Merge branch 'for_paulus' of git://git.kernel.org/pub/scm/linux/kernel/git/galak...
[linux-2.6] / drivers / char / drm / r300_cmdbuf.c
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  */
33
34 #include "drmP.h"
35 #include "drm.h"
36 #include "radeon_drm.h"
37 #include "radeon_drv.h"
38 #include "r300_reg.h"
39
40 #define R300_SIMULTANEOUS_CLIPRECTS             4
41
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 };
50
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         drm_clip_rect_t box;
59         int nr;
60         int i;
61         RING_LOCALS;
62
63         nr = cmdbuf->nbox - n;
64         if (nr > R300_SIMULTANEOUS_CLIPRECTS)
65                 nr = R300_SIMULTANEOUS_CLIPRECTS;
66
67         DRM_DEBUG("%i cliprects\n", nr);
68
69         if (nr) {
70                 BEGIN_RING(6 + nr * 2);
71                 OUT_RING(CP_PACKET0(R300_RE_CLIPRECT_TL_0, nr * 2 - 1));
72
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 DRM_ERR(EFAULT);
78                         }
79
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;
92
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));
97                 }
98
99                 OUT_RING_REG(R300_RE_CLIPRECT_CNTL, r300_cliprect_cntl[nr - 1]);
100
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.
105                  */
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).
116                  *
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.
121                  *
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.
125                  */
126                 BEGIN_RING(2);
127                 OUT_RING_REG(R300_RE_CLIPRECT_CNTL, 0);
128                 ADVANCE_RING();
129         }
130
131         return 0;
132 }
133
134 static u8 r300_reg_flags[0x10000 >> 2];
135
136 void r300_init_reg_flags(void)
137 {
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);
143
144 #define MARK_SAFE               1
145 #define MARK_CHECK_OFFSET       2
146
147 #define ADD_RANGE(reg, count)   ADD_RANGE_MARK(reg, count, MARK_SAFE)
148
149         /* these match cmducs() command in r300_driver/r300/r300_cmdbuf.c */
150         ADD_RANGE(R300_SE_VPORT_XSCALE, 6);
151         ADD_RANGE(0x2080, 1);
152         ADD_RANGE(R300_SE_VTE_CNTL, 2);
153         ADD_RANGE(0x2134, 2);
154         ADD_RANGE(0x2140, 1);
155         ADD_RANGE(R300_VAP_INPUT_CNTL_0, 2);
156         ADD_RANGE(0x21DC, 1);
157         ADD_RANGE(0x221C, 1);
158         ADD_RANGE(0x2220, 4);
159         ADD_RANGE(0x2288, 1);
160         ADD_RANGE(R300_VAP_OUTPUT_VTX_FMT_0, 2);
161         ADD_RANGE(R300_VAP_PVS_CNTL_1, 3);
162         ADD_RANGE(R300_GB_ENABLE, 1);
163         ADD_RANGE(R300_GB_MSPOS0, 5);
164         ADD_RANGE(R300_TX_CNTL, 1);
165         ADD_RANGE(R300_TX_ENABLE, 1);
166         ADD_RANGE(0x4200, 4);
167         ADD_RANGE(0x4214, 1);
168         ADD_RANGE(R300_RE_POINTSIZE, 1);
169         ADD_RANGE(0x4230, 3);
170         ADD_RANGE(R300_RE_LINE_CNT, 1);
171         ADD_RANGE(0x4238, 1);
172         ADD_RANGE(0x4260, 3);
173         ADD_RANGE(0x4274, 4);
174         ADD_RANGE(0x4288, 5);
175         ADD_RANGE(0x42A0, 1);
176         ADD_RANGE(R300_RE_ZBIAS_T_FACTOR, 4);
177         ADD_RANGE(0x42B4, 1);
178         ADD_RANGE(R300_RE_CULL_CNTL, 1);
179         ADD_RANGE(0x42C0, 2);
180         ADD_RANGE(R300_RS_CNTL_0, 2);
181         ADD_RANGE(R300_RS_INTERP_0, 8);
182         ADD_RANGE(R300_RS_ROUTE_0, 8);
183         ADD_RANGE(0x43A4, 2);
184         ADD_RANGE(0x43E8, 1);
185         ADD_RANGE(R300_PFS_CNTL_0, 3);
186         ADD_RANGE(R300_PFS_NODE_0, 4);
187         ADD_RANGE(R300_PFS_TEXI_0, 64);
188         ADD_RANGE(0x46A4, 5);
189         ADD_RANGE(R300_PFS_INSTR0_0, 64);
190         ADD_RANGE(R300_PFS_INSTR1_0, 64);
191         ADD_RANGE(R300_PFS_INSTR2_0, 64);
192         ADD_RANGE(R300_PFS_INSTR3_0, 64);
193         ADD_RANGE(0x4BC0, 1);
194         ADD_RANGE(0x4BC8, 3);
195         ADD_RANGE(R300_PP_ALPHA_TEST, 2);
196         ADD_RANGE(0x4BD8, 1);
197         ADD_RANGE(R300_PFS_PARAM_0_X, 64);
198         ADD_RANGE(0x4E00, 1);
199         ADD_RANGE(R300_RB3D_CBLEND, 2);
200         ADD_RANGE(R300_RB3D_COLORMASK, 1);
201         ADD_RANGE(0x4E10, 3);
202         ADD_RANGE_MARK(R300_RB3D_COLOROFFSET0, 1, MARK_CHECK_OFFSET);   /* check offset */
203         ADD_RANGE(R300_RB3D_COLORPITCH0, 1);
204         ADD_RANGE(0x4E50, 9);
205         ADD_RANGE(0x4E88, 1);
206         ADD_RANGE(0x4EA0, 2);
207         ADD_RANGE(R300_RB3D_ZSTENCIL_CNTL_0, 3);
208         ADD_RANGE(0x4F10, 4);
209         ADD_RANGE_MARK(R300_RB3D_DEPTHOFFSET, 1, MARK_CHECK_OFFSET);    /* check offset */
210         ADD_RANGE(R300_RB3D_DEPTHPITCH, 1);
211         ADD_RANGE(0x4F28, 1);
212         ADD_RANGE(0x4F30, 2);
213         ADD_RANGE(0x4F44, 1);
214         ADD_RANGE(0x4F54, 1);
215
216         ADD_RANGE(R300_TX_FILTER_0, 16);
217         ADD_RANGE(R300_TX_FILTER1_0, 16);
218         ADD_RANGE(R300_TX_SIZE_0, 16);
219         ADD_RANGE(R300_TX_FORMAT_0, 16);
220         ADD_RANGE(R300_TX_PITCH_0, 16);
221         /* Texture offset is dangerous and needs more checking */
222         ADD_RANGE_MARK(R300_TX_OFFSET_0, 16, MARK_CHECK_OFFSET);
223         ADD_RANGE(R300_TX_CHROMA_KEY_0, 16);
224         ADD_RANGE(R300_TX_BORDER_COLOR_0, 16);
225
226         /* Sporadic registers used as primitives are emitted */
227         ADD_RANGE(0x4f18, 1);
228         ADD_RANGE(R300_RB3D_DSTCACHE_CTLSTAT, 1);
229         ADD_RANGE(R300_VAP_INPUT_ROUTE_0_0, 8);
230         ADD_RANGE(R300_VAP_INPUT_ROUTE_1_0, 8);
231
232 }
233
234 static __inline__ int r300_check_range(unsigned reg, int count)
235 {
236         int i;
237         if (reg & ~0xffff)
238                 return -1;
239         for (i = (reg >> 2); i < (reg >> 2) + count; i++)
240                 if (r300_reg_flags[i] != MARK_SAFE)
241                         return 1;
242         return 0;
243 }
244
245 /*
246  * we expect offsets passed to the framebuffer to be either within video 
247  * memory or within AGP space 
248  */
249 static __inline__ int r300_check_offset(drm_radeon_private_t *dev_priv,
250                                         u32 offset)
251 {
252         /* we realy want to check against end of video aperture
253            but this value is not being kept.
254            This code is correct for now (does the same thing as the
255            code that sets MC_FB_LOCATION) in radeon_cp.c */
256         if (offset >= dev_priv->fb_location &&
257             offset < (dev_priv->fb_location + dev_priv->fb_size))
258                 return 0;
259         if (offset >= dev_priv->gart_vm_start &&
260             offset < (dev_priv->gart_vm_start + dev_priv->gart_size))
261                 return 0;
262         return 1;
263 }
264
265 static __inline__ int r300_emit_carefully_checked_packet0(drm_radeon_private_t *
266                                                           dev_priv,
267                                                           drm_radeon_kcmd_buffer_t
268                                                           * cmdbuf,
269                                                           drm_r300_cmd_header_t
270                                                           header)
271 {
272         int reg;
273         int sz;
274         int i;
275         int values[64];
276         RING_LOCALS;
277
278         sz = header.packet0.count;
279         reg = (header.packet0.reghi << 8) | header.packet0.reglo;
280
281         if ((sz > 64) || (sz < 0)) {
282                 DRM_ERROR
283                     ("Cannot emit more than 64 values at a time (reg=%04x sz=%d)\n",
284                      reg, sz);
285                 return DRM_ERR(EINVAL);
286         }
287         for (i = 0; i < sz; i++) {
288                 values[i] = ((int *)cmdbuf->buf)[i];
289                 switch (r300_reg_flags[(reg >> 2) + i]) {
290                 case MARK_SAFE:
291                         break;
292                 case MARK_CHECK_OFFSET:
293                         if (r300_check_offset(dev_priv, (u32) values[i])) {
294                                 DRM_ERROR
295                                     ("Offset failed range check (reg=%04x sz=%d)\n",
296                                      reg, sz);
297                                 return DRM_ERR(EINVAL);
298                         }
299                         break;
300                 default:
301                         DRM_ERROR("Register %04x failed check as flag=%02x\n",
302                                   reg + i * 4, r300_reg_flags[(reg >> 2) + i]);
303                         return DRM_ERR(EINVAL);
304                 }
305         }
306
307         BEGIN_RING(1 + sz);
308         OUT_RING(CP_PACKET0(reg, sz - 1));
309         OUT_RING_TABLE(values, sz);
310         ADVANCE_RING();
311
312         cmdbuf->buf += sz * 4;
313         cmdbuf->bufsz -= sz * 4;
314
315         return 0;
316 }
317
318 /**
319  * Emits a packet0 setting arbitrary registers.
320  * Called by r300_do_cp_cmdbuf.
321  *
322  * Note that checks are performed on contents and addresses of the registers
323  */
324 static __inline__ int r300_emit_packet0(drm_radeon_private_t *dev_priv,
325                                         drm_radeon_kcmd_buffer_t *cmdbuf,
326                                         drm_r300_cmd_header_t header)
327 {
328         int reg;
329         int sz;
330         RING_LOCALS;
331
332         sz = header.packet0.count;
333         reg = (header.packet0.reghi << 8) | header.packet0.reglo;
334
335         if (!sz)
336                 return 0;
337
338         if (sz * 4 > cmdbuf->bufsz)
339                 return DRM_ERR(EINVAL);
340
341         if (reg + sz * 4 >= 0x10000) {
342                 DRM_ERROR("No such registers in hardware reg=%04x sz=%d\n", reg,
343                           sz);
344                 return DRM_ERR(EINVAL);
345         }
346
347         if (r300_check_range(reg, sz)) {
348                 /* go and check everything */
349                 return r300_emit_carefully_checked_packet0(dev_priv, cmdbuf,
350                                                            header);
351         }
352         /* the rest of the data is safe to emit, whatever the values the user passed */
353
354         BEGIN_RING(1 + sz);
355         OUT_RING(CP_PACKET0(reg, sz - 1));
356         OUT_RING_TABLE((int *)cmdbuf->buf, sz);
357         ADVANCE_RING();
358
359         cmdbuf->buf += sz * 4;
360         cmdbuf->bufsz -= sz * 4;
361
362         return 0;
363 }
364
365 /**
366  * Uploads user-supplied vertex program instructions or parameters onto
367  * the graphics card.
368  * Called by r300_do_cp_cmdbuf.
369  */
370 static __inline__ int r300_emit_vpu(drm_radeon_private_t *dev_priv,
371                                     drm_radeon_kcmd_buffer_t *cmdbuf,
372                                     drm_r300_cmd_header_t header)
373 {
374         int sz;
375         int addr;
376         RING_LOCALS;
377
378         sz = header.vpu.count;
379         addr = (header.vpu.adrhi << 8) | header.vpu.adrlo;
380
381         if (!sz)
382                 return 0;
383         if (sz * 16 > cmdbuf->bufsz)
384                 return DRM_ERR(EINVAL);
385
386         BEGIN_RING(5 + sz * 4);
387         /* Wait for VAP to come to senses.. */
388         /* there is no need to emit it multiple times, (only once before VAP is programmed,
389            but this optimization is for later */
390         OUT_RING_REG(R300_VAP_PVS_WAITIDLE, 0);
391         OUT_RING_REG(R300_VAP_PVS_UPLOAD_ADDRESS, addr);
392         OUT_RING(CP_PACKET0_TABLE(R300_VAP_PVS_UPLOAD_DATA, sz * 4 - 1));
393         OUT_RING_TABLE((int *)cmdbuf->buf, sz * 4);
394
395         ADVANCE_RING();
396
397         cmdbuf->buf += sz * 16;
398         cmdbuf->bufsz -= sz * 16;
399
400         return 0;
401 }
402
403 /**
404  * Emit a clear packet from userspace.
405  * Called by r300_emit_packet3.
406  */
407 static __inline__ int r300_emit_clear(drm_radeon_private_t *dev_priv,
408                                       drm_radeon_kcmd_buffer_t *cmdbuf)
409 {
410         RING_LOCALS;
411
412         if (8 * 4 > cmdbuf->bufsz)
413                 return DRM_ERR(EINVAL);
414
415         BEGIN_RING(10);
416         OUT_RING(CP_PACKET3(R200_3D_DRAW_IMMD_2, 8));
417         OUT_RING(R300_PRIM_TYPE_POINT | R300_PRIM_WALK_RING |
418                  (1 << R300_PRIM_NUM_VERTICES_SHIFT));
419         OUT_RING_TABLE((int *)cmdbuf->buf, 8);
420         ADVANCE_RING();
421
422         cmdbuf->buf += 8 * 4;
423         cmdbuf->bufsz -= 8 * 4;
424
425         return 0;
426 }
427
428 static __inline__ int r300_emit_3d_load_vbpntr(drm_radeon_private_t *dev_priv,
429                                                drm_radeon_kcmd_buffer_t *cmdbuf,
430                                                u32 header)
431 {
432         int count, i, k;
433 #define MAX_ARRAY_PACKET  64
434         u32 payload[MAX_ARRAY_PACKET];
435         u32 narrays;
436         RING_LOCALS;
437
438         count = (header >> 16) & 0x3fff;
439
440         if ((count + 1) > MAX_ARRAY_PACKET) {
441                 DRM_ERROR("Too large payload in 3D_LOAD_VBPNTR (count=%d)\n",
442                           count);
443                 return DRM_ERR(EINVAL);
444         }
445         memset(payload, 0, MAX_ARRAY_PACKET * 4);
446         memcpy(payload, cmdbuf->buf + 4, (count + 1) * 4);
447
448         /* carefully check packet contents */
449
450         narrays = payload[0];
451         k = 0;
452         i = 1;
453         while ((k < narrays) && (i < (count + 1))) {
454                 i++;            /* skip attribute field */
455                 if (r300_check_offset(dev_priv, payload[i])) {
456                         DRM_ERROR
457                             ("Offset failed range check (k=%d i=%d) while processing 3D_LOAD_VBPNTR packet.\n",
458                              k, i);
459                         return DRM_ERR(EINVAL);
460                 }
461                 k++;
462                 i++;
463                 if (k == narrays)
464                         break;
465                 /* have one more to process, they come in pairs */
466                 if (r300_check_offset(dev_priv, payload[i])) {
467                         DRM_ERROR
468                             ("Offset failed range check (k=%d i=%d) while processing 3D_LOAD_VBPNTR packet.\n",
469                              k, i);
470                         return DRM_ERR(EINVAL);
471                 }
472                 k++;
473                 i++;
474         }
475         /* do the counts match what we expect ? */
476         if ((k != narrays) || (i != (count + 1))) {
477                 DRM_ERROR
478                     ("Malformed 3D_LOAD_VBPNTR packet (k=%d i=%d narrays=%d count+1=%d).\n",
479                      k, i, narrays, count + 1);
480                 return DRM_ERR(EINVAL);
481         }
482
483         /* all clear, output packet */
484
485         BEGIN_RING(count + 2);
486         OUT_RING(header);
487         OUT_RING_TABLE(payload, count + 1);
488         ADVANCE_RING();
489
490         cmdbuf->buf += (count + 2) * 4;
491         cmdbuf->bufsz -= (count + 2) * 4;
492
493         return 0;
494 }
495
496 static __inline__ int r300_emit_bitblt_multi(drm_radeon_private_t *dev_priv,
497                                              drm_radeon_kcmd_buffer_t *cmdbuf)
498 {
499         u32 *cmd = (u32 *) cmdbuf->buf;
500         int count, ret;
501         RING_LOCALS;
502
503         count=(cmd[0]>>16) & 0x3fff;
504
505         if (cmd[0] & 0x8000) {
506                 u32 offset;
507
508                 if (cmd[1] & (RADEON_GMC_SRC_PITCH_OFFSET_CNTL 
509                               | RADEON_GMC_DST_PITCH_OFFSET_CNTL)) {
510                         offset = cmd[2] << 10;
511                         ret = r300_check_offset(dev_priv, offset);
512                         if (ret) {
513                                 DRM_ERROR("Invalid bitblt first offset is %08X\n", offset);
514                                 return DRM_ERR(EINVAL);
515                         }
516                 }
517
518                 if ((cmd[1] & RADEON_GMC_SRC_PITCH_OFFSET_CNTL) &&
519                     (cmd[1] & RADEON_GMC_DST_PITCH_OFFSET_CNTL)) {
520                         offset = cmd[3] << 10;
521                         ret = r300_check_offset(dev_priv, offset);
522                         if (ret) {
523                                 DRM_ERROR("Invalid bitblt second offset is %08X\n", offset);
524                                 return DRM_ERR(EINVAL);
525                         }
526                         
527                 }
528         }
529
530         BEGIN_RING(count+2);
531         OUT_RING(cmd[0]);
532         OUT_RING_TABLE((int *)(cmdbuf->buf + 4), count + 1);
533         ADVANCE_RING();
534
535         cmdbuf->buf += (count+2)*4;
536         cmdbuf->bufsz -= (count+2)*4;
537
538         return 0;
539 }
540
541 static __inline__ int r300_emit_raw_packet3(drm_radeon_private_t *dev_priv,
542                                             drm_radeon_kcmd_buffer_t *cmdbuf)
543 {
544         u32 header;
545         int count;
546         RING_LOCALS;
547
548         if (4 > cmdbuf->bufsz)
549                 return DRM_ERR(EINVAL);
550
551         /* Fixme !! This simply emits a packet without much checking.
552            We need to be smarter. */
553
554         /* obtain first word - actual packet3 header */
555         header = *(u32 *) cmdbuf->buf;
556
557         /* Is it packet 3 ? */
558         if ((header >> 30) != 0x3) {
559                 DRM_ERROR("Not a packet3 header (0x%08x)\n", header);
560                 return DRM_ERR(EINVAL);
561         }
562
563         count = (header >> 16) & 0x3fff;
564
565         /* Check again now that we know how much data to expect */
566         if ((count + 2) * 4 > cmdbuf->bufsz) {
567                 DRM_ERROR
568                     ("Expected packet3 of length %d but have only %d bytes left\n",
569                      (count + 2) * 4, cmdbuf->bufsz);
570                 return DRM_ERR(EINVAL);
571         }
572
573         /* Is it a packet type we know about ? */
574         switch (header & 0xff00) {
575         case RADEON_3D_LOAD_VBPNTR:     /* load vertex array pointers */
576                 return r300_emit_3d_load_vbpntr(dev_priv, cmdbuf, header);
577
578         case RADEON_CNTL_BITBLT_MULTI:
579                 return r300_emit_bitblt_multi(dev_priv, cmdbuf);
580
581         case RADEON_CP_3D_DRAW_IMMD_2:  /* triggers drawing using in-packet vertex data */
582         case RADEON_CP_3D_DRAW_VBUF_2:  /* triggers drawing of vertex buffers setup elsewhere */
583         case RADEON_CP_3D_DRAW_INDX_2:  /* triggers drawing using indices to vertex buffer */
584         case RADEON_CP_INDX_BUFFER:     /* DRAW_INDX_2 without INDX_BUFFER seems to lock up the gpu */
585         case RADEON_WAIT_FOR_IDLE:
586         case RADEON_CP_NOP:
587                 /* these packets are safe */
588                 break;
589         default:
590                 DRM_ERROR("Unknown packet3 header (0x%08x)\n", header);
591                 return DRM_ERR(EINVAL);
592         }
593
594         BEGIN_RING(count + 2);
595         OUT_RING(header);
596         OUT_RING_TABLE((int *)(cmdbuf->buf + 4), count + 1);
597         ADVANCE_RING();
598
599         cmdbuf->buf += (count + 2) * 4;
600         cmdbuf->bufsz -= (count + 2) * 4;
601
602         return 0;
603 }
604
605 /**
606  * Emit a rendering packet3 from userspace.
607  * Called by r300_do_cp_cmdbuf.
608  */
609 static __inline__ int r300_emit_packet3(drm_radeon_private_t *dev_priv,
610                                         drm_radeon_kcmd_buffer_t *cmdbuf,
611                                         drm_r300_cmd_header_t header)
612 {
613         int n;
614         int ret;
615         char *orig_buf = cmdbuf->buf;
616         int orig_bufsz = cmdbuf->bufsz;
617
618         /* This is a do-while-loop so that we run the interior at least once,
619          * even if cmdbuf->nbox is 0. Compare r300_emit_cliprects for rationale.
620          */
621         n = 0;
622         do {
623                 if (cmdbuf->nbox > R300_SIMULTANEOUS_CLIPRECTS) {
624                         ret = r300_emit_cliprects(dev_priv, cmdbuf, n);
625                         if (ret)
626                                 return ret;
627
628                         cmdbuf->buf = orig_buf;
629                         cmdbuf->bufsz = orig_bufsz;
630                 }
631
632                 switch (header.packet3.packet) {
633                 case R300_CMD_PACKET3_CLEAR:
634                         DRM_DEBUG("R300_CMD_PACKET3_CLEAR\n");
635                         ret = r300_emit_clear(dev_priv, cmdbuf);
636                         if (ret) {
637                                 DRM_ERROR("r300_emit_clear failed\n");
638                                 return ret;
639                         }
640                         break;
641
642                 case R300_CMD_PACKET3_RAW:
643                         DRM_DEBUG("R300_CMD_PACKET3_RAW\n");
644                         ret = r300_emit_raw_packet3(dev_priv, cmdbuf);
645                         if (ret) {
646                                 DRM_ERROR("r300_emit_raw_packet3 failed\n");
647                                 return ret;
648                         }
649                         break;
650
651                 default:
652                         DRM_ERROR("bad packet3 type %i at %p\n",
653                                   header.packet3.packet,
654                                   cmdbuf->buf - sizeof(header));
655                         return DRM_ERR(EINVAL);
656                 }
657
658                 n += R300_SIMULTANEOUS_CLIPRECTS;
659         } while (n < cmdbuf->nbox);
660
661         return 0;
662 }
663
664 /* Some of the R300 chips seem to be extremely touchy about the two registers
665  * that are configured in r300_pacify.
666  * Among the worst offenders seems to be the R300 ND (0x4E44): When userspace
667  * sends a command buffer that contains only state setting commands and a
668  * vertex program/parameter upload sequence, this will eventually lead to a
669  * lockup, unless the sequence is bracketed by calls to r300_pacify.
670  * So we should take great care to *always* call r300_pacify before
671  * *anything* 3D related, and again afterwards. This is what the
672  * call bracket in r300_do_cp_cmdbuf is for.
673  */
674
675 /**
676  * Emit the sequence to pacify R300.
677  */
678 static __inline__ void r300_pacify(drm_radeon_private_t *dev_priv)
679 {
680         RING_LOCALS;
681
682         BEGIN_RING(6);
683         OUT_RING(CP_PACKET0(R300_RB3D_DSTCACHE_CTLSTAT, 0));
684         OUT_RING(0xa);
685         OUT_RING(CP_PACKET0(0x4f18, 0));
686         OUT_RING(0x3);
687         OUT_RING(CP_PACKET3(RADEON_CP_NOP, 0));
688         OUT_RING(0x0);
689         ADVANCE_RING();
690 }
691
692 /**
693  * Called by r300_do_cp_cmdbuf to update the internal buffer age and state.
694  * The actual age emit is done by r300_do_cp_cmdbuf, which is why you must
695  * be careful about how this function is called.
696  */
697 static void r300_discard_buffer(drm_device_t * dev, drm_buf_t * buf)
698 {
699         drm_radeon_private_t *dev_priv = dev->dev_private;
700         drm_radeon_buf_priv_t *buf_priv = buf->dev_private;
701
702         buf_priv->age = ++dev_priv->sarea_priv->last_dispatch;
703         buf->pending = 1;
704         buf->used = 0;
705 }
706
707 static int r300_scratch(drm_radeon_private_t *dev_priv,
708                         drm_radeon_kcmd_buffer_t *cmdbuf,
709                         drm_r300_cmd_header_t header)
710 {
711         u32 *ref_age_base;
712         u32 i, buf_idx, h_pending;
713         RING_LOCALS;
714         
715         if (cmdbuf->bufsz < 
716             (sizeof(u64) + header.scratch.n_bufs * sizeof(buf_idx))) {
717                 return DRM_ERR(EINVAL);
718         }
719         
720         if (header.scratch.reg >= 5) {
721                 return DRM_ERR(EINVAL);
722         }
723         
724         dev_priv->scratch_ages[header.scratch.reg]++;
725         
726         ref_age_base = *(u32 **)cmdbuf->buf;
727         
728         cmdbuf->buf += sizeof(u64);
729         cmdbuf->bufsz -= sizeof(u64);
730         
731         for (i=0; i < header.scratch.n_bufs; i++) {
732                 buf_idx = *(u32 *)cmdbuf->buf;
733                 buf_idx *= 2; /* 8 bytes per buf */
734                 
735                 if (DRM_COPY_TO_USER(ref_age_base + buf_idx, &dev_priv->scratch_ages[header.scratch.reg], sizeof(u32))) {
736                         return DRM_ERR(EINVAL);
737                 }
738                                         
739                 if (DRM_COPY_FROM_USER(&h_pending, ref_age_base + buf_idx + 1, sizeof(u32))) {
740                         return DRM_ERR(EINVAL);
741                 }
742                                         
743                 if (h_pending == 0) {
744                         return DRM_ERR(EINVAL);
745                 }
746                                         
747                 h_pending--;
748                                                 
749                 if (DRM_COPY_TO_USER(ref_age_base + buf_idx + 1, &h_pending, sizeof(u32))) {
750                         return DRM_ERR(EINVAL);
751                 }
752                                         
753                 cmdbuf->buf += sizeof(buf_idx);
754                 cmdbuf->bufsz -= sizeof(buf_idx);
755         }
756         
757         BEGIN_RING(2);
758         OUT_RING(CP_PACKET0(RADEON_SCRATCH_REG0 + header.scratch.reg * 4, 0));
759         OUT_RING(dev_priv->scratch_ages[header.scratch.reg]);
760         ADVANCE_RING();
761         
762         return 0;
763 }
764
765 /**
766  * Parses and validates a user-supplied command buffer and emits appropriate
767  * commands on the DMA ring buffer.
768  * Called by the ioctl handler function radeon_cp_cmdbuf.
769  */
770 int r300_do_cp_cmdbuf(drm_device_t *dev,
771                       DRMFILE filp,
772                       drm_file_t *filp_priv,
773                       drm_radeon_kcmd_buffer_t *cmdbuf)
774 {
775         drm_radeon_private_t *dev_priv = dev->dev_private;
776         drm_device_dma_t *dma = dev->dma;
777         drm_buf_t *buf = NULL;
778         int emit_dispatch_age = 0;
779         int ret = 0;
780
781         DRM_DEBUG("\n");
782
783         /* See the comment above r300_emit_begin3d for why this call must be here,
784          * and what the cleanup gotos are for. */
785         r300_pacify(dev_priv);
786
787         if (cmdbuf->nbox <= R300_SIMULTANEOUS_CLIPRECTS) {
788                 ret = r300_emit_cliprects(dev_priv, cmdbuf, 0);
789                 if (ret)
790                         goto cleanup;
791         }
792
793         while (cmdbuf->bufsz >= sizeof(drm_r300_cmd_header_t)) {
794                 int idx;
795                 drm_r300_cmd_header_t header;
796
797                 header.u = *(unsigned int *)cmdbuf->buf;
798
799                 cmdbuf->buf += sizeof(header);
800                 cmdbuf->bufsz -= sizeof(header);
801
802                 switch (header.header.cmd_type) {
803                 case R300_CMD_PACKET0:
804                         DRM_DEBUG("R300_CMD_PACKET0\n");
805                         ret = r300_emit_packet0(dev_priv, cmdbuf, header);
806                         if (ret) {
807                                 DRM_ERROR("r300_emit_packet0 failed\n");
808                                 goto cleanup;
809                         }
810                         break;
811
812                 case R300_CMD_VPU:
813                         DRM_DEBUG("R300_CMD_VPU\n");
814                         ret = r300_emit_vpu(dev_priv, cmdbuf, header);
815                         if (ret) {
816                                 DRM_ERROR("r300_emit_vpu failed\n");
817                                 goto cleanup;
818                         }
819                         break;
820
821                 case R300_CMD_PACKET3:
822                         DRM_DEBUG("R300_CMD_PACKET3\n");
823                         ret = r300_emit_packet3(dev_priv, cmdbuf, header);
824                         if (ret) {
825                                 DRM_ERROR("r300_emit_packet3 failed\n");
826                                 goto cleanup;
827                         }
828                         break;
829
830                 case R300_CMD_END3D:
831                         DRM_DEBUG("R300_CMD_END3D\n");
832                         /* TODO:
833                            Ideally userspace driver should not need to issue this call,
834                            i.e. the drm driver should issue it automatically and prevent
835                            lockups.
836
837                            In practice, we do not understand why this call is needed and what
838                            it does (except for some vague guesses that it has to do with cache
839                            coherence) and so the user space driver does it.
840
841                            Once we are sure which uses prevent lockups the code could be moved
842                            into the kernel and the userspace driver will not
843                            need to use this command.
844
845                            Note that issuing this command does not hurt anything
846                            except, possibly, performance */
847                         r300_pacify(dev_priv);
848                         break;
849
850                 case R300_CMD_CP_DELAY:
851                         /* simple enough, we can do it here */
852                         DRM_DEBUG("R300_CMD_CP_DELAY\n");
853                         {
854                                 int i;
855                                 RING_LOCALS;
856
857                                 BEGIN_RING(header.delay.count);
858                                 for (i = 0; i < header.delay.count; i++)
859                                         OUT_RING(RADEON_CP_PACKET2);
860                                 ADVANCE_RING();
861                         }
862                         break;
863
864                 case R300_CMD_DMA_DISCARD:
865                         DRM_DEBUG("RADEON_CMD_DMA_DISCARD\n");
866                         idx = header.dma.buf_idx;
867                         if (idx < 0 || idx >= dma->buf_count) {
868                                 DRM_ERROR("buffer index %d (of %d max)\n",
869                                           idx, dma->buf_count - 1);
870                                 ret = DRM_ERR(EINVAL);
871                                 goto cleanup;
872                         }
873
874                         buf = dma->buflist[idx];
875                         if (buf->filp != filp || buf->pending) {
876                                 DRM_ERROR("bad buffer %p %p %d\n",
877                                           buf->filp, filp, buf->pending);
878                                 ret = DRM_ERR(EINVAL);
879                                 goto cleanup;
880                         }
881
882                         emit_dispatch_age = 1;
883                         r300_discard_buffer(dev, buf);
884                         break;
885
886                 case R300_CMD_WAIT:
887                         /* simple enough, we can do it here */
888                         DRM_DEBUG("R300_CMD_WAIT\n");
889                         if (header.wait.flags == 0)
890                                 break;  /* nothing to do */
891
892                         {
893                                 RING_LOCALS;
894
895                                 BEGIN_RING(2);
896                                 OUT_RING(CP_PACKET0(RADEON_WAIT_UNTIL, 0));
897                                 OUT_RING((header.wait.flags & 0xf) << 14);
898                                 ADVANCE_RING();
899                         }
900                         break;
901
902                 case R300_CMD_SCRATCH:
903                         DRM_DEBUG("R300_CMD_SCRATCH\n");
904                         ret = r300_scratch(dev_priv, cmdbuf, header);
905                         if (ret) {
906                                 DRM_ERROR("r300_scratch failed\n");
907                                 goto cleanup;
908                         }
909                         break;
910                         
911                 default:
912                         DRM_ERROR("bad cmd_type %i at %p\n",
913                                   header.header.cmd_type,
914                                   cmdbuf->buf - sizeof(header));
915                         ret = DRM_ERR(EINVAL);
916                         goto cleanup;
917                 }
918         }
919
920         DRM_DEBUG("END\n");
921
922       cleanup:
923         r300_pacify(dev_priv);
924
925         /* We emit the vertex buffer age here, outside the pacifier "brackets"
926          * for two reasons:
927          *  (1) This may coalesce multiple age emissions into a single one and
928          *  (2) more importantly, some chips lock up hard when scratch registers
929          *      are written inside the pacifier bracket.
930          */
931         if (emit_dispatch_age) {
932                 RING_LOCALS;
933
934                 /* Emit the vertex buffer age */
935                 BEGIN_RING(2);
936                 RADEON_DISPATCH_AGE(dev_priv->sarea_priv->last_dispatch);
937                 ADVANCE_RING();
938         }
939
940         COMMIT_RING();
941
942         return ret;
943 }