| 1 | // SPDX-License-Identifier: GPL-2.0 |
| 2 | |
| 3 | /* Copyright (c) 2012-2018, The Linux Foundation. All rights reserved. |
| 4 | * Copyright (C) 2019-2024 Linaro Ltd. |
| 5 | */ |
| 6 | |
| 7 | #include <linux/bitfield.h> |
| 8 | #include <linux/bits.h> |
| 9 | #include <linux/dma-direction.h> |
| 10 | #include <linux/refcount.h> |
| 11 | #include <linux/scatterlist.h> |
| 12 | #include <linux/types.h> |
| 13 | |
| 14 | #include "gsi.h" |
| 15 | #include "gsi_private.h" |
| 16 | #include "gsi_trans.h" |
| 17 | #include "ipa_cmd.h" |
| 18 | #include "ipa_data.h" |
| 19 | #include "ipa_gsi.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 operations in a single structure. |
| 26 | * (A "operation" 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 operations |
| 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 an operation (or set of them), a user of the GSI transaction |
| 34 | * interface allocates a transaction, indicating the number of TREs required |
| 35 | * (one per operation). 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 as early |
| 38 | * as possible. Any other resources that might be needed to complete a |
| 39 | * transaction are also allocated when the transaction is allocated. |
| 40 | * |
| 41 | * Operations performed as part of a transaction are represented in an array |
| 42 | * of Linux scatterlist structures, allocated with the transaction. These |
| 43 | * scatterlist structures are initialized by "adding" operations to the |
| 44 | * transaction. If a buffer in an operation must be mapped for DMA, this is |
| 45 | * done at the time it is added to the transaction. It is possible for a |
| 46 | * mapping error to occur when an operation is added. In this case the |
| 47 | * transaction should simply be freed; this correctly releases resources |
| 48 | * associated with the transaction. |
| 49 | * |
| 50 | * Once all operations have been successfully added to a transaction, the |
| 51 | * transaction is committed. Committing transfers ownership of the entire |
| 52 | * transaction to the GSI transaction core. The GSI transaction code |
| 53 | * formats the content of the scatterlist array into the channel ring |
| 54 | * buffer and informs the hardware that new TREs are available to process. |
| 55 | * |
| 56 | * The last TRE in each transaction is marked to interrupt the AP when the |
| 57 | * GSI hardware has completed it. Because transfers described by TREs are |
| 58 | * performed strictly in order, signaling the completion of just the last |
| 59 | * TRE in the transaction is sufficient to indicate the full transaction |
| 60 | * is complete. |
| 61 | * |
| 62 | * When a transaction is complete, ipa_gsi_trans_complete() is called by the |
| 63 | * GSI code into the IPA layer, allowing it to perform any final cleanup |
| 64 | * required before the transaction is freed. |
| 65 | */ |
| 66 | |
| 67 | /* Hardware values representing a transfer element type */ |
| 68 | enum gsi_tre_type { |
| 69 | GSI_RE_XFER = 0x2, |
| 70 | GSI_RE_IMMD_CMD = 0x3, |
| 71 | }; |
| 72 | |
| 73 | /* An entry in a channel ring */ |
| 74 | struct gsi_tre { |
| 75 | __le64 addr; /* DMA address */ |
| 76 | __le16 len_opcode; /* length in bytes or enum IPA_CMD_* */ |
| 77 | __le16 reserved; |
| 78 | __le32 flags; /* TRE_FLAGS_* */ |
| 79 | }; |
| 80 | |
| 81 | /* gsi_tre->flags mask values (in CPU byte order) */ |
| 82 | #define TRE_FLAGS_CHAIN_FMASK GENMASK(0, 0) |
| 83 | #define TRE_FLAGS_IEOT_FMASK GENMASK(9, 9) |
| 84 | #define TRE_FLAGS_BEI_FMASK GENMASK(10, 10) |
| 85 | #define TRE_FLAGS_TYPE_FMASK GENMASK(23, 16) |
| 86 | |
| 87 | int gsi_trans_pool_init(struct gsi_trans_pool *pool, size_t size, u32 count, |
| 88 | u32 max_alloc) |
| 89 | { |
| 90 | size_t alloc_size; |
| 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 | alloc_size = size_mul(factor1: count + max_alloc - 1, factor2: size); |
| 108 | alloc_size = kmalloc_size_roundup(size: alloc_size); |
| 109 | virt = kzalloc(alloc_size, GFP_KERNEL); |
| 110 | if (!virt) |
| 111 | return -ENOMEM; |
| 112 | |
| 113 | pool->base = virt; |
| 114 | /* If the allocator gave us any extra memory, use it */ |
| 115 | pool->count = alloc_size / size; |
| 116 | pool->free = 0; |
| 117 | pool->max_alloc = max_alloc; |
| 118 | pool->size = size; |
| 119 | pool->addr = 0; /* Only used for DMA pools */ |
| 120 | |
| 121 | return 0; |
| 122 | } |
| 123 | |
| 124 | void gsi_trans_pool_exit(struct gsi_trans_pool *pool) |
| 125 | { |
| 126 | kfree(objp: pool->base); |
| 127 | memset(pool, 0, sizeof(*pool)); |
| 128 | } |
| 129 | |
| 130 | /* Home-grown DMA pool. This way we can preallocate the pool, and guarantee |
| 131 | * allocations will succeed. The immediate commands in a transaction can |
| 132 | * require up to max_alloc elements from the pool. But we only allow |
| 133 | * allocation of a single element from a DMA pool at a time. |
| 134 | */ |
| 135 | int gsi_trans_pool_init_dma(struct device *dev, struct gsi_trans_pool *pool, |
| 136 | size_t size, u32 count, u32 max_alloc) |
| 137 | { |
| 138 | size_t total_size; |
| 139 | dma_addr_t addr; |
| 140 | void *virt; |
| 141 | |
| 142 | if (!size) |
| 143 | return -EINVAL; |
| 144 | if (count < max_alloc) |
| 145 | return -EINVAL; |
| 146 | if (!max_alloc) |
| 147 | return -EINVAL; |
| 148 | |
| 149 | /* Don't let allocations cross a power-of-two boundary */ |
| 150 | size = __roundup_pow_of_two(n: size); |
| 151 | total_size = (count + max_alloc - 1) * size; |
| 152 | |
| 153 | /* The allocator will give us a power-of-2 number of pages |
| 154 | * sufficient to satisfy our request. Round up our requested |
| 155 | * size to avoid any unused space in the allocation. This way |
| 156 | * gsi_trans_pool_exit_dma() can assume the total allocated |
| 157 | * size is exactly (count * size). |
| 158 | */ |
| 159 | total_size = PAGE_SIZE << get_order(size: total_size); |
| 160 | |
| 161 | virt = dma_alloc_coherent(dev, size: total_size, dma_handle: &addr, GFP_KERNEL); |
| 162 | if (!virt) |
| 163 | return -ENOMEM; |
| 164 | |
| 165 | pool->base = virt; |
| 166 | pool->count = total_size / size; |
| 167 | pool->free = 0; |
| 168 | pool->size = size; |
| 169 | pool->max_alloc = max_alloc; |
| 170 | pool->addr = addr; |
| 171 | |
| 172 | return 0; |
| 173 | } |
| 174 | |
| 175 | void gsi_trans_pool_exit_dma(struct device *dev, struct gsi_trans_pool *pool) |
| 176 | { |
| 177 | size_t total_size = pool->count * pool->size; |
| 178 | |
| 179 | dma_free_coherent(dev, size: total_size, cpu_addr: pool->base, dma_handle: pool->addr); |
| 180 | memset(pool, 0, sizeof(*pool)); |
| 181 | } |
| 182 | |
| 183 | /* Return the byte offset of the next free entry in the pool */ |
| 184 | static u32 gsi_trans_pool_alloc_common(struct gsi_trans_pool *pool, u32 count) |
| 185 | { |
| 186 | u32 offset; |
| 187 | |
| 188 | WARN_ON(!count); |
| 189 | WARN_ON(count > pool->max_alloc); |
| 190 | |
| 191 | /* Allocate from beginning if wrap would occur */ |
| 192 | if (count > pool->count - pool->free) |
| 193 | pool->free = 0; |
| 194 | |
| 195 | offset = pool->free * pool->size; |
| 196 | pool->free += count; |
| 197 | memset(pool->base + offset, 0, count * pool->size); |
| 198 | |
| 199 | return offset; |
| 200 | } |
| 201 | |
| 202 | /* Allocate a contiguous block of zeroed entries from a pool */ |
| 203 | void *gsi_trans_pool_alloc(struct gsi_trans_pool *pool, u32 count) |
| 204 | { |
| 205 | return pool->base + gsi_trans_pool_alloc_common(pool, count); |
| 206 | } |
| 207 | |
| 208 | /* Allocate a single zeroed entry from a DMA pool */ |
| 209 | void *gsi_trans_pool_alloc_dma(struct gsi_trans_pool *pool, dma_addr_t *addr) |
| 210 | { |
| 211 | u32 offset = gsi_trans_pool_alloc_common(pool, count: 1); |
| 212 | |
| 213 | *addr = pool->addr + offset; |
| 214 | |
| 215 | return pool->base + offset; |
| 216 | } |
| 217 | |
| 218 | /* Map a TRE ring entry index to the transaction it is associated with */ |
| 219 | static void gsi_trans_map(struct gsi_trans *trans, u32 index) |
| 220 | { |
| 221 | struct gsi_channel *channel = &trans->gsi->channel[trans->channel_id]; |
| 222 | |
| 223 | /* The completion event will indicate the last TRE used */ |
| 224 | index += trans->used_count - 1; |
| 225 | |
| 226 | /* Note: index *must* be used modulo the ring count here */ |
| 227 | channel->trans_info.map[index % channel->tre_ring.count] = trans; |
| 228 | } |
| 229 | |
| 230 | /* Return the transaction mapped to a given ring entry */ |
| 231 | struct gsi_trans * |
| 232 | gsi_channel_trans_mapped(struct gsi_channel *channel, u32 index) |
| 233 | { |
| 234 | /* Note: index *must* be used modulo the ring count here */ |
| 235 | return channel->trans_info.map[index % channel->tre_ring.count]; |
| 236 | } |
| 237 | |
| 238 | /* Return the oldest completed transaction for a channel (or null) */ |
| 239 | struct gsi_trans *gsi_channel_trans_complete(struct gsi_channel *channel) |
| 240 | { |
| 241 | struct gsi_trans_info *trans_info = &channel->trans_info; |
| 242 | u16 trans_id = trans_info->completed_id; |
| 243 | |
| 244 | if (trans_id == trans_info->pending_id) { |
| 245 | gsi_channel_update(channel); |
| 246 | if (trans_id == trans_info->pending_id) |
| 247 | return NULL; |
| 248 | } |
| 249 | |
| 250 | return &trans_info->trans[trans_id %= channel->tre_count]; |
| 251 | } |
| 252 | |
| 253 | /* Move a transaction from allocated to committed state */ |
| 254 | static void gsi_trans_move_committed(struct gsi_trans *trans) |
| 255 | { |
| 256 | struct gsi_channel *channel = &trans->gsi->channel[trans->channel_id]; |
| 257 | struct gsi_trans_info *trans_info = &channel->trans_info; |
| 258 | |
| 259 | /* This allocated transaction is now committed */ |
| 260 | trans_info->allocated_id++; |
| 261 | } |
| 262 | |
| 263 | /* Move committed transactions to pending state */ |
| 264 | static void gsi_trans_move_pending(struct gsi_trans *trans) |
| 265 | { |
| 266 | struct gsi_channel *channel = &trans->gsi->channel[trans->channel_id]; |
| 267 | struct gsi_trans_info *trans_info = &channel->trans_info; |
| 268 | u16 trans_index = trans - &trans_info->trans[0]; |
| 269 | u16 delta; |
| 270 | |
| 271 | /* These committed transactions are now pending */ |
| 272 | delta = trans_index - trans_info->committed_id + 1; |
| 273 | trans_info->committed_id += delta % channel->tre_count; |
| 274 | } |
| 275 | |
| 276 | /* Move pending transactions to completed state */ |
| 277 | void gsi_trans_move_complete(struct gsi_trans *trans) |
| 278 | { |
| 279 | struct gsi_channel *channel = &trans->gsi->channel[trans->channel_id]; |
| 280 | struct gsi_trans_info *trans_info = &channel->trans_info; |
| 281 | u16 trans_index = trans - trans_info->trans; |
| 282 | u16 delta; |
| 283 | |
| 284 | /* These pending transactions are now completed */ |
| 285 | delta = trans_index - trans_info->pending_id + 1; |
| 286 | delta %= channel->tre_count; |
| 287 | trans_info->pending_id += delta; |
| 288 | } |
| 289 | |
| 290 | /* Move a transaction from completed to polled state */ |
| 291 | void gsi_trans_move_polled(struct gsi_trans *trans) |
| 292 | { |
| 293 | struct gsi_channel *channel = &trans->gsi->channel[trans->channel_id]; |
| 294 | struct gsi_trans_info *trans_info = &channel->trans_info; |
| 295 | |
| 296 | /* This completed transaction is now polled */ |
| 297 | trans_info->completed_id++; |
| 298 | } |
| 299 | |
| 300 | /* Reserve some number of TREs on a channel. Returns true if successful */ |
| 301 | static bool |
| 302 | gsi_trans_tre_reserve(struct gsi_trans_info *trans_info, u32 tre_count) |
| 303 | { |
| 304 | int avail = atomic_read(v: &trans_info->tre_avail); |
| 305 | int new; |
| 306 | |
| 307 | do { |
| 308 | new = avail - (int)tre_count; |
| 309 | if (unlikely(new < 0)) |
| 310 | return false; |
| 311 | } while (!atomic_try_cmpxchg(v: &trans_info->tre_avail, old: &avail, new)); |
| 312 | |
| 313 | return true; |
| 314 | } |
| 315 | |
| 316 | /* Release previously-reserved TRE entries to a channel */ |
| 317 | static void |
| 318 | gsi_trans_tre_release(struct gsi_trans_info *trans_info, u32 tre_count) |
| 319 | { |
| 320 | atomic_add(i: tre_count, v: &trans_info->tre_avail); |
| 321 | } |
| 322 | |
| 323 | /* Return true if no transactions are allocated, false otherwise */ |
| 324 | bool gsi_channel_trans_idle(struct gsi *gsi, u32 channel_id) |
| 325 | { |
| 326 | u32 tre_max = gsi_channel_tre_max(gsi, channel_id); |
| 327 | struct gsi_trans_info *trans_info; |
| 328 | |
| 329 | trans_info = &gsi->channel[channel_id].trans_info; |
| 330 | |
| 331 | return atomic_read(v: &trans_info->tre_avail) == tre_max; |
| 332 | } |
| 333 | |
| 334 | /* Allocate a GSI transaction on a channel */ |
| 335 | struct gsi_trans *gsi_channel_trans_alloc(struct gsi *gsi, u32 channel_id, |
| 336 | u32 tre_count, |
| 337 | enum dma_data_direction direction) |
| 338 | { |
| 339 | struct gsi_channel *channel = &gsi->channel[channel_id]; |
| 340 | struct gsi_trans_info *trans_info; |
| 341 | struct gsi_trans *trans; |
| 342 | u16 trans_index; |
| 343 | |
| 344 | if (WARN_ON(tre_count > channel->trans_tre_max)) |
| 345 | return NULL; |
| 346 | |
| 347 | trans_info = &channel->trans_info; |
| 348 | |
| 349 | /* If we can't reserve the TREs for the transaction, we're done */ |
| 350 | if (!gsi_trans_tre_reserve(trans_info, tre_count)) |
| 351 | return NULL; |
| 352 | |
| 353 | trans_index = trans_info->free_id % channel->tre_count; |
| 354 | trans = &trans_info->trans[trans_index]; |
| 355 | memset(trans, 0, sizeof(*trans)); |
| 356 | |
| 357 | /* Initialize non-zero fields in the transaction */ |
| 358 | trans->gsi = gsi; |
| 359 | trans->channel_id = channel_id; |
| 360 | trans->rsvd_count = tre_count; |
| 361 | init_completion(x: &trans->completion); |
| 362 | |
| 363 | /* Allocate the scatterlist */ |
| 364 | trans->sgl = gsi_trans_pool_alloc(pool: &trans_info->sg_pool, count: tre_count); |
| 365 | sg_init_marker(sgl: trans->sgl, nents: tre_count); |
| 366 | |
| 367 | trans->direction = direction; |
| 368 | refcount_set(r: &trans->refcount, n: 1); |
| 369 | |
| 370 | /* This free transaction is now allocated */ |
| 371 | trans_info->free_id++; |
| 372 | |
| 373 | return trans; |
| 374 | } |
| 375 | |
| 376 | /* Free a previously-allocated transaction */ |
| 377 | void gsi_trans_free(struct gsi_trans *trans) |
| 378 | { |
| 379 | struct gsi_trans_info *trans_info; |
| 380 | |
| 381 | if (!refcount_dec_and_test(r: &trans->refcount)) |
| 382 | return; |
| 383 | |
| 384 | /* Unused transactions are allocated but never committed, pending, |
| 385 | * completed, or polled. |
| 386 | */ |
| 387 | trans_info = &trans->gsi->channel[trans->channel_id].trans_info; |
| 388 | if (!trans->used_count) { |
| 389 | trans_info->allocated_id++; |
| 390 | trans_info->committed_id++; |
| 391 | trans_info->pending_id++; |
| 392 | trans_info->completed_id++; |
| 393 | } else { |
| 394 | ipa_gsi_trans_release(trans); |
| 395 | } |
| 396 | |
| 397 | /* This transaction is now free */ |
| 398 | trans_info->polled_id++; |
| 399 | |
| 400 | /* Releasing the reserved TREs implicitly frees the sgl[] and |
| 401 | * (if present) info[] arrays, plus the transaction itself. |
| 402 | */ |
| 403 | gsi_trans_tre_release(trans_info, tre_count: trans->rsvd_count); |
| 404 | } |
| 405 | |
| 406 | /* Add an immediate command to a transaction */ |
| 407 | void gsi_trans_cmd_add(struct gsi_trans *trans, void *buf, u32 size, |
| 408 | dma_addr_t addr, enum ipa_cmd_opcode opcode) |
| 409 | { |
| 410 | u32 which = trans->used_count++; |
| 411 | struct scatterlist *sg; |
| 412 | |
| 413 | WARN_ON(which >= trans->rsvd_count); |
| 414 | |
| 415 | /* Commands are quite different from data transfer requests. |
| 416 | * Their payloads come from a pool whose memory is allocated |
| 417 | * using dma_alloc_coherent(). We therefore do *not* map them |
| 418 | * for DMA (unlike what we do for pages and skbs). |
| 419 | * |
| 420 | * When a transaction completes, the SGL is normally unmapped. |
| 421 | * A command transaction has direction DMA_NONE, which tells |
| 422 | * gsi_trans_complete() to skip the unmapping step. |
| 423 | * |
| 424 | * The only things we use directly in a command scatter/gather |
| 425 | * entry are the DMA address and length. We still need the SG |
| 426 | * table flags to be maintained though, so assign a NULL page |
| 427 | * pointer for that purpose. |
| 428 | */ |
| 429 | sg = &trans->sgl[which]; |
| 430 | sg_assign_page(sg, NULL); |
| 431 | sg_dma_address(sg) = addr; |
| 432 | sg_dma_len(sg) = size; |
| 433 | |
| 434 | trans->cmd_opcode[which] = opcode; |
| 435 | } |
| 436 | |
| 437 | /* Add a page transfer to a transaction. It will fill the only TRE. */ |
| 438 | int gsi_trans_page_add(struct gsi_trans *trans, struct page *page, u32 size, |
| 439 | u32 offset) |
| 440 | { |
| 441 | struct scatterlist *sg = &trans->sgl[0]; |
| 442 | int ret; |
| 443 | |
| 444 | if (WARN_ON(trans->rsvd_count != 1)) |
| 445 | return -EINVAL; |
| 446 | if (WARN_ON(trans->used_count)) |
| 447 | return -EINVAL; |
| 448 | |
| 449 | sg_set_page(sg, page, len: size, offset); |
| 450 | ret = dma_map_sg(trans->gsi->dev, sg, 1, trans->direction); |
| 451 | if (!ret) |
| 452 | return -ENOMEM; |
| 453 | |
| 454 | trans->used_count++; /* Transaction now owns the (DMA mapped) page */ |
| 455 | |
| 456 | return 0; |
| 457 | } |
| 458 | |
| 459 | /* Add an SKB transfer to a transaction. No other TREs will be used. */ |
| 460 | int gsi_trans_skb_add(struct gsi_trans *trans, struct sk_buff *skb) |
| 461 | { |
| 462 | struct scatterlist *sg = &trans->sgl[0]; |
| 463 | u32 used_count; |
| 464 | int ret; |
| 465 | |
| 466 | if (WARN_ON(trans->rsvd_count != 1)) |
| 467 | return -EINVAL; |
| 468 | if (WARN_ON(trans->used_count)) |
| 469 | return -EINVAL; |
| 470 | |
| 471 | /* skb->len will not be 0 (checked early) */ |
| 472 | ret = skb_to_sgvec(skb, sg, offset: 0, len: skb->len); |
| 473 | if (ret < 0) |
| 474 | return ret; |
| 475 | used_count = ret; |
| 476 | |
| 477 | ret = dma_map_sg(trans->gsi->dev, sg, used_count, trans->direction); |
| 478 | if (!ret) |
| 479 | return -ENOMEM; |
| 480 | |
| 481 | /* Transaction now owns the (DMA mapped) skb */ |
| 482 | trans->used_count += used_count; |
| 483 | |
| 484 | return 0; |
| 485 | } |
| 486 | |
| 487 | /* Compute the length/opcode value to use for a TRE */ |
| 488 | static __le16 gsi_tre_len_opcode(enum ipa_cmd_opcode opcode, u32 len) |
| 489 | { |
| 490 | return opcode == IPA_CMD_NONE ? cpu_to_le16((u16)len) |
| 491 | : cpu_to_le16((u16)opcode); |
| 492 | } |
| 493 | |
| 494 | /* Compute the flags value to use for a given TRE */ |
| 495 | static __le32 gsi_tre_flags(bool last_tre, bool bei, enum ipa_cmd_opcode opcode) |
| 496 | { |
| 497 | enum gsi_tre_type tre_type; |
| 498 | u32 tre_flags; |
| 499 | |
| 500 | tre_type = opcode == IPA_CMD_NONE ? GSI_RE_XFER : GSI_RE_IMMD_CMD; |
| 501 | tre_flags = u32_encode_bits(v: tre_type, TRE_FLAGS_TYPE_FMASK); |
| 502 | |
| 503 | /* Last TRE contains interrupt flags */ |
| 504 | if (last_tre) { |
| 505 | /* All transactions end in a transfer completion interrupt */ |
| 506 | tre_flags |= TRE_FLAGS_IEOT_FMASK; |
| 507 | /* Don't interrupt when outbound commands are acknowledged */ |
| 508 | if (bei) |
| 509 | tre_flags |= TRE_FLAGS_BEI_FMASK; |
| 510 | } else { /* All others indicate there's more to come */ |
| 511 | tre_flags |= TRE_FLAGS_CHAIN_FMASK; |
| 512 | } |
| 513 | |
| 514 | return cpu_to_le32(tre_flags); |
| 515 | } |
| 516 | |
| 517 | static void gsi_trans_tre_fill(struct gsi_tre *dest_tre, dma_addr_t addr, |
| 518 | u32 len, bool last_tre, bool bei, |
| 519 | enum ipa_cmd_opcode opcode) |
| 520 | { |
| 521 | struct gsi_tre tre; |
| 522 | |
| 523 | tre.addr = cpu_to_le64(addr); |
| 524 | tre.len_opcode = gsi_tre_len_opcode(opcode, len); |
| 525 | tre.reserved = 0; |
| 526 | tre.flags = gsi_tre_flags(last_tre, bei, opcode); |
| 527 | |
| 528 | /* ARM64 can write 16 bytes as a unit with a single instruction. |
| 529 | * Doing the assignment this way is an attempt to make that happen. |
| 530 | */ |
| 531 | *dest_tre = tre; |
| 532 | } |
| 533 | |
| 534 | /** |
| 535 | * __gsi_trans_commit() - Common GSI transaction commit code |
| 536 | * @trans: Transaction to commit |
| 537 | * @ring_db: Whether to tell the hardware about these queued transfers |
| 538 | * |
| 539 | * Formats channel ring TRE entries based on the content of the scatterlist. |
| 540 | * Maps a transaction pointer to the last ring entry used for the transaction, |
| 541 | * so it can be recovered when it completes. Moves the transaction to |
| 542 | * pending state. Finally, updates the channel ring pointer and optionally |
| 543 | * rings the doorbell. |
| 544 | */ |
| 545 | static void __gsi_trans_commit(struct gsi_trans *trans, bool ring_db) |
| 546 | { |
| 547 | struct gsi_channel *channel = &trans->gsi->channel[trans->channel_id]; |
| 548 | struct gsi_ring *tre_ring = &channel->tre_ring; |
| 549 | enum ipa_cmd_opcode opcode = IPA_CMD_NONE; |
| 550 | bool bei = channel->toward_ipa; |
| 551 | struct gsi_tre *dest_tre; |
| 552 | struct scatterlist *sg; |
| 553 | u32 byte_count = 0; |
| 554 | u8 *cmd_opcode; |
| 555 | u32 avail; |
| 556 | u32 i; |
| 557 | |
| 558 | WARN_ON(!trans->used_count); |
| 559 | |
| 560 | /* Consume the entries. If we cross the end of the ring while |
| 561 | * filling them we'll switch to the beginning to finish. |
| 562 | * If there is no info array we're doing a simple data |
| 563 | * transfer request, whose opcode is IPA_CMD_NONE. |
| 564 | */ |
| 565 | cmd_opcode = channel->command ? &trans->cmd_opcode[0] : NULL; |
| 566 | avail = tre_ring->count - tre_ring->index % tre_ring->count; |
| 567 | dest_tre = gsi_ring_virt(ring: tre_ring, index: tre_ring->index); |
| 568 | for_each_sg(trans->sgl, sg, trans->used_count, i) { |
| 569 | bool last_tre = i == trans->used_count - 1; |
| 570 | dma_addr_t addr = sg_dma_address(sg); |
| 571 | u32 len = sg_dma_len(sg); |
| 572 | |
| 573 | byte_count += len; |
| 574 | if (!avail--) |
| 575 | dest_tre = gsi_ring_virt(ring: tre_ring, index: 0); |
| 576 | if (cmd_opcode) |
| 577 | opcode = *cmd_opcode++; |
| 578 | |
| 579 | gsi_trans_tre_fill(dest_tre, addr, len, last_tre, bei, opcode); |
| 580 | dest_tre++; |
| 581 | } |
| 582 | /* Associate the TRE with the transaction */ |
| 583 | gsi_trans_map(trans, index: tre_ring->index); |
| 584 | |
| 585 | tre_ring->index += trans->used_count; |
| 586 | |
| 587 | trans->len = byte_count; |
| 588 | if (channel->toward_ipa) |
| 589 | gsi_trans_tx_committed(trans); |
| 590 | |
| 591 | gsi_trans_move_committed(trans); |
| 592 | |
| 593 | /* Ring doorbell if requested, or if all TREs are allocated */ |
| 594 | if (ring_db || !atomic_read(v: &channel->trans_info.tre_avail)) { |
| 595 | /* Report what we're handing off to hardware for TX channels */ |
| 596 | if (channel->toward_ipa) |
| 597 | gsi_trans_tx_queued(trans); |
| 598 | gsi_trans_move_pending(trans); |
| 599 | gsi_channel_doorbell(channel); |
| 600 | } |
| 601 | } |
| 602 | |
| 603 | /* Commit a GSI transaction */ |
| 604 | void gsi_trans_commit(struct gsi_trans *trans, bool ring_db) |
| 605 | { |
| 606 | if (trans->used_count) |
| 607 | __gsi_trans_commit(trans, ring_db); |
| 608 | else |
| 609 | gsi_trans_free(trans); |
| 610 | } |
| 611 | |
| 612 | /* Commit a GSI transaction and wait for it to complete */ |
| 613 | void gsi_trans_commit_wait(struct gsi_trans *trans) |
| 614 | { |
| 615 | if (!trans->used_count) |
| 616 | goto out_trans_free; |
| 617 | |
| 618 | refcount_inc(r: &trans->refcount); |
| 619 | |
| 620 | __gsi_trans_commit(trans, ring_db: true); |
| 621 | |
| 622 | wait_for_completion(&trans->completion); |
| 623 | |
| 624 | out_trans_free: |
| 625 | gsi_trans_free(trans); |
| 626 | } |
| 627 | |
| 628 | /* Process the completion of a transaction; called while polling */ |
| 629 | void gsi_trans_complete(struct gsi_trans *trans) |
| 630 | { |
| 631 | /* If the entire SGL was mapped when added, unmap it now */ |
| 632 | if (trans->direction != DMA_NONE) |
| 633 | dma_unmap_sg(trans->gsi->dev, trans->sgl, trans->used_count, |
| 634 | trans->direction); |
| 635 | |
| 636 | ipa_gsi_trans_complete(trans); |
| 637 | |
| 638 | complete(&trans->completion); |
| 639 | |
| 640 | gsi_trans_free(trans); |
| 641 | } |
| 642 | |
| 643 | /* Cancel a channel's pending transactions */ |
| 644 | void gsi_channel_trans_cancel_pending(struct gsi_channel *channel) |
| 645 | { |
| 646 | struct gsi_trans_info *trans_info = &channel->trans_info; |
| 647 | u16 trans_id = trans_info->pending_id; |
| 648 | |
| 649 | /* channel->gsi->mutex is held by caller */ |
| 650 | |
| 651 | /* If there are no pending transactions, we're done */ |
| 652 | if (trans_id == trans_info->committed_id) |
| 653 | return; |
| 654 | |
| 655 | /* Mark all pending transactions cancelled */ |
| 656 | do { |
| 657 | struct gsi_trans *trans; |
| 658 | |
| 659 | trans = &trans_info->trans[trans_id % channel->tre_count]; |
| 660 | trans->cancelled = true; |
| 661 | } while (++trans_id != trans_info->committed_id); |
| 662 | |
| 663 | /* All pending transactions are now completed */ |
| 664 | trans_info->pending_id = trans_info->committed_id; |
| 665 | |
| 666 | /* Schedule NAPI polling to complete the cancelled transactions */ |
| 667 | napi_schedule(n: &channel->napi); |
| 668 | } |
| 669 | |
| 670 | /* Issue a command to read a single byte from a channel */ |
| 671 | int gsi_trans_read_byte(struct gsi *gsi, u32 channel_id, dma_addr_t addr) |
| 672 | { |
| 673 | struct gsi_channel *channel = &gsi->channel[channel_id]; |
| 674 | struct gsi_ring *tre_ring = &channel->tre_ring; |
| 675 | struct gsi_trans_info *trans_info; |
| 676 | struct gsi_tre *dest_tre; |
| 677 | |
| 678 | trans_info = &channel->trans_info; |
| 679 | |
| 680 | /* First reserve the TRE, if possible */ |
| 681 | if (!gsi_trans_tre_reserve(trans_info, tre_count: 1)) |
| 682 | return -EBUSY; |
| 683 | |
| 684 | /* Now fill the reserved TRE and tell the hardware */ |
| 685 | |
| 686 | dest_tre = gsi_ring_virt(ring: tre_ring, index: tre_ring->index); |
| 687 | gsi_trans_tre_fill(dest_tre, addr, len: 1, last_tre: true, bei: false, opcode: IPA_CMD_NONE); |
| 688 | |
| 689 | tre_ring->index++; |
| 690 | gsi_channel_doorbell(channel); |
| 691 | |
| 692 | return 0; |
| 693 | } |
| 694 | |
| 695 | /* Mark a gsi_trans_read_byte() request done */ |
| 696 | void gsi_trans_read_byte_done(struct gsi *gsi, u32 channel_id) |
| 697 | { |
| 698 | struct gsi_channel *channel = &gsi->channel[channel_id]; |
| 699 | |
| 700 | gsi_trans_tre_release(trans_info: &channel->trans_info, tre_count: 1); |
| 701 | } |
| 702 | |
| 703 | /* Initialize a channel's GSI transaction info */ |
| 704 | int gsi_channel_trans_init(struct gsi *gsi, u32 channel_id) |
| 705 | { |
| 706 | struct gsi_channel *channel = &gsi->channel[channel_id]; |
| 707 | u32 tre_count = channel->tre_count; |
| 708 | struct gsi_trans_info *trans_info; |
| 709 | u32 tre_max; |
| 710 | int ret; |
| 711 | |
| 712 | /* Ensure the size of a channel element is what's expected */ |
| 713 | BUILD_BUG_ON(sizeof(struct gsi_tre) != GSI_RING_ELEMENT_SIZE); |
| 714 | |
| 715 | trans_info = &channel->trans_info; |
| 716 | |
| 717 | /* The tre_avail field is what ultimately limits the number of |
| 718 | * outstanding transactions and their resources. A transaction |
| 719 | * allocation succeeds only if the TREs available are sufficient |
| 720 | * for what the transaction might need. |
| 721 | */ |
| 722 | tre_max = gsi_channel_tre_max(gsi: channel->gsi, channel_id); |
| 723 | atomic_set(v: &trans_info->tre_avail, i: tre_max); |
| 724 | |
| 725 | /* We can't use more TREs than the number available in the ring. |
| 726 | * This limits the number of transactions that can be outstanding. |
| 727 | * Worst case is one TRE per transaction (but we actually limit |
| 728 | * it to something a little less than that). By allocating a |
| 729 | * power-of-two number of transactions we can use an index |
| 730 | * modulo that number to determine the next one that's free. |
| 731 | * Transactions are allocated one at a time. |
| 732 | */ |
| 733 | trans_info->trans = kcalloc(tre_count, sizeof(*trans_info->trans), |
| 734 | GFP_KERNEL); |
| 735 | if (!trans_info->trans) |
| 736 | return -ENOMEM; |
| 737 | trans_info->free_id = 0; /* all modulo channel->tre_count */ |
| 738 | trans_info->allocated_id = 0; |
| 739 | trans_info->committed_id = 0; |
| 740 | trans_info->pending_id = 0; |
| 741 | trans_info->completed_id = 0; |
| 742 | trans_info->polled_id = 0; |
| 743 | |
| 744 | /* A completion event contains a pointer to the TRE that caused |
| 745 | * the event (which will be the last one used by the transaction). |
| 746 | * Each entry in this map records the transaction associated |
| 747 | * with a corresponding completed TRE. |
| 748 | */ |
| 749 | trans_info->map = kcalloc(tre_count, sizeof(*trans_info->map), |
| 750 | GFP_KERNEL); |
| 751 | if (!trans_info->map) { |
| 752 | ret = -ENOMEM; |
| 753 | goto err_trans_free; |
| 754 | } |
| 755 | |
| 756 | /* A transaction uses a scatterlist array to represent the data |
| 757 | * transfers implemented by the transaction. Each scatterlist |
| 758 | * element is used to fill a single TRE when the transaction is |
| 759 | * committed. So we need as many scatterlist elements as the |
| 760 | * maximum number of TREs that can be outstanding. |
| 761 | */ |
| 762 | ret = gsi_trans_pool_init(pool: &trans_info->sg_pool, |
| 763 | size: sizeof(struct scatterlist), |
| 764 | count: tre_max, max_alloc: channel->trans_tre_max); |
| 765 | if (ret) |
| 766 | goto err_map_free; |
| 767 | |
| 768 | |
| 769 | return 0; |
| 770 | |
| 771 | err_map_free: |
| 772 | kfree(objp: trans_info->map); |
| 773 | err_trans_free: |
| 774 | kfree(objp: trans_info->trans); |
| 775 | |
| 776 | dev_err(gsi->dev, "error %d initializing channel %u transactions\n" , |
| 777 | ret, channel_id); |
| 778 | |
| 779 | return ret; |
| 780 | } |
| 781 | |
| 782 | /* Inverse of gsi_channel_trans_init() */ |
| 783 | void gsi_channel_trans_exit(struct gsi_channel *channel) |
| 784 | { |
| 785 | struct gsi_trans_info *trans_info = &channel->trans_info; |
| 786 | |
| 787 | gsi_trans_pool_exit(pool: &trans_info->sg_pool); |
| 788 | kfree(objp: trans_info->trans); |
| 789 | kfree(objp: trans_info->map); |
| 790 | } |
| 791 | |