x86: cpa: move clflush_cache_range()
[wrt350n-kernel.git] / drivers / char / drm / r300_cmdbuf.c
blob59b2944811c50c7c9e0269585ba525ce88a49a4e
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 #define R300_SIMULTANEOUS_CLIPRECTS 4
42 /* Values for R300_RE_CLIPRECT_CNTL depending on the number of cliprects
44 static const int r300_cliprect_cntl[4] = {
45 0xAAAA,
46 0xEEEE,
47 0xFEFE,
48 0xFFFE
51 /**
52 * Emit up to R300_SIMULTANEOUS_CLIPRECTS cliprects from the given command
53 * buffer, starting with index n.
55 static int r300_emit_cliprects(drm_radeon_private_t *dev_priv,
56 drm_radeon_kcmd_buffer_t *cmdbuf, int n)
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;
80 box.x1 =
81 (box.x1 +
82 R300_CLIPRECT_OFFSET) & R300_CLIPRECT_MASK;
83 box.y1 =
84 (box.y1 +
85 R300_CLIPRECT_OFFSET) & R300_CLIPRECT_MASK;
86 box.x2 =
87 (box.x2 +
88 R300_CLIPRECT_OFFSET) & R300_CLIPRECT_MASK;
89 box.y2 =
90 (box.y2 +
91 R300_CLIPRECT_OFFSET) & R300_CLIPRECT_MASK;
93 OUT_RING((box.x1 << R300_CLIPRECT_X_SHIFT) |
94 (box.y1 << R300_CLIPRECT_Y_SHIFT));
95 OUT_RING((box.x2 << R300_CLIPRECT_X_SHIFT) |
96 (box.y2 << R300_CLIPRECT_Y_SHIFT));
99 OUT_RING_REG(R300_RE_CLIPRECT_CNTL, r300_cliprect_cntl[nr - 1]);
101 /* TODO/SECURITY: Force scissors to a safe value, otherwise the
102 * client might be able to trample over memory.
103 * The impact should be very limited, but I'd rather be safe than
104 * sorry.
106 OUT_RING(CP_PACKET0(R300_RE_SCISSORS_TL, 1));
107 OUT_RING(0);
108 OUT_RING(R300_SCISSORS_X_MASK | R300_SCISSORS_Y_MASK);
109 ADVANCE_RING();
110 } else {
111 /* Why we allow zero cliprect rendering:
112 * There are some commands in a command buffer that must be submitted
113 * even when there are no cliprects, e.g. DMA buffer discard
114 * or state setting (though state setting could be avoided by
115 * simulating a loss of context).
117 * Now since the cmdbuf interface is so chaotic right now (and is
118 * bound to remain that way for a bit until things settle down),
119 * it is basically impossible to filter out the commands that are
120 * necessary and those that aren't.
122 * So I choose the safe way and don't do any filtering at all;
123 * instead, I simply set up the engine so that all rendering
124 * can't produce any fragments.
126 BEGIN_RING(2);
127 OUT_RING_REG(R300_RE_CLIPRECT_CNTL, 0);
128 ADVANCE_RING();
131 return 0;
134 static u8 r300_reg_flags[0x10000 >> 2];
136 void r300_init_reg_flags(void)
138 int i;
139 memset(r300_reg_flags, 0, 0x10000 >> 2);
140 #define ADD_RANGE_MARK(reg, count,mark) \
141 for(i=((reg)>>2);i<((reg)>>2)+(count);i++)\
142 r300_reg_flags[i]|=(mark);
144 #define MARK_SAFE 1
145 #define MARK_CHECK_OFFSET 2
147 #define ADD_RANGE(reg, count) ADD_RANGE_MARK(reg, count, MARK_SAFE)
149 /* these match cmducs() command in r300_driver/r300/r300_cmdbuf.c */
150 ADD_RANGE(R300_SE_VPORT_XSCALE, 6);
151 ADD_RANGE(R300_VAP_CNTL, 1);
152 ADD_RANGE(R300_SE_VTE_CNTL, 2);
153 ADD_RANGE(0x2134, 2);
154 ADD_RANGE(R300_VAP_CNTL_STATUS, 1);
155 ADD_RANGE(R300_VAP_INPUT_CNTL_0, 2);
156 ADD_RANGE(0x21DC, 1);
157 ADD_RANGE(R300_VAP_UNKNOWN_221C, 1);
158 ADD_RANGE(R300_VAP_CLIP_X_0, 4);
159 ADD_RANGE(R300_VAP_PVS_WAITIDLE, 1);
160 ADD_RANGE(R300_VAP_UNKNOWN_2288, 1);
161 ADD_RANGE(R300_VAP_OUTPUT_VTX_FMT_0, 2);
162 ADD_RANGE(R300_VAP_PVS_CNTL_1, 3);
163 ADD_RANGE(R300_GB_ENABLE, 1);
164 ADD_RANGE(R300_GB_MSPOS0, 5);
165 ADD_RANGE(R300_TX_CNTL, 1);
166 ADD_RANGE(R300_TX_ENABLE, 1);
167 ADD_RANGE(0x4200, 4);
168 ADD_RANGE(0x4214, 1);
169 ADD_RANGE(R300_RE_POINTSIZE, 1);
170 ADD_RANGE(0x4230, 3);
171 ADD_RANGE(R300_RE_LINE_CNT, 1);
172 ADD_RANGE(R300_RE_UNK4238, 1);
173 ADD_RANGE(0x4260, 3);
174 ADD_RANGE(R300_RE_SHADE, 4);
175 ADD_RANGE(R300_RE_POLYGON_MODE, 5);
176 ADD_RANGE(R300_RE_ZBIAS_CNTL, 1);
177 ADD_RANGE(R300_RE_ZBIAS_T_FACTOR, 4);
178 ADD_RANGE(R300_RE_OCCLUSION_CNTL, 1);
179 ADD_RANGE(R300_RE_CULL_CNTL, 1);
180 ADD_RANGE(0x42C0, 2);
181 ADD_RANGE(R300_RS_CNTL_0, 2);
182 ADD_RANGE(R300_RS_INTERP_0, 8);
183 ADD_RANGE(R300_RS_ROUTE_0, 8);
184 ADD_RANGE(0x43A4, 2);
185 ADD_RANGE(0x43E8, 1);
186 ADD_RANGE(R300_PFS_CNTL_0, 3);
187 ADD_RANGE(R300_PFS_NODE_0, 4);
188 ADD_RANGE(R300_PFS_TEXI_0, 64);
189 ADD_RANGE(0x46A4, 5);
190 ADD_RANGE(R300_PFS_INSTR0_0, 64);
191 ADD_RANGE(R300_PFS_INSTR1_0, 64);
192 ADD_RANGE(R300_PFS_INSTR2_0, 64);
193 ADD_RANGE(R300_PFS_INSTR3_0, 64);
194 ADD_RANGE(R300_RE_FOG_STATE, 1);
195 ADD_RANGE(R300_FOG_COLOR_R, 3);
196 ADD_RANGE(R300_PP_ALPHA_TEST, 2);
197 ADD_RANGE(0x4BD8, 1);
198 ADD_RANGE(R300_PFS_PARAM_0_X, 64);
199 ADD_RANGE(0x4E00, 1);
200 ADD_RANGE(R300_RB3D_CBLEND, 2);
201 ADD_RANGE(R300_RB3D_COLORMASK, 1);
202 ADD_RANGE(R300_RB3D_BLEND_COLOR, 3);
203 ADD_RANGE_MARK(R300_RB3D_COLOROFFSET0, 1, MARK_CHECK_OFFSET); /* check offset */
204 ADD_RANGE(R300_RB3D_COLORPITCH0, 1);
205 ADD_RANGE(0x4E50, 9);
206 ADD_RANGE(0x4E88, 1);
207 ADD_RANGE(0x4EA0, 2);
208 ADD_RANGE(R300_RB3D_ZSTENCIL_CNTL_0, 3);
209 ADD_RANGE(R300_RB3D_ZSTENCIL_FORMAT, 4);
210 ADD_RANGE_MARK(R300_RB3D_DEPTHOFFSET, 1, MARK_CHECK_OFFSET); /* check offset */
211 ADD_RANGE(R300_RB3D_DEPTHPITCH, 1);
212 ADD_RANGE(0x4F28, 1);
213 ADD_RANGE(0x4F30, 2);
214 ADD_RANGE(0x4F44, 1);
215 ADD_RANGE(0x4F54, 1);
217 ADD_RANGE(R300_TX_FILTER_0, 16);
218 ADD_RANGE(R300_TX_FILTER1_0, 16);
219 ADD_RANGE(R300_TX_SIZE_0, 16);
220 ADD_RANGE(R300_TX_FORMAT_0, 16);
221 ADD_RANGE(R300_TX_PITCH_0, 16);
222 /* Texture offset is dangerous and needs more checking */
223 ADD_RANGE_MARK(R300_TX_OFFSET_0, 16, MARK_CHECK_OFFSET);
224 ADD_RANGE(R300_TX_CHROMA_KEY_0, 16);
225 ADD_RANGE(R300_TX_BORDER_COLOR_0, 16);
227 /* Sporadic registers used as primitives are emitted */
228 ADD_RANGE(R300_RB3D_ZCACHE_CTLSTAT, 1);
229 ADD_RANGE(R300_RB3D_DSTCACHE_CTLSTAT, 1);
230 ADD_RANGE(R300_VAP_INPUT_ROUTE_0_0, 8);
231 ADD_RANGE(R300_VAP_INPUT_ROUTE_1_0, 8);
235 static __inline__ int r300_check_range(unsigned reg, int count)
237 int i;
238 if (reg & ~0xffff)
239 return -1;
240 for (i = (reg >> 2); i < (reg >> 2) + count; i++)
241 if (r300_reg_flags[i] != MARK_SAFE)
242 return 1;
243 return 0;
246 static __inline__ int r300_emit_carefully_checked_packet0(drm_radeon_private_t *
247 dev_priv,
248 drm_radeon_kcmd_buffer_t
249 * cmdbuf,
250 drm_r300_cmd_header_t
251 header)
253 int reg;
254 int sz;
255 int i;
256 int values[64];
257 RING_LOCALS;
259 sz = header.packet0.count;
260 reg = (header.packet0.reghi << 8) | header.packet0.reglo;
262 if ((sz > 64) || (sz < 0)) {
263 DRM_ERROR
264 ("Cannot emit more than 64 values at a time (reg=%04x sz=%d)\n",
265 reg, sz);
266 return -EINVAL;
268 for (i = 0; i < sz; i++) {
269 values[i] = ((int *)cmdbuf->buf)[i];
270 switch (r300_reg_flags[(reg >> 2) + i]) {
271 case MARK_SAFE:
272 break;
273 case MARK_CHECK_OFFSET:
274 if (!radeon_check_offset(dev_priv, (u32) values[i])) {
275 DRM_ERROR
276 ("Offset failed range check (reg=%04x sz=%d)\n",
277 reg, sz);
278 return -EINVAL;
280 break;
281 default:
282 DRM_ERROR("Register %04x failed check as flag=%02x\n",
283 reg + i * 4, r300_reg_flags[(reg >> 2) + i]);
284 return -EINVAL;
288 BEGIN_RING(1 + sz);
289 OUT_RING(CP_PACKET0(reg, sz - 1));
290 OUT_RING_TABLE(values, sz);
291 ADVANCE_RING();
293 cmdbuf->buf += sz * 4;
294 cmdbuf->bufsz -= sz * 4;
296 return 0;
300 * Emits a packet0 setting arbitrary registers.
301 * Called by r300_do_cp_cmdbuf.
303 * Note that checks are performed on contents and addresses of the registers
305 static __inline__ int r300_emit_packet0(drm_radeon_private_t *dev_priv,
306 drm_radeon_kcmd_buffer_t *cmdbuf,
307 drm_r300_cmd_header_t header)
309 int reg;
310 int sz;
311 RING_LOCALS;
313 sz = header.packet0.count;
314 reg = (header.packet0.reghi << 8) | header.packet0.reglo;
316 if (!sz)
317 return 0;
319 if (sz * 4 > cmdbuf->bufsz)
320 return -EINVAL;
322 if (reg + sz * 4 >= 0x10000) {
323 DRM_ERROR("No such registers in hardware reg=%04x sz=%d\n", reg,
324 sz);
325 return -EINVAL;
328 if (r300_check_range(reg, sz)) {
329 /* go and check everything */
330 return r300_emit_carefully_checked_packet0(dev_priv, cmdbuf,
331 header);
333 /* the rest of the data is safe to emit, whatever the values the user passed */
335 BEGIN_RING(1 + sz);
336 OUT_RING(CP_PACKET0(reg, sz - 1));
337 OUT_RING_TABLE((int *)cmdbuf->buf, sz);
338 ADVANCE_RING();
340 cmdbuf->buf += sz * 4;
341 cmdbuf->bufsz -= sz * 4;
343 return 0;
347 * Uploads user-supplied vertex program instructions or parameters onto
348 * the graphics card.
349 * Called by r300_do_cp_cmdbuf.
351 static __inline__ int r300_emit_vpu(drm_radeon_private_t *dev_priv,
352 drm_radeon_kcmd_buffer_t *cmdbuf,
353 drm_r300_cmd_header_t header)
355 int sz;
356 int addr;
357 RING_LOCALS;
359 sz = header.vpu.count;
360 addr = (header.vpu.adrhi << 8) | header.vpu.adrlo;
362 if (!sz)
363 return 0;
364 if (sz * 16 > cmdbuf->bufsz)
365 return -EINVAL;
367 BEGIN_RING(5 + sz * 4);
368 /* Wait for VAP to come to senses.. */
369 /* there is no need to emit it multiple times, (only once before VAP is programmed,
370 but this optimization is for later */
371 OUT_RING_REG(R300_VAP_PVS_WAITIDLE, 0);
372 OUT_RING_REG(R300_VAP_PVS_UPLOAD_ADDRESS, addr);
373 OUT_RING(CP_PACKET0_TABLE(R300_VAP_PVS_UPLOAD_DATA, sz * 4 - 1));
374 OUT_RING_TABLE((int *)cmdbuf->buf, sz * 4);
376 ADVANCE_RING();
378 cmdbuf->buf += sz * 16;
379 cmdbuf->bufsz -= sz * 16;
381 return 0;
385 * Emit a clear packet from userspace.
386 * Called by r300_emit_packet3.
388 static __inline__ int r300_emit_clear(drm_radeon_private_t *dev_priv,
389 drm_radeon_kcmd_buffer_t *cmdbuf)
391 RING_LOCALS;
393 if (8 * 4 > cmdbuf->bufsz)
394 return -EINVAL;
396 BEGIN_RING(10);
397 OUT_RING(CP_PACKET3(R200_3D_DRAW_IMMD_2, 8));
398 OUT_RING(R300_PRIM_TYPE_POINT | R300_PRIM_WALK_RING |
399 (1 << R300_PRIM_NUM_VERTICES_SHIFT));
400 OUT_RING_TABLE((int *)cmdbuf->buf, 8);
401 ADVANCE_RING();
403 cmdbuf->buf += 8 * 4;
404 cmdbuf->bufsz -= 8 * 4;
406 return 0;
409 static __inline__ int r300_emit_3d_load_vbpntr(drm_radeon_private_t *dev_priv,
410 drm_radeon_kcmd_buffer_t *cmdbuf,
411 u32 header)
413 int count, i, k;
414 #define MAX_ARRAY_PACKET 64
415 u32 payload[MAX_ARRAY_PACKET];
416 u32 narrays;
417 RING_LOCALS;
419 count = (header >> 16) & 0x3fff;
421 if ((count + 1) > MAX_ARRAY_PACKET) {
422 DRM_ERROR("Too large payload in 3D_LOAD_VBPNTR (count=%d)\n",
423 count);
424 return -EINVAL;
426 memset(payload, 0, MAX_ARRAY_PACKET * 4);
427 memcpy(payload, cmdbuf->buf + 4, (count + 1) * 4);
429 /* carefully check packet contents */
431 narrays = payload[0];
432 k = 0;
433 i = 1;
434 while ((k < narrays) && (i < (count + 1))) {
435 i++; /* skip attribute field */
436 if (!radeon_check_offset(dev_priv, payload[i])) {
437 DRM_ERROR
438 ("Offset failed range check (k=%d i=%d) while processing 3D_LOAD_VBPNTR packet.\n",
439 k, i);
440 return -EINVAL;
442 k++;
443 i++;
444 if (k == narrays)
445 break;
446 /* have one more to process, they come in pairs */
447 if (!radeon_check_offset(dev_priv, payload[i])) {
448 DRM_ERROR
449 ("Offset failed range check (k=%d i=%d) while processing 3D_LOAD_VBPNTR packet.\n",
450 k, i);
451 return -EINVAL;
453 k++;
454 i++;
456 /* do the counts match what we expect ? */
457 if ((k != narrays) || (i != (count + 1))) {
458 DRM_ERROR
459 ("Malformed 3D_LOAD_VBPNTR packet (k=%d i=%d narrays=%d count+1=%d).\n",
460 k, i, narrays, count + 1);
461 return -EINVAL;
464 /* all clear, output packet */
466 BEGIN_RING(count + 2);
467 OUT_RING(header);
468 OUT_RING_TABLE(payload, count + 1);
469 ADVANCE_RING();
471 cmdbuf->buf += (count + 2) * 4;
472 cmdbuf->bufsz -= (count + 2) * 4;
474 return 0;
477 static __inline__ int r300_emit_bitblt_multi(drm_radeon_private_t *dev_priv,
478 drm_radeon_kcmd_buffer_t *cmdbuf)
480 u32 *cmd = (u32 *) cmdbuf->buf;
481 int count, ret;
482 RING_LOCALS;
484 count=(cmd[0]>>16) & 0x3fff;
486 if (cmd[0] & 0x8000) {
487 u32 offset;
489 if (cmd[1] & (RADEON_GMC_SRC_PITCH_OFFSET_CNTL
490 | RADEON_GMC_DST_PITCH_OFFSET_CNTL)) {
491 offset = cmd[2] << 10;
492 ret = !radeon_check_offset(dev_priv, offset);
493 if (ret) {
494 DRM_ERROR("Invalid bitblt first offset is %08X\n", offset);
495 return -EINVAL;
499 if ((cmd[1] & RADEON_GMC_SRC_PITCH_OFFSET_CNTL) &&
500 (cmd[1] & RADEON_GMC_DST_PITCH_OFFSET_CNTL)) {
501 offset = cmd[3] << 10;
502 ret = !radeon_check_offset(dev_priv, offset);
503 if (ret) {
504 DRM_ERROR("Invalid bitblt second offset is %08X\n", offset);
505 return -EINVAL;
511 BEGIN_RING(count+2);
512 OUT_RING(cmd[0]);
513 OUT_RING_TABLE((int *)(cmdbuf->buf + 4), count + 1);
514 ADVANCE_RING();
516 cmdbuf->buf += (count+2)*4;
517 cmdbuf->bufsz -= (count+2)*4;
519 return 0;
522 static __inline__ int r300_emit_indx_buffer(drm_radeon_private_t *dev_priv,
523 drm_radeon_kcmd_buffer_t *cmdbuf)
525 u32 *cmd = (u32 *) cmdbuf->buf;
526 int count, ret;
527 RING_LOCALS;
529 count=(cmd[0]>>16) & 0x3fff;
531 if ((cmd[1] & 0x8000ffff) != 0x80000810) {
532 DRM_ERROR("Invalid indx_buffer reg address %08X\n", cmd[1]);
533 return -EINVAL;
535 ret = !radeon_check_offset(dev_priv, cmd[2]);
536 if (ret) {
537 DRM_ERROR("Invalid indx_buffer offset is %08X\n", cmd[2]);
538 return -EINVAL;
541 BEGIN_RING(count+2);
542 OUT_RING(cmd[0]);
543 OUT_RING_TABLE((int *)(cmdbuf->buf + 4), count + 1);
544 ADVANCE_RING();
546 cmdbuf->buf += (count+2)*4;
547 cmdbuf->bufsz -= (count+2)*4;
549 return 0;
552 static __inline__ int r300_emit_raw_packet3(drm_radeon_private_t *dev_priv,
553 drm_radeon_kcmd_buffer_t *cmdbuf)
555 u32 header;
556 int count;
557 RING_LOCALS;
559 if (4 > cmdbuf->bufsz)
560 return -EINVAL;
562 /* Fixme !! This simply emits a packet without much checking.
563 We need to be smarter. */
565 /* obtain first word - actual packet3 header */
566 header = *(u32 *) cmdbuf->buf;
568 /* Is it packet 3 ? */
569 if ((header >> 30) != 0x3) {
570 DRM_ERROR("Not a packet3 header (0x%08x)\n", header);
571 return -EINVAL;
574 count = (header >> 16) & 0x3fff;
576 /* Check again now that we know how much data to expect */
577 if ((count + 2) * 4 > cmdbuf->bufsz) {
578 DRM_ERROR
579 ("Expected packet3 of length %d but have only %d bytes left\n",
580 (count + 2) * 4, cmdbuf->bufsz);
581 return -EINVAL;
584 /* Is it a packet type we know about ? */
585 switch (header & 0xff00) {
586 case RADEON_3D_LOAD_VBPNTR: /* load vertex array pointers */
587 return r300_emit_3d_load_vbpntr(dev_priv, cmdbuf, header);
589 case RADEON_CNTL_BITBLT_MULTI:
590 return r300_emit_bitblt_multi(dev_priv, cmdbuf);
592 case RADEON_CP_INDX_BUFFER: /* DRAW_INDX_2 without INDX_BUFFER seems to lock up the gpu */
593 return r300_emit_indx_buffer(dev_priv, cmdbuf);
594 case RADEON_CP_3D_DRAW_IMMD_2: /* triggers drawing using in-packet vertex data */
595 case RADEON_CP_3D_DRAW_VBUF_2: /* triggers drawing of vertex buffers setup elsewhere */
596 case RADEON_CP_3D_DRAW_INDX_2: /* triggers drawing using indices to vertex buffer */
597 case RADEON_WAIT_FOR_IDLE:
598 case RADEON_CP_NOP:
599 /* these packets are safe */
600 break;
601 default:
602 DRM_ERROR("Unknown packet3 header (0x%08x)\n", header);
603 return -EINVAL;
606 BEGIN_RING(count + 2);
607 OUT_RING(header);
608 OUT_RING_TABLE((int *)(cmdbuf->buf + 4), count + 1);
609 ADVANCE_RING();
611 cmdbuf->buf += (count + 2) * 4;
612 cmdbuf->bufsz -= (count + 2) * 4;
614 return 0;
618 * Emit a rendering packet3 from userspace.
619 * Called by r300_do_cp_cmdbuf.
621 static __inline__ int r300_emit_packet3(drm_radeon_private_t *dev_priv,
622 drm_radeon_kcmd_buffer_t *cmdbuf,
623 drm_r300_cmd_header_t header)
625 int n;
626 int ret;
627 char *orig_buf = cmdbuf->buf;
628 int orig_bufsz = cmdbuf->bufsz;
630 /* This is a do-while-loop so that we run the interior at least once,
631 * even if cmdbuf->nbox is 0. Compare r300_emit_cliprects for rationale.
633 n = 0;
634 do {
635 if (cmdbuf->nbox > R300_SIMULTANEOUS_CLIPRECTS) {
636 ret = r300_emit_cliprects(dev_priv, cmdbuf, n);
637 if (ret)
638 return ret;
640 cmdbuf->buf = orig_buf;
641 cmdbuf->bufsz = orig_bufsz;
644 switch (header.packet3.packet) {
645 case R300_CMD_PACKET3_CLEAR:
646 DRM_DEBUG("R300_CMD_PACKET3_CLEAR\n");
647 ret = r300_emit_clear(dev_priv, cmdbuf);
648 if (ret) {
649 DRM_ERROR("r300_emit_clear failed\n");
650 return ret;
652 break;
654 case R300_CMD_PACKET3_RAW:
655 DRM_DEBUG("R300_CMD_PACKET3_RAW\n");
656 ret = r300_emit_raw_packet3(dev_priv, cmdbuf);
657 if (ret) {
658 DRM_ERROR("r300_emit_raw_packet3 failed\n");
659 return ret;
661 break;
663 default:
664 DRM_ERROR("bad packet3 type %i at %p\n",
665 header.packet3.packet,
666 cmdbuf->buf - sizeof(header));
667 return -EINVAL;
670 n += R300_SIMULTANEOUS_CLIPRECTS;
671 } while (n < cmdbuf->nbox);
673 return 0;
676 /* Some of the R300 chips seem to be extremely touchy about the two registers
677 * that are configured in r300_pacify.
678 * Among the worst offenders seems to be the R300 ND (0x4E44): When userspace
679 * sends a command buffer that contains only state setting commands and a
680 * vertex program/parameter upload sequence, this will eventually lead to a
681 * lockup, unless the sequence is bracketed by calls to r300_pacify.
682 * So we should take great care to *always* call r300_pacify before
683 * *anything* 3D related, and again afterwards. This is what the
684 * call bracket in r300_do_cp_cmdbuf is for.
688 * Emit the sequence to pacify R300.
690 static __inline__ void r300_pacify(drm_radeon_private_t *dev_priv)
692 RING_LOCALS;
694 BEGIN_RING(6);
695 OUT_RING(CP_PACKET0(R300_RB3D_DSTCACHE_CTLSTAT, 0));
696 OUT_RING(R300_RB3D_DSTCACHE_UNKNOWN_0A);
697 OUT_RING(CP_PACKET0(R300_RB3D_ZCACHE_CTLSTAT, 0));
698 OUT_RING(R300_RB3D_ZCACHE_UNKNOWN_03);
699 OUT_RING(CP_PACKET3(RADEON_CP_NOP, 0));
700 OUT_RING(0x0);
701 ADVANCE_RING();
705 * Called by r300_do_cp_cmdbuf to update the internal buffer age and state.
706 * The actual age emit is done by r300_do_cp_cmdbuf, which is why you must
707 * be careful about how this function is called.
709 static void r300_discard_buffer(struct drm_device * dev, struct drm_buf * buf)
711 drm_radeon_private_t *dev_priv = dev->dev_private;
712 drm_radeon_buf_priv_t *buf_priv = buf->dev_private;
714 buf_priv->age = ++dev_priv->sarea_priv->last_dispatch;
715 buf->pending = 1;
716 buf->used = 0;
719 static int r300_scratch(drm_radeon_private_t *dev_priv,
720 drm_radeon_kcmd_buffer_t *cmdbuf,
721 drm_r300_cmd_header_t header)
723 u32 *ref_age_base;
724 u32 i, buf_idx, h_pending;
725 RING_LOCALS;
727 if (cmdbuf->bufsz <
728 (sizeof(u64) + header.scratch.n_bufs * sizeof(buf_idx))) {
729 return -EINVAL;
732 if (header.scratch.reg >= 5) {
733 return -EINVAL;
736 dev_priv->scratch_ages[header.scratch.reg]++;
738 ref_age_base = (u32 *)(unsigned long)*((uint64_t *)cmdbuf->buf);
740 cmdbuf->buf += sizeof(u64);
741 cmdbuf->bufsz -= sizeof(u64);
743 for (i=0; i < header.scratch.n_bufs; i++) {
744 buf_idx = *(u32 *)cmdbuf->buf;
745 buf_idx *= 2; /* 8 bytes per buf */
747 if (DRM_COPY_TO_USER(ref_age_base + buf_idx, &dev_priv->scratch_ages[header.scratch.reg], sizeof(u32))) {
748 return -EINVAL;
751 if (DRM_COPY_FROM_USER(&h_pending, ref_age_base + buf_idx + 1, sizeof(u32))) {
752 return -EINVAL;
755 if (h_pending == 0) {
756 return -EINVAL;
759 h_pending--;
761 if (DRM_COPY_TO_USER(ref_age_base + buf_idx + 1, &h_pending, sizeof(u32))) {
762 return -EINVAL;
765 cmdbuf->buf += sizeof(buf_idx);
766 cmdbuf->bufsz -= sizeof(buf_idx);
769 BEGIN_RING(2);
770 OUT_RING( CP_PACKET0( RADEON_SCRATCH_REG0 + header.scratch.reg * 4, 0 ) );
771 OUT_RING( dev_priv->scratch_ages[header.scratch.reg] );
772 ADVANCE_RING();
774 return 0;
778 * Parses and validates a user-supplied command buffer and emits appropriate
779 * commands on the DMA ring buffer.
780 * Called by the ioctl handler function radeon_cp_cmdbuf.
782 int r300_do_cp_cmdbuf(struct drm_device *dev,
783 struct drm_file *file_priv,
784 drm_radeon_kcmd_buffer_t *cmdbuf)
786 drm_radeon_private_t *dev_priv = dev->dev_private;
787 struct drm_device_dma *dma = dev->dma;
788 struct drm_buf *buf = NULL;
789 int emit_dispatch_age = 0;
790 int ret = 0;
792 DRM_DEBUG("\n");
794 /* See the comment above r300_emit_begin3d for why this call must be here,
795 * and what the cleanup gotos are for. */
796 r300_pacify(dev_priv);
798 if (cmdbuf->nbox <= R300_SIMULTANEOUS_CLIPRECTS) {
799 ret = r300_emit_cliprects(dev_priv, cmdbuf, 0);
800 if (ret)
801 goto cleanup;
804 while (cmdbuf->bufsz >= sizeof(drm_r300_cmd_header_t)) {
805 int idx;
806 drm_r300_cmd_header_t header;
808 header.u = *(unsigned int *)cmdbuf->buf;
810 cmdbuf->buf += sizeof(header);
811 cmdbuf->bufsz -= sizeof(header);
813 switch (header.header.cmd_type) {
814 case R300_CMD_PACKET0:
815 DRM_DEBUG("R300_CMD_PACKET0\n");
816 ret = r300_emit_packet0(dev_priv, cmdbuf, header);
817 if (ret) {
818 DRM_ERROR("r300_emit_packet0 failed\n");
819 goto cleanup;
821 break;
823 case R300_CMD_VPU:
824 DRM_DEBUG("R300_CMD_VPU\n");
825 ret = r300_emit_vpu(dev_priv, cmdbuf, header);
826 if (ret) {
827 DRM_ERROR("r300_emit_vpu failed\n");
828 goto cleanup;
830 break;
832 case R300_CMD_PACKET3:
833 DRM_DEBUG("R300_CMD_PACKET3\n");
834 ret = r300_emit_packet3(dev_priv, cmdbuf, header);
835 if (ret) {
836 DRM_ERROR("r300_emit_packet3 failed\n");
837 goto cleanup;
839 break;
841 case R300_CMD_END3D:
842 DRM_DEBUG("R300_CMD_END3D\n");
843 /* TODO:
844 Ideally userspace driver should not need to issue this call,
845 i.e. the drm driver should issue it automatically and prevent
846 lockups.
848 In practice, we do not understand why this call is needed and what
849 it does (except for some vague guesses that it has to do with cache
850 coherence) and so the user space driver does it.
852 Once we are sure which uses prevent lockups the code could be moved
853 into the kernel and the userspace driver will not
854 need to use this command.
856 Note that issuing this command does not hurt anything
857 except, possibly, performance */
858 r300_pacify(dev_priv);
859 break;
861 case R300_CMD_CP_DELAY:
862 /* simple enough, we can do it here */
863 DRM_DEBUG("R300_CMD_CP_DELAY\n");
865 int i;
866 RING_LOCALS;
868 BEGIN_RING(header.delay.count);
869 for (i = 0; i < header.delay.count; i++)
870 OUT_RING(RADEON_CP_PACKET2);
871 ADVANCE_RING();
873 break;
875 case R300_CMD_DMA_DISCARD:
876 DRM_DEBUG("RADEON_CMD_DMA_DISCARD\n");
877 idx = header.dma.buf_idx;
878 if (idx < 0 || idx >= dma->buf_count) {
879 DRM_ERROR("buffer index %d (of %d max)\n",
880 idx, dma->buf_count - 1);
881 ret = -EINVAL;
882 goto cleanup;
885 buf = dma->buflist[idx];
886 if (buf->file_priv != file_priv || buf->pending) {
887 DRM_ERROR("bad buffer %p %p %d\n",
888 buf->file_priv, file_priv,
889 buf->pending);
890 ret = -EINVAL;
891 goto cleanup;
894 emit_dispatch_age = 1;
895 r300_discard_buffer(dev, buf);
896 break;
898 case R300_CMD_WAIT:
899 /* simple enough, we can do it here */
900 DRM_DEBUG("R300_CMD_WAIT\n");
901 if (header.wait.flags == 0)
902 break; /* nothing to do */
905 RING_LOCALS;
907 BEGIN_RING(2);
908 OUT_RING(CP_PACKET0(RADEON_WAIT_UNTIL, 0));
909 OUT_RING((header.wait.flags & 0xf) << 14);
910 ADVANCE_RING();
912 break;
914 case R300_CMD_SCRATCH:
915 DRM_DEBUG("R300_CMD_SCRATCH\n");
916 ret = r300_scratch(dev_priv, cmdbuf, header);
917 if (ret) {
918 DRM_ERROR("r300_scratch failed\n");
919 goto cleanup;
921 break;
923 default:
924 DRM_ERROR("bad cmd_type %i at %p\n",
925 header.header.cmd_type,
926 cmdbuf->buf - sizeof(header));
927 ret = -EINVAL;
928 goto cleanup;
932 DRM_DEBUG("END\n");
934 cleanup:
935 r300_pacify(dev_priv);
937 /* We emit the vertex buffer age here, outside the pacifier "brackets"
938 * for two reasons:
939 * (1) This may coalesce multiple age emissions into a single one and
940 * (2) more importantly, some chips lock up hard when scratch registers
941 * are written inside the pacifier bracket.
943 if (emit_dispatch_age) {
944 RING_LOCALS;
946 /* Emit the vertex buffer age */
947 BEGIN_RING(2);
948 RADEON_DISPATCH_AGE(dev_priv->sarea_priv->last_dispatch);
949 ADVANCE_RING();
952 COMMIT_RING();
954 return ret;