?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
drm/nouveau: fix kernel-doc comments
[drm/drm-misc.git] / drivers / ntb / ntb_transport.c
bloba22ea4a4b202bd63fe5e027a69d540551e15a80a
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 bool use_msi;
97 #ifdef CONFIG_NTB_MSI
98 module_param(use_msi, bool, 0644);
99 MODULE_PARM_DESC(use_msi, "Use MSI interrupts instead of doorbells");
100 #endif
101
102 static struct dentry *nt_debugfs_dir;
103
104 /* Only two-ports NTB devices are supported */
105 #define PIDX NTB_DEF_PEER_IDX
106
107 struct ntb_queue_entry {
108 /* ntb_queue list reference */
109 struct list_head entry;
110 /* pointers to data to be transferred */
111 void *cb_data;
112 void *buf;
113 unsigned int len;
114 unsigned int flags;
115 int retries;
116 int errors;
117 unsigned int tx_index;
118 unsigned int rx_index;
119
120 struct ntb_transport_qp *qp;
121 union {
122 struct ntb_payload_header __iomem *tx_hdr;
123 struct ntb_payload_header *rx_hdr;
124 };
125 };
126
127 struct ntb_rx_info {
128 unsigned int entry;
129 };
130
131 struct ntb_transport_qp {
132 struct ntb_transport_ctx *transport;
133 struct ntb_dev *ndev;
134 void *cb_data;
135 struct dma_chan *tx_dma_chan;
136 struct dma_chan *rx_dma_chan;
137
138 bool client_ready;
139 bool link_is_up;
140 bool active;
141
142 u8 qp_num; /* Only 64 QP's are allowed. 0-63 */
143 u64 qp_bit;
144
145 struct ntb_rx_info __iomem *rx_info;
146 struct ntb_rx_info *remote_rx_info;
147
148 void (*tx_handler)(struct ntb_transport_qp *qp, void *qp_data,
149 void *data, int len);
150 struct list_head tx_free_q;
151 spinlock_t ntb_tx_free_q_lock;
152 void __iomem *tx_mw;
153 phys_addr_t tx_mw_phys;
154 size_t tx_mw_size;
155 dma_addr_t tx_mw_dma_addr;
156 unsigned int tx_index;
157 unsigned int tx_max_entry;
158 unsigned int tx_max_frame;
159
160 void (*rx_handler)(struct ntb_transport_qp *qp, void *qp_data,
161 void *data, int len);
162 struct list_head rx_post_q;
163 struct list_head rx_pend_q;
164 struct list_head rx_free_q;
165 /* ntb_rx_q_lock: synchronize access to rx_XXXX_q */
166 spinlock_t ntb_rx_q_lock;
167 void *rx_buff;
168 unsigned int rx_index;
169 unsigned int rx_max_entry;
170 unsigned int rx_max_frame;
171 unsigned int rx_alloc_entry;
172 dma_cookie_t last_cookie;
173 struct tasklet_struct rxc_db_work;
174
175 void (*event_handler)(void *data, int status);
176 struct delayed_work link_work;
177 struct work_struct link_cleanup;
178
179 struct dentry *debugfs_dir;
180 struct dentry *debugfs_stats;
181
182 /* Stats */
183 u64 rx_bytes;
184 u64 rx_pkts;
185 u64 rx_ring_empty;
186 u64 rx_err_no_buf;
187 u64 rx_err_oflow;
188 u64 rx_err_ver;
189 u64 rx_memcpy;
190 u64 rx_async;
191 u64 tx_bytes;
192 u64 tx_pkts;
193 u64 tx_ring_full;
194 u64 tx_err_no_buf;
195 u64 tx_memcpy;
196 u64 tx_async;
197
198 bool use_msi;
199 int msi_irq;
200 struct ntb_msi_desc msi_desc;
201 struct ntb_msi_desc peer_msi_desc;
202 };
203
204 struct ntb_transport_mw {
205 phys_addr_t phys_addr;
206 resource_size_t phys_size;
207 void __iomem *vbase;
208 size_t xlat_size;
209 size_t buff_size;
210 size_t alloc_size;
211 void *alloc_addr;
212 void *virt_addr;
213 dma_addr_t dma_addr;
214 };
215
216 struct ntb_transport_client_dev {
217 struct list_head entry;
218 struct ntb_transport_ctx *nt;
219 struct device dev;
220 };
221
222 struct ntb_transport_ctx {
223 struct list_head entry;
224 struct list_head client_devs;
225
226 struct ntb_dev *ndev;
227
228 struct ntb_transport_mw *mw_vec;
229 struct ntb_transport_qp *qp_vec;
230 unsigned int mw_count;
231 unsigned int qp_count;
232 u64 qp_bitmap;
233 u64 qp_bitmap_free;
234
235 bool use_msi;
236 unsigned int msi_spad_offset;
237 u64 msi_db_mask;
238
239 bool link_is_up;
240 struct delayed_work link_work;
241 struct work_struct link_cleanup;
242
243 struct dentry *debugfs_node_dir;
244 };
245
246 enum {
247 DESC_DONE_FLAG = BIT(0),
248 LINK_DOWN_FLAG = BIT(1),
249 };
250
251 struct ntb_payload_header {
252 unsigned int ver;
253 unsigned int len;
254 unsigned int flags;
255 };
256
257 enum {
258 VERSION = 0,
259 QP_LINKS,
260 NUM_QPS,
261 NUM_MWS,
262 MW0_SZ_HIGH,
263 MW0_SZ_LOW,
264 };
265
266 #define dev_client_dev(__dev) \
267 container_of((__dev), struct ntb_transport_client_dev, dev)
268
269 #define drv_client(__drv) \
270 container_of((__drv), struct ntb_transport_client, driver)
271
272 #define QP_TO_MW(nt, qp) ((qp) % nt->mw_count)
273 #define NTB_QP_DEF_NUM_ENTRIES 100
274 #define NTB_LINK_DOWN_TIMEOUT 10
275
276 static void ntb_transport_rxc_db(unsigned long data);
277 static const struct ntb_ctx_ops ntb_transport_ops;
278 static struct ntb_client ntb_transport_client;
279 static int ntb_async_tx_submit(struct ntb_transport_qp *qp,
280 struct ntb_queue_entry *entry);
281 static void ntb_memcpy_tx(struct ntb_queue_entry *entry, void __iomem *offset);
282 static int ntb_async_rx_submit(struct ntb_queue_entry *entry, void *offset);
283 static void ntb_memcpy_rx(struct ntb_queue_entry *entry, void *offset);
284
285
286 static int ntb_transport_bus_match(struct device *dev,
287 const struct device_driver *drv)
288 {
289 return !strncmp(dev_name(dev), drv->name, strlen(drv->name));
290 }
291
292 static int ntb_transport_bus_probe(struct device *dev)
293 {
294 const struct ntb_transport_client *client;
295 int rc;
296
297 get_device(dev);
298
299 client = drv_client(dev->driver);
300 rc = client->probe(dev);
301 if (rc)
302 put_device(dev);
303
304 return rc;
305 }
306
307 static void ntb_transport_bus_remove(struct device *dev)
308 {
309 const struct ntb_transport_client *client;
310
311 client = drv_client(dev->driver);
312 client->remove(dev);
313
314 put_device(dev);
315 }
316
317 static const struct bus_type ntb_transport_bus = {
318 .name = "ntb_transport",
319 .match = ntb_transport_bus_match,
320 .probe = ntb_transport_bus_probe,
321 .remove = ntb_transport_bus_remove,
322 };
323
324 static LIST_HEAD(ntb_transport_list);
325
326 static int ntb_bus_init(struct ntb_transport_ctx *nt)
327 {
328 list_add_tail(&nt->entry, &ntb_transport_list);
329 return 0;
330 }
331
332 static void ntb_bus_remove(struct ntb_transport_ctx *nt)
333 {
334 struct ntb_transport_client_dev *client_dev, *cd;
335
336 list_for_each_entry_safe(client_dev, cd, &nt->client_devs, entry) {
337 dev_err(client_dev->dev.parent, "%s still attached to bus, removing\n",
338 dev_name(&client_dev->dev));
339 list_del(&client_dev->entry);
340 device_unregister(&client_dev->dev);
341 }
342
343 list_del(&nt->entry);
344 }
345
346 static void ntb_transport_client_release(struct device *dev)
347 {
348 struct ntb_transport_client_dev *client_dev;
349
350 client_dev = dev_client_dev(dev);
351 kfree(client_dev);
352 }
353
354 /**
355 * ntb_transport_unregister_client_dev - Unregister NTB client device
356 * @device_name: Name of NTB client device
357 *
358 * Unregister an NTB client device with the NTB transport layer
359 */
360 void ntb_transport_unregister_client_dev(char *device_name)
361 {
362 struct ntb_transport_client_dev *client, *cd;
363 struct ntb_transport_ctx *nt;
364
365 list_for_each_entry(nt, &ntb_transport_list, entry)
366 list_for_each_entry_safe(client, cd, &nt->client_devs, entry)
367 if (!strncmp(dev_name(&client->dev), device_name,
368 strlen(device_name))) {
369 list_del(&client->entry);
370 device_unregister(&client->dev);
371 }
372 }
373 EXPORT_SYMBOL_GPL(ntb_transport_unregister_client_dev);
374
375 /**
376 * ntb_transport_register_client_dev - Register NTB client device
377 * @device_name: Name of NTB client device
378 *
379 * Register an NTB client device with the NTB transport layer
380 *
381 * Returns: %0 on success or -errno code on error
382 */
383 int ntb_transport_register_client_dev(char *device_name)
384 {
385 struct ntb_transport_client_dev *client_dev;
386 struct ntb_transport_ctx *nt;
387 int node;
388 int rc, i = 0;
389
390 if (list_empty(&ntb_transport_list))
391 return -ENODEV;
392
393 list_for_each_entry(nt, &ntb_transport_list, entry) {
394 struct device *dev;
395
396 node = dev_to_node(&nt->ndev->dev);
397
398 client_dev = kzalloc_node(sizeof(*client_dev),
399 GFP_KERNEL, node);
400 if (!client_dev) {
401 rc = -ENOMEM;
402 goto err;
403 }
404
405 dev = &client_dev->dev;
406
407 /* setup and register client devices */
408 dev_set_name(dev, "%s%d", device_name, i);
409 dev->bus = &ntb_transport_bus;
410 dev->release = ntb_transport_client_release;
411 dev->parent = &nt->ndev->dev;
412
413 rc = device_register(dev);
414 if (rc) {
415 put_device(dev);
416 goto err;
417 }
418
419 list_add_tail(&client_dev->entry, &nt->client_devs);
420 i++;
421 }
422
423 return 0;
424
425 err:
426 ntb_transport_unregister_client_dev(device_name);
427
428 return rc;
429 }
430 EXPORT_SYMBOL_GPL(ntb_transport_register_client_dev);
431
432 /**
433 * ntb_transport_register_client - Register NTB client driver
434 * @drv: NTB client driver to be registered
435 *
436 * Register 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 int ntb_transport_register_client(struct ntb_transport_client *drv)
441 {
442 drv->driver.bus = &ntb_transport_bus;
443
444 if (list_empty(&ntb_transport_list))
445 return -ENODEV;
446
447 return driver_register(&drv->driver);
448 }
449 EXPORT_SYMBOL_GPL(ntb_transport_register_client);
450
451 /**
452 * ntb_transport_unregister_client - Unregister NTB client driver
453 * @drv: NTB client driver to be unregistered
454 *
455 * Unregister an NTB client driver with the NTB transport layer
456 *
457 * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
458 */
459 void ntb_transport_unregister_client(struct ntb_transport_client *drv)
460 {
461 driver_unregister(&drv->driver);
462 }
463 EXPORT_SYMBOL_GPL(ntb_transport_unregister_client);
464
465 static ssize_t debugfs_read(struct file *filp, char __user *ubuf, size_t count,
466 loff_t *offp)
467 {
468 struct ntb_transport_qp *qp;
469 char *buf;
470 ssize_t ret, out_offset, out_count;
471
472 qp = filp->private_data;
473
474 if (!qp || !qp->link_is_up)
475 return 0;
476
477 out_count = 1000;
478
479 buf = kmalloc(out_count, GFP_KERNEL);
480 if (!buf)
481 return -ENOMEM;
482
483 out_offset = 0;
484 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
485 "\nNTB QP stats:\n\n");
486 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
487 "rx_bytes - \t%llu\n", qp->rx_bytes);
488 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
489 "rx_pkts - \t%llu\n", qp->rx_pkts);
490 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
491 "rx_memcpy - \t%llu\n", qp->rx_memcpy);
492 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
493 "rx_async - \t%llu\n", qp->rx_async);
494 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
495 "rx_ring_empty - %llu\n", qp->rx_ring_empty);
496 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
497 "rx_err_no_buf - %llu\n", qp->rx_err_no_buf);
498 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
499 "rx_err_oflow - \t%llu\n", qp->rx_err_oflow);
500 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
501 "rx_err_ver - \t%llu\n", qp->rx_err_ver);
502 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
503 "rx_buff - \t0x%p\n", qp->rx_buff);
504 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
505 "rx_index - \t%u\n", qp->rx_index);
506 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
507 "rx_max_entry - \t%u\n", qp->rx_max_entry);
508 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
509 "rx_alloc_entry - \t%u\n\n", qp->rx_alloc_entry);
510
511 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
512 "tx_bytes - \t%llu\n", qp->tx_bytes);
513 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
514 "tx_pkts - \t%llu\n", qp->tx_pkts);
515 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
516 "tx_memcpy - \t%llu\n", qp->tx_memcpy);
517 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
518 "tx_async - \t%llu\n", qp->tx_async);
519 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
520 "tx_ring_full - \t%llu\n", qp->tx_ring_full);
521 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
522 "tx_err_no_buf - %llu\n", qp->tx_err_no_buf);
523 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
524 "tx_mw - \t0x%p\n", qp->tx_mw);
525 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
526 "tx_index (H) - \t%u\n", qp->tx_index);
527 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
528 "RRI (T) - \t%u\n",
529 qp->remote_rx_info->entry);
530 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
531 "tx_max_entry - \t%u\n", qp->tx_max_entry);
532 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
533 "free tx - \t%u\n",
534 ntb_transport_tx_free_entry(qp));
535
536 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
537 "\n");
538 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
539 "Using TX DMA - \t%s\n",
540 qp->tx_dma_chan ? "Yes" : "No");
541 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
542 "Using RX DMA - \t%s\n",
543 qp->rx_dma_chan ? "Yes" : "No");
544 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
545 "QP Link - \t%s\n",
546 qp->link_is_up ? "Up" : "Down");
547 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
548 "\n");
549
550 if (out_offset > out_count)
551 out_offset = out_count;
552
553 ret = simple_read_from_buffer(ubuf, count, offp, buf, out_offset);
554 kfree(buf);
555 return ret;
556 }
557
558 static const struct file_operations ntb_qp_debugfs_stats = {
559 .owner = THIS_MODULE,
560 .open = simple_open,
561 .read = debugfs_read,
562 };
563
564 static void ntb_list_add(spinlock_t *lock, struct list_head *entry,
565 struct list_head *list)
566 {
567 unsigned long flags;
568
569 spin_lock_irqsave(lock, flags);
570 list_add_tail(entry, list);
571 spin_unlock_irqrestore(lock, flags);
572 }
573
574 static struct ntb_queue_entry *ntb_list_rm(spinlock_t *lock,
575 struct list_head *list)
576 {
577 struct ntb_queue_entry *entry;
578 unsigned long flags;
579
580 spin_lock_irqsave(lock, flags);
581 if (list_empty(list)) {
582 entry = NULL;
583 goto out;
584 }
585 entry = list_first_entry(list, struct ntb_queue_entry, entry);
586 list_del(&entry->entry);
587
588 out:
589 spin_unlock_irqrestore(lock, flags);
590
591 return entry;
592 }
593
594 static struct ntb_queue_entry *ntb_list_mv(spinlock_t *lock,
595 struct list_head *list,
596 struct list_head *to_list)
597 {
598 struct ntb_queue_entry *entry;
599 unsigned long flags;
600
601 spin_lock_irqsave(lock, flags);
602
603 if (list_empty(list)) {
604 entry = NULL;
605 } else {
606 entry = list_first_entry(list, struct ntb_queue_entry, entry);
607 list_move_tail(&entry->entry, to_list);
608 }
609
610 spin_unlock_irqrestore(lock, flags);
611
612 return entry;
613 }
614
615 static int ntb_transport_setup_qp_mw(struct ntb_transport_ctx *nt,
616 unsigned int qp_num)
617 {
618 struct ntb_transport_qp *qp = &nt->qp_vec[qp_num];
619 struct ntb_transport_mw *mw;
620 struct ntb_dev *ndev = nt->ndev;
621 struct ntb_queue_entry *entry;
622 unsigned int rx_size, num_qps_mw;
623 unsigned int mw_num, mw_count, qp_count;
624 unsigned int i;
625 int node;
626
627 mw_count = nt->mw_count;
628 qp_count = nt->qp_count;
629
630 mw_num = QP_TO_MW(nt, qp_num);
631 mw = &nt->mw_vec[mw_num];
632
633 if (!mw->virt_addr)
634 return -ENOMEM;
635
636 if (mw_num < qp_count % mw_count)
637 num_qps_mw = qp_count / mw_count + 1;
638 else
639 num_qps_mw = qp_count / mw_count;
640
641 rx_size = (unsigned int)mw->xlat_size / num_qps_mw;
642 qp->rx_buff = mw->virt_addr + rx_size * (qp_num / mw_count);
643 rx_size -= sizeof(struct ntb_rx_info);
644
645 qp->remote_rx_info = qp->rx_buff + rx_size;
646
647 /* Due to housekeeping, there must be atleast 2 buffs */
648 qp->rx_max_frame = min(transport_mtu, rx_size / 2);
649 qp->rx_max_entry = rx_size / qp->rx_max_frame;
650 qp->rx_index = 0;
651
652 /*
653 * Checking to see if we have more entries than the default.
654 * We should add additional entries if that is the case so we
655 * can be in sync with the transport frames.
656 */
657 node = dev_to_node(&ndev->dev);
658 for (i = qp->rx_alloc_entry; i < qp->rx_max_entry; i++) {
659 entry = kzalloc_node(sizeof(*entry), GFP_KERNEL, node);
660 if (!entry)
661 return -ENOMEM;
662
663 entry->qp = qp;
664 ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry,
665 &qp->rx_free_q);
666 qp->rx_alloc_entry++;
667 }
668
669 qp->remote_rx_info->entry = qp->rx_max_entry - 1;
670
671 /* setup the hdr offsets with 0's */
672 for (i = 0; i < qp->rx_max_entry; i++) {
673 void *offset = (qp->rx_buff + qp->rx_max_frame * (i + 1) -
674 sizeof(struct ntb_payload_header));
675 memset(offset, 0, sizeof(struct ntb_payload_header));
676 }
677
678 qp->rx_pkts = 0;
679 qp->tx_pkts = 0;
680 qp->tx_index = 0;
681
682 return 0;
683 }
684
685 static irqreturn_t ntb_transport_isr(int irq, void *dev)
686 {
687 struct ntb_transport_qp *qp = dev;
688
689 tasklet_schedule(&qp->rxc_db_work);
690
691 return IRQ_HANDLED;
692 }
693
694 static void ntb_transport_setup_qp_peer_msi(struct ntb_transport_ctx *nt,
695 unsigned int qp_num)
696 {
697 struct ntb_transport_qp *qp = &nt->qp_vec[qp_num];
698 int spad = qp_num * 2 + nt->msi_spad_offset;
699
700 if (!nt->use_msi)
701 return;
702
703 if (spad >= ntb_spad_count(nt->ndev))
704 return;
705
706 qp->peer_msi_desc.addr_offset =
707 ntb_peer_spad_read(qp->ndev, PIDX, spad);
708 qp->peer_msi_desc.data =
709 ntb_peer_spad_read(qp->ndev, PIDX, spad + 1);
710
711 dev_dbg(&qp->ndev->pdev->dev, "QP%d Peer MSI addr=%x data=%x\n",
712 qp_num, qp->peer_msi_desc.addr_offset, qp->peer_msi_desc.data);
713
714 if (qp->peer_msi_desc.addr_offset) {
715 qp->use_msi = true;
716 dev_info(&qp->ndev->pdev->dev,
717 "Using MSI interrupts for QP%d\n", qp_num);
718 }
719 }
720
721 static void ntb_transport_setup_qp_msi(struct ntb_transport_ctx *nt,
722 unsigned int qp_num)
723 {
724 struct ntb_transport_qp *qp = &nt->qp_vec[qp_num];
725 int spad = qp_num * 2 + nt->msi_spad_offset;
726 int rc;
727
728 if (!nt->use_msi)
729 return;
730
731 if (spad >= ntb_spad_count(nt->ndev)) {
732 dev_warn_once(&qp->ndev->pdev->dev,
733 "Not enough SPADS to use MSI interrupts\n");
734 return;
735 }
736
737 ntb_spad_write(qp->ndev, spad, 0);
738 ntb_spad_write(qp->ndev, spad + 1, 0);
739
740 if (!qp->msi_irq) {
741 qp->msi_irq = ntbm_msi_request_irq(qp->ndev, ntb_transport_isr,
742 KBUILD_MODNAME, qp,
743 &qp->msi_desc);
744 if (qp->msi_irq < 0) {
745 dev_warn(&qp->ndev->pdev->dev,
746 "Unable to allocate MSI interrupt for qp%d\n",
747 qp_num);
748 return;
749 }
750 }
751
752 rc = ntb_spad_write(qp->ndev, spad, qp->msi_desc.addr_offset);
753 if (rc)
754 goto err_free_interrupt;
755
756 rc = ntb_spad_write(qp->ndev, spad + 1, qp->msi_desc.data);
757 if (rc)
758 goto err_free_interrupt;
759
760 dev_dbg(&qp->ndev->pdev->dev, "QP%d MSI %d addr=%x data=%x\n",
761 qp_num, qp->msi_irq, qp->msi_desc.addr_offset,
762 qp->msi_desc.data);
763
764 return;
765
766 err_free_interrupt:
767 devm_free_irq(&nt->ndev->dev, qp->msi_irq, qp);
768 }
769
770 static void ntb_transport_msi_peer_desc_changed(struct ntb_transport_ctx *nt)
771 {
772 int i;
773
774 dev_dbg(&nt->ndev->pdev->dev, "Peer MSI descriptors changed");
775
776 for (i = 0; i < nt->qp_count; i++)
777 ntb_transport_setup_qp_peer_msi(nt, i);
778 }
779
780 static void ntb_transport_msi_desc_changed(void *data)
781 {
782 struct ntb_transport_ctx *nt = data;
783 int i;
784
785 dev_dbg(&nt->ndev->pdev->dev, "MSI descriptors changed");
786
787 for (i = 0; i < nt->qp_count; i++)
788 ntb_transport_setup_qp_msi(nt, i);
789
790 ntb_peer_db_set(nt->ndev, nt->msi_db_mask);
791 }
792
793 static void ntb_free_mw(struct ntb_transport_ctx *nt, int num_mw)
794 {
795 struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
796 struct pci_dev *pdev = nt->ndev->pdev;
797
798 if (!mw->virt_addr)
799 return;
800
801 ntb_mw_clear_trans(nt->ndev, PIDX, num_mw);
802 dma_free_coherent(&pdev->dev, mw->alloc_size,
803 mw->alloc_addr, mw->dma_addr);
804 mw->xlat_size = 0;
805 mw->buff_size = 0;
806 mw->alloc_size = 0;
807 mw->alloc_addr = NULL;
808 mw->virt_addr = NULL;
809 }
810
811 static int ntb_alloc_mw_buffer(struct ntb_transport_mw *mw,
812 struct device *ntb_dev, size_t align)
813 {
814 dma_addr_t dma_addr;
815 void *alloc_addr, *virt_addr;
816 int rc;
817
818 /*
819 * The buffer here is allocated against the NTB device. The reason to
820 * use dma_alloc_*() call is to allocate a large IOVA contiguous buffer
821 * backing the NTB BAR for the remote host to write to. During receive
822 * processing, the data is being copied out of the receive buffer to
823 * the kernel skbuff. When a DMA device is being used, dma_map_page()
824 * is called on the kvaddr of the receive buffer (from dma_alloc_*())
825 * and remapped against the DMA device. It appears to be a double
826 * DMA mapping of buffers, but first is mapped to the NTB device and
827 * second is to the DMA device. DMA_ATTR_FORCE_CONTIGUOUS is necessary
828 * in order for the later dma_map_page() to not fail.
829 */
830 alloc_addr = dma_alloc_attrs(ntb_dev, mw->alloc_size,
831 &dma_addr, GFP_KERNEL,
832 DMA_ATTR_FORCE_CONTIGUOUS);
833 if (!alloc_addr) {
834 dev_err(ntb_dev, "Unable to alloc MW buff of size %zu\n",
835 mw->alloc_size);
836 return -ENOMEM;
837 }
838 virt_addr = alloc_addr;
839
840 /*
841 * we must ensure that the memory address allocated is BAR size
842 * aligned in order for the XLAT register to take the value. This
843 * is a requirement of the hardware. It is recommended to setup CMA
844 * for BAR sizes equal or greater than 4MB.
845 */
846 if (!IS_ALIGNED(dma_addr, align)) {
847 if (mw->alloc_size > mw->buff_size) {
848 virt_addr = PTR_ALIGN(alloc_addr, align);
849 dma_addr = ALIGN(dma_addr, align);
850 } else {
851 rc = -ENOMEM;
852 goto err;
853 }
854 }
855
856 mw->alloc_addr = alloc_addr;
857 mw->virt_addr = virt_addr;
858 mw->dma_addr = dma_addr;
859
860 return 0;
861
862 err:
863 dma_free_coherent(ntb_dev, mw->alloc_size, alloc_addr, dma_addr);
864
865 return rc;
866 }
867
868 static int ntb_set_mw(struct ntb_transport_ctx *nt, int num_mw,
869 resource_size_t size)
870 {
871 struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
872 struct pci_dev *pdev = nt->ndev->pdev;
873 size_t xlat_size, buff_size;
874 resource_size_t xlat_align;
875 resource_size_t xlat_align_size;
876 int rc;
877
878 if (!size)
879 return -EINVAL;
880
881 rc = ntb_mw_get_align(nt->ndev, PIDX, num_mw, &xlat_align,
882 &xlat_align_size, NULL);
883 if (rc)
884 return rc;
885
886 xlat_size = round_up(size, xlat_align_size);
887 buff_size = round_up(size, xlat_align);
888
889 /* No need to re-setup */
890 if (mw->xlat_size == xlat_size)
891 return 0;
892
893 if (mw->buff_size)
894 ntb_free_mw(nt, num_mw);
895
896 /* Alloc memory for receiving data. Must be aligned */
897 mw->xlat_size = xlat_size;
898 mw->buff_size = buff_size;
899 mw->alloc_size = buff_size;
900
901 rc = ntb_alloc_mw_buffer(mw, &pdev->dev, xlat_align);
902 if (rc) {
903 mw->alloc_size *= 2;
904 rc = ntb_alloc_mw_buffer(mw, &pdev->dev, xlat_align);
905 if (rc) {
906 dev_err(&pdev->dev,
907 "Unable to alloc aligned MW buff\n");
908 mw->xlat_size = 0;
909 mw->buff_size = 0;
910 mw->alloc_size = 0;
911 return rc;
912 }
913 }
914
915 /* Notify HW the memory location of the receive buffer */
916 rc = ntb_mw_set_trans(nt->ndev, PIDX, num_mw, mw->dma_addr,
917 mw->xlat_size);
918 if (rc) {
919 dev_err(&pdev->dev, "Unable to set mw%d translation", num_mw);
920 ntb_free_mw(nt, num_mw);
921 return -EIO;
922 }
923
924 return 0;
925 }
926
927 static void ntb_qp_link_context_reset(struct ntb_transport_qp *qp)
928 {
929 qp->link_is_up = false;
930 qp->active = false;
931
932 qp->tx_index = 0;
933 qp->rx_index = 0;
934 qp->rx_bytes = 0;
935 qp->rx_pkts = 0;
936 qp->rx_ring_empty = 0;
937 qp->rx_err_no_buf = 0;
938 qp->rx_err_oflow = 0;
939 qp->rx_err_ver = 0;
940 qp->rx_memcpy = 0;
941 qp->rx_async = 0;
942 qp->tx_bytes = 0;
943 qp->tx_pkts = 0;
944 qp->tx_ring_full = 0;
945 qp->tx_err_no_buf = 0;
946 qp->tx_memcpy = 0;
947 qp->tx_async = 0;
948 }
949
950 static void ntb_qp_link_down_reset(struct ntb_transport_qp *qp)
951 {
952 ntb_qp_link_context_reset(qp);
953 if (qp->remote_rx_info)
954 qp->remote_rx_info->entry = qp->rx_max_entry - 1;
955 }
956
957 static void ntb_qp_link_cleanup(struct ntb_transport_qp *qp)
958 {
959 struct ntb_transport_ctx *nt = qp->transport;
960 struct pci_dev *pdev = nt->ndev->pdev;
961
962 dev_info(&pdev->dev, "qp %d: Link Cleanup\n", qp->qp_num);
963
964 cancel_delayed_work_sync(&qp->link_work);
965 ntb_qp_link_down_reset(qp);
966
967 if (qp->event_handler)
968 qp->event_handler(qp->cb_data, qp->link_is_up);
969 }
970
971 static void ntb_qp_link_cleanup_work(struct work_struct *work)
972 {
973 struct ntb_transport_qp *qp = container_of(work,
974 struct ntb_transport_qp,
975 link_cleanup);
976 struct ntb_transport_ctx *nt = qp->transport;
977
978 ntb_qp_link_cleanup(qp);
979
980 if (nt->link_is_up)
981 schedule_delayed_work(&qp->link_work,
982 msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
983 }
984
985 static void ntb_qp_link_down(struct ntb_transport_qp *qp)
986 {
987 schedule_work(&qp->link_cleanup);
988 }
989
990 static void ntb_transport_link_cleanup(struct ntb_transport_ctx *nt)
991 {
992 struct ntb_transport_qp *qp;
993 u64 qp_bitmap_alloc;
994 unsigned int i, count;
995
996 qp_bitmap_alloc = nt->qp_bitmap & ~nt->qp_bitmap_free;
997
998 /* Pass along the info to any clients */
999 for (i = 0; i < nt->qp_count; i++)
1000 if (qp_bitmap_alloc & BIT_ULL(i)) {
1001 qp = &nt->qp_vec[i];
1002 ntb_qp_link_cleanup(qp);
1003 cancel_work_sync(&qp->link_cleanup);
1004 cancel_delayed_work_sync(&qp->link_work);
1005 }
1006
1007 if (!nt->link_is_up)
1008 cancel_delayed_work_sync(&nt->link_work);
1009
1010 for (i = 0; i < nt->mw_count; i++)
1011 ntb_free_mw(nt, i);
1012
1013 /* The scratchpad registers keep the values if the remote side
1014 * goes down, blast them now to give them a sane value the next
1015 * time they are accessed
1016 */
1017 count = ntb_spad_count(nt->ndev);
1018 for (i = 0; i < count; i++)
1019 ntb_spad_write(nt->ndev, i, 0);
1020 }
1021
1022 static void ntb_transport_link_cleanup_work(struct work_struct *work)
1023 {
1024 struct ntb_transport_ctx *nt =
1025 container_of(work, struct ntb_transport_ctx, link_cleanup);
1026
1027 ntb_transport_link_cleanup(nt);
1028 }
1029
1030 static void ntb_transport_event_callback(void *data)
1031 {
1032 struct ntb_transport_ctx *nt = data;
1033
1034 if (ntb_link_is_up(nt->ndev, NULL, NULL) == 1)
1035 schedule_delayed_work(&nt->link_work, 0);
1036 else
1037 schedule_work(&nt->link_cleanup);
1038 }
1039
1040 static void ntb_transport_link_work(struct work_struct *work)
1041 {
1042 struct ntb_transport_ctx *nt =
1043 container_of(work, struct ntb_transport_ctx, link_work.work);
1044 struct ntb_dev *ndev = nt->ndev;
1045 struct pci_dev *pdev = ndev->pdev;
1046 resource_size_t size;
1047 u32 val;
1048 int rc = 0, i, spad;
1049
1050 /* send the local info, in the opposite order of the way we read it */
1051
1052 if (nt->use_msi) {
1053 rc = ntb_msi_setup_mws(ndev);
1054 if (rc) {
1055 dev_warn(&pdev->dev,
1056 "Failed to register MSI memory window: %d\n",
1057 rc);
1058 nt->use_msi = false;
1059 }
1060 }
1061
1062 for (i = 0; i < nt->qp_count; i++)
1063 ntb_transport_setup_qp_msi(nt, i);
1064
1065 for (i = 0; i < nt->mw_count; i++) {
1066 size = nt->mw_vec[i].phys_size;
1067
1068 if (max_mw_size && size > max_mw_size)
1069 size = max_mw_size;
1070
1071 spad = MW0_SZ_HIGH + (i * 2);
1072 ntb_peer_spad_write(ndev, PIDX, spad, upper_32_bits(size));
1073
1074 spad = MW0_SZ_LOW + (i * 2);
1075 ntb_peer_spad_write(ndev, PIDX, spad, lower_32_bits(size));
1076 }
1077
1078 ntb_peer_spad_write(ndev, PIDX, NUM_MWS, nt->mw_count);
1079
1080 ntb_peer_spad_write(ndev, PIDX, NUM_QPS, nt->qp_count);
1081
1082 ntb_peer_spad_write(ndev, PIDX, VERSION, NTB_TRANSPORT_VERSION);
1083
1084 /* Query the remote side for its info */
1085 val = ntb_spad_read(ndev, VERSION);
1086 dev_dbg(&pdev->dev, "Remote version = %d\n", val);
1087 if (val != NTB_TRANSPORT_VERSION)
1088 goto out;
1089
1090 val = ntb_spad_read(ndev, NUM_QPS);
1091 dev_dbg(&pdev->dev, "Remote max number of qps = %d\n", val);
1092 if (val != nt->qp_count)
1093 goto out;
1094
1095 val = ntb_spad_read(ndev, NUM_MWS);
1096 dev_dbg(&pdev->dev, "Remote number of mws = %d\n", val);
1097 if (val != nt->mw_count)
1098 goto out;
1099
1100 for (i = 0; i < nt->mw_count; i++) {
1101 u64 val64;
1102
1103 val = ntb_spad_read(ndev, MW0_SZ_HIGH + (i * 2));
1104 val64 = (u64)val << 32;
1105
1106 val = ntb_spad_read(ndev, MW0_SZ_LOW + (i * 2));
1107 val64 |= val;
1108
1109 dev_dbg(&pdev->dev, "Remote MW%d size = %#llx\n", i, val64);
1110
1111 rc = ntb_set_mw(nt, i, val64);
1112 if (rc)
1113 goto out1;
1114 }
1115
1116 nt->link_is_up = true;
1117
1118 for (i = 0; i < nt->qp_count; i++) {
1119 struct ntb_transport_qp *qp = &nt->qp_vec[i];
1120
1121 ntb_transport_setup_qp_mw(nt, i);
1122 ntb_transport_setup_qp_peer_msi(nt, i);
1123
1124 if (qp->client_ready)
1125 schedule_delayed_work(&qp->link_work, 0);
1126 }
1127
1128 return;
1129
1130 out1:
1131 for (i = 0; i < nt->mw_count; i++)
1132 ntb_free_mw(nt, i);
1133
1134 /* if there's an actual failure, we should just bail */
1135 if (rc < 0)
1136 return;
1137
1138 out:
1139 if (ntb_link_is_up(ndev, NULL, NULL) == 1)
1140 schedule_delayed_work(&nt->link_work,
1141 msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
1142 }
1143
1144 static void ntb_qp_link_work(struct work_struct *work)
1145 {
1146 struct ntb_transport_qp *qp = container_of(work,
1147 struct ntb_transport_qp,
1148 link_work.work);
1149 struct pci_dev *pdev = qp->ndev->pdev;
1150 struct ntb_transport_ctx *nt = qp->transport;
1151 int val;
1152
1153 WARN_ON(!nt->link_is_up);
1154
1155 val = ntb_spad_read(nt->ndev, QP_LINKS);
1156
1157 ntb_peer_spad_write(nt->ndev, PIDX, QP_LINKS, val | BIT(qp->qp_num));
1158
1159 /* query remote spad for qp ready bits */
1160 dev_dbg_ratelimited(&pdev->dev, "Remote QP link status = %x\n", val);
1161
1162 /* See if the remote side is up */
1163 if (val & BIT(qp->qp_num)) {
1164 dev_info(&pdev->dev, "qp %d: Link Up\n", qp->qp_num);
1165 qp->link_is_up = true;
1166 qp->active = true;
1167
1168 if (qp->event_handler)
1169 qp->event_handler(qp->cb_data, qp->link_is_up);
1170
1171 if (qp->active)
1172 tasklet_schedule(&qp->rxc_db_work);
1173 } else if (nt->link_is_up)
1174 schedule_delayed_work(&qp->link_work,
1175 msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
1176 }
1177
1178 static int ntb_transport_init_queue(struct ntb_transport_ctx *nt,
1179 unsigned int qp_num)
1180 {
1181 struct ntb_transport_qp *qp;
1182 phys_addr_t mw_base;
1183 resource_size_t mw_size;
1184 unsigned int num_qps_mw, tx_size;
1185 unsigned int mw_num, mw_count, qp_count;
1186 u64 qp_offset;
1187
1188 mw_count = nt->mw_count;
1189 qp_count = nt->qp_count;
1190
1191 mw_num = QP_TO_MW(nt, qp_num);
1192
1193 qp = &nt->qp_vec[qp_num];
1194 qp->qp_num = qp_num;
1195 qp->transport = nt;
1196 qp->ndev = nt->ndev;
1197 qp->client_ready = false;
1198 qp->event_handler = NULL;
1199 ntb_qp_link_context_reset(qp);
1200
1201 if (mw_num < qp_count % mw_count)
1202 num_qps_mw = qp_count / mw_count + 1;
1203 else
1204 num_qps_mw = qp_count / mw_count;
1205
1206 mw_base = nt->mw_vec[mw_num].phys_addr;
1207 mw_size = nt->mw_vec[mw_num].phys_size;
1208
1209 if (max_mw_size && mw_size > max_mw_size)
1210 mw_size = max_mw_size;
1211
1212 tx_size = (unsigned int)mw_size / num_qps_mw;
1213 qp_offset = tx_size * (qp_num / mw_count);
1214
1215 qp->tx_mw_size = tx_size;
1216 qp->tx_mw = nt->mw_vec[mw_num].vbase + qp_offset;
1217 if (!qp->tx_mw)
1218 return -EINVAL;
1219
1220 qp->tx_mw_phys = mw_base + qp_offset;
1221 if (!qp->tx_mw_phys)
1222 return -EINVAL;
1223
1224 tx_size -= sizeof(struct ntb_rx_info);
1225 qp->rx_info = qp->tx_mw + tx_size;
1226
1227 /* Due to housekeeping, there must be atleast 2 buffs */
1228 qp->tx_max_frame = min(transport_mtu, tx_size / 2);
1229 qp->tx_max_entry = tx_size / qp->tx_max_frame;
1230
1231 if (nt->debugfs_node_dir) {
1232 char debugfs_name[4];
1233
1234 snprintf(debugfs_name, 4, "qp%d", qp_num);
1235 qp->debugfs_dir = debugfs_create_dir(debugfs_name,
1236 nt->debugfs_node_dir);
1237
1238 qp->debugfs_stats = debugfs_create_file("stats", S_IRUSR,
1239 qp->debugfs_dir, qp,
1240 &ntb_qp_debugfs_stats);
1241 } else {
1242 qp->debugfs_dir = NULL;
1243 qp->debugfs_stats = NULL;
1244 }
1245
1246 INIT_DELAYED_WORK(&qp->link_work, ntb_qp_link_work);
1247 INIT_WORK(&qp->link_cleanup, ntb_qp_link_cleanup_work);
1248
1249 spin_lock_init(&qp->ntb_rx_q_lock);
1250 spin_lock_init(&qp->ntb_tx_free_q_lock);
1251
1252 INIT_LIST_HEAD(&qp->rx_post_q);
1253 INIT_LIST_HEAD(&qp->rx_pend_q);
1254 INIT_LIST_HEAD(&qp->rx_free_q);
1255 INIT_LIST_HEAD(&qp->tx_free_q);
1256
1257 tasklet_init(&qp->rxc_db_work, ntb_transport_rxc_db,
1258 (unsigned long)qp);
1259
1260 return 0;
1261 }
1262
1263 static int ntb_transport_probe(struct ntb_client *self, struct ntb_dev *ndev)
1264 {
1265 struct ntb_transport_ctx *nt;
1266 struct ntb_transport_mw *mw;
1267 unsigned int mw_count, qp_count, spad_count, max_mw_count_for_spads;
1268 u64 qp_bitmap;
1269 int node;
1270 int rc, i;
1271
1272 mw_count = ntb_peer_mw_count(ndev);
1273
1274 if (!ndev->ops->mw_set_trans) {
1275 dev_err(&ndev->dev, "Inbound MW based NTB API is required\n");
1276 return -EINVAL;
1277 }
1278
1279 if (ntb_db_is_unsafe(ndev))
1280 dev_dbg(&ndev->dev,
1281 "doorbell is unsafe, proceed anyway...\n");
1282 if (ntb_spad_is_unsafe(ndev))
1283 dev_dbg(&ndev->dev,
1284 "scratchpad is unsafe, proceed anyway...\n");
1285
1286 if (ntb_peer_port_count(ndev) != NTB_DEF_PEER_CNT)
1287 dev_warn(&ndev->dev, "Multi-port NTB devices unsupported\n");
1288
1289 node = dev_to_node(&ndev->dev);
1290
1291 nt = kzalloc_node(sizeof(*nt), GFP_KERNEL, node);
1292 if (!nt)
1293 return -ENOMEM;
1294
1295 nt->ndev = ndev;
1296
1297 /*
1298 * If we are using MSI, and have at least one extra memory window,
1299 * we will reserve the last MW for the MSI window.
1300 */
1301 if (use_msi && mw_count > 1) {
1302 rc = ntb_msi_init(ndev, ntb_transport_msi_desc_changed);
1303 if (!rc) {
1304 mw_count -= 1;
1305 nt->use_msi = true;
1306 }
1307 }
1308
1309 spad_count = ntb_spad_count(ndev);
1310
1311 /* Limit the MW's based on the availability of scratchpads */
1312
1313 if (spad_count < NTB_TRANSPORT_MIN_SPADS) {
1314 nt->mw_count = 0;
1315 rc = -EINVAL;
1316 goto err;
1317 }
1318
1319 max_mw_count_for_spads = (spad_count - MW0_SZ_HIGH) / 2;
1320 nt->mw_count = min(mw_count, max_mw_count_for_spads);
1321
1322 nt->msi_spad_offset = nt->mw_count * 2 + MW0_SZ_HIGH;
1323
1324 nt->mw_vec = kcalloc_node(mw_count, sizeof(*nt->mw_vec),
1325 GFP_KERNEL, node);
1326 if (!nt->mw_vec) {
1327 rc = -ENOMEM;
1328 goto err;
1329 }
1330
1331 for (i = 0; i < mw_count; i++) {
1332 mw = &nt->mw_vec[i];
1333
1334 rc = ntb_peer_mw_get_addr(ndev, i, &mw->phys_addr,
1335 &mw->phys_size);
1336 if (rc)
1337 goto err1;
1338
1339 mw->vbase = ioremap_wc(mw->phys_addr, mw->phys_size);
1340 if (!mw->vbase) {
1341 rc = -ENOMEM;
1342 goto err1;
1343 }
1344
1345 mw->buff_size = 0;
1346 mw->xlat_size = 0;
1347 mw->virt_addr = NULL;
1348 mw->dma_addr = 0;
1349 }
1350
1351 qp_bitmap = ntb_db_valid_mask(ndev);
1352
1353 qp_count = ilog2(qp_bitmap);
1354 if (nt->use_msi) {
1355 qp_count -= 1;
1356 nt->msi_db_mask = 1 << qp_count;
1357 ntb_db_clear_mask(ndev, nt->msi_db_mask);
1358 }
1359
1360 if (max_num_clients && max_num_clients < qp_count)
1361 qp_count = max_num_clients;
1362 else if (nt->mw_count < qp_count)
1363 qp_count = nt->mw_count;
1364
1365 qp_bitmap &= BIT_ULL(qp_count) - 1;
1366
1367 nt->qp_count = qp_count;
1368 nt->qp_bitmap = qp_bitmap;
1369 nt->qp_bitmap_free = qp_bitmap;
1370
1371 nt->qp_vec = kcalloc_node(qp_count, sizeof(*nt->qp_vec),
1372 GFP_KERNEL, node);
1373 if (!nt->qp_vec) {
1374 rc = -ENOMEM;
1375 goto err1;
1376 }
1377
1378 if (nt_debugfs_dir) {
1379 nt->debugfs_node_dir =
1380 debugfs_create_dir(pci_name(ndev->pdev),
1381 nt_debugfs_dir);
1382 }
1383
1384 for (i = 0; i < qp_count; i++) {
1385 rc = ntb_transport_init_queue(nt, i);
1386 if (rc)
1387 goto err2;
1388 }
1389
1390 INIT_DELAYED_WORK(&nt->link_work, ntb_transport_link_work);
1391 INIT_WORK(&nt->link_cleanup, ntb_transport_link_cleanup_work);
1392
1393 rc = ntb_set_ctx(ndev, nt, &ntb_transport_ops);
1394 if (rc)
1395 goto err2;
1396
1397 INIT_LIST_HEAD(&nt->client_devs);
1398 rc = ntb_bus_init(nt);
1399 if (rc)
1400 goto err3;
1401
1402 nt->link_is_up = false;
1403 ntb_link_enable(ndev, NTB_SPEED_AUTO, NTB_WIDTH_AUTO);
1404 ntb_link_event(ndev);
1405
1406 return 0;
1407
1408 err3:
1409 ntb_clear_ctx(ndev);
1410 err2:
1411 kfree(nt->qp_vec);
1412 err1:
1413 while (i--) {
1414 mw = &nt->mw_vec[i];
1415 iounmap(mw->vbase);
1416 }
1417 kfree(nt->mw_vec);
1418 err:
1419 kfree(nt);
1420 return rc;
1421 }
1422
1423 static void ntb_transport_free(struct ntb_client *self, struct ntb_dev *ndev)
1424 {
1425 struct ntb_transport_ctx *nt = ndev->ctx;
1426 struct ntb_transport_qp *qp;
1427 u64 qp_bitmap_alloc;
1428 int i;
1429
1430 ntb_transport_link_cleanup(nt);
1431 cancel_work_sync(&nt->link_cleanup);
1432 cancel_delayed_work_sync(&nt->link_work);
1433
1434 qp_bitmap_alloc = nt->qp_bitmap & ~nt->qp_bitmap_free;
1435
1436 /* verify that all the qp's are freed */
1437 for (i = 0; i < nt->qp_count; i++) {
1438 qp = &nt->qp_vec[i];
1439 if (qp_bitmap_alloc & BIT_ULL(i))
1440 ntb_transport_free_queue(qp);
1441 debugfs_remove_recursive(qp->debugfs_dir);
1442 }
1443
1444 ntb_link_disable(ndev);
1445 ntb_clear_ctx(ndev);
1446
1447 ntb_bus_remove(nt);
1448
1449 for (i = nt->mw_count; i--; ) {
1450 ntb_free_mw(nt, i);
1451 iounmap(nt->mw_vec[i].vbase);
1452 }
1453
1454 kfree(nt->qp_vec);
1455 kfree(nt->mw_vec);
1456 kfree(nt);
1457 }
1458
1459 static void ntb_complete_rxc(struct ntb_transport_qp *qp)
1460 {
1461 struct ntb_queue_entry *entry;
1462 void *cb_data;
1463 unsigned int len;
1464 unsigned long irqflags;
1465
1466 spin_lock_irqsave(&qp->ntb_rx_q_lock, irqflags);
1467
1468 while (!list_empty(&qp->rx_post_q)) {
1469 entry = list_first_entry(&qp->rx_post_q,
1470 struct ntb_queue_entry, entry);
1471 if (!(entry->flags & DESC_DONE_FLAG))
1472 break;
1473
1474 entry->rx_hdr->flags = 0;
1475 iowrite32(entry->rx_index, &qp->rx_info->entry);
1476
1477 cb_data = entry->cb_data;
1478 len = entry->len;
1479
1480 list_move_tail(&entry->entry, &qp->rx_free_q);
1481
1482 spin_unlock_irqrestore(&qp->ntb_rx_q_lock, irqflags);
1483
1484 if (qp->rx_handler && qp->client_ready)
1485 qp->rx_handler(qp, qp->cb_data, cb_data, len);
1486
1487 spin_lock_irqsave(&qp->ntb_rx_q_lock, irqflags);
1488 }
1489
1490 spin_unlock_irqrestore(&qp->ntb_rx_q_lock, irqflags);
1491 }
1492
1493 static void ntb_rx_copy_callback(void *data,
1494 const struct dmaengine_result *res)
1495 {
1496 struct ntb_queue_entry *entry = data;
1497
1498 /* we need to check DMA results if we are using DMA */
1499 if (res) {
1500 enum dmaengine_tx_result dma_err = res->result;
1501
1502 switch (dma_err) {
1503 case DMA_TRANS_READ_FAILED:
1504 case DMA_TRANS_WRITE_FAILED:
1505 entry->errors++;
1506 fallthrough;
1507 case DMA_TRANS_ABORTED:
1508 {
1509 struct ntb_transport_qp *qp = entry->qp;
1510 void *offset = qp->rx_buff + qp->rx_max_frame *
1511 qp->rx_index;
1512
1513 ntb_memcpy_rx(entry, offset);
1514 qp->rx_memcpy++;
1515 return;
1516 }
1517
1518 case DMA_TRANS_NOERROR:
1519 default:
1520 break;
1521 }
1522 }
1523
1524 entry->flags |= DESC_DONE_FLAG;
1525
1526 ntb_complete_rxc(entry->qp);
1527 }
1528
1529 static void ntb_memcpy_rx(struct ntb_queue_entry *entry, void *offset)
1530 {
1531 void *buf = entry->buf;
1532 size_t len = entry->len;
1533
1534 memcpy(buf, offset, len);
1535
1536 /* Ensure that the data is fully copied out before clearing the flag */
1537 wmb();
1538
1539 ntb_rx_copy_callback(entry, NULL);
1540 }
1541
1542 static int ntb_async_rx_submit(struct ntb_queue_entry *entry, void *offset)
1543 {
1544 struct dma_async_tx_descriptor *txd;
1545 struct ntb_transport_qp *qp = entry->qp;
1546 struct dma_chan *chan = qp->rx_dma_chan;
1547 struct dma_device *device;
1548 size_t pay_off, buff_off, len;
1549 struct dmaengine_unmap_data *unmap;
1550 dma_cookie_t cookie;
1551 void *buf = entry->buf;
1552
1553 len = entry->len;
1554 device = chan->device;
1555 pay_off = (size_t)offset & ~PAGE_MASK;
1556 buff_off = (size_t)buf & ~PAGE_MASK;
1557
1558 if (!is_dma_copy_aligned(device, pay_off, buff_off, len))
1559 goto err;
1560
1561 unmap = dmaengine_get_unmap_data(device->dev, 2, GFP_NOWAIT);
1562 if (!unmap)
1563 goto err;
1564
1565 unmap->len = len;
1566 unmap->addr[0] = dma_map_page(device->dev, virt_to_page(offset),
1567 pay_off, len, DMA_TO_DEVICE);
1568 if (dma_mapping_error(device->dev, unmap->addr[0]))
1569 goto err_get_unmap;
1570
1571 unmap->to_cnt = 1;
1572
1573 unmap->addr[1] = dma_map_page(device->dev, virt_to_page(buf),
1574 buff_off, len, DMA_FROM_DEVICE);
1575 if (dma_mapping_error(device->dev, unmap->addr[1]))
1576 goto err_get_unmap;
1577
1578 unmap->from_cnt = 1;
1579
1580 txd = device->device_prep_dma_memcpy(chan, unmap->addr[1],
1581 unmap->addr[0], len,
1582 DMA_PREP_INTERRUPT);
1583 if (!txd)
1584 goto err_get_unmap;
1585
1586 txd->callback_result = ntb_rx_copy_callback;
1587 txd->callback_param = entry;
1588 dma_set_unmap(txd, unmap);
1589
1590 cookie = dmaengine_submit(txd);
1591 if (dma_submit_error(cookie))
1592 goto err_set_unmap;
1593
1594 dmaengine_unmap_put(unmap);
1595
1596 qp->last_cookie = cookie;
1597
1598 qp->rx_async++;
1599
1600 return 0;
1601
1602 err_set_unmap:
1603 dmaengine_unmap_put(unmap);
1604 err_get_unmap:
1605 dmaengine_unmap_put(unmap);
1606 err:
1607 return -ENXIO;
1608 }
1609
1610 static void ntb_async_rx(struct ntb_queue_entry *entry, void *offset)
1611 {
1612 struct ntb_transport_qp *qp = entry->qp;
1613 struct dma_chan *chan = qp->rx_dma_chan;
1614 int res;
1615
1616 if (!chan)
1617 goto err;
1618
1619 if (entry->len < copy_bytes)
1620 goto err;
1621
1622 res = ntb_async_rx_submit(entry, offset);
1623 if (res < 0)
1624 goto err;
1625
1626 if (!entry->retries)
1627 qp->rx_async++;
1628
1629 return;
1630
1631 err:
1632 ntb_memcpy_rx(entry, offset);
1633 qp->rx_memcpy++;
1634 }
1635
1636 static int ntb_process_rxc(struct ntb_transport_qp *qp)
1637 {
1638 struct ntb_payload_header *hdr;
1639 struct ntb_queue_entry *entry;
1640 void *offset;
1641
1642 offset = qp->rx_buff + qp->rx_max_frame * qp->rx_index;
1643 hdr = offset + qp->rx_max_frame - sizeof(struct ntb_payload_header);
1644
1645 dev_dbg(&qp->ndev->pdev->dev, "qp %d: RX ver %u len %d flags %x\n",
1646 qp->qp_num, hdr->ver, hdr->len, hdr->flags);
1647
1648 if (!(hdr->flags & DESC_DONE_FLAG)) {
1649 dev_dbg(&qp->ndev->pdev->dev, "done flag not set\n");
1650 qp->rx_ring_empty++;
1651 return -EAGAIN;
1652 }
1653
1654 if (hdr->flags & LINK_DOWN_FLAG) {
1655 dev_dbg(&qp->ndev->pdev->dev, "link down flag set\n");
1656 ntb_qp_link_down(qp);
1657 hdr->flags = 0;
1658 return -EAGAIN;
1659 }
1660
1661 if (hdr->ver != (u32)qp->rx_pkts) {
1662 dev_dbg(&qp->ndev->pdev->dev,
1663 "version mismatch, expected %llu - got %u\n",
1664 qp->rx_pkts, hdr->ver);
1665 qp->rx_err_ver++;
1666 return -EIO;
1667 }
1668
1669 entry = ntb_list_mv(&qp->ntb_rx_q_lock, &qp->rx_pend_q, &qp->rx_post_q);
1670 if (!entry) {
1671 dev_dbg(&qp->ndev->pdev->dev, "no receive buffer\n");
1672 qp->rx_err_no_buf++;
1673 return -EAGAIN;
1674 }
1675
1676 entry->rx_hdr = hdr;
1677 entry->rx_index = qp->rx_index;
1678
1679 if (hdr->len > entry->len) {
1680 dev_dbg(&qp->ndev->pdev->dev,
1681 "receive buffer overflow! Wanted %d got %d\n",
1682 hdr->len, entry->len);
1683 qp->rx_err_oflow++;
1684
1685 entry->len = -EIO;
1686 entry->flags |= DESC_DONE_FLAG;
1687
1688 ntb_complete_rxc(qp);
1689 } else {
1690 dev_dbg(&qp->ndev->pdev->dev,
1691 "RX OK index %u ver %u size %d into buf size %d\n",
1692 qp->rx_index, hdr->ver, hdr->len, entry->len);
1693
1694 qp->rx_bytes += hdr->len;
1695 qp->rx_pkts++;
1696
1697 entry->len = hdr->len;
1698
1699 ntb_async_rx(entry, offset);
1700 }
1701
1702 qp->rx_index++;
1703 qp->rx_index %= qp->rx_max_entry;
1704
1705 return 0;
1706 }
1707
1708 static void ntb_transport_rxc_db(unsigned long data)
1709 {
1710 struct ntb_transport_qp *qp = (void *)data;
1711 int rc, i;
1712
1713 dev_dbg(&qp->ndev->pdev->dev, "%s: doorbell %d received\n",
1714 __func__, qp->qp_num);
1715
1716 /* Limit the number of packets processed in a single interrupt to
1717 * provide fairness to others
1718 */
1719 for (i = 0; i < qp->rx_max_entry; i++) {
1720 rc = ntb_process_rxc(qp);
1721 if (rc)
1722 break;
1723 }
1724
1725 if (i && qp->rx_dma_chan)
1726 dma_async_issue_pending(qp->rx_dma_chan);
1727
1728 if (i == qp->rx_max_entry) {
1729 /* there is more work to do */
1730 if (qp->active)
1731 tasklet_schedule(&qp->rxc_db_work);
1732 } else if (ntb_db_read(qp->ndev) & BIT_ULL(qp->qp_num)) {
1733 /* the doorbell bit is set: clear it */
1734 ntb_db_clear(qp->ndev, BIT_ULL(qp->qp_num));
1735 /* ntb_db_read ensures ntb_db_clear write is committed */
1736 ntb_db_read(qp->ndev);
1737
1738 /* an interrupt may have arrived between finishing
1739 * ntb_process_rxc and clearing the doorbell bit:
1740 * there might be some more work to do.
1741 */
1742 if (qp->active)
1743 tasklet_schedule(&qp->rxc_db_work);
1744 }
1745 }
1746
1747 static void ntb_tx_copy_callback(void *data,
1748 const struct dmaengine_result *res)
1749 {
1750 struct ntb_queue_entry *entry = data;
1751 struct ntb_transport_qp *qp = entry->qp;
1752 struct ntb_payload_header __iomem *hdr = entry->tx_hdr;
1753
1754 /* we need to check DMA results if we are using DMA */
1755 if (res) {
1756 enum dmaengine_tx_result dma_err = res->result;
1757
1758 switch (dma_err) {
1759 case DMA_TRANS_READ_FAILED:
1760 case DMA_TRANS_WRITE_FAILED:
1761 entry->errors++;
1762 fallthrough;
1763 case DMA_TRANS_ABORTED:
1764 {
1765 void __iomem *offset =
1766 qp->tx_mw + qp->tx_max_frame *
1767 entry->tx_index;
1768
1769 /* resubmit via CPU */
1770 ntb_memcpy_tx(entry, offset);
1771 qp->tx_memcpy++;
1772 return;
1773 }
1774
1775 case DMA_TRANS_NOERROR:
1776 default:
1777 break;
1778 }
1779 }
1780
1781 iowrite32(entry->flags | DESC_DONE_FLAG, &hdr->flags);
1782
1783 if (qp->use_msi)
1784 ntb_msi_peer_trigger(qp->ndev, PIDX, &qp->peer_msi_desc);
1785 else
1786 ntb_peer_db_set(qp->ndev, BIT_ULL(qp->qp_num));
1787
1788 /* The entry length can only be zero if the packet is intended to be a
1789 * "link down" or similar. Since no payload is being sent in these
1790 * cases, there is nothing to add to the completion queue.
1791 */
1792 if (entry->len > 0) {
1793 qp->tx_bytes += entry->len;
1794
1795 if (qp->tx_handler)
1796 qp->tx_handler(qp, qp->cb_data, entry->cb_data,
1797 entry->len);
1798 }
1799
1800 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry, &qp->tx_free_q);
1801 }
1802
1803 static void ntb_memcpy_tx(struct ntb_queue_entry *entry, void __iomem *offset)
1804 {
1805 #ifdef ARCH_HAS_NOCACHE_UACCESS
1806 /*
1807 * Using non-temporal mov to improve performance on non-cached
1808 * writes, even though we aren't actually copying from user space.
1809 */
1810 __copy_from_user_inatomic_nocache(offset, entry->buf, entry->len);
1811 #else
1812 memcpy_toio(offset, entry->buf, entry->len);
1813 #endif
1814
1815 /* Ensure that the data is fully copied out before setting the flags */
1816 wmb();
1817
1818 ntb_tx_copy_callback(entry, NULL);
1819 }
1820
1821 static int ntb_async_tx_submit(struct ntb_transport_qp *qp,
1822 struct ntb_queue_entry *entry)
1823 {
1824 struct dma_async_tx_descriptor *txd;
1825 struct dma_chan *chan = qp->tx_dma_chan;
1826 struct dma_device *device;
1827 size_t len = entry->len;
1828 void *buf = entry->buf;
1829 size_t dest_off, buff_off;
1830 struct dmaengine_unmap_data *unmap;
1831 dma_addr_t dest;
1832 dma_cookie_t cookie;
1833
1834 device = chan->device;
1835 dest = qp->tx_mw_dma_addr + qp->tx_max_frame * entry->tx_index;
1836 buff_off = (size_t)buf & ~PAGE_MASK;
1837 dest_off = (size_t)dest & ~PAGE_MASK;
1838
1839 if (!is_dma_copy_aligned(device, buff_off, dest_off, len))
1840 goto err;
1841
1842 unmap = dmaengine_get_unmap_data(device->dev, 1, GFP_NOWAIT);
1843 if (!unmap)
1844 goto err;
1845
1846 unmap->len = len;
1847 unmap->addr[0] = dma_map_page(device->dev, virt_to_page(buf),
1848 buff_off, len, DMA_TO_DEVICE);
1849 if (dma_mapping_error(device->dev, unmap->addr[0]))
1850 goto err_get_unmap;
1851
1852 unmap->to_cnt = 1;
1853
1854 txd = device->device_prep_dma_memcpy(chan, dest, unmap->addr[0], len,
1855 DMA_PREP_INTERRUPT);
1856 if (!txd)
1857 goto err_get_unmap;
1858
1859 txd->callback_result = ntb_tx_copy_callback;
1860 txd->callback_param = entry;
1861 dma_set_unmap(txd, unmap);
1862
1863 cookie = dmaengine_submit(txd);
1864 if (dma_submit_error(cookie))
1865 goto err_set_unmap;
1866
1867 dmaengine_unmap_put(unmap);
1868
1869 dma_async_issue_pending(chan);
1870
1871 return 0;
1872 err_set_unmap:
1873 dmaengine_unmap_put(unmap);
1874 err_get_unmap:
1875 dmaengine_unmap_put(unmap);
1876 err:
1877 return -ENXIO;
1878 }
1879
1880 static void ntb_async_tx(struct ntb_transport_qp *qp,
1881 struct ntb_queue_entry *entry)
1882 {
1883 struct ntb_payload_header __iomem *hdr;
1884 struct dma_chan *chan = qp->tx_dma_chan;
1885 void __iomem *offset;
1886 int res;
1887
1888 entry->tx_index = qp->tx_index;
1889 offset = qp->tx_mw + qp->tx_max_frame * entry->tx_index;
1890 hdr = offset + qp->tx_max_frame - sizeof(struct ntb_payload_header);
1891 entry->tx_hdr = hdr;
1892
1893 iowrite32(entry->len, &hdr->len);
1894 iowrite32((u32)qp->tx_pkts, &hdr->ver);
1895
1896 if (!chan)
1897 goto err;
1898
1899 if (entry->len < copy_bytes)
1900 goto err;
1901
1902 res = ntb_async_tx_submit(qp, entry);
1903 if (res < 0)
1904 goto err;
1905
1906 if (!entry->retries)
1907 qp->tx_async++;
1908
1909 return;
1910
1911 err:
1912 ntb_memcpy_tx(entry, offset);
1913 qp->tx_memcpy++;
1914 }
1915
1916 static int ntb_process_tx(struct ntb_transport_qp *qp,
1917 struct ntb_queue_entry *entry)
1918 {
1919 if (!ntb_transport_tx_free_entry(qp)) {
1920 qp->tx_ring_full++;
1921 return -EAGAIN;
1922 }
1923
1924 if (entry->len > qp->tx_max_frame - sizeof(struct ntb_payload_header)) {
1925 if (qp->tx_handler)
1926 qp->tx_handler(qp, qp->cb_data, NULL, -EIO);
1927
1928 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
1929 &qp->tx_free_q);
1930 return 0;
1931 }
1932
1933 ntb_async_tx(qp, entry);
1934
1935 qp->tx_index++;
1936 qp->tx_index %= qp->tx_max_entry;
1937
1938 qp->tx_pkts++;
1939
1940 return 0;
1941 }
1942
1943 static void ntb_send_link_down(struct ntb_transport_qp *qp)
1944 {
1945 struct pci_dev *pdev = qp->ndev->pdev;
1946 struct ntb_queue_entry *entry;
1947 int i, rc;
1948
1949 if (!qp->link_is_up)
1950 return;
1951
1952 dev_info(&pdev->dev, "qp %d: Send Link Down\n", qp->qp_num);
1953
1954 for (i = 0; i < NTB_LINK_DOWN_TIMEOUT; i++) {
1955 entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
1956 if (entry)
1957 break;
1958 msleep(100);
1959 }
1960
1961 if (!entry)
1962 return;
1963
1964 entry->cb_data = NULL;
1965 entry->buf = NULL;
1966 entry->len = 0;
1967 entry->flags = LINK_DOWN_FLAG;
1968
1969 rc = ntb_process_tx(qp, entry);
1970 if (rc)
1971 dev_err(&pdev->dev, "ntb: QP%d unable to send linkdown msg\n",
1972 qp->qp_num);
1973
1974 ntb_qp_link_down_reset(qp);
1975 }
1976
1977 static bool ntb_dma_filter_fn(struct dma_chan *chan, void *node)
1978 {
1979 return dev_to_node(&chan->dev->device) == (int)(unsigned long)node;
1980 }
1981
1982 /**
1983 * ntb_transport_create_queue - Create a new NTB transport layer queue
1984 * @data: pointer for callback data
1985 * @client_dev: &struct device pointer
1986 * @handlers: pointer to various ntb queue (callback) handlers
1987 *
1988 * Create a new NTB transport layer queue and provide the queue with a callback
1989 * routine for both transmit and receive. The receive callback routine will be
1990 * used to pass up data when the transport has received it on the queue. The
1991 * transmit callback routine will be called when the transport has completed the
1992 * transmission of the data on the queue and the data is ready to be freed.
1993 *
1994 * RETURNS: pointer to newly created ntb_queue, NULL on error.
1995 */
1996 struct ntb_transport_qp *
1997 ntb_transport_create_queue(void *data, struct device *client_dev,
1998 const struct ntb_queue_handlers *handlers)
1999 {
2000 struct ntb_dev *ndev;
2001 struct pci_dev *pdev;
2002 struct ntb_transport_ctx *nt;
2003 struct ntb_queue_entry *entry;
2004 struct ntb_transport_qp *qp;
2005 u64 qp_bit;
2006 unsigned int free_queue;
2007 dma_cap_mask_t dma_mask;
2008 int node;
2009 int i;
2010
2011 ndev = dev_ntb(client_dev->parent);
2012 pdev = ndev->pdev;
2013 nt = ndev->ctx;
2014
2015 node = dev_to_node(&ndev->dev);
2016
2017 free_queue = ffs(nt->qp_bitmap_free);
2018 if (!free_queue)
2019 goto err;
2020
2021 /* decrement free_queue to make it zero based */
2022 free_queue--;
2023
2024 qp = &nt->qp_vec[free_queue];
2025 qp_bit = BIT_ULL(qp->qp_num);
2026
2027 nt->qp_bitmap_free &= ~qp_bit;
2028
2029 qp->cb_data = data;
2030 qp->rx_handler = handlers->rx_handler;
2031 qp->tx_handler = handlers->tx_handler;
2032 qp->event_handler = handlers->event_handler;
2033
2034 dma_cap_zero(dma_mask);
2035 dma_cap_set(DMA_MEMCPY, dma_mask);
2036
2037 if (use_dma) {
2038 qp->tx_dma_chan =
2039 dma_request_channel(dma_mask, ntb_dma_filter_fn,
2040 (void *)(unsigned long)node);
2041 if (!qp->tx_dma_chan)
2042 dev_info(&pdev->dev, "Unable to allocate TX DMA channel\n");
2043
2044 qp->rx_dma_chan =
2045 dma_request_channel(dma_mask, ntb_dma_filter_fn,
2046 (void *)(unsigned long)node);
2047 if (!qp->rx_dma_chan)
2048 dev_info(&pdev->dev, "Unable to allocate RX DMA channel\n");
2049 } else {
2050 qp->tx_dma_chan = NULL;
2051 qp->rx_dma_chan = NULL;
2052 }
2053
2054 qp->tx_mw_dma_addr = 0;
2055 if (qp->tx_dma_chan) {
2056 qp->tx_mw_dma_addr =
2057 dma_map_resource(qp->tx_dma_chan->device->dev,
2058 qp->tx_mw_phys, qp->tx_mw_size,
2059 DMA_FROM_DEVICE, 0);
2060 if (dma_mapping_error(qp->tx_dma_chan->device->dev,
2061 qp->tx_mw_dma_addr)) {
2062 qp->tx_mw_dma_addr = 0;
2063 goto err1;
2064 }
2065 }
2066
2067 dev_dbg(&pdev->dev, "Using %s memcpy for TX\n",
2068 qp->tx_dma_chan ? "DMA" : "CPU");
2069
2070 dev_dbg(&pdev->dev, "Using %s memcpy for RX\n",
2071 qp->rx_dma_chan ? "DMA" : "CPU");
2072
2073 for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) {
2074 entry = kzalloc_node(sizeof(*entry), GFP_KERNEL, node);
2075 if (!entry)
2076 goto err1;
2077
2078 entry->qp = qp;
2079 ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry,
2080 &qp->rx_free_q);
2081 }
2082 qp->rx_alloc_entry = NTB_QP_DEF_NUM_ENTRIES;
2083
2084 for (i = 0; i < qp->tx_max_entry; i++) {
2085 entry = kzalloc_node(sizeof(*entry), GFP_KERNEL, node);
2086 if (!entry)
2087 goto err2;
2088
2089 entry->qp = qp;
2090 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
2091 &qp->tx_free_q);
2092 }
2093
2094 ntb_db_clear(qp->ndev, qp_bit);
2095 ntb_db_clear_mask(qp->ndev, qp_bit);
2096
2097 dev_info(&pdev->dev, "NTB Transport QP %d created\n", qp->qp_num);
2098
2099 return qp;
2100
2101 err2:
2102 while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
2103 kfree(entry);
2104 err1:
2105 qp->rx_alloc_entry = 0;
2106 while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q)))
2107 kfree(entry);
2108 if (qp->tx_mw_dma_addr)
2109 dma_unmap_resource(qp->tx_dma_chan->device->dev,
2110 qp->tx_mw_dma_addr, qp->tx_mw_size,
2111 DMA_FROM_DEVICE, 0);
2112 if (qp->tx_dma_chan)
2113 dma_release_channel(qp->tx_dma_chan);
2114 if (qp->rx_dma_chan)
2115 dma_release_channel(qp->rx_dma_chan);
2116 nt->qp_bitmap_free |= qp_bit;
2117 err:
2118 return NULL;
2119 }
2120 EXPORT_SYMBOL_GPL(ntb_transport_create_queue);
2121
2122 /**
2123 * ntb_transport_free_queue - Frees NTB transport queue
2124 * @qp: NTB queue to be freed
2125 *
2126 * Frees NTB transport queue
2127 */
2128 void ntb_transport_free_queue(struct ntb_transport_qp *qp)
2129 {
2130 struct pci_dev *pdev;
2131 struct ntb_queue_entry *entry;
2132 u64 qp_bit;
2133
2134 if (!qp)
2135 return;
2136
2137 pdev = qp->ndev->pdev;
2138
2139 qp->active = false;
2140
2141 if (qp->tx_dma_chan) {
2142 struct dma_chan *chan = qp->tx_dma_chan;
2143 /* Putting the dma_chan to NULL will force any new traffic to be
2144 * processed by the CPU instead of the DAM engine
2145 */
2146 qp->tx_dma_chan = NULL;
2147
2148 /* Try to be nice and wait for any queued DMA engine
2149 * transactions to process before smashing it with a rock
2150 */
2151 dma_sync_wait(chan, qp->last_cookie);
2152 dmaengine_terminate_all(chan);
2153
2154 dma_unmap_resource(chan->device->dev,
2155 qp->tx_mw_dma_addr, qp->tx_mw_size,
2156 DMA_FROM_DEVICE, 0);
2157
2158 dma_release_channel(chan);
2159 }
2160
2161 if (qp->rx_dma_chan) {
2162 struct dma_chan *chan = qp->rx_dma_chan;
2163 /* Putting the dma_chan to NULL will force any new traffic to be
2164 * processed by the CPU instead of the DAM engine
2165 */
2166 qp->rx_dma_chan = NULL;
2167
2168 /* Try to be nice and wait for any queued DMA engine
2169 * transactions to process before smashing it with a rock
2170 */
2171 dma_sync_wait(chan, qp->last_cookie);
2172 dmaengine_terminate_all(chan);
2173 dma_release_channel(chan);
2174 }
2175
2176 qp_bit = BIT_ULL(qp->qp_num);
2177
2178 ntb_db_set_mask(qp->ndev, qp_bit);
2179 tasklet_kill(&qp->rxc_db_work);
2180
2181 cancel_delayed_work_sync(&qp->link_work);
2182
2183 qp->cb_data = NULL;
2184 qp->rx_handler = NULL;
2185 qp->tx_handler = NULL;
2186 qp->event_handler = NULL;
2187
2188 while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q)))
2189 kfree(entry);
2190
2191 while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_pend_q))) {
2192 dev_warn(&pdev->dev, "Freeing item from non-empty rx_pend_q\n");
2193 kfree(entry);
2194 }
2195
2196 while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_post_q))) {
2197 dev_warn(&pdev->dev, "Freeing item from non-empty rx_post_q\n");
2198 kfree(entry);
2199 }
2200
2201 while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
2202 kfree(entry);
2203
2204 qp->transport->qp_bitmap_free |= qp_bit;
2205
2206 dev_info(&pdev->dev, "NTB Transport QP %d freed\n", qp->qp_num);
2207 }
2208 EXPORT_SYMBOL_GPL(ntb_transport_free_queue);
2209
2210 /**
2211 * ntb_transport_rx_remove - Dequeues enqueued rx packet
2212 * @qp: NTB queue to be freed
2213 * @len: pointer to variable to write enqueued buffers length
2214 *
2215 * Dequeues unused buffers from receive queue. Should only be used during
2216 * shutdown of qp.
2217 *
2218 * RETURNS: NULL error value on error, or void* for success.
2219 */
2220 void *ntb_transport_rx_remove(struct ntb_transport_qp *qp, unsigned int *len)
2221 {
2222 struct ntb_queue_entry *entry;
2223 void *buf;
2224
2225 if (!qp || qp->client_ready)
2226 return NULL;
2227
2228 entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_pend_q);
2229 if (!entry)
2230 return NULL;
2231
2232 buf = entry->cb_data;
2233 *len = entry->len;
2234
2235 ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry, &qp->rx_free_q);
2236
2237 return buf;
2238 }
2239 EXPORT_SYMBOL_GPL(ntb_transport_rx_remove);
2240
2241 /**
2242 * ntb_transport_rx_enqueue - Enqueue a new NTB queue entry
2243 * @qp: NTB transport layer queue the entry is to be enqueued on
2244 * @cb: per buffer pointer for callback function to use
2245 * @data: pointer to data buffer that incoming packets will be copied into
2246 * @len: length of the data buffer
2247 *
2248 * Enqueue a new receive buffer onto the transport queue into which a NTB
2249 * payload can be received into.
2250 *
2251 * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
2252 */
2253 int ntb_transport_rx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
2254 unsigned int len)
2255 {
2256 struct ntb_queue_entry *entry;
2257
2258 if (!qp)
2259 return -EINVAL;
2260
2261 entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q);
2262 if (!entry)
2263 return -ENOMEM;
2264
2265 entry->cb_data = cb;
2266 entry->buf = data;
2267 entry->len = len;
2268 entry->flags = 0;
2269 entry->retries = 0;
2270 entry->errors = 0;
2271 entry->rx_index = 0;
2272
2273 ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry, &qp->rx_pend_q);
2274
2275 if (qp->active)
2276 tasklet_schedule(&qp->rxc_db_work);
2277
2278 return 0;
2279 }
2280 EXPORT_SYMBOL_GPL(ntb_transport_rx_enqueue);
2281
2282 /**
2283 * ntb_transport_tx_enqueue - Enqueue a new NTB queue entry
2284 * @qp: NTB transport layer queue the entry is to be enqueued on
2285 * @cb: per buffer pointer for callback function to use
2286 * @data: pointer to data buffer that will be sent
2287 * @len: length of the data buffer
2288 *
2289 * Enqueue a new transmit buffer onto the transport queue from which a NTB
2290 * payload will be transmitted. This assumes that a lock is being held to
2291 * serialize access to the qp.
2292 *
2293 * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
2294 */
2295 int ntb_transport_tx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
2296 unsigned int len)
2297 {
2298 struct ntb_queue_entry *entry;
2299 int rc;
2300
2301 if (!qp || !len)
2302 return -EINVAL;
2303
2304 /* If the qp link is down already, just ignore. */
2305 if (!qp->link_is_up)
2306 return 0;
2307
2308 entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
2309 if (!entry) {
2310 qp->tx_err_no_buf++;
2311 return -EBUSY;
2312 }
2313
2314 entry->cb_data = cb;
2315 entry->buf = data;
2316 entry->len = len;
2317 entry->flags = 0;
2318 entry->errors = 0;
2319 entry->retries = 0;
2320 entry->tx_index = 0;
2321
2322 rc = ntb_process_tx(qp, entry);
2323 if (rc)
2324 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
2325 &qp->tx_free_q);
2326
2327 return rc;
2328 }
2329 EXPORT_SYMBOL_GPL(ntb_transport_tx_enqueue);
2330
2331 /**
2332 * ntb_transport_link_up - Notify NTB transport of client readiness to use queue
2333 * @qp: NTB transport layer queue to be enabled
2334 *
2335 * Notify NTB transport layer of client readiness to use queue
2336 */
2337 void ntb_transport_link_up(struct ntb_transport_qp *qp)
2338 {
2339 if (!qp)
2340 return;
2341
2342 qp->client_ready = true;
2343
2344 if (qp->transport->link_is_up)
2345 schedule_delayed_work(&qp->link_work, 0);
2346 }
2347 EXPORT_SYMBOL_GPL(ntb_transport_link_up);
2348
2349 /**
2350 * ntb_transport_link_down - Notify NTB transport to no longer enqueue data
2351 * @qp: NTB transport layer queue to be disabled
2352 *
2353 * Notify NTB transport layer of client's desire to no longer receive data on
2354 * transport queue specified. It is the client's responsibility to ensure all
2355 * entries on queue are purged or otherwise handled appropriately.
2356 */
2357 void ntb_transport_link_down(struct ntb_transport_qp *qp)
2358 {
2359 int val;
2360
2361 if (!qp)
2362 return;
2363
2364 qp->client_ready = false;
2365
2366 val = ntb_spad_read(qp->ndev, QP_LINKS);
2367
2368 ntb_peer_spad_write(qp->ndev, PIDX, QP_LINKS, val & ~BIT(qp->qp_num));
2369
2370 if (qp->link_is_up)
2371 ntb_send_link_down(qp);
2372 else
2373 cancel_delayed_work_sync(&qp->link_work);
2374 }
2375 EXPORT_SYMBOL_GPL(ntb_transport_link_down);
2376
2377 /**
2378 * ntb_transport_link_query - Query transport link state
2379 * @qp: NTB transport layer queue to be queried
2380 *
2381 * Query connectivity to the remote system of the NTB transport queue
2382 *
2383 * RETURNS: true for link up or false for link down
2384 */
2385 bool ntb_transport_link_query(struct ntb_transport_qp *qp)
2386 {
2387 if (!qp)
2388 return false;
2389
2390 return qp->link_is_up;
2391 }
2392 EXPORT_SYMBOL_GPL(ntb_transport_link_query);
2393
2394 /**
2395 * ntb_transport_qp_num - Query the qp number
2396 * @qp: NTB transport layer queue to be queried
2397 *
2398 * Query qp number of the NTB transport queue
2399 *
2400 * RETURNS: a zero based number specifying the qp number
2401 */
2402 unsigned char ntb_transport_qp_num(struct ntb_transport_qp *qp)
2403 {
2404 if (!qp)
2405 return 0;
2406
2407 return qp->qp_num;
2408 }
2409 EXPORT_SYMBOL_GPL(ntb_transport_qp_num);
2410
2411 /**
2412 * ntb_transport_max_size - Query the max payload size of a qp
2413 * @qp: NTB transport layer queue to be queried
2414 *
2415 * Query the maximum payload size permissible on the given qp
2416 *
2417 * RETURNS: the max payload size of a qp
2418 */
2419 unsigned int ntb_transport_max_size(struct ntb_transport_qp *qp)
2420 {
2421 unsigned int max_size;
2422 unsigned int copy_align;
2423 struct dma_chan *rx_chan, *tx_chan;
2424
2425 if (!qp)
2426 return 0;
2427
2428 rx_chan = qp->rx_dma_chan;
2429 tx_chan = qp->tx_dma_chan;
2430
2431 copy_align = max(rx_chan ? rx_chan->device->copy_align : 0,
2432 tx_chan ? tx_chan->device->copy_align : 0);
2433
2434 /* If DMA engine usage is possible, try to find the max size for that */
2435 max_size = qp->tx_max_frame - sizeof(struct ntb_payload_header);
2436 max_size = round_down(max_size, 1 << copy_align);
2437
2438 return max_size;
2439 }
2440 EXPORT_SYMBOL_GPL(ntb_transport_max_size);
2441
2442 unsigned int ntb_transport_tx_free_entry(struct ntb_transport_qp *qp)
2443 {
2444 unsigned int head = qp->tx_index;
2445 unsigned int tail = qp->remote_rx_info->entry;
2446
2447 return tail >= head ? tail - head : qp->tx_max_entry + tail - head;
2448 }
2449 EXPORT_SYMBOL_GPL(ntb_transport_tx_free_entry);
2450
2451 static void ntb_transport_doorbell_callback(void *data, int vector)
2452 {
2453 struct ntb_transport_ctx *nt = data;
2454 struct ntb_transport_qp *qp;
2455 u64 db_bits;
2456 unsigned int qp_num;
2457
2458 if (ntb_db_read(nt->ndev) & nt->msi_db_mask) {
2459 ntb_transport_msi_peer_desc_changed(nt);
2460 ntb_db_clear(nt->ndev, nt->msi_db_mask);
2461 }
2462
2463 db_bits = (nt->qp_bitmap & ~nt->qp_bitmap_free &
2464 ntb_db_vector_mask(nt->ndev, vector));
2465
2466 while (db_bits) {
2467 qp_num = __ffs(db_bits);
2468 qp = &nt->qp_vec[qp_num];
2469
2470 if (qp->active)
2471 tasklet_schedule(&qp->rxc_db_work);
2472
2473 db_bits &= ~BIT_ULL(qp_num);
2474 }
2475 }
2476
2477 static const struct ntb_ctx_ops ntb_transport_ops = {
2478 .link_event = ntb_transport_event_callback,
2479 .db_event = ntb_transport_doorbell_callback,
2480 };
2481
2482 static struct ntb_client ntb_transport_client = {
2483 .ops = {
2484 .probe = ntb_transport_probe,
2485 .remove = ntb_transport_free,
2486 },
2487 };
2488
2489 static int __init ntb_transport_init(void)
2490 {
2491 int rc;
2492
2493 pr_info("%s, version %s\n", NTB_TRANSPORT_DESC, NTB_TRANSPORT_VER);
2494
2495 if (debugfs_initialized())
2496 nt_debugfs_dir = debugfs_create_dir(KBUILD_MODNAME, NULL);
2497
2498 rc = bus_register(&ntb_transport_bus);
2499 if (rc)
2500 goto err_bus;
2501
2502 rc = ntb_register_client(&ntb_transport_client);
2503 if (rc)
2504 goto err_client;
2505
2506 return 0;
2507
2508 err_client:
2509 bus_unregister(&ntb_transport_bus);
2510 err_bus:
2511 debugfs_remove_recursive(nt_debugfs_dir);
2512 return rc;
2513 }
2514 module_init(ntb_transport_init);
2515
2516 static void __exit ntb_transport_exit(void)
2517 {
2518 ntb_unregister_client(&ntb_transport_client);
2519 bus_unregister(&ntb_transport_bus);
2520 debugfs_remove_recursive(nt_debugfs_dir);
2521 }
2522 module_exit(ntb_transport_exit);
Kernel DRM miscellaneous fixes and cross-tree changes