hw/vfio/ccw.c

   1 /*
   2  * vfio based subchannel assignment support
   3  *
   4  * Copyright 2017 IBM Corp.
   5  * Author(s): Dong Jia Shi <bjsdjshi@linux.vnet.ibm.com>
   6  *            Xiao Feng Ren <renxiaof@linux.vnet.ibm.com>
   7  *            Pierre Morel <pmorel@linux.vnet.ibm.com>
   8  *
   9  * This work is licensed under the terms of the GNU GPL, version 2 or (at
  10  * your option) any later version. See the COPYING file in the top-level
  11  * directory.
  12  */
  13
  14 #include "qemu/osdep.h"
  15 #include <linux/vfio.h>
  16 #include <linux/vfio_ccw.h>
  17 #include <sys/ioctl.h>
  18
  19 #include "qapi/error.h"
  20 #include "hw/sysbus.h"
  21 #include "hw/vfio/vfio.h"
  22 #include "hw/vfio/vfio-common.h"
  23 #include "hw/s390x/s390-ccw.h"
  24 #include "hw/s390x/ccw-device.h"
  25 #include "exec/address-spaces.h"
  26 #include "qemu/error-report.h"
  27
  28 #define TYPE_VFIO_CCW "vfio-ccw"
  29 typedef struct VFIOCCWDevice {
  30     S390CCWDevice cdev;
  31     VFIODevice vdev;
  32     uint64_t io_region_size;
  33     uint64_t io_region_offset;
  34     struct ccw_io_region *io_region;
  35     EventNotifier io_notifier;
  36     bool force_orb_pfch;
  37     bool warned_orb_pfch;
  38 } VFIOCCWDevice;
  39
  40 static inline void warn_once_pfch(VFIOCCWDevice *vcdev, SubchDev *sch,
  41                                   const char *msg)
  42 {
  43     warn_report_once_cond(&vcdev->warned_orb_pfch,
  44                           "vfio-ccw (devno %x.%x.%04x): %s",
  45                           sch->cssid, sch->ssid, sch->devno, msg);
  46 }
  47
  48 static void vfio_ccw_compute_needs_reset(VFIODevice *vdev)
  49 {
  50     vdev->needs_reset = false;
  51 }
  52
  53 /*
  54  * We don't need vfio_hot_reset_multi and vfio_eoi operations for
  55  * vfio_ccw device now.
  56  */
  57 struct VFIODeviceOps vfio_ccw_ops = {
  58     .vfio_compute_needs_reset = vfio_ccw_compute_needs_reset,
  59 };
  60
  61 static IOInstEnding vfio_ccw_handle_request(SubchDev *sch)
  62 {
  63     S390CCWDevice *cdev = sch->driver_data;
  64     VFIOCCWDevice *vcdev = DO_UPCAST(VFIOCCWDevice, cdev, cdev);
  65     struct ccw_io_region *region = vcdev->io_region;
  66     int ret;
  67
  68     if (!(sch->orb.ctrl0 & ORB_CTRL0_MASK_PFCH)) {
  69         if (!(vcdev->force_orb_pfch)) {
  70             warn_once_pfch(vcdev, sch, "requires PFCH flag set");
  71             sch_gen_unit_exception(sch);
  72             css_inject_io_interrupt(sch);
  73             return IOINST_CC_EXPECTED;
  74         } else {
  75             sch->orb.ctrl0 |= ORB_CTRL0_MASK_PFCH;
  76             warn_once_pfch(vcdev, sch, "PFCH flag forced");
  77         }
  78     }
  79
  80     QEMU_BUILD_BUG_ON(sizeof(region->orb_area) != sizeof(ORB));
  81     QEMU_BUILD_BUG_ON(sizeof(region->scsw_area) != sizeof(SCSW));
  82     QEMU_BUILD_BUG_ON(sizeof(region->irb_area) != sizeof(IRB));
  83
  84     memset(region, 0, sizeof(*region));
  85
  86     memcpy(region->orb_area, &sch->orb, sizeof(ORB));
  87     memcpy(region->scsw_area, &sch->curr_status.scsw, sizeof(SCSW));
  88
  89 again:
  90     ret = pwrite(vcdev->vdev.fd, region,
  91                  vcdev->io_region_size, vcdev->io_region_offset);
  92     if (ret != vcdev->io_region_size) {
  93         if (errno == EAGAIN) {
  94             goto again;
  95         }
  96         error_report("vfio-ccw: wirte I/O region failed with errno=%d", errno);
  97         ret = -errno;
  98     } else {
  99         ret = region->ret_code;
 100     }
 101     switch (ret) {
 102     case 0:
 103         return IOINST_CC_EXPECTED;
 104     case -EBUSY:
 105         return IOINST_CC_BUSY;
 106     case -ENODEV:
 107     case -EACCES:
 108         return IOINST_CC_NOT_OPERATIONAL;
 109     case -EFAULT:
 110     default:
 111         sch_gen_unit_exception(sch);
 112         css_inject_io_interrupt(sch);
 113         return IOINST_CC_EXPECTED;
 114     }
 115 }
 116
 117 static void vfio_ccw_reset(DeviceState *dev)
 118 {
 119     CcwDevice *ccw_dev = DO_UPCAST(CcwDevice, parent_obj, dev);
 120     S390CCWDevice *cdev = DO_UPCAST(S390CCWDevice, parent_obj, ccw_dev);
 121     VFIOCCWDevice *vcdev = DO_UPCAST(VFIOCCWDevice, cdev, cdev);
 122
 123     ioctl(vcdev->vdev.fd, VFIO_DEVICE_RESET);
 124 }
 125
 126 static void vfio_ccw_io_notifier_handler(void *opaque)
 127 {
 128     VFIOCCWDevice *vcdev = opaque;
 129     struct ccw_io_region *region = vcdev->io_region;
 130     S390CCWDevice *cdev = S390_CCW_DEVICE(vcdev);
 131     CcwDevice *ccw_dev = CCW_DEVICE(cdev);
 132     SubchDev *sch = ccw_dev->sch;
 133     SCHIB *schib = &sch->curr_status;
 134     SCSW s;
 135     IRB irb;
 136     int size;
 137
 138     if (!event_notifier_test_and_clear(&vcdev->io_notifier)) {
 139         return;
 140     }
 141
 142     size = pread(vcdev->vdev.fd, region, vcdev->io_region_size,
 143                  vcdev->io_region_offset);
 144     if (size == -1) {
 145         switch (errno) {
 146         case ENODEV:
 147             /* Generate a deferred cc 3 condition. */
 148             schib->scsw.flags |= SCSW_FLAGS_MASK_CC;
 149             schib->scsw.ctrl &= ~SCSW_CTRL_MASK_STCTL;
 150             schib->scsw.ctrl |= (SCSW_STCTL_ALERT | SCSW_STCTL_STATUS_PEND);
 151             goto read_err;
 152         case EFAULT:
 153             /* Memory problem, generate channel data check. */
 154             schib->scsw.ctrl &= ~SCSW_ACTL_START_PEND;
 155             schib->scsw.cstat = SCSW_CSTAT_DATA_CHECK;
 156             schib->scsw.ctrl &= ~SCSW_CTRL_MASK_STCTL;
 157             schib->scsw.ctrl |= SCSW_STCTL_PRIMARY | SCSW_STCTL_SECONDARY |
 158                        SCSW_STCTL_ALERT | SCSW_STCTL_STATUS_PEND;
 159             goto read_err;
 160         default:
 161             /* Error, generate channel program check. */
 162             schib->scsw.ctrl &= ~SCSW_ACTL_START_PEND;
 163             schib->scsw.cstat = SCSW_CSTAT_PROG_CHECK;
 164             schib->scsw.ctrl &= ~SCSW_CTRL_MASK_STCTL;
 165             schib->scsw.ctrl |= SCSW_STCTL_PRIMARY | SCSW_STCTL_SECONDARY |
 166                        SCSW_STCTL_ALERT | SCSW_STCTL_STATUS_PEND;
 167             goto read_err;
 168         }
 169     } else if (size != vcdev->io_region_size) {
 170         /* Information transfer error, generate channel-control check. */
 171         schib->scsw.ctrl &= ~SCSW_ACTL_START_PEND;
 172         schib->scsw.cstat = SCSW_CSTAT_CHN_CTRL_CHK;
 173         schib->scsw.ctrl &= ~SCSW_CTRL_MASK_STCTL;
 174         schib->scsw.ctrl |= SCSW_STCTL_PRIMARY | SCSW_STCTL_SECONDARY |
 175                    SCSW_STCTL_ALERT | SCSW_STCTL_STATUS_PEND;
 176         goto read_err;
 177     }
 178
 179     memcpy(&irb, region->irb_area, sizeof(IRB));
 180
 181     /* Update control block via irb. */
 182     s = schib->scsw;
 183     copy_scsw_to_guest(&s, &irb.scsw);
 184     schib->scsw = s;
 185
 186     /* If a uint check is pending, copy sense data. */
 187     if ((schib->scsw.dstat & SCSW_DSTAT_UNIT_CHECK) &&
 188         (schib->pmcw.chars & PMCW_CHARS_MASK_CSENSE)) {
 189         memcpy(sch->sense_data, irb.ecw, sizeof(irb.ecw));
 190     }
 191
 192 read_err:
 193     css_inject_io_interrupt(sch);
 194 }
 195
 196 static void vfio_ccw_register_io_notifier(VFIOCCWDevice *vcdev, Error **errp)
 197 {
 198     VFIODevice *vdev = &vcdev->vdev;
 199     struct vfio_irq_info *irq_info;
 200     struct vfio_irq_set *irq_set;
 201     size_t argsz;
 202     int32_t *pfd;
 203
 204     if (vdev->num_irqs < VFIO_CCW_IO_IRQ_INDEX + 1) {
 205         error_setg(errp, "vfio: unexpected number of io irqs %u",
 206                    vdev->num_irqs);
 207         return;
 208     }
 209
 210     argsz = sizeof(*irq_info);
 211     irq_info = g_malloc0(argsz);
 212     irq_info->index = VFIO_CCW_IO_IRQ_INDEX;
 213     irq_info->argsz = argsz;
 214     if (ioctl(vdev->fd, VFIO_DEVICE_GET_IRQ_INFO,
 215               irq_info) < 0 || irq_info->count < 1) {
 216         error_setg_errno(errp, errno, "vfio: Error getting irq info");
 217         goto out_free_info;
 218     }
 219
 220     if (event_notifier_init(&vcdev->io_notifier, 0)) {
 221         error_setg_errno(errp, errno,
 222                          "vfio: Unable to init event notifier for IO");
 223         goto out_free_info;
 224     }
 225
 226     argsz = sizeof(*irq_set) + sizeof(*pfd);
 227     irq_set = g_malloc0(argsz);
 228     irq_set->argsz = argsz;
 229     irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD |
 230                      VFIO_IRQ_SET_ACTION_TRIGGER;
 231     irq_set->index = VFIO_CCW_IO_IRQ_INDEX;
 232     irq_set->start = 0;
 233     irq_set->count = 1;
 234     pfd = (int32_t *) &irq_set->data;
 235
 236     *pfd = event_notifier_get_fd(&vcdev->io_notifier);
 237     qemu_set_fd_handler(*pfd, vfio_ccw_io_notifier_handler, NULL, vcdev);
 238     if (ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, irq_set)) {
 239         error_setg(errp, "vfio: Failed to set up io notification");
 240         qemu_set_fd_handler(*pfd, NULL, NULL, vcdev);
 241         event_notifier_cleanup(&vcdev->io_notifier);
 242     }
 243
 244     g_free(irq_set);
 245
 246 out_free_info:
 247     g_free(irq_info);
 248 }
 249
 250 static void vfio_ccw_unregister_io_notifier(VFIOCCWDevice *vcdev)
 251 {
 252     struct vfio_irq_set *irq_set;
 253     size_t argsz;
 254     int32_t *pfd;
 255
 256     argsz = sizeof(*irq_set) + sizeof(*pfd);
 257     irq_set = g_malloc0(argsz);
 258     irq_set->argsz = argsz;
 259     irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD |
 260                      VFIO_IRQ_SET_ACTION_TRIGGER;
 261     irq_set->index = VFIO_CCW_IO_IRQ_INDEX;
 262     irq_set->start = 0;
 263     irq_set->count = 1;
 264     pfd = (int32_t *) &irq_set->data;
 265     *pfd = -1;
 266
 267     if (ioctl(vcdev->vdev.fd, VFIO_DEVICE_SET_IRQS, irq_set)) {
 268         error_report("vfio: Failed to de-assign device io fd: %m");
 269     }
 270
 271     qemu_set_fd_handler(event_notifier_get_fd(&vcdev->io_notifier),
 272                         NULL, NULL, vcdev);
 273     event_notifier_cleanup(&vcdev->io_notifier);
 274
 275     g_free(irq_set);
 276 }
 277
 278 static void vfio_ccw_get_region(VFIOCCWDevice *vcdev, Error **errp)
 279 {
 280     VFIODevice *vdev = &vcdev->vdev;
 281     struct vfio_region_info *info;
 282     int ret;
 283
 284     /* Sanity check device */
 285     if (!(vdev->flags & VFIO_DEVICE_FLAGS_CCW)) {
 286         error_setg(errp, "vfio: Um, this isn't a vfio-ccw device");
 287         return;
 288     }
 289
 290     if (vdev->num_regions < VFIO_CCW_CONFIG_REGION_INDEX + 1) {
 291         error_setg(errp, "vfio: Unexpected number of the I/O region %u",
 292                    vdev->num_regions);
 293         return;
 294     }
 295
 296     ret = vfio_get_region_info(vdev, VFIO_CCW_CONFIG_REGION_INDEX, &info);
 297     if (ret) {
 298         error_setg_errno(errp, -ret, "vfio: Error getting config info");
 299         return;
 300     }
 301
 302     vcdev->io_region_size = info->size;
 303     if (sizeof(*vcdev->io_region) != vcdev->io_region_size) {
 304         error_setg(errp, "vfio: Unexpected size of the I/O region");
 305         g_free(info);
 306         return;
 307     }
 308
 309     vcdev->io_region_offset = info->offset;
 310     vcdev->io_region = g_malloc0(info->size);
 311
 312     g_free(info);
 313 }
 314
 315 static void vfio_ccw_put_region(VFIOCCWDevice *vcdev)
 316 {
 317     g_free(vcdev->io_region);
 318 }
 319
 320 static void vfio_ccw_put_device(VFIOCCWDevice *vcdev)
 321 {
 322     g_free(vcdev->vdev.name);
 323     vfio_put_base_device(&vcdev->vdev);
 324 }
 325
 326 static void vfio_ccw_get_device(VFIOGroup *group, VFIOCCWDevice *vcdev,
 327                                 Error **errp)
 328 {
 329     char *name = g_strdup_printf("%x.%x.%04x", vcdev->cdev.hostid.cssid,
 330                                  vcdev->cdev.hostid.ssid,
 331                                  vcdev->cdev.hostid.devid);
 332     VFIODevice *vbasedev;
 333
 334     QLIST_FOREACH(vbasedev, &group->device_list, next) {
 335         if (strcmp(vbasedev->name, name) == 0) {
 336             error_setg(errp, "vfio: subchannel %s has already been attached",
 337                        name);
 338             goto out_err;
 339         }
 340     }
 341
 342     /*
 343      * All vfio-ccw devices are believed to operate in a way compatible with
 344      * memory ballooning, ie. pages pinned in the host are in the current
 345      * working set of the guest driver and therefore never overlap with pages
 346      * available to the guest balloon driver.  This needs to be set before
 347      * vfio_get_device() for vfio common to handle the balloon inhibitor.
 348      */
 349     vcdev->vdev.balloon_allowed = true;
 350
 351     if (vfio_get_device(group, vcdev->cdev.mdevid, &vcdev->vdev, errp)) {
 352         goto out_err;
 353     }
 354
 355     vcdev->vdev.ops = &vfio_ccw_ops;
 356     vcdev->vdev.type = VFIO_DEVICE_TYPE_CCW;
 357     vcdev->vdev.name = name;
 358     vcdev->vdev.dev = &vcdev->cdev.parent_obj.parent_obj;
 359
 360     return;
 361
 362 out_err:
 363     g_free(name);
 364 }
 365
 366 static VFIOGroup *vfio_ccw_get_group(S390CCWDevice *cdev, Error **errp)
 367 {
 368     char *tmp, group_path[PATH_MAX];
 369     ssize_t len;
 370     int groupid;
 371
 372     tmp = g_strdup_printf("/sys/bus/css/devices/%x.%x.%04x/%s/iommu_group",
 373                           cdev->hostid.cssid, cdev->hostid.ssid,
 374                           cdev->hostid.devid, cdev->mdevid);
 375     len = readlink(tmp, group_path, sizeof(group_path));
 376     g_free(tmp);
 377
 378     if (len <= 0 || len >= sizeof(group_path)) {
 379         error_setg(errp, "vfio: no iommu_group found");
 380         return NULL;
 381     }
 382
 383     group_path[len] = 0;
 384
 385     if (sscanf(basename(group_path), "%d", &groupid) != 1) {
 386         error_setg(errp, "vfio: failed to read %s", group_path);
 387         return NULL;
 388     }
 389
 390     return vfio_get_group(groupid, &address_space_memory, errp);
 391 }
 392
 393 static void vfio_ccw_realize(DeviceState *dev, Error **errp)
 394 {
 395     VFIOGroup *group;
 396     CcwDevice *ccw_dev = DO_UPCAST(CcwDevice, parent_obj, dev);
 397     S390CCWDevice *cdev = DO_UPCAST(S390CCWDevice, parent_obj, ccw_dev);
 398     VFIOCCWDevice *vcdev = DO_UPCAST(VFIOCCWDevice, cdev, cdev);
 399     S390CCWDeviceClass *cdc = S390_CCW_DEVICE_GET_CLASS(cdev);
 400     Error *err = NULL;
 401
 402     /* Call the class init function for subchannel. */
 403     if (cdc->realize) {
 404         cdc->realize(cdev, vcdev->vdev.sysfsdev, &err);
 405         if (err) {
 406             goto out_err_propagate;
 407         }
 408     }
 409
 410     group = vfio_ccw_get_group(cdev, &err);
 411     if (!group) {
 412         goto out_group_err;
 413     }
 414
 415     vfio_ccw_get_device(group, vcdev, &err);
 416     if (err) {
 417         goto out_device_err;
 418     }
 419
 420     vfio_ccw_get_region(vcdev, &err);
 421     if (err) {
 422         goto out_region_err;
 423     }
 424
 425     vfio_ccw_register_io_notifier(vcdev, &err);
 426     if (err) {
 427         goto out_notifier_err;
 428     }
 429
 430     return;
 431
 432 out_notifier_err:
 433     vfio_ccw_put_region(vcdev);
 434 out_region_err:
 435     vfio_ccw_put_device(vcdev);
 436 out_device_err:
 437     vfio_put_group(group);
 438 out_group_err:
 439     if (cdc->unrealize) {
 440         cdc->unrealize(cdev, NULL);
 441     }
 442 out_err_propagate:
 443     error_propagate(errp, err);
 444 }
 445
 446 static void vfio_ccw_unrealize(DeviceState *dev, Error **errp)
 447 {
 448     CcwDevice *ccw_dev = DO_UPCAST(CcwDevice, parent_obj, dev);
 449     S390CCWDevice *cdev = DO_UPCAST(S390CCWDevice, parent_obj, ccw_dev);
 450     VFIOCCWDevice *vcdev = DO_UPCAST(VFIOCCWDevice, cdev, cdev);
 451     S390CCWDeviceClass *cdc = S390_CCW_DEVICE_GET_CLASS(cdev);
 452     VFIOGroup *group = vcdev->vdev.group;
 453
 454     vfio_ccw_unregister_io_notifier(vcdev);
 455     vfio_ccw_put_region(vcdev);
 456     vfio_ccw_put_device(vcdev);
 457     vfio_put_group(group);
 458
 459     if (cdc->unrealize) {
 460         cdc->unrealize(cdev, errp);
 461     }
 462 }
 463
 464 static Property vfio_ccw_properties[] = {
 465     DEFINE_PROP_STRING("sysfsdev", VFIOCCWDevice, vdev.sysfsdev),
 466     DEFINE_PROP_BOOL("force-orb-pfch", VFIOCCWDevice, force_orb_pfch, false),
 467     DEFINE_PROP_END_OF_LIST(),
 468 };
 469
 470 static const VMStateDescription vfio_ccw_vmstate = {
 471     .name = TYPE_VFIO_CCW,
 472     .unmigratable = 1,
 473 };
 474
 475 static void vfio_ccw_class_init(ObjectClass *klass, void *data)
 476 {
 477     DeviceClass *dc = DEVICE_CLASS(klass);
 478     S390CCWDeviceClass *cdc = S390_CCW_DEVICE_CLASS(klass);
 479
 480     dc->props = vfio_ccw_properties;
 481     dc->vmsd = &vfio_ccw_vmstate;
 482     dc->desc = "VFIO-based subchannel assignment";
 483     set_bit(DEVICE_CATEGORY_MISC, dc->categories);
 484     dc->realize = vfio_ccw_realize;
 485     dc->unrealize = vfio_ccw_unrealize;
 486     dc->reset = vfio_ccw_reset;
 487
 488     cdc->handle_request = vfio_ccw_handle_request;
 489 }
 490
 491 static const TypeInfo vfio_ccw_info = {
 492     .name = TYPE_VFIO_CCW,
 493     .parent = TYPE_S390_CCW,
 494     .instance_size = sizeof(VFIOCCWDevice),
 495     .class_init = vfio_ccw_class_init,
 496 };
 497
 498 static void register_vfio_ccw_type(void)
 499 {
 500     type_register_static(&vfio_ccw_info);
 501 }
 502
 503 type_init(register_vfio_ccw_type)