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->trbs) {
  65                 xhci_dbg(xhci, "Freeing DMA segment at %p (virtual) 0x%llx (DMA)\n",
  66                                 seg->trbs, (unsigned long long)seg->dma);
  67                 dma_pool_free(xhci->segment_pool, seg->trbs, seg->dma);
  68                 seg->trbs = NULL;
  69         }
  70         xhci_dbg(xhci, "Freeing priv segment structure at %p\n", seg);
  71         kfree(seg);
  72 }
  73
  74 /*
  75  * Make the prev segment point to the next segment.
  76  *
  77  * Change the last TRB in the prev segment to be a Link TRB which points to the
  78  * DMA address of the next segment.  The caller needs to set any Link TRB
  79  * related flags, such as End TRB, Toggle Cycle, and no snoop.
  80  */
  81 static void xhci_link_segments(struct xhci_hcd *xhci, struct xhci_segment *prev,
  82                 struct xhci_segment *next, bool link_trbs, bool isoc)
  83 {
  84         u32 val;
  85
  86         if (!prev || !next)
  87                 return;
  88         prev->next = next;
  89         if (link_trbs) {
  90                 prev->trbs[TRBS_PER_SEGMENT-1].link.segment_ptr =
  91                         cpu_to_le64(next->dma);
  92
  93                 /* Set the last TRB in the segment to have a TRB type ID of Link TRB */
  94                 val = le32_to_cpu(prev->trbs[TRBS_PER_SEGMENT-1].link.control);
  95                 val &= ~TRB_TYPE_BITMASK;
  96                 val |= TRB_TYPE(TRB_LINK);
  97                 /* Always set the chain bit with 0.95 hardware */
  98                 /* Set chain bit for isoc rings on AMD 0.96 host */
  99                 if (xhci_link_trb_quirk(xhci) ||
 100                                 (isoc && (xhci->quirks & XHCI_AMD_0x96_HOST)))
 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)
 116                 return;
 117         if (ring->first_seg) {
 118                 first_seg = ring->first_seg;
 119                 seg = first_seg->next;
 120                 xhci_dbg(xhci, "Freeing ring at %p\n", ring);
 121                 while (seg != first_seg) {
 122                         struct xhci_segment *next = seg->next;
 123                         xhci_segment_free(xhci, seg);
 124                         seg = next;
 125                 }
 126                 xhci_segment_free(xhci, first_seg);
 127                 ring->first_seg = NULL;
 128         }
 129         kfree(ring);
 130 }
 131
 132 static void xhci_initialize_ring_info(struct xhci_ring *ring)
 133 {
 134         /* The ring is empty, so the enqueue pointer == dequeue pointer */
 135         ring->enqueue = ring->first_seg->trbs;
 136         ring->enq_seg = ring->first_seg;
 137         ring->dequeue = ring->enqueue;
 138         ring->deq_seg = ring->first_seg;
 139         /* The ring is initialized to 0. The producer must write 1 to the cycle
 140          * bit to handover ownership of the TRB, so PCS = 1.  The consumer must
 141          * compare CCS to the cycle bit to check ownership, so CCS = 1.
 142          */
 143         ring->cycle_state = 1;
 144         /* Not necessary for new rings, but needed for re-initialized rings */
 145         ring->enq_updates = 0;
 146         ring->deq_updates = 0;
 147 }
 148
 149 /**
 150  * Create a new ring with zero or more segments.
 151  *
 152  * Link each segment together into a ring.
 153  * Set the end flag and the cycle toggle bit on the last segment.
 154  * See section 4.9.1 and figures 15 and 16.
 155  */
 156 static struct xhci_ring *xhci_ring_alloc(struct xhci_hcd *xhci,
 157                 unsigned int num_segs, bool link_trbs, bool isoc, gfp_t flags)
 158 {
 159         struct xhci_ring        *ring;
 160         struct xhci_segment     *prev;
 161
 162         ring = kzalloc(sizeof *(ring), flags);
 163         xhci_dbg(xhci, "Allocating ring at %p\n", ring);
 164         if (!ring)
 165                 return NULL;
 166
 167         INIT_LIST_HEAD(&ring->td_list);
 168         if (num_segs == 0)
 169                 return ring;
 170
 171         ring->first_seg = xhci_segment_alloc(xhci, flags);
 172         if (!ring->first_seg)
 173                 goto fail;
 174         num_segs--;
 175
 176         prev = ring->first_seg;
 177         while (num_segs > 0) {
 178                 struct xhci_segment     *next;
 179
 180                 next = xhci_segment_alloc(xhci, flags);
 181                 if (!next) {
 182                         prev = ring->first_seg;
 183                         while (prev) {
 184                                 next = prev->next;
 185                                 xhci_segment_free(xhci, prev);
 186                                 prev = next;
 187                         }
 188                         goto fail;
 189                 }
 190                 xhci_link_segments(xhci, prev, next, link_trbs, isoc);
 191
 192                 prev = next;
 193                 num_segs--;
 194         }
 195         xhci_link_segments(xhci, prev, ring->first_seg, link_trbs, isoc);
 196
 197         if (link_trbs) {
 198                 /* See section 4.9.2.1 and 6.4.4.1 */
 199                 prev->trbs[TRBS_PER_SEGMENT-1].link.control |=
 200                         cpu_to_le32(LINK_TOGGLE);
 201                 xhci_dbg(xhci, "Wrote link toggle flag to"
 202                                 " segment %p (virtual), 0x%llx (DMA)\n",
 203                                 prev, (unsigned long long)prev->dma);
 204         }
 205         xhci_initialize_ring_info(ring);
 206         return ring;
 207
 208 fail:
 209         kfree(ring);
 210         return NULL;
 211 }
 212
 213 void xhci_free_or_cache_endpoint_ring(struct xhci_hcd *xhci,
 214                 struct xhci_virt_device *virt_dev,
 215                 unsigned int ep_index)
 216 {
 217         int rings_cached;
 218
 219         rings_cached = virt_dev->num_rings_cached;
 220         if (rings_cached < XHCI_MAX_RINGS_CACHED) {
 221                 virt_dev->ring_cache[rings_cached] =
 222                         virt_dev->eps[ep_index].ring;
 223                 virt_dev->num_rings_cached++;
 224                 xhci_dbg(xhci, "Cached old ring, "
 225                                 "%d ring%s cached\n",
 226                                 virt_dev->num_rings_cached,
 227                                 (virt_dev->num_rings_cached > 1) ? "s" : "");
 228         } else {
 229                 xhci_ring_free(xhci, virt_dev->eps[ep_index].ring);
 230                 xhci_dbg(xhci, "Ring cache full (%d rings), "
 231                                 "freeing ring\n",
 232                                 virt_dev->num_rings_cached);
 233         }
 234         virt_dev->eps[ep_index].ring = NULL;
 235 }
 236
 237 /* Zero an endpoint ring (except for link TRBs) and move the enqueue and dequeue
 238  * pointers to the beginning of the ring.
 239  */
 240 static void xhci_reinit_cached_ring(struct xhci_hcd *xhci,
 241                 struct xhci_ring *ring, bool isoc)
 242 {
 243         struct xhci_segment     *seg = ring->first_seg;
 244         do {
 245                 memset(seg->trbs, 0,
 246                                 sizeof(union xhci_trb)*TRBS_PER_SEGMENT);
 247                 /* All endpoint rings have link TRBs */
 248                 xhci_link_segments(xhci, seg, seg->next, 1, isoc);
 249                 seg = seg->next;
 250         } while (seg != ring->first_seg);
 251         xhci_initialize_ring_info(ring);
 252         /* td list should be empty since all URBs have been cancelled,
 253          * but just in case...
 254          */
 255         INIT_LIST_HEAD(&ring->td_list);
 256 }
 257
 258 #define CTX_SIZE(_hcc) (HCC_64BYTE_CONTEXT(_hcc) ? 64 : 32)
 259
 260 static struct xhci_container_ctx *xhci_alloc_container_ctx(struct xhci_hcd *xhci,
 261                                                     int type, gfp_t flags)
 262 {
 263         struct xhci_container_ctx *ctx = kzalloc(sizeof(*ctx), flags);
 264         if (!ctx)
 265                 return NULL;
 266
 267         BUG_ON((type != XHCI_CTX_TYPE_DEVICE) && (type != XHCI_CTX_TYPE_INPUT));
 268         ctx->type = type;
 269         ctx->size = HCC_64BYTE_CONTEXT(xhci->hcc_params) ? 2048 : 1024;
 270         if (type == XHCI_CTX_TYPE_INPUT)
 271                 ctx->size += CTX_SIZE(xhci->hcc_params);
 272
 273         ctx->bytes = dma_pool_alloc(xhci->device_pool, flags, &ctx->dma);
 274         if (!ctx->bytes) {
 275                 kfree(ctx);
 276                 return NULL;
 277         }
 278         memset(ctx->bytes, 0, ctx->size);
 279         return ctx;
 280 }
 281
 282 static void xhci_free_container_ctx(struct xhci_hcd *xhci,
 283                              struct xhci_container_ctx *ctx)
 284 {
 285         if (!ctx)
 286                 return;
 287         dma_pool_free(xhci->device_pool, ctx->bytes, ctx->dma);
 288         kfree(ctx);
 289 }
 290
 291 struct xhci_input_control_ctx *xhci_get_input_control_ctx(struct xhci_hcd *xhci,
 292                                               struct xhci_container_ctx *ctx)
 293 {
 294         BUG_ON(ctx->type != XHCI_CTX_TYPE_INPUT);
 295         return (struct xhci_input_control_ctx *)ctx->bytes;
 296 }
 297
 298 struct xhci_slot_ctx *xhci_get_slot_ctx(struct xhci_hcd *xhci,
 299                                         struct xhci_container_ctx *ctx)
 300 {
 301         if (ctx->type == XHCI_CTX_TYPE_DEVICE)
 302                 return (struct xhci_slot_ctx *)ctx->bytes;
 303
 304         return (struct xhci_slot_ctx *)
 305                 (ctx->bytes + CTX_SIZE(xhci->hcc_params));
 306 }
 307
 308 struct xhci_ep_ctx *xhci_get_ep_ctx(struct xhci_hcd *xhci,
 309                                     struct xhci_container_ctx *ctx,
 310                                     unsigned int ep_index)
 311 {
 312         /* increment ep index by offset of start of ep ctx array */
 313         ep_index++;
 314         if (ctx->type == XHCI_CTX_TYPE_INPUT)
 315                 ep_index++;
 316
 317         return (struct xhci_ep_ctx *)
 318                 (ctx->bytes + (ep_index * CTX_SIZE(xhci->hcc_params)));
 319 }
 320
 321
 322 /***************** Streams structures manipulation *************************/
 323
 324 static void xhci_free_stream_ctx(struct xhci_hcd *xhci,
 325                 unsigned int num_stream_ctxs,
 326                 struct xhci_stream_ctx *stream_ctx, dma_addr_t dma)
 327 {
 328         struct pci_dev *pdev = to_pci_dev(xhci_to_hcd(xhci)->self.controller);
 329
 330         if (num_stream_ctxs > MEDIUM_STREAM_ARRAY_SIZE)
 331                 dma_free_coherent(&pdev->dev,
 332                                 sizeof(struct xhci_stream_ctx)*num_stream_ctxs,
 333                                 stream_ctx, dma);
 334         else if (num_stream_ctxs <= SMALL_STREAM_ARRAY_SIZE)
 335                 return dma_pool_free(xhci->small_streams_pool,
 336                                 stream_ctx, dma);
 337         else
 338                 return dma_pool_free(xhci->medium_streams_pool,
 339                                 stream_ctx, dma);
 340 }
 341
 342 /*
 343  * The stream context array for each endpoint with bulk streams enabled can
 344  * vary in size, based on:
 345  *  - how many streams the endpoint supports,
 346  *  - the maximum primary stream array size the host controller supports,
 347  *  - and how many streams the device driver asks for.
 348  *
 349  * The stream context array must be a power of 2, and can be as small as
 350  * 64 bytes or as large as 1MB.
 351  */
 352 static struct xhci_stream_ctx *xhci_alloc_stream_ctx(struct xhci_hcd *xhci,
 353                 unsigned int num_stream_ctxs, dma_addr_t *dma,
 354                 gfp_t mem_flags)
 355 {
 356         struct pci_dev *pdev = to_pci_dev(xhci_to_hcd(xhci)->self.controller);
 357
 358         if (num_stream_ctxs > MEDIUM_STREAM_ARRAY_SIZE)
 359                 return dma_alloc_coherent(&pdev->dev,
 360                                 sizeof(struct xhci_stream_ctx)*num_stream_ctxs,
 361                                 dma, mem_flags);
 362         else if (num_stream_ctxs <= SMALL_STREAM_ARRAY_SIZE)
 363                 return dma_pool_alloc(xhci->small_streams_pool,
 364                                 mem_flags, dma);
 365         else
 366                 return dma_pool_alloc(xhci->medium_streams_pool,
 367                                 mem_flags, dma);
 368 }
 369
 370 struct xhci_ring *xhci_dma_to_transfer_ring(
 371                 struct xhci_virt_ep *ep,
 372                 u64 address)
 373 {
 374         if (ep->ep_state & EP_HAS_STREAMS)
 375                 return radix_tree_lookup(&ep->stream_info->trb_address_map,
 376                                 address >> SEGMENT_SHIFT);
 377         return ep->ring;
 378 }
 379
 380 /* Only use this when you know stream_info is valid */
 381 #ifdef CONFIG_USB_XHCI_HCD_DEBUGGING
 382 static struct xhci_ring *dma_to_stream_ring(
 383                 struct xhci_stream_info *stream_info,
 384                 u64 address)
 385 {
 386         return radix_tree_lookup(&stream_info->trb_address_map,
 387                         address >> SEGMENT_SHIFT);
 388 }
 389 #endif  /* CONFIG_USB_XHCI_HCD_DEBUGGING */
 390
 391 struct xhci_ring *xhci_stream_id_to_ring(
 392                 struct xhci_virt_device *dev,
 393                 unsigned int ep_index,
 394                 unsigned int stream_id)
 395 {
 396         struct xhci_virt_ep *ep = &dev->eps[ep_index];
 397
 398         if (stream_id == 0)
 399                 return ep->ring;
 400         if (!ep->stream_info)
 401                 return NULL;
 402
 403         if (stream_id > ep->stream_info->num_streams)
 404                 return NULL;
 405         return ep->stream_info->stream_rings[stream_id];
 406 }
 407
 408 #ifdef CONFIG_USB_XHCI_HCD_DEBUGGING
 409 static int xhci_test_radix_tree(struct xhci_hcd *xhci,
 410                 unsigned int num_streams,
 411                 struct xhci_stream_info *stream_info)
 412 {
 413         u32 cur_stream;
 414         struct xhci_ring *cur_ring;
 415         u64 addr;
 416
 417         for (cur_stream = 1; cur_stream < num_streams; cur_stream++) {
 418                 struct xhci_ring *mapped_ring;
 419                 int trb_size = sizeof(union xhci_trb);
 420
 421                 cur_ring = stream_info->stream_rings[cur_stream];
 422                 for (addr = cur_ring->first_seg->dma;
 423                                 addr < cur_ring->first_seg->dma + SEGMENT_SIZE;
 424                                 addr += trb_size) {
 425                         mapped_ring = dma_to_stream_ring(stream_info, addr);
 426                         if (cur_ring != mapped_ring) {
 427                                 xhci_warn(xhci, "WARN: DMA address 0x%08llx "
 428                                                 "didn't map to stream ID %u; "
 429                                                 "mapped to ring %p\n",
 430                                                 (unsigned long long) addr,
 431                                                 cur_stream,
 432                                                 mapped_ring);
 433                                 return -EINVAL;
 434                         }
 435                 }
 436                 /* One TRB after the end of the ring segment shouldn't return a
 437                  * pointer to the current ring (although it may be a part of a
 438                  * different ring).
 439                  */
 440                 mapped_ring = dma_to_stream_ring(stream_info, addr);
 441                 if (mapped_ring != cur_ring) {
 442                         /* One TRB before should also fail */
 443                         addr = cur_ring->first_seg->dma - trb_size;
 444                         mapped_ring = dma_to_stream_ring(stream_info, addr);
 445                 }
 446                 if (mapped_ring == cur_ring) {
 447                         xhci_warn(xhci, "WARN: Bad DMA address 0x%08llx "
 448                                         "mapped to valid stream ID %u; "
 449                                         "mapped ring = %p\n",
 450                                         (unsigned long long) addr,
 451                                         cur_stream,
 452                                         mapped_ring);
 453                         return -EINVAL;
 454                 }
 455         }
 456         return 0;
 457 }
 458 #endif  /* CONFIG_USB_XHCI_HCD_DEBUGGING */
 459
 460 /*
 461  * Change an endpoint's internal structure so it supports stream IDs.  The
 462  * number of requested streams includes stream 0, which cannot be used by device
 463  * drivers.
 464  *
 465  * The number of stream contexts in the stream context array may be bigger than
 466  * the number of streams the driver wants to use.  This is because the number of
 467  * stream context array entries must be a power of two.
 468  *
 469  * We need a radix tree for mapping physical addresses of TRBs to which stream
 470  * ID they belong to.  We need to do this because the host controller won't tell
 471  * us which stream ring the TRB came from.  We could store the stream ID in an
 472  * event data TRB, but that doesn't help us for the cancellation case, since the
 473  * endpoint may stop before it reaches that event data TRB.
 474  *
 475  * The radix tree maps the upper portion of the TRB DMA address to a ring
 476  * segment that has the same upper portion of DMA addresses.  For example, say I
 477  * have segments of size 1KB, that are always 64-byte aligned.  A segment may
 478  * start at 0x10c91000 and end at 0x10c913f0.  If I use the upper 10 bits, the
 479  * key to the stream ID is 0x43244.  I can use the DMA address of the TRB to
 480  * pass the radix tree a key to get the right stream ID:
 481  *
 482  *      0x10c90fff >> 10 = 0x43243
 483  *      0x10c912c0 >> 10 = 0x43244
 484  *      0x10c91400 >> 10 = 0x43245
 485  *
 486  * Obviously, only those TRBs with DMA addresses that are within the segment
 487  * will make the radix tree return the stream ID for that ring.
 488  *
 489  * Caveats for the radix tree:
 490  *
 491  * The radix tree uses an unsigned long as a key pair.  On 32-bit systems, an
 492  * unsigned long will be 32-bits; on a 64-bit system an unsigned long will be
 493  * 64-bits.  Since we only request 32-bit DMA addresses, we can use that as the
 494  * key on 32-bit or 64-bit systems (it would also be fine if we asked for 64-bit
 495  * PCI DMA addresses on a 64-bit system).  There might be a problem on 32-bit
 496  * extended systems (where the DMA address can be bigger than 32-bits),
 497  * if we allow the PCI dma mask to be bigger than 32-bits.  So don't do that.
 498  */
 499 struct xhci_stream_info *xhci_alloc_stream_info(struct xhci_hcd *xhci,
 500                 unsigned int num_stream_ctxs,
 501                 unsigned int num_streams, gfp_t mem_flags)
 502 {
 503         struct xhci_stream_info *stream_info;
 504         u32 cur_stream;
 505         struct xhci_ring *cur_ring;
 506         unsigned long key;
 507         u64 addr;
 508         int ret;
 509
 510         xhci_dbg(xhci, "Allocating %u streams and %u "
 511                         "stream context array entries.\n",
 512                         num_streams, num_stream_ctxs);
 513         if (xhci->cmd_ring_reserved_trbs == MAX_RSVD_CMD_TRBS) {
 514                 xhci_dbg(xhci, "Command ring has no reserved TRBs available\n");
 515                 return NULL;
 516         }
 517         xhci->cmd_ring_reserved_trbs++;
 518
 519         stream_info = kzalloc(sizeof(struct xhci_stream_info), mem_flags);
 520         if (!stream_info)
 521                 goto cleanup_trbs;
 522
 523         stream_info->num_streams = num_streams;
 524         stream_info->num_stream_ctxs = num_stream_ctxs;
 525
 526         /* Initialize the array of virtual pointers to stream rings. */
 527         stream_info->stream_rings = kzalloc(
 528                         sizeof(struct xhci_ring *)*num_streams,
 529                         mem_flags);
 530         if (!stream_info->stream_rings)
 531                 goto cleanup_info;
 532
 533         /* Initialize the array of DMA addresses for stream rings for the HW. */
 534         stream_info->stream_ctx_array = xhci_alloc_stream_ctx(xhci,
 535                         num_stream_ctxs, &stream_info->ctx_array_dma,
 536                         mem_flags);
 537         if (!stream_info->stream_ctx_array)
 538                 goto cleanup_ctx;
 539         memset(stream_info->stream_ctx_array, 0,
 540                         sizeof(struct xhci_stream_ctx)*num_stream_ctxs);
 541
 542         /* Allocate everything needed to free the stream rings later */
 543         stream_info->free_streams_command =
 544                 xhci_alloc_command(xhci, true, true, mem_flags);
 545         if (!stream_info->free_streams_command)
 546                 goto cleanup_ctx;
 547
 548         INIT_RADIX_TREE(&stream_info->trb_address_map, GFP_ATOMIC);
 549
 550         /* Allocate rings for all the streams that the driver will use,
 551          * and add their segment DMA addresses to the radix tree.
 552          * Stream 0 is reserved.
 553          */
 554         for (cur_stream = 1; cur_stream < num_streams; cur_stream++) {
 555                 stream_info->stream_rings[cur_stream] =
 556                         xhci_ring_alloc(xhci, 1, true, false, mem_flags);
 557                 cur_ring = stream_info->stream_rings[cur_stream];
 558                 if (!cur_ring)
 559                         goto cleanup_rings;
 560                 cur_ring->stream_id = cur_stream;
 561                 /* Set deq ptr, cycle bit, and stream context type */
 562                 addr = cur_ring->first_seg->dma |
 563                         SCT_FOR_CTX(SCT_PRI_TR) |
 564                         cur_ring->cycle_state;
 565                 stream_info->stream_ctx_array[cur_stream].stream_ring =
 566                         cpu_to_le64(addr);
 567                 xhci_dbg(xhci, "Setting stream %d ring ptr to 0x%08llx\n",
 568                                 cur_stream, (unsigned long long) addr);
 569
 570                 key = (unsigned long)
 571                         (cur_ring->first_seg->dma >> SEGMENT_SHIFT);
 572                 ret = radix_tree_insert(&stream_info->trb_address_map,
 573                                 key, cur_ring);
 574                 if (ret) {
 575                         xhci_ring_free(xhci, cur_ring);
 576                         stream_info->stream_rings[cur_stream] = NULL;
 577                         goto cleanup_rings;
 578                 }
 579         }
 580         /* Leave the other unused stream ring pointers in the stream context
 581          * array initialized to zero.  This will cause the xHC to give us an
 582          * error if the device asks for a stream ID we don't have setup (if it
 583          * was any other way, the host controller would assume the ring is
 584          * "empty" and wait forever for data to be queued to that stream ID).
 585          */
 586 #if XHCI_DEBUG
 587         /* Do a little test on the radix tree to make sure it returns the
 588          * correct values.
 589          */
 590         if (xhci_test_radix_tree(xhci, num_streams, stream_info))
 591                 goto cleanup_rings;
 592 #endif
 593
 594         return stream_info;
 595
 596 cleanup_rings:
 597         for (cur_stream = 1; cur_stream < num_streams; cur_stream++) {
 598                 cur_ring = stream_info->stream_rings[cur_stream];
 599                 if (cur_ring) {
 600                         addr = cur_ring->first_seg->dma;
 601                         radix_tree_delete(&stream_info->trb_address_map,
 602                                         addr >> SEGMENT_SHIFT);
 603                         xhci_ring_free(xhci, cur_ring);
 604                         stream_info->stream_rings[cur_stream] = NULL;
 605                 }
 606         }
 607         xhci_free_command(xhci, stream_info->free_streams_command);
 608 cleanup_ctx:
 609         kfree(stream_info->stream_rings);
 610 cleanup_info:
 611         kfree(stream_info);
 612 cleanup_trbs:
 613         xhci->cmd_ring_reserved_trbs--;
 614         return NULL;
 615 }
 616 /*
 617  * Sets the MaxPStreams field and the Linear Stream Array field.
 618  * Sets the dequeue pointer to the stream context array.
 619  */
 620 void xhci_setup_streams_ep_input_ctx(struct xhci_hcd *xhci,
 621                 struct xhci_ep_ctx *ep_ctx,
 622                 struct xhci_stream_info *stream_info)
 623 {
 624         u32 max_primary_streams;
 625         /* MaxPStreams is the number of stream context array entries, not the
 626          * number we're actually using.  Must be in 2^(MaxPstreams + 1) format.
 627          * fls(0) = 0, fls(0x1) = 1, fls(0x10) = 2, fls(0x100) = 3, etc.
 628          */
 629         max_primary_streams = fls(stream_info->num_stream_ctxs) - 2;
 630         xhci_dbg(xhci, "Setting number of stream ctx array entries to %u\n",
 631                         1 << (max_primary_streams + 1));
 632         ep_ctx->ep_info &= cpu_to_le32(~EP_MAXPSTREAMS_MASK);
 633         ep_ctx->ep_info |= cpu_to_le32(EP_MAXPSTREAMS(max_primary_streams)
 634                                        | EP_HAS_LSA);
 635         ep_ctx->deq  = cpu_to_le64(stream_info->ctx_array_dma);
 636 }
 637
 638 /*
 639  * Sets the MaxPStreams field and the Linear Stream Array field to 0.
 640  * Reinstalls the "normal" endpoint ring (at its previous dequeue mark,
 641  * not at the beginning of the ring).
 642  */
 643 void xhci_setup_no_streams_ep_input_ctx(struct xhci_hcd *xhci,
 644                 struct xhci_ep_ctx *ep_ctx,
 645                 struct xhci_virt_ep *ep)
 646 {
 647         dma_addr_t addr;
 648         ep_ctx->ep_info &= cpu_to_le32(~(EP_MAXPSTREAMS_MASK | EP_HAS_LSA));
 649         addr = xhci_trb_virt_to_dma(ep->ring->deq_seg, ep->ring->dequeue);
 650         ep_ctx->deq  = cpu_to_le64(addr | ep->ring->cycle_state);
 651 }
 652
 653 /* Frees all stream contexts associated with the endpoint,
 654  *
 655  * Caller should fix the endpoint context streams fields.
 656  */
 657 void xhci_free_stream_info(struct xhci_hcd *xhci,
 658                 struct xhci_stream_info *stream_info)
 659 {
 660         int cur_stream;
 661         struct xhci_ring *cur_ring;
 662         dma_addr_t addr;
 663
 664         if (!stream_info)
 665                 return;
 666
 667         for (cur_stream = 1; cur_stream < stream_info->num_streams;
 668                         cur_stream++) {
 669                 cur_ring = stream_info->stream_rings[cur_stream];
 670                 if (cur_ring) {
 671                         addr = cur_ring->first_seg->dma;
 672                         radix_tree_delete(&stream_info->trb_address_map,
 673                                         addr >> SEGMENT_SHIFT);
 674                         xhci_ring_free(xhci, cur_ring);
 675                         stream_info->stream_rings[cur_stream] = NULL;
 676                 }
 677         }
 678         xhci_free_command(xhci, stream_info->free_streams_command);
 679         xhci->cmd_ring_reserved_trbs--;
 680         if (stream_info->stream_ctx_array)
 681                 xhci_free_stream_ctx(xhci,
 682                                 stream_info->num_stream_ctxs,
 683                                 stream_info->stream_ctx_array,
 684                                 stream_info->ctx_array_dma);
 685
 686         if (stream_info)
 687                 kfree(stream_info->stream_rings);
 688         kfree(stream_info);
 689 }
 690
 691
 692 /***************** Device context manipulation *************************/
 693
 694 static void xhci_init_endpoint_timer(struct xhci_hcd *xhci,
 695                 struct xhci_virt_ep *ep)
 696 {
 697         init_timer(&ep->stop_cmd_timer);
 698         ep->stop_cmd_timer.data = (unsigned long) ep;
 699         ep->stop_cmd_timer.function = xhci_stop_endpoint_command_watchdog;
 700         ep->xhci = xhci;
 701 }
 702
 703 static void xhci_free_tt_info(struct xhci_hcd *xhci,
 704                 struct xhci_virt_device *virt_dev,
 705                 int slot_id)
 706 {
 707         struct list_head *tt_list_head;
 708         struct xhci_tt_bw_info *tt_info, *next;
 709         bool slot_found = false;
 710
 711         /* If the device never made it past the Set Address stage,
 712          * it may not have the real_port set correctly.
 713          */
 714         if (virt_dev->real_port == 0 ||
 715                         virt_dev->real_port > HCS_MAX_PORTS(xhci->hcs_params1)) {
 716                 xhci_dbg(xhci, "Bad real port.\n");
 717                 return;
 718         }
 719
 720         tt_list_head = &(xhci->rh_bw[virt_dev->real_port - 1].tts);
 721         list_for_each_entry_safe(tt_info, next, tt_list_head, tt_list) {
 722                 /* Multi-TT hubs will have more than one entry */
 723                 if (tt_info->slot_id == slot_id) {
 724                         slot_found = true;
 725                         list_del(&tt_info->tt_list);
 726                         kfree(tt_info);
 727                 } else if (slot_found) {
 728                         break;
 729                 }
 730         }
 731 }
 732
 733 int xhci_alloc_tt_info(struct xhci_hcd *xhci,
 734                 struct xhci_virt_device *virt_dev,
 735                 struct usb_device *hdev,
 736                 struct usb_tt *tt, gfp_t mem_flags)
 737 {
 738         struct xhci_tt_bw_info          *tt_info;
 739         unsigned int                    num_ports;
 740         int                             i, j;
 741
 742         if (!tt->multi)
 743                 num_ports = 1;
 744         else
 745                 num_ports = hdev->maxchild;
 746
 747         for (i = 0; i < num_ports; i++, tt_info++) {
 748                 struct xhci_interval_bw_table *bw_table;
 749
 750                 tt_info = kzalloc(sizeof(*tt_info), mem_flags);
 751                 if (!tt_info)
 752                         goto free_tts;
 753                 INIT_LIST_HEAD(&tt_info->tt_list);
 754                 list_add(&tt_info->tt_list,
 755                                 &xhci->rh_bw[virt_dev->real_port - 1].tts);
 756                 tt_info->slot_id = virt_dev->udev->slot_id;
 757                 if (tt->multi)
 758                         tt_info->ttport = i+1;
 759                 bw_table = &tt_info->bw_table;
 760                 for (j = 0; j < XHCI_MAX_INTERVAL; j++)
 761                         INIT_LIST_HEAD(&bw_table->interval_bw[j].endpoints);
 762         }
 763         return 0;
 764
 765 free_tts:
 766         xhci_free_tt_info(xhci, virt_dev, virt_dev->udev->slot_id);
 767         return -ENOMEM;
 768 }
 769
 770
 771 /* All the xhci_tds in the ring's TD list should be freed at this point.
 772  * Should be called with xhci->lock held if there is any chance the TT lists
 773  * will be manipulated by the configure endpoint, allocate device, or update
 774  * hub functions while this function is removing the TT entries from the list.
 775  */
 776 void xhci_free_virt_device(struct xhci_hcd *xhci, int slot_id)
 777 {
 778         struct xhci_virt_device *dev;
 779         int i;
 780         int old_active_eps = 0;
 781
 782         /* Slot ID 0 is reserved */
 783         if (slot_id == 0 || !xhci->devs[slot_id])
 784                 return;
 785
 786         dev = xhci->devs[slot_id];
 787         xhci->dcbaa->dev_context_ptrs[slot_id] = 0;
 788         if (!dev)
 789                 return;
 790
 791         if (dev->tt_info)
 792                 old_active_eps = dev->tt_info->active_eps;
 793
 794         for (i = 0; i < 31; ++i) {
 795                 if (dev->eps[i].ring)
 796                         xhci_ring_free(xhci, dev->eps[i].ring);
 797                 if (dev->eps[i].stream_info)
 798                         xhci_free_stream_info(xhci,
 799                                         dev->eps[i].stream_info);
 800                 /* Endpoints on the TT/root port lists should have been removed
 801                  * when usb_disable_device() was called for the device.
 802                  * We can't drop them anyway, because the udev might have gone
 803                  * away by this point, and we can't tell what speed it was.
 804                  */
 805                 if (!list_empty(&dev->eps[i].bw_endpoint_list))
 806                         xhci_warn(xhci, "Slot %u endpoint %u "
 807                                         "not removed from BW list!\n",
 808                                         slot_id, i);
 809         }
 810         /* If this is a hub, free the TT(s) from the TT list */
 811         xhci_free_tt_info(xhci, dev, slot_id);
 812         /* If necessary, update the number of active TTs on this root port */
 813         xhci_update_tt_active_eps(xhci, dev, old_active_eps);
 814
 815         if (dev->ring_cache) {
 816                 for (i = 0; i < dev->num_rings_cached; i++)
 817                         xhci_ring_free(xhci, dev->ring_cache[i]);
 818                 kfree(dev->ring_cache);
 819         }
 820
 821         if (dev->in_ctx)
 822                 xhci_free_container_ctx(xhci, dev->in_ctx);
 823         if (dev->out_ctx)
 824                 xhci_free_container_ctx(xhci, dev->out_ctx);
 825
 826         kfree(xhci->devs[slot_id]);
 827         xhci->devs[slot_id] = NULL;
 828 }
 829
 830 /*
 831  * Free a virt_device structure.
 832  * If the virt_device added a tt_info (a hub) and has children pointing to
 833  * that tt_info, then free the child first. Recursive.
 834  * We can't rely on udev at this point to find child-parent relationships.
 835  */
 836 void xhci_free_virt_devices_depth_first(struct xhci_hcd *xhci, int slot_id)
 837 {
 838         struct xhci_virt_device *vdev;
 839         struct list_head *tt_list_head;
 840         struct xhci_tt_bw_info *tt_info, *next;
 841         int i;
 842
 843         vdev = xhci->devs[slot_id];
 844         if (!vdev)
 845                 return;
 846
 847         tt_list_head = &(xhci->rh_bw[vdev->real_port - 1].tts);
 848         list_for_each_entry_safe(tt_info, next, tt_list_head, tt_list) {
 849                 /* is this a hub device that added a tt_info to the tts list */
 850                 if (tt_info->slot_id == slot_id) {
 851                         /* are any devices using this tt_info? */
 852                         for (i = 1; i < HCS_MAX_SLOTS(xhci->hcs_params1); i++) {
 853                                 vdev = xhci->devs[i];
 854                                 if (vdev && (vdev->tt_info == tt_info))
 855                                         xhci_free_virt_devices_depth_first(
 856                                                 xhci, i);
 857                         }
 858                 }
 859         }
 860         /* we are now at a leaf device */
 861         xhci_free_virt_device(xhci, slot_id);
 862 }
 863
 864 int xhci_alloc_virt_device(struct xhci_hcd *xhci, int slot_id,
 865                 struct usb_device *udev, gfp_t flags)
 866 {
 867         struct xhci_virt_device *dev;
 868         int i;
 869
 870         /* Slot ID 0 is reserved */
 871         if (slot_id == 0 || xhci->devs[slot_id]) {
 872                 xhci_warn(xhci, "Bad Slot ID %d\n", slot_id);
 873                 return 0;
 874         }
 875
 876         xhci->devs[slot_id] = kzalloc(sizeof(*xhci->devs[slot_id]), flags);
 877         if (!xhci->devs[slot_id])
 878                 return 0;
 879         dev = xhci->devs[slot_id];
 880
 881         /* Allocate the (output) device context that will be used in the HC. */
 882         dev->out_ctx = xhci_alloc_container_ctx(xhci, XHCI_CTX_TYPE_DEVICE, flags);
 883         if (!dev->out_ctx)
 884                 goto fail;
 885
 886         xhci_dbg(xhci, "Slot %d output ctx = 0x%llx (dma)\n", slot_id,
 887                         (unsigned long long)dev->out_ctx->dma);
 888
 889         /* Allocate the (input) device context for address device command */
 890         dev->in_ctx = xhci_alloc_container_ctx(xhci, XHCI_CTX_TYPE_INPUT, flags);
 891         if (!dev->in_ctx)
 892                 goto fail;
 893
 894         xhci_dbg(xhci, "Slot %d input ctx = 0x%llx (dma)\n", slot_id,
 895                         (unsigned long long)dev->in_ctx->dma);
 896
 897         /* Initialize the cancellation list and watchdog timers for each ep */
 898         for (i = 0; i < 31; i++) {
 899                 xhci_init_endpoint_timer(xhci, &dev->eps[i]);
 900                 INIT_LIST_HEAD(&dev->eps[i].cancelled_td_list);
 901                 INIT_LIST_HEAD(&dev->eps[i].bw_endpoint_list);
 902         }
 903
 904         /* Allocate endpoint 0 ring */
 905         dev->eps[0].ring = xhci_ring_alloc(xhci, 1, true, false, flags);
 906         if (!dev->eps[0].ring)
 907                 goto fail;
 908
 909         /* Allocate pointers to the ring cache */
 910         dev->ring_cache = kzalloc(
 911                         sizeof(struct xhci_ring *)*XHCI_MAX_RINGS_CACHED,
 912                         flags);
 913         if (!dev->ring_cache)
 914                 goto fail;
 915         dev->num_rings_cached = 0;
 916
 917         init_completion(&dev->cmd_completion);
 918         INIT_LIST_HEAD(&dev->cmd_list);
 919         dev->udev = udev;
 920
 921         /* Point to output device context in dcbaa. */
 922         xhci->dcbaa->dev_context_ptrs[slot_id] = cpu_to_le64(dev->out_ctx->dma);
 923         xhci_dbg(xhci, "Set slot id %d dcbaa entry %p to 0x%llx\n",
 924                  slot_id,
 925                  &xhci->dcbaa->dev_context_ptrs[slot_id],
 926                  le64_to_cpu(xhci->dcbaa->dev_context_ptrs[slot_id]));
 927
 928         return 1;
 929 fail:
 930         xhci_free_virt_device(xhci, slot_id);
 931         return 0;
 932 }
 933
 934 void xhci_copy_ep0_dequeue_into_input_ctx(struct xhci_hcd *xhci,
 935                 struct usb_device *udev)
 936 {
 937         struct xhci_virt_device *virt_dev;
 938         struct xhci_ep_ctx      *ep0_ctx;
 939         struct xhci_ring        *ep_ring;
 940
 941         virt_dev = xhci->devs[udev->slot_id];
 942         ep0_ctx = xhci_get_ep_ctx(xhci, virt_dev->in_ctx, 0);
 943         ep_ring = virt_dev->eps[0].ring;
 944         /*
 945          * FIXME we don't keep track of the dequeue pointer very well after a
 946          * Set TR dequeue pointer, so we're setting the dequeue pointer of the
 947          * host to our enqueue pointer.  This should only be called after a
 948          * configured device has reset, so all control transfers should have
 949          * been completed or cancelled before the reset.
 950          */
 951         ep0_ctx->deq = cpu_to_le64(xhci_trb_virt_to_dma(ep_ring->enq_seg,
 952                                                         ep_ring->enqueue)
 953                                    | ep_ring->cycle_state);
 954 }
 955
 956 /*
 957  * The xHCI roothub may have ports of differing speeds in any order in the port
 958  * status registers.  xhci->port_array provides an array of the port speed for
 959  * each offset into the port status registers.
 960  *
 961  * The xHCI hardware wants to know the roothub port number that the USB device
 962  * is attached to (or the roothub port its ancestor hub is attached to).  All we
 963  * know is the index of that port under either the USB 2.0 or the USB 3.0
 964  * roothub, but that doesn't give us the real index into the HW port status
 965  * registers.  Scan through the xHCI roothub port array, looking for the Nth
 966  * entry of the correct port speed.  Return the port number of that entry.
 967  */
 968 static u32 xhci_find_real_port_number(struct xhci_hcd *xhci,
 969                 struct usb_device *udev)
 970 {
 971         struct usb_device *top_dev;
 972         unsigned int num_similar_speed_ports;
 973         unsigned int faked_port_num;
 974         int i;
 975
 976         for (top_dev = udev; top_dev->parent && top_dev->parent->parent;
 977                         top_dev = top_dev->parent)
 978                 /* Found device below root hub */;
 979         faked_port_num = top_dev->portnum;
 980         for (i = 0, num_similar_speed_ports = 0;
 981                         i < HCS_MAX_PORTS(xhci->hcs_params1); i++) {
 982                 u8 port_speed = xhci->port_array[i];
 983
 984                 /*
 985                  * Skip ports that don't have known speeds, or have duplicate
 986                  * Extended Capabilities port speed entries.
 987                  */
 988                 if (port_speed == 0 || port_speed == DUPLICATE_ENTRY)
 989                         continue;
 990
 991                 /*
 992                  * USB 3.0 ports are always under a USB 3.0 hub.  USB 2.0 and
 993                  * 1.1 ports are under the USB 2.0 hub.  If the port speed
 994                  * matches the device speed, it's a similar speed port.
 995                  */
 996                 if ((port_speed == 0x03) == (udev->speed == USB_SPEED_SUPER))
 997                         num_similar_speed_ports++;
 998                 if (num_similar_speed_ports == faked_port_num)
 999                         /* Roothub ports are numbered from 1 to N */
1000                         return i+1;
1001         }
1002         return 0;
1003 }
1004
1005 /* Setup an xHCI virtual device for a Set Address command */
1006 int xhci_setup_addressable_virt_dev(struct xhci_hcd *xhci, struct usb_device *udev)
1007 {
1008         struct xhci_virt_device *dev;
1009         struct xhci_ep_ctx      *ep0_ctx;
1010         struct xhci_slot_ctx    *slot_ctx;
1011         u32                     port_num;
1012         struct usb_device *top_dev;
1013
1014         dev = xhci->devs[udev->slot_id];
1015         /* Slot ID 0 is reserved */
1016         if (udev->slot_id == 0 || !dev) {
1017                 xhci_warn(xhci, "Slot ID %d is not assigned to this device\n",
1018                                 udev->slot_id);
1019                 return -EINVAL;
1020         }
1021         ep0_ctx = xhci_get_ep_ctx(xhci, dev->in_ctx, 0);
1022         slot_ctx = xhci_get_slot_ctx(xhci, dev->in_ctx);
1023
1024         /* 3) Only the control endpoint is valid - one endpoint context */
1025         slot_ctx->dev_info |= cpu_to_le32(LAST_CTX(1) | udev->route);
1026         switch (udev->speed) {
1027         case USB_SPEED_SUPER:
1028                 slot_ctx->dev_info |= cpu_to_le32(SLOT_SPEED_SS);
1029                 break;
1030         case USB_SPEED_HIGH:
1031                 slot_ctx->dev_info |= cpu_to_le32(SLOT_SPEED_HS);
1032                 break;
1033         case USB_SPEED_FULL:
1034                 slot_ctx->dev_info |= cpu_to_le32(SLOT_SPEED_FS);
1035                 break;
1036         case USB_SPEED_LOW:
1037                 slot_ctx->dev_info |= cpu_to_le32(SLOT_SPEED_LS);
1038                 break;
1039         case USB_SPEED_WIRELESS:
1040                 xhci_dbg(xhci, "FIXME xHCI doesn't support wireless speeds\n");
1041                 return -EINVAL;
1042                 break;
1043         default:
1044                 /* Speed was set earlier, this shouldn't happen. */
1045                 BUG();
1046         }
1047         /* Find the root hub port this device is under */
1048         port_num = xhci_find_real_port_number(xhci, udev);
1049         if (!port_num)
1050                 return -EINVAL;
1051         slot_ctx->dev_info2 |= cpu_to_le32(ROOT_HUB_PORT(port_num));
1052         /* Set the port number in the virtual_device to the faked port number */
1053         for (top_dev = udev; top_dev->parent && top_dev->parent->parent;
1054                         top_dev = top_dev->parent)
1055                 /* Found device below root hub */;
1056         dev->fake_port = top_dev->portnum;
1057         dev->real_port = port_num;
1058         xhci_dbg(xhci, "Set root hub portnum to %d\n", port_num);
1059         xhci_dbg(xhci, "Set fake root hub portnum to %d\n", dev->fake_port);
1060
1061         /* Find the right bandwidth table that this device will be a part of.
1062          * If this is a full speed device attached directly to a root port (or a
1063          * decendent of one), it counts as a primary bandwidth domain, not a
1064          * secondary bandwidth domain under a TT.  An xhci_tt_info structure
1065          * will never be created for the HS root hub.
1066          */
1067         if (!udev->tt || !udev->tt->hub->parent) {
1068                 dev->bw_table = &xhci->rh_bw[port_num - 1].bw_table;
1069         } else {
1070                 struct xhci_root_port_bw_info *rh_bw;
1071                 struct xhci_tt_bw_info *tt_bw;
1072
1073                 rh_bw = &xhci->rh_bw[port_num - 1];
1074                 /* Find the right TT. */
1075                 list_for_each_entry(tt_bw, &rh_bw->tts, tt_list) {
1076                         if (tt_bw->slot_id != udev->tt->hub->slot_id)
1077                                 continue;
1078
1079                         if (!dev->udev->tt->multi ||
1080                                         (udev->tt->multi &&
1081                                          tt_bw->ttport == dev->udev->ttport)) {
1082                                 dev->bw_table = &tt_bw->bw_table;
1083                                 dev->tt_info = tt_bw;
1084                                 break;
1085                         }
1086                 }
1087                 if (!dev->tt_info)
1088                         xhci_warn(xhci, "WARN: Didn't find a matching TT\n");
1089         }
1090
1091         /* Is this a LS/FS device under an external HS hub? */
1092         if (udev->tt && udev->tt->hub->parent) {
1093                 slot_ctx->tt_info = cpu_to_le32(udev->tt->hub->slot_id |
1094                                                 (udev->ttport << 8));
1095                 if (udev->tt->multi)
1096                         slot_ctx->dev_info |= cpu_to_le32(DEV_MTT);
1097         }
1098         xhci_dbg(xhci, "udev->tt = %p\n", udev->tt);
1099         xhci_dbg(xhci, "udev->ttport = 0x%x\n", udev->ttport);
1100
1101         /* Step 4 - ring already allocated */
1102         /* Step 5 */
1103         ep0_ctx->ep_info2 = cpu_to_le32(EP_TYPE(CTRL_EP));
1104         /*
1105          * XXX: Not sure about wireless USB devices.
1106          */
1107         switch (udev->speed) {
1108         case USB_SPEED_SUPER:
1109                 ep0_ctx->ep_info2 |= cpu_to_le32(MAX_PACKET(512));
1110                 break;
1111         case USB_SPEED_HIGH:
1112         /* USB core guesses at a 64-byte max packet first for FS devices */
1113         case USB_SPEED_FULL:
1114                 ep0_ctx->ep_info2 |= cpu_to_le32(MAX_PACKET(64));
1115                 break;
1116         case USB_SPEED_LOW:
1117                 ep0_ctx->ep_info2 |= cpu_to_le32(MAX_PACKET(8));
1118                 break;
1119         case USB_SPEED_WIRELESS:
1120                 xhci_dbg(xhci, "FIXME xHCI doesn't support wireless speeds\n");
1121                 return -EINVAL;
1122                 break;
1123         default:
1124                 /* New speed? */
1125                 BUG();
1126         }
1127         /* EP 0 can handle "burst" sizes of 1, so Max Burst Size field is 0 */
1128         ep0_ctx->ep_info2 |= cpu_to_le32(MAX_BURST(0) | ERROR_COUNT(3));
1129
1130         ep0_ctx->deq = cpu_to_le64(dev->eps[0].ring->first_seg->dma |
1131                                    dev->eps[0].ring->cycle_state);
1132
1133         /* Steps 7 and 8 were done in xhci_alloc_virt_device() */
1134
1135         return 0;
1136 }
1137
1138 /*
1139  * Convert interval expressed as 2^(bInterval - 1) == interval into
1140  * straight exponent value 2^n == interval.
1141  *
1142  */
1143 static unsigned int xhci_parse_exponent_interval(struct usb_device *udev,
1144                 struct usb_host_endpoint *ep)
1145 {
1146         unsigned int interval;
1147
1148         interval = clamp_val(ep->desc.bInterval, 1, 16) - 1;
1149         if (interval != ep->desc.bInterval - 1)
1150                 dev_warn(&udev->dev,
1151                          "ep %#x - rounding interval to %d %sframes\n",
1152                          ep->desc.bEndpointAddress,
1153                          1 << interval,
1154                          udev->speed == USB_SPEED_FULL ? "" : "micro");
1155
1156         if (udev->speed == USB_SPEED_FULL) {
1157                 /*
1158                  * Full speed isoc endpoints specify interval in frames,
1159                  * not microframes. We are using microframes everywhere,
1160                  * so adjust accordingly.
1161                  */
1162                 interval += 3;  /* 1 frame = 2^3 uframes */
1163         }
1164
1165         return interval;
1166 }
1167
1168 /*
1169  * Convert bInterval expressed in microframes (in 1-255 range) to exponent of
1170  * microframes, rounded down to nearest power of 2.
1171  */
1172 static unsigned int xhci_microframes_to_exponent(struct usb_device *udev,
1173                 struct usb_host_endpoint *ep, unsigned int desc_interval,
1174                 unsigned int min_exponent, unsigned int max_exponent)
1175 {
1176         unsigned int interval;
1177
1178         interval = fls(desc_interval) - 1;
1179         interval = clamp_val(interval, min_exponent, max_exponent);
1180         if ((1 << interval) != desc_interval)
1181                 dev_warn(&udev->dev,
1182                          "ep %#x - rounding interval to %d microframes, ep desc says %d microframes\n",
1183                          ep->desc.bEndpointAddress,
1184                          1 << interval,
1185                          desc_interval);
1186
1187         return interval;
1188 }
1189
1190 static unsigned int xhci_parse_microframe_interval(struct usb_device *udev,
1191                 struct usb_host_endpoint *ep)
1192 {
1193         if (ep->desc.bInterval == 0)
1194                 return 0;
1195         return xhci_microframes_to_exponent(udev, ep,
1196                         ep->desc.bInterval, 0, 15);
1197 }
1198
1199
1200 static unsigned int xhci_parse_frame_interval(struct usb_device *udev,
1201                 struct usb_host_endpoint *ep)
1202 {
1203         return xhci_microframes_to_exponent(udev, ep,
1204                         ep->desc.bInterval * 8, 3, 10);
1205 }
1206
1207 /* Return the polling or NAK interval.
1208  *
1209  * The polling interval is expressed in "microframes".  If xHCI's Interval field
1210  * is set to N, it will service the endpoint every 2^(Interval)*125us.
1211  *
1212  * The NAK interval is one NAK per 1 to 255 microframes, or no NAKs if interval
1213  * is set to 0.
1214  */
1215 static unsigned int xhci_get_endpoint_interval(struct usb_device *udev,
1216                 struct usb_host_endpoint *ep)
1217 {
1218         unsigned int interval = 0;
1219
1220         switch (udev->speed) {
1221         case USB_SPEED_HIGH:
1222                 /* Max NAK rate */
1223                 if (usb_endpoint_xfer_control(&ep->desc) ||
1224                     usb_endpoint_xfer_bulk(&ep->desc)) {
1225                         interval = xhci_parse_microframe_interval(udev, ep);
1226                         break;
1227                 }
1228                 /* Fall through - SS and HS isoc/int have same decoding */
1229
1230         case USB_SPEED_SUPER:
1231                 if (usb_endpoint_xfer_int(&ep->desc) ||
1232                     usb_endpoint_xfer_isoc(&ep->desc)) {
1233                         interval = xhci_parse_exponent_interval(udev, ep);
1234                 }
1235                 break;
1236
1237         case USB_SPEED_FULL:
1238                 if (usb_endpoint_xfer_isoc(&ep->desc)) {
1239                         interval = xhci_parse_exponent_interval(udev, ep);
1240                         break;
1241                 }
1242                 /*
1243                  * Fall through for interrupt endpoint interval decoding
1244                  * since it uses the same rules as low speed interrupt
1245                  * endpoints.
1246                  */
1247
1248         case USB_SPEED_LOW:
1249                 if (usb_endpoint_xfer_int(&ep->desc) ||
1250                     usb_endpoint_xfer_isoc(&ep->desc)) {
1251
1252                         interval = xhci_parse_frame_interval(udev, ep);
1253                 }
1254                 break;
1255
1256         default:
1257                 BUG();
1258         }
1259         return EP_INTERVAL(interval);
1260 }
1261
1262 /* The "Mult" field in the endpoint context is only set for SuperSpeed isoc eps.
1263  * High speed endpoint descriptors can define "the number of additional
1264  * transaction opportunities per microframe", but that goes in the Max Burst
1265  * endpoint context field.
1266  */
1267 static u32 xhci_get_endpoint_mult(struct usb_device *udev,
1268                 struct usb_host_endpoint *ep)
1269 {
1270         if (udev->speed != USB_SPEED_SUPER ||
1271                         !usb_endpoint_xfer_isoc(&ep->desc))
1272                 return 0;
1273         return ep->ss_ep_comp.bmAttributes;
1274 }
1275
1276 static u32 xhci_get_endpoint_type(struct usb_device *udev,
1277                 struct usb_host_endpoint *ep)
1278 {
1279         int in;
1280         u32 type;
1281
1282         in = usb_endpoint_dir_in(&ep->desc);
1283         if (usb_endpoint_xfer_control(&ep->desc)) {
1284                 type = EP_TYPE(CTRL_EP);
1285         } else if (usb_endpoint_xfer_bulk(&ep->desc)) {
1286                 if (in)
1287                         type = EP_TYPE(BULK_IN_EP);
1288                 else
1289                         type = EP_TYPE(BULK_OUT_EP);
1290         } else if (usb_endpoint_xfer_isoc(&ep->desc)) {
1291                 if (in)
1292                         type = EP_TYPE(ISOC_IN_EP);
1293                 else
1294                         type = EP_TYPE(ISOC_OUT_EP);
1295         } else if (usb_endpoint_xfer_int(&ep->desc)) {
1296                 if (in)
1297                         type = EP_TYPE(INT_IN_EP);
1298                 else
1299                         type = EP_TYPE(INT_OUT_EP);
1300         } else {
1301                 BUG();
1302         }
1303         return type;
1304 }
1305
1306 /* Return the maximum endpoint service interval time (ESIT) payload.
1307  * Basically, this is the maxpacket size, multiplied by the burst size
1308  * and mult size.
1309  */
1310 static u32 xhci_get_max_esit_payload(struct xhci_hcd *xhci,
1311                 struct usb_device *udev,
1312                 struct usb_host_endpoint *ep)
1313 {
1314         int max_burst;
1315         int max_packet;
1316
1317         /* Only applies for interrupt or isochronous endpoints */
1318         if (usb_endpoint_xfer_control(&ep->desc) ||
1319                         usb_endpoint_xfer_bulk(&ep->desc))
1320                 return 0;
1321
1322         if (udev->speed == USB_SPEED_SUPER)
1323                 return le16_to_cpu(ep->ss_ep_comp.wBytesPerInterval);
1324
1325         max_packet = GET_MAX_PACKET(usb_endpoint_maxp(&ep->desc));
1326         max_burst = (usb_endpoint_maxp(&ep->desc) & 0x1800) >> 11;
1327         /* A 0 in max burst means 1 transfer per ESIT */
1328         return max_packet * (max_burst + 1);
1329 }
1330
1331 /* Set up an endpoint with one ring segment.  Do not allocate stream rings.
1332  * Drivers will have to call usb_alloc_streams() to do that.
1333  */
1334 int xhci_endpoint_init(struct xhci_hcd *xhci,
1335                 struct xhci_virt_device *virt_dev,
1336                 struct usb_device *udev,
1337                 struct usb_host_endpoint *ep,
1338                 gfp_t mem_flags)
1339 {
1340         unsigned int ep_index;
1341         struct xhci_ep_ctx *ep_ctx;
1342         struct xhci_ring *ep_ring;
1343         unsigned int max_packet;
1344         unsigned int max_burst;
1345         u32 max_esit_payload;
1346
1347         ep_index = xhci_get_endpoint_index(&ep->desc);
1348         ep_ctx = xhci_get_ep_ctx(xhci, virt_dev->in_ctx, ep_index);
1349
1350         /* Set up the endpoint ring */
1351         /*
1352          * Isochronous endpoint ring needs bigger size because one isoc URB
1353          * carries multiple packets and it will insert multiple tds to the
1354          * ring.
1355          * This should be replaced with dynamic ring resizing in the future.
1356          */
1357         if (usb_endpoint_xfer_isoc(&ep->desc))
1358                 virt_dev->eps[ep_index].new_ring =
1359                         xhci_ring_alloc(xhci, 8, true, true, mem_flags);
1360         else
1361                 virt_dev->eps[ep_index].new_ring =
1362                         xhci_ring_alloc(xhci, 1, true, false, mem_flags);
1363         if (!virt_dev->eps[ep_index].new_ring) {
1364                 /* Attempt to use the ring cache */
1365                 if (virt_dev->num_rings_cached == 0)
1366                         return -ENOMEM;
1367                 virt_dev->num_rings_cached--;
1368                 virt_dev->eps[ep_index].new_ring =
1369                         virt_dev->ring_cache[virt_dev->num_rings_cached];
1370                 virt_dev->ring_cache[virt_dev->num_rings_cached] = NULL;
1371                 xhci_reinit_cached_ring(xhci, virt_dev->eps[ep_index].new_ring,
1372                         usb_endpoint_xfer_isoc(&ep->desc) ? true : false);
1373         }
1374         virt_dev->eps[ep_index].skip = false;
1375         ep_ring = virt_dev->eps[ep_index].new_ring;
1376         ep_ctx->deq = cpu_to_le64(ep_ring->first_seg->dma | ep_ring->cycle_state);
1377
1378         ep_ctx->ep_info = cpu_to_le32(xhci_get_endpoint_interval(udev, ep)
1379                                       | EP_MULT(xhci_get_endpoint_mult(udev, ep)));
1380
1381         /* FIXME dig Mult and streams info out of ep companion desc */
1382
1383         /* Allow 3 retries for everything but isoc;
1384          * CErr shall be set to 0 for Isoch endpoints.
1385          */
1386         if (!usb_endpoint_xfer_isoc(&ep->desc))
1387                 ep_ctx->ep_info2 = cpu_to_le32(ERROR_COUNT(3));
1388         else
1389                 ep_ctx->ep_info2 = cpu_to_le32(ERROR_COUNT(0));
1390
1391         ep_ctx->ep_info2 |= cpu_to_le32(xhci_get_endpoint_type(udev, ep));
1392
1393         /* Set the max packet size and max burst */
1394         max_packet = GET_MAX_PACKET(usb_endpoint_maxp(&ep->desc));
1395         max_burst = 0;
1396         switch (udev->speed) {
1397         case USB_SPEED_SUPER:
1398                 /* dig out max burst from ep companion desc */
1399                 max_burst = ep->ss_ep_comp.bMaxBurst;
1400                 break;
1401         case USB_SPEED_HIGH:
1402                 /* Some devices get this wrong */
1403                 if (usb_endpoint_xfer_bulk(&ep->desc))
1404                         max_packet = 512;
1405                 /* bits 11:12 specify the number of additional transaction
1406                  * opportunities per microframe (USB 2.0, section 9.6.6)
1407                  */
1408                 if (usb_endpoint_xfer_isoc(&ep->desc) ||
1409                                 usb_endpoint_xfer_int(&ep->desc)) {
1410                         max_burst = (usb_endpoint_maxp(&ep->desc)
1411                                      & 0x1800) >> 11;
1412                 }
1413                 break;
1414         case USB_SPEED_FULL:
1415         case USB_SPEED_LOW:
1416                 break;
1417         default:
1418                 BUG();
1419         }
1420         ep_ctx->ep_info2 |= cpu_to_le32(MAX_PACKET(max_packet) |
1421                         MAX_BURST(max_burst));
1422         max_esit_payload = xhci_get_max_esit_payload(xhci, udev, ep);
1423         ep_ctx->tx_info = cpu_to_le32(MAX_ESIT_PAYLOAD_FOR_EP(max_esit_payload));
1424
1425         /*
1426          * XXX no idea how to calculate the average TRB buffer length for bulk
1427          * endpoints, as the driver gives us no clue how big each scatter gather
1428          * list entry (or buffer) is going to be.
1429          *
1430          * For isochronous and interrupt endpoints, we set it to the max
1431          * available, until we have new API in the USB core to allow drivers to
1432          * declare how much bandwidth they actually need.
1433          *
1434          * Normally, it would be calculated by taking the total of the buffer
1435          * lengths in the TD and then dividing by the number of TRBs in a TD,
1436          * including link TRBs, No-op TRBs, and Event data TRBs.  Since we don't
1437          * use Event Data TRBs, and we don't chain in a link TRB on short
1438          * transfers, we're basically dividing by 1.
1439          *
1440          * xHCI 1.0 and 1.1 specification indicates that the Average TRB Length
1441          * should be set to 8 for control endpoints.
1442          */
1443         if (usb_endpoint_xfer_control(&ep->desc) && xhci->hci_version >= 0x100)
1444                 ep_ctx->tx_info |= cpu_to_le32(AVG_TRB_LENGTH_FOR_EP(8));
1445         else
1446                 ep_ctx->tx_info |=
1447                          cpu_to_le32(AVG_TRB_LENGTH_FOR_EP(max_esit_payload));
1448
1449         /* FIXME Debug endpoint context */
1450         return 0;
1451 }
1452
1453 void xhci_endpoint_zero(struct xhci_hcd *xhci,
1454                 struct xhci_virt_device *virt_dev,
1455                 struct usb_host_endpoint *ep)
1456 {
1457         unsigned int ep_index;
1458         struct xhci_ep_ctx *ep_ctx;
1459
1460         ep_index = xhci_get_endpoint_index(&ep->desc);
1461         ep_ctx = xhci_get_ep_ctx(xhci, virt_dev->in_ctx, ep_index);
1462
1463         ep_ctx->ep_info = 0;
1464         ep_ctx->ep_info2 = 0;
1465         ep_ctx->deq = 0;
1466         ep_ctx->tx_info = 0;
1467         /* Don't free the endpoint ring until the set interface or configuration
1468          * request succeeds.
1469          */
1470 }
1471
1472 void xhci_clear_endpoint_bw_info(struct xhci_bw_info *bw_info)
1473 {
1474         bw_info->ep_interval = 0;
1475         bw_info->mult = 0;
1476         bw_info->num_packets = 0;
1477         bw_info->max_packet_size = 0;
1478         bw_info->type = 0;
1479         bw_info->max_esit_payload = 0;
1480 }
1481
1482 void xhci_update_bw_info(struct xhci_hcd *xhci,
1483                 struct xhci_container_ctx *in_ctx,
1484                 struct xhci_input_control_ctx *ctrl_ctx,
1485                 struct xhci_virt_device *virt_dev)
1486 {
1487         struct xhci_bw_info *bw_info;
1488         struct xhci_ep_ctx *ep_ctx;
1489         unsigned int ep_type;
1490         int i;
1491
1492         for (i = 1; i < 31; ++i) {
1493                 bw_info = &virt_dev->eps[i].bw_info;
1494
1495                 /* We can't tell what endpoint type is being dropped, but
1496                  * unconditionally clearing the bandwidth info for non-periodic
1497                  * endpoints should be harmless because the info will never be
1498                  * set in the first place.
1499                  */
1500                 if (!EP_IS_ADDED(ctrl_ctx, i) && EP_IS_DROPPED(ctrl_ctx, i)) {
1501                         /* Dropped endpoint */
1502                         xhci_clear_endpoint_bw_info(bw_info);
1503                         continue;
1504                 }
1505
1506                 if (EP_IS_ADDED(ctrl_ctx, i)) {
1507                         ep_ctx = xhci_get_ep_ctx(xhci, in_ctx, i);
1508                         ep_type = CTX_TO_EP_TYPE(le32_to_cpu(ep_ctx->ep_info2));
1509
1510                         /* Ignore non-periodic endpoints */
1511                         if (ep_type != ISOC_OUT_EP && ep_type != INT_OUT_EP &&
1512                                         ep_type != ISOC_IN_EP &&
1513                                         ep_type != INT_IN_EP)
1514                                 continue;
1515
1516                         /* Added or changed endpoint */
1517                         bw_info->ep_interval = CTX_TO_EP_INTERVAL(
1518                                         le32_to_cpu(ep_ctx->ep_info));
1519                         /* Number of packets and mult are zero-based in the
1520                          * input context, but we want one-based for the
1521                          * interval table.
1522                          */
1523                         bw_info->mult = CTX_TO_EP_MULT(
1524                                         le32_to_cpu(ep_ctx->ep_info)) + 1;
1525                         bw_info->num_packets = CTX_TO_MAX_BURST(
1526                                         le32_to_cpu(ep_ctx->ep_info2)) + 1;
1527                         bw_info->max_packet_size = MAX_PACKET_DECODED(
1528                                         le32_to_cpu(ep_ctx->ep_info2));
1529                         bw_info->type = ep_type;
1530                         bw_info->max_esit_payload = CTX_TO_MAX_ESIT_PAYLOAD(
1531                                         le32_to_cpu(ep_ctx->tx_info));
1532                 }
1533         }
1534 }
1535
1536 /* Copy output xhci_ep_ctx to the input xhci_ep_ctx copy.
1537  * Useful when you want to change one particular aspect of the endpoint and then
1538  * issue a configure endpoint command.
1539  */
1540 void xhci_endpoint_copy(struct xhci_hcd *xhci,
1541                 struct xhci_container_ctx *in_ctx,
1542                 struct xhci_container_ctx *out_ctx,
1543                 unsigned int ep_index)
1544 {
1545         struct xhci_ep_ctx *out_ep_ctx;
1546         struct xhci_ep_ctx *in_ep_ctx;
1547
1548         out_ep_ctx = xhci_get_ep_ctx(xhci, out_ctx, ep_index);
1549         in_ep_ctx = xhci_get_ep_ctx(xhci, in_ctx, ep_index);
1550
1551         in_ep_ctx->ep_info = out_ep_ctx->ep_info;
1552         in_ep_ctx->ep_info2 = out_ep_ctx->ep_info2;
1553         in_ep_ctx->deq = out_ep_ctx->deq;
1554         in_ep_ctx->tx_info = out_ep_ctx->tx_info;
1555 }
1556
1557 /* Copy output xhci_slot_ctx to the input xhci_slot_ctx.
1558  * Useful when you want to change one particular aspect of the endpoint and then
1559  * issue a configure endpoint command.  Only the context entries field matters,
1560  * but we'll copy the whole thing anyway.
1561  */
1562 void xhci_slot_copy(struct xhci_hcd *xhci,
1563                 struct xhci_container_ctx *in_ctx,
1564                 struct xhci_container_ctx *out_ctx)
1565 {
1566         struct xhci_slot_ctx *in_slot_ctx;
1567         struct xhci_slot_ctx *out_slot_ctx;
1568
1569         in_slot_ctx = xhci_get_slot_ctx(xhci, in_ctx);
1570         out_slot_ctx = xhci_get_slot_ctx(xhci, out_ctx);
1571
1572         in_slot_ctx->dev_info = out_slot_ctx->dev_info;
1573         in_slot_ctx->dev_info2 = out_slot_ctx->dev_info2;
1574         in_slot_ctx->tt_info = out_slot_ctx->tt_info;
1575         in_slot_ctx->dev_state = out_slot_ctx->dev_state;
1576 }
1577
1578 /* Set up the scratchpad buffer array and scratchpad buffers, if needed. */
1579 static int scratchpad_alloc(struct xhci_hcd *xhci, gfp_t flags)
1580 {
1581         int i;
1582         struct device *dev = xhci_to_hcd(xhci)->self.controller;
1583         int num_sp = HCS_MAX_SCRATCHPAD(xhci->hcs_params2);
1584
1585         xhci_dbg(xhci, "Allocating %d scratchpad buffers\n", num_sp);
1586
1587         if (!num_sp)
1588                 return 0;
1589
1590         xhci->scratchpad = kzalloc(sizeof(*xhci->scratchpad), flags);
1591         if (!xhci->scratchpad)
1592                 goto fail_sp;
1593
1594         xhci->scratchpad->sp_array = dma_alloc_coherent(dev,
1595                                      num_sp * sizeof(u64),
1596                                      &xhci->scratchpad->sp_dma, flags);
1597         if (!xhci->scratchpad->sp_array)
1598                 goto fail_sp2;
1599
1600         xhci->scratchpad->sp_buffers = kzalloc(sizeof(void *) * num_sp, flags);
1601         if (!xhci->scratchpad->sp_buffers)
1602                 goto fail_sp3;
1603
1604         xhci->scratchpad->sp_dma_buffers =
1605                 kzalloc(sizeof(dma_addr_t) * num_sp, flags);
1606
1607         if (!xhci->scratchpad->sp_dma_buffers)
1608                 goto fail_sp4;
1609
1610         xhci->dcbaa->dev_context_ptrs[0] = cpu_to_le64(xhci->scratchpad->sp_dma);
1611         for (i = 0; i < num_sp; i++) {
1612                 dma_addr_t dma;
1613                 void *buf = dma_alloc_coherent(dev, xhci->page_size, &dma,
1614                                 flags);
1615                 if (!buf)
1616                         goto fail_sp5;
1617
1618                 xhci->scratchpad->sp_array[i] = dma;
1619                 xhci->scratchpad->sp_buffers[i] = buf;
1620                 xhci->scratchpad->sp_dma_buffers[i] = dma;
1621         }
1622
1623         return 0;
1624
1625  fail_sp5:
1626         for (i = i - 1; i >= 0; i--) {
1627                 dma_free_coherent(dev, xhci->page_size,
1628                                     xhci->scratchpad->sp_buffers[i],
1629                                     xhci->scratchpad->sp_dma_buffers[i]);
1630         }
1631         kfree(xhci->scratchpad->sp_dma_buffers);
1632
1633  fail_sp4:
1634         kfree(xhci->scratchpad->sp_buffers);
1635
1636  fail_sp3:
1637         dma_free_coherent(dev, num_sp * sizeof(u64),
1638                             xhci->scratchpad->sp_array,
1639                             xhci->scratchpad->sp_dma);
1640
1641  fail_sp2:
1642         kfree(xhci->scratchpad);
1643         xhci->scratchpad = NULL;
1644
1645  fail_sp:
1646         return -ENOMEM;
1647 }
1648
1649 static void scratchpad_free(struct xhci_hcd *xhci)
1650 {
1651         int num_sp;
1652         int i;
1653         struct pci_dev  *pdev = to_pci_dev(xhci_to_hcd(xhci)->self.controller);
1654
1655         if (!xhci->scratchpad)
1656                 return;
1657
1658         num_sp = HCS_MAX_SCRATCHPAD(xhci->hcs_params2);
1659
1660         for (i = 0; i < num_sp; i++) {
1661                 dma_free_coherent(&pdev->dev, xhci->page_size,
1662                                     xhci->scratchpad->sp_buffers[i],
1663                                     xhci->scratchpad->sp_dma_buffers[i]);
1664         }
1665         kfree(xhci->scratchpad->sp_dma_buffers);
1666         kfree(xhci->scratchpad->sp_buffers);
1667         dma_free_coherent(&pdev->dev, num_sp * sizeof(u64),
1668                             xhci->scratchpad->sp_array,
1669                             xhci->scratchpad->sp_dma);
1670         kfree(xhci->scratchpad);
1671         xhci->scratchpad = NULL;
1672 }
1673
1674 struct xhci_command *xhci_alloc_command(struct xhci_hcd *xhci,
1675                 bool allocate_in_ctx, bool allocate_completion,
1676                 gfp_t mem_flags)
1677 {
1678         struct xhci_command *command;
1679
1680         command = kzalloc(sizeof(*command), mem_flags);
1681         if (!command)
1682                 return NULL;
1683
1684         if (allocate_in_ctx) {
1685                 command->in_ctx =
1686                         xhci_alloc_container_ctx(xhci, XHCI_CTX_TYPE_INPUT,
1687                                         mem_flags);
1688                 if (!command->in_ctx) {
1689                         kfree(command);
1690                         return NULL;
1691                 }
1692         }
1693
1694         if (allocate_completion) {
1695                 command->completion =
1696                         kzalloc(sizeof(struct completion), mem_flags);
1697                 if (!command->completion) {
1698                         xhci_free_container_ctx(xhci, command->in_ctx);
1699                         kfree(command);
1700                         return NULL;
1701                 }
1702                 init_completion(command->completion);
1703         }
1704
1705         command->status = 0;
1706         INIT_LIST_HEAD(&command->cmd_list);
1707         return command;
1708 }
1709
1710 void xhci_urb_free_priv(struct xhci_hcd *xhci, struct urb_priv *urb_priv)
1711 {
1712         if (urb_priv) {
1713                 kfree(urb_priv->td[0]);
1714                 kfree(urb_priv);
1715         }
1716 }
1717
1718 void xhci_free_command(struct xhci_hcd *xhci,
1719                 struct xhci_command *command)
1720 {
1721         xhci_free_container_ctx(xhci,
1722                         command->in_ctx);
1723         kfree(command->completion);
1724         kfree(command);
1725 }
1726
1727 void xhci_mem_cleanup(struct xhci_hcd *xhci)
1728 {
1729         struct pci_dev  *pdev = to_pci_dev(xhci_to_hcd(xhci)->self.controller);
1730         struct dev_info *dev_info, *next;
1731         struct xhci_cd  *cur_cd, *next_cd;
1732         unsigned long   flags;
1733         int size;
1734         int i, j, num_ports;
1735
1736         /* Free the Event Ring Segment Table and the actual Event Ring */
1737         size = sizeof(struct xhci_erst_entry)*(xhci->erst.num_entries);
1738         if (xhci->erst.entries)
1739                 dma_free_coherent(&pdev->dev, size,
1740                                 xhci->erst.entries, xhci->erst.erst_dma_addr);
1741         xhci->erst.entries = NULL;
1742         xhci_dbg(xhci, "Freed ERST\n");
1743         if (xhci->event_ring)
1744                 xhci_ring_free(xhci, xhci->event_ring);
1745         xhci->event_ring = NULL;
1746         xhci_dbg(xhci, "Freed event ring\n");
1747
1748         xhci->cmd_ring_reserved_trbs = 0;
1749         if (xhci->cmd_ring)
1750                 xhci_ring_free(xhci, xhci->cmd_ring);
1751         xhci->cmd_ring = NULL;
1752         xhci_dbg(xhci, "Freed command ring\n");
1753         list_for_each_entry_safe(cur_cd, next_cd,
1754                         &xhci->cancel_cmd_list, cancel_cmd_list) {
1755                 list_del(&cur_cd->cancel_cmd_list);
1756                 kfree(cur_cd);
1757         }
1758
1759         num_ports = HCS_MAX_PORTS(xhci->hcs_params1);
1760         for (i = 0; i < num_ports && xhci->rh_bw; i++) {
1761                 struct xhci_interval_bw_table *bwt = &xhci->rh_bw[i].bw_table;
1762                 for (j = 0; j < XHCI_MAX_INTERVAL; j++) {
1763                         struct list_head *ep = &bwt->interval_bw[j].endpoints;
1764                         while (!list_empty(ep))
1765                                 list_del_init(ep->next);
1766                 }
1767         }
1768
1769         for (i = HCS_MAX_SLOTS(xhci->hcs_params1); i > 0; i--)
1770                 xhci_free_virt_devices_depth_first(xhci, i);
1771
1772         if (xhci->segment_pool)
1773                 dma_pool_destroy(xhci->segment_pool);
1774         xhci->segment_pool = NULL;
1775         xhci_dbg(xhci, "Freed segment pool\n");
1776
1777         if (xhci->device_pool)
1778                 dma_pool_destroy(xhci->device_pool);
1779         xhci->device_pool = NULL;
1780         xhci_dbg(xhci, "Freed device context pool\n");
1781
1782         if (xhci->small_streams_pool)
1783                 dma_pool_destroy(xhci->small_streams_pool);
1784         xhci->small_streams_pool = NULL;
1785         xhci_dbg(xhci, "Freed small stream array pool\n");
1786
1787         if (xhci->medium_streams_pool)
1788                 dma_pool_destroy(xhci->medium_streams_pool);
1789         xhci->medium_streams_pool = NULL;
1790         xhci_dbg(xhci, "Freed medium stream array pool\n");
1791
1792         if (xhci->dcbaa)
1793                 dma_free_coherent(&pdev->dev, sizeof(*xhci->dcbaa),
1794                                 xhci->dcbaa, xhci->dcbaa->dma);
1795         xhci->dcbaa = NULL;
1796
1797         scratchpad_free(xhci);
1798
1799         spin_lock_irqsave(&xhci->lock, flags);
1800         list_for_each_entry_safe(dev_info, next, &xhci->lpm_failed_devs, list) {
1801                 list_del(&dev_info->list);
1802                 kfree(dev_info);
1803         }
1804         spin_unlock_irqrestore(&xhci->lock, flags);
1805
1806         if (!xhci->rh_bw)
1807                 goto no_bw;
1808
1809         for (i = 0; i < num_ports; i++) {
1810                 struct xhci_tt_bw_info *tt, *n;
1811                 list_for_each_entry_safe(tt, n, &xhci->rh_bw[i].tts, tt_list) {
1812                         list_del(&tt->tt_list);
1813                         kfree(tt);
1814                 }
1815         }
1816
1817 no_bw:
1818         xhci->num_usb2_ports = 0;
1819         xhci->num_usb3_ports = 0;
1820         xhci->num_active_eps = 0;
1821         kfree(xhci->usb2_ports);
1822         kfree(xhci->usb3_ports);
1823         kfree(xhci->port_array);
1824         kfree(xhci->rh_bw);
1825
1826         xhci->usb2_ports = NULL;
1827         xhci->usb3_ports = NULL;
1828         xhci->port_array = NULL;
1829         xhci->rh_bw = NULL;
1830
1831         xhci->page_size = 0;
1832         xhci->page_shift = 0;
1833         xhci->bus_state[0].bus_suspended = 0;
1834         xhci->bus_state[1].bus_suspended = 0;
1835 }
1836
1837 static int xhci_test_trb_in_td(struct xhci_hcd *xhci,
1838                 struct xhci_segment *input_seg,
1839                 union xhci_trb *start_trb,
1840                 union xhci_trb *end_trb,
1841                 dma_addr_t input_dma,
1842                 struct xhci_segment *result_seg,
1843                 char *test_name, int test_number)
1844 {
1845         unsigned long long start_dma;
1846         unsigned long long end_dma;
1847         struct xhci_segment *seg;
1848
1849         start_dma = xhci_trb_virt_to_dma(input_seg, start_trb);
1850         end_dma = xhci_trb_virt_to_dma(input_seg, end_trb);
1851
1852         seg = trb_in_td(input_seg, start_trb, end_trb, input_dma);
1853         if (seg != result_seg) {
1854                 xhci_warn(xhci, "WARN: %s TRB math test %d failed!\n",
1855                                 test_name, test_number);
1856                 xhci_warn(xhci, "Tested TRB math w/ seg %p and "
1857                                 "input DMA 0x%llx\n",
1858                                 input_seg,
1859                                 (unsigned long long) input_dma);
1860                 xhci_warn(xhci, "starting TRB %p (0x%llx DMA), "
1861                                 "ending TRB %p (0x%llx DMA)\n",
1862                                 start_trb, start_dma,
1863                                 end_trb, end_dma);
1864                 xhci_warn(xhci, "Expected seg %p, got seg %p\n",
1865                                 result_seg, seg);
1866                 return -1;
1867         }
1868         return 0;
1869 }
1870
1871 /* TRB math checks for xhci_trb_in_td(), using the command and event rings. */
1872 static int xhci_check_trb_in_td_math(struct xhci_hcd *xhci, gfp_t mem_flags)
1873 {
1874         struct {
1875                 dma_addr_t              input_dma;
1876                 struct xhci_segment     *result_seg;
1877         } simple_test_vector [] = {
1878                 /* A zeroed DMA field should fail */
1879                 { 0, NULL },
1880                 /* One TRB before the ring start should fail */
1881                 { xhci->event_ring->first_seg->dma - 16, NULL },
1882                 /* One byte before the ring start should fail */
1883                 { xhci->event_ring->first_seg->dma - 1, NULL },
1884                 /* Starting TRB should succeed */
1885                 { xhci->event_ring->first_seg->dma, xhci->event_ring->first_seg },
1886                 /* Ending TRB should succeed */
1887                 { xhci->event_ring->first_seg->dma + (TRBS_PER_SEGMENT - 1)*16,
1888                         xhci->event_ring->first_seg },
1889                 /* One byte after the ring end should fail */
1890                 { xhci->event_ring->first_seg->dma + (TRBS_PER_SEGMENT - 1)*16 + 1, NULL },
1891                 /* One TRB after the ring end should fail */
1892                 { xhci->event_ring->first_seg->dma + (TRBS_PER_SEGMENT)*16, NULL },
1893                 /* An address of all ones should fail */
1894                 { (dma_addr_t) (~0), NULL },
1895         };
1896         struct {
1897                 struct xhci_segment     *input_seg;
1898                 union xhci_trb          *start_trb;
1899                 union xhci_trb          *end_trb;
1900                 dma_addr_t              input_dma;
1901                 struct xhci_segment     *result_seg;
1902         } complex_test_vector [] = {
1903                 /* Test feeding a valid DMA address from a different ring */
1904                 {       .input_seg = xhci->event_ring->first_seg,
1905                         .start_trb = xhci->event_ring->first_seg->trbs,
1906                         .end_trb = &xhci->event_ring->first_seg->trbs[TRBS_PER_SEGMENT - 1],
1907                         .input_dma = xhci->cmd_ring->first_seg->dma,
1908                         .result_seg = NULL,
1909                 },
1910                 /* Test feeding a valid end TRB from a different ring */
1911                 {       .input_seg = xhci->event_ring->first_seg,
1912                         .start_trb = xhci->event_ring->first_seg->trbs,
1913                         .end_trb = &xhci->cmd_ring->first_seg->trbs[TRBS_PER_SEGMENT - 1],
1914                         .input_dma = xhci->cmd_ring->first_seg->dma,
1915                         .result_seg = NULL,
1916                 },
1917                 /* Test feeding a valid start and end TRB from a different ring */
1918                 {       .input_seg = xhci->event_ring->first_seg,
1919                         .start_trb = xhci->cmd_ring->first_seg->trbs,
1920                         .end_trb = &xhci->cmd_ring->first_seg->trbs[TRBS_PER_SEGMENT - 1],
1921                         .input_dma = xhci->cmd_ring->first_seg->dma,
1922                         .result_seg = NULL,
1923                 },
1924                 /* TRB in this ring, but after this TD */
1925                 {       .input_seg = xhci->event_ring->first_seg,
1926                         .start_trb = &xhci->event_ring->first_seg->trbs[0],
1927                         .end_trb = &xhci->event_ring->first_seg->trbs[3],
1928                         .input_dma = xhci->event_ring->first_seg->dma + 4*16,
1929                         .result_seg = NULL,
1930                 },
1931                 /* TRB in this ring, but before this TD */
1932                 {       .input_seg = xhci->event_ring->first_seg,
1933                         .start_trb = &xhci->event_ring->first_seg->trbs[3],
1934                         .end_trb = &xhci->event_ring->first_seg->trbs[6],
1935                         .input_dma = xhci->event_ring->first_seg->dma + 2*16,
1936                         .result_seg = NULL,
1937                 },
1938                 /* TRB in this ring, but after this wrapped TD */
1939                 {       .input_seg = xhci->event_ring->first_seg,
1940                         .start_trb = &xhci->event_ring->first_seg->trbs[TRBS_PER_SEGMENT - 3],
1941                         .end_trb = &xhci->event_ring->first_seg->trbs[1],
1942                         .input_dma = xhci->event_ring->first_seg->dma + 2*16,
1943                         .result_seg = NULL,
1944                 },
1945                 /* TRB in this ring, but before this wrapped TD */
1946                 {       .input_seg = xhci->event_ring->first_seg,
1947                         .start_trb = &xhci->event_ring->first_seg->trbs[TRBS_PER_SEGMENT - 3],
1948                         .end_trb = &xhci->event_ring->first_seg->trbs[1],
1949                         .input_dma = xhci->event_ring->first_seg->dma + (TRBS_PER_SEGMENT - 4)*16,
1950                         .result_seg = NULL,
1951                 },
1952                 /* TRB not in this ring, and we have a wrapped TD */
1953                 {       .input_seg = xhci->event_ring->first_seg,
1954                         .start_trb = &xhci->event_ring->first_seg->trbs[TRBS_PER_SEGMENT - 3],
1955                         .end_trb = &xhci->event_ring->first_seg->trbs[1],
1956                         .input_dma = xhci->cmd_ring->first_seg->dma + 2*16,
1957                         .result_seg = NULL,
1958                 },
1959         };
1960
1961         unsigned int num_tests;
1962         int i, ret;
1963
1964         num_tests = ARRAY_SIZE(simple_test_vector);
1965         for (i = 0; i < num_tests; i++) {
1966                 ret = xhci_test_trb_in_td(xhci,
1967                                 xhci->event_ring->first_seg,
1968                                 xhci->event_ring->first_seg->trbs,
1969                                 &xhci->event_ring->first_seg->trbs[TRBS_PER_SEGMENT - 1],
1970                                 simple_test_vector[i].input_dma,
1971                                 simple_test_vector[i].result_seg,
1972                                 "Simple", i);
1973                 if (ret < 0)
1974                         return ret;
1975         }
1976
1977         num_tests = ARRAY_SIZE(complex_test_vector);
1978         for (i = 0; i < num_tests; i++) {
1979                 ret = xhci_test_trb_in_td(xhci,
1980                                 complex_test_vector[i].input_seg,
1981                                 complex_test_vector[i].start_trb,
1982                                 complex_test_vector[i].end_trb,
1983                                 complex_test_vector[i].input_dma,
1984                                 complex_test_vector[i].result_seg,
1985                                 "Complex", i);
1986                 if (ret < 0)
1987                         return ret;
1988         }
1989         xhci_dbg(xhci, "TRB math tests passed.\n");
1990         return 0;
1991 }
1992
1993 static void xhci_set_hc_event_deq(struct xhci_hcd *xhci)
1994 {
1995         u64 temp;
1996         dma_addr_t deq;
1997
1998         deq = xhci_trb_virt_to_dma(xhci->event_ring->deq_seg,
1999                         xhci->event_ring->dequeue);
2000         if (deq == 0 && !in_interrupt())
2001                 xhci_warn(xhci, "WARN something wrong with SW event ring "
2002                                 "dequeue ptr.\n");
2003         /* Update HC event ring dequeue pointer */
2004         temp = xhci_read_64(xhci, &xhci->ir_set->erst_dequeue);
2005         temp &= ERST_PTR_MASK;
2006         /* Don't clear the EHB bit (which is RW1C) because
2007          * there might be more events to service.
2008          */
2009         temp &= ~ERST_EHB;
2010         xhci_dbg(xhci, "// Write event ring dequeue pointer, "
2011                         "preserving EHB bit\n");
2012         xhci_write_64(xhci, ((u64) deq & (u64) ~ERST_PTR_MASK) | temp,
2013                         &xhci->ir_set->erst_dequeue);
2014 }
2015
2016 static void xhci_add_in_port(struct xhci_hcd *xhci, unsigned int num_ports,
2017                 __le32 __iomem *addr, u8 major_revision)
2018 {
2019         u32 temp, port_offset, port_count;
2020         int i;
2021
2022         if (major_revision > 0x03) {
2023                 xhci_warn(xhci, "Ignoring unknown port speed, "
2024                                 "Ext Cap %p, revision = 0x%x\n",
2025                                 addr, major_revision);
2026                 /* Ignoring port protocol we can't understand. FIXME */
2027                 return;
2028         }
2029
2030         /* Port offset and count in the third dword, see section 7.2 */
2031         temp = xhci_readl(xhci, addr + 2);
2032         port_offset = XHCI_EXT_PORT_OFF(temp);
2033         port_count = XHCI_EXT_PORT_COUNT(temp);
2034         xhci_dbg(xhci, "Ext Cap %p, port offset = %u, "
2035                         "count = %u, revision = 0x%x\n",
2036                         addr, port_offset, port_count, major_revision);
2037         /* Port count includes the current port offset */
2038         if (port_offset == 0 || (port_offset + port_count - 1) > num_ports)
2039                 /* WTF? "Valid values are ‘1’ to MaxPorts" */
2040                 return;
2041
2042         /* Check the host's USB2 LPM capability */
2043         if ((xhci->hci_version == 0x96) && (major_revision != 0x03) &&
2044                         (temp & XHCI_L1C)) {
2045                 xhci_dbg(xhci, "xHCI 0.96: support USB2 software lpm\n");
2046                 xhci->sw_lpm_support = 1;
2047         }
2048
2049         if ((xhci->hci_version >= 0x100) && (major_revision != 0x03)) {
2050                 xhci_dbg(xhci, "xHCI 1.0: support USB2 software lpm\n");
2051                 xhci->sw_lpm_support = 1;
2052                 if (temp & XHCI_HLC) {
2053                         xhci_dbg(xhci, "xHCI 1.0: support USB2 hardware lpm\n");
2054                         xhci->hw_lpm_support = 1;
2055                 }
2056         }
2057
2058         port_offset--;
2059         for (i = port_offset; i < (port_offset + port_count); i++) {
2060                 /* Duplicate entry.  Ignore the port if the revisions differ. */
2061                 if (xhci->port_array[i] != 0) {
2062                         xhci_warn(xhci, "Duplicate port entry, Ext Cap %p,"
2063                                         " port %u\n", addr, i);
2064                         xhci_warn(xhci, "Port was marked as USB %u, "
2065                                         "duplicated as USB %u\n",
2066                                         xhci->port_array[i], major_revision);
2067                         /* Only adjust the roothub port counts if we haven't
2068                          * found a similar duplicate.
2069                          */
2070                         if (xhci->port_array[i] != major_revision &&
2071                                 xhci->port_array[i] != DUPLICATE_ENTRY) {
2072                                 if (xhci->port_array[i] == 0x03)
2073                                         xhci->num_usb3_ports--;
2074                                 else
2075                                         xhci->num_usb2_ports--;
2076                                 xhci->port_array[i] = DUPLICATE_ENTRY;
2077                         }
2078                         /* FIXME: Should we disable the port? */
2079                         continue;
2080                 }
2081                 xhci->port_array[i] = major_revision;
2082                 if (major_revision == 0x03)
2083                         xhci->num_usb3_ports++;
2084                 else
2085                         xhci->num_usb2_ports++;
2086         }
2087         /* FIXME: Should we disable ports not in the Extended Capabilities? */
2088 }
2089
2090 /*
2091  * Scan the Extended Capabilities for the "Supported Protocol Capabilities" that
2092  * specify what speeds each port is supposed to be.  We can't count on the port
2093  * speed bits in the PORTSC register being correct until a device is connected,
2094  * but we need to set up the two fake roothubs with the correct number of USB
2095  * 3.0 and USB 2.0 ports at host controller initialization time.
2096  */
2097 static int xhci_setup_port_arrays(struct xhci_hcd *xhci, gfp_t flags)
2098 {
2099         __le32 __iomem *addr;
2100         u32 offset;
2101         unsigned int num_ports;
2102         int i, j, port_index;
2103
2104         addr = &xhci->cap_regs->hcc_params;
2105         offset = XHCI_HCC_EXT_CAPS(xhci_readl(xhci, addr));
2106         if (offset == 0) {
2107                 xhci_err(xhci, "No Extended Capability registers, "
2108                                 "unable to set up roothub.\n");
2109                 return -ENODEV;
2110         }
2111
2112         num_ports = HCS_MAX_PORTS(xhci->hcs_params1);
2113         xhci->port_array = kzalloc(sizeof(*xhci->port_array)*num_ports, flags);
2114         if (!xhci->port_array)
2115                 return -ENOMEM;
2116
2117         xhci->rh_bw = kzalloc(sizeof(*xhci->rh_bw)*num_ports, flags);
2118         if (!xhci->rh_bw)
2119                 return -ENOMEM;
2120         for (i = 0; i < num_ports; i++) {
2121                 struct xhci_interval_bw_table *bw_table;
2122
2123                 INIT_LIST_HEAD(&xhci->rh_bw[i].tts);
2124                 bw_table = &xhci->rh_bw[i].bw_table;
2125                 for (j = 0; j < XHCI_MAX_INTERVAL; j++)
2126                         INIT_LIST_HEAD(&bw_table->interval_bw[j].endpoints);
2127         }
2128
2129         /*
2130          * For whatever reason, the first capability offset is from the
2131          * capability register base, not from the HCCPARAMS register.
2132          * See section 5.3.6 for offset calculation.
2133          */
2134         addr = &xhci->cap_regs->hc_capbase + offset;
2135         while (1) {
2136                 u32 cap_id;
2137
2138                 cap_id = xhci_readl(xhci, addr);
2139                 if (XHCI_EXT_CAPS_ID(cap_id) == XHCI_EXT_CAPS_PROTOCOL)
2140                         xhci_add_in_port(xhci, num_ports, addr,
2141                                         (u8) XHCI_EXT_PORT_MAJOR(cap_id));
2142                 offset = XHCI_EXT_CAPS_NEXT(cap_id);
2143                 if (!offset || (xhci->num_usb2_ports + xhci->num_usb3_ports)
2144                                 == num_ports)
2145                         break;
2146                 /*
2147                  * Once you're into the Extended Capabilities, the offset is
2148                  * always relative to the register holding the offset.
2149                  */
2150                 addr += offset;
2151         }
2152
2153         if (xhci->num_usb2_ports == 0 && xhci->num_usb3_ports == 0) {
2154                 xhci_warn(xhci, "No ports on the roothubs?\n");
2155                 return -ENODEV;
2156         }
2157         xhci_dbg(xhci, "Found %u USB 2.0 ports and %u USB 3.0 ports.\n",
2158                         xhci->num_usb2_ports, xhci->num_usb3_ports);
2159
2160         /* Place limits on the number of roothub ports so that the hub
2161          * descriptors aren't longer than the USB core will allocate.
2162          */
2163         if (xhci->num_usb3_ports > 15) {
2164                 xhci_dbg(xhci, "Limiting USB 3.0 roothub ports to 15.\n");
2165                 xhci->num_usb3_ports = 15;
2166         }
2167         if (xhci->num_usb2_ports > USB_MAXCHILDREN) {
2168                 xhci_dbg(xhci, "Limiting USB 2.0 roothub ports to %u.\n",
2169                                 USB_MAXCHILDREN);
2170                 xhci->num_usb2_ports = USB_MAXCHILDREN;
2171         }
2172
2173         /*
2174          * Note we could have all USB 3.0 ports, or all USB 2.0 ports.
2175          * Not sure how the USB core will handle a hub with no ports...
2176          */
2177         if (xhci->num_usb2_ports) {
2178                 xhci->usb2_ports = kmalloc(sizeof(*xhci->usb2_ports)*
2179                                 xhci->num_usb2_ports, flags);
2180                 if (!xhci->usb2_ports)
2181                         return -ENOMEM;
2182
2183                 port_index = 0;
2184                 for (i = 0; i < num_ports; i++) {
2185                         if (xhci->port_array[i] == 0x03 ||
2186                                         xhci->port_array[i] == 0 ||
2187                                         xhci->port_array[i] == DUPLICATE_ENTRY)
2188                                 continue;
2189
2190                         xhci->usb2_ports[port_index] =
2191                                 &xhci->op_regs->port_status_base +
2192                                 NUM_PORT_REGS*i;
2193                         xhci_dbg(xhci, "USB 2.0 port at index %u, "
2194                                         "addr = %p\n", i,
2195                                         xhci->usb2_ports[port_index]);
2196                         port_index++;
2197                         if (port_index == xhci->num_usb2_ports)
2198                                 break;
2199                 }
2200         }
2201         if (xhci->num_usb3_ports) {
2202                 xhci->usb3_ports = kmalloc(sizeof(*xhci->usb3_ports)*
2203                                 xhci->num_usb3_ports, flags);
2204                 if (!xhci->usb3_ports)
2205                         return -ENOMEM;
2206
2207                 port_index = 0;
2208                 for (i = 0; i < num_ports; i++)
2209                         if (xhci->port_array[i] == 0x03) {
2210                                 xhci->usb3_ports[port_index] =
2211                                         &xhci->op_regs->port_status_base +
2212                                         NUM_PORT_REGS*i;
2213                                 xhci_dbg(xhci, "USB 3.0 port at index %u, "
2214                                                 "addr = %p\n", i,
2215                                                 xhci->usb3_ports[port_index]);
2216                                 port_index++;
2217                                 if (port_index == xhci->num_usb3_ports)
2218                                         break;
2219                         }
2220         }
2221         return 0;
2222 }
2223
2224 int xhci_mem_init(struct xhci_hcd *xhci, gfp_t flags)
2225 {
2226         dma_addr_t      dma;
2227         struct device   *dev = xhci_to_hcd(xhci)->self.controller;
2228         unsigned int    val, val2;
2229         u64             val_64;
2230         struct xhci_segment     *seg;
2231         u32 page_size;
2232         int i;
2233
2234         INIT_LIST_HEAD(&xhci->lpm_failed_devs);
2235         INIT_LIST_HEAD(&xhci->cancel_cmd_list);
2236
2237         page_size = xhci_readl(xhci, &xhci->op_regs->page_size);
2238         xhci_dbg(xhci, "Supported page size register = 0x%x\n", page_size);
2239         for (i = 0; i < 16; i++) {
2240                 if ((0x1 & page_size) != 0)
2241                         break;
2242                 page_size = page_size >> 1;
2243         }
2244         if (i < 16)
2245                 xhci_dbg(xhci, "Supported page size of %iK\n", (1 << (i+12)) / 1024);
2246         else
2247                 xhci_warn(xhci, "WARN: no supported page size\n");
2248         /* Use 4K pages, since that's common and the minimum the HC supports */
2249         xhci->page_shift = 12;
2250         xhci->page_size = 1 << xhci->page_shift;
2251         xhci_dbg(xhci, "HCD page size set to %iK\n", xhci->page_size / 1024);
2252
2253         /*
2254          * Program the Number of Device Slots Enabled field in the CONFIG
2255          * register with the max value of slots the HC can handle.
2256          */
2257         val = HCS_MAX_SLOTS(xhci_readl(xhci, &xhci->cap_regs->hcs_params1));
2258         xhci_dbg(xhci, "// xHC can handle at most %d device slots.\n",
2259                         (unsigned int) val);
2260         val2 = xhci_readl(xhci, &xhci->op_regs->config_reg);
2261         val |= (val2 & ~HCS_SLOTS_MASK);
2262         xhci_dbg(xhci, "// Setting Max device slots reg = 0x%x.\n",
2263                         (unsigned int) val);
2264         xhci_writel(xhci, val, &xhci->op_regs->config_reg);
2265
2266         /*
2267          * Section 5.4.8 - doorbell array must be
2268          * "physically contiguous and 64-byte (cache line) aligned".
2269          */
2270         xhci->dcbaa = dma_alloc_coherent(dev, sizeof(*xhci->dcbaa), &dma,
2271                         flags);
2272         if (!xhci->dcbaa)
2273                 goto fail;
2274         memset(xhci->dcbaa, 0, sizeof *(xhci->dcbaa));
2275         xhci->dcbaa->dma = dma;
2276         xhci_dbg(xhci, "// Device context base array address = 0x%llx (DMA), %p (virt)\n",
2277                         (unsigned long long)xhci->dcbaa->dma, xhci->dcbaa);
2278         xhci_write_64(xhci, dma, &xhci->op_regs->dcbaa_ptr);
2279
2280         /*
2281          * Initialize the ring segment pool.  The ring must be a contiguous
2282          * structure comprised of TRBs.  The TRBs must be 16 byte aligned,
2283          * however, the command ring segment needs 64-byte aligned segments,
2284          * so we pick the greater alignment need.
2285          */
2286         xhci->segment_pool = dma_pool_create("xHCI ring segments", dev,
2287                         SEGMENT_SIZE, 64, xhci->page_size);
2288
2289         /* See Table 46 and Note on Figure 55 */
2290         xhci->device_pool = dma_pool_create("xHCI input/output contexts", dev,
2291                         2112, 64, xhci->page_size);
2292         if (!xhci->segment_pool || !xhci->device_pool)
2293                 goto fail;
2294
2295         /* Linear stream context arrays don't have any boundary restrictions,
2296          * and only need to be 16-byte aligned.
2297          */
2298         xhci->small_streams_pool =
2299                 dma_pool_create("xHCI 256 byte stream ctx arrays",
2300                         dev, SMALL_STREAM_ARRAY_SIZE, 16, 0);
2301         xhci->medium_streams_pool =
2302                 dma_pool_create("xHCI 1KB stream ctx arrays",
2303                         dev, MEDIUM_STREAM_ARRAY_SIZE, 16, 0);
2304         /* Any stream context array bigger than MEDIUM_STREAM_ARRAY_SIZE
2305          * will be allocated with dma_alloc_coherent()
2306          */
2307
2308         if (!xhci->small_streams_pool || !xhci->medium_streams_pool)
2309                 goto fail;
2310
2311         /* Set up the command ring to have one segments for now. */
2312         xhci->cmd_ring = xhci_ring_alloc(xhci, 1, true, false, flags);
2313         if (!xhci->cmd_ring)
2314                 goto fail;
2315         xhci_dbg(xhci, "Allocated command ring at %p\n", xhci->cmd_ring);
2316         xhci_dbg(xhci, "First segment DMA is 0x%llx\n",
2317                         (unsigned long long)xhci->cmd_ring->first_seg->dma);
2318
2319         /* Set the address in the Command Ring Control register */
2320         val_64 = xhci_read_64(xhci, &xhci->op_regs->cmd_ring);
2321         val_64 = (val_64 & (u64) CMD_RING_RSVD_BITS) |
2322                 (xhci->cmd_ring->first_seg->dma & (u64) ~CMD_RING_RSVD_BITS) |
2323                 xhci->cmd_ring->cycle_state;
2324         xhci_dbg(xhci, "// Setting command ring address to 0x%x\n", val);
2325         xhci_write_64(xhci, val_64, &xhci->op_regs->cmd_ring);
2326         xhci_dbg_cmd_ptrs(xhci);
2327
2328         val = xhci_readl(xhci, &xhci->cap_regs->db_off);
2329         val &= DBOFF_MASK;
2330         xhci_dbg(xhci, "// Doorbell array is located at offset 0x%x"
2331                         " from cap regs base addr\n", val);
2332         xhci->dba = (void __iomem *) xhci->cap_regs + val;
2333         xhci_dbg_regs(xhci);
2334         xhci_print_run_regs(xhci);
2335         /* Set ir_set to interrupt register set 0 */
2336         xhci->ir_set = &xhci->run_regs->ir_set[0];
2337
2338         /*
2339          * Event ring setup: Allocate a normal ring, but also setup
2340          * the event ring segment table (ERST).  Section 4.9.3.
2341          */
2342         xhci_dbg(xhci, "// Allocating event ring\n");
2343         xhci->event_ring = xhci_ring_alloc(xhci, ERST_NUM_SEGS, false, false,
2344                                                 flags);
2345         if (!xhci->event_ring)
2346                 goto fail;
2347         if (xhci_check_trb_in_td_math(xhci, flags) < 0)
2348                 goto fail;
2349
2350         xhci->erst.entries = dma_alloc_coherent(dev,
2351                         sizeof(struct xhci_erst_entry) * ERST_NUM_SEGS, &dma,
2352                         flags);
2353         if (!xhci->erst.entries)
2354                 goto fail;
2355         xhci_dbg(xhci, "// Allocated event ring segment table at 0x%llx\n",
2356                         (unsigned long long)dma);
2357
2358         memset(xhci->erst.entries, 0, sizeof(struct xhci_erst_entry)*ERST_NUM_SEGS);
2359         xhci->erst.num_entries = ERST_NUM_SEGS;
2360         xhci->erst.erst_dma_addr = dma;
2361         xhci_dbg(xhci, "Set ERST to 0; private num segs = %i, virt addr = %p, dma addr = 0x%llx\n",
2362                         xhci->erst.num_entries,
2363                         xhci->erst.entries,
2364                         (unsigned long long)xhci->erst.erst_dma_addr);
2365
2366         /* set ring base address and size for each segment table entry */
2367         for (val = 0, seg = xhci->event_ring->first_seg; val < ERST_NUM_SEGS; val++) {
2368                 struct xhci_erst_entry *entry = &xhci->erst.entries[val];
2369                 entry->seg_addr = cpu_to_le64(seg->dma);
2370                 entry->seg_size = cpu_to_le32(TRBS_PER_SEGMENT);
2371                 entry->rsvd = 0;
2372                 seg = seg->next;
2373         }
2374
2375         /* set ERST count with the number of entries in the segment table */
2376         val = xhci_readl(xhci, &xhci->ir_set->erst_size);
2377         val &= ERST_SIZE_MASK;
2378         val |= ERST_NUM_SEGS;
2379         xhci_dbg(xhci, "// Write ERST size = %i to ir_set 0 (some bits preserved)\n",
2380                         val);
2381         xhci_writel(xhci, val, &xhci->ir_set->erst_size);
2382
2383         xhci_dbg(xhci, "// Set ERST entries to point to event ring.\n");
2384         /* set the segment table base address */
2385         xhci_dbg(xhci, "// Set ERST base address for ir_set 0 = 0x%llx\n",
2386                         (unsigned long long)xhci->erst.erst_dma_addr);
2387         val_64 = xhci_read_64(xhci, &xhci->ir_set->erst_base);
2388         val_64 &= ERST_PTR_MASK;
2389         val_64 |= (xhci->erst.erst_dma_addr & (u64) ~ERST_PTR_MASK);
2390         xhci_write_64(xhci, val_64, &xhci->ir_set->erst_base);
2391
2392         /* Set the event ring dequeue address */
2393         xhci_set_hc_event_deq(xhci);
2394         xhci_dbg(xhci, "Wrote ERST address to ir_set 0.\n");
2395         xhci_print_ir_set(xhci, 0);
2396
2397         /*
2398          * XXX: Might need to set the Interrupter Moderation Register to
2399          * something other than the default (~1ms minimum between interrupts).
2400          * See section 5.5.1.2.
2401          */
2402         init_completion(&xhci->addr_dev);
2403         for (i = 0; i < MAX_HC_SLOTS; ++i)
2404                 xhci->devs[i] = NULL;
2405         for (i = 0; i < USB_MAXCHILDREN; ++i) {
2406                 xhci->bus_state[0].resume_done[i] = 0;
2407                 xhci->bus_state[1].resume_done[i] = 0;
2408         }
2409
2410         if (scratchpad_alloc(xhci, flags))
2411                 goto fail;
2412         if (xhci_setup_port_arrays(xhci, flags))
2413                 goto fail;
2414
2415         return 0;
2416
2417 fail:
2418         xhci_warn(xhci, "Couldn't initialize memory\n");
2419         xhci_halt(xhci);
2420         xhci_reset(xhci);
2421         xhci_mem_cleanup(xhci);
2422         return -ENOMEM;
2423 }