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"
70 MODULE_DESCRIPTION(NTB_TRANSPORT_DESC);
71 MODULE_VERSION(NTB_TRANSPORT_VER);
72 MODULE_LICENSE("Dual BSD/GPL");
73 MODULE_AUTHOR("Intel Corporation");
75 static unsigned long max_mw_size;
76 module_param(max_mw_size, ulong, 0644);
77 MODULE_PARM_DESC(max_mw_size, "Limit size of large memory windows");
79 static unsigned int transport_mtu = 0x10000;
80 module_param(transport_mtu, uint, 0644);
81 MODULE_PARM_DESC(transport_mtu, "Maximum size of NTB transport packets");
83 static unsigned char max_num_clients;
84 module_param(max_num_clients, byte, 0644);
85 MODULE_PARM_DESC(max_num_clients, "Maximum number of NTB transport clients");
87 static unsigned int copy_bytes = 1024;
88 module_param(copy_bytes, uint, 0644);
89 MODULE_PARM_DESC(copy_bytes, "Threshold under which NTB will use the CPU to copy instead of DMA");
92 module_param(use_dma, bool, 0644);
93 MODULE_PARM_DESC(use_dma, "Use DMA engine to perform large data copy");
95 static struct dentry *nt_debugfs_dir;
97 struct ntb_queue_entry {
98 /* ntb_queue list reference */
99 struct list_head entry;
100 /* pointers to data to be transferred */
106 struct ntb_transport_qp *qp;
108 struct ntb_payload_header __iomem *tx_hdr;
109 struct ntb_payload_header *rx_hdr;
118 struct ntb_transport_qp {
119 struct ntb_transport_ctx *transport;
120 struct ntb_dev *ndev;
122 struct dma_chan *dma_chan;
127 u8 qp_num; /* Only 64 QP's are allowed. 0-63 */
130 struct ntb_rx_info __iomem *rx_info;
131 struct ntb_rx_info *remote_rx_info;
133 void (*tx_handler)(struct ntb_transport_qp *qp, void *qp_data,
134 void *data, int len);
135 struct list_head tx_free_q;
136 spinlock_t ntb_tx_free_q_lock;
138 dma_addr_t tx_mw_phys;
139 unsigned int tx_index;
140 unsigned int tx_max_entry;
141 unsigned int tx_max_frame;
143 void (*rx_handler)(struct ntb_transport_qp *qp, void *qp_data,
144 void *data, int len);
145 struct list_head rx_post_q;
146 struct list_head rx_pend_q;
147 struct list_head rx_free_q;
148 /* ntb_rx_q_lock: synchronize access to rx_XXXX_q */
149 spinlock_t ntb_rx_q_lock;
151 unsigned int rx_index;
152 unsigned int rx_max_entry;
153 unsigned int rx_max_frame;
154 dma_cookie_t last_cookie;
155 struct tasklet_struct rxc_db_work;
157 void (*event_handler)(void *data, int status);
158 struct delayed_work link_work;
159 struct work_struct link_cleanup;
161 struct dentry *debugfs_dir;
162 struct dentry *debugfs_stats;
181 struct ntb_transport_mw {
182 phys_addr_t phys_addr;
183 resource_size_t phys_size;
184 resource_size_t xlat_align;
185 resource_size_t xlat_align_size;
193 struct ntb_transport_client_dev {
194 struct list_head entry;
195 struct ntb_transport_ctx *nt;
199 struct ntb_transport_ctx {
200 struct list_head entry;
201 struct list_head client_devs;
203 struct ntb_dev *ndev;
205 struct ntb_transport_mw *mw_vec;
206 struct ntb_transport_qp *qp_vec;
207 unsigned int mw_count;
208 unsigned int qp_count;
213 struct delayed_work link_work;
214 struct work_struct link_cleanup;
218 DESC_DONE_FLAG = BIT(0),
219 LINK_DOWN_FLAG = BIT(1),
222 struct ntb_payload_header {
240 #define dev_client_dev(__dev) \
241 container_of((__dev), struct ntb_transport_client_dev, dev)
243 #define drv_client(__drv) \
244 container_of((__drv), struct ntb_transport_client, driver)
246 #define QP_TO_MW(nt, qp) ((qp) % nt->mw_count)
247 #define NTB_QP_DEF_NUM_ENTRIES 100
248 #define NTB_LINK_DOWN_TIMEOUT 10
250 static void ntb_transport_rxc_db(unsigned long data);
251 static const struct ntb_ctx_ops ntb_transport_ops;
252 static struct ntb_client ntb_transport_client;
254 static int ntb_transport_bus_match(struct device *dev,
255 struct device_driver *drv)
257 return !strncmp(dev_name(dev), drv->name, strlen(drv->name));
260 static int ntb_transport_bus_probe(struct device *dev)
262 const struct ntb_transport_client *client;
267 client = drv_client(dev->driver);
268 rc = client->probe(dev);
275 static int ntb_transport_bus_remove(struct device *dev)
277 const struct ntb_transport_client *client;
279 client = drv_client(dev->driver);
287 static struct bus_type ntb_transport_bus = {
288 .name = "ntb_transport",
289 .match = ntb_transport_bus_match,
290 .probe = ntb_transport_bus_probe,
291 .remove = ntb_transport_bus_remove,
294 static LIST_HEAD(ntb_transport_list);
296 static int ntb_bus_init(struct ntb_transport_ctx *nt)
298 list_add(&nt->entry, &ntb_transport_list);
302 static void ntb_bus_remove(struct ntb_transport_ctx *nt)
304 struct ntb_transport_client_dev *client_dev, *cd;
306 list_for_each_entry_safe(client_dev, cd, &nt->client_devs, entry) {
307 dev_err(client_dev->dev.parent, "%s still attached to bus, removing\n",
308 dev_name(&client_dev->dev));
309 list_del(&client_dev->entry);
310 device_unregister(&client_dev->dev);
313 list_del(&nt->entry);
316 static void ntb_transport_client_release(struct device *dev)
318 struct ntb_transport_client_dev *client_dev;
320 client_dev = dev_client_dev(dev);
325 * ntb_transport_unregister_client_dev - Unregister NTB client device
326 * @device_name: Name of NTB client device
328 * Unregister an NTB client device with the NTB transport layer
330 void ntb_transport_unregister_client_dev(char *device_name)
332 struct ntb_transport_client_dev *client, *cd;
333 struct ntb_transport_ctx *nt;
335 list_for_each_entry(nt, &ntb_transport_list, entry)
336 list_for_each_entry_safe(client, cd, &nt->client_devs, entry)
337 if (!strncmp(dev_name(&client->dev), device_name,
338 strlen(device_name))) {
339 list_del(&client->entry);
340 device_unregister(&client->dev);
343 EXPORT_SYMBOL_GPL(ntb_transport_unregister_client_dev);
346 * ntb_transport_register_client_dev - Register NTB client device
347 * @device_name: Name of NTB client device
349 * Register an NTB client device with the NTB transport layer
351 int ntb_transport_register_client_dev(char *device_name)
353 struct ntb_transport_client_dev *client_dev;
354 struct ntb_transport_ctx *nt;
358 if (list_empty(&ntb_transport_list))
361 list_for_each_entry(nt, &ntb_transport_list, entry) {
364 node = dev_to_node(&nt->ndev->dev);
366 client_dev = kzalloc_node(sizeof(*client_dev),
373 dev = &client_dev->dev;
375 /* setup and register client devices */
376 dev_set_name(dev, "%s%d", device_name, i);
377 dev->bus = &ntb_transport_bus;
378 dev->release = ntb_transport_client_release;
379 dev->parent = &nt->ndev->dev;
381 rc = device_register(dev);
387 list_add_tail(&client_dev->entry, &nt->client_devs);
394 ntb_transport_unregister_client_dev(device_name);
398 EXPORT_SYMBOL_GPL(ntb_transport_register_client_dev);
401 * ntb_transport_register_client - Register NTB client driver
402 * @drv: NTB client driver to be registered
404 * Register an NTB client driver with the NTB transport layer
406 * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
408 int ntb_transport_register_client(struct ntb_transport_client *drv)
410 drv->driver.bus = &ntb_transport_bus;
412 if (list_empty(&ntb_transport_list))
415 return driver_register(&drv->driver);
417 EXPORT_SYMBOL_GPL(ntb_transport_register_client);
420 * ntb_transport_unregister_client - Unregister NTB client driver
421 * @drv: NTB client driver to be unregistered
423 * Unregister an NTB client driver with the NTB transport layer
425 * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
427 void ntb_transport_unregister_client(struct ntb_transport_client *drv)
429 driver_unregister(&drv->driver);
431 EXPORT_SYMBOL_GPL(ntb_transport_unregister_client);
433 static ssize_t debugfs_read(struct file *filp, char __user *ubuf, size_t count,
436 struct ntb_transport_qp *qp;
438 ssize_t ret, out_offset, out_count;
442 buf = kmalloc(out_count, GFP_KERNEL);
446 qp = filp->private_data;
448 out_offset += snprintf(buf + out_offset, out_count - out_offset,
450 out_offset += snprintf(buf + out_offset, out_count - out_offset,
451 "rx_bytes - \t%llu\n", qp->rx_bytes);
452 out_offset += snprintf(buf + out_offset, out_count - out_offset,
453 "rx_pkts - \t%llu\n", qp->rx_pkts);
454 out_offset += snprintf(buf + out_offset, out_count - out_offset,
455 "rx_memcpy - \t%llu\n", qp->rx_memcpy);
456 out_offset += snprintf(buf + out_offset, out_count - out_offset,
457 "rx_async - \t%llu\n", qp->rx_async);
458 out_offset += snprintf(buf + out_offset, out_count - out_offset,
459 "rx_ring_empty - %llu\n", qp->rx_ring_empty);
460 out_offset += snprintf(buf + out_offset, out_count - out_offset,
461 "rx_err_no_buf - %llu\n", qp->rx_err_no_buf);
462 out_offset += snprintf(buf + out_offset, out_count - out_offset,
463 "rx_err_oflow - \t%llu\n", qp->rx_err_oflow);
464 out_offset += snprintf(buf + out_offset, out_count - out_offset,
465 "rx_err_ver - \t%llu\n", qp->rx_err_ver);
466 out_offset += snprintf(buf + out_offset, out_count - out_offset,
467 "rx_buff - \t%p\n", qp->rx_buff);
468 out_offset += snprintf(buf + out_offset, out_count - out_offset,
469 "rx_index - \t%u\n", qp->rx_index);
470 out_offset += snprintf(buf + out_offset, out_count - out_offset,
471 "rx_max_entry - \t%u\n", qp->rx_max_entry);
473 out_offset += snprintf(buf + out_offset, out_count - out_offset,
474 "tx_bytes - \t%llu\n", qp->tx_bytes);
475 out_offset += snprintf(buf + out_offset, out_count - out_offset,
476 "tx_pkts - \t%llu\n", qp->tx_pkts);
477 out_offset += snprintf(buf + out_offset, out_count - out_offset,
478 "tx_memcpy - \t%llu\n", qp->tx_memcpy);
479 out_offset += snprintf(buf + out_offset, out_count - out_offset,
480 "tx_async - \t%llu\n", qp->tx_async);
481 out_offset += snprintf(buf + out_offset, out_count - out_offset,
482 "tx_ring_full - \t%llu\n", qp->tx_ring_full);
483 out_offset += snprintf(buf + out_offset, out_count - out_offset,
484 "tx_err_no_buf - %llu\n", qp->tx_err_no_buf);
485 out_offset += snprintf(buf + out_offset, out_count - out_offset,
486 "tx_mw - \t%p\n", qp->tx_mw);
487 out_offset += snprintf(buf + out_offset, out_count - out_offset,
488 "tx_index - \t%u\n", qp->tx_index);
489 out_offset += snprintf(buf + out_offset, out_count - out_offset,
490 "tx_max_entry - \t%u\n", qp->tx_max_entry);
492 out_offset += snprintf(buf + out_offset, out_count - out_offset,
494 qp->link_is_up ? "Up" : "Down");
495 if (out_offset > out_count)
496 out_offset = out_count;
498 ret = simple_read_from_buffer(ubuf, count, offp, buf, out_offset);
503 static const struct file_operations ntb_qp_debugfs_stats = {
504 .owner = THIS_MODULE,
506 .read = debugfs_read,
509 static void ntb_list_add(spinlock_t *lock, struct list_head *entry,
510 struct list_head *list)
514 spin_lock_irqsave(lock, flags);
515 list_add_tail(entry, list);
516 spin_unlock_irqrestore(lock, flags);
519 static struct ntb_queue_entry *ntb_list_rm(spinlock_t *lock,
520 struct list_head *list)
522 struct ntb_queue_entry *entry;
525 spin_lock_irqsave(lock, flags);
526 if (list_empty(list)) {
530 entry = list_first_entry(list, struct ntb_queue_entry, entry);
531 list_del(&entry->entry);
533 spin_unlock_irqrestore(lock, flags);
538 static struct ntb_queue_entry *ntb_list_mv(spinlock_t *lock,
539 struct list_head *list,
540 struct list_head *to_list)
542 struct ntb_queue_entry *entry;
545 spin_lock_irqsave(lock, flags);
547 if (list_empty(list)) {
550 entry = list_first_entry(list, struct ntb_queue_entry, entry);
551 list_move_tail(&entry->entry, to_list);
554 spin_unlock_irqrestore(lock, flags);
559 static int ntb_transport_setup_qp_mw(struct ntb_transport_ctx *nt,
562 struct ntb_transport_qp *qp = &nt->qp_vec[qp_num];
563 struct ntb_transport_mw *mw;
564 unsigned int rx_size, num_qps_mw;
565 unsigned int mw_num, mw_count, qp_count;
568 mw_count = nt->mw_count;
569 qp_count = nt->qp_count;
571 mw_num = QP_TO_MW(nt, qp_num);
572 mw = &nt->mw_vec[mw_num];
577 if (qp_count % mw_count && mw_num + 1 < qp_count / mw_count)
578 num_qps_mw = qp_count / mw_count + 1;
580 num_qps_mw = qp_count / mw_count;
582 rx_size = (unsigned int)mw->xlat_size / num_qps_mw;
583 qp->rx_buff = mw->virt_addr + rx_size * qp_num / mw_count;
584 rx_size -= sizeof(struct ntb_rx_info);
586 qp->remote_rx_info = qp->rx_buff + rx_size;
588 /* Due to housekeeping, there must be atleast 2 buffs */
589 qp->rx_max_frame = min(transport_mtu, rx_size / 2);
590 qp->rx_max_entry = rx_size / qp->rx_max_frame;
593 qp->remote_rx_info->entry = qp->rx_max_entry - 1;
595 /* setup the hdr offsets with 0's */
596 for (i = 0; i < qp->rx_max_entry; i++) {
597 void *offset = (qp->rx_buff + qp->rx_max_frame * (i + 1) -
598 sizeof(struct ntb_payload_header));
599 memset(offset, 0, sizeof(struct ntb_payload_header));
609 static void ntb_free_mw(struct ntb_transport_ctx *nt, int num_mw)
611 struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
612 struct pci_dev *pdev = nt->ndev->pdev;
617 ntb_mw_clear_trans(nt->ndev, num_mw);
618 dma_free_coherent(&pdev->dev, mw->buff_size,
619 mw->virt_addr, mw->dma_addr);
622 mw->virt_addr = NULL;
625 static int ntb_set_mw(struct ntb_transport_ctx *nt, int num_mw,
628 struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
629 struct pci_dev *pdev = nt->ndev->pdev;
630 unsigned int xlat_size, buff_size;
633 xlat_size = round_up(size, mw->xlat_align_size);
634 buff_size = round_up(size, mw->xlat_align);
636 /* No need to re-setup */
637 if (mw->xlat_size == xlat_size)
641 ntb_free_mw(nt, num_mw);
643 /* Alloc memory for receiving data. Must be aligned */
644 mw->xlat_size = xlat_size;
645 mw->buff_size = buff_size;
647 mw->virt_addr = dma_alloc_coherent(&pdev->dev, buff_size,
648 &mw->dma_addr, GFP_KERNEL);
649 if (!mw->virt_addr) {
652 dev_err(&pdev->dev, "Unable to alloc MW buff of size %d\n",
658 * we must ensure that the memory address allocated is BAR size
659 * aligned in order for the XLAT register to take the value. This
660 * is a requirement of the hardware. It is recommended to setup CMA
661 * for BAR sizes equal or greater than 4MB.
663 if (!IS_ALIGNED(mw->dma_addr, mw->xlat_align)) {
664 dev_err(&pdev->dev, "DMA memory %pad is not aligned\n",
666 ntb_free_mw(nt, num_mw);
670 /* Notify HW the memory location of the receive buffer */
671 rc = ntb_mw_set_trans(nt->ndev, num_mw, mw->dma_addr, mw->xlat_size);
673 dev_err(&pdev->dev, "Unable to set mw%d translation", num_mw);
674 ntb_free_mw(nt, num_mw);
681 static void ntb_qp_link_down_reset(struct ntb_transport_qp *qp)
683 qp->link_is_up = false;
689 qp->rx_ring_empty = 0;
690 qp->rx_err_no_buf = 0;
691 qp->rx_err_oflow = 0;
697 qp->tx_ring_full = 0;
698 qp->tx_err_no_buf = 0;
703 static void ntb_qp_link_cleanup(struct ntb_transport_qp *qp)
705 struct ntb_transport_ctx *nt = qp->transport;
706 struct pci_dev *pdev = nt->ndev->pdev;
708 dev_info(&pdev->dev, "qp %d: Link Cleanup\n", qp->qp_num);
710 cancel_delayed_work_sync(&qp->link_work);
711 ntb_qp_link_down_reset(qp);
713 if (qp->event_handler)
714 qp->event_handler(qp->cb_data, qp->link_is_up);
717 static void ntb_qp_link_cleanup_work(struct work_struct *work)
719 struct ntb_transport_qp *qp = container_of(work,
720 struct ntb_transport_qp,
722 struct ntb_transport_ctx *nt = qp->transport;
724 ntb_qp_link_cleanup(qp);
727 schedule_delayed_work(&qp->link_work,
728 msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
731 static void ntb_qp_link_down(struct ntb_transport_qp *qp)
733 schedule_work(&qp->link_cleanup);
736 static void ntb_transport_link_cleanup(struct ntb_transport_ctx *nt)
738 struct ntb_transport_qp *qp;
742 qp_bitmap_alloc = nt->qp_bitmap & ~nt->qp_bitmap_free;
744 /* Pass along the info to any clients */
745 for (i = 0; i < nt->qp_count; i++)
746 if (qp_bitmap_alloc & BIT_ULL(i)) {
748 ntb_qp_link_cleanup(qp);
749 cancel_work_sync(&qp->link_cleanup);
750 cancel_delayed_work_sync(&qp->link_work);
754 cancel_delayed_work_sync(&nt->link_work);
756 /* The scratchpad registers keep the values if the remote side
757 * goes down, blast them now to give them a sane value the next
758 * time they are accessed
760 for (i = 0; i < MAX_SPAD; i++)
761 ntb_spad_write(nt->ndev, i, 0);
764 static void ntb_transport_link_cleanup_work(struct work_struct *work)
766 struct ntb_transport_ctx *nt =
767 container_of(work, struct ntb_transport_ctx, link_cleanup);
769 ntb_transport_link_cleanup(nt);
772 static void ntb_transport_event_callback(void *data)
774 struct ntb_transport_ctx *nt = data;
776 if (ntb_link_is_up(nt->ndev, NULL, NULL) == 1)
777 schedule_delayed_work(&nt->link_work, 0);
779 schedule_work(&nt->link_cleanup);
782 static void ntb_transport_link_work(struct work_struct *work)
784 struct ntb_transport_ctx *nt =
785 container_of(work, struct ntb_transport_ctx, link_work.work);
786 struct ntb_dev *ndev = nt->ndev;
787 struct pci_dev *pdev = ndev->pdev;
788 resource_size_t size;
792 /* send the local info, in the opposite order of the way we read it */
793 for (i = 0; i < nt->mw_count; i++) {
794 size = nt->mw_vec[i].phys_size;
796 if (max_mw_size && size > max_mw_size)
799 spad = MW0_SZ_HIGH + (i * 2);
800 ntb_peer_spad_write(ndev, spad, (u32)(size >> 32));
802 spad = MW0_SZ_LOW + (i * 2);
803 ntb_peer_spad_write(ndev, spad, (u32)size);
806 ntb_peer_spad_write(ndev, NUM_MWS, nt->mw_count);
808 ntb_peer_spad_write(ndev, NUM_QPS, nt->qp_count);
810 ntb_peer_spad_write(ndev, VERSION, NTB_TRANSPORT_VERSION);
812 /* Query the remote side for its info */
813 val = ntb_spad_read(ndev, VERSION);
814 dev_dbg(&pdev->dev, "Remote version = %d\n", val);
815 if (val != NTB_TRANSPORT_VERSION)
818 val = ntb_spad_read(ndev, NUM_QPS);
819 dev_dbg(&pdev->dev, "Remote max number of qps = %d\n", val);
820 if (val != nt->qp_count)
823 val = ntb_spad_read(ndev, NUM_MWS);
824 dev_dbg(&pdev->dev, "Remote number of mws = %d\n", val);
825 if (val != nt->mw_count)
828 for (i = 0; i < nt->mw_count; i++) {
831 val = ntb_spad_read(ndev, MW0_SZ_HIGH + (i * 2));
832 val64 = (u64)val << 32;
834 val = ntb_spad_read(ndev, MW0_SZ_LOW + (i * 2));
837 dev_dbg(&pdev->dev, "Remote MW%d size = %#llx\n", i, val64);
839 rc = ntb_set_mw(nt, i, val64);
844 nt->link_is_up = true;
846 for (i = 0; i < nt->qp_count; i++) {
847 struct ntb_transport_qp *qp = &nt->qp_vec[i];
849 ntb_transport_setup_qp_mw(nt, i);
851 if (qp->client_ready)
852 schedule_delayed_work(&qp->link_work, 0);
858 for (i = 0; i < nt->mw_count; i++)
861 if (ntb_link_is_up(ndev, NULL, NULL) == 1)
862 schedule_delayed_work(&nt->link_work,
863 msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
866 static void ntb_qp_link_work(struct work_struct *work)
868 struct ntb_transport_qp *qp = container_of(work,
869 struct ntb_transport_qp,
871 struct pci_dev *pdev = qp->ndev->pdev;
872 struct ntb_transport_ctx *nt = qp->transport;
875 WARN_ON(!nt->link_is_up);
877 val = ntb_spad_read(nt->ndev, QP_LINKS);
879 ntb_peer_spad_write(nt->ndev, QP_LINKS, val | BIT(qp->qp_num));
881 /* query remote spad for qp ready bits */
882 ntb_peer_spad_read(nt->ndev, QP_LINKS);
883 dev_dbg_ratelimited(&pdev->dev, "Remote QP link status = %x\n", val);
885 /* See if the remote side is up */
886 if (val & BIT(qp->qp_num)) {
887 dev_info(&pdev->dev, "qp %d: Link Up\n", qp->qp_num);
888 qp->link_is_up = true;
890 if (qp->event_handler)
891 qp->event_handler(qp->cb_data, qp->link_is_up);
892 } else if (nt->link_is_up)
893 schedule_delayed_work(&qp->link_work,
894 msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
897 static int ntb_transport_init_queue(struct ntb_transport_ctx *nt,
900 struct ntb_transport_qp *qp;
901 struct ntb_transport_mw *mw;
903 resource_size_t mw_size;
904 unsigned int num_qps_mw, tx_size;
905 unsigned int mw_num, mw_count, qp_count;
908 mw_count = nt->mw_count;
909 qp_count = nt->qp_count;
911 mw_num = QP_TO_MW(nt, qp_num);
912 mw = &nt->mw_vec[mw_num];
914 qp = &nt->qp_vec[qp_num];
918 qp->client_ready = false;
919 qp->event_handler = NULL;
920 ntb_qp_link_down_reset(qp);
922 if (qp_count % mw_count && mw_num + 1 < qp_count / mw_count)
923 num_qps_mw = qp_count / mw_count + 1;
925 num_qps_mw = qp_count / mw_count;
927 mw_base = nt->mw_vec[mw_num].phys_addr;
928 mw_size = nt->mw_vec[mw_num].phys_size;
930 tx_size = (unsigned int)mw_size / num_qps_mw;
931 qp_offset = tx_size * qp_num / mw_count;
933 qp->tx_mw = nt->mw_vec[mw_num].vbase + qp_offset;
937 qp->tx_mw_phys = mw_base + qp_offset;
941 tx_size -= sizeof(struct ntb_rx_info);
942 qp->rx_info = qp->tx_mw + tx_size;
944 /* Due to housekeeping, there must be atleast 2 buffs */
945 qp->tx_max_frame = min(transport_mtu, tx_size / 2);
946 qp->tx_max_entry = tx_size / qp->tx_max_frame;
948 if (nt_debugfs_dir) {
949 char debugfs_name[4];
951 snprintf(debugfs_name, 4, "qp%d", qp_num);
952 qp->debugfs_dir = debugfs_create_dir(debugfs_name,
955 qp->debugfs_stats = debugfs_create_file("stats", S_IRUSR,
957 &ntb_qp_debugfs_stats);
959 qp->debugfs_dir = NULL;
960 qp->debugfs_stats = NULL;
963 INIT_DELAYED_WORK(&qp->link_work, ntb_qp_link_work);
964 INIT_WORK(&qp->link_cleanup, ntb_qp_link_cleanup_work);
966 spin_lock_init(&qp->ntb_rx_q_lock);
967 spin_lock_init(&qp->ntb_tx_free_q_lock);
969 INIT_LIST_HEAD(&qp->rx_post_q);
970 INIT_LIST_HEAD(&qp->rx_pend_q);
971 INIT_LIST_HEAD(&qp->rx_free_q);
972 INIT_LIST_HEAD(&qp->tx_free_q);
974 tasklet_init(&qp->rxc_db_work, ntb_transport_rxc_db,
980 static int ntb_transport_probe(struct ntb_client *self, struct ntb_dev *ndev)
982 struct ntb_transport_ctx *nt;
983 struct ntb_transport_mw *mw;
984 unsigned int mw_count, qp_count;
989 if (ntb_db_is_unsafe(ndev))
991 "doorbell is unsafe, proceed anyway...\n");
992 if (ntb_spad_is_unsafe(ndev))
994 "scratchpad is unsafe, proceed anyway...\n");
996 node = dev_to_node(&ndev->dev);
998 nt = kzalloc_node(sizeof(*nt), GFP_KERNEL, node);
1004 mw_count = ntb_mw_count(ndev);
1006 nt->mw_count = mw_count;
1008 nt->mw_vec = kzalloc_node(mw_count * sizeof(*nt->mw_vec),
1015 for (i = 0; i < mw_count; i++) {
1016 mw = &nt->mw_vec[i];
1018 rc = ntb_mw_get_range(ndev, i, &mw->phys_addr, &mw->phys_size,
1019 &mw->xlat_align, &mw->xlat_align_size);
1023 mw->vbase = ioremap_wc(mw->phys_addr, mw->phys_size);
1031 mw->virt_addr = NULL;
1035 qp_bitmap = ntb_db_valid_mask(ndev);
1037 qp_count = ilog2(qp_bitmap);
1038 if (max_num_clients && max_num_clients < qp_count)
1039 qp_count = max_num_clients;
1040 else if (mw_count < qp_count)
1041 qp_count = mw_count;
1043 qp_bitmap &= BIT_ULL(qp_count) - 1;
1045 nt->qp_count = qp_count;
1046 nt->qp_bitmap = qp_bitmap;
1047 nt->qp_bitmap_free = qp_bitmap;
1049 nt->qp_vec = kzalloc_node(qp_count * sizeof(*nt->qp_vec),
1056 for (i = 0; i < qp_count; i++) {
1057 rc = ntb_transport_init_queue(nt, i);
1062 INIT_DELAYED_WORK(&nt->link_work, ntb_transport_link_work);
1063 INIT_WORK(&nt->link_cleanup, ntb_transport_link_cleanup_work);
1065 rc = ntb_set_ctx(ndev, nt, &ntb_transport_ops);
1069 INIT_LIST_HEAD(&nt->client_devs);
1070 rc = ntb_bus_init(nt);
1074 nt->link_is_up = false;
1075 ntb_link_enable(ndev, NTB_SPEED_AUTO, NTB_WIDTH_AUTO);
1076 ntb_link_event(ndev);
1081 ntb_clear_ctx(ndev);
1088 mw = &nt->mw_vec[i];
1096 static void ntb_transport_free(struct ntb_client *self, struct ntb_dev *ndev)
1098 struct ntb_transport_ctx *nt = ndev->ctx;
1099 struct ntb_transport_qp *qp;
1100 u64 qp_bitmap_alloc;
1103 ntb_transport_link_cleanup(nt);
1104 cancel_work_sync(&nt->link_cleanup);
1105 cancel_delayed_work_sync(&nt->link_work);
1107 qp_bitmap_alloc = nt->qp_bitmap & ~nt->qp_bitmap_free;
1109 /* verify that all the qp's are freed */
1110 for (i = 0; i < nt->qp_count; i++) {
1111 qp = &nt->qp_vec[i];
1112 if (qp_bitmap_alloc & BIT_ULL(i))
1113 ntb_transport_free_queue(qp);
1114 debugfs_remove_recursive(qp->debugfs_dir);
1117 ntb_link_disable(ndev);
1118 ntb_clear_ctx(ndev);
1122 for (i = nt->mw_count; i--; ) {
1124 iounmap(nt->mw_vec[i].vbase);
1132 static void ntb_complete_rxc(struct ntb_transport_qp *qp)
1134 struct ntb_queue_entry *entry;
1137 unsigned long irqflags;
1139 spin_lock_irqsave(&qp->ntb_rx_q_lock, irqflags);
1141 while (!list_empty(&qp->rx_post_q)) {
1142 entry = list_first_entry(&qp->rx_post_q,
1143 struct ntb_queue_entry, entry);
1144 if (!(entry->flags & DESC_DONE_FLAG))
1147 entry->rx_hdr->flags = 0;
1148 iowrite32(entry->index, &qp->rx_info->entry);
1150 cb_data = entry->cb_data;
1153 list_move_tail(&entry->entry, &qp->rx_free_q);
1155 spin_unlock_irqrestore(&qp->ntb_rx_q_lock, irqflags);
1157 if (qp->rx_handler && qp->client_ready)
1158 qp->rx_handler(qp, qp->cb_data, cb_data, len);
1160 spin_lock_irqsave(&qp->ntb_rx_q_lock, irqflags);
1163 spin_unlock_irqrestore(&qp->ntb_rx_q_lock, irqflags);
1166 static void ntb_rx_copy_callback(void *data)
1168 struct ntb_queue_entry *entry = data;
1170 entry->flags |= DESC_DONE_FLAG;
1172 ntb_complete_rxc(entry->qp);
1175 static void ntb_memcpy_rx(struct ntb_queue_entry *entry, void *offset)
1177 void *buf = entry->buf;
1178 size_t len = entry->len;
1180 memcpy(buf, offset, len);
1182 /* Ensure that the data is fully copied out before clearing the flag */
1185 ntb_rx_copy_callback(entry);
1188 static void ntb_async_rx(struct ntb_queue_entry *entry, void *offset)
1190 struct dma_async_tx_descriptor *txd;
1191 struct ntb_transport_qp *qp = entry->qp;
1192 struct dma_chan *chan = qp->dma_chan;
1193 struct dma_device *device;
1194 size_t pay_off, buff_off, len;
1195 struct dmaengine_unmap_data *unmap;
1196 dma_cookie_t cookie;
1197 void *buf = entry->buf;
1204 if (len < copy_bytes)
1207 device = chan->device;
1208 pay_off = (size_t)offset & ~PAGE_MASK;
1209 buff_off = (size_t)buf & ~PAGE_MASK;
1211 if (!is_dma_copy_aligned(device, pay_off, buff_off, len))
1214 unmap = dmaengine_get_unmap_data(device->dev, 2, GFP_NOWAIT);
1219 unmap->addr[0] = dma_map_page(device->dev, virt_to_page(offset),
1220 pay_off, len, DMA_TO_DEVICE);
1221 if (dma_mapping_error(device->dev, unmap->addr[0]))
1226 unmap->addr[1] = dma_map_page(device->dev, virt_to_page(buf),
1227 buff_off, len, DMA_FROM_DEVICE);
1228 if (dma_mapping_error(device->dev, unmap->addr[1]))
1231 unmap->from_cnt = 1;
1233 txd = device->device_prep_dma_memcpy(chan, unmap->addr[1],
1234 unmap->addr[0], len,
1235 DMA_PREP_INTERRUPT);
1239 txd->callback = ntb_rx_copy_callback;
1240 txd->callback_param = entry;
1241 dma_set_unmap(txd, unmap);
1243 cookie = dmaengine_submit(txd);
1244 if (dma_submit_error(cookie))
1247 dmaengine_unmap_put(unmap);
1249 qp->last_cookie = cookie;
1256 dmaengine_unmap_put(unmap);
1258 dmaengine_unmap_put(unmap);
1260 /* If the callbacks come out of order, the writing of the index to the
1261 * last completed will be out of order. This may result in the
1262 * receive stalling forever.
1264 dma_sync_wait(chan, qp->last_cookie);
1266 ntb_memcpy_rx(entry, offset);
1270 static int ntb_process_rxc(struct ntb_transport_qp *qp)
1272 struct ntb_payload_header *hdr;
1273 struct ntb_queue_entry *entry;
1276 offset = qp->rx_buff + qp->rx_max_frame * qp->rx_index;
1277 hdr = offset + qp->rx_max_frame - sizeof(struct ntb_payload_header);
1279 dev_dbg(&qp->ndev->pdev->dev, "qp %d: RX ver %u len %d flags %x\n",
1280 qp->qp_num, hdr->ver, hdr->len, hdr->flags);
1282 if (!(hdr->flags & DESC_DONE_FLAG)) {
1283 dev_dbg(&qp->ndev->pdev->dev, "done flag not set\n");
1284 qp->rx_ring_empty++;
1288 if (hdr->flags & LINK_DOWN_FLAG) {
1289 dev_dbg(&qp->ndev->pdev->dev, "link down flag set\n");
1290 ntb_qp_link_down(qp);
1295 if (hdr->ver != (u32)qp->rx_pkts) {
1296 dev_dbg(&qp->ndev->pdev->dev,
1297 "version mismatch, expected %llu - got %u\n",
1298 qp->rx_pkts, hdr->ver);
1303 entry = ntb_list_mv(&qp->ntb_rx_q_lock, &qp->rx_pend_q, &qp->rx_post_q);
1305 dev_dbg(&qp->ndev->pdev->dev, "no receive buffer\n");
1306 qp->rx_err_no_buf++;
1310 entry->rx_hdr = hdr;
1311 entry->index = qp->rx_index;
1313 if (hdr->len > entry->len) {
1314 dev_dbg(&qp->ndev->pdev->dev,
1315 "receive buffer overflow! Wanted %d got %d\n",
1316 hdr->len, entry->len);
1320 entry->flags |= DESC_DONE_FLAG;
1322 ntb_complete_rxc(qp);
1324 dev_dbg(&qp->ndev->pdev->dev,
1325 "RX OK index %u ver %u size %d into buf size %d\n",
1326 qp->rx_index, hdr->ver, hdr->len, entry->len);
1328 qp->rx_bytes += hdr->len;
1331 entry->len = hdr->len;
1333 ntb_async_rx(entry, offset);
1337 qp->rx_index %= qp->rx_max_entry;
1342 static void ntb_transport_rxc_db(unsigned long data)
1344 struct ntb_transport_qp *qp = (void *)data;
1347 dev_dbg(&qp->ndev->pdev->dev, "%s: doorbell %d received\n",
1348 __func__, qp->qp_num);
1350 /* Limit the number of packets processed in a single interrupt to
1351 * provide fairness to others
1353 for (i = 0; i < qp->rx_max_entry; i++) {
1354 rc = ntb_process_rxc(qp);
1359 if (i && qp->dma_chan)
1360 dma_async_issue_pending(qp->dma_chan);
1362 if (i == qp->rx_max_entry) {
1363 /* there is more work to do */
1364 tasklet_schedule(&qp->rxc_db_work);
1365 } else if (ntb_db_read(qp->ndev) & BIT_ULL(qp->qp_num)) {
1366 /* the doorbell bit is set: clear it */
1367 ntb_db_clear(qp->ndev, BIT_ULL(qp->qp_num));
1368 /* ntb_db_read ensures ntb_db_clear write is committed */
1369 ntb_db_read(qp->ndev);
1371 /* an interrupt may have arrived between finishing
1372 * ntb_process_rxc and clearing the doorbell bit:
1373 * there might be some more work to do.
1375 tasklet_schedule(&qp->rxc_db_work);
1379 static void ntb_tx_copy_callback(void *data)
1381 struct ntb_queue_entry *entry = data;
1382 struct ntb_transport_qp *qp = entry->qp;
1383 struct ntb_payload_header __iomem *hdr = entry->tx_hdr;
1385 iowrite32(entry->flags | DESC_DONE_FLAG, &hdr->flags);
1387 ntb_peer_db_set(qp->ndev, BIT_ULL(qp->qp_num));
1389 /* The entry length can only be zero if the packet is intended to be a
1390 * "link down" or similar. Since no payload is being sent in these
1391 * cases, there is nothing to add to the completion queue.
1393 if (entry->len > 0) {
1394 qp->tx_bytes += entry->len;
1397 qp->tx_handler(qp, qp->cb_data, entry->cb_data,
1401 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry, &qp->tx_free_q);
1404 static void ntb_memcpy_tx(struct ntb_queue_entry *entry, void __iomem *offset)
1406 #ifdef ARCH_HAS_NOCACHE_UACCESS
1408 * Using non-temporal mov to improve performance on non-cached
1409 * writes, even though we aren't actually copying from user space.
1411 __copy_from_user_inatomic_nocache(offset, entry->buf, entry->len);
1413 memcpy_toio(offset, entry->buf, entry->len);
1416 /* Ensure that the data is fully copied out before setting the flags */
1419 ntb_tx_copy_callback(entry);
1422 static void ntb_async_tx(struct ntb_transport_qp *qp,
1423 struct ntb_queue_entry *entry)
1425 struct ntb_payload_header __iomem *hdr;
1426 struct dma_async_tx_descriptor *txd;
1427 struct dma_chan *chan = qp->dma_chan;
1428 struct dma_device *device;
1429 size_t dest_off, buff_off;
1430 struct dmaengine_unmap_data *unmap;
1432 dma_cookie_t cookie;
1433 void __iomem *offset;
1434 size_t len = entry->len;
1435 void *buf = entry->buf;
1437 offset = qp->tx_mw + qp->tx_max_frame * qp->tx_index;
1438 hdr = offset + qp->tx_max_frame - sizeof(struct ntb_payload_header);
1439 entry->tx_hdr = hdr;
1441 iowrite32(entry->len, &hdr->len);
1442 iowrite32((u32)qp->tx_pkts, &hdr->ver);
1447 if (len < copy_bytes)
1450 device = chan->device;
1451 dest = qp->tx_mw_phys + qp->tx_max_frame * qp->tx_index;
1452 buff_off = (size_t)buf & ~PAGE_MASK;
1453 dest_off = (size_t)dest & ~PAGE_MASK;
1455 if (!is_dma_copy_aligned(device, buff_off, dest_off, len))
1458 unmap = dmaengine_get_unmap_data(device->dev, 1, GFP_NOWAIT);
1463 unmap->addr[0] = dma_map_page(device->dev, virt_to_page(buf),
1464 buff_off, len, DMA_TO_DEVICE);
1465 if (dma_mapping_error(device->dev, unmap->addr[0]))
1470 txd = device->device_prep_dma_memcpy(chan, dest, unmap->addr[0], len,
1471 DMA_PREP_INTERRUPT);
1475 txd->callback = ntb_tx_copy_callback;
1476 txd->callback_param = entry;
1477 dma_set_unmap(txd, unmap);
1479 cookie = dmaengine_submit(txd);
1480 if (dma_submit_error(cookie))
1483 dmaengine_unmap_put(unmap);
1485 dma_async_issue_pending(chan);
1490 dmaengine_unmap_put(unmap);
1492 dmaengine_unmap_put(unmap);
1494 ntb_memcpy_tx(entry, offset);
1498 static int ntb_process_tx(struct ntb_transport_qp *qp,
1499 struct ntb_queue_entry *entry)
1501 if (qp->tx_index == qp->remote_rx_info->entry) {
1506 if (entry->len > qp->tx_max_frame - sizeof(struct ntb_payload_header)) {
1508 qp->tx_handler(qp->cb_data, qp, NULL, -EIO);
1510 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
1515 ntb_async_tx(qp, entry);
1518 qp->tx_index %= qp->tx_max_entry;
1525 static void ntb_send_link_down(struct ntb_transport_qp *qp)
1527 struct pci_dev *pdev = qp->ndev->pdev;
1528 struct ntb_queue_entry *entry;
1531 if (!qp->link_is_up)
1534 dev_info(&pdev->dev, "qp %d: Send Link Down\n", qp->qp_num);
1536 for (i = 0; i < NTB_LINK_DOWN_TIMEOUT; i++) {
1537 entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
1546 entry->cb_data = NULL;
1549 entry->flags = LINK_DOWN_FLAG;
1551 rc = ntb_process_tx(qp, entry);
1553 dev_err(&pdev->dev, "ntb: QP%d unable to send linkdown msg\n",
1556 ntb_qp_link_down_reset(qp);
1559 static bool ntb_dma_filter_fn(struct dma_chan *chan, void *node)
1561 return dev_to_node(&chan->dev->device) == (int)(unsigned long)node;
1565 * ntb_transport_create_queue - Create a new NTB transport layer queue
1566 * @rx_handler: receive callback function
1567 * @tx_handler: transmit callback function
1568 * @event_handler: event callback function
1570 * Create a new NTB transport layer queue and provide the queue with a callback
1571 * routine for both transmit and receive. The receive callback routine will be
1572 * used to pass up data when the transport has received it on the queue. The
1573 * transmit callback routine will be called when the transport has completed the
1574 * transmission of the data on the queue and the data is ready to be freed.
1576 * RETURNS: pointer to newly created ntb_queue, NULL on error.
1578 struct ntb_transport_qp *
1579 ntb_transport_create_queue(void *data, struct device *client_dev,
1580 const struct ntb_queue_handlers *handlers)
1582 struct ntb_dev *ndev;
1583 struct pci_dev *pdev;
1584 struct ntb_transport_ctx *nt;
1585 struct ntb_queue_entry *entry;
1586 struct ntb_transport_qp *qp;
1588 unsigned int free_queue;
1589 dma_cap_mask_t dma_mask;
1593 ndev = dev_ntb(client_dev->parent);
1597 node = dev_to_node(&ndev->dev);
1599 free_queue = ffs(nt->qp_bitmap);
1603 /* decrement free_queue to make it zero based */
1606 qp = &nt->qp_vec[free_queue];
1607 qp_bit = BIT_ULL(qp->qp_num);
1609 nt->qp_bitmap_free &= ~qp_bit;
1612 qp->rx_handler = handlers->rx_handler;
1613 qp->tx_handler = handlers->tx_handler;
1614 qp->event_handler = handlers->event_handler;
1616 dma_cap_zero(dma_mask);
1617 dma_cap_set(DMA_MEMCPY, dma_mask);
1620 qp->dma_chan = dma_request_channel(dma_mask, ntb_dma_filter_fn,
1621 (void *)(unsigned long)node);
1623 dev_info(&pdev->dev, "Unable to allocate DMA channel\n");
1625 qp->dma_chan = NULL;
1627 dev_dbg(&pdev->dev, "Using %s memcpy\n", qp->dma_chan ? "DMA" : "CPU");
1629 for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) {
1630 entry = kzalloc_node(sizeof(*entry), GFP_ATOMIC, node);
1635 ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry,
1639 for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) {
1640 entry = kzalloc_node(sizeof(*entry), GFP_ATOMIC, node);
1645 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
1649 ntb_db_clear(qp->ndev, qp_bit);
1650 ntb_db_clear_mask(qp->ndev, qp_bit);
1652 dev_info(&pdev->dev, "NTB Transport QP %d created\n", qp->qp_num);
1657 while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
1660 while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q)))
1663 dma_release_channel(qp->dma_chan);
1664 nt->qp_bitmap_free |= qp_bit;
1668 EXPORT_SYMBOL_GPL(ntb_transport_create_queue);
1671 * ntb_transport_free_queue - Frees NTB transport queue
1672 * @qp: NTB queue to be freed
1674 * Frees NTB transport queue
1676 void ntb_transport_free_queue(struct ntb_transport_qp *qp)
1678 struct ntb_transport_ctx *nt = qp->transport;
1679 struct pci_dev *pdev;
1680 struct ntb_queue_entry *entry;
1686 pdev = qp->ndev->pdev;
1689 struct dma_chan *chan = qp->dma_chan;
1690 /* Putting the dma_chan to NULL will force any new traffic to be
1691 * processed by the CPU instead of the DAM engine
1693 qp->dma_chan = NULL;
1695 /* Try to be nice and wait for any queued DMA engine
1696 * transactions to process before smashing it with a rock
1698 dma_sync_wait(chan, qp->last_cookie);
1699 dmaengine_terminate_all(chan);
1700 dma_release_channel(chan);
1703 qp_bit = BIT_ULL(qp->qp_num);
1705 ntb_db_set_mask(qp->ndev, qp_bit);
1706 tasklet_disable(&qp->rxc_db_work);
1708 cancel_delayed_work_sync(&qp->link_work);
1711 qp->rx_handler = NULL;
1712 qp->tx_handler = NULL;
1713 qp->event_handler = NULL;
1715 while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q)))
1718 while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_pend_q))) {
1719 dev_warn(&pdev->dev, "Freeing item from non-empty rx_pend_q\n");
1723 while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_post_q))) {
1724 dev_warn(&pdev->dev, "Freeing item from non-empty rx_post_q\n");
1728 while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
1731 nt->qp_bitmap_free |= qp_bit;
1733 dev_info(&pdev->dev, "NTB Transport QP %d freed\n", qp->qp_num);
1735 EXPORT_SYMBOL_GPL(ntb_transport_free_queue);
1738 * ntb_transport_rx_remove - Dequeues enqueued rx packet
1739 * @qp: NTB queue to be freed
1740 * @len: pointer to variable to write enqueued buffers length
1742 * Dequeues unused buffers from receive queue. Should only be used during
1745 * RETURNS: NULL error value on error, or void* for success.
1747 void *ntb_transport_rx_remove(struct ntb_transport_qp *qp, unsigned int *len)
1749 struct ntb_queue_entry *entry;
1752 if (!qp || qp->client_ready)
1755 entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_pend_q);
1759 buf = entry->cb_data;
1762 ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry, &qp->rx_free_q);
1766 EXPORT_SYMBOL_GPL(ntb_transport_rx_remove);
1769 * ntb_transport_rx_enqueue - Enqueue a new NTB queue entry
1770 * @qp: NTB transport layer queue the entry is to be enqueued on
1771 * @cb: per buffer pointer for callback function to use
1772 * @data: pointer to data buffer that incoming packets will be copied into
1773 * @len: length of the data buffer
1775 * Enqueue a new receive buffer onto the transport queue into which a NTB
1776 * payload can be received into.
1778 * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
1780 int ntb_transport_rx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
1783 struct ntb_queue_entry *entry;
1788 entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q);
1792 entry->cb_data = cb;
1797 ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry, &qp->rx_pend_q);
1799 tasklet_schedule(&qp->rxc_db_work);
1803 EXPORT_SYMBOL_GPL(ntb_transport_rx_enqueue);
1806 * ntb_transport_tx_enqueue - Enqueue a new NTB queue entry
1807 * @qp: NTB transport layer queue the entry is to be enqueued on
1808 * @cb: per buffer pointer for callback function to use
1809 * @data: pointer to data buffer that will be sent
1810 * @len: length of the data buffer
1812 * Enqueue a new transmit buffer onto the transport queue from which a NTB
1813 * payload will be transmitted. This assumes that a lock is being held to
1814 * serialize access to the qp.
1816 * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
1818 int ntb_transport_tx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
1821 struct ntb_queue_entry *entry;
1824 if (!qp || !qp->link_is_up || !len)
1827 entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
1829 qp->tx_err_no_buf++;
1833 entry->cb_data = cb;
1838 rc = ntb_process_tx(qp, entry);
1840 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
1845 EXPORT_SYMBOL_GPL(ntb_transport_tx_enqueue);
1848 * ntb_transport_link_up - Notify NTB transport of client readiness to use queue
1849 * @qp: NTB transport layer queue to be enabled
1851 * Notify NTB transport layer of client readiness to use queue
1853 void ntb_transport_link_up(struct ntb_transport_qp *qp)
1858 qp->client_ready = true;
1860 if (qp->transport->link_is_up)
1861 schedule_delayed_work(&qp->link_work, 0);
1863 EXPORT_SYMBOL_GPL(ntb_transport_link_up);
1866 * ntb_transport_link_down - Notify NTB transport to no longer enqueue data
1867 * @qp: NTB transport layer queue to be disabled
1869 * Notify NTB transport layer of client's desire to no longer receive data on
1870 * transport queue specified. It is the client's responsibility to ensure all
1871 * entries on queue are purged or otherwise handled appropriately.
1873 void ntb_transport_link_down(struct ntb_transport_qp *qp)
1875 struct pci_dev *pdev;
1881 pdev = qp->ndev->pdev;
1882 qp->client_ready = false;
1884 val = ntb_spad_read(qp->ndev, QP_LINKS);
1886 ntb_peer_spad_write(qp->ndev, QP_LINKS,
1887 val & ~BIT(qp->qp_num));
1890 ntb_send_link_down(qp);
1892 cancel_delayed_work_sync(&qp->link_work);
1894 EXPORT_SYMBOL_GPL(ntb_transport_link_down);
1897 * ntb_transport_link_query - Query transport link state
1898 * @qp: NTB transport layer queue to be queried
1900 * Query connectivity to the remote system of the NTB transport queue
1902 * RETURNS: true for link up or false for link down
1904 bool ntb_transport_link_query(struct ntb_transport_qp *qp)
1909 return qp->link_is_up;
1911 EXPORT_SYMBOL_GPL(ntb_transport_link_query);
1914 * ntb_transport_qp_num - Query the qp number
1915 * @qp: NTB transport layer queue to be queried
1917 * Query qp number of the NTB transport queue
1919 * RETURNS: a zero based number specifying the qp number
1921 unsigned char ntb_transport_qp_num(struct ntb_transport_qp *qp)
1928 EXPORT_SYMBOL_GPL(ntb_transport_qp_num);
1931 * ntb_transport_max_size - Query the max payload size of a qp
1932 * @qp: NTB transport layer queue to be queried
1934 * Query the maximum payload size permissible on the given qp
1936 * RETURNS: the max payload size of a qp
1938 unsigned int ntb_transport_max_size(struct ntb_transport_qp *qp)
1946 return qp->tx_max_frame - sizeof(struct ntb_payload_header);
1948 /* If DMA engine usage is possible, try to find the max size for that */
1949 max = qp->tx_max_frame - sizeof(struct ntb_payload_header);
1950 max -= max % (1 << qp->dma_chan->device->copy_align);
1954 EXPORT_SYMBOL_GPL(ntb_transport_max_size);
1956 static void ntb_transport_doorbell_callback(void *data, int vector)
1958 struct ntb_transport_ctx *nt = data;
1959 struct ntb_transport_qp *qp;
1961 unsigned int qp_num;
1963 db_bits = (nt->qp_bitmap & ~nt->qp_bitmap_free &
1964 ntb_db_vector_mask(nt->ndev, vector));
1967 qp_num = __ffs(db_bits);
1968 qp = &nt->qp_vec[qp_num];
1970 tasklet_schedule(&qp->rxc_db_work);
1972 db_bits &= ~BIT_ULL(qp_num);
1976 static const struct ntb_ctx_ops ntb_transport_ops = {
1977 .link_event = ntb_transport_event_callback,
1978 .db_event = ntb_transport_doorbell_callback,
1981 static struct ntb_client ntb_transport_client = {
1983 .probe = ntb_transport_probe,
1984 .remove = ntb_transport_free,
1988 static int __init ntb_transport_init(void)
1992 pr_info("%s, version %s\n", NTB_TRANSPORT_DESC, NTB_TRANSPORT_VER);
1994 if (debugfs_initialized())
1995 nt_debugfs_dir = debugfs_create_dir(KBUILD_MODNAME, NULL);
1997 rc = bus_register(&ntb_transport_bus);
2001 rc = ntb_register_client(&ntb_transport_client);
2008 bus_unregister(&ntb_transport_bus);
2010 debugfs_remove_recursive(nt_debugfs_dir);
2013 module_init(ntb_transport_init);
2015 static void __exit ntb_transport_exit(void)
2017 debugfs_remove_recursive(nt_debugfs_dir);
2019 ntb_unregister_client(&ntb_transport_client);
2020 bus_unregister(&ntb_transport_bus);
2022 module_exit(ntb_transport_exit);