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(bool *warned, const char *fmt, ...)
  41 {
  42     va_list ap;
  43
  44     if (!warned || *warned) {
  45         return;
  46     }
  47     *warned = true;
  48     va_start(ap, fmt);
  49     warn_vreport(fmt, ap);
  50     va_end(ap);
  51 }
  52
  53 static inline void warn_once_pfch(VFIOCCWDevice *vcdev, SubchDev *sch,
  54                                   const char *msg)
  55 {
  56     warn_once(&vcdev->warned_orb_pfch, "vfio-ccw (devno %x.%x.%04x): %s",
  57               sch->cssid, sch->ssid, sch->devno, msg);
  58 }
  59
  60 static void vfio_ccw_compute_needs_reset(VFIODevice *vdev)
  61 {
  62     vdev->needs_reset = false;
  63 }
  64
  65 /*
  66  * We don't need vfio_hot_reset_multi and vfio_eoi operations for
  67  * vfio_ccw device now.
  68  */
  69 struct VFIODeviceOps vfio_ccw_ops = {
  70     .vfio_compute_needs_reset = vfio_ccw_compute_needs_reset,
  71 };
  72
  73 static IOInstEnding vfio_ccw_handle_request(SubchDev *sch)
  74 {
  75     S390CCWDevice *cdev = sch->driver_data;
  76     VFIOCCWDevice *vcdev = DO_UPCAST(VFIOCCWDevice, cdev, cdev);
  77     struct ccw_io_region *region = vcdev->io_region;
  78     int ret;
  79
  80     if (!(sch->orb.ctrl0 & ORB_CTRL0_MASK_PFCH)) {
  81         if (!(vcdev->force_orb_pfch)) {
  82             warn_once_pfch(vcdev, sch, "requires PFCH flag set");
  83             sch_gen_unit_exception(sch);
  84             css_inject_io_interrupt(sch);
  85             return IOINST_CC_EXPECTED;
  86         } else {
  87             sch->orb.ctrl0 |= ORB_CTRL0_MASK_PFCH;
  88             warn_once_pfch(vcdev, sch, "PFCH flag forced");
  89         }
  90     }
  91
  92     QEMU_BUILD_BUG_ON(sizeof(region->orb_area) != sizeof(ORB));
  93     QEMU_BUILD_BUG_ON(sizeof(region->scsw_area) != sizeof(SCSW));
  94     QEMU_BUILD_BUG_ON(sizeof(region->irb_area) != sizeof(IRB));
  95
  96     memset(region, 0, sizeof(*region));
  97
  98     memcpy(region->orb_area, &sch->orb, sizeof(ORB));
  99     memcpy(region->scsw_area, &sch->curr_status.scsw, sizeof(SCSW));
 100
 101 again:
 102     ret = pwrite(vcdev->vdev.fd, region,
 103                  vcdev->io_region_size, vcdev->io_region_offset);
 104     if (ret != vcdev->io_region_size) {
 105         if (errno == EAGAIN) {
 106             goto again;
 107         }
 108         error_report("vfio-ccw: wirte I/O region failed with errno=%d", errno);
 109         ret = -errno;
 110     } else {
 111         ret = region->ret_code;
 112     }
 113     switch (ret) {
 114     case 0:
 115         return IOINST_CC_EXPECTED;
 116     case -EBUSY:
 117         return IOINST_CC_BUSY;
 118     case -ENODEV:
 119     case -EACCES:
 120         return IOINST_CC_NOT_OPERATIONAL;
 121     case -EFAULT:
 122     default:
 123         sch_gen_unit_exception(sch);
 124         css_inject_io_interrupt(sch);
 125         return IOINST_CC_EXPECTED;
 126     }
 127 }
 128
 129 static void vfio_ccw_reset(DeviceState *dev)
 130 {
 131     CcwDevice *ccw_dev = DO_UPCAST(CcwDevice, parent_obj, dev);
 132     S390CCWDevice *cdev = DO_UPCAST(S390CCWDevice, parent_obj, ccw_dev);
 133     VFIOCCWDevice *vcdev = DO_UPCAST(VFIOCCWDevice, cdev, cdev);
 134
 135     ioctl(vcdev->vdev.fd, VFIO_DEVICE_RESET);
 136 }
 137
 138 static void vfio_ccw_io_notifier_handler(void *opaque)
 139 {
 140     VFIOCCWDevice *vcdev = opaque;
 141     struct ccw_io_region *region = vcdev->io_region;
 142     S390CCWDevice *cdev = S390_CCW_DEVICE(vcdev);
 143     CcwDevice *ccw_dev = CCW_DEVICE(cdev);
 144     SubchDev *sch = ccw_dev->sch;
 145     SCSW *s = &sch->curr_status.scsw;
 146     PMCW *p = &sch->curr_status.pmcw;
 147     IRB irb;
 148     int size;
 149
 150     if (!event_notifier_test_and_clear(&vcdev->io_notifier)) {
 151         return;
 152     }
 153
 154     size = pread(vcdev->vdev.fd, region, vcdev->io_region_size,
 155                  vcdev->io_region_offset);
 156     if (size == -1) {
 157         switch (errno) {
 158         case ENODEV:
 159             /* Generate a deferred cc 3 condition. */
 160             s->flags |= SCSW_FLAGS_MASK_CC;
 161             s->ctrl &= ~SCSW_CTRL_MASK_STCTL;
 162             s->ctrl |= (SCSW_STCTL_ALERT | SCSW_STCTL_STATUS_PEND);
 163             goto read_err;
 164         case EFAULT:
 165             /* Memory problem, generate channel data check. */
 166             s->ctrl &= ~SCSW_ACTL_START_PEND;
 167             s->cstat = SCSW_CSTAT_DATA_CHECK;
 168             s->ctrl &= ~SCSW_CTRL_MASK_STCTL;
 169             s->ctrl |= SCSW_STCTL_PRIMARY | SCSW_STCTL_SECONDARY |
 170                        SCSW_STCTL_ALERT | SCSW_STCTL_STATUS_PEND;
 171             goto read_err;
 172         default:
 173             /* Error, generate channel program check. */
 174             s->ctrl &= ~SCSW_ACTL_START_PEND;
 175             s->cstat = SCSW_CSTAT_PROG_CHECK;
 176             s->ctrl &= ~SCSW_CTRL_MASK_STCTL;
 177             s->ctrl |= SCSW_STCTL_PRIMARY | SCSW_STCTL_SECONDARY |
 178                        SCSW_STCTL_ALERT | SCSW_STCTL_STATUS_PEND;
 179             goto read_err;
 180         }
 181     } else if (size != vcdev->io_region_size) {
 182         /* Information transfer error, generate channel-control check. */
 183         s->ctrl &= ~SCSW_ACTL_START_PEND;
 184         s->cstat = SCSW_CSTAT_CHN_CTRL_CHK;
 185         s->ctrl &= ~SCSW_CTRL_MASK_STCTL;
 186         s->ctrl |= SCSW_STCTL_PRIMARY | SCSW_STCTL_SECONDARY |
 187                    SCSW_STCTL_ALERT | SCSW_STCTL_STATUS_PEND;
 188         goto read_err;
 189     }
 190
 191     memcpy(&irb, region->irb_area, sizeof(IRB));
 192
 193     /* Update control block via irb. */
 194     copy_scsw_to_guest(s, &irb.scsw);
 195
 196     /* If a uint check is pending, copy sense data. */
 197     if ((s->dstat & SCSW_DSTAT_UNIT_CHECK) &&
 198         (p->chars & PMCW_CHARS_MASK_CSENSE)) {
 199         memcpy(sch->sense_data, irb.ecw, sizeof(irb.ecw));
 200     }
 201
 202 read_err:
 203     css_inject_io_interrupt(sch);
 204 }
 205
 206 static void vfio_ccw_register_io_notifier(VFIOCCWDevice *vcdev, Error **errp)
 207 {
 208     VFIODevice *vdev = &vcdev->vdev;
 209     struct vfio_irq_info *irq_info;
 210     struct vfio_irq_set *irq_set;
 211     size_t argsz;
 212     int32_t *pfd;
 213
 214     if (vdev->num_irqs < VFIO_CCW_IO_IRQ_INDEX + 1) {
 215         error_setg(errp, "vfio: unexpected number of io irqs %u",
 216                    vdev->num_irqs);
 217         return;
 218     }
 219
 220     argsz = sizeof(*irq_info);
 221     irq_info = g_malloc0(argsz);
 222     irq_info->index = VFIO_CCW_IO_IRQ_INDEX;
 223     irq_info->argsz = argsz;
 224     if (ioctl(vdev->fd, VFIO_DEVICE_GET_IRQ_INFO,
 225               irq_info) < 0 || irq_info->count < 1) {
 226         error_setg_errno(errp, errno, "vfio: Error getting irq info");
 227         goto out_free_info;
 228     }
 229
 230     if (event_notifier_init(&vcdev->io_notifier, 0)) {
 231         error_setg_errno(errp, errno,
 232                          "vfio: Unable to init event notifier for IO");
 233         goto out_free_info;
 234     }
 235
 236     argsz = sizeof(*irq_set) + sizeof(*pfd);
 237     irq_set = g_malloc0(argsz);
 238     irq_set->argsz = argsz;
 239     irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD |
 240                      VFIO_IRQ_SET_ACTION_TRIGGER;
 241     irq_set->index = VFIO_CCW_IO_IRQ_INDEX;
 242     irq_set->start = 0;
 243     irq_set->count = 1;
 244     pfd = (int32_t *) &irq_set->data;
 245
 246     *pfd = event_notifier_get_fd(&vcdev->io_notifier);
 247     qemu_set_fd_handler(*pfd, vfio_ccw_io_notifier_handler, NULL, vcdev);
 248     if (ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, irq_set)) {
 249         error_setg(errp, "vfio: Failed to set up io notification");
 250         qemu_set_fd_handler(*pfd, NULL, NULL, vcdev);
 251         event_notifier_cleanup(&vcdev->io_notifier);
 252     }
 253
 254     g_free(irq_set);
 255
 256 out_free_info:
 257     g_free(irq_info);
 258 }
 259
 260 static void vfio_ccw_unregister_io_notifier(VFIOCCWDevice *vcdev)
 261 {
 262     struct vfio_irq_set *irq_set;
 263     size_t argsz;
 264     int32_t *pfd;
 265
 266     argsz = sizeof(*irq_set) + sizeof(*pfd);
 267     irq_set = g_malloc0(argsz);
 268     irq_set->argsz = argsz;
 269     irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD |
 270                      VFIO_IRQ_SET_ACTION_TRIGGER;
 271     irq_set->index = VFIO_CCW_IO_IRQ_INDEX;
 272     irq_set->start = 0;
 273     irq_set->count = 1;
 274     pfd = (int32_t *) &irq_set->data;
 275     *pfd = -1;
 276
 277     if (ioctl(vcdev->vdev.fd, VFIO_DEVICE_SET_IRQS, irq_set)) {
 278         error_report("vfio: Failed to de-assign device io fd: %m");
 279     }
 280
 281     qemu_set_fd_handler(event_notifier_get_fd(&vcdev->io_notifier),
 282                         NULL, NULL, vcdev);
 283     event_notifier_cleanup(&vcdev->io_notifier);
 284
 285     g_free(irq_set);
 286 }
 287
 288 static void vfio_ccw_get_region(VFIOCCWDevice *vcdev, Error **errp)
 289 {
 290     VFIODevice *vdev = &vcdev->vdev;
 291     struct vfio_region_info *info;
 292     int ret;
 293
 294     /* Sanity check device */
 295     if (!(vdev->flags & VFIO_DEVICE_FLAGS_CCW)) {
 296         error_setg(errp, "vfio: Um, this isn't a vfio-ccw device");
 297         return;
 298     }
 299
 300     if (vdev->num_regions < VFIO_CCW_CONFIG_REGION_INDEX + 1) {
 301         error_setg(errp, "vfio: Unexpected number of the I/O region %u",
 302                    vdev->num_regions);
 303         return;
 304     }
 305
 306     ret = vfio_get_region_info(vdev, VFIO_CCW_CONFIG_REGION_INDEX, &info);
 307     if (ret) {
 308         error_setg_errno(errp, -ret, "vfio: Error getting config info");
 309         return;
 310     }
 311
 312     vcdev->io_region_size = info->size;
 313     if (sizeof(*vcdev->io_region) != vcdev->io_region_size) {
 314         error_setg(errp, "vfio: Unexpected size of the I/O region");
 315         g_free(info);
 316         return;
 317     }
 318
 319     vcdev->io_region_offset = info->offset;
 320     vcdev->io_region = g_malloc0(info->size);
 321
 322     g_free(info);
 323 }
 324
 325 static void vfio_ccw_put_region(VFIOCCWDevice *vcdev)
 326 {
 327     g_free(vcdev->io_region);
 328 }
 329
 330 static void vfio_ccw_put_device(VFIOCCWDevice *vcdev)
 331 {
 332     g_free(vcdev->vdev.name);
 333     vfio_put_base_device(&vcdev->vdev);
 334 }
 335
 336 static void vfio_ccw_get_device(VFIOGroup *group, VFIOCCWDevice *vcdev,
 337                                 Error **errp)
 338 {
 339     char *name = g_strdup_printf("%x.%x.%04x", vcdev->cdev.hostid.cssid,
 340                                  vcdev->cdev.hostid.ssid,
 341                                  vcdev->cdev.hostid.devid);
 342     VFIODevice *vbasedev;
 343
 344     QLIST_FOREACH(vbasedev, &group->device_list, next) {
 345         if (strcmp(vbasedev->name, name) == 0) {
 346             error_setg(errp, "vfio: subchannel %s has already been attached",
 347                        name);
 348             goto out_err;
 349         }
 350     }
 351
 352     /*
 353      * All vfio-ccw devices are believed to operate in a way compatible with
 354      * memory ballooning, ie. pages pinned in the host are in the current
 355      * working set of the guest driver and therefore never overlap with pages
 356      * available to the guest balloon driver.  This needs to be set before
 357      * vfio_get_device() for vfio common to handle the balloon inhibitor.
 358      */
 359     vcdev->vdev.balloon_allowed = true;
 360
 361     if (vfio_get_device(group, vcdev->cdev.mdevid, &vcdev->vdev, errp)) {
 362         goto out_err;
 363     }
 364
 365     vcdev->vdev.ops = &vfio_ccw_ops;
 366     vcdev->vdev.type = VFIO_DEVICE_TYPE_CCW;
 367     vcdev->vdev.name = name;
 368     vcdev->vdev.dev = &vcdev->cdev.parent_obj.parent_obj;
 369
 370     return;
 371
 372 out_err:
 373     g_free(name);
 374 }
 375
 376 static VFIOGroup *vfio_ccw_get_group(S390CCWDevice *cdev, Error **errp)
 377 {
 378     char *tmp, group_path[PATH_MAX];
 379     ssize_t len;
 380     int groupid;
 381
 382     tmp = g_strdup_printf("/sys/bus/css/devices/%x.%x.%04x/%s/iommu_group",
 383                           cdev->hostid.cssid, cdev->hostid.ssid,
 384                           cdev->hostid.devid, cdev->mdevid);
 385     len = readlink(tmp, group_path, sizeof(group_path));
 386     g_free(tmp);
 387
 388     if (len <= 0 || len >= sizeof(group_path)) {
 389         error_setg(errp, "vfio: no iommu_group found");
 390         return NULL;
 391     }
 392
 393     group_path[len] = 0;
 394
 395     if (sscanf(basename(group_path), "%d", &groupid) != 1) {
 396         error_setg(errp, "vfio: failed to read %s", group_path);
 397         return NULL;
 398     }
 399
 400     return vfio_get_group(groupid, &address_space_memory, errp);
 401 }
 402
 403 static void vfio_ccw_realize(DeviceState *dev, Error **errp)
 404 {
 405     VFIOGroup *group;
 406     CcwDevice *ccw_dev = DO_UPCAST(CcwDevice, parent_obj, dev);
 407     S390CCWDevice *cdev = DO_UPCAST(S390CCWDevice, parent_obj, ccw_dev);
 408     VFIOCCWDevice *vcdev = DO_UPCAST(VFIOCCWDevice, cdev, cdev);
 409     S390CCWDeviceClass *cdc = S390_CCW_DEVICE_GET_CLASS(cdev);
 410     Error *err = NULL;
 411
 412     /* Call the class init function for subchannel. */
 413     if (cdc->realize) {
 414         cdc->realize(cdev, vcdev->vdev.sysfsdev, &err);
 415         if (err) {
 416             goto out_err_propagate;
 417         }
 418     }
 419
 420     group = vfio_ccw_get_group(cdev, &err);
 421     if (!group) {
 422         goto out_group_err;
 423     }
 424
 425     vfio_ccw_get_device(group, vcdev, &err);
 426     if (err) {
 427         goto out_device_err;
 428     }
 429
 430     vfio_ccw_get_region(vcdev, &err);
 431     if (err) {
 432         goto out_region_err;
 433     }
 434
 435     vfio_ccw_register_io_notifier(vcdev, &err);
 436     if (err) {
 437         goto out_notifier_err;
 438     }
 439
 440     return;
 441
 442 out_notifier_err:
 443     vfio_ccw_put_region(vcdev);
 444 out_region_err:
 445     vfio_ccw_put_device(vcdev);
 446 out_device_err:
 447     vfio_put_group(group);
 448 out_group_err:
 449     if (cdc->unrealize) {
 450         cdc->unrealize(cdev, NULL);
 451     }
 452 out_err_propagate:
 453     error_propagate(errp, err);
 454 }
 455
 456 static void vfio_ccw_unrealize(DeviceState *dev, Error **errp)
 457 {
 458     CcwDevice *ccw_dev = DO_UPCAST(CcwDevice, parent_obj, dev);
 459     S390CCWDevice *cdev = DO_UPCAST(S390CCWDevice, parent_obj, ccw_dev);
 460     VFIOCCWDevice *vcdev = DO_UPCAST(VFIOCCWDevice, cdev, cdev);
 461     S390CCWDeviceClass *cdc = S390_CCW_DEVICE_GET_CLASS(cdev);
 462     VFIOGroup *group = vcdev->vdev.group;
 463
 464     vfio_ccw_unregister_io_notifier(vcdev);
 465     vfio_ccw_put_region(vcdev);
 466     vfio_ccw_put_device(vcdev);
 467     vfio_put_group(group);
 468
 469     if (cdc->unrealize) {
 470         cdc->unrealize(cdev, errp);
 471     }
 472 }
 473
 474 static Property vfio_ccw_properties[] = {
 475     DEFINE_PROP_STRING("sysfsdev", VFIOCCWDevice, vdev.sysfsdev),
 476     DEFINE_PROP_BOOL("force-orb-pfch", VFIOCCWDevice, force_orb_pfch, false),
 477     DEFINE_PROP_END_OF_LIST(),
 478 };
 479
 480 static const VMStateDescription vfio_ccw_vmstate = {
 481     .name = TYPE_VFIO_CCW,
 482     .unmigratable = 1,
 483 };
 484
 485 static void vfio_ccw_class_init(ObjectClass *klass, void *data)
 486 {
 487     DeviceClass *dc = DEVICE_CLASS(klass);
 488     S390CCWDeviceClass *cdc = S390_CCW_DEVICE_CLASS(klass);
 489
 490     dc->props = vfio_ccw_properties;
 491     dc->vmsd = &vfio_ccw_vmstate;
 492     dc->desc = "VFIO-based subchannel assignment";
 493     set_bit(DEVICE_CATEGORY_MISC, dc->categories);
 494     dc->realize = vfio_ccw_realize;
 495     dc->unrealize = vfio_ccw_unrealize;
 496     dc->reset = vfio_ccw_reset;
 497
 498     cdc->handle_request = vfio_ccw_handle_request;
 499 }
 500
 501 static const TypeInfo vfio_ccw_info = {
 502     .name = TYPE_VFIO_CCW,
 503     .parent = TYPE_S390_CCW,
 504     .instance_size = sizeof(VFIOCCWDevice),
 505     .class_init = vfio_ccw_class_init,
 506 };
 507
 508 static void register_vfio_ccw_type(void)
 509 {
 510     type_register_static(&vfio_ccw_info);
 511 }
 512
 513 type_init(register_vfio_ccw_type)