drivers/nvme/host/apple.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Apple ANS NVM Express device driver
   4  * Copyright The Asahi Linux Contributors
   5  *
   6  * Based on the pci.c NVM Express device driver
   7  * Copyright (c) 2011-2014, Intel Corporation.
   8  * and on the rdma.c NVMe over Fabrics RDMA host code.
   9  * Copyright (c) 2015-2016 HGST, a Western Digital Company.
  10  */
  11
  12 #include <linux/async.h>
  13 #include <linux/blkdev.h>
  14 #include <linux/blk-mq.h>
  15 #include <linux/device.h>
  16 #include <linux/dma-mapping.h>
  17 #include <linux/dmapool.h>
  18 #include <linux/interrupt.h>
  19 #include <linux/io-64-nonatomic-lo-hi.h>
  20 #include <linux/io.h>
  21 #include <linux/iopoll.h>
  22 #include <linux/jiffies.h>
  23 #include <linux/mempool.h>
  24 #include <linux/module.h>
  25 #include <linux/of.h>
  26 #include <linux/of_platform.h>
  27 #include <linux/once.h>
  28 #include <linux/platform_device.h>
  29 #include <linux/pm_domain.h>
  30 #include <linux/soc/apple/rtkit.h>
  31 #include <linux/soc/apple/sart.h>
  32 #include <linux/reset.h>
  33 #include <linux/time64.h>
  34
  35 #include "nvme.h"
  36
  37 #define APPLE_ANS_BOOT_TIMEOUT    USEC_PER_SEC
  38 #define APPLE_ANS_MAX_QUEUE_DEPTH 64
  39
  40 #define APPLE_ANS_COPROC_CPU_CONTROL     0x44
  41 #define APPLE_ANS_COPROC_CPU_CONTROL_RUN BIT(4)
  42
  43 #define APPLE_ANS_ACQ_DB  0x1004
  44 #define APPLE_ANS_IOCQ_DB 0x100c
  45
  46 #define APPLE_ANS_MAX_PEND_CMDS_CTRL 0x1210
  47
  48 #define APPLE_ANS_BOOT_STATUS    0x1300
  49 #define APPLE_ANS_BOOT_STATUS_OK 0xde71ce55
  50
  51 #define APPLE_ANS_UNKNOWN_CTRL   0x24008
  52 #define APPLE_ANS_PRP_NULL_CHECK BIT(11)
  53
  54 #define APPLE_ANS_LINEAR_SQ_CTRL 0x24908
  55 #define APPLE_ANS_LINEAR_SQ_EN   BIT(0)
  56
  57 #define APPLE_ANS_LINEAR_ASQ_DB  0x2490c
  58 #define APPLE_ANS_LINEAR_IOSQ_DB 0x24910
  59
  60 #define APPLE_NVMMU_NUM_TCBS      0x28100
  61 #define APPLE_NVMMU_ASQ_TCB_BASE  0x28108
  62 #define APPLE_NVMMU_IOSQ_TCB_BASE 0x28110
  63 #define APPLE_NVMMU_TCB_INVAL     0x28118
  64 #define APPLE_NVMMU_TCB_STAT      0x28120
  65
  66 /*
  67  * This controller is a bit weird in the way command tags works: Both the
  68  * admin and the IO queue share the same tag space. Additionally, tags
  69  * cannot be higher than 0x40 which effectively limits the combined
  70  * queue depth to 0x40. Instead of wasting half of that on the admin queue
  71  * which gets much less traffic we instead reduce its size here.
  72  * The controller also doesn't support async event such that no space must
  73  * be reserved for NVME_NR_AEN_COMMANDS.
  74  */
  75 #define APPLE_NVME_AQ_DEPTH        2
  76 #define APPLE_NVME_AQ_MQ_TAG_DEPTH (APPLE_NVME_AQ_DEPTH - 1)
  77
  78 /*
  79  * These can be higher, but we need to ensure that any command doesn't
  80  * require an sg allocation that needs more than a page of data.
  81  */
  82 #define NVME_MAX_KB_SZ 4096
  83 #define NVME_MAX_SEGS  127
  84
  85 /*
  86  * This controller comes with an embedded IOMMU known as NVMMU.
  87  * The NVMMU is pointed to an array of TCBs indexed by the command tag.
  88  * Each command must be configured inside this structure before it's allowed
  89  * to execute, including commands that don't require DMA transfers.
  90  *
  91  * An exception to this are Apple's vendor-specific commands (opcode 0xD8 on the
  92  * admin queue): Those commands must still be added to the NVMMU but the DMA
  93  * buffers cannot be represented as PRPs and must instead be allowed using SART.
  94  *
  95  * Programming the PRPs to the same values as those in the submission queue
  96  * looks rather silly at first. This hardware is however designed for a kernel
  97  * that runs the NVMMU code in a higher exception level than the NVMe driver.
  98  * In that setting the NVMe driver first programs the submission queue entry
  99  * and then executes a hypercall to the code that is allowed to program the
 100  * NVMMU. The NVMMU driver then creates a shadow copy of the PRPs while
 101  * verifying that they don't point to kernel text, data, pagetables, or similar
 102  * protected areas before programming the TCB to point to this shadow copy.
 103  * Since Linux doesn't do any of that we may as well just point both the queue
 104  * and the TCB PRP pointer to the same memory.
 105  */
 106 struct apple_nvmmu_tcb {
 107         u8 opcode;
 108
 109 #define APPLE_ANS_TCB_DMA_FROM_DEVICE BIT(0)
 110 #define APPLE_ANS_TCB_DMA_TO_DEVICE   BIT(1)
 111         u8 dma_flags;
 112
 113         u8 command_id;
 114         u8 _unk0;
 115         __le16 length;
 116         u8 _unk1[18];
 117         __le64 prp1;
 118         __le64 prp2;
 119         u8 _unk2[16];
 120         u8 aes_iv[8];
 121         u8 _aes_unk[64];
 122 };
 123
 124 /*
 125  * The Apple NVMe controller only supports a single admin and a single IO queue
 126  * which are both limited to 64 entries and share a single interrupt.
 127  *
 128  * The completion queue works as usual. The submission "queue" instead is
 129  * an array indexed by the command tag on this hardware. Commands must also be
 130  * present in the NVMMU's tcb array. They are triggered by writing their tag to
 131  * a MMIO register.
 132  */
 133 struct apple_nvme_queue {
 134         struct nvme_command *sqes;
 135         struct nvme_completion *cqes;
 136         struct apple_nvmmu_tcb *tcbs;
 137
 138         dma_addr_t sq_dma_addr;
 139         dma_addr_t cq_dma_addr;
 140         dma_addr_t tcb_dma_addr;
 141
 142         u32 __iomem *sq_db;
 143         u32 __iomem *cq_db;
 144
 145         u16 cq_head;
 146         u8 cq_phase;
 147
 148         bool is_adminq;
 149         bool enabled;
 150 };
 151
 152 /*
 153  * The apple_nvme_iod describes the data in an I/O.
 154  *
 155  * The sg pointer contains the list of PRP chunk allocations in addition
 156  * to the actual struct scatterlist.
 157  */
 158 struct apple_nvme_iod {
 159         struct nvme_request req;
 160         struct nvme_command cmd;
 161         struct apple_nvme_queue *q;
 162         int npages; /* In the PRP list. 0 means small pool in use */
 163         int nents; /* Used in scatterlist */
 164         dma_addr_t first_dma;
 165         unsigned int dma_len; /* length of single DMA segment mapping */
 166         struct scatterlist *sg;
 167 };
 168
 169 struct apple_nvme {
 170         struct device *dev;
 171
 172         void __iomem *mmio_coproc;
 173         void __iomem *mmio_nvme;
 174
 175         struct device **pd_dev;
 176         struct device_link **pd_link;
 177         int pd_count;
 178
 179         struct apple_sart *sart;
 180         struct apple_rtkit *rtk;
 181         struct reset_control *reset;
 182
 183         struct dma_pool *prp_page_pool;
 184         struct dma_pool *prp_small_pool;
 185         mempool_t *iod_mempool;
 186
 187         struct nvme_ctrl ctrl;
 188         struct work_struct remove_work;
 189
 190         struct apple_nvme_queue adminq;
 191         struct apple_nvme_queue ioq;
 192
 193         struct blk_mq_tag_set admin_tagset;
 194         struct blk_mq_tag_set tagset;
 195
 196         int irq;
 197         spinlock_t lock;
 198 };
 199
 200 static_assert(sizeof(struct nvme_command) == 64);
 201 static_assert(sizeof(struct apple_nvmmu_tcb) == 128);
 202
 203 static inline struct apple_nvme *ctrl_to_apple_nvme(struct nvme_ctrl *ctrl)
 204 {
 205         return container_of(ctrl, struct apple_nvme, ctrl);
 206 }
 207
 208 static inline struct apple_nvme *queue_to_apple_nvme(struct apple_nvme_queue *q)
 209 {
 210         if (q->is_adminq)
 211                 return container_of(q, struct apple_nvme, adminq);
 212         else
 213                 return container_of(q, struct apple_nvme, ioq);
 214 }
 215
 216 static unsigned int apple_nvme_queue_depth(struct apple_nvme_queue *q)
 217 {
 218         if (q->is_adminq)
 219                 return APPLE_NVME_AQ_DEPTH;
 220         else
 221                 return APPLE_ANS_MAX_QUEUE_DEPTH;
 222 }
 223
 224 static void apple_nvme_rtkit_crashed(void *cookie)
 225 {
 226         struct apple_nvme *anv = cookie;
 227
 228         dev_warn(anv->dev, "RTKit crashed; unable to recover without a reboot");
 229         nvme_reset_ctrl(&anv->ctrl);
 230 }
 231
 232 static int apple_nvme_sart_dma_setup(void *cookie,
 233                                      struct apple_rtkit_shmem *bfr)
 234 {
 235         struct apple_nvme *anv = cookie;
 236         int ret;
 237
 238         if (bfr->iova)
 239                 return -EINVAL;
 240         if (!bfr->size)
 241                 return -EINVAL;
 242
 243         bfr->buffer =
 244                 dma_alloc_coherent(anv->dev, bfr->size, &bfr->iova, GFP_KERNEL);
 245         if (!bfr->buffer)
 246                 return -ENOMEM;
 247
 248         ret = apple_sart_add_allowed_region(anv->sart, bfr->iova, bfr->size);
 249         if (ret) {
 250                 dma_free_coherent(anv->dev, bfr->size, bfr->buffer, bfr->iova);
 251                 bfr->buffer = NULL;
 252                 return -ENOMEM;
 253         }
 254
 255         return 0;
 256 }
 257
 258 static void apple_nvme_sart_dma_destroy(void *cookie,
 259                                         struct apple_rtkit_shmem *bfr)
 260 {
 261         struct apple_nvme *anv = cookie;
 262
 263         apple_sart_remove_allowed_region(anv->sart, bfr->iova, bfr->size);
 264         dma_free_coherent(anv->dev, bfr->size, bfr->buffer, bfr->iova);
 265 }
 266
 267 static const struct apple_rtkit_ops apple_nvme_rtkit_ops = {
 268         .crashed = apple_nvme_rtkit_crashed,
 269         .shmem_setup = apple_nvme_sart_dma_setup,
 270         .shmem_destroy = apple_nvme_sart_dma_destroy,
 271 };
 272
 273 static void apple_nvmmu_inval(struct apple_nvme_queue *q, unsigned int tag)
 274 {
 275         struct apple_nvme *anv = queue_to_apple_nvme(q);
 276
 277         writel(tag, anv->mmio_nvme + APPLE_NVMMU_TCB_INVAL);
 278         if (readl(anv->mmio_nvme + APPLE_NVMMU_TCB_STAT))
 279                 dev_warn_ratelimited(anv->dev,
 280                                      "NVMMU TCB invalidation failed\n");
 281 }
 282
 283 static void apple_nvme_submit_cmd(struct apple_nvme_queue *q,
 284                                   struct nvme_command *cmd)
 285 {
 286         struct apple_nvme *anv = queue_to_apple_nvme(q);
 287         u32 tag = nvme_tag_from_cid(cmd->common.command_id);
 288         struct apple_nvmmu_tcb *tcb = &q->tcbs[tag];
 289
 290         tcb->opcode = cmd->common.opcode;
 291         tcb->prp1 = cmd->common.dptr.prp1;
 292         tcb->prp2 = cmd->common.dptr.prp2;
 293         tcb->length = cmd->rw.length;
 294         tcb->command_id = tag;
 295
 296         if (nvme_is_write(cmd))
 297                 tcb->dma_flags = APPLE_ANS_TCB_DMA_TO_DEVICE;
 298         else
 299                 tcb->dma_flags = APPLE_ANS_TCB_DMA_FROM_DEVICE;
 300
 301         memcpy(&q->sqes[tag], cmd, sizeof(*cmd));
 302
 303         /*
 304          * This lock here doesn't make much sense at a first glace but
 305          * removing it will result in occasional missed completetion
 306          * interrupts even though the commands still appear on the CQ.
 307          * It's unclear why this happens but our best guess is that
 308          * there is a bug in the firmware triggered when a new command
 309          * is issued while we're inside the irq handler between the
 310          * NVMMU invalidation (and making the tag available again)
 311          * and the final CQ update.
 312          */
 313         spin_lock_irq(&anv->lock);
 314         writel(tag, q->sq_db);
 315         spin_unlock_irq(&anv->lock);
 316 }
 317
 318 /*
 319  * From pci.c:
 320  * Will slightly overestimate the number of pages needed.  This is OK
 321  * as it only leads to a small amount of wasted memory for the lifetime of
 322  * the I/O.
 323  */
 324 static inline size_t apple_nvme_iod_alloc_size(void)
 325 {
 326         const unsigned int nprps = DIV_ROUND_UP(
 327                 NVME_MAX_KB_SZ + NVME_CTRL_PAGE_SIZE, NVME_CTRL_PAGE_SIZE);
 328         const int npages = DIV_ROUND_UP(8 * nprps, PAGE_SIZE - 8);
 329         const size_t alloc_size = sizeof(__le64 *) * npages +
 330                                   sizeof(struct scatterlist) * NVME_MAX_SEGS;
 331
 332         return alloc_size;
 333 }
 334
 335 static void **apple_nvme_iod_list(struct request *req)
 336 {
 337         struct apple_nvme_iod *iod = blk_mq_rq_to_pdu(req);
 338
 339         return (void **)(iod->sg + blk_rq_nr_phys_segments(req));
 340 }
 341
 342 static void apple_nvme_free_prps(struct apple_nvme *anv, struct request *req)
 343 {
 344         const int last_prp = NVME_CTRL_PAGE_SIZE / sizeof(__le64) - 1;
 345         struct apple_nvme_iod *iod = blk_mq_rq_to_pdu(req);
 346         dma_addr_t dma_addr = iod->first_dma;
 347         int i;
 348
 349         for (i = 0; i < iod->npages; i++) {
 350                 __le64 *prp_list = apple_nvme_iod_list(req)[i];
 351                 dma_addr_t next_dma_addr = le64_to_cpu(prp_list[last_prp]);
 352
 353                 dma_pool_free(anv->prp_page_pool, prp_list, dma_addr);
 354                 dma_addr = next_dma_addr;
 355         }
 356 }
 357
 358 static void apple_nvme_unmap_data(struct apple_nvme *anv, struct request *req)
 359 {
 360         struct apple_nvme_iod *iod = blk_mq_rq_to_pdu(req);
 361
 362         if (iod->dma_len) {
 363                 dma_unmap_page(anv->dev, iod->first_dma, iod->dma_len,
 364                                rq_dma_dir(req));
 365                 return;
 366         }
 367
 368         WARN_ON_ONCE(!iod->nents);
 369
 370         dma_unmap_sg(anv->dev, iod->sg, iod->nents, rq_dma_dir(req));
 371         if (iod->npages == 0)
 372                 dma_pool_free(anv->prp_small_pool, apple_nvme_iod_list(req)[0],
 373                               iod->first_dma);
 374         else
 375                 apple_nvme_free_prps(anv, req);
 376         mempool_free(iod->sg, anv->iod_mempool);
 377 }
 378
 379 static void apple_nvme_print_sgl(struct scatterlist *sgl, int nents)
 380 {
 381         int i;
 382         struct scatterlist *sg;
 383
 384         for_each_sg(sgl, sg, nents, i) {
 385                 dma_addr_t phys = sg_phys(sg);
 386
 387                 pr_warn("sg[%d] phys_addr:%pad offset:%d length:%d dma_address:%pad dma_length:%d\n",
 388                         i, &phys, sg->offset, sg->length, &sg_dma_address(sg),
 389                         sg_dma_len(sg));
 390         }
 391 }
 392
 393 static blk_status_t apple_nvme_setup_prps(struct apple_nvme *anv,
 394                                           struct request *req,
 395                                           struct nvme_rw_command *cmnd)
 396 {
 397         struct apple_nvme_iod *iod = blk_mq_rq_to_pdu(req);
 398         struct dma_pool *pool;
 399         int length = blk_rq_payload_bytes(req);
 400         struct scatterlist *sg = iod->sg;
 401         int dma_len = sg_dma_len(sg);
 402         u64 dma_addr = sg_dma_address(sg);
 403         int offset = dma_addr & (NVME_CTRL_PAGE_SIZE - 1);
 404         __le64 *prp_list;
 405         void **list = apple_nvme_iod_list(req);
 406         dma_addr_t prp_dma;
 407         int nprps, i;
 408
 409         length -= (NVME_CTRL_PAGE_SIZE - offset);
 410         if (length <= 0) {
 411                 iod->first_dma = 0;
 412                 goto done;
 413         }
 414
 415         dma_len -= (NVME_CTRL_PAGE_SIZE - offset);
 416         if (dma_len) {
 417                 dma_addr += (NVME_CTRL_PAGE_SIZE - offset);
 418         } else {
 419                 sg = sg_next(sg);
 420                 dma_addr = sg_dma_address(sg);
 421                 dma_len = sg_dma_len(sg);
 422         }
 423
 424         if (length <= NVME_CTRL_PAGE_SIZE) {
 425                 iod->first_dma = dma_addr;
 426                 goto done;
 427         }
 428
 429         nprps = DIV_ROUND_UP(length, NVME_CTRL_PAGE_SIZE);
 430         if (nprps <= (256 / 8)) {
 431                 pool = anv->prp_small_pool;
 432                 iod->npages = 0;
 433         } else {
 434                 pool = anv->prp_page_pool;
 435                 iod->npages = 1;
 436         }
 437
 438         prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma);
 439         if (!prp_list) {
 440                 iod->first_dma = dma_addr;
 441                 iod->npages = -1;
 442                 return BLK_STS_RESOURCE;
 443         }
 444         list[0] = prp_list;
 445         iod->first_dma = prp_dma;
 446         i = 0;
 447         for (;;) {
 448                 if (i == NVME_CTRL_PAGE_SIZE >> 3) {
 449                         __le64 *old_prp_list = prp_list;
 450
 451                         prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma);
 452                         if (!prp_list)
 453                                 goto free_prps;
 454                         list[iod->npages++] = prp_list;
 455                         prp_list[0] = old_prp_list[i - 1];
 456                         old_prp_list[i - 1] = cpu_to_le64(prp_dma);
 457                         i = 1;
 458                 }
 459                 prp_list[i++] = cpu_to_le64(dma_addr);
 460                 dma_len -= NVME_CTRL_PAGE_SIZE;
 461                 dma_addr += NVME_CTRL_PAGE_SIZE;
 462                 length -= NVME_CTRL_PAGE_SIZE;
 463                 if (length <= 0)
 464                         break;
 465                 if (dma_len > 0)
 466                         continue;
 467                 if (unlikely(dma_len < 0))
 468                         goto bad_sgl;
 469                 sg = sg_next(sg);
 470                 dma_addr = sg_dma_address(sg);
 471                 dma_len = sg_dma_len(sg);
 472         }
 473 done:
 474         cmnd->dptr.prp1 = cpu_to_le64(sg_dma_address(iod->sg));
 475         cmnd->dptr.prp2 = cpu_to_le64(iod->first_dma);
 476         return BLK_STS_OK;
 477 free_prps:
 478         apple_nvme_free_prps(anv, req);
 479         return BLK_STS_RESOURCE;
 480 bad_sgl:
 481         WARN(DO_ONCE(apple_nvme_print_sgl, iod->sg, iod->nents),
 482              "Invalid SGL for payload:%d nents:%d\n", blk_rq_payload_bytes(req),
 483              iod->nents);
 484         return BLK_STS_IOERR;
 485 }
 486
 487 static blk_status_t apple_nvme_setup_prp_simple(struct apple_nvme *anv,
 488                                                 struct request *req,
 489                                                 struct nvme_rw_command *cmnd,
 490                                                 struct bio_vec *bv)
 491 {
 492         struct apple_nvme_iod *iod = blk_mq_rq_to_pdu(req);
 493         unsigned int offset = bv->bv_offset & (NVME_CTRL_PAGE_SIZE - 1);
 494         unsigned int first_prp_len = NVME_CTRL_PAGE_SIZE - offset;
 495
 496         iod->first_dma = dma_map_bvec(anv->dev, bv, rq_dma_dir(req), 0);
 497         if (dma_mapping_error(anv->dev, iod->first_dma))
 498                 return BLK_STS_RESOURCE;
 499         iod->dma_len = bv->bv_len;
 500
 501         cmnd->dptr.prp1 = cpu_to_le64(iod->first_dma);
 502         if (bv->bv_len > first_prp_len)
 503                 cmnd->dptr.prp2 = cpu_to_le64(iod->first_dma + first_prp_len);
 504         return BLK_STS_OK;
 505 }
 506
 507 static blk_status_t apple_nvme_map_data(struct apple_nvme *anv,
 508                                         struct request *req,
 509                                         struct nvme_command *cmnd)
 510 {
 511         struct apple_nvme_iod *iod = blk_mq_rq_to_pdu(req);
 512         blk_status_t ret = BLK_STS_RESOURCE;
 513         int nr_mapped;
 514
 515         if (blk_rq_nr_phys_segments(req) == 1) {
 516                 struct bio_vec bv = req_bvec(req);
 517
 518                 if (bv.bv_offset + bv.bv_len <= NVME_CTRL_PAGE_SIZE * 2)
 519                         return apple_nvme_setup_prp_simple(anv, req, &cmnd->rw,
 520                                                            &bv);
 521         }
 522
 523         iod->dma_len = 0;
 524         iod->sg = mempool_alloc(anv->iod_mempool, GFP_ATOMIC);
 525         if (!iod->sg)
 526                 return BLK_STS_RESOURCE;
 527         sg_init_table(iod->sg, blk_rq_nr_phys_segments(req));
 528         iod->nents = blk_rq_map_sg(req->q, req, iod->sg);
 529         if (!iod->nents)
 530                 goto out_free_sg;
 531
 532         nr_mapped = dma_map_sg_attrs(anv->dev, iod->sg, iod->nents,
 533                                      rq_dma_dir(req), DMA_ATTR_NO_WARN);
 534         if (!nr_mapped)
 535                 goto out_free_sg;
 536
 537         ret = apple_nvme_setup_prps(anv, req, &cmnd->rw);
 538         if (ret != BLK_STS_OK)
 539                 goto out_unmap_sg;
 540         return BLK_STS_OK;
 541
 542 out_unmap_sg:
 543         dma_unmap_sg(anv->dev, iod->sg, iod->nents, rq_dma_dir(req));
 544 out_free_sg:
 545         mempool_free(iod->sg, anv->iod_mempool);
 546         return ret;
 547 }
 548
 549 static __always_inline void apple_nvme_unmap_rq(struct request *req)
 550 {
 551         struct apple_nvme_iod *iod = blk_mq_rq_to_pdu(req);
 552         struct apple_nvme *anv = queue_to_apple_nvme(iod->q);
 553
 554         if (blk_rq_nr_phys_segments(req))
 555                 apple_nvme_unmap_data(anv, req);
 556 }
 557
 558 static void apple_nvme_complete_rq(struct request *req)
 559 {
 560         apple_nvme_unmap_rq(req);
 561         nvme_complete_rq(req);
 562 }
 563
 564 static void apple_nvme_complete_batch(struct io_comp_batch *iob)
 565 {
 566         nvme_complete_batch(iob, apple_nvme_unmap_rq);
 567 }
 568
 569 static inline bool apple_nvme_cqe_pending(struct apple_nvme_queue *q)
 570 {
 571         struct nvme_completion *hcqe = &q->cqes[q->cq_head];
 572
 573         return (le16_to_cpu(READ_ONCE(hcqe->status)) & 1) == q->cq_phase;
 574 }
 575
 576 static inline struct blk_mq_tags *
 577 apple_nvme_queue_tagset(struct apple_nvme *anv, struct apple_nvme_queue *q)
 578 {
 579         if (q->is_adminq)
 580                 return anv->admin_tagset.tags[0];
 581         else
 582                 return anv->tagset.tags[0];
 583 }
 584
 585 static inline void apple_nvme_handle_cqe(struct apple_nvme_queue *q,
 586                                          struct io_comp_batch *iob, u16 idx)
 587 {
 588         struct apple_nvme *anv = queue_to_apple_nvme(q);
 589         struct nvme_completion *cqe = &q->cqes[idx];
 590         __u16 command_id = READ_ONCE(cqe->command_id);
 591         struct request *req;
 592
 593         apple_nvmmu_inval(q, command_id);
 594
 595         req = nvme_find_rq(apple_nvme_queue_tagset(anv, q), command_id);
 596         if (unlikely(!req)) {
 597                 dev_warn(anv->dev, "invalid id %d completed", command_id);
 598                 return;
 599         }
 600
 601         if (!nvme_try_complete_req(req, cqe->status, cqe->result) &&
 602             !blk_mq_add_to_batch(req, iob, nvme_req(req)->status,
 603                                  apple_nvme_complete_batch))
 604                 apple_nvme_complete_rq(req);
 605 }
 606
 607 static inline void apple_nvme_update_cq_head(struct apple_nvme_queue *q)
 608 {
 609         u32 tmp = q->cq_head + 1;
 610
 611         if (tmp == apple_nvme_queue_depth(q)) {
 612                 q->cq_head = 0;
 613                 q->cq_phase ^= 1;
 614         } else {
 615                 q->cq_head = tmp;
 616         }
 617 }
 618
 619 static bool apple_nvme_poll_cq(struct apple_nvme_queue *q,
 620                                struct io_comp_batch *iob)
 621 {
 622         bool found = false;
 623
 624         while (apple_nvme_cqe_pending(q)) {
 625                 found = true;
 626
 627                 /*
 628                  * load-load control dependency between phase and the rest of
 629                  * the cqe requires a full read memory barrier
 630                  */
 631                 dma_rmb();
 632                 apple_nvme_handle_cqe(q, iob, q->cq_head);
 633                 apple_nvme_update_cq_head(q);
 634         }
 635
 636         if (found)
 637                 writel(q->cq_head, q->cq_db);
 638
 639         return found;
 640 }
 641
 642 static bool apple_nvme_handle_cq(struct apple_nvme_queue *q, bool force)
 643 {
 644         bool found;
 645         DEFINE_IO_COMP_BATCH(iob);
 646
 647         if (!READ_ONCE(q->enabled) && !force)
 648                 return false;
 649
 650         found = apple_nvme_poll_cq(q, &iob);
 651
 652         if (!rq_list_empty(iob.req_list))
 653                 apple_nvme_complete_batch(&iob);
 654
 655         return found;
 656 }
 657
 658 static irqreturn_t apple_nvme_irq(int irq, void *data)
 659 {
 660         struct apple_nvme *anv = data;
 661         bool handled = false;
 662         unsigned long flags;
 663
 664         spin_lock_irqsave(&anv->lock, flags);
 665         if (apple_nvme_handle_cq(&anv->ioq, false))
 666                 handled = true;
 667         if (apple_nvme_handle_cq(&anv->adminq, false))
 668                 handled = true;
 669         spin_unlock_irqrestore(&anv->lock, flags);
 670
 671         if (handled)
 672                 return IRQ_HANDLED;
 673         return IRQ_NONE;
 674 }
 675
 676 static int apple_nvme_create_cq(struct apple_nvme *anv)
 677 {
 678         struct nvme_command c = {};
 679
 680         /*
 681          * Note: we (ab)use the fact that the prp fields survive if no data
 682          * is attached to the request.
 683          */
 684         c.create_cq.opcode = nvme_admin_create_cq;
 685         c.create_cq.prp1 = cpu_to_le64(anv->ioq.cq_dma_addr);
 686         c.create_cq.cqid = cpu_to_le16(1);
 687         c.create_cq.qsize = cpu_to_le16(APPLE_ANS_MAX_QUEUE_DEPTH - 1);
 688         c.create_cq.cq_flags = cpu_to_le16(NVME_QUEUE_PHYS_CONTIG | NVME_CQ_IRQ_ENABLED);
 689         c.create_cq.irq_vector = cpu_to_le16(0);
 690
 691         return nvme_submit_sync_cmd(anv->ctrl.admin_q, &c, NULL, 0);
 692 }
 693
 694 static int apple_nvme_remove_cq(struct apple_nvme *anv)
 695 {
 696         struct nvme_command c = {};
 697
 698         c.delete_queue.opcode = nvme_admin_delete_cq;
 699         c.delete_queue.qid = cpu_to_le16(1);
 700
 701         return nvme_submit_sync_cmd(anv->ctrl.admin_q, &c, NULL, 0);
 702 }
 703
 704 static int apple_nvme_create_sq(struct apple_nvme *anv)
 705 {
 706         struct nvme_command c = {};
 707
 708         /*
 709          * Note: we (ab)use the fact that the prp fields survive if no data
 710          * is attached to the request.
 711          */
 712         c.create_sq.opcode = nvme_admin_create_sq;
 713         c.create_sq.prp1 = cpu_to_le64(anv->ioq.sq_dma_addr);
 714         c.create_sq.sqid = cpu_to_le16(1);
 715         c.create_sq.qsize = cpu_to_le16(APPLE_ANS_MAX_QUEUE_DEPTH - 1);
 716         c.create_sq.sq_flags = cpu_to_le16(NVME_QUEUE_PHYS_CONTIG);
 717         c.create_sq.cqid = cpu_to_le16(1);
 718
 719         return nvme_submit_sync_cmd(anv->ctrl.admin_q, &c, NULL, 0);
 720 }
 721
 722 static int apple_nvme_remove_sq(struct apple_nvme *anv)
 723 {
 724         struct nvme_command c = {};
 725
 726         c.delete_queue.opcode = nvme_admin_delete_sq;
 727         c.delete_queue.qid = cpu_to_le16(1);
 728
 729         return nvme_submit_sync_cmd(anv->ctrl.admin_q, &c, NULL, 0);
 730 }
 731
 732 static blk_status_t apple_nvme_queue_rq(struct blk_mq_hw_ctx *hctx,
 733                                         const struct blk_mq_queue_data *bd)
 734 {
 735         struct nvme_ns *ns = hctx->queue->queuedata;
 736         struct apple_nvme_queue *q = hctx->driver_data;
 737         struct apple_nvme *anv = queue_to_apple_nvme(q);
 738         struct request *req = bd->rq;
 739         struct apple_nvme_iod *iod = blk_mq_rq_to_pdu(req);
 740         struct nvme_command *cmnd = &iod->cmd;
 741         blk_status_t ret;
 742
 743         iod->npages = -1;
 744         iod->nents = 0;
 745
 746         /*
 747          * We should not need to do this, but we're still using this to
 748          * ensure we can drain requests on a dying queue.
 749          */
 750         if (unlikely(!READ_ONCE(q->enabled)))
 751                 return BLK_STS_IOERR;
 752
 753         if (!nvme_check_ready(&anv->ctrl, req, true))
 754                 return nvme_fail_nonready_command(&anv->ctrl, req);
 755
 756         ret = nvme_setup_cmd(ns, req);
 757         if (ret)
 758                 return ret;
 759
 760         if (blk_rq_nr_phys_segments(req)) {
 761                 ret = apple_nvme_map_data(anv, req, cmnd);
 762                 if (ret)
 763                         goto out_free_cmd;
 764         }
 765
 766         blk_mq_start_request(req);
 767         apple_nvme_submit_cmd(q, cmnd);
 768         return BLK_STS_OK;
 769
 770 out_free_cmd:
 771         nvme_cleanup_cmd(req);
 772         return ret;
 773 }
 774
 775 static int apple_nvme_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 776                                 unsigned int hctx_idx)
 777 {
 778         hctx->driver_data = data;
 779         return 0;
 780 }
 781
 782 static int apple_nvme_init_request(struct blk_mq_tag_set *set,
 783                                    struct request *req, unsigned int hctx_idx,
 784                                    unsigned int numa_node)
 785 {
 786         struct apple_nvme_queue *q = set->driver_data;
 787         struct apple_nvme *anv = queue_to_apple_nvme(q);
 788         struct apple_nvme_iod *iod = blk_mq_rq_to_pdu(req);
 789         struct nvme_request *nreq = nvme_req(req);
 790
 791         iod->q = q;
 792         nreq->ctrl = &anv->ctrl;
 793         nreq->cmd = &iod->cmd;
 794
 795         return 0;
 796 }
 797
 798 static void apple_nvme_disable(struct apple_nvme *anv, bool shutdown)
 799 {
 800         u32 csts = readl(anv->mmio_nvme + NVME_REG_CSTS);
 801         bool dead = false, freeze = false;
 802         unsigned long flags;
 803
 804         if (apple_rtkit_is_crashed(anv->rtk))
 805                 dead = true;
 806         if (!(csts & NVME_CSTS_RDY))
 807                 dead = true;
 808         if (csts & NVME_CSTS_CFS)
 809                 dead = true;
 810
 811         if (anv->ctrl.state == NVME_CTRL_LIVE ||
 812             anv->ctrl.state == NVME_CTRL_RESETTING) {
 813                 freeze = true;
 814                 nvme_start_freeze(&anv->ctrl);
 815         }
 816
 817         /*
 818          * Give the controller a chance to complete all entered requests if
 819          * doing a safe shutdown.
 820          */
 821         if (!dead && shutdown && freeze)
 822                 nvme_wait_freeze_timeout(&anv->ctrl, NVME_IO_TIMEOUT);
 823
 824         nvme_stop_queues(&anv->ctrl);
 825
 826         if (!dead) {
 827                 if (READ_ONCE(anv->ioq.enabled)) {
 828                         apple_nvme_remove_sq(anv);
 829                         apple_nvme_remove_cq(anv);
 830                 }
 831
 832                 if (shutdown)
 833                         nvme_shutdown_ctrl(&anv->ctrl);
 834                 nvme_disable_ctrl(&anv->ctrl);
 835         }
 836
 837         WRITE_ONCE(anv->ioq.enabled, false);
 838         WRITE_ONCE(anv->adminq.enabled, false);
 839         mb(); /* ensure that nvme_queue_rq() sees that enabled is cleared */
 840         nvme_stop_admin_queue(&anv->ctrl);
 841
 842         /* last chance to complete any requests before nvme_cancel_request */
 843         spin_lock_irqsave(&anv->lock, flags);
 844         apple_nvme_handle_cq(&anv->ioq, true);
 845         apple_nvme_handle_cq(&anv->adminq, true);
 846         spin_unlock_irqrestore(&anv->lock, flags);
 847
 848         blk_mq_tagset_busy_iter(&anv->tagset, nvme_cancel_request, &anv->ctrl);
 849         blk_mq_tagset_busy_iter(&anv->admin_tagset, nvme_cancel_request,
 850                                 &anv->ctrl);
 851         blk_mq_tagset_wait_completed_request(&anv->tagset);
 852         blk_mq_tagset_wait_completed_request(&anv->admin_tagset);
 853
 854         /*
 855          * The driver will not be starting up queues again if shutting down so
 856          * must flush all entered requests to their failed completion to avoid
 857          * deadlocking blk-mq hot-cpu notifier.
 858          */
 859         if (shutdown) {
 860                 nvme_start_queues(&anv->ctrl);
 861                 nvme_start_admin_queue(&anv->ctrl);
 862         }
 863 }
 864
 865 static enum blk_eh_timer_return apple_nvme_timeout(struct request *req,
 866                                                    bool reserved)
 867 {
 868         struct apple_nvme_iod *iod = blk_mq_rq_to_pdu(req);
 869         struct apple_nvme_queue *q = iod->q;
 870         struct apple_nvme *anv = queue_to_apple_nvme(q);
 871         unsigned long flags;
 872         u32 csts = readl(anv->mmio_nvme + NVME_REG_CSTS);
 873
 874         if (anv->ctrl.state != NVME_CTRL_LIVE) {
 875                 /*
 876                  * From rdma.c:
 877                  * If we are resetting, connecting or deleting we should
 878                  * complete immediately because we may block controller
 879                  * teardown or setup sequence
 880                  * - ctrl disable/shutdown fabrics requests
 881                  * - connect requests
 882                  * - initialization admin requests
 883                  * - I/O requests that entered after unquiescing and
 884                  *   the controller stopped responding
 885                  *
 886                  * All other requests should be cancelled by the error
 887                  * recovery work, so it's fine that we fail it here.
 888                  */
 889                 dev_warn(anv->dev,
 890                          "I/O %d(aq:%d) timeout while not in live state\n",
 891                          req->tag, q->is_adminq);
 892                 if (blk_mq_request_started(req) &&
 893                     !blk_mq_request_completed(req)) {
 894                         nvme_req(req)->status = NVME_SC_HOST_ABORTED_CMD;
 895                         nvme_req(req)->flags |= NVME_REQ_CANCELLED;
 896                         blk_mq_complete_request(req);
 897                 }
 898                 return BLK_EH_DONE;
 899         }
 900
 901         /* check if we just missed an interrupt if we're still alive */
 902         if (!apple_rtkit_is_crashed(anv->rtk) && !(csts & NVME_CSTS_CFS)) {
 903                 spin_lock_irqsave(&anv->lock, flags);
 904                 apple_nvme_handle_cq(q, false);
 905                 spin_unlock_irqrestore(&anv->lock, flags);
 906                 if (blk_mq_request_completed(req)) {
 907                         dev_warn(anv->dev,
 908                                  "I/O %d(aq:%d) timeout: completion polled\n",
 909                                  req->tag, q->is_adminq);
 910                         return BLK_EH_DONE;
 911                 }
 912         }
 913
 914         /*
 915          * aborting commands isn't supported which leaves a full reset as our
 916          * only option here
 917          */
 918         dev_warn(anv->dev, "I/O %d(aq:%d) timeout: resetting controller\n",
 919                  req->tag, q->is_adminq);
 920         nvme_req(req)->flags |= NVME_REQ_CANCELLED;
 921         apple_nvme_disable(anv, false);
 922         nvme_reset_ctrl(&anv->ctrl);
 923         return BLK_EH_DONE;
 924 }
 925
 926 static int apple_nvme_poll(struct blk_mq_hw_ctx *hctx,
 927                            struct io_comp_batch *iob)
 928 {
 929         struct apple_nvme_queue *q = hctx->driver_data;
 930         struct apple_nvme *anv = queue_to_apple_nvme(q);
 931         bool found;
 932         unsigned long flags;
 933
 934         spin_lock_irqsave(&anv->lock, flags);
 935         found = apple_nvme_poll_cq(q, iob);
 936         spin_unlock_irqrestore(&anv->lock, flags);
 937
 938         return found;
 939 }
 940
 941 static const struct blk_mq_ops apple_nvme_mq_admin_ops = {
 942         .queue_rq = apple_nvme_queue_rq,
 943         .complete = apple_nvme_complete_rq,
 944         .init_hctx = apple_nvme_init_hctx,
 945         .init_request = apple_nvme_init_request,
 946         .timeout = apple_nvme_timeout,
 947 };
 948
 949 static const struct blk_mq_ops apple_nvme_mq_ops = {
 950         .queue_rq = apple_nvme_queue_rq,
 951         .complete = apple_nvme_complete_rq,
 952         .init_hctx = apple_nvme_init_hctx,
 953         .init_request = apple_nvme_init_request,
 954         .timeout = apple_nvme_timeout,
 955         .poll = apple_nvme_poll,
 956 };
 957
 958 static void apple_nvme_init_queue(struct apple_nvme_queue *q)
 959 {
 960         unsigned int depth = apple_nvme_queue_depth(q);
 961
 962         q->cq_head = 0;
 963         q->cq_phase = 1;
 964         memset(q->tcbs, 0,
 965                APPLE_ANS_MAX_QUEUE_DEPTH * sizeof(struct apple_nvmmu_tcb));
 966         memset(q->cqes, 0, depth * sizeof(struct nvme_completion));
 967         WRITE_ONCE(q->enabled, true);
 968         wmb(); /* ensure the first interrupt sees the initialization */
 969 }
 970
 971 static void apple_nvme_reset_work(struct work_struct *work)
 972 {
 973         unsigned int nr_io_queues = 1;
 974         int ret;
 975         u32 boot_status, aqa;
 976         struct apple_nvme *anv =
 977                 container_of(work, struct apple_nvme, ctrl.reset_work);
 978
 979         if (anv->ctrl.state != NVME_CTRL_RESETTING) {
 980                 dev_warn(anv->dev, "ctrl state %d is not RESETTING\n",
 981                          anv->ctrl.state);
 982                 ret = -ENODEV;
 983                 goto out;
 984         }
 985
 986         /* there's unfortunately no known way to recover if RTKit crashed :( */
 987         if (apple_rtkit_is_crashed(anv->rtk)) {
 988                 dev_err(anv->dev,
 989                         "RTKit has crashed without any way to recover.");
 990                 ret = -EIO;
 991                 goto out;
 992         }
 993
 994         if (anv->ctrl.ctrl_config & NVME_CC_ENABLE)
 995                 apple_nvme_disable(anv, false);
 996
 997         /* RTKit must be shut down cleanly for the (soft)-reset to work */
 998         if (apple_rtkit_is_running(anv->rtk)) {
 999                 dev_dbg(anv->dev, "Trying to shut down RTKit before reset.");
1000                 ret = apple_rtkit_shutdown(anv->rtk);
1001                 if (ret)
1002                         goto out;
1003         }
1004
1005         writel(0, anv->mmio_coproc + APPLE_ANS_COPROC_CPU_CONTROL);
1006
1007         ret = reset_control_assert(anv->reset);
1008         if (ret)
1009                 goto out;
1010
1011         ret = apple_rtkit_reinit(anv->rtk);
1012         if (ret)
1013                 goto out;
1014
1015         ret = reset_control_deassert(anv->reset);
1016         if (ret)
1017                 goto out;
1018
1019         writel(APPLE_ANS_COPROC_CPU_CONTROL_RUN,
1020                anv->mmio_coproc + APPLE_ANS_COPROC_CPU_CONTROL);
1021         ret = apple_rtkit_boot(anv->rtk);
1022         if (ret) {
1023                 dev_err(anv->dev, "ANS did not boot");
1024                 goto out;
1025         }
1026
1027         ret = readl_poll_timeout(anv->mmio_nvme + APPLE_ANS_BOOT_STATUS,
1028                                  boot_status,
1029                                  boot_status == APPLE_ANS_BOOT_STATUS_OK,
1030                                  USEC_PER_MSEC, APPLE_ANS_BOOT_TIMEOUT);
1031         if (ret) {
1032                 dev_err(anv->dev, "ANS did not initialize");
1033                 goto out;
1034         }
1035
1036         dev_dbg(anv->dev, "ANS booted successfully.");
1037
1038         /*
1039          * Limit the max command size to prevent iod->sg allocations going
1040          * over a single page.
1041          */
1042         anv->ctrl.max_hw_sectors = min_t(u32, NVME_MAX_KB_SZ << 1,
1043                                          dma_max_mapping_size(anv->dev) >> 9);
1044         anv->ctrl.max_segments = NVME_MAX_SEGS;
1045
1046         /*
1047          * Enable NVMMU and linear submission queues.
1048          * While we could keep those disabled and pretend this is slightly
1049          * more common NVMe controller we'd still need some quirks (e.g.
1050          * sq entries will be 128 bytes) and Apple might drop support for
1051          * that mode in the future.
1052          */
1053         writel(APPLE_ANS_LINEAR_SQ_EN,
1054                anv->mmio_nvme + APPLE_ANS_LINEAR_SQ_CTRL);
1055
1056         /* Allow as many pending command as possible for both queues */
1057         writel(APPLE_ANS_MAX_QUEUE_DEPTH | (APPLE_ANS_MAX_QUEUE_DEPTH << 16),
1058                anv->mmio_nvme + APPLE_ANS_MAX_PEND_CMDS_CTRL);
1059
1060         /* Setup the NVMMU for the maximum admin and IO queue depth */
1061         writel(APPLE_ANS_MAX_QUEUE_DEPTH - 1,
1062                anv->mmio_nvme + APPLE_NVMMU_NUM_TCBS);
1063
1064         /*
1065          * This is probably a chicken bit: without it all commands where any PRP
1066          * is set to zero (including those that don't use that field) fail and
1067          * the co-processor complains about "completed with err BAD_CMD-" or
1068          * a "NULL_PRP_PTR_ERR" in the syslog
1069          */
1070         writel(readl(anv->mmio_nvme + APPLE_ANS_UNKNOWN_CTRL) &
1071                        ~APPLE_ANS_PRP_NULL_CHECK,
1072                anv->mmio_nvme + APPLE_ANS_UNKNOWN_CTRL);
1073
1074         /* Setup the admin queue */
1075         aqa = APPLE_NVME_AQ_DEPTH - 1;
1076         aqa |= aqa << 16;
1077         writel(aqa, anv->mmio_nvme + NVME_REG_AQA);
1078         writeq(anv->adminq.sq_dma_addr, anv->mmio_nvme + NVME_REG_ASQ);
1079         writeq(anv->adminq.cq_dma_addr, anv->mmio_nvme + NVME_REG_ACQ);
1080
1081         /* Setup NVMMU for both queues */
1082         writeq(anv->adminq.tcb_dma_addr,
1083                anv->mmio_nvme + APPLE_NVMMU_ASQ_TCB_BASE);
1084         writeq(anv->ioq.tcb_dma_addr,
1085                anv->mmio_nvme + APPLE_NVMMU_IOSQ_TCB_BASE);
1086
1087         anv->ctrl.sqsize =
1088                 APPLE_ANS_MAX_QUEUE_DEPTH - 1; /* 0's based queue depth */
1089         anv->ctrl.cap = readq(anv->mmio_nvme + NVME_REG_CAP);
1090
1091         dev_dbg(anv->dev, "Enabling controller now");
1092         ret = nvme_enable_ctrl(&anv->ctrl);
1093         if (ret)
1094                 goto out;
1095
1096         dev_dbg(anv->dev, "Starting admin queue");
1097         apple_nvme_init_queue(&anv->adminq);
1098         nvme_start_admin_queue(&anv->ctrl);
1099
1100         if (!nvme_change_ctrl_state(&anv->ctrl, NVME_CTRL_CONNECTING)) {
1101                 dev_warn(anv->ctrl.device,
1102                          "failed to mark controller CONNECTING\n");
1103                 ret = -ENODEV;
1104                 goto out;
1105         }
1106
1107         ret = nvme_init_ctrl_finish(&anv->ctrl);
1108         if (ret)
1109                 goto out;
1110
1111         dev_dbg(anv->dev, "Creating IOCQ");
1112         ret = apple_nvme_create_cq(anv);
1113         if (ret)
1114                 goto out;
1115         dev_dbg(anv->dev, "Creating IOSQ");
1116         ret = apple_nvme_create_sq(anv);
1117         if (ret)
1118                 goto out_remove_cq;
1119
1120         apple_nvme_init_queue(&anv->ioq);
1121         nr_io_queues = 1;
1122         ret = nvme_set_queue_count(&anv->ctrl, &nr_io_queues);
1123         if (ret)
1124                 goto out_remove_sq;
1125         if (nr_io_queues != 1) {
1126                 ret = -ENXIO;
1127                 goto out_remove_sq;
1128         }
1129
1130         anv->ctrl.queue_count = nr_io_queues + 1;
1131
1132         nvme_start_queues(&anv->ctrl);
1133         nvme_wait_freeze(&anv->ctrl);
1134         blk_mq_update_nr_hw_queues(&anv->tagset, 1);
1135         nvme_unfreeze(&anv->ctrl);
1136
1137         if (!nvme_change_ctrl_state(&anv->ctrl, NVME_CTRL_LIVE)) {
1138                 dev_warn(anv->ctrl.device,
1139                          "failed to mark controller live state\n");
1140                 ret = -ENODEV;
1141                 goto out_remove_sq;
1142         }
1143
1144         nvme_start_ctrl(&anv->ctrl);
1145
1146         dev_dbg(anv->dev, "ANS boot and NVMe init completed.");
1147         return;
1148
1149 out_remove_sq:
1150         apple_nvme_remove_sq(anv);
1151 out_remove_cq:
1152         apple_nvme_remove_cq(anv);
1153 out:
1154         dev_warn(anv->ctrl.device, "Reset failure status: %d\n", ret);
1155         nvme_change_ctrl_state(&anv->ctrl, NVME_CTRL_DELETING);
1156         nvme_get_ctrl(&anv->ctrl);
1157         apple_nvme_disable(anv, false);
1158         nvme_kill_queues(&anv->ctrl);
1159         if (!queue_work(nvme_wq, &anv->remove_work))
1160                 nvme_put_ctrl(&anv->ctrl);
1161 }
1162
1163 static void apple_nvme_remove_dead_ctrl_work(struct work_struct *work)
1164 {
1165         struct apple_nvme *anv =
1166                 container_of(work, struct apple_nvme, remove_work);
1167
1168         nvme_put_ctrl(&anv->ctrl);
1169         device_release_driver(anv->dev);
1170 }
1171
1172 static int apple_nvme_reg_read32(struct nvme_ctrl *ctrl, u32 off, u32 *val)
1173 {
1174         *val = readl(ctrl_to_apple_nvme(ctrl)->mmio_nvme + off);
1175         return 0;
1176 }
1177
1178 static int apple_nvme_reg_write32(struct nvme_ctrl *ctrl, u32 off, u32 val)
1179 {
1180         writel(val, ctrl_to_apple_nvme(ctrl)->mmio_nvme + off);
1181         return 0;
1182 }
1183
1184 static int apple_nvme_reg_read64(struct nvme_ctrl *ctrl, u32 off, u64 *val)
1185 {
1186         *val = readq(ctrl_to_apple_nvme(ctrl)->mmio_nvme + off);
1187         return 0;
1188 }
1189
1190 static int apple_nvme_get_address(struct nvme_ctrl *ctrl, char *buf, int size)
1191 {
1192         struct device *dev = ctrl_to_apple_nvme(ctrl)->dev;
1193
1194         return snprintf(buf, size, "%s\n", dev_name(dev));
1195 }
1196
1197 static void apple_nvme_free_ctrl(struct nvme_ctrl *ctrl)
1198 {
1199         struct apple_nvme *anv = ctrl_to_apple_nvme(ctrl);
1200
1201         if (anv->ctrl.admin_q)
1202                 blk_put_queue(anv->ctrl.admin_q);
1203         put_device(anv->dev);
1204 }
1205
1206 static const struct nvme_ctrl_ops nvme_ctrl_ops = {
1207         .name = "apple-nvme",
1208         .module = THIS_MODULE,
1209         .flags = 0,
1210         .reg_read32 = apple_nvme_reg_read32,
1211         .reg_write32 = apple_nvme_reg_write32,
1212         .reg_read64 = apple_nvme_reg_read64,
1213         .free_ctrl = apple_nvme_free_ctrl,
1214         .get_address = apple_nvme_get_address,
1215 };
1216
1217 static void apple_nvme_async_probe(void *data, async_cookie_t cookie)
1218 {
1219         struct apple_nvme *anv = data;
1220
1221         flush_work(&anv->ctrl.reset_work);
1222         flush_work(&anv->ctrl.scan_work);
1223         nvme_put_ctrl(&anv->ctrl);
1224 }
1225
1226 static int apple_nvme_alloc_tagsets(struct apple_nvme *anv)
1227 {
1228         int ret;
1229
1230         anv->admin_tagset.ops = &apple_nvme_mq_admin_ops;
1231         anv->admin_tagset.nr_hw_queues = 1;
1232         anv->admin_tagset.queue_depth = APPLE_NVME_AQ_MQ_TAG_DEPTH;
1233         anv->admin_tagset.timeout = NVME_ADMIN_TIMEOUT;
1234         anv->admin_tagset.numa_node = NUMA_NO_NODE;
1235         anv->admin_tagset.cmd_size = sizeof(struct apple_nvme_iod);
1236         anv->admin_tagset.flags = BLK_MQ_F_NO_SCHED;
1237         anv->admin_tagset.driver_data = &anv->adminq;
1238
1239         ret = blk_mq_alloc_tag_set(&anv->admin_tagset);
1240         if (ret)
1241                 return ret;
1242         ret = devm_add_action_or_reset(anv->dev,
1243                                        (void (*)(void *))blk_mq_free_tag_set,
1244                                        &anv->admin_tagset);
1245         if (ret)
1246                 return ret;
1247
1248         anv->tagset.ops = &apple_nvme_mq_ops;
1249         anv->tagset.nr_hw_queues = 1;
1250         anv->tagset.nr_maps = 1;
1251         /*
1252          * Tags are used as an index to the NVMMU and must be unique across
1253          * both queues. The admin queue gets the first APPLE_NVME_AQ_DEPTH which
1254          * must be marked as reserved in the IO queue.
1255          */
1256         anv->tagset.reserved_tags = APPLE_NVME_AQ_DEPTH;
1257         anv->tagset.queue_depth = APPLE_ANS_MAX_QUEUE_DEPTH - 1;
1258         anv->tagset.timeout = NVME_IO_TIMEOUT;
1259         anv->tagset.numa_node = NUMA_NO_NODE;
1260         anv->tagset.cmd_size = sizeof(struct apple_nvme_iod);
1261         anv->tagset.flags = BLK_MQ_F_SHOULD_MERGE;
1262         anv->tagset.driver_data = &anv->ioq;
1263
1264         ret = blk_mq_alloc_tag_set(&anv->tagset);
1265         if (ret)
1266                 return ret;
1267         ret = devm_add_action_or_reset(
1268                 anv->dev, (void (*)(void *))blk_mq_free_tag_set, &anv->tagset);
1269         if (ret)
1270                 return ret;
1271
1272         anv->ctrl.admin_tagset = &anv->admin_tagset;
1273         anv->ctrl.tagset = &anv->tagset;
1274
1275         return 0;
1276 }
1277
1278 static int apple_nvme_queue_alloc(struct apple_nvme *anv,
1279                                   struct apple_nvme_queue *q)
1280 {
1281         unsigned int depth = apple_nvme_queue_depth(q);
1282
1283         q->cqes = dmam_alloc_coherent(anv->dev,
1284                                       depth * sizeof(struct nvme_completion),
1285                                       &q->cq_dma_addr, GFP_KERNEL);
1286         if (!q->cqes)
1287                 return -ENOMEM;
1288
1289         q->sqes = dmam_alloc_coherent(anv->dev,
1290                                       depth * sizeof(struct nvme_command),
1291                                       &q->sq_dma_addr, GFP_KERNEL);
1292         if (!q->sqes)
1293                 return -ENOMEM;
1294
1295         /*
1296          * We need the maximum queue depth here because the NVMMU only has a
1297          * single depth configuration shared between both queues.
1298          */
1299         q->tcbs = dmam_alloc_coherent(anv->dev,
1300                                       APPLE_ANS_MAX_QUEUE_DEPTH *
1301                                               sizeof(struct apple_nvmmu_tcb),
1302                                       &q->tcb_dma_addr, GFP_KERNEL);
1303         if (!q->tcbs)
1304                 return -ENOMEM;
1305
1306         /*
1307          * initialize phase to make sure the allocated and empty memory
1308          * doesn't look like a full cq already.
1309          */
1310         q->cq_phase = 1;
1311         return 0;
1312 }
1313
1314 static void apple_nvme_detach_genpd(struct apple_nvme *anv)
1315 {
1316         int i;
1317
1318         if (anv->pd_count <= 1)
1319                 return;
1320
1321         for (i = anv->pd_count - 1; i >= 0; i--) {
1322                 if (anv->pd_link[i])
1323                         device_link_del(anv->pd_link[i]);
1324                 if (!IS_ERR_OR_NULL(anv->pd_dev[i]))
1325                         dev_pm_domain_detach(anv->pd_dev[i], true);
1326         }
1327 }
1328
1329 static int apple_nvme_attach_genpd(struct apple_nvme *anv)
1330 {
1331         struct device *dev = anv->dev;
1332         int i;
1333
1334         anv->pd_count = of_count_phandle_with_args(
1335                 dev->of_node, "power-domains", "#power-domain-cells");
1336         if (anv->pd_count <= 1)
1337                 return 0;
1338
1339         anv->pd_dev = devm_kcalloc(dev, anv->pd_count, sizeof(*anv->pd_dev),
1340                                    GFP_KERNEL);
1341         if (!anv->pd_dev)
1342                 return -ENOMEM;
1343
1344         anv->pd_link = devm_kcalloc(dev, anv->pd_count, sizeof(*anv->pd_link),
1345                                     GFP_KERNEL);
1346         if (!anv->pd_link)
1347                 return -ENOMEM;
1348
1349         for (i = 0; i < anv->pd_count; i++) {
1350                 anv->pd_dev[i] = dev_pm_domain_attach_by_id(dev, i);
1351                 if (IS_ERR(anv->pd_dev[i])) {
1352                         apple_nvme_detach_genpd(anv);
1353                         return PTR_ERR(anv->pd_dev[i]);
1354                 }
1355
1356                 anv->pd_link[i] = device_link_add(dev, anv->pd_dev[i],
1357                                                   DL_FLAG_STATELESS |
1358                                                   DL_FLAG_PM_RUNTIME |
1359                                                   DL_FLAG_RPM_ACTIVE);
1360                 if (!anv->pd_link[i]) {
1361                         apple_nvme_detach_genpd(anv);
1362                         return -EINVAL;
1363                 }
1364         }
1365
1366         return 0;
1367 }
1368
1369 static int apple_nvme_probe(struct platform_device *pdev)
1370 {
1371         struct device *dev = &pdev->dev;
1372         struct apple_nvme *anv;
1373         int ret;
1374
1375         anv = devm_kzalloc(dev, sizeof(*anv), GFP_KERNEL);
1376         if (!anv)
1377                 return -ENOMEM;
1378
1379         anv->dev = get_device(dev);
1380         anv->adminq.is_adminq = true;
1381         platform_set_drvdata(pdev, anv);
1382
1383         ret = apple_nvme_attach_genpd(anv);
1384         if (ret < 0) {
1385                 dev_err_probe(dev, ret, "Failed to attach power domains");
1386                 goto put_dev;
1387         }
1388         if (dma_set_mask_and_coherent(dev, DMA_BIT_MASK(64))) {
1389                 ret = -ENXIO;
1390                 goto put_dev;
1391         }
1392
1393         anv->irq = platform_get_irq(pdev, 0);
1394         if (anv->irq < 0) {
1395                 ret = anv->irq;
1396                 goto put_dev;
1397         }
1398         if (!anv->irq) {
1399                 ret = -ENXIO;
1400                 goto put_dev;
1401         }
1402
1403         anv->mmio_coproc = devm_platform_ioremap_resource_byname(pdev, "ans");
1404         if (IS_ERR(anv->mmio_coproc)) {
1405                 ret = PTR_ERR(anv->mmio_coproc);
1406                 goto put_dev;
1407         }
1408         anv->mmio_nvme = devm_platform_ioremap_resource_byname(pdev, "nvme");
1409         if (IS_ERR(anv->mmio_nvme)) {
1410                 ret = PTR_ERR(anv->mmio_nvme);
1411                 goto put_dev;
1412         }
1413
1414         anv->adminq.sq_db = anv->mmio_nvme + APPLE_ANS_LINEAR_ASQ_DB;
1415         anv->adminq.cq_db = anv->mmio_nvme + APPLE_ANS_ACQ_DB;
1416         anv->ioq.sq_db = anv->mmio_nvme + APPLE_ANS_LINEAR_IOSQ_DB;
1417         anv->ioq.cq_db = anv->mmio_nvme + APPLE_ANS_IOCQ_DB;
1418
1419         anv->sart = devm_apple_sart_get(dev);
1420         if (IS_ERR(anv->sart)) {
1421                 ret = dev_err_probe(dev, PTR_ERR(anv->sart),
1422                                     "Failed to initialize SART");
1423                 goto put_dev;
1424         }
1425
1426         anv->reset = devm_reset_control_array_get_exclusive(anv->dev);
1427         if (IS_ERR(anv->reset)) {
1428                 ret = dev_err_probe(dev, PTR_ERR(anv->reset),
1429                                     "Failed to get reset control");
1430                 goto put_dev;
1431         }
1432
1433         INIT_WORK(&anv->ctrl.reset_work, apple_nvme_reset_work);
1434         INIT_WORK(&anv->remove_work, apple_nvme_remove_dead_ctrl_work);
1435         spin_lock_init(&anv->lock);
1436
1437         ret = apple_nvme_queue_alloc(anv, &anv->adminq);
1438         if (ret)
1439                 goto put_dev;
1440         ret = apple_nvme_queue_alloc(anv, &anv->ioq);
1441         if (ret)
1442                 goto put_dev;
1443
1444         anv->prp_page_pool = dmam_pool_create("prp list page", anv->dev,
1445                                               NVME_CTRL_PAGE_SIZE,
1446                                               NVME_CTRL_PAGE_SIZE, 0);
1447         if (!anv->prp_page_pool) {
1448                 ret = -ENOMEM;
1449                 goto put_dev;
1450         }
1451
1452         anv->prp_small_pool =
1453                 dmam_pool_create("prp list 256", anv->dev, 256, 256, 0);
1454         if (!anv->prp_small_pool) {
1455                 ret = -ENOMEM;
1456                 goto put_dev;
1457         }
1458
1459         WARN_ON_ONCE(apple_nvme_iod_alloc_size() > PAGE_SIZE);
1460         anv->iod_mempool =
1461                 mempool_create_kmalloc_pool(1, apple_nvme_iod_alloc_size());
1462         if (!anv->iod_mempool) {
1463                 ret = -ENOMEM;
1464                 goto put_dev;
1465         }
1466         ret = devm_add_action_or_reset(
1467                 anv->dev, (void (*)(void *))mempool_destroy, anv->iod_mempool);
1468         if (ret)
1469                 goto put_dev;
1470
1471         ret = apple_nvme_alloc_tagsets(anv);
1472         if (ret)
1473                 goto put_dev;
1474
1475         ret = devm_request_irq(anv->dev, anv->irq, apple_nvme_irq, 0,
1476                                "nvme-apple", anv);
1477         if (ret) {
1478                 dev_err_probe(dev, ret, "Failed to request IRQ");
1479                 goto put_dev;
1480         }
1481
1482         anv->rtk =
1483                 devm_apple_rtkit_init(dev, anv, NULL, 0, &apple_nvme_rtkit_ops);
1484         if (IS_ERR(anv->rtk)) {
1485                 ret = dev_err_probe(dev, PTR_ERR(anv->rtk),
1486                                     "Failed to initialize RTKit");
1487                 goto put_dev;
1488         }
1489
1490         ret = nvme_init_ctrl(&anv->ctrl, anv->dev, &nvme_ctrl_ops,
1491                              NVME_QUIRK_SKIP_CID_GEN);
1492         if (ret) {
1493                 dev_err_probe(dev, ret, "Failed to initialize nvme_ctrl");
1494                 goto put_dev;
1495         }
1496
1497         anv->ctrl.admin_q = blk_mq_init_queue(&anv->admin_tagset);
1498         if (IS_ERR(anv->ctrl.admin_q)) {
1499                 ret = -ENOMEM;
1500                 goto put_dev;
1501         }
1502
1503         if (!blk_get_queue(anv->ctrl.admin_q)) {
1504                 nvme_start_admin_queue(&anv->ctrl);
1505                 blk_cleanup_queue(anv->ctrl.admin_q);
1506                 anv->ctrl.admin_q = NULL;
1507                 ret = -ENODEV;
1508                 goto put_dev;
1509         }
1510
1511         nvme_reset_ctrl(&anv->ctrl);
1512         async_schedule(apple_nvme_async_probe, anv);
1513
1514         return 0;
1515
1516 put_dev:
1517         put_device(anv->dev);
1518         return ret;
1519 }
1520
1521 static int apple_nvme_remove(struct platform_device *pdev)
1522 {
1523         struct apple_nvme *anv = platform_get_drvdata(pdev);
1524
1525         nvme_change_ctrl_state(&anv->ctrl, NVME_CTRL_DELETING);
1526         flush_work(&anv->ctrl.reset_work);
1527         nvme_stop_ctrl(&anv->ctrl);
1528         nvme_remove_namespaces(&anv->ctrl);
1529         apple_nvme_disable(anv, true);
1530         nvme_uninit_ctrl(&anv->ctrl);
1531
1532         if (apple_rtkit_is_running(anv->rtk))
1533                 apple_rtkit_shutdown(anv->rtk);
1534
1535         apple_nvme_detach_genpd(anv);
1536
1537         return 0;
1538 }
1539
1540 static void apple_nvme_shutdown(struct platform_device *pdev)
1541 {
1542         struct apple_nvme *anv = platform_get_drvdata(pdev);
1543
1544         apple_nvme_disable(anv, true);
1545         if (apple_rtkit_is_running(anv->rtk))
1546                 apple_rtkit_shutdown(anv->rtk);
1547 }
1548
1549 static int apple_nvme_resume(struct device *dev)
1550 {
1551         struct apple_nvme *anv = dev_get_drvdata(dev);
1552
1553         return nvme_reset_ctrl(&anv->ctrl);
1554 }
1555
1556 static int apple_nvme_suspend(struct device *dev)
1557 {
1558         struct apple_nvme *anv = dev_get_drvdata(dev);
1559         int ret = 0;
1560
1561         apple_nvme_disable(anv, true);
1562
1563         if (apple_rtkit_is_running(anv->rtk))
1564                 ret = apple_rtkit_shutdown(anv->rtk);
1565
1566         writel(0, anv->mmio_coproc + APPLE_ANS_COPROC_CPU_CONTROL);
1567
1568         return ret;
1569 }
1570
1571 static DEFINE_SIMPLE_DEV_PM_OPS(apple_nvme_pm_ops, apple_nvme_suspend,
1572                                 apple_nvme_resume);
1573
1574 static const struct of_device_id apple_nvme_of_match[] = {
1575         { .compatible = "apple,nvme-ans2" },
1576         {},
1577 };
1578 MODULE_DEVICE_TABLE(of, apple_nvme_of_match);
1579
1580 static struct platform_driver apple_nvme_driver = {
1581         .driver = {
1582                 .name = "nvme-apple",
1583                 .of_match_table = apple_nvme_of_match,
1584                 .pm = pm_sleep_ptr(&apple_nvme_pm_ops),
1585         },
1586         .probe = apple_nvme_probe,
1587         .remove = apple_nvme_remove,
1588         .shutdown = apple_nvme_shutdown,
1589 };
1590 module_platform_driver(apple_nvme_driver);
1591
1592 MODULE_AUTHOR("Sven Peter <sven@svenpeter.dev>");
1593 MODULE_LICENSE("GPL");