1 /* $NetBSD: nouveau_nvkm_subdev_instmem_gk20a.c,v 1.10 2024/06/04 21:43:39 riastradh Exp $ */
2
3 /*
4 * Copyright (c) 2015, NVIDIA CORPORATION. All rights reserved.
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10 * and/or sell copies of the Software, and to permit persons to whom the
11 * Software is furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in
14 * all copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
21 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
22 * DEALINGS IN THE SOFTWARE.
23 */
24
25 /*
26 * GK20A does not have dedicated video memory, and to accurately represent this
27 * fact Nouveau will not create a RAM device for it. Therefore its instmem
28 * implementation must be done directly on top of system memory, while
29 * preserving coherency for read and write operations.
30 *
31 * Instmem can be allocated through two means:
32 * 1) If an IOMMU unit has been probed, the IOMMU API is used to make memory
33 * pages contiguous to the GPU. This is the preferred way.
34 * 2) If no IOMMU unit is probed, the DMA API is used to allocate physically
35 * contiguous memory.
36 *
37 * In both cases CPU read and writes are performed by creating a write-combined
38 * mapping. The GPU L2 cache must thus be flushed/invalidated when required. To
39 * be conservative we do this every time we acquire or release an instobj, but
40 * ideally L2 management should be handled at a higher level.
41 *
42 * To improve performance, CPU mappings are not removed upon instobj release.
43 * Instead they are placed into a LRU list to be recycled when the mapped space
44 * goes beyond a certain threshold. At the moment this limit is 1MB.
45 */
46 #include <sys/cdefs.h>
47 __KERNEL_RCSID(0, "$NetBSD: nouveau_nvkm_subdev_instmem_gk20a.c,v 1.10 2024/06/04 21:43:39 riastradh Exp $");
48
49 #include "priv.h"
50
51 #include <core/memory.h>
52 #include <core/tegra.h>
53 #include <subdev/ltc.h>
54 #include <subdev/mmu.h>
55
56 #include <linux/nbsd-namespace.h>
57
58 #ifdef __NetBSD__
59 # define __iomem __nvkm_memory_iomem
60 #endif
61
62 struct gk20a_instobj {
63 struct nvkm_memory memory;
64 struct nvkm_mm_node *mn;
65 struct gk20a_instmem *imem;
66
67 /* CPU mapping */
68 u32 *vaddr;
69 };
70 #define gk20a_instobj(p) container_of((p), struct gk20a_instobj, memory)
71
72 #ifndef __NetBSD__
73 /*
74 * Used for objects allocated using the DMA API
75 */
76 struct gk20a_instobj_dma {
77 struct gk20a_instobj base;
78
79 dma_addr_t handle;
80 struct nvkm_mm_node r;
81 };
82 #define gk20a_instobj_dma(p) \
83 container_of(gk20a_instobj(p), struct gk20a_instobj_dma, base)
84 #endif
85
86 /*
87 * Used for objects flattened using the IOMMU API
88 */
89 struct gk20a_instobj_iommu {
90 struct gk20a_instobj base;
91
92 /* to link into gk20a_instmem::vaddr_lru */
93 struct list_head vaddr_node;
94 /* how many clients are using vaddr? */
95 u32 use_cpt;
96
97 #ifdef __NetBSD__
98 struct nvkm_mm_node mm_node; /* XXX */
99 bus_dmamap_t map;
100 int nsegs;
101 bus_dma_segment_t segs[];
102 #else
103 /* will point to the higher half of pages */
104 dma_addr_t *dma_addrs;
105 /* array of base.mem->size pages (+ dma_addr_ts) */
106 struct page *pages[];
107 #endif
108 };
109 #define gk20a_instobj_iommu(p) \
110 container_of(gk20a_instobj(p), struct gk20a_instobj_iommu, base)
111
112 struct gk20a_instmem {
113 struct nvkm_instmem base;
114
115 /* protects vaddr_* and gk20a_instobj::vaddr* */
116 struct mutex lock;
117
118 /* CPU mappings LRU */
119 unsigned int vaddr_use;
120 unsigned int vaddr_max;
121 struct list_head vaddr_lru;
122
123 #ifdef __NetBSD__
124 bus_dma_tag_t dmat;
125 #else
126 /* Only used if IOMMU if present */
127 struct mutex *mm_mutex;
128 struct nvkm_mm *mm;
129 struct iommu_domain *domain;
130 unsigned long iommu_pgshift;
131 u16 iommu_bit;
132
133 /* Only used by DMA API */
134 unsigned long attrs;
135 #endif
136 };
137 #define gk20a_instmem(p) container_of((p), struct gk20a_instmem, base)
138
139 static enum nvkm_memory_target
gk20a_instobj_target(struct nvkm_memory * memory)140 gk20a_instobj_target(struct nvkm_memory *memory)
141 {
142 return NVKM_MEM_TARGET_NCOH;
143 }
144
145 static u8
gk20a_instobj_page(struct nvkm_memory * memory)146 gk20a_instobj_page(struct nvkm_memory *memory)
147 {
148 return 12;
149 }
150
151 static u64
gk20a_instobj_addr(struct nvkm_memory * memory)152 gk20a_instobj_addr(struct nvkm_memory *memory)
153 {
154 return (u64)gk20a_instobj(memory)->mn->offset << 12;
155 }
156
157 static u64
gk20a_instobj_size(struct nvkm_memory * memory)158 gk20a_instobj_size(struct nvkm_memory *memory)
159 {
160 return (u64)gk20a_instobj(memory)->mn->length << 12;
161 }
162
163 /*
164 * Recycle the vaddr of obj. Must be called with gk20a_instmem::lock held.
165 */
166 static void
gk20a_instobj_iommu_recycle_vaddr(struct gk20a_instobj_iommu * obj)167 gk20a_instobj_iommu_recycle_vaddr(struct gk20a_instobj_iommu *obj)
168 {
169 struct gk20a_instmem *imem = obj->base.imem;
170 /* there should not be any user left... */
171 WARN_ON(obj->use_cpt);
172 list_del(&obj->vaddr_node);
173 #ifdef __NetBSD__
174 bus_size_t size = nvkm_memory_size(&obj->base.memory);
175 bus_dmamem_unmap(imem->dmat, obj->base.vaddr, size);
176 #else
177 vunmap(obj->base.vaddr);
178 #endif
179 obj->base.vaddr = NULL;
180 imem->vaddr_use -= nvkm_memory_size(&obj->base.memory);
181 nvkm_debug(&imem->base.subdev, "vaddr used: %x/%x\n", imem->vaddr_use,
182 imem->vaddr_max);
183 }
184
185
186 /*
187 * Must be called while holding gk20a_instmem::lock
188 */
189 static void
gk20a_instmem_vaddr_gc(struct gk20a_instmem * imem,const u64 size)190 gk20a_instmem_vaddr_gc(struct gk20a_instmem *imem, const u64 size)
191 {
192 while (imem->vaddr_use + size > imem->vaddr_max) {
193 /* no candidate that can be unmapped, abort... */
194 if (list_empty(&imem->vaddr_lru))
195 break;
196
197 gk20a_instobj_iommu_recycle_vaddr(
198 list_first_entry(&imem->vaddr_lru,
199 struct gk20a_instobj_iommu, vaddr_node));
200 }
201 }
202
203 #ifndef __NetBSD__
204 static void __iomem *
gk20a_instobj_acquire_dma(struct nvkm_memory * memory)205 gk20a_instobj_acquire_dma(struct nvkm_memory *memory)
206 {
207 struct gk20a_instobj *node = gk20a_instobj(memory);
208 struct gk20a_instmem *imem = node->imem;
209 struct nvkm_ltc *ltc = imem->base.subdev.device->ltc;
210
211 nvkm_ltc_flush(ltc);
212
213 return node->vaddr;
214 }
215 #endif
216
217 static void __iomem *
gk20a_instobj_acquire_iommu(struct nvkm_memory * memory)218 gk20a_instobj_acquire_iommu(struct nvkm_memory *memory)
219 {
220 struct gk20a_instobj_iommu *node = gk20a_instobj_iommu(memory);
221 struct gk20a_instmem *imem = node->base.imem;
222 struct nvkm_ltc *ltc = imem->base.subdev.device->ltc;
223 const u64 size = nvkm_memory_size(memory);
224
225 nvkm_ltc_flush(ltc);
226
227 mutex_lock(&imem->lock);
228
229 if (node->base.vaddr) {
230 if (!node->use_cpt) {
231 /* remove from LRU list since mapping in use again */
232 list_del(&node->vaddr_node);
233 }
234 goto out;
235 }
236
237 /* try to free some address space if we reached the limit */
238 gk20a_instmem_vaddr_gc(imem, size);
239
240 /* map the pages */
241 #ifdef __NetBSD__
242 void *kva;
243 if (bus_dmamem_map(imem->dmat, node->segs, node->nsegs, size,
244 &kva, BUS_DMA_WAITOK|BUS_DMA_PREFETCHABLE))
245 node->base.vaddr = NULL;
246 else
247 node->base.vaddr = kva;
248 #else
249 node->base.vaddr = vmap(node->pages, size >> PAGE_SHIFT, VM_MAP,
250 pgprot_writecombine(PAGE_KERNEL));
251 #endif
252 if (!node->base.vaddr) {
253 nvkm_error(&imem->base.subdev, "cannot map instobj - "
254 "this is not going to end well...\n");
255 goto out;
256 }
257
258 imem->vaddr_use += size;
259 nvkm_debug(&imem->base.subdev, "vaddr used: %x/%x\n",
260 imem->vaddr_use, imem->vaddr_max);
261
262 out:
263 node->use_cpt++;
264 mutex_unlock(&imem->lock);
265
266 return node->base.vaddr;
267 }
268
269 #ifndef __NetBSD__
270 static void
gk20a_instobj_release_dma(struct nvkm_memory * memory)271 gk20a_instobj_release_dma(struct nvkm_memory *memory)
272 {
273 struct gk20a_instobj *node = gk20a_instobj(memory);
274 struct gk20a_instmem *imem = node->imem;
275 struct nvkm_ltc *ltc = imem->base.subdev.device->ltc;
276
277 /* in case we got a write-combined mapping */
278 wmb();
279 nvkm_ltc_invalidate(ltc);
280 }
281 #endif
282
283 static void
gk20a_instobj_release_iommu(struct nvkm_memory * memory)284 gk20a_instobj_release_iommu(struct nvkm_memory *memory)
285 {
286 struct gk20a_instobj_iommu *node = gk20a_instobj_iommu(memory);
287 struct gk20a_instmem *imem = node->base.imem;
288 struct nvkm_ltc *ltc = imem->base.subdev.device->ltc;
289
290 mutex_lock(&imem->lock);
291
292 /* we should at least have one user to release... */
293 if (WARN_ON(node->use_cpt == 0))
294 goto out;
295
296 /* add unused objs to the LRU list to recycle their mapping */
297 if (--node->use_cpt == 0)
298 list_add_tail(&node->vaddr_node, &imem->vaddr_lru);
299
300 out:
301 mutex_unlock(&imem->lock);
302
303 wmb();
304 nvkm_ltc_invalidate(ltc);
305 }
306
307 static u32
gk20a_instobj_rd32(struct nvkm_memory * memory,u64 offset)308 gk20a_instobj_rd32(struct nvkm_memory *memory, u64 offset)
309 {
310 struct gk20a_instobj *node = gk20a_instobj(memory);
311
312 return node->vaddr[offset / 4];
313 }
314
315 static void
gk20a_instobj_wr32(struct nvkm_memory * memory,u64 offset,u32 data)316 gk20a_instobj_wr32(struct nvkm_memory *memory, u64 offset, u32 data)
317 {
318 struct gk20a_instobj *node = gk20a_instobj(memory);
319
320 node->vaddr[offset / 4] = data;
321 }
322
323 static int
gk20a_instobj_map(struct nvkm_memory * memory,u64 offset,struct nvkm_vmm * vmm,struct nvkm_vma * vma,void * argv,u32 argc)324 gk20a_instobj_map(struct nvkm_memory *memory, u64 offset, struct nvkm_vmm *vmm,
325 struct nvkm_vma *vma, void *argv, u32 argc)
326 {
327 struct gk20a_instobj *node = gk20a_instobj(memory);
328 struct nvkm_vmm_map map = {
329 .memory = &node->memory,
330 .offset = offset,
331 .mem = node->mn,
332 };
333
334 return nvkm_vmm_map(vmm, vma, argv, argc, &map);
335 }
336
337 #ifndef __NetBSD__
338 static void *
gk20a_instobj_dtor_dma(struct nvkm_memory * memory)339 gk20a_instobj_dtor_dma(struct nvkm_memory *memory)
340 {
341 struct gk20a_instobj_dma *node = gk20a_instobj_dma(memory);
342 struct gk20a_instmem *imem = node->base.imem;
343 struct device *dev = imem->base.subdev.device->dev;
344
345 if (unlikely(!node->base.vaddr))
346 goto out;
347
348 dma_free_attrs(dev, (u64)node->base.mn->length << PAGE_SHIFT,
349 node->base.vaddr, node->handle, imem->attrs);
350
351 out:
352 return node;
353 }
354 #endif
355
356 static void *
gk20a_instobj_dtor_iommu(struct nvkm_memory * memory)357 gk20a_instobj_dtor_iommu(struct nvkm_memory *memory)
358 {
359 struct gk20a_instobj_iommu *node = gk20a_instobj_iommu(memory);
360 struct gk20a_instmem *imem = node->base.imem;
361 struct device *dev = imem->base.subdev.device->dev;
362 struct nvkm_mm_node *r = node->base.mn;
363 int i;
364
365 if (unlikely(!r))
366 goto out;
367
368 mutex_lock(&imem->lock);
369
370 /* vaddr has already been recycled */
371 if (node->base.vaddr)
372 gk20a_instobj_iommu_recycle_vaddr(node);
373
374 mutex_unlock(&imem->lock);
375
376 #ifdef __NetBSD__
377 __USE(i);
378 __USE(dev);
379 bus_dmamap_unload(imem->dmat, node->map);
380 bus_dmamap_destroy(imem->dmat, node->map);
381 bus_dmamem_free(imem->dmat, node->segs, node->nsegs);
382 #else
383 /* clear IOMMU bit to unmap pages */
384 r->offset &= ~BIT(imem->iommu_bit - imem->iommu_pgshift);
385
386 /* Unmap pages from GPU address space and free them */
387 for (i = 0; i < node->base.mn->length; i++) {
388 iommu_unmap(imem->domain,
389 (r->offset + i) << imem->iommu_pgshift, PAGE_SIZE);
390 dma_unmap_page(dev, node->dma_addrs[i], PAGE_SIZE,
391 DMA_BIDIRECTIONAL);
392 __free_page(node->pages[i]);
393 }
394
395 /* Release area from GPU address space */
396 mutex_lock(imem->mm_mutex);
397 nvkm_mm_free(imem->mm, &r);
398 mutex_unlock(imem->mm_mutex);
399 #endif
400
401 out:
402 return node;
403 }
404
405 #ifndef __NetBSD__
406 static const struct nvkm_memory_func
407 gk20a_instobj_func_dma = {
408 .dtor = gk20a_instobj_dtor_dma,
409 .target = gk20a_instobj_target,
410 .page = gk20a_instobj_page,
411 .addr = gk20a_instobj_addr,
412 .size = gk20a_instobj_size,
413 .acquire = gk20a_instobj_acquire_dma,
414 .release = gk20a_instobj_release_dma,
415 .map = gk20a_instobj_map,
416 };
417 #endif
418
419 static const struct nvkm_memory_func
420 gk20a_instobj_func_iommu = {
421 .dtor = gk20a_instobj_dtor_iommu,
422 .target = gk20a_instobj_target,
423 .page = gk20a_instobj_page,
424 .addr = gk20a_instobj_addr,
425 .size = gk20a_instobj_size,
426 .acquire = gk20a_instobj_acquire_iommu,
427 .release = gk20a_instobj_release_iommu,
428 .map = gk20a_instobj_map,
429 };
430
431 static const struct nvkm_memory_ptrs
432 gk20a_instobj_ptrs = {
433 .rd32 = gk20a_instobj_rd32,
434 .wr32 = gk20a_instobj_wr32,
435 };
436
437 #ifndef __NetBSD__
438 static int
gk20a_instobj_ctor_dma(struct gk20a_instmem * imem,u32 npages,u32 align,struct gk20a_instobj ** _node)439 gk20a_instobj_ctor_dma(struct gk20a_instmem *imem, u32 npages, u32 align,
440 struct gk20a_instobj **_node)
441 {
442 struct gk20a_instobj_dma *node;
443 struct nvkm_subdev *subdev = &imem->base.subdev;
444 struct device *dev = subdev->device->dev;
445
446 if (!(node = kzalloc(sizeof(*node), GFP_KERNEL)))
447 return -ENOMEM;
448 *_node = &node->base;
449
450 nvkm_memory_ctor(&gk20a_instobj_func_dma, &node->base.memory);
451 node->base.memory.ptrs = &gk20a_instobj_ptrs;
452
453 node->base.vaddr = dma_alloc_attrs(dev, npages << PAGE_SHIFT,
454 &node->handle, GFP_KERNEL,
455 imem->attrs);
456 if (!node->base.vaddr) {
457 nvkm_error(subdev, "cannot allocate DMA memory\n");
458 return -ENOMEM;
459 }
460
461 /* alignment check */
462 if (unlikely(node->handle & (align - 1)))
463 nvkm_warn(subdev,
464 "memory not aligned as requested: %pad (0x%x)\n",
465 &node->handle, align);
466
467 /* present memory for being mapped using small pages */
468 node->r.type = 12;
469 node->r.offset = node->handle >> 12;
470 node->r.length = (npages << PAGE_SHIFT) >> 12;
471
472 node->base.mn = &node->r;
473 return 0;
474 }
475 #endif
476
477 static int
gk20a_instobj_ctor_iommu(struct gk20a_instmem * imem,u32 npages,u32 align,struct gk20a_instobj ** _node)478 gk20a_instobj_ctor_iommu(struct gk20a_instmem *imem, u32 npages, u32 align,
479 struct gk20a_instobj **_node)
480 {
481 struct gk20a_instobj_iommu *node;
482 struct nvkm_subdev *subdev = &imem->base.subdev;
483 struct device *dev = subdev->device->dev;
484 struct nvkm_mm_node *r;
485 int ret;
486 int i;
487
488 /*
489 * despite their variable size, instmem allocations are small enough
490 * (< 1 page) to be handled by kzalloc
491 */
492 #ifdef __NetBSD__
493 node = kzalloc(struct_size(node, segs, npages), GFP_KERNEL);
494 if (node == NULL)
495 return -ENOMEM;
496 #else
497 if (!(node = kzalloc(sizeof(*node) + ((sizeof(node->pages[0]) +
498 sizeof(*node->dma_addrs)) * npages), GFP_KERNEL)))
499 return -ENOMEM;
500 #endif
501 *_node = &node->base;
502 #ifndef __NetBSD__
503 node->dma_addrs = (void *)(node->pages + npages);
504 #endif
505
506 nvkm_memory_ctor(&gk20a_instobj_func_iommu, &node->base.memory);
507 node->base.memory.ptrs = &gk20a_instobj_ptrs;
508
509 #ifdef __NetBSD__
510 bus_size_t nbytes = (bus_size_t)npages << PAGE_SHIFT;
511 __USE(i);
512 __USE(r);
513 __USE(dev);
514 /* XXX errno NetBSD->Linux */
515 ret = -bus_dmamem_alloc(imem->dmat, nbytes, PAGE_SIZE,
516 PAGE_SIZE, node->segs, npages, &node->nsegs, BUS_DMA_WAITOK);
517 if (ret)
518 fail0: goto out;
519 /* XXX errno NetBSD->Linux */
520 ret = -bus_dmamap_create(imem->dmat, nbytes, 1, nbytes, PAGE_SIZE,
521 BUS_DMA_WAITOK, &node->map);
522 if (ret) {
523 fail1: bus_dmamem_free(imem->dmat, node->segs, node->nsegs);
524 goto fail0;
525 }
526 /* XXX errno NetBSD->Linux */
527 ret = -bus_dmamap_load_raw(imem->dmat, node->map, node->segs,
528 node->nsegs, nbytes, BUS_DMA_WAITOK);
529 if (ret) {
530 fail2: __unused
531 bus_dmamap_destroy(imem->dmat, node->map);
532 goto fail1;
533 }
534 node->mm_node.type = 12; /* XXX ??? */
535 node->mm_node.offset = node->map->dm_segs[0].ds_addr;
536 node->mm_node.length = node->map->dm_segs[0].ds_len;
537 node->base.mn = &node->mm_node;
538 out:
539 #else
540 /* Allocate backing memory */
541 for (i = 0; i < npages; i++) {
542 struct page *p = alloc_page(GFP_KERNEL);
543 dma_addr_t dma_adr;
544
545 if (p == NULL) {
546 ret = -ENOMEM;
547 goto free_pages;
548 }
549 node->pages[i] = p;
550 dma_adr = dma_map_page(dev, p, 0, PAGE_SIZE, DMA_BIDIRECTIONAL);
551 if (dma_mapping_error(dev, dma_adr)) {
552 nvkm_error(subdev, "DMA mapping error!\n");
553 ret = -ENOMEM;
554 goto free_pages;
555 }
556 node->dma_addrs[i] = dma_adr;
557 }
558
559 mutex_lock(imem->mm_mutex);
560 /* Reserve area from GPU address space */
561 ret = nvkm_mm_head(imem->mm, 0, 1, npages, npages,
562 align >> imem->iommu_pgshift, &r);
563 mutex_unlock(imem->mm_mutex);
564 if (ret) {
565 nvkm_error(subdev, "IOMMU space is full!\n");
566 goto free_pages;
567 }
568
569 /* Map into GPU address space */
570 for (i = 0; i < npages; i++) {
571 u32 offset = (r->offset + i) << imem->iommu_pgshift;
572
573 ret = iommu_map(imem->domain, offset, node->dma_addrs[i],
574 PAGE_SIZE, IOMMU_READ | IOMMU_WRITE);
575 if (ret < 0) {
576 nvkm_error(subdev, "IOMMU mapping failure: %d\n", ret);
577
578 while (i-- > 0) {
579 offset -= PAGE_SIZE;
580 iommu_unmap(imem->domain, offset, PAGE_SIZE);
581 }
582 goto release_area;
583 }
584 }
585
586 /* IOMMU bit tells that an address is to be resolved through the IOMMU */
587 r->offset |= BIT(imem->iommu_bit - imem->iommu_pgshift);
588
589 node->base.mn = r;
590 return 0;
591
592 release_area:
593 mutex_lock(imem->mm_mutex);
594 nvkm_mm_free(imem->mm, &r);
595 mutex_unlock(imem->mm_mutex);
596
597 free_pages:
598 for (i = 0; i < npages && node->pages[i] != NULL; i++) {
599 dma_addr_t dma_addr = node->dma_addrs[i];
600 if (dma_addr)
601 dma_unmap_page(dev, dma_addr, PAGE_SIZE,
602 DMA_BIDIRECTIONAL);
603 __free_page(node->pages[i]);
604 }
605 #endif
606
607 return ret;
608 }
609
610 static int
gk20a_instobj_new(struct nvkm_instmem * base,u32 size,u32 align,bool zero,struct nvkm_memory ** pmemory)611 gk20a_instobj_new(struct nvkm_instmem *base, u32 size, u32 align, bool zero,
612 struct nvkm_memory **pmemory)
613 {
614 struct gk20a_instmem *imem = gk20a_instmem(base);
615 struct nvkm_subdev *subdev = &imem->base.subdev;
616 struct gk20a_instobj *node = NULL;
617 int ret = 0;
618
619 #ifdef __NetBSD__
620 nvkm_debug(subdev, "%s (%s): size: %x align: %x\n", __func__,
621 "bus_dma", size, align);
622 #else
623 nvkm_debug(subdev, "%s (%s): size: %x align: %x\n", __func__,
624 imem->domain ? "IOMMU" : "DMA", size, align);
625 #endif
626
627 /* Round size and align to page bounds */
628 size = max(roundup(size, PAGE_SIZE), PAGE_SIZE);
629 align = max(roundup(align, PAGE_SIZE), PAGE_SIZE);
630
631 #ifdef __NetBSD__
632 ret = gk20a_instobj_ctor_iommu(imem, size >> PAGE_SHIFT, align, &node);
633 #else
634 if (imem->domain)
635 ret = gk20a_instobj_ctor_iommu(imem, size >> PAGE_SHIFT,
636 align, &node);
637 else
638 ret = gk20a_instobj_ctor_dma(imem, size >> PAGE_SHIFT,
639 align, &node);
640 #endif
641 *pmemory = node ? &node->memory : NULL;
642 if (ret)
643 return ret;
644
645 node->imem = imem;
646
647 nvkm_debug(subdev, "alloc size: 0x%x, align: 0x%x, gaddr: 0x%"PRIx64"\n",
648 size, align, (u64)node->mn->offset << 12);
649
650 return 0;
651 }
652
653 static void *
gk20a_instmem_dtor(struct nvkm_instmem * base)654 gk20a_instmem_dtor(struct nvkm_instmem *base)
655 {
656 struct gk20a_instmem *imem = gk20a_instmem(base);
657
658 /* perform some sanity checks... */
659 if (!list_empty(&imem->vaddr_lru))
660 nvkm_warn(&base->subdev, "instobj LRU not empty!\n");
661
662 if (imem->vaddr_use != 0)
663 nvkm_warn(&base->subdev, "instobj vmap area not empty! "
664 "0x%x bytes still mapped\n", imem->vaddr_use);
665
666 mutex_destroy(&imem->lock);
667
668 return imem;
669 }
670
671 static const struct nvkm_instmem_func
672 gk20a_instmem = {
673 .dtor = gk20a_instmem_dtor,
674 .memory_new = gk20a_instobj_new,
675 .zero = false,
676 };
677
678 int
gk20a_instmem_new(struct nvkm_device * device,int index,struct nvkm_instmem ** pimem)679 gk20a_instmem_new(struct nvkm_device *device, int index,
680 struct nvkm_instmem **pimem)
681 {
682 #ifndef __NetBSD__
683 struct nvkm_device_tegra *tdev = device->func->tegra(device);
684 #endif
685 struct gk20a_instmem *imem;
686
687 if (!(imem = kzalloc(sizeof(*imem), GFP_KERNEL)))
688 return -ENOMEM;
689 nvkm_instmem_ctor(&gk20a_instmem, device, index, &imem->base);
690 mutex_init(&imem->lock);
691 *pimem = &imem->base;
692
693 /* do not allow more than 1MB of CPU-mapped instmem */
694 imem->vaddr_use = 0;
695 imem->vaddr_max = 0x100000;
696 INIT_LIST_HEAD(&imem->vaddr_lru);
697
698 #ifdef __NetBSD__
699 imem->dmat = device->func->dma_tag(device);
700 nvkm_info(&imem->base.subdev, "using bus_dma\n");
701 #else
702 if (tdev->iommu.domain) {
703 imem->mm_mutex = &tdev->iommu.mutex;
704 imem->mm = &tdev->iommu.mm;
705 imem->domain = tdev->iommu.domain;
706 imem->iommu_pgshift = tdev->iommu.pgshift;
707 imem->iommu_bit = tdev->func->iommu_bit;
708
709 nvkm_info(&imem->base.subdev, "using IOMMU\n");
710 } else {
711 imem->attrs = DMA_ATTR_NON_CONSISTENT |
712 DMA_ATTR_WEAK_ORDERING |
713 DMA_ATTR_WRITE_COMBINE;
714
715 nvkm_info(&imem->base.subdev, "using DMA API\n");
716 }
717 #endif
718
719 return 0;
720 }
721