Loading...
1/* via_dma.c -- DMA support for the VIA Unichrome/Pro
2 *
3 * Copyright 2003 Tungsten Graphics, Inc., Cedar Park, Texas.
4 * All Rights Reserved.
5 *
6 * Copyright 2004 Digeo, Inc., Palo Alto, CA, U.S.A.
7 * All Rights Reserved.
8 *
9 * Copyright 2004 The Unichrome project.
10 * All Rights Reserved.
11 *
12 * Permission is hereby granted, free of charge, to any person obtaining a
13 * copy of this software and associated documentation files (the "Software"),
14 * to deal in the Software without restriction, including without limitation
15 * the rights to use, copy, modify, merge, publish, distribute, sub license,
16 * and/or sell copies of the Software, and to permit persons to whom the
17 * Software is furnished to do so, subject to the following conditions:
18 *
19 * The above copyright notice and this permission notice (including the
20 * next paragraph) shall be included in all copies or substantial portions
21 * of the Software.
22 *
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
24 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
25 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
26 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
27 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
28 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
29 * USE OR OTHER DEALINGS IN THE SOFTWARE.
30 *
31 * Authors:
32 * Tungsten Graphics,
33 * Erdi Chen,
34 * Thomas Hellstrom.
35 */
36
37#include "drmP.h"
38#include "drm.h"
39#include "via_drm.h"
40#include "via_drv.h"
41#include "via_3d_reg.h"
42
43#define CMDBUF_ALIGNMENT_SIZE (0x100)
44#define CMDBUF_ALIGNMENT_MASK (0x0ff)
45
46/* defines for VIA 3D registers */
47#define VIA_REG_STATUS 0x400
48#define VIA_REG_TRANSET 0x43C
49#define VIA_REG_TRANSPACE 0x440
50
51/* VIA_REG_STATUS(0x400): Engine Status */
52#define VIA_CMD_RGTR_BUSY 0x00000080 /* Command Regulator is busy */
53#define VIA_2D_ENG_BUSY 0x00000001 /* 2D Engine is busy */
54#define VIA_3D_ENG_BUSY 0x00000002 /* 3D Engine is busy */
55#define VIA_VR_QUEUE_BUSY 0x00020000 /* Virtual Queue is busy */
56
57#define SetReg2DAGP(nReg, nData) { \
58 *((uint32_t *)(vb)) = ((nReg) >> 2) | HALCYON_HEADER1; \
59 *((uint32_t *)(vb) + 1) = (nData); \
60 vb = ((uint32_t *)vb) + 2; \
61 dev_priv->dma_low += 8; \
62}
63
64#define via_flush_write_combine() DRM_MEMORYBARRIER()
65
66#define VIA_OUT_RING_QW(w1, w2) do { \
67 *vb++ = (w1); \
68 *vb++ = (w2); \
69 dev_priv->dma_low += 8; \
70} while (0)
71
72static void via_cmdbuf_start(drm_via_private_t *dev_priv);
73static void via_cmdbuf_pause(drm_via_private_t *dev_priv);
74static void via_cmdbuf_reset(drm_via_private_t *dev_priv);
75static void via_cmdbuf_rewind(drm_via_private_t *dev_priv);
76static int via_wait_idle(drm_via_private_t *dev_priv);
77static void via_pad_cache(drm_via_private_t *dev_priv, int qwords);
78
79/*
80 * Free space in command buffer.
81 */
82
83static uint32_t via_cmdbuf_space(drm_via_private_t *dev_priv)
84{
85 uint32_t agp_base = dev_priv->dma_offset + (uint32_t) dev_priv->agpAddr;
86 uint32_t hw_addr = *(dev_priv->hw_addr_ptr) - agp_base;
87
88 return ((hw_addr <= dev_priv->dma_low) ?
89 (dev_priv->dma_high + hw_addr - dev_priv->dma_low) :
90 (hw_addr - dev_priv->dma_low));
91}
92
93/*
94 * How much does the command regulator lag behind?
95 */
96
97static uint32_t via_cmdbuf_lag(drm_via_private_t *dev_priv)
98{
99 uint32_t agp_base = dev_priv->dma_offset + (uint32_t) dev_priv->agpAddr;
100 uint32_t hw_addr = *(dev_priv->hw_addr_ptr) - agp_base;
101
102 return ((hw_addr <= dev_priv->dma_low) ?
103 (dev_priv->dma_low - hw_addr) :
104 (dev_priv->dma_wrap + dev_priv->dma_low - hw_addr));
105}
106
107/*
108 * Check that the given size fits in the buffer, otherwise wait.
109 */
110
111static inline int
112via_cmdbuf_wait(drm_via_private_t *dev_priv, unsigned int size)
113{
114 uint32_t agp_base = dev_priv->dma_offset + (uint32_t) dev_priv->agpAddr;
115 uint32_t cur_addr, hw_addr, next_addr;
116 volatile uint32_t *hw_addr_ptr;
117 uint32_t count;
118 hw_addr_ptr = dev_priv->hw_addr_ptr;
119 cur_addr = dev_priv->dma_low;
120 next_addr = cur_addr + size + 512 * 1024;
121 count = 1000000;
122 do {
123 hw_addr = *hw_addr_ptr - agp_base;
124 if (count-- == 0) {
125 DRM_ERROR
126 ("via_cmdbuf_wait timed out hw %x cur_addr %x next_addr %x\n",
127 hw_addr, cur_addr, next_addr);
128 return -1;
129 }
130 if ((cur_addr < hw_addr) && (next_addr >= hw_addr))
131 msleep(1);
132 } while ((cur_addr < hw_addr) && (next_addr >= hw_addr));
133 return 0;
134}
135
136/*
137 * Checks whether buffer head has reach the end. Rewind the ring buffer
138 * when necessary.
139 *
140 * Returns virtual pointer to ring buffer.
141 */
142
143static inline uint32_t *via_check_dma(drm_via_private_t * dev_priv,
144 unsigned int size)
145{
146 if ((dev_priv->dma_low + size + 4 * CMDBUF_ALIGNMENT_SIZE) >
147 dev_priv->dma_high) {
148 via_cmdbuf_rewind(dev_priv);
149 }
150 if (via_cmdbuf_wait(dev_priv, size) != 0)
151 return NULL;
152
153 return (uint32_t *) (dev_priv->dma_ptr + dev_priv->dma_low);
154}
155
156int via_dma_cleanup(struct drm_device *dev)
157{
158 if (dev->dev_private) {
159 drm_via_private_t *dev_priv =
160 (drm_via_private_t *) dev->dev_private;
161
162 if (dev_priv->ring.virtual_start) {
163 via_cmdbuf_reset(dev_priv);
164
165 drm_core_ioremapfree(&dev_priv->ring.map, dev);
166 dev_priv->ring.virtual_start = NULL;
167 }
168
169 }
170
171 return 0;
172}
173
174static int via_initialize(struct drm_device *dev,
175 drm_via_private_t *dev_priv,
176 drm_via_dma_init_t *init)
177{
178 if (!dev_priv || !dev_priv->mmio) {
179 DRM_ERROR("via_dma_init called before via_map_init\n");
180 return -EFAULT;
181 }
182
183 if (dev_priv->ring.virtual_start != NULL) {
184 DRM_ERROR("called again without calling cleanup\n");
185 return -EFAULT;
186 }
187
188 if (!dev->agp || !dev->agp->base) {
189 DRM_ERROR("called with no agp memory available\n");
190 return -EFAULT;
191 }
192
193 if (dev_priv->chipset == VIA_DX9_0) {
194 DRM_ERROR("AGP DMA is not supported on this chip\n");
195 return -EINVAL;
196 }
197
198 dev_priv->ring.map.offset = dev->agp->base + init->offset;
199 dev_priv->ring.map.size = init->size;
200 dev_priv->ring.map.type = 0;
201 dev_priv->ring.map.flags = 0;
202 dev_priv->ring.map.mtrr = 0;
203
204 drm_core_ioremap(&dev_priv->ring.map, dev);
205
206 if (dev_priv->ring.map.handle == NULL) {
207 via_dma_cleanup(dev);
208 DRM_ERROR("can not ioremap virtual address for"
209 " ring buffer\n");
210 return -ENOMEM;
211 }
212
213 dev_priv->ring.virtual_start = dev_priv->ring.map.handle;
214
215 dev_priv->dma_ptr = dev_priv->ring.virtual_start;
216 dev_priv->dma_low = 0;
217 dev_priv->dma_high = init->size;
218 dev_priv->dma_wrap = init->size;
219 dev_priv->dma_offset = init->offset;
220 dev_priv->last_pause_ptr = NULL;
221 dev_priv->hw_addr_ptr =
222 (volatile uint32_t *)((char *)dev_priv->mmio->handle +
223 init->reg_pause_addr);
224
225 via_cmdbuf_start(dev_priv);
226
227 return 0;
228}
229
230static int via_dma_init(struct drm_device *dev, void *data, struct drm_file *file_priv)
231{
232 drm_via_private_t *dev_priv = (drm_via_private_t *) dev->dev_private;
233 drm_via_dma_init_t *init = data;
234 int retcode = 0;
235
236 switch (init->func) {
237 case VIA_INIT_DMA:
238 if (!DRM_SUSER(DRM_CURPROC))
239 retcode = -EPERM;
240 else
241 retcode = via_initialize(dev, dev_priv, init);
242 break;
243 case VIA_CLEANUP_DMA:
244 if (!DRM_SUSER(DRM_CURPROC))
245 retcode = -EPERM;
246 else
247 retcode = via_dma_cleanup(dev);
248 break;
249 case VIA_DMA_INITIALIZED:
250 retcode = (dev_priv->ring.virtual_start != NULL) ?
251 0 : -EFAULT;
252 break;
253 default:
254 retcode = -EINVAL;
255 break;
256 }
257
258 return retcode;
259}
260
261static int via_dispatch_cmdbuffer(struct drm_device *dev, drm_via_cmdbuffer_t *cmd)
262{
263 drm_via_private_t *dev_priv;
264 uint32_t *vb;
265 int ret;
266
267 dev_priv = (drm_via_private_t *) dev->dev_private;
268
269 if (dev_priv->ring.virtual_start == NULL) {
270 DRM_ERROR("called without initializing AGP ring buffer.\n");
271 return -EFAULT;
272 }
273
274 if (cmd->size > VIA_PCI_BUF_SIZE)
275 return -ENOMEM;
276
277 if (DRM_COPY_FROM_USER(dev_priv->pci_buf, cmd->buf, cmd->size))
278 return -EFAULT;
279
280 /*
281 * Running this function on AGP memory is dead slow. Therefore
282 * we run it on a temporary cacheable system memory buffer and
283 * copy it to AGP memory when ready.
284 */
285
286 if ((ret =
287 via_verify_command_stream((uint32_t *) dev_priv->pci_buf,
288 cmd->size, dev, 1))) {
289 return ret;
290 }
291
292 vb = via_check_dma(dev_priv, (cmd->size < 0x100) ? 0x102 : cmd->size);
293 if (vb == NULL)
294 return -EAGAIN;
295
296 memcpy(vb, dev_priv->pci_buf, cmd->size);
297
298 dev_priv->dma_low += cmd->size;
299
300 /*
301 * Small submissions somehow stalls the CPU. (AGP cache effects?)
302 * pad to greater size.
303 */
304
305 if (cmd->size < 0x100)
306 via_pad_cache(dev_priv, (0x100 - cmd->size) >> 3);
307 via_cmdbuf_pause(dev_priv);
308
309 return 0;
310}
311
312int via_driver_dma_quiescent(struct drm_device *dev)
313{
314 drm_via_private_t *dev_priv = dev->dev_private;
315
316 if (!via_wait_idle(dev_priv))
317 return -EBUSY;
318 return 0;
319}
320
321static int via_flush_ioctl(struct drm_device *dev, void *data, struct drm_file *file_priv)
322{
323
324 LOCK_TEST_WITH_RETURN(dev, file_priv);
325
326 return via_driver_dma_quiescent(dev);
327}
328
329static int via_cmdbuffer(struct drm_device *dev, void *data, struct drm_file *file_priv)
330{
331 drm_via_cmdbuffer_t *cmdbuf = data;
332 int ret;
333
334 LOCK_TEST_WITH_RETURN(dev, file_priv);
335
336 DRM_DEBUG("buf %p size %lu\n", cmdbuf->buf, cmdbuf->size);
337
338 ret = via_dispatch_cmdbuffer(dev, cmdbuf);
339 return ret;
340}
341
342static int via_dispatch_pci_cmdbuffer(struct drm_device *dev,
343 drm_via_cmdbuffer_t *cmd)
344{
345 drm_via_private_t *dev_priv = dev->dev_private;
346 int ret;
347
348 if (cmd->size > VIA_PCI_BUF_SIZE)
349 return -ENOMEM;
350 if (DRM_COPY_FROM_USER(dev_priv->pci_buf, cmd->buf, cmd->size))
351 return -EFAULT;
352
353 if ((ret =
354 via_verify_command_stream((uint32_t *) dev_priv->pci_buf,
355 cmd->size, dev, 0))) {
356 return ret;
357 }
358
359 ret =
360 via_parse_command_stream(dev, (const uint32_t *)dev_priv->pci_buf,
361 cmd->size);
362 return ret;
363}
364
365static int via_pci_cmdbuffer(struct drm_device *dev, void *data, struct drm_file *file_priv)
366{
367 drm_via_cmdbuffer_t *cmdbuf = data;
368 int ret;
369
370 LOCK_TEST_WITH_RETURN(dev, file_priv);
371
372 DRM_DEBUG("buf %p size %lu\n", cmdbuf->buf, cmdbuf->size);
373
374 ret = via_dispatch_pci_cmdbuffer(dev, cmdbuf);
375 return ret;
376}
377
378static inline uint32_t *via_align_buffer(drm_via_private_t *dev_priv,
379 uint32_t * vb, int qw_count)
380{
381 for (; qw_count > 0; --qw_count)
382 VIA_OUT_RING_QW(HC_DUMMY, HC_DUMMY);
383 return vb;
384}
385
386/*
387 * This function is used internally by ring buffer management code.
388 *
389 * Returns virtual pointer to ring buffer.
390 */
391static inline uint32_t *via_get_dma(drm_via_private_t *dev_priv)
392{
393 return (uint32_t *) (dev_priv->dma_ptr + dev_priv->dma_low);
394}
395
396/*
397 * Hooks a segment of data into the tail of the ring-buffer by
398 * modifying the pause address stored in the buffer itself. If
399 * the regulator has already paused, restart it.
400 */
401static int via_hook_segment(drm_via_private_t *dev_priv,
402 uint32_t pause_addr_hi, uint32_t pause_addr_lo,
403 int no_pci_fire)
404{
405 int paused, count;
406 volatile uint32_t *paused_at = dev_priv->last_pause_ptr;
407 uint32_t reader, ptr;
408 uint32_t diff;
409
410 paused = 0;
411 via_flush_write_combine();
412 (void) *(volatile uint32_t *)(via_get_dma(dev_priv) - 1);
413
414 *paused_at = pause_addr_lo;
415 via_flush_write_combine();
416 (void) *paused_at;
417
418 reader = *(dev_priv->hw_addr_ptr);
419 ptr = ((volatile char *)paused_at - dev_priv->dma_ptr) +
420 dev_priv->dma_offset + (uint32_t) dev_priv->agpAddr + 4;
421
422 dev_priv->last_pause_ptr = via_get_dma(dev_priv) - 1;
423
424 /*
425 * If there is a possibility that the command reader will
426 * miss the new pause address and pause on the old one,
427 * In that case we need to program the new start address
428 * using PCI.
429 */
430
431 diff = (uint32_t) (ptr - reader) - dev_priv->dma_diff;
432 count = 10000000;
433 while (diff == 0 && count--) {
434 paused = (VIA_READ(0x41c) & 0x80000000);
435 if (paused)
436 break;
437 reader = *(dev_priv->hw_addr_ptr);
438 diff = (uint32_t) (ptr - reader) - dev_priv->dma_diff;
439 }
440
441 paused = VIA_READ(0x41c) & 0x80000000;
442
443 if (paused && !no_pci_fire) {
444 reader = *(dev_priv->hw_addr_ptr);
445 diff = (uint32_t) (ptr - reader) - dev_priv->dma_diff;
446 diff &= (dev_priv->dma_high - 1);
447 if (diff != 0 && diff < (dev_priv->dma_high >> 1)) {
448 DRM_ERROR("Paused at incorrect address. "
449 "0x%08x, 0x%08x 0x%08x\n",
450 ptr, reader, dev_priv->dma_diff);
451 } else if (diff == 0) {
452 /*
453 * There is a concern that these writes may stall the PCI bus
454 * if the GPU is not idle. However, idling the GPU first
455 * doesn't make a difference.
456 */
457
458 VIA_WRITE(VIA_REG_TRANSET, (HC_ParaType_PreCR << 16));
459 VIA_WRITE(VIA_REG_TRANSPACE, pause_addr_hi);
460 VIA_WRITE(VIA_REG_TRANSPACE, pause_addr_lo);
461 VIA_READ(VIA_REG_TRANSPACE);
462 }
463 }
464 return paused;
465}
466
467static int via_wait_idle(drm_via_private_t *dev_priv)
468{
469 int count = 10000000;
470
471 while (!(VIA_READ(VIA_REG_STATUS) & VIA_VR_QUEUE_BUSY) && --count)
472 ;
473
474 while (count && (VIA_READ(VIA_REG_STATUS) &
475 (VIA_CMD_RGTR_BUSY | VIA_2D_ENG_BUSY |
476 VIA_3D_ENG_BUSY)))
477 --count;
478 return count;
479}
480
481static uint32_t *via_align_cmd(drm_via_private_t *dev_priv, uint32_t cmd_type,
482 uint32_t addr, uint32_t *cmd_addr_hi,
483 uint32_t *cmd_addr_lo, int skip_wait)
484{
485 uint32_t agp_base;
486 uint32_t cmd_addr, addr_lo, addr_hi;
487 uint32_t *vb;
488 uint32_t qw_pad_count;
489
490 if (!skip_wait)
491 via_cmdbuf_wait(dev_priv, 2 * CMDBUF_ALIGNMENT_SIZE);
492
493 vb = via_get_dma(dev_priv);
494 VIA_OUT_RING_QW(HC_HEADER2 | ((VIA_REG_TRANSET >> 2) << 12) |
495 (VIA_REG_TRANSPACE >> 2), HC_ParaType_PreCR << 16);
496 agp_base = dev_priv->dma_offset + (uint32_t) dev_priv->agpAddr;
497 qw_pad_count = (CMDBUF_ALIGNMENT_SIZE >> 3) -
498 ((dev_priv->dma_low & CMDBUF_ALIGNMENT_MASK) >> 3);
499
500 cmd_addr = (addr) ? addr :
501 agp_base + dev_priv->dma_low - 8 + (qw_pad_count << 3);
502 addr_lo = ((HC_SubA_HAGPBpL << 24) | (cmd_type & HC_HAGPBpID_MASK) |
503 (cmd_addr & HC_HAGPBpL_MASK));
504 addr_hi = ((HC_SubA_HAGPBpH << 24) | (cmd_addr >> 24));
505
506 vb = via_align_buffer(dev_priv, vb, qw_pad_count - 1);
507 VIA_OUT_RING_QW(*cmd_addr_hi = addr_hi, *cmd_addr_lo = addr_lo);
508 return vb;
509}
510
511static void via_cmdbuf_start(drm_via_private_t *dev_priv)
512{
513 uint32_t pause_addr_lo, pause_addr_hi;
514 uint32_t start_addr, start_addr_lo;
515 uint32_t end_addr, end_addr_lo;
516 uint32_t command;
517 uint32_t agp_base;
518 uint32_t ptr;
519 uint32_t reader;
520 int count;
521
522 dev_priv->dma_low = 0;
523
524 agp_base = dev_priv->dma_offset + (uint32_t) dev_priv->agpAddr;
525 start_addr = agp_base;
526 end_addr = agp_base + dev_priv->dma_high;
527
528 start_addr_lo = ((HC_SubA_HAGPBstL << 24) | (start_addr & 0xFFFFFF));
529 end_addr_lo = ((HC_SubA_HAGPBendL << 24) | (end_addr & 0xFFFFFF));
530 command = ((HC_SubA_HAGPCMNT << 24) | (start_addr >> 24) |
531 ((end_addr & 0xff000000) >> 16));
532
533 dev_priv->last_pause_ptr =
534 via_align_cmd(dev_priv, HC_HAGPBpID_PAUSE, 0,
535 &pause_addr_hi, &pause_addr_lo, 1) - 1;
536
537 via_flush_write_combine();
538 (void) *(volatile uint32_t *)dev_priv->last_pause_ptr;
539
540 VIA_WRITE(VIA_REG_TRANSET, (HC_ParaType_PreCR << 16));
541 VIA_WRITE(VIA_REG_TRANSPACE, command);
542 VIA_WRITE(VIA_REG_TRANSPACE, start_addr_lo);
543 VIA_WRITE(VIA_REG_TRANSPACE, end_addr_lo);
544
545 VIA_WRITE(VIA_REG_TRANSPACE, pause_addr_hi);
546 VIA_WRITE(VIA_REG_TRANSPACE, pause_addr_lo);
547 DRM_WRITEMEMORYBARRIER();
548 VIA_WRITE(VIA_REG_TRANSPACE, command | HC_HAGPCMNT_MASK);
549 VIA_READ(VIA_REG_TRANSPACE);
550
551 dev_priv->dma_diff = 0;
552
553 count = 10000000;
554 while (!(VIA_READ(0x41c) & 0x80000000) && count--);
555
556 reader = *(dev_priv->hw_addr_ptr);
557 ptr = ((volatile char *)dev_priv->last_pause_ptr - dev_priv->dma_ptr) +
558 dev_priv->dma_offset + (uint32_t) dev_priv->agpAddr + 4;
559
560 /*
561 * This is the difference between where we tell the
562 * command reader to pause and where it actually pauses.
563 * This differs between hw implementation so we need to
564 * detect it.
565 */
566
567 dev_priv->dma_diff = ptr - reader;
568}
569
570static void via_pad_cache(drm_via_private_t *dev_priv, int qwords)
571{
572 uint32_t *vb;
573
574 via_cmdbuf_wait(dev_priv, qwords + 2);
575 vb = via_get_dma(dev_priv);
576 VIA_OUT_RING_QW(HC_HEADER2, HC_ParaType_NotTex << 16);
577 via_align_buffer(dev_priv, vb, qwords);
578}
579
580static inline void via_dummy_bitblt(drm_via_private_t *dev_priv)
581{
582 uint32_t *vb = via_get_dma(dev_priv);
583 SetReg2DAGP(0x0C, (0 | (0 << 16)));
584 SetReg2DAGP(0x10, 0 | (0 << 16));
585 SetReg2DAGP(0x0, 0x1 | 0x2000 | 0xAA000000);
586}
587
588static void via_cmdbuf_jump(drm_via_private_t *dev_priv)
589{
590 uint32_t agp_base;
591 uint32_t pause_addr_lo, pause_addr_hi;
592 uint32_t jump_addr_lo, jump_addr_hi;
593 volatile uint32_t *last_pause_ptr;
594 uint32_t dma_low_save1, dma_low_save2;
595
596 agp_base = dev_priv->dma_offset + (uint32_t) dev_priv->agpAddr;
597 via_align_cmd(dev_priv, HC_HAGPBpID_JUMP, 0, &jump_addr_hi,
598 &jump_addr_lo, 0);
599
600 dev_priv->dma_wrap = dev_priv->dma_low;
601
602 /*
603 * Wrap command buffer to the beginning.
604 */
605
606 dev_priv->dma_low = 0;
607 if (via_cmdbuf_wait(dev_priv, CMDBUF_ALIGNMENT_SIZE) != 0)
608 DRM_ERROR("via_cmdbuf_jump failed\n");
609
610 via_dummy_bitblt(dev_priv);
611 via_dummy_bitblt(dev_priv);
612
613 last_pause_ptr =
614 via_align_cmd(dev_priv, HC_HAGPBpID_PAUSE, 0, &pause_addr_hi,
615 &pause_addr_lo, 0) - 1;
616 via_align_cmd(dev_priv, HC_HAGPBpID_PAUSE, 0, &pause_addr_hi,
617 &pause_addr_lo, 0);
618
619 *last_pause_ptr = pause_addr_lo;
620 dma_low_save1 = dev_priv->dma_low;
621
622 /*
623 * Now, set a trap that will pause the regulator if it tries to rerun the old
624 * command buffer. (Which may happen if via_hook_segment detecs a command regulator pause
625 * and reissues the jump command over PCI, while the regulator has already taken the jump
626 * and actually paused at the current buffer end).
627 * There appears to be no other way to detect this condition, since the hw_addr_pointer
628 * does not seem to get updated immediately when a jump occurs.
629 */
630
631 last_pause_ptr =
632 via_align_cmd(dev_priv, HC_HAGPBpID_PAUSE, 0, &pause_addr_hi,
633 &pause_addr_lo, 0) - 1;
634 via_align_cmd(dev_priv, HC_HAGPBpID_PAUSE, 0, &pause_addr_hi,
635 &pause_addr_lo, 0);
636 *last_pause_ptr = pause_addr_lo;
637
638 dma_low_save2 = dev_priv->dma_low;
639 dev_priv->dma_low = dma_low_save1;
640 via_hook_segment(dev_priv, jump_addr_hi, jump_addr_lo, 0);
641 dev_priv->dma_low = dma_low_save2;
642 via_hook_segment(dev_priv, pause_addr_hi, pause_addr_lo, 0);
643}
644
645
646static void via_cmdbuf_rewind(drm_via_private_t *dev_priv)
647{
648 via_cmdbuf_jump(dev_priv);
649}
650
651static void via_cmdbuf_flush(drm_via_private_t *dev_priv, uint32_t cmd_type)
652{
653 uint32_t pause_addr_lo, pause_addr_hi;
654
655 via_align_cmd(dev_priv, cmd_type, 0, &pause_addr_hi, &pause_addr_lo, 0);
656 via_hook_segment(dev_priv, pause_addr_hi, pause_addr_lo, 0);
657}
658
659static void via_cmdbuf_pause(drm_via_private_t *dev_priv)
660{
661 via_cmdbuf_flush(dev_priv, HC_HAGPBpID_PAUSE);
662}
663
664static void via_cmdbuf_reset(drm_via_private_t *dev_priv)
665{
666 via_cmdbuf_flush(dev_priv, HC_HAGPBpID_STOP);
667 via_wait_idle(dev_priv);
668}
669
670/*
671 * User interface to the space and lag functions.
672 */
673
674static int via_cmdbuf_size(struct drm_device *dev, void *data, struct drm_file *file_priv)
675{
676 drm_via_cmdbuf_size_t *d_siz = data;
677 int ret = 0;
678 uint32_t tmp_size, count;
679 drm_via_private_t *dev_priv;
680
681 DRM_DEBUG("\n");
682 LOCK_TEST_WITH_RETURN(dev, file_priv);
683
684 dev_priv = (drm_via_private_t *) dev->dev_private;
685
686 if (dev_priv->ring.virtual_start == NULL) {
687 DRM_ERROR("called without initializing AGP ring buffer.\n");
688 return -EFAULT;
689 }
690
691 count = 1000000;
692 tmp_size = d_siz->size;
693 switch (d_siz->func) {
694 case VIA_CMDBUF_SPACE:
695 while (((tmp_size = via_cmdbuf_space(dev_priv)) < d_siz->size)
696 && --count) {
697 if (!d_siz->wait)
698 break;
699 }
700 if (!count) {
701 DRM_ERROR("VIA_CMDBUF_SPACE timed out.\n");
702 ret = -EAGAIN;
703 }
704 break;
705 case VIA_CMDBUF_LAG:
706 while (((tmp_size = via_cmdbuf_lag(dev_priv)) > d_siz->size)
707 && --count) {
708 if (!d_siz->wait)
709 break;
710 }
711 if (!count) {
712 DRM_ERROR("VIA_CMDBUF_LAG timed out.\n");
713 ret = -EAGAIN;
714 }
715 break;
716 default:
717 ret = -EFAULT;
718 }
719 d_siz->size = tmp_size;
720
721 return ret;
722}
723
724struct drm_ioctl_desc via_ioctls[] = {
725 DRM_IOCTL_DEF_DRV(VIA_ALLOCMEM, via_mem_alloc, DRM_AUTH),
726 DRM_IOCTL_DEF_DRV(VIA_FREEMEM, via_mem_free, DRM_AUTH),
727 DRM_IOCTL_DEF_DRV(VIA_AGP_INIT, via_agp_init, DRM_AUTH|DRM_MASTER),
728 DRM_IOCTL_DEF_DRV(VIA_FB_INIT, via_fb_init, DRM_AUTH|DRM_MASTER),
729 DRM_IOCTL_DEF_DRV(VIA_MAP_INIT, via_map_init, DRM_AUTH|DRM_MASTER),
730 DRM_IOCTL_DEF_DRV(VIA_DEC_FUTEX, via_decoder_futex, DRM_AUTH),
731 DRM_IOCTL_DEF_DRV(VIA_DMA_INIT, via_dma_init, DRM_AUTH),
732 DRM_IOCTL_DEF_DRV(VIA_CMDBUFFER, via_cmdbuffer, DRM_AUTH),
733 DRM_IOCTL_DEF_DRV(VIA_FLUSH, via_flush_ioctl, DRM_AUTH),
734 DRM_IOCTL_DEF_DRV(VIA_PCICMD, via_pci_cmdbuffer, DRM_AUTH),
735 DRM_IOCTL_DEF_DRV(VIA_CMDBUF_SIZE, via_cmdbuf_size, DRM_AUTH),
736 DRM_IOCTL_DEF_DRV(VIA_WAIT_IRQ, via_wait_irq, DRM_AUTH),
737 DRM_IOCTL_DEF_DRV(VIA_DMA_BLIT, via_dma_blit, DRM_AUTH),
738 DRM_IOCTL_DEF_DRV(VIA_BLIT_SYNC, via_dma_blit_sync, DRM_AUTH)
739};
740
741int via_max_ioctl = DRM_ARRAY_SIZE(via_ioctls);
1/* via_dma.c -- DMA support for the VIA Unichrome/Pro
2 *
3 * Copyright 2003 Tungsten Graphics, Inc., Cedar Park, Texas.
4 * All Rights Reserved.
5 *
6 * Copyright 2004 Digeo, Inc., Palo Alto, CA, U.S.A.
7 * All Rights Reserved.
8 *
9 * Copyright 2004 The Unichrome project.
10 * All Rights Reserved.
11 *
12 * Permission is hereby granted, free of charge, to any person obtaining a
13 * copy of this software and associated documentation files (the "Software"),
14 * to deal in the Software without restriction, including without limitation
15 * the rights to use, copy, modify, merge, publish, distribute, sub license,
16 * and/or sell copies of the Software, and to permit persons to whom the
17 * Software is furnished to do so, subject to the following conditions:
18 *
19 * The above copyright notice and this permission notice (including the
20 * next paragraph) shall be included in all copies or substantial portions
21 * of the Software.
22 *
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
24 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
25 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
26 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
27 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
28 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
29 * USE OR OTHER DEALINGS IN THE SOFTWARE.
30 *
31 * Authors:
32 * Tungsten Graphics,
33 * Erdi Chen,
34 * Thomas Hellstrom.
35 */
36
37#include <drm/drmP.h>
38#include <drm/via_drm.h>
39#include "via_drv.h"
40#include "via_3d_reg.h"
41
42#define CMDBUF_ALIGNMENT_SIZE (0x100)
43#define CMDBUF_ALIGNMENT_MASK (0x0ff)
44
45/* defines for VIA 3D registers */
46#define VIA_REG_STATUS 0x400
47#define VIA_REG_TRANSET 0x43C
48#define VIA_REG_TRANSPACE 0x440
49
50/* VIA_REG_STATUS(0x400): Engine Status */
51#define VIA_CMD_RGTR_BUSY 0x00000080 /* Command Regulator is busy */
52#define VIA_2D_ENG_BUSY 0x00000001 /* 2D Engine is busy */
53#define VIA_3D_ENG_BUSY 0x00000002 /* 3D Engine is busy */
54#define VIA_VR_QUEUE_BUSY 0x00020000 /* Virtual Queue is busy */
55
56#define SetReg2DAGP(nReg, nData) { \
57 *((uint32_t *)(vb)) = ((nReg) >> 2) | HALCYON_HEADER1; \
58 *((uint32_t *)(vb) + 1) = (nData); \
59 vb = ((uint32_t *)vb) + 2; \
60 dev_priv->dma_low += 8; \
61}
62
63#define via_flush_write_combine() mb()
64
65#define VIA_OUT_RING_QW(w1, w2) do { \
66 *vb++ = (w1); \
67 *vb++ = (w2); \
68 dev_priv->dma_low += 8; \
69} while (0)
70
71static void via_cmdbuf_start(drm_via_private_t *dev_priv);
72static void via_cmdbuf_pause(drm_via_private_t *dev_priv);
73static void via_cmdbuf_reset(drm_via_private_t *dev_priv);
74static void via_cmdbuf_rewind(drm_via_private_t *dev_priv);
75static int via_wait_idle(drm_via_private_t *dev_priv);
76static void via_pad_cache(drm_via_private_t *dev_priv, int qwords);
77
78/*
79 * Free space in command buffer.
80 */
81
82static uint32_t via_cmdbuf_space(drm_via_private_t *dev_priv)
83{
84 uint32_t agp_base = dev_priv->dma_offset + (uint32_t) dev_priv->agpAddr;
85 uint32_t hw_addr = *(dev_priv->hw_addr_ptr) - agp_base;
86
87 return ((hw_addr <= dev_priv->dma_low) ?
88 (dev_priv->dma_high + hw_addr - dev_priv->dma_low) :
89 (hw_addr - dev_priv->dma_low));
90}
91
92/*
93 * How much does the command regulator lag behind?
94 */
95
96static uint32_t via_cmdbuf_lag(drm_via_private_t *dev_priv)
97{
98 uint32_t agp_base = dev_priv->dma_offset + (uint32_t) dev_priv->agpAddr;
99 uint32_t hw_addr = *(dev_priv->hw_addr_ptr) - agp_base;
100
101 return ((hw_addr <= dev_priv->dma_low) ?
102 (dev_priv->dma_low - hw_addr) :
103 (dev_priv->dma_wrap + dev_priv->dma_low - hw_addr));
104}
105
106/*
107 * Check that the given size fits in the buffer, otherwise wait.
108 */
109
110static inline int
111via_cmdbuf_wait(drm_via_private_t *dev_priv, unsigned int size)
112{
113 uint32_t agp_base = dev_priv->dma_offset + (uint32_t) dev_priv->agpAddr;
114 uint32_t cur_addr, hw_addr, next_addr;
115 volatile uint32_t *hw_addr_ptr;
116 uint32_t count;
117 hw_addr_ptr = dev_priv->hw_addr_ptr;
118 cur_addr = dev_priv->dma_low;
119 next_addr = cur_addr + size + 512 * 1024;
120 count = 1000000;
121 do {
122 hw_addr = *hw_addr_ptr - agp_base;
123 if (count-- == 0) {
124 DRM_ERROR
125 ("via_cmdbuf_wait timed out hw %x cur_addr %x next_addr %x\n",
126 hw_addr, cur_addr, next_addr);
127 return -1;
128 }
129 if ((cur_addr < hw_addr) && (next_addr >= hw_addr))
130 msleep(1);
131 } while ((cur_addr < hw_addr) && (next_addr >= hw_addr));
132 return 0;
133}
134
135/*
136 * Checks whether buffer head has reach the end. Rewind the ring buffer
137 * when necessary.
138 *
139 * Returns virtual pointer to ring buffer.
140 */
141
142static inline uint32_t *via_check_dma(drm_via_private_t * dev_priv,
143 unsigned int size)
144{
145 if ((dev_priv->dma_low + size + 4 * CMDBUF_ALIGNMENT_SIZE) >
146 dev_priv->dma_high) {
147 via_cmdbuf_rewind(dev_priv);
148 }
149 if (via_cmdbuf_wait(dev_priv, size) != 0)
150 return NULL;
151
152 return (uint32_t *) (dev_priv->dma_ptr + dev_priv->dma_low);
153}
154
155int via_dma_cleanup(struct drm_device *dev)
156{
157 if (dev->dev_private) {
158 drm_via_private_t *dev_priv =
159 (drm_via_private_t *) dev->dev_private;
160
161 if (dev_priv->ring.virtual_start) {
162 via_cmdbuf_reset(dev_priv);
163
164 drm_legacy_ioremapfree(&dev_priv->ring.map, dev);
165 dev_priv->ring.virtual_start = NULL;
166 }
167
168 }
169
170 return 0;
171}
172
173static int via_initialize(struct drm_device *dev,
174 drm_via_private_t *dev_priv,
175 drm_via_dma_init_t *init)
176{
177 if (!dev_priv || !dev_priv->mmio) {
178 DRM_ERROR("via_dma_init called before via_map_init\n");
179 return -EFAULT;
180 }
181
182 if (dev_priv->ring.virtual_start != NULL) {
183 DRM_ERROR("called again without calling cleanup\n");
184 return -EFAULT;
185 }
186
187 if (!dev->agp || !dev->agp->base) {
188 DRM_ERROR("called with no agp memory available\n");
189 return -EFAULT;
190 }
191
192 if (dev_priv->chipset == VIA_DX9_0) {
193 DRM_ERROR("AGP DMA is not supported on this chip\n");
194 return -EINVAL;
195 }
196
197 dev_priv->ring.map.offset = dev->agp->base + init->offset;
198 dev_priv->ring.map.size = init->size;
199 dev_priv->ring.map.type = 0;
200 dev_priv->ring.map.flags = 0;
201 dev_priv->ring.map.mtrr = 0;
202
203 drm_legacy_ioremap(&dev_priv->ring.map, dev);
204
205 if (dev_priv->ring.map.handle == NULL) {
206 via_dma_cleanup(dev);
207 DRM_ERROR("can not ioremap virtual address for"
208 " ring buffer\n");
209 return -ENOMEM;
210 }
211
212 dev_priv->ring.virtual_start = dev_priv->ring.map.handle;
213
214 dev_priv->dma_ptr = dev_priv->ring.virtual_start;
215 dev_priv->dma_low = 0;
216 dev_priv->dma_high = init->size;
217 dev_priv->dma_wrap = init->size;
218 dev_priv->dma_offset = init->offset;
219 dev_priv->last_pause_ptr = NULL;
220 dev_priv->hw_addr_ptr =
221 (volatile uint32_t *)((char *)dev_priv->mmio->handle +
222 init->reg_pause_addr);
223
224 via_cmdbuf_start(dev_priv);
225
226 return 0;
227}
228
229static int via_dma_init(struct drm_device *dev, void *data, struct drm_file *file_priv)
230{
231 drm_via_private_t *dev_priv = (drm_via_private_t *) dev->dev_private;
232 drm_via_dma_init_t *init = data;
233 int retcode = 0;
234
235 switch (init->func) {
236 case VIA_INIT_DMA:
237 if (!capable(CAP_SYS_ADMIN))
238 retcode = -EPERM;
239 else
240 retcode = via_initialize(dev, dev_priv, init);
241 break;
242 case VIA_CLEANUP_DMA:
243 if (!capable(CAP_SYS_ADMIN))
244 retcode = -EPERM;
245 else
246 retcode = via_dma_cleanup(dev);
247 break;
248 case VIA_DMA_INITIALIZED:
249 retcode = (dev_priv->ring.virtual_start != NULL) ?
250 0 : -EFAULT;
251 break;
252 default:
253 retcode = -EINVAL;
254 break;
255 }
256
257 return retcode;
258}
259
260static int via_dispatch_cmdbuffer(struct drm_device *dev, drm_via_cmdbuffer_t *cmd)
261{
262 drm_via_private_t *dev_priv;
263 uint32_t *vb;
264 int ret;
265
266 dev_priv = (drm_via_private_t *) dev->dev_private;
267
268 if (dev_priv->ring.virtual_start == NULL) {
269 DRM_ERROR("called without initializing AGP ring buffer.\n");
270 return -EFAULT;
271 }
272
273 if (cmd->size > VIA_PCI_BUF_SIZE)
274 return -ENOMEM;
275
276 if (copy_from_user(dev_priv->pci_buf, cmd->buf, cmd->size))
277 return -EFAULT;
278
279 /*
280 * Running this function on AGP memory is dead slow. Therefore
281 * we run it on a temporary cacheable system memory buffer and
282 * copy it to AGP memory when ready.
283 */
284
285 if ((ret =
286 via_verify_command_stream((uint32_t *) dev_priv->pci_buf,
287 cmd->size, dev, 1))) {
288 return ret;
289 }
290
291 vb = via_check_dma(dev_priv, (cmd->size < 0x100) ? 0x102 : cmd->size);
292 if (vb == NULL)
293 return -EAGAIN;
294
295 memcpy(vb, dev_priv->pci_buf, cmd->size);
296
297 dev_priv->dma_low += cmd->size;
298
299 /*
300 * Small submissions somehow stalls the CPU. (AGP cache effects?)
301 * pad to greater size.
302 */
303
304 if (cmd->size < 0x100)
305 via_pad_cache(dev_priv, (0x100 - cmd->size) >> 3);
306 via_cmdbuf_pause(dev_priv);
307
308 return 0;
309}
310
311int via_driver_dma_quiescent(struct drm_device *dev)
312{
313 drm_via_private_t *dev_priv = dev->dev_private;
314
315 if (!via_wait_idle(dev_priv))
316 return -EBUSY;
317 return 0;
318}
319
320static int via_flush_ioctl(struct drm_device *dev, void *data, struct drm_file *file_priv)
321{
322
323 LOCK_TEST_WITH_RETURN(dev, file_priv);
324
325 return via_driver_dma_quiescent(dev);
326}
327
328static int via_cmdbuffer(struct drm_device *dev, void *data, struct drm_file *file_priv)
329{
330 drm_via_cmdbuffer_t *cmdbuf = data;
331 int ret;
332
333 LOCK_TEST_WITH_RETURN(dev, file_priv);
334
335 DRM_DEBUG("buf %p size %lu\n", cmdbuf->buf, cmdbuf->size);
336
337 ret = via_dispatch_cmdbuffer(dev, cmdbuf);
338 return ret;
339}
340
341static int via_dispatch_pci_cmdbuffer(struct drm_device *dev,
342 drm_via_cmdbuffer_t *cmd)
343{
344 drm_via_private_t *dev_priv = dev->dev_private;
345 int ret;
346
347 if (cmd->size > VIA_PCI_BUF_SIZE)
348 return -ENOMEM;
349 if (copy_from_user(dev_priv->pci_buf, cmd->buf, cmd->size))
350 return -EFAULT;
351
352 if ((ret =
353 via_verify_command_stream((uint32_t *) dev_priv->pci_buf,
354 cmd->size, dev, 0))) {
355 return ret;
356 }
357
358 ret =
359 via_parse_command_stream(dev, (const uint32_t *)dev_priv->pci_buf,
360 cmd->size);
361 return ret;
362}
363
364static int via_pci_cmdbuffer(struct drm_device *dev, void *data, struct drm_file *file_priv)
365{
366 drm_via_cmdbuffer_t *cmdbuf = data;
367 int ret;
368
369 LOCK_TEST_WITH_RETURN(dev, file_priv);
370
371 DRM_DEBUG("buf %p size %lu\n", cmdbuf->buf, cmdbuf->size);
372
373 ret = via_dispatch_pci_cmdbuffer(dev, cmdbuf);
374 return ret;
375}
376
377static inline uint32_t *via_align_buffer(drm_via_private_t *dev_priv,
378 uint32_t * vb, int qw_count)
379{
380 for (; qw_count > 0; --qw_count)
381 VIA_OUT_RING_QW(HC_DUMMY, HC_DUMMY);
382 return vb;
383}
384
385/*
386 * This function is used internally by ring buffer management code.
387 *
388 * Returns virtual pointer to ring buffer.
389 */
390static inline uint32_t *via_get_dma(drm_via_private_t *dev_priv)
391{
392 return (uint32_t *) (dev_priv->dma_ptr + dev_priv->dma_low);
393}
394
395/*
396 * Hooks a segment of data into the tail of the ring-buffer by
397 * modifying the pause address stored in the buffer itself. If
398 * the regulator has already paused, restart it.
399 */
400static int via_hook_segment(drm_via_private_t *dev_priv,
401 uint32_t pause_addr_hi, uint32_t pause_addr_lo,
402 int no_pci_fire)
403{
404 int paused, count;
405 volatile uint32_t *paused_at = dev_priv->last_pause_ptr;
406 uint32_t reader, ptr;
407 uint32_t diff;
408
409 paused = 0;
410 via_flush_write_combine();
411 (void) *(volatile uint32_t *)(via_get_dma(dev_priv) - 1);
412
413 *paused_at = pause_addr_lo;
414 via_flush_write_combine();
415 (void) *paused_at;
416
417 reader = *(dev_priv->hw_addr_ptr);
418 ptr = ((volatile char *)paused_at - dev_priv->dma_ptr) +
419 dev_priv->dma_offset + (uint32_t) dev_priv->agpAddr + 4;
420
421 dev_priv->last_pause_ptr = via_get_dma(dev_priv) - 1;
422
423 /*
424 * If there is a possibility that the command reader will
425 * miss the new pause address and pause on the old one,
426 * In that case we need to program the new start address
427 * using PCI.
428 */
429
430 diff = (uint32_t) (ptr - reader) - dev_priv->dma_diff;
431 count = 10000000;
432 while (diff == 0 && count--) {
433 paused = (VIA_READ(0x41c) & 0x80000000);
434 if (paused)
435 break;
436 reader = *(dev_priv->hw_addr_ptr);
437 diff = (uint32_t) (ptr - reader) - dev_priv->dma_diff;
438 }
439
440 paused = VIA_READ(0x41c) & 0x80000000;
441
442 if (paused && !no_pci_fire) {
443 reader = *(dev_priv->hw_addr_ptr);
444 diff = (uint32_t) (ptr - reader) - dev_priv->dma_diff;
445 diff &= (dev_priv->dma_high - 1);
446 if (diff != 0 && diff < (dev_priv->dma_high >> 1)) {
447 DRM_ERROR("Paused at incorrect address. "
448 "0x%08x, 0x%08x 0x%08x\n",
449 ptr, reader, dev_priv->dma_diff);
450 } else if (diff == 0) {
451 /*
452 * There is a concern that these writes may stall the PCI bus
453 * if the GPU is not idle. However, idling the GPU first
454 * doesn't make a difference.
455 */
456
457 VIA_WRITE(VIA_REG_TRANSET, (HC_ParaType_PreCR << 16));
458 VIA_WRITE(VIA_REG_TRANSPACE, pause_addr_hi);
459 VIA_WRITE(VIA_REG_TRANSPACE, pause_addr_lo);
460 VIA_READ(VIA_REG_TRANSPACE);
461 }
462 }
463 return paused;
464}
465
466static int via_wait_idle(drm_via_private_t *dev_priv)
467{
468 int count = 10000000;
469
470 while (!(VIA_READ(VIA_REG_STATUS) & VIA_VR_QUEUE_BUSY) && --count)
471 ;
472
473 while (count && (VIA_READ(VIA_REG_STATUS) &
474 (VIA_CMD_RGTR_BUSY | VIA_2D_ENG_BUSY |
475 VIA_3D_ENG_BUSY)))
476 --count;
477 return count;
478}
479
480static uint32_t *via_align_cmd(drm_via_private_t *dev_priv, uint32_t cmd_type,
481 uint32_t addr, uint32_t *cmd_addr_hi,
482 uint32_t *cmd_addr_lo, int skip_wait)
483{
484 uint32_t agp_base;
485 uint32_t cmd_addr, addr_lo, addr_hi;
486 uint32_t *vb;
487 uint32_t qw_pad_count;
488
489 if (!skip_wait)
490 via_cmdbuf_wait(dev_priv, 2 * CMDBUF_ALIGNMENT_SIZE);
491
492 vb = via_get_dma(dev_priv);
493 VIA_OUT_RING_QW(HC_HEADER2 | ((VIA_REG_TRANSET >> 2) << 12) |
494 (VIA_REG_TRANSPACE >> 2), HC_ParaType_PreCR << 16);
495 agp_base = dev_priv->dma_offset + (uint32_t) dev_priv->agpAddr;
496 qw_pad_count = (CMDBUF_ALIGNMENT_SIZE >> 3) -
497 ((dev_priv->dma_low & CMDBUF_ALIGNMENT_MASK) >> 3);
498
499 cmd_addr = (addr) ? addr :
500 agp_base + dev_priv->dma_low - 8 + (qw_pad_count << 3);
501 addr_lo = ((HC_SubA_HAGPBpL << 24) | (cmd_type & HC_HAGPBpID_MASK) |
502 (cmd_addr & HC_HAGPBpL_MASK));
503 addr_hi = ((HC_SubA_HAGPBpH << 24) | (cmd_addr >> 24));
504
505 vb = via_align_buffer(dev_priv, vb, qw_pad_count - 1);
506 VIA_OUT_RING_QW(*cmd_addr_hi = addr_hi, *cmd_addr_lo = addr_lo);
507 return vb;
508}
509
510static void via_cmdbuf_start(drm_via_private_t *dev_priv)
511{
512 uint32_t pause_addr_lo, pause_addr_hi;
513 uint32_t start_addr, start_addr_lo;
514 uint32_t end_addr, end_addr_lo;
515 uint32_t command;
516 uint32_t agp_base;
517 uint32_t ptr;
518 uint32_t reader;
519 int count;
520
521 dev_priv->dma_low = 0;
522
523 agp_base = dev_priv->dma_offset + (uint32_t) dev_priv->agpAddr;
524 start_addr = agp_base;
525 end_addr = agp_base + dev_priv->dma_high;
526
527 start_addr_lo = ((HC_SubA_HAGPBstL << 24) | (start_addr & 0xFFFFFF));
528 end_addr_lo = ((HC_SubA_HAGPBendL << 24) | (end_addr & 0xFFFFFF));
529 command = ((HC_SubA_HAGPCMNT << 24) | (start_addr >> 24) |
530 ((end_addr & 0xff000000) >> 16));
531
532 dev_priv->last_pause_ptr =
533 via_align_cmd(dev_priv, HC_HAGPBpID_PAUSE, 0,
534 &pause_addr_hi, &pause_addr_lo, 1) - 1;
535
536 via_flush_write_combine();
537 (void) *(volatile uint32_t *)dev_priv->last_pause_ptr;
538
539 VIA_WRITE(VIA_REG_TRANSET, (HC_ParaType_PreCR << 16));
540 VIA_WRITE(VIA_REG_TRANSPACE, command);
541 VIA_WRITE(VIA_REG_TRANSPACE, start_addr_lo);
542 VIA_WRITE(VIA_REG_TRANSPACE, end_addr_lo);
543
544 VIA_WRITE(VIA_REG_TRANSPACE, pause_addr_hi);
545 VIA_WRITE(VIA_REG_TRANSPACE, pause_addr_lo);
546 wmb();
547 VIA_WRITE(VIA_REG_TRANSPACE, command | HC_HAGPCMNT_MASK);
548 VIA_READ(VIA_REG_TRANSPACE);
549
550 dev_priv->dma_diff = 0;
551
552 count = 10000000;
553 while (!(VIA_READ(0x41c) & 0x80000000) && count--);
554
555 reader = *(dev_priv->hw_addr_ptr);
556 ptr = ((volatile char *)dev_priv->last_pause_ptr - dev_priv->dma_ptr) +
557 dev_priv->dma_offset + (uint32_t) dev_priv->agpAddr + 4;
558
559 /*
560 * This is the difference between where we tell the
561 * command reader to pause and where it actually pauses.
562 * This differs between hw implementation so we need to
563 * detect it.
564 */
565
566 dev_priv->dma_diff = ptr - reader;
567}
568
569static void via_pad_cache(drm_via_private_t *dev_priv, int qwords)
570{
571 uint32_t *vb;
572
573 via_cmdbuf_wait(dev_priv, qwords + 2);
574 vb = via_get_dma(dev_priv);
575 VIA_OUT_RING_QW(HC_HEADER2, HC_ParaType_NotTex << 16);
576 via_align_buffer(dev_priv, vb, qwords);
577}
578
579static inline void via_dummy_bitblt(drm_via_private_t *dev_priv)
580{
581 uint32_t *vb = via_get_dma(dev_priv);
582 SetReg2DAGP(0x0C, (0 | (0 << 16)));
583 SetReg2DAGP(0x10, 0 | (0 << 16));
584 SetReg2DAGP(0x0, 0x1 | 0x2000 | 0xAA000000);
585}
586
587static void via_cmdbuf_jump(drm_via_private_t *dev_priv)
588{
589 uint32_t agp_base;
590 uint32_t pause_addr_lo, pause_addr_hi;
591 uint32_t jump_addr_lo, jump_addr_hi;
592 volatile uint32_t *last_pause_ptr;
593 uint32_t dma_low_save1, dma_low_save2;
594
595 agp_base = dev_priv->dma_offset + (uint32_t) dev_priv->agpAddr;
596 via_align_cmd(dev_priv, HC_HAGPBpID_JUMP, 0, &jump_addr_hi,
597 &jump_addr_lo, 0);
598
599 dev_priv->dma_wrap = dev_priv->dma_low;
600
601 /*
602 * Wrap command buffer to the beginning.
603 */
604
605 dev_priv->dma_low = 0;
606 if (via_cmdbuf_wait(dev_priv, CMDBUF_ALIGNMENT_SIZE) != 0)
607 DRM_ERROR("via_cmdbuf_jump failed\n");
608
609 via_dummy_bitblt(dev_priv);
610 via_dummy_bitblt(dev_priv);
611
612 last_pause_ptr =
613 via_align_cmd(dev_priv, HC_HAGPBpID_PAUSE, 0, &pause_addr_hi,
614 &pause_addr_lo, 0) - 1;
615 via_align_cmd(dev_priv, HC_HAGPBpID_PAUSE, 0, &pause_addr_hi,
616 &pause_addr_lo, 0);
617
618 *last_pause_ptr = pause_addr_lo;
619 dma_low_save1 = dev_priv->dma_low;
620
621 /*
622 * Now, set a trap that will pause the regulator if it tries to rerun the old
623 * command buffer. (Which may happen if via_hook_segment detecs a command regulator pause
624 * and reissues the jump command over PCI, while the regulator has already taken the jump
625 * and actually paused at the current buffer end).
626 * There appears to be no other way to detect this condition, since the hw_addr_pointer
627 * does not seem to get updated immediately when a jump occurs.
628 */
629
630 last_pause_ptr =
631 via_align_cmd(dev_priv, HC_HAGPBpID_PAUSE, 0, &pause_addr_hi,
632 &pause_addr_lo, 0) - 1;
633 via_align_cmd(dev_priv, HC_HAGPBpID_PAUSE, 0, &pause_addr_hi,
634 &pause_addr_lo, 0);
635 *last_pause_ptr = pause_addr_lo;
636
637 dma_low_save2 = dev_priv->dma_low;
638 dev_priv->dma_low = dma_low_save1;
639 via_hook_segment(dev_priv, jump_addr_hi, jump_addr_lo, 0);
640 dev_priv->dma_low = dma_low_save2;
641 via_hook_segment(dev_priv, pause_addr_hi, pause_addr_lo, 0);
642}
643
644
645static void via_cmdbuf_rewind(drm_via_private_t *dev_priv)
646{
647 via_cmdbuf_jump(dev_priv);
648}
649
650static void via_cmdbuf_flush(drm_via_private_t *dev_priv, uint32_t cmd_type)
651{
652 uint32_t pause_addr_lo, pause_addr_hi;
653
654 via_align_cmd(dev_priv, cmd_type, 0, &pause_addr_hi, &pause_addr_lo, 0);
655 via_hook_segment(dev_priv, pause_addr_hi, pause_addr_lo, 0);
656}
657
658static void via_cmdbuf_pause(drm_via_private_t *dev_priv)
659{
660 via_cmdbuf_flush(dev_priv, HC_HAGPBpID_PAUSE);
661}
662
663static void via_cmdbuf_reset(drm_via_private_t *dev_priv)
664{
665 via_cmdbuf_flush(dev_priv, HC_HAGPBpID_STOP);
666 via_wait_idle(dev_priv);
667}
668
669/*
670 * User interface to the space and lag functions.
671 */
672
673static int via_cmdbuf_size(struct drm_device *dev, void *data, struct drm_file *file_priv)
674{
675 drm_via_cmdbuf_size_t *d_siz = data;
676 int ret = 0;
677 uint32_t tmp_size, count;
678 drm_via_private_t *dev_priv;
679
680 DRM_DEBUG("\n");
681 LOCK_TEST_WITH_RETURN(dev, file_priv);
682
683 dev_priv = (drm_via_private_t *) dev->dev_private;
684
685 if (dev_priv->ring.virtual_start == NULL) {
686 DRM_ERROR("called without initializing AGP ring buffer.\n");
687 return -EFAULT;
688 }
689
690 count = 1000000;
691 tmp_size = d_siz->size;
692 switch (d_siz->func) {
693 case VIA_CMDBUF_SPACE:
694 while (((tmp_size = via_cmdbuf_space(dev_priv)) < d_siz->size)
695 && --count) {
696 if (!d_siz->wait)
697 break;
698 }
699 if (!count) {
700 DRM_ERROR("VIA_CMDBUF_SPACE timed out.\n");
701 ret = -EAGAIN;
702 }
703 break;
704 case VIA_CMDBUF_LAG:
705 while (((tmp_size = via_cmdbuf_lag(dev_priv)) > d_siz->size)
706 && --count) {
707 if (!d_siz->wait)
708 break;
709 }
710 if (!count) {
711 DRM_ERROR("VIA_CMDBUF_LAG timed out.\n");
712 ret = -EAGAIN;
713 }
714 break;
715 default:
716 ret = -EFAULT;
717 }
718 d_siz->size = tmp_size;
719
720 return ret;
721}
722
723const struct drm_ioctl_desc via_ioctls[] = {
724 DRM_IOCTL_DEF_DRV(VIA_ALLOCMEM, via_mem_alloc, DRM_AUTH),
725 DRM_IOCTL_DEF_DRV(VIA_FREEMEM, via_mem_free, DRM_AUTH),
726 DRM_IOCTL_DEF_DRV(VIA_AGP_INIT, via_agp_init, DRM_AUTH|DRM_MASTER),
727 DRM_IOCTL_DEF_DRV(VIA_FB_INIT, via_fb_init, DRM_AUTH|DRM_MASTER),
728 DRM_IOCTL_DEF_DRV(VIA_MAP_INIT, via_map_init, DRM_AUTH|DRM_MASTER),
729 DRM_IOCTL_DEF_DRV(VIA_DEC_FUTEX, via_decoder_futex, DRM_AUTH),
730 DRM_IOCTL_DEF_DRV(VIA_DMA_INIT, via_dma_init, DRM_AUTH),
731 DRM_IOCTL_DEF_DRV(VIA_CMDBUFFER, via_cmdbuffer, DRM_AUTH),
732 DRM_IOCTL_DEF_DRV(VIA_FLUSH, via_flush_ioctl, DRM_AUTH),
733 DRM_IOCTL_DEF_DRV(VIA_PCICMD, via_pci_cmdbuffer, DRM_AUTH),
734 DRM_IOCTL_DEF_DRV(VIA_CMDBUF_SIZE, via_cmdbuf_size, DRM_AUTH),
735 DRM_IOCTL_DEF_DRV(VIA_WAIT_IRQ, via_wait_irq, DRM_AUTH),
736 DRM_IOCTL_DEF_DRV(VIA_DMA_BLIT, via_dma_blit, DRM_AUTH),
737 DRM_IOCTL_DEF_DRV(VIA_BLIT_SYNC, via_dma_blit_sync, DRM_AUTH)
738};
739
740int via_max_ioctl = ARRAY_SIZE(via_ioctls);