1 /* r300_cmdbuf.c -- Command buffer emission for R300 -*- linux-c -*-
2 *
3 * Copyright (C) The Weather Channel, Inc. 2002.
4 * Copyright (C) 2004 Nicolai Haehnle.
5 * All Rights Reserved.
6 *
7 * The Weather Channel (TM) funded Tungsten Graphics to develop the
8 * initial release of the Radeon 8500 driver under the XFree86 license.
9 * This notice must be preserved.
10 *
11 * Permission is hereby granted, free of charge, to any person obtaining a
12 * copy of this software and associated documentation files (the "Software"),
13 * to deal in the Software without restriction, including without limitation
14 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
15 * and/or sell copies of the Software, and to permit persons to whom the
16 * Software is furnished to do so, subject to the following conditions:
17 *
18 * The above copyright notice and this permission notice (including the next
19 * paragraph) shall be included in all copies or substantial portions of the
20 * Software.
21 *
22 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
23 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
24 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
25 * PRECISION INSIGHT AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
26 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
27 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
28 * DEALINGS IN THE SOFTWARE.
29 *
30 * Authors:
31 * Nicolai Haehnle <prefect_@gmx.net>
32 */
34 #include "drmP.h"
35 #include "drm.h"
36 #include "radeon_drm.h"
37 #include "radeon_drv.h"
38 #include "r300_reg.h"
40 #define R300_SIMULTANEOUS_CLIPRECTS 4
42 /* Values for R300_RE_CLIPRECT_CNTL depending on the number of cliprects
43 */
44 static const int r300_cliprect_cntl[4] = {
45 0xAAAA,
46 0xEEEE,
47 0xFEFE,
48 0xFFFE
49 };
51 /**
52 * Emit up to R300_SIMULTANEOUS_CLIPRECTS cliprects from the given command
53 * buffer, starting with index n.
54 */
55 static int r300_emit_cliprects(drm_radeon_private_t *dev_priv,
56 drm_radeon_kcmd_buffer_t *cmdbuf, int n)
57 {
58 struct drm_clip_rect box;
59 int nr;
60 int i;
61 RING_LOCALS;
63 nr = cmdbuf->nbox - n;
64 if (nr > R300_SIMULTANEOUS_CLIPRECTS)
65 nr = R300_SIMULTANEOUS_CLIPRECTS;
67 DRM_DEBUG("%i cliprects\n", nr);
69 if (nr) {
70 BEGIN_RING(6 + nr * 2);
71 OUT_RING(CP_PACKET0(R300_RE_CLIPRECT_TL_0, nr * 2 - 1));
73 for (i = 0; i < nr; ++i) {
74 if (DRM_COPY_FROM_USER_UNCHECKED
75 (&box, &cmdbuf->boxes[n + i], sizeof(box))) {
76 DRM_ERROR("copy cliprect faulted\n");
77 return -EFAULT;
78 }
80 box.x2--; /* Hardware expects inclusive bottom-right corner */
81 box.y2--;
83 if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_RV515) {
84 box.x1 = (box.x1) &
85 R300_CLIPRECT_MASK;
86 box.y1 = (box.y1) &
87 R300_CLIPRECT_MASK;
88 box.x2 = (box.x2) &
89 R300_CLIPRECT_MASK;
90 box.y2 = (box.y2) &
91 R300_CLIPRECT_MASK;
92 } else {
93 box.x1 = (box.x1 + R300_CLIPRECT_OFFSET) &
94 R300_CLIPRECT_MASK;
95 box.y1 = (box.y1 + R300_CLIPRECT_OFFSET) &
96 R300_CLIPRECT_MASK;
97 box.x2 = (box.x2 + R300_CLIPRECT_OFFSET) &
98 R300_CLIPRECT_MASK;
99 box.y2 = (box.y2 + R300_CLIPRECT_OFFSET) &
100 R300_CLIPRECT_MASK;
101 }
103 OUT_RING((box.x1 << R300_CLIPRECT_X_SHIFT) |
104 (box.y1 << R300_CLIPRECT_Y_SHIFT));
105 OUT_RING((box.x2 << R300_CLIPRECT_X_SHIFT) |
106 (box.y2 << R300_CLIPRECT_Y_SHIFT));
108 }
110 OUT_RING_REG(R300_RE_CLIPRECT_CNTL, r300_cliprect_cntl[nr - 1]);
112 /* TODO/SECURITY: Force scissors to a safe value, otherwise the
113 * client might be able to trample over memory.
114 * The impact should be very limited, but I'd rather be safe than
115 * sorry.
116 */
117 OUT_RING(CP_PACKET0(R300_RE_SCISSORS_TL, 1));
118 OUT_RING(0);
119 OUT_RING(R300_SCISSORS_X_MASK | R300_SCISSORS_Y_MASK);
120 ADVANCE_RING();
121 } else {
122 /* Why we allow zero cliprect rendering:
123 * There are some commands in a command buffer that must be submitted
124 * even when there are no cliprects, e.g. DMA buffer discard
125 * or state setting (though state setting could be avoided by
126 * simulating a loss of context).
127 *
128 * Now since the cmdbuf interface is so chaotic right now (and is
129 * bound to remain that way for a bit until things settle down),
130 * it is basically impossible to filter out the commands that are
131 * necessary and those that aren't.
132 *
133 * So I choose the safe way and don't do any filtering at all;
134 * instead, I simply set up the engine so that all rendering
135 * can't produce any fragments.
136 */
137 BEGIN_RING(2);
138 OUT_RING_REG(R300_RE_CLIPRECT_CNTL, 0);
139 ADVANCE_RING();
140 }
142 /* flus cache and wait idle clean after cliprect change */
143 BEGIN_RING(2);
144 OUT_RING(CP_PACKET0(R300_RB3D_DSTCACHE_CTLSTAT, 0));
145 OUT_RING(R300_RB3D_DC_FLUSH);
146 ADVANCE_RING();
147 BEGIN_RING(2);
148 OUT_RING(CP_PACKET0(RADEON_WAIT_UNTIL, 0));
149 OUT_RING(RADEON_WAIT_3D_IDLECLEAN);
150 ADVANCE_RING();
151 /* set flush flag */
152 dev_priv->track_flush |= RADEON_FLUSH_EMITED;
154 return 0;
155 }
157 static u8 r300_reg_flags[0x10000 >> 2];
159 void r300_init_reg_flags(struct drm_device *dev)
160 {
161 int i;
162 drm_radeon_private_t *dev_priv = dev->dev_private;
164 memset(r300_reg_flags, 0, 0x10000 >> 2);
165 #define ADD_RANGE_MARK(reg, count,mark) \
166 for(i=((reg)>>2);i<((reg)>>2)+(count);i++)\
167 r300_reg_flags[i]|=(mark);
169 #define MARK_SAFE 1
170 #define MARK_CHECK_OFFSET 2
172 #define ADD_RANGE(reg, count) ADD_RANGE_MARK(reg, count, MARK_SAFE)
174 /* these match cmducs() command in r300_driver/r300/r300_cmdbuf.c */
175 ADD_RANGE(R300_SE_VPORT_XSCALE, 6);
176 ADD_RANGE(R300_VAP_CNTL, 1);
177 ADD_RANGE(R300_SE_VTE_CNTL, 2);
178 ADD_RANGE(0x2134, 2);
179 ADD_RANGE(R300_VAP_CNTL_STATUS, 1);
180 ADD_RANGE(R300_VAP_INPUT_CNTL_0, 2);
181 ADD_RANGE(0x21DC, 1);
182 ADD_RANGE(R300_VAP_UNKNOWN_221C, 1);
183 ADD_RANGE(R300_VAP_CLIP_X_0, 4);
184 ADD_RANGE(R300_VAP_PVS_STATE_FLUSH_REG, 1);
185 ADD_RANGE(R300_VAP_UNKNOWN_2288, 1);
186 ADD_RANGE(R300_VAP_OUTPUT_VTX_FMT_0, 2);
187 ADD_RANGE(R300_VAP_PVS_CNTL_1, 3);
188 ADD_RANGE(R300_GB_ENABLE, 1);
189 ADD_RANGE(R300_GB_MSPOS0, 5);
190 ADD_RANGE(R300_TX_INVALTAGS, 1);
191 ADD_RANGE(R300_TX_ENABLE, 1);
192 ADD_RANGE(0x4200, 4);
193 ADD_RANGE(0x4214, 1);
194 ADD_RANGE(R300_RE_POINTSIZE, 1);
195 ADD_RANGE(0x4230, 3);
196 ADD_RANGE(R300_RE_LINE_CNT, 1);
197 ADD_RANGE(R300_RE_UNK4238, 1);
198 ADD_RANGE(0x4260, 3);
199 ADD_RANGE(R300_RE_SHADE, 4);
200 ADD_RANGE(R300_RE_POLYGON_MODE, 5);
201 ADD_RANGE(R300_RE_ZBIAS_CNTL, 1);
202 ADD_RANGE(R300_RE_ZBIAS_T_FACTOR, 4);
203 ADD_RANGE(R300_RE_OCCLUSION_CNTL, 1);
204 ADD_RANGE(R300_RE_CULL_CNTL, 1);
205 ADD_RANGE(0x42C0, 2);
206 ADD_RANGE(R300_RS_CNTL_0, 2);
208 ADD_RANGE(R300_SC_HYPERZ, 2);
209 ADD_RANGE(0x43E8, 1);
211 ADD_RANGE(0x46A4, 5);
213 ADD_RANGE(R300_RE_FOG_STATE, 1);
214 ADD_RANGE(R300_FOG_COLOR_R, 3);
215 ADD_RANGE(R300_PP_ALPHA_TEST, 2);
216 ADD_RANGE(0x4BD8, 1);
217 ADD_RANGE(R300_PFS_PARAM_0_X, 64);
218 ADD_RANGE(0x4E00, 1);
219 ADD_RANGE(R300_RB3D_CBLEND, 2);
220 ADD_RANGE(R300_RB3D_COLORMASK, 1);
221 ADD_RANGE(R300_RB3D_BLEND_COLOR, 3);
222 ADD_RANGE_MARK(R300_RB3D_COLOROFFSET0, 1, MARK_CHECK_OFFSET); /* check offset */
223 ADD_RANGE(R300_RB3D_COLORPITCH0, 1);
224 ADD_RANGE(0x4E50, 9);
225 ADD_RANGE(0x4E88, 1);
226 ADD_RANGE(0x4EA0, 2);
227 ADD_RANGE(R300_ZB_CNTL, 3);
228 ADD_RANGE(R300_ZB_FORMAT, 4);
229 ADD_RANGE_MARK(R300_ZB_DEPTHOFFSET, 1, MARK_CHECK_OFFSET); /* check offset */
230 ADD_RANGE(R300_ZB_DEPTHPITCH, 1);
231 ADD_RANGE(R300_ZB_DEPTHCLEARVALUE, 1);
232 ADD_RANGE(R300_ZB_ZMASK_OFFSET, 13);
234 ADD_RANGE(R300_TX_FILTER_0, 16);
235 ADD_RANGE(R300_TX_FILTER1_0, 16);
236 ADD_RANGE(R300_TX_SIZE_0, 16);
237 ADD_RANGE(R300_TX_FORMAT_0, 16);
238 ADD_RANGE(R300_TX_PITCH_0, 16);
239 /* Texture offset is dangerous and needs more checking */
240 ADD_RANGE_MARK(R300_TX_OFFSET_0, 16, MARK_CHECK_OFFSET);
241 ADD_RANGE(R300_TX_CHROMA_KEY_0, 16);
242 ADD_RANGE(R300_TX_BORDER_COLOR_0, 16);
244 /* Sporadic registers used as primitives are emitted */
245 ADD_RANGE(R300_ZB_ZCACHE_CTLSTAT, 1);
246 ADD_RANGE(R300_RB3D_DSTCACHE_CTLSTAT, 1);
247 ADD_RANGE(R300_VAP_INPUT_ROUTE_0_0, 8);
248 ADD_RANGE(R300_VAP_INPUT_ROUTE_1_0, 8);
250 if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_RV515) {
251 ADD_RANGE(R500_VAP_INDEX_OFFSET, 1);
252 ADD_RANGE(R500_US_CONFIG, 2);
253 ADD_RANGE(R500_US_CODE_ADDR, 3);
254 ADD_RANGE(R500_US_FC_CTRL, 1);
255 ADD_RANGE(R500_RS_IP_0, 16);
256 ADD_RANGE(R500_RS_INST_0, 16);
257 ADD_RANGE(R500_RB3D_COLOR_CLEAR_VALUE_AR, 2);
258 ADD_RANGE(R500_RB3D_CONSTANT_COLOR_AR, 2);
259 ADD_RANGE(R500_ZB_FIFO_SIZE, 2);
260 } else {
261 ADD_RANGE(R300_PFS_CNTL_0, 3);
262 ADD_RANGE(R300_PFS_NODE_0, 4);
263 ADD_RANGE(R300_PFS_TEXI_0, 64);
264 ADD_RANGE(R300_PFS_INSTR0_0, 64);
265 ADD_RANGE(R300_PFS_INSTR1_0, 64);
266 ADD_RANGE(R300_PFS_INSTR2_0, 64);
267 ADD_RANGE(R300_PFS_INSTR3_0, 64);
268 ADD_RANGE(R300_RS_INTERP_0, 8);
269 ADD_RANGE(R300_RS_ROUTE_0, 8);
271 }
272 }
274 static __inline__ int r300_check_range(unsigned reg, int count)
275 {
276 int i;
277 if (reg & ~0xffff)
278 return -1;
279 for (i = (reg >> 2); i < (reg >> 2) + count; i++)
280 if (r300_reg_flags[i] != MARK_SAFE)
281 return 1;
282 return 0;
283 }
285 static __inline__ int r300_emit_carefully_checked_packet0(drm_radeon_private_t *
286 dev_priv,
287 drm_radeon_kcmd_buffer_t
288 * cmdbuf,
289 drm_r300_cmd_header_t
290 header)
291 {
292 int reg;
293 int sz;
294 int i;
295 int values[64];
296 RING_LOCALS;
298 sz = header.packet0.count;
299 reg = (header.packet0.reghi << 8) | header.packet0.reglo;
301 if ((sz > 64) || (sz < 0)) {
302 DRM_ERROR
303 ("Cannot emit more than 64 values at a time (reg=%04x sz=%d)\n",
304 reg, sz);
305 return -EINVAL;
306 }
307 for (i = 0; i < sz; i++) {
308 values[i] = ((int *)cmdbuf->buf)[i];
309 switch (r300_reg_flags[(reg >> 2) + i]) {
310 case MARK_SAFE:
311 break;
312 case MARK_CHECK_OFFSET:
313 if (!radeon_check_offset(dev_priv, (u32) values[i])) {
314 DRM_ERROR
315 ("Offset failed range check (reg=%04x sz=%d)\n",
316 reg, sz);
317 return -EINVAL;
318 }
319 break;
320 default:
321 DRM_ERROR("Register %04x failed check as flag=%02x\n",
322 reg + i * 4, r300_reg_flags[(reg >> 2) + i]);
323 return -EINVAL;
324 }
325 }
327 BEGIN_RING(1 + sz);
328 OUT_RING(CP_PACKET0(reg, sz - 1));
329 OUT_RING_TABLE(values, sz);
330 ADVANCE_RING();
332 cmdbuf->buf += sz * 4;
333 cmdbuf->bufsz -= sz * 4;
335 return 0;
336 }
338 /**
339 * Emits a packet0 setting arbitrary registers.
340 * Called by r300_do_cp_cmdbuf.
341 *
342 * Note that checks are performed on contents and addresses of the registers
343 */
344 static __inline__ int r300_emit_packet0(drm_radeon_private_t *dev_priv,
345 drm_radeon_kcmd_buffer_t *cmdbuf,
346 drm_r300_cmd_header_t header)
347 {
348 int reg;
349 int sz;
350 RING_LOCALS;
352 sz = header.packet0.count;
353 reg = (header.packet0.reghi << 8) | header.packet0.reglo;
355 DRM_DEBUG("R300_CMD_PACKET0: reg %04x, sz %d\n", reg, sz);
356 if (!sz)
357 return 0;
359 if (sz * 4 > cmdbuf->bufsz)
360 return -EINVAL;
362 if (reg + sz * 4 >= 0x10000) {
363 DRM_ERROR("No such registers in hardware reg=%04x sz=%d\n", reg,
364 sz);
365 return -EINVAL;
366 }
368 if (r300_check_range(reg, sz)) {
369 /* go and check everything */
370 return r300_emit_carefully_checked_packet0(dev_priv, cmdbuf,
371 header);
372 }
373 /* the rest of the data is safe to emit, whatever the values the user passed */
375 BEGIN_RING(1 + sz);
376 OUT_RING(CP_PACKET0(reg, sz - 1));
377 OUT_RING_TABLE((int *)cmdbuf->buf, sz);
378 ADVANCE_RING();
380 cmdbuf->buf += sz * 4;
381 cmdbuf->bufsz -= sz * 4;
383 return 0;
384 }
386 /**
387 * Uploads user-supplied vertex program instructions or parameters onto
388 * the graphics card.
389 * Called by r300_do_cp_cmdbuf.
390 */
391 static __inline__ int r300_emit_vpu(drm_radeon_private_t *dev_priv,
392 drm_radeon_kcmd_buffer_t *cmdbuf,
393 drm_r300_cmd_header_t header)
394 {
395 int sz;
396 int addr;
397 RING_LOCALS;
399 sz = header.vpu.count;
400 addr = (header.vpu.adrhi << 8) | header.vpu.adrlo;
402 if (!sz)
403 return 0;
404 if (sz * 16 > cmdbuf->bufsz)
405 return -EINVAL;
407 /* VAP is very sensitive so we purge cache before we program it
408 * and we also flush its state before & after */
409 BEGIN_RING(6);
410 OUT_RING(CP_PACKET0(R300_RB3D_DSTCACHE_CTLSTAT, 0));
411 OUT_RING(R300_RB3D_DC_FLUSH);
412 OUT_RING(CP_PACKET0(RADEON_WAIT_UNTIL, 0));
413 OUT_RING(RADEON_WAIT_3D_IDLECLEAN);
414 OUT_RING(CP_PACKET0(R300_VAP_PVS_STATE_FLUSH_REG, 0));
415 OUT_RING(0);
416 ADVANCE_RING();
417 /* set flush flag */
418 dev_priv->track_flush |= RADEON_FLUSH_EMITED;
420 BEGIN_RING(3 + sz * 4);
421 OUT_RING_REG(R300_VAP_PVS_UPLOAD_ADDRESS, addr);
422 OUT_RING(CP_PACKET0_TABLE(R300_VAP_PVS_UPLOAD_DATA, sz * 4 - 1));
423 OUT_RING_TABLE((int *)cmdbuf->buf, sz * 4);
424 ADVANCE_RING();
426 BEGIN_RING(2);
427 OUT_RING(CP_PACKET0(R300_VAP_PVS_STATE_FLUSH_REG, 0));
428 OUT_RING(0);
429 ADVANCE_RING();
431 cmdbuf->buf += sz * 16;
432 cmdbuf->bufsz -= sz * 16;
434 return 0;
435 }
437 /**
438 * Emit a clear packet from userspace.
439 * Called by r300_emit_packet3.
440 */
441 static __inline__ int r300_emit_clear(drm_radeon_private_t *dev_priv,
442 drm_radeon_kcmd_buffer_t *cmdbuf)
443 {
444 RING_LOCALS;
446 if (8 * 4 > cmdbuf->bufsz)
447 return -EINVAL;
449 BEGIN_RING(10);
450 OUT_RING(CP_PACKET3(R200_3D_DRAW_IMMD_2, 8));
451 OUT_RING(R300_PRIM_TYPE_POINT | R300_PRIM_WALK_RING |
452 (1 << R300_PRIM_NUM_VERTICES_SHIFT));
453 OUT_RING_TABLE((int *)cmdbuf->buf, 8);
454 ADVANCE_RING();
456 BEGIN_RING(4);
457 OUT_RING(CP_PACKET0(R300_RB3D_DSTCACHE_CTLSTAT, 0));
458 OUT_RING(R300_RB3D_DC_FLUSH);
459 OUT_RING(CP_PACKET0(RADEON_WAIT_UNTIL, 0));
460 OUT_RING(RADEON_WAIT_3D_IDLECLEAN);
461 ADVANCE_RING();
462 /* set flush flag */
463 dev_priv->track_flush |= RADEON_FLUSH_EMITED;
465 cmdbuf->buf += 8 * 4;
466 cmdbuf->bufsz -= 8 * 4;
468 return 0;
469 }
471 static __inline__ int r300_emit_3d_load_vbpntr(drm_radeon_private_t *dev_priv,
472 drm_radeon_kcmd_buffer_t *cmdbuf,
473 u32 header)
474 {
475 int count, i, k;
476 #define MAX_ARRAY_PACKET 64
477 u32 payload[MAX_ARRAY_PACKET];
478 u32 narrays;
479 RING_LOCALS;
481 count = (header >> 16) & 0x3fff;
483 if ((count + 1) > MAX_ARRAY_PACKET) {
484 DRM_ERROR("Too large payload in 3D_LOAD_VBPNTR (count=%d)\n",
485 count);
486 return -EINVAL;
487 }
488 memset(payload, 0, MAX_ARRAY_PACKET * 4);
489 memcpy(payload, cmdbuf->buf + 4, (count + 1) * 4);
491 /* carefully check packet contents */
493 narrays = payload[0];
494 k = 0;
495 i = 1;
496 while ((k < narrays) && (i < (count + 1))) {
497 i++; /* skip attribute field */
498 if (!radeon_check_offset(dev_priv, payload[i])) {
499 DRM_ERROR
500 ("Offset failed range check (k=%d i=%d) while processing 3D_LOAD_VBPNTR packet.\n",
501 k, i);
502 return -EINVAL;
503 }
504 k++;
505 i++;
506 if (k == narrays)
507 break;
508 /* have one more to process, they come in pairs */
509 if (!radeon_check_offset(dev_priv, payload[i])) {
510 DRM_ERROR
511 ("Offset failed range check (k=%d i=%d) while processing 3D_LOAD_VBPNTR packet.\n",
512 k, i);
513 return -EINVAL;
514 }
515 k++;
516 i++;
517 }
518 /* do the counts match what we expect ? */
519 if ((k != narrays) || (i != (count + 1))) {
520 DRM_ERROR
521 ("Malformed 3D_LOAD_VBPNTR packet (k=%d i=%d narrays=%d count+1=%d).\n",
522 k, i, narrays, count + 1);
523 return -EINVAL;
524 }
526 /* all clear, output packet */
528 BEGIN_RING(count + 2);
529 OUT_RING(header);
530 OUT_RING_TABLE(payload, count + 1);
531 ADVANCE_RING();
533 cmdbuf->buf += (count + 2) * 4;
534 cmdbuf->bufsz -= (count + 2) * 4;
536 return 0;
537 }
539 static __inline__ int r300_emit_bitblt_multi(drm_radeon_private_t *dev_priv,
540 drm_radeon_kcmd_buffer_t *cmdbuf)
541 {
542 u32 *cmd = (u32 *) cmdbuf->buf;
543 int count, ret;
544 RING_LOCALS;
546 count=(cmd[0]>>16) & 0x3fff;
548 if (cmd[0] & 0x8000) {
549 u32 offset;
551 if (cmd[1] & (RADEON_GMC_SRC_PITCH_OFFSET_CNTL
552 | RADEON_GMC_DST_PITCH_OFFSET_CNTL)) {
553 offset = cmd[2] << 10;
554 ret = !radeon_check_offset(dev_priv, offset);
555 if (ret) {
556 DRM_ERROR("Invalid bitblt first offset is %08X\n", offset);
557 return -EINVAL;
558 }
559 }
561 if ((cmd[1] & RADEON_GMC_SRC_PITCH_OFFSET_CNTL) &&
562 (cmd[1] & RADEON_GMC_DST_PITCH_OFFSET_CNTL)) {
563 offset = cmd[3] << 10;
564 ret = !radeon_check_offset(dev_priv, offset);
565 if (ret) {
566 DRM_ERROR("Invalid bitblt second offset is %08X\n", offset);
567 return -EINVAL;
568 }
570 }
571 }
573 BEGIN_RING(count+2);
574 OUT_RING(cmd[0]);
575 OUT_RING_TABLE((int *)(cmdbuf->buf + 4), count + 1);
576 ADVANCE_RING();
578 cmdbuf->buf += (count+2)*4;
579 cmdbuf->bufsz -= (count+2)*4;
581 return 0;
582 }
584 static __inline__ int r300_emit_draw_indx_2(drm_radeon_private_t *dev_priv,
585 drm_radeon_kcmd_buffer_t *cmdbuf)
586 {
587 u32 *cmd;
588 int count;
589 int expected_count;
590 RING_LOCALS;
592 cmd = (u32 *) cmdbuf->buf;
593 count = (cmd[0]>>16) & 0x3fff;
594 expected_count = cmd[1] >> 16;
595 if (!(cmd[1] & R300_VAP_VF_CNTL__INDEX_SIZE_32bit))
596 expected_count = (expected_count+1)/2;
598 if (count && count != expected_count) {
599 DRM_ERROR("3D_DRAW_INDX_2: packet size %i, expected %i\n",
600 count, expected_count);
601 return -EINVAL;
602 }
604 BEGIN_RING(count+2);
605 OUT_RING(cmd[0]);
606 OUT_RING_TABLE((int *)(cmdbuf->buf + 4), count + 1);
607 ADVANCE_RING();
609 cmdbuf->buf += (count+2)*4;
610 cmdbuf->bufsz -= (count+2)*4;
612 if (!count) {
613 drm_r300_cmd_header_t header;
615 if (cmdbuf->bufsz < 4*4 + sizeof(header)) {
616 DRM_ERROR("3D_DRAW_INDX_2: expect subsequent INDX_BUFFER, but stream is too short.\n");
617 return -EINVAL;
618 }
620 header.u = *(unsigned int *)cmdbuf->buf;
622 cmdbuf->buf += sizeof(header);
623 cmdbuf->bufsz -= sizeof(header);
624 cmd = (u32 *) cmdbuf->buf;
626 if (header.header.cmd_type != R300_CMD_PACKET3 ||
627 header.packet3.packet != R300_CMD_PACKET3_RAW ||
628 cmd[0] != CP_PACKET3(RADEON_CP_INDX_BUFFER, 2)) {
629 DRM_ERROR("3D_DRAW_INDX_2: expect subsequent INDX_BUFFER.\n");
630 return -EINVAL;
631 }
633 if ((cmd[1] & 0x8000ffff) != 0x80000810) {
634 DRM_ERROR("Invalid indx_buffer reg address %08X\n", cmd[1]);
635 return -EINVAL;
636 }
637 if (!radeon_check_offset(dev_priv, cmd[2])) {
638 DRM_ERROR("Invalid indx_buffer offset is %08X\n", cmd[2]);
639 return -EINVAL;
640 }
641 if (cmd[3] != expected_count) {
642 DRM_ERROR("INDX_BUFFER: buffer size %i, expected %i\n",
643 cmd[3], expected_count);
644 return -EINVAL;
645 }
647 BEGIN_RING(4);
648 OUT_RING(cmd[0]);
649 OUT_RING_TABLE((int *)(cmdbuf->buf + 4), 3);
650 ADVANCE_RING();
652 cmdbuf->buf += 4*4;
653 cmdbuf->bufsz -= 4*4;
654 }
656 return 0;
657 }
659 static __inline__ int r300_emit_raw_packet3(drm_radeon_private_t *dev_priv,
660 drm_radeon_kcmd_buffer_t *cmdbuf)
661 {
662 u32 header;
663 int count;
664 RING_LOCALS;
666 if (4 > cmdbuf->bufsz)
667 return -EINVAL;
669 /* Fixme !! This simply emits a packet without much checking.
670 We need to be smarter. */
672 /* obtain first word - actual packet3 header */
673 header = *(u32 *) cmdbuf->buf;
675 /* Is it packet 3 ? */
676 if ((header >> 30) != 0x3) {
677 DRM_ERROR("Not a packet3 header (0x%08x)\n", header);
678 return -EINVAL;
679 }
681 count = (header >> 16) & 0x3fff;
683 /* Check again now that we know how much data to expect */
684 if ((count + 2) * 4 > cmdbuf->bufsz) {
685 DRM_ERROR
686 ("Expected packet3 of length %d but have only %d bytes left\n",
687 (count + 2) * 4, cmdbuf->bufsz);
688 return -EINVAL;
689 }
691 /* Is it a packet type we know about ? */
692 switch (header & 0xff00) {
693 case RADEON_3D_LOAD_VBPNTR: /* load vertex array pointers */
694 return r300_emit_3d_load_vbpntr(dev_priv, cmdbuf, header);
696 case RADEON_CNTL_BITBLT_MULTI:
697 return r300_emit_bitblt_multi(dev_priv, cmdbuf);
699 case RADEON_CP_INDX_BUFFER:
700 DRM_ERROR("packet3 INDX_BUFFER without preceding 3D_DRAW_INDX_2 is illegal.\n");
701 return -EINVAL;
702 case RADEON_CP_3D_DRAW_IMMD_2:
703 /* triggers drawing using in-packet vertex data */
704 case RADEON_CP_3D_DRAW_VBUF_2:
705 /* triggers drawing of vertex buffers setup elsewhere */
706 dev_priv->track_flush &= ~(RADEON_FLUSH_EMITED |
707 RADEON_PURGE_EMITED);
708 break;
709 case RADEON_CP_3D_DRAW_INDX_2:
710 /* triggers drawing using indices to vertex buffer */
711 /* whenever we send vertex we clear flush & purge */
712 dev_priv->track_flush &= ~(RADEON_FLUSH_EMITED |
713 RADEON_PURGE_EMITED);
714 return r300_emit_draw_indx_2(dev_priv, cmdbuf);
715 case RADEON_WAIT_FOR_IDLE:
716 case RADEON_CP_NOP:
717 /* these packets are safe */
718 break;
719 default:
720 DRM_ERROR("Unknown packet3 header (0x%08x)\n", header);
721 return -EINVAL;
722 }
724 BEGIN_RING(count + 2);
725 OUT_RING(header);
726 OUT_RING_TABLE((int *)(cmdbuf->buf + 4), count + 1);
727 ADVANCE_RING();
729 cmdbuf->buf += (count + 2) * 4;
730 cmdbuf->bufsz -= (count + 2) * 4;
732 return 0;
733 }
735 /**
736 * Emit a rendering packet3 from userspace.
737 * Called by r300_do_cp_cmdbuf.
738 */
739 static __inline__ int r300_emit_packet3(drm_radeon_private_t *dev_priv,
740 drm_radeon_kcmd_buffer_t *cmdbuf,
741 drm_r300_cmd_header_t header)
742 {
743 int n;
744 int ret;
745 char *orig_buf = cmdbuf->buf;
746 int orig_bufsz = cmdbuf->bufsz;
748 /* This is a do-while-loop so that we run the interior at least once,
749 * even if cmdbuf->nbox is 0. Compare r300_emit_cliprects for rationale.
750 */
751 n = 0;
752 do {
753 if (cmdbuf->nbox > R300_SIMULTANEOUS_CLIPRECTS) {
754 ret = r300_emit_cliprects(dev_priv, cmdbuf, n);
755 if (ret)
756 return ret;
758 cmdbuf->buf = orig_buf;
759 cmdbuf->bufsz = orig_bufsz;
760 }
762 switch (header.packet3.packet) {
763 case R300_CMD_PACKET3_CLEAR:
764 DRM_DEBUG("R300_CMD_PACKET3_CLEAR\n");
765 ret = r300_emit_clear(dev_priv, cmdbuf);
766 if (ret) {
767 DRM_ERROR("r300_emit_clear failed\n");
768 return ret;
769 }
770 break;
772 case R300_CMD_PACKET3_RAW:
773 DRM_DEBUG("R300_CMD_PACKET3_RAW\n");
774 ret = r300_emit_raw_packet3(dev_priv, cmdbuf);
775 if (ret) {
776 DRM_ERROR("r300_emit_raw_packet3 failed\n");
777 return ret;
778 }
779 break;
781 default:
782 DRM_ERROR("bad packet3 type %i at %p\n",
783 header.packet3.packet,
784 cmdbuf->buf - sizeof(header));
785 return -EINVAL;
786 }
788 n += R300_SIMULTANEOUS_CLIPRECTS;
789 } while (n < cmdbuf->nbox);
791 return 0;
792 }
794 /* Some of the R300 chips seem to be extremely touchy about the two registers
795 * that are configured in r300_pacify.
796 * Among the worst offenders seems to be the R300 ND (0x4E44): When userspace
797 * sends a command buffer that contains only state setting commands and a
798 * vertex program/parameter upload sequence, this will eventually lead to a
799 * lockup, unless the sequence is bracketed by calls to r300_pacify.
800 * So we should take great care to *always* call r300_pacify before
801 * *anything* 3D related, and again afterwards. This is what the
802 * call bracket in r300_do_cp_cmdbuf is for.
803 */
805 /**
806 * Emit the sequence to pacify R300.
807 */
808 static __inline__ void r300_pacify(drm_radeon_private_t *dev_priv)
809 {
810 uint32_t cache_z, cache_3d, cache_2d;
811 RING_LOCALS;
813 cache_z = R300_ZC_FLUSH;
814 cache_2d = R300_RB2D_DC_FLUSH;
815 cache_3d = R300_RB3D_DC_FLUSH;
816 if (!(dev_priv->track_flush & RADEON_PURGE_EMITED)) {
817 /* we can purge, primitive where draw since last purge */
818 cache_z |= R300_ZC_FREE;
819 cache_2d |= R300_RB2D_DC_FREE;
820 cache_3d |= R300_RB3D_DC_FREE;
821 }
823 /* flush & purge zbuffer */
824 BEGIN_RING(2);
825 OUT_RING(CP_PACKET0(R300_ZB_ZCACHE_CTLSTAT, 0));
826 OUT_RING(cache_z);
827 ADVANCE_RING();
828 /* flush & purge 3d */
829 BEGIN_RING(2);
830 OUT_RING(CP_PACKET0(R300_RB3D_DSTCACHE_CTLSTAT, 0));
831 OUT_RING(cache_3d);
832 ADVANCE_RING();
833 /* flush & purge texture */
834 BEGIN_RING(2);
835 OUT_RING(CP_PACKET0(R300_TX_INVALTAGS, 0));
836 OUT_RING(0);
837 ADVANCE_RING();
838 /* FIXME: is this one really needed ? */
839 BEGIN_RING(2);
840 OUT_RING(CP_PACKET0(R300_RB3D_AARESOLVE_CTL, 0));
841 OUT_RING(0);
842 ADVANCE_RING();
843 BEGIN_RING(2);
844 OUT_RING(CP_PACKET0(RADEON_WAIT_UNTIL, 0));
845 OUT_RING(RADEON_WAIT_3D_IDLECLEAN);
846 ADVANCE_RING();
847 /* flush & purge 2d through E2 as RB2D will trigger lockup */
848 BEGIN_RING(4);
849 OUT_RING(CP_PACKET0(R300_DSTCACHE_CTLSTAT, 0));
850 OUT_RING(cache_2d);
851 OUT_RING(CP_PACKET0(RADEON_WAIT_UNTIL, 0));
852 OUT_RING(RADEON_WAIT_2D_IDLECLEAN |
853 RADEON_WAIT_HOST_IDLECLEAN);
854 ADVANCE_RING();
855 /* set flush & purge flags */
856 dev_priv->track_flush |= RADEON_FLUSH_EMITED | RADEON_PURGE_EMITED;
857 }
859 /**
860 * Called by r300_do_cp_cmdbuf to update the internal buffer age and state.
861 * The actual age emit is done by r300_do_cp_cmdbuf, which is why you must
862 * be careful about how this function is called.
863 */
864 static void r300_discard_buffer(struct drm_device * dev, struct drm_buf * buf)
865 {
866 drm_radeon_private_t *dev_priv = dev->dev_private;
867 drm_radeon_buf_priv_t *buf_priv = buf->dev_private;
869 buf_priv->age = ++dev_priv->sarea_priv->last_dispatch;
870 buf->pending = 1;
871 buf->used = 0;
872 }
874 static void r300_cmd_wait(drm_radeon_private_t * dev_priv,
875 drm_r300_cmd_header_t header)
876 {
877 u32 wait_until;
878 RING_LOCALS;
880 if (!header.wait.flags)
881 return;
883 wait_until = 0;
885 switch(header.wait.flags) {
886 case R300_WAIT_2D:
887 wait_until = RADEON_WAIT_2D_IDLE;
888 break;
889 case R300_WAIT_3D:
890 wait_until = RADEON_WAIT_3D_IDLE;
891 break;
892 case R300_NEW_WAIT_2D_3D:
893 wait_until = RADEON_WAIT_2D_IDLE|RADEON_WAIT_3D_IDLE;
894 break;
895 case R300_NEW_WAIT_2D_2D_CLEAN:
896 wait_until = RADEON_WAIT_2D_IDLE|RADEON_WAIT_2D_IDLECLEAN;
897 break;
898 case R300_NEW_WAIT_3D_3D_CLEAN:
899 wait_until = RADEON_WAIT_3D_IDLE|RADEON_WAIT_3D_IDLECLEAN;
900 break;
901 case R300_NEW_WAIT_2D_2D_CLEAN_3D_3D_CLEAN:
902 wait_until = RADEON_WAIT_2D_IDLE|RADEON_WAIT_2D_IDLECLEAN;
903 wait_until |= RADEON_WAIT_3D_IDLE|RADEON_WAIT_3D_IDLECLEAN;
904 break;
905 default:
906 return;
907 }
909 BEGIN_RING(2);
910 OUT_RING(CP_PACKET0(RADEON_WAIT_UNTIL, 0));
911 OUT_RING(wait_until);
912 ADVANCE_RING();
913 }
915 static int r300_scratch(drm_radeon_private_t *dev_priv,
916 drm_radeon_kcmd_buffer_t *cmdbuf,
917 drm_r300_cmd_header_t header)
918 {
919 u32 *ref_age_base;
920 u32 i, buf_idx, h_pending;
921 RING_LOCALS;
923 if (cmdbuf->bufsz < sizeof(uint64_t) + header.scratch.n_bufs * sizeof(buf_idx) ) {
924 return -EINVAL;
925 }
927 if (header.scratch.reg >= 5) {
928 return -EINVAL;
929 }
931 dev_priv->scratch_ages[header.scratch.reg] ++;
933 ref_age_base = (u32 *)(unsigned long)*((uint64_t *)cmdbuf->buf);
935 cmdbuf->buf += sizeof(uint64_t);
936 cmdbuf->bufsz -= sizeof(uint64_t);
938 for (i=0; i < header.scratch.n_bufs; i++) {
939 buf_idx = *(u32 *)cmdbuf->buf;
940 buf_idx *= 2; /* 8 bytes per buf */
942 if (DRM_COPY_TO_USER(ref_age_base + buf_idx, &dev_priv->scratch_ages[header.scratch.reg], sizeof(u32))) {
943 return -EINVAL;
944 }
946 if (DRM_COPY_FROM_USER(&h_pending, ref_age_base + buf_idx + 1, sizeof(u32))) {
947 return -EINVAL;
948 }
950 if (h_pending == 0) {
951 return -EINVAL;
952 }
954 h_pending--;
956 if (DRM_COPY_TO_USER(ref_age_base + buf_idx + 1, &h_pending, sizeof(u32))) {
957 return -EINVAL;
958 }
960 cmdbuf->buf += sizeof(buf_idx);
961 cmdbuf->bufsz -= sizeof(buf_idx);
962 }
964 BEGIN_RING(2);
965 OUT_RING( CP_PACKET0( RADEON_SCRATCH_REG0 + header.scratch.reg * 4, 0 ) );
966 OUT_RING( dev_priv->scratch_ages[header.scratch.reg] );
967 ADVANCE_RING();
969 return 0;
970 }
972 /**
973 * Uploads user-supplied vertex program instructions or parameters onto
974 * the graphics card.
975 * Called by r300_do_cp_cmdbuf.
976 */
977 static __inline__ int r300_emit_r500fp(drm_radeon_private_t *dev_priv,
978 drm_radeon_kcmd_buffer_t *cmdbuf,
979 drm_r300_cmd_header_t header)
980 {
981 int sz;
982 int addr;
983 int type;
984 int clamp;
985 int stride;
986 RING_LOCALS;
988 sz = header.r500fp.count;
989 /* address is 9 bits 0 - 8, bit 1 of flags is part of address */
990 addr = ((header.r500fp.adrhi_flags & 1) << 8) | header.r500fp.adrlo;
992 type = !!(header.r500fp.adrhi_flags & R500FP_CONSTANT_TYPE);
993 clamp = !!(header.r500fp.adrhi_flags & R500FP_CONSTANT_CLAMP);
995 addr |= (type << 16);
996 addr |= (clamp << 17);
998 stride = type ? 4 : 6;
1000 DRM_DEBUG("r500fp %d %d type: %d\n", sz, addr, type);
1001 if (!sz)
1002 return 0;
1003 if (sz * stride * 4 > cmdbuf->bufsz)
1004 return -EINVAL;
1006 BEGIN_RING(3 + sz * stride);
1007 OUT_RING_REG(R500_GA_US_VECTOR_INDEX, addr);
1008 OUT_RING(CP_PACKET0_TABLE(R500_GA_US_VECTOR_DATA, sz * stride - 1));
1009 OUT_RING_TABLE((int *)cmdbuf->buf, sz * stride);
1011 ADVANCE_RING();
1013 cmdbuf->buf += sz * stride * 4;
1014 cmdbuf->bufsz -= sz * stride * 4;
1016 return 0;
1017 }
1020 /**
1021 * Parses and validates a user-supplied command buffer and emits appropriate
1022 * commands on the DMA ring buffer.
1023 * Called by the ioctl handler function radeon_cp_cmdbuf.
1024 */
1025 int r300_do_cp_cmdbuf(struct drm_device *dev,
1026 struct drm_file *file_priv,
1027 drm_radeon_kcmd_buffer_t *cmdbuf)
1028 {
1029 drm_radeon_private_t *dev_priv = dev->dev_private;
1030 struct drm_device_dma *dma = dev->dma;
1031 struct drm_buf *buf = NULL;
1032 int emit_dispatch_age = 0;
1033 int ret = 0;
1035 DRM_DEBUG("\n");
1037 /* pacify */
1038 r300_pacify(dev_priv);
1040 if (cmdbuf->nbox <= R300_SIMULTANEOUS_CLIPRECTS) {
1041 ret = r300_emit_cliprects(dev_priv, cmdbuf, 0);
1042 if (ret)
1043 goto cleanup;
1044 }
1046 while (cmdbuf->bufsz >= sizeof(drm_r300_cmd_header_t)) {
1047 int idx;
1048 drm_r300_cmd_header_t header;
1050 header.u = *(unsigned int *)cmdbuf->buf;
1052 cmdbuf->buf += sizeof(header);
1053 cmdbuf->bufsz -= sizeof(header);
1055 switch (header.header.cmd_type) {
1056 case R300_CMD_PACKET0:
1057 ret = r300_emit_packet0(dev_priv, cmdbuf, header);
1058 if (ret) {
1059 DRM_ERROR("r300_emit_packet0 failed\n");
1060 goto cleanup;
1061 }
1062 break;
1064 case R300_CMD_VPU:
1065 DRM_DEBUG("R300_CMD_VPU\n");
1066 ret = r300_emit_vpu(dev_priv, cmdbuf, header);
1067 if (ret) {
1068 DRM_ERROR("r300_emit_vpu failed\n");
1069 goto cleanup;
1070 }
1071 break;
1073 case R300_CMD_PACKET3:
1074 DRM_DEBUG("R300_CMD_PACKET3\n");
1075 ret = r300_emit_packet3(dev_priv, cmdbuf, header);
1076 if (ret) {
1077 DRM_ERROR("r300_emit_packet3 failed\n");
1078 goto cleanup;
1079 }
1080 break;
1082 case R300_CMD_END3D:
1083 DRM_DEBUG("R300_CMD_END3D\n");
1084 /* TODO:
1085 Ideally userspace driver should not need to issue this call,
1086 i.e. the drm driver should issue it automatically and prevent
1087 lockups.
1089 In practice, we do not understand why this call is needed and what
1090 it does (except for some vague guesses that it has to do with cache
1091 coherence) and so the user space driver does it.
1093 Once we are sure which uses prevent lockups the code could be moved
1094 into the kernel and the userspace driver will not
1095 need to use this command.
1097 Note that issuing this command does not hurt anything
1098 except, possibly, performance */
1099 r300_pacify(dev_priv);
1100 break;
1102 case R300_CMD_CP_DELAY:
1103 /* simple enough, we can do it here */
1104 DRM_DEBUG("R300_CMD_CP_DELAY\n");
1105 {
1106 int i;
1107 RING_LOCALS;
1109 BEGIN_RING(header.delay.count);
1110 for (i = 0; i < header.delay.count; i++)
1111 OUT_RING(RADEON_CP_PACKET2);
1112 ADVANCE_RING();
1113 }
1114 break;
1116 case R300_CMD_DMA_DISCARD:
1117 DRM_DEBUG("RADEON_CMD_DMA_DISCARD\n");
1118 idx = header.dma.buf_idx;
1119 if (idx < 0 || idx >= dma->buf_count) {
1120 DRM_ERROR("buffer index %d (of %d max)\n",
1121 idx, dma->buf_count - 1);
1122 ret = -EINVAL;
1123 goto cleanup;
1124 }
1126 buf = dma->buflist[idx];
1127 if (buf->file_priv != file_priv || buf->pending) {
1128 DRM_ERROR("bad buffer %p %p %d\n",
1129 buf->file_priv, file_priv,
1130 buf->pending);
1131 ret = -EINVAL;
1132 goto cleanup;
1133 }
1135 emit_dispatch_age = 1;
1136 r300_discard_buffer(dev, buf);
1137 break;
1139 case R300_CMD_WAIT:
1140 DRM_DEBUG("R300_CMD_WAIT\n");
1141 r300_cmd_wait(dev_priv, header);
1142 break;
1144 case R300_CMD_SCRATCH:
1145 DRM_DEBUG("R300_CMD_SCRATCH\n");
1146 ret = r300_scratch(dev_priv, cmdbuf, header);
1147 if (ret) {
1148 DRM_ERROR("r300_scratch failed\n");
1149 goto cleanup;
1150 }
1151 break;
1153 case R300_CMD_R500FP:
1154 if ((dev_priv->flags & RADEON_FAMILY_MASK) < CHIP_RV515) {
1155 DRM_ERROR("Calling r500 command on r300 card\n");
1156 ret = -EINVAL;
1157 goto cleanup;
1158 }
1159 DRM_DEBUG("R300_CMD_R500FP\n");
1160 ret = r300_emit_r500fp(dev_priv, cmdbuf, header);
1161 if (ret) {
1162 DRM_ERROR("r300_emit_r500fp failed\n");
1163 goto cleanup;
1164 }
1165 break;
1166 default:
1167 DRM_ERROR("bad cmd_type %i at %p\n",
1168 header.header.cmd_type,
1169 cmdbuf->buf - sizeof(header));
1170 ret = -EINVAL;
1171 goto cleanup;
1172 }
1173 }
1175 DRM_DEBUG("END\n");
1177 cleanup:
1178 r300_pacify(dev_priv);
1180 /* We emit the vertex buffer age here, outside the pacifier "brackets"
1181 * for two reasons:
1182 * (1) This may coalesce multiple age emissions into a single one and
1183 * (2) more importantly, some chips lock up hard when scratch registers
1184 * are written inside the pacifier bracket.
1185 */
1186 if (emit_dispatch_age) {
1187 RING_LOCALS;
1189 /* Emit the vertex buffer age */
1190 BEGIN_RING(2);
1191 RADEON_DISPATCH_AGE(dev_priv->sarea_priv->last_dispatch);
1192 ADVANCE_RING();
1193 }
1195 COMMIT_RING();
1197 return ret;
1198 }