]> Git Repo - linux.git/blob - drivers/gpu/host1x/job.c
Merge branch 'work.uaccess-unaligned' of git://git.kernel.org/pub/scm/linux/kernel...
[linux.git] / drivers / gpu / host1x / job.c
1 /*
2  * Tegra host1x Job
3  *
4  * Copyright (c) 2010-2015, NVIDIA Corporation.
5  *
6  * This program is free software; you can redistribute it and/or modify it
7  * under the terms and conditions of the GNU General Public License,
8  * version 2, as published by the Free Software Foundation.
9  *
10  * This program is distributed in the hope it will be useful, but WITHOUT
11  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
13  * more details.
14  *
15  * You should have received a copy of the GNU General Public License
16  * along with this program.  If not, see <http://www.gnu.org/licenses/>.
17  */
18
19 #include <linux/dma-mapping.h>
20 #include <linux/err.h>
21 #include <linux/host1x.h>
22 #include <linux/kref.h>
23 #include <linux/module.h>
24 #include <linux/scatterlist.h>
25 #include <linux/slab.h>
26 #include <linux/vmalloc.h>
27 #include <trace/events/host1x.h>
28
29 #include "channel.h"
30 #include "dev.h"
31 #include "job.h"
32 #include "syncpt.h"
33
34 #define HOST1X_WAIT_SYNCPT_OFFSET 0x8
35
36 struct host1x_job *host1x_job_alloc(struct host1x_channel *ch,
37                                     u32 num_cmdbufs, u32 num_relocs,
38                                     u32 num_waitchks)
39 {
40         struct host1x_job *job = NULL;
41         unsigned int num_unpins = num_cmdbufs + num_relocs;
42         u64 total;
43         void *mem;
44
45         /* Check that we're not going to overflow */
46         total = sizeof(struct host1x_job) +
47                 (u64)num_relocs * sizeof(struct host1x_reloc) +
48                 (u64)num_unpins * sizeof(struct host1x_job_unpin_data) +
49                 (u64)num_waitchks * sizeof(struct host1x_waitchk) +
50                 (u64)num_cmdbufs * sizeof(struct host1x_job_gather) +
51                 (u64)num_unpins * sizeof(dma_addr_t) +
52                 (u64)num_unpins * sizeof(u32 *);
53         if (total > ULONG_MAX)
54                 return NULL;
55
56         mem = job = kzalloc(total, GFP_KERNEL);
57         if (!job)
58                 return NULL;
59
60         kref_init(&job->ref);
61         job->channel = ch;
62
63         /* Redistribute memory to the structs  */
64         mem += sizeof(struct host1x_job);
65         job->relocarray = num_relocs ? mem : NULL;
66         mem += num_relocs * sizeof(struct host1x_reloc);
67         job->unpins = num_unpins ? mem : NULL;
68         mem += num_unpins * sizeof(struct host1x_job_unpin_data);
69         job->waitchk = num_waitchks ? mem : NULL;
70         mem += num_waitchks * sizeof(struct host1x_waitchk);
71         job->gathers = num_cmdbufs ? mem : NULL;
72         mem += num_cmdbufs * sizeof(struct host1x_job_gather);
73         job->addr_phys = num_unpins ? mem : NULL;
74
75         job->reloc_addr_phys = job->addr_phys;
76         job->gather_addr_phys = &job->addr_phys[num_relocs];
77
78         return job;
79 }
80 EXPORT_SYMBOL(host1x_job_alloc);
81
82 struct host1x_job *host1x_job_get(struct host1x_job *job)
83 {
84         kref_get(&job->ref);
85         return job;
86 }
87 EXPORT_SYMBOL(host1x_job_get);
88
89 static void job_free(struct kref *ref)
90 {
91         struct host1x_job *job = container_of(ref, struct host1x_job, ref);
92
93         kfree(job);
94 }
95
96 void host1x_job_put(struct host1x_job *job)
97 {
98         kref_put(&job->ref, job_free);
99 }
100 EXPORT_SYMBOL(host1x_job_put);
101
102 void host1x_job_add_gather(struct host1x_job *job, struct host1x_bo *bo,
103                            u32 words, u32 offset)
104 {
105         struct host1x_job_gather *cur_gather = &job->gathers[job->num_gathers];
106
107         cur_gather->words = words;
108         cur_gather->bo = bo;
109         cur_gather->offset = offset;
110         job->num_gathers++;
111 }
112 EXPORT_SYMBOL(host1x_job_add_gather);
113
114 /*
115  * NULL an already satisfied WAIT_SYNCPT host method, by patching its
116  * args in the command stream. The method data is changed to reference
117  * a reserved (never given out or incr) HOST1X_SYNCPT_RESERVED syncpt
118  * with a matching threshold value of 0, so is guaranteed to be popped
119  * by the host HW.
120  */
121 static void host1x_syncpt_patch_offset(struct host1x_syncpt *sp,
122                                        struct host1x_bo *h, u32 offset)
123 {
124         void *patch_addr = NULL;
125
126         /* patch the wait */
127         patch_addr = host1x_bo_kmap(h, offset >> PAGE_SHIFT);
128         if (patch_addr) {
129                 host1x_syncpt_patch_wait(sp,
130                                          patch_addr + (offset & ~PAGE_MASK));
131                 host1x_bo_kunmap(h, offset >> PAGE_SHIFT, patch_addr);
132         } else
133                 pr_err("Could not map cmdbuf for wait check\n");
134 }
135
136 /*
137  * Check driver supplied waitchk structs for syncpt thresholds
138  * that have already been satisfied and NULL the comparison (to
139  * avoid a wrap condition in the HW).
140  */
141 static int do_waitchks(struct host1x_job *job, struct host1x *host,
142                        struct host1x_job_gather *g)
143 {
144         struct host1x_bo *patch = g->bo;
145         int i;
146
147         /* compare syncpt vs wait threshold */
148         for (i = 0; i < job->num_waitchk; i++) {
149                 struct host1x_waitchk *wait = &job->waitchk[i];
150                 struct host1x_syncpt *sp =
151                         host1x_syncpt_get(host, wait->syncpt_id);
152
153                 /* validate syncpt id */
154                 if (wait->syncpt_id > host1x_syncpt_nb_pts(host))
155                         continue;
156
157                 /* skip all other gathers */
158                 if (patch != wait->bo)
159                         continue;
160
161                 trace_host1x_syncpt_wait_check(wait->bo, wait->offset,
162                                                wait->syncpt_id, wait->thresh,
163                                                host1x_syncpt_read_min(sp));
164
165                 if (host1x_syncpt_is_expired(sp, wait->thresh)) {
166                         dev_dbg(host->dev,
167                                 "drop WAIT id %u (%s) thresh 0x%x, min 0x%x\n",
168                                 wait->syncpt_id, sp->name, wait->thresh,
169                                 host1x_syncpt_read_min(sp));
170
171                         host1x_syncpt_patch_offset(sp, patch,
172                                                    g->offset + wait->offset);
173                 }
174
175                 wait->bo = NULL;
176         }
177
178         return 0;
179 }
180
181 static unsigned int pin_job(struct host1x *host, struct host1x_job *job)
182 {
183         unsigned int i;
184         int err;
185
186         job->num_unpins = 0;
187
188         for (i = 0; i < job->num_relocs; i++) {
189                 struct host1x_reloc *reloc = &job->relocarray[i];
190                 struct sg_table *sgt;
191                 dma_addr_t phys_addr;
192
193                 reloc->target.bo = host1x_bo_get(reloc->target.bo);
194                 if (!reloc->target.bo) {
195                         err = -EINVAL;
196                         goto unpin;
197                 }
198
199                 phys_addr = host1x_bo_pin(reloc->target.bo, &sgt);
200                 if (!phys_addr) {
201                         err = -EINVAL;
202                         goto unpin;
203                 }
204
205                 job->addr_phys[job->num_unpins] = phys_addr;
206                 job->unpins[job->num_unpins].bo = reloc->target.bo;
207                 job->unpins[job->num_unpins].sgt = sgt;
208                 job->num_unpins++;
209         }
210
211         for (i = 0; i < job->num_gathers; i++) {
212                 struct host1x_job_gather *g = &job->gathers[i];
213                 size_t gather_size = 0;
214                 struct scatterlist *sg;
215                 struct sg_table *sgt;
216                 dma_addr_t phys_addr;
217                 unsigned long shift;
218                 struct iova *alloc;
219                 unsigned int j;
220
221                 g->bo = host1x_bo_get(g->bo);
222                 if (!g->bo) {
223                         err = -EINVAL;
224                         goto unpin;
225                 }
226
227                 phys_addr = host1x_bo_pin(g->bo, &sgt);
228                 if (!phys_addr) {
229                         err = -EINVAL;
230                         goto unpin;
231                 }
232
233                 if (!IS_ENABLED(CONFIG_TEGRA_HOST1X_FIREWALL) && host->domain) {
234                         for_each_sg(sgt->sgl, sg, sgt->nents, j)
235                                 gather_size += sg->length;
236                         gather_size = iova_align(&host->iova, gather_size);
237
238                         shift = iova_shift(&host->iova);
239                         alloc = alloc_iova(&host->iova, gather_size >> shift,
240                                            host->iova_end >> shift, true);
241                         if (!alloc) {
242                                 err = -ENOMEM;
243                                 goto unpin;
244                         }
245
246                         err = iommu_map_sg(host->domain,
247                                         iova_dma_addr(&host->iova, alloc),
248                                         sgt->sgl, sgt->nents, IOMMU_READ);
249                         if (err == 0) {
250                                 __free_iova(&host->iova, alloc);
251                                 err = -EINVAL;
252                                 goto unpin;
253                         }
254
255                         job->addr_phys[job->num_unpins] =
256                                 iova_dma_addr(&host->iova, alloc);
257                         job->unpins[job->num_unpins].size = gather_size;
258                 } else {
259                         job->addr_phys[job->num_unpins] = phys_addr;
260                 }
261
262                 job->gather_addr_phys[i] = job->addr_phys[job->num_unpins];
263
264                 job->unpins[job->num_unpins].bo = g->bo;
265                 job->unpins[job->num_unpins].sgt = sgt;
266                 job->num_unpins++;
267         }
268
269         return 0;
270
271 unpin:
272         host1x_job_unpin(job);
273         return err;
274 }
275
276 static int do_relocs(struct host1x_job *job, struct host1x_job_gather *g)
277 {
278         int i = 0;
279         u32 last_page = ~0;
280         void *cmdbuf_page_addr = NULL;
281         struct host1x_bo *cmdbuf = g->bo;
282
283         /* pin & patch the relocs for one gather */
284         for (i = 0; i < job->num_relocs; i++) {
285                 struct host1x_reloc *reloc = &job->relocarray[i];
286                 u32 reloc_addr = (job->reloc_addr_phys[i] +
287                                   reloc->target.offset) >> reloc->shift;
288                 u32 *target;
289
290                 /* skip all other gathers */
291                 if (cmdbuf != reloc->cmdbuf.bo)
292                         continue;
293
294                 if (IS_ENABLED(CONFIG_TEGRA_HOST1X_FIREWALL)) {
295                         target = (u32 *)job->gather_copy_mapped +
296                                         reloc->cmdbuf.offset / sizeof(u32) +
297                                                 g->offset / sizeof(u32);
298                         goto patch_reloc;
299                 }
300
301                 if (last_page != reloc->cmdbuf.offset >> PAGE_SHIFT) {
302                         if (cmdbuf_page_addr)
303                                 host1x_bo_kunmap(cmdbuf, last_page,
304                                                  cmdbuf_page_addr);
305
306                         cmdbuf_page_addr = host1x_bo_kmap(cmdbuf,
307                                         reloc->cmdbuf.offset >> PAGE_SHIFT);
308                         last_page = reloc->cmdbuf.offset >> PAGE_SHIFT;
309
310                         if (unlikely(!cmdbuf_page_addr)) {
311                                 pr_err("Could not map cmdbuf for relocation\n");
312                                 return -ENOMEM;
313                         }
314                 }
315
316                 target = cmdbuf_page_addr + (reloc->cmdbuf.offset & ~PAGE_MASK);
317 patch_reloc:
318                 *target = reloc_addr;
319         }
320
321         if (cmdbuf_page_addr)
322                 host1x_bo_kunmap(cmdbuf, last_page, cmdbuf_page_addr);
323
324         return 0;
325 }
326
327 static bool check_reloc(struct host1x_reloc *reloc, struct host1x_bo *cmdbuf,
328                         unsigned int offset)
329 {
330         offset *= sizeof(u32);
331
332         if (reloc->cmdbuf.bo != cmdbuf || reloc->cmdbuf.offset != offset)
333                 return false;
334
335         /* relocation shift value validation isn't implemented yet */
336         if (reloc->shift)
337                 return false;
338
339         return true;
340 }
341
342 static bool check_wait(struct host1x_waitchk *wait, struct host1x_bo *cmdbuf,
343                        unsigned int offset)
344 {
345         offset *= sizeof(u32);
346
347         if (wait->bo != cmdbuf || wait->offset != offset)
348                 return false;
349
350         return true;
351 }
352
353 struct host1x_firewall {
354         struct host1x_job *job;
355         struct device *dev;
356
357         unsigned int num_relocs;
358         struct host1x_reloc *reloc;
359
360         unsigned int num_waitchks;
361         struct host1x_waitchk *waitchk;
362
363         struct host1x_bo *cmdbuf;
364         unsigned int offset;
365
366         u32 words;
367         u32 class;
368         u32 reg;
369         u32 mask;
370         u32 count;
371 };
372
373 static int check_register(struct host1x_firewall *fw, unsigned long offset)
374 {
375         if (!fw->job->is_addr_reg)
376                 return 0;
377
378         if (fw->job->is_addr_reg(fw->dev, fw->class, offset)) {
379                 if (!fw->num_relocs)
380                         return -EINVAL;
381
382                 if (!check_reloc(fw->reloc, fw->cmdbuf, fw->offset))
383                         return -EINVAL;
384
385                 fw->num_relocs--;
386                 fw->reloc++;
387         }
388
389         if (offset == HOST1X_WAIT_SYNCPT_OFFSET) {
390                 if (fw->class != HOST1X_CLASS_HOST1X)
391                         return -EINVAL;
392
393                 if (!fw->num_waitchks)
394                         return -EINVAL;
395
396                 if (!check_wait(fw->waitchk, fw->cmdbuf, fw->offset))
397                         return -EINVAL;
398
399                 fw->num_waitchks--;
400                 fw->waitchk++;
401         }
402
403         return 0;
404 }
405
406 static int check_class(struct host1x_firewall *fw, u32 class)
407 {
408         if (!fw->job->is_valid_class) {
409                 if (fw->class != class)
410                         return -EINVAL;
411         } else {
412                 if (!fw->job->is_valid_class(fw->class))
413                         return -EINVAL;
414         }
415
416         return 0;
417 }
418
419 static int check_mask(struct host1x_firewall *fw)
420 {
421         u32 mask = fw->mask;
422         u32 reg = fw->reg;
423         int ret;
424
425         while (mask) {
426                 if (fw->words == 0)
427                         return -EINVAL;
428
429                 if (mask & 1) {
430                         ret = check_register(fw, reg);
431                         if (ret < 0)
432                                 return ret;
433
434                         fw->words--;
435                         fw->offset++;
436                 }
437                 mask >>= 1;
438                 reg++;
439         }
440
441         return 0;
442 }
443
444 static int check_incr(struct host1x_firewall *fw)
445 {
446         u32 count = fw->count;
447         u32 reg = fw->reg;
448         int ret;
449
450         while (count) {
451                 if (fw->words == 0)
452                         return -EINVAL;
453
454                 ret = check_register(fw, reg);
455                 if (ret < 0)
456                         return ret;
457
458                 reg++;
459                 fw->words--;
460                 fw->offset++;
461                 count--;
462         }
463
464         return 0;
465 }
466
467 static int check_nonincr(struct host1x_firewall *fw)
468 {
469         u32 count = fw->count;
470         int ret;
471
472         while (count) {
473                 if (fw->words == 0)
474                         return -EINVAL;
475
476                 ret = check_register(fw, fw->reg);
477                 if (ret < 0)
478                         return ret;
479
480                 fw->words--;
481                 fw->offset++;
482                 count--;
483         }
484
485         return 0;
486 }
487
488 static int validate(struct host1x_firewall *fw, struct host1x_job_gather *g)
489 {
490         u32 *cmdbuf_base = (u32 *)fw->job->gather_copy_mapped +
491                 (g->offset / sizeof(u32));
492         u32 job_class = fw->class;
493         int err = 0;
494
495         fw->words = g->words;
496         fw->cmdbuf = g->bo;
497         fw->offset = 0;
498
499         while (fw->words && !err) {
500                 u32 word = cmdbuf_base[fw->offset];
501                 u32 opcode = (word & 0xf0000000) >> 28;
502
503                 fw->mask = 0;
504                 fw->reg = 0;
505                 fw->count = 0;
506                 fw->words--;
507                 fw->offset++;
508
509                 switch (opcode) {
510                 case 0:
511                         fw->class = word >> 6 & 0x3ff;
512                         fw->mask = word & 0x3f;
513                         fw->reg = word >> 16 & 0xfff;
514                         err = check_class(fw, job_class);
515                         if (!err)
516                                 err = check_mask(fw);
517                         if (err)
518                                 goto out;
519                         break;
520                 case 1:
521                         fw->reg = word >> 16 & 0xfff;
522                         fw->count = word & 0xffff;
523                         err = check_incr(fw);
524                         if (err)
525                                 goto out;
526                         break;
527
528                 case 2:
529                         fw->reg = word >> 16 & 0xfff;
530                         fw->count = word & 0xffff;
531                         err = check_nonincr(fw);
532                         if (err)
533                                 goto out;
534                         break;
535
536                 case 3:
537                         fw->mask = word & 0xffff;
538                         fw->reg = word >> 16 & 0xfff;
539                         err = check_mask(fw);
540                         if (err)
541                                 goto out;
542                         break;
543                 case 4:
544                 case 14:
545                         break;
546                 default:
547                         err = -EINVAL;
548                         break;
549                 }
550         }
551
552 out:
553         return err;
554 }
555
556 static inline int copy_gathers(struct host1x_job *job, struct device *dev)
557 {
558         struct host1x_firewall fw;
559         size_t size = 0;
560         size_t offset = 0;
561         int i;
562
563         fw.job = job;
564         fw.dev = dev;
565         fw.reloc = job->relocarray;
566         fw.num_relocs = job->num_relocs;
567         fw.waitchk = job->waitchk;
568         fw.num_waitchks = job->num_waitchk;
569         fw.class = job->class;
570
571         for (i = 0; i < job->num_gathers; i++) {
572                 struct host1x_job_gather *g = &job->gathers[i];
573
574                 size += g->words * sizeof(u32);
575         }
576
577         /*
578          * Try a non-blocking allocation from a higher priority pools first,
579          * as awaiting for the allocation here is a major performance hit.
580          */
581         job->gather_copy_mapped = dma_alloc_wc(dev, size, &job->gather_copy,
582                                                GFP_NOWAIT);
583
584         /* the higher priority allocation failed, try the generic-blocking */
585         if (!job->gather_copy_mapped)
586                 job->gather_copy_mapped = dma_alloc_wc(dev, size,
587                                                        &job->gather_copy,
588                                                        GFP_KERNEL);
589         if (!job->gather_copy_mapped)
590                 return -ENOMEM;
591
592         job->gather_copy_size = size;
593
594         for (i = 0; i < job->num_gathers; i++) {
595                 struct host1x_job_gather *g = &job->gathers[i];
596                 void *gather;
597
598                 /* Copy the gather */
599                 gather = host1x_bo_mmap(g->bo);
600                 memcpy(job->gather_copy_mapped + offset, gather + g->offset,
601                        g->words * sizeof(u32));
602                 host1x_bo_munmap(g->bo, gather);
603
604                 /* Store the location in the buffer */
605                 g->base = job->gather_copy;
606                 g->offset = offset;
607
608                 /* Validate the job */
609                 if (validate(&fw, g))
610                         return -EINVAL;
611
612                 offset += g->words * sizeof(u32);
613         }
614
615         /* No relocs and waitchks should remain at this point */
616         if (fw.num_relocs || fw.num_waitchks)
617                 return -EINVAL;
618
619         return 0;
620 }
621
622 int host1x_job_pin(struct host1x_job *job, struct device *dev)
623 {
624         int err;
625         unsigned int i, j;
626         struct host1x *host = dev_get_drvdata(dev->parent);
627         DECLARE_BITMAP(waitchk_mask, host1x_syncpt_nb_pts(host));
628
629         bitmap_zero(waitchk_mask, host1x_syncpt_nb_pts(host));
630         for (i = 0; i < job->num_waitchk; i++) {
631                 u32 syncpt_id = job->waitchk[i].syncpt_id;
632
633                 if (syncpt_id < host1x_syncpt_nb_pts(host))
634                         set_bit(syncpt_id, waitchk_mask);
635         }
636
637         /* get current syncpt values for waitchk */
638         for_each_set_bit(i, waitchk_mask, host1x_syncpt_nb_pts(host))
639                 host1x_syncpt_load(host->syncpt + i);
640
641         /* pin memory */
642         err = pin_job(host, job);
643         if (err)
644                 goto out;
645
646         if (IS_ENABLED(CONFIG_TEGRA_HOST1X_FIREWALL)) {
647                 err = copy_gathers(job, dev);
648                 if (err)
649                         goto out;
650         }
651
652         /* patch gathers */
653         for (i = 0; i < job->num_gathers; i++) {
654                 struct host1x_job_gather *g = &job->gathers[i];
655
656                 /* process each gather mem only once */
657                 if (g->handled)
658                         continue;
659
660                 /* copy_gathers() sets gathers base if firewall is enabled */
661                 if (!IS_ENABLED(CONFIG_TEGRA_HOST1X_FIREWALL))
662                         g->base = job->gather_addr_phys[i];
663
664                 for (j = i + 1; j < job->num_gathers; j++) {
665                         if (job->gathers[j].bo == g->bo) {
666                                 job->gathers[j].handled = true;
667                                 job->gathers[j].base = g->base;
668                         }
669                 }
670
671                 err = do_relocs(job, g);
672                 if (err)
673                         break;
674
675                 err = do_waitchks(job, host, g);
676                 if (err)
677                         break;
678         }
679
680 out:
681         if (err)
682                 host1x_job_unpin(job);
683         wmb();
684
685         return err;
686 }
687 EXPORT_SYMBOL(host1x_job_pin);
688
689 void host1x_job_unpin(struct host1x_job *job)
690 {
691         struct host1x *host = dev_get_drvdata(job->channel->dev->parent);
692         unsigned int i;
693
694         for (i = 0; i < job->num_unpins; i++) {
695                 struct host1x_job_unpin_data *unpin = &job->unpins[i];
696
697                 if (!IS_ENABLED(CONFIG_TEGRA_HOST1X_FIREWALL) && host->domain) {
698                         iommu_unmap(host->domain, job->addr_phys[i],
699                                     unpin->size);
700                         free_iova(&host->iova,
701                                 iova_pfn(&host->iova, job->addr_phys[i]));
702                 }
703
704                 host1x_bo_unpin(unpin->bo, unpin->sgt);
705                 host1x_bo_put(unpin->bo);
706         }
707
708         job->num_unpins = 0;
709
710         if (job->gather_copy_size)
711                 dma_free_wc(job->channel->dev, job->gather_copy_size,
712                             job->gather_copy_mapped, job->gather_copy);
713 }
714 EXPORT_SYMBOL(host1x_job_unpin);
715
716 /*
717  * Debug routine used to dump job entries
718  */
719 void host1x_job_dump(struct device *dev, struct host1x_job *job)
720 {
721         dev_dbg(dev, "    SYNCPT_ID   %d\n", job->syncpt_id);
722         dev_dbg(dev, "    SYNCPT_VAL  %d\n", job->syncpt_end);
723         dev_dbg(dev, "    FIRST_GET   0x%x\n", job->first_get);
724         dev_dbg(dev, "    TIMEOUT     %d\n", job->timeout);
725         dev_dbg(dev, "    NUM_SLOTS   %d\n", job->num_slots);
726         dev_dbg(dev, "    NUM_HANDLES %d\n", job->num_unpins);
727 }
This page took 0.069576 seconds and 4 git commands to generate.