Full support for Ginger Console
[linux-ginger.git] / drivers / gpu / drm / radeon / r300_cmdbuf.c
blobcb2e470f97d4d100e073cf953b170a72ae1d6d51
1 /* r300_cmdbuf.c -- Command buffer emission for R300 -*- linux-c -*-
3 * Copyright (C) The Weather Channel, Inc. 2002.
4 * Copyright (C) 2004 Nicolai Haehnle.
5 * All Rights Reserved.
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.
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:
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.
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.
30 * Authors:
31 * Nicolai Haehnle <prefect_@gmx.net>
34 #include "drmP.h"
35 #include "drm.h"
36 #include "radeon_drm.h"
37 #include "radeon_drv.h"
38 #include "r300_reg.h"
40 #include <asm/unaligned.h>
42 #define R300_SIMULTANEOUS_CLIPRECTS 4
44 /* Values for R300_RE_CLIPRECT_CNTL depending on the number of cliprects
46 static const int r300_cliprect_cntl[4] = {
47 0xAAAA,
48 0xEEEE,
49 0xFEFE,
50 0xFFFE
53 /**
54 * Emit up to R300_SIMULTANEOUS_CLIPRECTS cliprects from the given command
55 * buffer, starting with index n.
57 static int r300_emit_cliprects(drm_radeon_private_t *dev_priv,
58 drm_radeon_kcmd_buffer_t *cmdbuf, int n)
60 struct drm_clip_rect box;
61 int nr;
62 int i;
63 RING_LOCALS;
65 nr = cmdbuf->nbox - n;
66 if (nr > R300_SIMULTANEOUS_CLIPRECTS)
67 nr = R300_SIMULTANEOUS_CLIPRECTS;
69 DRM_DEBUG("%i cliprects\n", nr);
71 if (nr) {
72 BEGIN_RING(6 + nr * 2);
73 OUT_RING(CP_PACKET0(R300_RE_CLIPRECT_TL_0, nr * 2 - 1));
75 for (i = 0; i < nr; ++i) {
76 if (DRM_COPY_FROM_USER_UNCHECKED
77 (&box, &cmdbuf->boxes[n + i], sizeof(box))) {
78 DRM_ERROR("copy cliprect faulted\n");
79 return -EFAULT;
82 box.x2--; /* Hardware expects inclusive bottom-right corner */
83 box.y2--;
85 if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_RV515) {
86 box.x1 = (box.x1) &
87 R300_CLIPRECT_MASK;
88 box.y1 = (box.y1) &
89 R300_CLIPRECT_MASK;
90 box.x2 = (box.x2) &
91 R300_CLIPRECT_MASK;
92 box.y2 = (box.y2) &
93 R300_CLIPRECT_MASK;
94 } else {
95 box.x1 = (box.x1 + R300_CLIPRECT_OFFSET) &
96 R300_CLIPRECT_MASK;
97 box.y1 = (box.y1 + R300_CLIPRECT_OFFSET) &
98 R300_CLIPRECT_MASK;
99 box.x2 = (box.x2 + R300_CLIPRECT_OFFSET) &
100 R300_CLIPRECT_MASK;
101 box.y2 = (box.y2 + R300_CLIPRECT_OFFSET) &
102 R300_CLIPRECT_MASK;
105 OUT_RING((box.x1 << R300_CLIPRECT_X_SHIFT) |
106 (box.y1 << R300_CLIPRECT_Y_SHIFT));
107 OUT_RING((box.x2 << R300_CLIPRECT_X_SHIFT) |
108 (box.y2 << R300_CLIPRECT_Y_SHIFT));
112 OUT_RING_REG(R300_RE_CLIPRECT_CNTL, r300_cliprect_cntl[nr - 1]);
114 /* TODO/SECURITY: Force scissors to a safe value, otherwise the
115 * client might be able to trample over memory.
116 * The impact should be very limited, but I'd rather be safe than
117 * sorry.
119 OUT_RING(CP_PACKET0(R300_RE_SCISSORS_TL, 1));
120 OUT_RING(0);
121 OUT_RING(R300_SCISSORS_X_MASK | R300_SCISSORS_Y_MASK);
122 ADVANCE_RING();
123 } else {
124 /* Why we allow zero cliprect rendering:
125 * There are some commands in a command buffer that must be submitted
126 * even when there are no cliprects, e.g. DMA buffer discard
127 * or state setting (though state setting could be avoided by
128 * simulating a loss of context).
130 * Now since the cmdbuf interface is so chaotic right now (and is
131 * bound to remain that way for a bit until things settle down),
132 * it is basically impossible to filter out the commands that are
133 * necessary and those that aren't.
135 * So I choose the safe way and don't do any filtering at all;
136 * instead, I simply set up the engine so that all rendering
137 * can't produce any fragments.
139 BEGIN_RING(2);
140 OUT_RING_REG(R300_RE_CLIPRECT_CNTL, 0);
141 ADVANCE_RING();
144 /* flus cache and wait idle clean after cliprect change */
145 BEGIN_RING(2);
146 OUT_RING(CP_PACKET0(R300_RB3D_DSTCACHE_CTLSTAT, 0));
147 OUT_RING(R300_RB3D_DC_FLUSH);
148 ADVANCE_RING();
149 BEGIN_RING(2);
150 OUT_RING(CP_PACKET0(RADEON_WAIT_UNTIL, 0));
151 OUT_RING(RADEON_WAIT_3D_IDLECLEAN);
152 ADVANCE_RING();
153 /* set flush flag */
154 dev_priv->track_flush |= RADEON_FLUSH_EMITED;
156 return 0;
159 static u8 r300_reg_flags[0x10000 >> 2];
161 void r300_init_reg_flags(struct drm_device *dev)
163 int i;
164 drm_radeon_private_t *dev_priv = dev->dev_private;
166 memset(r300_reg_flags, 0, 0x10000 >> 2);
167 #define ADD_RANGE_MARK(reg, count,mark) \
168 for(i=((reg)>>2);i<((reg)>>2)+(count);i++)\
169 r300_reg_flags[i]|=(mark);
171 #define MARK_SAFE 1
172 #define MARK_CHECK_OFFSET 2
174 #define ADD_RANGE(reg, count) ADD_RANGE_MARK(reg, count, MARK_SAFE)
176 /* these match cmducs() command in r300_driver/r300/r300_cmdbuf.c */
177 ADD_RANGE(R300_SE_VPORT_XSCALE, 6);
178 ADD_RANGE(R300_VAP_CNTL, 1);
179 ADD_RANGE(R300_SE_VTE_CNTL, 2);
180 ADD_RANGE(0x2134, 2);
181 ADD_RANGE(R300_VAP_CNTL_STATUS, 1);
182 ADD_RANGE(R300_VAP_INPUT_CNTL_0, 2);
183 ADD_RANGE(0x21DC, 1);
184 ADD_RANGE(R300_VAP_UNKNOWN_221C, 1);
185 ADD_RANGE(R300_VAP_CLIP_X_0, 4);
186 ADD_RANGE(R300_VAP_PVS_STATE_FLUSH_REG, 1);
187 ADD_RANGE(R300_VAP_UNKNOWN_2288, 1);
188 ADD_RANGE(R300_VAP_OUTPUT_VTX_FMT_0, 2);
189 ADD_RANGE(R300_VAP_PVS_CNTL_1, 3);
190 ADD_RANGE(R300_GB_ENABLE, 1);
191 ADD_RANGE(R300_GB_MSPOS0, 5);
192 ADD_RANGE(R300_TX_INVALTAGS, 1);
193 ADD_RANGE(R300_TX_ENABLE, 1);
194 ADD_RANGE(0x4200, 4);
195 ADD_RANGE(0x4214, 1);
196 ADD_RANGE(R300_RE_POINTSIZE, 1);
197 ADD_RANGE(0x4230, 3);
198 ADD_RANGE(R300_RE_LINE_CNT, 1);
199 ADD_RANGE(R300_RE_UNK4238, 1);
200 ADD_RANGE(0x4260, 3);
201 ADD_RANGE(R300_RE_SHADE, 4);
202 ADD_RANGE(R300_RE_POLYGON_MODE, 5);
203 ADD_RANGE(R300_RE_ZBIAS_CNTL, 1);
204 ADD_RANGE(R300_RE_ZBIAS_T_FACTOR, 4);
205 ADD_RANGE(R300_RE_OCCLUSION_CNTL, 1);
206 ADD_RANGE(R300_RE_CULL_CNTL, 1);
207 ADD_RANGE(0x42C0, 2);
208 ADD_RANGE(R300_RS_CNTL_0, 2);
210 ADD_RANGE(R300_SU_REG_DEST, 1);
211 if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV530)
212 ADD_RANGE(RV530_FG_ZBREG_DEST, 1);
214 ADD_RANGE(R300_SC_HYPERZ, 2);
215 ADD_RANGE(0x43E8, 1);
217 ADD_RANGE(0x46A4, 5);
219 ADD_RANGE(R300_RE_FOG_STATE, 1);
220 ADD_RANGE(R300_FOG_COLOR_R, 3);
221 ADD_RANGE(R300_PP_ALPHA_TEST, 2);
222 ADD_RANGE(0x4BD8, 1);
223 ADD_RANGE(R300_PFS_PARAM_0_X, 64);
224 ADD_RANGE(0x4E00, 1);
225 ADD_RANGE(R300_RB3D_CBLEND, 2);
226 ADD_RANGE(R300_RB3D_COLORMASK, 1);
227 ADD_RANGE(R300_RB3D_BLEND_COLOR, 3);
228 ADD_RANGE_MARK(R300_RB3D_COLOROFFSET0, 1, MARK_CHECK_OFFSET); /* check offset */
229 ADD_RANGE(R300_RB3D_COLORPITCH0, 1);
230 ADD_RANGE(0x4E50, 9);
231 ADD_RANGE(0x4E88, 1);
232 ADD_RANGE(0x4EA0, 2);
233 ADD_RANGE(R300_ZB_CNTL, 3);
234 ADD_RANGE(R300_ZB_FORMAT, 4);
235 ADD_RANGE_MARK(R300_ZB_DEPTHOFFSET, 1, MARK_CHECK_OFFSET); /* check offset */
236 ADD_RANGE(R300_ZB_DEPTHPITCH, 1);
237 ADD_RANGE(R300_ZB_DEPTHCLEARVALUE, 1);
238 ADD_RANGE(R300_ZB_ZMASK_OFFSET, 13);
239 ADD_RANGE(R300_ZB_ZPASS_DATA, 2); /* ZB_ZPASS_DATA, ZB_ZPASS_ADDR */
241 ADD_RANGE(R300_TX_FILTER_0, 16);
242 ADD_RANGE(R300_TX_FILTER1_0, 16);
243 ADD_RANGE(R300_TX_SIZE_0, 16);
244 ADD_RANGE(R300_TX_FORMAT_0, 16);
245 ADD_RANGE(R300_TX_PITCH_0, 16);
246 /* Texture offset is dangerous and needs more checking */
247 ADD_RANGE_MARK(R300_TX_OFFSET_0, 16, MARK_CHECK_OFFSET);
248 ADD_RANGE(R300_TX_CHROMA_KEY_0, 16);
249 ADD_RANGE(R300_TX_BORDER_COLOR_0, 16);
251 /* Sporadic registers used as primitives are emitted */
252 ADD_RANGE(R300_ZB_ZCACHE_CTLSTAT, 1);
253 ADD_RANGE(R300_RB3D_DSTCACHE_CTLSTAT, 1);
254 ADD_RANGE(R300_VAP_INPUT_ROUTE_0_0, 8);
255 ADD_RANGE(R300_VAP_INPUT_ROUTE_1_0, 8);
257 if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_RV515) {
258 ADD_RANGE(R500_VAP_INDEX_OFFSET, 1);
259 ADD_RANGE(R500_US_CONFIG, 2);
260 ADD_RANGE(R500_US_CODE_ADDR, 3);
261 ADD_RANGE(R500_US_FC_CTRL, 1);
262 ADD_RANGE(R500_RS_IP_0, 16);
263 ADD_RANGE(R500_RS_INST_0, 16);
264 ADD_RANGE(R500_RB3D_COLOR_CLEAR_VALUE_AR, 2);
265 ADD_RANGE(R500_RB3D_CONSTANT_COLOR_AR, 2);
266 ADD_RANGE(R500_ZB_FIFO_SIZE, 2);
267 } else {
268 ADD_RANGE(R300_PFS_CNTL_0, 3);
269 ADD_RANGE(R300_PFS_NODE_0, 4);
270 ADD_RANGE(R300_PFS_TEXI_0, 64);
271 ADD_RANGE(R300_PFS_INSTR0_0, 64);
272 ADD_RANGE(R300_PFS_INSTR1_0, 64);
273 ADD_RANGE(R300_PFS_INSTR2_0, 64);
274 ADD_RANGE(R300_PFS_INSTR3_0, 64);
275 ADD_RANGE(R300_RS_INTERP_0, 8);
276 ADD_RANGE(R300_RS_ROUTE_0, 8);
281 static __inline__ int r300_check_range(unsigned reg, int count)
283 int i;
284 if (reg & ~0xffff)
285 return -1;
286 for (i = (reg >> 2); i < (reg >> 2) + count; i++)
287 if (r300_reg_flags[i] != MARK_SAFE)
288 return 1;
289 return 0;
292 static __inline__ int r300_emit_carefully_checked_packet0(drm_radeon_private_t *
293 dev_priv,
294 drm_radeon_kcmd_buffer_t
295 * cmdbuf,
296 drm_r300_cmd_header_t
297 header)
299 int reg;
300 int sz;
301 int i;
302 int values[64];
303 RING_LOCALS;
305 sz = header.packet0.count;
306 reg = (header.packet0.reghi << 8) | header.packet0.reglo;
308 if ((sz > 64) || (sz < 0)) {
309 DRM_ERROR
310 ("Cannot emit more than 64 values at a time (reg=%04x sz=%d)\n",
311 reg, sz);
312 return -EINVAL;
314 for (i = 0; i < sz; i++) {
315 values[i] = ((int *)cmdbuf->buf)[i];
316 switch (r300_reg_flags[(reg >> 2) + i]) {
317 case MARK_SAFE:
318 break;
319 case MARK_CHECK_OFFSET:
320 if (!radeon_check_offset(dev_priv, (u32) values[i])) {
321 DRM_ERROR
322 ("Offset failed range check (reg=%04x sz=%d)\n",
323 reg, sz);
324 return -EINVAL;
326 break;
327 default:
328 DRM_ERROR("Register %04x failed check as flag=%02x\n",
329 reg + i * 4, r300_reg_flags[(reg >> 2) + i]);
330 return -EINVAL;
334 BEGIN_RING(1 + sz);
335 OUT_RING(CP_PACKET0(reg, sz - 1));
336 OUT_RING_TABLE(values, sz);
337 ADVANCE_RING();
339 cmdbuf->buf += sz * 4;
340 cmdbuf->bufsz -= sz * 4;
342 return 0;
346 * Emits a packet0 setting arbitrary registers.
347 * Called by r300_do_cp_cmdbuf.
349 * Note that checks are performed on contents and addresses of the registers
351 static __inline__ int r300_emit_packet0(drm_radeon_private_t *dev_priv,
352 drm_radeon_kcmd_buffer_t *cmdbuf,
353 drm_r300_cmd_header_t header)
355 int reg;
356 int sz;
357 RING_LOCALS;
359 sz = header.packet0.count;
360 reg = (header.packet0.reghi << 8) | header.packet0.reglo;
362 if (!sz)
363 return 0;
365 if (sz * 4 > cmdbuf->bufsz)
366 return -EINVAL;
368 if (reg + sz * 4 >= 0x10000) {
369 DRM_ERROR("No such registers in hardware reg=%04x sz=%d\n", reg,
370 sz);
371 return -EINVAL;
374 if (r300_check_range(reg, sz)) {
375 /* go and check everything */
376 return r300_emit_carefully_checked_packet0(dev_priv, cmdbuf,
377 header);
379 /* the rest of the data is safe to emit, whatever the values the user passed */
381 BEGIN_RING(1 + sz);
382 OUT_RING(CP_PACKET0(reg, sz - 1));
383 OUT_RING_TABLE((int *)cmdbuf->buf, sz);
384 ADVANCE_RING();
386 cmdbuf->buf += sz * 4;
387 cmdbuf->bufsz -= sz * 4;
389 return 0;
393 * Uploads user-supplied vertex program instructions or parameters onto
394 * the graphics card.
395 * Called by r300_do_cp_cmdbuf.
397 static __inline__ int r300_emit_vpu(drm_radeon_private_t *dev_priv,
398 drm_radeon_kcmd_buffer_t *cmdbuf,
399 drm_r300_cmd_header_t header)
401 int sz;
402 int addr;
403 RING_LOCALS;
405 sz = header.vpu.count;
406 addr = (header.vpu.adrhi << 8) | header.vpu.adrlo;
408 if (!sz)
409 return 0;
410 if (sz * 16 > cmdbuf->bufsz)
411 return -EINVAL;
413 /* VAP is very sensitive so we purge cache before we program it
414 * and we also flush its state before & after */
415 BEGIN_RING(6);
416 OUT_RING(CP_PACKET0(R300_RB3D_DSTCACHE_CTLSTAT, 0));
417 OUT_RING(R300_RB3D_DC_FLUSH);
418 OUT_RING(CP_PACKET0(RADEON_WAIT_UNTIL, 0));
419 OUT_RING(RADEON_WAIT_3D_IDLECLEAN);
420 OUT_RING(CP_PACKET0(R300_VAP_PVS_STATE_FLUSH_REG, 0));
421 OUT_RING(0);
422 ADVANCE_RING();
423 /* set flush flag */
424 dev_priv->track_flush |= RADEON_FLUSH_EMITED;
426 BEGIN_RING(3 + sz * 4);
427 OUT_RING_REG(R300_VAP_PVS_UPLOAD_ADDRESS, addr);
428 OUT_RING(CP_PACKET0_TABLE(R300_VAP_PVS_UPLOAD_DATA, sz * 4 - 1));
429 OUT_RING_TABLE((int *)cmdbuf->buf, sz * 4);
430 ADVANCE_RING();
432 BEGIN_RING(2);
433 OUT_RING(CP_PACKET0(R300_VAP_PVS_STATE_FLUSH_REG, 0));
434 OUT_RING(0);
435 ADVANCE_RING();
437 cmdbuf->buf += sz * 16;
438 cmdbuf->bufsz -= sz * 16;
440 return 0;
444 * Emit a clear packet from userspace.
445 * Called by r300_emit_packet3.
447 static __inline__ int r300_emit_clear(drm_radeon_private_t *dev_priv,
448 drm_radeon_kcmd_buffer_t *cmdbuf)
450 RING_LOCALS;
452 if (8 * 4 > cmdbuf->bufsz)
453 return -EINVAL;
455 BEGIN_RING(10);
456 OUT_RING(CP_PACKET3(R200_3D_DRAW_IMMD_2, 8));
457 OUT_RING(R300_PRIM_TYPE_POINT | R300_PRIM_WALK_RING |
458 (1 << R300_PRIM_NUM_VERTICES_SHIFT));
459 OUT_RING_TABLE((int *)cmdbuf->buf, 8);
460 ADVANCE_RING();
462 BEGIN_RING(4);
463 OUT_RING(CP_PACKET0(R300_RB3D_DSTCACHE_CTLSTAT, 0));
464 OUT_RING(R300_RB3D_DC_FLUSH);
465 OUT_RING(CP_PACKET0(RADEON_WAIT_UNTIL, 0));
466 OUT_RING(RADEON_WAIT_3D_IDLECLEAN);
467 ADVANCE_RING();
468 /* set flush flag */
469 dev_priv->track_flush |= RADEON_FLUSH_EMITED;
471 cmdbuf->buf += 8 * 4;
472 cmdbuf->bufsz -= 8 * 4;
474 return 0;
477 static __inline__ int r300_emit_3d_load_vbpntr(drm_radeon_private_t *dev_priv,
478 drm_radeon_kcmd_buffer_t *cmdbuf,
479 u32 header)
481 int count, i, k;
482 #define MAX_ARRAY_PACKET 64
483 u32 payload[MAX_ARRAY_PACKET];
484 u32 narrays;
485 RING_LOCALS;
487 count = (header >> 16) & 0x3fff;
489 if ((count + 1) > MAX_ARRAY_PACKET) {
490 DRM_ERROR("Too large payload in 3D_LOAD_VBPNTR (count=%d)\n",
491 count);
492 return -EINVAL;
494 memset(payload, 0, MAX_ARRAY_PACKET * 4);
495 memcpy(payload, cmdbuf->buf + 4, (count + 1) * 4);
497 /* carefully check packet contents */
499 narrays = payload[0];
500 k = 0;
501 i = 1;
502 while ((k < narrays) && (i < (count + 1))) {
503 i++; /* skip attribute field */
504 if (!radeon_check_offset(dev_priv, payload[i])) {
505 DRM_ERROR
506 ("Offset failed range check (k=%d i=%d) while processing 3D_LOAD_VBPNTR packet.\n",
507 k, i);
508 return -EINVAL;
510 k++;
511 i++;
512 if (k == narrays)
513 break;
514 /* have one more to process, they come in pairs */
515 if (!radeon_check_offset(dev_priv, payload[i])) {
516 DRM_ERROR
517 ("Offset failed range check (k=%d i=%d) while processing 3D_LOAD_VBPNTR packet.\n",
518 k, i);
519 return -EINVAL;
521 k++;
522 i++;
524 /* do the counts match what we expect ? */
525 if ((k != narrays) || (i != (count + 1))) {
526 DRM_ERROR
527 ("Malformed 3D_LOAD_VBPNTR packet (k=%d i=%d narrays=%d count+1=%d).\n",
528 k, i, narrays, count + 1);
529 return -EINVAL;
532 /* all clear, output packet */
534 BEGIN_RING(count + 2);
535 OUT_RING(header);
536 OUT_RING_TABLE(payload, count + 1);
537 ADVANCE_RING();
539 cmdbuf->buf += (count + 2) * 4;
540 cmdbuf->bufsz -= (count + 2) * 4;
542 return 0;
545 static __inline__ int r300_emit_bitblt_multi(drm_radeon_private_t *dev_priv,
546 drm_radeon_kcmd_buffer_t *cmdbuf)
548 u32 *cmd = (u32 *) cmdbuf->buf;
549 int count, ret;
550 RING_LOCALS;
552 count=(cmd[0]>>16) & 0x3fff;
554 if (cmd[0] & 0x8000) {
555 u32 offset;
557 if (cmd[1] & (RADEON_GMC_SRC_PITCH_OFFSET_CNTL
558 | RADEON_GMC_DST_PITCH_OFFSET_CNTL)) {
559 offset = cmd[2] << 10;
560 ret = !radeon_check_offset(dev_priv, offset);
561 if (ret) {
562 DRM_ERROR("Invalid bitblt first offset is %08X\n", offset);
563 return -EINVAL;
567 if ((cmd[1] & RADEON_GMC_SRC_PITCH_OFFSET_CNTL) &&
568 (cmd[1] & RADEON_GMC_DST_PITCH_OFFSET_CNTL)) {
569 offset = cmd[3] << 10;
570 ret = !radeon_check_offset(dev_priv, offset);
571 if (ret) {
572 DRM_ERROR("Invalid bitblt second offset is %08X\n", offset);
573 return -EINVAL;
579 BEGIN_RING(count+2);
580 OUT_RING(cmd[0]);
581 OUT_RING_TABLE((int *)(cmdbuf->buf + 4), count + 1);
582 ADVANCE_RING();
584 cmdbuf->buf += (count+2)*4;
585 cmdbuf->bufsz -= (count+2)*4;
587 return 0;
590 static __inline__ int r300_emit_draw_indx_2(drm_radeon_private_t *dev_priv,
591 drm_radeon_kcmd_buffer_t *cmdbuf)
593 u32 *cmd;
594 int count;
595 int expected_count;
596 RING_LOCALS;
598 cmd = (u32 *) cmdbuf->buf;
599 count = (cmd[0]>>16) & 0x3fff;
600 expected_count = cmd[1] >> 16;
601 if (!(cmd[1] & R300_VAP_VF_CNTL__INDEX_SIZE_32bit))
602 expected_count = (expected_count+1)/2;
604 if (count && count != expected_count) {
605 DRM_ERROR("3D_DRAW_INDX_2: packet size %i, expected %i\n",
606 count, expected_count);
607 return -EINVAL;
610 BEGIN_RING(count+2);
611 OUT_RING(cmd[0]);
612 OUT_RING_TABLE((int *)(cmdbuf->buf + 4), count + 1);
613 ADVANCE_RING();
615 cmdbuf->buf += (count+2)*4;
616 cmdbuf->bufsz -= (count+2)*4;
618 if (!count) {
619 drm_r300_cmd_header_t header;
621 if (cmdbuf->bufsz < 4*4 + sizeof(header)) {
622 DRM_ERROR("3D_DRAW_INDX_2: expect subsequent INDX_BUFFER, but stream is too short.\n");
623 return -EINVAL;
626 header.u = *(unsigned int *)cmdbuf->buf;
628 cmdbuf->buf += sizeof(header);
629 cmdbuf->bufsz -= sizeof(header);
630 cmd = (u32 *) cmdbuf->buf;
632 if (header.header.cmd_type != R300_CMD_PACKET3 ||
633 header.packet3.packet != R300_CMD_PACKET3_RAW ||
634 cmd[0] != CP_PACKET3(RADEON_CP_INDX_BUFFER, 2)) {
635 DRM_ERROR("3D_DRAW_INDX_2: expect subsequent INDX_BUFFER.\n");
636 return -EINVAL;
639 if ((cmd[1] & 0x8000ffff) != 0x80000810) {
640 DRM_ERROR("Invalid indx_buffer reg address %08X\n", cmd[1]);
641 return -EINVAL;
643 if (!radeon_check_offset(dev_priv, cmd[2])) {
644 DRM_ERROR("Invalid indx_buffer offset is %08X\n", cmd[2]);
645 return -EINVAL;
647 if (cmd[3] != expected_count) {
648 DRM_ERROR("INDX_BUFFER: buffer size %i, expected %i\n",
649 cmd[3], expected_count);
650 return -EINVAL;
653 BEGIN_RING(4);
654 OUT_RING(cmd[0]);
655 OUT_RING_TABLE((int *)(cmdbuf->buf + 4), 3);
656 ADVANCE_RING();
658 cmdbuf->buf += 4*4;
659 cmdbuf->bufsz -= 4*4;
662 return 0;
665 static __inline__ int r300_emit_raw_packet3(drm_radeon_private_t *dev_priv,
666 drm_radeon_kcmd_buffer_t *cmdbuf)
668 u32 header;
669 int count;
670 RING_LOCALS;
672 if (4 > cmdbuf->bufsz)
673 return -EINVAL;
675 /* Fixme !! This simply emits a packet without much checking.
676 We need to be smarter. */
678 /* obtain first word - actual packet3 header */
679 header = *(u32 *) cmdbuf->buf;
681 /* Is it packet 3 ? */
682 if ((header >> 30) != 0x3) {
683 DRM_ERROR("Not a packet3 header (0x%08x)\n", header);
684 return -EINVAL;
687 count = (header >> 16) & 0x3fff;
689 /* Check again now that we know how much data to expect */
690 if ((count + 2) * 4 > cmdbuf->bufsz) {
691 DRM_ERROR
692 ("Expected packet3 of length %d but have only %d bytes left\n",
693 (count + 2) * 4, cmdbuf->bufsz);
694 return -EINVAL;
697 /* Is it a packet type we know about ? */
698 switch (header & 0xff00) {
699 case RADEON_3D_LOAD_VBPNTR: /* load vertex array pointers */
700 return r300_emit_3d_load_vbpntr(dev_priv, cmdbuf, header);
702 case RADEON_CNTL_BITBLT_MULTI:
703 return r300_emit_bitblt_multi(dev_priv, cmdbuf);
705 case RADEON_CP_INDX_BUFFER:
706 DRM_ERROR("packet3 INDX_BUFFER without preceding 3D_DRAW_INDX_2 is illegal.\n");
707 return -EINVAL;
708 case RADEON_CP_3D_DRAW_IMMD_2:
709 /* triggers drawing using in-packet vertex data */
710 case RADEON_CP_3D_DRAW_VBUF_2:
711 /* triggers drawing of vertex buffers setup elsewhere */
712 dev_priv->track_flush &= ~(RADEON_FLUSH_EMITED |
713 RADEON_PURGE_EMITED);
714 break;
715 case RADEON_CP_3D_DRAW_INDX_2:
716 /* triggers drawing using indices to vertex buffer */
717 /* whenever we send vertex we clear flush & purge */
718 dev_priv->track_flush &= ~(RADEON_FLUSH_EMITED |
719 RADEON_PURGE_EMITED);
720 return r300_emit_draw_indx_2(dev_priv, cmdbuf);
721 case RADEON_WAIT_FOR_IDLE:
722 case RADEON_CP_NOP:
723 /* these packets are safe */
724 break;
725 default:
726 DRM_ERROR("Unknown packet3 header (0x%08x)\n", header);
727 return -EINVAL;
730 BEGIN_RING(count + 2);
731 OUT_RING(header);
732 OUT_RING_TABLE((int *)(cmdbuf->buf + 4), count + 1);
733 ADVANCE_RING();
735 cmdbuf->buf += (count + 2) * 4;
736 cmdbuf->bufsz -= (count + 2) * 4;
738 return 0;
742 * Emit a rendering packet3 from userspace.
743 * Called by r300_do_cp_cmdbuf.
745 static __inline__ int r300_emit_packet3(drm_radeon_private_t *dev_priv,
746 drm_radeon_kcmd_buffer_t *cmdbuf,
747 drm_r300_cmd_header_t header)
749 int n;
750 int ret;
751 char *orig_buf = cmdbuf->buf;
752 int orig_bufsz = cmdbuf->bufsz;
754 /* This is a do-while-loop so that we run the interior at least once,
755 * even if cmdbuf->nbox is 0. Compare r300_emit_cliprects for rationale.
757 n = 0;
758 do {
759 if (cmdbuf->nbox > R300_SIMULTANEOUS_CLIPRECTS) {
760 ret = r300_emit_cliprects(dev_priv, cmdbuf, n);
761 if (ret)
762 return ret;
764 cmdbuf->buf = orig_buf;
765 cmdbuf->bufsz = orig_bufsz;
768 switch (header.packet3.packet) {
769 case R300_CMD_PACKET3_CLEAR:
770 DRM_DEBUG("R300_CMD_PACKET3_CLEAR\n");
771 ret = r300_emit_clear(dev_priv, cmdbuf);
772 if (ret) {
773 DRM_ERROR("r300_emit_clear failed\n");
774 return ret;
776 break;
778 case R300_CMD_PACKET3_RAW:
779 DRM_DEBUG("R300_CMD_PACKET3_RAW\n");
780 ret = r300_emit_raw_packet3(dev_priv, cmdbuf);
781 if (ret) {
782 DRM_ERROR("r300_emit_raw_packet3 failed\n");
783 return ret;
785 break;
787 default:
788 DRM_ERROR("bad packet3 type %i at %p\n",
789 header.packet3.packet,
790 cmdbuf->buf - sizeof(header));
791 return -EINVAL;
794 n += R300_SIMULTANEOUS_CLIPRECTS;
795 } while (n < cmdbuf->nbox);
797 return 0;
800 /* Some of the R300 chips seem to be extremely touchy about the two registers
801 * that are configured in r300_pacify.
802 * Among the worst offenders seems to be the R300 ND (0x4E44): When userspace
803 * sends a command buffer that contains only state setting commands and a
804 * vertex program/parameter upload sequence, this will eventually lead to a
805 * lockup, unless the sequence is bracketed by calls to r300_pacify.
806 * So we should take great care to *always* call r300_pacify before
807 * *anything* 3D related, and again afterwards. This is what the
808 * call bracket in r300_do_cp_cmdbuf is for.
812 * Emit the sequence to pacify R300.
814 static __inline__ void r300_pacify(drm_radeon_private_t *dev_priv)
816 uint32_t cache_z, cache_3d, cache_2d;
817 RING_LOCALS;
819 cache_z = R300_ZC_FLUSH;
820 cache_2d = R300_RB2D_DC_FLUSH;
821 cache_3d = R300_RB3D_DC_FLUSH;
822 if (!(dev_priv->track_flush & RADEON_PURGE_EMITED)) {
823 /* we can purge, primitive where draw since last purge */
824 cache_z |= R300_ZC_FREE;
825 cache_2d |= R300_RB2D_DC_FREE;
826 cache_3d |= R300_RB3D_DC_FREE;
829 /* flush & purge zbuffer */
830 BEGIN_RING(2);
831 OUT_RING(CP_PACKET0(R300_ZB_ZCACHE_CTLSTAT, 0));
832 OUT_RING(cache_z);
833 ADVANCE_RING();
834 /* flush & purge 3d */
835 BEGIN_RING(2);
836 OUT_RING(CP_PACKET0(R300_RB3D_DSTCACHE_CTLSTAT, 0));
837 OUT_RING(cache_3d);
838 ADVANCE_RING();
839 /* flush & purge texture */
840 BEGIN_RING(2);
841 OUT_RING(CP_PACKET0(R300_TX_INVALTAGS, 0));
842 OUT_RING(0);
843 ADVANCE_RING();
844 /* FIXME: is this one really needed ? */
845 BEGIN_RING(2);
846 OUT_RING(CP_PACKET0(R300_RB3D_AARESOLVE_CTL, 0));
847 OUT_RING(0);
848 ADVANCE_RING();
849 BEGIN_RING(2);
850 OUT_RING(CP_PACKET0(RADEON_WAIT_UNTIL, 0));
851 OUT_RING(RADEON_WAIT_3D_IDLECLEAN);
852 ADVANCE_RING();
853 /* flush & purge 2d through E2 as RB2D will trigger lockup */
854 BEGIN_RING(4);
855 OUT_RING(CP_PACKET0(R300_DSTCACHE_CTLSTAT, 0));
856 OUT_RING(cache_2d);
857 OUT_RING(CP_PACKET0(RADEON_WAIT_UNTIL, 0));
858 OUT_RING(RADEON_WAIT_2D_IDLECLEAN |
859 RADEON_WAIT_HOST_IDLECLEAN);
860 ADVANCE_RING();
861 /* set flush & purge flags */
862 dev_priv->track_flush |= RADEON_FLUSH_EMITED | RADEON_PURGE_EMITED;
866 * Called by r300_do_cp_cmdbuf to update the internal buffer age and state.
867 * The actual age emit is done by r300_do_cp_cmdbuf, which is why you must
868 * be careful about how this function is called.
870 static void r300_discard_buffer(struct drm_device *dev, struct drm_master *master, struct drm_buf *buf)
872 drm_radeon_buf_priv_t *buf_priv = buf->dev_private;
873 struct drm_radeon_master_private *master_priv = master->driver_priv;
875 buf_priv->age = ++master_priv->sarea_priv->last_dispatch;
876 buf->pending = 1;
877 buf->used = 0;
880 static void r300_cmd_wait(drm_radeon_private_t * dev_priv,
881 drm_r300_cmd_header_t header)
883 u32 wait_until;
884 RING_LOCALS;
886 if (!header.wait.flags)
887 return;
889 wait_until = 0;
891 switch(header.wait.flags) {
892 case R300_WAIT_2D:
893 wait_until = RADEON_WAIT_2D_IDLE;
894 break;
895 case R300_WAIT_3D:
896 wait_until = RADEON_WAIT_3D_IDLE;
897 break;
898 case R300_NEW_WAIT_2D_3D:
899 wait_until = RADEON_WAIT_2D_IDLE|RADEON_WAIT_3D_IDLE;
900 break;
901 case R300_NEW_WAIT_2D_2D_CLEAN:
902 wait_until = RADEON_WAIT_2D_IDLE|RADEON_WAIT_2D_IDLECLEAN;
903 break;
904 case R300_NEW_WAIT_3D_3D_CLEAN:
905 wait_until = RADEON_WAIT_3D_IDLE|RADEON_WAIT_3D_IDLECLEAN;
906 break;
907 case R300_NEW_WAIT_2D_2D_CLEAN_3D_3D_CLEAN:
908 wait_until = RADEON_WAIT_2D_IDLE|RADEON_WAIT_2D_IDLECLEAN;
909 wait_until |= RADEON_WAIT_3D_IDLE|RADEON_WAIT_3D_IDLECLEAN;
910 break;
911 default:
912 return;
915 BEGIN_RING(2);
916 OUT_RING(CP_PACKET0(RADEON_WAIT_UNTIL, 0));
917 OUT_RING(wait_until);
918 ADVANCE_RING();
921 static int r300_scratch(drm_radeon_private_t *dev_priv,
922 drm_radeon_kcmd_buffer_t *cmdbuf,
923 drm_r300_cmd_header_t header)
925 u32 *ref_age_base;
926 u32 i, buf_idx, h_pending;
927 u64 ptr_addr;
928 RING_LOCALS;
930 if (cmdbuf->bufsz <
931 (sizeof(u64) + header.scratch.n_bufs * sizeof(buf_idx))) {
932 return -EINVAL;
935 if (header.scratch.reg >= 5) {
936 return -EINVAL;
939 dev_priv->scratch_ages[header.scratch.reg]++;
941 ptr_addr = get_unaligned((u64 *)cmdbuf->buf);
942 ref_age_base = (u32 *)(unsigned long)ptr_addr;
944 cmdbuf->buf += sizeof(u64);
945 cmdbuf->bufsz -= sizeof(u64);
947 for (i=0; i < header.scratch.n_bufs; i++) {
948 buf_idx = *(u32 *)cmdbuf->buf;
949 buf_idx *= 2; /* 8 bytes per buf */
951 if (DRM_COPY_TO_USER(ref_age_base + buf_idx, &dev_priv->scratch_ages[header.scratch.reg], sizeof(u32))) {
952 return -EINVAL;
955 if (DRM_COPY_FROM_USER(&h_pending, ref_age_base + buf_idx + 1, sizeof(u32))) {
956 return -EINVAL;
959 if (h_pending == 0) {
960 return -EINVAL;
963 h_pending--;
965 if (DRM_COPY_TO_USER(ref_age_base + buf_idx + 1, &h_pending, sizeof(u32))) {
966 return -EINVAL;
969 cmdbuf->buf += sizeof(buf_idx);
970 cmdbuf->bufsz -= sizeof(buf_idx);
973 BEGIN_RING(2);
974 OUT_RING( CP_PACKET0( RADEON_SCRATCH_REG0 + header.scratch.reg * 4, 0 ) );
975 OUT_RING( dev_priv->scratch_ages[header.scratch.reg] );
976 ADVANCE_RING();
978 return 0;
982 * Uploads user-supplied vertex program instructions or parameters onto
983 * the graphics card.
984 * Called by r300_do_cp_cmdbuf.
986 static inline int r300_emit_r500fp(drm_radeon_private_t *dev_priv,
987 drm_radeon_kcmd_buffer_t *cmdbuf,
988 drm_r300_cmd_header_t header)
990 int sz;
991 int addr;
992 int type;
993 int clamp;
994 int stride;
995 RING_LOCALS;
997 sz = header.r500fp.count;
998 /* address is 9 bits 0 - 8, bit 1 of flags is part of address */
999 addr = ((header.r500fp.adrhi_flags & 1) << 8) | header.r500fp.adrlo;
1001 type = !!(header.r500fp.adrhi_flags & R500FP_CONSTANT_TYPE);
1002 clamp = !!(header.r500fp.adrhi_flags & R500FP_CONSTANT_CLAMP);
1004 addr |= (type << 16);
1005 addr |= (clamp << 17);
1007 stride = type ? 4 : 6;
1009 DRM_DEBUG("r500fp %d %d type: %d\n", sz, addr, type);
1010 if (!sz)
1011 return 0;
1012 if (sz * stride * 4 > cmdbuf->bufsz)
1013 return -EINVAL;
1015 BEGIN_RING(3 + sz * stride);
1016 OUT_RING_REG(R500_GA_US_VECTOR_INDEX, addr);
1017 OUT_RING(CP_PACKET0_TABLE(R500_GA_US_VECTOR_DATA, sz * stride - 1));
1018 OUT_RING_TABLE((int *)cmdbuf->buf, sz * stride);
1020 ADVANCE_RING();
1022 cmdbuf->buf += sz * stride * 4;
1023 cmdbuf->bufsz -= sz * stride * 4;
1025 return 0;
1030 * Parses and validates a user-supplied command buffer and emits appropriate
1031 * commands on the DMA ring buffer.
1032 * Called by the ioctl handler function radeon_cp_cmdbuf.
1034 int r300_do_cp_cmdbuf(struct drm_device *dev,
1035 struct drm_file *file_priv,
1036 drm_radeon_kcmd_buffer_t *cmdbuf)
1038 drm_radeon_private_t *dev_priv = dev->dev_private;
1039 struct drm_radeon_master_private *master_priv = file_priv->master->driver_priv;
1040 struct drm_device_dma *dma = dev->dma;
1041 struct drm_buf *buf = NULL;
1042 int emit_dispatch_age = 0;
1043 int ret = 0;
1045 DRM_DEBUG("\n");
1047 /* pacify */
1048 r300_pacify(dev_priv);
1050 if (cmdbuf->nbox <= R300_SIMULTANEOUS_CLIPRECTS) {
1051 ret = r300_emit_cliprects(dev_priv, cmdbuf, 0);
1052 if (ret)
1053 goto cleanup;
1056 while (cmdbuf->bufsz >= sizeof(drm_r300_cmd_header_t)) {
1057 int idx;
1058 drm_r300_cmd_header_t header;
1060 header.u = *(unsigned int *)cmdbuf->buf;
1062 cmdbuf->buf += sizeof(header);
1063 cmdbuf->bufsz -= sizeof(header);
1065 switch (header.header.cmd_type) {
1066 case R300_CMD_PACKET0:
1067 DRM_DEBUG("R300_CMD_PACKET0\n");
1068 ret = r300_emit_packet0(dev_priv, cmdbuf, header);
1069 if (ret) {
1070 DRM_ERROR("r300_emit_packet0 failed\n");
1071 goto cleanup;
1073 break;
1075 case R300_CMD_VPU:
1076 DRM_DEBUG("R300_CMD_VPU\n");
1077 ret = r300_emit_vpu(dev_priv, cmdbuf, header);
1078 if (ret) {
1079 DRM_ERROR("r300_emit_vpu failed\n");
1080 goto cleanup;
1082 break;
1084 case R300_CMD_PACKET3:
1085 DRM_DEBUG("R300_CMD_PACKET3\n");
1086 ret = r300_emit_packet3(dev_priv, cmdbuf, header);
1087 if (ret) {
1088 DRM_ERROR("r300_emit_packet3 failed\n");
1089 goto cleanup;
1091 break;
1093 case R300_CMD_END3D:
1094 DRM_DEBUG("R300_CMD_END3D\n");
1095 /* TODO:
1096 Ideally userspace driver should not need to issue this call,
1097 i.e. the drm driver should issue it automatically and prevent
1098 lockups.
1100 In practice, we do not understand why this call is needed and what
1101 it does (except for some vague guesses that it has to do with cache
1102 coherence) and so the user space driver does it.
1104 Once we are sure which uses prevent lockups the code could be moved
1105 into the kernel and the userspace driver will not
1106 need to use this command.
1108 Note that issuing this command does not hurt anything
1109 except, possibly, performance */
1110 r300_pacify(dev_priv);
1111 break;
1113 case R300_CMD_CP_DELAY:
1114 /* simple enough, we can do it here */
1115 DRM_DEBUG("R300_CMD_CP_DELAY\n");
1117 int i;
1118 RING_LOCALS;
1120 BEGIN_RING(header.delay.count);
1121 for (i = 0; i < header.delay.count; i++)
1122 OUT_RING(RADEON_CP_PACKET2);
1123 ADVANCE_RING();
1125 break;
1127 case R300_CMD_DMA_DISCARD:
1128 DRM_DEBUG("RADEON_CMD_DMA_DISCARD\n");
1129 idx = header.dma.buf_idx;
1130 if (idx < 0 || idx >= dma->buf_count) {
1131 DRM_ERROR("buffer index %d (of %d max)\n",
1132 idx, dma->buf_count - 1);
1133 ret = -EINVAL;
1134 goto cleanup;
1137 buf = dma->buflist[idx];
1138 if (buf->file_priv != file_priv || buf->pending) {
1139 DRM_ERROR("bad buffer %p %p %d\n",
1140 buf->file_priv, file_priv,
1141 buf->pending);
1142 ret = -EINVAL;
1143 goto cleanup;
1146 emit_dispatch_age = 1;
1147 r300_discard_buffer(dev, file_priv->master, buf);
1148 break;
1150 case R300_CMD_WAIT:
1151 DRM_DEBUG("R300_CMD_WAIT\n");
1152 r300_cmd_wait(dev_priv, header);
1153 break;
1155 case R300_CMD_SCRATCH:
1156 DRM_DEBUG("R300_CMD_SCRATCH\n");
1157 ret = r300_scratch(dev_priv, cmdbuf, header);
1158 if (ret) {
1159 DRM_ERROR("r300_scratch failed\n");
1160 goto cleanup;
1162 break;
1164 case R300_CMD_R500FP:
1165 if ((dev_priv->flags & RADEON_FAMILY_MASK) < CHIP_RV515) {
1166 DRM_ERROR("Calling r500 command on r300 card\n");
1167 ret = -EINVAL;
1168 goto cleanup;
1170 DRM_DEBUG("R300_CMD_R500FP\n");
1171 ret = r300_emit_r500fp(dev_priv, cmdbuf, header);
1172 if (ret) {
1173 DRM_ERROR("r300_emit_r500fp failed\n");
1174 goto cleanup;
1176 break;
1177 default:
1178 DRM_ERROR("bad cmd_type %i at %p\n",
1179 header.header.cmd_type,
1180 cmdbuf->buf - sizeof(header));
1181 ret = -EINVAL;
1182 goto cleanup;
1186 DRM_DEBUG("END\n");
1188 cleanup:
1189 r300_pacify(dev_priv);
1191 /* We emit the vertex buffer age here, outside the pacifier "brackets"
1192 * for two reasons:
1193 * (1) This may coalesce multiple age emissions into a single one and
1194 * (2) more importantly, some chips lock up hard when scratch registers
1195 * are written inside the pacifier bracket.
1197 if (emit_dispatch_age) {
1198 RING_LOCALS;
1200 /* Emit the vertex buffer age */
1201 BEGIN_RING(2);
1202 RADEON_DISPATCH_AGE(master_priv->sarea_priv->last_dispatch);
1203 ADVANCE_RING();
1206 COMMIT_RING();
1208 return ret;