1 // SPDX-License-Identifier: GPL-2.0+ 2 /* 3 * Adjunct processor matrix VFIO device driver callbacks. 4 * 5 * Copyright IBM Corp. 2018 6 * 7 * Author(s): Tony Krowiak <akrowiak@linux.ibm.com> 8 * Halil Pasic <pasic@linux.ibm.com> 9 * Pierre Morel <pmorel@linux.ibm.com> 10 */ 11 #include <linux/string.h> 12 #include <linux/vfio.h> 13 #include <linux/device.h> 14 #include <linux/list.h> 15 #include <linux/ctype.h> 16 #include <linux/bitops.h> 17 #include <linux/kvm_host.h> 18 #include <linux/module.h> 19 #include <linux/uuid.h> 20 #include <asm/kvm.h> 21 #include <asm/zcrypt.h> 22 23 #include "vfio_ap_private.h" 24 #include "vfio_ap_debug.h" 25 26 #define VFIO_AP_MDEV_TYPE_HWVIRT "passthrough" 27 #define VFIO_AP_MDEV_NAME_HWVIRT "VFIO AP Passthrough Device" 28 29 #define AP_QUEUE_ASSIGNED "assigned" 30 #define AP_QUEUE_UNASSIGNED "unassigned" 31 #define AP_QUEUE_IN_USE "in use" 32 33 #define AP_RESET_INTERVAL 20 /* Reset sleep interval (20ms) */ 34 35 static int vfio_ap_mdev_reset_queues(struct ap_matrix_mdev *matrix_mdev); 36 static int vfio_ap_mdev_reset_qlist(struct list_head *qlist); 37 static struct vfio_ap_queue *vfio_ap_find_queue(int apqn); 38 static const struct vfio_device_ops vfio_ap_matrix_dev_ops; 39 static void vfio_ap_mdev_reset_queue(struct vfio_ap_queue *q); 40 41 /** 42 * get_update_locks_for_kvm: Acquire the locks required to dynamically update a 43 * KVM guest's APCB in the proper order. 44 * 45 * @kvm: a pointer to a struct kvm object containing the KVM guest's APCB. 46 * 47 * The proper locking order is: 48 * 1. matrix_dev->guests_lock: required to use the KVM pointer to update a KVM 49 * guest's APCB. 50 * 2. kvm->lock: required to update a guest's APCB 51 * 3. kvm->arch.crypto.pqap_hook_rwsem: required to update pqap_hook and 52 * serialize against PQAP intercepts 53 * 4. matrix_dev->mdevs_lock: required to access data stored in a matrix_mdev 54 * 55 * Note: If @kvm is NULL, the KVM lock and pqap_hook_rwsem will not be taken. 56 */ 57 static inline void get_update_locks_for_kvm(struct kvm *kvm) 58 { 59 mutex_lock(&matrix_dev->guests_lock); 60 if (kvm) { 61 mutex_lock(&kvm->lock); 62 down_write(&kvm->arch.crypto.pqap_hook_rwsem); 63 } 64 mutex_lock(&matrix_dev->mdevs_lock); 65 } 66 67 /** 68 * release_update_locks_for_kvm: Release the locks used to dynamically update a 69 * KVM guest's APCB in the proper order. 70 * 71 * @kvm: a pointer to a struct kvm object containing the KVM guest's APCB. 72 * 73 * The proper unlocking order is: 74 * 1. matrix_dev->mdevs_lock 75 * 2. kvm->arch.crypto.pqap_hook_rwsem 76 * 3. kvm->lock 77 * 4. matrix_dev->guests_lock 78 * 79 * Note: If @kvm is NULL, the KVM lock and pqap_hook_rwsem will not be released. 80 */ 81 static inline void release_update_locks_for_kvm(struct kvm *kvm) 82 { 83 mutex_unlock(&matrix_dev->mdevs_lock); 84 if (kvm) { 85 up_write(&kvm->arch.crypto.pqap_hook_rwsem); 86 mutex_unlock(&kvm->lock); 87 } 88 mutex_unlock(&matrix_dev->guests_lock); 89 } 90 91 /** 92 * get_update_locks_for_mdev: Acquire the locks required to dynamically update a 93 * KVM guest's APCB in the proper order. 94 * 95 * @matrix_mdev: a pointer to a struct ap_matrix_mdev object containing the AP 96 * configuration data to use to update a KVM guest's APCB. 97 * 98 * The proper locking order is: 99 * 1. matrix_dev->guests_lock: required to use the KVM pointer to update a KVM 100 * guest's APCB. 101 * 2. matrix_mdev->kvm->lock: required to update a guest's APCB 102 * 3. matrix_dev->mdevs_lock: required to access data stored in a matrix_mdev 103 * 104 * Note: If @matrix_mdev is NULL or is not attached to a KVM guest, the KVM 105 * lock will not be taken. 106 */ 107 static inline void get_update_locks_for_mdev(struct ap_matrix_mdev *matrix_mdev) 108 { 109 mutex_lock(&matrix_dev->guests_lock); 110 if (matrix_mdev && matrix_mdev->kvm) 111 mutex_lock(&matrix_mdev->kvm->lock); 112 mutex_lock(&matrix_dev->mdevs_lock); 113 } 114 115 /** 116 * release_update_locks_for_mdev: Release the locks used to dynamically update a 117 * KVM guest's APCB in the proper order. 118 * 119 * @matrix_mdev: a pointer to a struct ap_matrix_mdev object containing the AP 120 * configuration data to use to update a KVM guest's APCB. 121 * 122 * The proper unlocking order is: 123 * 1. matrix_dev->mdevs_lock 124 * 2. matrix_mdev->kvm->lock 125 * 3. matrix_dev->guests_lock 126 * 127 * Note: If @matrix_mdev is NULL or is not attached to a KVM guest, the KVM 128 * lock will not be released. 129 */ 130 static inline void release_update_locks_for_mdev(struct ap_matrix_mdev *matrix_mdev) 131 { 132 mutex_unlock(&matrix_dev->mdevs_lock); 133 if (matrix_mdev && matrix_mdev->kvm) 134 mutex_unlock(&matrix_mdev->kvm->lock); 135 mutex_unlock(&matrix_dev->guests_lock); 136 } 137 138 /** 139 * get_update_locks_by_apqn: Find the mdev to which an APQN is assigned and 140 * acquire the locks required to update the APCB of 141 * the KVM guest to which the mdev is attached. 142 * 143 * @apqn: the APQN of a queue device. 144 * 145 * The proper locking order is: 146 * 1. matrix_dev->guests_lock: required to use the KVM pointer to update a KVM 147 * guest's APCB. 148 * 2. matrix_mdev->kvm->lock: required to update a guest's APCB 149 * 3. matrix_dev->mdevs_lock: required to access data stored in a matrix_mdev 150 * 151 * Note: If @apqn is not assigned to a matrix_mdev, the matrix_mdev->kvm->lock 152 * will not be taken. 153 * 154 * Return: the ap_matrix_mdev object to which @apqn is assigned or NULL if @apqn 155 * is not assigned to an ap_matrix_mdev. 156 */ 157 static struct ap_matrix_mdev *get_update_locks_by_apqn(int apqn) 158 { 159 struct ap_matrix_mdev *matrix_mdev; 160 161 mutex_lock(&matrix_dev->guests_lock); 162 163 list_for_each_entry(matrix_mdev, &matrix_dev->mdev_list, node) { 164 if (test_bit_inv(AP_QID_CARD(apqn), matrix_mdev->matrix.apm) && 165 test_bit_inv(AP_QID_QUEUE(apqn), matrix_mdev->matrix.aqm)) { 166 if (matrix_mdev->kvm) 167 mutex_lock(&matrix_mdev->kvm->lock); 168 169 mutex_lock(&matrix_dev->mdevs_lock); 170 171 return matrix_mdev; 172 } 173 } 174 175 mutex_lock(&matrix_dev->mdevs_lock); 176 177 return NULL; 178 } 179 180 /** 181 * get_update_locks_for_queue: get the locks required to update the APCB of the 182 * KVM guest to which the matrix mdev linked to a 183 * vfio_ap_queue object is attached. 184 * 185 * @q: a pointer to a vfio_ap_queue object. 186 * 187 * The proper locking order is: 188 * 1. q->matrix_dev->guests_lock: required to use the KVM pointer to update a 189 * KVM guest's APCB. 190 * 2. q->matrix_mdev->kvm->lock: required to update a guest's APCB 191 * 3. matrix_dev->mdevs_lock: required to access data stored in matrix_mdev 192 * 193 * Note: if @queue is not linked to an ap_matrix_mdev object, the KVM lock 194 * will not be taken. 195 */ 196 static inline void get_update_locks_for_queue(struct vfio_ap_queue *q) 197 { 198 mutex_lock(&matrix_dev->guests_lock); 199 if (q->matrix_mdev && q->matrix_mdev->kvm) 200 mutex_lock(&q->matrix_mdev->kvm->lock); 201 mutex_lock(&matrix_dev->mdevs_lock); 202 } 203 204 /** 205 * vfio_ap_mdev_get_queue - retrieve a queue with a specific APQN from a 206 * hash table of queues assigned to a matrix mdev 207 * @matrix_mdev: the matrix mdev 208 * @apqn: The APQN of a queue device 209 * 210 * Return: the pointer to the vfio_ap_queue struct representing the queue or 211 * NULL if the queue is not assigned to @matrix_mdev 212 */ 213 static struct vfio_ap_queue *vfio_ap_mdev_get_queue( 214 struct ap_matrix_mdev *matrix_mdev, 215 int apqn) 216 { 217 struct vfio_ap_queue *q; 218 219 hash_for_each_possible(matrix_mdev->qtable.queues, q, mdev_qnode, 220 apqn) { 221 if (q && q->apqn == apqn) 222 return q; 223 } 224 225 return NULL; 226 } 227 228 /** 229 * vfio_ap_wait_for_irqclear - clears the IR bit or gives up after 5 tries 230 * @apqn: The AP Queue number 231 * 232 * Checks the IRQ bit for the status of this APQN using ap_tapq. 233 * Returns if the ap_tapq function succeeded and the bit is clear. 234 * Returns if ap_tapq function failed with invalid, deconfigured or 235 * checkstopped AP. 236 * Otherwise retries up to 5 times after waiting 20ms. 237 */ 238 static void vfio_ap_wait_for_irqclear(int apqn) 239 { 240 struct ap_queue_status status; 241 int retry = 5; 242 243 do { 244 status = ap_tapq(apqn, NULL); 245 switch (status.response_code) { 246 case AP_RESPONSE_NORMAL: 247 case AP_RESPONSE_RESET_IN_PROGRESS: 248 if (!status.irq_enabled) 249 return; 250 fallthrough; 251 case AP_RESPONSE_BUSY: 252 msleep(20); 253 break; 254 case AP_RESPONSE_Q_NOT_AVAIL: 255 case AP_RESPONSE_DECONFIGURED: 256 case AP_RESPONSE_CHECKSTOPPED: 257 default: 258 WARN_ONCE(1, "%s: tapq rc %02x: %04x\n", __func__, 259 status.response_code, apqn); 260 return; 261 } 262 } while (--retry); 263 264 WARN_ONCE(1, "%s: tapq rc %02x: %04x could not clear IR bit\n", 265 __func__, status.response_code, apqn); 266 } 267 268 /** 269 * vfio_ap_free_aqic_resources - free vfio_ap_queue resources 270 * @q: The vfio_ap_queue 271 * 272 * Unregisters the ISC in the GIB when the saved ISC not invalid. 273 * Unpins the guest's page holding the NIB when it exists. 274 * Resets the saved_iova and saved_isc to invalid values. 275 */ 276 static void vfio_ap_free_aqic_resources(struct vfio_ap_queue *q) 277 { 278 if (!q) 279 return; 280 if (q->saved_isc != VFIO_AP_ISC_INVALID) { 281 if (!WARN_ON(!q->matrix_mdev) && q->matrix_mdev->kvm) 282 kvm_s390_gisc_unregister(q->matrix_mdev->kvm, q->saved_isc); 283 q->saved_isc = VFIO_AP_ISC_INVALID; 284 } 285 if (q->saved_iova && !WARN_ON(!q->matrix_mdev)) { 286 vfio_unpin_pages(&q->matrix_mdev->vdev, q->saved_iova, 1); 287 q->saved_iova = 0; 288 } 289 } 290 291 /** 292 * vfio_ap_irq_disable - disables and clears an ap_queue interrupt 293 * @q: The vfio_ap_queue 294 * 295 * Uses ap_aqic to disable the interruption and in case of success, reset 296 * in progress or IRQ disable command already proceeded: calls 297 * vfio_ap_wait_for_irqclear() to check for the IRQ bit to be clear 298 * and calls vfio_ap_free_aqic_resources() to free the resources associated 299 * with the AP interrupt handling. 300 * 301 * In the case the AP is busy, or a reset is in progress, 302 * retries after 20ms, up to 5 times. 303 * 304 * Returns if ap_aqic function failed with invalid, deconfigured or 305 * checkstopped AP. 306 * 307 * Return: &struct ap_queue_status 308 */ 309 static struct ap_queue_status vfio_ap_irq_disable(struct vfio_ap_queue *q) 310 { 311 union ap_qirq_ctrl aqic_gisa = { .value = 0 }; 312 struct ap_queue_status status; 313 int retries = 5; 314 315 do { 316 status = ap_aqic(q->apqn, aqic_gisa, 0); 317 switch (status.response_code) { 318 case AP_RESPONSE_OTHERWISE_CHANGED: 319 case AP_RESPONSE_NORMAL: 320 vfio_ap_wait_for_irqclear(q->apqn); 321 goto end_free; 322 case AP_RESPONSE_RESET_IN_PROGRESS: 323 case AP_RESPONSE_BUSY: 324 msleep(20); 325 break; 326 case AP_RESPONSE_Q_NOT_AVAIL: 327 case AP_RESPONSE_DECONFIGURED: 328 case AP_RESPONSE_CHECKSTOPPED: 329 case AP_RESPONSE_INVALID_ADDRESS: 330 default: 331 /* All cases in default means AP not operational */ 332 WARN_ONCE(1, "%s: ap_aqic status %d\n", __func__, 333 status.response_code); 334 goto end_free; 335 } 336 } while (retries--); 337 338 WARN_ONCE(1, "%s: ap_aqic status %d\n", __func__, 339 status.response_code); 340 end_free: 341 vfio_ap_free_aqic_resources(q); 342 return status; 343 } 344 345 /** 346 * vfio_ap_validate_nib - validate a notification indicator byte (nib) address. 347 * 348 * @vcpu: the object representing the vcpu executing the PQAP(AQIC) instruction. 349 * @nib: the location for storing the nib address. 350 * 351 * When the PQAP(AQIC) instruction is executed, general register 2 contains the 352 * address of the notification indicator byte (nib) used for IRQ notification. 353 * This function parses and validates the nib from gr2. 354 * 355 * Return: returns zero if the nib address is a valid; otherwise, returns 356 * -EINVAL. 357 */ 358 static int vfio_ap_validate_nib(struct kvm_vcpu *vcpu, dma_addr_t *nib) 359 { 360 *nib = vcpu->run->s.regs.gprs[2]; 361 362 if (!*nib) 363 return -EINVAL; 364 if (!kvm_s390_is_gpa_in_memslot(vcpu->kvm, *nib)) 365 return -EINVAL; 366 367 return 0; 368 } 369 370 /** 371 * ensure_nib_shared() - Ensure the address of the NIB is secure and shared 372 * @addr: the physical (absolute) address of the NIB 373 * 374 * This function checks whether the NIB page, which has been pinned with 375 * vfio_pin_pages(), is a shared page belonging to a secure guest. 376 * 377 * It will call uv_pin_shared() on it; if the page was already pinned shared 378 * (i.e. if the NIB belongs to a secure guest and is shared), then 0 379 * (success) is returned. If the NIB was not shared, vfio_pin_pages() had 380 * exported it and now it does not belong to the secure guest anymore. In 381 * that case, an error is returned. 382 * 383 * Context: the NIB (at physical address @addr) has to be pinned with 384 * vfio_pin_pages() before calling this function. 385 * 386 * Return: 0 in case of success, otherwise an error < 0. 387 */ 388 static int ensure_nib_shared(unsigned long addr) 389 { 390 /* 391 * The nib has to be located in shared storage since guest and 392 * host access it. vfio_pin_pages() will do a pin shared and 393 * if that fails (possibly because it's not a shared page) it 394 * calls export. We try to do a second pin shared here so that 395 * the UV gives us an error code if we try to pin a non-shared 396 * page. 397 * 398 * If the page is already pinned shared the UV will return a success. 399 */ 400 return uv_pin_shared(addr); 401 } 402 403 /** 404 * vfio_ap_irq_enable - Enable Interruption for a APQN 405 * 406 * @q: the vfio_ap_queue holding AQIC parameters 407 * @isc: the guest ISC to register with the GIB interface 408 * @vcpu: the vcpu object containing the registers specifying the parameters 409 * passed to the PQAP(AQIC) instruction. 410 * 411 * Pin the NIB saved in *q 412 * Register the guest ISC to GIB interface and retrieve the 413 * host ISC to issue the host side PQAP/AQIC 414 * 415 * status.response_code may be set to AP_RESPONSE_INVALID_ADDRESS in case the 416 * vfio_pin_pages or kvm_s390_gisc_register failed. 417 * 418 * Otherwise return the ap_queue_status returned by the ap_aqic(), 419 * all retry handling will be done by the guest. 420 * 421 * Return: &struct ap_queue_status 422 */ 423 static struct ap_queue_status vfio_ap_irq_enable(struct vfio_ap_queue *q, 424 int isc, 425 struct kvm_vcpu *vcpu) 426 { 427 union ap_qirq_ctrl aqic_gisa = { .value = 0 }; 428 struct ap_queue_status status = {}; 429 struct kvm_s390_gisa *gisa; 430 struct page *h_page; 431 int nisc; 432 struct kvm *kvm; 433 phys_addr_t h_nib; 434 dma_addr_t nib; 435 int ret; 436 437 /* Verify that the notification indicator byte address is valid */ 438 if (vfio_ap_validate_nib(vcpu, &nib)) { 439 VFIO_AP_DBF_WARN("%s: invalid NIB address: nib=%pad, apqn=%#04x\n", 440 __func__, &nib, q->apqn); 441 442 status.response_code = AP_RESPONSE_INVALID_ADDRESS; 443 return status; 444 } 445 446 /* The pin will probably be successful even if the NIB was not shared */ 447 ret = vfio_pin_pages(&q->matrix_mdev->vdev, nib, 1, 448 IOMMU_READ | IOMMU_WRITE, &h_page); 449 switch (ret) { 450 case 1: 451 break; 452 default: 453 VFIO_AP_DBF_WARN("%s: vfio_pin_pages failed: rc=%d," 454 "nib=%pad, apqn=%#04x\n", 455 __func__, ret, &nib, q->apqn); 456 457 status.response_code = AP_RESPONSE_INVALID_ADDRESS; 458 return status; 459 } 460 461 kvm = q->matrix_mdev->kvm; 462 gisa = kvm->arch.gisa_int.origin; 463 464 h_nib = page_to_phys(h_page) | (nib & ~PAGE_MASK); 465 aqic_gisa.gisc = isc; 466 467 /* NIB in non-shared storage is a rc 6 for PV guests */ 468 if (kvm_s390_pv_cpu_is_protected(vcpu) && 469 ensure_nib_shared(h_nib & PAGE_MASK)) { 470 vfio_unpin_pages(&q->matrix_mdev->vdev, nib, 1); 471 status.response_code = AP_RESPONSE_INVALID_ADDRESS; 472 return status; 473 } 474 475 nisc = kvm_s390_gisc_register(kvm, isc); 476 if (nisc < 0) { 477 VFIO_AP_DBF_WARN("%s: gisc registration failed: nisc=%d, isc=%d, apqn=%#04x\n", 478 __func__, nisc, isc, q->apqn); 479 480 vfio_unpin_pages(&q->matrix_mdev->vdev, nib, 1); 481 status.response_code = AP_RESPONSE_INVALID_ADDRESS; 482 return status; 483 } 484 485 aqic_gisa.isc = nisc; 486 aqic_gisa.ir = 1; 487 aqic_gisa.gisa = virt_to_phys(gisa) >> 4; 488 489 status = ap_aqic(q->apqn, aqic_gisa, h_nib); 490 switch (status.response_code) { 491 case AP_RESPONSE_NORMAL: 492 /* See if we did clear older IRQ configuration */ 493 vfio_ap_free_aqic_resources(q); 494 q->saved_iova = nib; 495 q->saved_isc = isc; 496 break; 497 case AP_RESPONSE_OTHERWISE_CHANGED: 498 /* We could not modify IRQ settings: clear new configuration */ 499 ret = kvm_s390_gisc_unregister(kvm, isc); 500 if (ret) 501 VFIO_AP_DBF_WARN("%s: kvm_s390_gisc_unregister: rc=%d isc=%d, apqn=%#04x\n", 502 __func__, ret, isc, q->apqn); 503 vfio_unpin_pages(&q->matrix_mdev->vdev, nib, 1); 504 break; 505 default: 506 pr_warn("%s: apqn %04x: response: %02x\n", __func__, q->apqn, 507 status.response_code); 508 vfio_ap_irq_disable(q); 509 break; 510 } 511 512 if (status.response_code != AP_RESPONSE_NORMAL) { 513 VFIO_AP_DBF_WARN("%s: PQAP(AQIC) failed with status=%#02x: " 514 "zone=%#x, ir=%#x, gisc=%#x, f=%#x," 515 "gisa=%#x, isc=%#x, apqn=%#04x\n", 516 __func__, status.response_code, 517 aqic_gisa.zone, aqic_gisa.ir, aqic_gisa.gisc, 518 aqic_gisa.gf, aqic_gisa.gisa, aqic_gisa.isc, 519 q->apqn); 520 } 521 522 return status; 523 } 524 525 /** 526 * vfio_ap_le_guid_to_be_uuid - convert a little endian guid array into an array 527 * of big endian elements that can be passed by 528 * value to an s390dbf sprintf event function to 529 * format a UUID string. 530 * 531 * @guid: the object containing the little endian guid 532 * @uuid: a six-element array of long values that can be passed by value as 533 * arguments for a formatting string specifying a UUID. 534 * 535 * The S390 Debug Feature (s390dbf) allows the use of "%s" in the sprintf 536 * event functions if the memory for the passed string is available as long as 537 * the debug feature exists. Since a mediated device can be removed at any 538 * time, it's name can not be used because %s passes the reference to the string 539 * in memory and the reference will go stale once the device is removed . 540 * 541 * The s390dbf string formatting function allows a maximum of 9 arguments for a 542 * message to be displayed in the 'sprintf' view. In order to use the bytes 543 * comprising the mediated device's UUID to display the mediated device name, 544 * they will have to be converted into an array whose elements can be passed by 545 * value to sprintf. For example: 546 * 547 * guid array: { 83, 78, 17, 62, bb, f1, f0, 47, 91, 4d, 32, a2, 2e, 3a, 88, 04 } 548 * mdev name: 62177883-f1bb-47f0-914d-32a22e3a8804 549 * array returned: { 62177883, f1bb, 47f0, 914d, 32a2, 2e3a8804 } 550 * formatting string: "%08lx-%04lx-%04lx-%04lx-%02lx%04lx" 551 */ 552 static void vfio_ap_le_guid_to_be_uuid(guid_t *guid, unsigned long *uuid) 553 { 554 /* 555 * The input guid is ordered in little endian, so it needs to be 556 * reordered for displaying a UUID as a string. This specifies the 557 * guid indices in proper order. 558 */ 559 uuid[0] = le32_to_cpup((__le32 *)guid); 560 uuid[1] = le16_to_cpup((__le16 *)&guid->b[4]); 561 uuid[2] = le16_to_cpup((__le16 *)&guid->b[6]); 562 uuid[3] = *((__u16 *)&guid->b[8]); 563 uuid[4] = *((__u16 *)&guid->b[10]); 564 uuid[5] = *((__u32 *)&guid->b[12]); 565 } 566 567 /** 568 * handle_pqap - PQAP instruction callback 569 * 570 * @vcpu: The vcpu on which we received the PQAP instruction 571 * 572 * Get the general register contents to initialize internal variables. 573 * REG[0]: APQN 574 * REG[1]: IR and ISC 575 * REG[2]: NIB 576 * 577 * Response.status may be set to following Response Code: 578 * - AP_RESPONSE_Q_NOT_AVAIL: if the queue is not available 579 * - AP_RESPONSE_DECONFIGURED: if the queue is not configured 580 * - AP_RESPONSE_NORMAL (0) : in case of success 581 * Check vfio_ap_setirq() and vfio_ap_clrirq() for other possible RC. 582 * We take the matrix_dev lock to ensure serialization on queues and 583 * mediated device access. 584 * 585 * Return: 0 if we could handle the request inside KVM. 586 * Otherwise, returns -EOPNOTSUPP to let QEMU handle the fault. 587 */ 588 static int handle_pqap(struct kvm_vcpu *vcpu) 589 { 590 uint64_t status; 591 uint16_t apqn; 592 unsigned long uuid[6]; 593 struct vfio_ap_queue *q; 594 struct ap_queue_status qstatus = { 595 .response_code = AP_RESPONSE_Q_NOT_AVAIL, }; 596 struct ap_matrix_mdev *matrix_mdev; 597 598 apqn = vcpu->run->s.regs.gprs[0] & 0xffff; 599 600 /* If we do not use the AIV facility just go to userland */ 601 if (!(vcpu->arch.sie_block->eca & ECA_AIV)) { 602 VFIO_AP_DBF_WARN("%s: AIV facility not installed: apqn=0x%04x, eca=0x%04x\n", 603 __func__, apqn, vcpu->arch.sie_block->eca); 604 605 return -EOPNOTSUPP; 606 } 607 608 mutex_lock(&matrix_dev->mdevs_lock); 609 610 if (!vcpu->kvm->arch.crypto.pqap_hook) { 611 VFIO_AP_DBF_WARN("%s: PQAP(AQIC) hook not registered with the vfio_ap driver: apqn=0x%04x\n", 612 __func__, apqn); 613 614 goto out_unlock; 615 } 616 617 matrix_mdev = container_of(vcpu->kvm->arch.crypto.pqap_hook, 618 struct ap_matrix_mdev, pqap_hook); 619 620 /* If the there is no guest using the mdev, there is nothing to do */ 621 if (!matrix_mdev->kvm) { 622 vfio_ap_le_guid_to_be_uuid(&matrix_mdev->mdev->uuid, uuid); 623 VFIO_AP_DBF_WARN("%s: mdev %08lx-%04lx-%04lx-%04lx-%04lx%08lx not in use: apqn=0x%04x\n", 624 __func__, uuid[0], uuid[1], uuid[2], 625 uuid[3], uuid[4], uuid[5], apqn); 626 goto out_unlock; 627 } 628 629 q = vfio_ap_mdev_get_queue(matrix_mdev, apqn); 630 if (!q) { 631 VFIO_AP_DBF_WARN("%s: Queue %02x.%04x not bound to the vfio_ap driver\n", 632 __func__, AP_QID_CARD(apqn), 633 AP_QID_QUEUE(apqn)); 634 goto out_unlock; 635 } 636 637 status = vcpu->run->s.regs.gprs[1]; 638 639 /* If IR bit(16) is set we enable the interrupt */ 640 if ((status >> (63 - 16)) & 0x01) 641 qstatus = vfio_ap_irq_enable(q, status & 0x07, vcpu); 642 else 643 qstatus = vfio_ap_irq_disable(q); 644 645 out_unlock: 646 memcpy(&vcpu->run->s.regs.gprs[1], &qstatus, sizeof(qstatus)); 647 vcpu->run->s.regs.gprs[1] >>= 32; 648 mutex_unlock(&matrix_dev->mdevs_lock); 649 return 0; 650 } 651 652 static void vfio_ap_matrix_init(struct ap_config_info *info, 653 struct ap_matrix *matrix) 654 { 655 matrix->apm_max = info->apxa ? info->na : 63; 656 matrix->aqm_max = info->apxa ? info->nd : 15; 657 matrix->adm_max = info->apxa ? info->nd : 15; 658 } 659 660 static void signal_guest_ap_cfg_changed(struct ap_matrix_mdev *matrix_mdev) 661 { 662 if (matrix_mdev->cfg_chg_trigger) 663 eventfd_signal(matrix_mdev->cfg_chg_trigger); 664 } 665 666 static void vfio_ap_mdev_update_guest_apcb(struct ap_matrix_mdev *matrix_mdev) 667 { 668 if (matrix_mdev->kvm) { 669 kvm_arch_crypto_set_masks(matrix_mdev->kvm, 670 matrix_mdev->shadow_apcb.apm, 671 matrix_mdev->shadow_apcb.aqm, 672 matrix_mdev->shadow_apcb.adm); 673 674 signal_guest_ap_cfg_changed(matrix_mdev); 675 } 676 } 677 678 static bool vfio_ap_mdev_filter_cdoms(struct ap_matrix_mdev *matrix_mdev) 679 { 680 DECLARE_BITMAP(prev_shadow_adm, AP_DOMAINS); 681 682 bitmap_copy(prev_shadow_adm, matrix_mdev->shadow_apcb.adm, AP_DOMAINS); 683 bitmap_and(matrix_mdev->shadow_apcb.adm, matrix_mdev->matrix.adm, 684 (unsigned long *)matrix_dev->info.adm, AP_DOMAINS); 685 686 return !bitmap_equal(prev_shadow_adm, matrix_mdev->shadow_apcb.adm, 687 AP_DOMAINS); 688 } 689 690 static bool _queue_passable(struct vfio_ap_queue *q) 691 { 692 if (!q) 693 return false; 694 695 switch (q->reset_status.response_code) { 696 case AP_RESPONSE_NORMAL: 697 case AP_RESPONSE_DECONFIGURED: 698 case AP_RESPONSE_CHECKSTOPPED: 699 return true; 700 default: 701 return false; 702 } 703 } 704 705 /* 706 * vfio_ap_mdev_filter_matrix - filter the APQNs assigned to the matrix mdev 707 * to ensure no queue devices are passed through to 708 * the guest that are not bound to the vfio_ap 709 * device driver. 710 * 711 * @matrix_mdev: the matrix mdev whose matrix is to be filtered. 712 * @apm_filtered: a 256-bit bitmap for storing the APIDs filtered from the 713 * guest's AP configuration that are still in the host's AP 714 * configuration. 715 * 716 * Note: If an APQN referencing a queue device that is not bound to the vfio_ap 717 * driver, its APID will be filtered from the guest's APCB. The matrix 718 * structure precludes filtering an individual APQN, so its APID will be 719 * filtered. Consequently, all queues associated with the adapter that 720 * are in the host's AP configuration must be reset. If queues are 721 * subsequently made available again to the guest, they should re-appear 722 * in a reset state 723 * 724 * Return: a boolean value indicating whether the KVM guest's APCB was changed 725 * by the filtering or not. 726 */ 727 static bool vfio_ap_mdev_filter_matrix(struct ap_matrix_mdev *matrix_mdev, 728 unsigned long *apm_filtered) 729 { 730 unsigned long apid, apqi, apqn; 731 DECLARE_BITMAP(prev_shadow_apm, AP_DEVICES); 732 DECLARE_BITMAP(prev_shadow_aqm, AP_DOMAINS); 733 734 bitmap_copy(prev_shadow_apm, matrix_mdev->shadow_apcb.apm, AP_DEVICES); 735 bitmap_copy(prev_shadow_aqm, matrix_mdev->shadow_apcb.aqm, AP_DOMAINS); 736 vfio_ap_matrix_init(&matrix_dev->info, &matrix_mdev->shadow_apcb); 737 bitmap_clear(apm_filtered, 0, AP_DEVICES); 738 739 /* 740 * Copy the adapters, domains and control domains to the shadow_apcb 741 * from the matrix mdev, but only those that are assigned to the host's 742 * AP configuration. 743 */ 744 bitmap_and(matrix_mdev->shadow_apcb.apm, matrix_mdev->matrix.apm, 745 (unsigned long *)matrix_dev->info.apm, AP_DEVICES); 746 bitmap_and(matrix_mdev->shadow_apcb.aqm, matrix_mdev->matrix.aqm, 747 (unsigned long *)matrix_dev->info.aqm, AP_DOMAINS); 748 749 for_each_set_bit_inv(apid, matrix_mdev->shadow_apcb.apm, AP_DEVICES) { 750 for_each_set_bit_inv(apqi, matrix_mdev->shadow_apcb.aqm, 751 AP_DOMAINS) { 752 /* 753 * If the APQN is not bound to the vfio_ap device 754 * driver, then we can't assign it to the guest's 755 * AP configuration. The AP architecture won't 756 * allow filtering of a single APQN, so let's filter 757 * the APID since an adapter represents a physical 758 * hardware device. 759 */ 760 apqn = AP_MKQID(apid, apqi); 761 if (!_queue_passable(vfio_ap_mdev_get_queue(matrix_mdev, apqn))) { 762 clear_bit_inv(apid, matrix_mdev->shadow_apcb.apm); 763 764 /* 765 * If the adapter was previously plugged into 766 * the guest, let's let the caller know that 767 * the APID was filtered. 768 */ 769 if (test_bit_inv(apid, prev_shadow_apm)) 770 set_bit_inv(apid, apm_filtered); 771 772 break; 773 } 774 } 775 } 776 777 return !bitmap_equal(prev_shadow_apm, matrix_mdev->shadow_apcb.apm, 778 AP_DEVICES) || 779 !bitmap_equal(prev_shadow_aqm, matrix_mdev->shadow_apcb.aqm, 780 AP_DOMAINS); 781 } 782 783 static int vfio_ap_mdev_init_dev(struct vfio_device *vdev) 784 { 785 struct ap_matrix_mdev *matrix_mdev = 786 container_of(vdev, struct ap_matrix_mdev, vdev); 787 788 matrix_mdev->mdev = to_mdev_device(vdev->dev); 789 vfio_ap_matrix_init(&matrix_dev->info, &matrix_mdev->matrix); 790 matrix_mdev->pqap_hook = handle_pqap; 791 vfio_ap_matrix_init(&matrix_dev->info, &matrix_mdev->shadow_apcb); 792 hash_init(matrix_mdev->qtable.queues); 793 794 return 0; 795 } 796 797 static int vfio_ap_mdev_probe(struct mdev_device *mdev) 798 { 799 struct ap_matrix_mdev *matrix_mdev; 800 int ret; 801 802 matrix_mdev = vfio_alloc_device(ap_matrix_mdev, vdev, &mdev->dev, 803 &vfio_ap_matrix_dev_ops); 804 if (IS_ERR(matrix_mdev)) 805 return PTR_ERR(matrix_mdev); 806 807 ret = vfio_register_emulated_iommu_dev(&matrix_mdev->vdev); 808 if (ret) 809 goto err_put_vdev; 810 811 /* 812 * Take the matrix_dev->guests_lock mutex before adding the matrix_mdev 813 * to the mdev_list. All functions that traverse the list must also hold 814 * this lock to guard against additions to or removals from the list 815 * while it is being traversed. 816 */ 817 mutex_lock(&matrix_dev->guests_lock); 818 dev_set_drvdata(&mdev->dev, matrix_mdev); 819 list_add(&matrix_mdev->node, &matrix_dev->mdev_list); 820 mutex_unlock(&matrix_dev->guests_lock); 821 return 0; 822 823 err_put_vdev: 824 vfio_put_device(&matrix_mdev->vdev); 825 return ret; 826 } 827 828 static void vfio_ap_mdev_link_queue(struct ap_matrix_mdev *matrix_mdev, 829 struct vfio_ap_queue *q) 830 { 831 if (!q || vfio_ap_mdev_get_queue(matrix_mdev, q->apqn)) 832 return; 833 834 q->matrix_mdev = matrix_mdev; 835 hash_add(matrix_mdev->qtable.queues, &q->mdev_qnode, q->apqn); 836 } 837 838 static void vfio_ap_mdev_link_apqn(struct ap_matrix_mdev *matrix_mdev, int apqn) 839 { 840 struct vfio_ap_queue *q; 841 842 q = vfio_ap_find_queue(apqn); 843 vfio_ap_mdev_link_queue(matrix_mdev, q); 844 } 845 846 static void vfio_ap_unlink_queue_fr_mdev(struct vfio_ap_queue *q) 847 { 848 hash_del(&q->mdev_qnode); 849 } 850 851 static void vfio_ap_unlink_mdev_fr_queue(struct vfio_ap_queue *q) 852 { 853 q->matrix_mdev = NULL; 854 } 855 856 static void vfio_ap_mdev_unlink_fr_queues(struct ap_matrix_mdev *matrix_mdev) 857 { 858 struct vfio_ap_queue *q; 859 unsigned long apid, apqi; 860 861 for_each_set_bit_inv(apid, matrix_mdev->matrix.apm, AP_DEVICES) { 862 for_each_set_bit_inv(apqi, matrix_mdev->matrix.aqm, 863 AP_DOMAINS) { 864 q = vfio_ap_mdev_get_queue(matrix_mdev, 865 AP_MKQID(apid, apqi)); 866 if (q) 867 q->matrix_mdev = NULL; 868 } 869 } 870 } 871 872 static void vfio_ap_mdev_remove(struct mdev_device *mdev) 873 { 874 struct ap_matrix_mdev *matrix_mdev = dev_get_drvdata(&mdev->dev); 875 876 vfio_unregister_group_dev(&matrix_mdev->vdev); 877 878 mutex_lock(&matrix_dev->guests_lock); 879 mutex_lock(&matrix_dev->mdevs_lock); 880 vfio_ap_mdev_reset_queues(matrix_mdev); 881 vfio_ap_mdev_unlink_fr_queues(matrix_mdev); 882 list_del(&matrix_mdev->node); 883 mutex_unlock(&matrix_dev->mdevs_lock); 884 mutex_unlock(&matrix_dev->guests_lock); 885 vfio_put_device(&matrix_mdev->vdev); 886 } 887 888 #define MDEV_SHARING_ERR "Userspace may not assign queue %02lx.%04lx to mdev: already assigned to %s" 889 890 #define MDEV_IN_USE_ERR "Can not reserve queue %02lx.%04lx for host driver: in use by mdev" 891 892 static void vfio_ap_mdev_log_sharing_err(struct ap_matrix_mdev *assignee, 893 struct ap_matrix_mdev *assigned_to, 894 unsigned long *apm, unsigned long *aqm) 895 { 896 unsigned long apid, apqi; 897 898 for_each_set_bit_inv(apid, apm, AP_DEVICES) { 899 for_each_set_bit_inv(apqi, aqm, AP_DOMAINS) { 900 dev_warn(mdev_dev(assignee->mdev), MDEV_SHARING_ERR, 901 apid, apqi, dev_name(mdev_dev(assigned_to->mdev))); 902 } 903 } 904 } 905 906 static void vfio_ap_mdev_log_in_use_err(struct ap_matrix_mdev *assignee, 907 unsigned long *apm, unsigned long *aqm) 908 { 909 unsigned long apid, apqi; 910 911 for_each_set_bit_inv(apid, apm, AP_DEVICES) { 912 for_each_set_bit_inv(apqi, aqm, AP_DOMAINS) 913 dev_warn(mdev_dev(assignee->mdev), MDEV_IN_USE_ERR, apid, apqi); 914 } 915 } 916 917 /** 918 * vfio_ap_mdev_verify_no_sharing - verify APQNs are not shared by matrix mdevs 919 * 920 * @assignee: the matrix mdev to which @mdev_apm and @mdev_aqm are being 921 * assigned; or, NULL if this function was called by the AP bus 922 * driver in_use callback to verify none of the APQNs being reserved 923 * for the host device driver are in use by a vfio_ap mediated device 924 * @mdev_apm: mask indicating the APIDs of the APQNs to be verified 925 * @mdev_aqm: mask indicating the APQIs of the APQNs to be verified 926 * 927 * Verifies that each APQN derived from the Cartesian product of APIDs 928 * represented by the bits set in @mdev_apm and the APQIs of the bits set in 929 * @mdev_aqm is not assigned to a mediated device other than the mdev to which 930 * the APQN is being assigned (@assignee). AP queue sharing is not allowed. 931 * 932 * Return: 0 if the APQNs are not shared; otherwise return -EADDRINUSE. 933 */ 934 static int vfio_ap_mdev_verify_no_sharing(struct ap_matrix_mdev *assignee, 935 unsigned long *mdev_apm, 936 unsigned long *mdev_aqm) 937 { 938 struct ap_matrix_mdev *assigned_to; 939 DECLARE_BITMAP(apm, AP_DEVICES); 940 DECLARE_BITMAP(aqm, AP_DOMAINS); 941 942 list_for_each_entry(assigned_to, &matrix_dev->mdev_list, node) { 943 /* 944 * If the mdev to which the mdev_apm and mdev_aqm is being 945 * assigned is the same as the mdev being verified 946 */ 947 if (assignee == assigned_to) 948 continue; 949 950 memset(apm, 0, sizeof(apm)); 951 memset(aqm, 0, sizeof(aqm)); 952 953 /* 954 * We work on full longs, as we can only exclude the leftover 955 * bits in non-inverse order. The leftover is all zeros. 956 */ 957 if (!bitmap_and(apm, mdev_apm, assigned_to->matrix.apm, AP_DEVICES)) 958 continue; 959 960 if (!bitmap_and(aqm, mdev_aqm, assigned_to->matrix.aqm, AP_DOMAINS)) 961 continue; 962 963 if (assignee) 964 vfio_ap_mdev_log_sharing_err(assignee, assigned_to, apm, aqm); 965 else 966 vfio_ap_mdev_log_in_use_err(assigned_to, apm, aqm); 967 968 return -EADDRINUSE; 969 } 970 971 return 0; 972 } 973 974 /** 975 * vfio_ap_mdev_validate_masks - verify that the APQNs assigned to the mdev are 976 * not reserved for the default zcrypt driver and 977 * are not assigned to another mdev. 978 * 979 * @matrix_mdev: the mdev to which the APQNs being validated are assigned. 980 * 981 * Return: One of the following values: 982 * o the error returned from the ap_apqn_in_matrix_owned_by_def_drv() function, 983 * most likely -EBUSY indicating the ap_attr_mutex lock is already held. 984 * o EADDRNOTAVAIL if an APQN assigned to @matrix_mdev is reserved for the 985 * zcrypt default driver. 986 * o EADDRINUSE if an APQN assigned to @matrix_mdev is assigned to another mdev 987 * o A zero indicating validation succeeded. 988 */ 989 static int vfio_ap_mdev_validate_masks(struct ap_matrix_mdev *matrix_mdev) 990 { 991 if (ap_apqn_in_matrix_owned_by_def_drv(matrix_mdev->matrix.apm, 992 matrix_mdev->matrix.aqm)) 993 return -EADDRNOTAVAIL; 994 995 return vfio_ap_mdev_verify_no_sharing(matrix_mdev, 996 matrix_mdev->matrix.apm, 997 matrix_mdev->matrix.aqm); 998 } 999 1000 static void vfio_ap_mdev_link_adapter(struct ap_matrix_mdev *matrix_mdev, 1001 unsigned long apid) 1002 { 1003 unsigned long apqi; 1004 1005 for_each_set_bit_inv(apqi, matrix_mdev->matrix.aqm, AP_DOMAINS) 1006 vfio_ap_mdev_link_apqn(matrix_mdev, 1007 AP_MKQID(apid, apqi)); 1008 } 1009 1010 static void collect_queues_to_reset(struct ap_matrix_mdev *matrix_mdev, 1011 unsigned long apid, 1012 struct list_head *qlist) 1013 { 1014 struct vfio_ap_queue *q; 1015 unsigned long apqi; 1016 1017 for_each_set_bit_inv(apqi, matrix_mdev->shadow_apcb.aqm, AP_DOMAINS) { 1018 q = vfio_ap_mdev_get_queue(matrix_mdev, AP_MKQID(apid, apqi)); 1019 if (q) 1020 list_add_tail(&q->reset_qnode, qlist); 1021 } 1022 } 1023 1024 static void reset_queues_for_apid(struct ap_matrix_mdev *matrix_mdev, 1025 unsigned long apid) 1026 { 1027 struct list_head qlist; 1028 1029 INIT_LIST_HEAD(&qlist); 1030 collect_queues_to_reset(matrix_mdev, apid, &qlist); 1031 vfio_ap_mdev_reset_qlist(&qlist); 1032 } 1033 1034 static int reset_queues_for_apids(struct ap_matrix_mdev *matrix_mdev, 1035 unsigned long *apm_reset) 1036 { 1037 struct list_head qlist; 1038 unsigned long apid; 1039 1040 if (bitmap_empty(apm_reset, AP_DEVICES)) 1041 return 0; 1042 1043 INIT_LIST_HEAD(&qlist); 1044 1045 for_each_set_bit_inv(apid, apm_reset, AP_DEVICES) 1046 collect_queues_to_reset(matrix_mdev, apid, &qlist); 1047 1048 return vfio_ap_mdev_reset_qlist(&qlist); 1049 } 1050 1051 /** 1052 * assign_adapter_store - parses the APID from @buf and sets the 1053 * corresponding bit in the mediated matrix device's APM 1054 * 1055 * @dev: the matrix device 1056 * @attr: the mediated matrix device's assign_adapter attribute 1057 * @buf: a buffer containing the AP adapter number (APID) to 1058 * be assigned 1059 * @count: the number of bytes in @buf 1060 * 1061 * Return: the number of bytes processed if the APID is valid; otherwise, 1062 * returns one of the following errors: 1063 * 1064 * 1. -EINVAL 1065 * The APID is not a valid number 1066 * 1067 * 2. -ENODEV 1068 * The APID exceeds the maximum value configured for the system 1069 * 1070 * 3. -EADDRNOTAVAIL 1071 * An APQN derived from the cross product of the APID being assigned 1072 * and the APQIs previously assigned is not bound to the vfio_ap device 1073 * driver; or, if no APQIs have yet been assigned, the APID is not 1074 * contained in an APQN bound to the vfio_ap device driver. 1075 * 1076 * 4. -EADDRINUSE 1077 * An APQN derived from the cross product of the APID being assigned 1078 * and the APQIs previously assigned is being used by another mediated 1079 * matrix device 1080 * 1081 * 5. -EAGAIN 1082 * A lock required to validate the mdev's AP configuration could not 1083 * be obtained. 1084 */ 1085 static ssize_t assign_adapter_store(struct device *dev, 1086 struct device_attribute *attr, 1087 const char *buf, size_t count) 1088 { 1089 int ret; 1090 unsigned long apid; 1091 DECLARE_BITMAP(apm_filtered, AP_DEVICES); 1092 struct ap_matrix_mdev *matrix_mdev = dev_get_drvdata(dev); 1093 1094 mutex_lock(&ap_attr_mutex); 1095 get_update_locks_for_mdev(matrix_mdev); 1096 1097 ret = kstrtoul(buf, 0, &apid); 1098 if (ret) 1099 goto done; 1100 1101 if (apid > matrix_mdev->matrix.apm_max) { 1102 ret = -ENODEV; 1103 goto done; 1104 } 1105 1106 if (test_bit_inv(apid, matrix_mdev->matrix.apm)) { 1107 ret = count; 1108 goto done; 1109 } 1110 1111 set_bit_inv(apid, matrix_mdev->matrix.apm); 1112 1113 ret = vfio_ap_mdev_validate_masks(matrix_mdev); 1114 if (ret) { 1115 clear_bit_inv(apid, matrix_mdev->matrix.apm); 1116 goto done; 1117 } 1118 1119 vfio_ap_mdev_link_adapter(matrix_mdev, apid); 1120 1121 if (vfio_ap_mdev_filter_matrix(matrix_mdev, apm_filtered)) { 1122 vfio_ap_mdev_update_guest_apcb(matrix_mdev); 1123 reset_queues_for_apids(matrix_mdev, apm_filtered); 1124 } 1125 1126 ret = count; 1127 done: 1128 release_update_locks_for_mdev(matrix_mdev); 1129 mutex_unlock(&ap_attr_mutex); 1130 1131 return ret; 1132 } 1133 static DEVICE_ATTR_WO(assign_adapter); 1134 1135 static struct vfio_ap_queue 1136 *vfio_ap_unlink_apqn_fr_mdev(struct ap_matrix_mdev *matrix_mdev, 1137 unsigned long apid, unsigned long apqi) 1138 { 1139 struct vfio_ap_queue *q = NULL; 1140 1141 q = vfio_ap_mdev_get_queue(matrix_mdev, AP_MKQID(apid, apqi)); 1142 /* If the queue is assigned to the matrix mdev, unlink it. */ 1143 if (q) 1144 vfio_ap_unlink_queue_fr_mdev(q); 1145 1146 return q; 1147 } 1148 1149 /** 1150 * vfio_ap_mdev_unlink_adapter - unlink all queues associated with unassigned 1151 * adapter from the matrix mdev to which the 1152 * adapter was assigned. 1153 * @matrix_mdev: the matrix mediated device to which the adapter was assigned. 1154 * @apid: the APID of the unassigned adapter. 1155 * @qlist: list for storing queues associated with unassigned adapter that 1156 * need to be reset. 1157 */ 1158 static void vfio_ap_mdev_unlink_adapter(struct ap_matrix_mdev *matrix_mdev, 1159 unsigned long apid, 1160 struct list_head *qlist) 1161 { 1162 unsigned long apqi; 1163 struct vfio_ap_queue *q; 1164 1165 for_each_set_bit_inv(apqi, matrix_mdev->matrix.aqm, AP_DOMAINS) { 1166 q = vfio_ap_unlink_apqn_fr_mdev(matrix_mdev, apid, apqi); 1167 1168 if (q && qlist) { 1169 if (test_bit_inv(apid, matrix_mdev->shadow_apcb.apm) && 1170 test_bit_inv(apqi, matrix_mdev->shadow_apcb.aqm)) 1171 list_add_tail(&q->reset_qnode, qlist); 1172 } 1173 } 1174 } 1175 1176 static void vfio_ap_mdev_hot_unplug_adapters(struct ap_matrix_mdev *matrix_mdev, 1177 unsigned long *apids) 1178 { 1179 struct vfio_ap_queue *q, *tmpq; 1180 struct list_head qlist; 1181 unsigned long apid; 1182 bool apcb_update = false; 1183 1184 INIT_LIST_HEAD(&qlist); 1185 1186 for_each_set_bit_inv(apid, apids, AP_DEVICES) { 1187 vfio_ap_mdev_unlink_adapter(matrix_mdev, apid, &qlist); 1188 1189 if (test_bit_inv(apid, matrix_mdev->shadow_apcb.apm)) { 1190 clear_bit_inv(apid, matrix_mdev->shadow_apcb.apm); 1191 apcb_update = true; 1192 } 1193 } 1194 1195 /* Only update apcb if needed to avoid impacting guest */ 1196 if (apcb_update) 1197 vfio_ap_mdev_update_guest_apcb(matrix_mdev); 1198 1199 vfio_ap_mdev_reset_qlist(&qlist); 1200 1201 list_for_each_entry_safe(q, tmpq, &qlist, reset_qnode) { 1202 vfio_ap_unlink_mdev_fr_queue(q); 1203 list_del(&q->reset_qnode); 1204 } 1205 } 1206 1207 static void vfio_ap_mdev_hot_unplug_adapter(struct ap_matrix_mdev *matrix_mdev, 1208 unsigned long apid) 1209 { 1210 DECLARE_BITMAP(apids, AP_DEVICES); 1211 1212 bitmap_zero(apids, AP_DEVICES); 1213 set_bit_inv(apid, apids); 1214 vfio_ap_mdev_hot_unplug_adapters(matrix_mdev, apids); 1215 } 1216 1217 /** 1218 * unassign_adapter_store - parses the APID from @buf and clears the 1219 * corresponding bit in the mediated matrix device's APM 1220 * 1221 * @dev: the matrix device 1222 * @attr: the mediated matrix device's unassign_adapter attribute 1223 * @buf: a buffer containing the adapter number (APID) to be unassigned 1224 * @count: the number of bytes in @buf 1225 * 1226 * Return: the number of bytes processed if the APID is valid; otherwise, 1227 * returns one of the following errors: 1228 * -EINVAL if the APID is not a number 1229 * -ENODEV if the APID it exceeds the maximum value configured for the 1230 * system 1231 */ 1232 static ssize_t unassign_adapter_store(struct device *dev, 1233 struct device_attribute *attr, 1234 const char *buf, size_t count) 1235 { 1236 int ret; 1237 unsigned long apid; 1238 struct ap_matrix_mdev *matrix_mdev = dev_get_drvdata(dev); 1239 1240 get_update_locks_for_mdev(matrix_mdev); 1241 1242 ret = kstrtoul(buf, 0, &apid); 1243 if (ret) 1244 goto done; 1245 1246 if (apid > matrix_mdev->matrix.apm_max) { 1247 ret = -ENODEV; 1248 goto done; 1249 } 1250 1251 if (!test_bit_inv(apid, matrix_mdev->matrix.apm)) { 1252 ret = count; 1253 goto done; 1254 } 1255 1256 clear_bit_inv((unsigned long)apid, matrix_mdev->matrix.apm); 1257 vfio_ap_mdev_hot_unplug_adapter(matrix_mdev, apid); 1258 ret = count; 1259 done: 1260 release_update_locks_for_mdev(matrix_mdev); 1261 return ret; 1262 } 1263 static DEVICE_ATTR_WO(unassign_adapter); 1264 1265 static void vfio_ap_mdev_link_domain(struct ap_matrix_mdev *matrix_mdev, 1266 unsigned long apqi) 1267 { 1268 unsigned long apid; 1269 1270 for_each_set_bit_inv(apid, matrix_mdev->matrix.apm, AP_DEVICES) 1271 vfio_ap_mdev_link_apqn(matrix_mdev, 1272 AP_MKQID(apid, apqi)); 1273 } 1274 1275 /** 1276 * assign_domain_store - parses the APQI from @buf and sets the 1277 * corresponding bit in the mediated matrix device's AQM 1278 * 1279 * @dev: the matrix device 1280 * @attr: the mediated matrix device's assign_domain attribute 1281 * @buf: a buffer containing the AP queue index (APQI) of the domain to 1282 * be assigned 1283 * @count: the number of bytes in @buf 1284 * 1285 * Return: the number of bytes processed if the APQI is valid; otherwise returns 1286 * one of the following errors: 1287 * 1288 * 1. -EINVAL 1289 * The APQI is not a valid number 1290 * 1291 * 2. -ENODEV 1292 * The APQI exceeds the maximum value configured for the system 1293 * 1294 * 3. -EADDRNOTAVAIL 1295 * An APQN derived from the cross product of the APQI being assigned 1296 * and the APIDs previously assigned is not bound to the vfio_ap device 1297 * driver; or, if no APIDs have yet been assigned, the APQI is not 1298 * contained in an APQN bound to the vfio_ap device driver. 1299 * 1300 * 4. -EADDRINUSE 1301 * An APQN derived from the cross product of the APQI being assigned 1302 * and the APIDs previously assigned is being used by another mediated 1303 * matrix device 1304 * 1305 * 5. -EAGAIN 1306 * The lock required to validate the mdev's AP configuration could not 1307 * be obtained. 1308 */ 1309 static ssize_t assign_domain_store(struct device *dev, 1310 struct device_attribute *attr, 1311 const char *buf, size_t count) 1312 { 1313 int ret; 1314 unsigned long apqi; 1315 DECLARE_BITMAP(apm_filtered, AP_DEVICES); 1316 struct ap_matrix_mdev *matrix_mdev = dev_get_drvdata(dev); 1317 1318 mutex_lock(&ap_attr_mutex); 1319 get_update_locks_for_mdev(matrix_mdev); 1320 1321 ret = kstrtoul(buf, 0, &apqi); 1322 if (ret) 1323 goto done; 1324 1325 if (apqi > matrix_mdev->matrix.aqm_max) { 1326 ret = -ENODEV; 1327 goto done; 1328 } 1329 1330 if (test_bit_inv(apqi, matrix_mdev->matrix.aqm)) { 1331 ret = count; 1332 goto done; 1333 } 1334 1335 set_bit_inv(apqi, matrix_mdev->matrix.aqm); 1336 1337 ret = vfio_ap_mdev_validate_masks(matrix_mdev); 1338 if (ret) { 1339 clear_bit_inv(apqi, matrix_mdev->matrix.aqm); 1340 goto done; 1341 } 1342 1343 vfio_ap_mdev_link_domain(matrix_mdev, apqi); 1344 1345 if (vfio_ap_mdev_filter_matrix(matrix_mdev, apm_filtered)) { 1346 vfio_ap_mdev_update_guest_apcb(matrix_mdev); 1347 reset_queues_for_apids(matrix_mdev, apm_filtered); 1348 } 1349 1350 ret = count; 1351 done: 1352 release_update_locks_for_mdev(matrix_mdev); 1353 mutex_unlock(&ap_attr_mutex); 1354 1355 return ret; 1356 } 1357 static DEVICE_ATTR_WO(assign_domain); 1358 1359 static void vfio_ap_mdev_unlink_domain(struct ap_matrix_mdev *matrix_mdev, 1360 unsigned long apqi, 1361 struct list_head *qlist) 1362 { 1363 unsigned long apid; 1364 struct vfio_ap_queue *q; 1365 1366 for_each_set_bit_inv(apid, matrix_mdev->matrix.apm, AP_DEVICES) { 1367 q = vfio_ap_unlink_apqn_fr_mdev(matrix_mdev, apid, apqi); 1368 1369 if (q && qlist) { 1370 if (test_bit_inv(apid, matrix_mdev->shadow_apcb.apm) && 1371 test_bit_inv(apqi, matrix_mdev->shadow_apcb.aqm)) 1372 list_add_tail(&q->reset_qnode, qlist); 1373 } 1374 } 1375 } 1376 1377 static void vfio_ap_mdev_hot_unplug_domains(struct ap_matrix_mdev *matrix_mdev, 1378 unsigned long *apqis) 1379 { 1380 struct vfio_ap_queue *q, *tmpq; 1381 struct list_head qlist; 1382 unsigned long apqi; 1383 bool apcb_update = false; 1384 1385 INIT_LIST_HEAD(&qlist); 1386 1387 for_each_set_bit_inv(apqi, apqis, AP_DOMAINS) { 1388 vfio_ap_mdev_unlink_domain(matrix_mdev, apqi, &qlist); 1389 1390 if (test_bit_inv(apqi, matrix_mdev->shadow_apcb.aqm)) { 1391 clear_bit_inv(apqi, matrix_mdev->shadow_apcb.aqm); 1392 apcb_update = true; 1393 } 1394 } 1395 1396 /* Only update apcb if needed to avoid impacting guest */ 1397 if (apcb_update) 1398 vfio_ap_mdev_update_guest_apcb(matrix_mdev); 1399 1400 vfio_ap_mdev_reset_qlist(&qlist); 1401 1402 list_for_each_entry_safe(q, tmpq, &qlist, reset_qnode) { 1403 vfio_ap_unlink_mdev_fr_queue(q); 1404 list_del(&q->reset_qnode); 1405 } 1406 } 1407 1408 static void vfio_ap_mdev_hot_unplug_domain(struct ap_matrix_mdev *matrix_mdev, 1409 unsigned long apqi) 1410 { 1411 DECLARE_BITMAP(apqis, AP_DOMAINS); 1412 1413 bitmap_zero(apqis, AP_DEVICES); 1414 set_bit_inv(apqi, apqis); 1415 vfio_ap_mdev_hot_unplug_domains(matrix_mdev, apqis); 1416 } 1417 1418 /** 1419 * unassign_domain_store - parses the APQI from @buf and clears the 1420 * corresponding bit in the mediated matrix device's AQM 1421 * 1422 * @dev: the matrix device 1423 * @attr: the mediated matrix device's unassign_domain attribute 1424 * @buf: a buffer containing the AP queue index (APQI) of the domain to 1425 * be unassigned 1426 * @count: the number of bytes in @buf 1427 * 1428 * Return: the number of bytes processed if the APQI is valid; otherwise, 1429 * returns one of the following errors: 1430 * -EINVAL if the APQI is not a number 1431 * -ENODEV if the APQI exceeds the maximum value configured for the system 1432 */ 1433 static ssize_t unassign_domain_store(struct device *dev, 1434 struct device_attribute *attr, 1435 const char *buf, size_t count) 1436 { 1437 int ret; 1438 unsigned long apqi; 1439 struct ap_matrix_mdev *matrix_mdev = dev_get_drvdata(dev); 1440 1441 get_update_locks_for_mdev(matrix_mdev); 1442 1443 ret = kstrtoul(buf, 0, &apqi); 1444 if (ret) 1445 goto done; 1446 1447 if (apqi > matrix_mdev->matrix.aqm_max) { 1448 ret = -ENODEV; 1449 goto done; 1450 } 1451 1452 if (!test_bit_inv(apqi, matrix_mdev->matrix.aqm)) { 1453 ret = count; 1454 goto done; 1455 } 1456 1457 clear_bit_inv((unsigned long)apqi, matrix_mdev->matrix.aqm); 1458 vfio_ap_mdev_hot_unplug_domain(matrix_mdev, apqi); 1459 ret = count; 1460 1461 done: 1462 release_update_locks_for_mdev(matrix_mdev); 1463 return ret; 1464 } 1465 static DEVICE_ATTR_WO(unassign_domain); 1466 1467 /** 1468 * assign_control_domain_store - parses the domain ID from @buf and sets 1469 * the corresponding bit in the mediated matrix device's ADM 1470 * 1471 * @dev: the matrix device 1472 * @attr: the mediated matrix device's assign_control_domain attribute 1473 * @buf: a buffer containing the domain ID to be assigned 1474 * @count: the number of bytes in @buf 1475 * 1476 * Return: the number of bytes processed if the domain ID is valid; otherwise, 1477 * returns one of the following errors: 1478 * -EINVAL if the ID is not a number 1479 * -ENODEV if the ID exceeds the maximum value configured for the system 1480 */ 1481 static ssize_t assign_control_domain_store(struct device *dev, 1482 struct device_attribute *attr, 1483 const char *buf, size_t count) 1484 { 1485 int ret; 1486 unsigned long id; 1487 struct ap_matrix_mdev *matrix_mdev = dev_get_drvdata(dev); 1488 1489 get_update_locks_for_mdev(matrix_mdev); 1490 1491 ret = kstrtoul(buf, 0, &id); 1492 if (ret) 1493 goto done; 1494 1495 if (id > matrix_mdev->matrix.adm_max) { 1496 ret = -ENODEV; 1497 goto done; 1498 } 1499 1500 if (test_bit_inv(id, matrix_mdev->matrix.adm)) { 1501 ret = count; 1502 goto done; 1503 } 1504 1505 /* Set the bit in the ADM (bitmask) corresponding to the AP control 1506 * domain number (id). The bits in the mask, from most significant to 1507 * least significant, correspond to IDs 0 up to the one less than the 1508 * number of control domains that can be assigned. 1509 */ 1510 set_bit_inv(id, matrix_mdev->matrix.adm); 1511 if (vfio_ap_mdev_filter_cdoms(matrix_mdev)) 1512 vfio_ap_mdev_update_guest_apcb(matrix_mdev); 1513 1514 ret = count; 1515 done: 1516 release_update_locks_for_mdev(matrix_mdev); 1517 return ret; 1518 } 1519 static DEVICE_ATTR_WO(assign_control_domain); 1520 1521 /** 1522 * unassign_control_domain_store - parses the domain ID from @buf and 1523 * clears the corresponding bit in the mediated matrix device's ADM 1524 * 1525 * @dev: the matrix device 1526 * @attr: the mediated matrix device's unassign_control_domain attribute 1527 * @buf: a buffer containing the domain ID to be unassigned 1528 * @count: the number of bytes in @buf 1529 * 1530 * Return: the number of bytes processed if the domain ID is valid; otherwise, 1531 * returns one of the following errors: 1532 * -EINVAL if the ID is not a number 1533 * -ENODEV if the ID exceeds the maximum value configured for the system 1534 */ 1535 static ssize_t unassign_control_domain_store(struct device *dev, 1536 struct device_attribute *attr, 1537 const char *buf, size_t count) 1538 { 1539 int ret; 1540 unsigned long domid; 1541 struct ap_matrix_mdev *matrix_mdev = dev_get_drvdata(dev); 1542 1543 get_update_locks_for_mdev(matrix_mdev); 1544 1545 ret = kstrtoul(buf, 0, &domid); 1546 if (ret) 1547 goto done; 1548 1549 if (domid > matrix_mdev->matrix.adm_max) { 1550 ret = -ENODEV; 1551 goto done; 1552 } 1553 1554 if (!test_bit_inv(domid, matrix_mdev->matrix.adm)) { 1555 ret = count; 1556 goto done; 1557 } 1558 1559 clear_bit_inv(domid, matrix_mdev->matrix.adm); 1560 1561 if (test_bit_inv(domid, matrix_mdev->shadow_apcb.adm)) { 1562 clear_bit_inv(domid, matrix_mdev->shadow_apcb.adm); 1563 vfio_ap_mdev_update_guest_apcb(matrix_mdev); 1564 } 1565 1566 ret = count; 1567 done: 1568 release_update_locks_for_mdev(matrix_mdev); 1569 return ret; 1570 } 1571 static DEVICE_ATTR_WO(unassign_control_domain); 1572 1573 static ssize_t control_domains_show(struct device *dev, 1574 struct device_attribute *dev_attr, 1575 char *buf) 1576 { 1577 unsigned long id; 1578 struct ap_matrix_mdev *matrix_mdev = dev_get_drvdata(dev); 1579 unsigned long max_domid = matrix_mdev->matrix.adm_max; 1580 int nchars = 0; 1581 1582 mutex_lock(&matrix_dev->mdevs_lock); 1583 for_each_set_bit_inv(id, matrix_mdev->matrix.adm, max_domid + 1) 1584 nchars += sysfs_emit_at(buf, nchars, "%04lx\n", id); 1585 mutex_unlock(&matrix_dev->mdevs_lock); 1586 1587 return nchars; 1588 } 1589 static DEVICE_ATTR_RO(control_domains); 1590 1591 static ssize_t vfio_ap_mdev_matrix_show(struct ap_matrix *matrix, char *buf) 1592 { 1593 unsigned long apid; 1594 unsigned long apqi; 1595 unsigned long apid1; 1596 unsigned long apqi1; 1597 unsigned long napm_bits = matrix->apm_max + 1; 1598 unsigned long naqm_bits = matrix->aqm_max + 1; 1599 int nchars = 0; 1600 1601 apid1 = find_first_bit_inv(matrix->apm, napm_bits); 1602 apqi1 = find_first_bit_inv(matrix->aqm, naqm_bits); 1603 1604 if ((apid1 < napm_bits) && (apqi1 < naqm_bits)) { 1605 for_each_set_bit_inv(apid, matrix->apm, napm_bits) { 1606 for_each_set_bit_inv(apqi, matrix->aqm, naqm_bits) 1607 nchars += sysfs_emit_at(buf, nchars, "%02lx.%04lx\n", apid, apqi); 1608 } 1609 } else if (apid1 < napm_bits) { 1610 for_each_set_bit_inv(apid, matrix->apm, napm_bits) 1611 nchars += sysfs_emit_at(buf, nchars, "%02lx.\n", apid); 1612 } else if (apqi1 < naqm_bits) { 1613 for_each_set_bit_inv(apqi, matrix->aqm, naqm_bits) 1614 nchars += sysfs_emit_at(buf, nchars, ".%04lx\n", apqi); 1615 } 1616 1617 return nchars; 1618 } 1619 1620 static ssize_t matrix_show(struct device *dev, struct device_attribute *attr, 1621 char *buf) 1622 { 1623 ssize_t nchars; 1624 struct ap_matrix_mdev *matrix_mdev = dev_get_drvdata(dev); 1625 1626 mutex_lock(&matrix_dev->mdevs_lock); 1627 nchars = vfio_ap_mdev_matrix_show(&matrix_mdev->matrix, buf); 1628 mutex_unlock(&matrix_dev->mdevs_lock); 1629 1630 return nchars; 1631 } 1632 static DEVICE_ATTR_RO(matrix); 1633 1634 static ssize_t guest_matrix_show(struct device *dev, 1635 struct device_attribute *attr, char *buf) 1636 { 1637 ssize_t nchars; 1638 struct ap_matrix_mdev *matrix_mdev = dev_get_drvdata(dev); 1639 1640 mutex_lock(&matrix_dev->mdevs_lock); 1641 nchars = vfio_ap_mdev_matrix_show(&matrix_mdev->shadow_apcb, buf); 1642 mutex_unlock(&matrix_dev->mdevs_lock); 1643 1644 return nchars; 1645 } 1646 static DEVICE_ATTR_RO(guest_matrix); 1647 1648 static ssize_t write_ap_bitmap(unsigned long *bitmap, char *buf, int offset, char sep) 1649 { 1650 return sysfs_emit_at(buf, offset, "0x%016lx%016lx%016lx%016lx%c", 1651 bitmap[0], bitmap[1], bitmap[2], bitmap[3], sep); 1652 } 1653 1654 static ssize_t ap_config_show(struct device *dev, struct device_attribute *attr, 1655 char *buf) 1656 { 1657 struct ap_matrix_mdev *matrix_mdev = dev_get_drvdata(dev); 1658 int idx = 0; 1659 1660 idx += write_ap_bitmap(matrix_mdev->matrix.apm, buf, idx, ','); 1661 idx += write_ap_bitmap(matrix_mdev->matrix.aqm, buf, idx, ','); 1662 idx += write_ap_bitmap(matrix_mdev->matrix.adm, buf, idx, '\n'); 1663 1664 return idx; 1665 } 1666 1667 /* Number of characters needed for a complete hex mask representing the bits in .. */ 1668 #define AP_DEVICES_STRLEN (AP_DEVICES / 4 + 3) 1669 #define AP_DOMAINS_STRLEN (AP_DOMAINS / 4 + 3) 1670 #define AP_CONFIG_STRLEN (AP_DEVICES_STRLEN + 2 * AP_DOMAINS_STRLEN) 1671 1672 static int parse_bitmap(char **strbufptr, unsigned long *bitmap, int nbits) 1673 { 1674 char *curmask; 1675 1676 curmask = strsep(strbufptr, ",\n"); 1677 if (!curmask) 1678 return -EINVAL; 1679 1680 bitmap_clear(bitmap, 0, nbits); 1681 return ap_hex2bitmap(curmask, bitmap, nbits); 1682 } 1683 1684 static int ap_matrix_overflow_check(struct ap_matrix_mdev *matrix_mdev) 1685 { 1686 unsigned long bit; 1687 1688 for_each_set_bit_inv(bit, matrix_mdev->matrix.apm, AP_DEVICES) { 1689 if (bit > matrix_mdev->matrix.apm_max) 1690 return -ENODEV; 1691 } 1692 1693 for_each_set_bit_inv(bit, matrix_mdev->matrix.aqm, AP_DOMAINS) { 1694 if (bit > matrix_mdev->matrix.aqm_max) 1695 return -ENODEV; 1696 } 1697 1698 for_each_set_bit_inv(bit, matrix_mdev->matrix.adm, AP_DOMAINS) { 1699 if (bit > matrix_mdev->matrix.adm_max) 1700 return -ENODEV; 1701 } 1702 1703 return 0; 1704 } 1705 1706 static void ap_matrix_copy(struct ap_matrix *dst, struct ap_matrix *src) 1707 { 1708 /* This check works around false positive gcc -Wstringop-overread */ 1709 if (!src) 1710 return; 1711 1712 bitmap_copy(dst->apm, src->apm, AP_DEVICES); 1713 bitmap_copy(dst->aqm, src->aqm, AP_DOMAINS); 1714 bitmap_copy(dst->adm, src->adm, AP_DOMAINS); 1715 } 1716 1717 static ssize_t ap_config_store(struct device *dev, struct device_attribute *attr, 1718 const char *buf, size_t count) 1719 { 1720 struct ap_matrix_mdev *matrix_mdev = dev_get_drvdata(dev); 1721 struct ap_matrix m_new, m_old, m_added, m_removed; 1722 DECLARE_BITMAP(apm_filtered, AP_DEVICES); 1723 unsigned long newbit; 1724 char *newbuf, *rest; 1725 int rc = count; 1726 bool do_update; 1727 1728 newbuf = kstrndup(buf, AP_CONFIG_STRLEN, GFP_KERNEL); 1729 if (!newbuf) 1730 return -ENOMEM; 1731 rest = newbuf; 1732 1733 mutex_lock(&ap_attr_mutex); 1734 get_update_locks_for_mdev(matrix_mdev); 1735 1736 /* Save old state */ 1737 ap_matrix_copy(&m_old, &matrix_mdev->matrix); 1738 if (parse_bitmap(&rest, m_new.apm, AP_DEVICES) || 1739 parse_bitmap(&rest, m_new.aqm, AP_DOMAINS) || 1740 parse_bitmap(&rest, m_new.adm, AP_DOMAINS)) { 1741 rc = -EINVAL; 1742 goto out; 1743 } 1744 1745 bitmap_andnot(m_removed.apm, m_old.apm, m_new.apm, AP_DEVICES); 1746 bitmap_andnot(m_removed.aqm, m_old.aqm, m_new.aqm, AP_DOMAINS); 1747 bitmap_andnot(m_added.apm, m_new.apm, m_old.apm, AP_DEVICES); 1748 bitmap_andnot(m_added.aqm, m_new.aqm, m_old.aqm, AP_DOMAINS); 1749 1750 /* Need new bitmaps in matrix_mdev for validation */ 1751 ap_matrix_copy(&matrix_mdev->matrix, &m_new); 1752 1753 /* Ensure new state is valid, else undo new state */ 1754 rc = vfio_ap_mdev_validate_masks(matrix_mdev); 1755 if (rc) { 1756 ap_matrix_copy(&matrix_mdev->matrix, &m_old); 1757 goto out; 1758 } 1759 rc = ap_matrix_overflow_check(matrix_mdev); 1760 if (rc) { 1761 ap_matrix_copy(&matrix_mdev->matrix, &m_old); 1762 goto out; 1763 } 1764 rc = count; 1765 1766 /* Need old bitmaps in matrix_mdev for unplug/unlink */ 1767 ap_matrix_copy(&matrix_mdev->matrix, &m_old); 1768 1769 /* Unlink removed adapters/domains */ 1770 vfio_ap_mdev_hot_unplug_adapters(matrix_mdev, m_removed.apm); 1771 vfio_ap_mdev_hot_unplug_domains(matrix_mdev, m_removed.aqm); 1772 1773 /* Need new bitmaps in matrix_mdev for linking new adapters/domains */ 1774 ap_matrix_copy(&matrix_mdev->matrix, &m_new); 1775 1776 /* Link newly added adapters */ 1777 for_each_set_bit_inv(newbit, m_added.apm, AP_DEVICES) 1778 vfio_ap_mdev_link_adapter(matrix_mdev, newbit); 1779 1780 for_each_set_bit_inv(newbit, m_added.aqm, AP_DOMAINS) 1781 vfio_ap_mdev_link_domain(matrix_mdev, newbit); 1782 1783 /* filter resources not bound to vfio-ap */ 1784 do_update = vfio_ap_mdev_filter_matrix(matrix_mdev, apm_filtered); 1785 do_update |= vfio_ap_mdev_filter_cdoms(matrix_mdev); 1786 1787 /* Apply changes to shadow apbc if things changed */ 1788 if (do_update) { 1789 vfio_ap_mdev_update_guest_apcb(matrix_mdev); 1790 reset_queues_for_apids(matrix_mdev, apm_filtered); 1791 } 1792 out: 1793 release_update_locks_for_mdev(matrix_mdev); 1794 mutex_unlock(&ap_attr_mutex); 1795 kfree(newbuf); 1796 return rc; 1797 } 1798 static DEVICE_ATTR_RW(ap_config); 1799 1800 static struct attribute *vfio_ap_mdev_attrs[] = { 1801 &dev_attr_assign_adapter.attr, 1802 &dev_attr_unassign_adapter.attr, 1803 &dev_attr_assign_domain.attr, 1804 &dev_attr_unassign_domain.attr, 1805 &dev_attr_assign_control_domain.attr, 1806 &dev_attr_unassign_control_domain.attr, 1807 &dev_attr_ap_config.attr, 1808 &dev_attr_control_domains.attr, 1809 &dev_attr_matrix.attr, 1810 &dev_attr_guest_matrix.attr, 1811 NULL, 1812 }; 1813 1814 static struct attribute_group vfio_ap_mdev_attr_group = { 1815 .attrs = vfio_ap_mdev_attrs 1816 }; 1817 1818 static const struct attribute_group *vfio_ap_mdev_attr_groups[] = { 1819 &vfio_ap_mdev_attr_group, 1820 NULL 1821 }; 1822 1823 /** 1824 * vfio_ap_mdev_set_kvm - sets all data for @matrix_mdev that are needed 1825 * to manage AP resources for the guest whose state is represented by 1826 * @kvm_file 1827 * 1828 * @matrix_mdev: a mediated matrix device 1829 * @kvm_file: the KVM VM file this vfio device is associated with 1830 * 1831 * Return: 0 if no other mediated matrix device has a reference to the VM; 1832 * otherwise, returns an -EPERM. 1833 */ 1834 static int vfio_ap_mdev_set_kvm(struct ap_matrix_mdev *matrix_mdev, 1835 struct file *kvm_file) 1836 { 1837 struct kvm *kvm; 1838 1839 if (!kvm_file) 1840 return -ENOENT; 1841 1842 kvm = file_to_kvm_s390(kvm_file); 1843 if (!kvm) 1844 return -ENOENT; 1845 1846 if (kvm->arch.crypto.crycbd) { 1847 get_update_locks_for_kvm(kvm); 1848 if (kvm->arch.crypto.pqap_hook) { 1849 release_update_locks_for_kvm(kvm); 1850 return -EPERM; 1851 } 1852 kvm->arch.crypto.pqap_hook = &matrix_mdev->pqap_hook; 1853 1854 matrix_mdev->kvm = kvm; 1855 vfio_ap_mdev_update_guest_apcb(matrix_mdev); 1856 release_update_locks_for_kvm(kvm); 1857 } 1858 1859 return 0; 1860 } 1861 1862 static void unmap_iova(struct ap_matrix_mdev *matrix_mdev, u64 iova, u64 length) 1863 { 1864 struct ap_queue_table *qtable = &matrix_mdev->qtable; 1865 struct vfio_ap_queue *q; 1866 int loop_cursor; 1867 1868 hash_for_each(qtable->queues, loop_cursor, q, mdev_qnode) { 1869 if (q->saved_iova >= iova && q->saved_iova < iova + length) 1870 vfio_ap_irq_disable(q); 1871 } 1872 } 1873 1874 static void vfio_ap_mdev_dma_unmap(struct vfio_device *vdev, u64 iova, 1875 u64 length) 1876 { 1877 struct ap_matrix_mdev *matrix_mdev = 1878 container_of(vdev, struct ap_matrix_mdev, vdev); 1879 1880 mutex_lock(&matrix_dev->mdevs_lock); 1881 1882 unmap_iova(matrix_mdev, iova, length); 1883 1884 mutex_unlock(&matrix_dev->mdevs_lock); 1885 } 1886 1887 /** 1888 * vfio_ap_mdev_unset_kvm - performs clean-up of resources no longer needed 1889 * by @matrix_mdev. 1890 * 1891 * @matrix_mdev: a matrix mediated device 1892 */ 1893 static void vfio_ap_mdev_unset_kvm(struct ap_matrix_mdev *matrix_mdev) 1894 { 1895 struct kvm *kvm = matrix_mdev->kvm; 1896 1897 if (kvm && kvm->arch.crypto.crycbd) { 1898 get_update_locks_for_kvm(kvm); 1899 kvm->arch.crypto.pqap_hook = NULL; 1900 1901 kvm_arch_crypto_clear_masks(kvm); 1902 vfio_ap_mdev_reset_queues(matrix_mdev); 1903 matrix_mdev->kvm = NULL; 1904 1905 release_update_locks_for_kvm(kvm); 1906 } 1907 } 1908 1909 static struct vfio_ap_queue *vfio_ap_find_queue(int apqn) 1910 { 1911 struct ap_queue *queue; 1912 struct vfio_ap_queue *q = NULL; 1913 1914 queue = ap_get_qdev(apqn); 1915 if (!queue) 1916 return NULL; 1917 1918 if (queue->ap_dev.device.driver == &matrix_dev->vfio_ap_drv->driver) 1919 q = dev_get_drvdata(&queue->ap_dev.device); 1920 1921 put_device(&queue->ap_dev.device); 1922 1923 return q; 1924 } 1925 1926 static int apq_status_check(int apqn, struct ap_queue_status *status) 1927 { 1928 switch (status->response_code) { 1929 case AP_RESPONSE_NORMAL: 1930 case AP_RESPONSE_DECONFIGURED: 1931 case AP_RESPONSE_CHECKSTOPPED: 1932 return 0; 1933 case AP_RESPONSE_RESET_IN_PROGRESS: 1934 case AP_RESPONSE_BUSY: 1935 return -EBUSY; 1936 case AP_RESPONSE_ASSOC_SECRET_NOT_UNIQUE: 1937 case AP_RESPONSE_ASSOC_FAILED: 1938 /* 1939 * These asynchronous response codes indicate a PQAP(AAPQ) 1940 * instruction to associate a secret with the guest failed. All 1941 * subsequent AP instructions will end with the asynchronous 1942 * response code until the AP queue is reset; so, let's return 1943 * a value indicating a reset needs to be performed again. 1944 */ 1945 return -EAGAIN; 1946 case AP_RESPONSE_Q_NOT_AVAIL: 1947 return -ENODEV; 1948 default: 1949 WARN(true, 1950 "failed to verify reset of queue %02x.%04x: TAPQ rc=%u\n", 1951 AP_QID_CARD(apqn), AP_QID_QUEUE(apqn), 1952 status->response_code); 1953 return -EIO; 1954 } 1955 } 1956 1957 #define WAIT_MSG "Waited %dms for reset of queue %02x.%04x (%u, %u, %u)" 1958 1959 static void apq_reset_check(struct work_struct *reset_work) 1960 { 1961 int ret = -EBUSY, elapsed = 0; 1962 struct ap_queue_status status; 1963 struct vfio_ap_queue *q; 1964 1965 q = container_of(reset_work, struct vfio_ap_queue, reset_work); 1966 memcpy(&status, &q->reset_status, sizeof(status)); 1967 while (true) { 1968 msleep(AP_RESET_INTERVAL); 1969 elapsed += AP_RESET_INTERVAL; 1970 status = ap_tapq(q->apqn, NULL); 1971 ret = apq_status_check(q->apqn, &status); 1972 if (ret == -EIO) 1973 return; 1974 if (ret == -ENODEV) { 1975 vfio_ap_free_aqic_resources(q); 1976 return; 1977 } 1978 if (ret == -EBUSY) { 1979 pr_notice_ratelimited(WAIT_MSG, elapsed, 1980 AP_QID_CARD(q->apqn), 1981 AP_QID_QUEUE(q->apqn), 1982 status.response_code, 1983 status.queue_empty, 1984 status.irq_enabled); 1985 } else { 1986 if (q->reset_status.response_code == AP_RESPONSE_RESET_IN_PROGRESS || 1987 q->reset_status.response_code == AP_RESPONSE_BUSY || 1988 q->reset_status.response_code == AP_RESPONSE_STATE_CHANGE_IN_PROGRESS || 1989 ret == -EAGAIN) { 1990 status = ap_zapq(q->apqn, 0); 1991 memcpy(&q->reset_status, &status, sizeof(status)); 1992 continue; 1993 } 1994 if (q->saved_isc != VFIO_AP_ISC_INVALID) 1995 vfio_ap_free_aqic_resources(q); 1996 break; 1997 } 1998 } 1999 } 2000 2001 static void vfio_ap_mdev_reset_queue(struct vfio_ap_queue *q) 2002 { 2003 struct ap_queue_status status; 2004 2005 if (!q) 2006 return; 2007 status = ap_zapq(q->apqn, 0); 2008 memcpy(&q->reset_status, &status, sizeof(status)); 2009 switch (status.response_code) { 2010 case AP_RESPONSE_NORMAL: 2011 case AP_RESPONSE_RESET_IN_PROGRESS: 2012 case AP_RESPONSE_BUSY: 2013 case AP_RESPONSE_STATE_CHANGE_IN_PROGRESS: 2014 /* 2015 * Let's verify whether the ZAPQ completed successfully on a work queue. 2016 */ 2017 queue_work(system_long_wq, &q->reset_work); 2018 break; 2019 case AP_RESPONSE_DECONFIGURED: 2020 case AP_RESPONSE_CHECKSTOPPED: 2021 case AP_RESPONSE_Q_NOT_AVAIL: 2022 vfio_ap_free_aqic_resources(q); 2023 break; 2024 default: 2025 WARN(true, 2026 "PQAP/ZAPQ for %02x.%04x failed with invalid rc=%u\n", 2027 AP_QID_CARD(q->apqn), AP_QID_QUEUE(q->apqn), 2028 status.response_code); 2029 } 2030 } 2031 2032 static int vfio_ap_mdev_reset_queues(struct ap_matrix_mdev *matrix_mdev) 2033 { 2034 int ret = 0, loop_cursor; 2035 struct vfio_ap_queue *q; 2036 2037 hash_for_each(matrix_mdev->qtable.queues, loop_cursor, q, mdev_qnode) 2038 vfio_ap_mdev_reset_queue(q); 2039 2040 hash_for_each(matrix_mdev->qtable.queues, loop_cursor, q, mdev_qnode) { 2041 flush_work(&q->reset_work); 2042 2043 if (q->reset_status.response_code) 2044 ret = -EIO; 2045 } 2046 2047 return ret; 2048 } 2049 2050 static int vfio_ap_mdev_reset_qlist(struct list_head *qlist) 2051 { 2052 int ret = 0; 2053 struct vfio_ap_queue *q; 2054 2055 list_for_each_entry(q, qlist, reset_qnode) 2056 vfio_ap_mdev_reset_queue(q); 2057 2058 list_for_each_entry(q, qlist, reset_qnode) { 2059 flush_work(&q->reset_work); 2060 2061 if (q->reset_status.response_code) 2062 ret = -EIO; 2063 } 2064 2065 return ret; 2066 } 2067 2068 static int vfio_ap_mdev_open_device(struct vfio_device *vdev) 2069 { 2070 struct ap_matrix_mdev *matrix_mdev = 2071 container_of(vdev, struct ap_matrix_mdev, vdev); 2072 2073 if (!vdev->kvm) 2074 return -EINVAL; 2075 2076 return vfio_ap_mdev_set_kvm(matrix_mdev, vdev->kvm); 2077 } 2078 2079 static void vfio_ap_mdev_close_device(struct vfio_device *vdev) 2080 { 2081 struct ap_matrix_mdev *matrix_mdev = 2082 container_of(vdev, struct ap_matrix_mdev, vdev); 2083 2084 vfio_ap_mdev_unset_kvm(matrix_mdev); 2085 } 2086 2087 static void vfio_ap_mdev_request(struct vfio_device *vdev, unsigned int count) 2088 { 2089 struct device *dev = vdev->dev; 2090 struct ap_matrix_mdev *matrix_mdev; 2091 2092 matrix_mdev = container_of(vdev, struct ap_matrix_mdev, vdev); 2093 2094 get_update_locks_for_mdev(matrix_mdev); 2095 2096 if (matrix_mdev->kvm) { 2097 kvm_arch_crypto_clear_masks(matrix_mdev->kvm); 2098 signal_guest_ap_cfg_changed(matrix_mdev); 2099 } 2100 2101 if (matrix_mdev->req_trigger) { 2102 if (!(count % 10)) 2103 dev_notice_ratelimited(dev, 2104 "Relaying device request to user (#%u)\n", 2105 count); 2106 2107 eventfd_signal(matrix_mdev->req_trigger); 2108 } else if (count == 0) { 2109 dev_notice(dev, 2110 "No device request registered, blocked until released by user\n"); 2111 } 2112 2113 release_update_locks_for_mdev(matrix_mdev); 2114 } 2115 2116 static int vfio_ap_mdev_get_device_info(unsigned long arg) 2117 { 2118 unsigned long minsz; 2119 struct vfio_device_info info; 2120 2121 minsz = offsetofend(struct vfio_device_info, num_irqs); 2122 2123 if (copy_from_user(&info, (void __user *)arg, minsz)) 2124 return -EFAULT; 2125 2126 if (info.argsz < minsz) 2127 return -EINVAL; 2128 2129 info.flags = VFIO_DEVICE_FLAGS_AP | VFIO_DEVICE_FLAGS_RESET; 2130 info.num_regions = 0; 2131 info.num_irqs = VFIO_AP_NUM_IRQS; 2132 2133 return copy_to_user((void __user *)arg, &info, minsz) ? -EFAULT : 0; 2134 } 2135 2136 static ssize_t vfio_ap_get_irq_info(unsigned long arg) 2137 { 2138 unsigned long minsz; 2139 struct vfio_irq_info info; 2140 2141 minsz = offsetofend(struct vfio_irq_info, count); 2142 2143 if (copy_from_user(&info, (void __user *)arg, minsz)) 2144 return -EFAULT; 2145 2146 if (info.argsz < minsz || info.index >= VFIO_AP_NUM_IRQS) 2147 return -EINVAL; 2148 2149 switch (info.index) { 2150 case VFIO_AP_REQ_IRQ_INDEX: 2151 info.count = 1; 2152 info.flags = VFIO_IRQ_INFO_EVENTFD; 2153 break; 2154 case VFIO_AP_CFG_CHG_IRQ_INDEX: 2155 info.count = 1; 2156 info.flags = VFIO_IRQ_INFO_EVENTFD; 2157 break; 2158 default: 2159 return -EINVAL; 2160 } 2161 2162 return copy_to_user((void __user *)arg, &info, minsz) ? -EFAULT : 0; 2163 } 2164 2165 static int vfio_ap_irq_set_init(struct vfio_irq_set *irq_set, unsigned long arg) 2166 { 2167 int ret; 2168 size_t data_size; 2169 unsigned long minsz; 2170 2171 minsz = offsetofend(struct vfio_irq_set, count); 2172 2173 if (copy_from_user(irq_set, (void __user *)arg, minsz)) 2174 return -EFAULT; 2175 2176 ret = vfio_set_irqs_validate_and_prepare(irq_set, 1, VFIO_AP_NUM_IRQS, 2177 &data_size); 2178 if (ret) 2179 return ret; 2180 2181 if (!(irq_set->flags & VFIO_IRQ_SET_ACTION_TRIGGER)) 2182 return -EINVAL; 2183 2184 return 0; 2185 } 2186 2187 static int vfio_ap_set_request_irq(struct ap_matrix_mdev *matrix_mdev, 2188 unsigned long arg) 2189 { 2190 s32 fd; 2191 void __user *data; 2192 unsigned long minsz; 2193 struct eventfd_ctx *req_trigger; 2194 2195 minsz = offsetofend(struct vfio_irq_set, count); 2196 data = (void __user *)(arg + minsz); 2197 2198 if (get_user(fd, (s32 __user *)data)) 2199 return -EFAULT; 2200 2201 if (fd == -1) { 2202 if (matrix_mdev->req_trigger) 2203 eventfd_ctx_put(matrix_mdev->req_trigger); 2204 matrix_mdev->req_trigger = NULL; 2205 } else if (fd >= 0) { 2206 req_trigger = eventfd_ctx_fdget(fd); 2207 if (IS_ERR(req_trigger)) 2208 return PTR_ERR(req_trigger); 2209 2210 if (matrix_mdev->req_trigger) 2211 eventfd_ctx_put(matrix_mdev->req_trigger); 2212 2213 matrix_mdev->req_trigger = req_trigger; 2214 } else { 2215 return -EINVAL; 2216 } 2217 2218 return 0; 2219 } 2220 2221 static int vfio_ap_set_cfg_change_irq(struct ap_matrix_mdev *matrix_mdev, unsigned long arg) 2222 { 2223 s32 fd; 2224 void __user *data; 2225 unsigned long minsz; 2226 struct eventfd_ctx *cfg_chg_trigger; 2227 2228 minsz = offsetofend(struct vfio_irq_set, count); 2229 data = (void __user *)(arg + minsz); 2230 2231 if (get_user(fd, (s32 __user *)data)) 2232 return -EFAULT; 2233 2234 if (fd == -1) { 2235 if (matrix_mdev->cfg_chg_trigger) 2236 eventfd_ctx_put(matrix_mdev->cfg_chg_trigger); 2237 matrix_mdev->cfg_chg_trigger = NULL; 2238 } else if (fd >= 0) { 2239 cfg_chg_trigger = eventfd_ctx_fdget(fd); 2240 if (IS_ERR(cfg_chg_trigger)) 2241 return PTR_ERR(cfg_chg_trigger); 2242 2243 if (matrix_mdev->cfg_chg_trigger) 2244 eventfd_ctx_put(matrix_mdev->cfg_chg_trigger); 2245 2246 matrix_mdev->cfg_chg_trigger = cfg_chg_trigger; 2247 } else { 2248 return -EINVAL; 2249 } 2250 2251 return 0; 2252 } 2253 2254 static int vfio_ap_set_irqs(struct ap_matrix_mdev *matrix_mdev, 2255 unsigned long arg) 2256 { 2257 int ret; 2258 struct vfio_irq_set irq_set; 2259 2260 ret = vfio_ap_irq_set_init(&irq_set, arg); 2261 if (ret) 2262 return ret; 2263 2264 switch (irq_set.flags & VFIO_IRQ_SET_DATA_TYPE_MASK) { 2265 case VFIO_IRQ_SET_DATA_EVENTFD: 2266 switch (irq_set.index) { 2267 case VFIO_AP_REQ_IRQ_INDEX: 2268 return vfio_ap_set_request_irq(matrix_mdev, arg); 2269 case VFIO_AP_CFG_CHG_IRQ_INDEX: 2270 return vfio_ap_set_cfg_change_irq(matrix_mdev, arg); 2271 default: 2272 return -EINVAL; 2273 } 2274 default: 2275 return -EINVAL; 2276 } 2277 } 2278 2279 static ssize_t vfio_ap_mdev_ioctl(struct vfio_device *vdev, 2280 unsigned int cmd, unsigned long arg) 2281 { 2282 struct ap_matrix_mdev *matrix_mdev = 2283 container_of(vdev, struct ap_matrix_mdev, vdev); 2284 int ret; 2285 2286 mutex_lock(&matrix_dev->mdevs_lock); 2287 switch (cmd) { 2288 case VFIO_DEVICE_GET_INFO: 2289 ret = vfio_ap_mdev_get_device_info(arg); 2290 break; 2291 case VFIO_DEVICE_RESET: 2292 ret = vfio_ap_mdev_reset_queues(matrix_mdev); 2293 break; 2294 case VFIO_DEVICE_GET_IRQ_INFO: 2295 ret = vfio_ap_get_irq_info(arg); 2296 break; 2297 case VFIO_DEVICE_SET_IRQS: 2298 ret = vfio_ap_set_irqs(matrix_mdev, arg); 2299 break; 2300 default: 2301 ret = -EOPNOTSUPP; 2302 break; 2303 } 2304 mutex_unlock(&matrix_dev->mdevs_lock); 2305 2306 return ret; 2307 } 2308 2309 static struct ap_matrix_mdev *vfio_ap_mdev_for_queue(struct vfio_ap_queue *q) 2310 { 2311 struct ap_matrix_mdev *matrix_mdev; 2312 unsigned long apid = AP_QID_CARD(q->apqn); 2313 unsigned long apqi = AP_QID_QUEUE(q->apqn); 2314 2315 lockdep_assert_held(&matrix_dev->guests_lock); 2316 2317 list_for_each_entry(matrix_mdev, &matrix_dev->mdev_list, node) { 2318 if (test_bit_inv(apid, matrix_mdev->matrix.apm) && 2319 test_bit_inv(apqi, matrix_mdev->matrix.aqm)) 2320 return matrix_mdev; 2321 } 2322 2323 return NULL; 2324 } 2325 2326 static ssize_t status_show(struct device *dev, 2327 struct device_attribute *attr, 2328 char *buf) 2329 { 2330 ssize_t nchars = 0; 2331 struct vfio_ap_queue *q; 2332 unsigned long apid, apqi; 2333 struct ap_matrix_mdev *matrix_mdev; 2334 struct ap_device *apdev = to_ap_dev(dev); 2335 2336 mutex_lock(&matrix_dev->guests_lock); 2337 mutex_lock(&matrix_dev->mdevs_lock); 2338 q = dev_get_drvdata(&apdev->device); 2339 2340 /* 2341 * Make sure the drvdata has been set before proceeding. There is a 2342 * possibility that the drvdata was not set if the vfio_ap_queue object 2343 * could not be allocated when the queue device was probed. In that case, 2344 * the locks used in vfio_ap_mdev_probe_queue() are released prior to 2345 * removing the sysfs status attribute to avoid a lockdep 2346 * splat. That opens a very small window where the status attribute is 2347 * still available without the vfio_ap_queue object having been 2348 * stored in the device drvdata. In that case, indicate the queue is not 2349 * assigned. 2350 */ 2351 if (!q) { 2352 nchars = sysfs_emit(buf, "%s\n", AP_QUEUE_UNASSIGNED); 2353 goto done; 2354 } 2355 2356 matrix_mdev = vfio_ap_mdev_for_queue(q); 2357 2358 /* If the queue is assigned to the matrix mediated device, then 2359 * determine whether it is passed through to a guest; otherwise, 2360 * indicate that it is unassigned. 2361 */ 2362 if (matrix_mdev) { 2363 apid = AP_QID_CARD(q->apqn); 2364 apqi = AP_QID_QUEUE(q->apqn); 2365 /* 2366 * If the queue is passed through to the guest, then indicate 2367 * that it is in use; otherwise, indicate that it is 2368 * merely assigned to a matrix mediated device. 2369 */ 2370 if (matrix_mdev->kvm && 2371 test_bit_inv(apid, matrix_mdev->shadow_apcb.apm) && 2372 test_bit_inv(apqi, matrix_mdev->shadow_apcb.aqm)) 2373 nchars = sysfs_emit(buf, "%s\n", AP_QUEUE_IN_USE); 2374 else 2375 nchars = sysfs_emit(buf, "%s\n", AP_QUEUE_ASSIGNED); 2376 } else { 2377 nchars = sysfs_emit(buf, "%s\n", AP_QUEUE_UNASSIGNED); 2378 } 2379 2380 done: 2381 mutex_unlock(&matrix_dev->mdevs_lock); 2382 mutex_unlock(&matrix_dev->guests_lock); 2383 2384 return nchars; 2385 } 2386 2387 static DEVICE_ATTR_RO(status); 2388 2389 static struct attribute *vfio_queue_attrs[] = { 2390 &dev_attr_status.attr, 2391 NULL, 2392 }; 2393 2394 static const struct attribute_group vfio_queue_attr_group = { 2395 .attrs = vfio_queue_attrs, 2396 }; 2397 2398 static const struct vfio_device_ops vfio_ap_matrix_dev_ops = { 2399 .init = vfio_ap_mdev_init_dev, 2400 .open_device = vfio_ap_mdev_open_device, 2401 .close_device = vfio_ap_mdev_close_device, 2402 .ioctl = vfio_ap_mdev_ioctl, 2403 .dma_unmap = vfio_ap_mdev_dma_unmap, 2404 .bind_iommufd = vfio_iommufd_emulated_bind, 2405 .unbind_iommufd = vfio_iommufd_emulated_unbind, 2406 .attach_ioas = vfio_iommufd_emulated_attach_ioas, 2407 .detach_ioas = vfio_iommufd_emulated_detach_ioas, 2408 .request = vfio_ap_mdev_request 2409 }; 2410 2411 static struct mdev_driver vfio_ap_matrix_driver = { 2412 .device_api = VFIO_DEVICE_API_AP_STRING, 2413 .max_instances = MAX_ZDEV_ENTRIES_EXT, 2414 .driver = { 2415 .name = "vfio_ap_mdev", 2416 .owner = THIS_MODULE, 2417 .mod_name = KBUILD_MODNAME, 2418 .dev_groups = vfio_ap_mdev_attr_groups, 2419 }, 2420 .probe = vfio_ap_mdev_probe, 2421 .remove = vfio_ap_mdev_remove, 2422 }; 2423 2424 int vfio_ap_mdev_register(void) 2425 { 2426 int ret; 2427 2428 ret = mdev_register_driver(&vfio_ap_matrix_driver); 2429 if (ret) 2430 return ret; 2431 2432 matrix_dev->mdev_type.sysfs_name = VFIO_AP_MDEV_TYPE_HWVIRT; 2433 matrix_dev->mdev_type.pretty_name = VFIO_AP_MDEV_NAME_HWVIRT; 2434 matrix_dev->mdev_types = &matrix_dev->mdev_type; 2435 ret = mdev_register_parent(&matrix_dev->parent, &matrix_dev->device, 2436 &vfio_ap_matrix_driver, 2437 &matrix_dev->mdev_types, 1); 2438 if (ret) 2439 goto err_driver; 2440 return 0; 2441 2442 err_driver: 2443 mdev_unregister_driver(&vfio_ap_matrix_driver); 2444 return ret; 2445 } 2446 2447 void vfio_ap_mdev_unregister(void) 2448 { 2449 mdev_unregister_parent(&matrix_dev->parent); 2450 mdev_unregister_driver(&vfio_ap_matrix_driver); 2451 } 2452 2453 int vfio_ap_mdev_probe_queue(struct ap_device *apdev) 2454 { 2455 int ret, apqn; 2456 struct vfio_ap_queue *q; 2457 DECLARE_BITMAP(apm_filtered, AP_DEVICES); 2458 struct ap_matrix_mdev *matrix_mdev; 2459 2460 apqn = to_ap_queue(&apdev->device)->qid; 2461 matrix_mdev = get_update_locks_by_apqn(apqn); 2462 2463 ret = sysfs_create_group(&apdev->device.kobj, &vfio_queue_attr_group); 2464 if (ret) 2465 goto err_release_locks; 2466 2467 q = kzalloc_obj(*q); 2468 if (!q) { 2469 ret = -ENOMEM; 2470 goto err_remove_group; 2471 } 2472 2473 q->apqn = apqn; 2474 q->saved_isc = VFIO_AP_ISC_INVALID; 2475 memset(&q->reset_status, 0, sizeof(q->reset_status)); 2476 INIT_WORK(&q->reset_work, apq_reset_check); 2477 2478 if (matrix_mdev) { 2479 vfio_ap_mdev_link_queue(matrix_mdev, q); 2480 2481 /* 2482 * If we're in the process of handling the adding of adapters or 2483 * domains to the host's AP configuration, then let the 2484 * vfio_ap device driver's on_scan_complete callback filter the 2485 * matrix and update the guest's AP configuration after all of 2486 * the new queue devices are probed. 2487 */ 2488 if (!bitmap_empty(matrix_mdev->apm_add, AP_DEVICES) || 2489 !bitmap_empty(matrix_mdev->aqm_add, AP_DOMAINS)) 2490 goto done; 2491 2492 if (vfio_ap_mdev_filter_matrix(matrix_mdev, apm_filtered)) { 2493 vfio_ap_mdev_update_guest_apcb(matrix_mdev); 2494 reset_queues_for_apids(matrix_mdev, apm_filtered); 2495 } 2496 } 2497 2498 done: 2499 dev_set_drvdata(&apdev->device, q); 2500 release_update_locks_for_mdev(matrix_mdev); 2501 2502 return ret; 2503 2504 err_remove_group: 2505 release_update_locks_for_mdev(matrix_mdev); 2506 sysfs_remove_group(&apdev->device.kobj, &vfio_queue_attr_group); 2507 return ret; 2508 2509 err_release_locks: 2510 release_update_locks_for_mdev(matrix_mdev); 2511 return ret; 2512 } 2513 2514 void vfio_ap_mdev_remove_queue(struct ap_device *apdev) 2515 { 2516 unsigned long apid, apqi; 2517 struct vfio_ap_queue *q; 2518 struct ap_matrix_mdev *matrix_mdev; 2519 2520 sysfs_remove_group(&apdev->device.kobj, &vfio_queue_attr_group); 2521 q = dev_get_drvdata(&apdev->device); 2522 get_update_locks_for_queue(q); 2523 matrix_mdev = q->matrix_mdev; 2524 apid = AP_QID_CARD(q->apqn); 2525 apqi = AP_QID_QUEUE(q->apqn); 2526 2527 if (matrix_mdev) { 2528 /* If the queue is assigned to the guest's AP configuration */ 2529 if (test_bit_inv(apid, matrix_mdev->shadow_apcb.apm) && 2530 test_bit_inv(apqi, matrix_mdev->shadow_apcb.aqm)) { 2531 /* 2532 * Since the queues are defined via a matrix of adapters 2533 * and domains, it is not possible to hot unplug a 2534 * single queue; so, let's unplug the adapter. 2535 */ 2536 clear_bit_inv(apid, matrix_mdev->shadow_apcb.apm); 2537 vfio_ap_mdev_update_guest_apcb(matrix_mdev); 2538 reset_queues_for_apid(matrix_mdev, apid); 2539 goto done; 2540 } 2541 } 2542 2543 /* 2544 * If the queue is not in the host's AP configuration, then resetting 2545 * it will fail with response code 01, (APQN not valid); so, let's make 2546 * sure it is in the host's config. If it is not, free the KVM GISC 2547 * resources. 2548 */ 2549 if (test_bit_inv(apid, (unsigned long *)matrix_dev->info.apm) && 2550 test_bit_inv(apqi, (unsigned long *)matrix_dev->info.aqm)) { 2551 vfio_ap_mdev_reset_queue(q); 2552 flush_work(&q->reset_work); 2553 } else { 2554 vfio_ap_free_aqic_resources(q); 2555 } 2556 2557 done: 2558 if (matrix_mdev) 2559 vfio_ap_unlink_queue_fr_mdev(q); 2560 2561 dev_set_drvdata(&apdev->device, NULL); 2562 kfree(q); 2563 release_update_locks_for_mdev(matrix_mdev); 2564 } 2565 2566 /** 2567 * vfio_ap_mdev_resource_in_use: check whether any of a set of APQNs is 2568 * assigned to a mediated device under the control 2569 * of the vfio_ap device driver. 2570 * 2571 * @apm: a bitmap specifying a set of APIDs comprising the APQNs to check. 2572 * @aqm: a bitmap specifying a set of APQIs comprising the APQNs to check. 2573 * 2574 * Return: 2575 * * -EADDRINUSE if one or more of the APQNs specified via @apm/@aqm are 2576 * assigned to a mediated device under the control of the vfio_ap 2577 * device driver. 2578 * * Otherwise, return 0. 2579 */ 2580 int vfio_ap_mdev_resource_in_use(unsigned long *apm, unsigned long *aqm) 2581 { 2582 int ret; 2583 2584 mutex_lock(&matrix_dev->guests_lock); 2585 mutex_lock(&matrix_dev->mdevs_lock); 2586 ret = vfio_ap_mdev_verify_no_sharing(NULL, apm, aqm); 2587 mutex_unlock(&matrix_dev->mdevs_lock); 2588 mutex_unlock(&matrix_dev->guests_lock); 2589 2590 return ret; 2591 } 2592 2593 /** 2594 * vfio_ap_mdev_hot_unplug_cfg - hot unplug the adapters, domains and control 2595 * domains that have been removed from the host's 2596 * AP configuration from a guest. 2597 * 2598 * @matrix_mdev: an ap_matrix_mdev object attached to a KVM guest. 2599 * @aprem: the adapters that have been removed from the host's AP configuration 2600 * @aqrem: the domains that have been removed from the host's AP configuration 2601 * @cdrem: the control domains that have been removed from the host's AP 2602 * configuration. 2603 */ 2604 static void vfio_ap_mdev_hot_unplug_cfg(struct ap_matrix_mdev *matrix_mdev, 2605 unsigned long *aprem, 2606 unsigned long *aqrem, 2607 unsigned long *cdrem) 2608 { 2609 bool do_hotplug = false; 2610 2611 if (bitmap_intersects(matrix_mdev->shadow_apcb.apm, aprem, AP_DEVICES)) { 2612 bitmap_andnot(matrix_mdev->shadow_apcb.apm, 2613 matrix_mdev->shadow_apcb.apm, 2614 aprem, AP_DEVICES); 2615 do_hotplug = true; 2616 } 2617 2618 if (bitmap_intersects(matrix_mdev->shadow_apcb.aqm, aqrem, AP_DOMAINS)) { 2619 bitmap_andnot(matrix_mdev->shadow_apcb.aqm, 2620 matrix_mdev->shadow_apcb.aqm, 2621 aqrem, AP_DOMAINS); 2622 do_hotplug = true; 2623 } 2624 2625 if (bitmap_intersects(matrix_mdev->shadow_apcb.adm, cdrem, AP_DOMAINS)) { 2626 bitmap_andnot(matrix_mdev->shadow_apcb.adm, 2627 matrix_mdev->shadow_apcb.adm, 2628 cdrem, AP_DOMAINS); 2629 do_hotplug = true; 2630 } 2631 2632 if (do_hotplug) 2633 vfio_ap_mdev_update_guest_apcb(matrix_mdev); 2634 } 2635 2636 /** 2637 * vfio_ap_mdev_cfg_remove - determines which guests are using the adapters, 2638 * domains and control domains that have been removed 2639 * from the host AP configuration and unplugs them 2640 * from those guests. 2641 * 2642 * @ap_remove: bitmap specifying which adapters have been removed from the host 2643 * config. 2644 * @aq_remove: bitmap specifying which domains have been removed from the host 2645 * config. 2646 * @cd_remove: bitmap specifying which control domains have been removed from 2647 * the host config. 2648 */ 2649 static void vfio_ap_mdev_cfg_remove(unsigned long *ap_remove, 2650 unsigned long *aq_remove, 2651 unsigned long *cd_remove) 2652 { 2653 struct ap_matrix_mdev *matrix_mdev; 2654 DECLARE_BITMAP(aprem, AP_DEVICES); 2655 DECLARE_BITMAP(aqrem, AP_DOMAINS); 2656 DECLARE_BITMAP(cdrem, AP_DOMAINS); 2657 int do_remove; 2658 2659 /* 2660 * It is safe to traverse this list here because the 2661 * required guard - matrix_dev->guests_lock - is taken in the 2662 * vfio_ap_on_cfg_changed function prior to this function getting 2663 * called. 2664 */ 2665 list_for_each_entry(matrix_mdev, &matrix_dev->mdev_list, node) { 2666 /* 2667 * The mdevs_lock must be held to access fields within matrix_mdev, 2668 * and kvm->lock must be taken before mdevs_lock to satisfy the lock 2669 * ordering requirement and prevent a lockdep splat. 2670 */ 2671 if (matrix_mdev->kvm) 2672 mutex_lock(&matrix_mdev->kvm->lock); 2673 mutex_lock(&matrix_dev->mdevs_lock); 2674 2675 do_remove = bitmap_and(aprem, ap_remove, 2676 matrix_mdev->matrix.apm, 2677 AP_DEVICES); 2678 do_remove |= bitmap_and(aqrem, aq_remove, 2679 matrix_mdev->matrix.aqm, 2680 AP_DOMAINS); 2681 do_remove |= bitmap_and(cdrem, cd_remove, 2682 matrix_mdev->matrix.adm, 2683 AP_DOMAINS); 2684 2685 if (do_remove) 2686 vfio_ap_mdev_hot_unplug_cfg(matrix_mdev, aprem, aqrem, 2687 cdrem); 2688 2689 mutex_unlock(&matrix_dev->mdevs_lock); 2690 if (matrix_mdev->kvm) 2691 mutex_unlock(&matrix_mdev->kvm->lock); 2692 } 2693 } 2694 2695 /** 2696 * vfio_ap_mdev_on_cfg_remove - responds to the removal of adapters, domains and 2697 * control domains from the host AP configuration 2698 * by unplugging them from the guests that are 2699 * using them. 2700 * @cur_config_info: the current host AP configuration information 2701 * @prev_config_info: the previous host AP configuration information 2702 */ 2703 static void vfio_ap_mdev_on_cfg_remove(struct ap_config_info *cur_config_info, 2704 struct ap_config_info *prev_config_info) 2705 { 2706 int do_remove; 2707 DECLARE_BITMAP(aprem, AP_DEVICES); 2708 DECLARE_BITMAP(aqrem, AP_DOMAINS); 2709 DECLARE_BITMAP(cdrem, AP_DOMAINS); 2710 2711 do_remove = bitmap_andnot(aprem, 2712 (unsigned long *)prev_config_info->apm, 2713 (unsigned long *)cur_config_info->apm, 2714 AP_DEVICES); 2715 do_remove |= bitmap_andnot(aqrem, 2716 (unsigned long *)prev_config_info->aqm, 2717 (unsigned long *)cur_config_info->aqm, 2718 AP_DEVICES); 2719 do_remove |= bitmap_andnot(cdrem, 2720 (unsigned long *)prev_config_info->adm, 2721 (unsigned long *)cur_config_info->adm, 2722 AP_DEVICES); 2723 2724 if (do_remove) 2725 vfio_ap_mdev_cfg_remove(aprem, aqrem, cdrem); 2726 } 2727 2728 /** 2729 * vfio_ap_filter_apid_by_qtype: filter APIDs from an AP mask for adapters that 2730 * are older than AP type 10 (CEX4). 2731 * @apm: a bitmap of the APIDs to examine 2732 * @aqm: a bitmap of the APQIs of the queues to query for the AP type. 2733 */ 2734 static void vfio_ap_filter_apid_by_qtype(unsigned long *apm, unsigned long *aqm) 2735 { 2736 bool apid_cleared; 2737 struct ap_queue_status status; 2738 unsigned long apid, apqi; 2739 struct ap_tapq_hwinfo info; 2740 2741 for_each_set_bit_inv(apid, apm, AP_DEVICES) { 2742 apid_cleared = false; 2743 2744 for_each_set_bit_inv(apqi, aqm, AP_DOMAINS) { 2745 status = ap_test_queue(AP_MKQID(apid, apqi), 1, &info); 2746 switch (status.response_code) { 2747 /* 2748 * According to the architecture in each case 2749 * below, the queue's info should be filled. 2750 */ 2751 case AP_RESPONSE_NORMAL: 2752 case AP_RESPONSE_RESET_IN_PROGRESS: 2753 case AP_RESPONSE_DECONFIGURED: 2754 case AP_RESPONSE_CHECKSTOPPED: 2755 case AP_RESPONSE_BUSY: 2756 /* 2757 * The vfio_ap device driver only 2758 * supports CEX4 and newer adapters, so 2759 * remove the APID if the adapter is 2760 * older than a CEX4. 2761 */ 2762 if (info.at < AP_DEVICE_TYPE_CEX4) { 2763 clear_bit_inv(apid, apm); 2764 apid_cleared = true; 2765 } 2766 2767 break; 2768 2769 default: 2770 /* 2771 * If we don't know the adapter type, 2772 * clear its APID since it can't be 2773 * determined whether the vfio_ap 2774 * device driver supports it. 2775 */ 2776 clear_bit_inv(apid, apm); 2777 apid_cleared = true; 2778 break; 2779 } 2780 2781 /* 2782 * If we've already cleared the APID from the apm, there 2783 * is no need to continue examining the remainin AP 2784 * queues to determine the type of the adapter. 2785 */ 2786 if (apid_cleared) 2787 continue; 2788 } 2789 } 2790 } 2791 2792 /** 2793 * vfio_ap_mdev_cfg_add - store bitmaps specifying the adapters, domains and 2794 * control domains that have been added to the host's 2795 * AP configuration for each matrix mdev to which they 2796 * are assigned. 2797 * 2798 * @apm_add: a bitmap specifying the adapters that have been added to the AP 2799 * configuration. 2800 * @aqm_add: a bitmap specifying the domains that have been added to the AP 2801 * configuration. 2802 * @adm_add: a bitmap specifying the control domains that have been added to the 2803 * AP configuration. 2804 */ 2805 static void vfio_ap_mdev_cfg_add(unsigned long *apm_add, unsigned long *aqm_add, 2806 unsigned long *adm_add) 2807 { 2808 struct ap_matrix_mdev *matrix_mdev; 2809 2810 if (list_empty(&matrix_dev->mdev_list)) 2811 return; 2812 2813 vfio_ap_filter_apid_by_qtype(apm_add, aqm_add); 2814 2815 /* 2816 * It is safe to traverse this list here because the 2817 * required guard - matrix_dev->guests_lock - is taken in the 2818 * vfio_ap_on_cfg_changed function prior to this function getting 2819 * called. 2820 */ 2821 list_for_each_entry(matrix_mdev, &matrix_dev->mdev_list, node) { 2822 /* 2823 * The mdevs_lock must be held in order to access fields 2824 * within matrix_mdev 2825 */ 2826 mutex_lock(&matrix_dev->mdevs_lock); 2827 2828 bitmap_and(matrix_mdev->apm_add, 2829 matrix_mdev->matrix.apm, apm_add, AP_DEVICES); 2830 bitmap_and(matrix_mdev->aqm_add, 2831 matrix_mdev->matrix.aqm, aqm_add, AP_DOMAINS); 2832 bitmap_and(matrix_mdev->adm_add, 2833 matrix_mdev->matrix.adm, adm_add, AP_DEVICES); 2834 2835 mutex_unlock(&matrix_dev->mdevs_lock); 2836 } 2837 } 2838 2839 /** 2840 * vfio_ap_mdev_on_cfg_add - responds to the addition of adapters, domains and 2841 * control domains to the host AP configuration 2842 * by updating the bitmaps that specify what adapters, 2843 * domains and control domains have been added so they 2844 * can be hot plugged into the guest when the AP bus 2845 * scan completes (see vfio_ap_on_scan_complete 2846 * function). 2847 * @cur_config_info: the current AP configuration information 2848 * @prev_config_info: the previous AP configuration information 2849 */ 2850 static void vfio_ap_mdev_on_cfg_add(struct ap_config_info *cur_config_info, 2851 struct ap_config_info *prev_config_info) 2852 { 2853 bool do_add; 2854 DECLARE_BITMAP(apm_add, AP_DEVICES); 2855 DECLARE_BITMAP(aqm_add, AP_DOMAINS); 2856 DECLARE_BITMAP(adm_add, AP_DOMAINS); 2857 2858 do_add = bitmap_andnot(apm_add, 2859 (unsigned long *)cur_config_info->apm, 2860 (unsigned long *)prev_config_info->apm, 2861 AP_DEVICES); 2862 do_add |= bitmap_andnot(aqm_add, 2863 (unsigned long *)cur_config_info->aqm, 2864 (unsigned long *)prev_config_info->aqm, 2865 AP_DOMAINS); 2866 do_add |= bitmap_andnot(adm_add, 2867 (unsigned long *)cur_config_info->adm, 2868 (unsigned long *)prev_config_info->adm, 2869 AP_DOMAINS); 2870 2871 if (do_add) 2872 vfio_ap_mdev_cfg_add(apm_add, aqm_add, adm_add); 2873 } 2874 2875 /** 2876 * vfio_ap_on_cfg_changed - handles notification of changes to the host AP 2877 * configuration. 2878 * 2879 * @cur_cfg_info: the current host AP configuration 2880 * @prev_cfg_info: the previous host AP configuration 2881 */ 2882 void vfio_ap_on_cfg_changed(struct ap_config_info *cur_cfg_info, 2883 struct ap_config_info *prev_cfg_info) 2884 { 2885 if (!cur_cfg_info || !prev_cfg_info) 2886 return; 2887 2888 /* 2889 * Take the guests_lock mutex here to guard access to the 2890 * matrix_dev->mdev_list in the two functions called below. 2891 */ 2892 mutex_lock(&matrix_dev->guests_lock); 2893 2894 vfio_ap_mdev_on_cfg_remove(cur_cfg_info, prev_cfg_info); 2895 vfio_ap_mdev_on_cfg_add(cur_cfg_info, prev_cfg_info); 2896 memcpy(&matrix_dev->info, cur_cfg_info, sizeof(*cur_cfg_info)); 2897 2898 mutex_unlock(&matrix_dev->guests_lock); 2899 } 2900 2901 static void vfio_ap_mdev_hot_plug_cfg(struct ap_matrix_mdev *matrix_mdev) 2902 { 2903 DECLARE_BITMAP(apm_filtered, AP_DEVICES); 2904 bool filter_domains, filter_adapters, filter_cdoms, do_hotplug = false; 2905 2906 /* 2907 * Zero out the apm_filtered bitmap in case there are no adapters or 2908 * domains to be added, but only control domains. In that case, 2909 * vfio_ap_mdev_filter_matrix() - which initializes apm_filtered - will 2910 * not get called and the reset_queues_for_apids will crash because it 2911 * will access an uninitialized bitmap. 2912 */ 2913 bitmap_zero(apm_filtered, AP_DEVICES); 2914 2915 filter_adapters = bitmap_intersects(matrix_mdev->matrix.apm, 2916 matrix_mdev->apm_add, AP_DEVICES); 2917 filter_domains = bitmap_intersects(matrix_mdev->matrix.aqm, 2918 matrix_mdev->aqm_add, AP_DOMAINS); 2919 filter_cdoms = bitmap_intersects(matrix_mdev->matrix.adm, 2920 matrix_mdev->adm_add, AP_DOMAINS); 2921 2922 if (filter_adapters || filter_domains) 2923 do_hotplug = vfio_ap_mdev_filter_matrix(matrix_mdev, apm_filtered); 2924 2925 if (filter_cdoms) 2926 do_hotplug |= vfio_ap_mdev_filter_cdoms(matrix_mdev); 2927 2928 if (do_hotplug) 2929 vfio_ap_mdev_update_guest_apcb(matrix_mdev); 2930 2931 reset_queues_for_apids(matrix_mdev, apm_filtered); 2932 } 2933 2934 void vfio_ap_on_scan_complete(struct ap_config_info *new_config_info, 2935 struct ap_config_info *old_config_info) 2936 { 2937 struct ap_matrix_mdev *matrix_mdev; 2938 2939 mutex_lock(&matrix_dev->guests_lock); 2940 2941 list_for_each_entry(matrix_mdev, &matrix_dev->mdev_list, node) { 2942 /* 2943 * The mdevs_lock must be held to access fields within matrix_mdev, 2944 * and kvm->lock must be taken before mdevs_lock to satisfy the lock 2945 * ordering requirement and prevent a lockdep splat. 2946 */ 2947 if (matrix_mdev->kvm) 2948 mutex_lock(&matrix_mdev->kvm->lock); 2949 mutex_lock(&matrix_dev->mdevs_lock); 2950 2951 if (bitmap_empty(matrix_mdev->apm_add, AP_DEVICES) && 2952 bitmap_empty(matrix_mdev->aqm_add, AP_DOMAINS) && 2953 bitmap_empty(matrix_mdev->adm_add, AP_DOMAINS)) 2954 goto do_unlock; 2955 2956 vfio_ap_mdev_hot_plug_cfg(matrix_mdev); 2957 bitmap_clear(matrix_mdev->apm_add, 0, AP_DEVICES); 2958 bitmap_clear(matrix_mdev->aqm_add, 0, AP_DOMAINS); 2959 bitmap_clear(matrix_mdev->adm_add, 0, AP_DOMAINS); 2960 2961 do_unlock: 2962 mutex_unlock(&matrix_dev->mdevs_lock); 2963 if (matrix_mdev->kvm) 2964 mutex_unlock(&matrix_mdev->kvm->lock); 2965 } 2966 2967 mutex_unlock(&matrix_dev->guests_lock); 2968 } 2969