1/* SPDX-License-Identifier: GPL-2.0-only */
2/*
3 * Copyright (C) 2013 Red Hat
4 * Author: Rob Clark <robdclark@gmail.com>
5 */
6
7#ifndef __MSM_GEM_H__
8#define __MSM_GEM_H__
9
10#include "msm_mmu.h"
11#include <linux/kref.h>
12#include <linux/dma-resv.h>
13#include "drm/drm_exec.h"
14#include "drm/drm_gpuvm.h"
15#include "drm/gpu_scheduler.h"
16#include "msm_drv.h"
17
18/* Make all GEM related WARN_ON()s ratelimited.. when things go wrong they
19 * tend to go wrong 1000s of times in a short timespan.
20 */
21#define GEM_WARN_ON(x) WARN_RATELIMIT(x, "%s", __stringify(x))
22
23/* Additional internal-use only BO flags: */
24#define MSM_BO_STOLEN 0x10000000 /* try to use stolen/splash memory */
25#define MSM_BO_MAP_PRIV 0x20000000 /* use IOMMU_PRIV when mapping */
26
27/**
28 * struct msm_gem_vm_log_entry - An entry in the VM log
29 *
30 * For userspace managed VMs, a log of recent VM updates is tracked and
31 * captured in GPU devcore dumps, to aid debugging issues caused by (for
32 * example) incorrectly synchronized VM updates
33 */
34struct msm_gem_vm_log_entry {
35 const char *op;
36 uint64_t iova;
37 uint64_t range;
38 int queue_id;
39};
40
41/**
42 * struct msm_gem_vm - VM object
43 *
44 * A VM object representing a GPU (or display or GMU or ...) virtual address
45 * space.
46 *
47 * In the case of GPU, if per-process address spaces are supported, the address
48 * space is split into two VMs, which map to TTBR0 and TTBR1 in the SMMU. TTBR0
49 * is used for userspace objects, and is unique per msm_context/drm_file, while
50 * TTBR1 is the same for all processes. (The kernel controlled ringbuffer and
51 * a few other kernel controlled buffers live in TTBR1.)
52 *
53 * The GPU TTBR0 vm can be managed by userspace or by the kernel, depending on
54 * whether userspace supports VM_BIND. All other vm's are managed by the kernel.
55 * (Managed by kernel means the kernel is responsible for VA allocation.)
56 *
57 * Note that because VM_BIND allows a given BO to be mapped multiple times in
58 * a VM, and therefore have multiple VMA's in a VM, there is an extra object
59 * provided by drm_gpuvm infrastructure.. the drm_gpuvm_bo, which is not
60 * embedded in any larger driver structure. The GEM object holds a list of
61 * drm_gpuvm_bo, which in turn holds a list of msm_gem_vma. A linked vma
62 * holds a reference to the vm_bo, and drops it when the vma is unlinked.
63 * So we just need to call drm_gpuvm_bo_obtain() to return a ref to an
64 * existing vm_bo, or create a new one. Once the vma is linked, the ref
65 * to the vm_bo can be dropped (since the vma is holding one).
66 */
67struct msm_gem_vm {
68 /** @base: Inherit from drm_gpuvm. */
69 struct drm_gpuvm base;
70
71 /**
72 * @sched: Scheduler used for asynchronous VM_BIND request.
73 *
74 * Unused for kernel managed VMs (where all operations are synchronous).
75 */
76 struct drm_gpu_scheduler sched;
77
78 /**
79 * @prealloc_throttle: Used to throttle VM_BIND ops if too much pre-
80 * allocated memory is in flight.
81 *
82 * Because we have to pre-allocate pgtable pages for the worst case
83 * (ie. new mappings do not share any PTEs with existing mappings)
84 * we could end up consuming a lot of resources transiently. The
85 * prealloc_throttle puts an upper bound on that.
86 */
87 struct {
88 /** @wait: Notified when preallocated resources are released */
89 wait_queue_head_t wait;
90
91 /**
92 * @in_flight: The # of preallocated pgtable pages in-flight
93 * for queued VM_BIND jobs.
94 */
95 atomic_t in_flight;
96 } prealloc_throttle;
97
98 /**
99 * @mm: Memory management for kernel managed VA allocations
100 *
101 * Only used for kernel managed VMs, unused for user managed VMs.
102 *
103 * Protected by vm lock. See msm_gem_lock_vm_and_obj(), for ex.
104 */
105 struct drm_mm mm;
106
107 /** @mmu: The mmu object which manages the pgtables */
108 struct msm_mmu *mmu;
109
110 /** @mmu_lock: Protects access to the mmu */
111 struct mutex mmu_lock;
112
113 /**
114 * @pid: For address spaces associated with a specific process, this
115 * will be non-NULL:
116 */
117 struct pid *pid;
118
119 /** @last_fence: Fence for last pending work scheduled on the VM */
120 struct dma_fence *last_fence;
121
122 /** @log: A log of recent VM updates */
123 struct msm_gem_vm_log_entry *log;
124
125 /** @log_shift: length of @log is (1 << @log_shift) */
126 uint32_t log_shift;
127
128 /** @log_idx: index of next @log entry to write */
129 uint32_t log_idx;
130
131 /** @faults: the number of GPU hangs associated with this address space */
132 int faults;
133
134 /** @managed: is this a kernel managed VM? */
135 bool managed;
136
137 /**
138 * @unusable: True if the VM has turned unusable because something
139 * bad happened during an asynchronous request.
140 *
141 * We don't try to recover from such failures, because this implies
142 * informing userspace about the specific operation that failed, and
143 * hoping the userspace driver can replay things from there. This all
144 * sounds very complicated for little gain.
145 *
146 * Instead, we should just flag the VM as unusable, and fail any
147 * further request targeting this VM.
148 *
149 * As an analogy, this would be mapped to a VK_ERROR_DEVICE_LOST
150 * situation, where the logical device needs to be re-created.
151 */
152 bool unusable;
153};
154#define to_msm_vm(x) container_of(x, struct msm_gem_vm, base)
155
156struct drm_gpuvm *
157msm_gem_vm_create(struct drm_device *drm, struct msm_mmu *mmu, const char *name,
158 u64 va_start, u64 va_size, bool managed);
159
160void msm_gem_vm_close(struct drm_gpuvm *gpuvm);
161void msm_gem_vm_unusable(struct drm_gpuvm *gpuvm);
162
163struct msm_fence_context;
164
165#define MSM_VMA_DUMP (DRM_GPUVA_USERBITS << 0)
166
167/**
168 * struct msm_gem_vma - a VMA mapping
169 *
170 * Represents a combination of a GEM object plus a VM.
171 */
172struct msm_gem_vma {
173 /** @base: inherit from drm_gpuva */
174 struct drm_gpuva base;
175
176 /**
177 * @node: mm node for VA allocation
178 *
179 * Only used by kernel managed VMs
180 */
181 struct drm_mm_node node;
182
183 /** @mapped: Is this VMA mapped? */
184 bool mapped;
185};
186#define to_msm_vma(x) container_of(x, struct msm_gem_vma, base)
187
188struct drm_gpuva *
189msm_gem_vma_new(struct drm_gpuvm *vm, struct drm_gem_object *obj,
190 u64 offset, u64 range_start, u64 range_end);
191void msm_gem_vma_unmap(struct drm_gpuva *vma, const char *reason);
192int msm_gem_vma_map(struct drm_gpuva *vma, int prot, struct sg_table *sgt);
193void msm_gem_vma_close(struct drm_gpuva *vma);
194
195struct msm_gem_object {
196 struct drm_gem_object base;
197
198 uint32_t flags;
199
200 /**
201 * madv: are the backing pages purgeable?
202 *
203 * Protected by obj lock and LRU lock
204 */
205 uint8_t madv;
206
207 /**
208 * count of active vmap'ing
209 */
210 uint8_t vmap_count;
211
212 /**
213 * Node in list of all objects (mainly for debugfs, protected by
214 * priv->obj_lock
215 */
216 struct list_head node;
217
218 struct page **pages;
219 struct sg_table *sgt;
220 void *vaddr;
221
222 char name[32]; /* Identifier to print for the debugfs files */
223
224 /* userspace metadata backchannel */
225 void *metadata;
226 u32 metadata_size;
227
228 /**
229 * pin_count: Number of times the pages are pinned
230 *
231 * Protected by LRU lock.
232 */
233 int pin_count;
234
235 /**
236 * @vma_ref: Reference count of VMA users.
237 *
238 * With the vm_bo/vma holding a reference to the GEM object, we'd
239 * otherwise have to actively tear down a VMA when, for example,
240 * a buffer is unpinned for scanout, vs. the pre-drm_gpuvm approach
241 * where a VMA did not hold a reference to the BO, but instead was
242 * implicitly torn down when the BO was freed.
243 *
244 * To regain the lazy VMA teardown, we use the @vma_ref. It is
245 * incremented for any of the following:
246 *
247 * 1) the BO is exported as a dma_buf
248 * 2) the BO has open userspace handle
249 *
250 * All of those conditions will hold an reference to the BO,
251 * preventing it from being freed. So lazily keeping around the
252 * VMA will not prevent the BO from being freed. (Or rather, the
253 * reference loop is harmless in this case.)
254 *
255 * When the @vma_ref drops to zero, then kms->vm VMA will be
256 * torn down.
257 */
258 atomic_t vma_ref;
259};
260#define to_msm_bo(x) container_of(x, struct msm_gem_object, base)
261
262void msm_gem_vma_get(struct drm_gem_object *obj);
263void msm_gem_vma_put(struct drm_gem_object *obj);
264
265uint64_t msm_gem_mmap_offset(struct drm_gem_object *obj);
266int msm_gem_prot(struct drm_gem_object *obj);
267int msm_gem_pin_vma_locked(struct drm_gem_object *obj, struct drm_gpuva *vma);
268void msm_gem_unpin_locked(struct drm_gem_object *obj);
269void msm_gem_unpin_active(struct drm_gem_object *obj);
270struct drm_gpuva *msm_gem_get_vma_locked(struct drm_gem_object *obj,
271 struct drm_gpuvm *vm);
272int msm_gem_get_iova(struct drm_gem_object *obj, struct drm_gpuvm *vm,
273 uint64_t *iova);
274int msm_gem_set_iova(struct drm_gem_object *obj, struct drm_gpuvm *vm,
275 uint64_t iova);
276int msm_gem_get_and_pin_iova_range(struct drm_gem_object *obj,
277 struct drm_gpuvm *vm, uint64_t *iova,
278 u64 range_start, u64 range_end);
279int msm_gem_get_and_pin_iova(struct drm_gem_object *obj, struct drm_gpuvm *vm,
280 uint64_t *iova);
281void msm_gem_unpin_iova(struct drm_gem_object *obj, struct drm_gpuvm *vm);
282void msm_gem_pin_obj_locked(struct drm_gem_object *obj);
283struct page **msm_gem_get_pages_locked(struct drm_gem_object *obj, unsigned madv);
284struct page **msm_gem_pin_pages_locked(struct drm_gem_object *obj);
285void msm_gem_unpin_pages_locked(struct drm_gem_object *obj);
286int msm_gem_dumb_create(struct drm_file *file, struct drm_device *dev,
287 struct drm_mode_create_dumb *args);
288int msm_gem_dumb_map_offset(struct drm_file *file, struct drm_device *dev,
289 uint32_t handle, uint64_t *offset);
290void *msm_gem_get_vaddr_locked(struct drm_gem_object *obj);
291void *msm_gem_get_vaddr(struct drm_gem_object *obj);
292void *msm_gem_get_vaddr_active(struct drm_gem_object *obj);
293void msm_gem_put_vaddr_locked(struct drm_gem_object *obj);
294void msm_gem_put_vaddr(struct drm_gem_object *obj);
295int msm_gem_madvise(struct drm_gem_object *obj, unsigned madv);
296bool msm_gem_active(struct drm_gem_object *obj);
297int msm_gem_cpu_prep(struct drm_gem_object *obj, uint32_t op, ktime_t *timeout);
298int msm_gem_cpu_fini(struct drm_gem_object *obj);
299int msm_gem_new_handle(struct drm_device *dev, struct drm_file *file,
300 size_t size, uint32_t flags, uint32_t *handle, char *name);
301struct drm_gem_object *msm_gem_new(struct drm_device *dev,
302 size_t size, uint32_t flags);
303void *msm_gem_kernel_new(struct drm_device *dev, size_t size, uint32_t flags,
304 struct drm_gpuvm *vm, struct drm_gem_object **bo,
305 uint64_t *iova);
306void msm_gem_kernel_put(struct drm_gem_object *bo, struct drm_gpuvm *vm);
307struct drm_gem_object *msm_gem_import(struct drm_device *dev,
308 struct dma_buf *dmabuf, struct sg_table *sgt);
309__printf(2, 3)
310void msm_gem_object_set_name(struct drm_gem_object *bo, const char *fmt, ...);
311
312#ifdef CONFIG_DEBUG_FS
313struct msm_gem_stats {
314 struct {
315 unsigned count;
316 size_t size;
317 } all, active, resident, purgeable, purged;
318};
319
320void msm_gem_describe(struct drm_gem_object *obj, struct seq_file *m,
321 struct msm_gem_stats *stats);
322void msm_gem_describe_objects(struct list_head *list, struct seq_file *m);
323#endif
324
325static inline void
326msm_gem_lock(struct drm_gem_object *obj)
327{
328 dma_resv_lock(obj: obj->resv, NULL);
329}
330
331static inline bool __must_check
332msm_gem_trylock(struct drm_gem_object *obj)
333{
334 return dma_resv_trylock(obj: obj->resv);
335}
336
337static inline int
338msm_gem_lock_interruptible(struct drm_gem_object *obj)
339{
340 return dma_resv_lock_interruptible(obj: obj->resv, NULL);
341}
342
343static inline void
344msm_gem_unlock(struct drm_gem_object *obj)
345{
346 dma_resv_unlock(obj: obj->resv);
347}
348
349/**
350 * msm_gem_lock_vm_and_obj() - Helper to lock an obj + VM
351 * @exec: the exec context helper which will be initalized
352 * @obj: the GEM object to lock
353 * @vm: the VM to lock
354 *
355 * Operations which modify a VM frequently need to lock both the VM and
356 * the object being mapped/unmapped/etc. This helper uses drm_exec to
357 * acquire both locks, dealing with potential deadlock/backoff scenarios
358 * which arise when multiple locks are involved.
359 */
360static inline int
361msm_gem_lock_vm_and_obj(struct drm_exec *exec,
362 struct drm_gem_object *obj,
363 struct drm_gpuvm *vm)
364{
365 int ret = 0;
366
367 drm_exec_init(exec, flags: 0, nr: 2);
368 drm_exec_until_all_locked (exec) {
369 ret = drm_exec_lock_obj(exec, drm_gpuvm_resv_obj(vm));
370 if (!ret && (obj->resv != drm_gpuvm_resv(vm)))
371 ret = drm_exec_lock_obj(exec, obj);
372 drm_exec_retry_on_contention(exec);
373 if (GEM_WARN_ON(ret))
374 break;
375 }
376
377 return ret;
378}
379
380static inline void
381msm_gem_assert_locked(struct drm_gem_object *obj)
382{
383 /*
384 * Destroying the object is a special case.. msm_gem_free_object()
385 * calls many things that WARN_ON if the obj lock is not held. But
386 * acquiring the obj lock in msm_gem_free_object() can cause a
387 * locking order inversion between reservation_ww_class_mutex and
388 * fs_reclaim.
389 *
390 * This deadlock is not actually possible, because no one should
391 * be already holding the lock when msm_gem_free_object() is called.
392 * Unfortunately lockdep is not aware of this detail. So when the
393 * refcount drops to zero, we pretend it is already locked.
394 */
395 lockdep_assert_once(
396 (kref_read(&obj->refcount) == 0) ||
397 (lockdep_is_held(&obj->resv->lock.base) != LOCK_STATE_NOT_HELD)
398 );
399}
400
401/* imported/exported objects are not purgeable: */
402static inline bool is_unpurgeable(struct msm_gem_object *msm_obj)
403{
404 return drm_gem_is_imported(obj: &msm_obj->base) || msm_obj->pin_count;
405}
406
407static inline bool is_purgeable(struct msm_gem_object *msm_obj)
408{
409 return (msm_obj->madv == MSM_MADV_DONTNEED) && msm_obj->sgt &&
410 !is_unpurgeable(msm_obj);
411}
412
413static inline bool is_vunmapable(struct msm_gem_object *msm_obj)
414{
415 msm_gem_assert_locked(obj: &msm_obj->base);
416 return (msm_obj->vmap_count == 0) && msm_obj->vaddr;
417}
418
419static inline bool is_unevictable(struct msm_gem_object *msm_obj)
420{
421 return is_unpurgeable(msm_obj) || msm_obj->vaddr;
422}
423
424void msm_gem_purge(struct drm_gem_object *obj);
425void msm_gem_evict(struct drm_gem_object *obj);
426void msm_gem_vunmap(struct drm_gem_object *obj);
427
428/* Created per submit-ioctl, to track bo's and cmdstream bufs, etc,
429 * associated with the cmdstream submission for synchronization (and
430 * make it easier to unwind when things go wrong, etc).
431 */
432struct msm_gem_submit {
433 struct drm_sched_job base;
434 struct kref ref;
435 struct drm_device *dev;
436 struct msm_gpu *gpu;
437 struct drm_gpuvm *vm;
438 struct list_head node; /* node in ring submit list */
439 struct drm_exec exec;
440 uint32_t seqno; /* Sequence number of the submit on the ring */
441
442 /* Hw fence, which is created when the scheduler executes the job, and
443 * is signaled when the hw finishes (via seqno write from cmdstream)
444 */
445 struct dma_fence *hw_fence;
446
447 /* Userspace visible fence, which is signaled by the scheduler after
448 * the hw_fence is signaled.
449 */
450 struct dma_fence *user_fence;
451
452 int fence_id; /* key into queue->fence_idr */
453 struct msm_gpu_submitqueue *queue;
454 struct pid *pid; /* submitting process */
455 bool bos_pinned : 1;
456 bool fault_dumped:1;/* Limit devcoredump dumping to one per submit */
457 bool in_rb : 1; /* "sudo" mode, copy cmds into RB */
458 struct msm_ringbuffer *ring;
459 unsigned int nr_cmds;
460 unsigned int nr_bos;
461 u32 ident; /* A "identifier" for the submit for logging */
462 struct {
463 uint32_t type;
464 uint32_t size; /* in dwords */
465 uint64_t iova;
466 uint32_t offset;/* in dwords */
467 uint32_t idx; /* cmdstream buffer idx in bos[] */
468 uint32_t nr_relocs;
469 struct drm_msm_gem_submit_reloc *relocs;
470 } *cmd; /* array of size nr_cmds */
471 struct {
472 uint32_t flags;
473 union {
474 struct drm_gem_object *obj;
475 uint32_t handle;
476 };
477 struct drm_gpuvm_bo *vm_bo;
478 uint64_t iova;
479 } bos[];
480};
481
482static inline struct msm_gem_submit *to_msm_submit(struct drm_sched_job *job)
483{
484 return container_of(job, struct msm_gem_submit, base);
485}
486
487void __msm_gem_submit_destroy(struct kref *kref);
488
489static inline void msm_gem_submit_get(struct msm_gem_submit *submit)
490{
491 kref_get(kref: &submit->ref);
492}
493
494static inline void msm_gem_submit_put(struct msm_gem_submit *submit)
495{
496 kref_put(kref: &submit->ref, release: __msm_gem_submit_destroy);
497}
498
499void msm_submit_retire(struct msm_gem_submit *submit);
500
501#endif /* __MSM_GEM_H__ */
502

source code of linux/drivers/gpu/drm/msm/msm_gem.h