xref: /linux/drivers/gpu/drm/xe/xe_migrate_doc.h (revision 2672031b20f6681514bef14ddcfe8c62c2757d11)
1 /* SPDX-License-Identifier: MIT */
2 /*
3  * Copyright © 2022 Intel Corporation
4  */
5 
6 #ifndef _XE_MIGRATE_DOC_H_
7 #define _XE_MIGRATE_DOC_H_
8 
9 /**
10  * DOC: Migrate Layer
11  *
12  * The XE migrate layer is used generate jobs which can copy memory (eviction),
13  * clear memory, or program tables (binds). This layer exists in every GT, has
14  * a migrate engine, and uses a special VM for all generated jobs.
15  *
16  * Special VM details
17  * ==================
18  *
19  * The special VM is configured with a page structure where we can dynamically
20  * map BOs which need to be copied and cleared, dynamically map other VM's page
21  * table BOs for updates, and identity map the entire device's VRAM with 1 GB
22  * pages.
23  *
24  * Currently the page structure consists of 32 physical pages with 16 being
25  * reserved for BO mapping during copies and clear, 1 reserved for kernel binds,
26  * several pages are needed to setup the identity mappings (exact number based
27  * on how many bits of address space the device has), and the rest are reserved
28  * user bind operations.
29  *
30  * TODO: Diagram of layout
31  *
32  * Bind jobs
33  * =========
34  *
35  * A bind job consist of two batches and runs either on the migrate engine
36  * (kernel binds) or the bind engine passed in (user binds). In both cases the
37  * VM of the engine is the migrate VM.
38  *
39  * The first batch is used to update the migration VM page structure to point to
40  * the bind VM page table BOs which need to be updated. A physical page is
41  * required for this. If it is a user bind, the page is allocated from pool of
42  * pages reserved user bind operations with drm_suballoc managing this pool. If
43  * it is a kernel bind, the page reserved for kernel binds is used.
44  *
45  * The first batch is only required for devices without VRAM as when the device
46  * has VRAM the bind VM page table BOs are in VRAM and the identity mapping can
47  * be used.
48  *
49  * The second batch is used to program page table updated in the bind VM. Why
50  * not just one batch? Well the TLBs need to be invalidated between these two
51  * batches and that only can be done from the ring.
52  *
53  * When the bind job complete, the page allocated is returned the pool of pages
54  * reserved for user bind operations if a user bind. No need do this for kernel
55  * binds as the reserved kernel page is serially used by each job.
56  *
57  * Copy / clear jobs
58  * =================
59  *
60  * A copy or clear job consist of two batches and runs on the migrate engine.
61  *
62  * Like binds, the first batch is used update the migration VM page structure.
63  * In copy jobs, we need to map the source and destination of the BO into page
64  * the structure. In clear jobs, we just need to add 1 mapping of BO into the
65  * page structure. We use the 16 reserved pages in migration VM for mappings,
66  * this gives us a maximum copy size of 16 MB and maximum clear size of 32 MB.
67  *
68  * The second batch is used do either do the copy or clear. Again similar to
69  * binds, two batches are required as the TLBs need to be invalidated from the
70  * ring between the batches.
71  *
72  * More than one job will be generated if the BO is larger than maximum copy /
73  * clear size.
74  *
75  * Future work
76  * ===========
77  *
78  * Update copy and clear code to use identity mapped VRAM.
79  *
80  * Can we rework the use of the pages async binds to use all the entries in each
81  * page?
82  *
83  * Using large pages for sysmem mappings.
84  *
85  * Is it possible to identity map the sysmem? We should explore this.
86  */
87 
88 #endif
89