1 /* 2 * Block driver for s390 storage class memory. 3 * 4 * Copyright IBM Corp. 2012 5 * Author(s): Sebastian Ott <sebott@linux.vnet.ibm.com> 6 */ 7 8 #define KMSG_COMPONENT "scm_block" 9 #define pr_fmt(fmt) KMSG_COMPONENT ": " fmt 10 11 #include <linux/interrupt.h> 12 #include <linux/spinlock.h> 13 #include <linux/mempool.h> 14 #include <linux/module.h> 15 #include <linux/blkdev.h> 16 #include <linux/blk-mq.h> 17 #include <linux/genhd.h> 18 #include <linux/slab.h> 19 #include <linux/list.h> 20 #include <asm/eadm.h> 21 #include "scm_blk.h" 22 23 debug_info_t *scm_debug; 24 static int scm_major; 25 static mempool_t *aidaw_pool; 26 static DEFINE_SPINLOCK(list_lock); 27 static LIST_HEAD(inactive_requests); 28 static unsigned int nr_requests = 64; 29 static unsigned int nr_requests_per_io = 8; 30 static atomic_t nr_devices = ATOMIC_INIT(0); 31 module_param(nr_requests, uint, S_IRUGO); 32 MODULE_PARM_DESC(nr_requests, "Number of parallel requests."); 33 34 module_param(nr_requests_per_io, uint, S_IRUGO); 35 MODULE_PARM_DESC(nr_requests_per_io, "Number of requests per IO."); 36 37 MODULE_DESCRIPTION("Block driver for s390 storage class memory."); 38 MODULE_LICENSE("GPL"); 39 MODULE_ALIAS("scm:scmdev*"); 40 41 static void __scm_free_rq(struct scm_request *scmrq) 42 { 43 struct aob_rq_header *aobrq = to_aobrq(scmrq); 44 45 free_page((unsigned long) scmrq->aob); 46 kfree(scmrq->request); 47 kfree(aobrq); 48 } 49 50 static void scm_free_rqs(void) 51 { 52 struct list_head *iter, *safe; 53 struct scm_request *scmrq; 54 55 spin_lock_irq(&list_lock); 56 list_for_each_safe(iter, safe, &inactive_requests) { 57 scmrq = list_entry(iter, struct scm_request, list); 58 list_del(&scmrq->list); 59 __scm_free_rq(scmrq); 60 } 61 spin_unlock_irq(&list_lock); 62 63 mempool_destroy(aidaw_pool); 64 } 65 66 static int __scm_alloc_rq(void) 67 { 68 struct aob_rq_header *aobrq; 69 struct scm_request *scmrq; 70 71 aobrq = kzalloc(sizeof(*aobrq) + sizeof(*scmrq), GFP_KERNEL); 72 if (!aobrq) 73 return -ENOMEM; 74 75 scmrq = (void *) aobrq->data; 76 scmrq->aob = (void *) get_zeroed_page(GFP_DMA); 77 if (!scmrq->aob) 78 goto free; 79 80 scmrq->request = kcalloc(nr_requests_per_io, sizeof(scmrq->request[0]), 81 GFP_KERNEL); 82 if (!scmrq->request) 83 goto free; 84 85 INIT_LIST_HEAD(&scmrq->list); 86 spin_lock_irq(&list_lock); 87 list_add(&scmrq->list, &inactive_requests); 88 spin_unlock_irq(&list_lock); 89 90 return 0; 91 free: 92 __scm_free_rq(scmrq); 93 return -ENOMEM; 94 } 95 96 static int scm_alloc_rqs(unsigned int nrqs) 97 { 98 int ret = 0; 99 100 aidaw_pool = mempool_create_page_pool(max(nrqs/8, 1U), 0); 101 if (!aidaw_pool) 102 return -ENOMEM; 103 104 while (nrqs-- && !ret) 105 ret = __scm_alloc_rq(); 106 107 return ret; 108 } 109 110 static struct scm_request *scm_request_fetch(void) 111 { 112 struct scm_request *scmrq = NULL; 113 114 spin_lock_irq(&list_lock); 115 if (list_empty(&inactive_requests)) 116 goto out; 117 scmrq = list_first_entry(&inactive_requests, struct scm_request, list); 118 list_del(&scmrq->list); 119 out: 120 spin_unlock_irq(&list_lock); 121 return scmrq; 122 } 123 124 static void scm_request_done(struct scm_request *scmrq) 125 { 126 unsigned long flags; 127 struct msb *msb; 128 u64 aidaw; 129 int i; 130 131 for (i = 0; i < nr_requests_per_io && scmrq->request[i]; i++) { 132 msb = &scmrq->aob->msb[i]; 133 aidaw = msb->data_addr; 134 135 if ((msb->flags & MSB_FLAG_IDA) && aidaw && 136 IS_ALIGNED(aidaw, PAGE_SIZE)) 137 mempool_free(virt_to_page(aidaw), aidaw_pool); 138 } 139 140 spin_lock_irqsave(&list_lock, flags); 141 list_add(&scmrq->list, &inactive_requests); 142 spin_unlock_irqrestore(&list_lock, flags); 143 } 144 145 static bool scm_permit_request(struct scm_blk_dev *bdev, struct request *req) 146 { 147 return rq_data_dir(req) != WRITE || bdev->state != SCM_WR_PROHIBIT; 148 } 149 150 static inline struct aidaw *scm_aidaw_alloc(void) 151 { 152 struct page *page = mempool_alloc(aidaw_pool, GFP_ATOMIC); 153 154 return page ? page_address(page) : NULL; 155 } 156 157 static inline unsigned long scm_aidaw_bytes(struct aidaw *aidaw) 158 { 159 unsigned long _aidaw = (unsigned long) aidaw; 160 unsigned long bytes = ALIGN(_aidaw, PAGE_SIZE) - _aidaw; 161 162 return (bytes / sizeof(*aidaw)) * PAGE_SIZE; 163 } 164 165 struct aidaw *scm_aidaw_fetch(struct scm_request *scmrq, unsigned int bytes) 166 { 167 struct aidaw *aidaw; 168 169 if (scm_aidaw_bytes(scmrq->next_aidaw) >= bytes) 170 return scmrq->next_aidaw; 171 172 aidaw = scm_aidaw_alloc(); 173 if (aidaw) 174 memset(aidaw, 0, PAGE_SIZE); 175 return aidaw; 176 } 177 178 static int scm_request_prepare(struct scm_request *scmrq) 179 { 180 struct scm_blk_dev *bdev = scmrq->bdev; 181 struct scm_device *scmdev = bdev->gendisk->private_data; 182 int pos = scmrq->aob->request.msb_count; 183 struct msb *msb = &scmrq->aob->msb[pos]; 184 struct request *req = scmrq->request[pos]; 185 struct req_iterator iter; 186 struct aidaw *aidaw; 187 struct bio_vec bv; 188 189 aidaw = scm_aidaw_fetch(scmrq, blk_rq_bytes(req)); 190 if (!aidaw) 191 return -ENOMEM; 192 193 msb->bs = MSB_BS_4K; 194 scmrq->aob->request.msb_count++; 195 msb->scm_addr = scmdev->address + ((u64) blk_rq_pos(req) << 9); 196 msb->oc = (rq_data_dir(req) == READ) ? MSB_OC_READ : MSB_OC_WRITE; 197 msb->flags |= MSB_FLAG_IDA; 198 msb->data_addr = (u64) aidaw; 199 200 rq_for_each_segment(bv, req, iter) { 201 WARN_ON(bv.bv_offset); 202 msb->blk_count += bv.bv_len >> 12; 203 aidaw->data_addr = (u64) page_address(bv.bv_page); 204 aidaw++; 205 } 206 207 scmrq->next_aidaw = aidaw; 208 return 0; 209 } 210 211 static inline void scm_request_set(struct scm_request *scmrq, 212 struct request *req) 213 { 214 scmrq->request[scmrq->aob->request.msb_count] = req; 215 } 216 217 static inline void scm_request_init(struct scm_blk_dev *bdev, 218 struct scm_request *scmrq) 219 { 220 struct aob_rq_header *aobrq = to_aobrq(scmrq); 221 struct aob *aob = scmrq->aob; 222 223 memset(scmrq->request, 0, 224 nr_requests_per_io * sizeof(scmrq->request[0])); 225 memset(aob, 0, sizeof(*aob)); 226 aobrq->scmdev = bdev->scmdev; 227 aob->request.cmd_code = ARQB_CMD_MOVE; 228 aob->request.data = (u64) aobrq; 229 scmrq->bdev = bdev; 230 scmrq->retries = 4; 231 scmrq->error = BLK_STS_OK; 232 /* We don't use all msbs - place aidaws at the end of the aob page. */ 233 scmrq->next_aidaw = (void *) &aob->msb[nr_requests_per_io]; 234 } 235 236 static void scm_request_requeue(struct scm_request *scmrq) 237 { 238 struct scm_blk_dev *bdev = scmrq->bdev; 239 int i; 240 241 for (i = 0; i < nr_requests_per_io && scmrq->request[i]; i++) 242 blk_mq_requeue_request(scmrq->request[i], false); 243 244 atomic_dec(&bdev->queued_reqs); 245 scm_request_done(scmrq); 246 blk_mq_kick_requeue_list(bdev->rq); 247 } 248 249 static void scm_request_finish(struct scm_request *scmrq) 250 { 251 struct scm_blk_dev *bdev = scmrq->bdev; 252 int i; 253 254 for (i = 0; i < nr_requests_per_io && scmrq->request[i]; i++) { 255 if (scmrq->error) 256 blk_mq_end_request(scmrq->request[i], scmrq->error); 257 else 258 blk_mq_complete_request(scmrq->request[i]); 259 } 260 261 atomic_dec(&bdev->queued_reqs); 262 scm_request_done(scmrq); 263 } 264 265 static void scm_request_start(struct scm_request *scmrq) 266 { 267 struct scm_blk_dev *bdev = scmrq->bdev; 268 269 atomic_inc(&bdev->queued_reqs); 270 if (eadm_start_aob(scmrq->aob)) { 271 SCM_LOG(5, "no subchannel"); 272 scm_request_requeue(scmrq); 273 } 274 } 275 276 struct scm_queue { 277 struct scm_request *scmrq; 278 spinlock_t lock; 279 }; 280 281 static blk_status_t scm_blk_request(struct blk_mq_hw_ctx *hctx, 282 const struct blk_mq_queue_data *qd) 283 { 284 struct scm_device *scmdev = hctx->queue->queuedata; 285 struct scm_blk_dev *bdev = dev_get_drvdata(&scmdev->dev); 286 struct scm_queue *sq = hctx->driver_data; 287 struct request *req = qd->rq; 288 struct scm_request *scmrq; 289 290 spin_lock(&sq->lock); 291 if (!scm_permit_request(bdev, req)) { 292 spin_unlock(&sq->lock); 293 return BLK_STS_RESOURCE; 294 } 295 296 scmrq = sq->scmrq; 297 if (!scmrq) { 298 scmrq = scm_request_fetch(); 299 if (!scmrq) { 300 SCM_LOG(5, "no request"); 301 spin_unlock(&sq->lock); 302 return BLK_STS_RESOURCE; 303 } 304 scm_request_init(bdev, scmrq); 305 sq->scmrq = scmrq; 306 } 307 scm_request_set(scmrq, req); 308 309 if (scm_request_prepare(scmrq)) { 310 SCM_LOG(5, "aidaw alloc failed"); 311 scm_request_set(scmrq, NULL); 312 313 if (scmrq->aob->request.msb_count) 314 scm_request_start(scmrq); 315 316 sq->scmrq = NULL; 317 spin_unlock(&sq->lock); 318 return BLK_STS_RESOURCE; 319 } 320 blk_mq_start_request(req); 321 322 if (qd->last || scmrq->aob->request.msb_count == nr_requests_per_io) { 323 scm_request_start(scmrq); 324 sq->scmrq = NULL; 325 } 326 spin_unlock(&sq->lock); 327 return BLK_STS_OK; 328 } 329 330 static int scm_blk_init_hctx(struct blk_mq_hw_ctx *hctx, void *data, 331 unsigned int idx) 332 { 333 struct scm_queue *qd = kzalloc(sizeof(*qd), GFP_KERNEL); 334 335 if (!qd) 336 return -ENOMEM; 337 338 spin_lock_init(&qd->lock); 339 hctx->driver_data = qd; 340 341 return 0; 342 } 343 344 static void scm_blk_exit_hctx(struct blk_mq_hw_ctx *hctx, unsigned int idx) 345 { 346 struct scm_queue *qd = hctx->driver_data; 347 348 WARN_ON(qd->scmrq); 349 kfree(hctx->driver_data); 350 hctx->driver_data = NULL; 351 } 352 353 static void __scmrq_log_error(struct scm_request *scmrq) 354 { 355 struct aob *aob = scmrq->aob; 356 357 if (scmrq->error == BLK_STS_TIMEOUT) 358 SCM_LOG(1, "Request timeout"); 359 else { 360 SCM_LOG(1, "Request error"); 361 SCM_LOG_HEX(1, &aob->response, sizeof(aob->response)); 362 } 363 if (scmrq->retries) 364 SCM_LOG(1, "Retry request"); 365 else 366 pr_err("An I/O operation to SCM failed with rc=%d\n", 367 scmrq->error); 368 } 369 370 static void scm_blk_handle_error(struct scm_request *scmrq) 371 { 372 struct scm_blk_dev *bdev = scmrq->bdev; 373 unsigned long flags; 374 375 if (scmrq->error != BLK_STS_IOERR) 376 goto restart; 377 378 /* For -EIO the response block is valid. */ 379 switch (scmrq->aob->response.eqc) { 380 case EQC_WR_PROHIBIT: 381 spin_lock_irqsave(&bdev->lock, flags); 382 if (bdev->state != SCM_WR_PROHIBIT) 383 pr_info("%lx: Write access to the SCM increment is suspended\n", 384 (unsigned long) bdev->scmdev->address); 385 bdev->state = SCM_WR_PROHIBIT; 386 spin_unlock_irqrestore(&bdev->lock, flags); 387 goto requeue; 388 default: 389 break; 390 } 391 392 restart: 393 if (!eadm_start_aob(scmrq->aob)) 394 return; 395 396 requeue: 397 scm_request_requeue(scmrq); 398 } 399 400 void scm_blk_irq(struct scm_device *scmdev, void *data, blk_status_t error) 401 { 402 struct scm_request *scmrq = data; 403 404 scmrq->error = error; 405 if (error) { 406 __scmrq_log_error(scmrq); 407 if (scmrq->retries-- > 0) { 408 scm_blk_handle_error(scmrq); 409 return; 410 } 411 } 412 413 scm_request_finish(scmrq); 414 } 415 416 static void scm_blk_request_done(struct request *req) 417 { 418 blk_mq_end_request(req, 0); 419 } 420 421 static const struct block_device_operations scm_blk_devops = { 422 .owner = THIS_MODULE, 423 }; 424 425 static const struct blk_mq_ops scm_mq_ops = { 426 .queue_rq = scm_blk_request, 427 .complete = scm_blk_request_done, 428 .init_hctx = scm_blk_init_hctx, 429 .exit_hctx = scm_blk_exit_hctx, 430 }; 431 432 int scm_blk_dev_setup(struct scm_blk_dev *bdev, struct scm_device *scmdev) 433 { 434 unsigned int devindex, nr_max_blk; 435 struct request_queue *rq; 436 int len, ret; 437 438 devindex = atomic_inc_return(&nr_devices) - 1; 439 /* scma..scmz + scmaa..scmzz */ 440 if (devindex > 701) { 441 ret = -ENODEV; 442 goto out; 443 } 444 445 bdev->scmdev = scmdev; 446 bdev->state = SCM_OPER; 447 spin_lock_init(&bdev->lock); 448 atomic_set(&bdev->queued_reqs, 0); 449 450 bdev->tag_set.ops = &scm_mq_ops; 451 bdev->tag_set.nr_hw_queues = nr_requests; 452 bdev->tag_set.queue_depth = nr_requests_per_io * nr_requests; 453 bdev->tag_set.flags = BLK_MQ_F_SHOULD_MERGE; 454 455 ret = blk_mq_alloc_tag_set(&bdev->tag_set); 456 if (ret) 457 goto out; 458 459 rq = blk_mq_init_queue(&bdev->tag_set); 460 if (IS_ERR(rq)) { 461 ret = PTR_ERR(rq); 462 goto out_tag; 463 } 464 bdev->rq = rq; 465 nr_max_blk = min(scmdev->nr_max_block, 466 (unsigned int) (PAGE_SIZE / sizeof(struct aidaw))); 467 468 blk_queue_logical_block_size(rq, 1 << 12); 469 blk_queue_max_hw_sectors(rq, nr_max_blk << 3); /* 8 * 512 = blk_size */ 470 blk_queue_max_segments(rq, nr_max_blk); 471 queue_flag_set_unlocked(QUEUE_FLAG_NONROT, rq); 472 queue_flag_clear_unlocked(QUEUE_FLAG_ADD_RANDOM, rq); 473 474 bdev->gendisk = alloc_disk(SCM_NR_PARTS); 475 if (!bdev->gendisk) { 476 ret = -ENOMEM; 477 goto out_queue; 478 } 479 rq->queuedata = scmdev; 480 bdev->gendisk->private_data = scmdev; 481 bdev->gendisk->fops = &scm_blk_devops; 482 bdev->gendisk->queue = rq; 483 bdev->gendisk->major = scm_major; 484 bdev->gendisk->first_minor = devindex * SCM_NR_PARTS; 485 486 len = snprintf(bdev->gendisk->disk_name, DISK_NAME_LEN, "scm"); 487 if (devindex > 25) { 488 len += snprintf(bdev->gendisk->disk_name + len, 489 DISK_NAME_LEN - len, "%c", 490 'a' + (devindex / 26) - 1); 491 devindex = devindex % 26; 492 } 493 snprintf(bdev->gendisk->disk_name + len, DISK_NAME_LEN - len, "%c", 494 'a' + devindex); 495 496 /* 512 byte sectors */ 497 set_capacity(bdev->gendisk, scmdev->size >> 9); 498 device_add_disk(&scmdev->dev, bdev->gendisk); 499 return 0; 500 501 out_queue: 502 blk_cleanup_queue(rq); 503 out_tag: 504 blk_mq_free_tag_set(&bdev->tag_set); 505 out: 506 atomic_dec(&nr_devices); 507 return ret; 508 } 509 510 void scm_blk_dev_cleanup(struct scm_blk_dev *bdev) 511 { 512 del_gendisk(bdev->gendisk); 513 blk_cleanup_queue(bdev->gendisk->queue); 514 blk_mq_free_tag_set(&bdev->tag_set); 515 put_disk(bdev->gendisk); 516 } 517 518 void scm_blk_set_available(struct scm_blk_dev *bdev) 519 { 520 unsigned long flags; 521 522 spin_lock_irqsave(&bdev->lock, flags); 523 if (bdev->state == SCM_WR_PROHIBIT) 524 pr_info("%lx: Write access to the SCM increment is restored\n", 525 (unsigned long) bdev->scmdev->address); 526 bdev->state = SCM_OPER; 527 spin_unlock_irqrestore(&bdev->lock, flags); 528 } 529 530 static bool __init scm_blk_params_valid(void) 531 { 532 if (!nr_requests_per_io || nr_requests_per_io > 64) 533 return false; 534 535 return true; 536 } 537 538 static int __init scm_blk_init(void) 539 { 540 int ret = -EINVAL; 541 542 if (!scm_blk_params_valid()) 543 goto out; 544 545 ret = register_blkdev(0, "scm"); 546 if (ret < 0) 547 goto out; 548 549 scm_major = ret; 550 ret = scm_alloc_rqs(nr_requests); 551 if (ret) 552 goto out_free; 553 554 scm_debug = debug_register("scm_log", 16, 1, 16); 555 if (!scm_debug) { 556 ret = -ENOMEM; 557 goto out_free; 558 } 559 560 debug_register_view(scm_debug, &debug_hex_ascii_view); 561 debug_set_level(scm_debug, 2); 562 563 ret = scm_drv_init(); 564 if (ret) 565 goto out_dbf; 566 567 return ret; 568 569 out_dbf: 570 debug_unregister(scm_debug); 571 out_free: 572 scm_free_rqs(); 573 unregister_blkdev(scm_major, "scm"); 574 out: 575 return ret; 576 } 577 module_init(scm_blk_init); 578 579 static void __exit scm_blk_cleanup(void) 580 { 581 scm_drv_cleanup(); 582 debug_unregister(scm_debug); 583 scm_free_rqs(); 584 unregister_blkdev(scm_major, "scm"); 585 } 586 module_exit(scm_blk_cleanup); 587