xref: /linux/drivers/gpu/drm/v3d/v3d_mmu.c (revision bba2c3615bd6cfee7456d1130f2e6b01b3f4e9ba)
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 <drm/drm_print.h>
22 
23 #include "v3d_drv.h"
24 #include "v3d_regs.h"
25 
26 /* Note: All PTEs for the 64KB bigpage or 1MB superpage must be filled
27  * with the bigpage/superpage bit set.
28  */
29 #define V3D_PTE_SUPERPAGE BIT(31)
30 #define V3D_PTE_BIGPAGE BIT(30)
31 #define V3D_PTE_WRITEABLE BIT(29)
32 #define V3D_PTE_VALID BIT(28)
33 
34 static bool v3d_mmu_is_aligned(u32 page, u32 page_address, size_t alignment)
35 {
36 	return IS_ALIGNED(page, alignment >> V3D_MMU_PAGE_SHIFT) &&
37 		IS_ALIGNED(page_address, alignment >> V3D_MMU_PAGE_SHIFT);
38 }
39 
40 /*
41  * Issue the MMUC flush and TLB clear unconditionally. The caller must
42  * already know that V3D is reachable. In particular, this is used from
43  * the runtime resume callback.
44  */
45 static int v3d_mmu_flush_all_locked(struct v3d_dev *v3d)
46 {
47 	int ret;
48 
49 	V3D_WRITE(V3D_MMUC_CONTROL, V3D_MMUC_CONTROL_FLUSH |
50 		  V3D_MMUC_CONTROL_ENABLE);
51 
52 	ret = wait_for(!(V3D_READ(V3D_MMUC_CONTROL) &
53 			 V3D_MMUC_CONTROL_FLUSHING), 100);
54 	if (ret) {
55 		dev_err(v3d->drm.dev, "MMUC flush wait idle failed\n");
56 		return ret;
57 	}
58 
59 	V3D_WRITE(V3D_MMU_CTL, V3D_READ(V3D_MMU_CTL) |
60 		  V3D_MMU_CTL_TLB_CLEAR);
61 
62 	ret = wait_for(!(V3D_READ(V3D_MMU_CTL) &
63 			 V3D_MMU_CTL_TLB_CLEARING), 100);
64 	if (ret)
65 		dev_err(v3d->drm.dev, "MMU TLB clear wait idle failed\n");
66 
67 	return ret;
68 }
69 
70 int v3d_mmu_flush_all(struct v3d_dev *v3d)
71 {
72 	int ret;
73 
74 	/* Flush the PTs only if we're already awake */
75 	if (!pm_runtime_get_if_active(v3d->drm.dev))
76 		return 0;
77 
78 	ret = v3d_mmu_flush_all_locked(v3d);
79 
80 	v3d_pm_runtime_put(v3d);
81 	return ret;
82 }
83 
84 int v3d_mmu_set_page_table(struct v3d_dev *v3d)
85 {
86 	V3D_WRITE(V3D_MMU_PT_PA_BASE, v3d->pt_paddr >> V3D_MMU_PAGE_SHIFT);
87 	V3D_WRITE(V3D_MMU_CTL,
88 		  V3D_MMU_CTL_ENABLE |
89 		  V3D_MMU_CTL_PT_INVALID_ENABLE |
90 		  V3D_MMU_CTL_PT_INVALID_ABORT |
91 		  V3D_MMU_CTL_PT_INVALID_INT |
92 		  V3D_MMU_CTL_WRITE_VIOLATION_ABORT |
93 		  V3D_MMU_CTL_WRITE_VIOLATION_INT |
94 		  V3D_MMU_CTL_CAP_EXCEEDED_ABORT |
95 		  V3D_MMU_CTL_CAP_EXCEEDED_INT);
96 	V3D_WRITE(V3D_MMU_ILLEGAL_ADDR,
97 		  (v3d->mmu_scratch_paddr >> V3D_MMU_PAGE_SHIFT) |
98 		  V3D_MMU_ILLEGAL_ADDR_ENABLE);
99 	V3D_WRITE(V3D_MMUC_CONTROL, V3D_MMUC_CONTROL_ENABLE);
100 
101 	return v3d_mmu_flush_all_locked(v3d);
102 }
103 
104 void v3d_mmu_insert_ptes(struct v3d_bo *bo)
105 {
106 	struct drm_gem_shmem_object *shmem_obj = &bo->base;
107 	struct v3d_dev *v3d = to_v3d_dev(shmem_obj->base.dev);
108 	u32 page = bo->node.start;
109 	struct scatterlist *sgl;
110 	unsigned int count;
111 
112 	for_each_sgtable_dma_sg(shmem_obj->sgt, sgl, count) {
113 		dma_addr_t dma_addr = sg_dma_address(sgl);
114 		u32 pfn = dma_addr >> V3D_MMU_PAGE_SHIFT;
115 		unsigned int len = sg_dma_len(sgl);
116 
117 		while (len > 0) {
118 			u32 page_prot = V3D_PTE_WRITEABLE | V3D_PTE_VALID;
119 			u32 page_address = page_prot | pfn;
120 			unsigned int i, page_size;
121 
122 			BUG_ON(pfn + V3D_PAGE_FACTOR >= BIT(24));
123 
124 			if (len >= SZ_1M &&
125 			    v3d_mmu_is_aligned(page, page_address, SZ_1M)) {
126 				page_size = SZ_1M;
127 				page_address |= V3D_PTE_SUPERPAGE;
128 			} else if (len >= SZ_64K &&
129 				   v3d_mmu_is_aligned(page, page_address, SZ_64K)) {
130 				page_size = SZ_64K;
131 				page_address |= V3D_PTE_BIGPAGE;
132 			} else {
133 				page_size = SZ_4K;
134 			}
135 
136 			for (i = 0; i < page_size >> V3D_MMU_PAGE_SHIFT; i++) {
137 				v3d->pt[page++] = page_address + i;
138 				pfn++;
139 			}
140 
141 			len -= page_size;
142 		}
143 	}
144 
145 	WARN_ON_ONCE(page - bo->node.start !=
146 		     shmem_obj->base.size >> V3D_MMU_PAGE_SHIFT);
147 
148 	if (v3d_mmu_flush_all(v3d))
149 		drm_err(&v3d->drm, "MMU flush timeout\n");
150 }
151 
152 void v3d_mmu_remove_ptes(struct v3d_bo *bo)
153 {
154 	struct v3d_dev *v3d = to_v3d_dev(bo->base.base.dev);
155 	u32 npages = bo->base.base.size >> V3D_MMU_PAGE_SHIFT;
156 	u32 page;
157 
158 	for (page = bo->node.start; page < bo->node.start + npages; page++)
159 		v3d->pt[page] = 0;
160 
161 	if (v3d_mmu_flush_all(v3d))
162 		drm_err(&v3d->drm, "MMU flush timeout\n");
163 }
164