Home | History | Annotate | Line # | Download | only in amdgpu
      1 /*	$NetBSD: amdgpu_bo_list.c,v 1.9 2021/12/19 12:02:39 riastradh Exp $	*/
      2 
      3 /*
      4  * Copyright 2015 Advanced Micro Devices, Inc.
      5  * All Rights Reserved.
      6  *
      7  * Permission is hereby granted, free of charge, to any person obtaining a
      8  * copy of this software and associated documentation files (the
      9  * "Software"), to deal in the Software without restriction, including
     10  * without limitation the rights to use, copy, modify, merge, publish,
     11  * distribute, sub license, and/or sell copies of the Software, and to
     12  * permit persons to whom the Software is furnished to do so, subject to
     13  * the following conditions:
     14  *
     15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
     16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
     17  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
     18  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
     19  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
     20  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
     21  * USE OR OTHER DEALINGS IN THE SOFTWARE.
     22  *
     23  * The above copyright notice and this permission notice (including the
     24  * next paragraph) shall be included in all copies or substantial portions
     25  * of the Software.
     26  *
     27  */
     28 /*
     29  * Authors:
     30  *    Christian Knig <deathsimple (at) vodafone.de>
     31  */
     32 
     33 #include <sys/cdefs.h>
     34 __KERNEL_RCSID(0, "$NetBSD: amdgpu_bo_list.c,v 1.9 2021/12/19 12:02:39 riastradh Exp $");
     35 
     36 #include <linux/uaccess.h>
     37 
     38 #include "amdgpu.h"
     39 #include "amdgpu_trace.h"
     40 
     41 #include <linux/nbsd-namespace.h>
     42 
     43 #define AMDGPU_BO_LIST_MAX_PRIORITY	32u
     44 #define AMDGPU_BO_LIST_NUM_BUCKETS	(AMDGPU_BO_LIST_MAX_PRIORITY + 1)
     45 
     46 static void amdgpu_bo_list_free_rcu(struct rcu_head *rcu)
     47 {
     48 	struct amdgpu_bo_list *list = container_of(rcu, struct amdgpu_bo_list,
     49 						   rhead);
     50 
     51 	kvfree(list);
     52 }
     53 
     54 static void amdgpu_bo_list_free(struct kref *ref)
     55 {
     56 	struct amdgpu_bo_list *list = container_of(ref, struct amdgpu_bo_list,
     57 						   refcount);
     58 	struct amdgpu_bo_list_entry *e;
     59 
     60 	amdgpu_bo_list_for_each_entry(e, list) {
     61 		struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
     62 
     63 		amdgpu_bo_unref(&bo);
     64 	}
     65 
     66 	call_rcu(&list->rhead, amdgpu_bo_list_free_rcu);
     67 }
     68 
     69 int amdgpu_bo_list_create(struct amdgpu_device *adev, struct drm_file *filp,
     70 			  struct drm_amdgpu_bo_list_entry *info,
     71 			  unsigned num_entries, struct amdgpu_bo_list **result)
     72 {
     73 	unsigned last_entry = 0, first_userptr = num_entries;
     74 	struct amdgpu_bo_list_entry *array;
     75 	struct amdgpu_bo_list *list;
     76 	uint64_t total_size = 0;
     77 	size_t size;
     78 	unsigned i;
     79 	int r;
     80 
     81 	if (num_entries > (SIZE_MAX - sizeof(struct amdgpu_bo_list))
     82 				/ sizeof(struct amdgpu_bo_list_entry))
     83 		return -EINVAL;
     84 
     85 	size = sizeof(struct amdgpu_bo_list);
     86 	size += num_entries * sizeof(struct amdgpu_bo_list_entry);
     87 	list = kvmalloc(size, GFP_KERNEL);
     88 	if (!list)
     89 		return -ENOMEM;
     90 
     91 	kref_init(&list->refcount);
     92 	list->gds_obj = NULL;
     93 	list->gws_obj = NULL;
     94 	list->oa_obj = NULL;
     95 
     96 	array = amdgpu_bo_list_array_entry(list, 0);
     97 	memset(array, 0, num_entries * sizeof(struct amdgpu_bo_list_entry));
     98 
     99 	for (i = 0; i < num_entries; ++i) {
    100 		struct amdgpu_bo_list_entry *entry;
    101 		struct drm_gem_object *gobj;
    102 		struct amdgpu_bo *bo;
    103 #ifdef __NetBSD__
    104 		struct vmspace *usermm;
    105 #else
    106 		struct mm_struct *usermm;
    107 #endif
    108 
    109 		gobj = drm_gem_object_lookup(filp, info[i].bo_handle);
    110 		if (!gobj) {
    111 			r = -ENOENT;
    112 			goto error_free;
    113 		}
    114 
    115 		bo = amdgpu_bo_ref(gem_to_amdgpu_bo(gobj));
    116 		drm_gem_object_put_unlocked(gobj);
    117 
    118 		usermm = amdgpu_ttm_tt_get_usermm(bo->tbo.ttm);
    119 		if (usermm) {
    120 #ifdef __NetBSD__		/* XXX amdgpu userptr */
    121 			amdgpu_bo_unref(&bo);
    122 			r = -EPERM;
    123 			goto error_free;
    124 #else
    125 			if (usermm != current->mm) {
    126 				amdgpu_bo_unref(&bo);
    127 				r = -EPERM;
    128 				goto error_free;
    129 			}
    130 			entry = &array[--first_userptr];
    131 #endif
    132 		} else {
    133 			entry = &array[last_entry++];
    134 		}
    135 
    136 		entry->priority = min(info[i].bo_priority,
    137 				      AMDGPU_BO_LIST_MAX_PRIORITY);
    138 		entry->tv.bo = &bo->tbo;
    139 
    140 		if (bo->preferred_domains == AMDGPU_GEM_DOMAIN_GDS)
    141 			list->gds_obj = bo;
    142 		if (bo->preferred_domains == AMDGPU_GEM_DOMAIN_GWS)
    143 			list->gws_obj = bo;
    144 		if (bo->preferred_domains == AMDGPU_GEM_DOMAIN_OA)
    145 			list->oa_obj = bo;
    146 
    147 		total_size += amdgpu_bo_size(bo);
    148 		trace_amdgpu_bo_list_set(list, bo);
    149 	}
    150 
    151 	list->first_userptr = first_userptr;
    152 	list->num_entries = num_entries;
    153 
    154 	trace_amdgpu_cs_bo_status(list->num_entries, total_size);
    155 
    156 	*result = list;
    157 	return 0;
    158 
    159 error_free:
    160 	for (i = 0; i < last_entry; ++i) {
    161 		struct amdgpu_bo *bo = ttm_to_amdgpu_bo(array[i].tv.bo);
    162 
    163 		amdgpu_bo_unref(&bo);
    164 	}
    165 	for (i = first_userptr; i < num_entries; ++i) {
    166 		struct amdgpu_bo *bo = ttm_to_amdgpu_bo(array[i].tv.bo);
    167 
    168 		amdgpu_bo_unref(&bo);
    169 	}
    170 	kvfree(list);
    171 	return r;
    172 
    173 }
    174 
    175 static void amdgpu_bo_list_destroy(struct amdgpu_fpriv *fpriv, int id)
    176 {
    177 	struct amdgpu_bo_list *list;
    178 
    179 	mutex_lock(&fpriv->bo_list_lock);
    180 	list = idr_remove(&fpriv->bo_list_handles, id);
    181 	mutex_unlock(&fpriv->bo_list_lock);
    182 	if (list)
    183 		kref_put(&list->refcount, amdgpu_bo_list_free);
    184 }
    185 
    186 int amdgpu_bo_list_get(struct amdgpu_fpriv *fpriv, int id,
    187 		       struct amdgpu_bo_list **result)
    188 {
    189 	rcu_read_lock();
    190 	*result = idr_find(&fpriv->bo_list_handles, id);
    191 
    192 	if (*result && kref_get_unless_zero(&(*result)->refcount)) {
    193 		rcu_read_unlock();
    194 		return 0;
    195 	}
    196 
    197 	rcu_read_unlock();
    198 	return -ENOENT;
    199 }
    200 
    201 void amdgpu_bo_list_get_list(struct amdgpu_bo_list *list,
    202 			     struct list_head *validated)
    203 {
    204 	/* This is based on the bucket sort with O(n) time complexity.
    205 	 * An item with priority "i" is added to bucket[i]. The lists are then
    206 	 * concatenated in descending order.
    207 	 */
    208 	struct list_head bucket[AMDGPU_BO_LIST_NUM_BUCKETS];
    209 	struct amdgpu_bo_list_entry *e;
    210 	unsigned i;
    211 
    212 	for (i = 0; i < AMDGPU_BO_LIST_NUM_BUCKETS; i++)
    213 		INIT_LIST_HEAD(&bucket[i]);
    214 
    215 	/* Since buffers which appear sooner in the relocation list are
    216 	 * likely to be used more often than buffers which appear later
    217 	 * in the list, the sort mustn't change the ordering of buffers
    218 	 * with the same priority, i.e. it must be stable.
    219 	 */
    220 	amdgpu_bo_list_for_each_entry(e, list) {
    221 		struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
    222 		unsigned priority = e->priority;
    223 
    224 		if (!bo->parent)
    225 			list_add_tail(&e->tv.head, &bucket[priority]);
    226 
    227 		e->user_pages = NULL;
    228 	}
    229 
    230 	/* Connect the sorted buckets in the output list. */
    231 	for (i = 0; i < AMDGPU_BO_LIST_NUM_BUCKETS; i++)
    232 		list_splice(&bucket[i], validated);
    233 }
    234 
    235 void amdgpu_bo_list_put(struct amdgpu_bo_list *list)
    236 {
    237 	kref_put(&list->refcount, amdgpu_bo_list_free);
    238 }
    239 
    240 int amdgpu_bo_create_list_entry_array(struct drm_amdgpu_bo_list_in *in,
    241 				      struct drm_amdgpu_bo_list_entry **info_param)
    242 {
    243 	const void __user *uptr = u64_to_user_ptr(in->bo_info_ptr);
    244 	const uint32_t info_size = sizeof(struct drm_amdgpu_bo_list_entry);
    245 	struct drm_amdgpu_bo_list_entry *info;
    246 	int r;
    247 
    248 	info = kvmalloc_array(in->bo_number, info_size, GFP_KERNEL);
    249 	if (!info)
    250 		return -ENOMEM;
    251 
    252 	/* copy the handle array from userspace to a kernel buffer */
    253 	r = -EFAULT;
    254 	if (likely(info_size == in->bo_info_size)) {
    255 		unsigned long bytes = in->bo_number *
    256 			in->bo_info_size;
    257 
    258 		if (copy_from_user(info, uptr, bytes))
    259 			goto error_free;
    260 
    261 	} else {
    262 		unsigned long bytes = min(in->bo_info_size, info_size);
    263 		unsigned i;
    264 
    265 		memset(info, 0, in->bo_number * info_size);
    266 		for (i = 0; i < in->bo_number; ++i) {
    267 			if (copy_from_user(&info[i], uptr, bytes))
    268 				goto error_free;
    269 
    270 			uptr += in->bo_info_size;
    271 		}
    272 	}
    273 
    274 	*info_param = info;
    275 	return 0;
    276 
    277 error_free:
    278 	kvfree(info);
    279 	return r;
    280 }
    281 
    282 int amdgpu_bo_list_ioctl(struct drm_device *dev, void *data,
    283 				struct drm_file *filp)
    284 {
    285 	struct amdgpu_device *adev = dev->dev_private;
    286 	struct amdgpu_fpriv *fpriv = filp->driver_priv;
    287 	union drm_amdgpu_bo_list *args = data;
    288 	uint32_t handle = args->in.list_handle;
    289 	struct drm_amdgpu_bo_list_entry *info = NULL;
    290 	struct amdgpu_bo_list *list, *old;
    291 	int r;
    292 
    293 	r = amdgpu_bo_create_list_entry_array(&args->in, &info);
    294 	if (r)
    295 		return r;
    296 
    297 	switch (args->in.operation) {
    298 	case AMDGPU_BO_LIST_OP_CREATE:
    299 		r = amdgpu_bo_list_create(adev, filp, info, args->in.bo_number,
    300 					  &list);
    301 		if (r)
    302 			goto error_free;
    303 
    304 		idr_preload(GFP_KERNEL);
    305 		mutex_lock(&fpriv->bo_list_lock);
    306 		r = idr_alloc(&fpriv->bo_list_handles, list, 1, 0, GFP_KERNEL);
    307 		mutex_unlock(&fpriv->bo_list_lock);
    308 		idr_preload_end();
    309 		if (r < 0) {
    310 			goto error_put_list;
    311 		}
    312 
    313 		handle = r;
    314 		break;
    315 
    316 	case AMDGPU_BO_LIST_OP_DESTROY:
    317 		amdgpu_bo_list_destroy(fpriv, handle);
    318 		handle = 0;
    319 		break;
    320 
    321 	case AMDGPU_BO_LIST_OP_UPDATE:
    322 		r = amdgpu_bo_list_create(adev, filp, info, args->in.bo_number,
    323 					  &list);
    324 		if (r)
    325 			goto error_free;
    326 
    327 		mutex_lock(&fpriv->bo_list_lock);
    328 		old = idr_replace(&fpriv->bo_list_handles, list, handle);
    329 		mutex_unlock(&fpriv->bo_list_lock);
    330 
    331 		if (IS_ERR(old)) {
    332 			r = PTR_ERR(old);
    333 			goto error_put_list;
    334 		}
    335 
    336 		amdgpu_bo_list_put(old);
    337 		break;
    338 
    339 	default:
    340 		r = -EINVAL;
    341 		goto error_free;
    342 	}
    343 
    344 	memset(args, 0, sizeof(*args));
    345 	args->out.list_handle = handle;
    346 	kvfree(info);
    347 
    348 	return 0;
    349 
    350 error_put_list:
    351 	amdgpu_bo_list_put(list);
    352 
    353 error_free:
    354 	kvfree(info);
    355 	return r;
    356 }
    357