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 <linux/delay.h>
38#include <linux/uaccess.h>
39
40#include <drm/drm.h>
41#include <drm/drm_agpsupport.h>
42#include <drm/drm_device.h>
43#include <drm/drm_file.h>
44#include <drm/via_drm.h>
45
46#include "via_drv.h"
47#include "via_3d_reg.h"
48
49#define CMDBUF_ALIGNMENT_SIZE (0x100)
50#define CMDBUF_ALIGNMENT_MASK (0x0ff)
51
52/* defines for VIA 3D registers */
53#define VIA_REG_STATUS 0x400
54#define VIA_REG_TRANSET 0x43C
55#define VIA_REG_TRANSPACE 0x440
56
57/* VIA_REG_STATUS(0x400): Engine Status */
58#define VIA_CMD_RGTR_BUSY 0x00000080 /* Command Regulator is busy */
59#define VIA_2D_ENG_BUSY 0x00000001 /* 2D Engine is busy */
60#define VIA_3D_ENG_BUSY 0x00000002 /* 3D Engine is busy */
61#define VIA_VR_QUEUE_BUSY 0x00020000 /* Virtual Queue is busy */
62
63#define SetReg2DAGP(nReg, nData) { \
64 *((uint32_t *)(vb)) = ((nReg) >> 2) | HALCYON_HEADER1; \
65 *((uint32_t *)(vb) + 1) = (nData); \
66 vb = ((uint32_t *)vb) + 2; \
67 dev_priv->dma_low += 8; \
68}
69
70#define via_flush_write_combine() mb()
71
72#define VIA_OUT_RING_QW(w1, w2) do { \
73 *vb++ = (w1); \
74 *vb++ = (w2); \
75 dev_priv->dma_low += 8; \
76} while (0)
77
78static void via_cmdbuf_start(drm_via_private_t *dev_priv);
79static void via_cmdbuf_pause(drm_via_private_t *dev_priv);
80static void via_cmdbuf_reset(drm_via_private_t *dev_priv);
81static void via_cmdbuf_rewind(drm_via_private_t *dev_priv);
82static int via_wait_idle(drm_via_private_t *dev_priv);
83static void via_pad_cache(drm_via_private_t *dev_priv, int qwords);
84
85/*
86 * Free space in command buffer.
87 */
88
89static uint32_t via_cmdbuf_space(drm_via_private_t *dev_priv)
90{
91 uint32_t agp_base = dev_priv->dma_offset + (uint32_t) dev_priv->agpAddr;
92 uint32_t hw_addr = *(dev_priv->hw_addr_ptr) - agp_base;
93
94 return ((hw_addr <= dev_priv->dma_low) ?
95 (dev_priv->dma_high + hw_addr - dev_priv->dma_low) :
96 (hw_addr - dev_priv->dma_low));
97}
98
99/*
100 * How much does the command regulator lag behind?
101 */
102
103static uint32_t via_cmdbuf_lag(drm_via_private_t *dev_priv)
104{
105 uint32_t agp_base = dev_priv->dma_offset + (uint32_t) dev_priv->agpAddr;
106 uint32_t hw_addr = *(dev_priv->hw_addr_ptr) - agp_base;
107
108 return ((hw_addr <= dev_priv->dma_low) ?
109 (dev_priv->dma_low - hw_addr) :
110 (dev_priv->dma_wrap + dev_priv->dma_low - hw_addr));
111}
112
113/*
114 * Check that the given size fits in the buffer, otherwise wait.
115 */
116
117static inline int
118via_cmdbuf_wait(drm_via_private_t *dev_priv, unsigned int size)
119{
120 uint32_t agp_base = dev_priv->dma_offset + (uint32_t) dev_priv->agpAddr;
121 uint32_t cur_addr, hw_addr, next_addr;
122 volatile uint32_t *hw_addr_ptr;
123 uint32_t count;
124 hw_addr_ptr = dev_priv->hw_addr_ptr;
125 cur_addr = dev_priv->dma_low;
126 next_addr = cur_addr + size + 512 * 1024;
127 count = 1000000;
128 do {
129 hw_addr = *hw_addr_ptr - agp_base;
130 if (count-- == 0) {
131 DRM_ERROR
132 ("via_cmdbuf_wait timed out hw %x cur_addr %x next_addr %x\n",
133 hw_addr, cur_addr, next_addr);
134 return -1;
135 }
136 if ((cur_addr < hw_addr) && (next_addr >= hw_addr))
137 msleep(1);
138 } while ((cur_addr < hw_addr) && (next_addr >= hw_addr));
139 return 0;
140}
141
142/*
143 * Checks whether buffer head has reach the end. Rewind the ring buffer
144 * when necessary.
145 *
146 * Returns virtual pointer to ring buffer.
147 */
148
149static inline uint32_t *via_check_dma(drm_via_private_t * dev_priv,
150 unsigned int size)
151{
152 if ((dev_priv->dma_low + size + 4 * CMDBUF_ALIGNMENT_SIZE) >
153 dev_priv->dma_high) {
154 via_cmdbuf_rewind(dev_priv);
155 }
156 if (via_cmdbuf_wait(dev_priv, size) != 0)
157 return NULL;
158
159 return (uint32_t *) (dev_priv->dma_ptr + dev_priv->dma_low);
160}
161
162int via_dma_cleanup(struct drm_device *dev)
163{
164 if (dev->dev_private) {
165 drm_via_private_t *dev_priv =
166 (drm_via_private_t *) dev->dev_private;
167
168 if (dev_priv->ring.virtual_start) {
169 via_cmdbuf_reset(dev_priv);
170
171 drm_legacy_ioremapfree(&dev_priv->ring.map, dev);
172 dev_priv->ring.virtual_start = NULL;
173 }
174
175 }
176
177 return 0;
178}
179
180static int via_initialize(struct drm_device *dev,
181 drm_via_private_t *dev_priv,
182 drm_via_dma_init_t *init)
183{
184 if (!dev_priv || !dev_priv->mmio) {
185 DRM_ERROR("via_dma_init called before via_map_init\n");
186 return -EFAULT;
187 }
188
189 if (dev_priv->ring.virtual_start != NULL) {
190 DRM_ERROR("called again without calling cleanup\n");
191 return -EFAULT;
192 }
193
194 if (!dev->agp || !dev->agp->base) {
195 DRM_ERROR("called with no agp memory available\n");
196 return -EFAULT;
197 }
198
199 if (dev_priv->chipset == VIA_DX9_0) {
200 DRM_ERROR("AGP DMA is not supported on this chip\n");
201 return -EINVAL;
202 }
203
204 dev_priv->ring.map.offset = dev->agp->base + init->offset;
205 dev_priv->ring.map.size = init->size;
206 dev_priv->ring.map.type = 0;
207 dev_priv->ring.map.flags = 0;
208 dev_priv->ring.map.mtrr = 0;
209
210 drm_legacy_ioremap(&dev_priv->ring.map, dev);
211
212 if (dev_priv->ring.map.handle == NULL) {
213 via_dma_cleanup(dev);
214 DRM_ERROR("can not ioremap virtual address for"
215 " ring buffer\n");
216 return -ENOMEM;
217 }
218
219 dev_priv->ring.virtual_start = dev_priv->ring.map.handle;
220
221 dev_priv->dma_ptr = dev_priv->ring.virtual_start;
222 dev_priv->dma_low = 0;
223 dev_priv->dma_high = init->size;
224 dev_priv->dma_wrap = init->size;
225 dev_priv->dma_offset = init->offset;
226 dev_priv->last_pause_ptr = NULL;
227 dev_priv->hw_addr_ptr =
228 (volatile uint32_t *)((char *)dev_priv->mmio->handle +
229 init->reg_pause_addr);
230
231 via_cmdbuf_start(dev_priv);
232
233 return 0;
234}
235
236static int via_dma_init(struct drm_device *dev, void *data, struct drm_file *file_priv)
237{
238 drm_via_private_t *dev_priv = (drm_via_private_t *) dev->dev_private;
239 drm_via_dma_init_t *init = data;
240 int retcode = 0;
241
242 switch (init->func) {
243 case VIA_INIT_DMA:
244 if (!capable(CAP_SYS_ADMIN))
245 retcode = -EPERM;
246 else
247 retcode = via_initialize(dev, dev_priv, init);
248 break;
249 case VIA_CLEANUP_DMA:
250 if (!capable(CAP_SYS_ADMIN))
251 retcode = -EPERM;
252 else
253 retcode = via_dma_cleanup(dev);
254 break;
255 case VIA_DMA_INITIALIZED:
256 retcode = (dev_priv->ring.virtual_start != NULL) ?
257 0 : -EFAULT;
258 break;
259 default:
260 retcode = -EINVAL;
261 break;
262 }
263
264 return retcode;
265}
266
267static int via_dispatch_cmdbuffer(struct drm_device *dev, drm_via_cmdbuffer_t *cmd)
268{
269 drm_via_private_t *dev_priv;
270 uint32_t *vb;
271 int ret;
272
273 dev_priv = (drm_via_private_t *) dev->dev_private;
274
275 if (dev_priv->ring.virtual_start == NULL) {
276 DRM_ERROR("called without initializing AGP ring buffer.\n");
277 return -EFAULT;
278 }
279
280 if (cmd->size > VIA_PCI_BUF_SIZE)
281 return -ENOMEM;
282
283 if (copy_from_user(dev_priv->pci_buf, cmd->buf, cmd->size))
284 return -EFAULT;
285
286 /*
287 * Running this function on AGP memory is dead slow. Therefore
288 * we run it on a temporary cacheable system memory buffer and
289 * copy it to AGP memory when ready.
290 */
291
292 if ((ret =
293 via_verify_command_stream((uint32_t *) dev_priv->pci_buf,
294 cmd->size, dev, 1))) {
295 return ret;
296 }
297
298 vb = via_check_dma(dev_priv, (cmd->size < 0x100) ? 0x102 : cmd->size);
299 if (vb == NULL)
300 return -EAGAIN;
301
302 memcpy(vb, dev_priv->pci_buf, cmd->size);
303
304 dev_priv->dma_low += cmd->size;
305
306 /*
307 * Small submissions somehow stalls the CPU. (AGP cache effects?)
308 * pad to greater size.
309 */
310
311 if (cmd->size < 0x100)
312 via_pad_cache(dev_priv, (0x100 - cmd->size) >> 3);
313 via_cmdbuf_pause(dev_priv);
314
315 return 0;
316}
317
318int via_driver_dma_quiescent(struct drm_device *dev)
319{
320 drm_via_private_t *dev_priv = dev->dev_private;
321
322 if (!via_wait_idle(dev_priv))
323 return -EBUSY;
324 return 0;
325}
326
327static int via_flush_ioctl(struct drm_device *dev, void *data, struct drm_file *file_priv)
328{
329
330 LOCK_TEST_WITH_RETURN(dev, file_priv);
331
332 return via_driver_dma_quiescent(dev);
333}
334
335static int via_cmdbuffer(struct drm_device *dev, void *data, struct drm_file *file_priv)
336{
337 drm_via_cmdbuffer_t *cmdbuf = data;
338 int ret;
339
340 LOCK_TEST_WITH_RETURN(dev, file_priv);
341
342 DRM_DEBUG("buf %p size %lu\n", cmdbuf->buf, cmdbuf->size);
343
344 ret = via_dispatch_cmdbuffer(dev, cmdbuf);
345 return ret;
346}
347
348static int via_dispatch_pci_cmdbuffer(struct drm_device *dev,
349 drm_via_cmdbuffer_t *cmd)
350{
351 drm_via_private_t *dev_priv = dev->dev_private;
352 int ret;
353
354 if (cmd->size > VIA_PCI_BUF_SIZE)
355 return -ENOMEM;
356 if (copy_from_user(dev_priv->pci_buf, cmd->buf, cmd->size))
357 return -EFAULT;
358
359 if ((ret =
360 via_verify_command_stream((uint32_t *) dev_priv->pci_buf,
361 cmd->size, dev, 0))) {
362 return ret;
363 }
364
365 ret =
366 via_parse_command_stream(dev, (const uint32_t *)dev_priv->pci_buf,
367 cmd->size);
368 return ret;
369}
370
371static int via_pci_cmdbuffer(struct drm_device *dev, void *data, struct drm_file *file_priv)
372{
373 drm_via_cmdbuffer_t *cmdbuf = data;
374 int ret;
375
376 LOCK_TEST_WITH_RETURN(dev, file_priv);
377
378 DRM_DEBUG("buf %p size %lu\n", cmdbuf->buf, cmdbuf->size);
379
380 ret = via_dispatch_pci_cmdbuffer(dev, cmdbuf);
381 return ret;
382}
383
384static inline uint32_t *via_align_buffer(drm_via_private_t *dev_priv,
385 uint32_t * vb, int qw_count)
386{
387 for (; qw_count > 0; --qw_count)
388 VIA_OUT_RING_QW(HC_DUMMY, HC_DUMMY);
389 return vb;
390}
391
392/*
393 * This function is used internally by ring buffer management code.
394 *
395 * Returns virtual pointer to ring buffer.
396 */
397static inline uint32_t *via_get_dma(drm_via_private_t *dev_priv)
398{
399 return (uint32_t *) (dev_priv->dma_ptr + dev_priv->dma_low);
400}
401
402/*
403 * Hooks a segment of data into the tail of the ring-buffer by
404 * modifying the pause address stored in the buffer itself. If
405 * the regulator has already paused, restart it.
406 */
407static int via_hook_segment(drm_via_private_t *dev_priv,
408 uint32_t pause_addr_hi, uint32_t pause_addr_lo,
409 int no_pci_fire)
410{
411 int paused, count;
412 volatile uint32_t *paused_at = dev_priv->last_pause_ptr;
413 uint32_t reader, ptr;
414 uint32_t diff;
415
416 paused = 0;
417 via_flush_write_combine();
418 (void) *(volatile uint32_t *)(via_get_dma(dev_priv) - 1);
419
420 *paused_at = pause_addr_lo;
421 via_flush_write_combine();
422 (void) *paused_at;
423
424 reader = *(dev_priv->hw_addr_ptr);
425 ptr = ((volatile char *)paused_at - dev_priv->dma_ptr) +
426 dev_priv->dma_offset + (uint32_t) dev_priv->agpAddr + 4;
427
428 dev_priv->last_pause_ptr = via_get_dma(dev_priv) - 1;
429
430 /*
431 * If there is a possibility that the command reader will
432 * miss the new pause address and pause on the old one,
433 * In that case we need to program the new start address
434 * using PCI.
435 */
436
437 diff = (uint32_t) (ptr - reader) - dev_priv->dma_diff;
438 count = 10000000;
439 while (diff == 0 && count--) {
440 paused = (via_read(dev_priv, 0x41c) & 0x80000000);
441 if (paused)
442 break;
443 reader = *(dev_priv->hw_addr_ptr);
444 diff = (uint32_t) (ptr - reader) - dev_priv->dma_diff;
445 }
446
447 paused = via_read(dev_priv, 0x41c) & 0x80000000;
448
449 if (paused && !no_pci_fire) {
450 reader = *(dev_priv->hw_addr_ptr);
451 diff = (uint32_t) (ptr - reader) - dev_priv->dma_diff;
452 diff &= (dev_priv->dma_high - 1);
453 if (diff != 0 && diff < (dev_priv->dma_high >> 1)) {
454 DRM_ERROR("Paused at incorrect address. "
455 "0x%08x, 0x%08x 0x%08x\n",
456 ptr, reader, dev_priv->dma_diff);
457 } else if (diff == 0) {
458 /*
459 * There is a concern that these writes may stall the PCI bus
460 * if the GPU is not idle. However, idling the GPU first
461 * doesn't make a difference.
462 */
463
464 via_write(dev_priv, VIA_REG_TRANSET, (HC_ParaType_PreCR << 16));
465 via_write(dev_priv, VIA_REG_TRANSPACE, pause_addr_hi);
466 via_write(dev_priv, VIA_REG_TRANSPACE, pause_addr_lo);
467 via_read(dev_priv, VIA_REG_TRANSPACE);
468 }
469 }
470 return paused;
471}
472
473static int via_wait_idle(drm_via_private_t *dev_priv)
474{
475 int count = 10000000;
476
477 while (!(via_read(dev_priv, VIA_REG_STATUS) & VIA_VR_QUEUE_BUSY) && --count)
478 ;
479
480 while (count && (via_read(dev_priv, VIA_REG_STATUS) &
481 (VIA_CMD_RGTR_BUSY | VIA_2D_ENG_BUSY |
482 VIA_3D_ENG_BUSY)))
483 --count;
484 return count;
485}
486
487static uint32_t *via_align_cmd(drm_via_private_t *dev_priv, uint32_t cmd_type,
488 uint32_t addr, uint32_t *cmd_addr_hi,
489 uint32_t *cmd_addr_lo, int skip_wait)
490{
491 uint32_t agp_base;
492 uint32_t cmd_addr, addr_lo, addr_hi;
493 uint32_t *vb;
494 uint32_t qw_pad_count;
495
496 if (!skip_wait)
497 via_cmdbuf_wait(dev_priv, 2 * CMDBUF_ALIGNMENT_SIZE);
498
499 vb = via_get_dma(dev_priv);
500 VIA_OUT_RING_QW(HC_HEADER2 | ((VIA_REG_TRANSET >> 2) << 12) |
501 (VIA_REG_TRANSPACE >> 2), HC_ParaType_PreCR << 16);
502 agp_base = dev_priv->dma_offset + (uint32_t) dev_priv->agpAddr;
503 qw_pad_count = (CMDBUF_ALIGNMENT_SIZE >> 3) -
504 ((dev_priv->dma_low & CMDBUF_ALIGNMENT_MASK) >> 3);
505
506 cmd_addr = (addr) ? addr :
507 agp_base + dev_priv->dma_low - 8 + (qw_pad_count << 3);
508 addr_lo = ((HC_SubA_HAGPBpL << 24) | (cmd_type & HC_HAGPBpID_MASK) |
509 (cmd_addr & HC_HAGPBpL_MASK));
510 addr_hi = ((HC_SubA_HAGPBpH << 24) | (cmd_addr >> 24));
511
512 vb = via_align_buffer(dev_priv, vb, qw_pad_count - 1);
513 VIA_OUT_RING_QW(*cmd_addr_hi = addr_hi, *cmd_addr_lo = addr_lo);
514 return vb;
515}
516
517static void via_cmdbuf_start(drm_via_private_t *dev_priv)
518{
519 uint32_t pause_addr_lo, pause_addr_hi;
520 uint32_t start_addr, start_addr_lo;
521 uint32_t end_addr, end_addr_lo;
522 uint32_t command;
523 uint32_t agp_base;
524 uint32_t ptr;
525 uint32_t reader;
526 int count;
527
528 dev_priv->dma_low = 0;
529
530 agp_base = dev_priv->dma_offset + (uint32_t) dev_priv->agpAddr;
531 start_addr = agp_base;
532 end_addr = agp_base + dev_priv->dma_high;
533
534 start_addr_lo = ((HC_SubA_HAGPBstL << 24) | (start_addr & 0xFFFFFF));
535 end_addr_lo = ((HC_SubA_HAGPBendL << 24) | (end_addr & 0xFFFFFF));
536 command = ((HC_SubA_HAGPCMNT << 24) | (start_addr >> 24) |
537 ((end_addr & 0xff000000) >> 16));
538
539 dev_priv->last_pause_ptr =
540 via_align_cmd(dev_priv, HC_HAGPBpID_PAUSE, 0,
541 &pause_addr_hi, &pause_addr_lo, 1) - 1;
542
543 via_flush_write_combine();
544 (void) *(volatile uint32_t *)dev_priv->last_pause_ptr;
545
546 via_write(dev_priv, VIA_REG_TRANSET, (HC_ParaType_PreCR << 16));
547 via_write(dev_priv, VIA_REG_TRANSPACE, command);
548 via_write(dev_priv, VIA_REG_TRANSPACE, start_addr_lo);
549 via_write(dev_priv, VIA_REG_TRANSPACE, end_addr_lo);
550
551 via_write(dev_priv, VIA_REG_TRANSPACE, pause_addr_hi);
552 via_write(dev_priv, VIA_REG_TRANSPACE, pause_addr_lo);
553 wmb();
554 via_write(dev_priv, VIA_REG_TRANSPACE, command | HC_HAGPCMNT_MASK);
555 via_read(dev_priv, VIA_REG_TRANSPACE);
556
557 dev_priv->dma_diff = 0;
558
559 count = 10000000;
560 while (!(via_read(dev_priv, 0x41c) & 0x80000000) && count--);
561
562 reader = *(dev_priv->hw_addr_ptr);
563 ptr = ((volatile char *)dev_priv->last_pause_ptr - dev_priv->dma_ptr) +
564 dev_priv->dma_offset + (uint32_t) dev_priv->agpAddr + 4;
565
566 /*
567 * This is the difference between where we tell the
568 * command reader to pause and where it actually pauses.
569 * This differs between hw implementation so we need to
570 * detect it.
571 */
572
573 dev_priv->dma_diff = ptr - reader;
574}
575
576static void via_pad_cache(drm_via_private_t *dev_priv, int qwords)
577{
578 uint32_t *vb;
579
580 via_cmdbuf_wait(dev_priv, qwords + 2);
581 vb = via_get_dma(dev_priv);
582 VIA_OUT_RING_QW(HC_HEADER2, HC_ParaType_NotTex << 16);
583 via_align_buffer(dev_priv, vb, qwords);
584}
585
586static inline void via_dummy_bitblt(drm_via_private_t *dev_priv)
587{
588 uint32_t *vb = via_get_dma(dev_priv);
589 SetReg2DAGP(0x0C, (0 | (0 << 16)));
590 SetReg2DAGP(0x10, 0 | (0 << 16));
591 SetReg2DAGP(0x0, 0x1 | 0x2000 | 0xAA000000);
592}
593
594static void via_cmdbuf_jump(drm_via_private_t *dev_priv)
595{
596 uint32_t agp_base;
597 uint32_t pause_addr_lo, pause_addr_hi;
598 uint32_t jump_addr_lo, jump_addr_hi;
599 volatile uint32_t *last_pause_ptr;
600 uint32_t dma_low_save1, dma_low_save2;
601
602 agp_base = dev_priv->dma_offset + (uint32_t) dev_priv->agpAddr;
603 via_align_cmd(dev_priv, HC_HAGPBpID_JUMP, 0, &jump_addr_hi,
604 &jump_addr_lo, 0);
605
606 dev_priv->dma_wrap = dev_priv->dma_low;
607
608 /*
609 * Wrap command buffer to the beginning.
610 */
611
612 dev_priv->dma_low = 0;
613 if (via_cmdbuf_wait(dev_priv, CMDBUF_ALIGNMENT_SIZE) != 0)
614 DRM_ERROR("via_cmdbuf_jump failed\n");
615
616 via_dummy_bitblt(dev_priv);
617 via_dummy_bitblt(dev_priv);
618
619 last_pause_ptr =
620 via_align_cmd(dev_priv, HC_HAGPBpID_PAUSE, 0, &pause_addr_hi,
621 &pause_addr_lo, 0) - 1;
622 via_align_cmd(dev_priv, HC_HAGPBpID_PAUSE, 0, &pause_addr_hi,
623 &pause_addr_lo, 0);
624
625 *last_pause_ptr = pause_addr_lo;
626 dma_low_save1 = dev_priv->dma_low;
627
628 /*
629 * Now, set a trap that will pause the regulator if it tries to rerun the old
630 * command buffer. (Which may happen if via_hook_segment detecs a command regulator pause
631 * and reissues the jump command over PCI, while the regulator has already taken the jump
632 * and actually paused at the current buffer end).
633 * There appears to be no other way to detect this condition, since the hw_addr_pointer
634 * does not seem to get updated immediately when a jump occurs.
635 */
636
637 last_pause_ptr =
638 via_align_cmd(dev_priv, HC_HAGPBpID_PAUSE, 0, &pause_addr_hi,
639 &pause_addr_lo, 0) - 1;
640 via_align_cmd(dev_priv, HC_HAGPBpID_PAUSE, 0, &pause_addr_hi,
641 &pause_addr_lo, 0);
642 *last_pause_ptr = pause_addr_lo;
643
644 dma_low_save2 = dev_priv->dma_low;
645 dev_priv->dma_low = dma_low_save1;
646 via_hook_segment(dev_priv, jump_addr_hi, jump_addr_lo, 0);
647 dev_priv->dma_low = dma_low_save2;
648 via_hook_segment(dev_priv, pause_addr_hi, pause_addr_lo, 0);
649}
650
651
652static void via_cmdbuf_rewind(drm_via_private_t *dev_priv)
653{
654 via_cmdbuf_jump(dev_priv);
655}
656
657static void via_cmdbuf_flush(drm_via_private_t *dev_priv, uint32_t cmd_type)
658{
659 uint32_t pause_addr_lo, pause_addr_hi;
660
661 via_align_cmd(dev_priv, cmd_type, 0, &pause_addr_hi, &pause_addr_lo, 0);
662 via_hook_segment(dev_priv, pause_addr_hi, pause_addr_lo, 0);
663}
664
665static void via_cmdbuf_pause(drm_via_private_t *dev_priv)
666{
667 via_cmdbuf_flush(dev_priv, HC_HAGPBpID_PAUSE);
668}
669
670static void via_cmdbuf_reset(drm_via_private_t *dev_priv)
671{
672 via_cmdbuf_flush(dev_priv, HC_HAGPBpID_STOP);
673 via_wait_idle(dev_priv);
674}
675
676/*
677 * User interface to the space and lag functions.
678 */
679
680static int via_cmdbuf_size(struct drm_device *dev, void *data, struct drm_file *file_priv)
681{
682 drm_via_cmdbuf_size_t *d_siz = data;
683 int ret = 0;
684 uint32_t tmp_size, count;
685 drm_via_private_t *dev_priv;
686
687 DRM_DEBUG("\n");
688 LOCK_TEST_WITH_RETURN(dev, file_priv);
689
690 dev_priv = (drm_via_private_t *) dev->dev_private;
691
692 if (dev_priv->ring.virtual_start == NULL) {
693 DRM_ERROR("called without initializing AGP ring buffer.\n");
694 return -EFAULT;
695 }
696
697 count = 1000000;
698 tmp_size = d_siz->size;
699 switch (d_siz->func) {
700 case VIA_CMDBUF_SPACE:
701 while (((tmp_size = via_cmdbuf_space(dev_priv)) < d_siz->size)
702 && --count) {
703 if (!d_siz->wait)
704 break;
705 }
706 if (!count) {
707 DRM_ERROR("VIA_CMDBUF_SPACE timed out.\n");
708 ret = -EAGAIN;
709 }
710 break;
711 case VIA_CMDBUF_LAG:
712 while (((tmp_size = via_cmdbuf_lag(dev_priv)) > d_siz->size)
713 && --count) {
714 if (!d_siz->wait)
715 break;
716 }
717 if (!count) {
718 DRM_ERROR("VIA_CMDBUF_LAG timed out.\n");
719 ret = -EAGAIN;
720 }
721 break;
722 default:
723 ret = -EFAULT;
724 }
725 d_siz->size = tmp_size;
726
727 return ret;
728}
729
730const struct drm_ioctl_desc via_ioctls[] = {
731 DRM_IOCTL_DEF_DRV(VIA_ALLOCMEM, via_mem_alloc, DRM_AUTH),
732 DRM_IOCTL_DEF_DRV(VIA_FREEMEM, via_mem_free, DRM_AUTH),
733 DRM_IOCTL_DEF_DRV(VIA_AGP_INIT, via_agp_init, DRM_AUTH|DRM_MASTER),
734 DRM_IOCTL_DEF_DRV(VIA_FB_INIT, via_fb_init, DRM_AUTH|DRM_MASTER),
735 DRM_IOCTL_DEF_DRV(VIA_MAP_INIT, via_map_init, DRM_AUTH|DRM_MASTER),
736 DRM_IOCTL_DEF_DRV(VIA_DEC_FUTEX, via_decoder_futex, DRM_AUTH),
737 DRM_IOCTL_DEF_DRV(VIA_DMA_INIT, via_dma_init, DRM_AUTH),
738 DRM_IOCTL_DEF_DRV(VIA_CMDBUFFER, via_cmdbuffer, DRM_AUTH),
739 DRM_IOCTL_DEF_DRV(VIA_FLUSH, via_flush_ioctl, DRM_AUTH),
740 DRM_IOCTL_DEF_DRV(VIA_PCICMD, via_pci_cmdbuffer, DRM_AUTH),
741 DRM_IOCTL_DEF_DRV(VIA_CMDBUF_SIZE, via_cmdbuf_size, DRM_AUTH),
742 DRM_IOCTL_DEF_DRV(VIA_WAIT_IRQ, via_wait_irq, DRM_AUTH),
743 DRM_IOCTL_DEF_DRV(VIA_DMA_BLIT, via_dma_blit, DRM_AUTH),
744 DRM_IOCTL_DEF_DRV(VIA_BLIT_SYNC, via_dma_blit_sync, DRM_AUTH)
745};
746
747int via_max_ioctl = ARRAY_SIZE(via_ioctls);