drivers/usb/host/xhci-mem.c

   1 /*
   2  * xHCI host controller driver
   3  *
   4  * Copyright (C) 2008 Intel Corp.
   5  *
   6  * Author: Sarah Sharp
   7  * Some code borrowed from the Linux EHCI driver.
   8  *
   9  * This program is free software; you can redistribute it and/or modify
  10  * it under the terms of the GNU General Public License version 2 as
  11  * published by the Free Software Foundation.
  12  *
  13  * This program is distributed in the hope that it will be useful, but
  14  * WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
  15  * or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  16  * for more details.
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * along with this program; if not, write to the Free Software Foundation,
  20  * Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  21  */
  22
  23 #include <linux/usb.h>
  24 #include <linux/pci.h>
  25 #include <linux/slab.h>
  26 #include <linux/dmapool.h>
  27
  28 #include "xhci.h"
  29
  30 /*
  31  * Allocates a generic ring segment from the ring pool, sets the dma address,
  32  * initializes the segment to zero, and sets the private next pointer to NULL.
  33  *
  34  * Section 4.11.1.1:
  35  * "All components of all Command and Transfer TRBs shall be initialized to '0'"
  36  */
  37 static struct xhci_segment *xhci_segment_alloc(struct xhci_hcd *xhci, gfp_t flags)
  38 {
  39         struct xhci_segment *seg;
  40         dma_addr_t      dma;
  41
  42         seg = kzalloc(sizeof *seg, flags);
  43         if (!seg)
  44                 return NULL;
  45         xhci_dbg(xhci, "Allocating priv segment structure at %p\n", seg);
  46
  47         seg->trbs = dma_pool_alloc(xhci->segment_pool, flags, &dma);
  48         if (!seg->trbs) {
  49                 kfree(seg);
  50                 return NULL;
  51         }
  52         xhci_dbg(xhci, "// Allocating segment at %p (virtual) 0x%llx (DMA)\n",
  53                         seg->trbs, (unsigned long long)dma);
  54
  55         memset(seg->trbs, 0, SEGMENT_SIZE);
  56         seg->dma = dma;
  57         seg->next = NULL;
  58
  59         return seg;
  60 }
  61
  62 static void xhci_segment_free(struct xhci_hcd *xhci, struct xhci_segment *seg)
  63 {
  64         if (!seg)
  65                 return;
  66         if (seg->trbs) {
  67                 xhci_dbg(xhci, "Freeing DMA segment at %p (virtual) 0x%llx (DMA)\n",
  68                                 seg->trbs, (unsigned long long)seg->dma);
  69                 dma_pool_free(xhci->segment_pool, seg->trbs, seg->dma);
  70                 seg->trbs = NULL;
  71         }
  72         xhci_dbg(xhci, "Freeing priv segment structure at %p\n", seg);
  73         kfree(seg);
  74 }
  75
  76 /*
  77  * Make the prev segment point to the next segment.
  78  *
  79  * Change the last TRB in the prev segment to be a Link TRB which points to the
  80  * DMA address of the next segment.  The caller needs to set any Link TRB
  81  * related flags, such as End TRB, Toggle Cycle, and no snoop.
  82  */
  83 static void xhci_link_segments(struct xhci_hcd *xhci, struct xhci_segment *prev,
  84                 struct xhci_segment *next, bool link_trbs)
  85 {
  86         u32 val;
  87
  88         if (!prev || !next)
  89                 return;
  90         prev->next = next;
  91         if (link_trbs) {
  92                 prev->trbs[TRBS_PER_SEGMENT-1].link.segment_ptr =
  93                         cpu_to_le64(next->dma);
  94
  95                 /* Set the last TRB in the segment to have a TRB type ID of Link TRB */
  96                 val = le32_to_cpu(prev->trbs[TRBS_PER_SEGMENT-1].link.control);
  97                 val &= ~TRB_TYPE_BITMASK;
  98                 val |= TRB_TYPE(TRB_LINK);
  99                 /* Always set the chain bit with 0.95 hardware */
 100                 if (xhci_link_trb_quirk(xhci))
 101                         val |= TRB_CHAIN;
 102                 prev->trbs[TRBS_PER_SEGMENT-1].link.control = cpu_to_le32(val);
 103         }
 104         xhci_dbg(xhci, "Linking segment 0x%llx to segment 0x%llx (DMA)\n",
 105                         (unsigned long long)prev->dma,
 106                         (unsigned long long)next->dma);
 107 }
 108
 109 /* XXX: Do we need the hcd structure in all these functions? */
 110 void xhci_ring_free(struct xhci_hcd *xhci, struct xhci_ring *ring)
 111 {
 112         struct xhci_segment *seg;
 113         struct xhci_segment *first_seg;
 114
 115         if (!ring || !ring->first_seg)
 116                 return;
 117         first_seg = ring->first_seg;
 118         seg = first_seg->next;
 119         xhci_dbg(xhci, "Freeing ring at %p\n", ring);
 120         while (seg != first_seg) {
 121                 struct xhci_segment *next = seg->next;
 122                 xhci_segment_free(xhci, seg);
 123                 seg = next;
 124         }
 125         xhci_segment_free(xhci, first_seg);
 126         ring->first_seg = NULL;
 127         kfree(ring);
 128 }
 129
 130 static void xhci_initialize_ring_info(struct xhci_ring *ring)
 131 {
 132         /* The ring is empty, so the enqueue pointer == dequeue pointer */
 133         ring->enqueue = ring->first_seg->trbs;
 134         ring->enq_seg = ring->first_seg;
 135         ring->dequeue = ring->enqueue;
 136         ring->deq_seg = ring->first_seg;
 137         /* The ring is initialized to 0. The producer must write 1 to the cycle
 138          * bit to handover ownership of the TRB, so PCS = 1.  The consumer must
 139          * compare CCS to the cycle bit to check ownership, so CCS = 1.
 140          */
 141         ring->cycle_state = 1;
 142         /* Not necessary for new rings, but needed for re-initialized rings */
 143         ring->enq_updates = 0;
 144         ring->deq_updates = 0;
 145 }
 146
 147 /**
 148  * Create a new ring with zero or more segments.
 149  *
 150  * Link each segment together into a ring.
 151  * Set the end flag and the cycle toggle bit on the last segment.
 152  * See section 4.9.1 and figures 15 and 16.
 153  */
 154 static struct xhci_ring *xhci_ring_alloc(struct xhci_hcd *xhci,
 155                 unsigned int num_segs, bool link_trbs, gfp_t flags)
 156 {
 157         struct xhci_ring        *ring;
 158         struct xhci_segment     *prev;
 159
 160         ring = kzalloc(sizeof *(ring), flags);
 161         xhci_dbg(xhci, "Allocating ring at %p\n", ring);
 162         if (!ring)
 163                 return NULL;
 164
 165         INIT_LIST_HEAD(&ring->td_list);
 166         if (num_segs == 0)
 167                 return ring;
 168
 169         ring->first_seg = xhci_segment_alloc(xhci, flags);
 170         if (!ring->first_seg)
 171                 goto fail;
 172         num_segs--;
 173
 174         prev = ring->first_seg;
 175         while (num_segs > 0) {
 176                 struct xhci_segment     *next;
 177
 178                 next = xhci_segment_alloc(xhci, flags);
 179                 if (!next)
 180                         goto fail;
 181                 xhci_link_segments(xhci, prev, next, link_trbs);
 182
 183                 prev = next;
 184                 num_segs--;
 185         }
 186         xhci_link_segments(xhci, prev, ring->first_seg, link_trbs);
 187
 188         if (link_trbs) {
 189                 /* See section 4.9.2.1 and 6.4.4.1 */
 190                 prev->trbs[TRBS_PER_SEGMENT-1].link.control |=
 191                         cpu_to_le32(LINK_TOGGLE);
 192                 xhci_dbg(xhci, "Wrote link toggle flag to"
 193                                 " segment %p (virtual), 0x%llx (DMA)\n",
 194                                 prev, (unsigned long long)prev->dma);
 195         }
 196         xhci_initialize_ring_info(ring);
 197         return ring;
 198
 199 fail:
 200         xhci_ring_free(xhci, ring);
 201         return NULL;
 202 }
 203
 204 void xhci_free_or_cache_endpoint_ring(struct xhci_hcd *xhci,
 205                 struct xhci_virt_device *virt_dev,
 206                 unsigned int ep_index)
 207 {
 208         int rings_cached;
 209
 210         rings_cached = virt_dev->num_rings_cached;
 211         if (rings_cached < XHCI_MAX_RINGS_CACHED) {
 212                 virt_dev->ring_cache[rings_cached] =
 213                         virt_dev->eps[ep_index].ring;
 214                 virt_dev->num_rings_cached++;
 215                 xhci_dbg(xhci, "Cached old ring, "
 216                                 "%d ring%s cached\n",
 217                                 virt_dev->num_rings_cached,
 218                                 (virt_dev->num_rings_cached > 1) ? "s" : "");
 219         } else {
 220                 xhci_ring_free(xhci, virt_dev->eps[ep_index].ring);
 221                 xhci_dbg(xhci, "Ring cache full (%d rings), "
 222                                 "freeing ring\n",
 223                                 virt_dev->num_rings_cached);
 224         }
 225         virt_dev->eps[ep_index].ring = NULL;
 226 }
 227
 228 /* Zero an endpoint ring (except for link TRBs) and move the enqueue and dequeue
 229  * pointers to the beginning of the ring.
 230  */
 231 static void xhci_reinit_cached_ring(struct xhci_hcd *xhci,
 232                 struct xhci_ring *ring)
 233 {
 234         struct xhci_segment     *seg = ring->first_seg;
 235         do {
 236                 memset(seg->trbs, 0,
 237                                 sizeof(union xhci_trb)*TRBS_PER_SEGMENT);
 238                 /* All endpoint rings have link TRBs */
 239                 xhci_link_segments(xhci, seg, seg->next, 1);
 240                 seg = seg->next;
 241         } while (seg != ring->first_seg);
 242         xhci_initialize_ring_info(ring);
 243         /* td list should be empty since all URBs have been cancelled,
 244          * but just in case...
 245          */
 246         INIT_LIST_HEAD(&ring->td_list);
 247 }
 248
 249 #define CTX_SIZE(_hcc) (HCC_64BYTE_CONTEXT(_hcc) ? 64 : 32)
 250
 251 static struct xhci_container_ctx *xhci_alloc_container_ctx(struct xhci_hcd *xhci,
 252                                                     int type, gfp_t flags)
 253 {
 254         struct xhci_container_ctx *ctx = kzalloc(sizeof(*ctx), flags);
 255         if (!ctx)
 256                 return NULL;
 257
 258         BUG_ON((type != XHCI_CTX_TYPE_DEVICE) && (type != XHCI_CTX_TYPE_INPUT));
 259         ctx->type = type;
 260         ctx->size = HCC_64BYTE_CONTEXT(xhci->hcc_params) ? 2048 : 1024;
 261         if (type == XHCI_CTX_TYPE_INPUT)
 262                 ctx->size += CTX_SIZE(xhci->hcc_params);
 263
 264         ctx->bytes = dma_pool_alloc(xhci->device_pool, flags, &ctx->dma);
 265         memset(ctx->bytes, 0, ctx->size);
 266         return ctx;
 267 }
 268
 269 static void xhci_free_container_ctx(struct xhci_hcd *xhci,
 270                              struct xhci_container_ctx *ctx)
 271 {
 272         if (!ctx)
 273                 return;
 274         dma_pool_free(xhci->device_pool, ctx->bytes, ctx->dma);
 275         kfree(ctx);
 276 }
 277
 278 struct xhci_input_control_ctx *xhci_get_input_control_ctx(struct xhci_hcd *xhci,
 279                                               struct xhci_container_ctx *ctx)
 280 {
 281         BUG_ON(ctx->type != XHCI_CTX_TYPE_INPUT);
 282         return (struct xhci_input_control_ctx *)ctx->bytes;
 283 }
 284
 285 struct xhci_slot_ctx *xhci_get_slot_ctx(struct xhci_hcd *xhci,
 286                                         struct xhci_container_ctx *ctx)
 287 {
 288         if (ctx->type == XHCI_CTX_TYPE_DEVICE)
 289                 return (struct xhci_slot_ctx *)ctx->bytes;
 290
 291         return (struct xhci_slot_ctx *)
 292                 (ctx->bytes + CTX_SIZE(xhci->hcc_params));
 293 }
 294
 295 struct xhci_ep_ctx *xhci_get_ep_ctx(struct xhci_hcd *xhci,
 296                                     struct xhci_container_ctx *ctx,
 297                                     unsigned int ep_index)
 298 {
 299         /* increment ep index by offset of start of ep ctx array */
 300         ep_index++;
 301         if (ctx->type == XHCI_CTX_TYPE_INPUT)
 302                 ep_index++;
 303
 304         return (struct xhci_ep_ctx *)
 305                 (ctx->bytes + (ep_index * CTX_SIZE(xhci->hcc_params)));
 306 }
 307
 308
 309 /***************** Streams structures manipulation *************************/
 310
 311 static void xhci_free_stream_ctx(struct xhci_hcd *xhci,
 312                 unsigned int num_stream_ctxs,
 313                 struct xhci_stream_ctx *stream_ctx, dma_addr_t dma)
 314 {
 315         struct pci_dev *pdev = to_pci_dev(xhci_to_hcd(xhci)->self.controller);
 316
 317         if (num_stream_ctxs > MEDIUM_STREAM_ARRAY_SIZE)
 318                 pci_free_consistent(pdev,
 319                                 sizeof(struct xhci_stream_ctx)*num_stream_ctxs,
 320                                 stream_ctx, dma);
 321         else if (num_stream_ctxs <= SMALL_STREAM_ARRAY_SIZE)
 322                 return dma_pool_free(xhci->small_streams_pool,
 323                                 stream_ctx, dma);
 324         else
 325                 return dma_pool_free(xhci->medium_streams_pool,
 326                                 stream_ctx, dma);
 327 }
 328
 329 /*
 330  * The stream context array for each endpoint with bulk streams enabled can
 331  * vary in size, based on:
 332  *  - how many streams the endpoint supports,
 333  *  - the maximum primary stream array size the host controller supports,
 334  *  - and how many streams the device driver asks for.
 335  *
 336  * The stream context array must be a power of 2, and can be as small as
 337  * 64 bytes or as large as 1MB.
 338  */
 339 static struct xhci_stream_ctx *xhci_alloc_stream_ctx(struct xhci_hcd *xhci,
 340                 unsigned int num_stream_ctxs, dma_addr_t *dma,
 341                 gfp_t mem_flags)
 342 {
 343         struct pci_dev *pdev = to_pci_dev(xhci_to_hcd(xhci)->self.controller);
 344
 345         if (num_stream_ctxs > MEDIUM_STREAM_ARRAY_SIZE)
 346                 return pci_alloc_consistent(pdev,
 347                                 sizeof(struct xhci_stream_ctx)*num_stream_ctxs,
 348                                 dma);
 349         else if (num_stream_ctxs <= SMALL_STREAM_ARRAY_SIZE)
 350                 return dma_pool_alloc(xhci->small_streams_pool,
 351                                 mem_flags, dma);
 352         else
 353                 return dma_pool_alloc(xhci->medium_streams_pool,
 354                                 mem_flags, dma);
 355 }
 356
 357 struct xhci_ring *xhci_dma_to_transfer_ring(
 358                 struct xhci_virt_ep *ep,
 359                 u64 address)
 360 {
 361         if (ep->ep_state & EP_HAS_STREAMS)
 362                 return radix_tree_lookup(&ep->stream_info->trb_address_map,
 363                                 address >> SEGMENT_SHIFT);
 364         return ep->ring;
 365 }
 366
 367 /* Only use this when you know stream_info is valid */
 368 #ifdef CONFIG_USB_XHCI_HCD_DEBUGGING
 369 static struct xhci_ring *dma_to_stream_ring(
 370                 struct xhci_stream_info *stream_info,
 371                 u64 address)
 372 {
 373         return radix_tree_lookup(&stream_info->trb_address_map,
 374                         address >> SEGMENT_SHIFT);
 375 }
 376 #endif  /* CONFIG_USB_XHCI_HCD_DEBUGGING */
 377
 378 struct xhci_ring *xhci_stream_id_to_ring(
 379                 struct xhci_virt_device *dev,
 380                 unsigned int ep_index,
 381                 unsigned int stream_id)
 382 {
 383         struct xhci_virt_ep *ep = &dev->eps[ep_index];
 384
 385         if (stream_id == 0)
 386                 return ep->ring;
 387         if (!ep->stream_info)
 388                 return NULL;
 389
 390         if (stream_id > ep->stream_info->num_streams)
 391                 return NULL;
 392         return ep->stream_info->stream_rings[stream_id];
 393 }
 394
 395 #ifdef CONFIG_USB_XHCI_HCD_DEBUGGING
 396 static int xhci_test_radix_tree(struct xhci_hcd *xhci,
 397                 unsigned int num_streams,
 398                 struct xhci_stream_info *stream_info)
 399 {
 400         u32 cur_stream;
 401         struct xhci_ring *cur_ring;
 402         u64 addr;
 403
 404         for (cur_stream = 1; cur_stream < num_streams; cur_stream++) {
 405                 struct xhci_ring *mapped_ring;
 406                 int trb_size = sizeof(union xhci_trb);
 407
 408                 cur_ring = stream_info->stream_rings[cur_stream];
 409                 for (addr = cur_ring->first_seg->dma;
 410                                 addr < cur_ring->first_seg->dma + SEGMENT_SIZE;
 411                                 addr += trb_size) {
 412                         mapped_ring = dma_to_stream_ring(stream_info, addr);
 413                         if (cur_ring != mapped_ring) {
 414                                 xhci_warn(xhci, "WARN: DMA address 0x%08llx "
 415                                                 "didn't map to stream ID %u; "
 416                                                 "mapped to ring %p\n",
 417                                                 (unsigned long long) addr,
 418                                                 cur_stream,
 419                                                 mapped_ring);
 420                                 return -EINVAL;
 421                         }
 422                 }
 423                 /* One TRB after the end of the ring segment shouldn't return a
 424                  * pointer to the current ring (although it may be a part of a
 425                  * different ring).
 426                  */
 427                 mapped_ring = dma_to_stream_ring(stream_info, addr);
 428                 if (mapped_ring != cur_ring) {
 429                         /* One TRB before should also fail */
 430                         addr = cur_ring->first_seg->dma - trb_size;
 431                         mapped_ring = dma_to_stream_ring(stream_info, addr);
 432                 }
 433                 if (mapped_ring == cur_ring) {
 434                         xhci_warn(xhci, "WARN: Bad DMA address 0x%08llx "
 435                                         "mapped to valid stream ID %u; "
 436                                         "mapped ring = %p\n",
 437                                         (unsigned long long) addr,
 438                                         cur_stream,
 439                                         mapped_ring);
 440                         return -EINVAL;
 441                 }
 442         }
 443         return 0;
 444 }
 445 #endif  /* CONFIG_USB_XHCI_HCD_DEBUGGING */
 446
 447 /*
 448  * Change an endpoint's internal structure so it supports stream IDs.  The
 449  * number of requested streams includes stream 0, which cannot be used by device
 450  * drivers.
 451  *
 452  * The number of stream contexts in the stream context array may be bigger than
 453  * the number of streams the driver wants to use.  This is because the number of
 454  * stream context array entries must be a power of two.
 455  *
 456  * We need a radix tree for mapping physical addresses of TRBs to which stream
 457  * ID they belong to.  We need to do this because the host controller won't tell
 458  * us which stream ring the TRB came from.  We could store the stream ID in an
 459  * event data TRB, but that doesn't help us for the cancellation case, since the
 460  * endpoint may stop before it reaches that event data TRB.
 461  *
 462  * The radix tree maps the upper portion of the TRB DMA address to a ring
 463  * segment that has the same upper portion of DMA addresses.  For example, say I
 464  * have segments of size 1KB, that are always 64-byte aligned.  A segment may
 465  * start at 0x10c91000 and end at 0x10c913f0.  If I use the upper 10 bits, the
 466  * key to the stream ID is 0x43244.  I can use the DMA address of the TRB to
 467  * pass the radix tree a key to get the right stream ID:
 468  *
 469  *      0x10c90fff >> 10 = 0x43243
 470  *      0x10c912c0 >> 10 = 0x43244
 471  *      0x10c91400 >> 10 = 0x43245
 472  *
 473  * Obviously, only those TRBs with DMA addresses that are within the segment
 474  * will make the radix tree return the stream ID for that ring.
 475  *
 476  * Caveats for the radix tree:
 477  *
 478  * The radix tree uses an unsigned long as a key pair.  On 32-bit systems, an
 479  * unsigned long will be 32-bits; on a 64-bit system an unsigned long will be
 480  * 64-bits.  Since we only request 32-bit DMA addresses, we can use that as the
 481  * key on 32-bit or 64-bit systems (it would also be fine if we asked for 64-bit
 482  * PCI DMA addresses on a 64-bit system).  There might be a problem on 32-bit
 483  * extended systems (where the DMA address can be bigger than 32-bits),
 484  * if we allow the PCI dma mask to be bigger than 32-bits.  So don't do that.
 485  */
 486 struct xhci_stream_info *xhci_alloc_stream_info(struct xhci_hcd *xhci,
 487                 unsigned int num_stream_ctxs,
 488                 unsigned int num_streams, gfp_t mem_flags)
 489 {
 490         struct xhci_stream_info *stream_info;
 491         u32 cur_stream;
 492         struct xhci_ring *cur_ring;
 493         unsigned long key;
 494         u64 addr;
 495         int ret;
 496
 497         xhci_dbg(xhci, "Allocating %u streams and %u "
 498                         "stream context array entries.\n",
 499                         num_streams, num_stream_ctxs);
 500         if (xhci->cmd_ring_reserved_trbs == MAX_RSVD_CMD_TRBS) {
 501                 xhci_dbg(xhci, "Command ring has no reserved TRBs available\n");
 502                 return NULL;
 503         }
 504         xhci->cmd_ring_reserved_trbs++;
 505
 506         stream_info = kzalloc(sizeof(struct xhci_stream_info), mem_flags);
 507         if (!stream_info)
 508                 goto cleanup_trbs;
 509
 510         stream_info->num_streams = num_streams;
 511         stream_info->num_stream_ctxs = num_stream_ctxs;
 512
 513         /* Initialize the array of virtual pointers to stream rings. */
 514         stream_info->stream_rings = kzalloc(
 515                         sizeof(struct xhci_ring *)*num_streams,
 516                         mem_flags);
 517         if (!stream_info->stream_rings)
 518                 goto cleanup_info;
 519
 520         /* Initialize the array of DMA addresses for stream rings for the HW. */
 521         stream_info->stream_ctx_array = xhci_alloc_stream_ctx(xhci,
 522                         num_stream_ctxs, &stream_info->ctx_array_dma,
 523                         mem_flags);
 524         if (!stream_info->stream_ctx_array)
 525                 goto cleanup_ctx;
 526         memset(stream_info->stream_ctx_array, 0,
 527                         sizeof(struct xhci_stream_ctx)*num_stream_ctxs);
 528
 529         /* Allocate everything needed to free the stream rings later */
 530         stream_info->free_streams_command =
 531                 xhci_alloc_command(xhci, true, true, mem_flags);
 532         if (!stream_info->free_streams_command)
 533                 goto cleanup_ctx;
 534
 535         INIT_RADIX_TREE(&stream_info->trb_address_map, GFP_ATOMIC);
 536
 537         /* Allocate rings for all the streams that the driver will use,
 538          * and add their segment DMA addresses to the radix tree.
 539          * Stream 0 is reserved.
 540          */
 541         for (cur_stream = 1; cur_stream < num_streams; cur_stream++) {
 542                 stream_info->stream_rings[cur_stream] =
 543                         xhci_ring_alloc(xhci, 1, true, mem_flags);
 544                 cur_ring = stream_info->stream_rings[cur_stream];
 545                 if (!cur_ring)
 546                         goto cleanup_rings;
 547                 cur_ring->stream_id = cur_stream;
 548                 /* Set deq ptr, cycle bit, and stream context type */
 549                 addr = cur_ring->first_seg->dma |
 550                         SCT_FOR_CTX(SCT_PRI_TR) |
 551                         cur_ring->cycle_state;
 552                 stream_info->stream_ctx_array[cur_stream].stream_ring =
 553                         cpu_to_le64(addr);
 554                 xhci_dbg(xhci, "Setting stream %d ring ptr to 0x%08llx\n",
 555                                 cur_stream, (unsigned long long) addr);
 556
 557                 key = (unsigned long)
 558                         (cur_ring->first_seg->dma >> SEGMENT_SHIFT);
 559                 ret = radix_tree_insert(&stream_info->trb_address_map,
 560                                 key, cur_ring);
 561                 if (ret) {
 562                         xhci_ring_free(xhci, cur_ring);
 563                         stream_info->stream_rings[cur_stream] = NULL;
 564                         goto cleanup_rings;
 565                 }
 566         }
 567         /* Leave the other unused stream ring pointers in the stream context
 568          * array initialized to zero.  This will cause the xHC to give us an
 569          * error if the device asks for a stream ID we don't have setup (if it
 570          * was any other way, the host controller would assume the ring is
 571          * "empty" and wait forever for data to be queued to that stream ID).
 572          */
 573 #if XHCI_DEBUG
 574         /* Do a little test on the radix tree to make sure it returns the
 575          * correct values.
 576          */
 577         if (xhci_test_radix_tree(xhci, num_streams, stream_info))
 578                 goto cleanup_rings;
 579 #endif
 580
 581         return stream_info;
 582
 583 cleanup_rings:
 584         for (cur_stream = 1; cur_stream < num_streams; cur_stream++) {
 585                 cur_ring = stream_info->stream_rings[cur_stream];
 586                 if (cur_ring) {
 587                         addr = cur_ring->first_seg->dma;
 588                         radix_tree_delete(&stream_info->trb_address_map,
 589                                         addr >> SEGMENT_SHIFT);
 590                         xhci_ring_free(xhci, cur_ring);
 591                         stream_info->stream_rings[cur_stream] = NULL;
 592                 }
 593         }
 594         xhci_free_command(xhci, stream_info->free_streams_command);
 595 cleanup_ctx:
 596         kfree(stream_info->stream_rings);
 597 cleanup_info:
 598         kfree(stream_info);
 599 cleanup_trbs:
 600         xhci->cmd_ring_reserved_trbs--;
 601         return NULL;
 602 }
 603 /*
 604  * Sets the MaxPStreams field and the Linear Stream Array field.
 605  * Sets the dequeue pointer to the stream context array.
 606  */
 607 void xhci_setup_streams_ep_input_ctx(struct xhci_hcd *xhci,
 608                 struct xhci_ep_ctx *ep_ctx,
 609                 struct xhci_stream_info *stream_info)
 610 {
 611         u32 max_primary_streams;
 612         /* MaxPStreams is the number of stream context array entries, not the
 613          * number we're actually using.  Must be in 2^(MaxPstreams + 1) format.
 614          * fls(0) = 0, fls(0x1) = 1, fls(0x10) = 2, fls(0x100) = 3, etc.
 615          */
 616         max_primary_streams = fls(stream_info->num_stream_ctxs) - 2;
 617         xhci_dbg(xhci, "Setting number of stream ctx array entries to %u\n",
 618                         1 << (max_primary_streams + 1));
 619         ep_ctx->ep_info &= cpu_to_le32(~EP_MAXPSTREAMS_MASK);
 620         ep_ctx->ep_info |= cpu_to_le32(EP_MAXPSTREAMS(max_primary_streams)
 621                                        | EP_HAS_LSA);
 622         ep_ctx->deq  = cpu_to_le64(stream_info->ctx_array_dma);
 623 }
 624
 625 /*
 626  * Sets the MaxPStreams field and the Linear Stream Array field to 0.
 627  * Reinstalls the "normal" endpoint ring (at its previous dequeue mark,
 628  * not at the beginning of the ring).
 629  */
 630 void xhci_setup_no_streams_ep_input_ctx(struct xhci_hcd *xhci,
 631                 struct xhci_ep_ctx *ep_ctx,
 632                 struct xhci_virt_ep *ep)
 633 {
 634         dma_addr_t addr;
 635         ep_ctx->ep_info &= cpu_to_le32(~(EP_MAXPSTREAMS_MASK | EP_HAS_LSA));
 636         addr = xhci_trb_virt_to_dma(ep->ring->deq_seg, ep->ring->dequeue);
 637         ep_ctx->deq  = cpu_to_le64(addr | ep->ring->cycle_state);
 638 }
 639
 640 /* Frees all stream contexts associated with the endpoint,
 641  *
 642  * Caller should fix the endpoint context streams fields.
 643  */
 644 void xhci_free_stream_info(struct xhci_hcd *xhci,
 645                 struct xhci_stream_info *stream_info)
 646 {
 647         int cur_stream;
 648         struct xhci_ring *cur_ring;
 649         dma_addr_t addr;
 650
 651         if (!stream_info)
 652                 return;
 653
 654         for (cur_stream = 1; cur_stream < stream_info->num_streams;
 655                         cur_stream++) {
 656                 cur_ring = stream_info->stream_rings[cur_stream];
 657                 if (cur_ring) {
 658                         addr = cur_ring->first_seg->dma;
 659                         radix_tree_delete(&stream_info->trb_address_map,
 660                                         addr >> SEGMENT_SHIFT);
 661                         xhci_ring_free(xhci, cur_ring);
 662                         stream_info->stream_rings[cur_stream] = NULL;
 663                 }
 664         }
 665         xhci_free_command(xhci, stream_info->free_streams_command);
 666         xhci->cmd_ring_reserved_trbs--;
 667         if (stream_info->stream_ctx_array)
 668                 xhci_free_stream_ctx(xhci,
 669                                 stream_info->num_stream_ctxs,
 670                                 stream_info->stream_ctx_array,
 671                                 stream_info->ctx_array_dma);
 672
 673         if (stream_info)
 674                 kfree(stream_info->stream_rings);
 675         kfree(stream_info);
 676 }
 677
 678
 679 /***************** Device context manipulation *************************/
 680
 681 static void xhci_init_endpoint_timer(struct xhci_hcd *xhci,
 682                 struct xhci_virt_ep *ep)
 683 {
 684         init_timer(&ep->stop_cmd_timer);
 685         ep->stop_cmd_timer.data = (unsigned long) ep;
 686         ep->stop_cmd_timer.function = xhci_stop_endpoint_command_watchdog;
 687         ep->xhci = xhci;
 688 }
 689
 690 /* All the xhci_tds in the ring's TD list should be freed at this point */
 691 void xhci_free_virt_device(struct xhci_hcd *xhci, int slot_id)
 692 {
 693         struct xhci_virt_device *dev;
 694         int i;
 695
 696         /* Slot ID 0 is reserved */
 697         if (slot_id == 0 || !xhci->devs[slot_id])
 698                 return;
 699
 700         dev = xhci->devs[slot_id];
 701         xhci->dcbaa->dev_context_ptrs[slot_id] = 0;
 702         if (!dev)
 703                 return;
 704
 705         for (i = 0; i < 31; ++i) {
 706                 if (dev->eps[i].ring)
 707                         xhci_ring_free(xhci, dev->eps[i].ring);
 708                 if (dev->eps[i].stream_info)
 709                         xhci_free_stream_info(xhci,
 710                                         dev->eps[i].stream_info);
 711         }
 712
 713         if (dev->ring_cache) {
 714                 for (i = 0; i < dev->num_rings_cached; i++)
 715                         xhci_ring_free(xhci, dev->ring_cache[i]);
 716                 kfree(dev->ring_cache);
 717         }
 718
 719         if (dev->in_ctx)
 720                 xhci_free_container_ctx(xhci, dev->in_ctx);
 721         if (dev->out_ctx)
 722                 xhci_free_container_ctx(xhci, dev->out_ctx);
 723
 724         kfree(xhci->devs[slot_id]);
 725         xhci->devs[slot_id] = NULL;
 726 }
 727
 728 int xhci_alloc_virt_device(struct xhci_hcd *xhci, int slot_id,
 729                 struct usb_device *udev, gfp_t flags)
 730 {
 731         struct xhci_virt_device *dev;
 732         int i;
 733
 734         /* Slot ID 0 is reserved */
 735         if (slot_id == 0 || xhci->devs[slot_id]) {
 736                 xhci_warn(xhci, "Bad Slot ID %d\n", slot_id);
 737                 return 0;
 738         }
 739
 740         xhci->devs[slot_id] = kzalloc(sizeof(*xhci->devs[slot_id]), flags);
 741         if (!xhci->devs[slot_id])
 742                 return 0;
 743         dev = xhci->devs[slot_id];
 744
 745         /* Allocate the (output) device context that will be used in the HC. */
 746         dev->out_ctx = xhci_alloc_container_ctx(xhci, XHCI_CTX_TYPE_DEVICE, flags);
 747         if (!dev->out_ctx)
 748                 goto fail;
 749
 750         xhci_dbg(xhci, "Slot %d output ctx = 0x%llx (dma)\n", slot_id,
 751                         (unsigned long long)dev->out_ctx->dma);
 752
 753         /* Allocate the (input) device context for address device command */
 754         dev->in_ctx = xhci_alloc_container_ctx(xhci, XHCI_CTX_TYPE_INPUT, flags);
 755         if (!dev->in_ctx)
 756                 goto fail;
 757
 758         xhci_dbg(xhci, "Slot %d input ctx = 0x%llx (dma)\n", slot_id,
 759                         (unsigned long long)dev->in_ctx->dma);
 760
 761         /* Initialize the cancellation list and watchdog timers for each ep */
 762         for (i = 0; i < 31; i++) {
 763                 xhci_init_endpoint_timer(xhci, &dev->eps[i]);
 764                 INIT_LIST_HEAD(&dev->eps[i].cancelled_td_list);
 765         }
 766
 767         /* Allocate endpoint 0 ring */
 768         dev->eps[0].ring = xhci_ring_alloc(xhci, 1, true, flags);
 769         if (!dev->eps[0].ring)
 770                 goto fail;
 771
 772         /* Allocate pointers to the ring cache */
 773         dev->ring_cache = kzalloc(
 774                         sizeof(struct xhci_ring *)*XHCI_MAX_RINGS_CACHED,
 775                         flags);
 776         if (!dev->ring_cache)
 777                 goto fail;
 778         dev->num_rings_cached = 0;
 779
 780         init_completion(&dev->cmd_completion);
 781         INIT_LIST_HEAD(&dev->cmd_list);
 782         dev->udev = udev;
 783
 784         /* Point to output device context in dcbaa. */
 785         xhci->dcbaa->dev_context_ptrs[slot_id] = cpu_to_le64(dev->out_ctx->dma);
 786         xhci_dbg(xhci, "Set slot id %d dcbaa entry %p to 0x%llx\n",
 787                  slot_id,
 788                  &xhci->dcbaa->dev_context_ptrs[slot_id],
 789                  le64_to_cpu(xhci->dcbaa->dev_context_ptrs[slot_id]));
 790
 791         return 1;
 792 fail:
 793         xhci_free_virt_device(xhci, slot_id);
 794         return 0;
 795 }
 796
 797 void xhci_copy_ep0_dequeue_into_input_ctx(struct xhci_hcd *xhci,
 798                 struct usb_device *udev)
 799 {
 800         struct xhci_virt_device *virt_dev;
 801         struct xhci_ep_ctx      *ep0_ctx;
 802         struct xhci_ring        *ep_ring;
 803
 804         virt_dev = xhci->devs[udev->slot_id];
 805         ep0_ctx = xhci_get_ep_ctx(xhci, virt_dev->in_ctx, 0);
 806         ep_ring = virt_dev->eps[0].ring;
 807         /*
 808          * FIXME we don't keep track of the dequeue pointer very well after a
 809          * Set TR dequeue pointer, so we're setting the dequeue pointer of the
 810          * host to our enqueue pointer.  This should only be called after a
 811          * configured device has reset, so all control transfers should have
 812          * been completed or cancelled before the reset.
 813          */
 814         ep0_ctx->deq = cpu_to_le64(xhci_trb_virt_to_dma(ep_ring->enq_seg,
 815                                                         ep_ring->enqueue)
 816                                    | ep_ring->cycle_state);
 817 }
 818
 819 /*
 820  * The xHCI roothub may have ports of differing speeds in any order in the port
 821  * status registers.  xhci->port_array provides an array of the port speed for
 822  * each offset into the port status registers.
 823  *
 824  * The xHCI hardware wants to know the roothub port number that the USB device
 825  * is attached to (or the roothub port its ancestor hub is attached to).  All we
 826  * know is the index of that port under either the USB 2.0 or the USB 3.0
 827  * roothub, but that doesn't give us the real index into the HW port status
 828  * registers.  Scan through the xHCI roothub port array, looking for the Nth
 829  * entry of the correct port speed.  Return the port number of that entry.
 830  */
 831 static u32 xhci_find_real_port_number(struct xhci_hcd *xhci,
 832                 struct usb_device *udev)
 833 {
 834         struct usb_device *top_dev;
 835         unsigned int num_similar_speed_ports;
 836         unsigned int faked_port_num;
 837         int i;
 838
 839         for (top_dev = udev; top_dev->parent && top_dev->parent->parent;
 840                         top_dev = top_dev->parent)
 841                 /* Found device below root hub */;
 842         faked_port_num = top_dev->portnum;
 843         for (i = 0, num_similar_speed_ports = 0;
 844                         i < HCS_MAX_PORTS(xhci->hcs_params1); i++) {
 845                 u8 port_speed = xhci->port_array[i];
 846
 847                 /*
 848                  * Skip ports that don't have known speeds, or have duplicate
 849                  * Extended Capabilities port speed entries.
 850                  */
 851                 if (port_speed == 0 || port_speed == DUPLICATE_ENTRY)
 852                         continue;
 853
 854                 /*
 855                  * USB 3.0 ports are always under a USB 3.0 hub.  USB 2.0 and
 856                  * 1.1 ports are under the USB 2.0 hub.  If the port speed
 857                  * matches the device speed, it's a similar speed port.
 858                  */
 859                 if ((port_speed == 0x03) == (udev->speed == USB_SPEED_SUPER))
 860                         num_similar_speed_ports++;
 861                 if (num_similar_speed_ports == faked_port_num)
 862                         /* Roothub ports are numbered from 1 to N */
 863                         return i+1;
 864         }
 865         return 0;
 866 }
 867
 868 /* Setup an xHCI virtual device for a Set Address command */
 869 int xhci_setup_addressable_virt_dev(struct xhci_hcd *xhci, struct usb_device *udev)
 870 {
 871         struct xhci_virt_device *dev;
 872         struct xhci_ep_ctx      *ep0_ctx;
 873         struct xhci_slot_ctx    *slot_ctx;
 874         struct xhci_input_control_ctx *ctrl_ctx;
 875         u32                     port_num;
 876         struct usb_device *top_dev;
 877
 878         dev = xhci->devs[udev->slot_id];
 879         /* Slot ID 0 is reserved */
 880         if (udev->slot_id == 0 || !dev) {
 881                 xhci_warn(xhci, "Slot ID %d is not assigned to this device\n",
 882                                 udev->slot_id);
 883                 return -EINVAL;
 884         }
 885         ep0_ctx = xhci_get_ep_ctx(xhci, dev->in_ctx, 0);
 886         ctrl_ctx = xhci_get_input_control_ctx(xhci, dev->in_ctx);
 887         slot_ctx = xhci_get_slot_ctx(xhci, dev->in_ctx);
 888
 889         /* 2) New slot context and endpoint 0 context are valid*/
 890         ctrl_ctx->add_flags = cpu_to_le32(SLOT_FLAG | EP0_FLAG);
 891
 892         /* 3) Only the control endpoint is valid - one endpoint context */
 893         slot_ctx->dev_info |= cpu_to_le32(LAST_CTX(1) | udev->route);
 894         switch (udev->speed) {
 895         case USB_SPEED_SUPER:
 896                 slot_ctx->dev_info |= cpu_to_le32(SLOT_SPEED_SS);
 897                 break;
 898         case USB_SPEED_HIGH:
 899                 slot_ctx->dev_info |= cpu_to_le32(SLOT_SPEED_HS);
 900                 break;
 901         case USB_SPEED_FULL:
 902                 slot_ctx->dev_info |= cpu_to_le32(SLOT_SPEED_FS);
 903                 break;
 904         case USB_SPEED_LOW:
 905                 slot_ctx->dev_info |= cpu_to_le32(SLOT_SPEED_LS);
 906                 break;
 907         case USB_SPEED_WIRELESS:
 908                 xhci_dbg(xhci, "FIXME xHCI doesn't support wireless speeds\n");
 909                 return -EINVAL;
 910                 break;
 911         default:
 912                 /* Speed was set earlier, this shouldn't happen. */
 913                 BUG();
 914         }
 915         /* Find the root hub port this device is under */
 916         port_num = xhci_find_real_port_number(xhci, udev);
 917         if (!port_num)
 918                 return -EINVAL;
 919         slot_ctx->dev_info2 |= cpu_to_le32(ROOT_HUB_PORT(port_num));
 920         /* Set the port number in the virtual_device to the faked port number */
 921         for (top_dev = udev; top_dev->parent && top_dev->parent->parent;
 922                         top_dev = top_dev->parent)
 923                 /* Found device below root hub */;
 924         dev->port = top_dev->portnum;
 925         xhci_dbg(xhci, "Set root hub portnum to %d\n", port_num);
 926         xhci_dbg(xhci, "Set fake root hub portnum to %d\n", dev->port);
 927
 928         /* Is this a LS/FS device under an external HS hub? */
 929         if (udev->tt && udev->tt->hub->parent) {
 930                 slot_ctx->tt_info = cpu_to_le32(udev->tt->hub->slot_id |
 931                                                 (udev->ttport << 8));
 932                 if (udev->tt->multi)
 933                         slot_ctx->dev_info |= cpu_to_le32(DEV_MTT);
 934         }
 935         xhci_dbg(xhci, "udev->tt = %p\n", udev->tt);
 936         xhci_dbg(xhci, "udev->ttport = 0x%x\n", udev->ttport);
 937
 938         /* Step 4 - ring already allocated */
 939         /* Step 5 */
 940         ep0_ctx->ep_info2 = cpu_to_le32(EP_TYPE(CTRL_EP));
 941         /*
 942          * XXX: Not sure about wireless USB devices.
 943          */
 944         switch (udev->speed) {
 945         case USB_SPEED_SUPER:
 946                 ep0_ctx->ep_info2 |= cpu_to_le32(MAX_PACKET(512));
 947                 break;
 948         case USB_SPEED_HIGH:
 949         /* USB core guesses at a 64-byte max packet first for FS devices */
 950         case USB_SPEED_FULL:
 951                 ep0_ctx->ep_info2 |= cpu_to_le32(MAX_PACKET(64));
 952                 break;
 953         case USB_SPEED_LOW:
 954                 ep0_ctx->ep_info2 |= cpu_to_le32(MAX_PACKET(8));
 955                 break;
 956         case USB_SPEED_WIRELESS:
 957                 xhci_dbg(xhci, "FIXME xHCI doesn't support wireless speeds\n");
 958                 return -EINVAL;
 959                 break;
 960         default:
 961                 /* New speed? */
 962                 BUG();
 963         }
 964         /* EP 0 can handle "burst" sizes of 1, so Max Burst Size field is 0 */
 965         ep0_ctx->ep_info2 |= cpu_to_le32(MAX_BURST(0) | ERROR_COUNT(3));
 966
 967         ep0_ctx->deq = cpu_to_le64(dev->eps[0].ring->first_seg->dma |
 968                                    dev->eps[0].ring->cycle_state);
 969
 970         /* Steps 7 and 8 were done in xhci_alloc_virt_device() */
 971
 972         return 0;
 973 }
 974
 975 /*
 976  * Convert interval expressed as 2^(bInterval - 1) == interval into
 977  * straight exponent value 2^n == interval.
 978  *
 979  */
 980 static unsigned int xhci_parse_exponent_interval(struct usb_device *udev,
 981                 struct usb_host_endpoint *ep)
 982 {
 983         unsigned int interval;
 984
 985         interval = clamp_val(ep->desc.bInterval, 1, 16) - 1;
 986         if (interval != ep->desc.bInterval - 1)
 987                 dev_warn(&udev->dev,
 988                          "ep %#x - rounding interval to %d %sframes\n",
 989                          ep->desc.bEndpointAddress,
 990                          1 << interval,
 991                          udev->speed == USB_SPEED_FULL ? "" : "micro");
 992
 993         if (udev->speed == USB_SPEED_FULL) {
 994                 /*
 995                  * Full speed isoc endpoints specify interval in frames,
 996                  * not microframes. We are using microframes everywhere,
 997                  * so adjust accordingly.
 998                  */
 999                 interval += 3;  /* 1 frame = 2^3 uframes */
1000         }
1001
1002         return interval;
1003 }
1004
1005 /*
1006  * Convert bInterval expressed in frames (in 1-255 range) to exponent of
1007  * microframes, rounded down to nearest power of 2.
1008  */
1009 static unsigned int xhci_parse_frame_interval(struct usb_device *udev,
1010                 struct usb_host_endpoint *ep)
1011 {
1012         unsigned int interval;
1013
1014         interval = fls(8 * ep->desc.bInterval) - 1;
1015         interval = clamp_val(interval, 3, 10);
1016         if ((1 << interval) != 8 * ep->desc.bInterval)
1017                 dev_warn(&udev->dev,
1018                          "ep %#x - rounding interval to %d microframes, ep desc says %d microframes\n",
1019                          ep->desc.bEndpointAddress,
1020                          1 << interval,
1021                          8 * ep->desc.bInterval);
1022
1023         return interval;
1024 }
1025
1026 /* Return the polling or NAK interval.
1027  *
1028  * The polling interval is expressed in "microframes".  If xHCI's Interval field
1029  * is set to N, it will service the endpoint every 2^(Interval)*125us.
1030  *
1031  * The NAK interval is one NAK per 1 to 255 microframes, or no NAKs if interval
1032  * is set to 0.
1033  */
1034 static unsigned int xhci_get_endpoint_interval(struct usb_device *udev,
1035                 struct usb_host_endpoint *ep)
1036 {
1037         unsigned int interval = 0;
1038
1039         switch (udev->speed) {
1040         case USB_SPEED_HIGH:
1041                 /* Max NAK rate */
1042                 if (usb_endpoint_xfer_control(&ep->desc) ||
1043                     usb_endpoint_xfer_bulk(&ep->desc)) {
1044                         interval = ep->desc.bInterval;
1045                         break;
1046                 }
1047                 /* Fall through - SS and HS isoc/int have same decoding */
1048
1049         case USB_SPEED_SUPER:
1050                 if (usb_endpoint_xfer_int(&ep->desc) ||
1051                     usb_endpoint_xfer_isoc(&ep->desc)) {
1052                         interval = xhci_parse_exponent_interval(udev, ep);
1053                 }
1054                 break;
1055
1056         case USB_SPEED_FULL:
1057                 if (usb_endpoint_xfer_isoc(&ep->desc)) {
1058                         interval = xhci_parse_exponent_interval(udev, ep);
1059                         break;
1060                 }
1061                 /*
1062                  * Fall through for interrupt endpoint interval decoding
1063                  * since it uses the same rules as low speed interrupt
1064                  * endpoints.
1065                  */
1066
1067         case USB_SPEED_LOW:
1068                 if (usb_endpoint_xfer_int(&ep->desc) ||
1069                     usb_endpoint_xfer_isoc(&ep->desc)) {
1070
1071                         interval = xhci_parse_frame_interval(udev, ep);
1072                 }
1073                 break;
1074
1075         default:
1076                 BUG();
1077         }
1078         return EP_INTERVAL(interval);
1079 }
1080
1081 /* The "Mult" field in the endpoint context is only set for SuperSpeed isoc eps.
1082  * High speed endpoint descriptors can define "the number of additional
1083  * transaction opportunities per microframe", but that goes in the Max Burst
1084  * endpoint context field.
1085  */
1086 static u32 xhci_get_endpoint_mult(struct usb_device *udev,
1087                 struct usb_host_endpoint *ep)
1088 {
1089         if (udev->speed != USB_SPEED_SUPER ||
1090                         !usb_endpoint_xfer_isoc(&ep->desc))
1091                 return 0;
1092         return ep->ss_ep_comp.bmAttributes;
1093 }
1094
1095 static u32 xhci_get_endpoint_type(struct usb_device *udev,
1096                 struct usb_host_endpoint *ep)
1097 {
1098         int in;
1099         u32 type;
1100
1101         in = usb_endpoint_dir_in(&ep->desc);
1102         if (usb_endpoint_xfer_control(&ep->desc)) {
1103                 type = EP_TYPE(CTRL_EP);
1104         } else if (usb_endpoint_xfer_bulk(&ep->desc)) {
1105                 if (in)
1106                         type = EP_TYPE(BULK_IN_EP);
1107                 else
1108                         type = EP_TYPE(BULK_OUT_EP);
1109         } else if (usb_endpoint_xfer_isoc(&ep->desc)) {
1110                 if (in)
1111                         type = EP_TYPE(ISOC_IN_EP);
1112                 else
1113                         type = EP_TYPE(ISOC_OUT_EP);
1114         } else if (usb_endpoint_xfer_int(&ep->desc)) {
1115                 if (in)
1116                         type = EP_TYPE(INT_IN_EP);
1117                 else
1118                         type = EP_TYPE(INT_OUT_EP);
1119         } else {
1120                 BUG();
1121         }
1122         return type;
1123 }
1124
1125 /* Return the maximum endpoint service interval time (ESIT) payload.
1126  * Basically, this is the maxpacket size, multiplied by the burst size
1127  * and mult size.
1128  */
1129 static u32 xhci_get_max_esit_payload(struct xhci_hcd *xhci,
1130                 struct usb_device *udev,
1131                 struct usb_host_endpoint *ep)
1132 {
1133         int max_burst;
1134         int max_packet;
1135
1136         /* Only applies for interrupt or isochronous endpoints */
1137         if (usb_endpoint_xfer_control(&ep->desc) ||
1138                         usb_endpoint_xfer_bulk(&ep->desc))
1139                 return 0;
1140
1141         if (udev->speed == USB_SPEED_SUPER)
1142                 return le16_to_cpu(ep->ss_ep_comp.wBytesPerInterval);
1143
1144         max_packet = GET_MAX_PACKET(le16_to_cpu(ep->desc.wMaxPacketSize));
1145         max_burst = (le16_to_cpu(ep->desc.wMaxPacketSize) & 0x1800) >> 11;
1146         /* A 0 in max burst means 1 transfer per ESIT */
1147         return max_packet * (max_burst + 1);
1148 }
1149
1150 /* Set up an endpoint with one ring segment.  Do not allocate stream rings.
1151  * Drivers will have to call usb_alloc_streams() to do that.
1152  */
1153 int xhci_endpoint_init(struct xhci_hcd *xhci,
1154                 struct xhci_virt_device *virt_dev,
1155                 struct usb_device *udev,
1156                 struct usb_host_endpoint *ep,
1157                 gfp_t mem_flags)
1158 {
1159         unsigned int ep_index;
1160         struct xhci_ep_ctx *ep_ctx;
1161         struct xhci_ring *ep_ring;
1162         unsigned int max_packet;
1163         unsigned int max_burst;
1164         u32 max_esit_payload;
1165
1166         ep_index = xhci_get_endpoint_index(&ep->desc);
1167         ep_ctx = xhci_get_ep_ctx(xhci, virt_dev->in_ctx, ep_index);
1168
1169         /* Set up the endpoint ring */
1170         /*
1171          * Isochronous endpoint ring needs bigger size because one isoc URB
1172          * carries multiple packets and it will insert multiple tds to the
1173          * ring.
1174          * This should be replaced with dynamic ring resizing in the future.
1175          */
1176         if (usb_endpoint_xfer_isoc(&ep->desc))
1177                 virt_dev->eps[ep_index].new_ring =
1178                         xhci_ring_alloc(xhci, 8, true, mem_flags);
1179         else
1180                 virt_dev->eps[ep_index].new_ring =
1181                         xhci_ring_alloc(xhci, 1, true, mem_flags);
1182         if (!virt_dev->eps[ep_index].new_ring) {
1183                 /* Attempt to use the ring cache */
1184                 if (virt_dev->num_rings_cached == 0)
1185                         return -ENOMEM;
1186                 virt_dev->eps[ep_index].new_ring =
1187                         virt_dev->ring_cache[virt_dev->num_rings_cached];
1188                 virt_dev->ring_cache[virt_dev->num_rings_cached] = NULL;
1189                 virt_dev->num_rings_cached--;
1190                 xhci_reinit_cached_ring(xhci, virt_dev->eps[ep_index].new_ring);
1191         }
1192         virt_dev->eps[ep_index].skip = false;
1193         ep_ring = virt_dev->eps[ep_index].new_ring;
1194         ep_ctx->deq = cpu_to_le64(ep_ring->first_seg->dma | ep_ring->cycle_state);
1195
1196         ep_ctx->ep_info = cpu_to_le32(xhci_get_endpoint_interval(udev, ep)
1197                                       | EP_MULT(xhci_get_endpoint_mult(udev, ep)));
1198
1199         /* FIXME dig Mult and streams info out of ep companion desc */
1200
1201         /* Allow 3 retries for everything but isoc;
1202          * CErr shall be set to 0 for Isoch endpoints.
1203          */
1204         if (!usb_endpoint_xfer_isoc(&ep->desc))
1205                 ep_ctx->ep_info2 = cpu_to_le32(ERROR_COUNT(3));
1206         else
1207                 ep_ctx->ep_info2 = cpu_to_le32(ERROR_COUNT(0));
1208
1209         ep_ctx->ep_info2 |= cpu_to_le32(xhci_get_endpoint_type(udev, ep));
1210
1211         /* Set the max packet size and max burst */
1212         switch (udev->speed) {
1213         case USB_SPEED_SUPER:
1214                 max_packet = le16_to_cpu(ep->desc.wMaxPacketSize);
1215                 ep_ctx->ep_info2 |= cpu_to_le32(MAX_PACKET(max_packet));
1216                 /* dig out max burst from ep companion desc */
1217                 max_packet = ep->ss_ep_comp.bMaxBurst;
1218                 ep_ctx->ep_info2 |= cpu_to_le32(MAX_BURST(max_packet));
1219                 break;
1220         case USB_SPEED_HIGH:
1221                 /* bits 11:12 specify the number of additional transaction
1222                  * opportunities per microframe (USB 2.0, section 9.6.6)
1223                  */
1224                 if (usb_endpoint_xfer_isoc(&ep->desc) ||
1225                                 usb_endpoint_xfer_int(&ep->desc)) {
1226                         max_burst = (le16_to_cpu(ep->desc.wMaxPacketSize)
1227                                      & 0x1800) >> 11;
1228                         ep_ctx->ep_info2 |= cpu_to_le32(MAX_BURST(max_burst));
1229                 }
1230                 /* Fall through */
1231         case USB_SPEED_FULL:
1232         case USB_SPEED_LOW:
1233                 max_packet = GET_MAX_PACKET(le16_to_cpu(ep->desc.wMaxPacketSize));
1234                 ep_ctx->ep_info2 |= cpu_to_le32(MAX_PACKET(max_packet));
1235                 break;
1236         default:
1237                 BUG();
1238         }
1239         max_esit_payload = xhci_get_max_esit_payload(xhci, udev, ep);
1240         ep_ctx->tx_info = cpu_to_le32(MAX_ESIT_PAYLOAD_FOR_EP(max_esit_payload));
1241
1242         /*
1243          * XXX no idea how to calculate the average TRB buffer length for bulk
1244          * endpoints, as the driver gives us no clue how big each scatter gather
1245          * list entry (or buffer) is going to be.
1246          *
1247          * For isochronous and interrupt endpoints, we set it to the max
1248          * available, until we have new API in the USB core to allow drivers to
1249          * declare how much bandwidth they actually need.
1250          *
1251          * Normally, it would be calculated by taking the total of the buffer
1252          * lengths in the TD and then dividing by the number of TRBs in a TD,
1253          * including link TRBs, No-op TRBs, and Event data TRBs.  Since we don't
1254          * use Event Data TRBs, and we don't chain in a link TRB on short
1255          * transfers, we're basically dividing by 1.
1256          *
1257          * xHCI 1.0 specification indicates that the Average TRB Length should
1258          * be set to 8 for control endpoints.
1259          */
1260         if (usb_endpoint_xfer_control(&ep->desc) && xhci->hci_version == 0x100)
1261                 ep_ctx->tx_info |= cpu_to_le32(AVG_TRB_LENGTH_FOR_EP(8));
1262         else
1263                 ep_ctx->tx_info |=
1264                          cpu_to_le32(AVG_TRB_LENGTH_FOR_EP(max_esit_payload));
1265
1266         /* FIXME Debug endpoint context */
1267         return 0;
1268 }
1269
1270 void xhci_endpoint_zero(struct xhci_hcd *xhci,
1271                 struct xhci_virt_device *virt_dev,
1272                 struct usb_host_endpoint *ep)
1273 {
1274         unsigned int ep_index;
1275         struct xhci_ep_ctx *ep_ctx;
1276
1277         ep_index = xhci_get_endpoint_index(&ep->desc);
1278         ep_ctx = xhci_get_ep_ctx(xhci, virt_dev->in_ctx, ep_index);
1279
1280         ep_ctx->ep_info = 0;
1281         ep_ctx->ep_info2 = 0;
1282         ep_ctx->deq = 0;
1283         ep_ctx->tx_info = 0;
1284         /* Don't free the endpoint ring until the set interface or configuration
1285          * request succeeds.
1286          */
1287 }
1288
1289 /* Copy output xhci_ep_ctx to the input xhci_ep_ctx copy.
1290  * Useful when you want to change one particular aspect of the endpoint and then
1291  * issue a configure endpoint command.
1292  */
1293 void xhci_endpoint_copy(struct xhci_hcd *xhci,
1294                 struct xhci_container_ctx *in_ctx,
1295                 struct xhci_container_ctx *out_ctx,
1296                 unsigned int ep_index)
1297 {
1298         struct xhci_ep_ctx *out_ep_ctx;
1299         struct xhci_ep_ctx *in_ep_ctx;
1300
1301         out_ep_ctx = xhci_get_ep_ctx(xhci, out_ctx, ep_index);
1302         in_ep_ctx = xhci_get_ep_ctx(xhci, in_ctx, ep_index);
1303
1304         in_ep_ctx->ep_info = out_ep_ctx->ep_info;
1305         in_ep_ctx->ep_info2 = out_ep_ctx->ep_info2;
1306         in_ep_ctx->deq = out_ep_ctx->deq;
1307         in_ep_ctx->tx_info = out_ep_ctx->tx_info;
1308 }
1309
1310 /* Copy output xhci_slot_ctx to the input xhci_slot_ctx.
1311  * Useful when you want to change one particular aspect of the endpoint and then
1312  * issue a configure endpoint command.  Only the context entries field matters,
1313  * but we'll copy the whole thing anyway.
1314  */
1315 void xhci_slot_copy(struct xhci_hcd *xhci,
1316                 struct xhci_container_ctx *in_ctx,
1317                 struct xhci_container_ctx *out_ctx)
1318 {
1319         struct xhci_slot_ctx *in_slot_ctx;
1320         struct xhci_slot_ctx *out_slot_ctx;
1321
1322         in_slot_ctx = xhci_get_slot_ctx(xhci, in_ctx);
1323         out_slot_ctx = xhci_get_slot_ctx(xhci, out_ctx);
1324
1325         in_slot_ctx->dev_info = out_slot_ctx->dev_info;
1326         in_slot_ctx->dev_info2 = out_slot_ctx->dev_info2;
1327         in_slot_ctx->tt_info = out_slot_ctx->tt_info;
1328         in_slot_ctx->dev_state = out_slot_ctx->dev_state;
1329 }
1330
1331 /* Set up the scratchpad buffer array and scratchpad buffers, if needed. */
1332 static int scratchpad_alloc(struct xhci_hcd *xhci, gfp_t flags)
1333 {
1334         int i;
1335         struct device *dev = xhci_to_hcd(xhci)->self.controller;
1336         int num_sp = HCS_MAX_SCRATCHPAD(xhci->hcs_params2);
1337
1338         xhci_dbg(xhci, "Allocating %d scratchpad buffers\n", num_sp);
1339
1340         if (!num_sp)
1341                 return 0;
1342
1343         xhci->scratchpad = kzalloc(sizeof(*xhci->scratchpad), flags);
1344         if (!xhci->scratchpad)
1345                 goto fail_sp;
1346
1347         xhci->scratchpad->sp_array =
1348                 pci_alloc_consistent(to_pci_dev(dev),
1349                                      num_sp * sizeof(u64),
1350                                      &xhci->scratchpad->sp_dma);
1351         if (!xhci->scratchpad->sp_array)
1352                 goto fail_sp2;
1353
1354         xhci->scratchpad->sp_buffers = kzalloc(sizeof(void *) * num_sp, flags);
1355         if (!xhci->scratchpad->sp_buffers)
1356                 goto fail_sp3;
1357
1358         xhci->scratchpad->sp_dma_buffers =
1359                 kzalloc(sizeof(dma_addr_t) * num_sp, flags);
1360
1361         if (!xhci->scratchpad->sp_dma_buffers)
1362                 goto fail_sp4;
1363
1364         xhci->dcbaa->dev_context_ptrs[0] = cpu_to_le64(xhci->scratchpad->sp_dma);
1365         for (i = 0; i < num_sp; i++) {
1366                 dma_addr_t dma;
1367                 void *buf = pci_alloc_consistent(to_pci_dev(dev),
1368                                                  xhci->page_size, &dma);
1369                 if (!buf)
1370                         goto fail_sp5;
1371
1372                 xhci->scratchpad->sp_array[i] = dma;
1373                 xhci->scratchpad->sp_buffers[i] = buf;
1374                 xhci->scratchpad->sp_dma_buffers[i] = dma;
1375         }
1376
1377         return 0;
1378
1379  fail_sp5:
1380         for (i = i - 1; i >= 0; i--) {
1381                 pci_free_consistent(to_pci_dev(dev), xhci->page_size,
1382                                     xhci->scratchpad->sp_buffers[i],
1383                                     xhci->scratchpad->sp_dma_buffers[i]);
1384         }
1385         kfree(xhci->scratchpad->sp_dma_buffers);
1386
1387  fail_sp4:
1388         kfree(xhci->scratchpad->sp_buffers);
1389
1390  fail_sp3:
1391         pci_free_consistent(to_pci_dev(dev), num_sp * sizeof(u64),
1392                             xhci->scratchpad->sp_array,
1393                             xhci->scratchpad->sp_dma);
1394
1395  fail_sp2:
1396         kfree(xhci->scratchpad);
1397         xhci->scratchpad = NULL;
1398
1399  fail_sp:
1400         return -ENOMEM;
1401 }
1402
1403 static void scratchpad_free(struct xhci_hcd *xhci)
1404 {
1405         int num_sp;
1406         int i;
1407         struct pci_dev  *pdev = to_pci_dev(xhci_to_hcd(xhci)->self.controller);
1408
1409         if (!xhci->scratchpad)
1410                 return;
1411
1412         num_sp = HCS_MAX_SCRATCHPAD(xhci->hcs_params2);
1413
1414         for (i = 0; i < num_sp; i++) {
1415                 pci_free_consistent(pdev, xhci->page_size,
1416                                     xhci->scratchpad->sp_buffers[i],
1417                                     xhci->scratchpad->sp_dma_buffers[i]);
1418         }
1419         kfree(xhci->scratchpad->sp_dma_buffers);
1420         kfree(xhci->scratchpad->sp_buffers);
1421         pci_free_consistent(pdev, num_sp * sizeof(u64),
1422                             xhci->scratchpad->sp_array,
1423                             xhci->scratchpad->sp_dma);
1424         kfree(xhci->scratchpad);
1425         xhci->scratchpad = NULL;
1426 }
1427
1428 struct xhci_command *xhci_alloc_command(struct xhci_hcd *xhci,
1429                 bool allocate_in_ctx, bool allocate_completion,
1430                 gfp_t mem_flags)
1431 {
1432         struct xhci_command *command;
1433
1434         command = kzalloc(sizeof(*command), mem_flags);
1435         if (!command)
1436                 return NULL;
1437
1438         if (allocate_in_ctx) {
1439                 command->in_ctx =
1440                         xhci_alloc_container_ctx(xhci, XHCI_CTX_TYPE_INPUT,
1441                                         mem_flags);
1442                 if (!command->in_ctx) {
1443                         kfree(command);
1444                         return NULL;
1445                 }
1446         }
1447
1448         if (allocate_completion) {
1449                 command->completion =
1450                         kzalloc(sizeof(struct completion), mem_flags);
1451                 if (!command->completion) {
1452                         xhci_free_container_ctx(xhci, command->in_ctx);
1453                         kfree(command);
1454                         return NULL;
1455                 }
1456                 init_completion(command->completion);
1457         }
1458
1459         command->status = 0;
1460         INIT_LIST_HEAD(&command->cmd_list);
1461         return command;
1462 }
1463
1464 void xhci_urb_free_priv(struct xhci_hcd *xhci, struct urb_priv *urb_priv)
1465 {
1466         int last;
1467
1468         if (!urb_priv)
1469                 return;
1470
1471         last = urb_priv->length - 1;
1472         if (last >= 0) {
1473                 int     i;
1474                 for (i = 0; i <= last; i++)
1475                         kfree(urb_priv->td[i]);
1476         }
1477         kfree(urb_priv);
1478 }
1479
1480 void xhci_free_command(struct xhci_hcd *xhci,
1481                 struct xhci_command *command)
1482 {
1483         xhci_free_container_ctx(xhci,
1484                         command->in_ctx);
1485         kfree(command->completion);
1486         kfree(command);
1487 }
1488
1489 void xhci_mem_cleanup(struct xhci_hcd *xhci)
1490 {
1491         struct pci_dev  *pdev = to_pci_dev(xhci_to_hcd(xhci)->self.controller);
1492         int size;
1493         int i;
1494
1495         /* Free the Event Ring Segment Table and the actual Event Ring */
1496         if (xhci->ir_set) {
1497                 xhci_writel(xhci, 0, &xhci->ir_set->erst_size);
1498                 xhci_write_64(xhci, 0, &xhci->ir_set->erst_base);
1499                 xhci_write_64(xhci, 0, &xhci->ir_set->erst_dequeue);
1500         }
1501         size = sizeof(struct xhci_erst_entry)*(xhci->erst.num_entries);
1502         if (xhci->erst.entries)
1503                 pci_free_consistent(pdev, size,
1504                                 xhci->erst.entries, xhci->erst.erst_dma_addr);
1505         xhci->erst.entries = NULL;
1506         xhci_dbg(xhci, "Freed ERST\n");
1507         if (xhci->event_ring)
1508                 xhci_ring_free(xhci, xhci->event_ring);
1509         xhci->event_ring = NULL;
1510         xhci_dbg(xhci, "Freed event ring\n");
1511
1512         xhci_write_64(xhci, 0, &xhci->op_regs->cmd_ring);
1513         if (xhci->cmd_ring)
1514                 xhci_ring_free(xhci, xhci->cmd_ring);
1515         xhci->cmd_ring = NULL;
1516         xhci_dbg(xhci, "Freed command ring\n");
1517
1518         for (i = 1; i < MAX_HC_SLOTS; ++i)
1519                 xhci_free_virt_device(xhci, i);
1520
1521         if (xhci->segment_pool)
1522                 dma_pool_destroy(xhci->segment_pool);
1523         xhci->segment_pool = NULL;
1524         xhci_dbg(xhci, "Freed segment pool\n");
1525
1526         if (xhci->device_pool)
1527                 dma_pool_destroy(xhci->device_pool);
1528         xhci->device_pool = NULL;
1529         xhci_dbg(xhci, "Freed device context pool\n");
1530
1531         if (xhci->small_streams_pool)
1532                 dma_pool_destroy(xhci->small_streams_pool);
1533         xhci->small_streams_pool = NULL;
1534         xhci_dbg(xhci, "Freed small stream array pool\n");
1535
1536         if (xhci->medium_streams_pool)
1537                 dma_pool_destroy(xhci->medium_streams_pool);
1538         xhci->medium_streams_pool = NULL;
1539         xhci_dbg(xhci, "Freed medium stream array pool\n");
1540
1541         xhci_write_64(xhci, 0, &xhci->op_regs->dcbaa_ptr);
1542         if (xhci->dcbaa)
1543                 pci_free_consistent(pdev, sizeof(*xhci->dcbaa),
1544                                 xhci->dcbaa, xhci->dcbaa->dma);
1545         xhci->dcbaa = NULL;
1546
1547         scratchpad_free(xhci);
1548
1549         xhci->num_usb2_ports = 0;
1550         xhci->num_usb3_ports = 0;
1551         kfree(xhci->usb2_ports);
1552         kfree(xhci->usb3_ports);
1553         kfree(xhci->port_array);
1554
1555         xhci->page_size = 0;
1556         xhci->page_shift = 0;
1557         xhci->bus_state[0].bus_suspended = 0;
1558         xhci->bus_state[1].bus_suspended = 0;
1559 }
1560
1561 static int xhci_test_trb_in_td(struct xhci_hcd *xhci,
1562                 struct xhci_segment *input_seg,
1563                 union xhci_trb *start_trb,
1564                 union xhci_trb *end_trb,
1565                 dma_addr_t input_dma,
1566                 struct xhci_segment *result_seg,
1567                 char *test_name, int test_number)
1568 {
1569         unsigned long long start_dma;
1570         unsigned long long end_dma;
1571         struct xhci_segment *seg;
1572
1573         start_dma = xhci_trb_virt_to_dma(input_seg, start_trb);
1574         end_dma = xhci_trb_virt_to_dma(input_seg, end_trb);
1575
1576         seg = trb_in_td(input_seg, start_trb, end_trb, input_dma);
1577         if (seg != result_seg) {
1578                 xhci_warn(xhci, "WARN: %s TRB math test %d failed!\n",
1579                                 test_name, test_number);
1580                 xhci_warn(xhci, "Tested TRB math w/ seg %p and "
1581                                 "input DMA 0x%llx\n",
1582                                 input_seg,
1583                                 (unsigned long long) input_dma);
1584                 xhci_warn(xhci, "starting TRB %p (0x%llx DMA), "
1585                                 "ending TRB %p (0x%llx DMA)\n",
1586                                 start_trb, start_dma,
1587                                 end_trb, end_dma);
1588                 xhci_warn(xhci, "Expected seg %p, got seg %p\n",
1589                                 result_seg, seg);
1590                 return -1;
1591         }
1592         return 0;
1593 }
1594
1595 /* TRB math checks for xhci_trb_in_td(), using the command and event rings. */
1596 static int xhci_check_trb_in_td_math(struct xhci_hcd *xhci, gfp_t mem_flags)
1597 {
1598         struct {
1599                 dma_addr_t              input_dma;
1600                 struct xhci_segment     *result_seg;
1601         } simple_test_vector [] = {
1602                 /* A zeroed DMA field should fail */
1603                 { 0, NULL },
1604                 /* One TRB before the ring start should fail */
1605                 { xhci->event_ring->first_seg->dma - 16, NULL },
1606                 /* One byte before the ring start should fail */
1607                 { xhci->event_ring->first_seg->dma - 1, NULL },
1608                 /* Starting TRB should succeed */
1609                 { xhci->event_ring->first_seg->dma, xhci->event_ring->first_seg },
1610                 /* Ending TRB should succeed */
1611                 { xhci->event_ring->first_seg->dma + (TRBS_PER_SEGMENT - 1)*16,
1612                         xhci->event_ring->first_seg },
1613                 /* One byte after the ring end should fail */
1614                 { xhci->event_ring->first_seg->dma + (TRBS_PER_SEGMENT - 1)*16 + 1, NULL },
1615                 /* One TRB after the ring end should fail */
1616                 { xhci->event_ring->first_seg->dma + (TRBS_PER_SEGMENT)*16, NULL },
1617                 /* An address of all ones should fail */
1618                 { (dma_addr_t) (~0), NULL },
1619         };
1620         struct {
1621                 struct xhci_segment     *input_seg;
1622                 union xhci_trb          *start_trb;
1623                 union xhci_trb          *end_trb;
1624                 dma_addr_t              input_dma;
1625                 struct xhci_segment     *result_seg;
1626         } complex_test_vector [] = {
1627                 /* Test feeding a valid DMA address from a different ring */
1628                 {       .input_seg = xhci->event_ring->first_seg,
1629                         .start_trb = xhci->event_ring->first_seg->trbs,
1630                         .end_trb = &xhci->event_ring->first_seg->trbs[TRBS_PER_SEGMENT - 1],
1631                         .input_dma = xhci->cmd_ring->first_seg->dma,
1632                         .result_seg = NULL,
1633                 },
1634                 /* Test feeding a valid end TRB from a different ring */
1635                 {       .input_seg = xhci->event_ring->first_seg,
1636                         .start_trb = xhci->event_ring->first_seg->trbs,
1637                         .end_trb = &xhci->cmd_ring->first_seg->trbs[TRBS_PER_SEGMENT - 1],
1638                         .input_dma = xhci->cmd_ring->first_seg->dma,
1639                         .result_seg = NULL,
1640                 },
1641                 /* Test feeding a valid start and end TRB from a different ring */
1642                 {       .input_seg = xhci->event_ring->first_seg,
1643                         .start_trb = xhci->cmd_ring->first_seg->trbs,
1644                         .end_trb = &xhci->cmd_ring->first_seg->trbs[TRBS_PER_SEGMENT - 1],
1645                         .input_dma = xhci->cmd_ring->first_seg->dma,
1646                         .result_seg = NULL,
1647                 },
1648                 /* TRB in this ring, but after this TD */
1649                 {       .input_seg = xhci->event_ring->first_seg,
1650                         .start_trb = &xhci->event_ring->first_seg->trbs[0],
1651                         .end_trb = &xhci->event_ring->first_seg->trbs[3],
1652                         .input_dma = xhci->event_ring->first_seg->dma + 4*16,
1653                         .result_seg = NULL,
1654                 },
1655                 /* TRB in this ring, but before this TD */
1656                 {       .input_seg = xhci->event_ring->first_seg,
1657                         .start_trb = &xhci->event_ring->first_seg->trbs[3],
1658                         .end_trb = &xhci->event_ring->first_seg->trbs[6],
1659                         .input_dma = xhci->event_ring->first_seg->dma + 2*16,
1660                         .result_seg = NULL,
1661                 },
1662                 /* TRB in this ring, but after this wrapped TD */
1663                 {       .input_seg = xhci->event_ring->first_seg,
1664                         .start_trb = &xhci->event_ring->first_seg->trbs[TRBS_PER_SEGMENT - 3],
1665                         .end_trb = &xhci->event_ring->first_seg->trbs[1],
1666                         .input_dma = xhci->event_ring->first_seg->dma + 2*16,
1667                         .result_seg = NULL,
1668                 },
1669                 /* TRB in this ring, but before this wrapped TD */
1670                 {       .input_seg = xhci->event_ring->first_seg,
1671                         .start_trb = &xhci->event_ring->first_seg->trbs[TRBS_PER_SEGMENT - 3],
1672                         .end_trb = &xhci->event_ring->first_seg->trbs[1],
1673                         .input_dma = xhci->event_ring->first_seg->dma + (TRBS_PER_SEGMENT - 4)*16,
1674                         .result_seg = NULL,
1675                 },
1676                 /* TRB not in this ring, and we have a wrapped TD */
1677                 {       .input_seg = xhci->event_ring->first_seg,
1678                         .start_trb = &xhci->event_ring->first_seg->trbs[TRBS_PER_SEGMENT - 3],
1679                         .end_trb = &xhci->event_ring->first_seg->trbs[1],
1680                         .input_dma = xhci->cmd_ring->first_seg->dma + 2*16,
1681                         .result_seg = NULL,
1682                 },
1683         };
1684
1685         unsigned int num_tests;
1686         int i, ret;
1687
1688         num_tests = ARRAY_SIZE(simple_test_vector);
1689         for (i = 0; i < num_tests; i++) {
1690                 ret = xhci_test_trb_in_td(xhci,
1691                                 xhci->event_ring->first_seg,
1692                                 xhci->event_ring->first_seg->trbs,
1693                                 &xhci->event_ring->first_seg->trbs[TRBS_PER_SEGMENT - 1],
1694                                 simple_test_vector[i].input_dma,
1695                                 simple_test_vector[i].result_seg,
1696                                 "Simple", i);
1697                 if (ret < 0)
1698                         return ret;
1699         }
1700
1701         num_tests = ARRAY_SIZE(complex_test_vector);
1702         for (i = 0; i < num_tests; i++) {
1703                 ret = xhci_test_trb_in_td(xhci,
1704                                 complex_test_vector[i].input_seg,
1705                                 complex_test_vector[i].start_trb,
1706                                 complex_test_vector[i].end_trb,
1707                                 complex_test_vector[i].input_dma,
1708                                 complex_test_vector[i].result_seg,
1709                                 "Complex", i);
1710                 if (ret < 0)
1711                         return ret;
1712         }
1713         xhci_dbg(xhci, "TRB math tests passed.\n");
1714         return 0;
1715 }
1716
1717 static void xhci_set_hc_event_deq(struct xhci_hcd *xhci)
1718 {
1719         u64 temp;
1720         dma_addr_t deq;
1721
1722         deq = xhci_trb_virt_to_dma(xhci->event_ring->deq_seg,
1723                         xhci->event_ring->dequeue);
1724         if (deq == 0 && !in_interrupt())
1725                 xhci_warn(xhci, "WARN something wrong with SW event ring "
1726                                 "dequeue ptr.\n");
1727         /* Update HC event ring dequeue pointer */
1728         temp = xhci_read_64(xhci, &xhci->ir_set->erst_dequeue);
1729         temp &= ERST_PTR_MASK;
1730         /* Don't clear the EHB bit (which is RW1C) because
1731          * there might be more events to service.
1732          */
1733         temp &= ~ERST_EHB;
1734         xhci_dbg(xhci, "// Write event ring dequeue pointer, "
1735                         "preserving EHB bit\n");
1736         xhci_write_64(xhci, ((u64) deq & (u64) ~ERST_PTR_MASK) | temp,
1737                         &xhci->ir_set->erst_dequeue);
1738 }
1739
1740 static void xhci_add_in_port(struct xhci_hcd *xhci, unsigned int num_ports,
1741                 __le32 __iomem *addr, u8 major_revision)
1742 {
1743         u32 temp, port_offset, port_count;
1744         int i;
1745
1746         if (major_revision > 0x03) {
1747                 xhci_warn(xhci, "Ignoring unknown port speed, "
1748                                 "Ext Cap %p, revision = 0x%x\n",
1749                                 addr, major_revision);
1750                 /* Ignoring port protocol we can't understand. FIXME */
1751                 return;
1752         }
1753
1754         /* Port offset and count in the third dword, see section 7.2 */
1755         temp = xhci_readl(xhci, addr + 2);
1756         port_offset = XHCI_EXT_PORT_OFF(temp);
1757         port_count = XHCI_EXT_PORT_COUNT(temp);
1758         xhci_dbg(xhci, "Ext Cap %p, port offset = %u, "
1759                         "count = %u, revision = 0x%x\n",
1760                         addr, port_offset, port_count, major_revision);
1761         /* Port count includes the current port offset */
1762         if (port_offset == 0 || (port_offset + port_count - 1) > num_ports)
1763                 /* WTF? "Valid values are ‘1’ to MaxPorts" */
1764                 return;
1765         port_offset--;
1766         for (i = port_offset; i < (port_offset + port_count); i++) {
1767                 /* Duplicate entry.  Ignore the port if the revisions differ. */
1768                 if (xhci->port_array[i] != 0) {
1769                         xhci_warn(xhci, "Duplicate port entry, Ext Cap %p,"
1770                                         " port %u\n", addr, i);
1771                         xhci_warn(xhci, "Port was marked as USB %u, "
1772                                         "duplicated as USB %u\n",
1773                                         xhci->port_array[i], major_revision);
1774                         /* Only adjust the roothub port counts if we haven't
1775                          * found a similar duplicate.
1776                          */
1777                         if (xhci->port_array[i] != major_revision &&
1778                                 xhci->port_array[i] != DUPLICATE_ENTRY) {
1779                                 if (xhci->port_array[i] == 0x03)
1780                                         xhci->num_usb3_ports--;
1781                                 else
1782                                         xhci->num_usb2_ports--;
1783                                 xhci->port_array[i] = DUPLICATE_ENTRY;
1784                         }
1785                         /* FIXME: Should we disable the port? */
1786                         continue;
1787                 }
1788                 xhci->port_array[i] = major_revision;
1789                 if (major_revision == 0x03)
1790                         xhci->num_usb3_ports++;
1791                 else
1792                         xhci->num_usb2_ports++;
1793         }
1794         /* FIXME: Should we disable ports not in the Extended Capabilities? */
1795 }
1796
1797 /*
1798  * Scan the Extended Capabilities for the "Supported Protocol Capabilities" that
1799  * specify what speeds each port is supposed to be.  We can't count on the port
1800  * speed bits in the PORTSC register being correct until a device is connected,
1801  * but we need to set up the two fake roothubs with the correct number of USB
1802  * 3.0 and USB 2.0 ports at host controller initialization time.
1803  */
1804 static int xhci_setup_port_arrays(struct xhci_hcd *xhci, gfp_t flags)
1805 {
1806         __le32 __iomem *addr;
1807         u32 offset;
1808         unsigned int num_ports;
1809         int i, port_index;
1810
1811         addr = &xhci->cap_regs->hcc_params;
1812         offset = XHCI_HCC_EXT_CAPS(xhci_readl(xhci, addr));
1813         if (offset == 0) {
1814                 xhci_err(xhci, "No Extended Capability registers, "
1815                                 "unable to set up roothub.\n");
1816                 return -ENODEV;
1817         }
1818
1819         num_ports = HCS_MAX_PORTS(xhci->hcs_params1);
1820         xhci->port_array = kzalloc(sizeof(*xhci->port_array)*num_ports, flags);
1821         if (!xhci->port_array)
1822                 return -ENOMEM;
1823
1824         /*
1825          * For whatever reason, the first capability offset is from the
1826          * capability register base, not from the HCCPARAMS register.
1827          * See section 5.3.6 for offset calculation.
1828          */
1829         addr = &xhci->cap_regs->hc_capbase + offset;
1830         while (1) {
1831                 u32 cap_id;
1832
1833                 cap_id = xhci_readl(xhci, addr);
1834                 if (XHCI_EXT_CAPS_ID(cap_id) == XHCI_EXT_CAPS_PROTOCOL)
1835                         xhci_add_in_port(xhci, num_ports, addr,
1836                                         (u8) XHCI_EXT_PORT_MAJOR(cap_id));
1837                 offset = XHCI_EXT_CAPS_NEXT(cap_id);
1838                 if (!offset || (xhci->num_usb2_ports + xhci->num_usb3_ports)
1839                                 == num_ports)
1840                         break;
1841                 /*
1842                  * Once you're into the Extended Capabilities, the offset is
1843                  * always relative to the register holding the offset.
1844                  */
1845                 addr += offset;
1846         }
1847
1848         if (xhci->num_usb2_ports == 0 && xhci->num_usb3_ports == 0) {
1849                 xhci_warn(xhci, "No ports on the roothubs?\n");
1850                 return -ENODEV;
1851         }
1852         xhci_dbg(xhci, "Found %u USB 2.0 ports and %u USB 3.0 ports.\n",
1853                         xhci->num_usb2_ports, xhci->num_usb3_ports);
1854
1855         /* Place limits on the number of roothub ports so that the hub
1856          * descriptors aren't longer than the USB core will allocate.
1857          */
1858         if (xhci->num_usb3_ports > 15) {
1859                 xhci_dbg(xhci, "Limiting USB 3.0 roothub ports to 15.\n");
1860                 xhci->num_usb3_ports = 15;
1861         }
1862         if (xhci->num_usb2_ports > USB_MAXCHILDREN) {
1863                 xhci_dbg(xhci, "Limiting USB 2.0 roothub ports to %u.\n",
1864                                 USB_MAXCHILDREN);
1865                 xhci->num_usb2_ports = USB_MAXCHILDREN;
1866         }
1867
1868         /*
1869          * Note we could have all USB 3.0 ports, or all USB 2.0 ports.
1870          * Not sure how the USB core will handle a hub with no ports...
1871          */
1872         if (xhci->num_usb2_ports) {
1873                 xhci->usb2_ports = kmalloc(sizeof(*xhci->usb2_ports)*
1874                                 xhci->num_usb2_ports, flags);
1875                 if (!xhci->usb2_ports)
1876                         return -ENOMEM;
1877
1878                 port_index = 0;
1879                 for (i = 0; i < num_ports; i++) {
1880                         if (xhci->port_array[i] == 0x03 ||
1881                                         xhci->port_array[i] == 0 ||
1882                                         xhci->port_array[i] == DUPLICATE_ENTRY)
1883                                 continue;
1884
1885                         xhci->usb2_ports[port_index] =
1886                                 &xhci->op_regs->port_status_base +
1887                                 NUM_PORT_REGS*i;
1888                         xhci_dbg(xhci, "USB 2.0 port at index %u, "
1889                                         "addr = %p\n", i,
1890                                         xhci->usb2_ports[port_index]);
1891                         port_index++;
1892                         if (port_index == xhci->num_usb2_ports)
1893                                 break;
1894                 }
1895         }
1896         if (xhci->num_usb3_ports) {
1897                 xhci->usb3_ports = kmalloc(sizeof(*xhci->usb3_ports)*
1898                                 xhci->num_usb3_ports, flags);
1899                 if (!xhci->usb3_ports)
1900                         return -ENOMEM;
1901
1902                 port_index = 0;
1903                 for (i = 0; i < num_ports; i++)
1904                         if (xhci->port_array[i] == 0x03) {
1905                                 xhci->usb3_ports[port_index] =
1906                                         &xhci->op_regs->port_status_base +
1907                                         NUM_PORT_REGS*i;
1908                                 xhci_dbg(xhci, "USB 3.0 port at index %u, "
1909                                                 "addr = %p\n", i,
1910                                                 xhci->usb3_ports[port_index]);
1911                                 port_index++;
1912                                 if (port_index == xhci->num_usb3_ports)
1913                                         break;
1914                         }
1915         }
1916         return 0;
1917 }
1918
1919 int xhci_mem_init(struct xhci_hcd *xhci, gfp_t flags)
1920 {
1921         dma_addr_t      dma;
1922         struct device   *dev = xhci_to_hcd(xhci)->self.controller;
1923         unsigned int    val, val2;
1924         u64             val_64;
1925         struct xhci_segment     *seg;
1926         u32 page_size;
1927         int i;
1928
1929         page_size = xhci_readl(xhci, &xhci->op_regs->page_size);
1930         xhci_dbg(xhci, "Supported page size register = 0x%x\n", page_size);
1931         for (i = 0; i < 16; i++) {
1932                 if ((0x1 & page_size) != 0)
1933                         break;
1934                 page_size = page_size >> 1;
1935         }
1936         if (i < 16)
1937                 xhci_dbg(xhci, "Supported page size of %iK\n", (1 << (i+12)) / 1024);
1938         else
1939                 xhci_warn(xhci, "WARN: no supported page size\n");
1940         /* Use 4K pages, since that's common and the minimum the HC supports */
1941         xhci->page_shift = 12;
1942         xhci->page_size = 1 << xhci->page_shift;
1943         xhci_dbg(xhci, "HCD page size set to %iK\n", xhci->page_size / 1024);
1944
1945         /*
1946          * Program the Number of Device Slots Enabled field in the CONFIG
1947          * register with the max value of slots the HC can handle.
1948          */
1949         val = HCS_MAX_SLOTS(xhci_readl(xhci, &xhci->cap_regs->hcs_params1));
1950         xhci_dbg(xhci, "// xHC can handle at most %d device slots.\n",
1951                         (unsigned int) val);
1952         val2 = xhci_readl(xhci, &xhci->op_regs->config_reg);
1953         val |= (val2 & ~HCS_SLOTS_MASK);
1954         xhci_dbg(xhci, "// Setting Max device slots reg = 0x%x.\n",
1955                         (unsigned int) val);
1956         xhci_writel(xhci, val, &xhci->op_regs->config_reg);
1957
1958         /*
1959          * Section 5.4.8 - doorbell array must be
1960          * "physically contiguous and 64-byte (cache line) aligned".
1961          */
1962         xhci->dcbaa = pci_alloc_consistent(to_pci_dev(dev),
1963                         sizeof(*xhci->dcbaa), &dma);
1964         if (!xhci->dcbaa)
1965                 goto fail;
1966         memset(xhci->dcbaa, 0, sizeof *(xhci->dcbaa));
1967         xhci->dcbaa->dma = dma;
1968         xhci_dbg(xhci, "// Device context base array address = 0x%llx (DMA), %p (virt)\n",
1969                         (unsigned long long)xhci->dcbaa->dma, xhci->dcbaa);
1970         xhci_write_64(xhci, dma, &xhci->op_regs->dcbaa_ptr);
1971
1972         /*
1973          * Initialize the ring segment pool.  The ring must be a contiguous
1974          * structure comprised of TRBs.  The TRBs must be 16 byte aligned,
1975          * however, the command ring segment needs 64-byte aligned segments,
1976          * so we pick the greater alignment need.
1977          */
1978         xhci->segment_pool = dma_pool_create("xHCI ring segments", dev,
1979                         SEGMENT_SIZE, 64, xhci->page_size);
1980
1981         /* See Table 46 and Note on Figure 55 */
1982         xhci->device_pool = dma_pool_create("xHCI input/output contexts", dev,
1983                         2112, 64, xhci->page_size);
1984         if (!xhci->segment_pool || !xhci->device_pool)
1985                 goto fail;
1986
1987         /* Linear stream context arrays don't have any boundary restrictions,
1988          * and only need to be 16-byte aligned.
1989          */
1990         xhci->small_streams_pool =
1991                 dma_pool_create("xHCI 256 byte stream ctx arrays",
1992                         dev, SMALL_STREAM_ARRAY_SIZE, 16, 0);
1993         xhci->medium_streams_pool =
1994                 dma_pool_create("xHCI 1KB stream ctx arrays",
1995                         dev, MEDIUM_STREAM_ARRAY_SIZE, 16, 0);
1996         /* Any stream context array bigger than MEDIUM_STREAM_ARRAY_SIZE
1997          * will be allocated with pci_alloc_consistent()
1998          */
1999
2000         if (!xhci->small_streams_pool || !xhci->medium_streams_pool)
2001                 goto fail;
2002
2003         /* Set up the command ring to have one segments for now. */
2004         xhci->cmd_ring = xhci_ring_alloc(xhci, 1, true, flags);
2005         if (!xhci->cmd_ring)
2006                 goto fail;
2007         xhci_dbg(xhci, "Allocated command ring at %p\n", xhci->cmd_ring);
2008         xhci_dbg(xhci, "First segment DMA is 0x%llx\n",
2009                         (unsigned long long)xhci->cmd_ring->first_seg->dma);
2010
2011         /* Set the address in the Command Ring Control register */
2012         val_64 = xhci_read_64(xhci, &xhci->op_regs->cmd_ring);
2013         val_64 = (val_64 & (u64) CMD_RING_RSVD_BITS) |
2014                 (xhci->cmd_ring->first_seg->dma & (u64) ~CMD_RING_RSVD_BITS) |
2015                 xhci->cmd_ring->cycle_state;
2016         xhci_dbg(xhci, "// Setting command ring address to 0x%x\n", val);
2017         xhci_write_64(xhci, val_64, &xhci->op_regs->cmd_ring);
2018         xhci_dbg_cmd_ptrs(xhci);
2019
2020         val = xhci_readl(xhci, &xhci->cap_regs->db_off);
2021         val &= DBOFF_MASK;
2022         xhci_dbg(xhci, "// Doorbell array is located at offset 0x%x"
2023                         " from cap regs base addr\n", val);
2024         xhci->dba = (void __iomem *) xhci->cap_regs + val;
2025         xhci_dbg_regs(xhci);
2026         xhci_print_run_regs(xhci);
2027         /* Set ir_set to interrupt register set 0 */
2028         xhci->ir_set = &xhci->run_regs->ir_set[0];
2029
2030         /*
2031          * Event ring setup: Allocate a normal ring, but also setup
2032          * the event ring segment table (ERST).  Section 4.9.3.
2033          */
2034         xhci_dbg(xhci, "// Allocating event ring\n");
2035         xhci->event_ring = xhci_ring_alloc(xhci, ERST_NUM_SEGS, false, flags);
2036         if (!xhci->event_ring)
2037                 goto fail;
2038         if (xhci_check_trb_in_td_math(xhci, flags) < 0)
2039                 goto fail;
2040
2041         xhci->erst.entries = pci_alloc_consistent(to_pci_dev(dev),
2042                         sizeof(struct xhci_erst_entry)*ERST_NUM_SEGS, &dma);
2043         if (!xhci->erst.entries)
2044                 goto fail;
2045         xhci_dbg(xhci, "// Allocated event ring segment table at 0x%llx\n",
2046                         (unsigned long long)dma);
2047
2048         memset(xhci->erst.entries, 0, sizeof(struct xhci_erst_entry)*ERST_NUM_SEGS);
2049         xhci->erst.num_entries = ERST_NUM_SEGS;
2050         xhci->erst.erst_dma_addr = dma;
2051         xhci_dbg(xhci, "Set ERST to 0; private num segs = %i, virt addr = %p, dma addr = 0x%llx\n",
2052                         xhci->erst.num_entries,
2053                         xhci->erst.entries,
2054                         (unsigned long long)xhci->erst.erst_dma_addr);
2055
2056         /* set ring base address and size for each segment table entry */
2057         for (val = 0, seg = xhci->event_ring->first_seg; val < ERST_NUM_SEGS; val++) {
2058                 struct xhci_erst_entry *entry = &xhci->erst.entries[val];
2059                 entry->seg_addr = cpu_to_le64(seg->dma);
2060                 entry->seg_size = cpu_to_le32(TRBS_PER_SEGMENT);
2061                 entry->rsvd = 0;
2062                 seg = seg->next;
2063         }
2064
2065         /* set ERST count with the number of entries in the segment table */
2066         val = xhci_readl(xhci, &xhci->ir_set->erst_size);
2067         val &= ERST_SIZE_MASK;
2068         val |= ERST_NUM_SEGS;
2069         xhci_dbg(xhci, "// Write ERST size = %i to ir_set 0 (some bits preserved)\n",
2070                         val);
2071         xhci_writel(xhci, val, &xhci->ir_set->erst_size);
2072
2073         xhci_dbg(xhci, "// Set ERST entries to point to event ring.\n");
2074         /* set the segment table base address */
2075         xhci_dbg(xhci, "// Set ERST base address for ir_set 0 = 0x%llx\n",
2076                         (unsigned long long)xhci->erst.erst_dma_addr);
2077         val_64 = xhci_read_64(xhci, &xhci->ir_set->erst_base);
2078         val_64 &= ERST_PTR_MASK;
2079         val_64 |= (xhci->erst.erst_dma_addr & (u64) ~ERST_PTR_MASK);
2080         xhci_write_64(xhci, val_64, &xhci->ir_set->erst_base);
2081
2082         /* Set the event ring dequeue address */
2083         xhci_set_hc_event_deq(xhci);
2084         xhci_dbg(xhci, "Wrote ERST address to ir_set 0.\n");
2085         xhci_print_ir_set(xhci, 0);
2086
2087         /*
2088          * XXX: Might need to set the Interrupter Moderation Register to
2089          * something other than the default (~1ms minimum between interrupts).
2090          * See section 5.5.1.2.
2091          */
2092         init_completion(&xhci->addr_dev);
2093         for (i = 0; i < MAX_HC_SLOTS; ++i)
2094                 xhci->devs[i] = NULL;
2095         for (i = 0; i < USB_MAXCHILDREN; ++i) {
2096                 xhci->bus_state[0].resume_done[i] = 0;
2097                 xhci->bus_state[1].resume_done[i] = 0;
2098         }
2099
2100         if (scratchpad_alloc(xhci, flags))
2101                 goto fail;
2102         if (xhci_setup_port_arrays(xhci, flags))
2103                 goto fail;
2104
2105         return 0;
2106
2107 fail:
2108         xhci_warn(xhci, "Couldn't initialize memory\n");
2109         xhci_mem_cleanup(xhci);
2110         return -ENOMEM;
2111 }