Stephen Warren | 49626ea | 2016-07-18 12:17:11 -0600 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2016, NVIDIA CORPORATION. |
| 3 | * |
| 4 | * SPDX-License-Identifier: GPL-2.0 |
| 5 | */ |
| 6 | |
| 7 | #include <common.h> |
| 8 | #include <asm/io.h> |
| 9 | #include <asm/arch-tegra/ivc.h> |
| 10 | |
| 11 | #define TEGRA_IVC_ALIGN 64 |
| 12 | |
| 13 | /* |
| 14 | * IVC channel reset protocol. |
| 15 | * |
| 16 | * Each end uses its tx_channel.state to indicate its synchronization state. |
| 17 | */ |
| 18 | enum ivc_state { |
| 19 | /* |
| 20 | * This value is zero for backwards compatibility with services that |
| 21 | * assume channels to be initially zeroed. Such channels are in an |
| 22 | * initially valid state, but cannot be asynchronously reset, and must |
| 23 | * maintain a valid state at all times. |
| 24 | * |
| 25 | * The transmitting end can enter the established state from the sync or |
| 26 | * ack state when it observes the receiving endpoint in the ack or |
| 27 | * established state, indicating that has cleared the counters in our |
| 28 | * rx_channel. |
| 29 | */ |
| 30 | ivc_state_established = 0, |
| 31 | |
| 32 | /* |
| 33 | * If an endpoint is observed in the sync state, the remote endpoint is |
| 34 | * allowed to clear the counters it owns asynchronously with respect to |
| 35 | * the current endpoint. Therefore, the current endpoint is no longer |
| 36 | * allowed to communicate. |
| 37 | */ |
| 38 | ivc_state_sync, |
| 39 | |
| 40 | /* |
| 41 | * When the transmitting end observes the receiving end in the sync |
| 42 | * state, it can clear the w_count and r_count and transition to the ack |
| 43 | * state. If the remote endpoint observes us in the ack state, it can |
| 44 | * return to the established state once it has cleared its counters. |
| 45 | */ |
| 46 | ivc_state_ack |
| 47 | }; |
| 48 | |
| 49 | /* |
| 50 | * This structure is divided into two-cache aligned parts, the first is only |
| 51 | * written through the tx_channel pointer, while the second is only written |
| 52 | * through the rx_channel pointer. This delineates ownership of the cache lines, |
| 53 | * which is critical to performance and necessary in non-cache coherent |
| 54 | * implementations. |
| 55 | */ |
| 56 | struct tegra_ivc_channel_header { |
| 57 | union { |
| 58 | /* fields owned by the transmitting end */ |
| 59 | struct { |
| 60 | uint32_t w_count; |
| 61 | uint32_t state; |
| 62 | }; |
| 63 | uint8_t w_align[TEGRA_IVC_ALIGN]; |
| 64 | }; |
| 65 | union { |
| 66 | /* fields owned by the receiving end */ |
| 67 | uint32_t r_count; |
| 68 | uint8_t r_align[TEGRA_IVC_ALIGN]; |
| 69 | }; |
| 70 | }; |
| 71 | |
| 72 | static inline void tegra_ivc_invalidate_counter(struct tegra_ivc *ivc, |
| 73 | struct tegra_ivc_channel_header *h, |
| 74 | ulong offset) |
| 75 | { |
| 76 | ulong base = ((ulong)h) + offset; |
| 77 | invalidate_dcache_range(base, base + TEGRA_IVC_ALIGN); |
| 78 | } |
| 79 | |
| 80 | static inline void tegra_ivc_flush_counter(struct tegra_ivc *ivc, |
| 81 | struct tegra_ivc_channel_header *h, |
| 82 | ulong offset) |
| 83 | { |
| 84 | ulong base = ((ulong)h) + offset; |
| 85 | flush_dcache_range(base, base + TEGRA_IVC_ALIGN); |
| 86 | } |
| 87 | |
| 88 | static inline ulong tegra_ivc_frame_addr(struct tegra_ivc *ivc, |
| 89 | struct tegra_ivc_channel_header *h, |
| 90 | uint32_t frame) |
| 91 | { |
| 92 | BUG_ON(frame >= ivc->nframes); |
| 93 | |
| 94 | return ((ulong)h) + sizeof(struct tegra_ivc_channel_header) + |
| 95 | (ivc->frame_size * frame); |
| 96 | } |
| 97 | |
| 98 | static inline void *tegra_ivc_frame_pointer(struct tegra_ivc *ivc, |
| 99 | struct tegra_ivc_channel_header *ch, |
| 100 | uint32_t frame) |
| 101 | { |
| 102 | return (void *)tegra_ivc_frame_addr(ivc, ch, frame); |
| 103 | } |
| 104 | |
| 105 | static inline void tegra_ivc_invalidate_frame(struct tegra_ivc *ivc, |
| 106 | struct tegra_ivc_channel_header *h, |
| 107 | unsigned frame) |
| 108 | { |
| 109 | ulong base = tegra_ivc_frame_addr(ivc, h, frame); |
| 110 | invalidate_dcache_range(base, base + ivc->frame_size); |
| 111 | } |
| 112 | |
| 113 | static inline void tegra_ivc_flush_frame(struct tegra_ivc *ivc, |
| 114 | struct tegra_ivc_channel_header *h, |
| 115 | unsigned frame) |
| 116 | { |
| 117 | ulong base = tegra_ivc_frame_addr(ivc, h, frame); |
| 118 | flush_dcache_range(base, base + ivc->frame_size); |
| 119 | } |
| 120 | |
| 121 | static inline int tegra_ivc_channel_empty(struct tegra_ivc *ivc, |
| 122 | struct tegra_ivc_channel_header *ch) |
| 123 | { |
| 124 | /* |
| 125 | * This function performs multiple checks on the same values with |
| 126 | * security implications, so create snapshots with ACCESS_ONCE() to |
| 127 | * ensure that these checks use the same values. |
| 128 | */ |
| 129 | uint32_t w_count = ACCESS_ONCE(ch->w_count); |
| 130 | uint32_t r_count = ACCESS_ONCE(ch->r_count); |
| 131 | |
| 132 | /* |
| 133 | * Perform an over-full check to prevent denial of service attacks where |
| 134 | * a server could be easily fooled into believing that there's an |
| 135 | * extremely large number of frames ready, since receivers are not |
| 136 | * expected to check for full or over-full conditions. |
| 137 | * |
| 138 | * Although the channel isn't empty, this is an invalid case caused by |
| 139 | * a potentially malicious peer, so returning empty is safer, because it |
| 140 | * gives the impression that the channel has gone silent. |
| 141 | */ |
| 142 | if (w_count - r_count > ivc->nframes) |
| 143 | return 1; |
| 144 | |
| 145 | return w_count == r_count; |
| 146 | } |
| 147 | |
| 148 | static inline int tegra_ivc_channel_full(struct tegra_ivc *ivc, |
| 149 | struct tegra_ivc_channel_header *ch) |
| 150 | { |
| 151 | /* |
| 152 | * Invalid cases where the counters indicate that the queue is over |
| 153 | * capacity also appear full. |
| 154 | */ |
| 155 | return (ACCESS_ONCE(ch->w_count) - ACCESS_ONCE(ch->r_count)) >= |
| 156 | ivc->nframes; |
| 157 | } |
| 158 | |
| 159 | static inline void tegra_ivc_advance_rx(struct tegra_ivc *ivc) |
| 160 | { |
| 161 | ACCESS_ONCE(ivc->rx_channel->r_count) = |
| 162 | ACCESS_ONCE(ivc->rx_channel->r_count) + 1; |
| 163 | |
| 164 | if (ivc->r_pos == ivc->nframes - 1) |
| 165 | ivc->r_pos = 0; |
| 166 | else |
| 167 | ivc->r_pos++; |
| 168 | } |
| 169 | |
| 170 | static inline void tegra_ivc_advance_tx(struct tegra_ivc *ivc) |
| 171 | { |
| 172 | ACCESS_ONCE(ivc->tx_channel->w_count) = |
| 173 | ACCESS_ONCE(ivc->tx_channel->w_count) + 1; |
| 174 | |
| 175 | if (ivc->w_pos == ivc->nframes - 1) |
| 176 | ivc->w_pos = 0; |
| 177 | else |
| 178 | ivc->w_pos++; |
| 179 | } |
| 180 | |
| 181 | static inline int tegra_ivc_check_read(struct tegra_ivc *ivc) |
| 182 | { |
| 183 | ulong offset; |
| 184 | |
| 185 | /* |
| 186 | * tx_channel->state is set locally, so it is not synchronized with |
| 187 | * state from the remote peer. The remote peer cannot reset its |
| 188 | * transmit counters until we've acknowledged its synchronization |
| 189 | * request, so no additional synchronization is required because an |
| 190 | * asynchronous transition of rx_channel->state to ivc_state_ack is not |
| 191 | * allowed. |
| 192 | */ |
| 193 | if (ivc->tx_channel->state != ivc_state_established) |
| 194 | return -ECONNRESET; |
| 195 | |
| 196 | /* |
| 197 | * Avoid unnecessary invalidations when performing repeated accesses to |
| 198 | * an IVC channel by checking the old queue pointers first. |
| 199 | * Synchronization is only necessary when these pointers indicate empty |
| 200 | * or full. |
| 201 | */ |
| 202 | if (!tegra_ivc_channel_empty(ivc, ivc->rx_channel)) |
| 203 | return 0; |
| 204 | |
| 205 | offset = offsetof(struct tegra_ivc_channel_header, w_count); |
| 206 | tegra_ivc_invalidate_counter(ivc, ivc->rx_channel, offset); |
| 207 | return tegra_ivc_channel_empty(ivc, ivc->rx_channel) ? -ENOMEM : 0; |
| 208 | } |
| 209 | |
| 210 | static inline int tegra_ivc_check_write(struct tegra_ivc *ivc) |
| 211 | { |
| 212 | ulong offset; |
| 213 | |
| 214 | if (ivc->tx_channel->state != ivc_state_established) |
| 215 | return -ECONNRESET; |
| 216 | |
| 217 | if (!tegra_ivc_channel_full(ivc, ivc->tx_channel)) |
| 218 | return 0; |
| 219 | |
| 220 | offset = offsetof(struct tegra_ivc_channel_header, r_count); |
| 221 | tegra_ivc_invalidate_counter(ivc, ivc->tx_channel, offset); |
| 222 | return tegra_ivc_channel_full(ivc, ivc->tx_channel) ? -ENOMEM : 0; |
| 223 | } |
| 224 | |
| 225 | static inline uint32_t tegra_ivc_channel_avail_count(struct tegra_ivc *ivc, |
| 226 | struct tegra_ivc_channel_header *ch) |
| 227 | { |
| 228 | /* |
| 229 | * This function isn't expected to be used in scenarios where an |
| 230 | * over-full situation can lead to denial of service attacks. See the |
| 231 | * comment in tegra_ivc_channel_empty() for an explanation about |
| 232 | * special over-full considerations. |
| 233 | */ |
| 234 | return ACCESS_ONCE(ch->w_count) - ACCESS_ONCE(ch->r_count); |
| 235 | } |
| 236 | |
| 237 | int tegra_ivc_read_get_next_frame(struct tegra_ivc *ivc, void **frame) |
| 238 | { |
| 239 | int result = tegra_ivc_check_read(ivc); |
| 240 | if (result < 0) |
| 241 | return result; |
| 242 | |
| 243 | /* |
| 244 | * Order observation of w_pos potentially indicating new data before |
| 245 | * data read. |
| 246 | */ |
| 247 | mb(); |
| 248 | |
| 249 | tegra_ivc_invalidate_frame(ivc, ivc->rx_channel, ivc->r_pos); |
| 250 | *frame = tegra_ivc_frame_pointer(ivc, ivc->rx_channel, ivc->r_pos); |
| 251 | |
| 252 | return 0; |
| 253 | } |
| 254 | |
| 255 | int tegra_ivc_read_advance(struct tegra_ivc *ivc) |
| 256 | { |
| 257 | ulong offset; |
| 258 | int result; |
| 259 | |
| 260 | /* |
| 261 | * No read barriers or synchronization here: the caller is expected to |
| 262 | * have already observed the channel non-empty. This check is just to |
| 263 | * catch programming errors. |
| 264 | */ |
| 265 | result = tegra_ivc_check_read(ivc); |
| 266 | if (result) |
| 267 | return result; |
| 268 | |
| 269 | tegra_ivc_advance_rx(ivc); |
| 270 | offset = offsetof(struct tegra_ivc_channel_header, r_count); |
| 271 | tegra_ivc_flush_counter(ivc, ivc->rx_channel, offset); |
| 272 | |
| 273 | /* |
| 274 | * Ensure our write to r_pos occurs before our read from w_pos. |
| 275 | */ |
| 276 | mb(); |
| 277 | |
| 278 | offset = offsetof(struct tegra_ivc_channel_header, w_count); |
| 279 | tegra_ivc_invalidate_counter(ivc, ivc->rx_channel, offset); |
| 280 | |
| 281 | if (tegra_ivc_channel_avail_count(ivc, ivc->rx_channel) == |
| 282 | ivc->nframes - 1) |
| 283 | ivc->notify(ivc); |
| 284 | |
| 285 | return 0; |
| 286 | } |
| 287 | |
| 288 | int tegra_ivc_write_get_next_frame(struct tegra_ivc *ivc, void **frame) |
| 289 | { |
| 290 | int result = tegra_ivc_check_write(ivc); |
| 291 | if (result) |
| 292 | return result; |
| 293 | |
| 294 | *frame = tegra_ivc_frame_pointer(ivc, ivc->tx_channel, ivc->w_pos); |
| 295 | |
| 296 | return 0; |
| 297 | } |
| 298 | |
| 299 | int tegra_ivc_write_advance(struct tegra_ivc *ivc) |
| 300 | { |
| 301 | ulong offset; |
| 302 | int result; |
| 303 | |
| 304 | result = tegra_ivc_check_write(ivc); |
| 305 | if (result) |
| 306 | return result; |
| 307 | |
| 308 | tegra_ivc_flush_frame(ivc, ivc->tx_channel, ivc->w_pos); |
| 309 | |
| 310 | /* |
| 311 | * Order any possible stores to the frame before update of w_pos. |
| 312 | */ |
| 313 | mb(); |
| 314 | |
| 315 | tegra_ivc_advance_tx(ivc); |
| 316 | offset = offsetof(struct tegra_ivc_channel_header, w_count); |
| 317 | tegra_ivc_flush_counter(ivc, ivc->tx_channel, offset); |
| 318 | |
| 319 | /* |
| 320 | * Ensure our write to w_pos occurs before our read from r_pos. |
| 321 | */ |
| 322 | mb(); |
| 323 | |
| 324 | offset = offsetof(struct tegra_ivc_channel_header, r_count); |
| 325 | tegra_ivc_invalidate_counter(ivc, ivc->tx_channel, offset); |
| 326 | |
| 327 | if (tegra_ivc_channel_avail_count(ivc, ivc->tx_channel) == 1) |
| 328 | ivc->notify(ivc); |
| 329 | |
| 330 | return 0; |
| 331 | } |
| 332 | |
| 333 | /* |
| 334 | * =============================================================== |
| 335 | * IVC State Transition Table - see tegra_ivc_channel_notified() |
| 336 | * =============================================================== |
| 337 | * |
| 338 | * local remote action |
| 339 | * ----- ------ ----------------------------------- |
| 340 | * SYNC EST <none> |
| 341 | * SYNC ACK reset counters; move to EST; notify |
| 342 | * SYNC SYNC reset counters; move to ACK; notify |
| 343 | * ACK EST move to EST; notify |
| 344 | * ACK ACK move to EST; notify |
| 345 | * ACK SYNC reset counters; move to ACK; notify |
| 346 | * EST EST <none> |
| 347 | * EST ACK <none> |
| 348 | * EST SYNC reset counters; move to ACK; notify |
| 349 | * |
| 350 | * =============================================================== |
| 351 | */ |
| 352 | int tegra_ivc_channel_notified(struct tegra_ivc *ivc) |
| 353 | { |
| 354 | ulong offset; |
| 355 | enum ivc_state peer_state; |
| 356 | |
| 357 | /* Copy the receiver's state out of shared memory. */ |
| 358 | offset = offsetof(struct tegra_ivc_channel_header, w_count); |
| 359 | tegra_ivc_invalidate_counter(ivc, ivc->rx_channel, offset); |
| 360 | peer_state = ACCESS_ONCE(ivc->rx_channel->state); |
| 361 | |
| 362 | if (peer_state == ivc_state_sync) { |
| 363 | /* |
| 364 | * Order observation of ivc_state_sync before stores clearing |
| 365 | * tx_channel. |
| 366 | */ |
| 367 | mb(); |
| 368 | |
| 369 | /* |
| 370 | * Reset tx_channel counters. The remote end is in the SYNC |
| 371 | * state and won't make progress until we change our state, |
| 372 | * so the counters are not in use at this time. |
| 373 | */ |
| 374 | ivc->tx_channel->w_count = 0; |
| 375 | ivc->rx_channel->r_count = 0; |
| 376 | |
| 377 | ivc->w_pos = 0; |
| 378 | ivc->r_pos = 0; |
| 379 | |
| 380 | /* |
| 381 | * Ensure that counters appear cleared before new state can be |
| 382 | * observed. |
| 383 | */ |
| 384 | mb(); |
| 385 | |
| 386 | /* |
| 387 | * Move to ACK state. We have just cleared our counters, so it |
| 388 | * is now safe for the remote end to start using these values. |
| 389 | */ |
| 390 | ivc->tx_channel->state = ivc_state_ack; |
| 391 | offset = offsetof(struct tegra_ivc_channel_header, w_count); |
| 392 | tegra_ivc_flush_counter(ivc, ivc->tx_channel, offset); |
| 393 | |
| 394 | /* |
| 395 | * Notify remote end to observe state transition. |
| 396 | */ |
| 397 | ivc->notify(ivc); |
| 398 | } else if (ivc->tx_channel->state == ivc_state_sync && |
| 399 | peer_state == ivc_state_ack) { |
| 400 | /* |
| 401 | * Order observation of ivc_state_sync before stores clearing |
| 402 | * tx_channel. |
| 403 | */ |
| 404 | mb(); |
| 405 | |
| 406 | /* |
| 407 | * Reset tx_channel counters. The remote end is in the ACK |
| 408 | * state and won't make progress until we change our state, |
| 409 | * so the counters are not in use at this time. |
| 410 | */ |
| 411 | ivc->tx_channel->w_count = 0; |
| 412 | ivc->rx_channel->r_count = 0; |
| 413 | |
| 414 | ivc->w_pos = 0; |
| 415 | ivc->r_pos = 0; |
| 416 | |
| 417 | /* |
| 418 | * Ensure that counters appear cleared before new state can be |
| 419 | * observed. |
| 420 | */ |
| 421 | mb(); |
| 422 | |
| 423 | /* |
| 424 | * Move to ESTABLISHED state. We know that the remote end has |
| 425 | * already cleared its counters, so it is safe to start |
| 426 | * writing/reading on this channel. |
| 427 | */ |
| 428 | ivc->tx_channel->state = ivc_state_established; |
| 429 | offset = offsetof(struct tegra_ivc_channel_header, w_count); |
| 430 | tegra_ivc_flush_counter(ivc, ivc->tx_channel, offset); |
| 431 | |
| 432 | /* |
| 433 | * Notify remote end to observe state transition. |
| 434 | */ |
| 435 | ivc->notify(ivc); |
| 436 | } else if (ivc->tx_channel->state == ivc_state_ack) { |
| 437 | /* |
| 438 | * At this point, we have observed the peer to be in either |
| 439 | * the ACK or ESTABLISHED state. Next, order observation of |
| 440 | * peer state before storing to tx_channel. |
| 441 | */ |
| 442 | mb(); |
| 443 | |
| 444 | /* |
| 445 | * Move to ESTABLISHED state. We know that we have previously |
| 446 | * cleared our counters, and we know that the remote end has |
| 447 | * cleared its counters, so it is safe to start writing/reading |
| 448 | * on this channel. |
| 449 | */ |
| 450 | ivc->tx_channel->state = ivc_state_established; |
| 451 | offset = offsetof(struct tegra_ivc_channel_header, w_count); |
| 452 | tegra_ivc_flush_counter(ivc, ivc->tx_channel, offset); |
| 453 | |
| 454 | /* |
| 455 | * Notify remote end to observe state transition. |
| 456 | */ |
| 457 | ivc->notify(ivc); |
| 458 | } else { |
| 459 | /* |
| 460 | * There is no need to handle any further action. Either the |
| 461 | * channel is already fully established, or we are waiting for |
| 462 | * the remote end to catch up with our current state. Refer |
| 463 | * to the diagram in "IVC State Transition Table" above. |
| 464 | */ |
| 465 | } |
| 466 | |
| 467 | if (ivc->tx_channel->state != ivc_state_established) |
| 468 | return -EAGAIN; |
| 469 | |
| 470 | return 0; |
| 471 | } |
| 472 | |
| 473 | void tegra_ivc_channel_reset(struct tegra_ivc *ivc) |
| 474 | { |
| 475 | ulong offset; |
| 476 | |
| 477 | ivc->tx_channel->state = ivc_state_sync; |
| 478 | offset = offsetof(struct tegra_ivc_channel_header, w_count); |
| 479 | tegra_ivc_flush_counter(ivc, ivc->tx_channel, offset); |
| 480 | ivc->notify(ivc); |
| 481 | } |
| 482 | |
| 483 | static int check_ivc_params(ulong qbase1, ulong qbase2, uint32_t nframes, |
| 484 | uint32_t frame_size) |
| 485 | { |
| 486 | int ret = 0; |
| 487 | |
| 488 | BUG_ON(offsetof(struct tegra_ivc_channel_header, w_count) & |
| 489 | (TEGRA_IVC_ALIGN - 1)); |
| 490 | BUG_ON(offsetof(struct tegra_ivc_channel_header, r_count) & |
| 491 | (TEGRA_IVC_ALIGN - 1)); |
| 492 | BUG_ON(sizeof(struct tegra_ivc_channel_header) & |
| 493 | (TEGRA_IVC_ALIGN - 1)); |
| 494 | |
| 495 | if ((uint64_t)nframes * (uint64_t)frame_size >= 0x100000000) { |
| 496 | error("tegra_ivc: nframes * frame_size overflows\n"); |
| 497 | return -EINVAL; |
| 498 | } |
| 499 | |
| 500 | /* |
| 501 | * The headers must at least be aligned enough for counters |
| 502 | * to be accessed atomically. |
| 503 | */ |
| 504 | if ((qbase1 & (TEGRA_IVC_ALIGN - 1)) || |
| 505 | (qbase2 & (TEGRA_IVC_ALIGN - 1))) { |
| 506 | error("tegra_ivc: channel start not aligned\n"); |
| 507 | return -EINVAL; |
| 508 | } |
| 509 | |
| 510 | if (frame_size & (TEGRA_IVC_ALIGN - 1)) { |
| 511 | error("tegra_ivc: frame size not adequately aligned\n"); |
| 512 | return -EINVAL; |
| 513 | } |
| 514 | |
| 515 | if (qbase1 < qbase2) { |
| 516 | if (qbase1 + frame_size * nframes > qbase2) |
| 517 | ret = -EINVAL; |
| 518 | } else { |
| 519 | if (qbase2 + frame_size * nframes > qbase1) |
| 520 | ret = -EINVAL; |
| 521 | } |
| 522 | |
| 523 | if (ret) { |
| 524 | error("tegra_ivc: queue regions overlap\n"); |
| 525 | return ret; |
| 526 | } |
| 527 | |
| 528 | return 0; |
| 529 | } |
| 530 | |
| 531 | int tegra_ivc_init(struct tegra_ivc *ivc, ulong rx_base, ulong tx_base, |
| 532 | uint32_t nframes, uint32_t frame_size, |
| 533 | void (*notify)(struct tegra_ivc *)) |
| 534 | { |
| 535 | int ret; |
| 536 | |
| 537 | if (!ivc) |
| 538 | return -EINVAL; |
| 539 | |
| 540 | ret = check_ivc_params(rx_base, tx_base, nframes, frame_size); |
| 541 | if (ret) |
| 542 | return ret; |
| 543 | |
| 544 | ivc->rx_channel = (struct tegra_ivc_channel_header *)rx_base; |
| 545 | ivc->tx_channel = (struct tegra_ivc_channel_header *)tx_base; |
| 546 | ivc->w_pos = 0; |
| 547 | ivc->r_pos = 0; |
| 548 | ivc->nframes = nframes; |
| 549 | ivc->frame_size = frame_size; |
| 550 | ivc->notify = notify; |
| 551 | |
| 552 | return 0; |
| 553 | } |