io-pgtable-dart.c

  1// SPDX-License-Identifier: GPL-2.0-only
  2/*
  3 * Apple DART page table allocator.
  4 *
  5 * Copyright (C) 2022 The Asahi Linux Contributors
  6 *
  7 * Based on io-pgtable-arm.
  8 *
  9 * Copyright (C) 2014 ARM Limited
 10 *
 11 * Author: Will Deacon <will.deacon@arm.com>
 12 */
 13
 14#define pr_fmt(fmt)	"dart io-pgtable: " fmt
 15
 16#include <linux/atomic.h>
 17#include <linux/bitfield.h>
 18#include <linux/bitops.h>
 19#include <linux/io-pgtable.h>
 20#include <linux/kernel.h>
 21#include <linux/sizes.h>
 22#include <linux/slab.h>
 23#include <linux/types.h>
 24
 25#include <asm/barrier.h>
 26#include "iommu-pages.h"
 27
 28#define DART1_MAX_ADDR_BITS	36
 29
 30#define DART_MAX_TABLE_BITS	2
 31#define DART_MAX_TABLES		BIT(DART_MAX_TABLE_BITS)
 32#define DART_MAX_LEVELS		4 /* Includes TTBR level */
 33
 34/* Struct accessors */
 35#define io_pgtable_to_data(x)						\
 36	container_of((x), struct dart_io_pgtable, iop)
 37
 38#define io_pgtable_ops_to_data(x)					\
 39	io_pgtable_to_data(io_pgtable_ops_to_pgtable(x))
 40
 41#define DART_GRANULE(d)						\
 42	(sizeof(dart_iopte) << (d)->bits_per_level)
 43#define DART_PTES_PER_TABLE(d)					\
 44	(DART_GRANULE(d) >> ilog2(sizeof(dart_iopte)))
 45
 46#define APPLE_DART_PTE_SUBPAGE_START   GENMASK_ULL(63, 52)
 47#define APPLE_DART_PTE_SUBPAGE_END     GENMASK_ULL(51, 40)
 48
 49#define APPLE_DART1_PADDR_MASK	GENMASK_ULL(35, 12)
 50#define APPLE_DART2_PADDR_MASK	GENMASK_ULL(37, 10)
 51#define APPLE_DART2_PADDR_SHIFT	(4)
 52
 53/* Apple DART1 protection bits */
 54#define APPLE_DART1_PTE_PROT_NO_READ	BIT(8)
 55#define APPLE_DART1_PTE_PROT_NO_WRITE	BIT(7)
 56#define APPLE_DART1_PTE_PROT_SP_DIS	BIT(1)
 57
 58/* Apple DART2 protection bits */
 59#define APPLE_DART2_PTE_PROT_NO_READ	BIT(3)
 60#define APPLE_DART2_PTE_PROT_NO_WRITE	BIT(2)
 61#define APPLE_DART2_PTE_PROT_NO_CACHE	BIT(1)
 62
 63/* marks PTE as valid */
 64#define APPLE_DART_PTE_VALID		BIT(0)
 65
 66/* IOPTE accessors */
 67#define iopte_deref(pte, d) __va(iopte_to_paddr(pte, d))
 68
 69struct dart_io_pgtable {
 70	struct io_pgtable	iop;
 71
 72	int			levels;
 73	int			tbl_bits;
 74	int			bits_per_level;
 75
 76	void			*pgd[DART_MAX_TABLES];
 77};
 78
 79typedef u64 dart_iopte;
 80
 81
 82static dart_iopte paddr_to_iopte(phys_addr_t paddr,
 83				     struct dart_io_pgtable *data)
 84{
 85	dart_iopte pte;
 86
 87	if (data->iop.fmt == APPLE_DART)
 88		return paddr & APPLE_DART1_PADDR_MASK;
 89
 90	/* format is APPLE_DART2 */
 91	pte = paddr >> APPLE_DART2_PADDR_SHIFT;
 92	pte &= APPLE_DART2_PADDR_MASK;
 93
 94	return pte;
 95}
 96
 97static phys_addr_t iopte_to_paddr(dart_iopte pte,
 98				  struct dart_io_pgtable *data)
 99{
100	u64 paddr;
101
102	if (data->iop.fmt == APPLE_DART)
103		return pte & APPLE_DART1_PADDR_MASK;
104
105	/* format is APPLE_DART2 */
106	paddr = pte & APPLE_DART2_PADDR_MASK;
107	paddr <<= APPLE_DART2_PADDR_SHIFT;
108
109	return paddr;
110}
111
112static int dart_init_pte(struct dart_io_pgtable *data,
113			     unsigned long iova, phys_addr_t paddr,
114			     dart_iopte prot, int num_entries,
115			     dart_iopte *ptep)
116{
117	int i;
118	dart_iopte pte = prot;
119	size_t sz = data->iop.cfg.pgsize_bitmap;
120
121	for (i = 0; i < num_entries; i++)
122		if (ptep[i] & APPLE_DART_PTE_VALID) {
123			/* We require an unmap first */
124			WARN_ON(ptep[i] & APPLE_DART_PTE_VALID);
125			return -EEXIST;
126		}
127
128	/* subpage protection: always allow access to the entire page */
129	pte |= FIELD_PREP(APPLE_DART_PTE_SUBPAGE_START, 0);
130	pte |= FIELD_PREP(APPLE_DART_PTE_SUBPAGE_END, 0xfff);
131
132	pte |= APPLE_DART_PTE_VALID;
133
134	for (i = 0; i < num_entries; i++)
135		ptep[i] = pte | paddr_to_iopte(paddr + i * sz, data);
136
137	return 0;
138}
139
140static dart_iopte dart_install_table(dart_iopte *table,
141					     dart_iopte *ptep,
142					     dart_iopte curr,
143					     struct dart_io_pgtable *data)
144{
145	dart_iopte old, new;
146
147	new = paddr_to_iopte(__pa(table), data) | APPLE_DART_PTE_VALID;
148
149	/*
150	 * Ensure the table itself is visible before its PTE can be.
151	 * Whilst we could get away with cmpxchg64_release below, this
152	 * doesn't have any ordering semantics when !CONFIG_SMP.
153	 */
154	dma_wmb();
155
156	old = cmpxchg64_relaxed(ptep, curr, new);
157
158	return old;
159}
160
161static int dart_get_index(struct dart_io_pgtable *data, unsigned long iova, int level)
162{
163	return (iova >> (level * data->bits_per_level + ilog2(sizeof(dart_iopte)))) &
164		((1 << data->bits_per_level) - 1);
165}
166
167static int dart_get_last_index(struct dart_io_pgtable *data, unsigned long iova)
168{
169
170	return (iova >> (data->bits_per_level + ilog2(sizeof(dart_iopte)))) &
171		 ((1 << data->bits_per_level) - 1);
172}
173
174static dart_iopte *dart_get_last(struct dart_io_pgtable *data, unsigned long iova)
175{
176	dart_iopte pte, *ptep;
177	int level = data->levels;
178	int tbl = dart_get_index(data, iova, level);
179
180	if (tbl >= (1 << data->tbl_bits))
181		return NULL;
182
183	ptep = data->pgd[tbl];
184	if (!ptep)
185		return NULL;
186
187	while (--level > 1) {
188		ptep += dart_get_index(data, iova, level);
189		pte = READ_ONCE(*ptep);
190
191		/* Valid entry? */
192		if (!pte)
193			return NULL;
194
195		/* Deref to get next level table */
196		ptep = iopte_deref(pte, data);
197	}
198
199	return ptep;
200}
201
202static dart_iopte dart_prot_to_pte(struct dart_io_pgtable *data,
203					   int prot)
204{
205	dart_iopte pte = 0;
206
207	if (data->iop.fmt == APPLE_DART) {
208		pte |= APPLE_DART1_PTE_PROT_SP_DIS;
209		if (!(prot & IOMMU_WRITE))
210			pte |= APPLE_DART1_PTE_PROT_NO_WRITE;
211		if (!(prot & IOMMU_READ))
212			pte |= APPLE_DART1_PTE_PROT_NO_READ;
213	}
214	if (data->iop.fmt == APPLE_DART2) {
215		if (!(prot & IOMMU_WRITE))
216			pte |= APPLE_DART2_PTE_PROT_NO_WRITE;
217		if (!(prot & IOMMU_READ))
218			pte |= APPLE_DART2_PTE_PROT_NO_READ;
219		if (!(prot & IOMMU_CACHE))
220			pte |= APPLE_DART2_PTE_PROT_NO_CACHE;
221	}
222
223	return pte;
224}
225
226static int dart_map_pages(struct io_pgtable_ops *ops, unsigned long iova,
227			      phys_addr_t paddr, size_t pgsize, size_t pgcount,
228			      int iommu_prot, gfp_t gfp, size_t *mapped)
229{
230	struct dart_io_pgtable *data = io_pgtable_ops_to_data(ops);
231	struct io_pgtable_cfg *cfg = &data->iop.cfg;
232	size_t tblsz = DART_GRANULE(data);
233	int ret = 0, tbl, num_entries, max_entries, map_idx_start;
234	dart_iopte pte, *cptep, *ptep;
235	dart_iopte prot;
236	int level = data->levels;
237
238	if (WARN_ON(pgsize != cfg->pgsize_bitmap))
239		return -EINVAL;
240
241	if (WARN_ON(paddr >> cfg->oas))
242		return -ERANGE;
243
244	if (!(iommu_prot & (IOMMU_READ | IOMMU_WRITE)))
245		return -EINVAL;
246
247	tbl = dart_get_index(data, iova, level);
248
249	if (tbl >= (1 << data->tbl_bits))
250		return -ENOMEM;
251
252	ptep = data->pgd[tbl];
253	while (--level > 1) {
254		ptep += dart_get_index(data, iova, level);
255		pte = READ_ONCE(*ptep);
256
257		/* no table present */
258		if (!pte) {
259			cptep = iommu_alloc_pages_sz(gfp, tblsz);
260			if (!cptep)
261				return -ENOMEM;
262
263			pte = dart_install_table(cptep, ptep, 0, data);
264			if (pte)
265				iommu_free_pages(cptep);
266
267			/* L2 table is present (now) */
268			pte = READ_ONCE(*ptep);
269		}
270
271		ptep = iopte_deref(pte, data);
272	}
273
274	/* install a leaf entries into L2 table */
275	prot = dart_prot_to_pte(data, iommu_prot);
276	map_idx_start = dart_get_last_index(data, iova);
277	max_entries = DART_PTES_PER_TABLE(data) - map_idx_start;
278	num_entries = min_t(int, pgcount, max_entries);
279	ptep += map_idx_start;
280	ret = dart_init_pte(data, iova, paddr, prot, num_entries, ptep);
281	if (!ret && mapped)
282		*mapped += num_entries * pgsize;
283
284	/*
285	 * Synchronise all PTE updates for the new mapping before there's
286	 * a chance for anything to kick off a table walk for the new iova.
287	 */
288	wmb();
289
290	return ret;
291}
292
293static size_t dart_unmap_pages(struct io_pgtable_ops *ops, unsigned long iova,
294				   size_t pgsize, size_t pgcount,
295				   struct iommu_iotlb_gather *gather)
296{
297	struct dart_io_pgtable *data = io_pgtable_ops_to_data(ops);
298	struct io_pgtable_cfg *cfg = &data->iop.cfg;
299	int i = 0, num_entries, max_entries, unmap_idx_start;
300	dart_iopte pte, *ptep;
301
302	if (WARN_ON(pgsize != cfg->pgsize_bitmap || !pgcount))
303		return 0;
304
305	ptep = dart_get_last(data, iova);
306
307	/* Valid L2 IOPTE pointer? */
308	if (WARN_ON(!ptep))
309		return 0;
310
311	unmap_idx_start = dart_get_last_index(data, iova);
312	ptep += unmap_idx_start;
313
314	max_entries = DART_PTES_PER_TABLE(data) - unmap_idx_start;
315	num_entries = min_t(int, pgcount, max_entries);
316
317	while (i < num_entries) {
318		pte = READ_ONCE(*ptep);
319		if (WARN_ON(!pte))
320			break;
321
322		/* clear pte */
323		*ptep = 0;
324
325		if (!iommu_iotlb_gather_queued(gather))
326			io_pgtable_tlb_add_page(&data->iop, gather,
327						iova + i * pgsize, pgsize);
328
329		ptep++;
330		i++;
331	}
332
333	return i * pgsize;
334}
335
336static phys_addr_t dart_iova_to_phys(struct io_pgtable_ops *ops,
337					 unsigned long iova)
338{
339	struct dart_io_pgtable *data = io_pgtable_ops_to_data(ops);
340	dart_iopte pte, *ptep;
341
342	ptep = dart_get_last(data, iova);
343
344	/* Valid L2 IOPTE pointer? */
345	if (!ptep)
346		return 0;
347
348	ptep += dart_get_last_index(data, iova);
349
350	pte = READ_ONCE(*ptep);
351	/* Found translation */
352	if (pte) {
353		iova &= (data->iop.cfg.pgsize_bitmap - 1);
354		return iopte_to_paddr(pte, data) | iova;
355	}
356
357	/* Ran out of page tables to walk */
358	return 0;
359}
360
361static struct dart_io_pgtable *
362dart_alloc_pgtable(struct io_pgtable_cfg *cfg)
363{
364	struct dart_io_pgtable *data;
365	int levels, max_tbl_bits, tbl_bits, bits_per_level, va_bits, pg_shift;
366
367	/*
368	 * Old 4K page DARTs can use up to 4 top-level tables.
369	 * Newer ones only ever use a maximum of 1.
370	 */
371	if (cfg->pgsize_bitmap == SZ_4K)
372		max_tbl_bits = DART_MAX_TABLE_BITS;
373	else
374		max_tbl_bits = 0;
375
376	pg_shift = __ffs(cfg->pgsize_bitmap);
377	bits_per_level = pg_shift - ilog2(sizeof(dart_iopte));
378
379	va_bits = cfg->ias - pg_shift;
380
381	levels = max_t(int, 2, (va_bits - max_tbl_bits + bits_per_level - 1) / bits_per_level);
382
383	if (levels > (DART_MAX_LEVELS - 1))
384		return NULL;
385
386	tbl_bits = max_t(int, 0, va_bits - (bits_per_level * levels));
387
388	if (tbl_bits > max_tbl_bits)
389		return NULL;
390
391	data = kzalloc(sizeof(*data), GFP_KERNEL);
392	if (!data)
393		return NULL;
394
395	data->levels = levels + 1; /* Table level counts as one level */
396	data->tbl_bits = tbl_bits;
397	data->bits_per_level = bits_per_level;
398
399	data->iop.ops = (struct io_pgtable_ops) {
400		.map_pages	= dart_map_pages,
401		.unmap_pages	= dart_unmap_pages,
402		.iova_to_phys	= dart_iova_to_phys,
403	};
404
405	return data;
406}
407
408static struct io_pgtable *
409apple_dart_alloc_pgtable(struct io_pgtable_cfg *cfg, void *cookie)
410{
411	struct dart_io_pgtable *data;
412	int i;
413
414	if (!cfg->coherent_walk)
415		return NULL;
416
417	if (cfg->oas != 36 && cfg->oas != 42)
418		return NULL;
419
420	if (cfg->ias > cfg->oas)
421		return NULL;
422
423	if (!(cfg->pgsize_bitmap == SZ_4K || cfg->pgsize_bitmap == SZ_16K))
424		return NULL;
425
426	data = dart_alloc_pgtable(cfg);
427	if (!data)
428		return NULL;
429
430	cfg->apple_dart_cfg.n_ttbrs = 1 << data->tbl_bits;
431	cfg->apple_dart_cfg.n_levels = data->levels;
432
433	for (i = 0; i < cfg->apple_dart_cfg.n_ttbrs; ++i) {
434		data->pgd[i] =
435			iommu_alloc_pages_sz(GFP_KERNEL, DART_GRANULE(data));
436		if (!data->pgd[i])
437			goto out_free_data;
438		cfg->apple_dart_cfg.ttbr[i] = virt_to_phys(data->pgd[i]);
439	}
440
441	return &data->iop;
442
443out_free_data:
444	while (--i >= 0) {
445		iommu_free_pages(data->pgd[i]);
446	}
447	kfree(data);
448	return NULL;
449}
450
451static void apple_dart_free_pgtables(struct dart_io_pgtable *data, dart_iopte *ptep, int level)
452{
453	dart_iopte *end;
454	dart_iopte *start = ptep;
455
456	if (level > 1) {
457		end = (void *)ptep + DART_GRANULE(data);
458
459		while (ptep != end) {
460			dart_iopte pte = *ptep++;
461
462			if (pte)
463				apple_dart_free_pgtables(data, iopte_deref(pte, data), level - 1);
464		}
465	}
466	iommu_free_pages(start);
467}
468
469static void apple_dart_free_pgtable(struct io_pgtable *iop)
470{
471	struct dart_io_pgtable *data = io_pgtable_to_data(iop);
472	int i;
473
474	for (i = 0; i < (1 << data->tbl_bits) && data->pgd[i]; ++i)
475		apple_dart_free_pgtables(data, data->pgd[i], data->levels - 1);
476
477	kfree(data);
478}
479
480struct io_pgtable_init_fns io_pgtable_apple_dart_init_fns = {
481	.alloc	= apple_dart_alloc_pgtable,
482	.free	= apple_dart_free_pgtable,
483};