xref: /openbmc/linux/drivers/gpu/drm/v3d/v3d_mmu.c (revision 675aaf05)
1 // SPDX-License-Identifier: GPL-2.0+
2 /* Copyright (C) 2017-2018 Broadcom */
3 
4 /**
5  * DOC: Broadcom V3D MMU
6  *
7  * The V3D 3.x hardware (compared to VC4) now includes an MMU.  It has
8  * a single level of page tables for the V3D's 4GB address space to
9  * map to AXI bus addresses, thus it could need up to 4MB of
10  * physically contiguous memory to store the PTEs.
11  *
12  * Because the 4MB of contiguous memory for page tables is precious,
13  * and switching between them is expensive, we load all BOs into the
14  * same 4GB address space.
15  *
16  * To protect clients from each other, we should use the GMP to
17  * quickly mask out (at 128kb granularity) what pages are available to
18  * each client.  This is not yet implemented.
19  */
20 
21 #include "v3d_drv.h"
22 #include "v3d_regs.h"
23 
24 #define V3D_MMU_PAGE_SHIFT 12
25 
26 /* Note: All PTEs for the 1MB superpage must be filled with the
27  * superpage bit set.
28  */
29 #define V3D_PTE_SUPERPAGE BIT(31)
30 #define V3D_PTE_WRITEABLE BIT(29)
31 #define V3D_PTE_VALID BIT(28)
32 
33 static int v3d_mmu_flush_all(struct v3d_dev *v3d)
34 {
35 	int ret;
36 
37 	/* Make sure that another flush isn't already running when we
38 	 * start this one.
39 	 */
40 	ret = wait_for(!(V3D_READ(V3D_MMU_CTL) &
41 			 V3D_MMU_CTL_TLB_CLEARING), 100);
42 	if (ret)
43 		dev_err(v3d->dev, "TLB clear wait idle pre-wait failed\n");
44 
45 	V3D_WRITE(V3D_MMU_CTL, V3D_READ(V3D_MMU_CTL) |
46 		  V3D_MMU_CTL_TLB_CLEAR);
47 
48 	V3D_WRITE(V3D_MMUC_CONTROL,
49 		  V3D_MMUC_CONTROL_FLUSH |
50 		  V3D_MMUC_CONTROL_ENABLE);
51 
52 	ret = wait_for(!(V3D_READ(V3D_MMU_CTL) &
53 			 V3D_MMU_CTL_TLB_CLEARING), 100);
54 	if (ret) {
55 		dev_err(v3d->dev, "TLB clear wait idle failed\n");
56 		return ret;
57 	}
58 
59 	ret = wait_for(!(V3D_READ(V3D_MMUC_CONTROL) &
60 			 V3D_MMUC_CONTROL_FLUSHING), 100);
61 	if (ret)
62 		dev_err(v3d->dev, "MMUC flush wait idle failed\n");
63 
64 	return ret;
65 }
66 
67 int v3d_mmu_set_page_table(struct v3d_dev *v3d)
68 {
69 	V3D_WRITE(V3D_MMU_PT_PA_BASE, v3d->pt_paddr >> V3D_MMU_PAGE_SHIFT);
70 	V3D_WRITE(V3D_MMU_CTL,
71 		  V3D_MMU_CTL_ENABLE |
72 		  V3D_MMU_CTL_PT_INVALID |
73 		  V3D_MMU_CTL_PT_INVALID_ABORT |
74 		  V3D_MMU_CTL_WRITE_VIOLATION_ABORT |
75 		  V3D_MMU_CTL_CAP_EXCEEDED_ABORT);
76 	V3D_WRITE(V3D_MMU_ILLEGAL_ADDR,
77 		  (v3d->mmu_scratch_paddr >> V3D_MMU_PAGE_SHIFT) |
78 		  V3D_MMU_ILLEGAL_ADDR_ENABLE);
79 	V3D_WRITE(V3D_MMUC_CONTROL, V3D_MMUC_CONTROL_ENABLE);
80 
81 	return v3d_mmu_flush_all(v3d);
82 }
83 
84 void v3d_mmu_insert_ptes(struct v3d_bo *bo)
85 {
86 	struct drm_gem_shmem_object *shmem_obj = &bo->base;
87 	struct v3d_dev *v3d = to_v3d_dev(shmem_obj->base.dev);
88 	u32 page = bo->node.start;
89 	u32 page_prot = V3D_PTE_WRITEABLE | V3D_PTE_VALID;
90 	unsigned int count;
91 	struct scatterlist *sgl;
92 
93 	for_each_sg(shmem_obj->sgt->sgl, sgl, shmem_obj->sgt->nents, count) {
94 		u32 page_address = sg_dma_address(sgl) >> V3D_MMU_PAGE_SHIFT;
95 		u32 pte = page_prot | page_address;
96 		u32 i;
97 
98 		BUG_ON(page_address + (sg_dma_len(sgl) >> V3D_MMU_PAGE_SHIFT) >=
99 		       BIT(24));
100 
101 		for (i = 0; i < sg_dma_len(sgl) >> V3D_MMU_PAGE_SHIFT; i++)
102 			v3d->pt[page++] = pte + i;
103 	}
104 
105 	WARN_ON_ONCE(page - bo->node.start !=
106 		     shmem_obj->base.size >> V3D_MMU_PAGE_SHIFT);
107 
108 	if (v3d_mmu_flush_all(v3d))
109 		dev_err(v3d->dev, "MMU flush timeout\n");
110 }
111 
112 void v3d_mmu_remove_ptes(struct v3d_bo *bo)
113 {
114 	struct v3d_dev *v3d = to_v3d_dev(bo->base.base.dev);
115 	u32 npages = bo->base.base.size >> V3D_MMU_PAGE_SHIFT;
116 	u32 page;
117 
118 	for (page = bo->node.start; page < bo->node.start + npages; page++)
119 		v3d->pt[page] = 0;
120 
121 	if (v3d_mmu_flush_all(v3d))
122 		dev_err(v3d->dev, "MMU flush timeout\n");
123 }
124