1 // SPDX-License-Identifier: GPL-2.0 2 3 /* Copyright (c) 2012-2018, The Linux Foundation. All rights reserved. 4 * Copyright (C) 2019-2020 Linaro Ltd. 5 */ 6 7 #include <linux/types.h> 8 #include <linux/bits.h> 9 #include <linux/bitfield.h> 10 #include <linux/refcount.h> 11 #include <linux/scatterlist.h> 12 #include <linux/dma-direction.h> 13 14 #include "gsi.h" 15 #include "gsi_private.h" 16 #include "gsi_trans.h" 17 #include "ipa_gsi.h" 18 #include "ipa_data.h" 19 #include "ipa_cmd.h" 20 21 /** 22 * DOC: GSI Transactions 23 * 24 * A GSI transaction abstracts the behavior of a GSI channel by representing 25 * everything about a related group of IPA commands in a single structure. 26 * (A "command" in this sense is either a data transfer or an IPA immediate 27 * command.) Most details of interaction with the GSI hardware are managed 28 * by the GSI transaction core, allowing users to simply describe commands 29 * to be performed. When a transaction has completed a callback function 30 * (dependent on the type of endpoint associated with the channel) allows 31 * cleanup of resources associated with the transaction. 32 * 33 * To perform a command (or set of them), a user of the GSI transaction 34 * interface allocates a transaction, indicating the number of TREs required 35 * (one per command). If sufficient TREs are available, they are reserved 36 * for use in the transaction and the allocation succeeds. This way 37 * exhaustion of the available TREs in a channel ring is detected 38 * as early as possible. All resources required to complete a transaction 39 * are allocated at transaction allocation time. 40 * 41 * Commands performed as part of a transaction are represented in an array 42 * of Linux scatterlist structures. This array is allocated with the 43 * transaction, and its entries are initialized using standard scatterlist 44 * functions (such as sg_set_buf() or skb_to_sgvec()). 45 * 46 * Once a transaction's scatterlist structures have been initialized, the 47 * transaction is committed. The caller is responsible for mapping buffers 48 * for DMA if necessary, and this should be done *before* allocating 49 * the transaction. Between a successful allocation and commit of a 50 * transaction no errors should occur. 51 * 52 * Committing transfers ownership of the entire transaction to the GSI 53 * transaction core. The GSI transaction code formats the content of 54 * the scatterlist array into the channel ring buffer and informs the 55 * hardware that new TREs are available to process. 56 * 57 * The last TRE in each transaction is marked to interrupt the AP when the 58 * GSI hardware has completed it. Because transfers described by TREs are 59 * performed strictly in order, signaling the completion of just the last 60 * TRE in the transaction is sufficient to indicate the full transaction 61 * is complete. 62 * 63 * When a transaction is complete, ipa_gsi_trans_complete() is called by the 64 * GSI code into the IPA layer, allowing it to perform any final cleanup 65 * required before the transaction is freed. 66 */ 67 68 /* Hardware values representing a transfer element type */ 69 enum gsi_tre_type { 70 GSI_RE_XFER = 0x2, 71 GSI_RE_IMMD_CMD = 0x3, 72 }; 73 74 /* An entry in a channel ring */ 75 struct gsi_tre { 76 __le64 addr; /* DMA address */ 77 __le16 len_opcode; /* length in bytes or enum IPA_CMD_* */ 78 __le16 reserved; 79 __le32 flags; /* TRE_FLAGS_* */ 80 }; 81 82 /* gsi_tre->flags mask values (in CPU byte order) */ 83 #define TRE_FLAGS_CHAIN_FMASK GENMASK(0, 0) 84 #define TRE_FLAGS_IEOT_FMASK GENMASK(9, 9) 85 #define TRE_FLAGS_BEI_FMASK GENMASK(10, 10) 86 #define TRE_FLAGS_TYPE_FMASK GENMASK(23, 16) 87 88 int gsi_trans_pool_init(struct gsi_trans_pool *pool, size_t size, u32 count, 89 u32 max_alloc) 90 { 91 void *virt; 92 93 if (!size) 94 return -EINVAL; 95 if (count < max_alloc) 96 return -EINVAL; 97 if (!max_alloc) 98 return -EINVAL; 99 100 /* By allocating a few extra entries in our pool (one less 101 * than the maximum number that will be requested in a 102 * single allocation), we can always satisfy requests without 103 * ever worrying about straddling the end of the pool array. 104 * If there aren't enough entries starting at the free index, 105 * we just allocate free entries from the beginning of the pool. 106 */ 107 virt = kcalloc(count + max_alloc - 1, size, GFP_KERNEL); 108 if (!virt) 109 return -ENOMEM; 110 111 pool->base = virt; 112 /* If the allocator gave us any extra memory, use it */ 113 pool->count = ksize(pool->base) / size; 114 pool->free = 0; 115 pool->max_alloc = max_alloc; 116 pool->size = size; 117 pool->addr = 0; /* Only used for DMA pools */ 118 119 return 0; 120 } 121 122 void gsi_trans_pool_exit(struct gsi_trans_pool *pool) 123 { 124 kfree(pool->base); 125 memset(pool, 0, sizeof(*pool)); 126 } 127 128 /* Allocate the requested number of (zeroed) entries from the pool */ 129 /* Home-grown DMA pool. This way we can preallocate and use the tre_count 130 * to guarantee allocations will succeed. Even though we specify max_alloc 131 * (and it can be more than one), we only allow allocation of a single 132 * element from a DMA pool. 133 */ 134 int gsi_trans_pool_init_dma(struct device *dev, struct gsi_trans_pool *pool, 135 size_t size, u32 count, u32 max_alloc) 136 { 137 size_t total_size; 138 dma_addr_t addr; 139 void *virt; 140 141 if (!size) 142 return -EINVAL; 143 if (count < max_alloc) 144 return -EINVAL; 145 if (!max_alloc) 146 return -EINVAL; 147 148 /* Don't let allocations cross a power-of-two boundary */ 149 size = __roundup_pow_of_two(size); 150 total_size = (count + max_alloc - 1) * size; 151 152 /* The allocator will give us a power-of-2 number of pages 153 * sufficient to satisfy our request. Round up our requested 154 * size to avoid any unused space in the allocation. This way 155 * gsi_trans_pool_exit_dma() can assume the total allocated 156 * size is exactly (count * size). 157 */ 158 total_size = get_order(total_size) << PAGE_SHIFT; 159 160 virt = dma_alloc_coherent(dev, total_size, &addr, GFP_KERNEL); 161 if (!virt) 162 return -ENOMEM; 163 164 pool->base = virt; 165 pool->count = total_size / size; 166 pool->free = 0; 167 pool->size = size; 168 pool->max_alloc = max_alloc; 169 pool->addr = addr; 170 171 return 0; 172 } 173 174 void gsi_trans_pool_exit_dma(struct device *dev, struct gsi_trans_pool *pool) 175 { 176 size_t total_size = pool->count * pool->size; 177 178 dma_free_coherent(dev, total_size, pool->base, pool->addr); 179 memset(pool, 0, sizeof(*pool)); 180 } 181 182 /* Return the byte offset of the next free entry in the pool */ 183 static u32 gsi_trans_pool_alloc_common(struct gsi_trans_pool *pool, u32 count) 184 { 185 u32 offset; 186 187 WARN_ON(!count); 188 WARN_ON(count > pool->max_alloc); 189 190 /* Allocate from beginning if wrap would occur */ 191 if (count > pool->count - pool->free) 192 pool->free = 0; 193 194 offset = pool->free * pool->size; 195 pool->free += count; 196 memset(pool->base + offset, 0, count * pool->size); 197 198 return offset; 199 } 200 201 /* Allocate a contiguous block of zeroed entries from a pool */ 202 void *gsi_trans_pool_alloc(struct gsi_trans_pool *pool, u32 count) 203 { 204 return pool->base + gsi_trans_pool_alloc_common(pool, count); 205 } 206 207 /* Allocate a single zeroed entry from a DMA pool */ 208 void *gsi_trans_pool_alloc_dma(struct gsi_trans_pool *pool, dma_addr_t *addr) 209 { 210 u32 offset = gsi_trans_pool_alloc_common(pool, 1); 211 212 *addr = pool->addr + offset; 213 214 return pool->base + offset; 215 } 216 217 /* Map a TRE ring entry index to the transaction it is associated with */ 218 static void gsi_trans_map(struct gsi_trans *trans, u32 index) 219 { 220 struct gsi_channel *channel = &trans->gsi->channel[trans->channel_id]; 221 222 /* The completion event will indicate the last TRE used */ 223 index += trans->used_count - 1; 224 225 /* Note: index *must* be used modulo the ring count here */ 226 channel->trans_info.map[index % channel->tre_ring.count] = trans; 227 } 228 229 /* Return the transaction mapped to a given ring entry */ 230 struct gsi_trans * 231 gsi_channel_trans_mapped(struct gsi_channel *channel, u32 index) 232 { 233 /* Note: index *must* be used modulo the ring count here */ 234 return channel->trans_info.map[index % channel->tre_ring.count]; 235 } 236 237 /* Return the oldest completed transaction for a channel (or null) */ 238 struct gsi_trans *gsi_channel_trans_complete(struct gsi_channel *channel) 239 { 240 struct gsi_trans_info *trans_info = &channel->trans_info; 241 u16 trans_id = trans_info->completed_id; 242 243 if (trans_id == trans_info->pending_id) 244 return NULL; 245 246 return &trans_info->trans[trans_id %= channel->tre_count]; 247 } 248 249 /* Move a transaction from the allocated list to the committed list */ 250 static void gsi_trans_move_committed(struct gsi_trans *trans) 251 { 252 struct gsi_channel *channel = &trans->gsi->channel[trans->channel_id]; 253 struct gsi_trans_info *trans_info = &channel->trans_info; 254 255 spin_lock_bh(&trans_info->spinlock); 256 257 list_move_tail(&trans->links, &trans_info->committed); 258 259 spin_unlock_bh(&trans_info->spinlock); 260 261 /* This allocated transaction is now committed */ 262 trans_info->allocated_id++; 263 } 264 265 /* Move transactions from the committed list to the pending list */ 266 static void gsi_trans_move_pending(struct gsi_trans *trans) 267 { 268 struct gsi_channel *channel = &trans->gsi->channel[trans->channel_id]; 269 struct gsi_trans_info *trans_info = &channel->trans_info; 270 u16 trans_index = trans - &trans_info->trans[0]; 271 struct list_head list; 272 u16 delta; 273 274 spin_lock_bh(&trans_info->spinlock); 275 276 /* Move this transaction and all predecessors to the pending list */ 277 list_cut_position(&list, &trans_info->committed, &trans->links); 278 list_splice_tail(&list, &trans_info->pending); 279 280 spin_unlock_bh(&trans_info->spinlock); 281 282 /* These committed transactions are now pending */ 283 delta = trans_index - trans_info->committed_id + 1; 284 trans_info->committed_id += delta % channel->tre_count; 285 } 286 287 /* Move a transaction and all of its predecessors from the pending list 288 * to the completed list. 289 */ 290 void gsi_trans_move_complete(struct gsi_trans *trans) 291 { 292 struct gsi_channel *channel = &trans->gsi->channel[trans->channel_id]; 293 struct gsi_trans_info *trans_info = &channel->trans_info; 294 u16 trans_index = trans - trans_info->trans; 295 struct list_head list; 296 u16 delta; 297 298 spin_lock_bh(&trans_info->spinlock); 299 300 /* Move this transaction and all predecessors to completed list */ 301 list_cut_position(&list, &trans_info->pending, &trans->links); 302 list_splice_tail(&list, &trans_info->complete); 303 304 spin_unlock_bh(&trans_info->spinlock); 305 306 /* These pending transactions are now completed */ 307 delta = trans_index - trans_info->pending_id + 1; 308 delta %= channel->tre_count; 309 trans_info->pending_id += delta; 310 } 311 312 /* Move a transaction from the completed list to the polled list */ 313 void gsi_trans_move_polled(struct gsi_trans *trans) 314 { 315 struct gsi_channel *channel = &trans->gsi->channel[trans->channel_id]; 316 struct gsi_trans_info *trans_info = &channel->trans_info; 317 318 spin_lock_bh(&trans_info->spinlock); 319 320 list_move_tail(&trans->links, &trans_info->polled); 321 322 spin_unlock_bh(&trans_info->spinlock); 323 324 /* This completed transaction is now polled */ 325 trans_info->completed_id++; 326 } 327 328 /* Reserve some number of TREs on a channel. Returns true if successful */ 329 static bool 330 gsi_trans_tre_reserve(struct gsi_trans_info *trans_info, u32 tre_count) 331 { 332 int avail = atomic_read(&trans_info->tre_avail); 333 int new; 334 335 do { 336 new = avail - (int)tre_count; 337 if (unlikely(new < 0)) 338 return false; 339 } while (!atomic_try_cmpxchg(&trans_info->tre_avail, &avail, new)); 340 341 return true; 342 } 343 344 /* Release previously-reserved TRE entries to a channel */ 345 static void 346 gsi_trans_tre_release(struct gsi_trans_info *trans_info, u32 tre_count) 347 { 348 atomic_add(tre_count, &trans_info->tre_avail); 349 } 350 351 /* Return true if no transactions are allocated, false otherwise */ 352 bool gsi_channel_trans_idle(struct gsi *gsi, u32 channel_id) 353 { 354 u32 tre_max = gsi_channel_tre_max(gsi, channel_id); 355 struct gsi_trans_info *trans_info; 356 357 trans_info = &gsi->channel[channel_id].trans_info; 358 359 return atomic_read(&trans_info->tre_avail) == tre_max; 360 } 361 362 /* Allocate a GSI transaction on a channel */ 363 struct gsi_trans *gsi_channel_trans_alloc(struct gsi *gsi, u32 channel_id, 364 u32 tre_count, 365 enum dma_data_direction direction) 366 { 367 struct gsi_channel *channel = &gsi->channel[channel_id]; 368 struct gsi_trans_info *trans_info; 369 struct gsi_trans *trans; 370 u16 trans_index; 371 372 if (WARN_ON(tre_count > channel->trans_tre_max)) 373 return NULL; 374 375 trans_info = &channel->trans_info; 376 377 /* If we can't reserve the TREs for the transaction, we're done */ 378 if (!gsi_trans_tre_reserve(trans_info, tre_count)) 379 return NULL; 380 381 trans_index = trans_info->free_id % channel->tre_count; 382 trans = &trans_info->trans[trans_index]; 383 memset(trans, 0, sizeof(*trans)); 384 385 /* Initialize non-zero fields in the transaction */ 386 trans->gsi = gsi; 387 trans->channel_id = channel_id; 388 trans->rsvd_count = tre_count; 389 init_completion(&trans->completion); 390 391 /* Allocate the scatterlist */ 392 trans->sgl = gsi_trans_pool_alloc(&trans_info->sg_pool, tre_count); 393 sg_init_marker(trans->sgl, tre_count); 394 395 trans->direction = direction; 396 refcount_set(&trans->refcount, 1); 397 398 /* This free transaction will now be allocated */ 399 trans_info->free_id++; 400 401 spin_lock_bh(&trans_info->spinlock); 402 403 list_add_tail(&trans->links, &trans_info->alloc); 404 405 spin_unlock_bh(&trans_info->spinlock); 406 407 return trans; 408 } 409 410 /* Free a previously-allocated transaction */ 411 void gsi_trans_free(struct gsi_trans *trans) 412 { 413 refcount_t *refcount = &trans->refcount; 414 struct gsi_trans_info *trans_info; 415 bool last; 416 417 /* We must hold the lock to release the last reference */ 418 if (refcount_dec_not_one(refcount)) 419 return; 420 421 trans_info = &trans->gsi->channel[trans->channel_id].trans_info; 422 423 spin_lock_bh(&trans_info->spinlock); 424 425 /* Reference might have been added before we got the lock */ 426 last = refcount_dec_and_test(refcount); 427 if (last) 428 list_del(&trans->links); 429 430 spin_unlock_bh(&trans_info->spinlock); 431 432 if (!last) 433 return; 434 435 /* Unused transactions are allocated but never committed, pending, 436 * completed, or polled. 437 */ 438 if (!trans->used_count) { 439 trans_info->allocated_id++; 440 trans_info->committed_id++; 441 trans_info->pending_id++; 442 trans_info->completed_id++; 443 } else { 444 ipa_gsi_trans_release(trans); 445 } 446 447 /* This transaction is now free */ 448 trans_info->polled_id++; 449 450 /* Releasing the reserved TREs implicitly frees the sgl[] and 451 * (if present) info[] arrays, plus the transaction itself. 452 */ 453 gsi_trans_tre_release(trans_info, trans->rsvd_count); 454 } 455 456 /* Add an immediate command to a transaction */ 457 void gsi_trans_cmd_add(struct gsi_trans *trans, void *buf, u32 size, 458 dma_addr_t addr, enum ipa_cmd_opcode opcode) 459 { 460 u32 which = trans->used_count++; 461 struct scatterlist *sg; 462 463 WARN_ON(which >= trans->rsvd_count); 464 465 /* Commands are quite different from data transfer requests. 466 * Their payloads come from a pool whose memory is allocated 467 * using dma_alloc_coherent(). We therefore do *not* map them 468 * for DMA (unlike what we do for pages and skbs). 469 * 470 * When a transaction completes, the SGL is normally unmapped. 471 * A command transaction has direction DMA_NONE, which tells 472 * gsi_trans_complete() to skip the unmapping step. 473 * 474 * The only things we use directly in a command scatter/gather 475 * entry are the DMA address and length. We still need the SG 476 * table flags to be maintained though, so assign a NULL page 477 * pointer for that purpose. 478 */ 479 sg = &trans->sgl[which]; 480 sg_assign_page(sg, NULL); 481 sg_dma_address(sg) = addr; 482 sg_dma_len(sg) = size; 483 484 trans->cmd_opcode[which] = opcode; 485 } 486 487 /* Add a page transfer to a transaction. It will fill the only TRE. */ 488 int gsi_trans_page_add(struct gsi_trans *trans, struct page *page, u32 size, 489 u32 offset) 490 { 491 struct scatterlist *sg = &trans->sgl[0]; 492 int ret; 493 494 if (WARN_ON(trans->rsvd_count != 1)) 495 return -EINVAL; 496 if (WARN_ON(trans->used_count)) 497 return -EINVAL; 498 499 sg_set_page(sg, page, size, offset); 500 ret = dma_map_sg(trans->gsi->dev, sg, 1, trans->direction); 501 if (!ret) 502 return -ENOMEM; 503 504 trans->used_count++; /* Transaction now owns the (DMA mapped) page */ 505 506 return 0; 507 } 508 509 /* Add an SKB transfer to a transaction. No other TREs will be used. */ 510 int gsi_trans_skb_add(struct gsi_trans *trans, struct sk_buff *skb) 511 { 512 struct scatterlist *sg = &trans->sgl[0]; 513 u32 used_count; 514 int ret; 515 516 if (WARN_ON(trans->rsvd_count != 1)) 517 return -EINVAL; 518 if (WARN_ON(trans->used_count)) 519 return -EINVAL; 520 521 /* skb->len will not be 0 (checked early) */ 522 ret = skb_to_sgvec(skb, sg, 0, skb->len); 523 if (ret < 0) 524 return ret; 525 used_count = ret; 526 527 ret = dma_map_sg(trans->gsi->dev, sg, used_count, trans->direction); 528 if (!ret) 529 return -ENOMEM; 530 531 /* Transaction now owns the (DMA mapped) skb */ 532 trans->used_count += used_count; 533 534 return 0; 535 } 536 537 /* Compute the length/opcode value to use for a TRE */ 538 static __le16 gsi_tre_len_opcode(enum ipa_cmd_opcode opcode, u32 len) 539 { 540 return opcode == IPA_CMD_NONE ? cpu_to_le16((u16)len) 541 : cpu_to_le16((u16)opcode); 542 } 543 544 /* Compute the flags value to use for a given TRE */ 545 static __le32 gsi_tre_flags(bool last_tre, bool bei, enum ipa_cmd_opcode opcode) 546 { 547 enum gsi_tre_type tre_type; 548 u32 tre_flags; 549 550 tre_type = opcode == IPA_CMD_NONE ? GSI_RE_XFER : GSI_RE_IMMD_CMD; 551 tre_flags = u32_encode_bits(tre_type, TRE_FLAGS_TYPE_FMASK); 552 553 /* Last TRE contains interrupt flags */ 554 if (last_tre) { 555 /* All transactions end in a transfer completion interrupt */ 556 tre_flags |= TRE_FLAGS_IEOT_FMASK; 557 /* Don't interrupt when outbound commands are acknowledged */ 558 if (bei) 559 tre_flags |= TRE_FLAGS_BEI_FMASK; 560 } else { /* All others indicate there's more to come */ 561 tre_flags |= TRE_FLAGS_CHAIN_FMASK; 562 } 563 564 return cpu_to_le32(tre_flags); 565 } 566 567 static void gsi_trans_tre_fill(struct gsi_tre *dest_tre, dma_addr_t addr, 568 u32 len, bool last_tre, bool bei, 569 enum ipa_cmd_opcode opcode) 570 { 571 struct gsi_tre tre; 572 573 tre.addr = cpu_to_le64(addr); 574 tre.len_opcode = gsi_tre_len_opcode(opcode, len); 575 tre.reserved = 0; 576 tre.flags = gsi_tre_flags(last_tre, bei, opcode); 577 578 /* ARM64 can write 16 bytes as a unit with a single instruction. 579 * Doing the assignment this way is an attempt to make that happen. 580 */ 581 *dest_tre = tre; 582 } 583 584 /** 585 * __gsi_trans_commit() - Common GSI transaction commit code 586 * @trans: Transaction to commit 587 * @ring_db: Whether to tell the hardware about these queued transfers 588 * 589 * Formats channel ring TRE entries based on the content of the scatterlist. 590 * Maps a transaction pointer to the last ring entry used for the transaction, 591 * so it can be recovered when it completes. Moves the transaction to the 592 * pending list. Finally, updates the channel ring pointer and optionally 593 * rings the doorbell. 594 */ 595 static void __gsi_trans_commit(struct gsi_trans *trans, bool ring_db) 596 { 597 struct gsi_channel *channel = &trans->gsi->channel[trans->channel_id]; 598 struct gsi_ring *tre_ring = &channel->tre_ring; 599 enum ipa_cmd_opcode opcode = IPA_CMD_NONE; 600 bool bei = channel->toward_ipa; 601 struct gsi_tre *dest_tre; 602 struct scatterlist *sg; 603 u32 byte_count = 0; 604 u8 *cmd_opcode; 605 u32 avail; 606 u32 i; 607 608 WARN_ON(!trans->used_count); 609 610 /* Consume the entries. If we cross the end of the ring while 611 * filling them we'll switch to the beginning to finish. 612 * If there is no info array we're doing a simple data 613 * transfer request, whose opcode is IPA_CMD_NONE. 614 */ 615 cmd_opcode = channel->command ? &trans->cmd_opcode[0] : NULL; 616 avail = tre_ring->count - tre_ring->index % tre_ring->count; 617 dest_tre = gsi_ring_virt(tre_ring, tre_ring->index); 618 for_each_sg(trans->sgl, sg, trans->used_count, i) { 619 bool last_tre = i == trans->used_count - 1; 620 dma_addr_t addr = sg_dma_address(sg); 621 u32 len = sg_dma_len(sg); 622 623 byte_count += len; 624 if (!avail--) 625 dest_tre = gsi_ring_virt(tre_ring, 0); 626 if (cmd_opcode) 627 opcode = *cmd_opcode++; 628 629 gsi_trans_tre_fill(dest_tre, addr, len, last_tre, bei, opcode); 630 dest_tre++; 631 } 632 /* Associate the TRE with the transaction */ 633 gsi_trans_map(trans, tre_ring->index); 634 635 tre_ring->index += trans->used_count; 636 637 trans->len = byte_count; 638 if (channel->toward_ipa) 639 gsi_trans_tx_committed(trans); 640 641 gsi_trans_move_committed(trans); 642 643 /* Ring doorbell if requested, or if all TREs are allocated */ 644 if (ring_db || !atomic_read(&channel->trans_info.tre_avail)) { 645 /* Report what we're handing off to hardware for TX channels */ 646 if (channel->toward_ipa) 647 gsi_trans_tx_queued(trans); 648 gsi_trans_move_pending(trans); 649 gsi_channel_doorbell(channel); 650 } 651 } 652 653 /* Commit a GSI transaction */ 654 void gsi_trans_commit(struct gsi_trans *trans, bool ring_db) 655 { 656 if (trans->used_count) 657 __gsi_trans_commit(trans, ring_db); 658 else 659 gsi_trans_free(trans); 660 } 661 662 /* Commit a GSI transaction and wait for it to complete */ 663 void gsi_trans_commit_wait(struct gsi_trans *trans) 664 { 665 if (!trans->used_count) 666 goto out_trans_free; 667 668 refcount_inc(&trans->refcount); 669 670 __gsi_trans_commit(trans, true); 671 672 wait_for_completion(&trans->completion); 673 674 out_trans_free: 675 gsi_trans_free(trans); 676 } 677 678 /* Process the completion of a transaction; called while polling */ 679 void gsi_trans_complete(struct gsi_trans *trans) 680 { 681 /* If the entire SGL was mapped when added, unmap it now */ 682 if (trans->direction != DMA_NONE) 683 dma_unmap_sg(trans->gsi->dev, trans->sgl, trans->used_count, 684 trans->direction); 685 686 ipa_gsi_trans_complete(trans); 687 688 complete(&trans->completion); 689 690 gsi_trans_free(trans); 691 } 692 693 /* Cancel a channel's pending transactions */ 694 void gsi_channel_trans_cancel_pending(struct gsi_channel *channel) 695 { 696 struct gsi_trans_info *trans_info = &channel->trans_info; 697 u16 trans_id = trans_info->pending_id; 698 699 /* channel->gsi->mutex is held by caller */ 700 spin_lock_bh(&trans_info->spinlock); 701 702 list_splice_tail_init(&trans_info->pending, &trans_info->complete); 703 704 spin_unlock_bh(&trans_info->spinlock); 705 706 /* If there are no pending transactions, we're done */ 707 if (trans_id == trans_info->committed_id) 708 return; 709 710 /* Mark all pending transactions cancelled */ 711 do { 712 struct gsi_trans *trans; 713 714 trans = &trans_info->trans[trans_id % channel->tre_count]; 715 trans->cancelled = true; 716 } while (++trans_id != trans_info->committed_id); 717 718 /* All pending transactions are now completed */ 719 trans_info->pending_id = trans_info->committed_id; 720 721 /* Schedule NAPI polling to complete the cancelled transactions */ 722 napi_schedule(&channel->napi); 723 } 724 725 /* Issue a command to read a single byte from a channel */ 726 int gsi_trans_read_byte(struct gsi *gsi, u32 channel_id, dma_addr_t addr) 727 { 728 struct gsi_channel *channel = &gsi->channel[channel_id]; 729 struct gsi_ring *tre_ring = &channel->tre_ring; 730 struct gsi_trans_info *trans_info; 731 struct gsi_tre *dest_tre; 732 733 trans_info = &channel->trans_info; 734 735 /* First reserve the TRE, if possible */ 736 if (!gsi_trans_tre_reserve(trans_info, 1)) 737 return -EBUSY; 738 739 /* Now fill the reserved TRE and tell the hardware */ 740 741 dest_tre = gsi_ring_virt(tre_ring, tre_ring->index); 742 gsi_trans_tre_fill(dest_tre, addr, 1, true, false, IPA_CMD_NONE); 743 744 tre_ring->index++; 745 gsi_channel_doorbell(channel); 746 747 return 0; 748 } 749 750 /* Mark a gsi_trans_read_byte() request done */ 751 void gsi_trans_read_byte_done(struct gsi *gsi, u32 channel_id) 752 { 753 struct gsi_channel *channel = &gsi->channel[channel_id]; 754 755 gsi_trans_tre_release(&channel->trans_info, 1); 756 } 757 758 /* Initialize a channel's GSI transaction info */ 759 int gsi_channel_trans_init(struct gsi *gsi, u32 channel_id) 760 { 761 struct gsi_channel *channel = &gsi->channel[channel_id]; 762 u32 tre_count = channel->tre_count; 763 struct gsi_trans_info *trans_info; 764 u32 tre_max; 765 int ret; 766 767 /* Ensure the size of a channel element is what's expected */ 768 BUILD_BUG_ON(sizeof(struct gsi_tre) != GSI_RING_ELEMENT_SIZE); 769 770 trans_info = &channel->trans_info; 771 772 /* The tre_avail field is what ultimately limits the number of 773 * outstanding transactions and their resources. A transaction 774 * allocation succeeds only if the TREs available are sufficient 775 * for what the transaction might need. 776 */ 777 tre_max = gsi_channel_tre_max(channel->gsi, channel_id); 778 atomic_set(&trans_info->tre_avail, tre_max); 779 780 /* We can't use more TREs than the number available in the ring. 781 * This limits the number of transactions that can be outstanding. 782 * Worst case is one TRE per transaction (but we actually limit 783 * it to something a little less than that). By allocating a 784 * power-of-two number of transactions we can use an index 785 * modulo that number to determine the next one that's free. 786 * Transactions are allocated one at a time. 787 */ 788 trans_info->trans = kcalloc(tre_count, sizeof(*trans_info->trans), 789 GFP_KERNEL); 790 if (!trans_info->trans) 791 return -ENOMEM; 792 trans_info->free_id = 0; /* all modulo channel->tre_count */ 793 trans_info->allocated_id = 0; 794 trans_info->committed_id = 0; 795 trans_info->pending_id = 0; 796 trans_info->completed_id = 0; 797 trans_info->polled_id = 0; 798 799 /* A completion event contains a pointer to the TRE that caused 800 * the event (which will be the last one used by the transaction). 801 * Each entry in this map records the transaction associated 802 * with a corresponding completed TRE. 803 */ 804 trans_info->map = kcalloc(tre_count, sizeof(*trans_info->map), 805 GFP_KERNEL); 806 if (!trans_info->map) { 807 ret = -ENOMEM; 808 goto err_trans_free; 809 } 810 811 /* A transaction uses a scatterlist array to represent the data 812 * transfers implemented by the transaction. Each scatterlist 813 * element is used to fill a single TRE when the transaction is 814 * committed. So we need as many scatterlist elements as the 815 * maximum number of TREs that can be outstanding. 816 */ 817 ret = gsi_trans_pool_init(&trans_info->sg_pool, 818 sizeof(struct scatterlist), 819 tre_max, channel->trans_tre_max); 820 if (ret) 821 goto err_map_free; 822 823 spin_lock_init(&trans_info->spinlock); 824 INIT_LIST_HEAD(&trans_info->alloc); 825 INIT_LIST_HEAD(&trans_info->committed); 826 INIT_LIST_HEAD(&trans_info->pending); 827 INIT_LIST_HEAD(&trans_info->complete); 828 INIT_LIST_HEAD(&trans_info->polled); 829 830 return 0; 831 832 err_map_free: 833 kfree(trans_info->map); 834 err_trans_free: 835 kfree(trans_info->trans); 836 837 dev_err(gsi->dev, "error %d initializing channel %u transactions\n", 838 ret, channel_id); 839 840 return ret; 841 } 842 843 /* Inverse of gsi_channel_trans_init() */ 844 void gsi_channel_trans_exit(struct gsi_channel *channel) 845 { 846 struct gsi_trans_info *trans_info = &channel->trans_info; 847 848 gsi_trans_pool_exit(&trans_info->sg_pool); 849 kfree(trans_info->trans); 850 kfree(trans_info->map); 851 } 852