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	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 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(new: &nt->entry, head: &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(entry: &client_dev->entry);
340	device_unregister(dev: &client_dev->dev);
341	}
342
343	list_del(entry: &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(objp: 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(dev: &client->dev), device_name,
368	strlen(device_name))) {
369	list_del(entry: &client->entry);
370	device_unregister(dev: &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	int ntb_transport_register_client_dev(char *device_name)
382	{
383	struct ntb_transport_client_dev *client_dev;
384	struct ntb_transport_ctx *nt;
385	int node;
386	int rc, i = 0;
387
388	if (list_empty(head: &ntb_transport_list))
389	return -ENODEV;
390
391	list_for_each_entry(nt, &ntb_transport_list, entry) {
392	struct device *dev;
393
394	node = dev_to_node(dev: &nt->ndev->dev);
395
396	client_dev = kzalloc_node(size: sizeof(*client_dev),
397	GFP_KERNEL, node);
398	if (!client_dev) {
399	rc = -ENOMEM;
400	goto err;
401	}
402
403	dev = &client_dev->dev;
404
405	/* setup and register client devices */
406	dev_set_name(dev, name: "%s%d", device_name, i);
407	dev->bus = &ntb_transport_bus;
408	dev->release = ntb_transport_client_release;
409	dev->parent = &nt->ndev->dev;
410
411	rc = device_register(dev);
412	if (rc) {
413	put_device(dev);
414	goto err;
415	}
416
417	list_add_tail(new: &client_dev->entry, head: &nt->client_devs);
418	i++;
419	}
420
421	return 0;
422
423	err:
424	ntb_transport_unregister_client_dev(device_name);
425
426	return rc;
427	}
428	EXPORT_SYMBOL_GPL(ntb_transport_register_client_dev);
429
430	/**
431	* ntb_transport_register_client - Register NTB client driver
432	* @drv: NTB client driver to be registered
433	*
434	* Register an NTB client driver with the NTB transport layer
435	*
436	* RETURNS: An appropriate -ERRNO error value on error, or zero for success.
437	*/
438	int ntb_transport_register_client(struct ntb_transport_client *drv)
439	{
440	drv->driver.bus = &ntb_transport_bus;
441
442	if (list_empty(head: &ntb_transport_list))
443	return -ENODEV;
444
445	return driver_register(drv: &drv->driver);
446	}
447	EXPORT_SYMBOL_GPL(ntb_transport_register_client);
448
449	/**
450	* ntb_transport_unregister_client - Unregister NTB client driver
451	* @drv: NTB client driver to be unregistered
452	*
453	* Unregister an NTB client driver with the NTB transport layer
454	*
455	* RETURNS: An appropriate -ERRNO error value on error, or zero for success.
456	*/
457	void ntb_transport_unregister_client(struct ntb_transport_client *drv)
458	{
459	driver_unregister(drv: &drv->driver);
460	}
461	EXPORT_SYMBOL_GPL(ntb_transport_unregister_client);
462
463	static ssize_t debugfs_read(struct file *filp, char __user *ubuf, size_t count,
464	loff_t *offp)
465	{
466	struct ntb_transport_qp *qp;
467	char *buf;
468	ssize_t ret, out_offset, out_count;
469
470	qp = filp->private_data;
471
472	if (!qp || !qp->link_is_up)
473	return 0;
474
475	out_count = 1000;
476
477	buf = kmalloc(size: out_count, GFP_KERNEL);
478	if (!buf)
479	return -ENOMEM;
480
481	out_offset = 0;
482	out_offset += scnprintf(buf: buf + out_offset, size: out_count - out_offset,
483	fmt: "\nNTB QP stats:\n\n");
484	out_offset += scnprintf(buf: buf + out_offset, size: out_count - out_offset,
485	fmt: "rx_bytes - \t%llu\n", qp->rx_bytes);
486	out_offset += scnprintf(buf: buf + out_offset, size: out_count - out_offset,
487	fmt: "rx_pkts - \t%llu\n", qp->rx_pkts);
488	out_offset += scnprintf(buf: buf + out_offset, size: out_count - out_offset,
489	fmt: "rx_memcpy - \t%llu\n", qp->rx_memcpy);
490	out_offset += scnprintf(buf: buf + out_offset, size: out_count - out_offset,
491	fmt: "rx_async - \t%llu\n", qp->rx_async);
492	out_offset += scnprintf(buf: buf + out_offset, size: out_count - out_offset,
493	fmt: "rx_ring_empty - %llu\n", qp->rx_ring_empty);
494	out_offset += scnprintf(buf: buf + out_offset, size: out_count - out_offset,
495	fmt: "rx_err_no_buf - %llu\n", qp->rx_err_no_buf);
496	out_offset += scnprintf(buf: buf + out_offset, size: out_count - out_offset,
497	fmt: "rx_err_oflow - \t%llu\n", qp->rx_err_oflow);
498	out_offset += scnprintf(buf: buf + out_offset, size: out_count - out_offset,
499	fmt: "rx_err_ver - \t%llu\n", qp->rx_err_ver);
500	out_offset += scnprintf(buf: buf + out_offset, size: out_count - out_offset,
501	fmt: "rx_buff - \t0x%p\n", qp->rx_buff);
502	out_offset += scnprintf(buf: buf + out_offset, size: out_count - out_offset,
503	fmt: "rx_index - \t%u\n", qp->rx_index);
504	out_offset += scnprintf(buf: buf + out_offset, size: out_count - out_offset,
505	fmt: "rx_max_entry - \t%u\n", qp->rx_max_entry);
506	out_offset += scnprintf(buf: buf + out_offset, size: out_count - out_offset,
507	fmt: "rx_alloc_entry - \t%u\n\n", qp->rx_alloc_entry);
508
509	out_offset += scnprintf(buf: buf + out_offset, size: out_count - out_offset,
510	fmt: "tx_bytes - \t%llu\n", qp->tx_bytes);
511	out_offset += scnprintf(buf: buf + out_offset, size: out_count - out_offset,
512	fmt: "tx_pkts - \t%llu\n", qp->tx_pkts);
513	out_offset += scnprintf(buf: buf + out_offset, size: out_count - out_offset,
514	fmt: "tx_memcpy - \t%llu\n", qp->tx_memcpy);
515	out_offset += scnprintf(buf: buf + out_offset, size: out_count - out_offset,
516	fmt: "tx_async - \t%llu\n", qp->tx_async);
517	out_offset += scnprintf(buf: buf + out_offset, size: out_count - out_offset,
518	fmt: "tx_ring_full - \t%llu\n", qp->tx_ring_full);
519	out_offset += scnprintf(buf: buf + out_offset, size: out_count - out_offset,
520	fmt: "tx_err_no_buf - %llu\n", qp->tx_err_no_buf);
521	out_offset += scnprintf(buf: buf + out_offset, size: out_count - out_offset,
522	fmt: "tx_mw - \t0x%p\n", qp->tx_mw);
523	out_offset += scnprintf(buf: buf + out_offset, size: out_count - out_offset,
524	fmt: "tx_index (H) - \t%u\n", qp->tx_index);
525	out_offset += scnprintf(buf: buf + out_offset, size: out_count - out_offset,
526	fmt: "RRI (T) - \t%u\n",
527	qp->remote_rx_info->entry);
528	out_offset += scnprintf(buf: buf + out_offset, size: out_count - out_offset,
529	fmt: "tx_max_entry - \t%u\n", qp->tx_max_entry);
530	out_offset += scnprintf(buf: buf + out_offset, size: out_count - out_offset,
531	fmt: "free tx - \t%u\n",
532	ntb_transport_tx_free_entry(qp));
533
534	out_offset += scnprintf(buf: buf + out_offset, size: out_count - out_offset,
535	fmt: "\n");
536	out_offset += scnprintf(buf: buf + out_offset, size: out_count - out_offset,
537	fmt: "Using TX DMA - \t%s\n",
538	qp->tx_dma_chan ? "Yes" : "No");
539	out_offset += scnprintf(buf: buf + out_offset, size: out_count - out_offset,
540	fmt: "Using RX DMA - \t%s\n",
541	qp->rx_dma_chan ? "Yes" : "No");
542	out_offset += scnprintf(buf: buf + out_offset, size: out_count - out_offset,
543	fmt: "QP Link - \t%s\n",
544	qp->link_is_up ? "Up" : "Down");
545	out_offset += scnprintf(buf: buf + out_offset, size: out_count - out_offset,
546	fmt: "\n");
547
548	if (out_offset > out_count)
549	out_offset = out_count;
550
551	ret = simple_read_from_buffer(to: ubuf, count, ppos: offp, from: buf, available: out_offset);
552	kfree(objp: buf);
553	return ret;
554	}
555
556	static const struct file_operations ntb_qp_debugfs_stats = {
557	.owner = THIS_MODULE,
558	.open = simple_open,
559	.read = debugfs_read,
560	};
561
562	static void ntb_list_add(spinlock_t *lock, struct list_head *entry,
563	struct list_head *list)
564	{
565	unsigned long flags;
566
567	spin_lock_irqsave(lock, flags);
568	list_add_tail(new: entry, head: list);
569	spin_unlock_irqrestore(lock, flags);
570	}
571
572	static struct ntb_queue_entry *ntb_list_rm(spinlock_t *lock,
573	struct list_head *list)
574	{
575	struct ntb_queue_entry *entry;
576	unsigned long flags;
577
578	spin_lock_irqsave(lock, flags);
579	if (list_empty(head: list)) {
580	entry = NULL;
581	goto out;
582	}
583	entry = list_first_entry(list, struct ntb_queue_entry, entry);
584	list_del(entry: &entry->entry);
585
586	out:
587	spin_unlock_irqrestore(lock, flags);
588
589	return entry;
590	}
591
592	static struct ntb_queue_entry *ntb_list_mv(spinlock_t *lock,
593	struct list_head *list,
594	struct list_head *to_list)
595	{
596	struct ntb_queue_entry *entry;
597	unsigned long flags;
598
599	spin_lock_irqsave(lock, flags);
600
601	if (list_empty(head: list)) {
602	entry = NULL;
603	} else {
604	entry = list_first_entry(list, struct ntb_queue_entry, entry);
605	list_move_tail(list: &entry->entry, head: to_list);
606	}
607
608	spin_unlock_irqrestore(lock, flags);
609
610	return entry;
611	}
612
613	static int ntb_transport_setup_qp_mw(struct ntb_transport_ctx *nt,
614	unsigned int qp_num)
615	{
616	struct ntb_transport_qp *qp = &nt->qp_vec[qp_num];
617	struct ntb_transport_mw *mw;
618	struct ntb_dev *ndev = nt->ndev;
619	struct ntb_queue_entry *entry;
620	unsigned int rx_size, num_qps_mw;
621	unsigned int mw_num, mw_count, qp_count;
622	unsigned int i;
623	int node;
624
625	mw_count = nt->mw_count;
626	qp_count = nt->qp_count;
627
628	mw_num = QP_TO_MW(nt, qp_num);
629	mw = &nt->mw_vec[mw_num];
630
631	if (!mw->virt_addr)
632	return -ENOMEM;
633
634	if (mw_num < qp_count % mw_count)
635	num_qps_mw = qp_count / mw_count + 1;
636	else
637	num_qps_mw = qp_count / mw_count;
638
639	rx_size = (unsigned int)mw->xlat_size / num_qps_mw;
640	qp->rx_buff = mw->virt_addr + rx_size * (qp_num / mw_count);
641	rx_size -= sizeof(struct ntb_rx_info);
642
643	qp->remote_rx_info = qp->rx_buff + rx_size;
644
645	/* Due to housekeeping, there must be atleast 2 buffs */
646	qp->rx_max_frame = min(transport_mtu, rx_size / 2);
647	qp->rx_max_entry = rx_size / qp->rx_max_frame;
648	qp->rx_index = 0;
649
650	/*
651	* Checking to see if we have more entries than the default.
652	* We should add additional entries if that is the case so we
653	* can be in sync with the transport frames.
654	*/
655	node = dev_to_node(dev: &ndev->dev);
656	for (i = qp->rx_alloc_entry; i < qp->rx_max_entry; i++) {
657	entry = kzalloc_node(size: sizeof(*entry), GFP_KERNEL, node);
658	if (!entry)
659	return -ENOMEM;
660
661	entry->qp = qp;
662	ntb_list_add(lock: &qp->ntb_rx_q_lock, entry: &entry->entry,
663	list: &qp->rx_free_q);
664	qp->rx_alloc_entry++;
665	}
666
667	qp->remote_rx_info->entry = qp->rx_max_entry - 1;
668
669	/* setup the hdr offsets with 0's */
670	for (i = 0; i < qp->rx_max_entry; i++) {
671	void *offset = (qp->rx_buff + qp->rx_max_frame * (i + 1) -
672	sizeof(struct ntb_payload_header));
673	memset(offset, 0, sizeof(struct ntb_payload_header));
674	}
675
676	qp->rx_pkts = 0;
677	qp->tx_pkts = 0;
678	qp->tx_index = 0;
679
680	return 0;
681	}
682
683	static irqreturn_t ntb_transport_isr(int irq, void *dev)
684	{
685	struct ntb_transport_qp *qp = dev;
686
687	tasklet_schedule(t: &qp->rxc_db_work);
688
689	return IRQ_HANDLED;
690	}
691
692	static void ntb_transport_setup_qp_peer_msi(struct ntb_transport_ctx *nt,
693	unsigned int qp_num)
694	{
695	struct ntb_transport_qp *qp = &nt->qp_vec[qp_num];
696	int spad = qp_num * 2 + nt->msi_spad_offset;
697
698	if (!nt->use_msi)
699	return;
700
701	if (spad >= ntb_spad_count(ntb: nt->ndev))
702	return;
703
704	qp->peer_msi_desc.addr_offset =
705	ntb_peer_spad_read(ntb: qp->ndev, PIDX, sidx: spad);
706	qp->peer_msi_desc.data =
707	ntb_peer_spad_read(ntb: qp->ndev, PIDX, sidx: spad + 1);
708
709	dev_dbg(&qp->ndev->pdev->dev, "QP%d Peer MSI addr=%x data=%x\n",
710	qp_num, qp->peer_msi_desc.addr_offset, qp->peer_msi_desc.data);
711
712	if (qp->peer_msi_desc.addr_offset) {
713	qp->use_msi = true;
714	dev_info(&qp->ndev->pdev->dev,
715	"Using MSI interrupts for QP%d\n", qp_num);
716	}
717	}
718
719	static void ntb_transport_setup_qp_msi(struct ntb_transport_ctx *nt,
720	unsigned int qp_num)
721	{
722	struct ntb_transport_qp *qp = &nt->qp_vec[qp_num];
723	int spad = qp_num * 2 + nt->msi_spad_offset;
724	int rc;
725
726	if (!nt->use_msi)
727	return;
728
729	if (spad >= ntb_spad_count(ntb: nt->ndev)) {
730	dev_warn_once(&qp->ndev->pdev->dev,
731	"Not enough SPADS to use MSI interrupts\n");
732	return;
733	}
734
735	ntb_spad_write(ntb: qp->ndev, sidx: spad, val: 0);
736	ntb_spad_write(ntb: qp->ndev, sidx: spad + 1, val: 0);
737
738	if (!qp->msi_irq) {
739	qp->msi_irq = ntbm_msi_request_irq(ntb: qp->ndev, handler: ntb_transport_isr,
740	KBUILD_MODNAME, dev_id: qp,
741	msi_desc: &qp->msi_desc);
742	if (qp->msi_irq < 0) {
743	dev_warn(&qp->ndev->pdev->dev,
744	"Unable to allocate MSI interrupt for qp%d\n",
745	qp_num);
746	return;
747	}
748	}
749
750	rc = ntb_spad_write(ntb: qp->ndev, sidx: spad, val: qp->msi_desc.addr_offset);
751	if (rc)
752	goto err_free_interrupt;
753
754	rc = ntb_spad_write(ntb: qp->ndev, sidx: spad + 1, val: qp->msi_desc.data);
755	if (rc)
756	goto err_free_interrupt;
757
758	dev_dbg(&qp->ndev->pdev->dev, "QP%d MSI %d addr=%x data=%x\n",
759	qp_num, qp->msi_irq, qp->msi_desc.addr_offset,
760	qp->msi_desc.data);
761
762	return;
763
764	err_free_interrupt:
765	devm_free_irq(dev: &nt->ndev->dev, irq: qp->msi_irq, dev_id: qp);
766	}
767
768	static void ntb_transport_msi_peer_desc_changed(struct ntb_transport_ctx *nt)
769	{
770	int i;
771
772	dev_dbg(&nt->ndev->pdev->dev, "Peer MSI descriptors changed");
773
774	for (i = 0; i < nt->qp_count; i++)
775	ntb_transport_setup_qp_peer_msi(nt, qp_num: i);
776	}
777
778	static void ntb_transport_msi_desc_changed(void *data)
779	{
780	struct ntb_transport_ctx *nt = data;
781	int i;
782
783	dev_dbg(&nt->ndev->pdev->dev, "MSI descriptors changed");
784
785	for (i = 0; i < nt->qp_count; i++)
786	ntb_transport_setup_qp_msi(nt, qp_num: i);
787
788	ntb_peer_db_set(ntb: nt->ndev, db_bits: nt->msi_db_mask);
789	}
790
791	static void ntb_free_mw(struct ntb_transport_ctx *nt, int num_mw)
792	{
793	struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
794	struct pci_dev *pdev = nt->ndev->pdev;
795
796	if (!mw->virt_addr)
797	return;
798
799	ntb_mw_clear_trans(ntb: nt->ndev, PIDX, widx: num_mw);
800	dma_free_coherent(dev: &pdev->dev, size: mw->alloc_size,
801	cpu_addr: mw->alloc_addr, dma_handle: mw->dma_addr);
802	mw->xlat_size = 0;
803	mw->buff_size = 0;
804	mw->alloc_size = 0;
805	mw->alloc_addr = NULL;
806	mw->virt_addr = NULL;
807	}
808
809	static int ntb_alloc_mw_buffer(struct ntb_transport_mw *mw,
810	struct device *dma_dev, size_t align)
811	{
812	dma_addr_t dma_addr;
813	void *alloc_addr, *virt_addr;
814	int rc;
815
816	alloc_addr = dma_alloc_coherent(dev: dma_dev, size: mw->alloc_size,
817	dma_handle: &dma_addr, GFP_KERNEL);
818	if (!alloc_addr) {
819	dev_err(dma_dev, "Unable to alloc MW buff of size %zu\n",
820	mw->alloc_size);
821	return -ENOMEM;
822	}
823	virt_addr = alloc_addr;
824
825	/*
826	* we must ensure that the memory address allocated is BAR size
827	* aligned in order for the XLAT register to take the value. This
828	* is a requirement of the hardware. It is recommended to setup CMA
829	* for BAR sizes equal or greater than 4MB.
830	*/
831	if (!IS_ALIGNED(dma_addr, align)) {
832	if (mw->alloc_size > mw->buff_size) {
833	virt_addr = PTR_ALIGN(alloc_addr, align);
834	dma_addr = ALIGN(dma_addr, align);
835	} else {
836	rc = -ENOMEM;
837	goto err;
838	}
839	}
840
841	mw->alloc_addr = alloc_addr;
842	mw->virt_addr = virt_addr;
843	mw->dma_addr = dma_addr;
844
845	return 0;
846
847	err:
848	dma_free_coherent(dev: dma_dev, size: mw->alloc_size, cpu_addr: alloc_addr, dma_handle: dma_addr);
849
850	return rc;
851	}
852
853	static int ntb_set_mw(struct ntb_transport_ctx *nt, int num_mw,
854	resource_size_t size)
855	{
856	struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
857	struct pci_dev *pdev = nt->ndev->pdev;
858	size_t xlat_size, buff_size;
859	resource_size_t xlat_align;
860	resource_size_t xlat_align_size;
861	int rc;
862
863	if (!size)
864	return -EINVAL;
865
866	rc = ntb_mw_get_align(ntb: nt->ndev, PIDX, widx: num_mw, addr_align: &xlat_align,
867	size_align: &xlat_align_size, NULL);
868	if (rc)
869	return rc;
870
871	xlat_size = round_up(size, xlat_align_size);
872	buff_size = round_up(size, xlat_align);
873
874	/* No need to re-setup */
875	if (mw->xlat_size == xlat_size)
876	return 0;
877
878	if (mw->buff_size)
879	ntb_free_mw(nt, num_mw);
880
881	/* Alloc memory for receiving data. Must be aligned */
882	mw->xlat_size = xlat_size;
883	mw->buff_size = buff_size;
884	mw->alloc_size = buff_size;
885
886	rc = ntb_alloc_mw_buffer(mw, dma_dev: &pdev->dev, align: xlat_align);
887	if (rc) {
888	mw->alloc_size *= 2;
889	rc = ntb_alloc_mw_buffer(mw, dma_dev: &pdev->dev, align: xlat_align);
890	if (rc) {
891	dev_err(&pdev->dev,
892	"Unable to alloc aligned MW buff\n");
893	mw->xlat_size = 0;
894	mw->buff_size = 0;
895	mw->alloc_size = 0;
896	return rc;
897	}
898	}
899
900	/* Notify HW the memory location of the receive buffer */
901	rc = ntb_mw_set_trans(ntb: nt->ndev, PIDX, widx: num_mw, addr: mw->dma_addr,
902	size: mw->xlat_size);
903	if (rc) {
904	dev_err(&pdev->dev, "Unable to set mw%d translation", num_mw);
905	ntb_free_mw(nt, num_mw);
906	return -EIO;
907	}
908
909	return 0;
910	}
911
912	static void ntb_qp_link_context_reset(struct ntb_transport_qp *qp)
913	{
914	qp->link_is_up = false;
915	qp->active = false;
916
917	qp->tx_index = 0;
918	qp->rx_index = 0;
919	qp->rx_bytes = 0;
920	qp->rx_pkts = 0;
921	qp->rx_ring_empty = 0;
922	qp->rx_err_no_buf = 0;
923	qp->rx_err_oflow = 0;
924	qp->rx_err_ver = 0;
925	qp->rx_memcpy = 0;
926	qp->rx_async = 0;
927	qp->tx_bytes = 0;
928	qp->tx_pkts = 0;
929	qp->tx_ring_full = 0;
930	qp->tx_err_no_buf = 0;
931	qp->tx_memcpy = 0;
932	qp->tx_async = 0;
933	}
934
935	static void ntb_qp_link_down_reset(struct ntb_transport_qp *qp)
936	{
937	ntb_qp_link_context_reset(qp);
938	if (qp->remote_rx_info)
939	qp->remote_rx_info->entry = qp->rx_max_entry - 1;
940	}
941
942	static void ntb_qp_link_cleanup(struct ntb_transport_qp *qp)
943	{
944	struct ntb_transport_ctx *nt = qp->transport;
945	struct pci_dev *pdev = nt->ndev->pdev;
946
947	dev_info(&pdev->dev, "qp %d: Link Cleanup\n", qp->qp_num);
948
949	cancel_delayed_work_sync(dwork: &qp->link_work);
950	ntb_qp_link_down_reset(qp);
951
952	if (qp->event_handler)
953	qp->event_handler(qp->cb_data, qp->link_is_up);
954	}
955
956	static void ntb_qp_link_cleanup_work(struct work_struct *work)
957	{
958	struct ntb_transport_qp *qp = container_of(work,
959	struct ntb_transport_qp,
960	link_cleanup);
961	struct ntb_transport_ctx *nt = qp->transport;
962
963	ntb_qp_link_cleanup(qp);
964
965	if (nt->link_is_up)
966	schedule_delayed_work(dwork: &qp->link_work,
967	delay: msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
968	}
969
970	static void ntb_qp_link_down(struct ntb_transport_qp *qp)
971	{
972	schedule_work(work: &qp->link_cleanup);
973	}
974
975	static void ntb_transport_link_cleanup(struct ntb_transport_ctx *nt)
976	{
977	struct ntb_transport_qp *qp;
978	u64 qp_bitmap_alloc;
979	unsigned int i, count;
980
981	qp_bitmap_alloc = nt->qp_bitmap & ~nt->qp_bitmap_free;
982
983	/* Pass along the info to any clients */
984	for (i = 0; i < nt->qp_count; i++)
985	if (qp_bitmap_alloc & BIT_ULL(i)) {
986	qp = &nt->qp_vec[i];
987	ntb_qp_link_cleanup(qp);
988	cancel_work_sync(work: &qp->link_cleanup);
989	cancel_delayed_work_sync(dwork: &qp->link_work);
990	}
991
992	if (!nt->link_is_up)
993	cancel_delayed_work_sync(dwork: &nt->link_work);
994
995	for (i = 0; i < nt->mw_count; i++)
996	ntb_free_mw(nt, num_mw: i);
997
998	/* The scratchpad registers keep the values if the remote side
999	* goes down, blast them now to give them a sane value the next
1000	* time they are accessed
1001	*/
1002	count = ntb_spad_count(ntb: nt->ndev);
1003	for (i = 0; i < count; i++)
1004	ntb_spad_write(ntb: nt->ndev, sidx: i, val: 0);
1005	}
1006
1007	static void ntb_transport_link_cleanup_work(struct work_struct *work)
1008	{
1009	struct ntb_transport_ctx *nt =
1010	container_of(work, struct ntb_transport_ctx, link_cleanup);
1011
1012	ntb_transport_link_cleanup(nt);
1013	}
1014
1015	static void ntb_transport_event_callback(void *data)
1016	{
1017	struct ntb_transport_ctx *nt = data;
1018
1019	if (ntb_link_is_up(ntb: nt->ndev, NULL, NULL) == 1)
1020	schedule_delayed_work(dwork: &nt->link_work, delay: 0);
1021	else
1022	schedule_work(work: &nt->link_cleanup);
1023	}
1024
1025	static void ntb_transport_link_work(struct work_struct *work)
1026	{
1027	struct ntb_transport_ctx *nt =
1028	container_of(work, struct ntb_transport_ctx, link_work.work);
1029	struct ntb_dev *ndev = nt->ndev;
1030	struct pci_dev *pdev = ndev->pdev;
1031	resource_size_t size;
1032	u32 val;
1033	int rc = 0, i, spad;
1034
1035	/* send the local info, in the opposite order of the way we read it */
1036
1037	if (nt->use_msi) {
1038	rc = ntb_msi_setup_mws(ntb: ndev);
1039	if (rc) {
1040	dev_warn(&pdev->dev,
1041	"Failed to register MSI memory window: %d\n",
1042	rc);
1043	nt->use_msi = false;
1044	}
1045	}
1046
1047	for (i = 0; i < nt->qp_count; i++)
1048	ntb_transport_setup_qp_msi(nt, qp_num: i);
1049
1050	for (i = 0; i < nt->mw_count; i++) {
1051	size = nt->mw_vec[i].phys_size;
1052
1053	if (max_mw_size && size > max_mw_size)
1054	size = max_mw_size;
1055
1056	spad = MW0_SZ_HIGH + (i * 2);
1057	ntb_peer_spad_write(ntb: ndev, PIDX, sidx: spad, upper_32_bits(size));
1058
1059	spad = MW0_SZ_LOW + (i * 2);
1060	ntb_peer_spad_write(ntb: ndev, PIDX, sidx: spad, lower_32_bits(size));
1061	}
1062
1063	ntb_peer_spad_write(ntb: ndev, PIDX, sidx: NUM_MWS, val: nt->mw_count);
1064
1065	ntb_peer_spad_write(ntb: ndev, PIDX, sidx: NUM_QPS, val: nt->qp_count);
1066
1067	ntb_peer_spad_write(ntb: ndev, PIDX, sidx: VERSION, NTB_TRANSPORT_VERSION);
1068
1069	/* Query the remote side for its info */
1070	val = ntb_spad_read(ntb: ndev, sidx: VERSION);
1071	dev_dbg(&pdev->dev, "Remote version = %d\n", val);
1072	if (val != NTB_TRANSPORT_VERSION)
1073	goto out;
1074
1075	val = ntb_spad_read(ntb: ndev, sidx: NUM_QPS);
1076	dev_dbg(&pdev->dev, "Remote max number of qps = %d\n", val);
1077	if (val != nt->qp_count)
1078	goto out;
1079
1080	val = ntb_spad_read(ntb: ndev, sidx: NUM_MWS);
1081	dev_dbg(&pdev->dev, "Remote number of mws = %d\n", val);
1082	if (val != nt->mw_count)
1083	goto out;
1084
1085	for (i = 0; i < nt->mw_count; i++) {
1086	u64 val64;
1087
1088	val = ntb_spad_read(ntb: ndev, sidx: MW0_SZ_HIGH + (i * 2));
1089	val64 = (u64)val << 32;
1090
1091	val = ntb_spad_read(ntb: ndev, sidx: MW0_SZ_LOW + (i * 2));
1092	val64 |= val;
1093
1094	dev_dbg(&pdev->dev, "Remote MW%d size = %#llx\n", i, val64);
1095
1096	rc = ntb_set_mw(nt, num_mw: i, size: val64);
1097	if (rc)
1098	goto out1;
1099	}
1100
1101	nt->link_is_up = true;
1102
1103	for (i = 0; i < nt->qp_count; i++) {
1104	struct ntb_transport_qp *qp = &nt->qp_vec[i];
1105
1106	ntb_transport_setup_qp_mw(nt, qp_num: i);
1107	ntb_transport_setup_qp_peer_msi(nt, qp_num: i);
1108
1109	if (qp->client_ready)
1110	schedule_delayed_work(dwork: &qp->link_work, delay: 0);
1111	}
1112
1113	return;
1114
1115	out1:
1116	for (i = 0; i < nt->mw_count; i++)
1117	ntb_free_mw(nt, num_mw: i);
1118
1119	/* if there's an actual failure, we should just bail */
1120	if (rc < 0)
1121	return;
1122
1123	out:
1124	if (ntb_link_is_up(ntb: ndev, NULL, NULL) == 1)
1125	schedule_delayed_work(dwork: &nt->link_work,
1126	delay: msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
1127	}
1128
1129	static void ntb_qp_link_work(struct work_struct *work)
1130	{
1131	struct ntb_transport_qp *qp = container_of(work,
1132	struct ntb_transport_qp,
1133	link_work.work);
1134	struct pci_dev *pdev = qp->ndev->pdev;
1135	struct ntb_transport_ctx *nt = qp->transport;
1136	int val;
1137
1138	WARN_ON(!nt->link_is_up);
1139
1140	val = ntb_spad_read(ntb: nt->ndev, sidx: QP_LINKS);
1141
1142	ntb_peer_spad_write(ntb: nt->ndev, PIDX, sidx: QP_LINKS, val: val | BIT(qp->qp_num));
1143
1144	/* query remote spad for qp ready bits */
1145	dev_dbg_ratelimited(&pdev->dev, "Remote QP link status = %x\n", val);
1146
1147	/* See if the remote side is up */
1148	if (val & BIT(qp->qp_num)) {
1149	dev_info(&pdev->dev, "qp %d: Link Up\n", qp->qp_num);
1150	qp->link_is_up = true;
1151	qp->active = true;
1152
1153	if (qp->event_handler)
1154	qp->event_handler(qp->cb_data, qp->link_is_up);
1155
1156	if (qp->active)
1157	tasklet_schedule(t: &qp->rxc_db_work);
1158	} else if (nt->link_is_up)
1159	schedule_delayed_work(dwork: &qp->link_work,
1160	delay: msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
1161	}
1162
1163	static int ntb_transport_init_queue(struct ntb_transport_ctx *nt,
1164	unsigned int qp_num)
1165	{
1166	struct ntb_transport_qp *qp;
1167	phys_addr_t mw_base;
1168	resource_size_t mw_size;
1169	unsigned int num_qps_mw, tx_size;
1170	unsigned int mw_num, mw_count, qp_count;
1171	u64 qp_offset;
1172
1173	mw_count = nt->mw_count;
1174	qp_count = nt->qp_count;
1175
1176	mw_num = QP_TO_MW(nt, qp_num);
1177
1178	qp = &nt->qp_vec[qp_num];
1179	qp->qp_num = qp_num;
1180	qp->transport = nt;
1181	qp->ndev = nt->ndev;
1182	qp->client_ready = false;
1183	qp->event_handler = NULL;
1184	ntb_qp_link_context_reset(qp);
1185
1186	if (mw_num < qp_count % mw_count)
1187	num_qps_mw = qp_count / mw_count + 1;
1188	else
1189	num_qps_mw = qp_count / mw_count;
1190
1191	mw_base = nt->mw_vec[mw_num].phys_addr;
1192	mw_size = nt->mw_vec[mw_num].phys_size;
1193
1194	if (max_mw_size && mw_size > max_mw_size)
1195	mw_size = max_mw_size;
1196
1197	tx_size = (unsigned int)mw_size / num_qps_mw;
1198	qp_offset = tx_size * (qp_num / mw_count);
1199
1200	qp->tx_mw_size = tx_size;
1201	qp->tx_mw = nt->mw_vec[mw_num].vbase + qp_offset;
1202	if (!qp->tx_mw)
1203	return -EINVAL;
1204
1205	qp->tx_mw_phys = mw_base + qp_offset;
1206	if (!qp->tx_mw_phys)
1207	return -EINVAL;
1208
1209	tx_size -= sizeof(struct ntb_rx_info);
1210	qp->rx_info = qp->tx_mw + tx_size;
1211
1212	/* Due to housekeeping, there must be atleast 2 buffs */
1213	qp->tx_max_frame = min(transport_mtu, tx_size / 2);
1214	qp->tx_max_entry = tx_size / qp->tx_max_frame;
1215
1216	if (nt->debugfs_node_dir) {
1217	char debugfs_name[4];
1218
1219	snprintf(buf: debugfs_name, size: 4, fmt: "qp%d", qp_num);
1220	qp->debugfs_dir = debugfs_create_dir(name: debugfs_name,
1221	parent: nt->debugfs_node_dir);
1222
1223	qp->debugfs_stats = debugfs_create_file(name: "stats", S_IRUSR,
1224	parent: qp->debugfs_dir, data: qp,
1225	fops: &ntb_qp_debugfs_stats);
1226	} else {
1227	qp->debugfs_dir = NULL;
1228	qp->debugfs_stats = NULL;
1229	}
1230
1231	INIT_DELAYED_WORK(&qp->link_work, ntb_qp_link_work);
1232	INIT_WORK(&qp->link_cleanup, ntb_qp_link_cleanup_work);
1233
1234	spin_lock_init(&qp->ntb_rx_q_lock);
1235	spin_lock_init(&qp->ntb_tx_free_q_lock);
1236
1237	INIT_LIST_HEAD(list: &qp->rx_post_q);
1238	INIT_LIST_HEAD(list: &qp->rx_pend_q);
1239	INIT_LIST_HEAD(list: &qp->rx_free_q);
1240	INIT_LIST_HEAD(list: &qp->tx_free_q);
1241
1242	tasklet_init(t: &qp->rxc_db_work, func: ntb_transport_rxc_db,
1243	data: (unsigned long)qp);
1244
1245	return 0;
1246	}
1247
1248	static int ntb_transport_probe(struct ntb_client *self, struct ntb_dev *ndev)
1249	{
1250	struct ntb_transport_ctx *nt;
1251	struct ntb_transport_mw *mw;
1252	unsigned int mw_count, qp_count, spad_count, max_mw_count_for_spads;
1253	u64 qp_bitmap;
1254	int node;
1255	int rc, i;
1256
1257	mw_count = ntb_peer_mw_count(ntb: ndev);
1258
1259	if (!ndev->ops->mw_set_trans) {
1260	dev_err(&ndev->dev, "Inbound MW based NTB API is required\n");
1261	return -EINVAL;
1262	}
1263
1264	if (ntb_db_is_unsafe(ntb: ndev))
1265	dev_dbg(&ndev->dev,
1266	"doorbell is unsafe, proceed anyway...\n");
1267	if (ntb_spad_is_unsafe(ntb: ndev))
1268	dev_dbg(&ndev->dev,
1269	"scratchpad is unsafe, proceed anyway...\n");
1270
1271	if (ntb_peer_port_count(ntb: ndev) != NTB_DEF_PEER_CNT)
1272	dev_warn(&ndev->dev, "Multi-port NTB devices unsupported\n");
1273
1274	node = dev_to_node(dev: &ndev->dev);
1275
1276	nt = kzalloc_node(size: sizeof(*nt), GFP_KERNEL, node);
1277	if (!nt)
1278	return -ENOMEM;
1279
1280	nt->ndev = ndev;
1281
1282	/*
1283	* If we are using MSI, and have at least one extra memory window,
1284	* we will reserve the last MW for the MSI window.
1285	*/
1286	if (use_msi && mw_count > 1) {
1287	rc = ntb_msi_init(ntb: ndev, desc_changed: ntb_transport_msi_desc_changed);
1288	if (!rc) {
1289	mw_count -= 1;
1290	nt->use_msi = true;
1291	}
1292	}
1293
1294	spad_count = ntb_spad_count(ntb: ndev);
1295
1296	/* Limit the MW's based on the availability of scratchpads */
1297
1298	if (spad_count < NTB_TRANSPORT_MIN_SPADS) {
1299	nt->mw_count = 0;
1300	rc = -EINVAL;
1301	goto err;
1302	}
1303
1304	max_mw_count_for_spads = (spad_count - MW0_SZ_HIGH) / 2;
1305	nt->mw_count = min(mw_count, max_mw_count_for_spads);
1306
1307	nt->msi_spad_offset = nt->mw_count * 2 + MW0_SZ_HIGH;
1308
1309	nt->mw_vec = kcalloc_node(n: mw_count, size: sizeof(*nt->mw_vec),
1310	GFP_KERNEL, node);
1311	if (!nt->mw_vec) {
1312	rc = -ENOMEM;
1313	goto err;
1314	}
1315
1316	for (i = 0; i < mw_count; i++) {
1317	mw = &nt->mw_vec[i];
1318
1319	rc = ntb_peer_mw_get_addr(ntb: ndev, widx: i, base: &mw->phys_addr,
1320	size: &mw->phys_size);
1321	if (rc)
1322	goto err1;
1323
1324	mw->vbase = ioremap_wc(offset: mw->phys_addr, size: mw->phys_size);
1325	if (!mw->vbase) {
1326	rc = -ENOMEM;
1327	goto err1;
1328	}
1329
1330	mw->buff_size = 0;
1331	mw->xlat_size = 0;
1332	mw->virt_addr = NULL;
1333	mw->dma_addr = 0;
1334	}
1335
1336	qp_bitmap = ntb_db_valid_mask(ntb: ndev);
1337
1338	qp_count = ilog2(qp_bitmap);
1339	if (nt->use_msi) {
1340	qp_count -= 1;
1341	nt->msi_db_mask = 1 << qp_count;
1342	ntb_db_clear_mask(ntb: ndev, db_bits: nt->msi_db_mask);
1343	}
1344
1345	if (max_num_clients && max_num_clients < qp_count)
1346	qp_count = max_num_clients;
1347	else if (nt->mw_count < qp_count)
1348	qp_count = nt->mw_count;
1349
1350	qp_bitmap &= BIT_ULL(qp_count) - 1;
1351
1352	nt->qp_count = qp_count;
1353	nt->qp_bitmap = qp_bitmap;
1354	nt->qp_bitmap_free = qp_bitmap;
1355
1356	nt->qp_vec = kcalloc_node(n: qp_count, size: sizeof(*nt->qp_vec),
1357	GFP_KERNEL, node);
1358	if (!nt->qp_vec) {
1359	rc = -ENOMEM;
1360	goto err1;
1361	}
1362
1363	if (nt_debugfs_dir) {
1364	nt->debugfs_node_dir =
1365	debugfs_create_dir(name: pci_name(pdev: ndev->pdev),
1366	parent: nt_debugfs_dir);
1367	}
1368
1369	for (i = 0; i < qp_count; i++) {
1370	rc = ntb_transport_init_queue(nt, qp_num: i);
1371	if (rc)
1372	goto err2;
1373	}
1374
1375	INIT_DELAYED_WORK(&nt->link_work, ntb_transport_link_work);
1376	INIT_WORK(&nt->link_cleanup, ntb_transport_link_cleanup_work);
1377
1378	rc = ntb_set_ctx(ntb: ndev, ctx: nt, ctx_ops: &ntb_transport_ops);
1379	if (rc)
1380	goto err2;
1381
1382	INIT_LIST_HEAD(list: &nt->client_devs);
1383	rc = ntb_bus_init(nt);
1384	if (rc)
1385	goto err3;
1386
1387	nt->link_is_up = false;
1388	ntb_link_enable(ntb: ndev, max_speed: NTB_SPEED_AUTO, max_width: NTB_WIDTH_AUTO);
1389	ntb_link_event(ntb: ndev);
1390
1391	return 0;
1392
1393	err3:
1394	ntb_clear_ctx(ntb: ndev);
1395	err2:
1396	kfree(objp: nt->qp_vec);
1397	err1:
1398	while (i--) {
1399	mw = &nt->mw_vec[i];
1400	iounmap(addr: mw->vbase);
1401	}
1402	kfree(objp: nt->mw_vec);
1403	err:
1404	kfree(objp: nt);
1405	return rc;
1406	}
1407
1408	static void ntb_transport_free(struct ntb_client *self, struct ntb_dev *ndev)
1409	{
1410	struct ntb_transport_ctx *nt = ndev->ctx;
1411	struct ntb_transport_qp *qp;
1412	u64 qp_bitmap_alloc;
1413	int i;
1414
1415	ntb_transport_link_cleanup(nt);
1416	cancel_work_sync(work: &nt->link_cleanup);
1417	cancel_delayed_work_sync(dwork: &nt->link_work);
1418
1419	qp_bitmap_alloc = nt->qp_bitmap & ~nt->qp_bitmap_free;
1420
1421	/* verify that all the qp's are freed */
1422	for (i = 0; i < nt->qp_count; i++) {
1423	qp = &nt->qp_vec[i];
1424	if (qp_bitmap_alloc & BIT_ULL(i))
1425	ntb_transport_free_queue(qp);
1426	debugfs_remove_recursive(dentry: qp->debugfs_dir);
1427	}
1428
1429	ntb_link_disable(ntb: ndev);
1430	ntb_clear_ctx(ntb: ndev);
1431
1432	ntb_bus_remove(nt);
1433
1434	for (i = nt->mw_count; i--; ) {
1435	ntb_free_mw(nt, num_mw: i);
1436	iounmap(addr: nt->mw_vec[i].vbase);
1437	}
1438
1439	kfree(objp: nt->qp_vec);
1440	kfree(objp: nt->mw_vec);
1441	kfree(objp: nt);
1442	}
1443
1444	static void ntb_complete_rxc(struct ntb_transport_qp *qp)
1445	{
1446	struct ntb_queue_entry *entry;
1447	void *cb_data;
1448	unsigned int len;
1449	unsigned long irqflags;
1450
1451	spin_lock_irqsave(&qp->ntb_rx_q_lock, irqflags);
1452
1453	while (!list_empty(head: &qp->rx_post_q)) {
1454	entry = list_first_entry(&qp->rx_post_q,
1455	struct ntb_queue_entry, entry);
1456	if (!(entry->flags & DESC_DONE_FLAG))
1457	break;
1458
1459	entry->rx_hdr->flags = 0;
1460	iowrite32(entry->rx_index, &qp->rx_info->entry);
1461
1462	cb_data = entry->cb_data;
1463	len = entry->len;
1464
1465	list_move_tail(list: &entry->entry, head: &qp->rx_free_q);
1466
1467	spin_unlock_irqrestore(lock: &qp->ntb_rx_q_lock, flags: irqflags);
1468
1469	if (qp->rx_handler && qp->client_ready)
1470	qp->rx_handler(qp, qp->cb_data, cb_data, len);
1471
1472	spin_lock_irqsave(&qp->ntb_rx_q_lock, irqflags);
1473	}
1474
1475	spin_unlock_irqrestore(lock: &qp->ntb_rx_q_lock, flags: irqflags);
1476	}
1477
1478	static void ntb_rx_copy_callback(void *data,
1479	const struct dmaengine_result *res)
1480	{
1481	struct ntb_queue_entry *entry = data;
1482
1483	/* we need to check DMA results if we are using DMA */
1484	if (res) {
1485	enum dmaengine_tx_result dma_err = res->result;
1486
1487	switch (dma_err) {
1488	case DMA_TRANS_READ_FAILED:
1489	case DMA_TRANS_WRITE_FAILED:
1490	entry->errors++;
1491	fallthrough;
1492	case DMA_TRANS_ABORTED:
1493	{
1494	struct ntb_transport_qp *qp = entry->qp;
1495	void *offset = qp->rx_buff + qp->rx_max_frame *
1496	qp->rx_index;
1497
1498	ntb_memcpy_rx(entry, offset);
1499	qp->rx_memcpy++;
1500	return;
1501	}
1502
1503	case DMA_TRANS_NOERROR:
1504	default:
1505	break;
1506	}
1507	}
1508
1509	entry->flags |= DESC_DONE_FLAG;
1510
1511	ntb_complete_rxc(qp: entry->qp);
1512	}
1513
1514	static void ntb_memcpy_rx(struct ntb_queue_entry *entry, void *offset)
1515	{
1516	void *buf = entry->buf;
1517	size_t len = entry->len;
1518
1519	memcpy(buf, offset, len);
1520
1521	/* Ensure that the data is fully copied out before clearing the flag */
1522	wmb();
1523
1524	ntb_rx_copy_callback(data: entry, NULL);
1525	}
1526
1527	static int ntb_async_rx_submit(struct ntb_queue_entry *entry, void *offset)
1528	{
1529	struct dma_async_tx_descriptor *txd;
1530	struct ntb_transport_qp *qp = entry->qp;
1531	struct dma_chan *chan = qp->rx_dma_chan;
1532	struct dma_device *device;
1533	size_t pay_off, buff_off, len;
1534	struct dmaengine_unmap_data *unmap;
1535	dma_cookie_t cookie;
1536	void *buf = entry->buf;
1537
1538	len = entry->len;
1539	device = chan->device;
1540	pay_off = (size_t)offset & ~PAGE_MASK;
1541	buff_off = (size_t)buf & ~PAGE_MASK;
1542
1543	if (!is_dma_copy_aligned(dev: device, off1: pay_off, off2: buff_off, len))
1544	goto err;
1545
1546	unmap = dmaengine_get_unmap_data(dev: device->dev, nr: 2, GFP_NOWAIT);
1547	if (!unmap)
1548	goto err;
1549
1550	unmap->len = len;
1551	unmap->addr[0] = dma_map_page(device->dev, virt_to_page(offset),
1552	pay_off, len, DMA_TO_DEVICE);
1553	if (dma_mapping_error(dev: device->dev, dma_addr: unmap->addr[0]))
1554	goto err_get_unmap;
1555
1556	unmap->to_cnt = 1;
1557
1558	unmap->addr[1] = dma_map_page(device->dev, virt_to_page(buf),
1559	buff_off, len, DMA_FROM_DEVICE);
1560	if (dma_mapping_error(dev: device->dev, dma_addr: unmap->addr[1]))
1561	goto err_get_unmap;
1562
1563	unmap->from_cnt = 1;
1564
1565	txd = device->device_prep_dma_memcpy(chan, unmap->addr[1],
1566	unmap->addr[0], len,
1567	DMA_PREP_INTERRUPT);
1568	if (!txd)
1569	goto err_get_unmap;
1570
1571	txd->callback_result = ntb_rx_copy_callback;
1572	txd->callback_param = entry;
1573	dma_set_unmap(tx: txd, unmap);
1574
1575	cookie = dmaengine_submit(desc: txd);
1576	if (dma_submit_error(cookie))
1577	goto err_set_unmap;
1578
1579	dmaengine_unmap_put(unmap);
1580
1581	qp->last_cookie = cookie;
1582
1583	qp->rx_async++;
1584
1585	return 0;
1586
1587	err_set_unmap:
1588	dmaengine_unmap_put(unmap);
1589	err_get_unmap:
1590	dmaengine_unmap_put(unmap);
1591	err:
1592	return -ENXIO;
1593	}
1594
1595	static void ntb_async_rx(struct ntb_queue_entry *entry, void *offset)
1596	{
1597	struct ntb_transport_qp *qp = entry->qp;
1598	struct dma_chan *chan = qp->rx_dma_chan;
1599	int res;
1600
1601	if (!chan)
1602	goto err;
1603
1604	if (entry->len < copy_bytes)
1605	goto err;
1606
1607	res = ntb_async_rx_submit(entry, offset);
1608	if (res < 0)
1609	goto err;
1610
1611	if (!entry->retries)
1612	qp->rx_async++;
1613
1614	return;
1615
1616	err:
1617	ntb_memcpy_rx(entry, offset);
1618	qp->rx_memcpy++;
1619	}
1620
1621	static int ntb_process_rxc(struct ntb_transport_qp *qp)
1622	{
1623	struct ntb_payload_header *hdr;
1624	struct ntb_queue_entry *entry;
1625	void *offset;
1626
1627	offset = qp->rx_buff + qp->rx_max_frame * qp->rx_index;
1628	hdr = offset + qp->rx_max_frame - sizeof(struct ntb_payload_header);
1629
1630	dev_dbg(&qp->ndev->pdev->dev, "qp %d: RX ver %u len %d flags %x\n",
1631	qp->qp_num, hdr->ver, hdr->len, hdr->flags);
1632
1633	if (!(hdr->flags & DESC_DONE_FLAG)) {
1634	dev_dbg(&qp->ndev->pdev->dev, "done flag not set\n");
1635	qp->rx_ring_empty++;
1636	return -EAGAIN;
1637	}
1638
1639	if (hdr->flags & LINK_DOWN_FLAG) {
1640	dev_dbg(&qp->ndev->pdev->dev, "link down flag set\n");
1641	ntb_qp_link_down(qp);
1642	hdr->flags = 0;
1643	return -EAGAIN;
1644	}
1645
1646	if (hdr->ver != (u32)qp->rx_pkts) {
1647	dev_dbg(&qp->ndev->pdev->dev,
1648	"version mismatch, expected %llu - got %u\n",
1649	qp->rx_pkts, hdr->ver);
1650	qp->rx_err_ver++;
1651	return -EIO;
1652	}
1653
1654	entry = ntb_list_mv(lock: &qp->ntb_rx_q_lock, list: &qp->rx_pend_q, to_list: &qp->rx_post_q);
1655	if (!entry) {
1656	dev_dbg(&qp->ndev->pdev->dev, "no receive buffer\n");
1657	qp->rx_err_no_buf++;
1658	return -EAGAIN;
1659	}
1660
1661	entry->rx_hdr = hdr;
1662	entry->rx_index = qp->rx_index;
1663
1664	if (hdr->len > entry->len) {
1665	dev_dbg(&qp->ndev->pdev->dev,
1666	"receive buffer overflow! Wanted %d got %d\n",
1667	hdr->len, entry->len);
1668	qp->rx_err_oflow++;
1669
1670	entry->len = -EIO;
1671	entry->flags |= DESC_DONE_FLAG;
1672
1673	ntb_complete_rxc(qp);
1674	} else {
1675	dev_dbg(&qp->ndev->pdev->dev,
1676	"RX OK index %u ver %u size %d into buf size %d\n",
1677	qp->rx_index, hdr->ver, hdr->len, entry->len);
1678
1679	qp->rx_bytes += hdr->len;
1680	qp->rx_pkts++;
1681
1682	entry->len = hdr->len;
1683
1684	ntb_async_rx(entry, offset);
1685	}
1686
1687	qp->rx_index++;
1688	qp->rx_index %= qp->rx_max_entry;
1689
1690	return 0;
1691	}
1692
1693	static void ntb_transport_rxc_db(unsigned long data)
1694	{
1695	struct ntb_transport_qp *qp = (void *)data;
1696	int rc, i;
1697
1698	dev_dbg(&qp->ndev->pdev->dev, "%s: doorbell %d received\n",
1699	__func__, qp->qp_num);
1700
1701	/* Limit the number of packets processed in a single interrupt to
1702	* provide fairness to others
1703	*/
1704	for (i = 0; i < qp->rx_max_entry; i++) {
1705	rc = ntb_process_rxc(qp);
1706	if (rc)
1707	break;
1708	}
1709
1710	if (i && qp->rx_dma_chan)
1711	dma_async_issue_pending(chan: qp->rx_dma_chan);
1712
1713	if (i == qp->rx_max_entry) {
1714	/* there is more work to do */
1715	if (qp->active)
1716	tasklet_schedule(t: &qp->rxc_db_work);
1717	} else if (ntb_db_read(ntb: qp->ndev) & BIT_ULL(qp->qp_num)) {
1718	/* the doorbell bit is set: clear it */
1719	ntb_db_clear(ntb: qp->ndev, BIT_ULL(qp->qp_num));
1720	/* ntb_db_read ensures ntb_db_clear write is committed */
1721	ntb_db_read(ntb: qp->ndev);
1722
1723	/* an interrupt may have arrived between finishing
1724	* ntb_process_rxc and clearing the doorbell bit:
1725	* there might be some more work to do.
1726	*/
1727	if (qp->active)
1728	tasklet_schedule(t: &qp->rxc_db_work);
1729	}
1730	}
1731
1732	static void ntb_tx_copy_callback(void *data,
1733	const struct dmaengine_result *res)
1734	{
1735	struct ntb_queue_entry *entry = data;
1736	struct ntb_transport_qp *qp = entry->qp;
1737	struct ntb_payload_header __iomem *hdr = entry->tx_hdr;
1738
1739	/* we need to check DMA results if we are using DMA */
1740	if (res) {
1741	enum dmaengine_tx_result dma_err = res->result;
1742
1743	switch (dma_err) {
1744	case DMA_TRANS_READ_FAILED:
1745	case DMA_TRANS_WRITE_FAILED:
1746	entry->errors++;
1747	fallthrough;
1748	case DMA_TRANS_ABORTED:
1749	{
1750	void __iomem *offset =
1751	qp->tx_mw + qp->tx_max_frame *
1752	entry->tx_index;
1753
1754	/* resubmit via CPU */
1755	ntb_memcpy_tx(entry, offset);
1756	qp->tx_memcpy++;
1757	return;
1758	}
1759
1760	case DMA_TRANS_NOERROR:
1761	default:
1762	break;
1763	}
1764	}
1765
1766	iowrite32(entry->flags | DESC_DONE_FLAG, &hdr->flags);
1767
1768	if (qp->use_msi)
1769	ntb_msi_peer_trigger(ntb: qp->ndev, PIDX, desc: &qp->peer_msi_desc);
1770	else
1771	ntb_peer_db_set(ntb: qp->ndev, BIT_ULL(qp->qp_num));
1772
1773	/* The entry length can only be zero if the packet is intended to be a
1774	* "link down" or similar. Since no payload is being sent in these
1775	* cases, there is nothing to add to the completion queue.
1776	*/
1777	if (entry->len > 0) {
1778	qp->tx_bytes += entry->len;
1779
1780	if (qp->tx_handler)
1781	qp->tx_handler(qp, qp->cb_data, entry->cb_data,
1782	entry->len);
1783	}
1784
1785	ntb_list_add(lock: &qp->ntb_tx_free_q_lock, entry: &entry->entry, list: &qp->tx_free_q);
1786	}
1787
1788	static void ntb_memcpy_tx(struct ntb_queue_entry *entry, void __iomem *offset)
1789	{
1790	#ifdef ARCH_HAS_NOCACHE_UACCESS
1791	/*
1792	* Using non-temporal mov to improve performance on non-cached
1793	* writes, even though we aren't actually copying from user space.
1794	*/
1795	__copy_from_user_inatomic_nocache(dst: offset, src: entry->buf, size: entry->len);
1796	#else
1797	memcpy_toio(offset, entry->buf, entry->len);
1798	#endif
1799
1800	/* Ensure that the data is fully copied out before setting the flags */
1801	wmb();
1802
1803	ntb_tx_copy_callback(data: entry, NULL);
1804	}
1805
1806	static int ntb_async_tx_submit(struct ntb_transport_qp *qp,
1807	struct ntb_queue_entry *entry)
1808	{
1809	struct dma_async_tx_descriptor *txd;
1810	struct dma_chan *chan = qp->tx_dma_chan;
1811	struct dma_device *device;
1812	size_t len = entry->len;
1813	void *buf = entry->buf;
1814	size_t dest_off, buff_off;
1815	struct dmaengine_unmap_data *unmap;
1816	dma_addr_t dest;
1817	dma_cookie_t cookie;
1818
1819	device = chan->device;
1820	dest = qp->tx_mw_dma_addr + qp->tx_max_frame * entry->tx_index;
1821	buff_off = (size_t)buf & ~PAGE_MASK;
1822	dest_off = (size_t)dest & ~PAGE_MASK;
1823
1824	if (!is_dma_copy_aligned(dev: device, off1: buff_off, off2: dest_off, len))
1825	goto err;
1826
1827	unmap = dmaengine_get_unmap_data(dev: device->dev, nr: 1, GFP_NOWAIT);
1828	if (!unmap)
1829	goto err;
1830
1831	unmap->len = len;
1832	unmap->addr[0] = dma_map_page(device->dev, virt_to_page(buf),
1833	buff_off, len, DMA_TO_DEVICE);
1834	if (dma_mapping_error(dev: device->dev, dma_addr: unmap->addr[0]))
1835	goto err_get_unmap;
1836
1837	unmap->to_cnt = 1;
1838
1839	txd = device->device_prep_dma_memcpy(chan, dest, unmap->addr[0], len,
1840	DMA_PREP_INTERRUPT);
1841	if (!txd)
1842	goto err_get_unmap;
1843
1844	txd->callback_result = ntb_tx_copy_callback;
1845	txd->callback_param = entry;
1846	dma_set_unmap(tx: txd, unmap);
1847
1848	cookie = dmaengine_submit(desc: txd);
1849	if (dma_submit_error(cookie))
1850	goto err_set_unmap;
1851
1852	dmaengine_unmap_put(unmap);
1853
1854	dma_async_issue_pending(chan);
1855
1856	return 0;
1857	err_set_unmap:
1858	dmaengine_unmap_put(unmap);
1859	err_get_unmap:
1860	dmaengine_unmap_put(unmap);
1861	err:
1862	return -ENXIO;
1863	}
1864
1865	static void ntb_async_tx(struct ntb_transport_qp *qp,
1866	struct ntb_queue_entry *entry)
1867	{
1868	struct ntb_payload_header __iomem *hdr;
1869	struct dma_chan *chan = qp->tx_dma_chan;
1870	void __iomem *offset;
1871	int res;
1872
1873	entry->tx_index = qp->tx_index;
1874	offset = qp->tx_mw + qp->tx_max_frame * entry->tx_index;
1875	hdr = offset + qp->tx_max_frame - sizeof(struct ntb_payload_header);
1876	entry->tx_hdr = hdr;
1877
1878	iowrite32(entry->len, &hdr->len);
1879	iowrite32((u32)qp->tx_pkts, &hdr->ver);
1880
1881	if (!chan)
1882	goto err;
1883
1884	if (entry->len < copy_bytes)
1885	goto err;
1886
1887	res = ntb_async_tx_submit(qp, entry);
1888	if (res < 0)
1889	goto err;
1890
1891	if (!entry->retries)
1892	qp->tx_async++;
1893
1894	return;
1895
1896	err:
1897	ntb_memcpy_tx(entry, offset);
1898	qp->tx_memcpy++;
1899	}
1900
1901	static int ntb_process_tx(struct ntb_transport_qp *qp,
1902	struct ntb_queue_entry *entry)
1903	{
1904	if (!ntb_transport_tx_free_entry(qp)) {
1905	qp->tx_ring_full++;
1906	return -EAGAIN;
1907	}
1908
1909	if (entry->len > qp->tx_max_frame - sizeof(struct ntb_payload_header)) {
1910	if (qp->tx_handler)
1911	qp->tx_handler(qp, qp->cb_data, NULL, -EIO);
1912
1913	ntb_list_add(lock: &qp->ntb_tx_free_q_lock, entry: &entry->entry,
1914	list: &qp->tx_free_q);
1915	return 0;
1916	}
1917
1918	ntb_async_tx(qp, entry);
1919
1920	qp->tx_index++;
1921	qp->tx_index %= qp->tx_max_entry;
1922
1923	qp->tx_pkts++;
1924
1925	return 0;
1926	}
1927
1928	static void ntb_send_link_down(struct ntb_transport_qp *qp)
1929	{
1930	struct pci_dev *pdev = qp->ndev->pdev;
1931	struct ntb_queue_entry *entry;
1932	int i, rc;
1933
1934	if (!qp->link_is_up)
1935	return;
1936
1937	dev_info(&pdev->dev, "qp %d: Send Link Down\n", qp->qp_num);
1938
1939	for (i = 0; i < NTB_LINK_DOWN_TIMEOUT; i++) {
1940	entry = ntb_list_rm(lock: &qp->ntb_tx_free_q_lock, list: &qp->tx_free_q);
1941	if (entry)
1942	break;
1943	msleep(msecs: 100);
1944	}
1945
1946	if (!entry)
1947	return;
1948
1949	entry->cb_data = NULL;
1950	entry->buf = NULL;
1951	entry->len = 0;
1952	entry->flags = LINK_DOWN_FLAG;
1953
1954	rc = ntb_process_tx(qp, entry);
1955	if (rc)
1956	dev_err(&pdev->dev, "ntb: QP%d unable to send linkdown msg\n",
1957	qp->qp_num);
1958
1959	ntb_qp_link_down_reset(qp);
1960	}
1961
1962	static bool ntb_dma_filter_fn(struct dma_chan *chan, void *node)
1963	{
1964	return dev_to_node(dev: &chan->dev->device) == (int)(unsigned long)node;
1965	}
1966
1967	/**
1968	* ntb_transport_create_queue - Create a new NTB transport layer queue
1969	* @rx_handler: receive callback function
1970	* @tx_handler: transmit callback function
1971	* @event_handler: event callback function
1972	*
1973	* Create a new NTB transport layer queue and provide the queue with a callback
1974	* routine for both transmit and receive. The receive callback routine will be
1975	* used to pass up data when the transport has received it on the queue. The
1976	* transmit callback routine will be called when the transport has completed the
1977	* transmission of the data on the queue and the data is ready to be freed.
1978	*
1979	* RETURNS: pointer to newly created ntb_queue, NULL on error.
1980	*/
1981	struct ntb_transport_qp *
1982	ntb_transport_create_queue(void *data, struct device *client_dev,
1983	const struct ntb_queue_handlers *handlers)
1984	{
1985	struct ntb_dev *ndev;
1986	struct pci_dev *pdev;
1987	struct ntb_transport_ctx *nt;
1988	struct ntb_queue_entry *entry;
1989	struct ntb_transport_qp *qp;
1990	u64 qp_bit;
1991	unsigned int free_queue;
1992	dma_cap_mask_t dma_mask;
1993	int node;
1994	int i;
1995
1996	ndev = dev_ntb(client_dev->parent);
1997	pdev = ndev->pdev;
1998	nt = ndev->ctx;
1999
2000	node = dev_to_node(dev: &ndev->dev);
2001
2002	free_queue = ffs(nt->qp_bitmap_free);
2003	if (!free_queue)
2004	goto err;
2005
2006	/* decrement free_queue to make it zero based */
2007	free_queue--;
2008
2009	qp = &nt->qp_vec[free_queue];
2010	qp_bit = BIT_ULL(qp->qp_num);
2011
2012	nt->qp_bitmap_free &= ~qp_bit;
2013
2014	qp->cb_data = data;
2015	qp->rx_handler = handlers->rx_handler;
2016	qp->tx_handler = handlers->tx_handler;
2017	qp->event_handler = handlers->event_handler;
2018
2019	dma_cap_zero(dma_mask);
2020	dma_cap_set(DMA_MEMCPY, dma_mask);
2021
2022	if (use_dma) {
2023	qp->tx_dma_chan =
2024	dma_request_channel(dma_mask, ntb_dma_filter_fn,
2025	(void *)(unsigned long)node);
2026	if (!qp->tx_dma_chan)
2027	dev_info(&pdev->dev, "Unable to allocate TX DMA channel\n");
2028
2029	qp->rx_dma_chan =
2030	dma_request_channel(dma_mask, ntb_dma_filter_fn,
2031	(void *)(unsigned long)node);
2032	if (!qp->rx_dma_chan)
2033	dev_info(&pdev->dev, "Unable to allocate RX DMA channel\n");
2034	} else {
2035	qp->tx_dma_chan = NULL;
2036	qp->rx_dma_chan = NULL;
2037	}
2038
2039	qp->tx_mw_dma_addr = 0;
2040	if (qp->tx_dma_chan) {
2041	qp->tx_mw_dma_addr =
2042	dma_map_resource(dev: qp->tx_dma_chan->device->dev,
2043	phys_addr: qp->tx_mw_phys, size: qp->tx_mw_size,
2044	dir: DMA_FROM_DEVICE, attrs: 0);
2045	if (dma_mapping_error(dev: qp->tx_dma_chan->device->dev,
2046	dma_addr: qp->tx_mw_dma_addr)) {
2047	qp->tx_mw_dma_addr = 0;
2048	goto err1;
2049	}
2050	}
2051
2052	dev_dbg(&pdev->dev, "Using %s memcpy for TX\n",
2053	qp->tx_dma_chan ? "DMA" : "CPU");
2054
2055	dev_dbg(&pdev->dev, "Using %s memcpy for RX\n",
2056	qp->rx_dma_chan ? "DMA" : "CPU");
2057
2058	for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) {
2059	entry = kzalloc_node(size: sizeof(*entry), GFP_KERNEL, node);
2060	if (!entry)
2061	goto err1;
2062
2063	entry->qp = qp;
2064	ntb_list_add(lock: &qp->ntb_rx_q_lock, entry: &entry->entry,
2065	list: &qp->rx_free_q);
2066	}
2067	qp->rx_alloc_entry = NTB_QP_DEF_NUM_ENTRIES;
2068
2069	for (i = 0; i < qp->tx_max_entry; i++) {
2070	entry = kzalloc_node(size: sizeof(*entry), GFP_KERNEL, node);
2071	if (!entry)
2072	goto err2;
2073
2074	entry->qp = qp;
2075	ntb_list_add(lock: &qp->ntb_tx_free_q_lock, entry: &entry->entry,
2076	list: &qp->tx_free_q);
2077	}
2078
2079	ntb_db_clear(ntb: qp->ndev, db_bits: qp_bit);
2080	ntb_db_clear_mask(ntb: qp->ndev, db_bits: qp_bit);
2081
2082	dev_info(&pdev->dev, "NTB Transport QP %d created\n", qp->qp_num);
2083
2084	return qp;
2085
2086	err2:
2087	while ((entry = ntb_list_rm(lock: &qp->ntb_tx_free_q_lock, list: &qp->tx_free_q)))
2088	kfree(objp: entry);
2089	err1:
2090	qp->rx_alloc_entry = 0;
2091	while ((entry = ntb_list_rm(lock: &qp->ntb_rx_q_lock, list: &qp->rx_free_q)))
2092	kfree(objp: entry);
2093	if (qp->tx_mw_dma_addr)
2094	dma_unmap_resource(dev: qp->tx_dma_chan->device->dev,
2095	addr: qp->tx_mw_dma_addr, size: qp->tx_mw_size,
2096	dir: DMA_FROM_DEVICE, attrs: 0);
2097	if (qp->tx_dma_chan)
2098	dma_release_channel(chan: qp->tx_dma_chan);
2099	if (qp->rx_dma_chan)
2100	dma_release_channel(chan: qp->rx_dma_chan);
2101	nt->qp_bitmap_free |= qp_bit;
2102	err:
2103	return NULL;
2104	}
2105	EXPORT_SYMBOL_GPL(ntb_transport_create_queue);
2106
2107	/**
2108	* ntb_transport_free_queue - Frees NTB transport queue
2109	* @qp: NTB queue to be freed
2110	*
2111	* Frees NTB transport queue
2112	*/
2113	void ntb_transport_free_queue(struct ntb_transport_qp *qp)
2114	{
2115	struct pci_dev *pdev;
2116	struct ntb_queue_entry *entry;
2117	u64 qp_bit;
2118
2119	if (!qp)
2120	return;
2121
2122	pdev = qp->ndev->pdev;
2123
2124	qp->active = false;
2125
2126	if (qp->tx_dma_chan) {
2127	struct dma_chan *chan = qp->tx_dma_chan;
2128	/* Putting the dma_chan to NULL will force any new traffic to be
2129	* processed by the CPU instead of the DAM engine
2130	*/
2131	qp->tx_dma_chan = NULL;
2132
2133	/* Try to be nice and wait for any queued DMA engine
2134	* transactions to process before smashing it with a rock
2135	*/
2136	dma_sync_wait(chan, cookie: qp->last_cookie);
2137	dmaengine_terminate_all(chan);
2138
2139	dma_unmap_resource(dev: chan->device->dev,
2140	addr: qp->tx_mw_dma_addr, size: qp->tx_mw_size,
2141	dir: DMA_FROM_DEVICE, attrs: 0);
2142
2143	dma_release_channel(chan);
2144	}
2145
2146	if (qp->rx_dma_chan) {
2147	struct dma_chan *chan = qp->rx_dma_chan;
2148	/* Putting the dma_chan to NULL will force any new traffic to be
2149	* processed by the CPU instead of the DAM engine
2150	*/
2151	qp->rx_dma_chan = NULL;
2152
2153	/* Try to be nice and wait for any queued DMA engine
2154	* transactions to process before smashing it with a rock
2155	*/
2156	dma_sync_wait(chan, cookie: qp->last_cookie);
2157	dmaengine_terminate_all(chan);
2158	dma_release_channel(chan);
2159	}
2160
2161	qp_bit = BIT_ULL(qp->qp_num);
2162
2163	ntb_db_set_mask(ntb: qp->ndev, db_bits: qp_bit);
2164	tasklet_kill(t: &qp->rxc_db_work);
2165
2166	cancel_delayed_work_sync(dwork: &qp->link_work);
2167
2168	qp->cb_data = NULL;
2169	qp->rx_handler = NULL;
2170	qp->tx_handler = NULL;
2171	qp->event_handler = NULL;
2172
2173	while ((entry = ntb_list_rm(lock: &qp->ntb_rx_q_lock, list: &qp->rx_free_q)))
2174	kfree(objp: entry);
2175
2176	while ((entry = ntb_list_rm(lock: &qp->ntb_rx_q_lock, list: &qp->rx_pend_q))) {
2177	dev_warn(&pdev->dev, "Freeing item from non-empty rx_pend_q\n");
2178	kfree(objp: entry);
2179	}
2180
2181	while ((entry = ntb_list_rm(lock: &qp->ntb_rx_q_lock, list: &qp->rx_post_q))) {
2182	dev_warn(&pdev->dev, "Freeing item from non-empty rx_post_q\n");
2183	kfree(objp: entry);
2184	}
2185
2186	while ((entry = ntb_list_rm(lock: &qp->ntb_tx_free_q_lock, list: &qp->tx_free_q)))
2187	kfree(objp: entry);
2188
2189	qp->transport->qp_bitmap_free |= qp_bit;
2190
2191	dev_info(&pdev->dev, "NTB Transport QP %d freed\n", qp->qp_num);
2192	}
2193	EXPORT_SYMBOL_GPL(ntb_transport_free_queue);
2194
2195	/**
2196	* ntb_transport_rx_remove - Dequeues enqueued rx packet
2197	* @qp: NTB queue to be freed
2198	* @len: pointer to variable to write enqueued buffers length
2199	*
2200	* Dequeues unused buffers from receive queue. Should only be used during
2201	* shutdown of qp.
2202	*
2203	* RETURNS: NULL error value on error, or void* for success.
2204	*/
2205	void *ntb_transport_rx_remove(struct ntb_transport_qp *qp, unsigned int *len)
2206	{
2207	struct ntb_queue_entry *entry;
2208	void *buf;
2209
2210	if (!qp || qp->client_ready)
2211	return NULL;
2212
2213	entry = ntb_list_rm(lock: &qp->ntb_rx_q_lock, list: &qp->rx_pend_q);
2214	if (!entry)
2215	return NULL;
2216
2217	buf = entry->cb_data;
2218	*len = entry->len;
2219
2220	ntb_list_add(lock: &qp->ntb_rx_q_lock, entry: &entry->entry, list: &qp->rx_free_q);
2221
2222	return buf;
2223	}
2224	EXPORT_SYMBOL_GPL(ntb_transport_rx_remove);
2225
2226	/**
2227	* ntb_transport_rx_enqueue - Enqueue a new NTB queue entry
2228	* @qp: NTB transport layer queue the entry is to be enqueued on
2229	* @cb: per buffer pointer for callback function to use
2230	* @data: pointer to data buffer that incoming packets will be copied into
2231	* @len: length of the data buffer
2232	*
2233	* Enqueue a new receive buffer onto the transport queue into which a NTB
2234	* payload can be received into.
2235	*
2236	* RETURNS: An appropriate -ERRNO error value on error, or zero for success.
2237	*/
2238	int ntb_transport_rx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
2239	unsigned int len)
2240	{
2241	struct ntb_queue_entry *entry;
2242
2243	if (!qp)
2244	return -EINVAL;
2245
2246	entry = ntb_list_rm(lock: &qp->ntb_rx_q_lock, list: &qp->rx_free_q);
2247	if (!entry)
2248	return -ENOMEM;
2249
2250	entry->cb_data = cb;
2251	entry->buf = data;
2252	entry->len = len;
2253	entry->flags = 0;
2254	entry->retries = 0;
2255	entry->errors = 0;
2256	entry->rx_index = 0;
2257
2258	ntb_list_add(lock: &qp->ntb_rx_q_lock, entry: &entry->entry, list: &qp->rx_pend_q);
2259
2260	if (qp->active)
2261	tasklet_schedule(t: &qp->rxc_db_work);
2262
2263	return 0;
2264	}
2265	EXPORT_SYMBOL_GPL(ntb_transport_rx_enqueue);
2266
2267	/**
2268	* ntb_transport_tx_enqueue - Enqueue a new NTB queue entry
2269	* @qp: NTB transport layer queue the entry is to be enqueued on
2270	* @cb: per buffer pointer for callback function to use
2271	* @data: pointer to data buffer that will be sent
2272	* @len: length of the data buffer
2273	*
2274	* Enqueue a new transmit buffer onto the transport queue from which a NTB
2275	* payload will be transmitted. This assumes that a lock is being held to
2276	* serialize access to the qp.
2277	*
2278	* RETURNS: An appropriate -ERRNO error value on error, or zero for success.
2279	*/
2280	int ntb_transport_tx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
2281	unsigned int len)
2282	{
2283	struct ntb_queue_entry *entry;
2284	int rc;
2285
2286	if (!qp || !len)
2287	return -EINVAL;
2288
2289	/* If the qp link is down already, just ignore. */
2290	if (!qp->link_is_up)
2291	return 0;
2292
2293	entry = ntb_list_rm(lock: &qp->ntb_tx_free_q_lock, list: &qp->tx_free_q);
2294	if (!entry) {
2295	qp->tx_err_no_buf++;
2296	return -EBUSY;
2297	}
2298
2299	entry->cb_data = cb;
2300	entry->buf = data;
2301	entry->len = len;
2302	entry->flags = 0;
2303	entry->errors = 0;
2304	entry->retries = 0;
2305	entry->tx_index = 0;
2306
2307	rc = ntb_process_tx(qp, entry);
2308	if (rc)
2309	ntb_list_add(lock: &qp->ntb_tx_free_q_lock, entry: &entry->entry,
2310	list: &qp->tx_free_q);
2311
2312	return rc;
2313	}
2314	EXPORT_SYMBOL_GPL(ntb_transport_tx_enqueue);
2315
2316	/**
2317	* ntb_transport_link_up - Notify NTB transport of client readiness to use queue
2318	* @qp: NTB transport layer queue to be enabled
2319	*
2320	* Notify NTB transport layer of client readiness to use queue
2321	*/
2322	void ntb_transport_link_up(struct ntb_transport_qp *qp)
2323	{
2324	if (!qp)
2325	return;
2326
2327	qp->client_ready = true;
2328
2329	if (qp->transport->link_is_up)
2330	schedule_delayed_work(dwork: &qp->link_work, delay: 0);
2331	}
2332	EXPORT_SYMBOL_GPL(ntb_transport_link_up);
2333
2334	/**
2335	* ntb_transport_link_down - Notify NTB transport to no longer enqueue data
2336	* @qp: NTB transport layer queue to be disabled
2337	*
2338	* Notify NTB transport layer of client's desire to no longer receive data on
2339	* transport queue specified. It is the client's responsibility to ensure all
2340	* entries on queue are purged or otherwise handled appropriately.
2341	*/
2342	void ntb_transport_link_down(struct ntb_transport_qp *qp)
2343	{
2344	int val;
2345
2346	if (!qp)
2347	return;
2348
2349	qp->client_ready = false;
2350
2351	val = ntb_spad_read(ntb: qp->ndev, sidx: QP_LINKS);
2352
2353	ntb_peer_spad_write(ntb: qp->ndev, PIDX, sidx: QP_LINKS, val: val & ~BIT(qp->qp_num));
2354
2355	if (qp->link_is_up)
2356	ntb_send_link_down(qp);
2357	else
2358	cancel_delayed_work_sync(dwork: &qp->link_work);
2359	}
2360	EXPORT_SYMBOL_GPL(ntb_transport_link_down);
2361
2362	/**
2363	* ntb_transport_link_query - Query transport link state
2364	* @qp: NTB transport layer queue to be queried
2365	*
2366	* Query connectivity to the remote system of the NTB transport queue
2367	*
2368	* RETURNS: true for link up or false for link down
2369	*/
2370	bool ntb_transport_link_query(struct ntb_transport_qp *qp)
2371	{
2372	if (!qp)
2373	return false;
2374
2375	return qp->link_is_up;
2376	}
2377	EXPORT_SYMBOL_GPL(ntb_transport_link_query);
2378
2379	/**
2380	* ntb_transport_qp_num - Query the qp number
2381	* @qp: NTB transport layer queue to be queried
2382	*
2383	* Query qp number of the NTB transport queue
2384	*
2385	* RETURNS: a zero based number specifying the qp number
2386	*/
2387	unsigned char ntb_transport_qp_num(struct ntb_transport_qp *qp)
2388	{
2389	if (!qp)
2390	return 0;
2391
2392	return qp->qp_num;
2393	}
2394	EXPORT_SYMBOL_GPL(ntb_transport_qp_num);
2395
2396	/**
2397	* ntb_transport_max_size - Query the max payload size of a qp
2398	* @qp: NTB transport layer queue to be queried
2399	*
2400	* Query the maximum payload size permissible on the given qp
2401	*
2402	* RETURNS: the max payload size of a qp
2403	*/
2404	unsigned int ntb_transport_max_size(struct ntb_transport_qp *qp)
2405	{
2406	unsigned int max_size;
2407	unsigned int copy_align;
2408	struct dma_chan *rx_chan, *tx_chan;
2409
2410	if (!qp)
2411	return 0;
2412
2413	rx_chan = qp->rx_dma_chan;
2414	tx_chan = qp->tx_dma_chan;
2415
2416	copy_align = max(rx_chan ? rx_chan->device->copy_align : 0,
2417	tx_chan ? tx_chan->device->copy_align : 0);
2418
2419	/* If DMA engine usage is possible, try to find the max size for that */
2420	max_size = qp->tx_max_frame - sizeof(struct ntb_payload_header);
2421	max_size = round_down(max_size, 1 << copy_align);
2422
2423	return max_size;
2424	}
2425	EXPORT_SYMBOL_GPL(ntb_transport_max_size);
2426
2427	unsigned int ntb_transport_tx_free_entry(struct ntb_transport_qp *qp)
2428	{
2429	unsigned int head = qp->tx_index;
2430	unsigned int tail = qp->remote_rx_info->entry;
2431
2432	return tail >= head ? tail - head : qp->tx_max_entry + tail - head;
2433	}
2434	EXPORT_SYMBOL_GPL(ntb_transport_tx_free_entry);
2435
2436	static void ntb_transport_doorbell_callback(void *data, int vector)
2437	{
2438	struct ntb_transport_ctx *nt = data;
2439	struct ntb_transport_qp *qp;
2440	u64 db_bits;
2441	unsigned int qp_num;
2442
2443	if (ntb_db_read(ntb: nt->ndev) & nt->msi_db_mask) {
2444	ntb_transport_msi_peer_desc_changed(nt);
2445	ntb_db_clear(ntb: nt->ndev, db_bits: nt->msi_db_mask);
2446	}
2447
2448	db_bits = (nt->qp_bitmap & ~nt->qp_bitmap_free &
2449	ntb_db_vector_mask(ntb: nt->ndev, vector));
2450
2451	while (db_bits) {
2452	qp_num = __ffs(db_bits);
2453	qp = &nt->qp_vec[qp_num];
2454
2455	if (qp->active)
2456	tasklet_schedule(t: &qp->rxc_db_work);
2457
2458	db_bits &= ~BIT_ULL(qp_num);
2459	}
2460	}
2461
2462	static const struct ntb_ctx_ops ntb_transport_ops = {
2463	.link_event = ntb_transport_event_callback,
2464	.db_event = ntb_transport_doorbell_callback,
2465	};
2466
2467	static struct ntb_client ntb_transport_client = {
2468	.ops = {
2469	.probe = ntb_transport_probe,
2470	.remove = ntb_transport_free,
2471	},
2472	};
2473
2474	static int __init ntb_transport_init(void)
2475	{
2476	int rc;
2477
2478	pr_info("%s, version %s\n", NTB_TRANSPORT_DESC, NTB_TRANSPORT_VER);
2479
2480	if (debugfs_initialized())
2481	nt_debugfs_dir = debugfs_create_dir(KBUILD_MODNAME, NULL);
2482
2483	rc = bus_register(bus: &ntb_transport_bus);
2484	if (rc)
2485	goto err_bus;
2486
2487	rc = ntb_register_client(&ntb_transport_client);
2488	if (rc)
2489	goto err_client;
2490
2491	return 0;
2492
2493	err_client:
2494	bus_unregister(bus: &ntb_transport_bus);
2495	err_bus:
2496	debugfs_remove_recursive(dentry: nt_debugfs_dir);
2497	return rc;
2498	}
2499	module_init(ntb_transport_init);
2500
2501	static void __exit ntb_transport_exit(void)
2502	{
2503	ntb_unregister_client(client: &ntb_transport_client);
2504	bus_unregister(bus: &ntb_transport_bus);
2505	debugfs_remove_recursive(dentry: nt_debugfs_dir);
2506	}
2507	module_exit(ntb_transport_exit);
2508