Repository navigation
Expand file tree
/
Copy pathNativeAdrenoMemory.cpp
More file actions
572 lines (571 loc) · 33.4 KB
/
Copy pathNativeAdrenoMemory.cpp
File metadata and controls
572 lines (571 loc) · 33.4 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
/* SPDX-License-Identifier: MIT */
#include "NativeAdrenoMemory.h"
#include "NativeAdrenoPool.h"
/* Where a call that never returns is stuck: the free/unmap/release path stores a stage number here, the board page shows it. */
extern "C" {volatile uint32_t native_adreno_trace_stage;}
#include <IOKit/IOBufferMemoryDescriptor.h>
#include <IOKit/IOLib.h>
#include <kern/task.h>
extern "C" boolean_t PE_parse_boot_argn(const char *,void *,int);
static bool outputWidth(unsigned bits)
{return bits==32 || bits==36 || bits==40 || bits==42 || bits==44 || bits==48;}
/* Where the last import stopped (10..19, 19 = done), and for a refused page the physical address and offset: the board log shows them. */
enum : IOReturn {kNativeAdrenoProtectionFailure=2}; /* KERN_PROTECTION_FAILURE, as prepare() reports a range that cannot be wired for writing */
extern "C" {volatile uint32_t native_adreno_import_stage;volatile uint64_t native_adreno_import_fail_pa,native_adreno_import_fail_offset;}
/* BO memory for the board log: [0] user BOs from the system (allocateSystem), [1] their spans, [2] refused, [3] the last refusal's status,
* [4] MiB of them live, [5] the peak, [6] kernel-mapped BOs (allocate), [7] import spans. */
extern "C" {volatile uint32_t native_adreno_bo_memory[8];}
/* Contiguous BO runs: [0] made, [1] refused by the free memory (those BOs fell back to pageable pages). */
extern "C" {volatile uint32_t native_adreno_bo_contig[2];}
/* Who holds the system BOs, for the board log: a row a process {pid, live KiB, live BOs, peak KiB}, the 24 processes that hold the
* most kept (a process that released everything gives its row up); and the live BOs by size, {count, KiB} for sizes up to 64 KiB,
* 1 MiB, 4 MiB, 16 MiB, 64 MiB and above. Both change under the driver's lock, as systemBytes does. */
extern "C" {volatile uint32_t native_adreno_bo_owners[24][4];volatile uint32_t native_adreno_bo_sizes[6][2];int proc_selfpid(void);}
/* Imported memory the driver keeps wired: an import pins another mapping's pages (prepare()) from importTask until release(), whether
* or not the GPU maps them. [0] imports alive, [1] KiB of them, [2] the peak KiB, [3] imports made so far. native_adreno_import_owners
* holds the importing processes as native_adreno_bo_owners holds the allocating ones: {pid, live KiB, live imports, peak KiB}. A range
* imported twice counts twice, although its pages are wired once. Both change under the driver's lock. */
extern "C" {volatile uint32_t native_adreno_import_memory[4];volatile uint32_t native_adreno_import_owners[24][4];}
namespace {
uint64_t systemBytes,systemPeak;
bool pressure;
unsigned sizeClass(uint64_t bytes)
{
static const uint64_t limit[5]={UINT64_C(64)<<10,UINT64_C(1)<<20,UINT64_C(4)<<20,UINT64_C(16)<<20,UINT64_C(64)<<20};
unsigned c=0;while(c<5 && bytes>limit[c])++c;
return c;
}
}
void NativeAdrenoNoteHolder(volatile uint32_t (*table)[4],int32_t pid,int64_t bytes)
{
const uint32_t kib=(uint32_t)((uint64_t)(bytes<0?-bytes:bytes)>>10);
unsigned row=24,spare=24,least=0;
for(unsigned i=0;i<24;++i) {
if(table[i][2] && table[i][0]==(uint32_t)pid) {row=i;break;}
if(!table[i][2] && spare==24)spare=i;
if(table[i][1]<table[least][1])least=i;
}
if(bytes<0) {
if(row==24)return;
table[row][1]=table[row][1]>kib?table[row][1]-kib:0;
table[row][2]=table[row][2]-1;
return;
}
if(row==24) {
/* A new holder takes a free row, else the row of the smallest holder once it holds more than that one. */
row=spare!=24?spare:least;
if(spare==24 && table[least][1]>=kib)return;
table[row][0]=(uint32_t)pid;table[row][1]=0;table[row][2]=0;table[row][3]=0;
}
table[row][1]=table[row][1]+kib;
table[row][2]=table[row][2]+1;
if(table[row][1]>table[row][3])table[row][3]=table[row][1];
}
namespace {
void noteSystem(int64_t bytes,int32_t pid)
{
systemBytes+=bytes;if(systemBytes>systemPeak)systemPeak=systemBytes;
native_adreno_bo_memory[4]=(uint32_t)(systemBytes>>20);native_adreno_bo_memory[5]=(uint32_t)(systemPeak>>20);
const unsigned c=sizeClass((uint64_t)(bytes<0?-bytes:bytes));
const uint32_t kib=(uint32_t)((uint64_t)(bytes<0?-bytes:bytes)>>10);
if(bytes>0) {native_adreno_bo_sizes[c][0]=native_adreno_bo_sizes[c][0]+1;native_adreno_bo_sizes[c][1]=native_adreno_bo_sizes[c][1]+kib;}
else {
native_adreno_bo_sizes[c][0]=native_adreno_bo_sizes[c][0]?native_adreno_bo_sizes[c][0]-1:0;
native_adreno_bo_sizes[c][1]=native_adreno_bo_sizes[c][1]>kib?native_adreno_bo_sizes[c][1]-kib:0;
}
NativeAdrenoNoteHolder(native_adreno_bo_owners,pid,bytes);
}
void noteImport(int64_t bytes,int32_t pid)
{
static uint64_t importBytes,importPeak;
importBytes+=bytes;if(importBytes>importPeak)importPeak=importBytes;
if(bytes>0) {native_adreno_import_memory[0]=native_adreno_import_memory[0]+1;native_adreno_import_memory[3]=native_adreno_import_memory[3]+1;}
else native_adreno_import_memory[0]=native_adreno_import_memory[0]?native_adreno_import_memory[0]-1:0;
native_adreno_import_memory[1]=(uint32_t)(importBytes>>10);native_adreno_import_memory[2]=(uint32_t)(importPeak>>10);
NativeAdrenoNoteHolder(native_adreno_import_owners,pid,bytes);
}
}
bool NativeAdrenoMemory::memoryPressure() {return pressure;}
/* Which DRAM pages this driver's own BOs hold, a bit a page over the kernel's physical range: an import (another task's pages) aliases
* one of them only where a bit is set, so proving an import exclusive costs its own pages, not the ranges mapped so far (every BO is
* scattered pages: a pairwise check over them ran out of budget on each import). Made on first use; without it every page counts as
* owned (the full alias check). */
extern "C" uint64_t gPhysBase,gPhysSize;
namespace {
uint64_t *ownedBits;
/* The bits of the kernel's own range, then those of the first 4 GiB below it (DRAM a kernel adopts outside its range). */
const uint64_t kOwnedLowPages=UINT64_C(1)<<20;
uint64_t *ownedMap()
{
static bool tried;
if(!ownedBits && !tried) {tried=true;ownedBits=static_cast<uint64_t *>(IOMallocZero((size_t)((gPhysSize/4096+kOwnedLowPages+63)/64*8)));}
return ownedBits;
}
/* The page's bit; false for a page that has none (neither in the kernel's range nor below it within 4 GiB). */
bool ownedIndex(uint64_t page,uint64_t *at)
{
if(page>=gPhysBase) {
if(page-gPhysBase>=gPhysSize)return false;
*at=(page-gPhysBase)/4096;
return true;
}
if(page>>32)return false;
*at=gPhysSize/4096+page/4096;
return true;
}
void noteOwned(const NativeAdrenoPhysicalSpan *spans,size_t count,bool owned)
{
uint64_t *bits=ownedMap();
if(!bits)return;
for(size_t i=0;i<count;++i)for(uint64_t page=spans[i].physical;page<spans[i].physical+spans[i].bytes;page+=4096) {
uint64_t at=0;
if(!ownedIndex(page,&at))continue;
const uint64_t bit=UINT64_C(1)<<(at%64);
if(owned)__atomic_fetch_or(&bits[at/64],bit,__ATOMIC_RELAXED);else __atomic_fetch_and(&bits[at/64],~bit,__ATOMIC_RELAXED);
}
}
}
bool NativeAdrenoMemory::ownsAnyPage(const NativeAdrenoPhysicalSpan *spans,size_t count)
{
const uint64_t *bits=ownedMap();
if(!bits)return true;
for(size_t i=0;i<count;++i)for(uint64_t page=spans[i].physical;page<spans[i].physical+spans[i].bytes;page+=4096) {
uint64_t at=0;
if(!ownedIndex(page,&at))continue;
if(__atomic_load_n(&bits[at/64],__ATOMIC_RELAXED)>>(at%64)&1)return true;
}
return false;
}
/* The descriptor's pages as spans (neighbours merged), each below `limit` and, for a contiguous allocation, following the first. */
IOReturn NativeAdrenoMemory::takeSegments(IOBufferMemoryDescriptor *descriptor,size_t bytes,uint64_t limit,bool contiguous)
{
span_capacity_=bytes/4096;
spans_=IONewData(NativeAdrenoPhysicalSpan,span_capacity_);
if(!spans_)return kIOReturnNoMemory;
for(size_t offset=0;offset<bytes;offset+=4096) {
IOByteCount length=0;
addr64_t physical=descriptor->getPhysicalSegment(offset,&length,kIOMemoryMapperNone);
if(!physical || (physical&4095) || length<4096 || physical>=limit || 4096>limit-physical)return kIOReturnUnsupported;
if(contiguous && offset && physical!=spans_[0].physical+offset)return kIOReturnUnsupported;
if(span_count_ && physical==spans_[span_count_-1].physical+spans_[span_count_-1].bytes)
spans_[span_count_-1].bytes+=4096;
else spans_[span_count_++]={offset,physical,4096};
}
return kIOReturnSuccess;
}
/* The grant's runs as spans, each below `limit`. */
IOReturn NativeAdrenoMemory::takeRuns(const NativeAdrenoPoolGrant *grant,size_t bytes,uint64_t limit)
{
if(!grant->count || grant->bytes!=bytes)return kIOReturnUnsupported;
span_capacity_=grant->count;
spans_=IONewData(NativeAdrenoPhysicalSpan,span_capacity_);
if(!spans_)return kIOReturnNoMemory;
uint64_t offset=0;
for(uint32_t i=0;i<grant->count;++i) {
const NativeAdrenoPoolRun &run=grant->runs[i];
if(!run.physical || !run.bytes || ((run.physical|run.bytes)&4095) || run.bytes>bytes-offset || run.physical>=limit ||
run.bytes>limit-run.physical)return kIOReturnUnsupported;
spans_[span_count_++]={offset,run.physical,run.bytes};
offset+=run.bytes;
}
return offset==bytes?kIOReturnSuccess:kIOReturnUnsupported;
}
/* The first kernel allocation that failed: [0] where (1 no descriptor, 2 prepare, 3 after it), [1] status, [2] bytes, [3] address
* bits, [4] contiguous, [5] failures so far. */
extern "C" {volatile uint32_t native_adreno_kernel_alloc_fail[6];}
static void noteKernelFailure(uint32_t where,IOReturn status,size_t bytes,unsigned bits,bool contiguous)
{
volatile uint32_t *f=native_adreno_kernel_alloc_fail;
if(!f[5]) {f[0]=where;f[1]=(uint32_t)status;f[2]=(uint32_t)bytes;f[3]=bits;f[4]=contiguous;}
f[5]=f[5]+1;
}
IOReturn NativeAdrenoMemory::allocate(size_t bytes,unsigned bits,bool contiguous,size_t budget)
{
if(descriptor_ || grant_ || spans_ || quarantined_)return kIOReturnBusy;
if(!bytes || (bytes&4095) || !outputWidth(bits) || !budget)
return kIOReturnBadArgument;
if(bytes>budget)return kIOReturnNoResources;
NativeAdrenoCpuState cpu{};
IOReturn status=NativeAdrenoQueryCpu4K(&cpu);if(status!=kIOReturnSuccess)return status;
uint64_t limit=UINT64_C(1)<<bits;
IOOptionBits options=static_cast<IOOptionBits>(kIODirectionInOut)|static_cast<IOOptionBits>(kIOMemoryMapperNone);
if(contiguous)options|=kIOMemoryPhysicallyContiguous;
descriptor_=IOBufferMemoryDescriptor::inTaskWithPhysicalMask(kernel_task,options,bytes,(limit-1)&~UINT64_C(4095));
if(!descriptor_) {noteKernelFailure(1,kIOReturnNoMemory,bytes,bits,contiguous);return kIOReturnNoMemory;}
status=descriptor_->prepare();
if(status!=kIOReturnSuccess) {noteKernelFailure(2,status,bytes,bits,contiguous);descriptor_->release();descriptor_=nullptr;return status;}
prepared_=true;bytes_=bytes;mapping_=descriptor_->getBytesNoCopy();
status=kIOReturnUnsupported;
if(!mapping_ || ((uintptr_t)mapping_&4095) || bytes>UINTPTR_MAX-(uintptr_t)mapping_ ||
descriptor_->getLength()!=bytes)goto failed;
status=takeSegments(descriptor_,bytes,limit,contiguous);
if(status!=kIOReturnSuccess)goto failed;
bzero(mapping_,bytes);
status=NativeAdrenoMaintainOwnedPages(mapping_,bytes,NativeAdrenoClean);
if(status!=kIOReturnSuccess)goto failed;
NativeAdrenoNotePhase(NativeAdrenoPhaseAllocate,bytes);
noteOwned(spans_,span_count_,true);owned_=true;
native_adreno_bo_memory[6]=native_adreno_bo_memory[6]+1;
return kIOReturnSuccess;
failed:
noteKernelFailure(3,status,bytes,bits,contiguous);
/* release() deliberately retains descriptor/metadata if complete fails.
* Preserve the allocation's original failure instead of reporting cleanup
* as success. No object has been exposed to a GPU or userspace yet. */
(void)release();return status;
}
/* No task: the memory is mapped nowhere until a client view maps it. The cache mode is the memory entry's, so every mapping of it (the
* client views) and the pages themselves take it. */
static IOBufferMemoryDescriptor *systemDescriptor(size_t bytes,bool write_combined,uint64_t *made=nullptr,uint64_t *wired=nullptr)
{
IOOptionBits options=static_cast<IOOptionBits>(kIODirectionInOut)|static_cast<IOOptionBits>(kIOMemoryMapperNone)|
static_cast<IOOptionBits>(kIOMemoryPageable);
if(write_combined)options|=kIOMapWriteCombineCache;
/* Pageable pages arrive one by one (a BO averaged 69 spans); from `gpu-bo-contig` KiB up to 128 KiB a BO first asks for one
* physically contiguous run (one span, one table walk), and a request the free memory cannot serve falls back to pageable pages.
* A contiguous request the kernel cannot serve scans every free page before it gives up: in ibb10331f24 124 such searches (0.1 to
* 64 MiB) took 163 ms each at the median and 21.7 s in all, synchronously in the caller (the render server's and the widget
* renderer's allocations, the stalls at an app launch and on the widget page), while every success was 64 to 128 KiB. So larger
* BOs never ask, and after a failure no request of that size or larger asks again for 5 s. */
static uint32_t contigKiB=UINT32_MAX;
if(contigKiB==UINT32_MAX) {contigKiB=0;PE_parse_boot_argn("gpu-bo-contig",&contigKiB,sizeof contigKiB);}
static const size_t contigMax=size_t(128)*1024;
static size_t failedBytes=0;
static uint64_t failedAt=0,backoff=0;
if(!backoff)nanoseconds_to_absolutetime(UINT64_C(5000000000),&backoff);
const uint64_t now=mach_absolute_time();
if(failedBytes && now-failedAt>=backoff)failedBytes=0;
if(contigKiB && bytes>=size_t(contigKiB)*1024 && bytes<=contigMax && !(failedBytes && bytes>=failedBytes)) {
IOBufferMemoryDescriptor *run=IOBufferMemoryDescriptor::inTaskWithPhysicalMask(nullptr,
(options&~static_cast<IOOptionBits>(kIOMemoryPageable))|static_cast<IOOptionBits>(kIOMemoryPhysicallyContiguous),bytes,UINT64_C(0xfffffffff000));
if(run && run->prepare()==kIOReturnSuccess) {native_adreno_bo_contig[0]=native_adreno_bo_contig[0]+1;return run;}
if(run)run->release();
native_adreno_bo_contig[1]=native_adreno_bo_contig[1]+1;
if(!failedBytes || bytes<failedBytes)failedBytes=bytes;
failedAt=now;
}
IOBufferMemoryDescriptor *descriptor=IOBufferMemoryDescriptor::inTaskWithOptions(nullptr,options,bytes,4096);
if(made)*made=mach_absolute_time();
if(descriptor && descriptor->prepare()!=kIOReturnSuccess) {descriptor->release();descriptor=nullptr;}
if(wired)*wired=mach_absolute_time();
return descriptor;
}
IOReturn NativeAdrenoMemory::prepareSystem(size_t bytes,bool write_combined,NativeAdrenoSystemPages *out,unsigned bits)
{
if(!out || out->held() || !bytes || (bytes&4095) || (bits && !outputWidth(bits)))return kIOReturnBadArgument;
/* Pool pages first: taken and zero-filled here, outside the driver's lock. Nothing to clean, no cache holds them. */
if(bits) {
NativeAdrenoPoolGrant *grant=nullptr;
if(NativeAdrenoPoolAllocate(bytes,UINT64_C(1)<<bits,write_combined,&grant)==kIOReturnSuccess) {
out->grant=grant;out->bytes=bytes;out->write_combined=write_combined;out->cleaned=true;
out->made=out->wired=mach_absolute_time();
return kIOReturnSuccess;
}
}
uint64_t made=0,wired=0;
IOBufferMemoryDescriptor *descriptor=systemDescriptor(bytes,write_combined,&made,&wired);
if(!descriptor)return kIOReturnNoMemory;
if(!made)made=wired=mach_absolute_time(); /* a contiguous run: made and wired in one step */
if(descriptor->getLength()!=bytes) {descriptor->complete();descriptor->release();return kIOReturnUnsupported;}
bool cleaned=false;
/* Default-attribute pages may still hold the zero fill in a CPU's cache: written back here, by physical runs, as allocateSystem would. */
if(!write_combined) {
cleaned=true;
for(size_t offset=0;offset<bytes && cleaned;) {
IOByteCount length=0;
const addr64_t physical=descriptor->getPhysicalSegment(offset,&length,kIOMemoryMapperNone);
if(!physical || (physical&4095) || length<4096) {cleaned=false;break;}
if(length>bytes-offset)length=bytes-offset;
length&=~IOByteCount(4095);
if(NativeAdrenoMaintainPhysicalPages(physical,length,NativeAdrenoClean)!=kIOReturnSuccess)cleaned=false;
offset+=length;
}
if(cleaned)NativeAdrenoNotePhase(NativeAdrenoPhaseAllocate,bytes);
}
out->descriptor=descriptor;out->bytes=bytes;out->write_combined=write_combined;out->cleaned=cleaned;out->made=made;out->wired=wired;
return kIOReturnSuccess;
}
void NativeAdrenoMemory::discardSystem(NativeAdrenoSystemPages *pages)
{
if(!pages || !pages->held())return;
if(pages->descriptor) {pages->descriptor->complete();pages->descriptor->release();}
/* The BO's reference to its pool pages: they go back once no client view holds them either. */
if(pages->grant)NativeAdrenoPoolRelease(pages->grant);
*pages=NativeAdrenoSystemPages{};
}
IOReturn NativeAdrenoMemory::allocateSystem(size_t bytes,unsigned bits,size_t budget,bool write_combined,NativeAdrenoSystemPages *prepared)
{
if(descriptor_ || grant_ || spans_ || quarantined_)return kIOReturnBusy;
if(!bytes || (bytes&4095) || !outputWidth(bits) || !budget)return kIOReturnBadArgument;
if(bytes>budget)return kIOReturnNoResources;
NativeAdrenoCpuState cpu{};
IOReturn status=NativeAdrenoQueryCpu4K(&cpu);if(status!=kIOReturnSuccess)return status;
/* Pages prepared outside the lock for this very request are taken over; any other request makes its own. */
bool cleaned=false;
if(prepared && prepared->held() && prepared->bytes==bytes && prepared->write_combined==write_combined) {
descriptor_=prepared->descriptor;grant_=prepared->grant;cleaned=prepared->cleaned;*prepared=NativeAdrenoSystemPages{};
} else if(prepared || NativeAdrenoPoolAllocate(bytes,UINT64_C(1)<<bits,write_combined,&grant_)!=kIOReturnSuccess) {
/* A caller that offers prepared pages holds the driver's lock: the pool's zero fill maps its run in the kernel map and is not
* done under it (prepareSystem already asked the pool, outside the lock). */
grant_=nullptr;descriptor_=systemDescriptor(bytes,write_combined);
}
if(grant_) {
/* Pool pages: the grant's runs are the spans. A grant that does not fit this client's physical limit after all (it was
* prepared for another width) goes back and the VM's pages take its place. */
const IOReturn taken=takeRuns(grant_,bytes,UINT64_C(1)<<bits);
if(taken!=kIOReturnSuccess) {
if(spans_)IODeleteData(spans_,NativeAdrenoPhysicalSpan,span_capacity_);
spans_=nullptr;span_count_=span_capacity_=0;
NativeAdrenoPoolRelease(grant_);grant_=nullptr;cleaned=false;
descriptor_=systemDescriptor(bytes,write_combined);
}
}
status=descriptor_ || grant_?kIOReturnSuccess:kIOReturnNoMemory;
if(status==kIOReturnSuccess) {
system_=true;prepared_=true;bytes_=bytes;write_combined_=write_combined;owner_pid_=proc_selfpid();noteSystem((int64_t)bytes,owner_pid_);
if(descriptor_)status=descriptor_->getLength()==bytes?takeSegments(descriptor_,bytes,UINT64_C(1)<<bits,false):kIOReturnUnsupported;
/* Default-attribute pages may still hold the zero fill in this CPU's cache: written back before the GPU reads them. */
if(status==kIOReturnSuccess && descriptor_ && !write_combined && !cleaned) {
status=maintainImported(0,bytes,NativeAdrenoClean);
NativeAdrenoNotePhase(NativeAdrenoPhaseAllocate,bytes);
}
if(status!=kIOReturnSuccess) {write_combined_=false;(void)release();}
}
if(status!=kIOReturnSuccess) {
pressure=true;
native_adreno_bo_memory[2]=native_adreno_bo_memory[2]+1;native_adreno_bo_memory[3]=(uint32_t)status;
return status;
}
pressure=false;
noteOwned(spans_,span_count_,true);owned_=true;
native_adreno_bo_memory[0]=native_adreno_bo_memory[0]+1;native_adreno_bo_memory[1]=native_adreno_bo_memory[1]+(uint32_t)span_count_;
return kIOReturnSuccess;
}
/* An import holds the pages through a memory entry of its own (kIOMemoryPersistent), which references the VM object: the owner may unmap
* its range while the GPU still uses it (Core Animation frees an image's Shmem once its no-copy buffer is released, before the runtime frees
* the import), and the object then lives on until the import completes instead of being torn down under a wired UPL (int-233..235: the render
* thread stuck in vm_deallocate). A range no memory entry can be made of is imported as before; [0] counts the persistent imports, [1] those. */
extern "C" {volatile uint32_t native_adreno_import_persistent[2];}
/* Imports whose pages were all non-cacheable ([1]: no CPU maintenance for them) or not ([0]). */
extern "C" {volatile uint32_t native_adreno_pool_stats_uncached[2];}
static IOMemoryDescriptor *importRange(uint64_t address,size_t bytes,IODirection direction,void *task)
{
const IOOptionBits options=static_cast<IOOptionBits>(direction)|static_cast<IOOptionBits>(kIOMemoryMapperNone);
IOMemoryDescriptor *range=IOMemoryDescriptor::withAddressRange(address,bytes,options|kIOMemoryPersistent,static_cast<task_t>(task));
if(range) {native_adreno_import_persistent[0]=native_adreno_import_persistent[0]+1;return range;}
range=IOMemoryDescriptor::withAddressRange(address,bytes,options,static_cast<task_t>(task));
if(range)native_adreno_import_persistent[1]=native_adreno_import_persistent[1]+1;
return range;
}
IOReturn NativeAdrenoMemory::importTask(void *task,uint64_t address,size_t bytes,unsigned bits,size_t budget)
{
native_adreno_import_stage=10;
if(descriptor_ || grant_ || imported_ || spans_ || quarantined_)return kIOReturnBusy;
native_adreno_import_stage=11;
if(!task || !address || (address&4095) || !bytes || (bytes&4095) || address>UINT64_MAX-bytes || !outputWidth(bits) ||
!budget)return kIOReturnBadArgument;
native_adreno_import_stage=12;
if(bytes>budget)return kIOReturnNoResources;
NativeAdrenoCpuState cpu{};
native_adreno_import_stage=13;
IOReturn status=NativeAdrenoQueryCpu4K(&cpu);if(status!=kIOReturnSuccess)return status;
const uint64_t limit=UINT64_C(1)<<bits;
imported_=importRange(address,bytes,kIODirectionInOut,task);
native_adreno_import_stage=14;
if(!imported_)return kIOReturnNoMemory;
native_adreno_import_stage=15;
status=imported_->prepare();
if(status==kNativeAdrenoProtectionFailure) {
/* Read-only memory of the owner (a mapped file, constant data): cannot be pinned for writing. Pin it for reading only; the GPU then maps it
* read-only. Metal's no-copy buffers over such memory are only ever read by shaders. */
imported_->release();imported_=nullptr;
imported_=importRange(address,bytes,kIODirectionOut,task);
if(!imported_)return kIOReturnNoMemory;
status=imported_->prepare();
if(status==kIOReturnSuccess)read_only_=true;
}
if(status!=kIOReturnSuccess) {imported_->release();imported_=nullptr;return status;}
prepared_=true;bytes_=bytes;
status=kIOReturnUnsupported;
native_adreno_import_stage=16;
if(imported_->getLength()!=bytes)goto failed;
native_adreno_import_stage=17;
span_capacity_=bytes/4096;
spans_=IONewData(NativeAdrenoPhysicalSpan,span_capacity_);
if(!spans_) {status=kIOReturnNoMemory;goto failed;}
for(size_t offset=0;offset<bytes;offset+=4096) {
IOByteCount length=0;
addr64_t physical=imported_->getPhysicalSegment(offset,&length,kIOMemoryMapperNone);
if(!physical || (physical&4095) || length<4096 || physical>=limit || 4096>limit-physical)
{native_adreno_import_stage=18;native_adreno_import_fail_pa=physical;native_adreno_import_fail_offset=offset;
status=kIOReturnUnsupported;goto failed;}
/* A range that maps pool pages (a client view of a pool BO) is not imported: the pin of such a range holds no page, so the
* pool could give the pages to another client while this import still maps them for the GPU. */
if(NativeAdrenoPoolContains(physical,4096))
{native_adreno_import_stage=18;native_adreno_import_fail_pa=physical;native_adreno_import_fail_offset=offset;
status=kIOReturnNotPermitted;goto failed;}
if(span_count_ && physical==spans_[span_count_-1].physical+spans_[span_count_-1].bytes)
spans_[span_count_-1].bytes+=4096;
else spans_[span_count_++]={offset,physical,4096};
}
native_adreno_import_stage=19;
owner_uncached_=NativeAdrenoPagesUncached(spans_,span_count_);
native_adreno_pool_stats_uncached[owner_uncached_]=native_adreno_pool_stats_uncached[owner_uncached_]+1;
native_adreno_bo_memory[7]=native_adreno_bo_memory[7]+(uint32_t)span_count_;
owner_pid_=proc_selfpid();import_noted_=true;noteImport((int64_t)bytes,owner_pid_);
return kIOReturnSuccess;
failed:
(void)release();return status;
}
IOReturn NativeAdrenoMemory::borrow(void *mapping,uint64_t physical,size_t bytes)
{
if(descriptor_ || grant_ || imported_ || spans_ || quarantined_ || borrowed_)return kIOReturnBusy;
if(!mapping || ((uintptr_t)mapping&4095) || !physical || (physical&4095) || !bytes || (bytes&4095))return kIOReturnBadArgument;
spans_=IONewData(NativeAdrenoPhysicalSpan,1);
if(!spans_)return kIOReturnNoMemory;
span_capacity_=1;span_count_=1;spans_[0]={0,physical,bytes};
mapping_=mapping;bytes_=bytes;prepared_=true;borrowed_=true;
bzero(mapping_,bytes);
return kIOReturnSuccess;
}
IOReturn NativeAdrenoMemory::makeWriteCombined()
{
if(!descriptor_ || system_ || imported_ || !prepared_ || quarantined_ || device_references_ || write_combined_ || !mapping_ || !span_count_)
return kIOReturnNotReady;
IOReturn status=NativeAdrenoMaintainOwnedPages(mapping_,bytes_,NativeAdrenoCleanInvalidate);
if(status)return status;
NativeAdrenoNotePhase(NativeAdrenoPhaseAllocate,bytes_);
status=NativeAdrenoSetPagesCacheMode(spans_,span_count_,true);
if(status)return status;
write_combined_=true;return kIOReturnSuccess;
}
IOReturn NativeAdrenoMemory::declareSynchronized()
{
if(!descriptor_ || system_ || imported_ || borrowed_ || !prepared_ || quarantined_ || device_references_ || !mapping_ || !span_count_)
return kIOReturnNotReady;
const IOReturn status=NativeAdrenoMaintainOwnedPages(mapping_,bytes_,NativeAdrenoCleanInvalidate);
if(status)return status;
NativeAdrenoNotePhase(NativeAdrenoPhaseAllocate,bytes_);
synchronized_=true;return kIOReturnSuccess;
}
/* Frees of an owned BO whose first pages a user task still maps: [0] count, [1] the last mapping's page number. */
extern "C" {volatile uint32_t native_adreno_free_user_mapped[2];}
extern "C" unsigned int native_pmap_user_mappings(ppnum_t,void **,uint64_t *,char *) __attribute__((weak_import));
IOReturn NativeAdrenoMemory::release()
{
native_adreno_trace_stage=30;
if(device_references_ || quarantined_)return kIOReturnBusy;
/* System memory's pages go back to the VM with the default attribute (vm_object_reap), as they came. */
if(system_)write_combined_=false;
if(write_combined_) {
/* The pages go back to the VM cacheable, as they came. */
const IOReturn status=NativeAdrenoSetPagesCacheMode(spans_,span_count_,false);
if(status) {quarantined_=true;return status;}
write_combined_=false;
}
if(imported_) {
if(prepared_) {
IOReturn status=imported_->complete();
if(status!=kIOReturnSuccess) {quarantined_=true;return status;}
prepared_=false;
}
imported_->release();imported_=nullptr;
}
if(descriptor_ && !system_ && native_pmap_user_mappings && spans_) {
/* A page a user task still maps (a BO view over physical pages holds no reference to them) is counted before the pages are freed:
* the first page of up to 64 spans, then which process maps it. XNU frees such a page under the mapping (pmap_assert_free). A view
* of system memory maps its VM object, which keeps the pages until the view goes. */
for(size_t i=0;i<span_count_ && i<64;++i) {
char name[17]={};void *pmap=nullptr;uint64_t va=0;
if(native_pmap_user_mappings((ppnum_t)(spans_[i].physical>>12),&pmap,&va,name)) {
native_adreno_free_user_mapped[0]=native_adreno_free_user_mapped[0]+1;
native_adreno_free_user_mapped[1]=(uint32_t)(va>>12);
break;
}
}
}
if(descriptor_) {
if(prepared_) {
native_adreno_trace_stage=31;
IOReturn status=descriptor_->complete();
native_adreno_trace_stage=32;
if(status!=kIOReturnSuccess) {quarantined_=true;return status;}
prepared_=false;
}
native_adreno_trace_stage=33;
descriptor_->release();descriptor_=nullptr;
native_adreno_trace_stage=34;
}
/* Pool pages: the BO's reference goes; the pages return to the pool once no client view holds the grant. */
if(grant_) {NativeAdrenoPoolRelease(grant_);grant_=nullptr;prepared_=false;}
if(owned_) {noteOwned(spans_,span_count_,false);owned_=false;}
if(system_) {noteSystem(-(int64_t)bytes_,owner_pid_);system_=false;}
if(import_noted_) {noteImport(-(int64_t)bytes_,owner_pid_);import_noted_=false;}
if(borrowed_) {borrowed_=false;prepared_=false;}
if(spans_)IODeleteData(spans_,NativeAdrenoPhysicalSpan,span_capacity_);
mapping_=nullptr;spans_=nullptr;bytes_=span_count_=span_capacity_=0;read_only_=false;owner_uncached_=false;synchronized_=false;
return kIOReturnSuccess;
}
IOReturn NativeAdrenoMemory::releaseSystemInto(NativeAdrenoSystemPages *out)
{
if(!out || out->held() || !system_ || !(descriptor_ || grant_) || !prepared_ || imported_ || device_references_ || quarantined_)return release();
/* The GPU no longer maps the pages (the free unmapped them first) and every client view holds the descriptor itself, or its own
* reference to the pool grant. */
out->descriptor=descriptor_;out->grant=grant_;out->bytes=bytes_;out->write_combined=write_combined_;out->cleaned=false;
descriptor_=nullptr;grant_=nullptr;prepared_=false;
return release();
}
IOReturn NativeAdrenoMemory::holdForDevice()
{
if(!backed() || !prepared_ || !span_count_ || quarantined_)return kIOReturnNotReady;
if(device_references_==UINT32_MAX)return kIOReturnNoResources;
++device_references_;return kIOReturnSuccess;
}
IOReturn NativeAdrenoMemory::retireAfterQuiescence()
{
if(!device_references_)return kIOReturnBadArgument;
if(quarantined_)return kIOReturnNotReady;
--device_references_;return kIOReturnSuccess;
}
IOReturn NativeAdrenoMemory::publishForDevice(size_t offset,size_t bytes,bool writes,bool include_write_combined,unsigned phase)
{
if(!backed() || !prepared_ || quarantined_)return kIOReturnNotReady;
if(!bytes || ((offset|bytes)&4095) || offset>bytes_ || bytes>bytes_-offset)return kIOReturnBadArgument;
/* Pages every CPU mapping holds non-cacheable carry no line to clean. */
if(!holdsCpuLines(include_write_combined))return kIOReturnSuccess;
if(imported_) {NativeAdrenoNotePhase(phase,bytes);return maintainImported(offset,bytes,writes?NativeAdrenoCleanInvalidate:NativeAdrenoClean);}
if(write_combined_) { /* no CPU mapping should cache these pages; if one does, reach its lines through the physical pages */
NativeAdrenoNotePhase(phase,bytes);
return maintainImported(offset,bytes,writes?NativeAdrenoCleanInvalidate:NativeAdrenoClean);
}
NativeAdrenoNotePhase(phase,bytes);
if(!mapping_)return maintainImported(offset,bytes,writes?NativeAdrenoCleanInvalidate:NativeAdrenoClean); /* system memory */
return NativeAdrenoMaintainOwnedPages(static_cast<unsigned char *>(mapping_)+offset,bytes,
writes?NativeAdrenoCleanInvalidate:NativeAdrenoClean);
}
IOReturn NativeAdrenoMemory::acquireDeviceWrites(size_t offset,size_t bytes,unsigned phase)
{
if(!backed() || !prepared_ || quarantined_)return kIOReturnNotReady;
if(!bytes || ((offset|bytes)&4095) || offset>bytes_ || bytes>bytes_-offset)return kIOReturnBadArgument;
if(grant_ || (imported_ && owner_uncached_))return kIOReturnSuccess;
/* Imported pages may hold the owner's unrelated dirty lines next to the GPU's writes: clean as well as invalidate. */
if(imported_) {NativeAdrenoNotePhase(phase,bytes);return maintainImported(offset,bytes,NativeAdrenoCleanInvalidate);}
if(write_combined_)return kIOReturnSuccess;
NativeAdrenoNotePhase(phase,bytes);
if(!mapping_)return maintainImported(offset,bytes,NativeAdrenoInvalidate); /* system memory */
return NativeAdrenoMaintainOwnedPages(static_cast<unsigned char *>(mapping_)+offset,bytes,NativeAdrenoInvalidate);
}
/* Imported pages have no kernel mapping here (and write-combined ones only a non-cacheable one); maintain each span's part of
* [offset, offset+bytes) by physical address. */
IOReturn NativeAdrenoMemory::maintainImported(size_t offset,size_t bytes,NativeAdrenoCacheOperation op)
{
for(size_t i=0;i<span_count_;++i) {
const NativeAdrenoPhysicalSpan &span=spans_[i];
const size_t begin=offset>span.offset?offset:span.offset;
const size_t end=offset+bytes<span.offset+span.bytes?offset+bytes:span.offset+span.bytes;
if(begin>=end)continue;
const IOReturn status=NativeAdrenoMaintainPhysicalPages(span.physical+(begin-span.offset),end-begin,op);
if(status)return status;
}
return kIOReturnSuccess;
}