src/gallium/winsys/svga/drm/vmw_context.c

   1 /**********************************************************
   2  * Copyright 2009 VMware, Inc.  All rights reserved.
   3  *
   4  * Permission is hereby granted, free of charge, to any person
   5  * obtaining a copy of this software and associated documentation
   6  * files (the "Software"), to deal in the Software without
   7  * restriction, including without limitation the rights to use, copy,
   8  * modify, merge, publish, distribute, sublicense, and/or sell copies
   9  * of the Software, and to permit persons to whom the Software is
  10  * furnished to do so, subject to the following conditions:
  11  *
  12  * The above copyright notice and this permission notice shall be
  13  * included in all copies or substantial portions of the Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  16  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  17  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  18  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  19  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  20  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  21  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  22  * SOFTWARE.
  23  *
  24  **********************************************************/
  25
  26
  27 #include "svga_cmd.h"
  28
  29 #include "util/u_debug.h"
  30 #include "util/u_memory.h"
  31 #include "util/u_debug_stack.h"
  32 #include "pipebuffer/pb_buffer.h"
  33 #include "pipebuffer/pb_validate.h"
  34
  35 #include "svga_winsys.h"
  36 #include "vmw_context.h"
  37 #include "vmw_screen.h"
  38 #include "vmw_buffer.h"
  39 #include "vmw_surface.h"
  40 #include "vmw_fence.h"
  41
  42 #define VMW_COMMAND_SIZE (64*1024)
  43 #define VMW_SURFACE_RELOCS (1024)
  44 #define VMW_REGION_RELOCS (512)
  45
  46 #define VMW_MUST_FLUSH_STACK 8
  47
  48 struct vmw_region_relocation
  49 {
  50    struct SVGAGuestPtr *where;
  51    struct pb_buffer *buffer;
  52    /* TODO: put offset info inside where */
  53    uint32 offset;
  54 };
  55
  56 struct vmw_svga_winsys_context
  57 {
  58    struct svga_winsys_context base;
  59
  60    struct vmw_winsys_screen *vws;
  61
  62 #ifdef DEBUG
  63    boolean must_flush;
  64    struct debug_stack_frame must_flush_stack[VMW_MUST_FLUSH_STACK];
  65 #endif
  66
  67    struct {
  68       uint8_t buffer[VMW_COMMAND_SIZE];
  69       uint32_t size;
  70       uint32_t used;
  71       uint32_t reserved;
  72    } command;
  73
  74    struct {
  75       struct vmw_svga_winsys_surface *handles[VMW_SURFACE_RELOCS];
  76       uint32_t size;
  77       uint32_t used;
  78       uint32_t staged;
  79       uint32_t reserved;
  80    } surface;
  81
  82    struct {
  83       struct vmw_region_relocation relocs[VMW_REGION_RELOCS];
  84       uint32_t size;
  85       uint32_t used;
  86       uint32_t staged;
  87       uint32_t reserved;
  88    } region;
  89
  90    struct pb_validate *validate;
  91
  92    uint32_t last_fence;
  93
  94    /**
  95     * The amount of GMR that is referred by the commands currently batched
  96     * in the context.
  97     */
  98    uint32_t seen_regions;
  99
 100    /**
 101     * Whether this context should fail to reserve more commands, not because it
 102     * ran out of command space, but because a substantial ammount of GMR was
 103     * referred.
 104     */
 105    boolean preemptive_flush;
 106 };
 107
 108
 109 static INLINE struct vmw_svga_winsys_context *
 110 vmw_svga_winsys_context(struct svga_winsys_context *swc)
 111 {
 112    assert(swc);
 113    return (struct vmw_svga_winsys_context *)swc;
 114 }
 115
 116
 117 static INLINE unsigned
 118 vmw_translate_to_pb_flags(unsigned flags)
 119 {
 120    unsigned f = 0;
 121    if (flags & SVGA_RELOC_READ)
 122       f |= PB_USAGE_GPU_READ;
 123
 124    if (flags & SVGA_RELOC_WRITE)
 125       f |= PB_USAGE_GPU_WRITE;
 126
 127    return f;
 128 }
 129
 130 static enum pipe_error
 131 vmw_swc_flush(struct svga_winsys_context *swc,
 132               struct pipe_fence_handle **pfence)
 133 {
 134    struct vmw_svga_winsys_context *vswc = vmw_svga_winsys_context(swc);
 135    struct pipe_fence_handle *fence = NULL;
 136    unsigned i;
 137    enum pipe_error ret;
 138
 139    ret = pb_validate_validate(vswc->validate);
 140    assert(ret == PIPE_OK);
 141    if(ret == PIPE_OK) {
 142
 143       /* Apply relocations */
 144       for(i = 0; i < vswc->region.used; ++i) {
 145          struct vmw_region_relocation *reloc = &vswc->region.relocs[i];
 146          struct SVGAGuestPtr ptr;
 147
 148          if(!vmw_gmr_bufmgr_region_ptr(reloc->buffer, &ptr))
 149             assert(0);
 150
 151          ptr.offset += reloc->offset;
 152
 153          *reloc->where = ptr;
 154       }
 155
 156       if (vswc->command.used)
 157          vmw_ioctl_command(vswc->vws,
 158                            vswc->command.buffer,
 159                            vswc->command.used,
 160                            &vswc->last_fence);
 161
 162       fence = vmw_pipe_fence(vswc->last_fence);
 163
 164       pb_validate_fence(vswc->validate, fence);
 165    }
 166
 167    vswc->command.used = 0;
 168    vswc->command.reserved = 0;
 169
 170    for(i = 0; i < vswc->surface.used + vswc->surface.staged; ++i) {
 171       struct vmw_svga_winsys_surface *vsurf =
 172          vswc->surface.handles[i];
 173       p_atomic_dec(&vsurf->validated);
 174       vmw_svga_winsys_surface_reference(&vswc->surface.handles[i], NULL);
 175    }
 176
 177    vswc->surface.used = 0;
 178    vswc->surface.reserved = 0;
 179
 180    for(i = 0; i < vswc->region.used + vswc->region.staged; ++i) {
 181       pb_reference(&vswc->region.relocs[i].buffer, NULL);
 182    }
 183
 184    vswc->region.used = 0;
 185    vswc->region.reserved = 0;
 186
 187 #ifdef DEBUG
 188    vswc->must_flush = FALSE;
 189 #endif
 190    vswc->preemptive_flush = FALSE;
 191    vswc->seen_regions = 0;
 192
 193    if(pfence)
 194       *pfence = fence;
 195
 196    return ret;
 197 }
 198
 199
 200 static void *
 201 vmw_swc_reserve(struct svga_winsys_context *swc,
 202                 uint32_t nr_bytes, uint32_t nr_relocs )
 203 {
 204    struct vmw_svga_winsys_context *vswc = vmw_svga_winsys_context(swc);
 205
 206 #ifdef DEBUG
 207    /* Check if somebody forgot to check the previous failure */
 208    if(vswc->must_flush) {
 209       debug_printf("Forgot to flush:\n");
 210       debug_backtrace_dump(vswc->must_flush_stack, VMW_MUST_FLUSH_STACK);
 211       assert(!vswc->must_flush);
 212    }
 213 #endif
 214
 215    assert(nr_bytes <= vswc->command.size);
 216    if(nr_bytes > vswc->command.size)
 217       return NULL;
 218
 219    if(vswc->preemptive_flush ||
 220       vswc->command.used + nr_bytes > vswc->command.size ||
 221       vswc->surface.used + nr_relocs > vswc->surface.size ||
 222       vswc->region.used + nr_relocs > vswc->region.size) {
 223 #ifdef DEBUG
 224       vswc->must_flush = TRUE;
 225       debug_backtrace_capture(vswc->must_flush_stack, 1,
 226                               VMW_MUST_FLUSH_STACK);
 227 #endif
 228       return NULL;
 229    }
 230
 231    assert(vswc->command.used + nr_bytes <= vswc->command.size);
 232    assert(vswc->surface.used + nr_relocs <= vswc->surface.size);
 233    assert(vswc->region.used + nr_relocs <= vswc->region.size);
 234
 235    vswc->command.reserved = nr_bytes;
 236    vswc->surface.reserved = nr_relocs;
 237    vswc->surface.staged = 0;
 238    vswc->region.reserved = nr_relocs;
 239    vswc->region.staged = 0;
 240
 241    return vswc->command.buffer + vswc->command.used;
 242 }
 243
 244
 245 static void
 246 vmw_swc_surface_relocation(struct svga_winsys_context *swc,
 247                            uint32 *where,
 248                            struct svga_winsys_surface *surface,
 249                            unsigned flags)
 250 {
 251    struct vmw_svga_winsys_context *vswc = vmw_svga_winsys_context(swc);
 252    struct vmw_svga_winsys_surface *vsurf;
 253
 254    if(!surface) {
 255       *where = SVGA3D_INVALID_ID;
 256       return;
 257    }
 258
 259    assert(vswc->surface.staged < vswc->surface.reserved);
 260
 261    vsurf = vmw_svga_winsys_surface(surface);
 262
 263    *where = vsurf->sid;
 264
 265    vmw_svga_winsys_surface_reference(&vswc->surface.handles[vswc->surface.used + vswc->surface.staged], vsurf);
 266    p_atomic_inc(&vsurf->validated);
 267    ++vswc->surface.staged;
 268 }
 269
 270
 271 static void
 272 vmw_swc_region_relocation(struct svga_winsys_context *swc,
 273                           struct SVGAGuestPtr *where,
 274                           struct svga_winsys_buffer *buffer,
 275                           uint32 offset,
 276                           unsigned flags)
 277 {
 278    struct vmw_svga_winsys_context *vswc = vmw_svga_winsys_context(swc);
 279    struct vmw_region_relocation *reloc;
 280    unsigned translated_flags;
 281    enum pipe_error ret;
 282
 283    assert(vswc->region.staged < vswc->region.reserved);
 284
 285    reloc = &vswc->region.relocs[vswc->region.used + vswc->region.staged];
 286    reloc->where = where;
 287    pb_reference(&reloc->buffer, vmw_pb_buffer(buffer));
 288    reloc->offset = offset;
 289
 290    ++vswc->region.staged;
 291
 292    translated_flags = vmw_translate_to_pb_flags(flags);
 293    ret = pb_validate_add_buffer(vswc->validate, reloc->buffer, translated_flags);
 294    /* TODO: Update pipebuffer to reserve buffers and not fail here */
 295    assert(ret == PIPE_OK);
 296
 297    /*
 298     * Flush preemptively the FIFO commands to keep the GMR working set within
 299     * the GMR pool size.
 300     *
 301     * This is necessary for applications like SPECviewperf that generate huge
 302     * amounts of immediate vertex data, so that we don't pile up too much of
 303     * that vertex data neither in the guest nor in the host.
 304     *
 305     * Note that in the current implementation if a region is referred twice in
 306     * a command stream, it will be accounted twice. We could detect repeated
 307     * regions and count only once, but there is no incentive to do that, since
 308     * regions are typically short-lived; always referred in a single command;
 309     * and at the worst we just flush the commands a bit sooner, which for the
 310     * SVGA virtual device it's not a performance issue since flushing commands
 311     * to the FIFO won't cause flushing in the host.
 312     */
 313    vswc->seen_regions += reloc->buffer->base.size;
 314    if(vswc->seen_regions >= VMW_GMR_POOL_SIZE/2)
 315       vswc->preemptive_flush = TRUE;
 316 }
 317
 318
 319 static void
 320 vmw_swc_commit(struct svga_winsys_context *swc)
 321 {
 322    struct vmw_svga_winsys_context *vswc = vmw_svga_winsys_context(swc);
 323
 324    assert(vswc->command.reserved);
 325    assert(vswc->command.used + vswc->command.reserved <= vswc->command.size);
 326    vswc->command.used += vswc->command.reserved;
 327    vswc->command.reserved = 0;
 328
 329    assert(vswc->surface.staged <= vswc->surface.reserved);
 330    assert(vswc->surface.used + vswc->surface.staged <= vswc->surface.size);
 331    vswc->surface.used += vswc->surface.staged;
 332    vswc->surface.staged = 0;
 333    vswc->surface.reserved = 0;
 334
 335    assert(vswc->region.staged <= vswc->region.reserved);
 336    assert(vswc->region.used + vswc->region.staged <= vswc->region.size);
 337    vswc->region.used += vswc->region.staged;
 338    vswc->region.staged = 0;
 339    vswc->region.reserved = 0;
 340 }
 341
 342
 343 static void
 344 vmw_swc_destroy(struct svga_winsys_context *swc)
 345 {
 346    struct vmw_svga_winsys_context *vswc = vmw_svga_winsys_context(swc);
 347    unsigned i;
 348
 349    for(i = 0; i < vswc->region.used; ++i) {
 350       pb_reference(&vswc->region.relocs[i].buffer, NULL);
 351    }
 352
 353    for(i = 0; i < vswc->surface.used; ++i) {
 354       p_atomic_dec(&vswc->surface.handles[i]->validated);
 355       vmw_svga_winsys_surface_reference(&vswc->surface.handles[i], NULL);
 356    }
 357    pb_validate_destroy(vswc->validate);
 358    vmw_ioctl_context_destroy(vswc->vws, swc->cid);
 359    FREE(vswc);
 360 }
 361
 362
 363 struct svga_winsys_context *
 364 vmw_svga_winsys_context_create(struct svga_winsys_screen *sws)
 365 {
 366    struct vmw_winsys_screen *vws = vmw_winsys_screen(sws);
 367    struct vmw_svga_winsys_context *vswc;
 368
 369    vswc = CALLOC_STRUCT(vmw_svga_winsys_context);
 370    if(!vswc)
 371       return NULL;
 372
 373    vswc->base.destroy = vmw_swc_destroy;
 374    vswc->base.reserve = vmw_swc_reserve;
 375    vswc->base.surface_relocation = vmw_swc_surface_relocation;
 376    vswc->base.region_relocation = vmw_swc_region_relocation;
 377    vswc->base.commit = vmw_swc_commit;
 378    vswc->base.flush = vmw_swc_flush;
 379
 380    vswc->base.cid = vmw_ioctl_context_create(vws);
 381
 382    vswc->vws = vws;
 383
 384    vswc->command.size = VMW_COMMAND_SIZE;
 385    vswc->surface.size = VMW_SURFACE_RELOCS;
 386    vswc->region.size = VMW_REGION_RELOCS;
 387
 388    vswc->validate = pb_validate_create();
 389    if(!vswc->validate) {
 390       FREE(vswc);
 391       return NULL;
 392    }
 393
 394    return &vswc->base;
 395 }
 396
 397