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");
98 module_param(use_msi, bool, 0644);
99 MODULE_PARM_DESC(use_msi, "Use MSI interrupts instead of doorbells");
102 static struct dentry *nt_debugfs_dir;
104 /* Only two-ports NTB devices are supported */
105 #define PIDX NTB_DEF_PEER_IDX
107 struct ntb_queue_entry {
108 /* ntb_queue list reference */
109 struct list_head entry;
110 /* pointers to data to be transferred */
117 unsigned int tx_index;
118 unsigned int rx_index;
120 struct ntb_transport_qp *qp;
122 struct ntb_payload_header __iomem *tx_hdr;
123 struct ntb_payload_header *rx_hdr;
131 struct ntb_transport_qp {
132 struct ntb_transport_ctx *transport;
133 struct ntb_dev *ndev;
135 struct dma_chan *tx_dma_chan;
136 struct dma_chan *rx_dma_chan;
142 u8 qp_num; /* Only 64 QP's are allowed. 0-63 */
145 struct ntb_rx_info __iomem *rx_info;
146 struct ntb_rx_info *remote_rx_info;
148 void (*tx_handler)(struct ntb_transport_qp *qp, void *qp_data,
149 void *data, int len);
150 struct list_head tx_free_q;
151 spinlock_t ntb_tx_free_q_lock;
153 phys_addr_t tx_mw_phys;
155 dma_addr_t tx_mw_dma_addr;
156 unsigned int tx_index;
157 unsigned int tx_max_entry;
158 unsigned int tx_max_frame;
160 void (*rx_handler)(struct ntb_transport_qp *qp, void *qp_data,
161 void *data, int len);
162 struct list_head rx_post_q;
163 struct list_head rx_pend_q;
164 struct list_head rx_free_q;
165 /* ntb_rx_q_lock: synchronize access to rx_XXXX_q */
166 spinlock_t ntb_rx_q_lock;
168 unsigned int rx_index;
169 unsigned int rx_max_entry;
170 unsigned int rx_max_frame;
171 unsigned int rx_alloc_entry;
172 dma_cookie_t last_cookie;
173 struct tasklet_struct rxc_db_work;
175 void (*event_handler)(void *data, int status);
176 struct delayed_work link_work;
177 struct work_struct link_cleanup;
179 struct dentry *debugfs_dir;
180 struct dentry *debugfs_stats;
200 struct ntb_msi_desc msi_desc;
201 struct ntb_msi_desc peer_msi_desc;
204 struct ntb_transport_mw {
205 phys_addr_t phys_addr;
206 resource_size_t phys_size;
216 struct ntb_transport_client_dev {
217 struct list_head entry;
218 struct ntb_transport_ctx *nt;
222 struct ntb_transport_ctx {
223 struct list_head entry;
224 struct list_head client_devs;
226 struct ntb_dev *ndev;
228 struct ntb_transport_mw *mw_vec;
229 struct ntb_transport_qp *qp_vec;
230 unsigned int mw_count;
231 unsigned int qp_count;
236 unsigned int msi_spad_offset;
240 struct delayed_work link_work;
241 struct work_struct link_cleanup;
243 struct dentry *debugfs_node_dir;
247 DESC_DONE_FLAG = BIT(0),
248 LINK_DOWN_FLAG = BIT(1),
251 struct ntb_payload_header {
266 #define dev_client_dev(__dev) \
267 container_of((__dev), struct ntb_transport_client_dev, dev)
269 #define drv_client(__drv) \
270 container_of((__drv), struct ntb_transport_client, driver)
272 #define QP_TO_MW(nt, qp) ((qp) % nt->mw_count)
273 #define NTB_QP_DEF_NUM_ENTRIES 100
274 #define NTB_LINK_DOWN_TIMEOUT 10
276 static void ntb_transport_rxc_db(unsigned long data);
277 static const struct ntb_ctx_ops ntb_transport_ops;
278 static struct ntb_client ntb_transport_client;
279 static int ntb_async_tx_submit(struct ntb_transport_qp *qp,
280 struct ntb_queue_entry *entry);
281 static void ntb_memcpy_tx(struct ntb_queue_entry *entry, void __iomem *offset);
282 static int ntb_async_rx_submit(struct ntb_queue_entry *entry, void *offset);
283 static void ntb_memcpy_rx(struct ntb_queue_entry *entry, void *offset);
286 static int ntb_transport_bus_match(struct device *dev,
287 struct device_driver *drv)
289 return !strncmp(dev_name(dev), drv->name, strlen(drv->name));
292 static int ntb_transport_bus_probe(struct device *dev)
294 const struct ntb_transport_client *client;
299 client = drv_client(dev->driver);
300 rc = client->probe(dev);
307 static int ntb_transport_bus_remove(struct device *dev)
309 const struct ntb_transport_client *client;
311 client = drv_client(dev->driver);
319 static struct bus_type ntb_transport_bus = {
320 .name = "ntb_transport",
321 .match = ntb_transport_bus_match,
322 .probe = ntb_transport_bus_probe,
323 .remove = ntb_transport_bus_remove,
326 static LIST_HEAD(ntb_transport_list);
328 static int ntb_bus_init(struct ntb_transport_ctx *nt)
330 list_add_tail(&nt->entry, &ntb_transport_list);
334 static void ntb_bus_remove(struct ntb_transport_ctx *nt)
336 struct ntb_transport_client_dev *client_dev, *cd;
338 list_for_each_entry_safe(client_dev, cd, &nt->client_devs, entry) {
339 dev_err(client_dev->dev.parent, "%s still attached to bus, removing\n",
340 dev_name(&client_dev->dev));
341 list_del(&client_dev->entry);
342 device_unregister(&client_dev->dev);
345 list_del(&nt->entry);
348 static void ntb_transport_client_release(struct device *dev)
350 struct ntb_transport_client_dev *client_dev;
352 client_dev = dev_client_dev(dev);
357 * ntb_transport_unregister_client_dev - Unregister NTB client device
358 * @device_name: Name of NTB client device
360 * Unregister an NTB client device with the NTB transport layer
362 void ntb_transport_unregister_client_dev(char *device_name)
364 struct ntb_transport_client_dev *client, *cd;
365 struct ntb_transport_ctx *nt;
367 list_for_each_entry(nt, &ntb_transport_list, entry)
368 list_for_each_entry_safe(client, cd, &nt->client_devs, entry)
369 if (!strncmp(dev_name(&client->dev), device_name,
370 strlen(device_name))) {
371 list_del(&client->entry);
372 device_unregister(&client->dev);
375 EXPORT_SYMBOL_GPL(ntb_transport_unregister_client_dev);
378 * ntb_transport_register_client_dev - Register NTB client device
379 * @device_name: Name of NTB client device
381 * Register an NTB client device with the NTB transport layer
383 int ntb_transport_register_client_dev(char *device_name)
385 struct ntb_transport_client_dev *client_dev;
386 struct ntb_transport_ctx *nt;
390 if (list_empty(&ntb_transport_list))
393 list_for_each_entry(nt, &ntb_transport_list, entry) {
396 node = dev_to_node(&nt->ndev->dev);
398 client_dev = kzalloc_node(sizeof(*client_dev),
405 dev = &client_dev->dev;
407 /* setup and register client devices */
408 dev_set_name(dev, "%s%d", device_name, i);
409 dev->bus = &ntb_transport_bus;
410 dev->release = ntb_transport_client_release;
411 dev->parent = &nt->ndev->dev;
413 rc = device_register(dev);
419 list_add_tail(&client_dev->entry, &nt->client_devs);
426 ntb_transport_unregister_client_dev(device_name);
430 EXPORT_SYMBOL_GPL(ntb_transport_register_client_dev);
433 * ntb_transport_register_client - Register NTB client driver
434 * @drv: NTB client driver to be registered
436 * Register an NTB client driver with the NTB transport layer
438 * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
440 int ntb_transport_register_client(struct ntb_transport_client *drv)
442 drv->driver.bus = &ntb_transport_bus;
444 if (list_empty(&ntb_transport_list))
447 return driver_register(&drv->driver);
449 EXPORT_SYMBOL_GPL(ntb_transport_register_client);
452 * ntb_transport_unregister_client - Unregister NTB client driver
453 * @drv: NTB client driver to be unregistered
455 * Unregister an NTB client driver with the NTB transport layer
457 * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
459 void ntb_transport_unregister_client(struct ntb_transport_client *drv)
461 driver_unregister(&drv->driver);
463 EXPORT_SYMBOL_GPL(ntb_transport_unregister_client);
465 static ssize_t debugfs_read(struct file *filp, char __user *ubuf, size_t count,
468 struct ntb_transport_qp *qp;
470 ssize_t ret, out_offset, out_count;
472 qp = filp->private_data;
474 if (!qp || !qp->link_is_up)
479 buf = kmalloc(out_count, GFP_KERNEL);
484 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
485 "\nNTB QP stats:\n\n");
486 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
487 "rx_bytes - \t%llu\n", qp->rx_bytes);
488 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
489 "rx_pkts - \t%llu\n", qp->rx_pkts);
490 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
491 "rx_memcpy - \t%llu\n", qp->rx_memcpy);
492 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
493 "rx_async - \t%llu\n", qp->rx_async);
494 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
495 "rx_ring_empty - %llu\n", qp->rx_ring_empty);
496 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
497 "rx_err_no_buf - %llu\n", qp->rx_err_no_buf);
498 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
499 "rx_err_oflow - \t%llu\n", qp->rx_err_oflow);
500 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
501 "rx_err_ver - \t%llu\n", qp->rx_err_ver);
502 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
503 "rx_buff - \t0x%p\n", qp->rx_buff);
504 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
505 "rx_index - \t%u\n", qp->rx_index);
506 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
507 "rx_max_entry - \t%u\n", qp->rx_max_entry);
508 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
509 "rx_alloc_entry - \t%u\n\n", qp->rx_alloc_entry);
511 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
512 "tx_bytes - \t%llu\n", qp->tx_bytes);
513 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
514 "tx_pkts - \t%llu\n", qp->tx_pkts);
515 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
516 "tx_memcpy - \t%llu\n", qp->tx_memcpy);
517 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
518 "tx_async - \t%llu\n", qp->tx_async);
519 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
520 "tx_ring_full - \t%llu\n", qp->tx_ring_full);
521 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
522 "tx_err_no_buf - %llu\n", qp->tx_err_no_buf);
523 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
524 "tx_mw - \t0x%p\n", qp->tx_mw);
525 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
526 "tx_index (H) - \t%u\n", qp->tx_index);
527 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
529 qp->remote_rx_info->entry);
530 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
531 "tx_max_entry - \t%u\n", qp->tx_max_entry);
532 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
534 ntb_transport_tx_free_entry(qp));
536 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
538 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
539 "Using TX DMA - \t%s\n",
540 qp->tx_dma_chan ? "Yes" : "No");
541 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
542 "Using RX DMA - \t%s\n",
543 qp->rx_dma_chan ? "Yes" : "No");
544 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
546 qp->link_is_up ? "Up" : "Down");
547 out_offset += scnprintf(buf + out_offset, out_count - out_offset,
550 if (out_offset > out_count)
551 out_offset = out_count;
553 ret = simple_read_from_buffer(ubuf, count, offp, buf, out_offset);
558 static const struct file_operations ntb_qp_debugfs_stats = {
559 .owner = THIS_MODULE,
561 .read = debugfs_read,
564 static void ntb_list_add(spinlock_t *lock, struct list_head *entry,
565 struct list_head *list)
569 spin_lock_irqsave(lock, flags);
570 list_add_tail(entry, list);
571 spin_unlock_irqrestore(lock, flags);
574 static struct ntb_queue_entry *ntb_list_rm(spinlock_t *lock,
575 struct list_head *list)
577 struct ntb_queue_entry *entry;
580 spin_lock_irqsave(lock, flags);
581 if (list_empty(list)) {
585 entry = list_first_entry(list, struct ntb_queue_entry, entry);
586 list_del(&entry->entry);
589 spin_unlock_irqrestore(lock, flags);
594 static struct ntb_queue_entry *ntb_list_mv(spinlock_t *lock,
595 struct list_head *list,
596 struct list_head *to_list)
598 struct ntb_queue_entry *entry;
601 spin_lock_irqsave(lock, flags);
603 if (list_empty(list)) {
606 entry = list_first_entry(list, struct ntb_queue_entry, entry);
607 list_move_tail(&entry->entry, to_list);
610 spin_unlock_irqrestore(lock, flags);
615 static int ntb_transport_setup_qp_mw(struct ntb_transport_ctx *nt,
618 struct ntb_transport_qp *qp = &nt->qp_vec[qp_num];
619 struct ntb_transport_mw *mw;
620 struct ntb_dev *ndev = nt->ndev;
621 struct ntb_queue_entry *entry;
622 unsigned int rx_size, num_qps_mw;
623 unsigned int mw_num, mw_count, qp_count;
627 mw_count = nt->mw_count;
628 qp_count = nt->qp_count;
630 mw_num = QP_TO_MW(nt, qp_num);
631 mw = &nt->mw_vec[mw_num];
636 if (mw_num < qp_count % mw_count)
637 num_qps_mw = qp_count / mw_count + 1;
639 num_qps_mw = qp_count / mw_count;
641 rx_size = (unsigned int)mw->xlat_size / num_qps_mw;
642 qp->rx_buff = mw->virt_addr + rx_size * (qp_num / mw_count);
643 rx_size -= sizeof(struct ntb_rx_info);
645 qp->remote_rx_info = qp->rx_buff + rx_size;
647 /* Due to housekeeping, there must be atleast 2 buffs */
648 qp->rx_max_frame = min(transport_mtu, rx_size / 2);
649 qp->rx_max_entry = rx_size / qp->rx_max_frame;
653 * Checking to see if we have more entries than the default.
654 * We should add additional entries if that is the case so we
655 * can be in sync with the transport frames.
657 node = dev_to_node(&ndev->dev);
658 for (i = qp->rx_alloc_entry; i < qp->rx_max_entry; i++) {
659 entry = kzalloc_node(sizeof(*entry), GFP_KERNEL, node);
664 ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry,
666 qp->rx_alloc_entry++;
669 qp->remote_rx_info->entry = qp->rx_max_entry - 1;
671 /* setup the hdr offsets with 0's */
672 for (i = 0; i < qp->rx_max_entry; i++) {
673 void *offset = (qp->rx_buff + qp->rx_max_frame * (i + 1) -
674 sizeof(struct ntb_payload_header));
675 memset(offset, 0, sizeof(struct ntb_payload_header));
685 static irqreturn_t ntb_transport_isr(int irq, void *dev)
687 struct ntb_transport_qp *qp = dev;
689 tasklet_schedule(&qp->rxc_db_work);
694 static void ntb_transport_setup_qp_peer_msi(struct ntb_transport_ctx *nt,
697 struct ntb_transport_qp *qp = &nt->qp_vec[qp_num];
698 int spad = qp_num * 2 + nt->msi_spad_offset;
703 if (spad >= ntb_spad_count(nt->ndev))
706 qp->peer_msi_desc.addr_offset =
707 ntb_peer_spad_read(qp->ndev, PIDX, spad);
708 qp->peer_msi_desc.data =
709 ntb_peer_spad_read(qp->ndev, PIDX, spad + 1);
711 dev_dbg(&qp->ndev->pdev->dev, "QP%d Peer MSI addr=%x data=%x\n",
712 qp_num, qp->peer_msi_desc.addr_offset, qp->peer_msi_desc.data);
714 if (qp->peer_msi_desc.addr_offset) {
716 dev_info(&qp->ndev->pdev->dev,
717 "Using MSI interrupts for QP%d\n", qp_num);
721 static void ntb_transport_setup_qp_msi(struct ntb_transport_ctx *nt,
724 struct ntb_transport_qp *qp = &nt->qp_vec[qp_num];
725 int spad = qp_num * 2 + nt->msi_spad_offset;
731 if (spad >= ntb_spad_count(nt->ndev)) {
732 dev_warn_once(&qp->ndev->pdev->dev,
733 "Not enough SPADS to use MSI interrupts\n");
737 ntb_spad_write(qp->ndev, spad, 0);
738 ntb_spad_write(qp->ndev, spad + 1, 0);
741 qp->msi_irq = ntbm_msi_request_irq(qp->ndev, ntb_transport_isr,
744 if (qp->msi_irq < 0) {
745 dev_warn(&qp->ndev->pdev->dev,
746 "Unable to allocate MSI interrupt for qp%d\n",
752 rc = ntb_spad_write(qp->ndev, spad, qp->msi_desc.addr_offset);
754 goto err_free_interrupt;
756 rc = ntb_spad_write(qp->ndev, spad + 1, qp->msi_desc.data);
758 goto err_free_interrupt;
760 dev_dbg(&qp->ndev->pdev->dev, "QP%d MSI %d addr=%x data=%x\n",
761 qp_num, qp->msi_irq, qp->msi_desc.addr_offset,
767 devm_free_irq(&nt->ndev->dev, qp->msi_irq, qp);
770 static void ntb_transport_msi_peer_desc_changed(struct ntb_transport_ctx *nt)
774 dev_dbg(&nt->ndev->pdev->dev, "Peer MSI descriptors changed");
776 for (i = 0; i < nt->qp_count; i++)
777 ntb_transport_setup_qp_peer_msi(nt, i);
780 static void ntb_transport_msi_desc_changed(void *data)
782 struct ntb_transport_ctx *nt = data;
785 dev_dbg(&nt->ndev->pdev->dev, "MSI descriptors changed");
787 for (i = 0; i < nt->qp_count; i++)
788 ntb_transport_setup_qp_msi(nt, i);
790 ntb_peer_db_set(nt->ndev, nt->msi_db_mask);
793 static void ntb_free_mw(struct ntb_transport_ctx *nt, int num_mw)
795 struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
796 struct pci_dev *pdev = nt->ndev->pdev;
801 ntb_mw_clear_trans(nt->ndev, PIDX, num_mw);
802 dma_free_coherent(&pdev->dev, mw->alloc_size,
803 mw->alloc_addr, mw->dma_addr);
807 mw->alloc_addr = NULL;
808 mw->virt_addr = NULL;
811 static int ntb_alloc_mw_buffer(struct ntb_transport_mw *mw,
812 struct device *dma_dev, size_t align)
815 void *alloc_addr, *virt_addr;
818 alloc_addr = dma_alloc_coherent(dma_dev, mw->alloc_size,
819 &dma_addr, GFP_KERNEL);
821 dev_err(dma_dev, "Unable to alloc MW buff of size %zu\n",
825 virt_addr = alloc_addr;
828 * we must ensure that the memory address allocated is BAR size
829 * aligned in order for the XLAT register to take the value. This
830 * is a requirement of the hardware. It is recommended to setup CMA
831 * for BAR sizes equal or greater than 4MB.
833 if (!IS_ALIGNED(dma_addr, align)) {
834 if (mw->alloc_size > mw->buff_size) {
835 virt_addr = PTR_ALIGN(alloc_addr, align);
836 dma_addr = ALIGN(dma_addr, align);
843 mw->alloc_addr = alloc_addr;
844 mw->virt_addr = virt_addr;
845 mw->dma_addr = dma_addr;
850 dma_free_coherent(dma_dev, mw->alloc_size, alloc_addr, dma_addr);
855 static int ntb_set_mw(struct ntb_transport_ctx *nt, int num_mw,
856 resource_size_t size)
858 struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
859 struct pci_dev *pdev = nt->ndev->pdev;
860 size_t xlat_size, buff_size;
861 resource_size_t xlat_align;
862 resource_size_t xlat_align_size;
868 rc = ntb_mw_get_align(nt->ndev, PIDX, num_mw, &xlat_align,
869 &xlat_align_size, NULL);
873 xlat_size = round_up(size, xlat_align_size);
874 buff_size = round_up(size, xlat_align);
876 /* No need to re-setup */
877 if (mw->xlat_size == xlat_size)
881 ntb_free_mw(nt, num_mw);
883 /* Alloc memory for receiving data. Must be aligned */
884 mw->xlat_size = xlat_size;
885 mw->buff_size = buff_size;
886 mw->alloc_size = buff_size;
888 rc = ntb_alloc_mw_buffer(mw, &pdev->dev, xlat_align);
891 rc = ntb_alloc_mw_buffer(mw, &pdev->dev, xlat_align);
894 "Unable to alloc aligned MW buff\n");
902 /* Notify HW the memory location of the receive buffer */
903 rc = ntb_mw_set_trans(nt->ndev, PIDX, num_mw, mw->dma_addr,
906 dev_err(&pdev->dev, "Unable to set mw%d translation", num_mw);
907 ntb_free_mw(nt, num_mw);
914 static void ntb_qp_link_down_reset(struct ntb_transport_qp *qp)
916 qp->link_is_up = false;
923 qp->rx_ring_empty = 0;
924 qp->rx_err_no_buf = 0;
925 qp->rx_err_oflow = 0;
931 qp->tx_ring_full = 0;
932 qp->tx_err_no_buf = 0;
937 static void ntb_qp_link_cleanup(struct ntb_transport_qp *qp)
939 struct ntb_transport_ctx *nt = qp->transport;
940 struct pci_dev *pdev = nt->ndev->pdev;
942 dev_info(&pdev->dev, "qp %d: Link Cleanup\n", qp->qp_num);
944 cancel_delayed_work_sync(&qp->link_work);
945 ntb_qp_link_down_reset(qp);
947 if (qp->event_handler)
948 qp->event_handler(qp->cb_data, qp->link_is_up);
951 static void ntb_qp_link_cleanup_work(struct work_struct *work)
953 struct ntb_transport_qp *qp = container_of(work,
954 struct ntb_transport_qp,
956 struct ntb_transport_ctx *nt = qp->transport;
958 ntb_qp_link_cleanup(qp);
961 schedule_delayed_work(&qp->link_work,
962 msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
965 static void ntb_qp_link_down(struct ntb_transport_qp *qp)
967 schedule_work(&qp->link_cleanup);
970 static void ntb_transport_link_cleanup(struct ntb_transport_ctx *nt)
972 struct ntb_transport_qp *qp;
974 unsigned int i, count;
976 qp_bitmap_alloc = nt->qp_bitmap & ~nt->qp_bitmap_free;
978 /* Pass along the info to any clients */
979 for (i = 0; i < nt->qp_count; i++)
980 if (qp_bitmap_alloc & BIT_ULL(i)) {
982 ntb_qp_link_cleanup(qp);
983 cancel_work_sync(&qp->link_cleanup);
984 cancel_delayed_work_sync(&qp->link_work);
988 cancel_delayed_work_sync(&nt->link_work);
990 for (i = 0; i < nt->mw_count; i++)
993 /* The scratchpad registers keep the values if the remote side
994 * goes down, blast them now to give them a sane value the next
995 * time they are accessed
997 count = ntb_spad_count(nt->ndev);
998 for (i = 0; i < count; i++)
999 ntb_spad_write(nt->ndev, i, 0);
1002 static void ntb_transport_link_cleanup_work(struct work_struct *work)
1004 struct ntb_transport_ctx *nt =
1005 container_of(work, struct ntb_transport_ctx, link_cleanup);
1007 ntb_transport_link_cleanup(nt);
1010 static void ntb_transport_event_callback(void *data)
1012 struct ntb_transport_ctx *nt = data;
1014 if (ntb_link_is_up(nt->ndev, NULL, NULL) == 1)
1015 schedule_delayed_work(&nt->link_work, 0);
1017 schedule_work(&nt->link_cleanup);
1020 static void ntb_transport_link_work(struct work_struct *work)
1022 struct ntb_transport_ctx *nt =
1023 container_of(work, struct ntb_transport_ctx, link_work.work);
1024 struct ntb_dev *ndev = nt->ndev;
1025 struct pci_dev *pdev = ndev->pdev;
1026 resource_size_t size;
1028 int rc = 0, i, spad;
1030 /* send the local info, in the opposite order of the way we read it */
1033 rc = ntb_msi_setup_mws(ndev);
1035 dev_warn(&pdev->dev,
1036 "Failed to register MSI memory window: %d\n",
1038 nt->use_msi = false;
1042 for (i = 0; i < nt->qp_count; i++)
1043 ntb_transport_setup_qp_msi(nt, i);
1045 for (i = 0; i < nt->mw_count; i++) {
1046 size = nt->mw_vec[i].phys_size;
1048 if (max_mw_size && size > max_mw_size)
1051 spad = MW0_SZ_HIGH + (i * 2);
1052 ntb_peer_spad_write(ndev, PIDX, spad, upper_32_bits(size));
1054 spad = MW0_SZ_LOW + (i * 2);
1055 ntb_peer_spad_write(ndev, PIDX, spad, lower_32_bits(size));
1058 ntb_peer_spad_write(ndev, PIDX, NUM_MWS, nt->mw_count);
1060 ntb_peer_spad_write(ndev, PIDX, NUM_QPS, nt->qp_count);
1062 ntb_peer_spad_write(ndev, PIDX, VERSION, NTB_TRANSPORT_VERSION);
1064 /* Query the remote side for its info */
1065 val = ntb_spad_read(ndev, VERSION);
1066 dev_dbg(&pdev->dev, "Remote version = %d\n", val);
1067 if (val != NTB_TRANSPORT_VERSION)
1070 val = ntb_spad_read(ndev, NUM_QPS);
1071 dev_dbg(&pdev->dev, "Remote max number of qps = %d\n", val);
1072 if (val != nt->qp_count)
1075 val = ntb_spad_read(ndev, NUM_MWS);
1076 dev_dbg(&pdev->dev, "Remote number of mws = %d\n", val);
1077 if (val != nt->mw_count)
1080 for (i = 0; i < nt->mw_count; i++) {
1083 val = ntb_spad_read(ndev, MW0_SZ_HIGH + (i * 2));
1084 val64 = (u64)val << 32;
1086 val = ntb_spad_read(ndev, MW0_SZ_LOW + (i * 2));
1089 dev_dbg(&pdev->dev, "Remote MW%d size = %#llx\n", i, val64);
1091 rc = ntb_set_mw(nt, i, val64);
1096 nt->link_is_up = true;
1098 for (i = 0; i < nt->qp_count; i++) {
1099 struct ntb_transport_qp *qp = &nt->qp_vec[i];
1101 ntb_transport_setup_qp_mw(nt, i);
1102 ntb_transport_setup_qp_peer_msi(nt, i);
1104 if (qp->client_ready)
1105 schedule_delayed_work(&qp->link_work, 0);
1111 for (i = 0; i < nt->mw_count; i++)
1114 /* if there's an actual failure, we should just bail */
1119 if (ntb_link_is_up(ndev, NULL, NULL) == 1)
1120 schedule_delayed_work(&nt->link_work,
1121 msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
1124 static void ntb_qp_link_work(struct work_struct *work)
1126 struct ntb_transport_qp *qp = container_of(work,
1127 struct ntb_transport_qp,
1129 struct pci_dev *pdev = qp->ndev->pdev;
1130 struct ntb_transport_ctx *nt = qp->transport;
1133 WARN_ON(!nt->link_is_up);
1135 val = ntb_spad_read(nt->ndev, QP_LINKS);
1137 ntb_peer_spad_write(nt->ndev, PIDX, QP_LINKS, val | BIT(qp->qp_num));
1139 /* query remote spad for qp ready bits */
1140 dev_dbg_ratelimited(&pdev->dev, "Remote QP link status = %x\n", val);
1142 /* See if the remote side is up */
1143 if (val & BIT(qp->qp_num)) {
1144 dev_info(&pdev->dev, "qp %d: Link Up\n", qp->qp_num);
1145 qp->link_is_up = true;
1148 if (qp->event_handler)
1149 qp->event_handler(qp->cb_data, qp->link_is_up);
1152 tasklet_schedule(&qp->rxc_db_work);
1153 } else if (nt->link_is_up)
1154 schedule_delayed_work(&qp->link_work,
1155 msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
1158 static int ntb_transport_init_queue(struct ntb_transport_ctx *nt,
1159 unsigned int qp_num)
1161 struct ntb_transport_qp *qp;
1162 phys_addr_t mw_base;
1163 resource_size_t mw_size;
1164 unsigned int num_qps_mw, tx_size;
1165 unsigned int mw_num, mw_count, qp_count;
1168 mw_count = nt->mw_count;
1169 qp_count = nt->qp_count;
1171 mw_num = QP_TO_MW(nt, qp_num);
1173 qp = &nt->qp_vec[qp_num];
1174 qp->qp_num = qp_num;
1176 qp->ndev = nt->ndev;
1177 qp->client_ready = false;
1178 qp->event_handler = NULL;
1179 ntb_qp_link_down_reset(qp);
1181 if (mw_num < qp_count % mw_count)
1182 num_qps_mw = qp_count / mw_count + 1;
1184 num_qps_mw = qp_count / mw_count;
1186 mw_base = nt->mw_vec[mw_num].phys_addr;
1187 mw_size = nt->mw_vec[mw_num].phys_size;
1189 if (max_mw_size && mw_size > max_mw_size)
1190 mw_size = max_mw_size;
1192 tx_size = (unsigned int)mw_size / num_qps_mw;
1193 qp_offset = tx_size * (qp_num / mw_count);
1195 qp->tx_mw_size = tx_size;
1196 qp->tx_mw = nt->mw_vec[mw_num].vbase + qp_offset;
1200 qp->tx_mw_phys = mw_base + qp_offset;
1201 if (!qp->tx_mw_phys)
1204 tx_size -= sizeof(struct ntb_rx_info);
1205 qp->rx_info = qp->tx_mw + tx_size;
1207 /* Due to housekeeping, there must be atleast 2 buffs */
1208 qp->tx_max_frame = min(transport_mtu, tx_size / 2);
1209 qp->tx_max_entry = tx_size / qp->tx_max_frame;
1211 if (nt->debugfs_node_dir) {
1212 char debugfs_name[4];
1214 snprintf(debugfs_name, 4, "qp%d", qp_num);
1215 qp->debugfs_dir = debugfs_create_dir(debugfs_name,
1216 nt->debugfs_node_dir);
1218 qp->debugfs_stats = debugfs_create_file("stats", S_IRUSR,
1219 qp->debugfs_dir, qp,
1220 &ntb_qp_debugfs_stats);
1222 qp->debugfs_dir = NULL;
1223 qp->debugfs_stats = NULL;
1226 INIT_DELAYED_WORK(&qp->link_work, ntb_qp_link_work);
1227 INIT_WORK(&qp->link_cleanup, ntb_qp_link_cleanup_work);
1229 spin_lock_init(&qp->ntb_rx_q_lock);
1230 spin_lock_init(&qp->ntb_tx_free_q_lock);
1232 INIT_LIST_HEAD(&qp->rx_post_q);
1233 INIT_LIST_HEAD(&qp->rx_pend_q);
1234 INIT_LIST_HEAD(&qp->rx_free_q);
1235 INIT_LIST_HEAD(&qp->tx_free_q);
1237 tasklet_init(&qp->rxc_db_work, ntb_transport_rxc_db,
1243 static int ntb_transport_probe(struct ntb_client *self, struct ntb_dev *ndev)
1245 struct ntb_transport_ctx *nt;
1246 struct ntb_transport_mw *mw;
1247 unsigned int mw_count, qp_count, spad_count, max_mw_count_for_spads;
1252 mw_count = ntb_peer_mw_count(ndev);
1254 if (!ndev->ops->mw_set_trans) {
1255 dev_err(&ndev->dev, "Inbound MW based NTB API is required\n");
1259 if (ntb_db_is_unsafe(ndev))
1261 "doorbell is unsafe, proceed anyway...\n");
1262 if (ntb_spad_is_unsafe(ndev))
1264 "scratchpad is unsafe, proceed anyway...\n");
1266 if (ntb_peer_port_count(ndev) != NTB_DEF_PEER_CNT)
1267 dev_warn(&ndev->dev, "Multi-port NTB devices unsupported\n");
1269 node = dev_to_node(&ndev->dev);
1271 nt = kzalloc_node(sizeof(*nt), GFP_KERNEL, node);
1278 * If we are using MSI, and have at least one extra memory window,
1279 * we will reserve the last MW for the MSI window.
1281 if (use_msi && mw_count > 1) {
1282 rc = ntb_msi_init(ndev, ntb_transport_msi_desc_changed);
1289 spad_count = ntb_spad_count(ndev);
1291 /* Limit the MW's based on the availability of scratchpads */
1293 if (spad_count < NTB_TRANSPORT_MIN_SPADS) {
1299 max_mw_count_for_spads = (spad_count - MW0_SZ_HIGH) / 2;
1300 nt->mw_count = min(mw_count, max_mw_count_for_spads);
1302 nt->msi_spad_offset = nt->mw_count * 2 + MW0_SZ_HIGH;
1304 nt->mw_vec = kcalloc_node(mw_count, sizeof(*nt->mw_vec),
1311 for (i = 0; i < mw_count; i++) {
1312 mw = &nt->mw_vec[i];
1314 rc = ntb_peer_mw_get_addr(ndev, i, &mw->phys_addr,
1319 mw->vbase = ioremap_wc(mw->phys_addr, mw->phys_size);
1327 mw->virt_addr = NULL;
1331 qp_bitmap = ntb_db_valid_mask(ndev);
1333 qp_count = ilog2(qp_bitmap);
1336 nt->msi_db_mask = 1 << qp_count;
1337 ntb_db_clear_mask(ndev, nt->msi_db_mask);
1340 if (max_num_clients && max_num_clients < qp_count)
1341 qp_count = max_num_clients;
1342 else if (nt->mw_count < qp_count)
1343 qp_count = nt->mw_count;
1345 qp_bitmap &= BIT_ULL(qp_count) - 1;
1347 nt->qp_count = qp_count;
1348 nt->qp_bitmap = qp_bitmap;
1349 nt->qp_bitmap_free = qp_bitmap;
1351 nt->qp_vec = kcalloc_node(qp_count, sizeof(*nt->qp_vec),
1358 if (nt_debugfs_dir) {
1359 nt->debugfs_node_dir =
1360 debugfs_create_dir(pci_name(ndev->pdev),
1364 for (i = 0; i < qp_count; i++) {
1365 rc = ntb_transport_init_queue(nt, i);
1370 INIT_DELAYED_WORK(&nt->link_work, ntb_transport_link_work);
1371 INIT_WORK(&nt->link_cleanup, ntb_transport_link_cleanup_work);
1373 rc = ntb_set_ctx(ndev, nt, &ntb_transport_ops);
1377 INIT_LIST_HEAD(&nt->client_devs);
1378 rc = ntb_bus_init(nt);
1382 nt->link_is_up = false;
1383 ntb_link_enable(ndev, NTB_SPEED_AUTO, NTB_WIDTH_AUTO);
1384 ntb_link_event(ndev);
1389 ntb_clear_ctx(ndev);
1394 mw = &nt->mw_vec[i];
1403 static void ntb_transport_free(struct ntb_client *self, struct ntb_dev *ndev)
1405 struct ntb_transport_ctx *nt = ndev->ctx;
1406 struct ntb_transport_qp *qp;
1407 u64 qp_bitmap_alloc;
1410 ntb_transport_link_cleanup(nt);
1411 cancel_work_sync(&nt->link_cleanup);
1412 cancel_delayed_work_sync(&nt->link_work);
1414 qp_bitmap_alloc = nt->qp_bitmap & ~nt->qp_bitmap_free;
1416 /* verify that all the qp's are freed */
1417 for (i = 0; i < nt->qp_count; i++) {
1418 qp = &nt->qp_vec[i];
1419 if (qp_bitmap_alloc & BIT_ULL(i))
1420 ntb_transport_free_queue(qp);
1421 debugfs_remove_recursive(qp->debugfs_dir);
1424 ntb_link_disable(ndev);
1425 ntb_clear_ctx(ndev);
1429 for (i = nt->mw_count; i--; ) {
1431 iounmap(nt->mw_vec[i].vbase);
1439 static void ntb_complete_rxc(struct ntb_transport_qp *qp)
1441 struct ntb_queue_entry *entry;
1444 unsigned long irqflags;
1446 spin_lock_irqsave(&qp->ntb_rx_q_lock, irqflags);
1448 while (!list_empty(&qp->rx_post_q)) {
1449 entry = list_first_entry(&qp->rx_post_q,
1450 struct ntb_queue_entry, entry);
1451 if (!(entry->flags & DESC_DONE_FLAG))
1454 entry->rx_hdr->flags = 0;
1455 iowrite32(entry->rx_index, &qp->rx_info->entry);
1457 cb_data = entry->cb_data;
1460 list_move_tail(&entry->entry, &qp->rx_free_q);
1462 spin_unlock_irqrestore(&qp->ntb_rx_q_lock, irqflags);
1464 if (qp->rx_handler && qp->client_ready)
1465 qp->rx_handler(qp, qp->cb_data, cb_data, len);
1467 spin_lock_irqsave(&qp->ntb_rx_q_lock, irqflags);
1470 spin_unlock_irqrestore(&qp->ntb_rx_q_lock, irqflags);
1473 static void ntb_rx_copy_callback(void *data,
1474 const struct dmaengine_result *res)
1476 struct ntb_queue_entry *entry = data;
1478 /* we need to check DMA results if we are using DMA */
1480 enum dmaengine_tx_result dma_err = res->result;
1483 case DMA_TRANS_READ_FAILED:
1484 case DMA_TRANS_WRITE_FAILED:
1487 case DMA_TRANS_ABORTED:
1489 struct ntb_transport_qp *qp = entry->qp;
1490 void *offset = qp->rx_buff + qp->rx_max_frame *
1493 ntb_memcpy_rx(entry, offset);
1498 case DMA_TRANS_NOERROR:
1504 entry->flags |= DESC_DONE_FLAG;
1506 ntb_complete_rxc(entry->qp);
1509 static void ntb_memcpy_rx(struct ntb_queue_entry *entry, void *offset)
1511 void *buf = entry->buf;
1512 size_t len = entry->len;
1514 memcpy(buf, offset, len);
1516 /* Ensure that the data is fully copied out before clearing the flag */
1519 ntb_rx_copy_callback(entry, NULL);
1522 static int ntb_async_rx_submit(struct ntb_queue_entry *entry, void *offset)
1524 struct dma_async_tx_descriptor *txd;
1525 struct ntb_transport_qp *qp = entry->qp;
1526 struct dma_chan *chan = qp->rx_dma_chan;
1527 struct dma_device *device;
1528 size_t pay_off, buff_off, len;
1529 struct dmaengine_unmap_data *unmap;
1530 dma_cookie_t cookie;
1531 void *buf = entry->buf;
1534 device = chan->device;
1535 pay_off = (size_t)offset & ~PAGE_MASK;
1536 buff_off = (size_t)buf & ~PAGE_MASK;
1538 if (!is_dma_copy_aligned(device, pay_off, buff_off, len))
1541 unmap = dmaengine_get_unmap_data(device->dev, 2, GFP_NOWAIT);
1546 unmap->addr[0] = dma_map_page(device->dev, virt_to_page(offset),
1547 pay_off, len, DMA_TO_DEVICE);
1548 if (dma_mapping_error(device->dev, unmap->addr[0]))
1553 unmap->addr[1] = dma_map_page(device->dev, virt_to_page(buf),
1554 buff_off, len, DMA_FROM_DEVICE);
1555 if (dma_mapping_error(device->dev, unmap->addr[1]))
1558 unmap->from_cnt = 1;
1560 txd = device->device_prep_dma_memcpy(chan, unmap->addr[1],
1561 unmap->addr[0], len,
1562 DMA_PREP_INTERRUPT);
1566 txd->callback_result = ntb_rx_copy_callback;
1567 txd->callback_param = entry;
1568 dma_set_unmap(txd, unmap);
1570 cookie = dmaengine_submit(txd);
1571 if (dma_submit_error(cookie))
1574 dmaengine_unmap_put(unmap);
1576 qp->last_cookie = cookie;
1583 dmaengine_unmap_put(unmap);
1585 dmaengine_unmap_put(unmap);
1590 static void ntb_async_rx(struct ntb_queue_entry *entry, void *offset)
1592 struct ntb_transport_qp *qp = entry->qp;
1593 struct dma_chan *chan = qp->rx_dma_chan;
1599 if (entry->len < copy_bytes)
1602 res = ntb_async_rx_submit(entry, offset);
1606 if (!entry->retries)
1612 ntb_memcpy_rx(entry, offset);
1616 static int ntb_process_rxc(struct ntb_transport_qp *qp)
1618 struct ntb_payload_header *hdr;
1619 struct ntb_queue_entry *entry;
1622 offset = qp->rx_buff + qp->rx_max_frame * qp->rx_index;
1623 hdr = offset + qp->rx_max_frame - sizeof(struct ntb_payload_header);
1625 dev_dbg(&qp->ndev->pdev->dev, "qp %d: RX ver %u len %d flags %x\n",
1626 qp->qp_num, hdr->ver, hdr->len, hdr->flags);
1628 if (!(hdr->flags & DESC_DONE_FLAG)) {
1629 dev_dbg(&qp->ndev->pdev->dev, "done flag not set\n");
1630 qp->rx_ring_empty++;
1634 if (hdr->flags & LINK_DOWN_FLAG) {
1635 dev_dbg(&qp->ndev->pdev->dev, "link down flag set\n");
1636 ntb_qp_link_down(qp);
1641 if (hdr->ver != (u32)qp->rx_pkts) {
1642 dev_dbg(&qp->ndev->pdev->dev,
1643 "version mismatch, expected %llu - got %u\n",
1644 qp->rx_pkts, hdr->ver);
1649 entry = ntb_list_mv(&qp->ntb_rx_q_lock, &qp->rx_pend_q, &qp->rx_post_q);
1651 dev_dbg(&qp->ndev->pdev->dev, "no receive buffer\n");
1652 qp->rx_err_no_buf++;
1656 entry->rx_hdr = hdr;
1657 entry->rx_index = qp->rx_index;
1659 if (hdr->len > entry->len) {
1660 dev_dbg(&qp->ndev->pdev->dev,
1661 "receive buffer overflow! Wanted %d got %d\n",
1662 hdr->len, entry->len);
1666 entry->flags |= DESC_DONE_FLAG;
1668 ntb_complete_rxc(qp);
1670 dev_dbg(&qp->ndev->pdev->dev,
1671 "RX OK index %u ver %u size %d into buf size %d\n",
1672 qp->rx_index, hdr->ver, hdr->len, entry->len);
1674 qp->rx_bytes += hdr->len;
1677 entry->len = hdr->len;
1679 ntb_async_rx(entry, offset);
1683 qp->rx_index %= qp->rx_max_entry;
1688 static void ntb_transport_rxc_db(unsigned long data)
1690 struct ntb_transport_qp *qp = (void *)data;
1693 dev_dbg(&qp->ndev->pdev->dev, "%s: doorbell %d received\n",
1694 __func__, qp->qp_num);
1696 /* Limit the number of packets processed in a single interrupt to
1697 * provide fairness to others
1699 for (i = 0; i < qp->rx_max_entry; i++) {
1700 rc = ntb_process_rxc(qp);
1705 if (i && qp->rx_dma_chan)
1706 dma_async_issue_pending(qp->rx_dma_chan);
1708 if (i == qp->rx_max_entry) {
1709 /* there is more work to do */
1711 tasklet_schedule(&qp->rxc_db_work);
1712 } else if (ntb_db_read(qp->ndev) & BIT_ULL(qp->qp_num)) {
1713 /* the doorbell bit is set: clear it */
1714 ntb_db_clear(qp->ndev, BIT_ULL(qp->qp_num));
1715 /* ntb_db_read ensures ntb_db_clear write is committed */
1716 ntb_db_read(qp->ndev);
1718 /* an interrupt may have arrived between finishing
1719 * ntb_process_rxc and clearing the doorbell bit:
1720 * there might be some more work to do.
1723 tasklet_schedule(&qp->rxc_db_work);
1727 static void ntb_tx_copy_callback(void *data,
1728 const struct dmaengine_result *res)
1730 struct ntb_queue_entry *entry = data;
1731 struct ntb_transport_qp *qp = entry->qp;
1732 struct ntb_payload_header __iomem *hdr = entry->tx_hdr;
1734 /* we need to check DMA results if we are using DMA */
1736 enum dmaengine_tx_result dma_err = res->result;
1739 case DMA_TRANS_READ_FAILED:
1740 case DMA_TRANS_WRITE_FAILED:
1743 case DMA_TRANS_ABORTED:
1745 void __iomem *offset =
1746 qp->tx_mw + qp->tx_max_frame *
1749 /* resubmit via CPU */
1750 ntb_memcpy_tx(entry, offset);
1755 case DMA_TRANS_NOERROR:
1761 iowrite32(entry->flags | DESC_DONE_FLAG, &hdr->flags);
1764 ntb_msi_peer_trigger(qp->ndev, PIDX, &qp->peer_msi_desc);
1766 ntb_peer_db_set(qp->ndev, BIT_ULL(qp->qp_num));
1768 /* The entry length can only be zero if the packet is intended to be a
1769 * "link down" or similar. Since no payload is being sent in these
1770 * cases, there is nothing to add to the completion queue.
1772 if (entry->len > 0) {
1773 qp->tx_bytes += entry->len;
1776 qp->tx_handler(qp, qp->cb_data, entry->cb_data,
1780 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry, &qp->tx_free_q);
1783 static void ntb_memcpy_tx(struct ntb_queue_entry *entry, void __iomem *offset)
1785 #ifdef ARCH_HAS_NOCACHE_UACCESS
1787 * Using non-temporal mov to improve performance on non-cached
1788 * writes, even though we aren't actually copying from user space.
1790 __copy_from_user_inatomic_nocache(offset, entry->buf, entry->len);
1792 memcpy_toio(offset, entry->buf, entry->len);
1795 /* Ensure that the data is fully copied out before setting the flags */
1798 ntb_tx_copy_callback(entry, NULL);
1801 static int ntb_async_tx_submit(struct ntb_transport_qp *qp,
1802 struct ntb_queue_entry *entry)
1804 struct dma_async_tx_descriptor *txd;
1805 struct dma_chan *chan = qp->tx_dma_chan;
1806 struct dma_device *device;
1807 size_t len = entry->len;
1808 void *buf = entry->buf;
1809 size_t dest_off, buff_off;
1810 struct dmaengine_unmap_data *unmap;
1812 dma_cookie_t cookie;
1814 device = chan->device;
1815 dest = qp->tx_mw_dma_addr + qp->tx_max_frame * entry->tx_index;
1816 buff_off = (size_t)buf & ~PAGE_MASK;
1817 dest_off = (size_t)dest & ~PAGE_MASK;
1819 if (!is_dma_copy_aligned(device, buff_off, dest_off, len))
1822 unmap = dmaengine_get_unmap_data(device->dev, 1, GFP_NOWAIT);
1827 unmap->addr[0] = dma_map_page(device->dev, virt_to_page(buf),
1828 buff_off, len, DMA_TO_DEVICE);
1829 if (dma_mapping_error(device->dev, unmap->addr[0]))
1834 txd = device->device_prep_dma_memcpy(chan, dest, unmap->addr[0], len,
1835 DMA_PREP_INTERRUPT);
1839 txd->callback_result = ntb_tx_copy_callback;
1840 txd->callback_param = entry;
1841 dma_set_unmap(txd, unmap);
1843 cookie = dmaengine_submit(txd);
1844 if (dma_submit_error(cookie))
1847 dmaengine_unmap_put(unmap);
1849 dma_async_issue_pending(chan);
1853 dmaengine_unmap_put(unmap);
1855 dmaengine_unmap_put(unmap);
1860 static void ntb_async_tx(struct ntb_transport_qp *qp,
1861 struct ntb_queue_entry *entry)
1863 struct ntb_payload_header __iomem *hdr;
1864 struct dma_chan *chan = qp->tx_dma_chan;
1865 void __iomem *offset;
1868 entry->tx_index = qp->tx_index;
1869 offset = qp->tx_mw + qp->tx_max_frame * entry->tx_index;
1870 hdr = offset + qp->tx_max_frame - sizeof(struct ntb_payload_header);
1871 entry->tx_hdr = hdr;
1873 iowrite32(entry->len, &hdr->len);
1874 iowrite32((u32)qp->tx_pkts, &hdr->ver);
1879 if (entry->len < copy_bytes)
1882 res = ntb_async_tx_submit(qp, entry);
1886 if (!entry->retries)
1892 ntb_memcpy_tx(entry, offset);
1896 static int ntb_process_tx(struct ntb_transport_qp *qp,
1897 struct ntb_queue_entry *entry)
1899 if (qp->tx_index == qp->remote_rx_info->entry) {
1904 if (entry->len > qp->tx_max_frame - sizeof(struct ntb_payload_header)) {
1906 qp->tx_handler(qp, qp->cb_data, NULL, -EIO);
1908 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
1913 ntb_async_tx(qp, entry);
1916 qp->tx_index %= qp->tx_max_entry;
1923 static void ntb_send_link_down(struct ntb_transport_qp *qp)
1925 struct pci_dev *pdev = qp->ndev->pdev;
1926 struct ntb_queue_entry *entry;
1929 if (!qp->link_is_up)
1932 dev_info(&pdev->dev, "qp %d: Send Link Down\n", qp->qp_num);
1934 for (i = 0; i < NTB_LINK_DOWN_TIMEOUT; i++) {
1935 entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
1944 entry->cb_data = NULL;
1947 entry->flags = LINK_DOWN_FLAG;
1949 rc = ntb_process_tx(qp, entry);
1951 dev_err(&pdev->dev, "ntb: QP%d unable to send linkdown msg\n",
1954 ntb_qp_link_down_reset(qp);
1957 static bool ntb_dma_filter_fn(struct dma_chan *chan, void *node)
1959 return dev_to_node(&chan->dev->device) == (int)(unsigned long)node;
1963 * ntb_transport_create_queue - Create a new NTB transport layer queue
1964 * @rx_handler: receive callback function
1965 * @tx_handler: transmit callback function
1966 * @event_handler: event callback function
1968 * Create a new NTB transport layer queue and provide the queue with a callback
1969 * routine for both transmit and receive. The receive callback routine will be
1970 * used to pass up data when the transport has received it on the queue. The
1971 * transmit callback routine will be called when the transport has completed the
1972 * transmission of the data on the queue and the data is ready to be freed.
1974 * RETURNS: pointer to newly created ntb_queue, NULL on error.
1976 struct ntb_transport_qp *
1977 ntb_transport_create_queue(void *data, struct device *client_dev,
1978 const struct ntb_queue_handlers *handlers)
1980 struct ntb_dev *ndev;
1981 struct pci_dev *pdev;
1982 struct ntb_transport_ctx *nt;
1983 struct ntb_queue_entry *entry;
1984 struct ntb_transport_qp *qp;
1986 unsigned int free_queue;
1987 dma_cap_mask_t dma_mask;
1991 ndev = dev_ntb(client_dev->parent);
1995 node = dev_to_node(&ndev->dev);
1997 free_queue = ffs(nt->qp_bitmap_free);
2001 /* decrement free_queue to make it zero based */
2004 qp = &nt->qp_vec[free_queue];
2005 qp_bit = BIT_ULL(qp->qp_num);
2007 nt->qp_bitmap_free &= ~qp_bit;
2010 qp->rx_handler = handlers->rx_handler;
2011 qp->tx_handler = handlers->tx_handler;
2012 qp->event_handler = handlers->event_handler;
2014 dma_cap_zero(dma_mask);
2015 dma_cap_set(DMA_MEMCPY, dma_mask);
2019 dma_request_channel(dma_mask, ntb_dma_filter_fn,
2020 (void *)(unsigned long)node);
2021 if (!qp->tx_dma_chan)
2022 dev_info(&pdev->dev, "Unable to allocate TX DMA channel\n");
2025 dma_request_channel(dma_mask, ntb_dma_filter_fn,
2026 (void *)(unsigned long)node);
2027 if (!qp->rx_dma_chan)
2028 dev_info(&pdev->dev, "Unable to allocate RX DMA channel\n");
2030 qp->tx_dma_chan = NULL;
2031 qp->rx_dma_chan = NULL;
2034 qp->tx_mw_dma_addr = 0;
2035 if (qp->tx_dma_chan) {
2036 qp->tx_mw_dma_addr =
2037 dma_map_resource(qp->tx_dma_chan->device->dev,
2038 qp->tx_mw_phys, qp->tx_mw_size,
2039 DMA_FROM_DEVICE, 0);
2040 if (dma_mapping_error(qp->tx_dma_chan->device->dev,
2041 qp->tx_mw_dma_addr)) {
2042 qp->tx_mw_dma_addr = 0;
2047 dev_dbg(&pdev->dev, "Using %s memcpy for TX\n",
2048 qp->tx_dma_chan ? "DMA" : "CPU");
2050 dev_dbg(&pdev->dev, "Using %s memcpy for RX\n",
2051 qp->rx_dma_chan ? "DMA" : "CPU");
2053 for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) {
2054 entry = kzalloc_node(sizeof(*entry), GFP_KERNEL, node);
2059 ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry,
2062 qp->rx_alloc_entry = NTB_QP_DEF_NUM_ENTRIES;
2064 for (i = 0; i < qp->tx_max_entry; i++) {
2065 entry = kzalloc_node(sizeof(*entry), GFP_KERNEL, node);
2070 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
2074 ntb_db_clear(qp->ndev, qp_bit);
2075 ntb_db_clear_mask(qp->ndev, qp_bit);
2077 dev_info(&pdev->dev, "NTB Transport QP %d created\n", qp->qp_num);
2082 while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
2085 qp->rx_alloc_entry = 0;
2086 while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q)))
2088 if (qp->tx_mw_dma_addr)
2089 dma_unmap_resource(qp->tx_dma_chan->device->dev,
2090 qp->tx_mw_dma_addr, qp->tx_mw_size,
2091 DMA_FROM_DEVICE, 0);
2092 if (qp->tx_dma_chan)
2093 dma_release_channel(qp->tx_dma_chan);
2094 if (qp->rx_dma_chan)
2095 dma_release_channel(qp->rx_dma_chan);
2096 nt->qp_bitmap_free |= qp_bit;
2100 EXPORT_SYMBOL_GPL(ntb_transport_create_queue);
2103 * ntb_transport_free_queue - Frees NTB transport queue
2104 * @qp: NTB queue to be freed
2106 * Frees NTB transport queue
2108 void ntb_transport_free_queue(struct ntb_transport_qp *qp)
2110 struct pci_dev *pdev;
2111 struct ntb_queue_entry *entry;
2117 pdev = qp->ndev->pdev;
2121 if (qp->tx_dma_chan) {
2122 struct dma_chan *chan = qp->tx_dma_chan;
2123 /* Putting the dma_chan to NULL will force any new traffic to be
2124 * processed by the CPU instead of the DAM engine
2126 qp->tx_dma_chan = NULL;
2128 /* Try to be nice and wait for any queued DMA engine
2129 * transactions to process before smashing it with a rock
2131 dma_sync_wait(chan, qp->last_cookie);
2132 dmaengine_terminate_all(chan);
2134 dma_unmap_resource(chan->device->dev,
2135 qp->tx_mw_dma_addr, qp->tx_mw_size,
2136 DMA_FROM_DEVICE, 0);
2138 dma_release_channel(chan);
2141 if (qp->rx_dma_chan) {
2142 struct dma_chan *chan = qp->rx_dma_chan;
2143 /* Putting the dma_chan to NULL will force any new traffic to be
2144 * processed by the CPU instead of the DAM engine
2146 qp->rx_dma_chan = NULL;
2148 /* Try to be nice and wait for any queued DMA engine
2149 * transactions to process before smashing it with a rock
2151 dma_sync_wait(chan, qp->last_cookie);
2152 dmaengine_terminate_all(chan);
2153 dma_release_channel(chan);
2156 qp_bit = BIT_ULL(qp->qp_num);
2158 ntb_db_set_mask(qp->ndev, qp_bit);
2159 tasklet_kill(&qp->rxc_db_work);
2161 cancel_delayed_work_sync(&qp->link_work);
2164 qp->rx_handler = NULL;
2165 qp->tx_handler = NULL;
2166 qp->event_handler = NULL;
2168 while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q)))
2171 while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_pend_q))) {
2172 dev_warn(&pdev->dev, "Freeing item from non-empty rx_pend_q\n");
2176 while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_post_q))) {
2177 dev_warn(&pdev->dev, "Freeing item from non-empty rx_post_q\n");
2181 while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
2184 qp->transport->qp_bitmap_free |= qp_bit;
2186 dev_info(&pdev->dev, "NTB Transport QP %d freed\n", qp->qp_num);
2188 EXPORT_SYMBOL_GPL(ntb_transport_free_queue);
2191 * ntb_transport_rx_remove - Dequeues enqueued rx packet
2192 * @qp: NTB queue to be freed
2193 * @len: pointer to variable to write enqueued buffers length
2195 * Dequeues unused buffers from receive queue. Should only be used during
2198 * RETURNS: NULL error value on error, or void* for success.
2200 void *ntb_transport_rx_remove(struct ntb_transport_qp *qp, unsigned int *len)
2202 struct ntb_queue_entry *entry;
2205 if (!qp || qp->client_ready)
2208 entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_pend_q);
2212 buf = entry->cb_data;
2215 ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry, &qp->rx_free_q);
2219 EXPORT_SYMBOL_GPL(ntb_transport_rx_remove);
2222 * ntb_transport_rx_enqueue - Enqueue a new NTB queue entry
2223 * @qp: NTB transport layer queue the entry is to be enqueued on
2224 * @cb: per buffer pointer for callback function to use
2225 * @data: pointer to data buffer that incoming packets will be copied into
2226 * @len: length of the data buffer
2228 * Enqueue a new receive buffer onto the transport queue into which a NTB
2229 * payload can be received into.
2231 * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
2233 int ntb_transport_rx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
2236 struct ntb_queue_entry *entry;
2241 entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q);
2245 entry->cb_data = cb;
2251 entry->rx_index = 0;
2253 ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry, &qp->rx_pend_q);
2256 tasklet_schedule(&qp->rxc_db_work);
2260 EXPORT_SYMBOL_GPL(ntb_transport_rx_enqueue);
2263 * ntb_transport_tx_enqueue - Enqueue a new NTB queue entry
2264 * @qp: NTB transport layer queue the entry is to be enqueued on
2265 * @cb: per buffer pointer for callback function to use
2266 * @data: pointer to data buffer that will be sent
2267 * @len: length of the data buffer
2269 * Enqueue a new transmit buffer onto the transport queue from which a NTB
2270 * payload will be transmitted. This assumes that a lock is being held to
2271 * serialize access to the qp.
2273 * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
2275 int ntb_transport_tx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
2278 struct ntb_queue_entry *entry;
2281 if (!qp || !qp->link_is_up || !len)
2284 entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
2286 qp->tx_err_no_buf++;
2290 entry->cb_data = cb;
2296 entry->tx_index = 0;
2298 rc = ntb_process_tx(qp, entry);
2300 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
2305 EXPORT_SYMBOL_GPL(ntb_transport_tx_enqueue);
2308 * ntb_transport_link_up - Notify NTB transport of client readiness to use queue
2309 * @qp: NTB transport layer queue to be enabled
2311 * Notify NTB transport layer of client readiness to use queue
2313 void ntb_transport_link_up(struct ntb_transport_qp *qp)
2318 qp->client_ready = true;
2320 if (qp->transport->link_is_up)
2321 schedule_delayed_work(&qp->link_work, 0);
2323 EXPORT_SYMBOL_GPL(ntb_transport_link_up);
2326 * ntb_transport_link_down - Notify NTB transport to no longer enqueue data
2327 * @qp: NTB transport layer queue to be disabled
2329 * Notify NTB transport layer of client's desire to no longer receive data on
2330 * transport queue specified. It is the client's responsibility to ensure all
2331 * entries on queue are purged or otherwise handled appropriately.
2333 void ntb_transport_link_down(struct ntb_transport_qp *qp)
2340 qp->client_ready = false;
2342 val = ntb_spad_read(qp->ndev, QP_LINKS);
2344 ntb_peer_spad_write(qp->ndev, PIDX, QP_LINKS, val & ~BIT(qp->qp_num));
2347 ntb_send_link_down(qp);
2349 cancel_delayed_work_sync(&qp->link_work);
2351 EXPORT_SYMBOL_GPL(ntb_transport_link_down);
2354 * ntb_transport_link_query - Query transport link state
2355 * @qp: NTB transport layer queue to be queried
2357 * Query connectivity to the remote system of the NTB transport queue
2359 * RETURNS: true for link up or false for link down
2361 bool ntb_transport_link_query(struct ntb_transport_qp *qp)
2366 return qp->link_is_up;
2368 EXPORT_SYMBOL_GPL(ntb_transport_link_query);
2371 * ntb_transport_qp_num - Query the qp number
2372 * @qp: NTB transport layer queue to be queried
2374 * Query qp number of the NTB transport queue
2376 * RETURNS: a zero based number specifying the qp number
2378 unsigned char ntb_transport_qp_num(struct ntb_transport_qp *qp)
2385 EXPORT_SYMBOL_GPL(ntb_transport_qp_num);
2388 * ntb_transport_max_size - Query the max payload size of a qp
2389 * @qp: NTB transport layer queue to be queried
2391 * Query the maximum payload size permissible on the given qp
2393 * RETURNS: the max payload size of a qp
2395 unsigned int ntb_transport_max_size(struct ntb_transport_qp *qp)
2397 unsigned int max_size;
2398 unsigned int copy_align;
2399 struct dma_chan *rx_chan, *tx_chan;
2404 rx_chan = qp->rx_dma_chan;
2405 tx_chan = qp->tx_dma_chan;
2407 copy_align = max(rx_chan ? rx_chan->device->copy_align : 0,
2408 tx_chan ? tx_chan->device->copy_align : 0);
2410 /* If DMA engine usage is possible, try to find the max size for that */
2411 max_size = qp->tx_max_frame - sizeof(struct ntb_payload_header);
2412 max_size = round_down(max_size, 1 << copy_align);
2416 EXPORT_SYMBOL_GPL(ntb_transport_max_size);
2418 unsigned int ntb_transport_tx_free_entry(struct ntb_transport_qp *qp)
2420 unsigned int head = qp->tx_index;
2421 unsigned int tail = qp->remote_rx_info->entry;
2423 return tail > head ? tail - head : qp->tx_max_entry + tail - head;
2425 EXPORT_SYMBOL_GPL(ntb_transport_tx_free_entry);
2427 static void ntb_transport_doorbell_callback(void *data, int vector)
2429 struct ntb_transport_ctx *nt = data;
2430 struct ntb_transport_qp *qp;
2432 unsigned int qp_num;
2434 if (ntb_db_read(nt->ndev) & nt->msi_db_mask) {
2435 ntb_transport_msi_peer_desc_changed(nt);
2436 ntb_db_clear(nt->ndev, nt->msi_db_mask);
2439 db_bits = (nt->qp_bitmap & ~nt->qp_bitmap_free &
2440 ntb_db_vector_mask(nt->ndev, vector));
2443 qp_num = __ffs(db_bits);
2444 qp = &nt->qp_vec[qp_num];
2447 tasklet_schedule(&qp->rxc_db_work);
2449 db_bits &= ~BIT_ULL(qp_num);
2453 static const struct ntb_ctx_ops ntb_transport_ops = {
2454 .link_event = ntb_transport_event_callback,
2455 .db_event = ntb_transport_doorbell_callback,
2458 static struct ntb_client ntb_transport_client = {
2460 .probe = ntb_transport_probe,
2461 .remove = ntb_transport_free,
2465 static int __init ntb_transport_init(void)
2469 pr_info("%s, version %s\n", NTB_TRANSPORT_DESC, NTB_TRANSPORT_VER);
2471 if (debugfs_initialized())
2472 nt_debugfs_dir = debugfs_create_dir(KBUILD_MODNAME, NULL);
2474 rc = bus_register(&ntb_transport_bus);
2478 rc = ntb_register_client(&ntb_transport_client);
2485 bus_unregister(&ntb_transport_bus);
2487 debugfs_remove_recursive(nt_debugfs_dir);
2490 module_init(ntb_transport_init);
2492 static void __exit ntb_transport_exit(void)
2494 ntb_unregister_client(&ntb_transport_client);
2495 bus_unregister(&ntb_transport_bus);
2496 debugfs_remove_recursive(nt_debugfs_dir);
2498 module_exit(ntb_transport_exit);