	u64 rx_err_no_buf;
179 	u64 rx_err_oflow;
180 	u64 rx_err_ver;
181 	u64 rx_memcpy;
182 	u64 rx_async;
183 	u64 tx_bytes;
184 	u64 tx_pkts;
185 	u64 tx_ring_full;
186 	u64 tx_err_no_buf;
187 	u64 tx_memcpy;
188 	u64 tx_async;
189 };
190 
191 struct ntb_transport_mw {
192 	phys_addr_t phys_addr;
193 	resource_size_t phys_size;
194 	void __iomem *vbase;
195 	size_t xlat_size;
196 	size_t buff_size;
197 	void *virt_addr;
198 	dma_addr_t dma_addr;
199 };
200 
201 struct ntb_transport_client_dev {
202 	struct list_head entry;
203 	struct ntb_transport_ctx *nt;
204 	struct device dev;
205 };
206 
207 struct ntb_transport_ctx {
208 	struct list_head entry;
209 	struct list_head client_devs;
210 
211 	struct ntb_dev *ndev;
212 
213 	struct ntb_transport_mw *mw_vec;
214 	struct ntb_transport_qp *qp_vec;
215 	unsigned int mw_count;
216 	unsigned int qp_count;
217 	u64 qp_bitmap;
218 	u64 qp_bitmap_free;
219 
220 	bool link_is_up;
221 	struct delayed_work link_work;
222 	struct work_struct link_cleanup;
223 
224 	struct dentry *debugfs_node_dir;
225 };
226 
227 enum {
228 	DESC_DONE_FLAG = BIT(0),
229 	LINK_DOWN_FLAG = BIT(1),
230 };
231 
232 struct ntb_payload_header {
233 	unsigned int ver;
234 	unsigned int len;
235 	unsigned int flags;
236 };
237 
238 enum {
239 	VERSION = 0,
240 	QP_LINKS,
241 	NUM_QPS,
242 	NUM_MWS,
243 	MW0_SZ_HIGH,
244 	MW0_SZ_LOW,
245 };
246 
247 #define dev_client_dev(__dev) \
248 	container_of((__dev), struct ntb_transport_client_dev, dev)
249 
250 #define drv_client(__drv) \
251 	container_of((__drv), struct ntb_transport_client, driver)
252 
253 #define QP_TO_MW(nt, qp)	((qp) % nt->mw_count)
254 #define NTB_QP_DEF_NUM_ENTRIES	100
255 #define NTB_LINK_DOWN_TIMEOUT	10
256 
257 static void ntb_transport_rxc_db(unsigned long data);
258 static const struct ntb_ctx_ops ntb_transport_ops;
259 static struct ntb_client ntb_transport_client;
260 static int ntb_async_tx_submit(struct ntb_transport_qp *qp,
261 			       struct ntb_queue_entry *entry);
262 static void ntb_memcpy_tx(struct ntb_queue_entry *entry, void __iomem *offset);
263 static int ntb_async_rx_submit(struct ntb_queue_entry *entry, void *offset);
264 static void ntb_memcpy_rx(struct ntb_queue_entry *entry, void *offset);
265 
266 
267 static int ntb_transport_bus_match(struct device *dev,
268 				   struct device_driver *drv)
269 {
270 	return !strncmp(dev_name(dev), drv->name, strlen(drv->name));
271 }
272 
273 static int ntb_transport_bus_probe(struct device *dev)
274 {
275 	const struct ntb_transport_client *client;
276 	int rc = -EINVAL;
277 
278 	get_device(dev);
279 
280 	client = drv_client(dev->driver);
281 	rc = client->probe(dev);
282 	if (rc)
283 		put_device(dev);
284 
285 	return rc;
286 }
287 
288 static int ntb_transport_bus_remove(struct device *dev)
289 {
290 	const struct ntb_transport_client *client;
291 
292 	client = drv_client(dev->driver);
293 	client->remove(dev);
294 
295 	put_device(dev);
296 
297 	return 0;
298 }
299 
300 static struct bus_type ntb_transport_bus = {
301 	.name = "ntb_transport",
302 	.match = ntb_transport_bus_match,
303 	.probe = ntb_transport_bus_probe,
304 	.remove = ntb_transport_bus_remove,
305 };
306 
307 static LIST_HEAD(ntb_transport_list);
308 
309 static int ntb_bus_init(struct ntb_transport_ctx *nt)
310 {
311 	list_add_tail(&nt->entry, &ntb_transport_list);
312 	return 0;
313 }
314 
315 static void ntb_bus_remove(struct ntb_transport_ctx *nt)
316 {
317 	struct ntb_transport_client_dev *client_dev, *cd;
318 
319 	list_for_each_entry_safe(client_dev, cd, &nt->client_devs, entry) {
320 		dev_err(client_dev->dev.parent, "%s still attached to bus, removing\n",
321 			dev_name(&client_dev->dev));
322 		list_del(&client_dev->entry);
323 		device_unregister(&client_dev->dev);
324 	}
325 
326 	list_del(&nt->entry);
327 }
328 
329 static void ntb_transport_client_release(struct device *dev)
330 {
331 	struct ntb_transport_client_dev *client_dev;
332 
333 	client_dev = dev_client_dev(dev);
334 	kfree(client_dev);
335 }
336 
337 /**
338  * ntb_transport_unregister_client_dev - Unregister NTB client device
339  * @device_name: Name of NTB client device
340  *
341  * Unregister an NTB client device with the NTB transport layer
342  */
343 void ntb_transport_unregister_client_dev(char *device_name)
344 {
345 	struct ntb_transport_client_dev *client, *cd;
346 	struct ntb_transport_ctx *nt;
347 
348 	list_for_each_entry(nt, &ntb_transport_list, entry)
349 		list_for_each_entry_safe(client, cd, &nt->client_devs, entry)
350 			if (!strncmp(dev_name(&client->dev), device_name,
351 				     strlen(device_name))) {
352 				list_del(&client->entry);
353 				device_unregister(&client->dev);
354 			}
355 }
356 EXPORT_SYMBOL_GPL(ntb_transport_unregister_client_dev);
357 
358 /**
359  * ntb_transport_register_client_dev - Register NTB client device
360  * @device_name: Name of NTB client device
361  *
362  * Register an NTB client device with the NTB transport layer
363  */
364 int ntb_transport_register_client_dev(char *device_name)
365 {
366 	struct ntb_transport_client_dev *client_dev;
367 	struct ntb_transport_ctx *nt;
368 	int node;
369 	int rc, i = 0;
370 
371 	if (list_empty(&ntb_transport_list))
372 		return -ENODEV;
373 
374 	list_for_each_entry(nt, &ntb_transport_list, entry) {
375 		struct device *dev;
376 
377 		node = dev_to_node(&nt->ndev->dev);
378 
379 		client_dev = kzalloc_node(sizeof(*client_dev),
380 					  GFP_KERNEL, node);
381 		if (!client_dev) {
382 			rc = -ENOMEM;
383 			goto err;
384 		}
385 
386 		dev = &client_dev->dev;
387 
388 		/* setup and register client devices */
389 		dev_set_name(dev, "%s%d", device_name, i);
390 		dev->bus = &ntb_transport_bus;
391 		dev->release = ntb_transport_client_release;
392 		dev->parent = &nt->ndev->dev;
393 
394 		rc = device_register(dev);
395 		if (rc) {
396 			kfree(client_dev);
397 			goto err;
398 		}
399 
400 		list_add_tail(&client_dev->entry, &nt->client_devs);
401 		i++;
402 	}
403 
404 	return 0;
405 
406 err:
407 	ntb_transport_unregister_client_dev(device_name);
408 
409 	return rc;
410 }
411 EXPORT_SYMBOL_GPL(ntb_transport_register_client_dev);
412 
413 /**
414  * ntb_transport_register_client - Register NTB client driver
415  * @drv: NTB client driver to be registered
416  *
417  * Register an NTB client driver with the NTB transport layer
418  *
419  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
420  */
421 int ntb_transport_register_client(struct ntb_transport_client *drv)
422 {
423 	drv->driver.bus = &ntb_transport_bus;
424 
425 	if (list_empty(&ntb_transport_list))
426 		return -ENODEV;
427 
428 	return driver_register(&drv->driver);
429 }
430 EXPORT_SYMBOL_GPL(ntb_transport_register_client);
431 
432 /**
433  * ntb_transport_unregister_client - Unregister NTB client driver
434  * @drv: NTB client driver to be unregistered
435  *
436  * Unregister an NTB client driver with the NTB transport layer
437  *
438  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
439  */
440 void ntb_transport_unregister_client(struct ntb_transport_client *drv)
441 {
442 	driver_unregister(&drv->driver);
443 }
444 EXPORT_SYMBOL_GPL(ntb_transport_unregister_client);
445 
446 static ssize_t debugfs_read(struct file *filp, char __user *ubuf, size_t count,
447 			    loff_t *offp)
448 {
449 	struct ntb_transport_qp *qp;
450 	char *buf;
451 	ssize_t ret, out_offset, out_count;
452 
453 	qp = filp->private_data;
454 
455 	if (!qp || !qp->link_is_up)
456 		return 0;
457 
458 	out_count = 1000;
459 
460 	buf = kmalloc(out_count, GFP_KERNEL);
461 	if (!buf)
462 		return -ENOMEM;
463 
464 	out_offset = 0;
465 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
466 			       "\nNTB QP stats:\n\n");
467 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
468 			       "rx_bytes - \t%llu\n", qp->rx_bytes);
469 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
470 			       "rx_pkts - \t%llu\n", qp->rx_pkts);
471 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
472 			       "rx_memcpy - \t%llu\n", qp->rx_memcpy);
473 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
474 			       "rx_async - \t%llu\n", qp->rx_async);
475 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
476 			       "rx_ring_empty - %llu\n", qp->rx_ring_empty);
477 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
478 			       "rx_err_no_buf - %llu\n", qp->rx_err_no_buf);
479 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
480 			       "rx_err_oflow - \t%llu\n", qp->rx_err_oflow);
481 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
482 			       "rx_err_ver - \t%llu\n", qp->rx_err_ver);
483 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
484 			       "rx_buff - \t0x%p\n", qp->rx_buff);
485 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
486 			       "rx_index - \t%u\n", qp->rx_index);
487 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
488 			       "rx_max_entry - \t%u\n", qp->rx_max_entry);
489 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
490 			       "rx_alloc_entry - \t%u\n\n", qp->rx_alloc_entry);
491 
492 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
493 			       "tx_bytes - \t%llu\n", qp->tx_bytes);
494 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
495 			       "tx_pkts - \t%llu\n", qp->tx_pkts);
496 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
497 			       "tx_memcpy - \t%llu\n", qp->tx_memcpy);
498 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
499 			       "tx_async - \t%llu\n", qp->tx_async);
500 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
501 			       "tx_ring_full - \t%llu\n", qp->tx_ring_full);
502 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
503 			       "tx_err_no_buf - %llu\n", qp->tx_err_no_buf);
504 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
505 			       "tx_mw - \t0x%p\n", qp->tx_mw);
506 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
507 			       "tx_index (H) - \t%u\n", qp->tx_index);
508 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
509 			       "RRI (T) - \t%u\n",
510 			       qp->remote_rx_info->entry);
511 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
512 			       "tx_max_entry - \t%u\n", qp->tx_max_entry);
513 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
514 			       "free tx - \t%u\n",
515 			       ntb_transport_tx_free_entry(qp));
516 
517 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
518 			       "\n");
519 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
520 			       "Using TX DMA - \t%s\n",
521 			       qp->tx_dma_chan ? "Yes" : "No");
522 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
523 			       "Using RX DMA - \t%s\n",
524 			       qp->rx_dma_chan ? "Yes" : "No");
525 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
526 			       "QP Link - \t%s\n",
527 			       qp->link_is_up ? "Up" : "Down");
528 	out_offset += snprintf(buf + out_offset, out_count - out_offset,
529 			       "\n");
530 
531 	if (out_offset > out_count)
532 		out_offset = out_count;
533 
534 	ret = simple_read_from_buffer(ubuf, count, offp, buf, out_offset);
535 	kfree(buf);
536 	return ret;
537 }
538 
539 static const struct file_operations ntb_qp_debugfs_stats = {
540 	.owner = THIS_MODULE,
541 	.open = simple_open,
542 	.read = debugfs_read,
543 };
544 
545 static void ntb_list_add(spinlock_t *lock, struct list_head *entry,
546 			 struct list_head *list)
547 {
548 	unsigned long flags;
549 
550 	spin_lock_irqsave(lock, flags);
551 	list_add_tail(entry, list);
552 	spin_unlock_irqrestore(lock, flags);
553 }
554 
555 static struct ntb_queue_entry *ntb_list_rm(spinlock_t *lock,
556 					   struct list_head *list)
557 {
558 	struct ntb_queue_entry *entry;
559 	unsigned long flags;
560 
561 	spin_lock_irqsave(lock, flags);
562 	if (list_empty(list)) {
563 		entry = NULL;
564 		goto out;
565 	}
566 	entry = list_first_entry(list, struct ntb_queue_entry, entry);
567 	list_del(&entry->entry);
568 
569 out:
570 	spin_unlock_irqrestore(lock, flags);
571 
572 	return entry;
573 }
574 
575 static struct ntb_queue_entry *ntb_list_mv(spinlock_t *lock,
576 					   struct list_head *list,
577 					   struct list_head *to_list)
578 {
579 	struct ntb_queue_entry *entry;
580 	unsigned long flags;
581 
582 	spin_lock_irqsave(lock, flags);
583 
584 	if (list_empty(list)) {
585 		entry = NULL;
586 	} else {
587 		entry = list_first_entry(list, struct ntb_queue_entry, entry);
588 		list_move_tail(&entry->entry, to_list);
589 	}
590 
591 	spin_unlock_irqrestore(lock, flags);
592 
593 	return entry;
594 }
595 
596 static int ntb_transport_setup_qp_mw(struct ntb_transport_ctx *nt,
597 				     unsigned int qp_num)
598 {
599 	struct ntb_transport_qp *qp = &nt->qp_vec[qp_num];
600 	struct ntb_transport_mw *mw;
601 	struct ntb_dev *ndev = nt->ndev;
602 	struct ntb_queue_entry *entry;
603 	unsigned int rx_size, num_qps_mw;
604 	unsigned int mw_num, mw_count, qp_count;
605 	unsigned int i;
606 	int node;
607 
608 	mw_count = nt->mw_count;
609 	qp_count = nt->qp_count;
610 
611 	mw_num = QP_TO_MW(nt, qp_num);
612 	mw = &nt->mw_vec[mw_num];
613 
614 	if (!mw->virt_addr)
615 		return -ENOMEM;
616 
617 	if (mw_num < qp_count % mw_count)
618 		num_qps_mw = qp_count / mw_count + 1;
619 	else
620 		num_qps_mw = qp_count / mw_count;
621 
622 	rx_size = (unsigned int)mw->xlat_size / num_qps_mw;
623 	qp->rx_buff = mw->virt_addr + rx_size * (qp_num / mw_count);
624 	rx_size -= sizeof(struct ntb_rx_info);
625 
626 	qp->remote_rx_info = qp->rx_buff + rx_size;
627 
628 	/* Due to housekeeping, there must be atleast 2 buffs */
629 	qp->rx_max_frame = min(transport_mtu, rx_size / 2);
630 	qp->rx_max_entry = rx_size / qp->rx_max_frame;
631 	qp->rx_index = 0;
632 
633 	/*
634 	 * Checking to see if we have more entries than the default.
635 	 * We should add additional entries if that is the case so we
636 	 * can be in sync with the transport frames.
637 	 */
638 	node = dev_to_node(&ndev->dev);
639 	for (i = qp->rx_alloc_entry; i < qp->rx_max_entry; i++) {
640 		entry = kzalloc_node(sizeof(*entry), GFP_ATOMIC, node);
641 		if (!entry)
642 			return -ENOMEM;
643 
644 		entry->qp = qp;
645 		ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry,
646 			     &qp->rx_free_q);
647 		qp->rx_alloc_entry++;
648 	}
649 
650 	qp->remote_rx_info->entry = qp->rx_max_entry - 1;
651 
652 	/* setup the hdr offsets with 0's */
653 	for (i = 0; i < qp->rx_max_entry; i++) {
654 		void *offset = (qp->rx_buff + qp->rx_max_frame * (i + 1) -
655 				sizeof(struct ntb_payload_header));
656 		memset(offset, 0, sizeof(struct ntb_payload_header));
657 	}
658 
659 	qp->rx_pkts = 0;
660 	qp->tx_pkts = 0;
661 	qp->tx_index = 0;
662 
663 	return 0;
664 }
665 
666 static void ntb_free_mw(struct ntb_transport_ctx *nt, int num_mw)
667 {
668 	struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
669 	struct pci_dev *pdev = nt->ndev->pdev;
670 
671 	if (!mw->virt_addr)
672 		return;
673 
674 	ntb_mw_clear_trans(nt->ndev, PIDX, num_mw);
675 	dma_free_coherent(&pdev->dev, mw->buff_size,
676 			  mw->virt_addr, mw->dma_addr);
677 	mw->xlat_size = 0;
678 	mw->buff_size = 0;
679 	mw->virt_addr = NULL;
680 }
681 
682 static int ntb_set_mw(struct ntb_transport_ctx *nt, int num_mw,
683 		      resource_size_t size)
684 {
685 	struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
686 	struct pci_dev *pdev = nt->ndev->pdev;
687 	size_t xlat_size, buff_size;
688 	resource_size_t xlat_align;
689 	resource_size_t xlat_align_size;
690 	int rc;
691 
692 	if (!size)
693 		return -EINVAL;
694 
695 	rc = ntb_mw_get_align(nt->ndev, PIDX, num_mw, &xlat_align,
696 			      &xlat_align_size, NULL);
697 	if (rc)
698 		return rc;
699 
700 	xlat_size = round_up(size, xlat_align_size);
701 	buff_size = round_up(size, xlat_align);
702 
703 	/* No need to re-setup */
704 	if (mw->xlat_size == xlat_size)
705 		return 0;
706 
707 	if (mw->buff_size)
708 		ntb_free_mw(nt, num_mw);
709 
710 	/* Alloc memory for receiving data.  Must be aligned */
711 	mw->xlat_size = xlat_size;
712 	mw->buff_size = buff_size;
713 
714 	mw->virt_addr = dma_alloc_coherent(&pdev->dev, buff_size,
715 					   &mw->dma_addr, GFP_KERNEL);
716 	if (!mw->virt_addr) {
717 		mw->xlat_size = 0;
718 		mw->buff_size = 0;
719 		dev_err(&pdev->dev, "Unable to alloc MW buff of size %zu\n",
720 			buff_size);
721 		return -ENOMEM;
722 	}
723 
724 	/*
725 	 * we must ensure that the memory address allocated is BAR size
726 	 * aligned in order for the XLAT register to take the value. This
727 	 * is a requirement of the hardware. It is recommended to setup CMA
728 	 * for BAR sizes equal or greater than 4MB.
729 	 */
730 	if (!IS_ALIGNED(mw->dma_addr, xlat_align)) {
731 		dev_err(&pdev->dev, "DMA memory %pad is not aligned\n",
732 			&mw->dma_addr);
733 		ntb_free_mw(nt, num_mw);
734 		return -ENOMEM;
735 	}
736 
737 	/* Notify HW the memory location of the receive buffer */
738 	rc = ntb_mw_set_trans(nt->ndev, PIDX, num_mw, mw->dma_addr,
739 			      mw->xlat_size);
740 	if (rc) {
741 		dev_err(&pdev->dev, "Unable to set mw%d translation", num_mw);
742 		ntb_free_mw(nt, num_mw);
743 		return -EIO;
744 	}
745 
746 	return 0;
747 }
748 
749 static void ntb_qp_link_down_reset(struct ntb_transport_qp *qp)
750 {
751 	qp->link_is_up = false;
752 	qp->active = false;
753 
754 	qp->tx_index = 0;
755 	qp->rx_index = 0;
756 	qp->rx_bytes = 0;
757 	qp->rx_pkts = 0;
758 	qp->rx_ring_empty = 0;
759 	qp->rx_err_no_buf = 0;
760 	qp->rx_err_oflow = 0;
761 	qp->rx_err_ver = 0;
762 	qp->rx_memcpy = 0;
763 	qp->rx_async = 0;
764 	qp->tx_bytes = 0;
765 	qp->tx_pkts = 0;
766 	qp->tx_ring_full = 0;
767 	qp->tx_err_no_buf = 0;
768 	qp->tx_memcpy = 0;
769 	qp->tx_async = 0;
770 }
771 
772 static void ntb_qp_link_cleanup(struct ntb_transport_qp *qp)
773 {
774 	struct ntb_transport_ctx *nt = qp->transport;
775 	struct pci_dev *pdev = nt->ndev->pdev;
776 
777 	dev_info(&pdev->dev, "qp %d: Link Cleanup\n", qp->qp_num);
778 
779 	cancel_delayed_work_sync(&qp->link_work);
780 	ntb_qp_link_down_reset(qp);
781 
782 	if (qp->event_handler)
783 		qp->event_handler(qp->cb_data, qp->link_is_up);
784 }
785 
786 static void ntb_qp_link_cleanup_work(struct work_struct *work)
787 {
788 	struct ntb_transport_qp *qp = container_of(work,
789 						   struct ntb_transport_qp,
790 						   link_cleanup);
791 	struct ntb_transport_ctx *nt = qp->transport;
792 
793 	ntb_qp_link_cleanup(qp);
794 
795 	if (nt->link_is_up)
796 		schedule_delayed_work(&qp->link_work,
797 				      msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
798 }
799 
800 static void ntb_qp_link_down(struct ntb_transport_qp *qp)
801 {
802 	schedule_work(&qp->link_cleanup);
803 }
804 
805 static void ntb_transport_link_cleanup(struct ntb_transport_ctx *nt)
806 {
807 	struct ntb_transport_qp *qp;
808 	u64 qp_bitmap_alloc;
809 	unsigned int i, count;
810 
811 	qp_bitmap_alloc = nt->qp_bitmap & ~nt->qp_bitmap_free;
812 
813 	/* Pass along the info to any clients */
814 	for (i = 0; i < nt->qp_count; i++)
815 		if (qp_bitmap_alloc & BIT_ULL(i)) {
816 			qp = &nt->qp_vec[i];
817 			ntb_qp_link_cleanup(qp);
818 			cancel_work_sync(&qp->link_cleanup);
819 			cancel_delayed_work_sync(&qp->link_work);
820 		}
821 
822 	if (!nt->link_is_up)
823 		cancel_delayed_work_sync(&nt->link_work);
824 
825 	/* The scratchpad registers keep the values if the remote side
826 	 * goes down, blast them now to give them a sane value the next
827 	 * time they are accessed
828 	 */
829 	count = ntb_spad_count(nt->ndev);
830 	for (i = 0; i < count; i++)
831 		ntb_spad_write(nt->ndev, i, 0);
832 }
833 
834 static void ntb_transport_link_cleanup_work(struct work_struct *work)
835 {
836 	struct ntb_transport_ctx *nt =
837 		container_of(work, struct ntb_transport_ctx, link_cleanup);
838 
839 	ntb_transport_link_cleanup(nt);
840 }
841 
842 static void ntb_transport_event_callback(void *data)
843 {
844 	struct ntb_transport_ctx *nt = data;
845 
846 	if (ntb_link_is_up(nt->ndev, NULL, NULL) == 1)
847 		schedule_delayed_work(&nt->link_work, 0);
848 	else
849 		schedule_work(&nt->link_cleanup);
850 }
851 
852 static void ntb_transport_link_work(struct work_struct *work)
853 {
854 	struct ntb_transport_ctx *nt =
855 		container_of(work, struct ntb_transport_ctx, link_work.work);
856 	struct ntb_dev *ndev = nt->ndev;
857 	struct pci_dev *pdev = ndev->pdev;
858 	resource_size_t size;
859 	u32 val;
860 	int rc = 0, i, spad;
861 
862 	/* send the local info, in the opposite order of the way we read it */
863 	for (i = 0; i < nt->mw_count; i++) {
864 		size = nt->mw_vec[i].phys_size;
865 
866 		if (max_mw_size && size > max_mw_size)
867 			size = max_mw_size;
868 
869 		spad = MW0_SZ_HIGH + (i * 2);
870 		ntb_peer_spad_write(ndev, PIDX, spad, upper_32_bits(size));
871 
872 		spad = MW0_SZ_LOW + (i * 2);
873 		ntb_peer_spad_write(ndev, PIDX, spad, lower_32_bits(size));
874 	}
875 
876 	ntb_peer_spad_write(ndev, PIDX, NUM_MWS, nt->mw_count);
877 
878 	ntb_peer_spad_write(ndev, PIDX, NUM_QPS, nt->qp_count);
879 
880 	ntb_peer_spad_write(ndev, PIDX, VERSION, NTB_TRANSPORT_VERSION);
881 
882 	/* Query the remote side for its info */
883 	val = ntb_spad_read(ndev, VERSION);
884 	dev_dbg(&pdev->dev, "Remote version = %d\n", val);
885 	if (val != NTB_TRANSPORT_VERSION)
886 		goto out;
887 
888 	val = ntb_spad_read(ndev, NUM_QPS);
889 	dev_dbg(&pdev->dev, "Remote max number of qps = %d\n", val);
890 	if (val != nt->qp_count)
891 		goto out;
892 
893 	val = ntb_spad_read(ndev, NUM_MWS);
894 	dev_dbg(&pdev->dev, "Remote number of mws = %d\n", val);
895 	if (val != nt->mw_count)
896 		goto out;
897 
898 	for (i = 0; i < nt->mw_count; i++) {
899 		u64 val64;
900 
901 		val = ntb_spad_read(ndev, MW0_SZ_HIGH + (i * 2));
902 		val64 = (u64)val << 32;
903 
904 		val = ntb_spad_read(ndev, MW0_SZ_LOW + (i * 2));
905 		val64 |= val;
906 
907 		dev_dbg(&pdev->dev, "Remote MW%d size = %#llx\n", i, val64);
908 
909 		rc = ntb_set_mw(nt, i, val64);
910 		if (rc)
911 			goto out1;
912 	}
913 
914 	nt->link_is_up = true;
915 
916 	for (i = 0; i < nt->qp_count; i++) {
917 		struct ntb_transport_qp *qp = &nt->qp_vec[i];
918 
919 		ntb_transport_setup_qp_mw(nt, i);
920 
921 		if (qp->client_ready)
922 			schedule_delayed_work(&qp->link_work, 0);
923 	}
924 
925 	return;
926 
927 out1:
928 	for (i = 0; i < nt->mw_count; i++)
929 		ntb_free_mw(nt, i);
930 
931 	/* if there's an actual failure, we should just bail */
932 	if (rc < 0)
933 		return;
934 
935 out:
936 	if (ntb_link_is_up(ndev, NULL, NULL) == 1)
937 		schedule_delayed_work(&nt->link_work,
938 				      msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
939 }
940 
941 static void ntb_qp_link_work(struct work_struct *work)
942 {
943 	struct ntb_transport_qp *qp = container_of(work,
944 						   struct ntb_transport_qp,
945 						   link_work.work);
946 	struct pci_dev *pdev = qp->ndev->pdev;
947 	struct ntb_transport_ctx *nt = qp->transport;
948 	int val;
949 
950 	WARN_ON(!nt->link_is_up);
951 
952 	val = ntb_spad_read(nt->ndev, QP_LINKS);
953 
954 	ntb_peer_spad_write(nt->ndev, PIDX, QP_LINKS, val | BIT(qp->qp_num));
955 
956 	/* query remote spad for qp ready bits */
957 	dev_dbg_ratelimited(&pdev->dev, "Remote QP link status = %x\n", val);
958 
959 	/* See if the remote side is up */
960 	if (val & BIT(qp->qp_num)) {
961 		dev_info(&pdev->dev, "qp %d: Link Up\n", qp->qp_num);
962 		qp->link_is_up = true;
963 		qp->active = true;
964 
965 		if (qp->event_handler)
966 			qp->event_handler(qp->cb_data, qp->link_is_up);
967 
968 		if (qp->active)
969 			tasklet_schedule(&qp->rxc_db_work);
970 	} else if (nt->link_is_up)
971 		schedule_delayed_work(&qp->link_work,
972 				      msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
973 }
974 
975 static int ntb_transport_init_queue(struct ntb_transport_ctx *nt,
976 				    unsigned int qp_num)
977 {
978 	struct ntb_transport_qp *qp;
979 	phys_addr_t mw_base;
980 	resource_size_t mw_size;
981 	unsigned int num_qps_mw, tx_size;
982 	unsigned int mw_num, mw_count, qp_count;
983 	u64 qp_offset;
984 
985 	mw_count = nt->mw_count;
986 	qp_count = nt->qp_count;
987 
988 	mw_num = QP_TO_MW(nt, qp_num);
989 
990 	qp = &nt->qp_vec[qp_num];
991 	qp->qp_num = qp_num;
992 	qp->transport = nt;
993 	qp->ndev = nt->ndev;
994 	qp->client_ready = false;
995 	qp->event_handler = NULL;
996 	ntb_qp_link_down_reset(qp);
997 
998 	if (mw_num < qp_count % mw_count)
999 		num_qps_mw = qp_count / mw_count + 1;
1000 	else
1001 		num_qps_mw = qp_count / mw_count;
1002 
1003 	mw_base = nt->mw_vec[mw_num].phys_addr;
1004 	mw_size = nt->mw_vec[mw_num].phys_size;
1005 
1006 	if (max_mw_size && mw_size > max_mw_size)
1007 		mw_size = max_mw_size;
1008 
1009 	tx_size = (unsigned int)mw_size / num_qps_mw;
1010 	qp_offset = tx_size * (qp_num / mw_count);
1011 
1012 	qp->tx_mw = nt->mw_vec[mw_num].vbase + qp_offset;
1013 	if (!qp->tx_mw)
1014 		return -EINVAL;
1015 
1016 	qp->tx_mw_phys = mw_base + qp_offset;
1017 	if (!qp->tx_mw_phys)
1018 		return -EINVAL;
1019 
1020 	tx_size -= sizeof(struct ntb_rx_info);
1021 	qp->rx_info = qp->tx_mw + tx_size;
1022 
1023 	/* Due to housekeeping, there must be atleast 2 buffs */
1024 	qp->tx_max_frame = min(transport_mtu, tx_size / 2);
1025 	qp->tx_max_entry = tx_size / qp->tx_max_frame;
1026 
1027 	if (nt->debugfs_node_dir) {
1028 		char debugfs_name[4];
1029 
1030 		snprintf(debugfs_name, 4, "qp%d", qp_num);
1031 		qp->debugfs_dir = debugfs_create_dir(debugfs_name,
1032 						     nt->debugfs_node_dir);
1033 
1034 		qp->debugfs_stats = debugfs_create_file("stats", S_IRUSR,
1035 							qp->debugfs_dir, qp,
1036 							&ntb_qp_debugfs_stats);
1037 	} else {
1038 		qp->debugfs_dir = NULL;
1039 		qp->debugfs_stats = NULL;
1040 	}
1041 
1042 	INIT_DELAYED_WORK(&qp->link_work, ntb_qp_link_work);
1043 	INIT_WORK(&qp->link_cleanup, ntb_qp_link_cleanup_work);
1044 
1045 	spin_lock_init(&qp->ntb_rx_q_lock);
1046 	spin_lock_init(&qp->ntb_tx_free_q_lock);
1047 
1048 	INIT_LIST_HEAD(&qp->rx_post_q);
1049 	INIT_LIST_HEAD(&qp->rx_pend_q);
1050 	INIT_LIST_HEAD(&qp->rx_free_q);
1051 	INIT_LIST_HEAD(&qp->tx_free_q);
1052 
1053 	tasklet_init(&qp->rxc_db_work, ntb_transport_rxc_db,
1054 		     (unsigned long)qp);
1055 
1056 	return 0;
1057 }
1058 
1059 static int ntb_transport_probe(struct ntb_client *self, struct ntb_dev *ndev)
1060 {
1061 	struct ntb_transport_ctx *nt;
1062 	struct ntb_transport_mw *mw;
1063 	unsigned int mw_count, qp_count, spad_count, max_mw_count_for_spads;
1064 	u64 qp_bitmap;
1065 	int node;
1066 	int rc, i;
1067 
1068 	mw_count = ntb_peer_mw_count(ndev);
1069 
1070 	if (!ndev->ops->mw_set_trans) {
1071 		dev_err(&ndev->dev, "Inbound MW based NTB API is required\n");
1072 		return -EINVAL;
1073 	}
1074 
1075 	if (ntb_db_is_unsafe(ndev))
1076 		dev_dbg(&ndev->dev,
1077 			"doorbell is unsafe, proceed anyway...\n");
1078 	if (ntb_spad_is_unsafe(ndev))
1079 		dev_dbg(&ndev->dev,
1080 			"scratchpad is unsafe, proceed anyway...\n");
1081 
1082 	if (ntb_peer_port_count(ndev) != NTB_DEF_PEER_CNT)
1083 		dev_warn(&ndev->dev, "Multi-port NTB devices unsupported\n");
1084 
1085 	node = dev_to_node(&ndev->dev);
1086 
1087 	nt = kzalloc_node(sizeof(*nt), GFP_KERNEL, node);
1088 	if (!nt)
1089 		return -ENOMEM;
1090 
1091 	nt->ndev = ndev;
1092 	spad_count = ntb_spad_count(ndev);
1093 
1094 	/* Limit the MW's based on the availability of scratchpads */
1095 
1096 	if (spad_count < NTB_TRANSPORT_MIN_SPADS) {
1097 		nt->mw_count = 0;
1098 		rc = -EINVAL;
1099 		goto err;
1100 	}
1101 
1102 	max_mw_count_for_spads = (spad_count - MW0_SZ_HIGH) / 2;
1103 	nt->mw_count = min(mw_count, max_mw_count_for_spads);
1104 
1105 	nt->mw_vec = kzalloc_node(mw_count * sizeof(*nt->mw_vec),
1106 				  GFP_KERNEL, node);
1107 	if (!nt->mw_vec) {
1108 		rc = -ENOMEM;
1109 		goto err;
1110 	}
1111 
1112 	for (i = 0; i < mw_count; i++) {
1113 		mw = &nt->mw_vec[i];
1114 
1115 		rc = ntb_peer_mw_get_addr(ndev, i, &mw->phys_addr,
1116 					  &mw->phys_size);
1117 		if (rc)
1118 			goto err1;
1119 
1120 		mw->vbase = ioremap_wc(mw->phys_addr, mw->phys_size);
1121 		if (!mw->vbase) {
1122 			rc = -ENOMEM;
1123 			goto err1;
1124 		}
1125 
1126 		mw->buff_size = 0;
1127 		mw->xlat_size = 0;
1128 		mw->virt_addr = NULL;
1129 		mw->dma_addr = 0;
1130 	}
1131 
1132 	qp_bitmap = ntb_db_valid_mask(ndev);
1133 
1134 	qp_count = ilog2(qp_bitmap);
1135 	if (max_num_clients && max_num_clients < qp_count)
1136 		qp_count = max_num_clients;
1137 	else if (nt->mw_count < qp_count)
1138 		qp_count = nt->mw_count;
1139 
1140 	qp_bitmap &= BIT_ULL(qp_count) - 1;
1141 
1142 	nt->qp_count = qp_count;
1143 	nt->qp_bitmap = qp_bitmap;
1144 	nt->qp_bitmap_free = qp_bitmap;
1145 
1146 	nt->qp_vec = kzalloc_node(qp_count * sizeof(*nt->qp_vec),
1147 				  GFP_KERNEL, node);
1148 	if (!nt->qp_vec) {
1149 		rc = -ENOMEM;
1150 		goto err1;
1151 	}
1152 
1153 	if (nt_debugfs_dir) {
1154 		nt->debugfs_node_dir =
1155 			debugfs_create_dir(pci_name(ndev->pdev),
1156 					   nt_debugfs_dir);
1157 	}
1158 
1159 	for (i = 0; i < qp_count; i++) {
1160 		rc = ntb_transport_init_queue(nt, i);
1161 		if (rc)
1162 			goto err2;
1163 	}
1164 
1165 	INIT_DELAYED_WORK(&nt->link_work, ntb_transport_link_work);
1166 	INIT_WORK(&nt->link_cleanup, ntb_transport_link_cleanup_work);
1167 
1168 	rc = ntb_set_ctx(ndev, nt, &ntb_transport_ops);
1169 	if (rc)
1170 		goto err2;
1171 
1172 	INIT_LIST_HEAD(&nt->client_devs);
1173 	rc = ntb_bus_init(nt);
1174 	if (rc)
1175 		goto err3;
1176 
1177 	nt->link_is_up = false;
1178 	ntb_link_enable(ndev, NTB_SPEED_AUTO, NTB_WIDTH_AUTO);
1179 	ntb_link_event(ndev);
1180 
1181 	return 0;
1182 
1183 err3:
1184 	ntb_clear_ctx(ndev);
1185 err2:
1186 	kfree(nt->qp_vec);
1187 err1:
1188 	while (i--) {
1189 		mw = &nt->mw_vec[i];
1190 		iounmap(mw->vbase);
1191 	}
1192 	kfree(nt->mw_vec);
1193 err:
1194 	kfree(nt);
1195 	return rc;
1196 }
1197 
1198 static void ntb_transport_free(struct ntb_client *self, struct ntb_dev *ndev)
1199 {
1200 	struct ntb_transport_ctx *nt = ndev->ctx;
1201 	struct ntb_transport_qp *qp;
1202 	u64 qp_bitmap_alloc;
1203 	int i;
1204 
1205 	ntb_transport_link_cleanup(nt);
1206 	cancel_work_sync(&nt->link_cleanup);
1207 	cancel_delayed_work_sync(&nt->link_work);
1208 
1209 	qp_bitmap_alloc = nt->qp_bitmap & ~nt->qp_bitmap_free;
1210 
1211 	/* verify that all the qp's are freed */
1212 	for (i = 0; i < nt->qp_count; i++) {
1213 		qp = &nt->qp_vec[i];
1214 		if (qp_bitmap_alloc & BIT_ULL(i))
1215 			ntb_transport_free_queue(qp);
1216 		debugfs_remove_recursive(qp->debugfs_dir);
1217 	}
1218 
1219 	ntb_link_disable(ndev);
1220 	ntb_clear_ctx(ndev);
1221 
1222 	ntb_bus_remove(nt);
1223 
1224 	for (i = nt->mw_count; i--; ) {
1225 		ntb_free_mw(nt, i);
1226 		iounmap(nt->mw_vec[i].vbase);
1227 	}
1228 
1229 	kfree(nt->qp_vec);
1230 	kfree(nt->mw_vec);
1231 	kfree(nt);
1232 }
1233 
1234 static void ntb_complete_rxc(struct ntb_transport_qp *qp)
1235 {
1236 	struct ntb_queue_entry *entry;
1237 	void *cb_data;
1238 	unsigned int len;
1239 	unsigned long irqflags;
1240 
1241 	spin_lock_irqsave(&qp->ntb_rx_q_lock, irqflags);
1242 
1243 	while (!list_empty(&qp->rx_post_q)) {
1244 		entry = list_first_entry(&qp->rx_post_q,
1245 					 struct ntb_queue_entry, entry);
1246 		if (!(entry->flags & DESC_DONE_FLAG))
1247 			break;
1248 
1249 		entry->rx_hdr->flags = 0;
1250 		iowrite32(entry->rx_index, &qp->rx_info->entry);
1251 
1252 		cb_data = entry->cb_data;
1253 		len = entry->len;
1254 
1255 		list_move_tail(&entry->entry, &qp->rx_free_q);
1256 
1257 		spin_unlock_irqrestore(&qp->ntb_rx_q_lock, irqflags);
1258 
1259 		if (qp->rx_handler && qp->client_ready)
1260 			qp->rx_handler(qp, qp->cb_data, cb_data, len);
1261 
1262 		spin_lock_irqsave(&qp->ntb_rx_q_lock, irqflags);
1263 	}
1264 
1265 	spin_unlock_irqrestore(&qp->ntb_rx_q_lock, irqflags);
1266 }
1267 
1268 static void ntb_rx_copy_callback(void *data,
1269 				 const struct dmaengine_result *res)
1270 {
1271 	struct ntb_queue_entry *entry = data;
1272 
1273 	/* we need to check DMA results if we are using DMA */
1274 	if (res) {
1275 		enum dmaengine_tx_result dma_err = res->result;
1276 
1277 		switch (dma_err) {
1278 		case DMA_TRANS_READ_FAILED:
1279 		case DMA_TRANS_WRITE_FAILED:
1280 			entry->errors++;
1281 		case DMA_TRANS_ABORTED:
1282 		{
1283 			struct ntb_transport_qp *qp = entry->qp;
1284 			void *offset = qp->rx_buff + qp->rx_max_frame *
1285 					qp->rx_index;
1286 
1287 			ntb_memcpy_rx(entry, offset);
1288 			qp->rx_memcpy++;
1289 			return;
1290 		}
1291 
1292 		case DMA_TRANS_NOERROR:
1293 		default:
1294 			break;
1295 		}
1296 	}
1297 
1298 	entry->flags |= DESC_DONE_FLAG;
1299 
1300 	ntb_complete_rxc(entry->qp);
1301 }
1302 
1303 static void ntb_memcpy_rx(struct ntb_queue_entry *entry, void *offset)
1304 {
1305 	void *buf = entry->buf;
1306 	size_t len = entry->len;
1307 
1308 	memcpy(buf, offset, len);
1309 
1310 	/* Ensure that the data is fully copied out before clearing the flag */
1311 	wmb();
1312 
1313 	ntb_rx_copy_callback(entry, NULL);
1314 }
1315 
1316 static int ntb_async_rx_submit(struct ntb_queue_entry *entry, void *offset)
1317 {
1318 	struct dma_async_tx_descriptor *txd;
1319 	struct ntb_transport_qp *qp = entry->qp;
1320 	struct dma_chan *chan = qp->rx_dma_chan;
1321 	struct dma_device *device;
1322 	size_t pay_off, buff_off, len;
1323 	struct dmaengine_unmap_data *unmap;
1324 	dma_cookie_t cookie;
1325 	void *buf = entry->buf;
1326 
1327 	len = entry->len;
1328 	device = chan->device;
1329 	pay_off = (size_t)offset & ~PAGE_MASK;
1330 	buff_off = (size_t)buf & ~PAGE_MASK;
1331 
1332 	if (!is_dma_copy_aligned(device, pay_off, buff_off, len))
1333 		goto err;
1334 
1335 	unmap = dmaengine_get_unmap_data(device->dev, 2, GFP_NOWAIT);
1336 	if (!unmap)
1337 		goto err;
1338 
1339 	unmap->len = len;
1340 	unmap->addr[0] = dma_map_page(device->dev, virt_to_page(offset),
1341 				      pay_off, len, DMA_TO_DEVICE);
1342 	if (dma_mapping_error(device->dev, unmap->addr[0]))
1343 		goto err_get_unmap;
1344 
1345 	unmap->to_cnt = 1;
1346 
1347 	unmap->addr[1] = dma_map_page(device->dev, virt_to_page(buf),
1348 				      buff_off, len, DMA_FROM_DEVICE);
1349 	if (dma_mapping_error(device->dev, unmap->addr[1]))
1350 		goto err_get_unmap;
1351 
1352 	unmap->from_cnt = 1;
1353 
1354 	txd = device->device_prep_dma_memcpy(chan, unmap->addr[1],
1355 					     unmap->addr[0], len,
1356 					     DMA_PREP_INTERRUPT);
1357 	if (!txd)
1358 		goto err_get_unmap;
1359 
1360 	txd->callback_result = ntb_rx_copy_callback;
1361 	txd->callback_param = entry;
1362 	dma_set_unmap(txd, unmap);
1363 
1364 	cookie = dmaengine_submit(txd);
1365 	if (dma_submit_error(cookie))
1366 		goto err_set_unmap;
1367 
1368 	dmaengine_unmap_put(unmap);
1369 
1370 	qp->last_cookie = cookie;
1371 
1372 	qp->rx_async++;
1373 
1374 	return 0;
1375 
1376 err_set_unmap:
1377 	dmaengine_unmap_put(unmap);
1378 err_get_unmap:
1379 	dmaengine_unmap_put(unmap);
1380 err:
1381 	return -ENXIO;
1382 }
1383 
1384 static void ntb_async_rx(struct ntb_queue_entry *entry, void *offset)
1385 {
1386 	struct ntb_transport_qp *qp = entry->qp;
1387 	struct dma_chan *chan = qp->rx_dma_chan;
1388 	int res;
1389 
1390 	if (!chan)
1391 		goto err;
1392 
1393 	if (entry->len < copy_bytes)
1394 		goto err;
1395 
1396 	res = ntb_async_rx_submit(entry, offset);
1397 	if (res < 0)
1398 		goto err;
1399 
1400 	if (!entry->retries)
1401 		qp->rx_async++;
1402 
1403 	return;
1404 
1405 err:
1406 	ntb_memcpy_rx(entry, offset);
1407 	qp->rx_memcpy++;
1408 }
1409 
1410 static int ntb_process_rxc(struct ntb_transport_qp *qp)
1411 {
1412 	struct ntb_payload_header *hdr;
1413 	struct ntb_queue_entry *entry;
1414 	void *offset;
1415 
1416 	offset = qp->rx_buff + qp->rx_max_frame * qp->rx_index;
1417 	hdr = offset + qp->rx_max_frame - sizeof(struct ntb_payload_header);
1418 
1419 	dev_dbg(&qp->ndev->pdev->dev, "qp %d: RX ver %u len %d flags %x\n",
1420 		qp->qp_num, hdr->ver, hdr->len, hdr->flags);
1421 
1422 	if (!(hdr->flags & DESC_DONE_FLAG)) {
1423 		dev_dbg(&qp->ndev->pdev->dev, "done flag not set\n");
1424 		qp->rx_ring_empty++;
1425 		return -EAGAIN;
1426 	}
1427 
1428 	if (hdr->flags & LINK_DOWN_FLAG) {
1429 		dev_dbg(&qp->ndev->pdev->dev, "link down flag set\n");
1430 		ntb_qp_link_down(qp);
1431 		hdr->flags = 0;
1432 		return -EAGAIN;
1433 	}
1434 
1435 	if (hdr->ver != (u32)qp->rx_pkts) {
1436 		dev_dbg(&qp->ndev->pdev->dev,
1437 			"version mismatch, expected %llu - got %u\n",
1438 			qp->rx_pkts, hdr->ver);
1439 		qp->rx_err_ver++;
1440 		return -EIO;
1441 	}
1442 
1443 	entry = ntb_list_mv(&qp->ntb_rx_q_lock, &qp->rx_pend_q, &qp->rx_post_q);
1444 	if (!entry) {
1445 		dev_dbg(&qp->ndev->pdev->dev, "no receive buffer\n");
1446 		qp->rx_err_no_buf++;
1447 		return -EAGAIN;
1448 	}
1449 
1450 	entry->rx_hdr = hdr;
1451 	entry->rx_index = qp->rx_index;
1452 
1453 	if (hdr->len > entry->len) {
1454 		dev_dbg(&qp->ndev->pdev->dev,
1455 			"receive buffer overflow! Wanted %d got %d\n",
1456 			hdr->len, entry->len);
1457 		qp->rx_err_oflow++;
1458 
1459 		entry->len = -EIO;
1460 		entry->flags |= DESC_DONE_FLAG;
1461 
1462 		ntb_complete_rxc(qp);
1463 	} else {
1464 		dev_dbg(&qp->ndev->pdev->dev,
1465 			"RX OK index %u ver %u size %d into buf size %d\n",
1466 			qp->rx_index, hdr->ver, hdr->len, entry->len);
1467 
1468 		qp->rx_bytes += hdr->len;
1469 		qp->rx_pkts++;
1470 
1471 		entry->len = hdr->len;
1472 
1473 		ntb_async_rx(entry, offset);
1474 	}
1475 
1476 	qp->rx_index++;
1477 	qp->rx_index %= qp->rx_max_entry;
1478 
1479 	return 0;
1480 }
1481 
1482 static void ntb_transport_rxc_db(unsigned long data)
1483 {
1484 	struct ntb_transport_qp *qp = (void *)data;
1485 	int rc, i;
1486 
1487 	dev_dbg(&qp->ndev->pdev->dev, "%s: doorbell %d received\n",
1488 		__func__, qp->qp_num);
1489 
1490 	/* Limit the number of packets processed in a single interrupt to
1491 	 * provide fairness to others
1492 	 */
1493 	for (i = 0; i < qp->rx_max_entry; i++) {
1494 		rc = ntb_process_rxc(qp);
1495 		if (rc)
1496 			break;
1497 	}
1498 
1499 	if (i && qp->rx_dma_chan)
1500 		dma_async_issue_pending(qp->rx_dma_chan);
1501 
1502 	if (i == qp->rx_max_entry) {
1503 		/* there is more work to do */
1504 		if (qp->active)
1505 			tasklet_schedule(&qp->rxc_db_work);
1506 	} else if (ntb_db_read(qp->ndev) & BIT_ULL(qp->qp_num)) {
1507 		/* the doorbell bit is set: clear it */
1508 		ntb_db_clear(qp->ndev, BIT_ULL(qp->qp_num));
1509 		/* ntb_db_read ensures ntb_db_clear write is committed */
1510 		ntb_db_read(qp->ndev);
1511 
1512 		/* an interrupt may have arrived between finishing
1513 		 * ntb_process_rxc and clearing the doorbell bit:
1514 		 * there might be some more work to do.
1515 		 */
1516 		if (qp->active)
1517 			tasklet_schedule(&qp->rxc_db_work);
1518 	}
1519 }
1520 
1521 static void ntb_tx_copy_callback(void *data,
1522 				 const struct dmaengine_result *res)
1523 {
1524 	struct ntb_queue_entry *entry = data;
1525 	struct ntb_transport_qp *qp = entry->qp;
1526 	struct ntb_payload_header __iomem *hdr = entry->tx_hdr;
1527 
1528 	/* we need to check DMA results if we are using DMA */
1529 	if (res) {
1530 		enum dmaengine_tx_result dma_err = res->result;
1531 
1532 		switch (dma_err) {
1533 		case DMA_TRANS_READ_FAILED:
1534 		case DMA_TRANS_WRITE_FAILED:
1535 			entry->errors++;
1536 		case DMA_TRANS_ABORTED:
1537 		{
1538 			void __iomem *offset =
1539 				qp->tx_mw + qp->tx_max_frame *
1540 				entry->tx_index;
1541 
1542 			/* resubmit via CPU */
1543 			ntb_memcpy_tx(entry, offset);
1544 			qp->tx_memcpy++;
1545 			return;
1546 		}
1547 
1548 		case DMA_TRANS_NOERROR:
1549 		default:
1550 			break;
1551 		}
1552 	}
1553 
1554 	iowrite32(entry->flags | DESC_DONE_FLAG, &hdr->flags);
1555 
1556 	ntb_peer_db_set(qp->ndev, BIT_ULL(qp->qp_num));
1557 
1558 	/* The entry length can only be zero if the packet is intended to be a
1559 	 * "link down" or similar.  Since no payload is being sent in these
1560 	 * cases, there is nothing to add to the completion queue.
1561 	 */
1562 	if (entry->len > 0) {
1563 		qp->tx_bytes += entry->len;
1564 
1565 		if (qp->tx_handler)
1566 			qp->tx_handler(qp, qp->cb_data, entry->cb_data,
1567 				       entry->len);
1568 	}
1569 
1570 	ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry, &qp->tx_free_q);
1571 }
1572 
1573 static void ntb_memcpy_tx(struct ntb_queue_entry *entry, void __iomem *offset)
1574 {
1575 #ifdef ARCH_HAS_NOCACHE_UACCESS
1576 	/*
1577 	 * Using non-temporal mov to improve performance on non-cached
1578 	 * writes, even though we aren't actually copying from user space.
1579 	 */
1580 	__copy_from_user_inatomic_nocache(offset, entry->buf, entry->len);
1581 #else
1582 	memcpy_toio(offset, entry->buf, entry->len);
1583 #endif
1584 
1585 	/* Ensure that the data is fully copied out before setting the flags */
1586 	wmb();
1587 
1588 	ntb_tx_copy_callback(entry, NULL);
1589 }
1590 
1591 static int ntb_async_tx_submit(struct ntb_transport_qp *qp,
1592 			       struct ntb_queue_entry *entry)
1593 {
1594 	struct dma_async_tx_descriptor *txd;
1595 	struct dma_chan *chan = qp->tx_dma_chan;
1596 	struct dma_device *device;
1597 	size_t len = entry->len;
1598 	void *buf = entry->buf;
1599 	size_t dest_off, buff_off;
1600 	struct dmaengine_unmap_data *unmap;
1601 	dma_addr_t dest;
1602 	dma_cookie_t cookie;
1603 
1604 	device = chan->device;
1605 	dest = qp->tx_mw_phys + qp->tx_max_frame * entry->tx_index;
1606 	buff_off = (size_t)buf & ~PAGE_MASK;
1607 	dest_off = (size_t)dest & ~PAGE_MASK;
1608 
1609 	if (!is_dma_copy_aligned(device, buff_off, dest_off, len))
1610 		goto err;
1611 
1612 	unmap = dmaengine_get_unmap_data(device->dev, 1, GFP_NOWAIT);
1613 	if (!unmap)
1614 		goto err;
1615 
1616 	unmap->len = len;
1617 	unmap->addr[0] = dma_map_page(device->dev, virt_to_page(buf),
1618 				      buff_off, len, DMA_TO_DEVICE);
1619 	if (dma_mapping_error(device->dev, unmap->addr[0]))
1620 		goto err_get_unmap;
1621 
1622 	unmap->to_cnt = 1;
1623 
1624 	txd = device->device_prep_dma_memcpy(chan, dest, unmap->addr[0], len,
1625 					     DMA_PREP_INTERRUPT);
1626 	if (!txd)
1627 		goto err_get_unmap;
1628 
1629 	txd->callback_result = ntb_tx_copy_callback;
1630 	txd->callback_param = entry;
1631 	dma_set_unmap(txd, unmap);
1632 
1633 	cookie = dmaengine_submit(txd);
1634 	if (dma_submit_error(cookie))
1635 		goto err_set_unmap;
1636 
1637 	dmaengine_unmap_put(unmap);
1638 
1639 	dma_async_issue_pending(chan);
1640 
1641 	return 0;
1642 err_set_unmap:
1643 	dmaengine_unmap_put(unmap);
1644 err_get_unmap:
1645 	dmaengine_unmap_put(unmap);
1646 err:
1647 	return -ENXIO;
1648 }
1649 
1650 static void ntb_async_tx(struct ntb_transport_qp *qp,
1651 			 struct ntb_queue_entry *entry)
1652 {
1653 	struct ntb_payload_header __iomem *hdr;
1654 	struct dma_chan *chan = qp->tx_dma_chan;
1655 	void __iomem *offset;
1656 	int res;
1657 
1658 	entry->tx_index = qp->tx_index;
1659 	offset = qp->tx_mw + qp->tx_max_frame * entry->tx_index;
1660 	hdr = offset + qp->tx_max_frame - sizeof(struct ntb_payload_header);
1661 	entry->tx_hdr = hdr;
1662 
1663 	iowrite32(entry->len, &hdr->len);
1664 	iowrite32((u32)qp->tx_pkts, &hdr->ver);
1665 
1666 	if (!chan)
1667 		goto err;
1668 
1669 	if (entry->len < copy_bytes)
1670 		goto err;
1671 
1672 	res = ntb_async_tx_submit(qp, entry);
1673 	if (res < 0)
1674 		goto err;
1675 
1676 	if (!entry->retries)
1677 		qp->tx_async++;
1678 
1679 	return;
1680 
1681 err:
1682 	ntb_memcpy_tx(entry, offset);
1683 	qp->tx_memcpy++;
1684 }
1685 
1686 static int ntb_process_tx(struct ntb_transport_qp *qp,
1687 			  struct ntb_queue_entry *entry)
1688 {
1689 	if (qp->tx_index == qp->remote_rx_info->entry) {
1690 		qp->tx_ring_full++;
1691 		return -EAGAIN;
1692 	}
1693 
1694 	if (entry->len > qp->tx_max_frame - sizeof(struct ntb_payload_header)) {
1695 		if (qp->tx_handler)
1696 			qp->tx_handler(qp, qp->cb_data, NULL, -EIO);
1697 
1698 		ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
1699 			     &qp->tx_free_q);
1700 		return 0;
1701 	}
1702 
1703 	ntb_async_tx(qp, entry);
1704 
1705 	qp->tx_index++;
1706 	qp->tx_index %= qp->tx_max_entry;
1707 
1708 	qp->tx_pkts++;
1709 
1710 	return 0;
1711 }
1712 
1713 static void ntb_send_link_down(struct ntb_transport_qp *qp)
1714 {
1715 	struct pci_dev *pdev = qp->ndev->pdev;
1716 	struct ntb_queue_entry *entry;
1717 	int i, rc;
1718 
1719 	if (!qp->link_is_up)
1720 		return;
1721 
1722 	dev_info(&pdev->dev, "qp %d: Send Link Down\n", qp->qp_num);
1723 
1724 	for (i = 0; i < NTB_LINK_DOWN_TIMEOUT; i++) {
1725 		entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
1726 		if (entry)
1727 			break;
1728 		msleep(100);
1729 	}
1730 
1731 	if (!entry)
1732 		return;
1733 
1734 	entry->cb_data = NULL;
1735 	entry->buf = NULL;
1736 	entry->len = 0;
1737 	entry->flags = LINK_DOWN_FLAG;
1738 
1739 	rc = ntb_process_tx(qp, entry);
1740 	if (rc)
1741 		dev_err(&pdev->dev, "ntb: QP%d unable to send linkdown msg\n",
1742 			qp->qp_num);
1743 
1744 	ntb_qp_link_down_reset(qp);
1745 }
1746 
1747 static bool ntb_dma_filter_fn(struct dma_chan *chan, void *node)
1748 {
1749 	return dev_to_node(&chan->dev->device) == (int)(unsigned long)node;
1750 }
1751 
1752 /**
1753  * ntb_transport_create_queue - Create a new NTB transport layer queue
1754  * @rx_handler: receive callback function
1755  * @tx_handler: transmit callback function
1756  * @event_handler: event callback function
1757  *
1758  * Create a new NTB transport layer queue and provide the queue with a callback
1759  * routine for both transmit and receive.  The receive callback routine will be
1760  * used to pass up data when the transport has received it on the queue.   The
1761  * transmit callback routine will be called when the transport has completed the
1762  * transmission of the data on the queue and the data is ready to be freed.
1763  *
1764  * RETURNS: pointer to newly created ntb_queue, NULL on error.
1765  */
1766 struct ntb_transport_qp *
1767 ntb_transport_create_queue(void *data, struct device *client_dev,
1768 			   const struct ntb_queue_handlers *handlers)
1769 {
1770 	struct ntb_dev *ndev;
1771 	struct pci_dev *pdev;
1772 	struct ntb_transport_ctx *nt;
1773 	struct ntb_queue_entry *entry;
1774 	struct ntb_transport_qp *qp;
1775 	u64 qp_bit;
1776 	unsigned int free_queue;
1777 	dma_cap_mask_t dma_mask;
1778 	int node;
1779 	int i;
1780 
1781 	ndev = dev_ntb(client_dev->parent);
1782 	pdev = ndev->pdev;
1783 	nt = ndev->ctx;
1784 
1785 	node = dev_to_node(&ndev->dev);
1786 
1787 	free_queue = ffs(nt->qp_bitmap_free);
1788 	if (!free_queue)
1789 		goto err;
1790 
1791 	/* decrement free_queue to make it zero based */
1792 	free_queue--;
1793 
1794 	qp = &nt->qp_vec[free_queue];
1795 	qp_bit = BIT_ULL(qp->qp_num);
1796 
1797 	nt->qp_bitmap_free &= ~qp_bit;
1798 
1799 	qp->cb_data = data;
1800 	qp->rx_handler = handlers->rx_handler;
1801 	qp->tx_handler = handlers->tx_handler;
1802 	qp->event_handler = handlers->event_handler;
1803 
1804 	dma_cap_zero(dma_mask);
1805 	dma_cap_set(DMA_MEMCPY, dma_mask);
1806 
1807 	if (use_dma) {
1808 		qp->tx_dma_chan =
1809 			dma_request_channel(dma_mask, ntb_dma_filter_fn,
1810 					    (void *)(unsigned long)node);
1811 		if (!qp->tx_dma_chan)
1812 			dev_info(&pdev->dev, "Unable to allocate TX DMA channel\n");
1813 
1814 		qp->rx_dma_chan =
1815 			dma_request_channel(dma_mask, ntb_dma_filter_fn,
1816 					    (void *)(unsigned long)node);
1817 		if (!qp->rx_dma_chan)
1818 			dev_info(&pdev->dev, "Unable to allocate RX DMA channel\n");
1819 	} else {
1820 		qp->tx_dma_chan = NULL;
1821 		qp->rx_dma_chan = NULL;
1822 	}
1823 
1824 	dev_dbg(&pdev->dev, "Using %s memcpy for TX\n",
1825 		qp->tx_dma_chan ? "DMA" : "CPU");
1826 
1827 	dev_dbg(&pdev->dev, "Using %s memcpy for RX\n",
1828 		qp->rx_dma_chan ? "DMA" : "CPU");
1829 
1830 	for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) {
1831 		entry = kzalloc_node(sizeof(*entry), GFP_ATOMIC, node);
1832 		if (!entry)
1833 			goto err1;
1834 
1835 		entry->qp = qp;
1836 		ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry,
1837 			     &qp->rx_free_q);
1838 	}
1839 	qp->rx_alloc_entry = NTB_QP_DEF_NUM_ENTRIES;
1840 
1841 	for (i = 0; i < qp->tx_max_entry; i++) {
1842 		entry = kzalloc_node(sizeof(*entry), GFP_ATOMIC, node);
1843 		if (!entry)
1844 			goto err2;
1845 
1846 		entry->qp = qp;
1847 		ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
1848 			     &qp->tx_free_q);
1849 	}
1850 
1851 	ntb_db_clear(qp->ndev, qp_bit);
1852 	ntb_db_clear_mask(qp->ndev, qp_bit);
1853 
1854 	dev_info(&pdev->dev, "NTB Transport QP %d created\n", qp->qp_num);
1855 
1856 	return qp;
1857 
1858 err2:
1859 	while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
1860 		kfree(entry);
1861 err1:
1862 	qp->rx_alloc_entry = 0;
1863 	while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q)))
1864 		kfree(entry);
1865 	if (qp->tx_dma_chan)
1866 		dma_release_channel(qp->tx_dma_chan);
1867 	if (qp->rx_dma_chan)
1868 		dma_release_channel(qp->rx_dma_chan);
1869 	nt->qp_bitmap_free |= qp_bit;
1870 err:
1871 	return NULL;
1872 }
1873 EXPORT_SYMBOL_GPL(ntb_transport_create_queue);
1874 
1875 /**
1876  * ntb_transport_free_queue - Frees NTB transport queue
1877  * @qp: NTB queue to be freed
1878  *
1879  * Frees NTB transport queue
1880  */
1881 void ntb_transport_free_queue(struct ntb_transport_qp *qp)
1882 {
1883 	struct pci_dev *pdev;
1884 	struct ntb_queue_entry *entry;
1885 	u64 qp_bit;
1886 
1887 	if (!qp)
1888 		return;
1889 
1890 	pdev = qp->ndev->pdev;
1891 
1892 	qp->active = false;
1893 
1894 	if (qp->tx_dma_chan) {
1895 		struct dma_chan *chan = qp->tx_dma_chan;
1896 		/* Putting the dma_chan to NULL will force any new traffic to be
1897 		 * processed by the CPU instead of the DAM engine
1898 		 */
1899 		qp->tx_dma_chan = NULL;
1900 
1901 		/* Try to be nice and wait for any queued DMA engine
1902 		 * transactions to process before smashing it with a rock
1903 		 */
1904 		dma_sync_wait(chan, qp->last_cookie);
1905 		dmaengine_terminate_all(chan);
1906 		dma_release_channel(chan);
1907 	}
1908 
1909 	if (qp->rx_dma_chan) {
1910 		struct dma_chan *chan = qp->rx_dma_chan;
1911 		/* Putting the dma_chan to NULL will force any new traffic to be
1912 		 * processed by the CPU instead of the DAM engine
1913 		 */
1914 		qp->rx_dma_chan = NULL;
1915 
1916 		/* Try to be nice and wait for any queued DMA engine
1917 		 * transactions to process before smashing it with a rock
1918 		 */
1919 		dma_sync_wait(chan, qp->last_cookie);
1920 		dmaengine_terminate_all(chan);
1921 		dma_release_channel(chan);
1922 	}
1923 
1924 	qp_bit = BIT_ULL(qp->qp_num);
1925 
1926 	ntb_db_set_mask(qp->ndev, qp_bit);
1927 	tasklet_kill(&qp->rxc_db_work);
1928 
1929 	cancel_delayed_work_sync(&qp->link_work);
1930 
1931 	qp->cb_data = NULL;
1932 	qp->rx_handler = NULL;
1933 	qp->tx_handler = NULL;
1934 	qp->event_handler = NULL;
1935 
1936 	while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q)))
1937 		kfree(entry);
1938 
1939 	while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_pend_q))) {
1940 		dev_warn(&pdev->dev, "Freeing item from non-empty rx_pend_q\n");
1941 		kfree(entry);
1942 	}
1943 
1944 	while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_post_q))) {
1945 		dev_warn(&pdev->dev, "Freeing item from non-empty rx_post_q\n");
1946 		kfree(entry);
1947 	}
1948 
1949 	while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
1950 		kfree(entry);
1951 
1952 	qp->transport->qp_bitmap_free |= qp_bit;
1953 
1954 	dev_info(&pdev->dev, "NTB Transport QP %d freed\n", qp->qp_num);
1955 }
1956 EXPORT_SYMBOL_GPL(ntb_transport_free_queue);
1957 
1958 /**
1959  * ntb_transport_rx_remove - Dequeues enqueued rx packet
1960  * @qp: NTB queue to be freed
1961  * @len: pointer to variable to write enqueued buffers length
1962  *
1963  * Dequeues unused buffers from receive queue.  Should only be used during
1964  * shutdown of qp.
1965  *
1966  * RETURNS: NULL error value on error, or void* for success.
1967  */
1968 void *ntb_transport_rx_remove(struct ntb_transport_qp *qp, unsigned int *len)
1969 {
1970 	struct ntb_queue_entry *entry;
1971 	void *buf;
1972 
1973 	if (!qp || qp->client_ready)
1974 		return NULL;
1975 
1976 	entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_pend_q);
1977 	if (!entry)
1978 		return NULL;
1979 
1980 	buf = entry->cb_data;
1981 	*len = entry->len;
1982 
1983 	ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry, &qp->rx_free_q);
1984 
1985 	return buf;
1986 }
1987 EXPORT_SYMBOL_GPL(ntb_transport_rx_remove);
1988 
1989 /**
1990  * ntb_transport_rx_enqueue - Enqueue a new NTB queue entry
1991  * @qp: NTB transport layer queue the entry is to be enqueued on
1992  * @cb: per buffer pointer for callback function to use
1993  * @data: pointer to data buffer that incoming packets will be copied into
1994  * @len: length of the data buffer
1995  *
1996  * Enqueue a new receive buffer onto the transport queue into which a NTB
1997  * payload can be received into.
1998  *
1999  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
2000  */
2001 int ntb_transport_rx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
2002 			     unsigned int len)
2003 {
2004 	struct ntb_queue_entry *entry;
2005 
2006 	if (!qp)
2007 		return -EINVAL;
2008 
2009 	entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q);
2010 	if (!entry)
2011 		return -ENOMEM;
2012 
2013 	entry->cb_data = cb;
2014 	entry->buf = data;
2015 	entry->len = len;
2016 	entry->flags = 0;
2017 	entry->retries = 0;
2018 	entry->errors = 0;
2019 	entry->rx_index = 0;
2020 
2021 	ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry, &qp->rx_pend_q);
2022 
2023 	if (qp->active)
2024 		tasklet_schedule(&qp->rxc_db_work);
2025 
2026 	return 0;
2027 }
2028 EXPORT_SYMBOL_GPL(ntb_transport_rx_enqueue);
2029 
2030 /**
2031  * ntb_transport_tx_enqueue - Enqueue a new NTB queue entry
2032  * @qp: NTB transport layer queue the entry is to be enqueued on
2033  * @cb: per buffer pointer for callback function to use
2034  * @data: pointer to data buffer that will be sent
2035  * @len: length of the data buffer
2036  *
2037  * Enqueue a new transmit buffer onto the transport queue from which a NTB
2038  * payload will be transmitted.  This assumes that a lock is being held to
2039  * serialize access to the qp.
2040  *
2041  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
2042  */
2043 int ntb_transport_tx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
2044 			     unsigned int len)
2045 {
2046 	struct ntb_queue_entry *entry;
2047 	int rc;
2048 
2049 	if (!qp || !qp->link_is_up || !len)
2050 		return -EINVAL;
2051 
2052 	entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
2053 	if (!entry) {
2054 		qp->tx_err_no_buf++;
2055 		return -EBUSY;
2056 	}
2057 
2058 	entry->cb_data = cb;
2059 	entry->buf = data;
2060 	entry->len = len;
2061 	entry->flags = 0;
2062 	entry->errors = 0;
2063 	entry->retries = 0;
2064 	entry->tx_index = 0;
2065 
2066 	rc = ntb_process_tx(qp, entry);
2067 	if (rc)
2068 		ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
2069 			     &qp->tx_free_q);
2070 
2071 	return rc;
2072 }
2073 EXPORT_SYMBOL_GPL(ntb_transport_tx_enqueue);
2074 
2075 /**
2076  * ntb_transport_link_up - Notify NTB transport of client readiness to use queue
2077  * @qp: NTB transport layer queue to be enabled
2078  *
2079  * Notify NTB transport layer of client readiness to use queue
2080  */
2081 void ntb_transport_link_up(struct ntb_transport_qp *qp)
2082 {
2083 	if (!qp)
2084 		return;
2085 
2086 	qp->client_ready = true;
2087 
2088 	if (qp->transport->link_is_up)
2089 		schedule_delayed_work(&qp->link_work, 0);
2090 }
2091 EXPORT_SYMBOL_GPL(ntb_transport_link_up);
2092 
2093 /**
2094  * ntb_transport_link_down - Notify NTB transport to no longer enqueue data
2095  * @qp: NTB transport layer queue to be disabled
2096  *
2097  * Notify NTB transport layer of client's desire to no longer receive data on
2098  * transport queue specified.  It is the client's responsibility to ensure all
2099  * entries on queue are purged or otherwise handled appropriately.
2100  */
2101 void ntb_transport_link_down(struct ntb_transport_qp *qp)
2102 {
2103 	int val;
2104 
2105 	if (!qp)
2106 		return;
2107 
2108 	qp->client_ready = false;
2109 
2110 	val = ntb_spad_read(qp->ndev, QP_LINKS);
2111 
2112 	ntb_peer_spad_write(qp->ndev, PIDX, QP_LINKS, val & ~BIT(qp->qp_num));
2113 
2114 	if (qp->link_is_up)
2115 		ntb_send_link_down(qp);
2116 	else
2117 		cancel_delayed_work_sync(&qp->link_work);
2118 }
2119 EXPORT_SYMBOL_GPL(ntb_transport_link_down);
2120 
2121 /**
2122  * ntb_transport_link_query - Query transport link state
2123  * @qp: NTB transport layer queue to be queried
2124  *
2125  * Query connectivity to the remote system of the NTB transport queue
2126  *
2127  * RETURNS: true for link up or false for link down
2128  */
2129 bool ntb_transport_link_query(struct ntb_transport_qp *qp)
2130 {
2131 	if (!qp)
2132 		return false;
2133 
2134 	return qp->link_is_up;
2135 }
2136 EXPORT_SYMBOL_GPL(ntb_transport_link_query);
2137 
2138 /**
2139  * ntb_transport_qp_num - Query the qp number
2140  * @qp: NTB transport layer queue to be queried
2141  *
2142  * Query qp number of the NTB transport queue
2143  *
2144  * RETURNS: a zero based number specifying the qp number
2145  */
2146 unsigned char ntb_transport_qp_num(struct ntb_transport_qp *qp)
2147 {
2148 	if (!qp)
2149 		return 0;
2150 
2151 	return qp->qp_num;
2152 }
2153 EXPORT_SYMBOL_GPL(ntb_transport_qp_num);
2154 
2155 /**
2156  * ntb_transport_max_size - Query the max payload size of a qp
2157  * @qp: NTB transport layer queue to be queried
2158  *
2159  * Query the maximum payload size permissible on the given qp
2160  *
2161  * RETURNS: the max payload size of a qp
2162  */
2163 unsigned int ntb_transport_max_size(struct ntb_transport_qp *qp)
2164 {
2165 	unsigned int max_size;
2166 	unsigned int copy_align;
2167 	struct dma_chan *rx_chan, *tx_chan;
2168 
2169 	if (!qp)
2170 		return 0;
2171 
2172 	rx_chan = qp->rx_dma_chan;
2173 	tx_chan = qp->tx_dma_chan;
2174 
2175 	copy_align = max(rx_chan ? rx_chan->device->copy_align : 0,
2176 			 tx_chan ? tx_chan->device->copy_align : 0);
2177 
2178 	/* If DMA engine usage is possible, try to find the max size for that */
2179 	max_size = qp->tx_max_frame - sizeof(struct ntb_payload_header);
2180 	max_size = round_down(max_size, 1 << copy_align);
2181 
2182 	return max_size;
2183 }
2184 EXPORT_SYMBOL_GPL(ntb_transport_max_size);
2185 
2186 unsigned int ntb_transport_tx_free_entry(struct ntb_transport_qp *qp)
2187 {
2188 	unsigned int head = qp->tx_index;
2189 	unsigned int tail = qp->remote_rx_info->entry;
2190 
2191 	return tail > head ? tail - head : qp->tx_max_entry + tail - head;
2192 }
2193 EXPORT_SYMBOL_GPL(ntb_transport_tx_free_entry);
2194 
2195 static void ntb_transport_doorbell_callback(void *data, int vector)
2196 {
2197 	struct ntb_transport_ctx *nt = data;
2198 	struct ntb_transport_qp *qp;
2199 	u64 db_bits;
2200 	unsigned int qp_num;
2201 
2202 	db_bits = (nt->qp_bitmap & ~nt->qp_bitmap_free &
2203 		   ntb_db_vector_mask(nt->ndev, vector));
2204 
2205 	while (db_bits) {
2206 		qp_num = __ffs(db_bits);
2207 		qp = &nt->qp_vec[qp_num];
2208 
2209 		if (qp->active)
2210 			tasklet_schedule(&qp->rxc_db_work);
2211 
2212 		db_bits &= ~BIT_ULL(qp_num);
2213 	}
2214 }
2215 
2216 static const struct ntb_ctx_ops ntb_transport_ops = {
2217 	.link_event = ntb_transport_event_callback,
2218 	.db_event = ntb_transport_doorbell_callback,
2219 };
2220 
2221 static struct ntb_client ntb_transport_client = {
2222 	.ops = {
2223 		.probe = ntb_transport_probe,
2224 		.remove = ntb_transport_free,
2225 	},
2226 };
2227 
2228 static int __init ntb_transport_init(void)
2229 {
2230 	int rc;
2231 
2232 	pr_info("%s, version %s\n", NTB_TRANSPORT_DESC, NTB_TRANSPORT_VER);
2233 
2234 	if (debugfs_initialized())
2235 		nt_debugfs_dir = debugfs_create_dir(KBUILD_MODNAME, NULL);
2236 
2237 	rc = bus_register(&ntb_transport_bus);
2238 	if (rc)
2239 		goto err_bus;
2240 
2241 	rc = ntb_register_client(&ntb_transport_client);
2242 	if (rc)
2243 		goto err_client;
2244 
2245 	return 0;
2246 
2247 err_client:
2248 	bus_unregister(&ntb_transport_bus);
2249 err_bus:
2250 	debugfs_remove_recursive(nt_debugfs_dir);
2251 	return rc;
2252 }
2253 module_init(ntb_transport_init);
2254 
2255 static void __exit ntb_transport_exit(void)
2256 {
2257 	ntb_unregister_client(&ntb_transport_client);
2258 	bus_unregister(&ntb_transport_bus);
2259 	debugfs_remove_recursive(nt_debugfs_dir);
2260 }
2261 module_exit(ntb_transport_exit);
2262