| 1 | /* |
| 2 | * Multifd qpl compression accelerator implementation |
| 3 | * |
| 4 | * Copyright (c) 2023 Intel Corporation |
| 5 | * |
| 6 | * Authors: |
| 7 | * Yuan Liu<yuan1.liu@intel.com> |
| 8 | * |
| 9 | * This work is licensed under the terms of the GNU GPL, version 2 or later. |
| 10 | * See the COPYING file in the top-level directory. |
| 11 | */ |
| 12 | |
| 13 | #include "qemu/osdep.h" |
| 14 | #include "qemu/module.h" |
| 15 | #include "qapi/error.h" |
| 16 | #include "qapi/qapi-types-migration.h" |
| 17 | #include "system/ramblock.h" |
| 18 | #include "multifd.h" |
| 19 | #include "qpl/qpl.h" |
| 20 | |
| 21 | /* Maximum number of retries to resubmit a job if IAA work queues are full */ |
| 22 | #define MAX_SUBMIT_RETRY_NUM (3) |
| 23 | |
| 24 | typedef struct { |
| 25 | /* the QPL hardware path job */ |
| 26 | qpl_job *job; |
| 27 | /* indicates if fallback to software path is required */ |
| 28 | bool fallback_sw_path; |
| 29 | /* output data from the software path */ |
| 30 | uint8_t *sw_output; |
| 31 | /* output data length from the software path */ |
| 32 | uint32_t sw_output_len; |
| 33 | } QplHwJob; |
| 34 | |
| 35 | typedef struct { |
| 36 | /* array of hardware jobs, the number of jobs equals the number pages */ |
| 37 | QplHwJob *hw_jobs; |
| 38 | /* the QPL software job for the slow path and software fallback */ |
| 39 | qpl_job *sw_job; |
| 40 | /* the number of pages that the QPL needs to process at one time */ |
| 41 | uint32_t page_num; |
| 42 | /* array of compressed page buffers */ |
| 43 | uint8_t *zbuf; |
| 44 | /* array of compressed page lengths */ |
| 45 | uint32_t *zlen; |
| 46 | /* the status of the hardware device */ |
| 47 | bool hw_avail; |
| 48 | } QplData; |
| 49 | |
| 50 | /** |
| 51 | * check_hw_avail: check if IAA hardware is available |
| 52 | * |
| 53 | * If the IAA hardware does not exist or is unavailable, |
| 54 | * the QPL hardware job initialization will fail. |
| 55 | * |
| 56 | * Returns true if IAA hardware is available, otherwise false. |
| 57 | * |
| 58 | * @job_size: indicates the hardware job size if hardware is available |
| 59 | */ |
| 60 | static bool check_hw_avail(uint32_t *job_size) |
| 61 | { |
| 62 | qpl_path_t path = qpl_path_hardware; |
| 63 | uint32_t size = 0; |
| 64 | qpl_job *job; |
| 65 | |
| 66 | if (qpl_get_job_size(path, &size) != QPL_STS_OK) { |
| 67 | return false; |
| 68 | } |
| 69 | assert(size > 0); |
| 70 | job = g_malloc0(size); |
| 71 | if (qpl_init_job(path, job) != QPL_STS_OK) { |
| 72 | g_free(job); |
| 73 | return false; |
| 74 | } |
| 75 | g_free(job); |
| 76 | *job_size = size; |
| 77 | return true; |
| 78 | } |
| 79 | |
| 80 | /** |
| 81 | * multifd_qpl_free_sw_job: clean up software job |
| 82 | * |
| 83 | * Free the software job resources. |
| 84 | * |
| 85 | * @qpl: pointer to the QplData structure |
| 86 | */ |
| 87 | static void multifd_qpl_free_sw_job(QplData *qpl) |
| 88 | { |
| 89 | assert(qpl); |
| 90 | if (qpl->sw_job) { |
| 91 | qpl_fini_job(qpl->sw_job); |
| 92 | g_free(qpl->sw_job); |
| 93 | qpl->sw_job = NULL; |
| 94 | } |
| 95 | } |
| 96 | |
| 97 | /** |
| 98 | * multifd_qpl_free_jobs: clean up hardware jobs |
| 99 | * |
| 100 | * Free all hardware job resources. |
| 101 | * |
| 102 | * @qpl: pointer to the QplData structure |
| 103 | */ |
| 104 | static void multifd_qpl_free_hw_job(QplData *qpl) |
| 105 | { |
| 106 | assert(qpl); |
| 107 | if (qpl->hw_jobs) { |
| 108 | for (int i = 0; i < qpl->page_num; i++) { |
| 109 | qpl_fini_job(qpl->hw_jobs[i].job); |
| 110 | g_free(qpl->hw_jobs[i].job); |
| 111 | qpl->hw_jobs[i].job = NULL; |
| 112 | } |
| 113 | g_free(qpl->hw_jobs); |
| 114 | qpl->hw_jobs = NULL; |
| 115 | } |
| 116 | } |
| 117 | |
| 118 | /** |
| 119 | * multifd_qpl_init_sw_job: initialize a software job |
| 120 | * |
| 121 | * Use the QPL software path to initialize a job |
| 122 | * |
| 123 | * @qpl: pointer to the QplData structure |
| 124 | * @errp: pointer to an error |
| 125 | */ |
| 126 | static int multifd_qpl_init_sw_job(QplData *qpl, Error **errp) |
| 127 | { |
| 128 | qpl_path_t path = qpl_path_software; |
| 129 | uint32_t size = 0; |
| 130 | qpl_job *job = NULL; |
| 131 | qpl_status status; |
| 132 | |
| 133 | status = qpl_get_job_size(path, &size); |
| 134 | if (status != QPL_STS_OK) { |
| 135 | error_setg(errp, "qpl_get_job_size failed with error %d", status); |
| 136 | return -1; |
| 137 | } |
| 138 | job = g_malloc0(size); |
| 139 | status = qpl_init_job(path, job); |
| 140 | if (status != QPL_STS_OK) { |
| 141 | error_setg(errp, "qpl_init_job failed with error %d", status); |
| 142 | g_free(job); |
| 143 | return -1; |
| 144 | } |
| 145 | qpl->sw_job = job; |
| 146 | return 0; |
| 147 | } |
| 148 | |
| 149 | /** |
| 150 | * multifd_qpl_init_jobs: initialize hardware jobs |
| 151 | * |
| 152 | * Use the QPL hardware path to initialize jobs |
| 153 | * |
| 154 | * @qpl: pointer to the QplData structure |
| 155 | * @size: the size of QPL hardware path job |
| 156 | * @errp: pointer to an error |
| 157 | */ |
| 158 | static void multifd_qpl_init_hw_job(QplData *qpl, uint32_t size, Error **errp) |
| 159 | { |
| 160 | qpl_path_t path = qpl_path_hardware; |
| 161 | qpl_job *job = NULL; |
| 162 | qpl_status status; |
| 163 | |
| 164 | qpl->hw_jobs = g_new0(QplHwJob, qpl->page_num); |
| 165 | for (int i = 0; i < qpl->page_num; i++) { |
| 166 | job = g_malloc0(size); |
| 167 | status = qpl_init_job(path, job); |
| 168 | /* the job initialization should succeed after check_hw_avail */ |
| 169 | assert(status == QPL_STS_OK); |
| 170 | qpl->hw_jobs[i].job = job; |
| 171 | } |
| 172 | } |
| 173 | |
| 174 | /** |
| 175 | * multifd_qpl_init: initialize QplData structure |
| 176 | * |
| 177 | * Allocate and initialize a QplData structure |
| 178 | * |
| 179 | * Returns a QplData pointer on success or NULL on error |
| 180 | * |
| 181 | * @num: the number of pages |
| 182 | * @size: the page size |
| 183 | * @errp: pointer to an error |
| 184 | */ |
| 185 | static QplData *multifd_qpl_init(uint32_t num, uint32_t size, Error **errp) |
| 186 | { |
| 187 | uint32_t job_size = 0; |
| 188 | QplData *qpl; |
| 189 | |
| 190 | qpl = g_new0(QplData, 1); |
| 191 | qpl->page_num = num; |
| 192 | if (multifd_qpl_init_sw_job(qpl, errp) != 0) { |
| 193 | g_free(qpl); |
| 194 | return NULL; |
| 195 | } |
| 196 | qpl->hw_avail = check_hw_avail(&job_size); |
| 197 | if (qpl->hw_avail) { |
| 198 | multifd_qpl_init_hw_job(qpl, job_size, errp); |
| 199 | } |
| 200 | qpl->zbuf = g_malloc0(size * num); |
| 201 | qpl->zlen = g_new0(uint32_t, num); |
| 202 | return qpl; |
| 203 | } |
| 204 | |
| 205 | /** |
| 206 | * multifd_qpl_deinit: clean up QplData structure |
| 207 | * |
| 208 | * Free jobs, buffers and the QplData structure |
| 209 | * |
| 210 | * @qpl: pointer to the QplData structure |
| 211 | */ |
| 212 | static void multifd_qpl_deinit(QplData *qpl) |
| 213 | { |
| 214 | if (qpl) { |
| 215 | multifd_qpl_free_sw_job(qpl); |
| 216 | multifd_qpl_free_hw_job(qpl); |
| 217 | g_free(qpl->zbuf); |
| 218 | g_free(qpl->zlen); |
| 219 | g_free(qpl); |
| 220 | } |
| 221 | } |
| 222 | |
| 223 | static int multifd_qpl_send_setup(MultiFDSendParams *p, Error **errp) |
| 224 | { |
| 225 | QplData *qpl; |
| 226 | uint32_t page_size = multifd_ram_page_size(); |
| 227 | uint32_t page_count = multifd_ram_page_count(); |
| 228 | |
| 229 | qpl = multifd_qpl_init(page_count, page_size, errp); |
| 230 | if (!qpl) { |
| 231 | return -1; |
| 232 | } |
| 233 | p->compress_data = qpl; |
| 234 | |
| 235 | /* |
| 236 | * the page will be compressed independently and sent using an IOV. The |
| 237 | * additional two IOVs are used to store packet header and compressed data |
| 238 | * length |
| 239 | */ |
| 240 | p->iov = g_new0(struct iovec, page_count + 2); |
| 241 | return 0; |
| 242 | } |
| 243 | |
| 244 | static void multifd_qpl_send_cleanup(MultiFDSendParams *p, Error **errp) |
| 245 | { |
| 246 | multifd_qpl_deinit(p->compress_data); |
| 247 | p->compress_data = NULL; |
| 248 | g_free(p->iov); |
| 249 | p->iov = NULL; |
| 250 | } |
| 251 | |
| 252 | /** |
| 253 | * multifd_qpl_prepare_job: prepare the job |
| 254 | * |
| 255 | * Set the QPL job parameters and properties. |
| 256 | * |
| 257 | * @job: pointer to the qpl_job structure |
| 258 | * @is_compression: indicates compression and decompression |
| 259 | * @input: pointer to the input data buffer |
| 260 | * @input_len: the length of the input data |
| 261 | * @output: pointer to the output data buffer |
| 262 | * @output_len: the length of the output data |
| 263 | */ |
| 264 | static void multifd_qpl_prepare_job(qpl_job *job, bool is_compression, |
| 265 | uint8_t *input, uint32_t input_len, |
| 266 | uint8_t *output, uint32_t output_len) |
| 267 | { |
| 268 | job->op = is_compression ? qpl_op_compress : qpl_op_decompress; |
| 269 | job->next_in_ptr = input; |
| 270 | job->next_out_ptr = output; |
| 271 | job->available_in = input_len; |
| 272 | job->available_out = output_len; |
| 273 | job->flags = QPL_FLAG_FIRST | QPL_FLAG_LAST | QPL_FLAG_OMIT_VERIFY; |
| 274 | /* only supports compression level 1 */ |
| 275 | job->level = 1; |
| 276 | } |
| 277 | |
| 278 | /** |
| 279 | * multifd_qpl_prepare_comp_job: prepare the compression job |
| 280 | * |
| 281 | * Set the compression job parameters and properties. |
| 282 | * |
| 283 | * @job: pointer to the qpl_job structure |
| 284 | * @input: pointer to the input data buffer |
| 285 | * @output: pointer to the output data buffer |
| 286 | * @size: the page size |
| 287 | */ |
| 288 | static void multifd_qpl_prepare_comp_job(qpl_job *job, uint8_t *input, |
| 289 | uint8_t *output, uint32_t size) |
| 290 | { |
| 291 | /* |
| 292 | * Set output length to less than the page size to force the job to |
| 293 | * fail in case it compresses to a larger size. We'll send that page |
| 294 | * without compression and skip the decompression operation on the |
| 295 | * destination. |
| 296 | */ |
| 297 | multifd_qpl_prepare_job(job, true, input, size, output, size - 1); |
| 298 | } |
| 299 | |
| 300 | /** |
| 301 | * multifd_qpl_prepare_decomp_job: prepare the decompression job |
| 302 | * |
| 303 | * Set the decompression job parameters and properties. |
| 304 | * |
| 305 | * @job: pointer to the qpl_job structure |
| 306 | * @input: pointer to the input data buffer |
| 307 | * @len: the length of the input data |
| 308 | * @output: pointer to the output data buffer |
| 309 | * @size: the page size |
| 310 | */ |
| 311 | static void multifd_qpl_prepare_decomp_job(qpl_job *job, uint8_t *input, |
| 312 | uint32_t len, uint8_t *output, |
| 313 | uint32_t size) |
| 314 | { |
| 315 | multifd_qpl_prepare_job(job, false, input, len, output, size); |
| 316 | } |
| 317 | |
| 318 | /** |
| 319 | * multifd_qpl_fill_iov: fill in the IOV |
| 320 | * |
| 321 | * Fill in the QPL packet IOV |
| 322 | * |
| 323 | * @p: Params for the channel being used |
| 324 | * @data: pointer to the IOV data |
| 325 | * @len: The length of the IOV data |
| 326 | */ |
| 327 | static void multifd_qpl_fill_iov(MultiFDSendParams *p, uint8_t *data, |
| 328 | uint32_t len) |
| 329 | { |
| 330 | p->iov[p->iovs_num].iov_base = data; |
| 331 | p->iov[p->iovs_num].iov_len = len; |
| 332 | p->iovs_num++; |
| 333 | p->next_packet_size += len; |
| 334 | } |
| 335 | |
| 336 | /** |
| 337 | * multifd_qpl_fill_packet: fill the compressed page into the QPL packet |
| 338 | * |
| 339 | * Fill the compressed page length and IOV into the QPL packet |
| 340 | * |
| 341 | * @idx: The index of the compressed length array |
| 342 | * @p: Params for the channel being used |
| 343 | * @data: pointer to the compressed page buffer |
| 344 | * @len: The length of the compressed page |
| 345 | */ |
| 346 | static void multifd_qpl_fill_packet(uint32_t idx, MultiFDSendParams *p, |
| 347 | uint8_t *data, uint32_t len) |
| 348 | { |
| 349 | QplData *qpl = p->compress_data; |
| 350 | |
| 351 | qpl->zlen[idx] = cpu_to_be32(len); |
| 352 | multifd_qpl_fill_iov(p, data, len); |
| 353 | } |
| 354 | |
| 355 | /** |
| 356 | * multifd_qpl_submit_job: submit a job to the hardware |
| 357 | * |
| 358 | * Submit a QPL hardware job to the IAA device |
| 359 | * |
| 360 | * Returns true if the job is submitted successfully, otherwise false. |
| 361 | * |
| 362 | * @job: pointer to the qpl_job structure |
| 363 | */ |
| 364 | static bool multifd_qpl_submit_job(qpl_job *job) |
| 365 | { |
| 366 | qpl_status status; |
| 367 | uint32_t num = 0; |
| 368 | |
| 369 | retry: |
| 370 | status = qpl_submit_job(job); |
| 371 | if (status == QPL_STS_QUEUES_ARE_BUSY_ERR) { |
| 372 | if (num < MAX_SUBMIT_RETRY_NUM) { |
| 373 | num++; |
| 374 | goto retry; |
| 375 | } |
| 376 | } |
| 377 | return (status == QPL_STS_OK); |
| 378 | } |
| 379 | |
| 380 | /** |
| 381 | * multifd_qpl_compress_pages_slow_path: compress pages using slow path |
| 382 | * |
| 383 | * Compress the pages using software. If compression fails, the uncompressed |
| 384 | * page will be sent. |
| 385 | * |
| 386 | * @p: Params for the channel being used |
| 387 | */ |
| 388 | static void multifd_qpl_compress_pages_slow_path(MultiFDSendParams *p) |
| 389 | { |
| 390 | QplData *qpl = p->compress_data; |
| 391 | MultiFDPages_t *pages = &p->data->u.ram; |
| 392 | uint32_t size = multifd_ram_page_size(); |
| 393 | qpl_job *job = qpl->sw_job; |
| 394 | uint8_t *zbuf = qpl->zbuf; |
| 395 | uint8_t *buf; |
| 396 | |
| 397 | for (int i = 0; i < pages->normal_num; i++) { |
| 398 | buf = pages->block->host + pages->offset[i]; |
| 399 | multifd_qpl_prepare_comp_job(job, buf, zbuf, size); |
| 400 | if (qpl_execute_job(job) == QPL_STS_OK) { |
| 401 | multifd_qpl_fill_packet(i, p, zbuf, job->total_out); |
| 402 | } else { |
| 403 | /* send the uncompressed page */ |
| 404 | multifd_qpl_fill_packet(i, p, buf, size); |
| 405 | } |
| 406 | zbuf += size; |
| 407 | } |
| 408 | } |
| 409 | |
| 410 | /** |
| 411 | * multifd_qpl_compress_pages: compress pages |
| 412 | * |
| 413 | * Submit the pages to the IAA hardware for compression. If hardware |
| 414 | * compression fails, it falls back to software compression. If software |
| 415 | * compression also fails, the uncompressed page is sent. |
| 416 | * |
| 417 | * @p: Params for the channel being used |
| 418 | */ |
| 419 | static void multifd_qpl_compress_pages(MultiFDSendParams *p) |
| 420 | { |
| 421 | QplData *qpl = p->compress_data; |
| 422 | MultiFDPages_t *pages = &p->data->u.ram; |
| 423 | uint32_t size = multifd_ram_page_size(); |
| 424 | QplHwJob *hw_job; |
| 425 | uint8_t *buf; |
| 426 | uint8_t *zbuf; |
| 427 | |
| 428 | for (int i = 0; i < pages->normal_num; i++) { |
| 429 | buf = pages->block->host + pages->offset[i]; |
| 430 | zbuf = qpl->zbuf + (size * i); |
| 431 | hw_job = &qpl->hw_jobs[i]; |
| 432 | multifd_qpl_prepare_comp_job(hw_job->job, buf, zbuf, size); |
| 433 | if (multifd_qpl_submit_job(hw_job->job)) { |
| 434 | hw_job->fallback_sw_path = false; |
| 435 | } else { |
| 436 | /* |
| 437 | * The IAA work queue is full, any immediate subsequent job |
| 438 | * submission is likely to fail, sending the page via the QPL |
| 439 | * software path at this point gives us a better chance of |
| 440 | * finding the queue open for the next pages. |
| 441 | */ |
| 442 | hw_job->fallback_sw_path = true; |
| 443 | multifd_qpl_prepare_comp_job(qpl->sw_job, buf, zbuf, size); |
| 444 | if (qpl_execute_job(qpl->sw_job) == QPL_STS_OK) { |
| 445 | hw_job->sw_output = zbuf; |
| 446 | hw_job->sw_output_len = qpl->sw_job->total_out; |
| 447 | } else { |
| 448 | hw_job->sw_output = buf; |
| 449 | hw_job->sw_output_len = size; |
| 450 | } |
| 451 | } |
| 452 | } |
| 453 | |
| 454 | for (int i = 0; i < pages->normal_num; i++) { |
| 455 | buf = pages->block->host + pages->offset[i]; |
| 456 | zbuf = qpl->zbuf + (size * i); |
| 457 | hw_job = &qpl->hw_jobs[i]; |
| 458 | if (hw_job->fallback_sw_path) { |
| 459 | multifd_qpl_fill_packet(i, p, hw_job->sw_output, |
| 460 | hw_job->sw_output_len); |
| 461 | continue; |
| 462 | } |
| 463 | if (qpl_wait_job(hw_job->job) == QPL_STS_OK) { |
| 464 | multifd_qpl_fill_packet(i, p, zbuf, hw_job->job->total_out); |
| 465 | } else { |
| 466 | /* send the uncompressed page */ |
| 467 | multifd_qpl_fill_packet(i, p, buf, size); |
| 468 | } |
| 469 | } |
| 470 | } |
| 471 | |
| 472 | static int multifd_qpl_send_prepare(MultiFDSendParams *p, Error **errp) |
| 473 | { |
| 474 | QplData *qpl = p->compress_data; |
| 475 | MultiFDPages_t *pages = &p->data->u.ram; |
| 476 | uint32_t len = 0; |
| 477 | |
| 478 | if (!multifd_send_prepare_common(p)) { |
| 479 | goto out; |
| 480 | } |
| 481 | |
| 482 | /* The first IOV is used to store the compressed page lengths */ |
| 483 | len = pages->normal_num * sizeof(uint32_t); |
| 484 | multifd_qpl_fill_iov(p, (uint8_t *) qpl->zlen, len); |
| 485 | if (qpl->hw_avail) { |
| 486 | multifd_qpl_compress_pages(p); |
| 487 | } else { |
| 488 | multifd_qpl_compress_pages_slow_path(p); |
| 489 | } |
| 490 | |
| 491 | out: |
| 492 | p->flags |= MULTIFD_FLAG_QPL; |
| 493 | multifd_send_fill_packet(p); |
| 494 | return 0; |
| 495 | } |
| 496 | |
| 497 | static int multifd_qpl_recv_setup(MultiFDRecvParams *p, Error **errp) |
| 498 | { |
| 499 | QplData *qpl; |
| 500 | uint32_t page_size = multifd_ram_page_size(); |
| 501 | uint32_t page_count = multifd_ram_page_count(); |
| 502 | |
| 503 | qpl = multifd_qpl_init(page_count, page_size, errp); |
| 504 | if (!qpl) { |
| 505 | return -1; |
| 506 | } |
| 507 | p->compress_data = qpl; |
| 508 | return 0; |
| 509 | } |
| 510 | |
| 511 | static void multifd_qpl_recv_cleanup(MultiFDRecvParams *p) |
| 512 | { |
| 513 | multifd_qpl_deinit(p->compress_data); |
| 514 | p->compress_data = NULL; |
| 515 | } |
| 516 | |
| 517 | /** |
| 518 | * multifd_qpl_process_and_check_job: process and check a QPL job |
| 519 | * |
| 520 | * Process the job and check whether the job output length is the |
| 521 | * same as the specified length |
| 522 | * |
| 523 | * Returns true if the job execution succeeded and the output length |
| 524 | * is equal to the specified length, otherwise false. |
| 525 | * |
| 526 | * @job: pointer to the qpl_job structure |
| 527 | * @is_hardware: indicates whether the job is a hardware job |
| 528 | * @len: Specified output length |
| 529 | * @errp: pointer to an error |
| 530 | */ |
| 531 | static bool multifd_qpl_process_and_check_job(qpl_job *job, bool is_hardware, |
| 532 | uint32_t len, Error **errp) |
| 533 | { |
| 534 | qpl_status status; |
| 535 | |
| 536 | status = (is_hardware ? qpl_wait_job(job) : qpl_execute_job(job)); |
| 537 | if (status != QPL_STS_OK) { |
| 538 | error_setg(errp, "qpl job failed with error %d", status); |
| 539 | return false; |
| 540 | } |
| 541 | if (job->total_out != len) { |
| 542 | error_setg(errp, "qpl decompressed len %u, expected len %u", |
| 543 | job->total_out, len); |
| 544 | return false; |
| 545 | } |
| 546 | return true; |
| 547 | } |
| 548 | |
| 549 | /** |
| 550 | * multifd_qpl_decompress_pages_slow_path: decompress pages using slow path |
| 551 | * |
| 552 | * Decompress the pages using software |
| 553 | * |
| 554 | * Returns 0 on success or -1 on error |
| 555 | * |
| 556 | * @p: Params for the channel being used |
| 557 | * @errp: pointer to an error |
| 558 | */ |
| 559 | static int multifd_qpl_decompress_pages_slow_path(MultiFDRecvParams *p, |
| 560 | Error **errp) |
| 561 | { |
| 562 | QplData *qpl = p->compress_data; |
| 563 | uint32_t size = multifd_ram_page_size(); |
| 564 | qpl_job *job = qpl->sw_job; |
| 565 | uint8_t *zbuf = qpl->zbuf; |
| 566 | uint8_t *addr; |
| 567 | uint32_t len; |
| 568 | |
| 569 | for (int i = 0; i < p->normal_num; i++) { |
| 570 | len = qpl->zlen[i]; |
| 571 | addr = p->host + p->normal[i]; |
| 572 | /* the page is uncompressed, load it */ |
| 573 | if (len == size) { |
| 574 | memcpy(addr, zbuf, size); |
| 575 | zbuf += size; |
| 576 | continue; |
| 577 | } |
| 578 | multifd_qpl_prepare_decomp_job(job, zbuf, len, addr, size); |
| 579 | if (!multifd_qpl_process_and_check_job(job, false, size, errp)) { |
| 580 | return -1; |
| 581 | } |
| 582 | zbuf += len; |
| 583 | } |
| 584 | return 0; |
| 585 | } |
| 586 | |
| 587 | /** |
| 588 | * multifd_qpl_decompress_pages: decompress pages |
| 589 | * |
| 590 | * Decompress the pages using the IAA hardware. If hardware |
| 591 | * decompression fails, it falls back to software decompression. |
| 592 | * |
| 593 | * Returns 0 on success or -1 on error |
| 594 | * |
| 595 | * @p: Params for the channel being used |
| 596 | * @errp: pointer to an error |
| 597 | */ |
| 598 | static int multifd_qpl_decompress_pages(MultiFDRecvParams *p, Error **errp) |
| 599 | { |
| 600 | QplData *qpl = p->compress_data; |
| 601 | uint32_t size = multifd_ram_page_size(); |
| 602 | uint8_t *zbuf = qpl->zbuf; |
| 603 | uint8_t *addr; |
| 604 | uint32_t len; |
| 605 | qpl_job *job; |
| 606 | |
| 607 | for (int i = 0; i < p->normal_num; i++) { |
| 608 | addr = p->host + p->normal[i]; |
| 609 | len = qpl->zlen[i]; |
| 610 | /* the page is uncompressed if received length equals the page size */ |
| 611 | if (len == size) { |
| 612 | memcpy(addr, zbuf, size); |
| 613 | zbuf += size; |
| 614 | continue; |
| 615 | } |
| 616 | |
| 617 | job = qpl->hw_jobs[i].job; |
| 618 | multifd_qpl_prepare_decomp_job(job, zbuf, len, addr, size); |
| 619 | if (multifd_qpl_submit_job(job)) { |
| 620 | qpl->hw_jobs[i].fallback_sw_path = false; |
| 621 | } else { |
| 622 | /* |
| 623 | * The IAA work queue is full, any immediate subsequent job |
| 624 | * submission is likely to fail, sending the page via the QPL |
| 625 | * software path at this point gives us a better chance of |
| 626 | * finding the queue open for the next pages. |
| 627 | */ |
| 628 | qpl->hw_jobs[i].fallback_sw_path = true; |
| 629 | job = qpl->sw_job; |
| 630 | multifd_qpl_prepare_decomp_job(job, zbuf, len, addr, size); |
| 631 | if (!multifd_qpl_process_and_check_job(job, false, size, errp)) { |
| 632 | return -1; |
| 633 | } |
| 634 | } |
| 635 | zbuf += len; |
| 636 | } |
| 637 | |
| 638 | for (int i = 0; i < p->normal_num; i++) { |
| 639 | /* ignore pages that have already been processed */ |
| 640 | if (qpl->zlen[i] == size || qpl->hw_jobs[i].fallback_sw_path) { |
| 641 | continue; |
| 642 | } |
| 643 | |
| 644 | job = qpl->hw_jobs[i].job; |
| 645 | if (!multifd_qpl_process_and_check_job(job, true, size, errp)) { |
| 646 | return -1; |
| 647 | } |
| 648 | } |
| 649 | return 0; |
| 650 | } |
| 651 | static int multifd_qpl_recv(MultiFDRecvParams *p, Error **errp) |
| 652 | { |
| 653 | QplData *qpl = p->compress_data; |
| 654 | uint32_t in_size = p->next_packet_size; |
| 655 | uint32_t flags = p->flags & MULTIFD_FLAG_COMPRESSION_MASK; |
| 656 | uint32_t len = 0; |
| 657 | uint32_t zbuf_len = 0; |
| 658 | int ret; |
| 659 | |
| 660 | if (flags != MULTIFD_FLAG_QPL) { |
| 661 | error_setg(errp, "multifd %u: flags received %x flags expected %x", |
| 662 | p->id, flags, MULTIFD_FLAG_QPL); |
| 663 | return -1; |
| 664 | } |
| 665 | multifd_recv_zero_page_process(p); |
| 666 | if (!p->normal_num) { |
| 667 | if (in_size != 0) { |
| 668 | error_setg(errp, "multifd %u: expected empty packet", p->id); |
| 669 | return -1; |
| 670 | } |
| 671 | return 0; |
| 672 | } |
| 673 | |
| 674 | /* read compressed page lengths */ |
| 675 | len = p->normal_num * sizeof(uint32_t); |
| 676 | if (len >= in_size) { |
| 677 | error_setg(errp, "multifd %u: header len %"PRIu32 |
| 678 | " >= packet size %"PRIu32, p->id, len, in_size); |
| 679 | return -1; |
| 680 | } |
| 681 | ret = qio_channel_read_all(p->c, (void *) qpl->zlen, len, errp); |
| 682 | if (ret != 0) { |
| 683 | return ret; |
| 684 | } |
| 685 | for (int i = 0; i < p->normal_num; i++) { |
| 686 | qpl->zlen[i] = be32_to_cpu(qpl->zlen[i]); |
| 687 | if (qpl->zlen[i] > multifd_ram_page_size()) { |
| 688 | error_setg(errp, "multifd %u: page %d compressed len %" |
| 689 | PRIu32" too large", p->id, i, qpl->zlen[i]); |
| 690 | return -1; |
| 691 | } |
| 692 | zbuf_len += qpl->zlen[i]; |
| 693 | ramblock_recv_bitmap_set_offset(p->block, p->normal[i]); |
| 694 | } |
| 695 | |
| 696 | /* read compressed pages */ |
| 697 | if (in_size != len + zbuf_len) { |
| 698 | error_setg(errp, "multifd %u: packet size %"PRIu32 |
| 699 | " != header %"PRIu32" + data %"PRIu32, |
| 700 | p->id, in_size, len, zbuf_len); |
| 701 | return -1; |
| 702 | } |
| 703 | ret = qio_channel_read_all(p->c, (void *) qpl->zbuf, zbuf_len, errp); |
| 704 | if (ret != 0) { |
| 705 | return ret; |
| 706 | } |
| 707 | |
| 708 | if (qpl->hw_avail) { |
| 709 | return multifd_qpl_decompress_pages(p, errp); |
| 710 | } |
| 711 | return multifd_qpl_decompress_pages_slow_path(p, errp); |
| 712 | } |
| 713 | |
| 714 | static const MultiFDMethods multifd_qpl_ops = { |
| 715 | .send_setup = multifd_qpl_send_setup, |
| 716 | .send_cleanup = multifd_qpl_send_cleanup, |
| 717 | .send_prepare = multifd_qpl_send_prepare, |
| 718 | .recv_setup = multifd_qpl_recv_setup, |
| 719 | .recv_cleanup = multifd_qpl_recv_cleanup, |
| 720 | .recv = multifd_qpl_recv, |
| 721 | }; |
| 722 | |
| 723 | static void multifd_qpl_register(void) |
| 724 | { |
| 725 | multifd_register_ops(MULTIFD_COMPRESSION_QPL, &multifd_qpl_ops); |
| 726 | } |
| 727 | |
| 728 | migration_init(multifd_qpl_register); |