src/gallium/drivers/virgl/virgl_resource.c

   1 /*
   2  * Copyright 2014, 2015 Red Hat.
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * on the rights to use, copy, modify, merge, publish, distribute, sub
   8  * license, and/or sell copies of the Software, and to permit persons to whom
   9  * the Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  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 AUTHOR(S) AND/OR THEIR 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 #include "util/u_format.h"
  24 #include "util/u_inlines.h"
  25 #include "util/u_memory.h"
  26 #include "virgl_context.h"
  27 #include "virgl_resource.h"
  28 #include "virgl_screen.h"
  29
  30 /* We need to flush to properly sync the transfer with the current cmdbuf.
  31  * But there are cases where the flushing can be skipped:
  32  *
  33  *  - synchronization is disabled
  34  *  - the resource is not referenced by the current cmdbuf
  35  */
  36 static bool virgl_res_needs_flush(struct virgl_context *vctx,
  37                                   struct virgl_transfer *trans)
  38 {
  39    struct virgl_winsys *vws = virgl_screen(vctx->base.screen)->vws;
  40    struct virgl_resource *res = virgl_resource(trans->base.resource);
  41
  42    if (trans->base.usage & PIPE_TRANSFER_UNSYNCHRONIZED)
  43       return false;
  44
  45    if (!vws->res_is_referenced(vws, vctx->cbuf, res->hw_res))
  46       return false;
  47
  48    return true;
  49 }
  50
  51 /* We need to read back from the host storage to make sure the guest storage
  52  * is up-to-date.  But there are cases where the readback can be skipped:
  53  *
  54  *  - the content can be discarded
  55  *  - the host storage is read-only
  56  *
  57  * Note that PIPE_TRANSFER_WRITE without discard bits requires readback.
  58  * PIPE_TRANSFER_READ becomes irrelevant.  PIPE_TRANSFER_UNSYNCHRONIZED and
  59  * PIPE_TRANSFER_FLUSH_EXPLICIT are also irrelevant.
  60  */
  61 static bool virgl_res_needs_readback(struct virgl_context *vctx,
  62                                      struct virgl_resource *res,
  63                                      unsigned usage, unsigned level)
  64 {
  65    if (usage & (PIPE_TRANSFER_DISCARD_RANGE |
  66                 PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE))
  67       return false;
  68
  69    if (res->clean_mask & (1 << level))
  70       return false;
  71
  72    return true;
  73 }
  74
  75 enum virgl_transfer_map_type
  76 virgl_resource_transfer_prepare(struct virgl_context *vctx,
  77                                 struct virgl_transfer *xfer)
  78 {
  79    struct virgl_winsys *vws = virgl_screen(vctx->base.screen)->vws;
  80    struct virgl_resource *res = virgl_resource(xfer->base.resource);
  81    enum virgl_transfer_map_type map_type = VIRGL_TRANSFER_MAP_HW_RES;
  82    bool flush;
  83    bool readback;
  84    bool wait;
  85
  86    /* there is no way to map the host storage currently */
  87    if (xfer->base.usage & PIPE_TRANSFER_MAP_DIRECTLY)
  88       return VIRGL_TRANSFER_MAP_ERROR;
  89
  90    /* We break the logic down into four steps
  91     *
  92     * step 1: determine the required operations independently
  93     * step 2: look for chances to skip the operations
  94     * step 3: resolve dependencies between the operations
  95     * step 4: execute the operations
  96     */
  97
  98    flush = virgl_res_needs_flush(vctx, xfer);
  99    readback = virgl_res_needs_readback(vctx, res, xfer->base.usage,
 100                                        xfer->base.level);
 101    /* We need to wait for all cmdbufs, current or previous, that access the
 102     * resource to finish unless synchronization is disabled.
 103     */
 104    wait = !(xfer->base.usage & PIPE_TRANSFER_UNSYNCHRONIZED);
 105
 106    /* When the transfer range consists of only uninitialized data, we can
 107     * assume the GPU is not accessing the range and readback is unnecessary.
 108     * We can proceed as if PIPE_TRANSFER_UNSYNCHRONIZED and
 109     * PIPE_TRANSFER_DISCARD_RANGE are set.
 110     */
 111    if (res->u.b.target == PIPE_BUFFER &&
 112          !util_ranges_intersect(&res->valid_buffer_range, xfer->base.box.x,
 113             xfer->base.box.x + xfer->base.box.width)) {
 114       flush = false;
 115       readback = false;
 116       wait = false;
 117    }
 118
 119    /* readback has some implications */
 120    if (readback) {
 121       /* Readback is yet another command and is transparent to the state
 122        * trackers.  It should be waited for in all cases, including when
 123        * PIPE_TRANSFER_UNSYNCHRONIZED is set.
 124        */
 125       wait = true;
 126
 127       /* When the transfer queue has pending writes to this transfer's region,
 128        * we have to flush before readback.
 129        */
 130       if (!flush && virgl_transfer_queue_is_queued(&vctx->queue, xfer))
 131          flush = true;
 132    }
 133
 134    /* XXX This is incorrect and will be removed.  Consider
 135     *
 136     *   glTexImage2D(..., data1);
 137     *   glDrawArrays();
 138     *   glFlush();
 139     *   glTexImage2D(..., data2);
 140     *
 141     * readback and flush are both false in the second glTexImage2D call.  The
 142     * draw call might end up seeing data2.  Same applies to buffers with
 143     * glBufferSubData.
 144     */
 145    wait = flush || readback;
 146
 147    if (flush)
 148       vctx->base.flush(&vctx->base, NULL, 0);
 149
 150    /* If we are not allowed to block, and we know that we will have to wait,
 151     * either because the resource is busy, or because it will become busy due
 152     * to a readback, return early to avoid performing an incomplete
 153     * transfer_get. Such an incomplete transfer_get may finish at any time,
 154     * during which another unsynchronized map could write to the resource
 155     * contents, leaving the contents in an undefined state.
 156     */
 157    if ((xfer->base.usage & PIPE_TRANSFER_DONTBLOCK) &&
 158        (readback || (wait && vws->resource_is_busy(vws, res->hw_res))))
 159       return VIRGL_TRANSFER_MAP_ERROR;
 160
 161    if (readback) {
 162       vws->transfer_get(vws, res->hw_res, &xfer->base.box, xfer->base.stride,
 163                         xfer->l_stride, xfer->offset, xfer->base.level);
 164    }
 165
 166    if (wait)
 167       vws->resource_wait(vws, res->hw_res);
 168
 169    return map_type;
 170 }
 171
 172 static struct pipe_resource *virgl_resource_create(struct pipe_screen *screen,
 173                                                    const struct pipe_resource *templ)
 174 {
 175    unsigned vbind;
 176    struct virgl_screen *vs = virgl_screen(screen);
 177    struct virgl_resource *res = CALLOC_STRUCT(virgl_resource);
 178
 179    res->u.b = *templ;
 180    res->u.b.screen = &vs->base;
 181    pipe_reference_init(&res->u.b.reference, 1);
 182    vbind = pipe_to_virgl_bind(vs, templ->bind);
 183    virgl_resource_layout(&res->u.b, &res->metadata);
 184    res->hw_res = vs->vws->resource_create(vs->vws, templ->target,
 185                                           templ->format, vbind,
 186                                           templ->width0,
 187                                           templ->height0,
 188                                           templ->depth0,
 189                                           templ->array_size,
 190                                           templ->last_level,
 191                                           templ->nr_samples,
 192                                           res->metadata.total_size);
 193    if (!res->hw_res) {
 194       FREE(res);
 195       return NULL;
 196    }
 197
 198    res->clean_mask = (1 << VR_MAX_TEXTURE_2D_LEVELS) - 1;
 199
 200    if (templ->target == PIPE_BUFFER) {
 201       util_range_init(&res->valid_buffer_range);
 202       virgl_buffer_init(res);
 203    } else {
 204       virgl_texture_init(res);
 205    }
 206
 207    return &res->u.b;
 208
 209 }
 210
 211 static struct pipe_resource *virgl_resource_from_handle(struct pipe_screen *screen,
 212                                                         const struct pipe_resource *templ,
 213                                                         struct winsys_handle *whandle,
 214                                                         unsigned usage)
 215 {
 216    struct virgl_screen *vs = virgl_screen(screen);
 217    if (templ->target == PIPE_BUFFER)
 218       return NULL;
 219
 220    struct virgl_resource *res = CALLOC_STRUCT(virgl_resource);
 221    res->u.b = *templ;
 222    res->u.b.screen = &vs->base;
 223    pipe_reference_init(&res->u.b.reference, 1);
 224
 225    res->hw_res = vs->vws->resource_create_from_handle(vs->vws, whandle);
 226    if (!res->hw_res) {
 227       FREE(res);
 228       return NULL;
 229    }
 230
 231    virgl_texture_init(res);
 232
 233    return &res->u.b;
 234 }
 235
 236 void virgl_init_screen_resource_functions(struct pipe_screen *screen)
 237 {
 238     screen->resource_create = virgl_resource_create;
 239     screen->resource_from_handle = virgl_resource_from_handle;
 240     screen->resource_get_handle = u_resource_get_handle_vtbl;
 241     screen->resource_destroy = u_resource_destroy_vtbl;
 242 }
 243
 244 static bool virgl_buffer_transfer_extend(struct pipe_context *ctx,
 245                                          struct pipe_resource *resource,
 246                                          unsigned usage,
 247                                          const struct pipe_box *box,
 248                                          const void *data)
 249 {
 250    struct virgl_context *vctx = virgl_context(ctx);
 251    struct virgl_resource *vbuf = virgl_resource(resource);
 252    struct virgl_transfer dummy_trans = { 0 };
 253    bool flush;
 254    struct virgl_transfer *queued;
 255
 256    /*
 257     * Attempts to short circuit the entire process of mapping and unmapping
 258     * a resource if there is an existing transfer that can be extended.
 259     * Pessimestically falls back if a flush is required.
 260     */
 261    dummy_trans.base.resource = resource;
 262    dummy_trans.base.usage = usage;
 263    dummy_trans.base.box = *box;
 264    dummy_trans.base.stride = vbuf->metadata.stride[0];
 265    dummy_trans.base.layer_stride = vbuf->metadata.layer_stride[0];
 266    dummy_trans.offset = box->x;
 267
 268    flush = virgl_res_needs_flush(vctx, &dummy_trans);
 269    if (flush && util_ranges_intersect(&vbuf->valid_buffer_range,
 270                                       box->x, box->x + box->width))
 271       return false;
 272
 273    queued = virgl_transfer_queue_extend(&vctx->queue, &dummy_trans);
 274    if (!queued || !queued->hw_res_map)
 275       return false;
 276
 277    memcpy(queued->hw_res_map + dummy_trans.offset, data, box->width);
 278    util_range_add(&vbuf->valid_buffer_range, box->x, box->x + box->width);
 279
 280    return true;
 281 }
 282
 283 static void virgl_buffer_subdata(struct pipe_context *pipe,
 284                                  struct pipe_resource *resource,
 285                                  unsigned usage, unsigned offset,
 286                                  unsigned size, const void *data)
 287 {
 288    struct pipe_transfer *transfer;
 289    uint8_t *map;
 290    struct pipe_box box;
 291
 292    assert(!(usage & PIPE_TRANSFER_READ));
 293
 294    /* the write flag is implicit by the nature of buffer_subdata */
 295    usage |= PIPE_TRANSFER_WRITE;
 296
 297    if (offset == 0 && size == resource->width0)
 298       usage |= PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE;
 299    else
 300       usage |= PIPE_TRANSFER_DISCARD_RANGE;
 301
 302    u_box_1d(offset, size, &box);
 303
 304    if (usage & PIPE_TRANSFER_DISCARD_RANGE &&
 305        virgl_buffer_transfer_extend(pipe, resource, usage, &box, data))
 306       return;
 307
 308    map = pipe->transfer_map(pipe, resource, 0, usage, &box, &transfer);
 309    if (map) {
 310       memcpy(map, data, size);
 311       pipe_transfer_unmap(pipe, transfer);
 312    }
 313 }
 314
 315 void virgl_init_context_resource_functions(struct pipe_context *ctx)
 316 {
 317     ctx->transfer_map = u_transfer_map_vtbl;
 318     ctx->transfer_flush_region = u_transfer_flush_region_vtbl;
 319     ctx->transfer_unmap = u_transfer_unmap_vtbl;
 320     ctx->buffer_subdata = virgl_buffer_subdata;
 321     ctx->texture_subdata = u_default_texture_subdata;
 322 }
 323
 324 void virgl_resource_layout(struct pipe_resource *pt,
 325                            struct virgl_resource_metadata *metadata)
 326 {
 327    unsigned level, nblocksy;
 328    unsigned width = pt->width0;
 329    unsigned height = pt->height0;
 330    unsigned depth = pt->depth0;
 331    unsigned buffer_size = 0;
 332
 333    for (level = 0; level <= pt->last_level; level++) {
 334       unsigned slices;
 335
 336       if (pt->target == PIPE_TEXTURE_CUBE)
 337          slices = 6;
 338       else if (pt->target == PIPE_TEXTURE_3D)
 339          slices = depth;
 340       else
 341          slices = pt->array_size;
 342
 343       nblocksy = util_format_get_nblocksy(pt->format, height);
 344       metadata->stride[level] = util_format_get_stride(pt->format, width);
 345       metadata->layer_stride[level] = nblocksy * metadata->stride[level];
 346       metadata->level_offset[level] = buffer_size;
 347
 348       buffer_size += slices * metadata->layer_stride[level];
 349
 350       width = u_minify(width, 1);
 351       height = u_minify(height, 1);
 352       depth = u_minify(depth, 1);
 353    }
 354
 355    if (pt->nr_samples <= 1)
 356       metadata->total_size = buffer_size;
 357    else /* don't create guest backing store for MSAA */
 358       metadata->total_size = 0;
 359 }
 360
 361 struct virgl_transfer *
 362 virgl_resource_create_transfer(struct slab_child_pool *pool,
 363                                struct pipe_resource *pres,
 364                                const struct virgl_resource_metadata *metadata,
 365                                unsigned level, unsigned usage,
 366                                const struct pipe_box *box)
 367 {
 368    struct virgl_transfer *trans;
 369    enum pipe_format format = pres->format;
 370    const unsigned blocksy = box->y / util_format_get_blockheight(format);
 371    const unsigned blocksx = box->x / util_format_get_blockwidth(format);
 372
 373    unsigned offset = metadata->level_offset[level];
 374    if (pres->target == PIPE_TEXTURE_CUBE ||
 375        pres->target == PIPE_TEXTURE_CUBE_ARRAY ||
 376        pres->target == PIPE_TEXTURE_3D ||
 377        pres->target == PIPE_TEXTURE_2D_ARRAY) {
 378       offset += box->z * metadata->layer_stride[level];
 379    }
 380    else if (pres->target == PIPE_TEXTURE_1D_ARRAY) {
 381       offset += box->z * metadata->stride[level];
 382       assert(box->y == 0);
 383    } else if (pres->target == PIPE_BUFFER) {
 384       assert(box->y == 0 && box->z == 0);
 385    } else {
 386       assert(box->z == 0);
 387    }
 388
 389    offset += blocksy * metadata->stride[level];
 390    offset += blocksx * util_format_get_blocksize(format);
 391
 392    trans = slab_alloc(pool);
 393    if (!trans)
 394       return NULL;
 395
 396    trans->base.resource = pres;
 397    trans->base.level = level;
 398    trans->base.usage = usage;
 399    trans->base.box = *box;
 400    trans->base.stride = metadata->stride[level];
 401    trans->base.layer_stride = metadata->layer_stride[level];
 402    trans->offset = offset;
 403    util_range_init(&trans->range);
 404
 405    if (trans->base.resource->target != PIPE_TEXTURE_3D &&
 406        trans->base.resource->target != PIPE_TEXTURE_CUBE &&
 407        trans->base.resource->target != PIPE_TEXTURE_1D_ARRAY &&
 408        trans->base.resource->target != PIPE_TEXTURE_2D_ARRAY &&
 409        trans->base.resource->target != PIPE_TEXTURE_CUBE_ARRAY)
 410       trans->l_stride = 0;
 411    else
 412       trans->l_stride = trans->base.layer_stride;
 413
 414    return trans;
 415 }
 416
 417 void virgl_resource_destroy_transfer(struct slab_child_pool *pool,
 418                                      struct virgl_transfer *trans)
 419 {
 420    util_range_destroy(&trans->range);
 421    slab_free(pool, trans);
 422 }
 423
 424 void virgl_resource_destroy(struct pipe_screen *screen,
 425                             struct pipe_resource *resource)
 426 {
 427    struct virgl_screen *vs = virgl_screen(screen);
 428    struct virgl_resource *res = virgl_resource(resource);
 429
 430    if (res->u.b.target == PIPE_BUFFER)
 431       util_range_destroy(&res->valid_buffer_range);
 432
 433    vs->vws->resource_unref(vs->vws, res->hw_res);
 434    FREE(res);
 435 }
 436
 437 boolean virgl_resource_get_handle(struct pipe_screen *screen,
 438                                   struct pipe_resource *resource,
 439                                   struct winsys_handle *whandle)
 440 {
 441    struct virgl_screen *vs = virgl_screen(screen);
 442    struct virgl_resource *res = virgl_resource(resource);
 443
 444    if (res->u.b.target == PIPE_BUFFER)
 445       return FALSE;
 446
 447    return vs->vws->resource_get_handle(vs->vws, res->hw_res,
 448                                        res->metadata.stride[0],
 449                                        whandle);
 450 }
 451
 452 void virgl_resource_dirty(struct virgl_resource *res, uint32_t level)
 453 {
 454    if (res) {
 455       if (res->u.b.target == PIPE_BUFFER)
 456          res->clean_mask &= ~1;
 457       else
 458          res->clean_mask &= ~(1 << level);
 459    }
 460 }