Back to home page

OSCL-LXR
 
 

     

0001 /*
0002  * This file is provided under a dual BSD/GPLv2 license.  When using or
0003  *   redistributing this file, you may do so under either license.
0004  *
0005  *   GPL LICENSE SUMMARY
0006  *
0007  *   Copyright(c) 2012 Intel Corporation. All rights reserved.
0008  *   Copyright (C) 2015 EMC Corporation. All Rights Reserved.
0009  *
0010  *   This program is free software; you can redistribute it and/or modify
0011  *   it under the terms of version 2 of the GNU General Public License as
0012  *   published by the Free Software Foundation.
0013  *
0014  *   BSD LICENSE
0015  *
0016  *   Copyright(c) 2012 Intel Corporation. All rights reserved.
0017  *   Copyright (C) 2015 EMC Corporation. All Rights Reserved.
0018  *
0019  *   Redistribution and use in source and binary forms, with or without
0020  *   modification, are permitted provided that the following conditions
0021  *   are met:
0022  *
0023  *     * Redistributions of source code must retain the above copyright
0024  *       notice, this list of conditions and the following disclaimer.
0025  *     * Redistributions in binary form must reproduce the above copy
0026  *       notice, this list of conditions and the following disclaimer in
0027  *       the documentation and/or other materials provided with the
0028  *       distribution.
0029  *     * Neither the name of Intel Corporation nor the names of its
0030  *       contributors may be used to endorse or promote products derived
0031  *       from this software without specific prior written permission.
0032  *
0033  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
0034  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
0035  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
0036  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
0037  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
0038  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
0039  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
0040  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
0041  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
0042  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
0043  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
0044  *
0045  * PCIe NTB Transport Linux driver
0046  *
0047  * Contact Information:
0048  * Jon Mason <jon.mason@intel.com>
0049  */
0050 #include <linux/debugfs.h>
0051 #include <linux/delay.h>
0052 #include <linux/dmaengine.h>
0053 #include <linux/dma-mapping.h>
0054 #include <linux/errno.h>
0055 #include <linux/export.h>
0056 #include <linux/interrupt.h>
0057 #include <linux/module.h>
0058 #include <linux/pci.h>
0059 #include <linux/slab.h>
0060 #include <linux/types.h>
0061 #include <linux/uaccess.h>
0062 #include "linux/ntb.h"
0063 #include "linux/ntb_transport.h"
0064 
0065 #define NTB_TRANSPORT_VERSION   4
0066 #define NTB_TRANSPORT_VER   "4"
0067 #define NTB_TRANSPORT_NAME  "ntb_transport"
0068 #define NTB_TRANSPORT_DESC  "Software Queue-Pair Transport over NTB"
0069 #define NTB_TRANSPORT_MIN_SPADS (MW0_SZ_HIGH + 2)
0070 
0071 MODULE_DESCRIPTION(NTB_TRANSPORT_DESC);
0072 MODULE_VERSION(NTB_TRANSPORT_VER);
0073 MODULE_LICENSE("Dual BSD/GPL");
0074 MODULE_AUTHOR("Intel Corporation");
0075 
0076 static unsigned long max_mw_size;
0077 module_param(max_mw_size, ulong, 0644);
0078 MODULE_PARM_DESC(max_mw_size, "Limit size of large memory windows");
0079 
0080 static unsigned int transport_mtu = 0x10000;
0081 module_param(transport_mtu, uint, 0644);
0082 MODULE_PARM_DESC(transport_mtu, "Maximum size of NTB transport packets");
0083 
0084 static unsigned char max_num_clients;
0085 module_param(max_num_clients, byte, 0644);
0086 MODULE_PARM_DESC(max_num_clients, "Maximum number of NTB transport clients");
0087 
0088 static unsigned int copy_bytes = 1024;
0089 module_param(copy_bytes, uint, 0644);
0090 MODULE_PARM_DESC(copy_bytes, "Threshold under which NTB will use the CPU to copy instead of DMA");
0091 
0092 static bool use_dma;
0093 module_param(use_dma, bool, 0644);
0094 MODULE_PARM_DESC(use_dma, "Use DMA engine to perform large data copy");
0095 
0096 static bool use_msi;
0097 #ifdef CONFIG_NTB_MSI
0098 module_param(use_msi, bool, 0644);
0099 MODULE_PARM_DESC(use_msi, "Use MSI interrupts instead of doorbells");
0100 #endif
0101 
0102 static struct dentry *nt_debugfs_dir;
0103 
0104 /* Only two-ports NTB devices are supported */
0105 #define PIDX        NTB_DEF_PEER_IDX
0106 
0107 struct ntb_queue_entry {
0108     /* ntb_queue list reference */
0109     struct list_head entry;
0110     /* pointers to data to be transferred */
0111     void *cb_data;
0112     void *buf;
0113     unsigned int len;
0114     unsigned int flags;
0115     int retries;
0116     int errors;
0117     unsigned int tx_index;
0118     unsigned int rx_index;
0119 
0120     struct ntb_transport_qp *qp;
0121     union {
0122         struct ntb_payload_header __iomem *tx_hdr;
0123         struct ntb_payload_header *rx_hdr;
0124     };
0125 };
0126 
0127 struct ntb_rx_info {
0128     unsigned int entry;
0129 };
0130 
0131 struct ntb_transport_qp {
0132     struct ntb_transport_ctx *transport;
0133     struct ntb_dev *ndev;
0134     void *cb_data;
0135     struct dma_chan *tx_dma_chan;
0136     struct dma_chan *rx_dma_chan;
0137 
0138     bool client_ready;
0139     bool link_is_up;
0140     bool active;
0141 
0142     u8 qp_num;  /* Only 64 QP's are allowed.  0-63 */
0143     u64 qp_bit;
0144 
0145     struct ntb_rx_info __iomem *rx_info;
0146     struct ntb_rx_info *remote_rx_info;
0147 
0148     void (*tx_handler)(struct ntb_transport_qp *qp, void *qp_data,
0149                void *data, int len);
0150     struct list_head tx_free_q;
0151     spinlock_t ntb_tx_free_q_lock;
0152     void __iomem *tx_mw;
0153     phys_addr_t tx_mw_phys;
0154     size_t tx_mw_size;
0155     dma_addr_t tx_mw_dma_addr;
0156     unsigned int tx_index;
0157     unsigned int tx_max_entry;
0158     unsigned int tx_max_frame;
0159 
0160     void (*rx_handler)(struct ntb_transport_qp *qp, void *qp_data,
0161                void *data, int len);
0162     struct list_head rx_post_q;
0163     struct list_head rx_pend_q;
0164     struct list_head rx_free_q;
0165     /* ntb_rx_q_lock: synchronize access to rx_XXXX_q */
0166     spinlock_t ntb_rx_q_lock;
0167     void *rx_buff;
0168     unsigned int rx_index;
0169     unsigned int rx_max_entry;
0170     unsigned int rx_max_frame;
0171     unsigned int rx_alloc_entry;
0172     dma_cookie_t last_cookie;
0173     struct tasklet_struct rxc_db_work;
0174 
0175     void (*event_handler)(void *data, int status);
0176     struct delayed_work link_work;
0177     struct work_struct link_cleanup;
0178 
0179     struct dentry *debugfs_dir;
0180     struct dentry *debugfs_stats;
0181 
0182     /* Stats */
0183     u64 rx_bytes;
0184     u64 rx_pkts;
0185     u64 rx_ring_empty;
0186     u64 rx_err_no_buf;
0187     u64 rx_err_oflow;
0188     u64 rx_err_ver;
0189     u64 rx_memcpy;
0190     u64 rx_async;
0191     u64 tx_bytes;
0192     u64 tx_pkts;
0193     u64 tx_ring_full;
0194     u64 tx_err_no_buf;
0195     u64 tx_memcpy;
0196     u64 tx_async;
0197 
0198     bool use_msi;
0199     int msi_irq;
0200     struct ntb_msi_desc msi_desc;
0201     struct ntb_msi_desc peer_msi_desc;
0202 };
0203 
0204 struct ntb_transport_mw {
0205     phys_addr_t phys_addr;
0206     resource_size_t phys_size;
0207     void __iomem *vbase;
0208     size_t xlat_size;
0209     size_t buff_size;
0210     size_t alloc_size;
0211     void *alloc_addr;
0212     void *virt_addr;
0213     dma_addr_t dma_addr;
0214 };
0215 
0216 struct ntb_transport_client_dev {
0217     struct list_head entry;
0218     struct ntb_transport_ctx *nt;
0219     struct device dev;
0220 };
0221 
0222 struct ntb_transport_ctx {
0223     struct list_head entry;
0224     struct list_head client_devs;
0225 
0226     struct ntb_dev *ndev;
0227 
0228     struct ntb_transport_mw *mw_vec;
0229     struct ntb_transport_qp *qp_vec;
0230     unsigned int mw_count;
0231     unsigned int qp_count;
0232     u64 qp_bitmap;
0233     u64 qp_bitmap_free;
0234 
0235     bool use_msi;
0236     unsigned int msi_spad_offset;
0237     u64 msi_db_mask;
0238 
0239     bool link_is_up;
0240     struct delayed_work link_work;
0241     struct work_struct link_cleanup;
0242 
0243     struct dentry *debugfs_node_dir;
0244 };
0245 
0246 enum {
0247     DESC_DONE_FLAG = BIT(0),
0248     LINK_DOWN_FLAG = BIT(1),
0249 };
0250 
0251 struct ntb_payload_header {
0252     unsigned int ver;
0253     unsigned int len;
0254     unsigned int flags;
0255 };
0256 
0257 enum {
0258     VERSION = 0,
0259     QP_LINKS,
0260     NUM_QPS,
0261     NUM_MWS,
0262     MW0_SZ_HIGH,
0263     MW0_SZ_LOW,
0264 };
0265 
0266 #define dev_client_dev(__dev) \
0267     container_of((__dev), struct ntb_transport_client_dev, dev)
0268 
0269 #define drv_client(__drv) \
0270     container_of((__drv), struct ntb_transport_client, driver)
0271 
0272 #define QP_TO_MW(nt, qp)    ((qp) % nt->mw_count)
0273 #define NTB_QP_DEF_NUM_ENTRIES  100
0274 #define NTB_LINK_DOWN_TIMEOUT   10
0275 
0276 static void ntb_transport_rxc_db(unsigned long data);
0277 static const struct ntb_ctx_ops ntb_transport_ops;
0278 static struct ntb_client ntb_transport_client;
0279 static int ntb_async_tx_submit(struct ntb_transport_qp *qp,
0280                    struct ntb_queue_entry *entry);
0281 static void ntb_memcpy_tx(struct ntb_queue_entry *entry, void __iomem *offset);
0282 static int ntb_async_rx_submit(struct ntb_queue_entry *entry, void *offset);
0283 static void ntb_memcpy_rx(struct ntb_queue_entry *entry, void *offset);
0284 
0285 
0286 static int ntb_transport_bus_match(struct device *dev,
0287                    struct device_driver *drv)
0288 {
0289     return !strncmp(dev_name(dev), drv->name, strlen(drv->name));
0290 }
0291 
0292 static int ntb_transport_bus_probe(struct device *dev)
0293 {
0294     const struct ntb_transport_client *client;
0295     int rc;
0296 
0297     get_device(dev);
0298 
0299     client = drv_client(dev->driver);
0300     rc = client->probe(dev);
0301     if (rc)
0302         put_device(dev);
0303 
0304     return rc;
0305 }
0306 
0307 static void ntb_transport_bus_remove(struct device *dev)
0308 {
0309     const struct ntb_transport_client *client;
0310 
0311     client = drv_client(dev->driver);
0312     client->remove(dev);
0313 
0314     put_device(dev);
0315 }
0316 
0317 static struct bus_type ntb_transport_bus = {
0318     .name = "ntb_transport",
0319     .match = ntb_transport_bus_match,
0320     .probe = ntb_transport_bus_probe,
0321     .remove = ntb_transport_bus_remove,
0322 };
0323 
0324 static LIST_HEAD(ntb_transport_list);
0325 
0326 static int ntb_bus_init(struct ntb_transport_ctx *nt)
0327 {
0328     list_add_tail(&nt->entry, &ntb_transport_list);
0329     return 0;
0330 }
0331 
0332 static void ntb_bus_remove(struct ntb_transport_ctx *nt)
0333 {
0334     struct ntb_transport_client_dev *client_dev, *cd;
0335 
0336     list_for_each_entry_safe(client_dev, cd, &nt->client_devs, entry) {
0337         dev_err(client_dev->dev.parent, "%s still attached to bus, removing\n",
0338             dev_name(&client_dev->dev));
0339         list_del(&client_dev->entry);
0340         device_unregister(&client_dev->dev);
0341     }
0342 
0343     list_del(&nt->entry);
0344 }
0345 
0346 static void ntb_transport_client_release(struct device *dev)
0347 {
0348     struct ntb_transport_client_dev *client_dev;
0349 
0350     client_dev = dev_client_dev(dev);
0351     kfree(client_dev);
0352 }
0353 
0354 /**
0355  * ntb_transport_unregister_client_dev - Unregister NTB client device
0356  * @device_name: Name of NTB client device
0357  *
0358  * Unregister an NTB client device with the NTB transport layer
0359  */
0360 void ntb_transport_unregister_client_dev(char *device_name)
0361 {
0362     struct ntb_transport_client_dev *client, *cd;
0363     struct ntb_transport_ctx *nt;
0364 
0365     list_for_each_entry(nt, &ntb_transport_list, entry)
0366         list_for_each_entry_safe(client, cd, &nt->client_devs, entry)
0367             if (!strncmp(dev_name(&client->dev), device_name,
0368                      strlen(device_name))) {
0369                 list_del(&client->entry);
0370                 device_unregister(&client->dev);
0371             }
0372 }
0373 EXPORT_SYMBOL_GPL(ntb_transport_unregister_client_dev);
0374 
0375 /**
0376  * ntb_transport_register_client_dev - Register NTB client device
0377  * @device_name: Name of NTB client device
0378  *
0379  * Register an NTB client device with the NTB transport layer
0380  */
0381 int ntb_transport_register_client_dev(char *device_name)
0382 {
0383     struct ntb_transport_client_dev *client_dev;
0384     struct ntb_transport_ctx *nt;
0385     int node;
0386     int rc, i = 0;
0387 
0388     if (list_empty(&ntb_transport_list))
0389         return -ENODEV;
0390 
0391     list_for_each_entry(nt, &ntb_transport_list, entry) {
0392         struct device *dev;
0393 
0394         node = dev_to_node(&nt->ndev->dev);
0395 
0396         client_dev = kzalloc_node(sizeof(*client_dev),
0397                       GFP_KERNEL, node);
0398         if (!client_dev) {
0399             rc = -ENOMEM;
0400             goto err;
0401         }
0402 
0403         dev = &client_dev->dev;
0404 
0405         /* setup and register client devices */
0406         dev_set_name(dev, "%s%d", device_name, i);
0407         dev->bus = &ntb_transport_bus;
0408         dev->release = ntb_transport_client_release;
0409         dev->parent = &nt->ndev->dev;
0410 
0411         rc = device_register(dev);
0412         if (rc) {
0413             kfree(client_dev);
0414             goto err;
0415         }
0416 
0417         list_add_tail(&client_dev->entry, &nt->client_devs);
0418         i++;
0419     }
0420 
0421     return 0;
0422 
0423 err:
0424     ntb_transport_unregister_client_dev(device_name);
0425 
0426     return rc;
0427 }
0428 EXPORT_SYMBOL_GPL(ntb_transport_register_client_dev);
0429 
0430 /**
0431  * ntb_transport_register_client - Register NTB client driver
0432  * @drv: NTB client driver to be registered
0433  *
0434  * Register an NTB client driver with the NTB transport layer
0435  *
0436  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
0437  */
0438 int ntb_transport_register_client(struct ntb_transport_client *drv)
0439 {
0440     drv->driver.bus = &ntb_transport_bus;
0441 
0442     if (list_empty(&ntb_transport_list))
0443         return -ENODEV;
0444 
0445     return driver_register(&drv->driver);
0446 }
0447 EXPORT_SYMBOL_GPL(ntb_transport_register_client);
0448 
0449 /**
0450  * ntb_transport_unregister_client - Unregister NTB client driver
0451  * @drv: NTB client driver to be unregistered
0452  *
0453  * Unregister an NTB client driver with the NTB transport layer
0454  *
0455  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
0456  */
0457 void ntb_transport_unregister_client(struct ntb_transport_client *drv)
0458 {
0459     driver_unregister(&drv->driver);
0460 }
0461 EXPORT_SYMBOL_GPL(ntb_transport_unregister_client);
0462 
0463 static ssize_t debugfs_read(struct file *filp, char __user *ubuf, size_t count,
0464                 loff_t *offp)
0465 {
0466     struct ntb_transport_qp *qp;
0467     char *buf;
0468     ssize_t ret, out_offset, out_count;
0469 
0470     qp = filp->private_data;
0471 
0472     if (!qp || !qp->link_is_up)
0473         return 0;
0474 
0475     out_count = 1000;
0476 
0477     buf = kmalloc(out_count, GFP_KERNEL);
0478     if (!buf)
0479         return -ENOMEM;
0480 
0481     out_offset = 0;
0482     out_offset += scnprintf(buf + out_offset, out_count - out_offset,
0483                    "\nNTB QP stats:\n\n");
0484     out_offset += scnprintf(buf + out_offset, out_count - out_offset,
0485                    "rx_bytes - \t%llu\n", qp->rx_bytes);
0486     out_offset += scnprintf(buf + out_offset, out_count - out_offset,
0487                    "rx_pkts - \t%llu\n", qp->rx_pkts);
0488     out_offset += scnprintf(buf + out_offset, out_count - out_offset,
0489                    "rx_memcpy - \t%llu\n", qp->rx_memcpy);
0490     out_offset += scnprintf(buf + out_offset, out_count - out_offset,
0491                    "rx_async - \t%llu\n", qp->rx_async);
0492     out_offset += scnprintf(buf + out_offset, out_count - out_offset,
0493                    "rx_ring_empty - %llu\n", qp->rx_ring_empty);
0494     out_offset += scnprintf(buf + out_offset, out_count - out_offset,
0495                    "rx_err_no_buf - %llu\n", qp->rx_err_no_buf);
0496     out_offset += scnprintf(buf + out_offset, out_count - out_offset,
0497                    "rx_err_oflow - \t%llu\n", qp->rx_err_oflow);
0498     out_offset += scnprintf(buf + out_offset, out_count - out_offset,
0499                    "rx_err_ver - \t%llu\n", qp->rx_err_ver);
0500     out_offset += scnprintf(buf + out_offset, out_count - out_offset,
0501                    "rx_buff - \t0x%p\n", qp->rx_buff);
0502     out_offset += scnprintf(buf + out_offset, out_count - out_offset,
0503                    "rx_index - \t%u\n", qp->rx_index);
0504     out_offset += scnprintf(buf + out_offset, out_count - out_offset,
0505                    "rx_max_entry - \t%u\n", qp->rx_max_entry);
0506     out_offset += scnprintf(buf + out_offset, out_count - out_offset,
0507                    "rx_alloc_entry - \t%u\n\n", qp->rx_alloc_entry);
0508 
0509     out_offset += scnprintf(buf + out_offset, out_count - out_offset,
0510                    "tx_bytes - \t%llu\n", qp->tx_bytes);
0511     out_offset += scnprintf(buf + out_offset, out_count - out_offset,
0512                    "tx_pkts - \t%llu\n", qp->tx_pkts);
0513     out_offset += scnprintf(buf + out_offset, out_count - out_offset,
0514                    "tx_memcpy - \t%llu\n", qp->tx_memcpy);
0515     out_offset += scnprintf(buf + out_offset, out_count - out_offset,
0516                    "tx_async - \t%llu\n", qp->tx_async);
0517     out_offset += scnprintf(buf + out_offset, out_count - out_offset,
0518                    "tx_ring_full - \t%llu\n", qp->tx_ring_full);
0519     out_offset += scnprintf(buf + out_offset, out_count - out_offset,
0520                    "tx_err_no_buf - %llu\n", qp->tx_err_no_buf);
0521     out_offset += scnprintf(buf + out_offset, out_count - out_offset,
0522                    "tx_mw - \t0x%p\n", qp->tx_mw);
0523     out_offset += scnprintf(buf + out_offset, out_count - out_offset,
0524                    "tx_index (H) - \t%u\n", qp->tx_index);
0525     out_offset += scnprintf(buf + out_offset, out_count - out_offset,
0526                    "RRI (T) - \t%u\n",
0527                    qp->remote_rx_info->entry);
0528     out_offset += scnprintf(buf + out_offset, out_count - out_offset,
0529                    "tx_max_entry - \t%u\n", qp->tx_max_entry);
0530     out_offset += scnprintf(buf + out_offset, out_count - out_offset,
0531                    "free tx - \t%u\n",
0532                    ntb_transport_tx_free_entry(qp));
0533 
0534     out_offset += scnprintf(buf + out_offset, out_count - out_offset,
0535                    "\n");
0536     out_offset += scnprintf(buf + out_offset, out_count - out_offset,
0537                    "Using TX DMA - \t%s\n",
0538                    qp->tx_dma_chan ? "Yes" : "No");
0539     out_offset += scnprintf(buf + out_offset, out_count - out_offset,
0540                    "Using RX DMA - \t%s\n",
0541                    qp->rx_dma_chan ? "Yes" : "No");
0542     out_offset += scnprintf(buf + out_offset, out_count - out_offset,
0543                    "QP Link - \t%s\n",
0544                    qp->link_is_up ? "Up" : "Down");
0545     out_offset += scnprintf(buf + out_offset, out_count - out_offset,
0546                    "\n");
0547 
0548     if (out_offset > out_count)
0549         out_offset = out_count;
0550 
0551     ret = simple_read_from_buffer(ubuf, count, offp, buf, out_offset);
0552     kfree(buf);
0553     return ret;
0554 }
0555 
0556 static const struct file_operations ntb_qp_debugfs_stats = {
0557     .owner = THIS_MODULE,
0558     .open = simple_open,
0559     .read = debugfs_read,
0560 };
0561 
0562 static void ntb_list_add(spinlock_t *lock, struct list_head *entry,
0563              struct list_head *list)
0564 {
0565     unsigned long flags;
0566 
0567     spin_lock_irqsave(lock, flags);
0568     list_add_tail(entry, list);
0569     spin_unlock_irqrestore(lock, flags);
0570 }
0571 
0572 static struct ntb_queue_entry *ntb_list_rm(spinlock_t *lock,
0573                        struct list_head *list)
0574 {
0575     struct ntb_queue_entry *entry;
0576     unsigned long flags;
0577 
0578     spin_lock_irqsave(lock, flags);
0579     if (list_empty(list)) {
0580         entry = NULL;
0581         goto out;
0582     }
0583     entry = list_first_entry(list, struct ntb_queue_entry, entry);
0584     list_del(&entry->entry);
0585 
0586 out:
0587     spin_unlock_irqrestore(lock, flags);
0588 
0589     return entry;
0590 }
0591 
0592 static struct ntb_queue_entry *ntb_list_mv(spinlock_t *lock,
0593                        struct list_head *list,
0594                        struct list_head *to_list)
0595 {
0596     struct ntb_queue_entry *entry;
0597     unsigned long flags;
0598 
0599     spin_lock_irqsave(lock, flags);
0600 
0601     if (list_empty(list)) {
0602         entry = NULL;
0603     } else {
0604         entry = list_first_entry(list, struct ntb_queue_entry, entry);
0605         list_move_tail(&entry->entry, to_list);
0606     }
0607 
0608     spin_unlock_irqrestore(lock, flags);
0609 
0610     return entry;
0611 }
0612 
0613 static int ntb_transport_setup_qp_mw(struct ntb_transport_ctx *nt,
0614                      unsigned int qp_num)
0615 {
0616     struct ntb_transport_qp *qp = &nt->qp_vec[qp_num];
0617     struct ntb_transport_mw *mw;
0618     struct ntb_dev *ndev = nt->ndev;
0619     struct ntb_queue_entry *entry;
0620     unsigned int rx_size, num_qps_mw;
0621     unsigned int mw_num, mw_count, qp_count;
0622     unsigned int i;
0623     int node;
0624 
0625     mw_count = nt->mw_count;
0626     qp_count = nt->qp_count;
0627 
0628     mw_num = QP_TO_MW(nt, qp_num);
0629     mw = &nt->mw_vec[mw_num];
0630 
0631     if (!mw->virt_addr)
0632         return -ENOMEM;
0633 
0634     if (mw_num < qp_count % mw_count)
0635         num_qps_mw = qp_count / mw_count + 1;
0636     else
0637         num_qps_mw = qp_count / mw_count;
0638 
0639     rx_size = (unsigned int)mw->xlat_size / num_qps_mw;
0640     qp->rx_buff = mw->virt_addr + rx_size * (qp_num / mw_count);
0641     rx_size -= sizeof(struct ntb_rx_info);
0642 
0643     qp->remote_rx_info = qp->rx_buff + rx_size;
0644 
0645     /* Due to housekeeping, there must be atleast 2 buffs */
0646     qp->rx_max_frame = min(transport_mtu, rx_size / 2);
0647     qp->rx_max_entry = rx_size / qp->rx_max_frame;
0648     qp->rx_index = 0;
0649 
0650     /*
0651      * Checking to see if we have more entries than the default.
0652      * We should add additional entries if that is the case so we
0653      * can be in sync with the transport frames.
0654      */
0655     node = dev_to_node(&ndev->dev);
0656     for (i = qp->rx_alloc_entry; i < qp->rx_max_entry; i++) {
0657         entry = kzalloc_node(sizeof(*entry), GFP_KERNEL, node);
0658         if (!entry)
0659             return -ENOMEM;
0660 
0661         entry->qp = qp;
0662         ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry,
0663                  &qp->rx_free_q);
0664         qp->rx_alloc_entry++;
0665     }
0666 
0667     qp->remote_rx_info->entry = qp->rx_max_entry - 1;
0668 
0669     /* setup the hdr offsets with 0's */
0670     for (i = 0; i < qp->rx_max_entry; i++) {
0671         void *offset = (qp->rx_buff + qp->rx_max_frame * (i + 1) -
0672                 sizeof(struct ntb_payload_header));
0673         memset(offset, 0, sizeof(struct ntb_payload_header));
0674     }
0675 
0676     qp->rx_pkts = 0;
0677     qp->tx_pkts = 0;
0678     qp->tx_index = 0;
0679 
0680     return 0;
0681 }
0682 
0683 static irqreturn_t ntb_transport_isr(int irq, void *dev)
0684 {
0685     struct ntb_transport_qp *qp = dev;
0686 
0687     tasklet_schedule(&qp->rxc_db_work);
0688 
0689     return IRQ_HANDLED;
0690 }
0691 
0692 static void ntb_transport_setup_qp_peer_msi(struct ntb_transport_ctx *nt,
0693                         unsigned int qp_num)
0694 {
0695     struct ntb_transport_qp *qp = &nt->qp_vec[qp_num];
0696     int spad = qp_num * 2 + nt->msi_spad_offset;
0697 
0698     if (!nt->use_msi)
0699         return;
0700 
0701     if (spad >= ntb_spad_count(nt->ndev))
0702         return;
0703 
0704     qp->peer_msi_desc.addr_offset =
0705         ntb_peer_spad_read(qp->ndev, PIDX, spad);
0706     qp->peer_msi_desc.data =
0707         ntb_peer_spad_read(qp->ndev, PIDX, spad + 1);
0708 
0709     dev_dbg(&qp->ndev->pdev->dev, "QP%d Peer MSI addr=%x data=%x\n",
0710         qp_num, qp->peer_msi_desc.addr_offset, qp->peer_msi_desc.data);
0711 
0712     if (qp->peer_msi_desc.addr_offset) {
0713         qp->use_msi = true;
0714         dev_info(&qp->ndev->pdev->dev,
0715              "Using MSI interrupts for QP%d\n", qp_num);
0716     }
0717 }
0718 
0719 static void ntb_transport_setup_qp_msi(struct ntb_transport_ctx *nt,
0720                        unsigned int qp_num)
0721 {
0722     struct ntb_transport_qp *qp = &nt->qp_vec[qp_num];
0723     int spad = qp_num * 2 + nt->msi_spad_offset;
0724     int rc;
0725 
0726     if (!nt->use_msi)
0727         return;
0728 
0729     if (spad >= ntb_spad_count(nt->ndev)) {
0730         dev_warn_once(&qp->ndev->pdev->dev,
0731                   "Not enough SPADS to use MSI interrupts\n");
0732         return;
0733     }
0734 
0735     ntb_spad_write(qp->ndev, spad, 0);
0736     ntb_spad_write(qp->ndev, spad + 1, 0);
0737 
0738     if (!qp->msi_irq) {
0739         qp->msi_irq = ntbm_msi_request_irq(qp->ndev, ntb_transport_isr,
0740                            KBUILD_MODNAME, qp,
0741                            &qp->msi_desc);
0742         if (qp->msi_irq < 0) {
0743             dev_warn(&qp->ndev->pdev->dev,
0744                  "Unable to allocate MSI interrupt for qp%d\n",
0745                  qp_num);
0746             return;
0747         }
0748     }
0749 
0750     rc = ntb_spad_write(qp->ndev, spad, qp->msi_desc.addr_offset);
0751     if (rc)
0752         goto err_free_interrupt;
0753 
0754     rc = ntb_spad_write(qp->ndev, spad + 1, qp->msi_desc.data);
0755     if (rc)
0756         goto err_free_interrupt;
0757 
0758     dev_dbg(&qp->ndev->pdev->dev, "QP%d MSI %d addr=%x data=%x\n",
0759         qp_num, qp->msi_irq, qp->msi_desc.addr_offset,
0760         qp->msi_desc.data);
0761 
0762     return;
0763 
0764 err_free_interrupt:
0765     devm_free_irq(&nt->ndev->dev, qp->msi_irq, qp);
0766 }
0767 
0768 static void ntb_transport_msi_peer_desc_changed(struct ntb_transport_ctx *nt)
0769 {
0770     int i;
0771 
0772     dev_dbg(&nt->ndev->pdev->dev, "Peer MSI descriptors changed");
0773 
0774     for (i = 0; i < nt->qp_count; i++)
0775         ntb_transport_setup_qp_peer_msi(nt, i);
0776 }
0777 
0778 static void ntb_transport_msi_desc_changed(void *data)
0779 {
0780     struct ntb_transport_ctx *nt = data;
0781     int i;
0782 
0783     dev_dbg(&nt->ndev->pdev->dev, "MSI descriptors changed");
0784 
0785     for (i = 0; i < nt->qp_count; i++)
0786         ntb_transport_setup_qp_msi(nt, i);
0787 
0788     ntb_peer_db_set(nt->ndev, nt->msi_db_mask);
0789 }
0790 
0791 static void ntb_free_mw(struct ntb_transport_ctx *nt, int num_mw)
0792 {
0793     struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
0794     struct pci_dev *pdev = nt->ndev->pdev;
0795 
0796     if (!mw->virt_addr)
0797         return;
0798 
0799     ntb_mw_clear_trans(nt->ndev, PIDX, num_mw);
0800     dma_free_coherent(&pdev->dev, mw->alloc_size,
0801               mw->alloc_addr, mw->dma_addr);
0802     mw->xlat_size = 0;
0803     mw->buff_size = 0;
0804     mw->alloc_size = 0;
0805     mw->alloc_addr = NULL;
0806     mw->virt_addr = NULL;
0807 }
0808 
0809 static int ntb_alloc_mw_buffer(struct ntb_transport_mw *mw,
0810                    struct device *dma_dev, size_t align)
0811 {
0812     dma_addr_t dma_addr;
0813     void *alloc_addr, *virt_addr;
0814     int rc;
0815 
0816     alloc_addr = dma_alloc_coherent(dma_dev, mw->alloc_size,
0817                     &dma_addr, GFP_KERNEL);
0818     if (!alloc_addr) {
0819         dev_err(dma_dev, "Unable to alloc MW buff of size %zu\n",
0820             mw->alloc_size);
0821         return -ENOMEM;
0822     }
0823     virt_addr = alloc_addr;
0824 
0825     /*
0826      * we must ensure that the memory address allocated is BAR size
0827      * aligned in order for the XLAT register to take the value. This
0828      * is a requirement of the hardware. It is recommended to setup CMA
0829      * for BAR sizes equal or greater than 4MB.
0830      */
0831     if (!IS_ALIGNED(dma_addr, align)) {
0832         if (mw->alloc_size > mw->buff_size) {
0833             virt_addr = PTR_ALIGN(alloc_addr, align);
0834             dma_addr = ALIGN(dma_addr, align);
0835         } else {
0836             rc = -ENOMEM;
0837             goto err;
0838         }
0839     }
0840 
0841     mw->alloc_addr = alloc_addr;
0842     mw->virt_addr = virt_addr;
0843     mw->dma_addr = dma_addr;
0844 
0845     return 0;
0846 
0847 err:
0848     dma_free_coherent(dma_dev, mw->alloc_size, alloc_addr, dma_addr);
0849 
0850     return rc;
0851 }
0852 
0853 static int ntb_set_mw(struct ntb_transport_ctx *nt, int num_mw,
0854               resource_size_t size)
0855 {
0856     struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
0857     struct pci_dev *pdev = nt->ndev->pdev;
0858     size_t xlat_size, buff_size;
0859     resource_size_t xlat_align;
0860     resource_size_t xlat_align_size;
0861     int rc;
0862 
0863     if (!size)
0864         return -EINVAL;
0865 
0866     rc = ntb_mw_get_align(nt->ndev, PIDX, num_mw, &xlat_align,
0867                   &xlat_align_size, NULL);
0868     if (rc)
0869         return rc;
0870 
0871     xlat_size = round_up(size, xlat_align_size);
0872     buff_size = round_up(size, xlat_align);
0873 
0874     /* No need to re-setup */
0875     if (mw->xlat_size == xlat_size)
0876         return 0;
0877 
0878     if (mw->buff_size)
0879         ntb_free_mw(nt, num_mw);
0880 
0881     /* Alloc memory for receiving data.  Must be aligned */
0882     mw->xlat_size = xlat_size;
0883     mw->buff_size = buff_size;
0884     mw->alloc_size = buff_size;
0885 
0886     rc = ntb_alloc_mw_buffer(mw, &pdev->dev, xlat_align);
0887     if (rc) {
0888         mw->alloc_size *= 2;
0889         rc = ntb_alloc_mw_buffer(mw, &pdev->dev, xlat_align);
0890         if (rc) {
0891             dev_err(&pdev->dev,
0892                 "Unable to alloc aligned MW buff\n");
0893             mw->xlat_size = 0;
0894             mw->buff_size = 0;
0895             mw->alloc_size = 0;
0896             return rc;
0897         }
0898     }
0899 
0900     /* Notify HW the memory location of the receive buffer */
0901     rc = ntb_mw_set_trans(nt->ndev, PIDX, num_mw, mw->dma_addr,
0902                   mw->xlat_size);
0903     if (rc) {
0904         dev_err(&pdev->dev, "Unable to set mw%d translation", num_mw);
0905         ntb_free_mw(nt, num_mw);
0906         return -EIO;
0907     }
0908 
0909     return 0;
0910 }
0911 
0912 static void ntb_qp_link_down_reset(struct ntb_transport_qp *qp)
0913 {
0914     qp->link_is_up = false;
0915     qp->active = false;
0916 
0917     qp->tx_index = 0;
0918     qp->rx_index = 0;
0919     qp->rx_bytes = 0;
0920     qp->rx_pkts = 0;
0921     qp->rx_ring_empty = 0;
0922     qp->rx_err_no_buf = 0;
0923     qp->rx_err_oflow = 0;
0924     qp->rx_err_ver = 0;
0925     qp->rx_memcpy = 0;
0926     qp->rx_async = 0;
0927     qp->tx_bytes = 0;
0928     qp->tx_pkts = 0;
0929     qp->tx_ring_full = 0;
0930     qp->tx_err_no_buf = 0;
0931     qp->tx_memcpy = 0;
0932     qp->tx_async = 0;
0933 }
0934 
0935 static void ntb_qp_link_cleanup(struct ntb_transport_qp *qp)
0936 {
0937     struct ntb_transport_ctx *nt = qp->transport;
0938     struct pci_dev *pdev = nt->ndev->pdev;
0939 
0940     dev_info(&pdev->dev, "qp %d: Link Cleanup\n", qp->qp_num);
0941 
0942     cancel_delayed_work_sync(&qp->link_work);
0943     ntb_qp_link_down_reset(qp);
0944 
0945     if (qp->event_handler)
0946         qp->event_handler(qp->cb_data, qp->link_is_up);
0947 }
0948 
0949 static void ntb_qp_link_cleanup_work(struct work_struct *work)
0950 {
0951     struct ntb_transport_qp *qp = container_of(work,
0952                            struct ntb_transport_qp,
0953                            link_cleanup);
0954     struct ntb_transport_ctx *nt = qp->transport;
0955 
0956     ntb_qp_link_cleanup(qp);
0957 
0958     if (nt->link_is_up)
0959         schedule_delayed_work(&qp->link_work,
0960                       msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
0961 }
0962 
0963 static void ntb_qp_link_down(struct ntb_transport_qp *qp)
0964 {
0965     schedule_work(&qp->link_cleanup);
0966 }
0967 
0968 static void ntb_transport_link_cleanup(struct ntb_transport_ctx *nt)
0969 {
0970     struct ntb_transport_qp *qp;
0971     u64 qp_bitmap_alloc;
0972     unsigned int i, count;
0973 
0974     qp_bitmap_alloc = nt->qp_bitmap & ~nt->qp_bitmap_free;
0975 
0976     /* Pass along the info to any clients */
0977     for (i = 0; i < nt->qp_count; i++)
0978         if (qp_bitmap_alloc & BIT_ULL(i)) {
0979             qp = &nt->qp_vec[i];
0980             ntb_qp_link_cleanup(qp);
0981             cancel_work_sync(&qp->link_cleanup);
0982             cancel_delayed_work_sync(&qp->link_work);
0983         }
0984 
0985     if (!nt->link_is_up)
0986         cancel_delayed_work_sync(&nt->link_work);
0987 
0988     for (i = 0; i < nt->mw_count; i++)
0989         ntb_free_mw(nt, i);
0990 
0991     /* The scratchpad registers keep the values if the remote side
0992      * goes down, blast them now to give them a sane value the next
0993      * time they are accessed
0994      */
0995     count = ntb_spad_count(nt->ndev);
0996     for (i = 0; i < count; i++)
0997         ntb_spad_write(nt->ndev, i, 0);
0998 }
0999 
1000 static void ntb_transport_link_cleanup_work(struct work_struct *work)
1001 {
1002     struct ntb_transport_ctx *nt =
1003         container_of(work, struct ntb_transport_ctx, link_cleanup);
1004 
1005     ntb_transport_link_cleanup(nt);
1006 }
1007 
1008 static void ntb_transport_event_callback(void *data)
1009 {
1010     struct ntb_transport_ctx *nt = data;
1011 
1012     if (ntb_link_is_up(nt->ndev, NULL, NULL) == 1)
1013         schedule_delayed_work(&nt->link_work, 0);
1014     else
1015         schedule_work(&nt->link_cleanup);
1016 }
1017 
1018 static void ntb_transport_link_work(struct work_struct *work)
1019 {
1020     struct ntb_transport_ctx *nt =
1021         container_of(work, struct ntb_transport_ctx, link_work.work);
1022     struct ntb_dev *ndev = nt->ndev;
1023     struct pci_dev *pdev = ndev->pdev;
1024     resource_size_t size;
1025     u32 val;
1026     int rc = 0, i, spad;
1027 
1028     /* send the local info, in the opposite order of the way we read it */
1029 
1030     if (nt->use_msi) {
1031         rc = ntb_msi_setup_mws(ndev);
1032         if (rc) {
1033             dev_warn(&pdev->dev,
1034                  "Failed to register MSI memory window: %d\n",
1035                  rc);
1036             nt->use_msi = false;
1037         }
1038     }
1039 
1040     for (i = 0; i < nt->qp_count; i++)
1041         ntb_transport_setup_qp_msi(nt, i);
1042 
1043     for (i = 0; i < nt->mw_count; i++) {
1044         size = nt->mw_vec[i].phys_size;
1045 
1046         if (max_mw_size && size > max_mw_size)
1047             size = max_mw_size;
1048 
1049         spad = MW0_SZ_HIGH + (i * 2);
1050         ntb_peer_spad_write(ndev, PIDX, spad, upper_32_bits(size));
1051 
1052         spad = MW0_SZ_LOW + (i * 2);
1053         ntb_peer_spad_write(ndev, PIDX, spad, lower_32_bits(size));
1054     }
1055 
1056     ntb_peer_spad_write(ndev, PIDX, NUM_MWS, nt->mw_count);
1057 
1058     ntb_peer_spad_write(ndev, PIDX, NUM_QPS, nt->qp_count);
1059 
1060     ntb_peer_spad_write(ndev, PIDX, VERSION, NTB_TRANSPORT_VERSION);
1061 
1062     /* Query the remote side for its info */
1063     val = ntb_spad_read(ndev, VERSION);
1064     dev_dbg(&pdev->dev, "Remote version = %d\n", val);
1065     if (val != NTB_TRANSPORT_VERSION)
1066         goto out;
1067 
1068     val = ntb_spad_read(ndev, NUM_QPS);
1069     dev_dbg(&pdev->dev, "Remote max number of qps = %d\n", val);
1070     if (val != nt->qp_count)
1071         goto out;
1072 
1073     val = ntb_spad_read(ndev, NUM_MWS);
1074     dev_dbg(&pdev->dev, "Remote number of mws = %d\n", val);
1075     if (val != nt->mw_count)
1076         goto out;
1077 
1078     for (i = 0; i < nt->mw_count; i++) {
1079         u64 val64;
1080 
1081         val = ntb_spad_read(ndev, MW0_SZ_HIGH + (i * 2));
1082         val64 = (u64)val << 32;
1083 
1084         val = ntb_spad_read(ndev, MW0_SZ_LOW + (i * 2));
1085         val64 |= val;
1086 
1087         dev_dbg(&pdev->dev, "Remote MW%d size = %#llx\n", i, val64);
1088 
1089         rc = ntb_set_mw(nt, i, val64);
1090         if (rc)
1091             goto out1;
1092     }
1093 
1094     nt->link_is_up = true;
1095 
1096     for (i = 0; i < nt->qp_count; i++) {
1097         struct ntb_transport_qp *qp = &nt->qp_vec[i];
1098 
1099         ntb_transport_setup_qp_mw(nt, i);
1100         ntb_transport_setup_qp_peer_msi(nt, i);
1101 
1102         if (qp->client_ready)
1103             schedule_delayed_work(&qp->link_work, 0);
1104     }
1105 
1106     return;
1107 
1108 out1:
1109     for (i = 0; i < nt->mw_count; i++)
1110         ntb_free_mw(nt, i);
1111 
1112     /* if there's an actual failure, we should just bail */
1113     if (rc < 0)
1114         return;
1115 
1116 out:
1117     if (ntb_link_is_up(ndev, NULL, NULL) == 1)
1118         schedule_delayed_work(&nt->link_work,
1119                       msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
1120 }
1121 
1122 static void ntb_qp_link_work(struct work_struct *work)
1123 {
1124     struct ntb_transport_qp *qp = container_of(work,
1125                            struct ntb_transport_qp,
1126                            link_work.work);
1127     struct pci_dev *pdev = qp->ndev->pdev;
1128     struct ntb_transport_ctx *nt = qp->transport;
1129     int val;
1130 
1131     WARN_ON(!nt->link_is_up);
1132 
1133     val = ntb_spad_read(nt->ndev, QP_LINKS);
1134 
1135     ntb_peer_spad_write(nt->ndev, PIDX, QP_LINKS, val | BIT(qp->qp_num));
1136 
1137     /* query remote spad for qp ready bits */
1138     dev_dbg_ratelimited(&pdev->dev, "Remote QP link status = %x\n", val);
1139 
1140     /* See if the remote side is up */
1141     if (val & BIT(qp->qp_num)) {
1142         dev_info(&pdev->dev, "qp %d: Link Up\n", qp->qp_num);
1143         qp->link_is_up = true;
1144         qp->active = true;
1145 
1146         if (qp->event_handler)
1147             qp->event_handler(qp->cb_data, qp->link_is_up);
1148 
1149         if (qp->active)
1150             tasklet_schedule(&qp->rxc_db_work);
1151     } else if (nt->link_is_up)
1152         schedule_delayed_work(&qp->link_work,
1153                       msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
1154 }
1155 
1156 static int ntb_transport_init_queue(struct ntb_transport_ctx *nt,
1157                     unsigned int qp_num)
1158 {
1159     struct ntb_transport_qp *qp;
1160     phys_addr_t mw_base;
1161     resource_size_t mw_size;
1162     unsigned int num_qps_mw, tx_size;
1163     unsigned int mw_num, mw_count, qp_count;
1164     u64 qp_offset;
1165 
1166     mw_count = nt->mw_count;
1167     qp_count = nt->qp_count;
1168 
1169     mw_num = QP_TO_MW(nt, qp_num);
1170 
1171     qp = &nt->qp_vec[qp_num];
1172     qp->qp_num = qp_num;
1173     qp->transport = nt;
1174     qp->ndev = nt->ndev;
1175     qp->client_ready = false;
1176     qp->event_handler = NULL;
1177     ntb_qp_link_down_reset(qp);
1178 
1179     if (mw_num < qp_count % mw_count)
1180         num_qps_mw = qp_count / mw_count + 1;
1181     else
1182         num_qps_mw = qp_count / mw_count;
1183 
1184     mw_base = nt->mw_vec[mw_num].phys_addr;
1185     mw_size = nt->mw_vec[mw_num].phys_size;
1186 
1187     if (max_mw_size && mw_size > max_mw_size)
1188         mw_size = max_mw_size;
1189 
1190     tx_size = (unsigned int)mw_size / num_qps_mw;
1191     qp_offset = tx_size * (qp_num / mw_count);
1192 
1193     qp->tx_mw_size = tx_size;
1194     qp->tx_mw = nt->mw_vec[mw_num].vbase + qp_offset;
1195     if (!qp->tx_mw)
1196         return -EINVAL;
1197 
1198     qp->tx_mw_phys = mw_base + qp_offset;
1199     if (!qp->tx_mw_phys)
1200         return -EINVAL;
1201 
1202     tx_size -= sizeof(struct ntb_rx_info);
1203     qp->rx_info = qp->tx_mw + tx_size;
1204 
1205     /* Due to housekeeping, there must be atleast 2 buffs */
1206     qp->tx_max_frame = min(transport_mtu, tx_size / 2);
1207     qp->tx_max_entry = tx_size / qp->tx_max_frame;
1208 
1209     if (nt->debugfs_node_dir) {
1210         char debugfs_name[4];
1211 
1212         snprintf(debugfs_name, 4, "qp%d", qp_num);
1213         qp->debugfs_dir = debugfs_create_dir(debugfs_name,
1214                              nt->debugfs_node_dir);
1215 
1216         qp->debugfs_stats = debugfs_create_file("stats", S_IRUSR,
1217                             qp->debugfs_dir, qp,
1218                             &ntb_qp_debugfs_stats);
1219     } else {
1220         qp->debugfs_dir = NULL;
1221         qp->debugfs_stats = NULL;
1222     }
1223 
1224     INIT_DELAYED_WORK(&qp->link_work, ntb_qp_link_work);
1225     INIT_WORK(&qp->link_cleanup, ntb_qp_link_cleanup_work);
1226 
1227     spin_lock_init(&qp->ntb_rx_q_lock);
1228     spin_lock_init(&qp->ntb_tx_free_q_lock);
1229 
1230     INIT_LIST_HEAD(&qp->rx_post_q);
1231     INIT_LIST_HEAD(&qp->rx_pend_q);
1232     INIT_LIST_HEAD(&qp->rx_free_q);
1233     INIT_LIST_HEAD(&qp->tx_free_q);
1234 
1235     tasklet_init(&qp->rxc_db_work, ntb_transport_rxc_db,
1236              (unsigned long)qp);
1237 
1238     return 0;
1239 }
1240 
1241 static int ntb_transport_probe(struct ntb_client *self, struct ntb_dev *ndev)
1242 {
1243     struct ntb_transport_ctx *nt;
1244     struct ntb_transport_mw *mw;
1245     unsigned int mw_count, qp_count, spad_count, max_mw_count_for_spads;
1246     u64 qp_bitmap;
1247     int node;
1248     int rc, i;
1249 
1250     mw_count = ntb_peer_mw_count(ndev);
1251 
1252     if (!ndev->ops->mw_set_trans) {
1253         dev_err(&ndev->dev, "Inbound MW based NTB API is required\n");
1254         return -EINVAL;
1255     }
1256 
1257     if (ntb_db_is_unsafe(ndev))
1258         dev_dbg(&ndev->dev,
1259             "doorbell is unsafe, proceed anyway...\n");
1260     if (ntb_spad_is_unsafe(ndev))
1261         dev_dbg(&ndev->dev,
1262             "scratchpad is unsafe, proceed anyway...\n");
1263 
1264     if (ntb_peer_port_count(ndev) != NTB_DEF_PEER_CNT)
1265         dev_warn(&ndev->dev, "Multi-port NTB devices unsupported\n");
1266 
1267     node = dev_to_node(&ndev->dev);
1268 
1269     nt = kzalloc_node(sizeof(*nt), GFP_KERNEL, node);
1270     if (!nt)
1271         return -ENOMEM;
1272 
1273     nt->ndev = ndev;
1274 
1275     /*
1276      * If we are using MSI, and have at least one extra memory window,
1277      * we will reserve the last MW for the MSI window.
1278      */
1279     if (use_msi && mw_count > 1) {
1280         rc = ntb_msi_init(ndev, ntb_transport_msi_desc_changed);
1281         if (!rc) {
1282             mw_count -= 1;
1283             nt->use_msi = true;
1284         }
1285     }
1286 
1287     spad_count = ntb_spad_count(ndev);
1288 
1289     /* Limit the MW's based on the availability of scratchpads */
1290 
1291     if (spad_count < NTB_TRANSPORT_MIN_SPADS) {
1292         nt->mw_count = 0;
1293         rc = -EINVAL;
1294         goto err;
1295     }
1296 
1297     max_mw_count_for_spads = (spad_count - MW0_SZ_HIGH) / 2;
1298     nt->mw_count = min(mw_count, max_mw_count_for_spads);
1299 
1300     nt->msi_spad_offset = nt->mw_count * 2 + MW0_SZ_HIGH;
1301 
1302     nt->mw_vec = kcalloc_node(mw_count, sizeof(*nt->mw_vec),
1303                   GFP_KERNEL, node);
1304     if (!nt->mw_vec) {
1305         rc = -ENOMEM;
1306         goto err;
1307     }
1308 
1309     for (i = 0; i < mw_count; i++) {
1310         mw = &nt->mw_vec[i];
1311 
1312         rc = ntb_peer_mw_get_addr(ndev, i, &mw->phys_addr,
1313                       &mw->phys_size);
1314         if (rc)
1315             goto err1;
1316 
1317         mw->vbase = ioremap_wc(mw->phys_addr, mw->phys_size);
1318         if (!mw->vbase) {
1319             rc = -ENOMEM;
1320             goto err1;
1321         }
1322 
1323         mw->buff_size = 0;
1324         mw->xlat_size = 0;
1325         mw->virt_addr = NULL;
1326         mw->dma_addr = 0;
1327     }
1328 
1329     qp_bitmap = ntb_db_valid_mask(ndev);
1330 
1331     qp_count = ilog2(qp_bitmap);
1332     if (nt->use_msi) {
1333         qp_count -= 1;
1334         nt->msi_db_mask = 1 << qp_count;
1335         ntb_db_clear_mask(ndev, nt->msi_db_mask);
1336     }
1337 
1338     if (max_num_clients && max_num_clients < qp_count)
1339         qp_count = max_num_clients;
1340     else if (nt->mw_count < qp_count)
1341         qp_count = nt->mw_count;
1342 
1343     qp_bitmap &= BIT_ULL(qp_count) - 1;
1344 
1345     nt->qp_count = qp_count;
1346     nt->qp_bitmap = qp_bitmap;
1347     nt->qp_bitmap_free = qp_bitmap;
1348 
1349     nt->qp_vec = kcalloc_node(qp_count, sizeof(*nt->qp_vec),
1350                   GFP_KERNEL, node);
1351     if (!nt->qp_vec) {
1352         rc = -ENOMEM;
1353         goto err1;
1354     }
1355 
1356     if (nt_debugfs_dir) {
1357         nt->debugfs_node_dir =
1358             debugfs_create_dir(pci_name(ndev->pdev),
1359                        nt_debugfs_dir);
1360     }
1361 
1362     for (i = 0; i < qp_count; i++) {
1363         rc = ntb_transport_init_queue(nt, i);
1364         if (rc)
1365             goto err2;
1366     }
1367 
1368     INIT_DELAYED_WORK(&nt->link_work, ntb_transport_link_work);
1369     INIT_WORK(&nt->link_cleanup, ntb_transport_link_cleanup_work);
1370 
1371     rc = ntb_set_ctx(ndev, nt, &ntb_transport_ops);
1372     if (rc)
1373         goto err2;
1374 
1375     INIT_LIST_HEAD(&nt->client_devs);
1376     rc = ntb_bus_init(nt);
1377     if (rc)
1378         goto err3;
1379 
1380     nt->link_is_up = false;
1381     ntb_link_enable(ndev, NTB_SPEED_AUTO, NTB_WIDTH_AUTO);
1382     ntb_link_event(ndev);
1383 
1384     return 0;
1385 
1386 err3:
1387     ntb_clear_ctx(ndev);
1388 err2:
1389     kfree(nt->qp_vec);
1390 err1:
1391     while (i--) {
1392         mw = &nt->mw_vec[i];
1393         iounmap(mw->vbase);
1394     }
1395     kfree(nt->mw_vec);
1396 err:
1397     kfree(nt);
1398     return rc;
1399 }
1400 
1401 static void ntb_transport_free(struct ntb_client *self, struct ntb_dev *ndev)
1402 {
1403     struct ntb_transport_ctx *nt = ndev->ctx;
1404     struct ntb_transport_qp *qp;
1405     u64 qp_bitmap_alloc;
1406     int i;
1407 
1408     ntb_transport_link_cleanup(nt);
1409     cancel_work_sync(&nt->link_cleanup);
1410     cancel_delayed_work_sync(&nt->link_work);
1411 
1412     qp_bitmap_alloc = nt->qp_bitmap & ~nt->qp_bitmap_free;
1413 
1414     /* verify that all the qp's are freed */
1415     for (i = 0; i < nt->qp_count; i++) {
1416         qp = &nt->qp_vec[i];
1417         if (qp_bitmap_alloc & BIT_ULL(i))
1418             ntb_transport_free_queue(qp);
1419         debugfs_remove_recursive(qp->debugfs_dir);
1420     }
1421 
1422     ntb_link_disable(ndev);
1423     ntb_clear_ctx(ndev);
1424 
1425     ntb_bus_remove(nt);
1426 
1427     for (i = nt->mw_count; i--; ) {
1428         ntb_free_mw(nt, i);
1429         iounmap(nt->mw_vec[i].vbase);
1430     }
1431 
1432     kfree(nt->qp_vec);
1433     kfree(nt->mw_vec);
1434     kfree(nt);
1435 }
1436 
1437 static void ntb_complete_rxc(struct ntb_transport_qp *qp)
1438 {
1439     struct ntb_queue_entry *entry;
1440     void *cb_data;
1441     unsigned int len;
1442     unsigned long irqflags;
1443 
1444     spin_lock_irqsave(&qp->ntb_rx_q_lock, irqflags);
1445 
1446     while (!list_empty(&qp->rx_post_q)) {
1447         entry = list_first_entry(&qp->rx_post_q,
1448                      struct ntb_queue_entry, entry);
1449         if (!(entry->flags & DESC_DONE_FLAG))
1450             break;
1451 
1452         entry->rx_hdr->flags = 0;
1453         iowrite32(entry->rx_index, &qp->rx_info->entry);
1454 
1455         cb_data = entry->cb_data;
1456         len = entry->len;
1457 
1458         list_move_tail(&entry->entry, &qp->rx_free_q);
1459 
1460         spin_unlock_irqrestore(&qp->ntb_rx_q_lock, irqflags);
1461 
1462         if (qp->rx_handler && qp->client_ready)
1463             qp->rx_handler(qp, qp->cb_data, cb_data, len);
1464 
1465         spin_lock_irqsave(&qp->ntb_rx_q_lock, irqflags);
1466     }
1467 
1468     spin_unlock_irqrestore(&qp->ntb_rx_q_lock, irqflags);
1469 }
1470 
1471 static void ntb_rx_copy_callback(void *data,
1472                  const struct dmaengine_result *res)
1473 {
1474     struct ntb_queue_entry *entry = data;
1475 
1476     /* we need to check DMA results if we are using DMA */
1477     if (res) {
1478         enum dmaengine_tx_result dma_err = res->result;
1479 
1480         switch (dma_err) {
1481         case DMA_TRANS_READ_FAILED:
1482         case DMA_TRANS_WRITE_FAILED:
1483             entry->errors++;
1484             fallthrough;
1485         case DMA_TRANS_ABORTED:
1486         {
1487             struct ntb_transport_qp *qp = entry->qp;
1488             void *offset = qp->rx_buff + qp->rx_max_frame *
1489                     qp->rx_index;
1490 
1491             ntb_memcpy_rx(entry, offset);
1492             qp->rx_memcpy++;
1493             return;
1494         }
1495 
1496         case DMA_TRANS_NOERROR:
1497         default:
1498             break;
1499         }
1500     }
1501 
1502     entry->flags |= DESC_DONE_FLAG;
1503 
1504     ntb_complete_rxc(entry->qp);
1505 }
1506 
1507 static void ntb_memcpy_rx(struct ntb_queue_entry *entry, void *offset)
1508 {
1509     void *buf = entry->buf;
1510     size_t len = entry->len;
1511 
1512     memcpy(buf, offset, len);
1513 
1514     /* Ensure that the data is fully copied out before clearing the flag */
1515     wmb();
1516 
1517     ntb_rx_copy_callback(entry, NULL);
1518 }
1519 
1520 static int ntb_async_rx_submit(struct ntb_queue_entry *entry, void *offset)
1521 {
1522     struct dma_async_tx_descriptor *txd;
1523     struct ntb_transport_qp *qp = entry->qp;
1524     struct dma_chan *chan = qp->rx_dma_chan;
1525     struct dma_device *device;
1526     size_t pay_off, buff_off, len;
1527     struct dmaengine_unmap_data *unmap;
1528     dma_cookie_t cookie;
1529     void *buf = entry->buf;
1530 
1531     len = entry->len;
1532     device = chan->device;
1533     pay_off = (size_t)offset & ~PAGE_MASK;
1534     buff_off = (size_t)buf & ~PAGE_MASK;
1535 
1536     if (!is_dma_copy_aligned(device, pay_off, buff_off, len))
1537         goto err;
1538 
1539     unmap = dmaengine_get_unmap_data(device->dev, 2, GFP_NOWAIT);
1540     if (!unmap)
1541         goto err;
1542 
1543     unmap->len = len;
1544     unmap->addr[0] = dma_map_page(device->dev, virt_to_page(offset),
1545                       pay_off, len, DMA_TO_DEVICE);
1546     if (dma_mapping_error(device->dev, unmap->addr[0]))
1547         goto err_get_unmap;
1548 
1549     unmap->to_cnt = 1;
1550 
1551     unmap->addr[1] = dma_map_page(device->dev, virt_to_page(buf),
1552                       buff_off, len, DMA_FROM_DEVICE);
1553     if (dma_mapping_error(device->dev, unmap->addr[1]))
1554         goto err_get_unmap;
1555 
1556     unmap->from_cnt = 1;
1557 
1558     txd = device->device_prep_dma_memcpy(chan, unmap->addr[1],
1559                          unmap->addr[0], len,
1560                          DMA_PREP_INTERRUPT);
1561     if (!txd)
1562         goto err_get_unmap;
1563 
1564     txd->callback_result = ntb_rx_copy_callback;
1565     txd->callback_param = entry;
1566     dma_set_unmap(txd, unmap);
1567 
1568     cookie = dmaengine_submit(txd);
1569     if (dma_submit_error(cookie))
1570         goto err_set_unmap;
1571 
1572     dmaengine_unmap_put(unmap);
1573 
1574     qp->last_cookie = cookie;
1575 
1576     qp->rx_async++;
1577 
1578     return 0;
1579 
1580 err_set_unmap:
1581     dmaengine_unmap_put(unmap);
1582 err_get_unmap:
1583     dmaengine_unmap_put(unmap);
1584 err:
1585     return -ENXIO;
1586 }
1587 
1588 static void ntb_async_rx(struct ntb_queue_entry *entry, void *offset)
1589 {
1590     struct ntb_transport_qp *qp = entry->qp;
1591     struct dma_chan *chan = qp->rx_dma_chan;
1592     int res;
1593 
1594     if (!chan)
1595         goto err;
1596 
1597     if (entry->len < copy_bytes)
1598         goto err;
1599 
1600     res = ntb_async_rx_submit(entry, offset);
1601     if (res < 0)
1602         goto err;
1603 
1604     if (!entry->retries)
1605         qp->rx_async++;
1606 
1607     return;
1608 
1609 err:
1610     ntb_memcpy_rx(entry, offset);
1611     qp->rx_memcpy++;
1612 }
1613 
1614 static int ntb_process_rxc(struct ntb_transport_qp *qp)
1615 {
1616     struct ntb_payload_header *hdr;
1617     struct ntb_queue_entry *entry;
1618     void *offset;
1619 
1620     offset = qp->rx_buff + qp->rx_max_frame * qp->rx_index;
1621     hdr = offset + qp->rx_max_frame - sizeof(struct ntb_payload_header);
1622 
1623     dev_dbg(&qp->ndev->pdev->dev, "qp %d: RX ver %u len %d flags %x\n",
1624         qp->qp_num, hdr->ver, hdr->len, hdr->flags);
1625 
1626     if (!(hdr->flags & DESC_DONE_FLAG)) {
1627         dev_dbg(&qp->ndev->pdev->dev, "done flag not set\n");
1628         qp->rx_ring_empty++;
1629         return -EAGAIN;
1630     }
1631 
1632     if (hdr->flags & LINK_DOWN_FLAG) {
1633         dev_dbg(&qp->ndev->pdev->dev, "link down flag set\n");
1634         ntb_qp_link_down(qp);
1635         hdr->flags = 0;
1636         return -EAGAIN;
1637     }
1638 
1639     if (hdr->ver != (u32)qp->rx_pkts) {
1640         dev_dbg(&qp->ndev->pdev->dev,
1641             "version mismatch, expected %llu - got %u\n",
1642             qp->rx_pkts, hdr->ver);
1643         qp->rx_err_ver++;
1644         return -EIO;
1645     }
1646 
1647     entry = ntb_list_mv(&qp->ntb_rx_q_lock, &qp->rx_pend_q, &qp->rx_post_q);
1648     if (!entry) {
1649         dev_dbg(&qp->ndev->pdev->dev, "no receive buffer\n");
1650         qp->rx_err_no_buf++;
1651         return -EAGAIN;
1652     }
1653 
1654     entry->rx_hdr = hdr;
1655     entry->rx_index = qp->rx_index;
1656 
1657     if (hdr->len > entry->len) {
1658         dev_dbg(&qp->ndev->pdev->dev,
1659             "receive buffer overflow! Wanted %d got %d\n",
1660             hdr->len, entry->len);
1661         qp->rx_err_oflow++;
1662 
1663         entry->len = -EIO;
1664         entry->flags |= DESC_DONE_FLAG;
1665 
1666         ntb_complete_rxc(qp);
1667     } else {
1668         dev_dbg(&qp->ndev->pdev->dev,
1669             "RX OK index %u ver %u size %d into buf size %d\n",
1670             qp->rx_index, hdr->ver, hdr->len, entry->len);
1671 
1672         qp->rx_bytes += hdr->len;
1673         qp->rx_pkts++;
1674 
1675         entry->len = hdr->len;
1676 
1677         ntb_async_rx(entry, offset);
1678     }
1679 
1680     qp->rx_index++;
1681     qp->rx_index %= qp->rx_max_entry;
1682 
1683     return 0;
1684 }
1685 
1686 static void ntb_transport_rxc_db(unsigned long data)
1687 {
1688     struct ntb_transport_qp *qp = (void *)data;
1689     int rc, i;
1690 
1691     dev_dbg(&qp->ndev->pdev->dev, "%s: doorbell %d received\n",
1692         __func__, qp->qp_num);
1693 
1694     /* Limit the number of packets processed in a single interrupt to
1695      * provide fairness to others
1696      */
1697     for (i = 0; i < qp->rx_max_entry; i++) {
1698         rc = ntb_process_rxc(qp);
1699         if (rc)
1700             break;
1701     }
1702 
1703     if (i && qp->rx_dma_chan)
1704         dma_async_issue_pending(qp->rx_dma_chan);
1705 
1706     if (i == qp->rx_max_entry) {
1707         /* there is more work to do */
1708         if (qp->active)
1709             tasklet_schedule(&qp->rxc_db_work);
1710     } else if (ntb_db_read(qp->ndev) & BIT_ULL(qp->qp_num)) {
1711         /* the doorbell bit is set: clear it */
1712         ntb_db_clear(qp->ndev, BIT_ULL(qp->qp_num));
1713         /* ntb_db_read ensures ntb_db_clear write is committed */
1714         ntb_db_read(qp->ndev);
1715 
1716         /* an interrupt may have arrived between finishing
1717          * ntb_process_rxc and clearing the doorbell bit:
1718          * there might be some more work to do.
1719          */
1720         if (qp->active)
1721             tasklet_schedule(&qp->rxc_db_work);
1722     }
1723 }
1724 
1725 static void ntb_tx_copy_callback(void *data,
1726                  const struct dmaengine_result *res)
1727 {
1728     struct ntb_queue_entry *entry = data;
1729     struct ntb_transport_qp *qp = entry->qp;
1730     struct ntb_payload_header __iomem *hdr = entry->tx_hdr;
1731 
1732     /* we need to check DMA results if we are using DMA */
1733     if (res) {
1734         enum dmaengine_tx_result dma_err = res->result;
1735 
1736         switch (dma_err) {
1737         case DMA_TRANS_READ_FAILED:
1738         case DMA_TRANS_WRITE_FAILED:
1739             entry->errors++;
1740             fallthrough;
1741         case DMA_TRANS_ABORTED:
1742         {
1743             void __iomem *offset =
1744                 qp->tx_mw + qp->tx_max_frame *
1745                 entry->tx_index;
1746 
1747             /* resubmit via CPU */
1748             ntb_memcpy_tx(entry, offset);
1749             qp->tx_memcpy++;
1750             return;
1751         }
1752 
1753         case DMA_TRANS_NOERROR:
1754         default:
1755             break;
1756         }
1757     }
1758 
1759     iowrite32(entry->flags | DESC_DONE_FLAG, &hdr->flags);
1760 
1761     if (qp->use_msi)
1762         ntb_msi_peer_trigger(qp->ndev, PIDX, &qp->peer_msi_desc);
1763     else
1764         ntb_peer_db_set(qp->ndev, BIT_ULL(qp->qp_num));
1765 
1766     /* The entry length can only be zero if the packet is intended to be a
1767      * "link down" or similar.  Since no payload is being sent in these
1768      * cases, there is nothing to add to the completion queue.
1769      */
1770     if (entry->len > 0) {
1771         qp->tx_bytes += entry->len;
1772 
1773         if (qp->tx_handler)
1774             qp->tx_handler(qp, qp->cb_data, entry->cb_data,
1775                        entry->len);
1776     }
1777 
1778     ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry, &qp->tx_free_q);
1779 }
1780 
1781 static void ntb_memcpy_tx(struct ntb_queue_entry *entry, void __iomem *offset)
1782 {
1783 #ifdef ARCH_HAS_NOCACHE_UACCESS
1784     /*
1785      * Using non-temporal mov to improve performance on non-cached
1786      * writes, even though we aren't actually copying from user space.
1787      */
1788     __copy_from_user_inatomic_nocache(offset, entry->buf, entry->len);
1789 #else
1790     memcpy_toio(offset, entry->buf, entry->len);
1791 #endif
1792 
1793     /* Ensure that the data is fully copied out before setting the flags */
1794     wmb();
1795 
1796     ntb_tx_copy_callback(entry, NULL);
1797 }
1798 
1799 static int ntb_async_tx_submit(struct ntb_transport_qp *qp,
1800                    struct ntb_queue_entry *entry)
1801 {
1802     struct dma_async_tx_descriptor *txd;
1803     struct dma_chan *chan = qp->tx_dma_chan;
1804     struct dma_device *device;
1805     size_t len = entry->len;
1806     void *buf = entry->buf;
1807     size_t dest_off, buff_off;
1808     struct dmaengine_unmap_data *unmap;
1809     dma_addr_t dest;
1810     dma_cookie_t cookie;
1811 
1812     device = chan->device;
1813     dest = qp->tx_mw_dma_addr + qp->tx_max_frame * entry->tx_index;
1814     buff_off = (size_t)buf & ~PAGE_MASK;
1815     dest_off = (size_t)dest & ~PAGE_MASK;
1816 
1817     if (!is_dma_copy_aligned(device, buff_off, dest_off, len))
1818         goto err;
1819 
1820     unmap = dmaengine_get_unmap_data(device->dev, 1, GFP_NOWAIT);
1821     if (!unmap)
1822         goto err;
1823 
1824     unmap->len = len;
1825     unmap->addr[0] = dma_map_page(device->dev, virt_to_page(buf),
1826                       buff_off, len, DMA_TO_DEVICE);
1827     if (dma_mapping_error(device->dev, unmap->addr[0]))
1828         goto err_get_unmap;
1829 
1830     unmap->to_cnt = 1;
1831 
1832     txd = device->device_prep_dma_memcpy(chan, dest, unmap->addr[0], len,
1833                          DMA_PREP_INTERRUPT);
1834     if (!txd)
1835         goto err_get_unmap;
1836 
1837     txd->callback_result = ntb_tx_copy_callback;
1838     txd->callback_param = entry;
1839     dma_set_unmap(txd, unmap);
1840 
1841     cookie = dmaengine_submit(txd);
1842     if (dma_submit_error(cookie))
1843         goto err_set_unmap;
1844 
1845     dmaengine_unmap_put(unmap);
1846 
1847     dma_async_issue_pending(chan);
1848 
1849     return 0;
1850 err_set_unmap:
1851     dmaengine_unmap_put(unmap);
1852 err_get_unmap:
1853     dmaengine_unmap_put(unmap);
1854 err:
1855     return -ENXIO;
1856 }
1857 
1858 static void ntb_async_tx(struct ntb_transport_qp *qp,
1859              struct ntb_queue_entry *entry)
1860 {
1861     struct ntb_payload_header __iomem *hdr;
1862     struct dma_chan *chan = qp->tx_dma_chan;
1863     void __iomem *offset;
1864     int res;
1865 
1866     entry->tx_index = qp->tx_index;
1867     offset = qp->tx_mw + qp->tx_max_frame * entry->tx_index;
1868     hdr = offset + qp->tx_max_frame - sizeof(struct ntb_payload_header);
1869     entry->tx_hdr = hdr;
1870 
1871     iowrite32(entry->len, &hdr->len);
1872     iowrite32((u32)qp->tx_pkts, &hdr->ver);
1873 
1874     if (!chan)
1875         goto err;
1876 
1877     if (entry->len < copy_bytes)
1878         goto err;
1879 
1880     res = ntb_async_tx_submit(qp, entry);
1881     if (res < 0)
1882         goto err;
1883 
1884     if (!entry->retries)
1885         qp->tx_async++;
1886 
1887     return;
1888 
1889 err:
1890     ntb_memcpy_tx(entry, offset);
1891     qp->tx_memcpy++;
1892 }
1893 
1894 static int ntb_process_tx(struct ntb_transport_qp *qp,
1895               struct ntb_queue_entry *entry)
1896 {
1897     if (qp->tx_index == qp->remote_rx_info->entry) {
1898         qp->tx_ring_full++;
1899         return -EAGAIN;
1900     }
1901 
1902     if (entry->len > qp->tx_max_frame - sizeof(struct ntb_payload_header)) {
1903         if (qp->tx_handler)
1904             qp->tx_handler(qp, qp->cb_data, NULL, -EIO);
1905 
1906         ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
1907                  &qp->tx_free_q);
1908         return 0;
1909     }
1910 
1911     ntb_async_tx(qp, entry);
1912 
1913     qp->tx_index++;
1914     qp->tx_index %= qp->tx_max_entry;
1915 
1916     qp->tx_pkts++;
1917 
1918     return 0;
1919 }
1920 
1921 static void ntb_send_link_down(struct ntb_transport_qp *qp)
1922 {
1923     struct pci_dev *pdev = qp->ndev->pdev;
1924     struct ntb_queue_entry *entry;
1925     int i, rc;
1926 
1927     if (!qp->link_is_up)
1928         return;
1929 
1930     dev_info(&pdev->dev, "qp %d: Send Link Down\n", qp->qp_num);
1931 
1932     for (i = 0; i < NTB_LINK_DOWN_TIMEOUT; i++) {
1933         entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
1934         if (entry)
1935             break;
1936         msleep(100);
1937     }
1938 
1939     if (!entry)
1940         return;
1941 
1942     entry->cb_data = NULL;
1943     entry->buf = NULL;
1944     entry->len = 0;
1945     entry->flags = LINK_DOWN_FLAG;
1946 
1947     rc = ntb_process_tx(qp, entry);
1948     if (rc)
1949         dev_err(&pdev->dev, "ntb: QP%d unable to send linkdown msg\n",
1950             qp->qp_num);
1951 
1952     ntb_qp_link_down_reset(qp);
1953 }
1954 
1955 static bool ntb_dma_filter_fn(struct dma_chan *chan, void *node)
1956 {
1957     return dev_to_node(&chan->dev->device) == (int)(unsigned long)node;
1958 }
1959 
1960 /**
1961  * ntb_transport_create_queue - Create a new NTB transport layer queue
1962  * @rx_handler: receive callback function
1963  * @tx_handler: transmit callback function
1964  * @event_handler: event callback function
1965  *
1966  * Create a new NTB transport layer queue and provide the queue with a callback
1967  * routine for both transmit and receive.  The receive callback routine will be
1968  * used to pass up data when the transport has received it on the queue.   The
1969  * transmit callback routine will be called when the transport has completed the
1970  * transmission of the data on the queue and the data is ready to be freed.
1971  *
1972  * RETURNS: pointer to newly created ntb_queue, NULL on error.
1973  */
1974 struct ntb_transport_qp *
1975 ntb_transport_create_queue(void *data, struct device *client_dev,
1976                const struct ntb_queue_handlers *handlers)
1977 {
1978     struct ntb_dev *ndev;
1979     struct pci_dev *pdev;
1980     struct ntb_transport_ctx *nt;
1981     struct ntb_queue_entry *entry;
1982     struct ntb_transport_qp *qp;
1983     u64 qp_bit;
1984     unsigned int free_queue;
1985     dma_cap_mask_t dma_mask;
1986     int node;
1987     int i;
1988 
1989     ndev = dev_ntb(client_dev->parent);
1990     pdev = ndev->pdev;
1991     nt = ndev->ctx;
1992 
1993     node = dev_to_node(&ndev->dev);
1994 
1995     free_queue = ffs(nt->qp_bitmap_free);
1996     if (!free_queue)
1997         goto err;
1998 
1999     /* decrement free_queue to make it zero based */
2000     free_queue--;
2001 
2002     qp = &nt->qp_vec[free_queue];
2003     qp_bit = BIT_ULL(qp->qp_num);
2004 
2005     nt->qp_bitmap_free &= ~qp_bit;
2006 
2007     qp->cb_data = data;
2008     qp->rx_handler = handlers->rx_handler;
2009     qp->tx_handler = handlers->tx_handler;
2010     qp->event_handler = handlers->event_handler;
2011 
2012     dma_cap_zero(dma_mask);
2013     dma_cap_set(DMA_MEMCPY, dma_mask);
2014 
2015     if (use_dma) {
2016         qp->tx_dma_chan =
2017             dma_request_channel(dma_mask, ntb_dma_filter_fn,
2018                         (void *)(unsigned long)node);
2019         if (!qp->tx_dma_chan)
2020             dev_info(&pdev->dev, "Unable to allocate TX DMA channel\n");
2021 
2022         qp->rx_dma_chan =
2023             dma_request_channel(dma_mask, ntb_dma_filter_fn,
2024                         (void *)(unsigned long)node);
2025         if (!qp->rx_dma_chan)
2026             dev_info(&pdev->dev, "Unable to allocate RX DMA channel\n");
2027     } else {
2028         qp->tx_dma_chan = NULL;
2029         qp->rx_dma_chan = NULL;
2030     }
2031 
2032     qp->tx_mw_dma_addr = 0;
2033     if (qp->tx_dma_chan) {
2034         qp->tx_mw_dma_addr =
2035             dma_map_resource(qp->tx_dma_chan->device->dev,
2036                      qp->tx_mw_phys, qp->tx_mw_size,
2037                      DMA_FROM_DEVICE, 0);
2038         if (dma_mapping_error(qp->tx_dma_chan->device->dev,
2039                       qp->tx_mw_dma_addr)) {
2040             qp->tx_mw_dma_addr = 0;
2041             goto err1;
2042         }
2043     }
2044 
2045     dev_dbg(&pdev->dev, "Using %s memcpy for TX\n",
2046         qp->tx_dma_chan ? "DMA" : "CPU");
2047 
2048     dev_dbg(&pdev->dev, "Using %s memcpy for RX\n",
2049         qp->rx_dma_chan ? "DMA" : "CPU");
2050 
2051     for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) {
2052         entry = kzalloc_node(sizeof(*entry), GFP_KERNEL, node);
2053         if (!entry)
2054             goto err1;
2055 
2056         entry->qp = qp;
2057         ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry,
2058                  &qp->rx_free_q);
2059     }
2060     qp->rx_alloc_entry = NTB_QP_DEF_NUM_ENTRIES;
2061 
2062     for (i = 0; i < qp->tx_max_entry; i++) {
2063         entry = kzalloc_node(sizeof(*entry), GFP_KERNEL, node);
2064         if (!entry)
2065             goto err2;
2066 
2067         entry->qp = qp;
2068         ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
2069                  &qp->tx_free_q);
2070     }
2071 
2072     ntb_db_clear(qp->ndev, qp_bit);
2073     ntb_db_clear_mask(qp->ndev, qp_bit);
2074 
2075     dev_info(&pdev->dev, "NTB Transport QP %d created\n", qp->qp_num);
2076 
2077     return qp;
2078 
2079 err2:
2080     while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
2081         kfree(entry);
2082 err1:
2083     qp->rx_alloc_entry = 0;
2084     while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q)))
2085         kfree(entry);
2086     if (qp->tx_mw_dma_addr)
2087         dma_unmap_resource(qp->tx_dma_chan->device->dev,
2088                    qp->tx_mw_dma_addr, qp->tx_mw_size,
2089                    DMA_FROM_DEVICE, 0);
2090     if (qp->tx_dma_chan)
2091         dma_release_channel(qp->tx_dma_chan);
2092     if (qp->rx_dma_chan)
2093         dma_release_channel(qp->rx_dma_chan);
2094     nt->qp_bitmap_free |= qp_bit;
2095 err:
2096     return NULL;
2097 }
2098 EXPORT_SYMBOL_GPL(ntb_transport_create_queue);
2099 
2100 /**
2101  * ntb_transport_free_queue - Frees NTB transport queue
2102  * @qp: NTB queue to be freed
2103  *
2104  * Frees NTB transport queue
2105  */
2106 void ntb_transport_free_queue(struct ntb_transport_qp *qp)
2107 {
2108     struct pci_dev *pdev;
2109     struct ntb_queue_entry *entry;
2110     u64 qp_bit;
2111 
2112     if (!qp)
2113         return;
2114 
2115     pdev = qp->ndev->pdev;
2116 
2117     qp->active = false;
2118 
2119     if (qp->tx_dma_chan) {
2120         struct dma_chan *chan = qp->tx_dma_chan;
2121         /* Putting the dma_chan to NULL will force any new traffic to be
2122          * processed by the CPU instead of the DAM engine
2123          */
2124         qp->tx_dma_chan = NULL;
2125 
2126         /* Try to be nice and wait for any queued DMA engine
2127          * transactions to process before smashing it with a rock
2128          */
2129         dma_sync_wait(chan, qp->last_cookie);
2130         dmaengine_terminate_all(chan);
2131 
2132         dma_unmap_resource(chan->device->dev,
2133                    qp->tx_mw_dma_addr, qp->tx_mw_size,
2134                    DMA_FROM_DEVICE, 0);
2135 
2136         dma_release_channel(chan);
2137     }
2138 
2139     if (qp->rx_dma_chan) {
2140         struct dma_chan *chan = qp->rx_dma_chan;
2141         /* Putting the dma_chan to NULL will force any new traffic to be
2142          * processed by the CPU instead of the DAM engine
2143          */
2144         qp->rx_dma_chan = NULL;
2145 
2146         /* Try to be nice and wait for any queued DMA engine
2147          * transactions to process before smashing it with a rock
2148          */
2149         dma_sync_wait(chan, qp->last_cookie);
2150         dmaengine_terminate_all(chan);
2151         dma_release_channel(chan);
2152     }
2153 
2154     qp_bit = BIT_ULL(qp->qp_num);
2155 
2156     ntb_db_set_mask(qp->ndev, qp_bit);
2157     tasklet_kill(&qp->rxc_db_work);
2158 
2159     cancel_delayed_work_sync(&qp->link_work);
2160 
2161     qp->cb_data = NULL;
2162     qp->rx_handler = NULL;
2163     qp->tx_handler = NULL;
2164     qp->event_handler = NULL;
2165 
2166     while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q)))
2167         kfree(entry);
2168 
2169     while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_pend_q))) {
2170         dev_warn(&pdev->dev, "Freeing item from non-empty rx_pend_q\n");
2171         kfree(entry);
2172     }
2173 
2174     while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_post_q))) {
2175         dev_warn(&pdev->dev, "Freeing item from non-empty rx_post_q\n");
2176         kfree(entry);
2177     }
2178 
2179     while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
2180         kfree(entry);
2181 
2182     qp->transport->qp_bitmap_free |= qp_bit;
2183 
2184     dev_info(&pdev->dev, "NTB Transport QP %d freed\n", qp->qp_num);
2185 }
2186 EXPORT_SYMBOL_GPL(ntb_transport_free_queue);
2187 
2188 /**
2189  * ntb_transport_rx_remove - Dequeues enqueued rx packet
2190  * @qp: NTB queue to be freed
2191  * @len: pointer to variable to write enqueued buffers length
2192  *
2193  * Dequeues unused buffers from receive queue.  Should only be used during
2194  * shutdown of qp.
2195  *
2196  * RETURNS: NULL error value on error, or void* for success.
2197  */
2198 void *ntb_transport_rx_remove(struct ntb_transport_qp *qp, unsigned int *len)
2199 {
2200     struct ntb_queue_entry *entry;
2201     void *buf;
2202 
2203     if (!qp || qp->client_ready)
2204         return NULL;
2205 
2206     entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_pend_q);
2207     if (!entry)
2208         return NULL;
2209 
2210     buf = entry->cb_data;
2211     *len = entry->len;
2212 
2213     ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry, &qp->rx_free_q);
2214 
2215     return buf;
2216 }
2217 EXPORT_SYMBOL_GPL(ntb_transport_rx_remove);
2218 
2219 /**
2220  * ntb_transport_rx_enqueue - Enqueue a new NTB queue entry
2221  * @qp: NTB transport layer queue the entry is to be enqueued on
2222  * @cb: per buffer pointer for callback function to use
2223  * @data: pointer to data buffer that incoming packets will be copied into
2224  * @len: length of the data buffer
2225  *
2226  * Enqueue a new receive buffer onto the transport queue into which a NTB
2227  * payload can be received into.
2228  *
2229  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
2230  */
2231 int ntb_transport_rx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
2232                  unsigned int len)
2233 {
2234     struct ntb_queue_entry *entry;
2235 
2236     if (!qp)
2237         return -EINVAL;
2238 
2239     entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q);
2240     if (!entry)
2241         return -ENOMEM;
2242 
2243     entry->cb_data = cb;
2244     entry->buf = data;
2245     entry->len = len;
2246     entry->flags = 0;
2247     entry->retries = 0;
2248     entry->errors = 0;
2249     entry->rx_index = 0;
2250 
2251     ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry, &qp->rx_pend_q);
2252 
2253     if (qp->active)
2254         tasklet_schedule(&qp->rxc_db_work);
2255 
2256     return 0;
2257 }
2258 EXPORT_SYMBOL_GPL(ntb_transport_rx_enqueue);
2259 
2260 /**
2261  * ntb_transport_tx_enqueue - Enqueue a new NTB queue entry
2262  * @qp: NTB transport layer queue the entry is to be enqueued on
2263  * @cb: per buffer pointer for callback function to use
2264  * @data: pointer to data buffer that will be sent
2265  * @len: length of the data buffer
2266  *
2267  * Enqueue a new transmit buffer onto the transport queue from which a NTB
2268  * payload will be transmitted.  This assumes that a lock is being held to
2269  * serialize access to the qp.
2270  *
2271  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
2272  */
2273 int ntb_transport_tx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
2274                  unsigned int len)
2275 {
2276     struct ntb_queue_entry *entry;
2277     int rc;
2278 
2279     if (!qp || !qp->link_is_up || !len)
2280         return -EINVAL;
2281 
2282     entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
2283     if (!entry) {
2284         qp->tx_err_no_buf++;
2285         return -EBUSY;
2286     }
2287 
2288     entry->cb_data = cb;
2289     entry->buf = data;
2290     entry->len = len;
2291     entry->flags = 0;
2292     entry->errors = 0;
2293     entry->retries = 0;
2294     entry->tx_index = 0;
2295 
2296     rc = ntb_process_tx(qp, entry);
2297     if (rc)
2298         ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
2299                  &qp->tx_free_q);
2300 
2301     return rc;
2302 }
2303 EXPORT_SYMBOL_GPL(ntb_transport_tx_enqueue);
2304 
2305 /**
2306  * ntb_transport_link_up - Notify NTB transport of client readiness to use queue
2307  * @qp: NTB transport layer queue to be enabled
2308  *
2309  * Notify NTB transport layer of client readiness to use queue
2310  */
2311 void ntb_transport_link_up(struct ntb_transport_qp *qp)
2312 {
2313     if (!qp)
2314         return;
2315 
2316     qp->client_ready = true;
2317 
2318     if (qp->transport->link_is_up)
2319         schedule_delayed_work(&qp->link_work, 0);
2320 }
2321 EXPORT_SYMBOL_GPL(ntb_transport_link_up);
2322 
2323 /**
2324  * ntb_transport_link_down - Notify NTB transport to no longer enqueue data
2325  * @qp: NTB transport layer queue to be disabled
2326  *
2327  * Notify NTB transport layer of client's desire to no longer receive data on
2328  * transport queue specified.  It is the client's responsibility to ensure all
2329  * entries on queue are purged or otherwise handled appropriately.
2330  */
2331 void ntb_transport_link_down(struct ntb_transport_qp *qp)
2332 {
2333     int val;
2334 
2335     if (!qp)
2336         return;
2337 
2338     qp->client_ready = false;
2339 
2340     val = ntb_spad_read(qp->ndev, QP_LINKS);
2341 
2342     ntb_peer_spad_write(qp->ndev, PIDX, QP_LINKS, val & ~BIT(qp->qp_num));
2343 
2344     if (qp->link_is_up)
2345         ntb_send_link_down(qp);
2346     else
2347         cancel_delayed_work_sync(&qp->link_work);
2348 }
2349 EXPORT_SYMBOL_GPL(ntb_transport_link_down);
2350 
2351 /**
2352  * ntb_transport_link_query - Query transport link state
2353  * @qp: NTB transport layer queue to be queried
2354  *
2355  * Query connectivity to the remote system of the NTB transport queue
2356  *
2357  * RETURNS: true for link up or false for link down
2358  */
2359 bool ntb_transport_link_query(struct ntb_transport_qp *qp)
2360 {
2361     if (!qp)
2362         return false;
2363 
2364     return qp->link_is_up;
2365 }
2366 EXPORT_SYMBOL_GPL(ntb_transport_link_query);
2367 
2368 /**
2369  * ntb_transport_qp_num - Query the qp number
2370  * @qp: NTB transport layer queue to be queried
2371  *
2372  * Query qp number of the NTB transport queue
2373  *
2374  * RETURNS: a zero based number specifying the qp number
2375  */
2376 unsigned char ntb_transport_qp_num(struct ntb_transport_qp *qp)
2377 {
2378     if (!qp)
2379         return 0;
2380 
2381     return qp->qp_num;
2382 }
2383 EXPORT_SYMBOL_GPL(ntb_transport_qp_num);
2384 
2385 /**
2386  * ntb_transport_max_size - Query the max payload size of a qp
2387  * @qp: NTB transport layer queue to be queried
2388  *
2389  * Query the maximum payload size permissible on the given qp
2390  *
2391  * RETURNS: the max payload size of a qp
2392  */
2393 unsigned int ntb_transport_max_size(struct ntb_transport_qp *qp)
2394 {
2395     unsigned int max_size;
2396     unsigned int copy_align;
2397     struct dma_chan *rx_chan, *tx_chan;
2398 
2399     if (!qp)
2400         return 0;
2401 
2402     rx_chan = qp->rx_dma_chan;
2403     tx_chan = qp->tx_dma_chan;
2404 
2405     copy_align = max(rx_chan ? rx_chan->device->copy_align : 0,
2406              tx_chan ? tx_chan->device->copy_align : 0);
2407 
2408     /* If DMA engine usage is possible, try to find the max size for that */
2409     max_size = qp->tx_max_frame - sizeof(struct ntb_payload_header);
2410     max_size = round_down(max_size, 1 << copy_align);
2411 
2412     return max_size;
2413 }
2414 EXPORT_SYMBOL_GPL(ntb_transport_max_size);
2415 
2416 unsigned int ntb_transport_tx_free_entry(struct ntb_transport_qp *qp)
2417 {
2418     unsigned int head = qp->tx_index;
2419     unsigned int tail = qp->remote_rx_info->entry;
2420 
2421     return tail > head ? tail - head : qp->tx_max_entry + tail - head;
2422 }
2423 EXPORT_SYMBOL_GPL(ntb_transport_tx_free_entry);
2424 
2425 static void ntb_transport_doorbell_callback(void *data, int vector)
2426 {
2427     struct ntb_transport_ctx *nt = data;
2428     struct ntb_transport_qp *qp;
2429     u64 db_bits;
2430     unsigned int qp_num;
2431 
2432     if (ntb_db_read(nt->ndev) & nt->msi_db_mask) {
2433         ntb_transport_msi_peer_desc_changed(nt);
2434         ntb_db_clear(nt->ndev, nt->msi_db_mask);
2435     }
2436 
2437     db_bits = (nt->qp_bitmap & ~nt->qp_bitmap_free &
2438            ntb_db_vector_mask(nt->ndev, vector));
2439 
2440     while (db_bits) {
2441         qp_num = __ffs(db_bits);
2442         qp = &nt->qp_vec[qp_num];
2443 
2444         if (qp->active)
2445             tasklet_schedule(&qp->rxc_db_work);
2446 
2447         db_bits &= ~BIT_ULL(qp_num);
2448     }
2449 }
2450 
2451 static const struct ntb_ctx_ops ntb_transport_ops = {
2452     .link_event = ntb_transport_event_callback,
2453     .db_event = ntb_transport_doorbell_callback,
2454 };
2455 
2456 static struct ntb_client ntb_transport_client = {
2457     .ops = {
2458         .probe = ntb_transport_probe,
2459         .remove = ntb_transport_free,
2460     },
2461 };
2462 
2463 static int __init ntb_transport_init(void)
2464 {
2465     int rc;
2466 
2467     pr_info("%s, version %s\n", NTB_TRANSPORT_DESC, NTB_TRANSPORT_VER);
2468 
2469     if (debugfs_initialized())
2470         nt_debugfs_dir = debugfs_create_dir(KBUILD_MODNAME, NULL);
2471 
2472     rc = bus_register(&ntb_transport_bus);
2473     if (rc)
2474         goto err_bus;
2475 
2476     rc = ntb_register_client(&ntb_transport_client);
2477     if (rc)
2478         goto err_client;
2479 
2480     return 0;
2481 
2482 err_client:
2483     bus_unregister(&ntb_transport_bus);
2484 err_bus:
2485     debugfs_remove_recursive(nt_debugfs_dir);
2486     return rc;
2487 }
2488 module_init(ntb_transport_init);
2489 
2490 static void __exit ntb_transport_exit(void)
2491 {
2492     ntb_unregister_client(&ntb_transport_client);
2493     bus_unregister(&ntb_transport_bus);
2494     debugfs_remove_recursive(nt_debugfs_dir);
2495 }
2496 module_exit(ntb_transport_exit);
[ Source navigation ] [ Diff markup ] [ Identifier search ] [ general search ]

This page was automatically generated by the 2.1.0 LXR engine.

The LXR team
Valid CSS 2.1! Valid HTML 4.01!