2 * This file is provided under a dual BSD/GPLv2 license. When using or
3 * redistributing this file, you may do so under either license.
7 * Copyright(c) 2012 Intel Corporation. All rights reserved.
8 * Copyright (C) 2015 EMC Corporation. All Rights Reserved.
10 * This program is free software; you can redistribute it and/or modify
11 * it under the terms of version 2 of the GNU General Public License as
12 * published by the Free Software Foundation.
16 * Copyright(c) 2012 Intel Corporation. All rights reserved.
17 * Copyright (C) 2015 EMC Corporation. All Rights Reserved.
19 * Redistribution and use in source and binary forms, with or without
20 * modification, are permitted provided that the following conditions
23 * * Redistributions of source code must retain the above copyright
24 * notice, this list of conditions and the following disclaimer.
25 * * Redistributions in binary form must reproduce the above copy
26 * notice, this list of conditions and the following disclaimer in
27 * the documentation and/or other materials provided with the
29 * * Neither the name of Intel Corporation nor the names of its
30 * contributors may be used to endorse or promote products derived
31 * from this software without specific prior written permission.
33 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
34 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
35 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
36 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
37 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
38 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
39 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
40 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
41 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
42 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
43 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
45 * PCIe NTB Transport Linux driver
47 * Contact Information:
48 * Jon Mason <jon.mason@intel.com>
50 #include <linux/debugfs.h>
51 #include <linux/delay.h>
52 #include <linux/dmaengine.h>
53 #include <linux/dma-mapping.h>
54 #include <linux/errno.h>
55 #include <linux/export.h>
56 #include <linux/interrupt.h>
57 #include <linux/module.h>
58 #include <linux/pci.h>
59 #include <linux/slab.h>
60 #include <linux/types.h>
61 #include <linux/uaccess.h>
62 #include "linux/ntb.h"
63 #include "linux/ntb_transport.h"
65 #define NTB_TRANSPORT_VERSION 4
66 #define NTB_TRANSPORT_VER "4"
67 #define NTB_TRANSPORT_NAME "ntb_transport"
68 #define NTB_TRANSPORT_DESC "Software Queue-Pair Transport over NTB"
69 #define NTB_TRANSPORT_MIN_SPADS (MW0_SZ_HIGH + 2)
71 MODULE_DESCRIPTION(NTB_TRANSPORT_DESC);
72 MODULE_VERSION(NTB_TRANSPORT_VER);
73 MODULE_LICENSE("Dual BSD/GPL");
74 MODULE_AUTHOR("Intel Corporation");
76 static unsigned long max_mw_size;
77 module_param(max_mw_size, ulong, 0644);
78 MODULE_PARM_DESC(max_mw_size, "Limit size of large memory windows");
80 static unsigned int transport_mtu = 0x10000;
81 module_param(transport_mtu, uint, 0644);
82 MODULE_PARM_DESC(transport_mtu, "Maximum size of NTB transport packets");
84 static unsigned char max_num_clients;
85 module_param(max_num_clients, byte, 0644);
86 MODULE_PARM_DESC(max_num_clients, "Maximum number of NTB transport clients");
88 static unsigned int copy_bytes = 1024;
89 module_param(copy_bytes, uint, 0644);
90 MODULE_PARM_DESC(copy_bytes, "Threshold under which NTB will use the CPU to copy instead of DMA");
93 module_param(use_dma, bool, 0644);
94 MODULE_PARM_DESC(use_dma, "Use DMA engine to perform large data copy");
96 static struct dentry *nt_debugfs_dir;
98 struct ntb_queue_entry {
99 /* ntb_queue list reference */
100 struct list_head entry;
101 /* pointers to data to be transferred */
108 unsigned int tx_index;
109 unsigned int rx_index;
111 struct ntb_transport_qp *qp;
113 struct ntb_payload_header __iomem *tx_hdr;
114 struct ntb_payload_header *rx_hdr;
122 struct ntb_transport_qp {
123 struct ntb_transport_ctx *transport;
124 struct ntb_dev *ndev;
126 struct dma_chan *tx_dma_chan;
127 struct dma_chan *rx_dma_chan;
133 u8 qp_num; /* Only 64 QP's are allowed. 0-63 */
136 struct ntb_rx_info __iomem *rx_info;
137 struct ntb_rx_info *remote_rx_info;
139 void (*tx_handler)(struct ntb_transport_qp *qp, void *qp_data,
140 void *data, int len);
141 struct list_head tx_free_q;
142 spinlock_t ntb_tx_free_q_lock;
144 dma_addr_t tx_mw_phys;
145 unsigned int tx_index;
146 unsigned int tx_max_entry;
147 unsigned int tx_max_frame;
149 void (*rx_handler)(struct ntb_transport_qp *qp, void *qp_data,
150 void *data, int len);
151 struct list_head rx_post_q;
152 struct list_head rx_pend_q;
153 struct list_head rx_free_q;
154 /* ntb_rx_q_lock: synchronize access to rx_XXXX_q */
155 spinlock_t ntb_rx_q_lock;
157 unsigned int rx_index;
158 unsigned int rx_max_entry;
159 unsigned int rx_max_frame;
160 unsigned int rx_alloc_entry;
161 dma_cookie_t last_cookie;
162 struct tasklet_struct rxc_db_work;
164 void (*event_handler)(void *data, int status);
165 struct delayed_work link_work;
166 struct work_struct link_cleanup;
168 struct dentry *debugfs_dir;
169 struct dentry *debugfs_stats;
188 struct ntb_transport_mw {
189 phys_addr_t phys_addr;
190 resource_size_t phys_size;
191 resource_size_t xlat_align;
192 resource_size_t xlat_align_size;
200 struct ntb_transport_client_dev {
201 struct list_head entry;
202 struct ntb_transport_ctx *nt;
206 struct ntb_transport_ctx {
207 struct list_head entry;
208 struct list_head client_devs;
210 struct ntb_dev *ndev;
212 struct ntb_transport_mw *mw_vec;
213 struct ntb_transport_qp *qp_vec;
214 unsigned int mw_count;
215 unsigned int qp_count;
220 struct delayed_work link_work;
221 struct work_struct link_cleanup;
223 struct dentry *debugfs_node_dir;
227 DESC_DONE_FLAG = BIT(0),
228 LINK_DOWN_FLAG = BIT(1),
231 struct ntb_payload_header {
246 #define dev_client_dev(__dev) \
247 container_of((__dev), struct ntb_transport_client_dev, dev)
249 #define drv_client(__drv) \
250 container_of((__drv), struct ntb_transport_client, driver)
252 #define QP_TO_MW(nt, qp) ((qp) % nt->mw_count)
253 #define NTB_QP_DEF_NUM_ENTRIES 100
254 #define NTB_LINK_DOWN_TIMEOUT 10
256 static void ntb_transport_rxc_db(unsigned long data);
257 static const struct ntb_ctx_ops ntb_transport_ops;
258 static struct ntb_client ntb_transport_client;
259 static int ntb_async_tx_submit(struct ntb_transport_qp *qp,
260 struct ntb_queue_entry *entry);
261 static void ntb_memcpy_tx(struct ntb_queue_entry *entry, void __iomem *offset);
262 static int ntb_async_rx_submit(struct ntb_queue_entry *entry, void *offset);
263 static void ntb_memcpy_rx(struct ntb_queue_entry *entry, void *offset);
266 static int ntb_transport_bus_match(struct device *dev,
267 struct device_driver *drv)
269 return !strncmp(dev_name(dev), drv->name, strlen(drv->name));
272 static int ntb_transport_bus_probe(struct device *dev)
274 const struct ntb_transport_client *client;
279 client = drv_client(dev->driver);
280 rc = client->probe(dev);
287 static int ntb_transport_bus_remove(struct device *dev)
289 const struct ntb_transport_client *client;
291 client = drv_client(dev->driver);
299 static struct bus_type ntb_transport_bus = {
300 .name = "ntb_transport",
301 .match = ntb_transport_bus_match,
302 .probe = ntb_transport_bus_probe,
303 .remove = ntb_transport_bus_remove,
306 static LIST_HEAD(ntb_transport_list);
308 static int ntb_bus_init(struct ntb_transport_ctx *nt)
310 list_add_tail(&nt->entry, &ntb_transport_list);
314 static void ntb_bus_remove(struct ntb_transport_ctx *nt)
316 struct ntb_transport_client_dev *client_dev, *cd;
318 list_for_each_entry_safe(client_dev, cd, &nt->client_devs, entry) {
319 dev_err(client_dev->dev.parent, "%s still attached to bus, removing\n",
320 dev_name(&client_dev->dev));
321 list_del(&client_dev->entry);
322 device_unregister(&client_dev->dev);
325 list_del(&nt->entry);
328 static void ntb_transport_client_release(struct device *dev)
330 struct ntb_transport_client_dev *client_dev;
332 client_dev = dev_client_dev(dev);
337 * ntb_transport_unregister_client_dev - Unregister NTB client device
338 * @device_name: Name of NTB client device
340 * Unregister an NTB client device with the NTB transport layer
342 void ntb_transport_unregister_client_dev(char *device_name)
344 struct ntb_transport_client_dev *client, *cd;
345 struct ntb_transport_ctx *nt;
347 list_for_each_entry(nt, &ntb_transport_list, entry)
348 list_for_each_entry_safe(client, cd, &nt->client_devs, entry)
349 if (!strncmp(dev_name(&client->dev), device_name,
350 strlen(device_name))) {
351 list_del(&client->entry);
352 device_unregister(&client->dev);
355 EXPORT_SYMBOL_GPL(ntb_transport_unregister_client_dev);
358 * ntb_transport_register_client_dev - Register NTB client device
359 * @device_name: Name of NTB client device
361 * Register an NTB client device with the NTB transport layer
363 int ntb_transport_register_client_dev(char *device_name)
365 struct ntb_transport_client_dev *client_dev;
366 struct ntb_transport_ctx *nt;
370 if (list_empty(&ntb_transport_list))
373 list_for_each_entry(nt, &ntb_transport_list, entry) {
376 node = dev_to_node(&nt->ndev->dev);
378 client_dev = kzalloc_node(sizeof(*client_dev),
385 dev = &client_dev->dev;
387 /* setup and register client devices */
388 dev_set_name(dev, "%s%d", device_name, i);
389 dev->bus = &ntb_transport_bus;
390 dev->release = ntb_transport_client_release;
391 dev->parent = &nt->ndev->dev;
393 rc = device_register(dev);
399 list_add_tail(&client_dev->entry, &nt->client_devs);
406 ntb_transport_unregister_client_dev(device_name);
410 EXPORT_SYMBOL_GPL(ntb_transport_register_client_dev);
413 * ntb_transport_register_client - Register NTB client driver
414 * @drv: NTB client driver to be registered
416 * Register an NTB client driver with the NTB transport layer
418 * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
420 int ntb_transport_register_client(struct ntb_transport_client *drv)
422 drv->driver.bus = &ntb_transport_bus;
424 if (list_empty(&ntb_transport_list))
427 return driver_register(&drv->driver);
429 EXPORT_SYMBOL_GPL(ntb_transport_register_client);
432 * ntb_transport_unregister_client - Unregister NTB client driver
433 * @drv: NTB client driver to be unregistered
435 * Unregister an NTB client driver with the NTB transport layer
437 * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
439 void ntb_transport_unregister_client(struct ntb_transport_client *drv)
441 driver_unregister(&drv->driver);
443 EXPORT_SYMBOL_GPL(ntb_transport_unregister_client);
445 static ssize_t debugfs_read(struct file *filp, char __user *ubuf, size_t count,
448 struct ntb_transport_qp *qp;
450 ssize_t ret, out_offset, out_count;
452 qp = filp->private_data;
454 if (!qp || !qp->link_is_up)
459 buf = kmalloc(out_count, GFP_KERNEL);
464 out_offset += snprintf(buf + out_offset, out_count - out_offset,
465 "\nNTB QP stats:\n\n");
466 out_offset += snprintf(buf + out_offset, out_count - out_offset,
467 "rx_bytes - \t%llu\n", qp->rx_bytes);
468 out_offset += snprintf(buf + out_offset, out_count - out_offset,
469 "rx_pkts - \t%llu\n", qp->rx_pkts);
470 out_offset += snprintf(buf + out_offset, out_count - out_offset,
471 "rx_memcpy - \t%llu\n", qp->rx_memcpy);
472 out_offset += snprintf(buf + out_offset, out_count - out_offset,
473 "rx_async - \t%llu\n", qp->rx_async);
474 out_offset += snprintf(buf + out_offset, out_count - out_offset,
475 "rx_ring_empty - %llu\n", qp->rx_ring_empty);
476 out_offset += snprintf(buf + out_offset, out_count - out_offset,
477 "rx_err_no_buf - %llu\n", qp->rx_err_no_buf);
478 out_offset += snprintf(buf + out_offset, out_count - out_offset,
479 "rx_err_oflow - \t%llu\n", qp->rx_err_oflow);
480 out_offset += snprintf(buf + out_offset, out_count - out_offset,
481 "rx_err_ver - \t%llu\n", qp->rx_err_ver);
482 out_offset += snprintf(buf + out_offset, out_count - out_offset,
483 "rx_buff - \t0x%p\n", qp->rx_buff);
484 out_offset += snprintf(buf + out_offset, out_count - out_offset,
485 "rx_index - \t%u\n", qp->rx_index);
486 out_offset += snprintf(buf + out_offset, out_count - out_offset,
487 "rx_max_entry - \t%u\n", qp->rx_max_entry);
488 out_offset += snprintf(buf + out_offset, out_count - out_offset,
489 "rx_alloc_entry - \t%u\n\n", qp->rx_alloc_entry);
491 out_offset += snprintf(buf + out_offset, out_count - out_offset,
492 "tx_bytes - \t%llu\n", qp->tx_bytes);
493 out_offset += snprintf(buf + out_offset, out_count - out_offset,
494 "tx_pkts - \t%llu\n", qp->tx_pkts);
495 out_offset += snprintf(buf + out_offset, out_count - out_offset,
496 "tx_memcpy - \t%llu\n", qp->tx_memcpy);
497 out_offset += snprintf(buf + out_offset, out_count - out_offset,
498 "tx_async - \t%llu\n", qp->tx_async);
499 out_offset += snprintf(buf + out_offset, out_count - out_offset,
500 "tx_ring_full - \t%llu\n", qp->tx_ring_full);
501 out_offset += snprintf(buf + out_offset, out_count - out_offset,
502 "tx_err_no_buf - %llu\n", qp->tx_err_no_buf);
503 out_offset += snprintf(buf + out_offset, out_count - out_offset,
504 "tx_mw - \t0x%p\n", qp->tx_mw);
505 out_offset += snprintf(buf + out_offset, out_count - out_offset,
506 "tx_index (H) - \t%u\n", qp->tx_index);
507 out_offset += snprintf(buf + out_offset, out_count - out_offset,
509 qp->remote_rx_info->entry);
510 out_offset += snprintf(buf + out_offset, out_count - out_offset,
511 "tx_max_entry - \t%u\n", qp->tx_max_entry);
512 out_offset += snprintf(buf + out_offset, out_count - out_offset,
514 ntb_transport_tx_free_entry(qp));
516 out_offset += snprintf(buf + out_offset, out_count - out_offset,
518 out_offset += snprintf(buf + out_offset, out_count - out_offset,
519 "Using TX DMA - \t%s\n",
520 qp->tx_dma_chan ? "Yes" : "No");
521 out_offset += snprintf(buf + out_offset, out_count - out_offset,
522 "Using RX DMA - \t%s\n",
523 qp->rx_dma_chan ? "Yes" : "No");
524 out_offset += snprintf(buf + out_offset, out_count - out_offset,
526 qp->link_is_up ? "Up" : "Down");
527 out_offset += snprintf(buf + out_offset, out_count - out_offset,
530 if (out_offset > out_count)
531 out_offset = out_count;
533 ret = simple_read_from_buffer(ubuf, count, offp, buf, out_offset);
538 static const struct file_operations ntb_qp_debugfs_stats = {
539 .owner = THIS_MODULE,
541 .read = debugfs_read,
544 static void ntb_list_add(spinlock_t *lock, struct list_head *entry,
545 struct list_head *list)
549 spin_lock_irqsave(lock, flags);
550 list_add_tail(entry, list);
551 spin_unlock_irqrestore(lock, flags);
554 static struct ntb_queue_entry *ntb_list_rm(spinlock_t *lock,
555 struct list_head *list)
557 struct ntb_queue_entry *entry;
560 spin_lock_irqsave(lock, flags);
561 if (list_empty(list)) {
565 entry = list_first_entry(list, struct ntb_queue_entry, entry);
566 list_del(&entry->entry);
569 spin_unlock_irqrestore(lock, flags);
574 static struct ntb_queue_entry *ntb_list_mv(spinlock_t *lock,
575 struct list_head *list,
576 struct list_head *to_list)
578 struct ntb_queue_entry *entry;
581 spin_lock_irqsave(lock, flags);
583 if (list_empty(list)) {
586 entry = list_first_entry(list, struct ntb_queue_entry, entry);
587 list_move_tail(&entry->entry, to_list);
590 spin_unlock_irqrestore(lock, flags);
595 static int ntb_transport_setup_qp_mw(struct ntb_transport_ctx *nt,
598 struct ntb_transport_qp *qp = &nt->qp_vec[qp_num];
599 struct ntb_transport_mw *mw;
600 struct ntb_dev *ndev = nt->ndev;
601 struct ntb_queue_entry *entry;
602 unsigned int rx_size, num_qps_mw;
603 unsigned int mw_num, mw_count, qp_count;
607 mw_count = nt->mw_count;
608 qp_count = nt->qp_count;
610 mw_num = QP_TO_MW(nt, qp_num);
611 mw = &nt->mw_vec[mw_num];
616 if (mw_num < qp_count % mw_count)
617 num_qps_mw = qp_count / mw_count + 1;
619 num_qps_mw = qp_count / mw_count;
621 rx_size = (unsigned int)mw->xlat_size / num_qps_mw;
622 qp->rx_buff = mw->virt_addr + rx_size * (qp_num / mw_count);
623 rx_size -= sizeof(struct ntb_rx_info);
625 qp->remote_rx_info = qp->rx_buff + rx_size;
627 /* Due to housekeeping, there must be atleast 2 buffs */
628 qp->rx_max_frame = min(transport_mtu, rx_size / 2);
629 qp->rx_max_entry = rx_size / qp->rx_max_frame;
633 * Checking to see if we have more entries than the default.
634 * We should add additional entries if that is the case so we
635 * can be in sync with the transport frames.
637 node = dev_to_node(&ndev->dev);
638 for (i = qp->rx_alloc_entry; i < qp->rx_max_entry; i++) {
639 entry = kzalloc_node(sizeof(*entry), GFP_ATOMIC, node);
644 ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry,
646 qp->rx_alloc_entry++;
649 qp->remote_rx_info->entry = qp->rx_max_entry - 1;
651 /* setup the hdr offsets with 0's */
652 for (i = 0; i < qp->rx_max_entry; i++) {
653 void *offset = (qp->rx_buff + qp->rx_max_frame * (i + 1) -
654 sizeof(struct ntb_payload_header));
655 memset(offset, 0, sizeof(struct ntb_payload_header));
665 static void ntb_free_mw(struct ntb_transport_ctx *nt, int num_mw)
667 struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
668 struct pci_dev *pdev = nt->ndev->pdev;
673 ntb_mw_clear_trans(nt->ndev, num_mw);
674 dma_free_coherent(&pdev->dev, mw->buff_size,
675 mw->virt_addr, mw->dma_addr);
678 mw->virt_addr = NULL;
681 static int ntb_set_mw(struct ntb_transport_ctx *nt, int num_mw,
682 resource_size_t size)
684 struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
685 struct pci_dev *pdev = nt->ndev->pdev;
686 size_t xlat_size, buff_size;
692 xlat_size = round_up(size, mw->xlat_align_size);
693 buff_size = round_up(size, mw->xlat_align);
695 /* No need to re-setup */
696 if (mw->xlat_size == xlat_size)
700 ntb_free_mw(nt, num_mw);
702 /* Alloc memory for receiving data. Must be aligned */
703 mw->xlat_size = xlat_size;
704 mw->buff_size = buff_size;
706 mw->virt_addr = dma_alloc_coherent(&pdev->dev, buff_size,
707 &mw->dma_addr, GFP_KERNEL);
708 if (!mw->virt_addr) {
711 dev_err(&pdev->dev, "Unable to alloc MW buff of size %zu\n",
717 * we must ensure that the memory address allocated is BAR size
718 * aligned in order for the XLAT register to take the value. This
719 * is a requirement of the hardware. It is recommended to setup CMA
720 * for BAR sizes equal or greater than 4MB.
722 if (!IS_ALIGNED(mw->dma_addr, mw->xlat_align)) {
723 dev_err(&pdev->dev, "DMA memory %pad is not aligned\n",
725 ntb_free_mw(nt, num_mw);
729 /* Notify HW the memory location of the receive buffer */
730 rc = ntb_mw_set_trans(nt->ndev, num_mw, mw->dma_addr, mw->xlat_size);
732 dev_err(&pdev->dev, "Unable to set mw%d translation", num_mw);
733 ntb_free_mw(nt, num_mw);
740 static void ntb_qp_link_down_reset(struct ntb_transport_qp *qp)
742 qp->link_is_up = false;
749 qp->rx_ring_empty = 0;
750 qp->rx_err_no_buf = 0;
751 qp->rx_err_oflow = 0;
757 qp->tx_ring_full = 0;
758 qp->tx_err_no_buf = 0;
763 static void ntb_qp_link_cleanup(struct ntb_transport_qp *qp)
765 struct ntb_transport_ctx *nt = qp->transport;
766 struct pci_dev *pdev = nt->ndev->pdev;
768 dev_info(&pdev->dev, "qp %d: Link Cleanup\n", qp->qp_num);
770 cancel_delayed_work_sync(&qp->link_work);
771 ntb_qp_link_down_reset(qp);
773 if (qp->event_handler)
774 qp->event_handler(qp->cb_data, qp->link_is_up);
777 static void ntb_qp_link_cleanup_work(struct work_struct *work)
779 struct ntb_transport_qp *qp = container_of(work,
780 struct ntb_transport_qp,
782 struct ntb_transport_ctx *nt = qp->transport;
784 ntb_qp_link_cleanup(qp);
787 schedule_delayed_work(&qp->link_work,
788 msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
791 static void ntb_qp_link_down(struct ntb_transport_qp *qp)
793 schedule_work(&qp->link_cleanup);
796 static void ntb_transport_link_cleanup(struct ntb_transport_ctx *nt)
798 struct ntb_transport_qp *qp;
800 unsigned int i, count;
802 qp_bitmap_alloc = nt->qp_bitmap & ~nt->qp_bitmap_free;
804 /* Pass along the info to any clients */
805 for (i = 0; i < nt->qp_count; i++)
806 if (qp_bitmap_alloc & BIT_ULL(i)) {
808 ntb_qp_link_cleanup(qp);
809 cancel_work_sync(&qp->link_cleanup);
810 cancel_delayed_work_sync(&qp->link_work);
814 cancel_delayed_work_sync(&nt->link_work);
816 /* The scratchpad registers keep the values if the remote side
817 * goes down, blast them now to give them a sane value the next
818 * time they are accessed
820 count = ntb_spad_count(nt->ndev);
821 for (i = 0; i < count; i++)
822 ntb_spad_write(nt->ndev, i, 0);
825 static void ntb_transport_link_cleanup_work(struct work_struct *work)
827 struct ntb_transport_ctx *nt =
828 container_of(work, struct ntb_transport_ctx, link_cleanup);
830 ntb_transport_link_cleanup(nt);
833 static void ntb_transport_event_callback(void *data)
835 struct ntb_transport_ctx *nt = data;
837 if (ntb_link_is_up(nt->ndev, NULL, NULL) == 1)
838 schedule_delayed_work(&nt->link_work, 0);
840 schedule_work(&nt->link_cleanup);
843 static void ntb_transport_link_work(struct work_struct *work)
845 struct ntb_transport_ctx *nt =
846 container_of(work, struct ntb_transport_ctx, link_work.work);
847 struct ntb_dev *ndev = nt->ndev;
848 struct pci_dev *pdev = ndev->pdev;
849 resource_size_t size;
853 /* send the local info, in the opposite order of the way we read it */
854 for (i = 0; i < nt->mw_count; i++) {
855 size = nt->mw_vec[i].phys_size;
857 if (max_mw_size && size > max_mw_size)
860 spad = MW0_SZ_HIGH + (i * 2);
861 ntb_peer_spad_write(ndev, spad, upper_32_bits(size));
863 spad = MW0_SZ_LOW + (i * 2);
864 ntb_peer_spad_write(ndev, spad, lower_32_bits(size));
867 ntb_peer_spad_write(ndev, NUM_MWS, nt->mw_count);
869 ntb_peer_spad_write(ndev, NUM_QPS, nt->qp_count);
871 ntb_peer_spad_write(ndev, VERSION, NTB_TRANSPORT_VERSION);
873 /* Query the remote side for its info */
874 val = ntb_spad_read(ndev, VERSION);
875 dev_dbg(&pdev->dev, "Remote version = %d\n", val);
876 if (val != NTB_TRANSPORT_VERSION)
879 val = ntb_spad_read(ndev, NUM_QPS);
880 dev_dbg(&pdev->dev, "Remote max number of qps = %d\n", val);
881 if (val != nt->qp_count)
884 val = ntb_spad_read(ndev, NUM_MWS);
885 dev_dbg(&pdev->dev, "Remote number of mws = %d\n", val);
886 if (val != nt->mw_count)
889 for (i = 0; i < nt->mw_count; i++) {
892 val = ntb_spad_read(ndev, MW0_SZ_HIGH + (i * 2));
893 val64 = (u64)val << 32;
895 val = ntb_spad_read(ndev, MW0_SZ_LOW + (i * 2));
898 dev_dbg(&pdev->dev, "Remote MW%d size = %#llx\n", i, val64);
900 rc = ntb_set_mw(nt, i, val64);
905 nt->link_is_up = true;
907 for (i = 0; i < nt->qp_count; i++) {
908 struct ntb_transport_qp *qp = &nt->qp_vec[i];
910 ntb_transport_setup_qp_mw(nt, i);
912 if (qp->client_ready)
913 schedule_delayed_work(&qp->link_work, 0);
919 for (i = 0; i < nt->mw_count; i++)
922 /* if there's an actual failure, we should just bail */
924 ntb_link_disable(ndev);
929 if (ntb_link_is_up(ndev, NULL, NULL) == 1)
930 schedule_delayed_work(&nt->link_work,
931 msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
934 static void ntb_qp_link_work(struct work_struct *work)
936 struct ntb_transport_qp *qp = container_of(work,
937 struct ntb_transport_qp,
939 struct pci_dev *pdev = qp->ndev->pdev;
940 struct ntb_transport_ctx *nt = qp->transport;
943 WARN_ON(!nt->link_is_up);
945 val = ntb_spad_read(nt->ndev, QP_LINKS);
947 ntb_peer_spad_write(nt->ndev, QP_LINKS, val | BIT(qp->qp_num));
949 /* query remote spad for qp ready bits */
950 dev_dbg_ratelimited(&pdev->dev, "Remote QP link status = %x\n", val);
952 /* See if the remote side is up */
953 if (val & BIT(qp->qp_num)) {
954 dev_info(&pdev->dev, "qp %d: Link Up\n", qp->qp_num);
955 qp->link_is_up = true;
958 if (qp->event_handler)
959 qp->event_handler(qp->cb_data, qp->link_is_up);
962 tasklet_schedule(&qp->rxc_db_work);
963 } else if (nt->link_is_up)
964 schedule_delayed_work(&qp->link_work,
965 msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
968 static int ntb_transport_init_queue(struct ntb_transport_ctx *nt,
971 struct ntb_transport_qp *qp;
973 resource_size_t mw_size;
974 unsigned int num_qps_mw, tx_size;
975 unsigned int mw_num, mw_count, qp_count;
978 mw_count = nt->mw_count;
979 qp_count = nt->qp_count;
981 mw_num = QP_TO_MW(nt, qp_num);
983 qp = &nt->qp_vec[qp_num];
987 qp->client_ready = false;
988 qp->event_handler = NULL;
989 ntb_qp_link_down_reset(qp);
991 if (mw_num < qp_count % mw_count)
992 num_qps_mw = qp_count / mw_count + 1;
994 num_qps_mw = qp_count / mw_count;
996 mw_base = nt->mw_vec[mw_num].phys_addr;
997 mw_size = nt->mw_vec[mw_num].phys_size;
999 tx_size = (unsigned int)mw_size / num_qps_mw;
1000 qp_offset = tx_size * (qp_num / mw_count);
1002 qp->tx_mw = nt->mw_vec[mw_num].vbase + qp_offset;
1006 qp->tx_mw_phys = mw_base + qp_offset;
1007 if (!qp->tx_mw_phys)
1010 tx_size -= sizeof(struct ntb_rx_info);
1011 qp->rx_info = qp->tx_mw + tx_size;
1013 /* Due to housekeeping, there must be atleast 2 buffs */
1014 qp->tx_max_frame = min(transport_mtu, tx_size / 2);
1015 qp->tx_max_entry = tx_size / qp->tx_max_frame;
1017 if (nt->debugfs_node_dir) {
1018 char debugfs_name[4];
1020 snprintf(debugfs_name, 4, "qp%d", qp_num);
1021 qp->debugfs_dir = debugfs_create_dir(debugfs_name,
1022 nt->debugfs_node_dir);
1024 qp->debugfs_stats = debugfs_create_file("stats", S_IRUSR,
1025 qp->debugfs_dir, qp,
1026 &ntb_qp_debugfs_stats);
1028 qp->debugfs_dir = NULL;
1029 qp->debugfs_stats = NULL;
1032 INIT_DELAYED_WORK(&qp->link_work, ntb_qp_link_work);
1033 INIT_WORK(&qp->link_cleanup, ntb_qp_link_cleanup_work);
1035 spin_lock_init(&qp->ntb_rx_q_lock);
1036 spin_lock_init(&qp->ntb_tx_free_q_lock);
1038 INIT_LIST_HEAD(&qp->rx_post_q);
1039 INIT_LIST_HEAD(&qp->rx_pend_q);
1040 INIT_LIST_HEAD(&qp->rx_free_q);
1041 INIT_LIST_HEAD(&qp->tx_free_q);
1043 tasklet_init(&qp->rxc_db_work, ntb_transport_rxc_db,
1049 static int ntb_transport_probe(struct ntb_client *self, struct ntb_dev *ndev)
1051 struct ntb_transport_ctx *nt;
1052 struct ntb_transport_mw *mw;
1053 unsigned int mw_count, qp_count, spad_count, max_mw_count_for_spads;
1058 mw_count = ntb_mw_count(ndev);
1060 if (ntb_db_is_unsafe(ndev))
1062 "doorbell is unsafe, proceed anyway...\n");
1063 if (ntb_spad_is_unsafe(ndev))
1065 "scratchpad is unsafe, proceed anyway...\n");
1067 node = dev_to_node(&ndev->dev);
1069 nt = kzalloc_node(sizeof(*nt), GFP_KERNEL, node);
1074 spad_count = ntb_spad_count(ndev);
1076 /* Limit the MW's based on the availability of scratchpads */
1078 if (spad_count < NTB_TRANSPORT_MIN_SPADS) {
1084 max_mw_count_for_spads = (spad_count - MW0_SZ_HIGH) / 2;
1085 nt->mw_count = min(mw_count, max_mw_count_for_spads);
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 (nt->mw_count < qp_count)
1120 qp_count = nt->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->rx_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,
1251 const struct dmaengine_result *res)
1253 struct ntb_queue_entry *entry = data;
1255 /* we need to check DMA results if we are using DMA */
1257 enum dmaengine_tx_result dma_err = res->result;
1260 case DMA_TRANS_READ_FAILED:
1261 case DMA_TRANS_WRITE_FAILED:
1263 case DMA_TRANS_ABORTED:
1265 struct ntb_transport_qp *qp = entry->qp;
1266 void *offset = qp->rx_buff + qp->rx_max_frame *
1269 ntb_memcpy_rx(entry, offset);
1274 case DMA_TRANS_NOERROR:
1280 entry->flags |= DESC_DONE_FLAG;
1282 ntb_complete_rxc(entry->qp);
1285 static void ntb_memcpy_rx(struct ntb_queue_entry *entry, void *offset)
1287 void *buf = entry->buf;
1288 size_t len = entry->len;
1290 memcpy(buf, offset, len);
1292 /* Ensure that the data is fully copied out before clearing the flag */
1295 ntb_rx_copy_callback(entry, NULL);
1298 static int ntb_async_rx_submit(struct ntb_queue_entry *entry, void *offset)
1300 struct dma_async_tx_descriptor *txd;
1301 struct ntb_transport_qp *qp = entry->qp;
1302 struct dma_chan *chan = qp->rx_dma_chan;
1303 struct dma_device *device;
1304 size_t pay_off, buff_off, len;
1305 struct dmaengine_unmap_data *unmap;
1306 dma_cookie_t cookie;
1307 void *buf = entry->buf;
1310 device = chan->device;
1311 pay_off = (size_t)offset & ~PAGE_MASK;
1312 buff_off = (size_t)buf & ~PAGE_MASK;
1314 if (!is_dma_copy_aligned(device, pay_off, buff_off, len))
1317 unmap = dmaengine_get_unmap_data(device->dev, 2, GFP_NOWAIT);
1322 unmap->addr[0] = dma_map_page(device->dev, virt_to_page(offset),
1323 pay_off, len, DMA_TO_DEVICE);
1324 if (dma_mapping_error(device->dev, unmap->addr[0]))
1329 unmap->addr[1] = dma_map_page(device->dev, virt_to_page(buf),
1330 buff_off, len, DMA_FROM_DEVICE);
1331 if (dma_mapping_error(device->dev, unmap->addr[1]))
1334 unmap->from_cnt = 1;
1336 txd = device->device_prep_dma_memcpy(chan, unmap->addr[1],
1337 unmap->addr[0], len,
1338 DMA_PREP_INTERRUPT);
1342 txd->callback_result = ntb_rx_copy_callback;
1343 txd->callback_param = entry;
1344 dma_set_unmap(txd, unmap);
1346 cookie = dmaengine_submit(txd);
1347 if (dma_submit_error(cookie))
1350 dmaengine_unmap_put(unmap);
1352 qp->last_cookie = cookie;
1359 dmaengine_unmap_put(unmap);
1361 dmaengine_unmap_put(unmap);
1366 static void ntb_async_rx(struct ntb_queue_entry *entry, void *offset)
1368 struct ntb_transport_qp *qp = entry->qp;
1369 struct dma_chan *chan = qp->rx_dma_chan;
1375 if (entry->len < copy_bytes)
1378 res = ntb_async_rx_submit(entry, offset);
1382 if (!entry->retries)
1388 ntb_memcpy_rx(entry, offset);
1392 static int ntb_process_rxc(struct ntb_transport_qp *qp)
1394 struct ntb_payload_header *hdr;
1395 struct ntb_queue_entry *entry;
1398 offset = qp->rx_buff + qp->rx_max_frame * qp->rx_index;
1399 hdr = offset + qp->rx_max_frame - sizeof(struct ntb_payload_header);
1401 dev_dbg(&qp->ndev->pdev->dev, "qp %d: RX ver %u len %d flags %x\n",
1402 qp->qp_num, hdr->ver, hdr->len, hdr->flags);
1404 if (!(hdr->flags & DESC_DONE_FLAG)) {
1405 dev_dbg(&qp->ndev->pdev->dev, "done flag not set\n");
1406 qp->rx_ring_empty++;
1410 if (hdr->flags & LINK_DOWN_FLAG) {
1411 dev_dbg(&qp->ndev->pdev->dev, "link down flag set\n");
1412 ntb_qp_link_down(qp);
1417 if (hdr->ver != (u32)qp->rx_pkts) {
1418 dev_dbg(&qp->ndev->pdev->dev,
1419 "version mismatch, expected %llu - got %u\n",
1420 qp->rx_pkts, hdr->ver);
1425 entry = ntb_list_mv(&qp->ntb_rx_q_lock, &qp->rx_pend_q, &qp->rx_post_q);
1427 dev_dbg(&qp->ndev->pdev->dev, "no receive buffer\n");
1428 qp->rx_err_no_buf++;
1432 entry->rx_hdr = hdr;
1433 entry->rx_index = qp->rx_index;
1435 if (hdr->len > entry->len) {
1436 dev_dbg(&qp->ndev->pdev->dev,
1437 "receive buffer overflow! Wanted %d got %d\n",
1438 hdr->len, entry->len);
1442 entry->flags |= DESC_DONE_FLAG;
1444 ntb_complete_rxc(qp);
1446 dev_dbg(&qp->ndev->pdev->dev,
1447 "RX OK index %u ver %u size %d into buf size %d\n",
1448 qp->rx_index, hdr->ver, hdr->len, entry->len);
1450 qp->rx_bytes += hdr->len;
1453 entry->len = hdr->len;
1455 ntb_async_rx(entry, offset);
1459 qp->rx_index %= qp->rx_max_entry;
1464 static void ntb_transport_rxc_db(unsigned long data)
1466 struct ntb_transport_qp *qp = (void *)data;
1469 dev_dbg(&qp->ndev->pdev->dev, "%s: doorbell %d received\n",
1470 __func__, qp->qp_num);
1472 /* Limit the number of packets processed in a single interrupt to
1473 * provide fairness to others
1475 for (i = 0; i < qp->rx_max_entry; i++) {
1476 rc = ntb_process_rxc(qp);
1481 if (i && qp->rx_dma_chan)
1482 dma_async_issue_pending(qp->rx_dma_chan);
1484 if (i == qp->rx_max_entry) {
1485 /* there is more work to do */
1487 tasklet_schedule(&qp->rxc_db_work);
1488 } else if (ntb_db_read(qp->ndev) & BIT_ULL(qp->qp_num)) {
1489 /* the doorbell bit is set: clear it */
1490 ntb_db_clear(qp->ndev, BIT_ULL(qp->qp_num));
1491 /* ntb_db_read ensures ntb_db_clear write is committed */
1492 ntb_db_read(qp->ndev);
1494 /* an interrupt may have arrived between finishing
1495 * ntb_process_rxc and clearing the doorbell bit:
1496 * there might be some more work to do.
1499 tasklet_schedule(&qp->rxc_db_work);
1503 static void ntb_tx_copy_callback(void *data,
1504 const struct dmaengine_result *res)
1506 struct ntb_queue_entry *entry = data;
1507 struct ntb_transport_qp *qp = entry->qp;
1508 struct ntb_payload_header __iomem *hdr = entry->tx_hdr;
1510 /* we need to check DMA results if we are using DMA */
1512 enum dmaengine_tx_result dma_err = res->result;
1515 case DMA_TRANS_READ_FAILED:
1516 case DMA_TRANS_WRITE_FAILED:
1518 case DMA_TRANS_ABORTED:
1520 void __iomem *offset =
1521 qp->tx_mw + qp->tx_max_frame *
1524 /* resubmit via CPU */
1525 ntb_memcpy_tx(entry, offset);
1530 case DMA_TRANS_NOERROR:
1536 iowrite32(entry->flags | DESC_DONE_FLAG, &hdr->flags);
1538 ntb_peer_db_set(qp->ndev, BIT_ULL(qp->qp_num));
1540 /* The entry length can only be zero if the packet is intended to be a
1541 * "link down" or similar. Since no payload is being sent in these
1542 * cases, there is nothing to add to the completion queue.
1544 if (entry->len > 0) {
1545 qp->tx_bytes += entry->len;
1548 qp->tx_handler(qp, qp->cb_data, entry->cb_data,
1552 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry, &qp->tx_free_q);
1555 static void ntb_memcpy_tx(struct ntb_queue_entry *entry, void __iomem *offset)
1557 #ifdef ARCH_HAS_NOCACHE_UACCESS
1559 * Using non-temporal mov to improve performance on non-cached
1560 * writes, even though we aren't actually copying from user space.
1562 __copy_from_user_inatomic_nocache(offset, entry->buf, entry->len);
1564 memcpy_toio(offset, entry->buf, entry->len);
1567 /* Ensure that the data is fully copied out before setting the flags */
1570 ntb_tx_copy_callback(entry, NULL);
1573 static int ntb_async_tx_submit(struct ntb_transport_qp *qp,
1574 struct ntb_queue_entry *entry)
1576 struct dma_async_tx_descriptor *txd;
1577 struct dma_chan *chan = qp->tx_dma_chan;
1578 struct dma_device *device;
1579 size_t len = entry->len;
1580 void *buf = entry->buf;
1581 size_t dest_off, buff_off;
1582 struct dmaengine_unmap_data *unmap;
1584 dma_cookie_t cookie;
1586 device = chan->device;
1587 dest = qp->tx_mw_phys + qp->tx_max_frame * entry->tx_index;
1588 buff_off = (size_t)buf & ~PAGE_MASK;
1589 dest_off = (size_t)dest & ~PAGE_MASK;
1591 if (!is_dma_copy_aligned(device, buff_off, dest_off, len))
1594 unmap = dmaengine_get_unmap_data(device->dev, 1, GFP_NOWAIT);
1599 unmap->addr[0] = dma_map_page(device->dev, virt_to_page(buf),
1600 buff_off, len, DMA_TO_DEVICE);
1601 if (dma_mapping_error(device->dev, unmap->addr[0]))
1606 txd = device->device_prep_dma_memcpy(chan, dest, unmap->addr[0], len,
1607 DMA_PREP_INTERRUPT);
1611 txd->callback_result = ntb_tx_copy_callback;
1612 txd->callback_param = entry;
1613 dma_set_unmap(txd, unmap);
1615 cookie = dmaengine_submit(txd);
1616 if (dma_submit_error(cookie))
1619 dmaengine_unmap_put(unmap);
1621 dma_async_issue_pending(chan);
1625 dmaengine_unmap_put(unmap);
1627 dmaengine_unmap_put(unmap);
1632 static void ntb_async_tx(struct ntb_transport_qp *qp,
1633 struct ntb_queue_entry *entry)
1635 struct ntb_payload_header __iomem *hdr;
1636 struct dma_chan *chan = qp->tx_dma_chan;
1637 void __iomem *offset;
1640 entry->tx_index = qp->tx_index;
1641 offset = qp->tx_mw + qp->tx_max_frame * entry->tx_index;
1642 hdr = offset + qp->tx_max_frame - sizeof(struct ntb_payload_header);
1643 entry->tx_hdr = hdr;
1645 iowrite32(entry->len, &hdr->len);
1646 iowrite32((u32)qp->tx_pkts, &hdr->ver);
1651 if (entry->len < copy_bytes)
1654 res = ntb_async_tx_submit(qp, entry);
1658 if (!entry->retries)
1664 ntb_memcpy_tx(entry, offset);
1668 static int ntb_process_tx(struct ntb_transport_qp *qp,
1669 struct ntb_queue_entry *entry)
1671 if (qp->tx_index == qp->remote_rx_info->entry) {
1676 if (entry->len > qp->tx_max_frame - sizeof(struct ntb_payload_header)) {
1678 qp->tx_handler(qp, qp->cb_data, NULL, -EIO);
1680 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
1685 ntb_async_tx(qp, entry);
1688 qp->tx_index %= qp->tx_max_entry;
1695 static void ntb_send_link_down(struct ntb_transport_qp *qp)
1697 struct pci_dev *pdev = qp->ndev->pdev;
1698 struct ntb_queue_entry *entry;
1701 if (!qp->link_is_up)
1704 dev_info(&pdev->dev, "qp %d: Send Link Down\n", qp->qp_num);
1706 for (i = 0; i < NTB_LINK_DOWN_TIMEOUT; i++) {
1707 entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
1716 entry->cb_data = NULL;
1719 entry->flags = LINK_DOWN_FLAG;
1721 rc = ntb_process_tx(qp, entry);
1723 dev_err(&pdev->dev, "ntb: QP%d unable to send linkdown msg\n",
1726 ntb_qp_link_down_reset(qp);
1729 static bool ntb_dma_filter_fn(struct dma_chan *chan, void *node)
1731 return dev_to_node(&chan->dev->device) == (int)(unsigned long)node;
1735 * ntb_transport_create_queue - Create a new NTB transport layer queue
1736 * @rx_handler: receive callback function
1737 * @tx_handler: transmit callback function
1738 * @event_handler: event callback function
1740 * Create a new NTB transport layer queue and provide the queue with a callback
1741 * routine for both transmit and receive. The receive callback routine will be
1742 * used to pass up data when the transport has received it on the queue. The
1743 * transmit callback routine will be called when the transport has completed the
1744 * transmission of the data on the queue and the data is ready to be freed.
1746 * RETURNS: pointer to newly created ntb_queue, NULL on error.
1748 struct ntb_transport_qp *
1749 ntb_transport_create_queue(void *data, struct device *client_dev,
1750 const struct ntb_queue_handlers *handlers)
1752 struct ntb_dev *ndev;
1753 struct pci_dev *pdev;
1754 struct ntb_transport_ctx *nt;
1755 struct ntb_queue_entry *entry;
1756 struct ntb_transport_qp *qp;
1758 unsigned int free_queue;
1759 dma_cap_mask_t dma_mask;
1763 ndev = dev_ntb(client_dev->parent);
1767 node = dev_to_node(&ndev->dev);
1769 free_queue = ffs(nt->qp_bitmap_free);
1773 /* decrement free_queue to make it zero based */
1776 qp = &nt->qp_vec[free_queue];
1777 qp_bit = BIT_ULL(qp->qp_num);
1779 nt->qp_bitmap_free &= ~qp_bit;
1782 qp->rx_handler = handlers->rx_handler;
1783 qp->tx_handler = handlers->tx_handler;
1784 qp->event_handler = handlers->event_handler;
1786 dma_cap_zero(dma_mask);
1787 dma_cap_set(DMA_MEMCPY, dma_mask);
1791 dma_request_channel(dma_mask, ntb_dma_filter_fn,
1792 (void *)(unsigned long)node);
1793 if (!qp->tx_dma_chan)
1794 dev_info(&pdev->dev, "Unable to allocate TX DMA channel\n");
1797 dma_request_channel(dma_mask, ntb_dma_filter_fn,
1798 (void *)(unsigned long)node);
1799 if (!qp->rx_dma_chan)
1800 dev_info(&pdev->dev, "Unable to allocate RX DMA channel\n");
1802 qp->tx_dma_chan = NULL;
1803 qp->rx_dma_chan = NULL;
1806 dev_dbg(&pdev->dev, "Using %s memcpy for TX\n",
1807 qp->tx_dma_chan ? "DMA" : "CPU");
1809 dev_dbg(&pdev->dev, "Using %s memcpy for RX\n",
1810 qp->rx_dma_chan ? "DMA" : "CPU");
1812 for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) {
1813 entry = kzalloc_node(sizeof(*entry), GFP_ATOMIC, node);
1818 ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry,
1821 qp->rx_alloc_entry = NTB_QP_DEF_NUM_ENTRIES;
1823 for (i = 0; i < qp->tx_max_entry; i++) {
1824 entry = kzalloc_node(sizeof(*entry), GFP_ATOMIC, node);
1829 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
1833 ntb_db_clear(qp->ndev, qp_bit);
1834 ntb_db_clear_mask(qp->ndev, qp_bit);
1836 dev_info(&pdev->dev, "NTB Transport QP %d created\n", qp->qp_num);
1841 while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
1844 qp->rx_alloc_entry = 0;
1845 while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q)))
1847 if (qp->tx_dma_chan)
1848 dma_release_channel(qp->tx_dma_chan);
1849 if (qp->rx_dma_chan)
1850 dma_release_channel(qp->rx_dma_chan);
1851 nt->qp_bitmap_free |= qp_bit;
1855 EXPORT_SYMBOL_GPL(ntb_transport_create_queue);
1858 * ntb_transport_free_queue - Frees NTB transport queue
1859 * @qp: NTB queue to be freed
1861 * Frees NTB transport queue
1863 void ntb_transport_free_queue(struct ntb_transport_qp *qp)
1865 struct pci_dev *pdev;
1866 struct ntb_queue_entry *entry;
1872 pdev = qp->ndev->pdev;
1876 if (qp->tx_dma_chan) {
1877 struct dma_chan *chan = qp->tx_dma_chan;
1878 /* Putting the dma_chan to NULL will force any new traffic to be
1879 * processed by the CPU instead of the DAM engine
1881 qp->tx_dma_chan = NULL;
1883 /* Try to be nice and wait for any queued DMA engine
1884 * transactions to process before smashing it with a rock
1886 dma_sync_wait(chan, qp->last_cookie);
1887 dmaengine_terminate_all(chan);
1888 dma_release_channel(chan);
1891 if (qp->rx_dma_chan) {
1892 struct dma_chan *chan = qp->rx_dma_chan;
1893 /* Putting the dma_chan to NULL will force any new traffic to be
1894 * processed by the CPU instead of the DAM engine
1896 qp->rx_dma_chan = NULL;
1898 /* Try to be nice and wait for any queued DMA engine
1899 * transactions to process before smashing it with a rock
1901 dma_sync_wait(chan, qp->last_cookie);
1902 dmaengine_terminate_all(chan);
1903 dma_release_channel(chan);
1906 qp_bit = BIT_ULL(qp->qp_num);
1908 ntb_db_set_mask(qp->ndev, qp_bit);
1909 tasklet_kill(&qp->rxc_db_work);
1911 cancel_delayed_work_sync(&qp->link_work);
1914 qp->rx_handler = NULL;
1915 qp->tx_handler = NULL;
1916 qp->event_handler = NULL;
1918 while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q)))
1921 while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_pend_q))) {
1922 dev_warn(&pdev->dev, "Freeing item from non-empty rx_pend_q\n");
1926 while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_post_q))) {
1927 dev_warn(&pdev->dev, "Freeing item from non-empty rx_post_q\n");
1931 while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
1934 qp->transport->qp_bitmap_free |= qp_bit;
1936 dev_info(&pdev->dev, "NTB Transport QP %d freed\n", qp->qp_num);
1938 EXPORT_SYMBOL_GPL(ntb_transport_free_queue);
1941 * ntb_transport_rx_remove - Dequeues enqueued rx packet
1942 * @qp: NTB queue to be freed
1943 * @len: pointer to variable to write enqueued buffers length
1945 * Dequeues unused buffers from receive queue. Should only be used during
1948 * RETURNS: NULL error value on error, or void* for success.
1950 void *ntb_transport_rx_remove(struct ntb_transport_qp *qp, unsigned int *len)
1952 struct ntb_queue_entry *entry;
1955 if (!qp || qp->client_ready)
1958 entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_pend_q);
1962 buf = entry->cb_data;
1965 ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry, &qp->rx_free_q);
1969 EXPORT_SYMBOL_GPL(ntb_transport_rx_remove);
1972 * ntb_transport_rx_enqueue - Enqueue a new NTB queue entry
1973 * @qp: NTB transport layer queue the entry is to be enqueued on
1974 * @cb: per buffer pointer for callback function to use
1975 * @data: pointer to data buffer that incoming packets will be copied into
1976 * @len: length of the data buffer
1978 * Enqueue a new receive buffer onto the transport queue into which a NTB
1979 * payload can be received into.
1981 * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
1983 int ntb_transport_rx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
1986 struct ntb_queue_entry *entry;
1991 entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q);
1995 entry->cb_data = cb;
2001 entry->rx_index = 0;
2003 ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry, &qp->rx_pend_q);
2006 tasklet_schedule(&qp->rxc_db_work);
2010 EXPORT_SYMBOL_GPL(ntb_transport_rx_enqueue);
2013 * ntb_transport_tx_enqueue - Enqueue a new NTB queue entry
2014 * @qp: NTB transport layer queue the entry is to be enqueued on
2015 * @cb: per buffer pointer for callback function to use
2016 * @data: pointer to data buffer that will be sent
2017 * @len: length of the data buffer
2019 * Enqueue a new transmit buffer onto the transport queue from which a NTB
2020 * payload will be transmitted. This assumes that a lock is being held to
2021 * serialize access to the qp.
2023 * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
2025 int ntb_transport_tx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
2028 struct ntb_queue_entry *entry;
2031 if (!qp || !qp->link_is_up || !len)
2034 entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
2036 qp->tx_err_no_buf++;
2040 entry->cb_data = cb;
2046 entry->tx_index = 0;
2048 rc = ntb_process_tx(qp, entry);
2050 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
2055 EXPORT_SYMBOL_GPL(ntb_transport_tx_enqueue);
2058 * ntb_transport_link_up - Notify NTB transport of client readiness to use queue
2059 * @qp: NTB transport layer queue to be enabled
2061 * Notify NTB transport layer of client readiness to use queue
2063 void ntb_transport_link_up(struct ntb_transport_qp *qp)
2068 qp->client_ready = true;
2070 if (qp->transport->link_is_up)
2071 schedule_delayed_work(&qp->link_work, 0);
2073 EXPORT_SYMBOL_GPL(ntb_transport_link_up);
2076 * ntb_transport_link_down - Notify NTB transport to no longer enqueue data
2077 * @qp: NTB transport layer queue to be disabled
2079 * Notify NTB transport layer of client's desire to no longer receive data on
2080 * transport queue specified. It is the client's responsibility to ensure all
2081 * entries on queue are purged or otherwise handled appropriately.
2083 void ntb_transport_link_down(struct ntb_transport_qp *qp)
2090 qp->client_ready = false;
2092 val = ntb_spad_read(qp->ndev, QP_LINKS);
2094 ntb_peer_spad_write(qp->ndev, QP_LINKS,
2095 val & ~BIT(qp->qp_num));
2098 ntb_send_link_down(qp);
2100 cancel_delayed_work_sync(&qp->link_work);
2102 EXPORT_SYMBOL_GPL(ntb_transport_link_down);
2105 * ntb_transport_link_query - Query transport link state
2106 * @qp: NTB transport layer queue to be queried
2108 * Query connectivity to the remote system of the NTB transport queue
2110 * RETURNS: true for link up or false for link down
2112 bool ntb_transport_link_query(struct ntb_transport_qp *qp)
2117 return qp->link_is_up;
2119 EXPORT_SYMBOL_GPL(ntb_transport_link_query);
2122 * ntb_transport_qp_num - Query the qp number
2123 * @qp: NTB transport layer queue to be queried
2125 * Query qp number of the NTB transport queue
2127 * RETURNS: a zero based number specifying the qp number
2129 unsigned char ntb_transport_qp_num(struct ntb_transport_qp *qp)
2136 EXPORT_SYMBOL_GPL(ntb_transport_qp_num);
2139 * ntb_transport_max_size - Query the max payload size of a qp
2140 * @qp: NTB transport layer queue to be queried
2142 * Query the maximum payload size permissible on the given qp
2144 * RETURNS: the max payload size of a qp
2146 unsigned int ntb_transport_max_size(struct ntb_transport_qp *qp)
2148 unsigned int max_size;
2149 unsigned int copy_align;
2150 struct dma_chan *rx_chan, *tx_chan;
2155 rx_chan = qp->rx_dma_chan;
2156 tx_chan = qp->tx_dma_chan;
2158 copy_align = max(rx_chan ? rx_chan->device->copy_align : 0,
2159 tx_chan ? tx_chan->device->copy_align : 0);
2161 /* If DMA engine usage is possible, try to find the max size for that */
2162 max_size = qp->tx_max_frame - sizeof(struct ntb_payload_header);
2163 max_size = round_down(max_size, 1 << copy_align);
2167 EXPORT_SYMBOL_GPL(ntb_transport_max_size);
2169 unsigned int ntb_transport_tx_free_entry(struct ntb_transport_qp *qp)
2171 unsigned int head = qp->tx_index;
2172 unsigned int tail = qp->remote_rx_info->entry;
2174 return tail > head ? tail - head : qp->tx_max_entry + tail - head;
2176 EXPORT_SYMBOL_GPL(ntb_transport_tx_free_entry);
2178 static void ntb_transport_doorbell_callback(void *data, int vector)
2180 struct ntb_transport_ctx *nt = data;
2181 struct ntb_transport_qp *qp;
2183 unsigned int qp_num;
2185 db_bits = (nt->qp_bitmap & ~nt->qp_bitmap_free &
2186 ntb_db_vector_mask(nt->ndev, vector));
2189 qp_num = __ffs(db_bits);
2190 qp = &nt->qp_vec[qp_num];
2193 tasklet_schedule(&qp->rxc_db_work);
2195 db_bits &= ~BIT_ULL(qp_num);
2199 static const struct ntb_ctx_ops ntb_transport_ops = {
2200 .link_event = ntb_transport_event_callback,
2201 .db_event = ntb_transport_doorbell_callback,
2204 static struct ntb_client ntb_transport_client = {
2206 .probe = ntb_transport_probe,
2207 .remove = ntb_transport_free,
2211 static int __init ntb_transport_init(void)
2215 pr_info("%s, version %s\n", NTB_TRANSPORT_DESC, NTB_TRANSPORT_VER);
2217 if (debugfs_initialized())
2218 nt_debugfs_dir = debugfs_create_dir(KBUILD_MODNAME, NULL);
2220 rc = bus_register(&ntb_transport_bus);
2224 rc = ntb_register_client(&ntb_transport_client);
2231 bus_unregister(&ntb_transport_bus);
2233 debugfs_remove_recursive(nt_debugfs_dir);
2236 module_init(ntb_transport_init);
2238 static void __exit ntb_transport_exit(void)
2240 ntb_unregister_client(&ntb_transport_client);
2241 bus_unregister(&ntb_transport_bus);
2242 debugfs_remove_recursive(nt_debugfs_dir);
2244 module_exit(ntb_transport_exit);