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 *tx_dma_chan;
123 struct dma_chan *rx_dma_chan;
129 u8 qp_num; /* Only 64 QP's are allowed. 0-63 */
132 struct ntb_rx_info __iomem *rx_info;
133 struct ntb_rx_info *remote_rx_info;
135 void (*tx_handler)(struct ntb_transport_qp *qp, void *qp_data,
136 void *data, int len);
137 struct list_head tx_free_q;
138 spinlock_t ntb_tx_free_q_lock;
140 dma_addr_t tx_mw_phys;
141 unsigned int tx_index;
142 unsigned int tx_max_entry;
143 unsigned int tx_max_frame;
145 void (*rx_handler)(struct ntb_transport_qp *qp, void *qp_data,
146 void *data, int len);
147 struct list_head rx_post_q;
148 struct list_head rx_pend_q;
149 struct list_head rx_free_q;
150 /* ntb_rx_q_lock: synchronize access to rx_XXXX_q */
151 spinlock_t ntb_rx_q_lock;
153 unsigned int rx_index;
154 unsigned int rx_max_entry;
155 unsigned int rx_max_frame;
156 unsigned int rx_alloc_entry;
157 dma_cookie_t last_cookie;
158 struct tasklet_struct rxc_db_work;
160 void (*event_handler)(void *data, int status);
161 struct delayed_work link_work;
162 struct work_struct link_cleanup;
164 struct dentry *debugfs_dir;
165 struct dentry *debugfs_stats;
186 struct ntb_transport_mw {
187 phys_addr_t phys_addr;
188 resource_size_t phys_size;
189 resource_size_t xlat_align;
190 resource_size_t xlat_align_size;
198 struct ntb_transport_client_dev {
199 struct list_head entry;
200 struct ntb_transport_ctx *nt;
204 struct ntb_transport_ctx {
205 struct list_head entry;
206 struct list_head client_devs;
208 struct ntb_dev *ndev;
210 struct ntb_transport_mw *mw_vec;
211 struct ntb_transport_qp *qp_vec;
212 unsigned int mw_count;
213 unsigned int qp_count;
218 struct delayed_work link_work;
219 struct work_struct link_cleanup;
221 struct dentry *debugfs_node_dir;
225 DESC_DONE_FLAG = BIT(0),
226 LINK_DOWN_FLAG = BIT(1),
229 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
256 #define DMA_RETRIES 20
257 #define DMA_OUT_RESOURCE_TO 50
259 static void ntb_transport_rxc_db(unsigned long data);
260 static const struct ntb_ctx_ops ntb_transport_ops;
261 static struct ntb_client ntb_transport_client;
263 static int ntb_transport_bus_match(struct device *dev,
264 struct device_driver *drv)
266 return !strncmp(dev_name(dev), drv->name, strlen(drv->name));
269 static int ntb_transport_bus_probe(struct device *dev)
271 const struct ntb_transport_client *client;
276 client = drv_client(dev->driver);
277 rc = client->probe(dev);
284 static int ntb_transport_bus_remove(struct device *dev)
286 const struct ntb_transport_client *client;
288 client = drv_client(dev->driver);
296 static struct bus_type ntb_transport_bus = {
297 .name = "ntb_transport",
298 .match = ntb_transport_bus_match,
299 .probe = ntb_transport_bus_probe,
300 .remove = ntb_transport_bus_remove,
303 static LIST_HEAD(ntb_transport_list);
305 static int ntb_bus_init(struct ntb_transport_ctx *nt)
307 list_add_tail(&nt->entry, &ntb_transport_list);
311 static void ntb_bus_remove(struct ntb_transport_ctx *nt)
313 struct ntb_transport_client_dev *client_dev, *cd;
315 list_for_each_entry_safe(client_dev, cd, &nt->client_devs, entry) {
316 dev_err(client_dev->dev.parent, "%s still attached to bus, removing\n",
317 dev_name(&client_dev->dev));
318 list_del(&client_dev->entry);
319 device_unregister(&client_dev->dev);
322 list_del(&nt->entry);
325 static void ntb_transport_client_release(struct device *dev)
327 struct ntb_transport_client_dev *client_dev;
329 client_dev = dev_client_dev(dev);
334 * ntb_transport_unregister_client_dev - Unregister NTB client device
335 * @device_name: Name of NTB client device
337 * Unregister an NTB client device with the NTB transport layer
339 void ntb_transport_unregister_client_dev(char *device_name)
341 struct ntb_transport_client_dev *client, *cd;
342 struct ntb_transport_ctx *nt;
344 list_for_each_entry(nt, &ntb_transport_list, entry)
345 list_for_each_entry_safe(client, cd, &nt->client_devs, entry)
346 if (!strncmp(dev_name(&client->dev), device_name,
347 strlen(device_name))) {
348 list_del(&client->entry);
349 device_unregister(&client->dev);
352 EXPORT_SYMBOL_GPL(ntb_transport_unregister_client_dev);
355 * ntb_transport_register_client_dev - Register NTB client device
356 * @device_name: Name of NTB client device
358 * Register an NTB client device with the NTB transport layer
360 int ntb_transport_register_client_dev(char *device_name)
362 struct ntb_transport_client_dev *client_dev;
363 struct ntb_transport_ctx *nt;
367 if (list_empty(&ntb_transport_list))
370 list_for_each_entry(nt, &ntb_transport_list, entry) {
373 node = dev_to_node(&nt->ndev->dev);
375 client_dev = kzalloc_node(sizeof(*client_dev),
382 dev = &client_dev->dev;
384 /* setup and register client devices */
385 dev_set_name(dev, "%s%d", device_name, i);
386 dev->bus = &ntb_transport_bus;
387 dev->release = ntb_transport_client_release;
388 dev->parent = &nt->ndev->dev;
390 rc = device_register(dev);
396 list_add_tail(&client_dev->entry, &nt->client_devs);
403 ntb_transport_unregister_client_dev(device_name);
407 EXPORT_SYMBOL_GPL(ntb_transport_register_client_dev);
410 * ntb_transport_register_client - Register NTB client driver
411 * @drv: NTB client driver to be registered
413 * Register an NTB client driver with the NTB transport layer
415 * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
417 int ntb_transport_register_client(struct ntb_transport_client *drv)
419 drv->driver.bus = &ntb_transport_bus;
421 if (list_empty(&ntb_transport_list))
424 return driver_register(&drv->driver);
426 EXPORT_SYMBOL_GPL(ntb_transport_register_client);
429 * ntb_transport_unregister_client - Unregister NTB client driver
430 * @drv: NTB client driver to be unregistered
432 * Unregister an NTB client driver with the NTB transport layer
434 * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
436 void ntb_transport_unregister_client(struct ntb_transport_client *drv)
438 driver_unregister(&drv->driver);
440 EXPORT_SYMBOL_GPL(ntb_transport_unregister_client);
442 static ssize_t debugfs_read(struct file *filp, char __user *ubuf, size_t count,
445 struct ntb_transport_qp *qp;
447 ssize_t ret, out_offset, out_count;
449 qp = filp->private_data;
451 if (!qp || !qp->link_is_up)
456 buf = kmalloc(out_count, GFP_KERNEL);
461 out_offset += snprintf(buf + out_offset, out_count - out_offset,
462 "\nNTB QP stats:\n\n");
463 out_offset += snprintf(buf + out_offset, out_count - out_offset,
464 "rx_bytes - \t%llu\n", qp->rx_bytes);
465 out_offset += snprintf(buf + out_offset, out_count - out_offset,
466 "rx_pkts - \t%llu\n", qp->rx_pkts);
467 out_offset += snprintf(buf + out_offset, out_count - out_offset,
468 "rx_memcpy - \t%llu\n", qp->rx_memcpy);
469 out_offset += snprintf(buf + out_offset, out_count - out_offset,
470 "rx_async - \t%llu\n", qp->rx_async);
471 out_offset += snprintf(buf + out_offset, out_count - out_offset,
472 "rx_ring_empty - %llu\n", qp->rx_ring_empty);
473 out_offset += snprintf(buf + out_offset, out_count - out_offset,
474 "rx_err_no_buf - %llu\n", qp->rx_err_no_buf);
475 out_offset += snprintf(buf + out_offset, out_count - out_offset,
476 "rx_err_oflow - \t%llu\n", qp->rx_err_oflow);
477 out_offset += snprintf(buf + out_offset, out_count - out_offset,
478 "rx_err_ver - \t%llu\n", qp->rx_err_ver);
479 out_offset += snprintf(buf + out_offset, out_count - out_offset,
480 "rx_buff - \t0x%p\n", qp->rx_buff);
481 out_offset += snprintf(buf + out_offset, out_count - out_offset,
482 "rx_index - \t%u\n", qp->rx_index);
483 out_offset += snprintf(buf + out_offset, out_count - out_offset,
484 "rx_max_entry - \t%u\n", qp->rx_max_entry);
485 out_offset += snprintf(buf + out_offset, out_count - out_offset,
486 "rx_alloc_entry - \t%u\n\n", qp->rx_alloc_entry);
488 out_offset += snprintf(buf + out_offset, out_count - out_offset,
489 "tx_bytes - \t%llu\n", qp->tx_bytes);
490 out_offset += snprintf(buf + out_offset, out_count - out_offset,
491 "tx_pkts - \t%llu\n", qp->tx_pkts);
492 out_offset += snprintf(buf + out_offset, out_count - out_offset,
493 "tx_memcpy - \t%llu\n", qp->tx_memcpy);
494 out_offset += snprintf(buf + out_offset, out_count - out_offset,
495 "tx_async - \t%llu\n", qp->tx_async);
496 out_offset += snprintf(buf + out_offset, out_count - out_offset,
497 "tx_ring_full - \t%llu\n", qp->tx_ring_full);
498 out_offset += snprintf(buf + out_offset, out_count - out_offset,
499 "tx_err_no_buf - %llu\n", qp->tx_err_no_buf);
500 out_offset += snprintf(buf + out_offset, out_count - out_offset,
501 "tx_mw - \t0x%p\n", qp->tx_mw);
502 out_offset += snprintf(buf + out_offset, out_count - out_offset,
503 "tx_index (H) - \t%u\n", qp->tx_index);
504 out_offset += snprintf(buf + out_offset, out_count - out_offset,
506 qp->remote_rx_info->entry);
507 out_offset += snprintf(buf + out_offset, out_count - out_offset,
508 "tx_max_entry - \t%u\n", qp->tx_max_entry);
509 out_offset += snprintf(buf + out_offset, out_count - out_offset,
511 ntb_transport_tx_free_entry(qp));
512 out_offset += snprintf(buf + out_offset, out_count - out_offset,
513 "DMA tx prep err - \t%llu\n",
514 qp->dma_tx_prep_err);
515 out_offset += snprintf(buf + out_offset, out_count - out_offset,
516 "DMA rx prep err - \t%llu\n",
517 qp->dma_rx_prep_err);
519 out_offset += snprintf(buf + out_offset, out_count - out_offset,
521 out_offset += snprintf(buf + out_offset, out_count - out_offset,
522 "Using TX DMA - \t%s\n",
523 qp->tx_dma_chan ? "Yes" : "No");
524 out_offset += snprintf(buf + out_offset, out_count - out_offset,
525 "Using RX DMA - \t%s\n",
526 qp->rx_dma_chan ? "Yes" : "No");
527 out_offset += snprintf(buf + out_offset, out_count - out_offset,
529 qp->link_is_up ? "Up" : "Down");
530 out_offset += snprintf(buf + out_offset, out_count - out_offset,
533 if (out_offset > out_count)
534 out_offset = out_count;
536 ret = simple_read_from_buffer(ubuf, count, offp, buf, out_offset);
541 static const struct file_operations ntb_qp_debugfs_stats = {
542 .owner = THIS_MODULE,
544 .read = debugfs_read,
547 static void ntb_list_add(spinlock_t *lock, struct list_head *entry,
548 struct list_head *list)
552 spin_lock_irqsave(lock, flags);
553 list_add_tail(entry, list);
554 spin_unlock_irqrestore(lock, flags);
557 static struct ntb_queue_entry *ntb_list_rm(spinlock_t *lock,
558 struct list_head *list)
560 struct ntb_queue_entry *entry;
563 spin_lock_irqsave(lock, flags);
564 if (list_empty(list)) {
568 entry = list_first_entry(list, struct ntb_queue_entry, entry);
569 list_del(&entry->entry);
572 spin_unlock_irqrestore(lock, flags);
577 static struct ntb_queue_entry *ntb_list_mv(spinlock_t *lock,
578 struct list_head *list,
579 struct list_head *to_list)
581 struct ntb_queue_entry *entry;
584 spin_lock_irqsave(lock, flags);
586 if (list_empty(list)) {
589 entry = list_first_entry(list, struct ntb_queue_entry, entry);
590 list_move_tail(&entry->entry, to_list);
593 spin_unlock_irqrestore(lock, flags);
598 static int ntb_transport_setup_qp_mw(struct ntb_transport_ctx *nt,
601 struct ntb_transport_qp *qp = &nt->qp_vec[qp_num];
602 struct ntb_transport_mw *mw;
603 struct ntb_dev *ndev = nt->ndev;
604 struct ntb_queue_entry *entry;
605 unsigned int rx_size, num_qps_mw;
606 unsigned int mw_num, mw_count, qp_count;
610 mw_count = nt->mw_count;
611 qp_count = nt->qp_count;
613 mw_num = QP_TO_MW(nt, qp_num);
614 mw = &nt->mw_vec[mw_num];
619 if (qp_count % mw_count && mw_num + 1 < qp_count / mw_count)
620 num_qps_mw = qp_count / mw_count + 1;
622 num_qps_mw = qp_count / mw_count;
624 rx_size = (unsigned int)mw->xlat_size / num_qps_mw;
625 qp->rx_buff = mw->virt_addr + rx_size * (qp_num / mw_count);
626 rx_size -= sizeof(struct ntb_rx_info);
628 qp->remote_rx_info = qp->rx_buff + rx_size;
630 /* Due to housekeeping, there must be atleast 2 buffs */
631 qp->rx_max_frame = min(transport_mtu, rx_size / 2);
632 qp->rx_max_entry = rx_size / qp->rx_max_frame;
636 * Checking to see if we have more entries than the default.
637 * We should add additional entries if that is the case so we
638 * can be in sync with the transport frames.
640 node = dev_to_node(&ndev->dev);
641 for (i = qp->rx_alloc_entry; i < qp->rx_max_entry; i++) {
642 entry = kzalloc_node(sizeof(*entry), GFP_ATOMIC, node);
647 ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry,
649 qp->rx_alloc_entry++;
652 qp->remote_rx_info->entry = qp->rx_max_entry - 1;
654 /* setup the hdr offsets with 0's */
655 for (i = 0; i < qp->rx_max_entry; i++) {
656 void *offset = (qp->rx_buff + qp->rx_max_frame * (i + 1) -
657 sizeof(struct ntb_payload_header));
658 memset(offset, 0, sizeof(struct ntb_payload_header));
668 static void ntb_free_mw(struct ntb_transport_ctx *nt, int num_mw)
670 struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
671 struct pci_dev *pdev = nt->ndev->pdev;
676 ntb_mw_clear_trans(nt->ndev, num_mw);
677 dma_free_coherent(&pdev->dev, mw->buff_size,
678 mw->virt_addr, mw->dma_addr);
681 mw->virt_addr = NULL;
684 static int ntb_set_mw(struct ntb_transport_ctx *nt, int num_mw,
685 resource_size_t size)
687 struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
688 struct pci_dev *pdev = nt->ndev->pdev;
689 size_t xlat_size, buff_size;
695 xlat_size = round_up(size, mw->xlat_align_size);
696 buff_size = round_up(size, mw->xlat_align);
698 /* No need to re-setup */
699 if (mw->xlat_size == xlat_size)
703 ntb_free_mw(nt, num_mw);
705 /* Alloc memory for receiving data. Must be aligned */
706 mw->xlat_size = xlat_size;
707 mw->buff_size = buff_size;
709 mw->virt_addr = dma_alloc_coherent(&pdev->dev, buff_size,
710 &mw->dma_addr, GFP_KERNEL);
711 if (!mw->virt_addr) {
714 dev_err(&pdev->dev, "Unable to alloc MW buff of size %zu\n",
720 * we must ensure that the memory address allocated is BAR size
721 * aligned in order for the XLAT register to take the value. This
722 * is a requirement of the hardware. It is recommended to setup CMA
723 * for BAR sizes equal or greater than 4MB.
725 if (!IS_ALIGNED(mw->dma_addr, mw->xlat_align)) {
726 dev_err(&pdev->dev, "DMA memory %pad is not aligned\n",
728 ntb_free_mw(nt, num_mw);
732 /* Notify HW the memory location of the receive buffer */
733 rc = ntb_mw_set_trans(nt->ndev, num_mw, mw->dma_addr, mw->xlat_size);
735 dev_err(&pdev->dev, "Unable to set mw%d translation", num_mw);
736 ntb_free_mw(nt, num_mw);
743 static void ntb_qp_link_down_reset(struct ntb_transport_qp *qp)
745 qp->link_is_up = false;
752 qp->rx_ring_empty = 0;
753 qp->rx_err_no_buf = 0;
754 qp->rx_err_oflow = 0;
760 qp->tx_ring_full = 0;
761 qp->tx_err_no_buf = 0;
764 qp->dma_tx_prep_err = 0;
765 qp->dma_rx_prep_err = 0;
768 static void ntb_qp_link_cleanup(struct ntb_transport_qp *qp)
770 struct ntb_transport_ctx *nt = qp->transport;
771 struct pci_dev *pdev = nt->ndev->pdev;
773 dev_info(&pdev->dev, "qp %d: Link Cleanup\n", qp->qp_num);
775 cancel_delayed_work_sync(&qp->link_work);
776 ntb_qp_link_down_reset(qp);
778 if (qp->event_handler)
779 qp->event_handler(qp->cb_data, qp->link_is_up);
782 static void ntb_qp_link_cleanup_work(struct work_struct *work)
784 struct ntb_transport_qp *qp = container_of(work,
785 struct ntb_transport_qp,
787 struct ntb_transport_ctx *nt = qp->transport;
789 ntb_qp_link_cleanup(qp);
792 schedule_delayed_work(&qp->link_work,
793 msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
796 static void ntb_qp_link_down(struct ntb_transport_qp *qp)
798 schedule_work(&qp->link_cleanup);
801 static void ntb_transport_link_cleanup(struct ntb_transport_ctx *nt)
803 struct ntb_transport_qp *qp;
807 qp_bitmap_alloc = nt->qp_bitmap & ~nt->qp_bitmap_free;
809 /* Pass along the info to any clients */
810 for (i = 0; i < nt->qp_count; i++)
811 if (qp_bitmap_alloc & BIT_ULL(i)) {
813 ntb_qp_link_cleanup(qp);
814 cancel_work_sync(&qp->link_cleanup);
815 cancel_delayed_work_sync(&qp->link_work);
819 cancel_delayed_work_sync(&nt->link_work);
821 /* The scratchpad registers keep the values if the remote side
822 * goes down, blast them now to give them a sane value the next
823 * time they are accessed
825 for (i = 0; i < MAX_SPAD; i++)
826 ntb_spad_write(nt->ndev, i, 0);
829 static void ntb_transport_link_cleanup_work(struct work_struct *work)
831 struct ntb_transport_ctx *nt =
832 container_of(work, struct ntb_transport_ctx, link_cleanup);
834 ntb_transport_link_cleanup(nt);
837 static void ntb_transport_event_callback(void *data)
839 struct ntb_transport_ctx *nt = data;
841 if (ntb_link_is_up(nt->ndev, NULL, NULL) == 1)
842 schedule_delayed_work(&nt->link_work, 0);
844 schedule_work(&nt->link_cleanup);
847 static void ntb_transport_link_work(struct work_struct *work)
849 struct ntb_transport_ctx *nt =
850 container_of(work, struct ntb_transport_ctx, link_work.work);
851 struct ntb_dev *ndev = nt->ndev;
852 struct pci_dev *pdev = ndev->pdev;
853 resource_size_t size;
857 /* send the local info, in the opposite order of the way we read it */
858 for (i = 0; i < nt->mw_count; i++) {
859 size = nt->mw_vec[i].phys_size;
861 if (max_mw_size && size > max_mw_size)
864 spad = MW0_SZ_HIGH + (i * 2);
865 ntb_peer_spad_write(ndev, spad, upper_32_bits(size));
867 spad = MW0_SZ_LOW + (i * 2);
868 ntb_peer_spad_write(ndev, spad, lower_32_bits(size));
871 ntb_peer_spad_write(ndev, NUM_MWS, nt->mw_count);
873 ntb_peer_spad_write(ndev, NUM_QPS, nt->qp_count);
875 ntb_peer_spad_write(ndev, VERSION, NTB_TRANSPORT_VERSION);
877 /* Query the remote side for its info */
878 val = ntb_spad_read(ndev, VERSION);
879 dev_dbg(&pdev->dev, "Remote version = %d\n", val);
880 if (val != NTB_TRANSPORT_VERSION)
883 val = ntb_spad_read(ndev, NUM_QPS);
884 dev_dbg(&pdev->dev, "Remote max number of qps = %d\n", val);
885 if (val != nt->qp_count)
888 val = ntb_spad_read(ndev, NUM_MWS);
889 dev_dbg(&pdev->dev, "Remote number of mws = %d\n", val);
890 if (val != nt->mw_count)
893 for (i = 0; i < nt->mw_count; i++) {
896 val = ntb_spad_read(ndev, MW0_SZ_HIGH + (i * 2));
897 val64 = (u64)val << 32;
899 val = ntb_spad_read(ndev, MW0_SZ_LOW + (i * 2));
902 dev_dbg(&pdev->dev, "Remote MW%d size = %#llx\n", i, val64);
904 rc = ntb_set_mw(nt, i, val64);
909 nt->link_is_up = true;
911 for (i = 0; i < nt->qp_count; i++) {
912 struct ntb_transport_qp *qp = &nt->qp_vec[i];
914 ntb_transport_setup_qp_mw(nt, i);
916 if (qp->client_ready)
917 schedule_delayed_work(&qp->link_work, 0);
923 for (i = 0; i < nt->mw_count; i++)
926 /* if there's an actual failure, we should just bail */
928 ntb_link_disable(ndev);
933 if (ntb_link_is_up(ndev, NULL, NULL) == 1)
934 schedule_delayed_work(&nt->link_work,
935 msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
938 static void ntb_qp_link_work(struct work_struct *work)
940 struct ntb_transport_qp *qp = container_of(work,
941 struct ntb_transport_qp,
943 struct pci_dev *pdev = qp->ndev->pdev;
944 struct ntb_transport_ctx *nt = qp->transport;
947 WARN_ON(!nt->link_is_up);
949 val = ntb_spad_read(nt->ndev, QP_LINKS);
951 ntb_peer_spad_write(nt->ndev, QP_LINKS, val | BIT(qp->qp_num));
953 /* query remote spad for qp ready bits */
954 ntb_peer_spad_read(nt->ndev, QP_LINKS);
955 dev_dbg_ratelimited(&pdev->dev, "Remote QP link status = %x\n", val);
957 /* See if the remote side is up */
958 if (val & BIT(qp->qp_num)) {
959 dev_info(&pdev->dev, "qp %d: Link Up\n", qp->qp_num);
960 qp->link_is_up = true;
963 if (qp->event_handler)
964 qp->event_handler(qp->cb_data, qp->link_is_up);
967 tasklet_schedule(&qp->rxc_db_work);
968 } else if (nt->link_is_up)
969 schedule_delayed_work(&qp->link_work,
970 msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
973 static int ntb_transport_init_queue(struct ntb_transport_ctx *nt,
976 struct ntb_transport_qp *qp;
978 resource_size_t mw_size;
979 unsigned int num_qps_mw, tx_size;
980 unsigned int mw_num, mw_count, qp_count;
983 mw_count = nt->mw_count;
984 qp_count = nt->qp_count;
986 mw_num = QP_TO_MW(nt, qp_num);
988 qp = &nt->qp_vec[qp_num];
992 qp->client_ready = false;
993 qp->event_handler = NULL;
994 ntb_qp_link_down_reset(qp);
996 if (qp_count % mw_count && mw_num + 1 < qp_count / mw_count)
997 num_qps_mw = qp_count / mw_count + 1;
999 num_qps_mw = qp_count / mw_count;
1001 mw_base = nt->mw_vec[mw_num].phys_addr;
1002 mw_size = nt->mw_vec[mw_num].phys_size;
1004 tx_size = (unsigned int)mw_size / num_qps_mw;
1005 qp_offset = tx_size * (qp_num / mw_count);
1007 qp->tx_mw = nt->mw_vec[mw_num].vbase + qp_offset;
1011 qp->tx_mw_phys = mw_base + qp_offset;
1012 if (!qp->tx_mw_phys)
1015 tx_size -= sizeof(struct ntb_rx_info);
1016 qp->rx_info = qp->tx_mw + tx_size;
1018 /* Due to housekeeping, there must be atleast 2 buffs */
1019 qp->tx_max_frame = min(transport_mtu, tx_size / 2);
1020 qp->tx_max_entry = tx_size / qp->tx_max_frame;
1022 if (nt->debugfs_node_dir) {
1023 char debugfs_name[4];
1025 snprintf(debugfs_name, 4, "qp%d", qp_num);
1026 qp->debugfs_dir = debugfs_create_dir(debugfs_name,
1027 nt->debugfs_node_dir);
1029 qp->debugfs_stats = debugfs_create_file("stats", S_IRUSR,
1030 qp->debugfs_dir, qp,
1031 &ntb_qp_debugfs_stats);
1033 qp->debugfs_dir = NULL;
1034 qp->debugfs_stats = NULL;
1037 INIT_DELAYED_WORK(&qp->link_work, ntb_qp_link_work);
1038 INIT_WORK(&qp->link_cleanup, ntb_qp_link_cleanup_work);
1040 spin_lock_init(&qp->ntb_rx_q_lock);
1041 spin_lock_init(&qp->ntb_tx_free_q_lock);
1043 INIT_LIST_HEAD(&qp->rx_post_q);
1044 INIT_LIST_HEAD(&qp->rx_pend_q);
1045 INIT_LIST_HEAD(&qp->rx_free_q);
1046 INIT_LIST_HEAD(&qp->tx_free_q);
1048 tasklet_init(&qp->rxc_db_work, ntb_transport_rxc_db,
1054 static int ntb_transport_probe(struct ntb_client *self, struct ntb_dev *ndev)
1056 struct ntb_transport_ctx *nt;
1057 struct ntb_transport_mw *mw;
1058 unsigned int mw_count, qp_count;
1063 mw_count = ntb_mw_count(ndev);
1064 if (ntb_spad_count(ndev) < (NUM_MWS + 1 + mw_count * 2)) {
1065 dev_err(&ndev->dev, "Not enough scratch pad registers for %s",
1066 NTB_TRANSPORT_NAME);
1070 if (ntb_db_is_unsafe(ndev))
1072 "doorbell is unsafe, proceed anyway...\n");
1073 if (ntb_spad_is_unsafe(ndev))
1075 "scratchpad is unsafe, proceed anyway...\n");
1077 node = dev_to_node(&ndev->dev);
1079 nt = kzalloc_node(sizeof(*nt), GFP_KERNEL, node);
1085 nt->mw_count = mw_count;
1087 nt->mw_vec = kzalloc_node(mw_count * sizeof(*nt->mw_vec),
1094 for (i = 0; i < mw_count; i++) {
1095 mw = &nt->mw_vec[i];
1097 rc = ntb_mw_get_range(ndev, i, &mw->phys_addr, &mw->phys_size,
1098 &mw->xlat_align, &mw->xlat_align_size);
1102 mw->vbase = ioremap_wc(mw->phys_addr, mw->phys_size);
1110 mw->virt_addr = NULL;
1114 qp_bitmap = ntb_db_valid_mask(ndev);
1116 qp_count = ilog2(qp_bitmap);
1117 if (max_num_clients && max_num_clients < qp_count)
1118 qp_count = max_num_clients;
1119 else if (mw_count < qp_count)
1120 qp_count = mw_count;
1122 qp_bitmap &= BIT_ULL(qp_count) - 1;
1124 nt->qp_count = qp_count;
1125 nt->qp_bitmap = qp_bitmap;
1126 nt->qp_bitmap_free = qp_bitmap;
1128 nt->qp_vec = kzalloc_node(qp_count * sizeof(*nt->qp_vec),
1135 if (nt_debugfs_dir) {
1136 nt->debugfs_node_dir =
1137 debugfs_create_dir(pci_name(ndev->pdev),
1141 for (i = 0; i < qp_count; i++) {
1142 rc = ntb_transport_init_queue(nt, i);
1147 INIT_DELAYED_WORK(&nt->link_work, ntb_transport_link_work);
1148 INIT_WORK(&nt->link_cleanup, ntb_transport_link_cleanup_work);
1150 rc = ntb_set_ctx(ndev, nt, &ntb_transport_ops);
1154 INIT_LIST_HEAD(&nt->client_devs);
1155 rc = ntb_bus_init(nt);
1159 nt->link_is_up = false;
1160 ntb_link_enable(ndev, NTB_SPEED_AUTO, NTB_WIDTH_AUTO);
1161 ntb_link_event(ndev);
1166 ntb_clear_ctx(ndev);
1171 mw = &nt->mw_vec[i];
1180 static void ntb_transport_free(struct ntb_client *self, struct ntb_dev *ndev)
1182 struct ntb_transport_ctx *nt = ndev->ctx;
1183 struct ntb_transport_qp *qp;
1184 u64 qp_bitmap_alloc;
1187 ntb_transport_link_cleanup(nt);
1188 cancel_work_sync(&nt->link_cleanup);
1189 cancel_delayed_work_sync(&nt->link_work);
1191 qp_bitmap_alloc = nt->qp_bitmap & ~nt->qp_bitmap_free;
1193 /* verify that all the qp's are freed */
1194 for (i = 0; i < nt->qp_count; i++) {
1195 qp = &nt->qp_vec[i];
1196 if (qp_bitmap_alloc & BIT_ULL(i))
1197 ntb_transport_free_queue(qp);
1198 debugfs_remove_recursive(qp->debugfs_dir);
1201 ntb_link_disable(ndev);
1202 ntb_clear_ctx(ndev);
1206 for (i = nt->mw_count; i--; ) {
1208 iounmap(nt->mw_vec[i].vbase);
1216 static void ntb_complete_rxc(struct ntb_transport_qp *qp)
1218 struct ntb_queue_entry *entry;
1221 unsigned long irqflags;
1223 spin_lock_irqsave(&qp->ntb_rx_q_lock, irqflags);
1225 while (!list_empty(&qp->rx_post_q)) {
1226 entry = list_first_entry(&qp->rx_post_q,
1227 struct ntb_queue_entry, entry);
1228 if (!(entry->flags & DESC_DONE_FLAG))
1231 entry->rx_hdr->flags = 0;
1232 iowrite32(entry->index, &qp->rx_info->entry);
1234 cb_data = entry->cb_data;
1237 list_move_tail(&entry->entry, &qp->rx_free_q);
1239 spin_unlock_irqrestore(&qp->ntb_rx_q_lock, irqflags);
1241 if (qp->rx_handler && qp->client_ready)
1242 qp->rx_handler(qp, qp->cb_data, cb_data, len);
1244 spin_lock_irqsave(&qp->ntb_rx_q_lock, irqflags);
1247 spin_unlock_irqrestore(&qp->ntb_rx_q_lock, irqflags);
1250 static void ntb_rx_copy_callback(void *data)
1252 struct ntb_queue_entry *entry = data;
1254 entry->flags |= DESC_DONE_FLAG;
1256 ntb_complete_rxc(entry->qp);
1259 static void ntb_memcpy_rx(struct ntb_queue_entry *entry, void *offset)
1261 void *buf = entry->buf;
1262 size_t len = entry->len;
1264 memcpy(buf, offset, len);
1266 /* Ensure that the data is fully copied out before clearing the flag */
1269 ntb_rx_copy_callback(entry);
1272 static void ntb_async_rx(struct ntb_queue_entry *entry, void *offset)
1274 struct dma_async_tx_descriptor *txd;
1275 struct ntb_transport_qp *qp = entry->qp;
1276 struct dma_chan *chan = qp->rx_dma_chan;
1277 struct dma_device *device;
1278 size_t pay_off, buff_off, len;
1279 struct dmaengine_unmap_data *unmap;
1280 dma_cookie_t cookie;
1281 void *buf = entry->buf;
1289 if (len < copy_bytes)
1292 device = chan->device;
1293 pay_off = (size_t)offset & ~PAGE_MASK;
1294 buff_off = (size_t)buf & ~PAGE_MASK;
1296 if (!is_dma_copy_aligned(device, pay_off, buff_off, len))
1299 unmap = dmaengine_get_unmap_data(device->dev, 2, GFP_NOWAIT);
1304 unmap->addr[0] = dma_map_page(device->dev, virt_to_page(offset),
1305 pay_off, len, DMA_TO_DEVICE);
1306 if (dma_mapping_error(device->dev, unmap->addr[0]))
1311 unmap->addr[1] = dma_map_page(device->dev, virt_to_page(buf),
1312 buff_off, len, DMA_FROM_DEVICE);
1313 if (dma_mapping_error(device->dev, unmap->addr[1]))
1316 unmap->from_cnt = 1;
1318 for (retries = 0; retries < DMA_RETRIES; retries++) {
1319 txd = device->device_prep_dma_memcpy(chan, unmap->addr[1],
1320 unmap->addr[0], len,
1321 DMA_PREP_INTERRUPT);
1325 set_current_state(TASK_INTERRUPTIBLE);
1326 schedule_timeout(DMA_OUT_RESOURCE_TO);
1330 qp->dma_rx_prep_err++;
1334 txd->callback = ntb_rx_copy_callback;
1335 txd->callback_param = entry;
1336 dma_set_unmap(txd, unmap);
1338 cookie = dmaengine_submit(txd);
1339 if (dma_submit_error(cookie))
1342 dmaengine_unmap_put(unmap);
1344 qp->last_cookie = cookie;
1351 dmaengine_unmap_put(unmap);
1353 dmaengine_unmap_put(unmap);
1355 ntb_memcpy_rx(entry, offset);
1359 static int ntb_process_rxc(struct ntb_transport_qp *qp)
1361 struct ntb_payload_header *hdr;
1362 struct ntb_queue_entry *entry;
1365 offset = qp->rx_buff + qp->rx_max_frame * qp->rx_index;
1366 hdr = offset + qp->rx_max_frame - sizeof(struct ntb_payload_header);
1368 dev_dbg(&qp->ndev->pdev->dev, "qp %d: RX ver %u len %d flags %x\n",
1369 qp->qp_num, hdr->ver, hdr->len, hdr->flags);
1371 if (!(hdr->flags & DESC_DONE_FLAG)) {
1372 dev_dbg(&qp->ndev->pdev->dev, "done flag not set\n");
1373 qp->rx_ring_empty++;
1377 if (hdr->flags & LINK_DOWN_FLAG) {
1378 dev_dbg(&qp->ndev->pdev->dev, "link down flag set\n");
1379 ntb_qp_link_down(qp);
1384 if (hdr->ver != (u32)qp->rx_pkts) {
1385 dev_dbg(&qp->ndev->pdev->dev,
1386 "version mismatch, expected %llu - got %u\n",
1387 qp->rx_pkts, hdr->ver);
1392 entry = ntb_list_mv(&qp->ntb_rx_q_lock, &qp->rx_pend_q, &qp->rx_post_q);
1394 dev_dbg(&qp->ndev->pdev->dev, "no receive buffer\n");
1395 qp->rx_err_no_buf++;
1399 entry->rx_hdr = hdr;
1400 entry->index = qp->rx_index;
1402 if (hdr->len > entry->len) {
1403 dev_dbg(&qp->ndev->pdev->dev,
1404 "receive buffer overflow! Wanted %d got %d\n",
1405 hdr->len, entry->len);
1409 entry->flags |= DESC_DONE_FLAG;
1411 ntb_complete_rxc(qp);
1413 dev_dbg(&qp->ndev->pdev->dev,
1414 "RX OK index %u ver %u size %d into buf size %d\n",
1415 qp->rx_index, hdr->ver, hdr->len, entry->len);
1417 qp->rx_bytes += hdr->len;
1420 entry->len = hdr->len;
1422 ntb_async_rx(entry, offset);
1426 qp->rx_index %= qp->rx_max_entry;
1431 static void ntb_transport_rxc_db(unsigned long data)
1433 struct ntb_transport_qp *qp = (void *)data;
1436 dev_dbg(&qp->ndev->pdev->dev, "%s: doorbell %d received\n",
1437 __func__, qp->qp_num);
1439 /* Limit the number of packets processed in a single interrupt to
1440 * provide fairness to others
1442 for (i = 0; i < qp->rx_max_entry; i++) {
1443 rc = ntb_process_rxc(qp);
1448 if (i && qp->rx_dma_chan)
1449 dma_async_issue_pending(qp->rx_dma_chan);
1451 if (i == qp->rx_max_entry) {
1452 /* there is more work to do */
1454 tasklet_schedule(&qp->rxc_db_work);
1455 } else if (ntb_db_read(qp->ndev) & BIT_ULL(qp->qp_num)) {
1456 /* the doorbell bit is set: clear it */
1457 ntb_db_clear(qp->ndev, BIT_ULL(qp->qp_num));
1458 /* ntb_db_read ensures ntb_db_clear write is committed */
1459 ntb_db_read(qp->ndev);
1461 /* an interrupt may have arrived between finishing
1462 * ntb_process_rxc and clearing the doorbell bit:
1463 * there might be some more work to do.
1466 tasklet_schedule(&qp->rxc_db_work);
1470 static void ntb_tx_copy_callback(void *data)
1472 struct ntb_queue_entry *entry = data;
1473 struct ntb_transport_qp *qp = entry->qp;
1474 struct ntb_payload_header __iomem *hdr = entry->tx_hdr;
1476 iowrite32(entry->flags | DESC_DONE_FLAG, &hdr->flags);
1478 ntb_peer_db_set(qp->ndev, BIT_ULL(qp->qp_num));
1480 /* The entry length can only be zero if the packet is intended to be a
1481 * "link down" or similar. Since no payload is being sent in these
1482 * cases, there is nothing to add to the completion queue.
1484 if (entry->len > 0) {
1485 qp->tx_bytes += entry->len;
1488 qp->tx_handler(qp, qp->cb_data, entry->cb_data,
1492 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry, &qp->tx_free_q);
1495 static void ntb_memcpy_tx(struct ntb_queue_entry *entry, void __iomem *offset)
1497 #ifdef ARCH_HAS_NOCACHE_UACCESS
1499 * Using non-temporal mov to improve performance on non-cached
1500 * writes, even though we aren't actually copying from user space.
1502 __copy_from_user_inatomic_nocache(offset, entry->buf, entry->len);
1504 memcpy_toio(offset, entry->buf, entry->len);
1507 /* Ensure that the data is fully copied out before setting the flags */
1510 ntb_tx_copy_callback(entry);
1513 static void ntb_async_tx(struct ntb_transport_qp *qp,
1514 struct ntb_queue_entry *entry)
1516 struct ntb_payload_header __iomem *hdr;
1517 struct dma_async_tx_descriptor *txd;
1518 struct dma_chan *chan = qp->tx_dma_chan;
1519 struct dma_device *device;
1520 size_t dest_off, buff_off;
1521 struct dmaengine_unmap_data *unmap;
1523 dma_cookie_t cookie;
1524 void __iomem *offset;
1525 size_t len = entry->len;
1526 void *buf = entry->buf;
1529 offset = qp->tx_mw + qp->tx_max_frame * qp->tx_index;
1530 hdr = offset + qp->tx_max_frame - sizeof(struct ntb_payload_header);
1531 entry->tx_hdr = hdr;
1533 iowrite32(entry->len, &hdr->len);
1534 iowrite32((u32)qp->tx_pkts, &hdr->ver);
1539 if (len < copy_bytes)
1542 device = chan->device;
1543 dest = qp->tx_mw_phys + qp->tx_max_frame * qp->tx_index;
1544 buff_off = (size_t)buf & ~PAGE_MASK;
1545 dest_off = (size_t)dest & ~PAGE_MASK;
1547 if (!is_dma_copy_aligned(device, buff_off, dest_off, len))
1550 unmap = dmaengine_get_unmap_data(device->dev, 1, GFP_NOWAIT);
1555 unmap->addr[0] = dma_map_page(device->dev, virt_to_page(buf),
1556 buff_off, len, DMA_TO_DEVICE);
1557 if (dma_mapping_error(device->dev, unmap->addr[0]))
1562 for (retries = 0; retries < DMA_RETRIES; retries++) {
1563 txd = device->device_prep_dma_memcpy(chan, dest, unmap->addr[0],
1564 len, DMA_PREP_INTERRUPT);
1568 set_current_state(TASK_INTERRUPTIBLE);
1569 schedule_timeout(DMA_OUT_RESOURCE_TO);
1573 qp->dma_tx_prep_err++;
1577 txd->callback = ntb_tx_copy_callback;
1578 txd->callback_param = entry;
1579 dma_set_unmap(txd, unmap);
1581 cookie = dmaengine_submit(txd);
1582 if (dma_submit_error(cookie))
1585 dmaengine_unmap_put(unmap);
1587 dma_async_issue_pending(chan);
1592 dmaengine_unmap_put(unmap);
1594 dmaengine_unmap_put(unmap);
1596 ntb_memcpy_tx(entry, offset);
1600 static int ntb_process_tx(struct ntb_transport_qp *qp,
1601 struct ntb_queue_entry *entry)
1603 if (qp->tx_index == qp->remote_rx_info->entry) {
1608 if (entry->len > qp->tx_max_frame - sizeof(struct ntb_payload_header)) {
1610 qp->tx_handler(qp, qp->cb_data, NULL, -EIO);
1612 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
1617 ntb_async_tx(qp, entry);
1620 qp->tx_index %= qp->tx_max_entry;
1627 static void ntb_send_link_down(struct ntb_transport_qp *qp)
1629 struct pci_dev *pdev = qp->ndev->pdev;
1630 struct ntb_queue_entry *entry;
1633 if (!qp->link_is_up)
1636 dev_info(&pdev->dev, "qp %d: Send Link Down\n", qp->qp_num);
1638 for (i = 0; i < NTB_LINK_DOWN_TIMEOUT; i++) {
1639 entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
1648 entry->cb_data = NULL;
1651 entry->flags = LINK_DOWN_FLAG;
1653 rc = ntb_process_tx(qp, entry);
1655 dev_err(&pdev->dev, "ntb: QP%d unable to send linkdown msg\n",
1658 ntb_qp_link_down_reset(qp);
1661 static bool ntb_dma_filter_fn(struct dma_chan *chan, void *node)
1663 return dev_to_node(&chan->dev->device) == (int)(unsigned long)node;
1667 * ntb_transport_create_queue - Create a new NTB transport layer queue
1668 * @rx_handler: receive callback function
1669 * @tx_handler: transmit callback function
1670 * @event_handler: event callback function
1672 * Create a new NTB transport layer queue and provide the queue with a callback
1673 * routine for both transmit and receive. The receive callback routine will be
1674 * used to pass up data when the transport has received it on the queue. The
1675 * transmit callback routine will be called when the transport has completed the
1676 * transmission of the data on the queue and the data is ready to be freed.
1678 * RETURNS: pointer to newly created ntb_queue, NULL on error.
1680 struct ntb_transport_qp *
1681 ntb_transport_create_queue(void *data, struct device *client_dev,
1682 const struct ntb_queue_handlers *handlers)
1684 struct ntb_dev *ndev;
1685 struct pci_dev *pdev;
1686 struct ntb_transport_ctx *nt;
1687 struct ntb_queue_entry *entry;
1688 struct ntb_transport_qp *qp;
1690 unsigned int free_queue;
1691 dma_cap_mask_t dma_mask;
1695 ndev = dev_ntb(client_dev->parent);
1699 node = dev_to_node(&ndev->dev);
1701 free_queue = ffs(nt->qp_bitmap);
1705 /* decrement free_queue to make it zero based */
1708 qp = &nt->qp_vec[free_queue];
1709 qp_bit = BIT_ULL(qp->qp_num);
1711 nt->qp_bitmap_free &= ~qp_bit;
1714 qp->rx_handler = handlers->rx_handler;
1715 qp->tx_handler = handlers->tx_handler;
1716 qp->event_handler = handlers->event_handler;
1718 dma_cap_zero(dma_mask);
1719 dma_cap_set(DMA_MEMCPY, dma_mask);
1723 dma_request_channel(dma_mask, ntb_dma_filter_fn,
1724 (void *)(unsigned long)node);
1725 if (!qp->tx_dma_chan)
1726 dev_info(&pdev->dev, "Unable to allocate TX DMA channel\n");
1729 dma_request_channel(dma_mask, ntb_dma_filter_fn,
1730 (void *)(unsigned long)node);
1731 if (!qp->rx_dma_chan)
1732 dev_info(&pdev->dev, "Unable to allocate RX DMA channel\n");
1734 qp->tx_dma_chan = NULL;
1735 qp->rx_dma_chan = NULL;
1738 dev_dbg(&pdev->dev, "Using %s memcpy for TX\n",
1739 qp->tx_dma_chan ? "DMA" : "CPU");
1741 dev_dbg(&pdev->dev, "Using %s memcpy for RX\n",
1742 qp->rx_dma_chan ? "DMA" : "CPU");
1744 for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) {
1745 entry = kzalloc_node(sizeof(*entry), GFP_ATOMIC, node);
1750 ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry,
1753 qp->rx_alloc_entry = NTB_QP_DEF_NUM_ENTRIES;
1755 for (i = 0; i < qp->tx_max_entry; i++) {
1756 entry = kzalloc_node(sizeof(*entry), GFP_ATOMIC, node);
1761 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
1765 ntb_db_clear(qp->ndev, qp_bit);
1766 ntb_db_clear_mask(qp->ndev, qp_bit);
1768 dev_info(&pdev->dev, "NTB Transport QP %d created\n", qp->qp_num);
1773 while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
1776 qp->rx_alloc_entry = 0;
1777 while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q)))
1779 if (qp->tx_dma_chan)
1780 dma_release_channel(qp->tx_dma_chan);
1781 if (qp->rx_dma_chan)
1782 dma_release_channel(qp->rx_dma_chan);
1783 nt->qp_bitmap_free |= qp_bit;
1787 EXPORT_SYMBOL_GPL(ntb_transport_create_queue);
1790 * ntb_transport_free_queue - Frees NTB transport queue
1791 * @qp: NTB queue to be freed
1793 * Frees NTB transport queue
1795 void ntb_transport_free_queue(struct ntb_transport_qp *qp)
1797 struct pci_dev *pdev;
1798 struct ntb_queue_entry *entry;
1804 pdev = qp->ndev->pdev;
1808 if (qp->tx_dma_chan) {
1809 struct dma_chan *chan = qp->tx_dma_chan;
1810 /* Putting the dma_chan to NULL will force any new traffic to be
1811 * processed by the CPU instead of the DAM engine
1813 qp->tx_dma_chan = NULL;
1815 /* Try to be nice and wait for any queued DMA engine
1816 * transactions to process before smashing it with a rock
1818 dma_sync_wait(chan, qp->last_cookie);
1819 dmaengine_terminate_all(chan);
1820 dma_release_channel(chan);
1823 if (qp->rx_dma_chan) {
1824 struct dma_chan *chan = qp->rx_dma_chan;
1825 /* Putting the dma_chan to NULL will force any new traffic to be
1826 * processed by the CPU instead of the DAM engine
1828 qp->rx_dma_chan = NULL;
1830 /* Try to be nice and wait for any queued DMA engine
1831 * transactions to process before smashing it with a rock
1833 dma_sync_wait(chan, qp->last_cookie);
1834 dmaengine_terminate_all(chan);
1835 dma_release_channel(chan);
1838 qp_bit = BIT_ULL(qp->qp_num);
1840 ntb_db_set_mask(qp->ndev, qp_bit);
1841 tasklet_kill(&qp->rxc_db_work);
1843 cancel_delayed_work_sync(&qp->link_work);
1846 qp->rx_handler = NULL;
1847 qp->tx_handler = NULL;
1848 qp->event_handler = NULL;
1850 while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q)))
1853 while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_pend_q))) {
1854 dev_warn(&pdev->dev, "Freeing item from non-empty rx_pend_q\n");
1858 while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_post_q))) {
1859 dev_warn(&pdev->dev, "Freeing item from non-empty rx_post_q\n");
1863 while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
1866 qp->transport->qp_bitmap_free |= qp_bit;
1868 dev_info(&pdev->dev, "NTB Transport QP %d freed\n", qp->qp_num);
1870 EXPORT_SYMBOL_GPL(ntb_transport_free_queue);
1873 * ntb_transport_rx_remove - Dequeues enqueued rx packet
1874 * @qp: NTB queue to be freed
1875 * @len: pointer to variable to write enqueued buffers length
1877 * Dequeues unused buffers from receive queue. Should only be used during
1880 * RETURNS: NULL error value on error, or void* for success.
1882 void *ntb_transport_rx_remove(struct ntb_transport_qp *qp, unsigned int *len)
1884 struct ntb_queue_entry *entry;
1887 if (!qp || qp->client_ready)
1890 entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_pend_q);
1894 buf = entry->cb_data;
1897 ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry, &qp->rx_free_q);
1901 EXPORT_SYMBOL_GPL(ntb_transport_rx_remove);
1904 * ntb_transport_rx_enqueue - Enqueue a new NTB queue entry
1905 * @qp: NTB transport layer queue the entry is to be enqueued on
1906 * @cb: per buffer pointer for callback function to use
1907 * @data: pointer to data buffer that incoming packets will be copied into
1908 * @len: length of the data buffer
1910 * Enqueue a new receive buffer onto the transport queue into which a NTB
1911 * payload can be received into.
1913 * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
1915 int ntb_transport_rx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
1918 struct ntb_queue_entry *entry;
1923 entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q);
1927 entry->cb_data = cb;
1932 ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry, &qp->rx_pend_q);
1935 tasklet_schedule(&qp->rxc_db_work);
1939 EXPORT_SYMBOL_GPL(ntb_transport_rx_enqueue);
1942 * ntb_transport_tx_enqueue - Enqueue a new NTB queue entry
1943 * @qp: NTB transport layer queue the entry is to be enqueued on
1944 * @cb: per buffer pointer for callback function to use
1945 * @data: pointer to data buffer that will be sent
1946 * @len: length of the data buffer
1948 * Enqueue a new transmit buffer onto the transport queue from which a NTB
1949 * payload will be transmitted. This assumes that a lock is being held to
1950 * serialize access to the qp.
1952 * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
1954 int ntb_transport_tx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
1957 struct ntb_queue_entry *entry;
1960 if (!qp || !qp->link_is_up || !len)
1963 entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
1965 qp->tx_err_no_buf++;
1969 entry->cb_data = cb;
1974 rc = ntb_process_tx(qp, entry);
1976 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
1981 EXPORT_SYMBOL_GPL(ntb_transport_tx_enqueue);
1984 * ntb_transport_link_up - Notify NTB transport of client readiness to use queue
1985 * @qp: NTB transport layer queue to be enabled
1987 * Notify NTB transport layer of client readiness to use queue
1989 void ntb_transport_link_up(struct ntb_transport_qp *qp)
1994 qp->client_ready = true;
1996 if (qp->transport->link_is_up)
1997 schedule_delayed_work(&qp->link_work, 0);
1999 EXPORT_SYMBOL_GPL(ntb_transport_link_up);
2002 * ntb_transport_link_down - Notify NTB transport to no longer enqueue data
2003 * @qp: NTB transport layer queue to be disabled
2005 * Notify NTB transport layer of client's desire to no longer receive data on
2006 * transport queue specified. It is the client's responsibility to ensure all
2007 * entries on queue are purged or otherwise handled appropriately.
2009 void ntb_transport_link_down(struct ntb_transport_qp *qp)
2016 qp->client_ready = false;
2018 val = ntb_spad_read(qp->ndev, QP_LINKS);
2020 ntb_peer_spad_write(qp->ndev, QP_LINKS,
2021 val & ~BIT(qp->qp_num));
2024 ntb_send_link_down(qp);
2026 cancel_delayed_work_sync(&qp->link_work);
2028 EXPORT_SYMBOL_GPL(ntb_transport_link_down);
2031 * ntb_transport_link_query - Query transport link state
2032 * @qp: NTB transport layer queue to be queried
2034 * Query connectivity to the remote system of the NTB transport queue
2036 * RETURNS: true for link up or false for link down
2038 bool ntb_transport_link_query(struct ntb_transport_qp *qp)
2043 return qp->link_is_up;
2045 EXPORT_SYMBOL_GPL(ntb_transport_link_query);
2048 * ntb_transport_qp_num - Query the qp number
2049 * @qp: NTB transport layer queue to be queried
2051 * Query qp number of the NTB transport queue
2053 * RETURNS: a zero based number specifying the qp number
2055 unsigned char ntb_transport_qp_num(struct ntb_transport_qp *qp)
2062 EXPORT_SYMBOL_GPL(ntb_transport_qp_num);
2065 * ntb_transport_max_size - Query the max payload size of a qp
2066 * @qp: NTB transport layer queue to be queried
2068 * Query the maximum payload size permissible on the given qp
2070 * RETURNS: the max payload size of a qp
2072 unsigned int ntb_transport_max_size(struct ntb_transport_qp *qp)
2074 unsigned int max_size;
2075 unsigned int copy_align;
2076 struct dma_chan *rx_chan, *tx_chan;
2081 rx_chan = qp->rx_dma_chan;
2082 tx_chan = qp->tx_dma_chan;
2084 copy_align = max(rx_chan ? rx_chan->device->copy_align : 0,
2085 tx_chan ? tx_chan->device->copy_align : 0);
2087 /* If DMA engine usage is possible, try to find the max size for that */
2088 max_size = qp->tx_max_frame - sizeof(struct ntb_payload_header);
2089 max_size = round_down(max_size, 1 << copy_align);
2093 EXPORT_SYMBOL_GPL(ntb_transport_max_size);
2095 unsigned int ntb_transport_tx_free_entry(struct ntb_transport_qp *qp)
2097 unsigned int head = qp->tx_index;
2098 unsigned int tail = qp->remote_rx_info->entry;
2100 return tail > head ? tail - head : qp->tx_max_entry + tail - head;
2102 EXPORT_SYMBOL_GPL(ntb_transport_tx_free_entry);
2104 static void ntb_transport_doorbell_callback(void *data, int vector)
2106 struct ntb_transport_ctx *nt = data;
2107 struct ntb_transport_qp *qp;
2109 unsigned int qp_num;
2111 db_bits = (nt->qp_bitmap & ~nt->qp_bitmap_free &
2112 ntb_db_vector_mask(nt->ndev, vector));
2115 qp_num = __ffs(db_bits);
2116 qp = &nt->qp_vec[qp_num];
2119 tasklet_schedule(&qp->rxc_db_work);
2121 db_bits &= ~BIT_ULL(qp_num);
2125 static const struct ntb_ctx_ops ntb_transport_ops = {
2126 .link_event = ntb_transport_event_callback,
2127 .db_event = ntb_transport_doorbell_callback,
2130 static struct ntb_client ntb_transport_client = {
2132 .probe = ntb_transport_probe,
2133 .remove = ntb_transport_free,
2137 static int __init ntb_transport_init(void)
2141 pr_info("%s, version %s\n", NTB_TRANSPORT_DESC, NTB_TRANSPORT_VER);
2143 if (debugfs_initialized())
2144 nt_debugfs_dir = debugfs_create_dir(KBUILD_MODNAME, NULL);
2146 rc = bus_register(&ntb_transport_bus);
2150 rc = ntb_register_client(&ntb_transport_client);
2157 bus_unregister(&ntb_transport_bus);
2159 debugfs_remove_recursive(nt_debugfs_dir);
2162 module_init(ntb_transport_init);
2164 static void __exit ntb_transport_exit(void)
2166 debugfs_remove_recursive(nt_debugfs_dir);
2168 ntb_unregister_client(&ntb_transport_client);
2169 bus_unregister(&ntb_transport_bus);
2171 module_exit(ntb_transport_exit);