xref: /linux/drivers/gpu/drm/msm/msm_iommu.c (revision fd7d598270724cc787982ea48bbe17ad383a8b7f)
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 #include <linux/adreno-smmu-priv.h>
8 #include <linux/io-pgtable.h>
9 #include "msm_drv.h"
10 #include "msm_mmu.h"
11 
12 struct msm_iommu {
13 	struct msm_mmu base;
14 	struct iommu_domain *domain;
15 	atomic_t pagetables;
16 };
17 
18 #define to_msm_iommu(x) container_of(x, struct msm_iommu, base)
19 
20 struct msm_iommu_pagetable {
21 	struct msm_mmu base;
22 	struct msm_mmu *parent;
23 	struct io_pgtable_ops *pgtbl_ops;
24 	unsigned long pgsize_bitmap;	/* Bitmap of page sizes in use */
25 	phys_addr_t ttbr;
26 	u32 asid;
27 };
28 static struct msm_iommu_pagetable *to_pagetable(struct msm_mmu *mmu)
29 {
30 	return container_of(mmu, struct msm_iommu_pagetable, base);
31 }
32 
33 /* based on iommu_pgsize() in iommu.c: */
34 static size_t calc_pgsize(struct msm_iommu_pagetable *pagetable,
35 			   unsigned long iova, phys_addr_t paddr,
36 			   size_t size, size_t *count)
37 {
38 	unsigned int pgsize_idx, pgsize_idx_next;
39 	unsigned long pgsizes;
40 	size_t offset, pgsize, pgsize_next;
41 	unsigned long addr_merge = paddr | iova;
42 
43 	/* Page sizes supported by the hardware and small enough for @size */
44 	pgsizes = pagetable->pgsize_bitmap & GENMASK(__fls(size), 0);
45 
46 	/* Constrain the page sizes further based on the maximum alignment */
47 	if (likely(addr_merge))
48 		pgsizes &= GENMASK(__ffs(addr_merge), 0);
49 
50 	/* Make sure we have at least one suitable page size */
51 	BUG_ON(!pgsizes);
52 
53 	/* Pick the biggest page size remaining */
54 	pgsize_idx = __fls(pgsizes);
55 	pgsize = BIT(pgsize_idx);
56 	if (!count)
57 		return pgsize;
58 
59 	/* Find the next biggest support page size, if it exists */
60 	pgsizes = pagetable->pgsize_bitmap & ~GENMASK(pgsize_idx, 0);
61 	if (!pgsizes)
62 		goto out_set_count;
63 
64 	pgsize_idx_next = __ffs(pgsizes);
65 	pgsize_next = BIT(pgsize_idx_next);
66 
67 	/*
68 	 * There's no point trying a bigger page size unless the virtual
69 	 * and physical addresses are similarly offset within the larger page.
70 	 */
71 	if ((iova ^ paddr) & (pgsize_next - 1))
72 		goto out_set_count;
73 
74 	/* Calculate the offset to the next page size alignment boundary */
75 	offset = pgsize_next - (addr_merge & (pgsize_next - 1));
76 
77 	/*
78 	 * If size is big enough to accommodate the larger page, reduce
79 	 * the number of smaller pages.
80 	 */
81 	if (offset + pgsize_next <= size)
82 		size = offset;
83 
84 out_set_count:
85 	*count = size >> pgsize_idx;
86 	return pgsize;
87 }
88 
89 static int msm_iommu_pagetable_unmap(struct msm_mmu *mmu, u64 iova,
90 		size_t size)
91 {
92 	struct msm_iommu_pagetable *pagetable = to_pagetable(mmu);
93 	struct io_pgtable_ops *ops = pagetable->pgtbl_ops;
94 
95 	while (size) {
96 		size_t unmapped, pgsize, count;
97 
98 		pgsize = calc_pgsize(pagetable, iova, iova, size, &count);
99 
100 		unmapped = ops->unmap_pages(ops, iova, pgsize, count, NULL);
101 		if (!unmapped)
102 			break;
103 
104 		iova += unmapped;
105 		size -= unmapped;
106 	}
107 
108 	iommu_flush_iotlb_all(to_msm_iommu(pagetable->parent)->domain);
109 
110 	return (size == 0) ? 0 : -EINVAL;
111 }
112 
113 static int msm_iommu_pagetable_map(struct msm_mmu *mmu, u64 iova,
114 		struct sg_table *sgt, size_t len, int prot)
115 {
116 	struct msm_iommu_pagetable *pagetable = to_pagetable(mmu);
117 	struct io_pgtable_ops *ops = pagetable->pgtbl_ops;
118 	struct scatterlist *sg;
119 	u64 addr = iova;
120 	unsigned int i;
121 
122 	for_each_sgtable_sg(sgt, sg, i) {
123 		size_t size = sg->length;
124 		phys_addr_t phys = sg_phys(sg);
125 
126 		while (size) {
127 			size_t pgsize, count, mapped = 0;
128 			int ret;
129 
130 			pgsize = calc_pgsize(pagetable, addr, phys, size, &count);
131 
132 			ret = ops->map_pages(ops, addr, phys, pgsize, count,
133 					     prot, GFP_KERNEL, &mapped);
134 
135 			/* map_pages could fail after mapping some of the pages,
136 			 * so update the counters before error handling.
137 			 */
138 			phys += mapped;
139 			addr += mapped;
140 			size -= mapped;
141 
142 			if (ret) {
143 				msm_iommu_pagetable_unmap(mmu, iova, addr - iova);
144 				return -EINVAL;
145 			}
146 		}
147 	}
148 
149 	return 0;
150 }
151 
152 static void msm_iommu_pagetable_destroy(struct msm_mmu *mmu)
153 {
154 	struct msm_iommu_pagetable *pagetable = to_pagetable(mmu);
155 	struct msm_iommu *iommu = to_msm_iommu(pagetable->parent);
156 	struct adreno_smmu_priv *adreno_smmu =
157 		dev_get_drvdata(pagetable->parent->dev);
158 
159 	/*
160 	 * If this is the last attached pagetable for the parent,
161 	 * disable TTBR0 in the arm-smmu driver
162 	 */
163 	if (atomic_dec_return(&iommu->pagetables) == 0)
164 		adreno_smmu->set_ttbr0_cfg(adreno_smmu->cookie, NULL);
165 
166 	free_io_pgtable_ops(pagetable->pgtbl_ops);
167 	kfree(pagetable);
168 }
169 
170 int msm_iommu_pagetable_params(struct msm_mmu *mmu,
171 		phys_addr_t *ttbr, int *asid)
172 {
173 	struct msm_iommu_pagetable *pagetable;
174 
175 	if (mmu->type != MSM_MMU_IOMMU_PAGETABLE)
176 		return -EINVAL;
177 
178 	pagetable = to_pagetable(mmu);
179 
180 	if (ttbr)
181 		*ttbr = pagetable->ttbr;
182 
183 	if (asid)
184 		*asid = pagetable->asid;
185 
186 	return 0;
187 }
188 
189 struct iommu_domain_geometry *msm_iommu_get_geometry(struct msm_mmu *mmu)
190 {
191 	struct msm_iommu *iommu = to_msm_iommu(mmu);
192 
193 	return &iommu->domain->geometry;
194 }
195 
196 static const struct msm_mmu_funcs pagetable_funcs = {
197 		.map = msm_iommu_pagetable_map,
198 		.unmap = msm_iommu_pagetable_unmap,
199 		.destroy = msm_iommu_pagetable_destroy,
200 };
201 
202 static void msm_iommu_tlb_flush_all(void *cookie)
203 {
204 }
205 
206 static void msm_iommu_tlb_flush_walk(unsigned long iova, size_t size,
207 		size_t granule, void *cookie)
208 {
209 }
210 
211 static void msm_iommu_tlb_add_page(struct iommu_iotlb_gather *gather,
212 		unsigned long iova, size_t granule, void *cookie)
213 {
214 }
215 
216 static const struct iommu_flush_ops null_tlb_ops = {
217 	.tlb_flush_all = msm_iommu_tlb_flush_all,
218 	.tlb_flush_walk = msm_iommu_tlb_flush_walk,
219 	.tlb_add_page = msm_iommu_tlb_add_page,
220 };
221 
222 static int msm_fault_handler(struct iommu_domain *domain, struct device *dev,
223 		unsigned long iova, int flags, void *arg);
224 
225 struct msm_mmu *msm_iommu_pagetable_create(struct msm_mmu *parent)
226 {
227 	struct adreno_smmu_priv *adreno_smmu = dev_get_drvdata(parent->dev);
228 	struct msm_iommu *iommu = to_msm_iommu(parent);
229 	struct msm_iommu_pagetable *pagetable;
230 	const struct io_pgtable_cfg *ttbr1_cfg = NULL;
231 	struct io_pgtable_cfg ttbr0_cfg;
232 	int ret;
233 
234 	/* Get the pagetable configuration from the domain */
235 	if (adreno_smmu->cookie)
236 		ttbr1_cfg = adreno_smmu->get_ttbr1_cfg(adreno_smmu->cookie);
237 
238 	/*
239 	 * If you hit this WARN_ONCE() you are probably missing an entry in
240 	 * qcom_smmu_impl_of_match[] in arm-smmu-qcom.c
241 	 */
242 	if (WARN_ONCE(!ttbr1_cfg, "No per-process page tables"))
243 		return ERR_PTR(-ENODEV);
244 
245 	pagetable = kzalloc(sizeof(*pagetable), GFP_KERNEL);
246 	if (!pagetable)
247 		return ERR_PTR(-ENOMEM);
248 
249 	msm_mmu_init(&pagetable->base, parent->dev, &pagetable_funcs,
250 		MSM_MMU_IOMMU_PAGETABLE);
251 
252 	/* Clone the TTBR1 cfg as starting point for TTBR0 cfg: */
253 	ttbr0_cfg = *ttbr1_cfg;
254 
255 	/* The incoming cfg will have the TTBR1 quirk enabled */
256 	ttbr0_cfg.quirks &= ~IO_PGTABLE_QUIRK_ARM_TTBR1;
257 	ttbr0_cfg.tlb = &null_tlb_ops;
258 
259 	pagetable->pgtbl_ops = alloc_io_pgtable_ops(ARM_64_LPAE_S1,
260 		&ttbr0_cfg, iommu->domain);
261 
262 	if (!pagetable->pgtbl_ops) {
263 		kfree(pagetable);
264 		return ERR_PTR(-ENOMEM);
265 	}
266 
267 	/*
268 	 * If this is the first pagetable that we've allocated, send it back to
269 	 * the arm-smmu driver as a trigger to set up TTBR0
270 	 */
271 	if (atomic_inc_return(&iommu->pagetables) == 1) {
272 		ret = adreno_smmu->set_ttbr0_cfg(adreno_smmu->cookie, &ttbr0_cfg);
273 		if (ret) {
274 			free_io_pgtable_ops(pagetable->pgtbl_ops);
275 			kfree(pagetable);
276 			return ERR_PTR(ret);
277 		}
278 	}
279 
280 	/* Needed later for TLB flush */
281 	pagetable->parent = parent;
282 	pagetable->pgsize_bitmap = ttbr0_cfg.pgsize_bitmap;
283 	pagetable->ttbr = ttbr0_cfg.arm_lpae_s1_cfg.ttbr;
284 
285 	/*
286 	 * TODO we would like each set of page tables to have a unique ASID
287 	 * to optimize TLB invalidation.  But iommu_flush_iotlb_all() will
288 	 * end up flushing the ASID used for TTBR1 pagetables, which is not
289 	 * what we want.  So for now just use the same ASID as TTBR1.
290 	 */
291 	pagetable->asid = 0;
292 
293 	return &pagetable->base;
294 }
295 
296 static int msm_fault_handler(struct iommu_domain *domain, struct device *dev,
297 		unsigned long iova, int flags, void *arg)
298 {
299 	struct msm_iommu *iommu = arg;
300 	struct msm_mmu *mmu = &iommu->base;
301 	struct adreno_smmu_priv *adreno_smmu = dev_get_drvdata(iommu->base.dev);
302 	struct adreno_smmu_fault_info info, *ptr = NULL;
303 
304 	if (adreno_smmu->get_fault_info) {
305 		adreno_smmu->get_fault_info(adreno_smmu->cookie, &info);
306 		ptr = &info;
307 	}
308 
309 	if (iommu->base.handler)
310 		return iommu->base.handler(iommu->base.arg, iova, flags, ptr);
311 
312 	pr_warn_ratelimited("*** fault: iova=%16lx, flags=%d\n", iova, flags);
313 
314 	if (mmu->funcs->resume_translation)
315 		mmu->funcs->resume_translation(mmu);
316 
317 	return 0;
318 }
319 
320 static void msm_iommu_resume_translation(struct msm_mmu *mmu)
321 {
322 	struct adreno_smmu_priv *adreno_smmu = dev_get_drvdata(mmu->dev);
323 
324 	if (adreno_smmu->resume_translation)
325 		adreno_smmu->resume_translation(adreno_smmu->cookie, true);
326 }
327 
328 static void msm_iommu_detach(struct msm_mmu *mmu)
329 {
330 	struct msm_iommu *iommu = to_msm_iommu(mmu);
331 
332 	iommu_detach_device(iommu->domain, mmu->dev);
333 }
334 
335 static int msm_iommu_map(struct msm_mmu *mmu, uint64_t iova,
336 		struct sg_table *sgt, size_t len, int prot)
337 {
338 	struct msm_iommu *iommu = to_msm_iommu(mmu);
339 	size_t ret;
340 
341 	/* The arm-smmu driver expects the addresses to be sign extended */
342 	if (iova & BIT_ULL(48))
343 		iova |= GENMASK_ULL(63, 49);
344 
345 	ret = iommu_map_sgtable(iommu->domain, iova, sgt, prot);
346 	WARN_ON(!ret);
347 
348 	return (ret == len) ? 0 : -EINVAL;
349 }
350 
351 static int msm_iommu_unmap(struct msm_mmu *mmu, uint64_t iova, size_t len)
352 {
353 	struct msm_iommu *iommu = to_msm_iommu(mmu);
354 
355 	if (iova & BIT_ULL(48))
356 		iova |= GENMASK_ULL(63, 49);
357 
358 	iommu_unmap(iommu->domain, iova, len);
359 
360 	return 0;
361 }
362 
363 static void msm_iommu_destroy(struct msm_mmu *mmu)
364 {
365 	struct msm_iommu *iommu = to_msm_iommu(mmu);
366 	iommu_domain_free(iommu->domain);
367 	kfree(iommu);
368 }
369 
370 static const struct msm_mmu_funcs funcs = {
371 		.detach = msm_iommu_detach,
372 		.map = msm_iommu_map,
373 		.unmap = msm_iommu_unmap,
374 		.destroy = msm_iommu_destroy,
375 		.resume_translation = msm_iommu_resume_translation,
376 };
377 
378 struct msm_mmu *msm_iommu_new(struct device *dev, unsigned long quirks)
379 {
380 	struct iommu_domain *domain;
381 	struct msm_iommu *iommu;
382 	int ret;
383 
384 	domain = iommu_domain_alloc(dev->bus);
385 	if (!domain)
386 		return NULL;
387 
388 	iommu_set_pgtable_quirks(domain, quirks);
389 
390 	iommu = kzalloc(sizeof(*iommu), GFP_KERNEL);
391 	if (!iommu) {
392 		iommu_domain_free(domain);
393 		return ERR_PTR(-ENOMEM);
394 	}
395 
396 	iommu->domain = domain;
397 	msm_mmu_init(&iommu->base, dev, &funcs, MSM_MMU_IOMMU);
398 
399 	atomic_set(&iommu->pagetables, 0);
400 
401 	ret = iommu_attach_device(iommu->domain, dev);
402 	if (ret) {
403 		iommu_domain_free(domain);
404 		kfree(iommu);
405 		return ERR_PTR(ret);
406 	}
407 
408 	return &iommu->base;
409 }
410 
411 struct msm_mmu *msm_iommu_gpu_new(struct device *dev, struct msm_gpu *gpu, unsigned long quirks)
412 {
413 	struct adreno_smmu_priv *adreno_smmu = dev_get_drvdata(dev);
414 	struct msm_iommu *iommu;
415 	struct msm_mmu *mmu;
416 
417 	mmu = msm_iommu_new(dev, quirks);
418 	if (IS_ERR_OR_NULL(mmu))
419 		return mmu;
420 
421 	iommu = to_msm_iommu(mmu);
422 	iommu_set_fault_handler(iommu->domain, msm_fault_handler, iommu);
423 
424 	/* Enable stall on iommu fault: */
425 	if (adreno_smmu->set_stall)
426 		adreno_smmu->set_stall(adreno_smmu->cookie, true);
427 
428 	return mmu;
429 }
430