1 // SPDX-License-Identifier: GPL-2.0-only
3 /* Copyright (c) 2019-2021, The Linux Foundation. All rights reserved. */
4 /* Copyright (c) 2021-2023 Qualcomm Innovation Center, Inc. All rights reserved. */
6 #include <linux/bitfield.h>
7 #include <linux/bits.h>
8 #include <linux/completion.h>
9 #include <linux/delay.h>
10 #include <linux/dma-buf.h>
11 #include <linux/dma-mapping.h>
12 #include <linux/interrupt.h>
13 #include <linux/kref.h>
14 #include <linux/list.h>
15 #include <linux/math64.h>
17 #include <linux/moduleparam.h>
18 #include <linux/scatterlist.h>
19 #include <linux/spinlock.h>
20 #include <linux/srcu.h>
21 #include <linux/types.h>
22 #include <linux/uaccess.h>
23 #include <linux/wait.h>
24 #include <drm/drm_file.h>
25 #include <drm/drm_gem.h>
26 #include <drm/drm_prime.h>
27 #include <drm/drm_print.h>
28 #include <uapi/drm/qaic_accel.h>
32 #define SEM_VAL_MASK GENMASK_ULL(11, 0)
33 #define SEM_INDEX_MASK GENMASK_ULL(4, 0)
34 #define BULK_XFER BIT(3)
35 #define GEN_COMPLETION BIT(4)
36 #define INBOUND_XFER 1
37 #define OUTBOUND_XFER 2
38 #define REQHP_OFF 0x0 /* we read this */
39 #define REQTP_OFF 0x4 /* we write this */
40 #define RSPHP_OFF 0x8 /* we write this */
41 #define RSPTP_OFF 0xc /* we read this */
43 #define ENCODE_SEM(val, index, sync, cmd, flags) \
45 FIELD_PREP(GENMASK(11, 0), (val)) | \
46 FIELD_PREP(GENMASK(20, 16), (index)) | \
47 FIELD_PREP(BIT(22), (sync)) | \
48 FIELD_PREP(GENMASK(26, 24), (cmd)) | \
49 FIELD_PREP(GENMASK(30, 29), (flags)) | \
50 FIELD_PREP(BIT(31), (cmd) ? 1 : 0); \
52 #define NUM_EVENTS 128
54 #define fifo_at(base, offset) ((base) + (offset) * get_dbc_req_elem_size())
56 static unsigned int wait_exec_default_timeout_ms = 5000; /* 5 sec default */
57 module_param(wait_exec_default_timeout_ms, uint, 0600);
58 MODULE_PARM_DESC(wait_exec_default_timeout_ms, "Default timeout for DRM_IOCTL_QAIC_WAIT_BO");
60 static unsigned int datapath_poll_interval_us = 100; /* 100 usec default */
61 module_param(datapath_poll_interval_us, uint, 0600);
62 MODULE_PARM_DESC(datapath_poll_interval_us,
63 "Amount of time to sleep between activity when datapath polling is enabled");
67 * A request ID is assigned to each memory handle going in DMA queue.
68 * As a single memory handle can enqueue multiple elements in DMA queue
69 * all of them will have the same request ID.
75 * Special encoded variable
76 * 7 0 - Do not force to generate MSI after DMA is completed
77 * 1 - Force to generate MSI after DMA is completed
79 * 4 1 - Generate completion element in the response queue
80 * 0 - No Completion Code
81 * 3 0 - DMA request is a Link list transfer
82 * 1 - DMA request is a Bulk transfer
84 * 1:0 00 - No DMA transfer involved
85 * 01 - DMA transfer is part of inbound transfer
86 * 10 - DMA transfer has outbound transfer
91 /* Source address for the transfer */
93 /* Destination address for the transfer */
95 /* Length of transfer request */
98 /* Doorbell address */
101 * Special encoded variable
102 * 7 1 - Doorbell(db) write
103 * 0 - No doorbell write
105 * 1:0 00 - 32 bit access, db address must be aligned to 32bit-boundary
106 * 01 - 16 bit access, db address must be aligned to 16bit-boundary
107 * 10 - 8 bit access, db address must be aligned to 8bit-boundary
113 /* 32 bit data written to doorbell address */
116 * Special encoded variable
117 * All the fields of sem_cmdX are passed from user and all are ORed
118 * together to form sem_cmd.
119 * 0:11 Semaphore value
121 * 20:16 Semaphore index
125 * 26:24 Semaphore command
127 * 29 Semaphore DMA out bound sync fence
128 * 30 Semaphore DMA in bound sync fence
129 * 31 Enable semaphore command
138 /* Request ID of the memory handle whose DMA transaction is completed */
140 /* Status of the DMA transaction. 0 : Success otherwise failure */
144 static inline bool bo_queued(struct qaic_bo *bo)
146 return !list_empty(&bo->xfer_list);
149 inline int get_dbc_req_elem_size(void)
151 return sizeof(struct dbc_req);
154 inline int get_dbc_rsp_elem_size(void)
156 return sizeof(struct dbc_rsp);
159 static void free_slice(struct kref *kref)
161 struct bo_slice *slice = container_of(kref, struct bo_slice, ref_count);
163 slice->bo->total_slice_nents -= slice->nents;
164 list_del(&slice->slice);
165 drm_gem_object_put(&slice->bo->base);
166 sg_free_table(slice->sgt);
172 static int clone_range_of_sgt_for_slice(struct qaic_device *qdev, struct sg_table **sgt_out,
173 struct sg_table *sgt_in, u64 size, u64 offset)
175 int total_len, len, nents, offf = 0, offl = 0;
176 struct scatterlist *sg, *sgn, *sgf, *sgl;
177 struct sg_table *sgt;
180 /* find out number of relevant nents needed for this mem */
186 size = size ? size : PAGE_SIZE;
187 for (sg = sgt_in->sgl; sg; sg = sg_next(sg)) {
188 len = sg_dma_len(sg);
192 if (offset >= total_len && offset < total_len + len) {
194 offf = offset - total_len;
198 if (offset + size >= total_len &&
199 offset + size <= total_len + len) {
201 offl = offset + size - total_len;
212 sgt = kzalloc(sizeof(*sgt), GFP_KERNEL);
218 ret = sg_alloc_table(sgt, nents, GFP_KERNEL);
222 /* copy relevant sg node and fix page and length */
224 for_each_sgtable_sg(sgt, sg, j) {
225 memcpy(sg, sgn, sizeof(*sg));
227 sg_dma_address(sg) += offf;
228 sg_dma_len(sg) -= offf;
229 sg_set_page(sg, sg_page(sgn), sg_dma_len(sg), offf);
234 sg_dma_len(sg) = offl - offf;
235 sg_set_page(sg, sg_page(sgn), offl - offf, offf);
252 static int encode_reqs(struct qaic_device *qdev, struct bo_slice *slice,
253 struct qaic_attach_slice_entry *req)
255 __le64 db_addr = cpu_to_le64(req->db_addr);
256 __le32 db_data = cpu_to_le32(req->db_data);
257 struct scatterlist *sg;
258 __u8 cmd = BULK_XFER;
265 cmd |= (slice->dir == DMA_TO_DEVICE ? INBOUND_XFER : OUTBOUND_XFER);
267 if (req->db_len && !IS_ALIGNED(req->db_addr, req->db_len / 8))
270 presync_sem = req->sem0.presync + req->sem1.presync + req->sem2.presync + req->sem3.presync;
274 presync_sem = req->sem0.presync << 0 | req->sem1.presync << 1 |
275 req->sem2.presync << 2 | req->sem3.presync << 3;
277 switch (req->db_len) {
288 db_len = 0; /* doorbell is not active for this command */
291 return -EINVAL; /* should never hit this */
295 * When we end up splitting up a single request (ie a buf slice) into
296 * multiple DMA requests, we have to manage the sync data carefully.
297 * There can only be one presync sem. That needs to be on every xfer
298 * so that the DMA engine doesn't transfer data before the receiver is
299 * ready. We only do the doorbell and postsync sems after the xfer.
300 * To guarantee previous xfers for the request are complete, we use a
303 dev_addr = req->dev_addr;
304 for_each_sgtable_sg(slice->sgt, sg, i) {
305 slice->reqs[i].cmd = cmd;
306 slice->reqs[i].src_addr = cpu_to_le64(slice->dir == DMA_TO_DEVICE ?
307 sg_dma_address(sg) : dev_addr);
308 slice->reqs[i].dest_addr = cpu_to_le64(slice->dir == DMA_TO_DEVICE ?
309 dev_addr : sg_dma_address(sg));
311 * sg_dma_len(sg) returns size of a DMA segment, maximum DMA
312 * segment size is set to UINT_MAX by qaic and hence return
313 * values of sg_dma_len(sg) can never exceed u32 range. So,
314 * by down sizing we are not corrupting the value.
316 slice->reqs[i].len = cpu_to_le32((u32)sg_dma_len(sg));
317 switch (presync_sem) {
319 slice->reqs[i].sem_cmd0 = cpu_to_le32(ENCODE_SEM(req->sem0.val,
326 slice->reqs[i].sem_cmd1 = cpu_to_le32(ENCODE_SEM(req->sem1.val,
333 slice->reqs[i].sem_cmd2 = cpu_to_le32(ENCODE_SEM(req->sem2.val,
340 slice->reqs[i].sem_cmd3 = cpu_to_le32(ENCODE_SEM(req->sem3.val,
347 dev_addr += sg_dma_len(sg);
349 /* add post transfer stuff to last segment */
351 slice->reqs[i].cmd |= GEN_COMPLETION;
352 slice->reqs[i].db_addr = db_addr;
353 slice->reqs[i].db_len = db_len;
354 slice->reqs[i].db_data = db_data;
356 * Add a fence if we have more than one request going to the hardware
357 * representing the entirety of the user request, and the user request
358 * has no presync condition.
359 * Fences are expensive, so we try to avoid them. We rely on the
360 * hardware behavior to avoid needing one when there is a presync
361 * condition. When a presync exists, all requests for that same
362 * presync will be queued into a fifo. Thus, since we queue the
363 * post xfer activity only on the last request we queue, the hardware
364 * will ensure that the last queued request is processed last, thus
365 * making sure the post xfer activity happens at the right time without
368 if (i && !presync_sem)
369 req->sem0.flags |= (slice->dir == DMA_TO_DEVICE ?
370 QAIC_SEM_INSYNCFENCE : QAIC_SEM_OUTSYNCFENCE);
371 slice->reqs[i].sem_cmd0 = cpu_to_le32(ENCODE_SEM(req->sem0.val, req->sem0.index,
372 req->sem0.presync, req->sem0.cmd,
374 slice->reqs[i].sem_cmd1 = cpu_to_le32(ENCODE_SEM(req->sem1.val, req->sem1.index,
375 req->sem1.presync, req->sem1.cmd,
377 slice->reqs[i].sem_cmd2 = cpu_to_le32(ENCODE_SEM(req->sem2.val, req->sem2.index,
378 req->sem2.presync, req->sem2.cmd,
380 slice->reqs[i].sem_cmd3 = cpu_to_le32(ENCODE_SEM(req->sem3.val, req->sem3.index,
381 req->sem3.presync, req->sem3.cmd,
387 static int qaic_map_one_slice(struct qaic_device *qdev, struct qaic_bo *bo,
388 struct qaic_attach_slice_entry *slice_ent)
390 struct sg_table *sgt = NULL;
391 struct bo_slice *slice;
394 ret = clone_range_of_sgt_for_slice(qdev, &sgt, bo->sgt, slice_ent->size, slice_ent->offset);
398 slice = kmalloc(sizeof(*slice), GFP_KERNEL);
404 slice->reqs = kcalloc(sgt->nents, sizeof(*slice->reqs), GFP_KERNEL);
410 slice->no_xfer = !slice_ent->size;
412 slice->nents = sgt->nents;
413 slice->dir = bo->dir;
415 slice->size = slice_ent->size;
416 slice->offset = slice_ent->offset;
418 ret = encode_reqs(qdev, slice, slice_ent);
422 bo->total_slice_nents += sgt->nents;
423 kref_init(&slice->ref_count);
424 drm_gem_object_get(&bo->base);
425 list_add_tail(&slice->slice, &bo->slices);
440 static int create_sgt(struct qaic_device *qdev, struct sg_table **sgt_out, u64 size)
442 struct scatterlist *sg;
443 struct sg_table *sgt;
454 nr_pages = DIV_ROUND_UP(size, PAGE_SIZE);
456 * calculate how much extra we are going to allocate, to remove
459 buf_extra = (PAGE_SIZE - size % PAGE_SIZE) % PAGE_SIZE;
460 max_order = min(MAX_PAGE_ORDER, get_order(size));
462 /* allocate a single page for book keeping */
468 pages = kvmalloc_array(nr_pages, sizeof(*pages) + sizeof(*pages_order), GFP_KERNEL);
473 pages_order = (void *)pages + sizeof(*pages) * nr_pages;
476 * Allocate requested memory using alloc_pages. It is possible to allocate
477 * the requested memory in multiple chunks by calling alloc_pages
478 * multiple times. Use SG table to handle multiple allocated pages.
481 while (nr_pages > 0) {
482 order = min(get_order(nr_pages * PAGE_SIZE), max_order);
484 pages[i] = alloc_pages(GFP_KERNEL | GFP_HIGHUSER |
485 __GFP_NOWARN | __GFP_ZERO |
486 (order ? __GFP_NORETRY : __GFP_RETRY_MAYFAIL),
492 goto free_partial_alloc;
497 pages_order[i] = order;
499 nr_pages -= 1 << order;
501 /* account for over allocation */
502 buf_extra += abs(nr_pages) * PAGE_SIZE;
506 sgt = kmalloc(sizeof(*sgt), GFP_KERNEL);
509 goto free_partial_alloc;
512 if (sg_alloc_table(sgt, i, GFP_KERNEL)) {
517 /* Populate the SG table with the allocated memory pages */
519 for (k = 0; k < i; k++, sg = sg_next(sg)) {
520 /* Last entry requires special handling */
522 sg_set_page(sg, pages[k], PAGE_SIZE << pages_order[k], 0);
524 sg_set_page(sg, pages[k], (PAGE_SIZE << pages_order[k]) - buf_extra, 0);
536 for (j = 0; j < i; j++)
537 __free_pages(pages[j], pages_order[j]);
544 static bool invalid_sem(struct qaic_sem *sem)
546 if (sem->val & ~SEM_VAL_MASK || sem->index & ~SEM_INDEX_MASK ||
547 !(sem->presync == 0 || sem->presync == 1) || sem->pad ||
548 sem->flags & ~(QAIC_SEM_INSYNCFENCE | QAIC_SEM_OUTSYNCFENCE) ||
549 sem->cmd > QAIC_SEM_WAIT_GT_0)
554 static int qaic_validate_req(struct qaic_device *qdev, struct qaic_attach_slice_entry *slice_ent,
555 u32 count, u64 total_size)
559 for (i = 0; i < count; i++) {
560 if (!(slice_ent[i].db_len == 32 || slice_ent[i].db_len == 16 ||
561 slice_ent[i].db_len == 8 || slice_ent[i].db_len == 0) ||
562 invalid_sem(&slice_ent[i].sem0) || invalid_sem(&slice_ent[i].sem1) ||
563 invalid_sem(&slice_ent[i].sem2) || invalid_sem(&slice_ent[i].sem3))
566 if (slice_ent[i].offset + slice_ent[i].size > total_size)
573 static void qaic_free_sgt(struct sg_table *sgt)
575 struct scatterlist *sg;
580 for (sg = sgt->sgl; sg; sg = sg_next(sg))
582 __free_pages(sg_page(sg), get_order(sg->length));
587 static void qaic_gem_print_info(struct drm_printer *p, unsigned int indent,
588 const struct drm_gem_object *obj)
590 struct qaic_bo *bo = to_qaic_bo(obj);
592 drm_printf_indent(p, indent, "BO DMA direction %d\n", bo->dir);
595 static const struct vm_operations_struct drm_vm_ops = {
596 .open = drm_gem_vm_open,
597 .close = drm_gem_vm_close,
600 static int qaic_gem_object_mmap(struct drm_gem_object *obj, struct vm_area_struct *vma)
602 struct qaic_bo *bo = to_qaic_bo(obj);
603 unsigned long offset = 0;
604 struct scatterlist *sg;
607 if (obj->import_attach)
610 for (sg = bo->sgt->sgl; sg; sg = sg_next(sg)) {
612 ret = remap_pfn_range(vma, vma->vm_start + offset, page_to_pfn(sg_page(sg)),
613 sg->length, vma->vm_page_prot);
616 offset += sg->length;
624 static void qaic_free_object(struct drm_gem_object *obj)
626 struct qaic_bo *bo = to_qaic_bo(obj);
628 if (obj->import_attach) {
629 /* DMABUF/PRIME Path */
630 drm_prime_gem_destroy(obj, NULL);
632 /* Private buffer allocation path */
633 qaic_free_sgt(bo->sgt);
636 mutex_destroy(&bo->lock);
637 drm_gem_object_release(obj);
641 static const struct drm_gem_object_funcs qaic_gem_funcs = {
642 .free = qaic_free_object,
643 .print_info = qaic_gem_print_info,
644 .mmap = qaic_gem_object_mmap,
645 .vm_ops = &drm_vm_ops,
648 static void qaic_init_bo(struct qaic_bo *bo, bool reinit)
652 reinit_completion(&bo->xfer_done);
654 mutex_init(&bo->lock);
655 init_completion(&bo->xfer_done);
657 complete_all(&bo->xfer_done);
658 INIT_LIST_HEAD(&bo->slices);
659 INIT_LIST_HEAD(&bo->xfer_list);
662 static struct qaic_bo *qaic_alloc_init_bo(void)
666 bo = kzalloc(sizeof(*bo), GFP_KERNEL);
668 return ERR_PTR(-ENOMEM);
670 qaic_init_bo(bo, false);
675 int qaic_create_bo_ioctl(struct drm_device *dev, void *data, struct drm_file *file_priv)
677 struct qaic_create_bo *args = data;
678 int usr_rcu_id, qdev_rcu_id;
679 struct drm_gem_object *obj;
680 struct qaic_device *qdev;
681 struct qaic_user *usr;
689 size = PAGE_ALIGN(args->size);
693 usr = file_priv->driver_priv;
694 usr_rcu_id = srcu_read_lock(&usr->qddev_lock);
697 goto unlock_usr_srcu;
700 qdev = usr->qddev->qdev;
701 qdev_rcu_id = srcu_read_lock(&qdev->dev_lock);
702 if (qdev->dev_state != QAIC_ONLINE) {
704 goto unlock_dev_srcu;
707 bo = qaic_alloc_init_bo();
710 goto unlock_dev_srcu;
714 drm_gem_private_object_init(dev, obj, size);
716 obj->funcs = &qaic_gem_funcs;
717 ret = create_sgt(qdev, &bo->sgt, size);
721 ret = drm_gem_create_mmap_offset(obj);
725 ret = drm_gem_handle_create(file_priv, obj, &args->handle);
729 bo->handle = args->handle;
730 drm_gem_object_put(obj);
731 srcu_read_unlock(&qdev->dev_lock, qdev_rcu_id);
732 srcu_read_unlock(&usr->qddev_lock, usr_rcu_id);
737 drm_gem_object_put(obj);
739 srcu_read_unlock(&qdev->dev_lock, qdev_rcu_id);
741 srcu_read_unlock(&usr->qddev_lock, usr_rcu_id);
745 int qaic_mmap_bo_ioctl(struct drm_device *dev, void *data, struct drm_file *file_priv)
747 struct qaic_mmap_bo *args = data;
748 int usr_rcu_id, qdev_rcu_id;
749 struct drm_gem_object *obj;
750 struct qaic_device *qdev;
751 struct qaic_user *usr;
754 usr = file_priv->driver_priv;
755 usr_rcu_id = srcu_read_lock(&usr->qddev_lock);
758 goto unlock_usr_srcu;
761 qdev = usr->qddev->qdev;
762 qdev_rcu_id = srcu_read_lock(&qdev->dev_lock);
763 if (qdev->dev_state != QAIC_ONLINE) {
765 goto unlock_dev_srcu;
768 obj = drm_gem_object_lookup(file_priv, args->handle);
771 goto unlock_dev_srcu;
774 args->offset = drm_vma_node_offset_addr(&obj->vma_node);
776 drm_gem_object_put(obj);
779 srcu_read_unlock(&qdev->dev_lock, qdev_rcu_id);
781 srcu_read_unlock(&usr->qddev_lock, usr_rcu_id);
785 struct drm_gem_object *qaic_gem_prime_import(struct drm_device *dev, struct dma_buf *dma_buf)
787 struct dma_buf_attachment *attach;
788 struct drm_gem_object *obj;
792 bo = qaic_alloc_init_bo();
799 get_dma_buf(dma_buf);
801 attach = dma_buf_attach(dma_buf, dev->dev);
802 if (IS_ERR(attach)) {
803 ret = PTR_ERR(attach);
807 if (!attach->dmabuf->size) {
809 goto size_align_fail;
812 drm_gem_private_object_init(dev, obj, attach->dmabuf->size);
814 * skipping dma_buf_map_attachment() as we do not know the direction
815 * just yet. Once the direction is known in the subsequent IOCTL to
816 * attach slicing, we can do it then.
819 obj->funcs = &qaic_gem_funcs;
820 obj->import_attach = attach;
821 obj->resv = dma_buf->resv;
826 dma_buf_detach(dma_buf, attach);
828 dma_buf_put(dma_buf);
834 static int qaic_prepare_import_bo(struct qaic_bo *bo, struct qaic_attach_slice_hdr *hdr)
836 struct drm_gem_object *obj = &bo->base;
837 struct sg_table *sgt;
840 sgt = dma_buf_map_attachment(obj->import_attach, hdr->dir);
851 static int qaic_prepare_export_bo(struct qaic_device *qdev, struct qaic_bo *bo,
852 struct qaic_attach_slice_hdr *hdr)
856 ret = dma_map_sgtable(&qdev->pdev->dev, bo->sgt, hdr->dir, 0);
863 static int qaic_prepare_bo(struct qaic_device *qdev, struct qaic_bo *bo,
864 struct qaic_attach_slice_hdr *hdr)
868 if (bo->base.import_attach)
869 ret = qaic_prepare_import_bo(bo, hdr);
871 ret = qaic_prepare_export_bo(qdev, bo, hdr);
873 bo->dbc = &qdev->dbc[hdr->dbc_id];
874 bo->nr_slice = hdr->count;
879 static void qaic_unprepare_import_bo(struct qaic_bo *bo)
881 dma_buf_unmap_attachment(bo->base.import_attach, bo->sgt, bo->dir);
885 static void qaic_unprepare_export_bo(struct qaic_device *qdev, struct qaic_bo *bo)
887 dma_unmap_sgtable(&qdev->pdev->dev, bo->sgt, bo->dir, 0);
890 static void qaic_unprepare_bo(struct qaic_device *qdev, struct qaic_bo *bo)
892 if (bo->base.import_attach)
893 qaic_unprepare_import_bo(bo);
895 qaic_unprepare_export_bo(qdev, bo);
902 static void qaic_free_slices_bo(struct qaic_bo *bo)
904 struct bo_slice *slice, *temp;
906 list_for_each_entry_safe(slice, temp, &bo->slices, slice)
907 kref_put(&slice->ref_count, free_slice);
908 if (WARN_ON_ONCE(bo->total_slice_nents != 0))
909 bo->total_slice_nents = 0;
913 static int qaic_attach_slicing_bo(struct qaic_device *qdev, struct qaic_bo *bo,
914 struct qaic_attach_slice_hdr *hdr,
915 struct qaic_attach_slice_entry *slice_ent)
919 for (i = 0; i < hdr->count; i++) {
920 ret = qaic_map_one_slice(qdev, bo, &slice_ent[i]);
922 qaic_free_slices_bo(bo);
927 if (bo->total_slice_nents > bo->dbc->nelem) {
928 qaic_free_slices_bo(bo);
935 int qaic_attach_slice_bo_ioctl(struct drm_device *dev, void *data, struct drm_file *file_priv)
937 struct qaic_attach_slice_entry *slice_ent;
938 struct qaic_attach_slice *args = data;
939 int rcu_id, usr_rcu_id, qdev_rcu_id;
940 struct dma_bridge_chan *dbc;
941 struct drm_gem_object *obj;
942 struct qaic_device *qdev;
943 unsigned long arg_size;
944 struct qaic_user *usr;
945 u8 __user *user_data;
949 if (args->hdr.count == 0)
952 arg_size = args->hdr.count * sizeof(*slice_ent);
953 if (arg_size / args->hdr.count != sizeof(*slice_ent))
956 if (!(args->hdr.dir == DMA_TO_DEVICE || args->hdr.dir == DMA_FROM_DEVICE))
962 usr = file_priv->driver_priv;
963 usr_rcu_id = srcu_read_lock(&usr->qddev_lock);
966 goto unlock_usr_srcu;
969 qdev = usr->qddev->qdev;
970 qdev_rcu_id = srcu_read_lock(&qdev->dev_lock);
971 if (qdev->dev_state != QAIC_ONLINE) {
973 goto unlock_dev_srcu;
976 if (args->hdr.dbc_id >= qdev->num_dbc) {
978 goto unlock_dev_srcu;
981 user_data = u64_to_user_ptr(args->data);
983 slice_ent = kzalloc(arg_size, GFP_KERNEL);
986 goto unlock_dev_srcu;
989 ret = copy_from_user(slice_ent, user_data, arg_size);
995 obj = drm_gem_object_lookup(file_priv, args->hdr.handle);
1001 ret = qaic_validate_req(qdev, slice_ent, args->hdr.count, obj->size);
1005 bo = to_qaic_bo(obj);
1006 ret = mutex_lock_interruptible(&bo->lock);
1015 dbc = &qdev->dbc[args->hdr.dbc_id];
1016 rcu_id = srcu_read_lock(&dbc->ch_lock);
1017 if (dbc->usr != usr) {
1019 goto unlock_ch_srcu;
1022 ret = qaic_prepare_bo(qdev, bo, &args->hdr);
1024 goto unlock_ch_srcu;
1026 ret = qaic_attach_slicing_bo(qdev, bo, &args->hdr, slice_ent);
1030 if (args->hdr.dir == DMA_TO_DEVICE)
1031 dma_sync_sgtable_for_cpu(&qdev->pdev->dev, bo->sgt, args->hdr.dir);
1034 list_add_tail(&bo->bo_list, &bo->dbc->bo_lists);
1035 srcu_read_unlock(&dbc->ch_lock, rcu_id);
1036 mutex_unlock(&bo->lock);
1038 srcu_read_unlock(&qdev->dev_lock, qdev_rcu_id);
1039 srcu_read_unlock(&usr->qddev_lock, usr_rcu_id);
1044 qaic_unprepare_bo(qdev, bo);
1046 srcu_read_unlock(&dbc->ch_lock, rcu_id);
1048 mutex_unlock(&bo->lock);
1050 drm_gem_object_put(obj);
1054 srcu_read_unlock(&qdev->dev_lock, qdev_rcu_id);
1056 srcu_read_unlock(&usr->qddev_lock, usr_rcu_id);
1060 static inline u32 fifo_space_avail(u32 head, u32 tail, u32 q_size)
1062 u32 avail = head - tail - 1;
1070 static inline int copy_exec_reqs(struct qaic_device *qdev, struct bo_slice *slice, u32 dbc_id,
1071 u32 head, u32 *ptail)
1073 struct dma_bridge_chan *dbc = &qdev->dbc[dbc_id];
1074 struct dbc_req *reqs = slice->reqs;
1078 avail = fifo_space_avail(head, tail, dbc->nelem);
1079 if (avail < slice->nents)
1082 if (tail + slice->nents > dbc->nelem) {
1083 avail = dbc->nelem - tail;
1084 avail = min_t(u32, avail, slice->nents);
1085 memcpy(fifo_at(dbc->req_q_base, tail), reqs, sizeof(*reqs) * avail);
1087 avail = slice->nents - avail;
1089 memcpy(dbc->req_q_base, reqs, sizeof(*reqs) * avail);
1091 memcpy(fifo_at(dbc->req_q_base, tail), reqs, sizeof(*reqs) * slice->nents);
1094 *ptail = (tail + slice->nents) % dbc->nelem;
1099 static inline int copy_partial_exec_reqs(struct qaic_device *qdev, struct bo_slice *slice,
1100 u64 resize, struct dma_bridge_chan *dbc, u32 head,
1103 struct dbc_req *reqs = slice->reqs;
1104 struct dbc_req *last_req;
1110 avail = fifo_space_avail(head, tail, dbc->nelem);
1113 * After this for loop is complete, first_n represents the index
1114 * of the last DMA request of this slice that needs to be
1115 * transferred after resizing and last_bytes represents DMA size
1118 last_bytes = resize;
1119 for (first_n = 0; first_n < slice->nents; first_n++)
1120 if (last_bytes > le32_to_cpu(reqs[first_n].len))
1121 last_bytes -= le32_to_cpu(reqs[first_n].len);
1125 if (avail < (first_n + 1))
1129 if (tail + first_n > dbc->nelem) {
1130 avail = dbc->nelem - tail;
1131 avail = min_t(u32, avail, first_n);
1132 memcpy(fifo_at(dbc->req_q_base, tail), reqs, sizeof(*reqs) * avail);
1133 last_req = reqs + avail;
1134 avail = first_n - avail;
1136 memcpy(dbc->req_q_base, last_req, sizeof(*reqs) * avail);
1138 memcpy(fifo_at(dbc->req_q_base, tail), reqs, sizeof(*reqs) * first_n);
1143 * Copy over the last entry. Here we need to adjust len to the left over
1144 * size, and set src and dst to the entry it is copied to.
1146 last_req = fifo_at(dbc->req_q_base, (tail + first_n) % dbc->nelem);
1147 memcpy(last_req, reqs + slice->nents - 1, sizeof(*reqs));
1150 * last_bytes holds size of a DMA segment, maximum DMA segment size is
1151 * set to UINT_MAX by qaic and hence last_bytes can never exceed u32
1152 * range. So, by down sizing we are not corrupting the value.
1154 last_req->len = cpu_to_le32((u32)last_bytes);
1155 last_req->src_addr = reqs[first_n].src_addr;
1156 last_req->dest_addr = reqs[first_n].dest_addr;
1158 /* Disable DMA transfer */
1159 last_req->cmd = GENMASK(7, 2) & reqs[first_n].cmd;
1161 *ptail = (tail + first_n + 1) % dbc->nelem;
1166 static int send_bo_list_to_device(struct qaic_device *qdev, struct drm_file *file_priv,
1167 struct qaic_execute_entry *exec, unsigned int count,
1168 bool is_partial, struct dma_bridge_chan *dbc, u32 head,
1171 struct qaic_partial_execute_entry *pexec = (struct qaic_partial_execute_entry *)exec;
1172 struct drm_gem_object *obj;
1173 struct bo_slice *slice;
1174 unsigned long flags;
1179 for (i = 0; i < count; i++) {
1181 * ref count will be decremented when the transfer of this
1182 * buffer is complete. It is inside dbc_irq_threaded_fn().
1184 obj = drm_gem_object_lookup(file_priv,
1185 is_partial ? pexec[i].handle : exec[i].handle);
1188 goto failed_to_send_bo;
1191 bo = to_qaic_bo(obj);
1192 ret = mutex_lock_interruptible(&bo->lock);
1194 goto failed_to_send_bo;
1201 if (is_partial && pexec[i].resize > bo->base.size) {
1206 spin_lock_irqsave(&dbc->xfer_lock, flags);
1207 if (bo_queued(bo)) {
1208 spin_unlock_irqrestore(&dbc->xfer_lock, flags);
1213 bo->req_id = dbc->next_req_id++;
1215 list_for_each_entry(slice, &bo->slices, slice) {
1216 for (j = 0; j < slice->nents; j++)
1217 slice->reqs[j].req_id = cpu_to_le16(bo->req_id);
1219 if (is_partial && (!pexec[i].resize || pexec[i].resize <= slice->offset))
1220 /* Configure the slice for no DMA transfer */
1221 ret = copy_partial_exec_reqs(qdev, slice, 0, dbc, head, tail);
1222 else if (is_partial && pexec[i].resize < slice->offset + slice->size)
1223 /* Configure the slice to be partially DMA transferred */
1224 ret = copy_partial_exec_reqs(qdev, slice,
1225 pexec[i].resize - slice->offset, dbc,
1228 ret = copy_exec_reqs(qdev, slice, dbc->id, head, tail);
1230 spin_unlock_irqrestore(&dbc->xfer_lock, flags);
1234 reinit_completion(&bo->xfer_done);
1235 list_add_tail(&bo->xfer_list, &dbc->xfer_list);
1236 spin_unlock_irqrestore(&dbc->xfer_lock, flags);
1237 dma_sync_sgtable_for_device(&qdev->pdev->dev, bo->sgt, bo->dir);
1238 mutex_unlock(&bo->lock);
1244 mutex_unlock(&bo->lock);
1247 drm_gem_object_put(obj);
1248 for (j = 0; j < i; j++) {
1249 spin_lock_irqsave(&dbc->xfer_lock, flags);
1250 bo = list_last_entry(&dbc->xfer_list, struct qaic_bo, xfer_list);
1252 list_del_init(&bo->xfer_list);
1253 spin_unlock_irqrestore(&dbc->xfer_lock, flags);
1254 dma_sync_sgtable_for_cpu(&qdev->pdev->dev, bo->sgt, bo->dir);
1255 drm_gem_object_put(obj);
1260 static void update_profiling_data(struct drm_file *file_priv,
1261 struct qaic_execute_entry *exec, unsigned int count,
1262 bool is_partial, u64 received_ts, u64 submit_ts, u32 queue_level)
1264 struct qaic_partial_execute_entry *pexec = (struct qaic_partial_execute_entry *)exec;
1265 struct drm_gem_object *obj;
1269 for (i = 0; i < count; i++) {
1271 * Since we already committed the BO to hardware, the only way
1272 * this should fail is a pending signal. We can't cancel the
1273 * submit to hardware, so we have to just skip the profiling
1274 * data. In case the signal is not fatal to the process, we
1275 * return success so that the user doesn't try to resubmit.
1277 obj = drm_gem_object_lookup(file_priv,
1278 is_partial ? pexec[i].handle : exec[i].handle);
1281 bo = to_qaic_bo(obj);
1282 bo->perf_stats.req_received_ts = received_ts;
1283 bo->perf_stats.req_submit_ts = submit_ts;
1284 bo->perf_stats.queue_level_before = queue_level;
1285 queue_level += bo->total_slice_nents;
1286 drm_gem_object_put(obj);
1290 static int __qaic_execute_bo_ioctl(struct drm_device *dev, void *data, struct drm_file *file_priv,
1293 struct qaic_execute *args = data;
1294 struct qaic_execute_entry *exec;
1295 struct dma_bridge_chan *dbc;
1296 int usr_rcu_id, qdev_rcu_id;
1297 struct qaic_device *qdev;
1298 struct qaic_user *usr;
1299 u8 __user *user_data;
1310 received_ts = ktime_get_ns();
1312 size = is_partial ? sizeof(struct qaic_partial_execute_entry) : sizeof(*exec);
1313 n = (unsigned long)size * args->hdr.count;
1314 if (args->hdr.count == 0 || n / args->hdr.count != size)
1317 user_data = u64_to_user_ptr(args->data);
1319 exec = kcalloc(args->hdr.count, size, GFP_KERNEL);
1323 if (copy_from_user(exec, user_data, n)) {
1328 usr = file_priv->driver_priv;
1329 usr_rcu_id = srcu_read_lock(&usr->qddev_lock);
1332 goto unlock_usr_srcu;
1335 qdev = usr->qddev->qdev;
1336 qdev_rcu_id = srcu_read_lock(&qdev->dev_lock);
1337 if (qdev->dev_state != QAIC_ONLINE) {
1339 goto unlock_dev_srcu;
1342 if (args->hdr.dbc_id >= qdev->num_dbc) {
1344 goto unlock_dev_srcu;
1347 dbc = &qdev->dbc[args->hdr.dbc_id];
1349 rcu_id = srcu_read_lock(&dbc->ch_lock);
1350 if (!dbc->usr || dbc->usr->handle != usr->handle) {
1352 goto release_ch_rcu;
1355 head = readl(dbc->dbc_base + REQHP_OFF);
1356 tail = readl(dbc->dbc_base + REQTP_OFF);
1358 if (head == U32_MAX || tail == U32_MAX) {
1359 /* PCI link error */
1361 goto release_ch_rcu;
1364 queue_level = head <= tail ? tail - head : dbc->nelem - (head - tail);
1366 ret = send_bo_list_to_device(qdev, file_priv, exec, args->hdr.count, is_partial, dbc,
1369 goto release_ch_rcu;
1371 /* Finalize commit to hardware */
1372 submit_ts = ktime_get_ns();
1373 writel(tail, dbc->dbc_base + REQTP_OFF);
1375 update_profiling_data(file_priv, exec, args->hdr.count, is_partial, received_ts,
1376 submit_ts, queue_level);
1378 if (datapath_polling)
1379 schedule_work(&dbc->poll_work);
1382 srcu_read_unlock(&dbc->ch_lock, rcu_id);
1384 srcu_read_unlock(&qdev->dev_lock, qdev_rcu_id);
1386 srcu_read_unlock(&usr->qddev_lock, usr_rcu_id);
1392 int qaic_execute_bo_ioctl(struct drm_device *dev, void *data, struct drm_file *file_priv)
1394 return __qaic_execute_bo_ioctl(dev, data, file_priv, false);
1397 int qaic_partial_execute_bo_ioctl(struct drm_device *dev, void *data, struct drm_file *file_priv)
1399 return __qaic_execute_bo_ioctl(dev, data, file_priv, true);
1403 * Our interrupt handling is a bit more complicated than a simple ideal, but
1406 * Each dbc has a completion queue. Entries in the queue correspond to DMA
1407 * requests which the device has processed. The hardware already has a built
1408 * in irq mitigation. When the device puts an entry into the queue, it will
1409 * only trigger an interrupt if the queue was empty. Therefore, when adding
1410 * the Nth event to a non-empty queue, the hardware doesn't trigger an
1411 * interrupt. This means the host doesn't get additional interrupts signaling
1412 * the same thing - the queue has something to process.
1413 * This behavior can be overridden in the DMA request.
1414 * This means that when the host receives an interrupt, it is required to
1417 * This behavior is what NAPI attempts to accomplish, although we can't use
1418 * NAPI as we don't have a netdev. We use threaded irqs instead.
1420 * However, there is a situation where the host drains the queue fast enough
1421 * that every event causes an interrupt. Typically this is not a problem as
1422 * the rate of events would be low. However, that is not the case with
1423 * lprnet for example. On an Intel Xeon D-2191 where we run 8 instances of
1424 * lprnet, the host receives roughly 80k interrupts per second from the device
1425 * (per /proc/interrupts). While NAPI documentation indicates the host should
1426 * just chug along, sadly that behavior causes instability in some hosts.
1428 * Therefore, we implement an interrupt disable scheme similar to NAPI. The
1429 * key difference is that we will delay after draining the queue for a small
1430 * time to allow additional events to come in via polling. Using the above
1431 * lprnet workload, this reduces the number of interrupts processed from
1432 * ~80k/sec to about 64 in 5 minutes and appears to solve the system
1435 irqreturn_t dbc_irq_handler(int irq, void *data)
1437 struct dma_bridge_chan *dbc = data;
1442 rcu_id = srcu_read_lock(&dbc->ch_lock);
1444 if (datapath_polling) {
1445 srcu_read_unlock(&dbc->ch_lock, rcu_id);
1447 * Normally datapath_polling will not have irqs enabled, but
1448 * when running with only one MSI the interrupt is shared with
1449 * MHI so it cannot be disabled. Return ASAP instead.
1455 srcu_read_unlock(&dbc->ch_lock, rcu_id);
1459 head = readl(dbc->dbc_base + RSPHP_OFF);
1460 if (head == U32_MAX) { /* PCI link error */
1461 srcu_read_unlock(&dbc->ch_lock, rcu_id);
1465 tail = readl(dbc->dbc_base + RSPTP_OFF);
1466 if (tail == U32_MAX) { /* PCI link error */
1467 srcu_read_unlock(&dbc->ch_lock, rcu_id);
1471 if (head == tail) { /* queue empty */
1472 srcu_read_unlock(&dbc->ch_lock, rcu_id);
1476 if (!dbc->qdev->single_msi)
1477 disable_irq_nosync(irq);
1478 srcu_read_unlock(&dbc->ch_lock, rcu_id);
1479 return IRQ_WAKE_THREAD;
1482 void irq_polling_work(struct work_struct *work)
1484 struct dma_bridge_chan *dbc = container_of(work, struct dma_bridge_chan, poll_work);
1485 unsigned long flags;
1490 rcu_id = srcu_read_lock(&dbc->ch_lock);
1493 if (dbc->qdev->dev_state != QAIC_ONLINE) {
1494 srcu_read_unlock(&dbc->ch_lock, rcu_id);
1498 srcu_read_unlock(&dbc->ch_lock, rcu_id);
1501 spin_lock_irqsave(&dbc->xfer_lock, flags);
1502 if (list_empty(&dbc->xfer_list)) {
1503 spin_unlock_irqrestore(&dbc->xfer_lock, flags);
1504 srcu_read_unlock(&dbc->ch_lock, rcu_id);
1507 spin_unlock_irqrestore(&dbc->xfer_lock, flags);
1509 head = readl(dbc->dbc_base + RSPHP_OFF);
1510 if (head == U32_MAX) { /* PCI link error */
1511 srcu_read_unlock(&dbc->ch_lock, rcu_id);
1515 tail = readl(dbc->dbc_base + RSPTP_OFF);
1516 if (tail == U32_MAX) { /* PCI link error */
1517 srcu_read_unlock(&dbc->ch_lock, rcu_id);
1522 irq_wake_thread(dbc->irq, dbc);
1523 srcu_read_unlock(&dbc->ch_lock, rcu_id);
1528 usleep_range(datapath_poll_interval_us, 2 * datapath_poll_interval_us);
1532 irqreturn_t dbc_irq_threaded_fn(int irq, void *data)
1534 struct dma_bridge_chan *dbc = data;
1535 int event_count = NUM_EVENTS;
1536 int delay_count = NUM_DELAYS;
1537 struct qaic_device *qdev;
1538 struct qaic_bo *bo, *i;
1539 struct dbc_rsp *rsp;
1540 unsigned long flags;
1547 rcu_id = srcu_read_lock(&dbc->ch_lock);
1550 head = readl(dbc->dbc_base + RSPHP_OFF);
1551 if (head == U32_MAX) /* PCI link error */
1557 event_count = NUM_EVENTS;
1562 * if this channel isn't assigned or gets unassigned during processing
1563 * we have nothing further to do
1568 tail = readl(dbc->dbc_base + RSPTP_OFF);
1569 if (tail == U32_MAX) /* PCI link error */
1572 if (head == tail) { /* queue empty */
1575 usleep_range(100, 200);
1576 goto read_fifo; /* check for a new event */
1581 delay_count = NUM_DELAYS;
1582 while (head != tail) {
1586 rsp = dbc->rsp_q_base + head * sizeof(*rsp);
1587 req_id = le16_to_cpu(rsp->req_id);
1588 status = le16_to_cpu(rsp->status);
1590 pci_dbg(qdev->pdev, "req_id %d failed with status %d\n", req_id, status);
1591 spin_lock_irqsave(&dbc->xfer_lock, flags);
1593 * A BO can receive multiple interrupts, since a BO can be
1594 * divided into multiple slices and a buffer receives as many
1595 * interrupts as slices. So until it receives interrupts for
1596 * all the slices we cannot mark that buffer complete.
1598 list_for_each_entry_safe(bo, i, &dbc->xfer_list, xfer_list) {
1599 if (bo->req_id == req_id)
1600 bo->nr_slice_xfer_done++;
1604 if (bo->nr_slice_xfer_done < bo->nr_slice)
1608 * At this point we have received all the interrupts for
1609 * BO, which means BO execution is complete.
1611 dma_sync_sgtable_for_cpu(&qdev->pdev->dev, bo->sgt, bo->dir);
1612 bo->nr_slice_xfer_done = 0;
1613 list_del_init(&bo->xfer_list);
1614 bo->perf_stats.req_processed_ts = ktime_get_ns();
1615 complete_all(&bo->xfer_done);
1616 drm_gem_object_put(&bo->base);
1619 spin_unlock_irqrestore(&dbc->xfer_lock, flags);
1620 head = (head + 1) % dbc->nelem;
1624 * Update the head pointer of response queue and let the device know
1625 * that we have consumed elements from the queue.
1627 writel(head, dbc->dbc_base + RSPHP_OFF);
1629 /* elements might have been put in the queue while we were processing */
1633 if (!qdev->single_msi && likely(!datapath_polling))
1635 else if (unlikely(datapath_polling))
1636 schedule_work(&dbc->poll_work);
1637 /* checking the fifo and enabling irqs is a race, missed event check */
1638 tail = readl(dbc->dbc_base + RSPTP_OFF);
1639 if (tail != U32_MAX && head != tail) {
1640 if (!qdev->single_msi && likely(!datapath_polling))
1641 disable_irq_nosync(irq);
1644 srcu_read_unlock(&dbc->ch_lock, rcu_id);
1648 srcu_read_unlock(&dbc->ch_lock, rcu_id);
1649 if (!qdev->single_msi && likely(!datapath_polling))
1651 else if (unlikely(datapath_polling))
1652 schedule_work(&dbc->poll_work);
1657 int qaic_wait_bo_ioctl(struct drm_device *dev, void *data, struct drm_file *file_priv)
1659 struct qaic_wait *args = data;
1660 int usr_rcu_id, qdev_rcu_id;
1661 struct dma_bridge_chan *dbc;
1662 struct drm_gem_object *obj;
1663 struct qaic_device *qdev;
1664 unsigned long timeout;
1665 struct qaic_user *usr;
1673 usr = file_priv->driver_priv;
1674 usr_rcu_id = srcu_read_lock(&usr->qddev_lock);
1677 goto unlock_usr_srcu;
1680 qdev = usr->qddev->qdev;
1681 qdev_rcu_id = srcu_read_lock(&qdev->dev_lock);
1682 if (qdev->dev_state != QAIC_ONLINE) {
1684 goto unlock_dev_srcu;
1687 if (args->dbc_id >= qdev->num_dbc) {
1689 goto unlock_dev_srcu;
1692 dbc = &qdev->dbc[args->dbc_id];
1694 rcu_id = srcu_read_lock(&dbc->ch_lock);
1695 if (dbc->usr != usr) {
1697 goto unlock_ch_srcu;
1700 obj = drm_gem_object_lookup(file_priv, args->handle);
1703 goto unlock_ch_srcu;
1706 bo = to_qaic_bo(obj);
1707 timeout = args->timeout ? args->timeout : wait_exec_default_timeout_ms;
1708 timeout = msecs_to_jiffies(timeout);
1709 ret = wait_for_completion_interruptible_timeout(&bo->xfer_done, timeout);
1721 drm_gem_object_put(obj);
1723 srcu_read_unlock(&dbc->ch_lock, rcu_id);
1725 srcu_read_unlock(&qdev->dev_lock, qdev_rcu_id);
1727 srcu_read_unlock(&usr->qddev_lock, usr_rcu_id);
1731 int qaic_perf_stats_bo_ioctl(struct drm_device *dev, void *data, struct drm_file *file_priv)
1733 struct qaic_perf_stats_entry *ent = NULL;
1734 struct qaic_perf_stats *args = data;
1735 int usr_rcu_id, qdev_rcu_id;
1736 struct drm_gem_object *obj;
1737 struct qaic_device *qdev;
1738 struct qaic_user *usr;
1742 usr = file_priv->driver_priv;
1743 usr_rcu_id = srcu_read_lock(&usr->qddev_lock);
1746 goto unlock_usr_srcu;
1749 qdev = usr->qddev->qdev;
1750 qdev_rcu_id = srcu_read_lock(&qdev->dev_lock);
1751 if (qdev->dev_state != QAIC_ONLINE) {
1753 goto unlock_dev_srcu;
1756 if (args->hdr.dbc_id >= qdev->num_dbc) {
1758 goto unlock_dev_srcu;
1761 ent = kcalloc(args->hdr.count, sizeof(*ent), GFP_KERNEL);
1764 goto unlock_dev_srcu;
1767 ret = copy_from_user(ent, u64_to_user_ptr(args->data), args->hdr.count * sizeof(*ent));
1773 for (i = 0; i < args->hdr.count; i++) {
1774 obj = drm_gem_object_lookup(file_priv, ent[i].handle);
1779 bo = to_qaic_bo(obj);
1781 * perf stats ioctl is called before wait ioctl is complete then
1782 * the latency information is invalid.
1784 if (bo->perf_stats.req_processed_ts < bo->perf_stats.req_submit_ts) {
1785 ent[i].device_latency_us = 0;
1787 ent[i].device_latency_us = div_u64((bo->perf_stats.req_processed_ts -
1788 bo->perf_stats.req_submit_ts), 1000);
1790 ent[i].submit_latency_us = div_u64((bo->perf_stats.req_submit_ts -
1791 bo->perf_stats.req_received_ts), 1000);
1792 ent[i].queue_level_before = bo->perf_stats.queue_level_before;
1793 ent[i].num_queue_element = bo->total_slice_nents;
1794 drm_gem_object_put(obj);
1797 if (copy_to_user(u64_to_user_ptr(args->data), ent, args->hdr.count * sizeof(*ent)))
1803 srcu_read_unlock(&qdev->dev_lock, qdev_rcu_id);
1805 srcu_read_unlock(&usr->qddev_lock, usr_rcu_id);
1809 static void detach_slice_bo(struct qaic_device *qdev, struct qaic_bo *bo)
1811 qaic_free_slices_bo(bo);
1812 qaic_unprepare_bo(qdev, bo);
1813 qaic_init_bo(bo, true);
1814 list_del(&bo->bo_list);
1815 drm_gem_object_put(&bo->base);
1818 int qaic_detach_slice_bo_ioctl(struct drm_device *dev, void *data, struct drm_file *file_priv)
1820 struct qaic_detach_slice *args = data;
1821 int rcu_id, usr_rcu_id, qdev_rcu_id;
1822 struct dma_bridge_chan *dbc;
1823 struct drm_gem_object *obj;
1824 struct qaic_device *qdev;
1825 struct qaic_user *usr;
1826 unsigned long flags;
1833 usr = file_priv->driver_priv;
1834 usr_rcu_id = srcu_read_lock(&usr->qddev_lock);
1837 goto unlock_usr_srcu;
1840 qdev = usr->qddev->qdev;
1841 qdev_rcu_id = srcu_read_lock(&qdev->dev_lock);
1842 if (qdev->dev_state != QAIC_ONLINE) {
1844 goto unlock_dev_srcu;
1847 obj = drm_gem_object_lookup(file_priv, args->handle);
1850 goto unlock_dev_srcu;
1853 bo = to_qaic_bo(obj);
1854 ret = mutex_lock_interruptible(&bo->lock);
1864 rcu_id = srcu_read_lock(&dbc->ch_lock);
1865 if (dbc->usr != usr) {
1867 goto unlock_ch_srcu;
1870 /* Check if BO is committed to H/W for DMA */
1871 spin_lock_irqsave(&dbc->xfer_lock, flags);
1872 if (bo_queued(bo)) {
1873 spin_unlock_irqrestore(&dbc->xfer_lock, flags);
1875 goto unlock_ch_srcu;
1877 spin_unlock_irqrestore(&dbc->xfer_lock, flags);
1879 detach_slice_bo(qdev, bo);
1882 srcu_read_unlock(&dbc->ch_lock, rcu_id);
1884 mutex_unlock(&bo->lock);
1886 drm_gem_object_put(obj);
1888 srcu_read_unlock(&qdev->dev_lock, qdev_rcu_id);
1890 srcu_read_unlock(&usr->qddev_lock, usr_rcu_id);
1894 static void empty_xfer_list(struct qaic_device *qdev, struct dma_bridge_chan *dbc)
1896 unsigned long flags;
1899 spin_lock_irqsave(&dbc->xfer_lock, flags);
1900 while (!list_empty(&dbc->xfer_list)) {
1901 bo = list_first_entry(&dbc->xfer_list, typeof(*bo), xfer_list);
1902 list_del_init(&bo->xfer_list);
1903 spin_unlock_irqrestore(&dbc->xfer_lock, flags);
1904 bo->nr_slice_xfer_done = 0;
1906 bo->perf_stats.req_received_ts = 0;
1907 bo->perf_stats.req_submit_ts = 0;
1908 bo->perf_stats.req_processed_ts = 0;
1909 bo->perf_stats.queue_level_before = 0;
1910 dma_sync_sgtable_for_cpu(&qdev->pdev->dev, bo->sgt, bo->dir);
1911 complete_all(&bo->xfer_done);
1912 drm_gem_object_put(&bo->base);
1913 spin_lock_irqsave(&dbc->xfer_lock, flags);
1915 spin_unlock_irqrestore(&dbc->xfer_lock, flags);
1918 int disable_dbc(struct qaic_device *qdev, u32 dbc_id, struct qaic_user *usr)
1920 if (!qdev->dbc[dbc_id].usr || qdev->dbc[dbc_id].usr->handle != usr->handle)
1923 qdev->dbc[dbc_id].usr = NULL;
1924 synchronize_srcu(&qdev->dbc[dbc_id].ch_lock);
1929 * enable_dbc - Enable the DBC. DBCs are disabled by removing the context of
1930 * user. Add user context back to DBC to enable it. This function trusts the
1931 * DBC ID passed and expects the DBC to be disabled.
1932 * @qdev: Qranium device handle
1933 * @dbc_id: ID of the DBC
1934 * @usr: User context
1936 void enable_dbc(struct qaic_device *qdev, u32 dbc_id, struct qaic_user *usr)
1938 qdev->dbc[dbc_id].usr = usr;
1941 void wakeup_dbc(struct qaic_device *qdev, u32 dbc_id)
1943 struct dma_bridge_chan *dbc = &qdev->dbc[dbc_id];
1946 empty_xfer_list(qdev, dbc);
1947 synchronize_srcu(&dbc->ch_lock);
1949 * Threads holding channel lock, may add more elements in the xfer_list.
1950 * Flush out these elements from xfer_list.
1952 empty_xfer_list(qdev, dbc);
1955 void release_dbc(struct qaic_device *qdev, u32 dbc_id)
1957 struct qaic_bo *bo, *bo_temp;
1958 struct dma_bridge_chan *dbc;
1960 dbc = &qdev->dbc[dbc_id];
1964 wakeup_dbc(qdev, dbc_id);
1966 dma_free_coherent(&qdev->pdev->dev, dbc->total_size, dbc->req_q_base, dbc->dma_addr);
1967 dbc->total_size = 0;
1968 dbc->req_q_base = NULL;
1973 list_for_each_entry_safe(bo, bo_temp, &dbc->bo_lists, bo_list) {
1974 drm_gem_object_get(&bo->base);
1975 mutex_lock(&bo->lock);
1976 detach_slice_bo(qdev, bo);
1977 mutex_unlock(&bo->lock);
1978 drm_gem_object_put(&bo->base);
1981 dbc->in_use = false;
1982 wake_up(&dbc->dbc_release);