src/mesa/drivers/dri/intel/intel_batchbuffer.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2006 Tungsten Graphics, Inc., Cedar Park, Texas.
   4  * All Rights Reserved.
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the
   8  * "Software"), to deal in the Software without restriction, including
   9  * without limitation the rights to use, copy, modify, merge, publish,
  10  * distribute, sub license, and/or sell copies of the Software, and to
  11  * permit persons to whom the Software is furnished to do so, subject to
  12  * the following conditions:
  13  *
  14  * The above copyright notice and this permission notice (including the
  15  * next paragraph) shall be included in all copies or substantial portions
  16  * of the Software.
  17  *
  18  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  19  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  20  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
  21  * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
  22  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
  23  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
  24  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  25  *
  26  **************************************************************************/
  27
  28 #include "intel_context.h"
  29 #include "intel_batchbuffer.h"
  30 #include "intel_decode.h"
  31 #include "intel_reg.h"
  32 #include "intel_bufmgr.h"
  33 #include "intel_buffers.h"
  34
  35 /* Relocations in kernel space:
  36  *    - pass dma buffer seperately
  37  *    - memory manager knows how to patch
  38  *    - pass list of dependent buffers
  39  *    - pass relocation list
  40  *
  41  * Either:
  42  *    - get back an offset for buffer to fire
  43  *    - memory manager knows how to fire buffer
  44  *
  45  * Really want the buffer to be AGP and pinned.
  46  *
  47  */
  48
  49 /* Cliprect fence: The highest fence protecting a dma buffer
  50  * containing explicit cliprect information.  Like the old drawable
  51  * lock but irq-driven.  X server must wait for this fence to expire
  52  * before changing cliprects [and then doing sw rendering?].  For
  53  * other dma buffers, the scheduler will grab current cliprect info
  54  * and mix into buffer.  X server must hold the lock while changing
  55  * cliprects???  Make per-drawable.  Need cliprects in shared memory
  56  * -- beats storing them with every cmd buffer in the queue.
  57  *
  58  * ==> X server must wait for this fence to expire before touching the
  59  * framebuffer with new cliprects.
  60  *
  61  * ==> Cliprect-dependent buffers associated with a
  62  * cliprect-timestamp.  All of the buffers associated with a timestamp
  63  * must go to hardware before any buffer with a newer timestamp.
  64  *
  65  * ==> Dma should be queued per-drawable for correct X/GL
  66  * synchronization.  Or can fences be used for this?
  67  *
  68  * Applies to: Blit operations, metaops, X server operations -- X
  69  * server automatically waits on its own dma to complete before
  70  * modifying cliprects ???
  71  */
  72
  73 void
  74 intel_batchbuffer_reset(struct intel_batchbuffer *batch)
  75 {
  76    struct intel_context *intel = batch->intel;
  77
  78    if (batch->buf != NULL) {
  79       dri_bo_unreference(batch->buf);
  80       batch->buf = NULL;
  81    }
  82
  83    if (!batch->buffer)
  84       batch->buffer = malloc (intel->maxBatchSize);
  85
  86    batch->buf = dri_bo_alloc(intel->bufmgr, "batchbuffer",
  87                              intel->maxBatchSize, 4096);
  88    if (batch->buffer)
  89       batch->map = batch->buffer;
  90    else {
  91       dri_bo_map(batch->buf, GL_TRUE);
  92       batch->map = batch->buf->virtual;
  93    }
  94    batch->size = intel->maxBatchSize;
  95    batch->ptr = batch->map;
  96    batch->dirty_state = ~0;
  97 }
  98
  99 struct intel_batchbuffer *
 100 intel_batchbuffer_alloc(struct intel_context *intel)
 101 {
 102    struct intel_batchbuffer *batch = calloc(sizeof(*batch), 1);
 103
 104    batch->intel = intel;
 105    intel_batchbuffer_reset(batch);
 106
 107    return batch;
 108 }
 109
 110 void
 111 intel_batchbuffer_free(struct intel_batchbuffer *batch)
 112 {
 113    if (batch->buffer)
 114       free (batch->buffer);
 115    else {
 116       if (batch->map) {
 117          dri_bo_unmap(batch->buf);
 118          batch->map = NULL;
 119       }
 120    }
 121    dri_bo_unreference(batch->buf);
 122    batch->buf = NULL;
 123    free(batch);
 124 }
 125
 126
 127
 128 /* TODO: Push this whole function into bufmgr.
 129  */
 130 static void
 131 do_flush_locked(struct intel_batchbuffer *batch, GLuint used)
 132 {
 133    struct intel_context *intel = batch->intel;
 134    int ret = 0;
 135    int x_off = 0, y_off = 0;
 136
 137    if (batch->buffer)
 138       dri_bo_subdata (batch->buf, 0, used, batch->buffer);
 139    else
 140       dri_bo_unmap(batch->buf);
 141
 142    batch->map = NULL;
 143    batch->ptr = NULL;
 144
 145    dri_bo_exec(batch->buf, used, NULL, 0, (x_off & 0xffff) | (y_off << 16));
 146
 147    if (INTEL_DEBUG & DEBUG_BATCH) {
 148       dri_bo_map(batch->buf, GL_FALSE);
 149       intel_decode(batch->buf->virtual, used / 4, batch->buf->offset,
 150                    intel->intelScreen->deviceID);
 151       dri_bo_unmap(batch->buf);
 152
 153       if (intel->vtbl.debug_batch != NULL)
 154          intel->vtbl.debug_batch(intel);
 155    }
 156
 157    if (ret != 0) {
 158       exit(1);
 159    }
 160    intel->vtbl.new_batch(intel);
 161 }
 162
 163 void
 164 _intel_batchbuffer_flush(struct intel_batchbuffer *batch, const char *file,
 165                          int line)
 166 {
 167    struct intel_context *intel = batch->intel;
 168    GLuint used = batch->ptr - batch->map;
 169
 170    if (intel->first_post_swapbuffers_batch == NULL) {
 171       intel->first_post_swapbuffers_batch = intel->batch->buf;
 172       drm_intel_bo_reference(intel->first_post_swapbuffers_batch);
 173    }
 174
 175    if (used == 0)
 176       return;
 177
 178    if (INTEL_DEBUG & DEBUG_BATCH)
 179       fprintf(stderr, "%s:%d: Batchbuffer flush with %db used\n", file, line,
 180               used);
 181
 182    batch->reserved_space = 0;
 183    /* Emit a flush if the bufmgr doesn't do it for us. */
 184    if (intel->always_flush_cache) {
 185       intel_batchbuffer_emit_mi_flush(batch);
 186       used = batch->ptr - batch->map;
 187    }
 188
 189    /* Round batchbuffer usage to 2 DWORDs. */
 190
 191    if ((used & 4) == 0) {
 192       *(GLuint *) (batch->ptr) = 0; /* noop */
 193       batch->ptr += 4;
 194       used = batch->ptr - batch->map;
 195    }
 196
 197    /* Mark the end of the buffer. */
 198    *(GLuint *) (batch->ptr) = MI_BATCH_BUFFER_END; /* noop */
 199    batch->ptr += 4;
 200    used = batch->ptr - batch->map;
 201
 202    /* Workaround for recursive batchbuffer flushing: If the window is
 203     * moved, we can get into a case where we try to flush during a
 204     * flush.  What happens is that when we try to grab the lock for
 205     * the first flush, we detect that the window moved which then
 206     * causes another flush (from the intel_draw_buffer() call in
 207     * intelUpdatePageFlipping()).  To work around this we reset the
 208     * batchbuffer tail pointer before trying to get the lock.  This
 209     * prevent the nested buffer flush, but a better fix would be to
 210     * avoid that in the first place. */
 211    batch->ptr = batch->map;
 212
 213    if (intel->vtbl.finish_batch)
 214       intel->vtbl.finish_batch(intel);
 215
 216    /* Check that we didn't just wrap our batchbuffer at a bad time. */
 217    assert(!intel->no_batch_wrap);
 218
 219    batch->reserved_space = BATCH_RESERVED;
 220
 221    /* TODO: Just pass the relocation list and dma buffer up to the
 222     * kernel.
 223     */
 224    do_flush_locked(batch, used);
 225
 226    if (INTEL_DEBUG & DEBUG_SYNC) {
 227       fprintf(stderr, "waiting for idle\n");
 228       dri_bo_map(batch->buf, GL_TRUE);
 229       dri_bo_unmap(batch->buf);
 230    }
 231
 232    /* Reset the buffer:
 233     */
 234    intel_batchbuffer_reset(batch);
 235 }
 236
 237
 238 /*  This is the only way buffers get added to the validate list.
 239  */
 240 GLboolean
 241 intel_batchbuffer_emit_reloc(struct intel_batchbuffer *batch,
 242                              dri_bo *buffer,
 243                              uint32_t read_domains, uint32_t write_domain,
 244                              uint32_t delta)
 245 {
 246    int ret;
 247
 248    if (batch->ptr - batch->map > batch->buf->size)
 249     _mesa_printf ("bad relocation ptr %p map %p offset %d size %d\n",
 250                   batch->ptr, batch->map, batch->ptr - batch->map, batch->buf->size);
 251    ret = dri_bo_emit_reloc(batch->buf, read_domains, write_domain,
 252                            delta, batch->ptr - batch->map, buffer);
 253
 254    /*
 255     * Using the old buffer offset, write in what the right data would be, in case
 256     * the buffer doesn't move and we can short-circuit the relocation processing
 257     * in the kernel
 258     */
 259    intel_batchbuffer_emit_dword (batch, buffer->offset + delta);
 260
 261    return GL_TRUE;
 262 }
 263
 264 void
 265 intel_batchbuffer_data(struct intel_batchbuffer *batch,
 266                        const void *data, GLuint bytes)
 267 {
 268    assert((bytes & 3) == 0);
 269    intel_batchbuffer_require_space(batch, bytes);
 270    __memcpy(batch->ptr, data, bytes);
 271    batch->ptr += bytes;
 272 }
 273
 274 /* Emit a pipelined flush to either flush render and texture cache for
 275  * reading from a FBO-drawn texture, or flush so that frontbuffer
 276  * render appears on the screen in DRI1.
 277  *
 278  * This is also used for the always_flush_cache driconf debug option.
 279  */
 280 void
 281 intel_batchbuffer_emit_mi_flush(struct intel_batchbuffer *batch)
 282 {
 283    struct intel_context *intel = batch->intel;
 284
 285    if (intel->gen >= 4) {
 286       BEGIN_BATCH(4);
 287       OUT_BATCH(_3DSTATE_PIPE_CONTROL |
 288                 PIPE_CONTROL_INSTRUCTION_FLUSH |
 289                 PIPE_CONTROL_WRITE_FLUSH |
 290                 PIPE_CONTROL_NO_WRITE);
 291       OUT_BATCH(0); /* write address */
 292       OUT_BATCH(0); /* write data */
 293       OUT_BATCH(0); /* write data */
 294       ADVANCE_BATCH();
 295    } else {
 296       BEGIN_BATCH(1);
 297       OUT_BATCH(MI_FLUSH);
 298       ADVANCE_BATCH();
 299    }
 300 }