lib/librte_eal/linuxapp/kni/kni_misc.c

   1 /*-
   2  * GPL LICENSE SUMMARY
   3  *
   4  *   Copyright(c) 2010-2014 Intel Corporation. All rights reserved.
   5  *
   6  *   This program is free software; you can redistribute it and/or modify
   7  *   it under the terms of version 2 of the GNU General Public License as
   8  *   published by the Free Software Foundation.
   9  *
  10  *   This program is distributed in the hope that it will be useful, but
  11  *   WITHOUT ANY WARRANTY; without even the implied warranty of
  12  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  *   General Public License for more details.
  14  *
  15  *   You should have received a copy of the GNU General Public License
  16  *   along with this program; if not, write to the Free Software
  17  *   Foundation, Inc., 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
  18  *   The full GNU General Public License is included in this distribution
  19  *   in the file called LICENSE.GPL.
  20  *
  21  *   Contact Information:
  22  *   Intel Corporation
  23  */
  24
  25 #include <linux/module.h>
  26 #include <linux/miscdevice.h>
  27 #include <linux/netdevice.h>
  28 #include <linux/pci.h>
  29 #include <linux/kthread.h>
  30 #include <linux/rwsem.h>
  31
  32 #include <exec-env/rte_kni_common.h>
  33 #include "kni_dev.h"
  34 #include <rte_config.h>
  35
  36 MODULE_LICENSE("Dual BSD/GPL");
  37 MODULE_AUTHOR("Intel Corporation");
  38 MODULE_DESCRIPTION("Kernel Module for managing kni devices");
  39
  40 #define KNI_RX_LOOP_NUM 1000
  41
  42 #define KNI_MAX_DEVICES 32
  43
  44 extern void kni_net_rx(struct kni_dev *kni);
  45 extern void kni_net_init(struct net_device *dev);
  46 extern void kni_net_config_lo_mode(char *lo_str);
  47 extern void kni_net_poll_resp(struct kni_dev *kni);
  48 extern void kni_set_ethtool_ops(struct net_device *netdev);
  49
  50 extern int ixgbe_kni_probe(struct pci_dev *pdev, struct net_device **lad_dev);
  51 extern void ixgbe_kni_remove(struct pci_dev *pdev);
  52 extern int igb_kni_probe(struct pci_dev *pdev, struct net_device **lad_dev);
  53 extern void igb_kni_remove(struct pci_dev *pdev);
  54
  55 static int kni_open(struct inode *inode, struct file *file);
  56 static int kni_release(struct inode *inode, struct file *file);
  57 static int kni_ioctl(struct inode *inode, unsigned int ioctl_num,
  58                                         unsigned long ioctl_param);
  59 static int kni_compat_ioctl(struct inode *inode, unsigned int ioctl_num,
  60                                                 unsigned long ioctl_param);
  61 static int kni_dev_remove(struct kni_dev *dev);
  62
  63 static int __init kni_parse_kthread_mode(void);
  64
  65 /* KNI processing for single kernel thread mode */
  66 static int kni_thread_single(void *unused);
  67 /* KNI processing for multiple kernel thread mode */
  68 static int kni_thread_multiple(void *param);
  69
  70 static struct file_operations kni_fops = {
  71         .owner = THIS_MODULE,
  72         .open = kni_open,
  73         .release = kni_release,
  74         .unlocked_ioctl = (void *)kni_ioctl,
  75         .compat_ioctl = (void *)kni_compat_ioctl,
  76 };
  77
  78 static struct miscdevice kni_misc = {
  79         .minor = MISC_DYNAMIC_MINOR,
  80         .name = KNI_DEVICE,
  81         .fops = &kni_fops,
  82 };
  83
  84 /* loopback mode */
  85 static char *lo_mode = NULL;
  86
  87 /* Kernel thread mode */
  88 static char *kthread_mode = NULL;
  89 static unsigned multiple_kthread_on = 0;
  90
  91 #define KNI_DEV_IN_USE_BIT_NUM 0 /* Bit number for device in use */
  92
  93 static volatile unsigned long device_in_use; /* device in use flag */
  94 static struct task_struct *kni_kthread;
  95
  96 /* kni list lock */
  97 static DECLARE_RWSEM(kni_list_lock);
  98
  99 /* kni list */
 100 static struct list_head kni_list_head = LIST_HEAD_INIT(kni_list_head);
 101
 102 static int __init
 103 kni_init(void)
 104 {
 105         KNI_PRINT("######## DPDK kni module loading ########\n");
 106
 107         if (kni_parse_kthread_mode() < 0) {
 108                 KNI_ERR("Invalid parameter for kthread_mode\n");
 109                 return -EINVAL;
 110         }
 111
 112         if (misc_register(&kni_misc) != 0) {
 113                 KNI_ERR("Misc registration failed\n");
 114                 return -EPERM;
 115         }
 116
 117         /* Clear the bit of device in use */
 118         clear_bit(KNI_DEV_IN_USE_BIT_NUM, &device_in_use);
 119
 120         /* Configure the lo mode according to the input parameter */
 121         kni_net_config_lo_mode(lo_mode);
 122
 123         KNI_PRINT("######## DPDK kni module loaded  ########\n");
 124
 125         return 0;
 126 }
 127
 128 static void __exit
 129 kni_exit(void)
 130 {
 131         misc_deregister(&kni_misc);
 132         KNI_PRINT("####### DPDK kni module unloaded  #######\n");
 133 }
 134
 135 static int __init
 136 kni_parse_kthread_mode(void)
 137 {
 138         if (!kthread_mode)
 139                 return 0;
 140
 141         if (strcmp(kthread_mode, "single") == 0)
 142                 return 0;
 143         else if (strcmp(kthread_mode, "multiple") == 0)
 144                 multiple_kthread_on = 1;
 145         else
 146                 return -1;
 147
 148         return 0;
 149 }
 150
 151 static int
 152 kni_open(struct inode *inode, struct file *file)
 153 {
 154         /* kni device can be opened by one user only, test and set bit */
 155         if (test_and_set_bit(KNI_DEV_IN_USE_BIT_NUM, &device_in_use))
 156                 return -EBUSY;
 157
 158         /* Create kernel thread for single mode */
 159         if (multiple_kthread_on == 0) {
 160                 KNI_PRINT("Single kernel thread for all KNI devices\n");
 161                 /* Create kernel thread for RX */
 162                 kni_kthread = kthread_run(kni_thread_single, NULL,
 163                                                 "kni_single");
 164                 if (IS_ERR(kni_kthread)) {
 165                         KNI_ERR("Unable to create kernel threaed\n");
 166                         return PTR_ERR(kni_kthread);
 167                 }
 168         } else
 169                 KNI_PRINT("Multiple kernel thread mode enabled\n");
 170
 171         KNI_PRINT("/dev/kni opened\n");
 172
 173         return 0;
 174 }
 175
 176 static int
 177 kni_release(struct inode *inode, struct file *file)
 178 {
 179         struct kni_dev *dev, *n;
 180
 181         /* Stop kernel thread for single mode */
 182         if (multiple_kthread_on == 0) {
 183                 /* Stop kernel thread */
 184                 kthread_stop(kni_kthread);
 185                 kni_kthread = NULL;
 186         }
 187
 188         down_write(&kni_list_lock);
 189         list_for_each_entry_safe(dev, n, &kni_list_head, list) {
 190                 /* Stop kernel thread for multiple mode */
 191                 if (multiple_kthread_on && dev->pthread != NULL) {
 192                         kthread_stop(dev->pthread);
 193                         dev->pthread = NULL;
 194                 }
 195
 196 #ifdef RTE_KNI_VHOST
 197                 kni_vhost_backend_release(dev);
 198 #endif
 199                 kni_dev_remove(dev);
 200                 list_del(&dev->list);
 201         }
 202         up_write(&kni_list_lock);
 203
 204         /* Clear the bit of device in use */
 205         clear_bit(KNI_DEV_IN_USE_BIT_NUM, &device_in_use);
 206
 207         KNI_PRINT("/dev/kni closed\n");
 208
 209         return 0;
 210 }
 211
 212 static int
 213 kni_thread_single(void *unused)
 214 {
 215         int j;
 216         struct kni_dev *dev;
 217
 218         while (!kthread_should_stop()) {
 219                 down_read(&kni_list_lock);
 220                 for (j = 0; j < KNI_RX_LOOP_NUM; j++) {
 221                         list_for_each_entry(dev, &kni_list_head, list) {
 222 #ifdef RTE_KNI_VHOST
 223                                 kni_chk_vhost_rx(dev);
 224 #else
 225                                 kni_net_rx(dev);
 226 #endif
 227                                 kni_net_poll_resp(dev);
 228                         }
 229                 }
 230                 up_read(&kni_list_lock);
 231 #ifdef RTE_KNI_PREEMPT_DEFAULT
 232                 /* reschedule out for a while */
 233                 schedule_timeout_interruptible(usecs_to_jiffies( \
 234                                 KNI_KTHREAD_RESCHEDULE_INTERVAL));
 235 #endif
 236         }
 237
 238         return 0;
 239 }
 240
 241 static int
 242 kni_thread_multiple(void *param)
 243 {
 244         int j;
 245         struct kni_dev *dev = (struct kni_dev *)param;
 246
 247         while (!kthread_should_stop()) {
 248                 for (j = 0; j < KNI_RX_LOOP_NUM; j++) {
 249 #ifdef RTE_KNI_VHOST
 250                         kni_chk_vhost_rx(dev);
 251 #else
 252                         kni_net_rx(dev);
 253 #endif
 254                         kni_net_poll_resp(dev);
 255                 }
 256 #ifdef RTE_KNI_PREEMPT_DEFAULT
 257                 schedule_timeout_interruptible(usecs_to_jiffies( \
 258                                 KNI_KTHREAD_RESCHEDULE_INTERVAL));
 259 #endif
 260         }
 261
 262         return 0;
 263 }
 264
 265 static int
 266 kni_dev_remove(struct kni_dev *dev)
 267 {
 268         if (!dev)
 269                 return -ENODEV;
 270
 271         switch (dev->device_id) {
 272         #define RTE_PCI_DEV_ID_DECL_IGB(vend, dev) case (dev):
 273         #include <rte_pci_dev_ids.h>
 274                 igb_kni_remove(dev->pci_dev);
 275                 break;
 276         #define RTE_PCI_DEV_ID_DECL_IXGBE(vend, dev) case (dev):
 277         #include <rte_pci_dev_ids.h>
 278                 ixgbe_kni_remove(dev->pci_dev);
 279                 break;
 280         default:
 281                 break;
 282         }
 283
 284         if (dev->net_dev) {
 285                 unregister_netdev(dev->net_dev);
 286                 free_netdev(dev->net_dev);
 287         }
 288
 289         return 0;
 290 }
 291
 292 static int
 293 kni_check_param(struct kni_dev *kni, struct rte_kni_device_info *dev)
 294 {
 295         if (!kni || !dev)
 296                 return -1;
 297
 298         /* Check if network name has been used */
 299         if (!strncmp(kni->name, dev->name, RTE_KNI_NAMESIZE)) {
 300                 KNI_ERR("KNI name %s duplicated\n", dev->name);
 301                 return -1;
 302         }
 303
 304         return 0;
 305 }
 306
 307 static int
 308 kni_ioctl_create(unsigned int ioctl_num, unsigned long ioctl_param)
 309 {
 310         int ret;
 311         struct rte_kni_device_info dev_info;
 312         struct pci_dev *pci = NULL;
 313         struct pci_dev *found_pci = NULL;
 314         struct net_device *net_dev = NULL;
 315         struct net_device *lad_dev = NULL;
 316         struct kni_dev *kni, *dev, *n;
 317         struct net *net;
 318
 319         printk(KERN_INFO "KNI: Creating kni...\n");
 320         /* Check the buffer size, to avoid warning */
 321         if (_IOC_SIZE(ioctl_num) > sizeof(dev_info))
 322                 return -EINVAL;
 323
 324         /* Copy kni info from user space */
 325         ret = copy_from_user(&dev_info, (void *)ioctl_param, sizeof(dev_info));
 326         if (ret) {
 327                 KNI_ERR("copy_from_user in kni_ioctl_create");
 328                 return -EIO;
 329         }
 330
 331         /**
 332          * Check if the cpu core id is valid for binding,
 333          * for multiple kernel thread mode.
 334          */
 335         if (multiple_kthread_on && dev_info.force_bind &&
 336                                 !cpu_online(dev_info.core_id)) {
 337                 KNI_ERR("cpu %u is not online\n", dev_info.core_id);
 338                 return -EINVAL;
 339         }
 340
 341         /* Check if it has been created */
 342         down_read(&kni_list_lock);
 343         list_for_each_entry_safe(dev, n, &kni_list_head, list) {
 344                 if (kni_check_param(dev, &dev_info) < 0) {
 345                         up_read(&kni_list_lock);
 346                         return -EINVAL;
 347                 }
 348         }
 349         up_read(&kni_list_lock);
 350
 351         net_dev = alloc_netdev(sizeof(struct kni_dev), dev_info.name,
 352 #ifdef NET_NAME_UNKNOWN
 353                                                         NET_NAME_UNKNOWN,
 354 #endif
 355                                                         kni_net_init);
 356         if (net_dev == NULL) {
 357                 KNI_ERR("error allocating device \"%s\"\n", dev_info.name);
 358                 return -EBUSY;
 359         }
 360
 361         net = get_net_ns_by_pid(task_pid_vnr(current));
 362         if (IS_ERR(net)) {
 363                 free_netdev(net_dev);
 364                 return PTR_ERR(net);
 365         }
 366         dev_net_set(net_dev, net);
 367         put_net(net);
 368
 369         kni = netdev_priv(net_dev);
 370
 371         kni->net_dev = net_dev;
 372         kni->group_id = dev_info.group_id;
 373         kni->core_id = dev_info.core_id;
 374         strncpy(kni->name, dev_info.name, RTE_KNI_NAMESIZE);
 375
 376         /* Translate user space info into kernel space info */
 377         kni->tx_q = phys_to_virt(dev_info.tx_phys);
 378         kni->rx_q = phys_to_virt(dev_info.rx_phys);
 379         kni->alloc_q = phys_to_virt(dev_info.alloc_phys);
 380         kni->free_q = phys_to_virt(dev_info.free_phys);
 381
 382         kni->req_q = phys_to_virt(dev_info.req_phys);
 383         kni->resp_q = phys_to_virt(dev_info.resp_phys);
 384         kni->sync_va = dev_info.sync_va;
 385         kni->sync_kva = phys_to_virt(dev_info.sync_phys);
 386
 387         kni->mbuf_kva = phys_to_virt(dev_info.mbuf_phys);
 388         kni->mbuf_va = dev_info.mbuf_va;
 389
 390 #ifdef RTE_KNI_VHOST
 391         kni->vhost_queue = NULL;
 392         kni->vq_status = BE_STOP;
 393 #endif
 394         kni->mbuf_size = dev_info.mbuf_size;
 395
 396         KNI_PRINT("tx_phys:      0x%016llx, tx_q addr:      0x%p\n",
 397                 (unsigned long long) dev_info.tx_phys, kni->tx_q);
 398         KNI_PRINT("rx_phys:      0x%016llx, rx_q addr:      0x%p\n",
 399                 (unsigned long long) dev_info.rx_phys, kni->rx_q);
 400         KNI_PRINT("alloc_phys:   0x%016llx, alloc_q addr:   0x%p\n",
 401                 (unsigned long long) dev_info.alloc_phys, kni->alloc_q);
 402         KNI_PRINT("free_phys:    0x%016llx, free_q addr:    0x%p\n",
 403                 (unsigned long long) dev_info.free_phys, kni->free_q);
 404         KNI_PRINT("req_phys:     0x%016llx, req_q addr:     0x%p\n",
 405                 (unsigned long long) dev_info.req_phys, kni->req_q);
 406         KNI_PRINT("resp_phys:    0x%016llx, resp_q addr:    0x%p\n",
 407                 (unsigned long long) dev_info.resp_phys, kni->resp_q);
 408         KNI_PRINT("mbuf_phys:    0x%016llx, mbuf_kva:       0x%p\n",
 409                 (unsigned long long) dev_info.mbuf_phys, kni->mbuf_kva);
 410         KNI_PRINT("mbuf_va:      0x%p\n", dev_info.mbuf_va);
 411         KNI_PRINT("mbuf_size:    %u\n", kni->mbuf_size);
 412
 413         KNI_DBG("PCI: %02x:%02x.%02x %04x:%04x\n",
 414                                         dev_info.bus,
 415                                         dev_info.devid,
 416                                         dev_info.function,
 417                                         dev_info.vendor_id,
 418                                         dev_info.device_id);
 419
 420         pci = pci_get_device(dev_info.vendor_id, dev_info.device_id, NULL);
 421
 422         /* Support Ethtool */
 423         while (pci) {
 424                 KNI_PRINT("pci_bus: %02x:%02x:%02x \n",
 425                                         pci->bus->number,
 426                                         PCI_SLOT(pci->devfn),
 427                                         PCI_FUNC(pci->devfn));
 428
 429                 if ((pci->bus->number == dev_info.bus) &&
 430                         (PCI_SLOT(pci->devfn) == dev_info.devid) &&
 431                         (PCI_FUNC(pci->devfn) == dev_info.function)) {
 432                         found_pci = pci;
 433                         switch (dev_info.device_id) {
 434                         #define RTE_PCI_DEV_ID_DECL_IGB(vend, dev) case (dev):
 435                         #include <rte_pci_dev_ids.h>
 436                                 ret = igb_kni_probe(found_pci, &lad_dev);
 437                                 break;
 438                         #define RTE_PCI_DEV_ID_DECL_IXGBE(vend, dev) \
 439                                                         case (dev):
 440                         #include <rte_pci_dev_ids.h>
 441                                 ret = ixgbe_kni_probe(found_pci, &lad_dev);
 442                                 break;
 443                         default:
 444                                 ret = -1;
 445                                 break;
 446                         }
 447
 448                         KNI_DBG("PCI found: pci=0x%p, lad_dev=0x%p\n",
 449                                                         pci, lad_dev);
 450                         if (ret == 0) {
 451                                 kni->lad_dev = lad_dev;
 452                                 kni_set_ethtool_ops(kni->net_dev);
 453                         } else {
 454                                 KNI_ERR("Device not supported by ethtool");
 455                                 kni->lad_dev = NULL;
 456                         }
 457
 458                         kni->pci_dev = found_pci;
 459                         kni->device_id = dev_info.device_id;
 460                         break;
 461                 }
 462                 pci = pci_get_device(dev_info.vendor_id,
 463                                 dev_info.device_id, pci);
 464         }
 465         if (pci)
 466                 pci_dev_put(pci);
 467
 468         ret = register_netdev(net_dev);
 469         if (ret) {
 470                 KNI_ERR("error %i registering device \"%s\"\n",
 471                                         ret, dev_info.name);
 472                 kni_dev_remove(kni);
 473                 return -ENODEV;
 474         }
 475
 476 #ifdef RTE_KNI_VHOST
 477         kni_vhost_init(kni);
 478 #endif
 479
 480         /**
 481          * Create a new kernel thread for multiple mode, set its core affinity,
 482          * and finally wake it up.
 483          */
 484         if (multiple_kthread_on) {
 485                 kni->pthread = kthread_create(kni_thread_multiple,
 486                                               (void *)kni,
 487                                               "kni_%s", kni->name);
 488                 if (IS_ERR(kni->pthread)) {
 489                         kni_dev_remove(kni);
 490                         return -ECANCELED;
 491                 }
 492                 if (dev_info.force_bind)
 493                         kthread_bind(kni->pthread, kni->core_id);
 494                 wake_up_process(kni->pthread);
 495         }
 496
 497         down_write(&kni_list_lock);
 498         list_add(&kni->list, &kni_list_head);
 499         up_write(&kni_list_lock);
 500
 501         return 0;
 502 }
 503
 504 static int
 505 kni_ioctl_release(unsigned int ioctl_num, unsigned long ioctl_param)
 506 {
 507         int ret = -EINVAL;
 508         struct kni_dev *dev, *n;
 509         struct rte_kni_device_info dev_info;
 510
 511         if (_IOC_SIZE(ioctl_num) > sizeof(dev_info))
 512                         return -EINVAL;
 513
 514         ret = copy_from_user(&dev_info, (void *)ioctl_param, sizeof(dev_info));
 515         if (ret) {
 516                 KNI_ERR("copy_from_user in kni_ioctl_release");
 517                 return -EIO;
 518         }
 519
 520         /* Release the network device according to its name */
 521         if (strlen(dev_info.name) == 0)
 522                 return ret;
 523
 524         down_write(&kni_list_lock);
 525         list_for_each_entry_safe(dev, n, &kni_list_head, list) {
 526                 if (strncmp(dev->name, dev_info.name, RTE_KNI_NAMESIZE) != 0)
 527                         continue;
 528
 529                 if (multiple_kthread_on && dev->pthread != NULL) {
 530                         kthread_stop(dev->pthread);
 531                         dev->pthread = NULL;
 532                 }
 533
 534 #ifdef RTE_KNI_VHOST
 535                 kni_vhost_backend_release(dev);
 536 #endif
 537                 kni_dev_remove(dev);
 538                 list_del(&dev->list);
 539                 ret = 0;
 540                 break;
 541         }
 542         up_write(&kni_list_lock);
 543         printk(KERN_INFO "KNI: %s release kni named %s\n",
 544                 (ret == 0 ? "Successfully" : "Unsuccessfully"), dev_info.name);
 545
 546         return ret;
 547 }
 548
 549 static int
 550 kni_ioctl(struct inode *inode,
 551         unsigned int ioctl_num,
 552         unsigned long ioctl_param)
 553 {
 554         int ret = -EINVAL;
 555
 556         KNI_DBG("IOCTL num=0x%0x param=0x%0lx \n", ioctl_num, ioctl_param);
 557
 558         /*
 559          * Switch according to the ioctl called
 560          */
 561         switch (_IOC_NR(ioctl_num)) {
 562         case _IOC_NR(RTE_KNI_IOCTL_TEST):
 563                 /* For test only, not used */
 564                 break;
 565         case _IOC_NR(RTE_KNI_IOCTL_CREATE):
 566                 ret = kni_ioctl_create(ioctl_num, ioctl_param);
 567                 break;
 568         case _IOC_NR(RTE_KNI_IOCTL_RELEASE):
 569                 ret = kni_ioctl_release(ioctl_num, ioctl_param);
 570                 break;
 571         default:
 572                 KNI_DBG("IOCTL default \n");
 573                 break;
 574         }
 575
 576         return ret;
 577 }
 578
 579 static int
 580 kni_compat_ioctl(struct inode *inode,
 581                 unsigned int ioctl_num,
 582                 unsigned long ioctl_param)
 583 {
 584         /* 32 bits app on 64 bits OS to be supported later */
 585         KNI_PRINT("Not implemented.\n");
 586
 587         return -EINVAL;
 588 }
 589
 590 module_init(kni_init);
 591 module_exit(kni_exit);
 592
 593 module_param(lo_mode, charp, S_IRUGO | S_IWUSR);
 594 MODULE_PARM_DESC(lo_mode,
 595 "KNI loopback mode (default=lo_mode_none):\n"
 596 "    lo_mode_none        Kernel loopback disabled\n"
 597 "    lo_mode_fifo        Enable kernel loopback with fifo\n"
 598 "    lo_mode_fifo_skb    Enable kernel loopback with fifo and skb buffer\n"
 599 "\n"
 600 );
 601
 602 module_param(kthread_mode, charp, S_IRUGO);
 603 MODULE_PARM_DESC(kthread_mode,
 604 "Kernel thread mode (default=single):\n"
 605 "    single    Single kernel thread mode enabled.\n"
 606 "    multiple  Multiple kernel thread mode enabled.\n"
 607 "\n"
 608 );