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.
7 * Copyright(c) 2012 Intel Corporation. All rights reserved.
8 * Copyright (C) 2015 EMC Corporation. All Rights Reserved.
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.
16 * Copyright(c) 2012 Intel Corporation. All rights reserved.
17 * Copyright (C) 2015 EMC Corporation. All Rights Reserved.
19 * Redistribution and use in source and binary forms, with or without
20 * modification, are permitted provided that the following conditions
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
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.
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.
45 * PCIe NTB Transport Linux driver
47 * Contact Information:
48 * Jon Mason <jon.mason@intel.com>
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"
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)
71 MODULE_DESCRIPTION(NTB_TRANSPORT_DESC
);
72 MODULE_VERSION(NTB_TRANSPORT_VER
);
73 MODULE_LICENSE("Dual BSD/GPL");
74 MODULE_AUTHOR("Intel Corporation");
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");
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");
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");
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");
93 module_param(use_dma
, bool, 0644);
94 MODULE_PARM_DESC(use_dma
, "Use DMA engine to perform large data copy");
96 static struct dentry
*nt_debugfs_dir
;
98 /* Only two-ports NTB devices are supported */
99 #define PIDX NTB_DEF_PEER_IDX
101 struct ntb_queue_entry
{
102 /* ntb_queue list reference */
103 struct list_head entry
;
104 /* pointers to data to be transferred */
111 unsigned int tx_index
;
112 unsigned int rx_index
;
114 struct ntb_transport_qp
*qp
;
116 struct ntb_payload_header __iomem
*tx_hdr
;
117 struct ntb_payload_header
*rx_hdr
;
125 struct ntb_transport_qp
{
126 struct ntb_transport_ctx
*transport
;
127 struct ntb_dev
*ndev
;
129 struct dma_chan
*tx_dma_chan
;
130 struct dma_chan
*rx_dma_chan
;
136 u8 qp_num
; /* Only 64 QP's are allowed. 0-63 */
139 struct ntb_rx_info __iomem
*rx_info
;
140 struct ntb_rx_info
*remote_rx_info
;
142 void (*tx_handler
)(struct ntb_transport_qp
*qp
, void *qp_data
,
143 void *data
, int len
);
144 struct list_head tx_free_q
;
145 spinlock_t ntb_tx_free_q_lock
;
147 dma_addr_t tx_mw_phys
;
148 unsigned int tx_index
;
149 unsigned int tx_max_entry
;
150 unsigned int tx_max_frame
;
152 void (*rx_handler
)(struct ntb_transport_qp
*qp
, void *qp_data
,
153 void *data
, int len
);
154 struct list_head rx_post_q
;
155 struct list_head rx_pend_q
;
156 struct list_head rx_free_q
;
157 /* ntb_rx_q_lock: synchronize access to rx_XXXX_q */
158 spinlock_t ntb_rx_q_lock
;
160 unsigned int rx_index
;
161 unsigned int rx_max_entry
;
162 unsigned int rx_max_frame
;
163 unsigned int rx_alloc_entry
;
164 dma_cookie_t last_cookie
;
165 struct tasklet_struct rxc_db_work
;
167 void (*event_handler
)(void *data
, int status
);
168 struct delayed_work link_work
;
169 struct work_struct link_cleanup
;
171 struct dentry
*debugfs_dir
;
172 struct dentry
*debugfs_stats
;
191 struct ntb_transport_mw
{
192 phys_addr_t phys_addr
;
193 resource_size_t phys_size
;
201 struct ntb_transport_client_dev
{
202 struct list_head entry
;
203 struct ntb_transport_ctx
*nt
;
207 struct ntb_transport_ctx
{
208 struct list_head entry
;
209 struct list_head client_devs
;
211 struct ntb_dev
*ndev
;
213 struct ntb_transport_mw
*mw_vec
;
214 struct ntb_transport_qp
*qp_vec
;
215 unsigned int mw_count
;
216 unsigned int qp_count
;
221 struct delayed_work link_work
;
222 struct work_struct link_cleanup
;
224 struct dentry
*debugfs_node_dir
;
228 DESC_DONE_FLAG
= BIT(0),
229 LINK_DOWN_FLAG
= BIT(1),
232 struct ntb_payload_header
{
247 #define dev_client_dev(__dev) \
248 container_of((__dev), struct ntb_transport_client_dev, dev)
250 #define drv_client(__drv) \
251 container_of((__drv), struct ntb_transport_client, driver)
253 #define QP_TO_MW(nt, qp) ((qp) % nt->mw_count)
254 #define NTB_QP_DEF_NUM_ENTRIES 100
255 #define NTB_LINK_DOWN_TIMEOUT 10
257 static void ntb_transport_rxc_db(unsigned long data
);
258 static const struct ntb_ctx_ops ntb_transport_ops
;
259 static struct ntb_client ntb_transport_client
;
260 static int ntb_async_tx_submit(struct ntb_transport_qp
*qp
,
261 struct ntb_queue_entry
*entry
);
262 static void ntb_memcpy_tx(struct ntb_queue_entry
*entry
, void __iomem
*offset
);
263 static int ntb_async_rx_submit(struct ntb_queue_entry
*entry
, void *offset
);
264 static void ntb_memcpy_rx(struct ntb_queue_entry
*entry
, void *offset
);
267 static int ntb_transport_bus_match(struct device
*dev
,
268 struct device_driver
*drv
)
270 return !strncmp(dev_name(dev
), drv
->name
, strlen(drv
->name
));
273 static int ntb_transport_bus_probe(struct device
*dev
)
275 const struct ntb_transport_client
*client
;
280 client
= drv_client(dev
->driver
);
281 rc
= client
->probe(dev
);
288 static int ntb_transport_bus_remove(struct device
*dev
)
290 const struct ntb_transport_client
*client
;
292 client
= drv_client(dev
->driver
);
300 static struct bus_type ntb_transport_bus
= {
301 .name
= "ntb_transport",
302 .match
= ntb_transport_bus_match
,
303 .probe
= ntb_transport_bus_probe
,
304 .remove
= ntb_transport_bus_remove
,
307 static LIST_HEAD(ntb_transport_list
);
309 static int ntb_bus_init(struct ntb_transport_ctx
*nt
)
311 list_add_tail(&nt
->entry
, &ntb_transport_list
);
315 static void ntb_bus_remove(struct ntb_transport_ctx
*nt
)
317 struct ntb_transport_client_dev
*client_dev
, *cd
;
319 list_for_each_entry_safe(client_dev
, cd
, &nt
->client_devs
, entry
) {
320 dev_err(client_dev
->dev
.parent
, "%s still attached to bus, removing\n",
321 dev_name(&client_dev
->dev
));
322 list_del(&client_dev
->entry
);
323 device_unregister(&client_dev
->dev
);
326 list_del(&nt
->entry
);
329 static void ntb_transport_client_release(struct device
*dev
)
331 struct ntb_transport_client_dev
*client_dev
;
333 client_dev
= dev_client_dev(dev
);
338 * ntb_transport_unregister_client_dev - Unregister NTB client device
339 * @device_name: Name of NTB client device
341 * Unregister an NTB client device with the NTB transport layer
343 void ntb_transport_unregister_client_dev(char *device_name
)
345 struct ntb_transport_client_dev
*client
, *cd
;
346 struct ntb_transport_ctx
*nt
;
348 list_for_each_entry(nt
, &ntb_transport_list
, entry
)
349 list_for_each_entry_safe(client
, cd
, &nt
->client_devs
, entry
)
350 if (!strncmp(dev_name(&client
->dev
), device_name
,
351 strlen(device_name
))) {
352 list_del(&client
->entry
);
353 device_unregister(&client
->dev
);
356 EXPORT_SYMBOL_GPL(ntb_transport_unregister_client_dev
);
359 * ntb_transport_register_client_dev - Register NTB client device
360 * @device_name: Name of NTB client device
362 * Register an NTB client device with the NTB transport layer
364 int ntb_transport_register_client_dev(char *device_name
)
366 struct ntb_transport_client_dev
*client_dev
;
367 struct ntb_transport_ctx
*nt
;
371 if (list_empty(&ntb_transport_list
))
374 list_for_each_entry(nt
, &ntb_transport_list
, entry
) {
377 node
= dev_to_node(&nt
->ndev
->dev
);
379 client_dev
= kzalloc_node(sizeof(*client_dev
),
386 dev
= &client_dev
->dev
;
388 /* setup and register client devices */
389 dev_set_name(dev
, "%s%d", device_name
, i
);
390 dev
->bus
= &ntb_transport_bus
;
391 dev
->release
= ntb_transport_client_release
;
392 dev
->parent
= &nt
->ndev
->dev
;
394 rc
= device_register(dev
);
400 list_add_tail(&client_dev
->entry
, &nt
->client_devs
);
407 ntb_transport_unregister_client_dev(device_name
);
411 EXPORT_SYMBOL_GPL(ntb_transport_register_client_dev
);
414 * ntb_transport_register_client - Register NTB client driver
415 * @drv: NTB client driver to be registered
417 * Register an NTB client driver with the NTB transport layer
419 * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
421 int ntb_transport_register_client(struct ntb_transport_client
*drv
)
423 drv
->driver
.bus
= &ntb_transport_bus
;
425 if (list_empty(&ntb_transport_list
))
428 return driver_register(&drv
->driver
);
430 EXPORT_SYMBOL_GPL(ntb_transport_register_client
);
433 * ntb_transport_unregister_client - Unregister NTB client driver
434 * @drv: NTB client driver to be unregistered
436 * Unregister an NTB client driver with the NTB transport layer
438 * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
440 void ntb_transport_unregister_client(struct ntb_transport_client
*drv
)
442 driver_unregister(&drv
->driver
);
444 EXPORT_SYMBOL_GPL(ntb_transport_unregister_client
);
446 static ssize_t
debugfs_read(struct file
*filp
, char __user
*ubuf
, size_t count
,
449 struct ntb_transport_qp
*qp
;
451 ssize_t ret
, out_offset
, out_count
;
453 qp
= filp
->private_data
;
455 if (!qp
|| !qp
->link_is_up
)
460 buf
= kmalloc(out_count
, GFP_KERNEL
);
465 out_offset
+= snprintf(buf
+ out_offset
, out_count
- out_offset
,
466 "\nNTB QP stats:\n\n");
467 out_offset
+= snprintf(buf
+ out_offset
, out_count
- out_offset
,
468 "rx_bytes - \t%llu\n", qp
->rx_bytes
);
469 out_offset
+= snprintf(buf
+ out_offset
, out_count
- out_offset
,
470 "rx_pkts - \t%llu\n", qp
->rx_pkts
);
471 out_offset
+= snprintf(buf
+ out_offset
, out_count
- out_offset
,
472 "rx_memcpy - \t%llu\n", qp
->rx_memcpy
);
473 out_offset
+= snprintf(buf
+ out_offset
, out_count
- out_offset
,
474 "rx_async - \t%llu\n", qp
->rx_async
);
475 out_offset
+= snprintf(buf
+ out_offset
, out_count
- out_offset
,
476 "rx_ring_empty - %llu\n", qp
->rx_ring_empty
);
477 out_offset
+= snprintf(buf
+ out_offset
, out_count
- out_offset
,
478 "rx_err_no_buf - %llu\n", qp
->rx_err_no_buf
);
479 out_offset
+= snprintf(buf
+ out_offset
, out_count
- out_offset
,
480 "rx_err_oflow - \t%llu\n", qp
->rx_err_oflow
);
481 out_offset
+= snprintf(buf
+ out_offset
, out_count
- out_offset
,
482 "rx_err_ver - \t%llu\n", qp
->rx_err_ver
);
483 out_offset
+= snprintf(buf
+ out_offset
, out_count
- out_offset
,
484 "rx_buff - \t0x%p\n", qp
->rx_buff
);
485 out_offset
+= snprintf(buf
+ out_offset
, out_count
- out_offset
,
486 "rx_index - \t%u\n", qp
->rx_index
);
487 out_offset
+= snprintf(buf
+ out_offset
, out_count
- out_offset
,
488 "rx_max_entry - \t%u\n", qp
->rx_max_entry
);
489 out_offset
+= snprintf(buf
+ out_offset
, out_count
- out_offset
,
490 "rx_alloc_entry - \t%u\n\n", qp
->rx_alloc_entry
);
492 out_offset
+= snprintf(buf
+ out_offset
, out_count
- out_offset
,
493 "tx_bytes - \t%llu\n", qp
->tx_bytes
);
494 out_offset
+= snprintf(buf
+ out_offset
, out_count
- out_offset
,
495 "tx_pkts - \t%llu\n", qp
->tx_pkts
);
496 out_offset
+= snprintf(buf
+ out_offset
, out_count
- out_offset
,
497 "tx_memcpy - \t%llu\n", qp
->tx_memcpy
);
498 out_offset
+= snprintf(buf
+ out_offset
, out_count
- out_offset
,
499 "tx_async - \t%llu\n", qp
->tx_async
);
500 out_offset
+= snprintf(buf
+ out_offset
, out_count
- out_offset
,
501 "tx_ring_full - \t%llu\n", qp
->tx_ring_full
);
502 out_offset
+= snprintf(buf
+ out_offset
, out_count
- out_offset
,
503 "tx_err_no_buf - %llu\n", qp
->tx_err_no_buf
);
504 out_offset
+= snprintf(buf
+ out_offset
, out_count
- out_offset
,
505 "tx_mw - \t0x%p\n", qp
->tx_mw
);
506 out_offset
+= snprintf(buf
+ out_offset
, out_count
- out_offset
,
507 "tx_index (H) - \t%u\n", qp
->tx_index
);
508 out_offset
+= snprintf(buf
+ out_offset
, out_count
- out_offset
,
510 qp
->remote_rx_info
->entry
);
511 out_offset
+= snprintf(buf
+ out_offset
, out_count
- out_offset
,
512 "tx_max_entry - \t%u\n", qp
->tx_max_entry
);
513 out_offset
+= snprintf(buf
+ out_offset
, out_count
- out_offset
,
515 ntb_transport_tx_free_entry(qp
));
517 out_offset
+= snprintf(buf
+ out_offset
, out_count
- out_offset
,
519 out_offset
+= snprintf(buf
+ out_offset
, out_count
- out_offset
,
520 "Using TX DMA - \t%s\n",
521 qp
->tx_dma_chan
? "Yes" : "No");
522 out_offset
+= snprintf(buf
+ out_offset
, out_count
- out_offset
,
523 "Using RX DMA - \t%s\n",
524 qp
->rx_dma_chan
? "Yes" : "No");
525 out_offset
+= snprintf(buf
+ out_offset
, out_count
- out_offset
,
527 qp
->link_is_up
? "Up" : "Down");
528 out_offset
+= snprintf(buf
+ out_offset
, out_count
- out_offset
,
531 if (out_offset
> out_count
)
532 out_offset
= out_count
;
534 ret
= simple_read_from_buffer(ubuf
, count
, offp
, buf
, out_offset
);
539 static const struct file_operations ntb_qp_debugfs_stats
= {
540 .owner
= THIS_MODULE
,
542 .read
= debugfs_read
,
545 static void ntb_list_add(spinlock_t
*lock
, struct list_head
*entry
,
546 struct list_head
*list
)
550 spin_lock_irqsave(lock
, flags
);
551 list_add_tail(entry
, list
);
552 spin_unlock_irqrestore(lock
, flags
);
555 static struct ntb_queue_entry
*ntb_list_rm(spinlock_t
*lock
,
556 struct list_head
*list
)
558 struct ntb_queue_entry
*entry
;
561 spin_lock_irqsave(lock
, flags
);
562 if (list_empty(list
)) {
566 entry
= list_first_entry(list
, struct ntb_queue_entry
, entry
);
567 list_del(&entry
->entry
);
570 spin_unlock_irqrestore(lock
, flags
);
575 static struct ntb_queue_entry
*ntb_list_mv(spinlock_t
*lock
,
576 struct list_head
*list
,
577 struct list_head
*to_list
)
579 struct ntb_queue_entry
*entry
;
582 spin_lock_irqsave(lock
, flags
);
584 if (list_empty(list
)) {
587 entry
= list_first_entry(list
, struct ntb_queue_entry
, entry
);
588 list_move_tail(&entry
->entry
, to_list
);
591 spin_unlock_irqrestore(lock
, flags
);
596 static int ntb_transport_setup_qp_mw(struct ntb_transport_ctx
*nt
,
599 struct ntb_transport_qp
*qp
= &nt
->qp_vec
[qp_num
];
600 struct ntb_transport_mw
*mw
;
601 struct ntb_dev
*ndev
= nt
->ndev
;
602 struct ntb_queue_entry
*entry
;
603 unsigned int rx_size
, num_qps_mw
;
604 unsigned int mw_num
, mw_count
, qp_count
;
608 mw_count
= nt
->mw_count
;
609 qp_count
= nt
->qp_count
;
611 mw_num
= QP_TO_MW(nt
, qp_num
);
612 mw
= &nt
->mw_vec
[mw_num
];
617 if (mw_num
< qp_count
% mw_count
)
618 num_qps_mw
= qp_count
/ mw_count
+ 1;
620 num_qps_mw
= qp_count
/ mw_count
;
622 rx_size
= (unsigned int)mw
->xlat_size
/ num_qps_mw
;
623 qp
->rx_buff
= mw
->virt_addr
+ rx_size
* (qp_num
/ mw_count
);
624 rx_size
-= sizeof(struct ntb_rx_info
);
626 qp
->remote_rx_info
= qp
->rx_buff
+ rx_size
;
628 /* Due to housekeeping, there must be atleast 2 buffs */
629 qp
->rx_max_frame
= min(transport_mtu
, rx_size
/ 2);
630 qp
->rx_max_entry
= rx_size
/ qp
->rx_max_frame
;
634 * Checking to see if we have more entries than the default.
635 * We should add additional entries if that is the case so we
636 * can be in sync with the transport frames.
638 node
= dev_to_node(&ndev
->dev
);
639 for (i
= qp
->rx_alloc_entry
; i
< qp
->rx_max_entry
; i
++) {
640 entry
= kzalloc_node(sizeof(*entry
), GFP_ATOMIC
, node
);
645 ntb_list_add(&qp
->ntb_rx_q_lock
, &entry
->entry
,
647 qp
->rx_alloc_entry
++;
650 qp
->remote_rx_info
->entry
= qp
->rx_max_entry
- 1;
652 /* setup the hdr offsets with 0's */
653 for (i
= 0; i
< qp
->rx_max_entry
; i
++) {
654 void *offset
= (qp
->rx_buff
+ qp
->rx_max_frame
* (i
+ 1) -
655 sizeof(struct ntb_payload_header
));
656 memset(offset
, 0, sizeof(struct ntb_payload_header
));
666 static void ntb_free_mw(struct ntb_transport_ctx
*nt
, int num_mw
)
668 struct ntb_transport_mw
*mw
= &nt
->mw_vec
[num_mw
];
669 struct pci_dev
*pdev
= nt
->ndev
->pdev
;
674 ntb_mw_clear_trans(nt
->ndev
, PIDX
, num_mw
);
675 dma_free_coherent(&pdev
->dev
, mw
->buff_size
,
676 mw
->virt_addr
, mw
->dma_addr
);
679 mw
->virt_addr
= NULL
;
682 static int ntb_set_mw(struct ntb_transport_ctx
*nt
, int num_mw
,
683 resource_size_t size
)
685 struct ntb_transport_mw
*mw
= &nt
->mw_vec
[num_mw
];
686 struct pci_dev
*pdev
= nt
->ndev
->pdev
;
687 size_t xlat_size
, buff_size
;
688 resource_size_t xlat_align
;
689 resource_size_t xlat_align_size
;
695 rc
= ntb_mw_get_align(nt
->ndev
, PIDX
, num_mw
, &xlat_align
,
696 &xlat_align_size
, NULL
);
700 xlat_size
= round_up(size
, xlat_align_size
);
701 buff_size
= round_up(size
, xlat_align
);
703 /* No need to re-setup */
704 if (mw
->xlat_size
== xlat_size
)
708 ntb_free_mw(nt
, num_mw
);
710 /* Alloc memory for receiving data. Must be aligned */
711 mw
->xlat_size
= xlat_size
;
712 mw
->buff_size
= buff_size
;
714 mw
->virt_addr
= dma_alloc_coherent(&pdev
->dev
, buff_size
,
715 &mw
->dma_addr
, GFP_KERNEL
);
716 if (!mw
->virt_addr
) {
719 dev_err(&pdev
->dev
, "Unable to alloc MW buff of size %zu\n",
725 * we must ensure that the memory address allocated is BAR size
726 * aligned in order for the XLAT register to take the value. This
727 * is a requirement of the hardware. It is recommended to setup CMA
728 * for BAR sizes equal or greater than 4MB.
730 if (!IS_ALIGNED(mw
->dma_addr
, xlat_align
)) {
731 dev_err(&pdev
->dev
, "DMA memory %pad is not aligned\n",
733 ntb_free_mw(nt
, num_mw
);
737 /* Notify HW the memory location of the receive buffer */
738 rc
= ntb_mw_set_trans(nt
->ndev
, PIDX
, num_mw
, mw
->dma_addr
,
741 dev_err(&pdev
->dev
, "Unable to set mw%d translation", num_mw
);
742 ntb_free_mw(nt
, num_mw
);
749 static void ntb_qp_link_down_reset(struct ntb_transport_qp
*qp
)
751 qp
->link_is_up
= false;
758 qp
->rx_ring_empty
= 0;
759 qp
->rx_err_no_buf
= 0;
760 qp
->rx_err_oflow
= 0;
766 qp
->tx_ring_full
= 0;
767 qp
->tx_err_no_buf
= 0;
772 static void ntb_qp_link_cleanup(struct ntb_transport_qp
*qp
)
774 struct ntb_transport_ctx
*nt
= qp
->transport
;
775 struct pci_dev
*pdev
= nt
->ndev
->pdev
;
777 dev_info(&pdev
->dev
, "qp %d: Link Cleanup\n", qp
->qp_num
);
779 cancel_delayed_work_sync(&qp
->link_work
);
780 ntb_qp_link_down_reset(qp
);
782 if (qp
->event_handler
)
783 qp
->event_handler(qp
->cb_data
, qp
->link_is_up
);
786 static void ntb_qp_link_cleanup_work(struct work_struct
*work
)
788 struct ntb_transport_qp
*qp
= container_of(work
,
789 struct ntb_transport_qp
,
791 struct ntb_transport_ctx
*nt
= qp
->transport
;
793 ntb_qp_link_cleanup(qp
);
796 schedule_delayed_work(&qp
->link_work
,
797 msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT
));
800 static void ntb_qp_link_down(struct ntb_transport_qp
*qp
)
802 schedule_work(&qp
->link_cleanup
);
805 static void ntb_transport_link_cleanup(struct ntb_transport_ctx
*nt
)
807 struct ntb_transport_qp
*qp
;
809 unsigned int i
, count
;
811 qp_bitmap_alloc
= nt
->qp_bitmap
& ~nt
->qp_bitmap_free
;
813 /* Pass along the info to any clients */
814 for (i
= 0; i
< nt
->qp_count
; i
++)
815 if (qp_bitmap_alloc
& BIT_ULL(i
)) {
817 ntb_qp_link_cleanup(qp
);
818 cancel_work_sync(&qp
->link_cleanup
);
819 cancel_delayed_work_sync(&qp
->link_work
);
823 cancel_delayed_work_sync(&nt
->link_work
);
825 /* The scratchpad registers keep the values if the remote side
826 * goes down, blast them now to give them a sane value the next
827 * time they are accessed
829 count
= ntb_spad_count(nt
->ndev
);
830 for (i
= 0; i
< count
; i
++)
831 ntb_spad_write(nt
->ndev
, i
, 0);
834 static void ntb_transport_link_cleanup_work(struct work_struct
*work
)
836 struct ntb_transport_ctx
*nt
=
837 container_of(work
, struct ntb_transport_ctx
, link_cleanup
);
839 ntb_transport_link_cleanup(nt
);
842 static void ntb_transport_event_callback(void *data
)
844 struct ntb_transport_ctx
*nt
= data
;
846 if (ntb_link_is_up(nt
->ndev
, NULL
, NULL
) == 1)
847 schedule_delayed_work(&nt
->link_work
, 0);
849 schedule_work(&nt
->link_cleanup
);
852 static void ntb_transport_link_work(struct work_struct
*work
)
854 struct ntb_transport_ctx
*nt
=
855 container_of(work
, struct ntb_transport_ctx
, link_work
.work
);
856 struct ntb_dev
*ndev
= nt
->ndev
;
857 struct pci_dev
*pdev
= ndev
->pdev
;
858 resource_size_t size
;
862 /* send the local info, in the opposite order of the way we read it */
863 for (i
= 0; i
< nt
->mw_count
; i
++) {
864 size
= nt
->mw_vec
[i
].phys_size
;
866 if (max_mw_size
&& size
> max_mw_size
)
869 spad
= MW0_SZ_HIGH
+ (i
* 2);
870 ntb_peer_spad_write(ndev
, PIDX
, spad
, upper_32_bits(size
));
872 spad
= MW0_SZ_LOW
+ (i
* 2);
873 ntb_peer_spad_write(ndev
, PIDX
, spad
, lower_32_bits(size
));
876 ntb_peer_spad_write(ndev
, PIDX
, NUM_MWS
, nt
->mw_count
);
878 ntb_peer_spad_write(ndev
, PIDX
, NUM_QPS
, nt
->qp_count
);
880 ntb_peer_spad_write(ndev
, PIDX
, VERSION
, NTB_TRANSPORT_VERSION
);
882 /* Query the remote side for its info */
883 val
= ntb_spad_read(ndev
, VERSION
);
884 dev_dbg(&pdev
->dev
, "Remote version = %d\n", val
);
885 if (val
!= NTB_TRANSPORT_VERSION
)
888 val
= ntb_spad_read(ndev
, NUM_QPS
);
889 dev_dbg(&pdev
->dev
, "Remote max number of qps = %d\n", val
);
890 if (val
!= nt
->qp_count
)
893 val
= ntb_spad_read(ndev
, NUM_MWS
);
894 dev_dbg(&pdev
->dev
, "Remote number of mws = %d\n", val
);
895 if (val
!= nt
->mw_count
)
898 for (i
= 0; i
< nt
->mw_count
; i
++) {
901 val
= ntb_spad_read(ndev
, MW0_SZ_HIGH
+ (i
* 2));
902 val64
= (u64
)val
<< 32;
904 val
= ntb_spad_read(ndev
, MW0_SZ_LOW
+ (i
* 2));
907 dev_dbg(&pdev
->dev
, "Remote MW%d size = %#llx\n", i
, val64
);
909 rc
= ntb_set_mw(nt
, i
, val64
);
914 nt
->link_is_up
= true;
916 for (i
= 0; i
< nt
->qp_count
; i
++) {
917 struct ntb_transport_qp
*qp
= &nt
->qp_vec
[i
];
919 ntb_transport_setup_qp_mw(nt
, i
);
921 if (qp
->client_ready
)
922 schedule_delayed_work(&qp
->link_work
, 0);
928 for (i
= 0; i
< nt
->mw_count
; i
++)
931 /* if there's an actual failure, we should just bail */
936 if (ntb_link_is_up(ndev
, NULL
, NULL
) == 1)
937 schedule_delayed_work(&nt
->link_work
,
938 msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT
));
941 static void ntb_qp_link_work(struct work_struct
*work
)
943 struct ntb_transport_qp
*qp
= container_of(work
,
944 struct ntb_transport_qp
,
946 struct pci_dev
*pdev
= qp
->ndev
->pdev
;
947 struct ntb_transport_ctx
*nt
= qp
->transport
;
950 WARN_ON(!nt
->link_is_up
);
952 val
= ntb_spad_read(nt
->ndev
, QP_LINKS
);
954 ntb_peer_spad_write(nt
->ndev
, PIDX
, QP_LINKS
, val
| BIT(qp
->qp_num
));
956 /* query remote spad for qp ready bits */
957 dev_dbg_ratelimited(&pdev
->dev
, "Remote QP link status = %x\n", val
);
959 /* See if the remote side is up */
960 if (val
& BIT(qp
->qp_num
)) {
961 dev_info(&pdev
->dev
, "qp %d: Link Up\n", qp
->qp_num
);
962 qp
->link_is_up
= true;
965 if (qp
->event_handler
)
966 qp
->event_handler(qp
->cb_data
, qp
->link_is_up
);
969 tasklet_schedule(&qp
->rxc_db_work
);
970 } else if (nt
->link_is_up
)
971 schedule_delayed_work(&qp
->link_work
,
972 msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT
));
975 static int ntb_transport_init_queue(struct ntb_transport_ctx
*nt
,
978 struct ntb_transport_qp
*qp
;
980 resource_size_t mw_size
;
981 unsigned int num_qps_mw
, tx_size
;
982 unsigned int mw_num
, mw_count
, qp_count
;
985 mw_count
= nt
->mw_count
;
986 qp_count
= nt
->qp_count
;
988 mw_num
= QP_TO_MW(nt
, qp_num
);
990 qp
= &nt
->qp_vec
[qp_num
];
994 qp
->client_ready
= false;
995 qp
->event_handler
= NULL
;
996 ntb_qp_link_down_reset(qp
);
998 if (mw_num
< qp_count
% mw_count
)
999 num_qps_mw
= qp_count
/ mw_count
+ 1;
1001 num_qps_mw
= qp_count
/ mw_count
;
1003 mw_base
= nt
->mw_vec
[mw_num
].phys_addr
;
1004 mw_size
= nt
->mw_vec
[mw_num
].phys_size
;
1006 tx_size
= (unsigned int)mw_size
/ num_qps_mw
;
1007 qp_offset
= tx_size
* (qp_num
/ mw_count
);
1009 qp
->tx_mw
= nt
->mw_vec
[mw_num
].vbase
+ qp_offset
;
1013 qp
->tx_mw_phys
= mw_base
+ qp_offset
;
1014 if (!qp
->tx_mw_phys
)
1017 tx_size
-= sizeof(struct ntb_rx_info
);
1018 qp
->rx_info
= qp
->tx_mw
+ tx_size
;
1020 /* Due to housekeeping, there must be atleast 2 buffs */
1021 qp
->tx_max_frame
= min(transport_mtu
, tx_size
/ 2);
1022 qp
->tx_max_entry
= tx_size
/ qp
->tx_max_frame
;
1024 if (nt
->debugfs_node_dir
) {
1025 char debugfs_name
[4];
1027 snprintf(debugfs_name
, 4, "qp%d", qp_num
);
1028 qp
->debugfs_dir
= debugfs_create_dir(debugfs_name
,
1029 nt
->debugfs_node_dir
);
1031 qp
->debugfs_stats
= debugfs_create_file("stats", S_IRUSR
,
1032 qp
->debugfs_dir
, qp
,
1033 &ntb_qp_debugfs_stats
);
1035 qp
->debugfs_dir
= NULL
;
1036 qp
->debugfs_stats
= NULL
;
1039 INIT_DELAYED_WORK(&qp
->link_work
, ntb_qp_link_work
);
1040 INIT_WORK(&qp
->link_cleanup
, ntb_qp_link_cleanup_work
);
1042 spin_lock_init(&qp
->ntb_rx_q_lock
);
1043 spin_lock_init(&qp
->ntb_tx_free_q_lock
);
1045 INIT_LIST_HEAD(&qp
->rx_post_q
);
1046 INIT_LIST_HEAD(&qp
->rx_pend_q
);
1047 INIT_LIST_HEAD(&qp
->rx_free_q
);
1048 INIT_LIST_HEAD(&qp
->tx_free_q
);
1050 tasklet_init(&qp
->rxc_db_work
, ntb_transport_rxc_db
,
1056 static int ntb_transport_probe(struct ntb_client
*self
, struct ntb_dev
*ndev
)
1058 struct ntb_transport_ctx
*nt
;
1059 struct ntb_transport_mw
*mw
;
1060 unsigned int mw_count
, qp_count
, spad_count
, max_mw_count_for_spads
;
1065 mw_count
= ntb_peer_mw_count(ndev
);
1067 if (!ndev
->ops
->mw_set_trans
) {
1068 dev_err(&ndev
->dev
, "Inbound MW based NTB API is required\n");
1072 if (ntb_db_is_unsafe(ndev
))
1074 "doorbell is unsafe, proceed anyway...\n");
1075 if (ntb_spad_is_unsafe(ndev
))
1077 "scratchpad is unsafe, proceed anyway...\n");
1079 if (ntb_peer_port_count(ndev
) != NTB_DEF_PEER_CNT
)
1080 dev_warn(&ndev
->dev
, "Multi-port NTB devices unsupported\n");
1082 node
= dev_to_node(&ndev
->dev
);
1084 nt
= kzalloc_node(sizeof(*nt
), GFP_KERNEL
, node
);
1089 spad_count
= ntb_spad_count(ndev
);
1091 /* Limit the MW's based on the availability of scratchpads */
1093 if (spad_count
< NTB_TRANSPORT_MIN_SPADS
) {
1099 max_mw_count_for_spads
= (spad_count
- MW0_SZ_HIGH
) / 2;
1100 nt
->mw_count
= min(mw_count
, max_mw_count_for_spads
);
1102 nt
->mw_vec
= kzalloc_node(mw_count
* sizeof(*nt
->mw_vec
),
1109 for (i
= 0; i
< mw_count
; i
++) {
1110 mw
= &nt
->mw_vec
[i
];
1112 rc
= ntb_peer_mw_get_addr(ndev
, i
, &mw
->phys_addr
,
1117 mw
->vbase
= ioremap_wc(mw
->phys_addr
, mw
->phys_size
);
1125 mw
->virt_addr
= NULL
;
1129 qp_bitmap
= ntb_db_valid_mask(ndev
);
1131 qp_count
= ilog2(qp_bitmap
);
1132 if (max_num_clients
&& max_num_clients
< qp_count
)
1133 qp_count
= max_num_clients
;
1134 else if (nt
->mw_count
< qp_count
)
1135 qp_count
= nt
->mw_count
;
1137 qp_bitmap
&= BIT_ULL(qp_count
) - 1;
1139 nt
->qp_count
= qp_count
;
1140 nt
->qp_bitmap
= qp_bitmap
;
1141 nt
->qp_bitmap_free
= qp_bitmap
;
1143 nt
->qp_vec
= kzalloc_node(qp_count
* sizeof(*nt
->qp_vec
),
1150 if (nt_debugfs_dir
) {
1151 nt
->debugfs_node_dir
=
1152 debugfs_create_dir(pci_name(ndev
->pdev
),
1156 for (i
= 0; i
< qp_count
; i
++) {
1157 rc
= ntb_transport_init_queue(nt
, i
);
1162 INIT_DELAYED_WORK(&nt
->link_work
, ntb_transport_link_work
);
1163 INIT_WORK(&nt
->link_cleanup
, ntb_transport_link_cleanup_work
);
1165 rc
= ntb_set_ctx(ndev
, nt
, &ntb_transport_ops
);
1169 INIT_LIST_HEAD(&nt
->client_devs
);
1170 rc
= ntb_bus_init(nt
);
1174 nt
->link_is_up
= false;
1175 ntb_link_enable(ndev
, NTB_SPEED_AUTO
, NTB_WIDTH_AUTO
);
1176 ntb_link_event(ndev
);
1181 ntb_clear_ctx(ndev
);
1186 mw
= &nt
->mw_vec
[i
];
1195 static void ntb_transport_free(struct ntb_client
*self
, struct ntb_dev
*ndev
)
1197 struct ntb_transport_ctx
*nt
= ndev
->ctx
;
1198 struct ntb_transport_qp
*qp
;
1199 u64 qp_bitmap_alloc
;
1202 ntb_transport_link_cleanup(nt
);
1203 cancel_work_sync(&nt
->link_cleanup
);
1204 cancel_delayed_work_sync(&nt
->link_work
);
1206 qp_bitmap_alloc
= nt
->qp_bitmap
& ~nt
->qp_bitmap_free
;
1208 /* verify that all the qp's are freed */
1209 for (i
= 0; i
< nt
->qp_count
; i
++) {
1210 qp
= &nt
->qp_vec
[i
];
1211 if (qp_bitmap_alloc
& BIT_ULL(i
))
1212 ntb_transport_free_queue(qp
);
1213 debugfs_remove_recursive(qp
->debugfs_dir
);
1216 ntb_link_disable(ndev
);
1217 ntb_clear_ctx(ndev
);
1221 for (i
= nt
->mw_count
; i
--; ) {
1223 iounmap(nt
->mw_vec
[i
].vbase
);
1231 static void ntb_complete_rxc(struct ntb_transport_qp
*qp
)
1233 struct ntb_queue_entry
*entry
;
1236 unsigned long irqflags
;
1238 spin_lock_irqsave(&qp
->ntb_rx_q_lock
, irqflags
);
1240 while (!list_empty(&qp
->rx_post_q
)) {
1241 entry
= list_first_entry(&qp
->rx_post_q
,
1242 struct ntb_queue_entry
, entry
);
1243 if (!(entry
->flags
& DESC_DONE_FLAG
))
1246 entry
->rx_hdr
->flags
= 0;
1247 iowrite32(entry
->rx_index
, &qp
->rx_info
->entry
);
1249 cb_data
= entry
->cb_data
;
1252 list_move_tail(&entry
->entry
, &qp
->rx_free_q
);
1254 spin_unlock_irqrestore(&qp
->ntb_rx_q_lock
, irqflags
);
1256 if (qp
->rx_handler
&& qp
->client_ready
)
1257 qp
->rx_handler(qp
, qp
->cb_data
, cb_data
, len
);
1259 spin_lock_irqsave(&qp
->ntb_rx_q_lock
, irqflags
);
1262 spin_unlock_irqrestore(&qp
->ntb_rx_q_lock
, irqflags
);
1265 static void ntb_rx_copy_callback(void *data
,
1266 const struct dmaengine_result
*res
)
1268 struct ntb_queue_entry
*entry
= data
;
1270 /* we need to check DMA results if we are using DMA */
1272 enum dmaengine_tx_result dma_err
= res
->result
;
1275 case DMA_TRANS_READ_FAILED
:
1276 case DMA_TRANS_WRITE_FAILED
:
1278 case DMA_TRANS_ABORTED
:
1280 struct ntb_transport_qp
*qp
= entry
->qp
;
1281 void *offset
= qp
->rx_buff
+ qp
->rx_max_frame
*
1284 ntb_memcpy_rx(entry
, offset
);
1289 case DMA_TRANS_NOERROR
:
1295 entry
->flags
|= DESC_DONE_FLAG
;
1297 ntb_complete_rxc(entry
->qp
);
1300 static void ntb_memcpy_rx(struct ntb_queue_entry
*entry
, void *offset
)
1302 void *buf
= entry
->buf
;
1303 size_t len
= entry
->len
;
1305 memcpy(buf
, offset
, len
);
1307 /* Ensure that the data is fully copied out before clearing the flag */
1310 ntb_rx_copy_callback(entry
, NULL
);
1313 static int ntb_async_rx_submit(struct ntb_queue_entry
*entry
, void *offset
)
1315 struct dma_async_tx_descriptor
*txd
;
1316 struct ntb_transport_qp
*qp
= entry
->qp
;
1317 struct dma_chan
*chan
= qp
->rx_dma_chan
;
1318 struct dma_device
*device
;
1319 size_t pay_off
, buff_off
, len
;
1320 struct dmaengine_unmap_data
*unmap
;
1321 dma_cookie_t cookie
;
1322 void *buf
= entry
->buf
;
1325 device
= chan
->device
;
1326 pay_off
= (size_t)offset
& ~PAGE_MASK
;
1327 buff_off
= (size_t)buf
& ~PAGE_MASK
;
1329 if (!is_dma_copy_aligned(device
, pay_off
, buff_off
, len
))
1332 unmap
= dmaengine_get_unmap_data(device
->dev
, 2, GFP_NOWAIT
);
1337 unmap
->addr
[0] = dma_map_page(device
->dev
, virt_to_page(offset
),
1338 pay_off
, len
, DMA_TO_DEVICE
);
1339 if (dma_mapping_error(device
->dev
, unmap
->addr
[0]))
1344 unmap
->addr
[1] = dma_map_page(device
->dev
, virt_to_page(buf
),
1345 buff_off
, len
, DMA_FROM_DEVICE
);
1346 if (dma_mapping_error(device
->dev
, unmap
->addr
[1]))
1349 unmap
->from_cnt
= 1;
1351 txd
= device
->device_prep_dma_memcpy(chan
, unmap
->addr
[1],
1352 unmap
->addr
[0], len
,
1353 DMA_PREP_INTERRUPT
);
1357 txd
->callback_result
= ntb_rx_copy_callback
;
1358 txd
->callback_param
= entry
;
1359 dma_set_unmap(txd
, unmap
);
1361 cookie
= dmaengine_submit(txd
);
1362 if (dma_submit_error(cookie
))
1365 dmaengine_unmap_put(unmap
);
1367 qp
->last_cookie
= cookie
;
1374 dmaengine_unmap_put(unmap
);
1376 dmaengine_unmap_put(unmap
);
1381 static void ntb_async_rx(struct ntb_queue_entry
*entry
, void *offset
)
1383 struct ntb_transport_qp
*qp
= entry
->qp
;
1384 struct dma_chan
*chan
= qp
->rx_dma_chan
;
1390 if (entry
->len
< copy_bytes
)
1393 res
= ntb_async_rx_submit(entry
, offset
);
1397 if (!entry
->retries
)
1403 ntb_memcpy_rx(entry
, offset
);
1407 static int ntb_process_rxc(struct ntb_transport_qp
*qp
)
1409 struct ntb_payload_header
*hdr
;
1410 struct ntb_queue_entry
*entry
;
1413 offset
= qp
->rx_buff
+ qp
->rx_max_frame
* qp
->rx_index
;
1414 hdr
= offset
+ qp
->rx_max_frame
- sizeof(struct ntb_payload_header
);
1416 dev_dbg(&qp
->ndev
->pdev
->dev
, "qp %d: RX ver %u len %d flags %x\n",
1417 qp
->qp_num
, hdr
->ver
, hdr
->len
, hdr
->flags
);
1419 if (!(hdr
->flags
& DESC_DONE_FLAG
)) {
1420 dev_dbg(&qp
->ndev
->pdev
->dev
, "done flag not set\n");
1421 qp
->rx_ring_empty
++;
1425 if (hdr
->flags
& LINK_DOWN_FLAG
) {
1426 dev_dbg(&qp
->ndev
->pdev
->dev
, "link down flag set\n");
1427 ntb_qp_link_down(qp
);
1432 if (hdr
->ver
!= (u32
)qp
->rx_pkts
) {
1433 dev_dbg(&qp
->ndev
->pdev
->dev
,
1434 "version mismatch, expected %llu - got %u\n",
1435 qp
->rx_pkts
, hdr
->ver
);
1440 entry
= ntb_list_mv(&qp
->ntb_rx_q_lock
, &qp
->rx_pend_q
, &qp
->rx_post_q
);
1442 dev_dbg(&qp
->ndev
->pdev
->dev
, "no receive buffer\n");
1443 qp
->rx_err_no_buf
++;
1447 entry
->rx_hdr
= hdr
;
1448 entry
->rx_index
= qp
->rx_index
;
1450 if (hdr
->len
> entry
->len
) {
1451 dev_dbg(&qp
->ndev
->pdev
->dev
,
1452 "receive buffer overflow! Wanted %d got %d\n",
1453 hdr
->len
, entry
->len
);
1457 entry
->flags
|= DESC_DONE_FLAG
;
1459 ntb_complete_rxc(qp
);
1461 dev_dbg(&qp
->ndev
->pdev
->dev
,
1462 "RX OK index %u ver %u size %d into buf size %d\n",
1463 qp
->rx_index
, hdr
->ver
, hdr
->len
, entry
->len
);
1465 qp
->rx_bytes
+= hdr
->len
;
1468 entry
->len
= hdr
->len
;
1470 ntb_async_rx(entry
, offset
);
1474 qp
->rx_index
%= qp
->rx_max_entry
;
1479 static void ntb_transport_rxc_db(unsigned long data
)
1481 struct ntb_transport_qp
*qp
= (void *)data
;
1484 dev_dbg(&qp
->ndev
->pdev
->dev
, "%s: doorbell %d received\n",
1485 __func__
, qp
->qp_num
);
1487 /* Limit the number of packets processed in a single interrupt to
1488 * provide fairness to others
1490 for (i
= 0; i
< qp
->rx_max_entry
; i
++) {
1491 rc
= ntb_process_rxc(qp
);
1496 if (i
&& qp
->rx_dma_chan
)
1497 dma_async_issue_pending(qp
->rx_dma_chan
);
1499 if (i
== qp
->rx_max_entry
) {
1500 /* there is more work to do */
1502 tasklet_schedule(&qp
->rxc_db_work
);
1503 } else if (ntb_db_read(qp
->ndev
) & BIT_ULL(qp
->qp_num
)) {
1504 /* the doorbell bit is set: clear it */
1505 ntb_db_clear(qp
->ndev
, BIT_ULL(qp
->qp_num
));
1506 /* ntb_db_read ensures ntb_db_clear write is committed */
1507 ntb_db_read(qp
->ndev
);
1509 /* an interrupt may have arrived between finishing
1510 * ntb_process_rxc and clearing the doorbell bit:
1511 * there might be some more work to do.
1514 tasklet_schedule(&qp
->rxc_db_work
);
1518 static void ntb_tx_copy_callback(void *data
,
1519 const struct dmaengine_result
*res
)
1521 struct ntb_queue_entry
*entry
= data
;
1522 struct ntb_transport_qp
*qp
= entry
->qp
;
1523 struct ntb_payload_header __iomem
*hdr
= entry
->tx_hdr
;
1525 /* we need to check DMA results if we are using DMA */
1527 enum dmaengine_tx_result dma_err
= res
->result
;
1530 case DMA_TRANS_READ_FAILED
:
1531 case DMA_TRANS_WRITE_FAILED
:
1533 case DMA_TRANS_ABORTED
:
1535 void __iomem
*offset
=
1536 qp
->tx_mw
+ qp
->tx_max_frame
*
1539 /* resubmit via CPU */
1540 ntb_memcpy_tx(entry
, offset
);
1545 case DMA_TRANS_NOERROR
:
1551 iowrite32(entry
->flags
| DESC_DONE_FLAG
, &hdr
->flags
);
1553 ntb_peer_db_set(qp
->ndev
, BIT_ULL(qp
->qp_num
));
1555 /* The entry length can only be zero if the packet is intended to be a
1556 * "link down" or similar. Since no payload is being sent in these
1557 * cases, there is nothing to add to the completion queue.
1559 if (entry
->len
> 0) {
1560 qp
->tx_bytes
+= entry
->len
;
1563 qp
->tx_handler(qp
, qp
->cb_data
, entry
->cb_data
,
1567 ntb_list_add(&qp
->ntb_tx_free_q_lock
, &entry
->entry
, &qp
->tx_free_q
);
1570 static void ntb_memcpy_tx(struct ntb_queue_entry
*entry
, void __iomem
*offset
)
1572 #ifdef ARCH_HAS_NOCACHE_UACCESS
1574 * Using non-temporal mov to improve performance on non-cached
1575 * writes, even though we aren't actually copying from user space.
1577 __copy_from_user_inatomic_nocache(offset
, entry
->buf
, entry
->len
);
1579 memcpy_toio(offset
, entry
->buf
, entry
->len
);
1582 /* Ensure that the data is fully copied out before setting the flags */
1585 ntb_tx_copy_callback(entry
, NULL
);
1588 static int ntb_async_tx_submit(struct ntb_transport_qp
*qp
,
1589 struct ntb_queue_entry
*entry
)
1591 struct dma_async_tx_descriptor
*txd
;
1592 struct dma_chan
*chan
= qp
->tx_dma_chan
;
1593 struct dma_device
*device
;
1594 size_t len
= entry
->len
;
1595 void *buf
= entry
->buf
;
1596 size_t dest_off
, buff_off
;
1597 struct dmaengine_unmap_data
*unmap
;
1599 dma_cookie_t cookie
;
1601 device
= chan
->device
;
1602 dest
= qp
->tx_mw_phys
+ qp
->tx_max_frame
* entry
->tx_index
;
1603 buff_off
= (size_t)buf
& ~PAGE_MASK
;
1604 dest_off
= (size_t)dest
& ~PAGE_MASK
;
1606 if (!is_dma_copy_aligned(device
, buff_off
, dest_off
, len
))
1609 unmap
= dmaengine_get_unmap_data(device
->dev
, 1, GFP_NOWAIT
);
1614 unmap
->addr
[0] = dma_map_page(device
->dev
, virt_to_page(buf
),
1615 buff_off
, len
, DMA_TO_DEVICE
);
1616 if (dma_mapping_error(device
->dev
, unmap
->addr
[0]))
1621 txd
= device
->device_prep_dma_memcpy(chan
, dest
, unmap
->addr
[0], len
,
1622 DMA_PREP_INTERRUPT
);
1626 txd
->callback_result
= ntb_tx_copy_callback
;
1627 txd
->callback_param
= entry
;
1628 dma_set_unmap(txd
, unmap
);
1630 cookie
= dmaengine_submit(txd
);
1631 if (dma_submit_error(cookie
))
1634 dmaengine_unmap_put(unmap
);
1636 dma_async_issue_pending(chan
);
1640 dmaengine_unmap_put(unmap
);
1642 dmaengine_unmap_put(unmap
);
1647 static void ntb_async_tx(struct ntb_transport_qp
*qp
,
1648 struct ntb_queue_entry
*entry
)
1650 struct ntb_payload_header __iomem
*hdr
;
1651 struct dma_chan
*chan
= qp
->tx_dma_chan
;
1652 void __iomem
*offset
;
1655 entry
->tx_index
= qp
->tx_index
;
1656 offset
= qp
->tx_mw
+ qp
->tx_max_frame
* entry
->tx_index
;
1657 hdr
= offset
+ qp
->tx_max_frame
- sizeof(struct ntb_payload_header
);
1658 entry
->tx_hdr
= hdr
;
1660 iowrite32(entry
->len
, &hdr
->len
);
1661 iowrite32((u32
)qp
->tx_pkts
, &hdr
->ver
);
1666 if (entry
->len
< copy_bytes
)
1669 res
= ntb_async_tx_submit(qp
, entry
);
1673 if (!entry
->retries
)
1679 ntb_memcpy_tx(entry
, offset
);
1683 static int ntb_process_tx(struct ntb_transport_qp
*qp
,
1684 struct ntb_queue_entry
*entry
)
1686 if (qp
->tx_index
== qp
->remote_rx_info
->entry
) {
1691 if (entry
->len
> qp
->tx_max_frame
- sizeof(struct ntb_payload_header
)) {
1693 qp
->tx_handler(qp
, qp
->cb_data
, NULL
, -EIO
);
1695 ntb_list_add(&qp
->ntb_tx_free_q_lock
, &entry
->entry
,
1700 ntb_async_tx(qp
, entry
);
1703 qp
->tx_index
%= qp
->tx_max_entry
;
1710 static void ntb_send_link_down(struct ntb_transport_qp
*qp
)
1712 struct pci_dev
*pdev
= qp
->ndev
->pdev
;
1713 struct ntb_queue_entry
*entry
;
1716 if (!qp
->link_is_up
)
1719 dev_info(&pdev
->dev
, "qp %d: Send Link Down\n", qp
->qp_num
);
1721 for (i
= 0; i
< NTB_LINK_DOWN_TIMEOUT
; i
++) {
1722 entry
= ntb_list_rm(&qp
->ntb_tx_free_q_lock
, &qp
->tx_free_q
);
1731 entry
->cb_data
= NULL
;
1734 entry
->flags
= LINK_DOWN_FLAG
;
1736 rc
= ntb_process_tx(qp
, entry
);
1738 dev_err(&pdev
->dev
, "ntb: QP%d unable to send linkdown msg\n",
1741 ntb_qp_link_down_reset(qp
);
1744 static bool ntb_dma_filter_fn(struct dma_chan
*chan
, void *node
)
1746 return dev_to_node(&chan
->dev
->device
) == (int)(unsigned long)node
;
1750 * ntb_transport_create_queue - Create a new NTB transport layer queue
1751 * @rx_handler: receive callback function
1752 * @tx_handler: transmit callback function
1753 * @event_handler: event callback function
1755 * Create a new NTB transport layer queue and provide the queue with a callback
1756 * routine for both transmit and receive. The receive callback routine will be
1757 * used to pass up data when the transport has received it on the queue. The
1758 * transmit callback routine will be called when the transport has completed the
1759 * transmission of the data on the queue and the data is ready to be freed.
1761 * RETURNS: pointer to newly created ntb_queue, NULL on error.
1763 struct ntb_transport_qp
*
1764 ntb_transport_create_queue(void *data
, struct device
*client_dev
,
1765 const struct ntb_queue_handlers
*handlers
)
1767 struct ntb_dev
*ndev
;
1768 struct pci_dev
*pdev
;
1769 struct ntb_transport_ctx
*nt
;
1770 struct ntb_queue_entry
*entry
;
1771 struct ntb_transport_qp
*qp
;
1773 unsigned int free_queue
;
1774 dma_cap_mask_t dma_mask
;
1778 ndev
= dev_ntb(client_dev
->parent
);
1782 node
= dev_to_node(&ndev
->dev
);
1784 free_queue
= ffs(nt
->qp_bitmap_free
);
1788 /* decrement free_queue to make it zero based */
1791 qp
= &nt
->qp_vec
[free_queue
];
1792 qp_bit
= BIT_ULL(qp
->qp_num
);
1794 nt
->qp_bitmap_free
&= ~qp_bit
;
1797 qp
->rx_handler
= handlers
->rx_handler
;
1798 qp
->tx_handler
= handlers
->tx_handler
;
1799 qp
->event_handler
= handlers
->event_handler
;
1801 dma_cap_zero(dma_mask
);
1802 dma_cap_set(DMA_MEMCPY
, dma_mask
);
1806 dma_request_channel(dma_mask
, ntb_dma_filter_fn
,
1807 (void *)(unsigned long)node
);
1808 if (!qp
->tx_dma_chan
)
1809 dev_info(&pdev
->dev
, "Unable to allocate TX DMA channel\n");
1812 dma_request_channel(dma_mask
, ntb_dma_filter_fn
,
1813 (void *)(unsigned long)node
);
1814 if (!qp
->rx_dma_chan
)
1815 dev_info(&pdev
->dev
, "Unable to allocate RX DMA channel\n");
1817 qp
->tx_dma_chan
= NULL
;
1818 qp
->rx_dma_chan
= NULL
;
1821 dev_dbg(&pdev
->dev
, "Using %s memcpy for TX\n",
1822 qp
->tx_dma_chan
? "DMA" : "CPU");
1824 dev_dbg(&pdev
->dev
, "Using %s memcpy for RX\n",
1825 qp
->rx_dma_chan
? "DMA" : "CPU");
1827 for (i
= 0; i
< NTB_QP_DEF_NUM_ENTRIES
; i
++) {
1828 entry
= kzalloc_node(sizeof(*entry
), GFP_ATOMIC
, node
);
1833 ntb_list_add(&qp
->ntb_rx_q_lock
, &entry
->entry
,
1836 qp
->rx_alloc_entry
= NTB_QP_DEF_NUM_ENTRIES
;
1838 for (i
= 0; i
< qp
->tx_max_entry
; i
++) {
1839 entry
= kzalloc_node(sizeof(*entry
), GFP_ATOMIC
, node
);
1844 ntb_list_add(&qp
->ntb_tx_free_q_lock
, &entry
->entry
,
1848 ntb_db_clear(qp
->ndev
, qp_bit
);
1849 ntb_db_clear_mask(qp
->ndev
, qp_bit
);
1851 dev_info(&pdev
->dev
, "NTB Transport QP %d created\n", qp
->qp_num
);
1856 while ((entry
= ntb_list_rm(&qp
->ntb_tx_free_q_lock
, &qp
->tx_free_q
)))
1859 qp
->rx_alloc_entry
= 0;
1860 while ((entry
= ntb_list_rm(&qp
->ntb_rx_q_lock
, &qp
->rx_free_q
)))
1862 if (qp
->tx_dma_chan
)
1863 dma_release_channel(qp
->tx_dma_chan
);
1864 if (qp
->rx_dma_chan
)
1865 dma_release_channel(qp
->rx_dma_chan
);
1866 nt
->qp_bitmap_free
|= qp_bit
;
1870 EXPORT_SYMBOL_GPL(ntb_transport_create_queue
);
1873 * ntb_transport_free_queue - Frees NTB transport queue
1874 * @qp: NTB queue to be freed
1876 * Frees NTB transport queue
1878 void ntb_transport_free_queue(struct ntb_transport_qp
*qp
)
1880 struct pci_dev
*pdev
;
1881 struct ntb_queue_entry
*entry
;
1887 pdev
= qp
->ndev
->pdev
;
1891 if (qp
->tx_dma_chan
) {
1892 struct dma_chan
*chan
= qp
->tx_dma_chan
;
1893 /* Putting the dma_chan to NULL will force any new traffic to be
1894 * processed by the CPU instead of the DAM engine
1896 qp
->tx_dma_chan
= NULL
;
1898 /* Try to be nice and wait for any queued DMA engine
1899 * transactions to process before smashing it with a rock
1901 dma_sync_wait(chan
, qp
->last_cookie
);
1902 dmaengine_terminate_all(chan
);
1903 dma_release_channel(chan
);
1906 if (qp
->rx_dma_chan
) {
1907 struct dma_chan
*chan
= qp
->rx_dma_chan
;
1908 /* Putting the dma_chan to NULL will force any new traffic to be
1909 * processed by the CPU instead of the DAM engine
1911 qp
->rx_dma_chan
= NULL
;
1913 /* Try to be nice and wait for any queued DMA engine
1914 * transactions to process before smashing it with a rock
1916 dma_sync_wait(chan
, qp
->last_cookie
);
1917 dmaengine_terminate_all(chan
);
1918 dma_release_channel(chan
);
1921 qp_bit
= BIT_ULL(qp
->qp_num
);
1923 ntb_db_set_mask(qp
->ndev
, qp_bit
);
1924 tasklet_kill(&qp
->rxc_db_work
);
1926 cancel_delayed_work_sync(&qp
->link_work
);
1929 qp
->rx_handler
= NULL
;
1930 qp
->tx_handler
= NULL
;
1931 qp
->event_handler
= NULL
;
1933 while ((entry
= ntb_list_rm(&qp
->ntb_rx_q_lock
, &qp
->rx_free_q
)))
1936 while ((entry
= ntb_list_rm(&qp
->ntb_rx_q_lock
, &qp
->rx_pend_q
))) {
1937 dev_warn(&pdev
->dev
, "Freeing item from non-empty rx_pend_q\n");
1941 while ((entry
= ntb_list_rm(&qp
->ntb_rx_q_lock
, &qp
->rx_post_q
))) {
1942 dev_warn(&pdev
->dev
, "Freeing item from non-empty rx_post_q\n");
1946 while ((entry
= ntb_list_rm(&qp
->ntb_tx_free_q_lock
, &qp
->tx_free_q
)))
1949 qp
->transport
->qp_bitmap_free
|= qp_bit
;
1951 dev_info(&pdev
->dev
, "NTB Transport QP %d freed\n", qp
->qp_num
);
1953 EXPORT_SYMBOL_GPL(ntb_transport_free_queue
);
1956 * ntb_transport_rx_remove - Dequeues enqueued rx packet
1957 * @qp: NTB queue to be freed
1958 * @len: pointer to variable to write enqueued buffers length
1960 * Dequeues unused buffers from receive queue. Should only be used during
1963 * RETURNS: NULL error value on error, or void* for success.
1965 void *ntb_transport_rx_remove(struct ntb_transport_qp
*qp
, unsigned int *len
)
1967 struct ntb_queue_entry
*entry
;
1970 if (!qp
|| qp
->client_ready
)
1973 entry
= ntb_list_rm(&qp
->ntb_rx_q_lock
, &qp
->rx_pend_q
);
1977 buf
= entry
->cb_data
;
1980 ntb_list_add(&qp
->ntb_rx_q_lock
, &entry
->entry
, &qp
->rx_free_q
);
1984 EXPORT_SYMBOL_GPL(ntb_transport_rx_remove
);
1987 * ntb_transport_rx_enqueue - Enqueue a new NTB queue entry
1988 * @qp: NTB transport layer queue the entry is to be enqueued on
1989 * @cb: per buffer pointer for callback function to use
1990 * @data: pointer to data buffer that incoming packets will be copied into
1991 * @len: length of the data buffer
1993 * Enqueue a new receive buffer onto the transport queue into which a NTB
1994 * payload can be received into.
1996 * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
1998 int ntb_transport_rx_enqueue(struct ntb_transport_qp
*qp
, void *cb
, void *data
,
2001 struct ntb_queue_entry
*entry
;
2006 entry
= ntb_list_rm(&qp
->ntb_rx_q_lock
, &qp
->rx_free_q
);
2010 entry
->cb_data
= cb
;
2016 entry
->rx_index
= 0;
2018 ntb_list_add(&qp
->ntb_rx_q_lock
, &entry
->entry
, &qp
->rx_pend_q
);
2021 tasklet_schedule(&qp
->rxc_db_work
);
2025 EXPORT_SYMBOL_GPL(ntb_transport_rx_enqueue
);
2028 * ntb_transport_tx_enqueue - Enqueue a new NTB queue entry
2029 * @qp: NTB transport layer queue the entry is to be enqueued on
2030 * @cb: per buffer pointer for callback function to use
2031 * @data: pointer to data buffer that will be sent
2032 * @len: length of the data buffer
2034 * Enqueue a new transmit buffer onto the transport queue from which a NTB
2035 * payload will be transmitted. This assumes that a lock is being held to
2036 * serialize access to the qp.
2038 * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
2040 int ntb_transport_tx_enqueue(struct ntb_transport_qp
*qp
, void *cb
, void *data
,
2043 struct ntb_queue_entry
*entry
;
2046 if (!qp
|| !qp
->link_is_up
|| !len
)
2049 entry
= ntb_list_rm(&qp
->ntb_tx_free_q_lock
, &qp
->tx_free_q
);
2051 qp
->tx_err_no_buf
++;
2055 entry
->cb_data
= cb
;
2061 entry
->tx_index
= 0;
2063 rc
= ntb_process_tx(qp
, entry
);
2065 ntb_list_add(&qp
->ntb_tx_free_q_lock
, &entry
->entry
,
2070 EXPORT_SYMBOL_GPL(ntb_transport_tx_enqueue
);
2073 * ntb_transport_link_up - Notify NTB transport of client readiness to use queue
2074 * @qp: NTB transport layer queue to be enabled
2076 * Notify NTB transport layer of client readiness to use queue
2078 void ntb_transport_link_up(struct ntb_transport_qp
*qp
)
2083 qp
->client_ready
= true;
2085 if (qp
->transport
->link_is_up
)
2086 schedule_delayed_work(&qp
->link_work
, 0);
2088 EXPORT_SYMBOL_GPL(ntb_transport_link_up
);
2091 * ntb_transport_link_down - Notify NTB transport to no longer enqueue data
2092 * @qp: NTB transport layer queue to be disabled
2094 * Notify NTB transport layer of client's desire to no longer receive data on
2095 * transport queue specified. It is the client's responsibility to ensure all
2096 * entries on queue are purged or otherwise handled appropriately.
2098 void ntb_transport_link_down(struct ntb_transport_qp
*qp
)
2105 qp
->client_ready
= false;
2107 val
= ntb_spad_read(qp
->ndev
, QP_LINKS
);
2109 ntb_peer_spad_write(qp
->ndev
, PIDX
, QP_LINKS
, val
& ~BIT(qp
->qp_num
));
2112 ntb_send_link_down(qp
);
2114 cancel_delayed_work_sync(&qp
->link_work
);
2116 EXPORT_SYMBOL_GPL(ntb_transport_link_down
);
2119 * ntb_transport_link_query - Query transport link state
2120 * @qp: NTB transport layer queue to be queried
2122 * Query connectivity to the remote system of the NTB transport queue
2124 * RETURNS: true for link up or false for link down
2126 bool ntb_transport_link_query(struct ntb_transport_qp
*qp
)
2131 return qp
->link_is_up
;
2133 EXPORT_SYMBOL_GPL(ntb_transport_link_query
);
2136 * ntb_transport_qp_num - Query the qp number
2137 * @qp: NTB transport layer queue to be queried
2139 * Query qp number of the NTB transport queue
2141 * RETURNS: a zero based number specifying the qp number
2143 unsigned char ntb_transport_qp_num(struct ntb_transport_qp
*qp
)
2150 EXPORT_SYMBOL_GPL(ntb_transport_qp_num
);
2153 * ntb_transport_max_size - Query the max payload size of a qp
2154 * @qp: NTB transport layer queue to be queried
2156 * Query the maximum payload size permissible on the given qp
2158 * RETURNS: the max payload size of a qp
2160 unsigned int ntb_transport_max_size(struct ntb_transport_qp
*qp
)
2162 unsigned int max_size
;
2163 unsigned int copy_align
;
2164 struct dma_chan
*rx_chan
, *tx_chan
;
2169 rx_chan
= qp
->rx_dma_chan
;
2170 tx_chan
= qp
->tx_dma_chan
;
2172 copy_align
= max(rx_chan
? rx_chan
->device
->copy_align
: 0,
2173 tx_chan
? tx_chan
->device
->copy_align
: 0);
2175 /* If DMA engine usage is possible, try to find the max size for that */
2176 max_size
= qp
->tx_max_frame
- sizeof(struct ntb_payload_header
);
2177 max_size
= round_down(max_size
, 1 << copy_align
);
2181 EXPORT_SYMBOL_GPL(ntb_transport_max_size
);
2183 unsigned int ntb_transport_tx_free_entry(struct ntb_transport_qp
*qp
)
2185 unsigned int head
= qp
->tx_index
;
2186 unsigned int tail
= qp
->remote_rx_info
->entry
;
2188 return tail
> head
? tail
- head
: qp
->tx_max_entry
+ tail
- head
;
2190 EXPORT_SYMBOL_GPL(ntb_transport_tx_free_entry
);
2192 static void ntb_transport_doorbell_callback(void *data
, int vector
)
2194 struct ntb_transport_ctx
*nt
= data
;
2195 struct ntb_transport_qp
*qp
;
2197 unsigned int qp_num
;
2199 db_bits
= (nt
->qp_bitmap
& ~nt
->qp_bitmap_free
&
2200 ntb_db_vector_mask(nt
->ndev
, vector
));
2203 qp_num
= __ffs(db_bits
);
2204 qp
= &nt
->qp_vec
[qp_num
];
2207 tasklet_schedule(&qp
->rxc_db_work
);
2209 db_bits
&= ~BIT_ULL(qp_num
);
2213 static const struct ntb_ctx_ops ntb_transport_ops
= {
2214 .link_event
= ntb_transport_event_callback
,
2215 .db_event
= ntb_transport_doorbell_callback
,
2218 static struct ntb_client ntb_transport_client
= {
2220 .probe
= ntb_transport_probe
,
2221 .remove
= ntb_transport_free
,
2225 static int __init
ntb_transport_init(void)
2229 pr_info("%s, version %s\n", NTB_TRANSPORT_DESC
, NTB_TRANSPORT_VER
);
2231 if (debugfs_initialized())
2232 nt_debugfs_dir
= debugfs_create_dir(KBUILD_MODNAME
, NULL
);
2234 rc
= bus_register(&ntb_transport_bus
);
2238 rc
= ntb_register_client(&ntb_transport_client
);
2245 bus_unregister(&ntb_transport_bus
);
2247 debugfs_remove_recursive(nt_debugfs_dir
);
2250 module_init(ntb_transport_init
);
2252 static void __exit
ntb_transport_exit(void)
2254 ntb_unregister_client(&ntb_transport_client
);
2255 bus_unregister(&ntb_transport_bus
);
2256 debugfs_remove_recursive(nt_debugfs_dir
);
2258 module_exit(ntb_transport_exit
);