1 #ifndef __NVKM_VMM_H__
2 #define __NVKM_VMM_H__
3 #include "priv.h"
4 #include <core/memory.h>
5 enum nvkm_memory_target;
6 
7 struct nvkm_vmm_pt {
8 	/* Some GPUs have a mapping level with a dual page tables to
9 	 * support large and small pages in the same address-range.
10 	 *
11 	 * We track the state of both page tables in one place, which
12 	 * is why there's multiple PT pointers/refcounts here.
13 	 */
14 	struct nvkm_mmu_pt *pt[2];
15 	u32 refs[2];
16 
17 	/* Page size handled by this PT.
18 	 *
19 	 * Tesla backend needs to know this when writinge PDEs,
20 	 * otherwise unnecessary.
21 	 */
22 	u8 page;
23 
24 	/* Entire page table sparse.
25 	 *
26 	 * Used to propagate sparseness to child page tables.
27 	 */
28 	bool sparse:1;
29 
30 	/* Tracking for page directories.
31 	 *
32 	 * The array is indexed by PDE, and will either point to the
33 	 * child page table, or indicate the PDE is marked as sparse.
34 	 **/
35 #define NVKM_VMM_PDE_INVALID(pde) IS_ERR_OR_NULL(pde)
36 #define NVKM_VMM_PDE_SPARSED(pde) IS_ERR(pde)
37 #define NVKM_VMM_PDE_SPARSE       ERR_PTR(-EBUSY)
38 	struct nvkm_vmm_pt **pde;
39 
40 	/* Tracking for dual page tables.
41 	 *
42 	 * There's one entry for each LPTE, keeping track of whether
43 	 * there are valid SPTEs in the same address-range.
44 	 *
45 	 * This information is used to manage LPTE state transitions.
46 	 */
47 #define NVKM_VMM_PTE_SPARSE 0x80
48 #define NVKM_VMM_PTE_VALID  0x40
49 #define NVKM_VMM_PTE_SPTES  0x3f
50 	u8 pte[];
51 };
52 
53 typedef void (*nvkm_vmm_pxe_func)(struct nvkm_vmm *,
54 				  struct nvkm_mmu_pt *, u32 ptei, u32 ptes);
55 typedef void (*nvkm_vmm_pde_func)(struct nvkm_vmm *,
56 				  struct nvkm_vmm_pt *, u32 pdei);
57 typedef void (*nvkm_vmm_pte_func)(struct nvkm_vmm *, struct nvkm_mmu_pt *,
58 				  u32 ptei, u32 ptes, struct nvkm_vmm_map *);
59 
60 struct nvkm_vmm_desc_func {
61 	nvkm_vmm_pxe_func invalid;
62 	nvkm_vmm_pxe_func unmap;
63 	nvkm_vmm_pxe_func sparse;
64 
65 	nvkm_vmm_pde_func pde;
66 
67 	nvkm_vmm_pte_func mem;
68 	nvkm_vmm_pte_func dma;
69 	nvkm_vmm_pte_func sgl;
70 
71 	nvkm_vmm_pte_func pfn;
72 	bool (*pfn_clear)(struct nvkm_vmm *, struct nvkm_mmu_pt *, u32 ptei, u32 ptes);
73 	nvkm_vmm_pxe_func pfn_unmap;
74 };
75 
76 extern const struct nvkm_vmm_desc_func gf100_vmm_pgd;
77 void gf100_vmm_pgd_pde(struct nvkm_vmm *, struct nvkm_vmm_pt *, u32);
78 extern const struct nvkm_vmm_desc_func gf100_vmm_pgt;
79 void gf100_vmm_pgt_unmap(struct nvkm_vmm *, struct nvkm_mmu_pt *, u32, u32);
80 void gf100_vmm_pgt_mem(struct nvkm_vmm *, struct nvkm_mmu_pt *, u32, u32,
81 		       struct nvkm_vmm_map *);
82 void gf100_vmm_pgt_dma(struct nvkm_vmm *, struct nvkm_mmu_pt *, u32, u32,
83 		       struct nvkm_vmm_map *);
84 void gf100_vmm_pgt_sgl(struct nvkm_vmm *, struct nvkm_mmu_pt *, u32, u32,
85 		       struct nvkm_vmm_map *);
86 
87 void gk104_vmm_lpt_invalid(struct nvkm_vmm *, struct nvkm_mmu_pt *, u32, u32);
88 
89 struct nvkm_vmm_desc {
90 	enum {
91 		PGD,
92 		PGT,
93 		SPT,
94 		LPT,
95 	} type;
96 	u8 bits;	/* VMA bits covered by PT. */
97 	u8 size;	/* Bytes-per-PTE. */
98 	u32 align;	/* PT address alignment. */
99 	const struct nvkm_vmm_desc_func *func;
100 };
101 
102 extern const struct nvkm_vmm_desc nv50_vmm_desc_12[];
103 extern const struct nvkm_vmm_desc nv50_vmm_desc_16[];
104 
105 extern const struct nvkm_vmm_desc gk104_vmm_desc_16_12[];
106 extern const struct nvkm_vmm_desc gk104_vmm_desc_16_16[];
107 extern const struct nvkm_vmm_desc gk104_vmm_desc_17_12[];
108 extern const struct nvkm_vmm_desc gk104_vmm_desc_17_17[];
109 
110 extern const struct nvkm_vmm_desc gm200_vmm_desc_16_12[];
111 extern const struct nvkm_vmm_desc gm200_vmm_desc_16_16[];
112 extern const struct nvkm_vmm_desc gm200_vmm_desc_17_12[];
113 extern const struct nvkm_vmm_desc gm200_vmm_desc_17_17[];
114 
115 extern const struct nvkm_vmm_desc gp100_vmm_desc_12[];
116 extern const struct nvkm_vmm_desc gp100_vmm_desc_16[];
117 
118 struct nvkm_vmm_page {
119 	u8 shift;
120 	const struct nvkm_vmm_desc *desc;
121 #define NVKM_VMM_PAGE_SPARSE                                               0x01
122 #define NVKM_VMM_PAGE_VRAM                                                 0x02
123 #define NVKM_VMM_PAGE_HOST                                                 0x04
124 #define NVKM_VMM_PAGE_COMP                                                 0x08
125 #define NVKM_VMM_PAGE_Sxxx                                (NVKM_VMM_PAGE_SPARSE)
126 #define NVKM_VMM_PAGE_xVxx                                  (NVKM_VMM_PAGE_VRAM)
127 #define NVKM_VMM_PAGE_SVxx             (NVKM_VMM_PAGE_Sxxx | NVKM_VMM_PAGE_VRAM)
128 #define NVKM_VMM_PAGE_xxHx                                  (NVKM_VMM_PAGE_HOST)
129 #define NVKM_VMM_PAGE_SxHx             (NVKM_VMM_PAGE_Sxxx | NVKM_VMM_PAGE_HOST)
130 #define NVKM_VMM_PAGE_xVHx             (NVKM_VMM_PAGE_xVxx | NVKM_VMM_PAGE_HOST)
131 #define NVKM_VMM_PAGE_SVHx             (NVKM_VMM_PAGE_SVxx | NVKM_VMM_PAGE_HOST)
132 #define NVKM_VMM_PAGE_xVxC             (NVKM_VMM_PAGE_xVxx | NVKM_VMM_PAGE_COMP)
133 #define NVKM_VMM_PAGE_SVxC             (NVKM_VMM_PAGE_SVxx | NVKM_VMM_PAGE_COMP)
134 #define NVKM_VMM_PAGE_xxHC             (NVKM_VMM_PAGE_xxHx | NVKM_VMM_PAGE_COMP)
135 #define NVKM_VMM_PAGE_SxHC             (NVKM_VMM_PAGE_SxHx | NVKM_VMM_PAGE_COMP)
136 	u8 type;
137 };
138 
139 struct nvkm_vmm_func {
140 	int (*join)(struct nvkm_vmm *, struct nvkm_memory *inst);
141 	void (*part)(struct nvkm_vmm *, struct nvkm_memory *inst);
142 
143 	int (*aper)(enum nvkm_memory_target);
144 	int (*valid)(struct nvkm_vmm *, void *argv, u32 argc,
145 		     struct nvkm_vmm_map *);
146 	void (*flush)(struct nvkm_vmm *, int depth);
147 
148 	int (*mthd)(struct nvkm_vmm *, struct nvkm_client *,
149 		    u32 mthd, void *argv, u32 argc);
150 
151 	void (*invalidate_pdb)(struct nvkm_vmm *, u64 addr);
152 
153 	u64 page_block;
154 	const struct nvkm_vmm_page page[];
155 };
156 
157 struct nvkm_vmm_join {
158 	struct nvkm_memory *inst;
159 	struct list_head head;
160 };
161 
162 int nvkm_vmm_new_(const struct nvkm_vmm_func *, struct nvkm_mmu *,
163 		  u32 pd_header, bool managed, u64 addr, u64 size,
164 		  struct lock_class_key *, const char *name,
165 		  struct nvkm_vmm **);
166 struct nvkm_vma *nvkm_vma_new(u64 addr, u64 size);
167 struct nvkm_vma *nvkm_vmm_node_search(struct nvkm_vmm *, u64 addr);
168 struct nvkm_vma *nvkm_vmm_node_split(struct nvkm_vmm *, struct nvkm_vma *,
169 				     u64 addr, u64 size);
170 int nvkm_vmm_get_locked(struct nvkm_vmm *, bool getref, bool mapref,
171 			bool sparse, u8 page, u8 align, u64 size,
172 			struct nvkm_vma **pvma);
173 void nvkm_vmm_put_locked(struct nvkm_vmm *, struct nvkm_vma *);
174 void nvkm_vmm_unmap_locked(struct nvkm_vmm *, struct nvkm_vma *, bool pfn);
175 void nvkm_vmm_unmap_region(struct nvkm_vmm *, struct nvkm_vma *);
176 
177 int nvkm_vmm_raw_get(struct nvkm_vmm *vmm, u64 addr, u64 size, u8 refd);
178 void nvkm_vmm_raw_put(struct nvkm_vmm *vmm, u64 addr, u64 size, u8 refd);
179 void nvkm_vmm_raw_unmap(struct nvkm_vmm *vmm, u64 addr, u64 size,
180 			bool sparse, u8 refd);
181 int nvkm_vmm_raw_sparse(struct nvkm_vmm *, u64 addr, u64 size, bool ref);
182 
183 static inline bool
nvkm_vmm_in_managed_range(struct nvkm_vmm * vmm,u64 start,u64 size)184 nvkm_vmm_in_managed_range(struct nvkm_vmm *vmm, u64 start, u64 size)
185 {
186 	u64 p_start = vmm->managed.p.addr;
187 	u64 p_end = p_start + vmm->managed.p.size;
188 	u64 n_start = vmm->managed.n.addr;
189 	u64 n_end = n_start + vmm->managed.n.size;
190 	u64 end = start + size;
191 
192 	if (start >= p_start && end <= p_end)
193 		return true;
194 
195 	if (start >= n_start && end <= n_end)
196 		return true;
197 
198 	return false;
199 }
200 
201 #define NVKM_VMM_PFN_ADDR                                 0xfffffffffffff000ULL
202 #define NVKM_VMM_PFN_ADDR_SHIFT                                              12
203 #define NVKM_VMM_PFN_APER                                 0x00000000000000f0ULL
204 #define NVKM_VMM_PFN_HOST                                 0x0000000000000000ULL
205 #define NVKM_VMM_PFN_VRAM                                 0x0000000000000010ULL
206 #define NVKM_VMM_PFN_A					  0x0000000000000004ULL
207 #define NVKM_VMM_PFN_W                                    0x0000000000000002ULL
208 #define NVKM_VMM_PFN_V                                    0x0000000000000001ULL
209 #define NVKM_VMM_PFN_NONE                                 0x0000000000000000ULL
210 
211 int nvkm_vmm_pfn_map(struct nvkm_vmm *, u8 page, u64 addr, u64 size, u64 *pfn);
212 int nvkm_vmm_pfn_unmap(struct nvkm_vmm *, u64 addr, u64 size);
213 
214 struct nvkm_vma *nvkm_vma_tail(struct nvkm_vma *, u64 tail);
215 
216 int nv04_vmm_new_(const struct nvkm_vmm_func *, struct nvkm_mmu *, u32,
217 		  bool, u64, u64, void *, u32, struct lock_class_key *,
218 		  const char *, struct nvkm_vmm **);
219 int nv04_vmm_valid(struct nvkm_vmm *, void *, u32, struct nvkm_vmm_map *);
220 
221 int nv50_vmm_join(struct nvkm_vmm *, struct nvkm_memory *);
222 void nv50_vmm_part(struct nvkm_vmm *, struct nvkm_memory *);
223 int nv50_vmm_valid(struct nvkm_vmm *, void *, u32, struct nvkm_vmm_map *);
224 void nv50_vmm_flush(struct nvkm_vmm *, int);
225 
226 int gf100_vmm_new_(const struct nvkm_vmm_func *, const struct nvkm_vmm_func *,
227 		   struct nvkm_mmu *, bool, u64, u64, void *, u32,
228 		   struct lock_class_key *, const char *, struct nvkm_vmm **);
229 int gf100_vmm_join_(struct nvkm_vmm *, struct nvkm_memory *, u64 base);
230 int gf100_vmm_join(struct nvkm_vmm *, struct nvkm_memory *);
231 void gf100_vmm_part(struct nvkm_vmm *, struct nvkm_memory *);
232 int gf100_vmm_aper(enum nvkm_memory_target);
233 int gf100_vmm_valid(struct nvkm_vmm *, void *, u32, struct nvkm_vmm_map *);
234 void gf100_vmm_flush(struct nvkm_vmm *, int);
235 void gf100_vmm_invalidate(struct nvkm_vmm *, u32 type);
236 void gf100_vmm_invalidate_pdb(struct nvkm_vmm *, u64 addr);
237 
238 int gk20a_vmm_aper(enum nvkm_memory_target);
239 
240 int gm200_vmm_new_(const struct nvkm_vmm_func *, const struct nvkm_vmm_func *,
241 		   struct nvkm_mmu *, bool, u64, u64, void *, u32,
242 		   struct lock_class_key *, const char *, struct nvkm_vmm **);
243 int gm200_vmm_join_(struct nvkm_vmm *, struct nvkm_memory *, u64 base);
244 int gm200_vmm_join(struct nvkm_vmm *, struct nvkm_memory *);
245 
246 int gp100_vmm_new_(const struct nvkm_vmm_func *,
247 		   struct nvkm_mmu *, bool, u64, u64, void *, u32,
248 		   struct lock_class_key *, const char *, struct nvkm_vmm **);
249 int gp100_vmm_join(struct nvkm_vmm *, struct nvkm_memory *);
250 int gp100_vmm_valid(struct nvkm_vmm *, void *, u32, struct nvkm_vmm_map *);
251 void gp100_vmm_flush(struct nvkm_vmm *, int);
252 int gp100_vmm_mthd(struct nvkm_vmm *, struct nvkm_client *, u32, void *, u32);
253 void gp100_vmm_invalidate_pdb(struct nvkm_vmm *, u64 addr);
254 
255 int gv100_vmm_join(struct nvkm_vmm *, struct nvkm_memory *);
256 
257 int nv04_vmm_new(struct nvkm_mmu *, bool, u64, u64, void *, u32,
258 		 struct lock_class_key *, const char *, struct nvkm_vmm **);
259 int nv41_vmm_new(struct nvkm_mmu *, bool, u64, u64, void *, u32,
260 		 struct lock_class_key *, const char *, struct nvkm_vmm **);
261 int nv44_vmm_new(struct nvkm_mmu *, bool, u64, u64, void *, u32,
262 		 struct lock_class_key *, const char *, struct nvkm_vmm **);
263 int nv50_vmm_new(struct nvkm_mmu *, bool, u64, u64, void *, u32,
264 		 struct lock_class_key *, const char *, struct nvkm_vmm **);
265 int mcp77_vmm_new(struct nvkm_mmu *, bool, u64, u64, void *, u32,
266 		  struct lock_class_key *, const char *, struct nvkm_vmm **);
267 int g84_vmm_new(struct nvkm_mmu *, bool, u64, u64, void *, u32,
268 		struct lock_class_key *, const char *, struct nvkm_vmm **);
269 int gf100_vmm_new(struct nvkm_mmu *, bool, u64, u64, void *, u32,
270 		  struct lock_class_key *, const char *, struct nvkm_vmm **);
271 int gk104_vmm_new(struct nvkm_mmu *, bool, u64, u64, void *, u32,
272 		  struct lock_class_key *, const char *, struct nvkm_vmm **);
273 int gk20a_vmm_new(struct nvkm_mmu *, bool, u64, u64, void *, u32,
274 		  struct lock_class_key *, const char *, struct nvkm_vmm **);
275 int gm200_vmm_new_fixed(struct nvkm_mmu *, bool, u64, u64, void *, u32,
276 			struct lock_class_key *, const char *,
277 			struct nvkm_vmm **);
278 int gm200_vmm_new(struct nvkm_mmu *, bool, u64, u64, void *, u32,
279 		  struct lock_class_key *, const char *,
280 		  struct nvkm_vmm **);
281 int gm20b_vmm_new_fixed(struct nvkm_mmu *, bool, u64, u64, void *, u32,
282 			struct lock_class_key *, const char *,
283 			struct nvkm_vmm **);
284 int gm20b_vmm_new(struct nvkm_mmu *, bool, u64, u64, void *, u32,
285 		  struct lock_class_key *, const char *,
286 		  struct nvkm_vmm **);
287 int gp100_vmm_new(struct nvkm_mmu *, bool, u64, u64, void *, u32,
288 		  struct lock_class_key *, const char *,
289 		  struct nvkm_vmm **);
290 int gp10b_vmm_new(struct nvkm_mmu *, bool, u64, u64, void *, u32,
291 		  struct lock_class_key *, const char *,
292 		  struct nvkm_vmm **);
293 int gv100_vmm_new(struct nvkm_mmu *, bool, u64, u64, void *, u32,
294 		  struct lock_class_key *, const char *,
295 		  struct nvkm_vmm **);
296 int tu102_vmm_new(struct nvkm_mmu *, bool, u64, u64, void *, u32,
297 		  struct lock_class_key *, const char *,
298 		  struct nvkm_vmm **);
299 
300 #define VMM_PRINT(l,v,p,f,a...) do {                                           \
301 	struct nvkm_vmm *_vmm = (v);                                           \
302 	if (CONFIG_NOUVEAU_DEBUG >= (l) && _vmm->debug >= (l)) {               \
303 		nvkm_printk_(&_vmm->mmu->subdev, 0, p, "%s: "f"\n",            \
304 			     _vmm->name, ##a);                                 \
305 	}                                                                      \
306 } while(0)
307 #define VMM_DEBUG(v,f,a...) VMM_PRINT(NV_DBG_DEBUG, (v), info, f, ##a)
308 #define VMM_TRACE(v,f,a...) VMM_PRINT(NV_DBG_TRACE, (v), info, f, ##a)
309 #define VMM_SPAM(v,f,a...)  VMM_PRINT(NV_DBG_SPAM , (v),  dbg, f, ##a)
310 
311 #define VMM_MAP_ITER(VMM,PT,PTEI,PTEN,MAP,FILL,BASE,SIZE,NEXT) do {            \
312 	nvkm_kmap((PT)->memory);                                               \
313 	while (PTEN) {                                                         \
314 		u64 _ptes = ((SIZE) - MAP->off) >> MAP->page->shift;           \
315 		u64 _addr = ((BASE) + MAP->off);                               \
316                                                                                \
317 		if (_ptes > PTEN) {                                            \
318 			MAP->off += PTEN << MAP->page->shift;                  \
319 			_ptes = PTEN;                                          \
320 		} else {                                                       \
321 			MAP->off = 0;                                          \
322 			NEXT;                                                  \
323 		}                                                              \
324                                                                                \
325 		VMM_SPAM(VMM, "ITER %08x %08x PTE(s)", PTEI, (u32)_ptes);      \
326                                                                                \
327 		FILL(VMM, PT, PTEI, _ptes, MAP, _addr);                        \
328 		PTEI += _ptes;                                                 \
329 		PTEN -= _ptes;                                                 \
330 	}                                                                      \
331 	nvkm_done((PT)->memory);                                               \
332 } while(0)
333 
334 #define VMM_MAP_ITER_MEM(VMM,PT,PTEI,PTEN,MAP,FILL)                            \
335 	VMM_MAP_ITER(VMM,PT,PTEI,PTEN,MAP,FILL,                                \
336 		     ((u64)MAP->mem->offset << NVKM_RAM_MM_SHIFT),             \
337 		     ((u64)MAP->mem->length << NVKM_RAM_MM_SHIFT),             \
338 		     (MAP->mem = MAP->mem->next))
339 #define VMM_MAP_ITER_DMA(VMM,PT,PTEI,PTEN,MAP,FILL)                            \
340 	VMM_MAP_ITER(VMM,PT,PTEI,PTEN,MAP,FILL,                                \
341 		     *MAP->dma, PAGE_SIZE, MAP->dma++)
342 #define VMM_MAP_ITER_SGL(VMM,PT,PTEI,PTEN,MAP,FILL)                            \
343 	VMM_MAP_ITER(VMM,PT,PTEI,PTEN,MAP,FILL,                                \
344 		     sg_dma_address(MAP->sgl), sg_dma_len(MAP->sgl),           \
345 		     (MAP->sgl = sg_next(MAP->sgl)))
346 
347 #define VMM_FO(m,o,d,c,b) nvkm_fo##b((m)->memory, (o), (d), (c))
348 #define VMM_WO(m,o,d,c,b) nvkm_wo##b((m)->memory, (o), (d))
349 #define VMM_XO(m,v,o,d,c,b,fn,f,a...) do {                                     \
350 	const u32 _pteo = (o); u##b _data = (d);                               \
351 	VMM_SPAM((v), "   %010llx "f, (m)->addr + _pteo, _data, ##a);          \
352 	VMM_##fn((m), (m)->base + _pteo, _data, (c), b);                       \
353 } while(0)
354 
355 #define VMM_WO032(m,v,o,d) VMM_XO((m),(v),(o),(d),  1, 32, WO, "%08x")
356 #define VMM_FO032(m,v,o,d,c)                                                   \
357 	VMM_XO((m),(v),(o),(d),(c), 32, FO, "%08x %08x", (c))
358 
359 #define VMM_WO064(m,v,o,d) VMM_XO((m),(v),(o),(d),  1, 64, WO, "%016llx")
360 #define VMM_FO064(m,v,o,d,c)                                                   \
361 	VMM_XO((m),(v),(o),(d),(c), 64, FO, "%016llx %08x", (c))
362 
363 #define VMM_XO128(m,v,o,lo,hi,c,f,a...) do {                                   \
364 	u32 _pteo = (o), _ptes = (c);                                          \
365 	const u64 _addr = (m)->addr + _pteo;                                   \
366 	VMM_SPAM((v), "   %010llx %016llx%016llx"f, _addr, (hi), (lo), ##a);   \
367 	while (_ptes--) {                                                      \
368 		nvkm_wo64((m)->memory, (m)->base + _pteo + 0, (lo));           \
369 		nvkm_wo64((m)->memory, (m)->base + _pteo + 8, (hi));           \
370 		_pteo += 0x10;                                                 \
371 	}                                                                      \
372 } while(0)
373 
374 #define VMM_WO128(m,v,o,lo,hi) VMM_XO128((m),(v),(o),(lo),(hi), 1, "")
375 #define VMM_FO128(m,v,o,lo,hi,c) do {                                          \
376 	nvkm_kmap((m)->memory);                                                \
377 	VMM_XO128((m),(v),(o),(lo),(hi),(c), " %08x", (c));                    \
378 	nvkm_done((m)->memory);                                                \
379 } while(0)
380 #endif
381