OSDN Git Service

Merge tag 'perf-urgent-2023-09-10' of git://git.kernel.org/pub/scm/linux/kernel/git...
[tomoyo/tomoyo-test1.git] / drivers / gpu / drm / nouveau / nouveau_chan.c
1 /*
2  * Copyright 2012 Red Hat Inc.
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining a
5  * copy of this software and associated documentation files (the "Software"),
6  * to deal in the Software without restriction, including without limitation
7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * Software is furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice shall be included in
12  * all copies or substantial portions of the Software.
13  *
14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20  * OTHER DEALINGS IN THE SOFTWARE.
21  *
22  * Authors: Ben Skeggs
23  */
24 #include <nvif/push006c.h>
25
26 #include <nvif/class.h>
27 #include <nvif/cl0002.h>
28 #include <nvif/if0020.h>
29
30 #include "nouveau_drv.h"
31 #include "nouveau_dma.h"
32 #include "nouveau_bo.h"
33 #include "nouveau_chan.h"
34 #include "nouveau_fence.h"
35 #include "nouveau_abi16.h"
36 #include "nouveau_vmm.h"
37 #include "nouveau_svm.h"
38
39 MODULE_PARM_DESC(vram_pushbuf, "Create DMA push buffers in VRAM");
40 int nouveau_vram_pushbuf;
41 module_param_named(vram_pushbuf, nouveau_vram_pushbuf, int, 0400);
42
43 void
44 nouveau_channel_kill(struct nouveau_channel *chan)
45 {
46         atomic_set(&chan->killed, 1);
47         if (chan->fence)
48                 nouveau_fence_context_kill(chan->fence, -ENODEV);
49 }
50
51 static int
52 nouveau_channel_killed(struct nvif_event *event, void *repv, u32 repc)
53 {
54         struct nouveau_channel *chan = container_of(event, typeof(*chan), kill);
55         struct nouveau_cli *cli = (void *)chan->user.client;
56
57         NV_PRINTK(warn, cli, "channel %d killed!\n", chan->chid);
58
59         if (unlikely(!atomic_read(&chan->killed)))
60                 nouveau_channel_kill(chan);
61
62         return NVIF_EVENT_DROP;
63 }
64
65 int
66 nouveau_channel_idle(struct nouveau_channel *chan)
67 {
68         if (likely(chan && chan->fence && !atomic_read(&chan->killed))) {
69                 struct nouveau_cli *cli = (void *)chan->user.client;
70                 struct nouveau_fence *fence = NULL;
71                 int ret;
72
73                 ret = nouveau_fence_new(&fence, chan);
74                 if (!ret) {
75                         ret = nouveau_fence_wait(fence, false, false);
76                         nouveau_fence_unref(&fence);
77                 }
78
79                 if (ret) {
80                         NV_PRINTK(err, cli, "failed to idle channel %d [%s]\n",
81                                   chan->chid, nvxx_client(&cli->base)->name);
82                         return ret;
83                 }
84         }
85         return 0;
86 }
87
88 void
89 nouveau_channel_del(struct nouveau_channel **pchan)
90 {
91         struct nouveau_channel *chan = *pchan;
92         if (chan) {
93                 struct nouveau_cli *cli = (void *)chan->user.client;
94
95                 if (chan->fence)
96                         nouveau_fence(chan->drm)->context_del(chan);
97
98                 if (cli)
99                         nouveau_svmm_part(chan->vmm->svmm, chan->inst);
100
101                 nvif_object_dtor(&chan->blit);
102                 nvif_object_dtor(&chan->nvsw);
103                 nvif_object_dtor(&chan->gart);
104                 nvif_object_dtor(&chan->vram);
105                 nvif_event_dtor(&chan->kill);
106                 nvif_object_dtor(&chan->user);
107                 nvif_mem_dtor(&chan->mem_userd);
108                 nvif_object_dtor(&chan->push.ctxdma);
109                 nouveau_vma_del(&chan->push.vma);
110                 nouveau_bo_unmap(chan->push.buffer);
111                 if (chan->push.buffer && chan->push.buffer->bo.pin_count)
112                         nouveau_bo_unpin(chan->push.buffer);
113                 nouveau_bo_ref(NULL, &chan->push.buffer);
114                 kfree(chan);
115         }
116         *pchan = NULL;
117 }
118
119 static void
120 nouveau_channel_kick(struct nvif_push *push)
121 {
122         struct nouveau_channel *chan = container_of(push, typeof(*chan), chan._push);
123         chan->dma.cur = chan->dma.cur + (chan->chan._push.cur - chan->chan._push.bgn);
124         FIRE_RING(chan);
125         chan->chan._push.bgn = chan->chan._push.cur;
126 }
127
128 static int
129 nouveau_channel_wait(struct nvif_push *push, u32 size)
130 {
131         struct nouveau_channel *chan = container_of(push, typeof(*chan), chan._push);
132         int ret;
133         chan->dma.cur = chan->dma.cur + (chan->chan._push.cur - chan->chan._push.bgn);
134         ret = RING_SPACE(chan, size);
135         if (ret == 0) {
136                 chan->chan._push.bgn = chan->chan._push.mem.object.map.ptr;
137                 chan->chan._push.bgn = chan->chan._push.bgn + chan->dma.cur;
138                 chan->chan._push.cur = chan->chan._push.bgn;
139                 chan->chan._push.end = chan->chan._push.bgn + size;
140         }
141         return ret;
142 }
143
144 static int
145 nouveau_channel_prep(struct nouveau_drm *drm, struct nvif_device *device,
146                      u32 size, struct nouveau_channel **pchan)
147 {
148         struct nouveau_cli *cli = (void *)device->object.client;
149         struct nv_dma_v0 args = {};
150         struct nouveau_channel *chan;
151         u32 target;
152         int ret;
153
154         chan = *pchan = kzalloc(sizeof(*chan), GFP_KERNEL);
155         if (!chan)
156                 return -ENOMEM;
157
158         chan->device = device;
159         chan->drm = drm;
160         chan->vmm = nouveau_cli_vmm(cli);
161         atomic_set(&chan->killed, 0);
162
163         /* allocate memory for dma push buffer */
164         target = NOUVEAU_GEM_DOMAIN_GART | NOUVEAU_GEM_DOMAIN_COHERENT;
165         if (nouveau_vram_pushbuf)
166                 target = NOUVEAU_GEM_DOMAIN_VRAM;
167
168         ret = nouveau_bo_new(cli, size, 0, target, 0, 0, NULL, NULL,
169                             &chan->push.buffer);
170         if (ret == 0) {
171                 ret = nouveau_bo_pin(chan->push.buffer, target, false);
172                 if (ret == 0)
173                         ret = nouveau_bo_map(chan->push.buffer);
174         }
175
176         if (ret) {
177                 nouveau_channel_del(pchan);
178                 return ret;
179         }
180
181         chan->chan._push.mem.object.parent = cli->base.object.parent;
182         chan->chan._push.mem.object.client = &cli->base;
183         chan->chan._push.mem.object.name = "chanPush";
184         chan->chan._push.mem.object.map.ptr = chan->push.buffer->kmap.virtual;
185         chan->chan._push.wait = nouveau_channel_wait;
186         chan->chan._push.kick = nouveau_channel_kick;
187         chan->chan.push = &chan->chan._push;
188
189         /* create dma object covering the *entire* memory space that the
190          * pushbuf lives in, this is because the GEM code requires that
191          * we be able to call out to other (indirect) push buffers
192          */
193         chan->push.addr = chan->push.buffer->offset;
194
195         if (device->info.family >= NV_DEVICE_INFO_V0_TESLA) {
196                 ret = nouveau_vma_new(chan->push.buffer, chan->vmm,
197                                       &chan->push.vma);
198                 if (ret) {
199                         nouveau_channel_del(pchan);
200                         return ret;
201                 }
202
203                 chan->push.addr = chan->push.vma->addr;
204
205                 if (device->info.family >= NV_DEVICE_INFO_V0_FERMI)
206                         return 0;
207
208                 args.target = NV_DMA_V0_TARGET_VM;
209                 args.access = NV_DMA_V0_ACCESS_VM;
210                 args.start = 0;
211                 args.limit = chan->vmm->vmm.limit - 1;
212         } else
213         if (chan->push.buffer->bo.resource->mem_type == TTM_PL_VRAM) {
214                 if (device->info.family == NV_DEVICE_INFO_V0_TNT) {
215                         /* nv04 vram pushbuf hack, retarget to its location in
216                          * the framebuffer bar rather than direct vram access..
217                          * nfi why this exists, it came from the -nv ddx.
218                          */
219                         args.target = NV_DMA_V0_TARGET_PCI;
220                         args.access = NV_DMA_V0_ACCESS_RDWR;
221                         args.start = nvxx_device(device)->func->
222                                 resource_addr(nvxx_device(device), 1);
223                         args.limit = args.start + device->info.ram_user - 1;
224                 } else {
225                         args.target = NV_DMA_V0_TARGET_VRAM;
226                         args.access = NV_DMA_V0_ACCESS_RDWR;
227                         args.start = 0;
228                         args.limit = device->info.ram_user - 1;
229                 }
230         } else {
231                 if (chan->drm->agp.bridge) {
232                         args.target = NV_DMA_V0_TARGET_AGP;
233                         args.access = NV_DMA_V0_ACCESS_RDWR;
234                         args.start = chan->drm->agp.base;
235                         args.limit = chan->drm->agp.base +
236                                      chan->drm->agp.size - 1;
237                 } else {
238                         args.target = NV_DMA_V0_TARGET_VM;
239                         args.access = NV_DMA_V0_ACCESS_RDWR;
240                         args.start = 0;
241                         args.limit = chan->vmm->vmm.limit - 1;
242                 }
243         }
244
245         ret = nvif_object_ctor(&device->object, "abi16PushCtxDma", 0,
246                                NV_DMA_FROM_MEMORY, &args, sizeof(args),
247                                &chan->push.ctxdma);
248         if (ret) {
249                 nouveau_channel_del(pchan);
250                 return ret;
251         }
252
253         return 0;
254 }
255
256 static int
257 nouveau_channel_ctor(struct nouveau_drm *drm, struct nvif_device *device, bool priv, u64 runm,
258                      struct nouveau_channel **pchan)
259 {
260         static const struct {
261                 s32 oclass;
262                 int version;
263         } hosts[] = {
264                 {  AMPERE_CHANNEL_GPFIFO_B, 0 },
265                 {  AMPERE_CHANNEL_GPFIFO_A, 0 },
266                 {  TURING_CHANNEL_GPFIFO_A, 0 },
267                 {   VOLTA_CHANNEL_GPFIFO_A, 0 },
268                 {  PASCAL_CHANNEL_GPFIFO_A, 0 },
269                 { MAXWELL_CHANNEL_GPFIFO_A, 0 },
270                 {  KEPLER_CHANNEL_GPFIFO_B, 0 },
271                 {  KEPLER_CHANNEL_GPFIFO_A, 0 },
272                 {   FERMI_CHANNEL_GPFIFO  , 0 },
273                 {     G82_CHANNEL_GPFIFO  , 0 },
274                 {    NV50_CHANNEL_GPFIFO  , 0 },
275                 {    NV40_CHANNEL_DMA     , 0 },
276                 {    NV17_CHANNEL_DMA     , 0 },
277                 {    NV10_CHANNEL_DMA     , 0 },
278                 {    NV03_CHANNEL_DMA     , 0 },
279                 {}
280         };
281         struct {
282                 struct nvif_chan_v0 chan;
283                 char name[TASK_COMM_LEN+16];
284         } args;
285         struct nouveau_cli *cli = (void *)device->object.client;
286         struct nouveau_channel *chan;
287         const u64 plength = 0x10000;
288         const u64 ioffset = plength;
289         const u64 ilength = 0x02000;
290         char name[TASK_COMM_LEN];
291         int cid, ret;
292         u64 size;
293
294         cid = nvif_mclass(&device->object, hosts);
295         if (cid < 0)
296                 return cid;
297
298         if (hosts[cid].oclass < NV50_CHANNEL_GPFIFO)
299                 size = plength;
300         else
301                 size = ioffset + ilength;
302
303         /* allocate dma push buffer */
304         ret = nouveau_channel_prep(drm, device, size, &chan);
305         *pchan = chan;
306         if (ret)
307                 return ret;
308
309         /* create channel object */
310         args.chan.version = 0;
311         args.chan.namelen = sizeof(args.name);
312         args.chan.runlist = __ffs64(runm);
313         args.chan.runq = 0;
314         args.chan.priv = priv;
315         args.chan.devm = BIT(0);
316         if (hosts[cid].oclass < NV50_CHANNEL_GPFIFO) {
317                 args.chan.vmm = 0;
318                 args.chan.ctxdma = nvif_handle(&chan->push.ctxdma);
319                 args.chan.offset = chan->push.addr;
320                 args.chan.length = 0;
321         } else {
322                 args.chan.vmm = nvif_handle(&chan->vmm->vmm.object);
323                 if (hosts[cid].oclass < FERMI_CHANNEL_GPFIFO)
324                         args.chan.ctxdma = nvif_handle(&chan->push.ctxdma);
325                 else
326                         args.chan.ctxdma = 0;
327                 args.chan.offset = ioffset + chan->push.addr;
328                 args.chan.length = ilength;
329         }
330         args.chan.huserd = 0;
331         args.chan.ouserd = 0;
332
333         /* allocate userd */
334         if (hosts[cid].oclass >= VOLTA_CHANNEL_GPFIFO_A) {
335                 ret = nvif_mem_ctor(&cli->mmu, "abi16ChanUSERD", NVIF_CLASS_MEM_GF100,
336                                     NVIF_MEM_VRAM | NVIF_MEM_COHERENT | NVIF_MEM_MAPPABLE,
337                                     0, PAGE_SIZE, NULL, 0, &chan->mem_userd);
338                 if (ret)
339                         return ret;
340
341                 args.chan.huserd = nvif_handle(&chan->mem_userd.object);
342                 args.chan.ouserd = 0;
343
344                 chan->userd = &chan->mem_userd.object;
345         } else {
346                 chan->userd = &chan->user;
347         }
348
349         get_task_comm(name, current);
350         snprintf(args.name, sizeof(args.name), "%s[%d]", name, task_pid_nr(current));
351
352         ret = nvif_object_ctor(&device->object, "abi16ChanUser", 0, hosts[cid].oclass,
353                                &args, sizeof(args), &chan->user);
354         if (ret) {
355                 nouveau_channel_del(pchan);
356                 return ret;
357         }
358
359         chan->runlist = args.chan.runlist;
360         chan->chid = args.chan.chid;
361         chan->inst = args.chan.inst;
362         chan->token = args.chan.token;
363         return 0;
364 }
365
366 static int
367 nouveau_channel_init(struct nouveau_channel *chan, u32 vram, u32 gart)
368 {
369         struct nvif_device *device = chan->device;
370         struct nouveau_drm *drm = chan->drm;
371         struct nv_dma_v0 args = {};
372         int ret, i;
373
374         ret = nvif_object_map(chan->userd, NULL, 0);
375         if (ret)
376                 return ret;
377
378         if (chan->user.oclass >= FERMI_CHANNEL_GPFIFO) {
379                 struct {
380                         struct nvif_event_v0 base;
381                         struct nvif_chan_event_v0 host;
382                 } args;
383
384                 args.host.version = 0;
385                 args.host.type = NVIF_CHAN_EVENT_V0_KILLED;
386
387                 ret = nvif_event_ctor(&chan->user, "abi16ChanKilled", chan->chid,
388                                       nouveau_channel_killed, false,
389                                       &args.base, sizeof(args), &chan->kill);
390                 if (ret == 0)
391                         ret = nvif_event_allow(&chan->kill);
392                 if (ret) {
393                         NV_ERROR(drm, "Failed to request channel kill "
394                                       "notification: %d\n", ret);
395                         return ret;
396                 }
397         }
398
399         /* allocate dma objects to cover all allowed vram, and gart */
400         if (device->info.family < NV_DEVICE_INFO_V0_FERMI) {
401                 if (device->info.family >= NV_DEVICE_INFO_V0_TESLA) {
402                         args.target = NV_DMA_V0_TARGET_VM;
403                         args.access = NV_DMA_V0_ACCESS_VM;
404                         args.start = 0;
405                         args.limit = chan->vmm->vmm.limit - 1;
406                 } else {
407                         args.target = NV_DMA_V0_TARGET_VRAM;
408                         args.access = NV_DMA_V0_ACCESS_RDWR;
409                         args.start = 0;
410                         args.limit = device->info.ram_user - 1;
411                 }
412
413                 ret = nvif_object_ctor(&chan->user, "abi16ChanVramCtxDma", vram,
414                                        NV_DMA_IN_MEMORY, &args, sizeof(args),
415                                        &chan->vram);
416                 if (ret)
417                         return ret;
418
419                 if (device->info.family >= NV_DEVICE_INFO_V0_TESLA) {
420                         args.target = NV_DMA_V0_TARGET_VM;
421                         args.access = NV_DMA_V0_ACCESS_VM;
422                         args.start = 0;
423                         args.limit = chan->vmm->vmm.limit - 1;
424                 } else
425                 if (chan->drm->agp.bridge) {
426                         args.target = NV_DMA_V0_TARGET_AGP;
427                         args.access = NV_DMA_V0_ACCESS_RDWR;
428                         args.start = chan->drm->agp.base;
429                         args.limit = chan->drm->agp.base +
430                                      chan->drm->agp.size - 1;
431                 } else {
432                         args.target = NV_DMA_V0_TARGET_VM;
433                         args.access = NV_DMA_V0_ACCESS_RDWR;
434                         args.start = 0;
435                         args.limit = chan->vmm->vmm.limit - 1;
436                 }
437
438                 ret = nvif_object_ctor(&chan->user, "abi16ChanGartCtxDma", gart,
439                                        NV_DMA_IN_MEMORY, &args, sizeof(args),
440                                        &chan->gart);
441                 if (ret)
442                         return ret;
443         }
444
445         /* initialise dma tracking parameters */
446         switch (chan->user.oclass & 0x00ff) {
447         case 0x006b:
448         case 0x006e:
449                 chan->user_put = 0x40;
450                 chan->user_get = 0x44;
451                 chan->dma.max = (0x10000 / 4) - 2;
452                 break;
453         default:
454                 chan->user_put = 0x40;
455                 chan->user_get = 0x44;
456                 chan->user_get_hi = 0x60;
457                 chan->dma.ib_base =  0x10000 / 4;
458                 chan->dma.ib_max  = (0x02000 / 8) - 1;
459                 chan->dma.ib_put  = 0;
460                 chan->dma.ib_free = chan->dma.ib_max - chan->dma.ib_put;
461                 chan->dma.max = chan->dma.ib_base;
462                 break;
463         }
464
465         chan->dma.put = 0;
466         chan->dma.cur = chan->dma.put;
467         chan->dma.free = chan->dma.max - chan->dma.cur;
468
469         ret = PUSH_WAIT(chan->chan.push, NOUVEAU_DMA_SKIPS);
470         if (ret)
471                 return ret;
472
473         for (i = 0; i < NOUVEAU_DMA_SKIPS; i++)
474                 PUSH_DATA(chan->chan.push, 0x00000000);
475
476         /* allocate software object class (used for fences on <= nv05) */
477         if (device->info.family < NV_DEVICE_INFO_V0_CELSIUS) {
478                 ret = nvif_object_ctor(&chan->user, "abi16NvswFence", 0x006e,
479                                        NVIF_CLASS_SW_NV04,
480                                        NULL, 0, &chan->nvsw);
481                 if (ret)
482                         return ret;
483
484                 ret = PUSH_WAIT(chan->chan.push, 2);
485                 if (ret)
486                         return ret;
487
488                 PUSH_NVSQ(chan->chan.push, NV_SW, 0x0000, chan->nvsw.handle);
489                 PUSH_KICK(chan->chan.push);
490         }
491
492         /* initialise synchronisation */
493         return nouveau_fence(chan->drm)->context_new(chan);
494 }
495
496 int
497 nouveau_channel_new(struct nouveau_drm *drm, struct nvif_device *device,
498                     bool priv, u64 runm, u32 vram, u32 gart, struct nouveau_channel **pchan)
499 {
500         struct nouveau_cli *cli = (void *)device->object.client;
501         int ret;
502
503         ret = nouveau_channel_ctor(drm, device, priv, runm, pchan);
504         if (ret) {
505                 NV_PRINTK(dbg, cli, "channel create, %d\n", ret);
506                 return ret;
507         }
508
509         ret = nouveau_channel_init(*pchan, vram, gart);
510         if (ret) {
511                 NV_PRINTK(err, cli, "channel failed to initialise, %d\n", ret);
512                 nouveau_channel_del(pchan);
513                 return ret;
514         }
515
516         ret = nouveau_svmm_join((*pchan)->vmm->svmm, (*pchan)->inst);
517         if (ret)
518                 nouveau_channel_del(pchan);
519
520         return ret;
521 }
522
523 void
524 nouveau_channels_fini(struct nouveau_drm *drm)
525 {
526         kfree(drm->runl);
527 }
528
529 int
530 nouveau_channels_init(struct nouveau_drm *drm)
531 {
532         struct {
533                 struct nv_device_info_v1 m;
534                 struct {
535                         struct nv_device_info_v1_data channels;
536                         struct nv_device_info_v1_data runlists;
537                 } v;
538         } args = {
539                 .m.version = 1,
540                 .m.count = sizeof(args.v) / sizeof(args.v.channels),
541                 .v.channels.mthd = NV_DEVICE_HOST_CHANNELS,
542                 .v.runlists.mthd = NV_DEVICE_HOST_RUNLISTS,
543         };
544         struct nvif_object *device = &drm->client.device.object;
545         int ret, i;
546
547         ret = nvif_object_mthd(device, NV_DEVICE_V0_INFO, &args, sizeof(args));
548         if (ret ||
549             args.v.runlists.mthd == NV_DEVICE_INFO_INVALID || !args.v.runlists.data ||
550             args.v.channels.mthd == NV_DEVICE_INFO_INVALID)
551                 return -ENODEV;
552
553         drm->chan_nr = drm->chan_total = args.v.channels.data;
554         drm->runl_nr = fls64(args.v.runlists.data);
555         drm->runl = kcalloc(drm->runl_nr, sizeof(*drm->runl), GFP_KERNEL);
556         if (!drm->runl)
557                 return -ENOMEM;
558
559         if (drm->chan_nr == 0) {
560                 for (i = 0; i < drm->runl_nr; i++) {
561                         if (!(args.v.runlists.data & BIT(i)))
562                                 continue;
563
564                         args.v.channels.mthd = NV_DEVICE_HOST_RUNLIST_CHANNELS;
565                         args.v.channels.data = i;
566
567                         ret = nvif_object_mthd(device, NV_DEVICE_V0_INFO, &args, sizeof(args));
568                         if (ret || args.v.channels.mthd == NV_DEVICE_INFO_INVALID)
569                                 return -ENODEV;
570
571                         drm->runl[i].chan_nr = args.v.channels.data;
572                         drm->runl[i].chan_id_base = drm->chan_total;
573                         drm->runl[i].context_base = dma_fence_context_alloc(drm->runl[i].chan_nr);
574
575                         drm->chan_total += drm->runl[i].chan_nr;
576                 }
577         } else {
578                 drm->runl[0].context_base = dma_fence_context_alloc(drm->chan_nr);
579                 for (i = 1; i < drm->runl_nr; i++)
580                         drm->runl[i].context_base = drm->runl[0].context_base;
581
582         }
583
584         return 0;
585 }