1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright 2021 Intel Corporation
7 #include <rte_malloc.h>
8 #include <rte_common.h>
10 #include <rte_prefetch.h>
12 #include "idxd_internal.h"
14 #define IDXD_PMD_NAME_STR "dmadev_idxd"
16 static __rte_always_inline rte_iova_t
17 __desc_idx_to_iova(struct idxd_dmadev *idxd, uint16_t n)
19 return idxd->desc_iova + (n * sizeof(struct idxd_hw_desc));
22 static __rte_always_inline void
23 __idxd_movdir64b(volatile void *dst, const struct idxd_hw_desc *src)
25 asm volatile (".byte 0x66, 0x0f, 0x38, 0xf8, 0x02"
27 : "a" (dst), "d" (src)
31 static __rte_always_inline void
32 __submit(struct idxd_dmadev *idxd)
34 rte_prefetch1(&idxd->batch_comp_ring[idxd->batch_idx_read]);
36 if (idxd->batch_size == 0)
39 /* write completion to batch comp ring */
40 rte_iova_t comp_addr = idxd->batch_iova +
41 (idxd->batch_idx_write * sizeof(struct idxd_completion));
43 if (idxd->batch_size == 1) {
44 /* submit batch directly */
45 struct idxd_hw_desc desc =
46 idxd->desc_ring[idxd->batch_start & idxd->desc_ring_mask];
47 desc.completion = comp_addr;
48 desc.op_flags |= IDXD_FLAG_REQUEST_COMPLETION;
49 _mm_sfence(); /* fence before writing desc to device */
50 __idxd_movdir64b(idxd->portal, &desc);
52 const struct idxd_hw_desc batch_desc = {
53 .op_flags = (idxd_op_batch << IDXD_CMD_OP_SHIFT) |
54 IDXD_FLAG_COMPLETION_ADDR_VALID |
55 IDXD_FLAG_REQUEST_COMPLETION,
56 .desc_addr = __desc_idx_to_iova(idxd,
57 idxd->batch_start & idxd->desc_ring_mask),
58 .completion = comp_addr,
59 .size = idxd->batch_size,
61 _mm_sfence(); /* fence before writing desc to device */
62 __idxd_movdir64b(idxd->portal, &batch_desc);
65 if (++idxd->batch_idx_write > idxd->max_batches)
66 idxd->batch_idx_write = 0;
68 idxd->stats.submitted += idxd->batch_size;
70 idxd->batch_start += idxd->batch_size;
72 idxd->batch_idx_ring[idxd->batch_idx_write] = idxd->batch_start;
73 _mm256_store_si256((void *)&idxd->batch_comp_ring[idxd->batch_idx_write],
74 _mm256_setzero_si256());
77 static __rte_always_inline int
78 __idxd_write_desc(struct idxd_dmadev *idxd,
79 const uint32_t op_flags,
85 uint16_t mask = idxd->desc_ring_mask;
86 uint16_t job_id = idxd->batch_start + idxd->batch_size;
87 /* we never wrap batches, so we only mask the start and allow start+size to overflow */
88 uint16_t write_idx = (idxd->batch_start & mask) + idxd->batch_size;
90 /* first check batch ring space then desc ring space */
91 if ((idxd->batch_idx_read == 0 && idxd->batch_idx_write == idxd->max_batches) ||
92 idxd->batch_idx_write + 1 == idxd->batch_idx_read)
94 if (((write_idx + 1) & mask) == (idxd->ids_returned & mask))
97 /* write desc. Note: descriptors don't wrap, but the completion address does */
98 const uint64_t op_flags64 = (uint64_t)(op_flags | IDXD_FLAG_COMPLETION_ADDR_VALID) << 32;
99 const uint64_t comp_addr = __desc_idx_to_iova(idxd, write_idx & mask);
100 _mm256_store_si256((void *)&idxd->desc_ring[write_idx],
101 _mm256_set_epi64x(dst, src, comp_addr, op_flags64));
102 _mm256_store_si256((void *)&idxd->desc_ring[write_idx].size,
103 _mm256_set_epi64x(0, 0, 0, size));
107 rte_prefetch0_write(&idxd->desc_ring[write_idx + 1]);
109 if (flags & RTE_DMA_OP_FLAG_SUBMIT)
116 idxd_enqueue_copy(void *dev_private, uint16_t qid __rte_unused, rte_iova_t src,
117 rte_iova_t dst, unsigned int length, uint64_t flags)
119 /* we can take advantage of the fact that the fence flag in dmadev and DSA are the same,
120 * but check it at compile time to be sure.
122 RTE_BUILD_BUG_ON(RTE_DMA_OP_FLAG_FENCE != IDXD_FLAG_FENCE);
123 uint32_t memmove = (idxd_op_memmove << IDXD_CMD_OP_SHIFT) |
124 IDXD_FLAG_CACHE_CONTROL | (flags & IDXD_FLAG_FENCE);
125 return __idxd_write_desc(dev_private, memmove, src, dst, length,
130 idxd_enqueue_fill(void *dev_private, uint16_t qid __rte_unused, uint64_t pattern,
131 rte_iova_t dst, unsigned int length, uint64_t flags)
133 uint32_t fill = (idxd_op_fill << IDXD_CMD_OP_SHIFT) |
134 IDXD_FLAG_CACHE_CONTROL | (flags & IDXD_FLAG_FENCE);
135 return __idxd_write_desc(dev_private, fill, pattern, dst, length,
140 idxd_submit(void *dev_private, uint16_t qid __rte_unused)
142 __submit(dev_private);
146 static enum rte_dma_status_code
147 get_comp_status(struct idxd_completion *c)
149 uint8_t st = c->status;
151 /* successful descriptors are not written back normally */
152 case IDXD_COMP_STATUS_INCOMPLETE:
153 case IDXD_COMP_STATUS_SUCCESS:
154 return RTE_DMA_STATUS_SUCCESSFUL;
155 case IDXD_COMP_STATUS_INVALID_OPCODE:
156 return RTE_DMA_STATUS_INVALID_OPCODE;
157 case IDXD_COMP_STATUS_INVALID_SIZE:
158 return RTE_DMA_STATUS_INVALID_LENGTH;
159 case IDXD_COMP_STATUS_SKIPPED:
160 return RTE_DMA_STATUS_NOT_ATTEMPTED;
162 return RTE_DMA_STATUS_ERROR_UNKNOWN;
167 idxd_vchan_status(const struct rte_dma_dev *dev, uint16_t vchan __rte_unused,
168 enum rte_dma_vchan_status *status)
170 struct idxd_dmadev *idxd = dev->fp_obj->dev_private;
171 uint16_t last_batch_write = idxd->batch_idx_write == 0 ? idxd->max_batches :
172 idxd->batch_idx_write - 1;
173 uint8_t bstatus = (idxd->batch_comp_ring[last_batch_write].status != 0);
175 /* An IDXD device will always be either active or idle.
176 * RTE_DMA_VCHAN_HALTED_ERROR is therefore not supported by IDXD.
178 *status = bstatus ? RTE_DMA_VCHAN_IDLE : RTE_DMA_VCHAN_ACTIVE;
183 static __rte_always_inline int
184 batch_ok(struct idxd_dmadev *idxd, uint16_t max_ops, enum rte_dma_status_code *status)
192 /* first check if there are any unreturned handles from last time */
193 if (idxd->ids_avail != idxd->ids_returned) {
194 ret = RTE_MIN((uint16_t)(idxd->ids_avail - idxd->ids_returned), max_ops);
195 idxd->ids_returned += ret;
197 memset(status, RTE_DMA_STATUS_SUCCESSFUL, ret * sizeof(*status));
201 if (idxd->batch_idx_read == idxd->batch_idx_write)
204 bstatus = idxd->batch_comp_ring[idxd->batch_idx_read].status;
205 /* now check if next batch is complete and successful */
206 if (bstatus == IDXD_COMP_STATUS_SUCCESS) {
207 /* since the batch idx ring stores the start of each batch, pre-increment to lookup
208 * start of next batch.
210 if (++idxd->batch_idx_read > idxd->max_batches)
211 idxd->batch_idx_read = 0;
212 idxd->ids_avail = idxd->batch_idx_ring[idxd->batch_idx_read];
214 ret = RTE_MIN((uint16_t)(idxd->ids_avail - idxd->ids_returned), max_ops);
215 idxd->ids_returned += ret;
217 memset(status, RTE_DMA_STATUS_SUCCESSFUL, ret * sizeof(*status));
220 /* check if batch is incomplete */
221 else if (bstatus == IDXD_COMP_STATUS_INCOMPLETE)
224 return -1; /* error case */
227 static inline uint16_t
228 batch_completed(struct idxd_dmadev *idxd, uint16_t max_ops, bool *has_error)
231 uint16_t b_start, b_end, next_batch;
233 int ret = batch_ok(idxd, max_ops, NULL);
237 /* ERROR case, not successful, not incomplete */
238 /* Get the batch size, and special case size 1.
239 * once we identify the actual failure job, return other jobs, then update
240 * the batch ring indexes to make it look like the first job of the batch has failed.
241 * Subsequent calls here will always return zero packets, and the error must be cleared by
242 * calling the completed_status() function.
244 next_batch = (idxd->batch_idx_read + 1);
245 if (next_batch > idxd->max_batches)
247 b_start = idxd->batch_idx_ring[idxd->batch_idx_read];
248 b_end = idxd->batch_idx_ring[next_batch];
250 if (b_end - b_start == 1) { /* not a batch */
255 for (i = b_start; i < b_end; i++) {
256 struct idxd_completion *c = (void *)&idxd->desc_ring[i & idxd->desc_ring_mask];
257 if (c->status > IDXD_COMP_STATUS_SUCCESS) /* ignore incomplete(0) and success(1) */
260 ret = RTE_MIN((uint16_t)(i - idxd->ids_returned), max_ops);
262 *has_error = true; /* we got up to the point of error */
263 idxd->ids_avail = idxd->ids_returned += ret;
265 /* to ensure we can call twice and just return 0, set start of batch to where we finished */
266 idxd->batch_comp_ring[idxd->batch_idx_read].completed_size -= ret;
267 idxd->batch_idx_ring[idxd->batch_idx_read] += ret;
268 if (idxd->batch_idx_ring[next_batch] - idxd->batch_idx_ring[idxd->batch_idx_read] == 1) {
269 /* copy over the descriptor status to the batch ring as if no batch */
270 uint16_t d_idx = idxd->batch_idx_ring[idxd->batch_idx_read] & idxd->desc_ring_mask;
271 struct idxd_completion *desc_comp = (void *)&idxd->desc_ring[d_idx];
272 idxd->batch_comp_ring[idxd->batch_idx_read].status = desc_comp->status;
279 batch_completed_status(struct idxd_dmadev *idxd, uint16_t max_ops, enum rte_dma_status_code *status)
283 int ret = batch_ok(idxd, max_ops, status);
287 /* ERROR case, not successful, not incomplete */
288 /* Get the batch size, and special case size 1.
290 next_batch = (idxd->batch_idx_read + 1);
291 if (next_batch > idxd->max_batches)
293 const uint16_t b_start = idxd->batch_idx_ring[idxd->batch_idx_read];
294 const uint16_t b_end = idxd->batch_idx_ring[next_batch];
295 const uint16_t b_len = b_end - b_start;
296 if (b_len == 1) {/* not a batch */
297 *status = get_comp_status(&idxd->batch_comp_ring[idxd->batch_idx_read]);
298 if (status != RTE_DMA_STATUS_SUCCESSFUL)
299 idxd->stats.errors++;
301 idxd->ids_returned++;
302 idxd->batch_idx_read = next_batch;
306 /* not a single-element batch, need to process more.
308 * 1. max_ops >= batch_size - can fit everything, simple case
309 * - loop through completed ops and then add on any not-attempted ones
310 * 2. max_ops < batch_size - can't fit everything, more complex case
311 * - loop through completed/incomplete and stop when hit max_ops
312 * - adjust the batch descriptor to update where we stopped, with appropriate bcount
313 * - if bcount is to be exactly 1, update the batch descriptor as it will be treated as
314 * non-batch next time.
316 const uint16_t bcount = idxd->batch_comp_ring[idxd->batch_idx_read].completed_size;
317 for (ret = 0; ret < b_len && ret < max_ops; ret++) {
318 struct idxd_completion *c = (void *)
319 &idxd->desc_ring[(b_start + ret) & idxd->desc_ring_mask];
320 status[ret] = (ret < bcount) ? get_comp_status(c) : RTE_DMA_STATUS_NOT_ATTEMPTED;
321 if (status[ret] != RTE_DMA_STATUS_SUCCESSFUL)
322 idxd->stats.errors++;
324 idxd->ids_avail = idxd->ids_returned += ret;
328 idxd->batch_idx_read = next_batch;
332 /* set up for next time, update existing batch descriptor & start idx at batch_idx_read */
333 idxd->batch_idx_ring[idxd->batch_idx_read] += ret;
335 /* we have only incomplete ones - set batch completed size to 0 */
336 struct idxd_completion *comp = &idxd->batch_comp_ring[idxd->batch_idx_read];
337 comp->completed_size = 0;
338 /* if there is only one descriptor left, job skipped so set flag appropriately */
339 if (b_len - ret == 1)
340 comp->status = IDXD_COMP_STATUS_SKIPPED;
342 struct idxd_completion *comp = &idxd->batch_comp_ring[idxd->batch_idx_read];
343 comp->completed_size -= ret;
344 /* if there is only one descriptor left, copy status info straight to desc */
345 if (comp->completed_size == 1) {
346 struct idxd_completion *c = (void *)
347 &idxd->desc_ring[(b_start + ret) & idxd->desc_ring_mask];
348 comp->status = c->status;
349 /* individual descs can be ok without writeback, but not batches */
350 if (comp->status == IDXD_COMP_STATUS_INCOMPLETE)
351 comp->status = IDXD_COMP_STATUS_SUCCESS;
352 } else if (bcount == b_len) {
353 /* check if we still have an error, and clear flag if not */
355 for (i = b_start + ret; i < b_end; i++) {
356 struct idxd_completion *c = (void *)
357 &idxd->desc_ring[i & idxd->desc_ring_mask];
358 if (c->status > IDXD_COMP_STATUS_SUCCESS)
361 if (i == b_end) /* no errors */
362 comp->status = IDXD_COMP_STATUS_SUCCESS;
370 idxd_completed(void *dev_private, uint16_t qid __rte_unused, uint16_t max_ops,
371 uint16_t *last_idx, bool *has_error)
373 struct idxd_dmadev *idxd = dev_private;
374 uint16_t batch, ret = 0;
377 batch = batch_completed(idxd, max_ops - ret, has_error);
379 } while (batch > 0 && *has_error == false);
381 idxd->stats.completed += ret;
382 *last_idx = idxd->ids_returned - 1;
387 idxd_completed_status(void *dev_private, uint16_t qid __rte_unused, uint16_t max_ops,
388 uint16_t *last_idx, enum rte_dma_status_code *status)
390 struct idxd_dmadev *idxd = dev_private;
391 uint16_t batch, ret = 0;
394 batch = batch_completed_status(idxd, max_ops - ret, &status[ret]);
398 idxd->stats.completed += ret;
399 *last_idx = idxd->ids_returned - 1;
404 idxd_dump(const struct rte_dma_dev *dev, FILE *f)
406 struct idxd_dmadev *idxd = dev->fp_obj->dev_private;
409 fprintf(f, "== IDXD Private Data ==\n");
410 fprintf(f, " Portal: %p\n", idxd->portal);
411 fprintf(f, " Config: { ring_size: %u }\n",
413 fprintf(f, " Batch ring (sz = %u, max_batches = %u):\n\t",
414 idxd->max_batches + 1, idxd->max_batches);
415 for (i = 0; i <= idxd->max_batches; i++) {
416 fprintf(f, " %u ", idxd->batch_idx_ring[i]);
417 if (i == idxd->batch_idx_read && i == idxd->batch_idx_write)
418 fprintf(f, "[rd ptr, wr ptr] ");
419 else if (i == idxd->batch_idx_read)
420 fprintf(f, "[rd ptr] ");
421 else if (i == idxd->batch_idx_write)
422 fprintf(f, "[wr ptr] ");
423 if (i == idxd->max_batches)
427 fprintf(f, " Curr batch: start = %u, size = %u\n", idxd->batch_start, idxd->batch_size);
428 fprintf(f, " IDS: avail = %u, returned: %u\n", idxd->ids_avail, idxd->ids_returned);
433 idxd_stats_get(const struct rte_dma_dev *dev, uint16_t vchan __rte_unused,
434 struct rte_dma_stats *stats, uint32_t stats_sz)
436 struct idxd_dmadev *idxd = dev->fp_obj->dev_private;
437 if (stats_sz < sizeof(*stats))
439 *stats = idxd->stats;
444 idxd_stats_reset(struct rte_dma_dev *dev, uint16_t vchan __rte_unused)
446 struct idxd_dmadev *idxd = dev->fp_obj->dev_private;
447 idxd->stats = (struct rte_dma_stats){0};
452 idxd_info_get(const struct rte_dma_dev *dev, struct rte_dma_info *info, uint32_t size)
454 struct idxd_dmadev *idxd = dev->fp_obj->dev_private;
456 if (size < sizeof(*info))
459 *info = (struct rte_dma_info) {
460 .dev_capa = RTE_DMA_CAPA_MEM_TO_MEM | RTE_DMA_CAPA_HANDLES_ERRORS |
461 RTE_DMA_CAPA_OPS_COPY | RTE_DMA_CAPA_OPS_FILL,
466 if (idxd->sva_support)
467 info->dev_capa |= RTE_DMA_CAPA_SVA;
472 idxd_burst_capacity(const void *dev_private, uint16_t vchan __rte_unused)
474 const struct idxd_dmadev *idxd = dev_private;
475 uint16_t write_idx = idxd->batch_start + idxd->batch_size;
478 /* Check for space in the batch ring */
479 if ((idxd->batch_idx_read == 0 && idxd->batch_idx_write == idxd->max_batches) ||
480 idxd->batch_idx_write + 1 == idxd->batch_idx_read)
483 /* For descriptors, check for wrap-around on write but not read */
484 if (idxd->ids_returned > write_idx)
485 write_idx += idxd->desc_ring_mask + 1;
486 used_space = write_idx - idxd->ids_returned;
488 return RTE_MIN((idxd->desc_ring_mask - used_space), idxd->max_batch_size);
492 idxd_configure(struct rte_dma_dev *dev __rte_unused, const struct rte_dma_conf *dev_conf,
495 if (sizeof(struct rte_dma_conf) != conf_sz)
498 if (dev_conf->nb_vchans != 1)
504 idxd_vchan_setup(struct rte_dma_dev *dev, uint16_t vchan __rte_unused,
505 const struct rte_dma_vchan_conf *qconf, uint32_t qconf_sz)
507 struct idxd_dmadev *idxd = dev->fp_obj->dev_private;
508 uint16_t max_desc = qconf->nb_desc;
510 if (sizeof(struct rte_dma_vchan_conf) != qconf_sz)
515 if (!rte_is_power_of_2(max_desc))
516 max_desc = rte_align32pow2(max_desc);
517 IDXD_PMD_DEBUG("DMA dev %u using %u descriptors", dev->data->dev_id, max_desc);
518 idxd->desc_ring_mask = max_desc - 1;
519 idxd->qcfg.nb_desc = max_desc;
521 /* in case we are reconfiguring a device, free any existing memory */
522 rte_free(idxd->desc_ring);
524 /* allocate the descriptor ring at 2x size as batches can't wrap */
525 idxd->desc_ring = rte_zmalloc(NULL, sizeof(*idxd->desc_ring) * max_desc * 2, 0);
526 if (idxd->desc_ring == NULL)
528 idxd->desc_iova = rte_mem_virt2iova(idxd->desc_ring);
530 idxd->batch_idx_read = 0;
531 idxd->batch_idx_write = 0;
532 idxd->batch_start = 0;
533 idxd->batch_size = 0;
534 idxd->ids_returned = 0;
537 memset(idxd->batch_comp_ring, 0, sizeof(*idxd->batch_comp_ring) *
538 (idxd->max_batches + 1));
543 idxd_dmadev_create(const char *name, struct rte_device *dev,
544 const struct idxd_dmadev *base_idxd,
545 const struct rte_dma_dev_ops *ops)
547 struct idxd_dmadev *idxd = NULL;
548 struct rte_dma_dev *dmadev = NULL;
551 RTE_BUILD_BUG_ON(sizeof(struct idxd_hw_desc) != 64);
552 RTE_BUILD_BUG_ON(offsetof(struct idxd_hw_desc, size) != 32);
553 RTE_BUILD_BUG_ON(sizeof(struct idxd_completion) != 32);
556 IDXD_PMD_ERR("Invalid name of the device!");
561 /* Allocate device structure */
562 dmadev = rte_dma_pmd_allocate(name, dev->numa_node, sizeof(struct idxd_dmadev));
563 if (dmadev == NULL) {
564 IDXD_PMD_ERR("Unable to allocate dma device");
568 dmadev->dev_ops = ops;
569 dmadev->device = dev;
571 dmadev->fp_obj->copy = idxd_enqueue_copy;
572 dmadev->fp_obj->fill = idxd_enqueue_fill;
573 dmadev->fp_obj->submit = idxd_submit;
574 dmadev->fp_obj->completed = idxd_completed;
575 dmadev->fp_obj->completed_status = idxd_completed_status;
576 dmadev->fp_obj->burst_capacity = idxd_burst_capacity;
578 idxd = dmadev->data->dev_private;
579 *idxd = *base_idxd; /* copy over the main fields already passed in */
580 idxd->dmadev = dmadev;
582 /* allocate batch index ring and completion ring.
583 * The +1 is because we can never fully use
584 * the ring, otherwise read == write means both full and empty.
586 idxd->batch_comp_ring = rte_zmalloc_socket(NULL, (sizeof(idxd->batch_idx_ring[0]) +
587 sizeof(idxd->batch_comp_ring[0])) * (idxd->max_batches + 1),
588 sizeof(idxd->batch_comp_ring[0]), dev->numa_node);
589 if (idxd->batch_comp_ring == NULL) {
590 IDXD_PMD_ERR("Unable to reserve memory for batch data\n");
594 idxd->batch_idx_ring = (void *)&idxd->batch_comp_ring[idxd->max_batches+1];
595 idxd->batch_iova = rte_mem_virt2iova(idxd->batch_comp_ring);
597 dmadev->fp_obj->dev_private = idxd;
599 idxd->dmadev->state = RTE_DMA_DEV_READY;
605 rte_dma_pmd_release(name);
610 int idxd_pmd_logtype;
612 RTE_LOG_REGISTER_DEFAULT(idxd_pmd_logtype, WARNING);