Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
kernel
os
linux
drm/panthor: Add a PANTHOR_BO_SYNC ioctl
This will be used by the UMD to synchronize CPU-cached mappings when
the UMD can't do it directly (no usermode cache maintenance instruction
on Arm32).
v2:
- Change the flags so they better match the drm_gem_shmem_sync()
semantics
v3:
- Add Steve's R-b
v4:
- No changes
v5:
- Drop Steve's R-b (the semantics changes call for a new review)
v6:
- Drop ret initialization in panthor_ioctl_bo_sync()
- Bail out early in panthor_ioctl_bo_sync() if ops.count is zero
- Drop unused PANTHOR_BO_SYNC_OP_FLAGS definition
v7:
- Hand-roll the sync logic (was previously provided by gem_shmem)
v8:
- Collect R-b
Signed-off-by: Faith Ekstrand <faith.ekstrand@collabora.com>
Reviewed-by: Steven Price <steven.price@arm.com>
Link: https://patch.msgid.link/20251208100841.730527-5-boris.brezillon@collabora.com
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
+179
-1
+40
-1
drivers/gpu/drm/panthor/panthor_drv.c
+40
-1
drivers/gpu/drm/panthor/panthor_drv.c
···
177
177
PANTHOR_UOBJ_DECL(struct drm_panthor_sync_op, timeline_value), \
178
178
PANTHOR_UOBJ_DECL(struct drm_panthor_queue_submit, syncs), \
179
179
PANTHOR_UOBJ_DECL(struct drm_panthor_queue_create, ringbuf_size), \
180
-
PANTHOR_UOBJ_DECL(struct drm_panthor_vm_bind_op, syncs))
180
+
PANTHOR_UOBJ_DECL(struct drm_panthor_vm_bind_op, syncs), \
181
+
PANTHOR_UOBJ_DECL(struct drm_panthor_bo_sync_op, size))
181
182
182
183
/**
183
184
* PANTHOR_UOBJ_SET() - Copy a kernel object to a user object.
···
1397
1396
return 0;
1398
1397
}
1399
1398
1399
+
static int panthor_ioctl_bo_sync(struct drm_device *ddev, void *data,
1400
+
struct drm_file *file)
1401
+
{
1402
+
struct drm_panthor_bo_sync *args = data;
1403
+
struct drm_panthor_bo_sync_op *ops;
1404
+
struct drm_gem_object *obj;
1405
+
int ret;
1406
+
1407
+
if (!args->ops.count)
1408
+
return 0;
1409
+
1410
+
ret = PANTHOR_UOBJ_GET_ARRAY(ops, &args->ops);
1411
+
if (ret)
1412
+
return ret;
1413
+
1414
+
for (u32 i = 0; i < args->ops.count; i++) {
1415
+
obj = drm_gem_object_lookup(file, ops[i].handle);
1416
+
if (!obj) {
1417
+
ret = -ENOENT;
1418
+
goto err_ops;
1419
+
}
1420
+
1421
+
ret = panthor_gem_sync(obj, ops[i].type, ops[i].offset,
1422
+
ops[i].size);
1423
+
1424
+
drm_gem_object_put(obj);
1425
+
1426
+
if (ret)
1427
+
goto err_ops;
1428
+
}
1429
+
1430
+
err_ops:
1431
+
kvfree(ops);
1432
+
1433
+
return ret;
1434
+
}
1435
+
1400
1436
static int
1401
1437
panthor_open(struct drm_device *ddev, struct drm_file *file)
1402
1438
{
···
1508
1470
PANTHOR_IOCTL(GROUP_SUBMIT, group_submit, DRM_RENDER_ALLOW),
1509
1471
PANTHOR_IOCTL(BO_SET_LABEL, bo_set_label, DRM_RENDER_ALLOW),
1510
1472
PANTHOR_IOCTL(SET_USER_MMIO_OFFSET, set_user_mmio_offset, DRM_RENDER_ALLOW),
1473
+
PANTHOR_IOCTL(BO_SYNC, bo_sync, DRM_RENDER_ALLOW),
1511
1474
};
1512
1475
1513
1476
static int panthor_mmap(struct file *filp, struct vm_area_struct *vma)
+85
drivers/gpu/drm/panthor/panthor_gem.c
+85
drivers/gpu/drm/panthor/panthor_gem.c
···
465
465
panthor_gem_bo_set_label(bo->obj, str);
466
466
}
467
467
468
+
int
469
+
panthor_gem_sync(struct drm_gem_object *obj, u32 type,
470
+
u64 offset, u64 size)
471
+
{
472
+
struct panthor_gem_object *bo = to_panthor_bo(obj);
473
+
struct drm_gem_shmem_object *shmem = &bo->base;
474
+
const struct drm_device *dev = shmem->base.dev;
475
+
struct sg_table *sgt;
476
+
struct scatterlist *sgl;
477
+
unsigned int count;
478
+
479
+
/* Make sure the range is in bounds. */
480
+
if (offset + size < offset || offset + size > shmem->base.size)
481
+
return -EINVAL;
482
+
483
+
/* Disallow CPU-cache maintenance on imported buffers. */
484
+
if (drm_gem_is_imported(&shmem->base))
485
+
return -EINVAL;
486
+
487
+
switch (type) {
488
+
case DRM_PANTHOR_BO_SYNC_CPU_CACHE_FLUSH:
489
+
case DRM_PANTHOR_BO_SYNC_CPU_CACHE_FLUSH_AND_INVALIDATE:
490
+
break;
491
+
492
+
default:
493
+
return -EINVAL;
494
+
}
495
+
496
+
/* Don't bother if it's WC-mapped */
497
+
if (shmem->map_wc)
498
+
return 0;
499
+
500
+
/* Nothing to do if the size is zero. */
501
+
if (size == 0)
502
+
return 0;
503
+
504
+
sgt = drm_gem_shmem_get_pages_sgt(shmem);
505
+
if (IS_ERR(sgt))
506
+
return PTR_ERR(sgt);
507
+
508
+
for_each_sgtable_dma_sg(sgt, sgl, count) {
509
+
if (size == 0)
510
+
break;
511
+
512
+
dma_addr_t paddr = sg_dma_address(sgl);
513
+
size_t len = sg_dma_len(sgl);
514
+
515
+
if (len <= offset) {
516
+
offset -= len;
517
+
continue;
518
+
}
519
+
520
+
paddr += offset;
521
+
len -= offset;
522
+
len = min_t(size_t, len, size);
523
+
size -= len;
524
+
offset = 0;
525
+
526
+
/* It's unclear whether dma_sync_xxx() is the right API to do CPU
527
+
* cache maintenance given an IOMMU can register their own
528
+
* implementation doing more than just CPU cache flushes/invalidation,
529
+
* and what we really care about here is CPU caches only, but that's
530
+
* the best we have that is both arch-agnostic and does at least the
531
+
* CPU cache maintenance on a <page,offset,size> tuple.
532
+
*
533
+
* Also, I wish we could do a single
534
+
*
535
+
* dma_sync_single_for_device(BIDIR)
536
+
*
537
+
* and get a flush+invalidate, but that's not how it's implemented
538
+
* in practice (at least on arm64), so we have to make it
539
+
*
540
+
* dma_sync_single_for_device(TO_DEVICE)
541
+
* dma_sync_single_for_cpu(FROM_DEVICE)
542
+
*
543
+
* for the flush+invalidate case.
544
+
*/
545
+
dma_sync_single_for_device(dev->dev, paddr, len, DMA_TO_DEVICE);
546
+
if (type == DRM_PANTHOR_BO_SYNC_CPU_CACHE_FLUSH_AND_INVALIDATE)
547
+
dma_sync_single_for_cpu(dev->dev, paddr, len, DMA_FROM_DEVICE);
548
+
}
549
+
550
+
return 0;
551
+
}
552
+
468
553
#ifdef CONFIG_DEBUG_FS
469
554
struct gem_size_totals {
470
555
size_t size;
+2
drivers/gpu/drm/panthor/panthor_gem.h
+2
drivers/gpu/drm/panthor/panthor_gem.h
···
148
148
149
149
void panthor_gem_bo_set_label(struct drm_gem_object *obj, const char *label);
150
150
void panthor_gem_kernel_bo_set_label(struct panthor_kernel_bo *bo, const char *label);
151
+
int panthor_gem_sync(struct drm_gem_object *obj,
152
+
u32 type, u64 offset, u64 size);
151
153
152
154
struct drm_gem_object *
153
155
panthor_gem_prime_import(struct drm_device *dev,
+52
include/uapi/drm/panthor_drm.h
+52
include/uapi/drm/panthor_drm.h
···
144
144
* pgoff_t size.
145
145
*/
146
146
DRM_PANTHOR_SET_USER_MMIO_OFFSET,
147
+
148
+
/** @DRM_PANTHOR_BO_SYNC: Sync BO data to/from the device */
149
+
DRM_PANTHOR_BO_SYNC,
147
150
};
148
151
149
152
/**
···
1077
1074
};
1078
1075
1079
1076
/**
1077
+
* enum drm_panthor_bo_sync_op_type - BO sync type
1078
+
*/
1079
+
enum drm_panthor_bo_sync_op_type {
1080
+
/** @DRM_PANTHOR_BO_SYNC_CPU_CACHE_FLUSH: Flush CPU caches. */
1081
+
DRM_PANTHOR_BO_SYNC_CPU_CACHE_FLUSH = 0,
1082
+
1083
+
/** @DRM_PANTHOR_BO_SYNC_CPU_CACHE_FLUSH_AND_INVALIDATE: Flush and invalidate CPU caches. */
1084
+
DRM_PANTHOR_BO_SYNC_CPU_CACHE_FLUSH_AND_INVALIDATE = 1,
1085
+
};
1086
+
1087
+
/**
1088
+
* struct drm_panthor_bo_sync_op - BO map sync op
1089
+
*/
1090
+
struct drm_panthor_bo_sync_op {
1091
+
/** @handle: Handle of the buffer object to sync. */
1092
+
__u32 handle;
1093
+
1094
+
/** @type: Type of operation. */
1095
+
__u32 type;
1096
+
1097
+
/**
1098
+
* @offset: Offset into the BO at which the sync range starts.
1099
+
*
1100
+
* This will be rounded down to the nearest cache line as needed.
1101
+
*/
1102
+
__u64 offset;
1103
+
1104
+
/**
1105
+
* @size: Size of the range to sync
1106
+
*
1107
+
* @size + @offset will be rounded up to the nearest cache line as
1108
+
* needed.
1109
+
*/
1110
+
__u64 size;
1111
+
};
1112
+
1113
+
/**
1114
+
* struct drm_panthor_bo_sync - BO map sync request
1115
+
*/
1116
+
struct drm_panthor_bo_sync {
1117
+
/**
1118
+
* @ops: Array of struct drm_panthor_bo_sync_op sync operations.
1119
+
*/
1120
+
struct drm_panthor_obj_array ops;
1121
+
};
1122
+
1123
+
/**
1080
1124
* DRM_IOCTL_PANTHOR() - Build a Panthor IOCTL number
1081
1125
* @__access: Access type. Must be R, W or RW.
1082
1126
* @__id: One of the DRM_PANTHOR_xxx id.
···
1169
1119
DRM_IOCTL_PANTHOR(WR, BO_SET_LABEL, bo_set_label),
1170
1120
DRM_IOCTL_PANTHOR_SET_USER_MMIO_OFFSET =
1171
1121
DRM_IOCTL_PANTHOR(WR, SET_USER_MMIO_OFFSET, set_user_mmio_offset),
1122
+
DRM_IOCTL_PANTHOR_BO_SYNC =
1123
+
DRM_IOCTL_PANTHOR(WR, BO_SYNC, bo_sync),
1172
1124
};
1173
1125
1174
1126
#if defined(__cplusplus)