xref: /rk3399_ARM-atf/lib/gpt_rme/gpt_rme.c (revision 88655be93c148f9849d8fd02d7edaf0cb7d04058)
1 /*
2  * Copyright (c) 2022-2025, Arm Limited. All rights reserved.
3  *
4  * SPDX-License-Identifier: BSD-3-Clause
5  */
6 
7 #include <assert.h>
8 #include <errno.h>
9 #include <inttypes.h>
10 #include <limits.h>
11 #include <stdint.h>
12 
13 #include <arch.h>
14 #include <arch_features.h>
15 #include <common/debug.h>
16 #include <lib/gpt_rme/gpt_rme.h>
17 #include <lib/smccc.h>
18 #include <lib/xlat_tables/xlat_tables_v2.h>
19 
20 #include "gpt_rme_private.h"
21 
22 #if !ENABLE_RME
23 #error "ENABLE_RME must be enabled to use the GPT library"
24 #endif
25 
26 /*
27  * Lookup T from PPS
28  *
29  *   PPS    Size    T
30  *   0b000  4GB     32
31  *   0b001  64GB    36
32  *   0b010  1TB     40
33  *   0b011  4TB     42
34  *   0b100  16TB    44
35  *   0b101  256TB   48
36  *   0b110  4PB     52
37  *
38  * See section 15.1.27 of the RME specification.
39  */
40 static const gpt_t_val_e gpt_t_lookup[] = {PPS_4GB_T, PPS_64GB_T,
41 					   PPS_1TB_T, PPS_4TB_T,
42 					   PPS_16TB_T, PPS_256TB_T,
43 					   PPS_4PB_T};
44 
45 /*
46  * Lookup P from PGS
47  *
48  *   PGS    Size    P
49  *   0b00   4KB     12
50  *   0b10   16KB    14
51  *   0b01   64KB    16
52  *
53  * Note that pgs=0b10 is 16KB and pgs=0b01 is 64KB, this is not a typo.
54  *
55  * See section 15.1.27 of the RME specification.
56  */
57 static const gpt_p_val_e gpt_p_lookup[] = {PGS_4KB_P, PGS_64KB_P, PGS_16KB_P};
58 
59 static void shatter_2mb(uintptr_t base, const gpi_info_t *gpi_info,
60 				uint64_t l1_desc);
61 static void shatter_32mb(uintptr_t base, const gpi_info_t *gpi_info,
62 				uint64_t l1_desc);
63 static void shatter_512mb(uintptr_t base, const gpi_info_t *gpi_info,
64 				uint64_t l1_desc);
65 
66 /*
67  * This structure contains GPT configuration data
68  */
69 typedef struct {
70 	uintptr_t plat_gpt_l0_base;
71 	gpccr_pps_e pps;
72 	gpt_t_val_e t;
73 	gpccr_pgs_e pgs;
74 	gpt_p_val_e p;
75 } gpt_config_t;
76 
77 static gpt_config_t gpt_config;
78 
79 /*
80  * Number of L1 entries in 2MB, depending on GPCCR_EL3.PGS:
81  * +-------+------------+
82  * |  PGS  | L1 entries |
83  * +-------+------------+
84  * |  4KB  |     32     |
85  * +-------+------------+
86  * |  16KB |     8      |
87  * +-------+------------+
88  * |  64KB |     2      |
89  * +-------+------------+
90  */
91 static unsigned int gpt_l1_cnt_2mb;
92 
93 /*
94  * Mask for the L1 index field, depending on
95  * GPCCR_EL3.L0GPTSZ and GPCCR_EL3.PGS:
96  * +---------+-------------------------------+
97  * |         |             PGS               |
98  * +---------+----------+----------+---------+
99  * | L0GPTSZ |   4KB    |   16KB   |   64KB  |
100  * +---------+----------+----------+---------+
101  * |  1GB    |  0x3FFF  |  0xFFF   |  0x3FF  |
102  * +---------+----------+----------+---------+
103  * |  16GB   | 0x3FFFF  |  0xFFFF  | 0x3FFF  |
104  * +---------+----------+----------+---------+
105  * |  64GB   | 0xFFFFF  | 0x3FFFF  | 0xFFFF  |
106  * +---------+----------+----------+---------+
107  * |  512GB  | 0x7FFFFF | 0x1FFFFF | 0x7FFFF |
108  * +---------+----------+----------+---------+
109  */
110 static uint64_t gpt_l1_index_mask;
111 
112 /* Number of 128-bit L1 entries in 2MB, 32MB and 512MB */
113 #define L1_QWORDS_2MB	(gpt_l1_cnt_2mb / 2U)
114 #define L1_QWORDS_32MB	(L1_QWORDS_2MB * 16U)
115 #define L1_QWORDS_512MB	(L1_QWORDS_32MB * 16U)
116 
117 /* Size in bytes of L1 entries in 2MB, 32MB */
118 #define L1_BYTES_2MB	(gpt_l1_cnt_2mb * sizeof(uint64_t))
119 #define L1_BYTES_32MB	(L1_BYTES_2MB * 16U)
120 
121 /* Get the index into the L1 table from a physical address */
122 #define GPT_L1_INDEX(_pa)	\
123 	(((_pa) >> (unsigned int)GPT_L1_IDX_SHIFT(gpt_config.p)) & gpt_l1_index_mask)
124 
125 /* This variable is used during initialization of the L1 tables */
126 static uintptr_t gpt_l1_tbl;
127 
128 /* These variables are used during runtime */
129 #if (RME_GPT_BITLOCK_BLOCK == 0)
130 /*
131  * The GPTs are protected by a global spinlock to ensure
132  * that multiple CPUs do not attempt to change the descriptors at once.
133  */
134 static spinlock_t gpt_lock;
135 
136 /* Lock/unlock macros for GPT entries
137  *
138  * Access to GPT is controlled by a global lock to ensure
139  * that no more than one CPU is allowed to make changes at any
140  * given time.
141  */
142 #define GPT_LOCK	spin_lock(&gpt_lock)
143 #define GPT_UNLOCK	spin_unlock(&gpt_lock)
144 #else
145 
146 /* Base address of bitlocks array */
147 static bitlock_t *gpt_bitlock;
148 
149 /*
150  * Access to a block of memory is controlled by a bitlock.
151  * Size of block = RME_GPT_BITLOCK_BLOCK * 512MB.
152  */
153 #define GPT_LOCK	bit_lock(gpi_info.lock, gpi_info.mask)
154 #define GPT_UNLOCK	bit_unlock(gpi_info.lock, gpi_info.mask)
155 #endif /* RME_GPT_BITLOCK_BLOCK */
156 
157 static void tlbi_page_dsbosh(uintptr_t base)
158 {
159 	/* Look-up table for invalidation TLBs for 4KB, 16KB and 64KB pages */
160 	static const gpt_tlbi_lookup_t tlbi_page_lookup[] = {
161 		{ tlbirpalos_4k, ~(SZ_4K - 1UL) },
162 		{ tlbirpalos_64k, ~(SZ_64K - 1UL) },
163 		{ tlbirpalos_16k, ~(SZ_16K - 1UL) }
164 	};
165 
166 	tlbi_page_lookup[gpt_config.pgs].function(
167 			base & tlbi_page_lookup[gpt_config.pgs].mask);
168 	dsbosh();
169 }
170 
171 /*
172  * Helper function to fill out GPI entries in a single L1 table
173  * with Granules or Contiguous descriptor.
174  *
175  * Parameters
176  *   l1			Pointer to 2MB, 32MB or 512MB aligned L1 table entry to fill out
177  *   l1_desc		GPT Granules or Contiguous descriptor set this range to
178  *   cnt		Number of double 128-bit L1 entries to fill
179  *
180  */
181 static void fill_desc(uint64_t *l1, uint64_t l1_desc, unsigned int cnt)
182 {
183 	uint128_t *l1_quad = (uint128_t *)l1;
184 	uint128_t l1_quad_desc = (uint128_t)l1_desc | ((uint128_t)l1_desc << 64);
185 
186 	VERBOSE("GPT: %s(%p 0x%"PRIx64" %u)\n", __func__, l1, l1_desc, cnt);
187 
188 	for (unsigned int i = 0U; i < cnt; i++) {
189 		*l1_quad++ = l1_quad_desc;
190 	}
191 }
192 
193 static void shatter_2mb(uintptr_t base, const gpi_info_t *gpi_info,
194 				uint64_t l1_desc)
195 {
196 	unsigned long idx = GPT_L1_INDEX(ALIGN_2MB(base));
197 
198 	VERBOSE("GPT: %s(0x%"PRIxPTR" 0x%"PRIx64")\n",
199 				__func__, base, l1_desc);
200 
201 	/* Convert 2MB Contiguous block to Granules */
202 	fill_desc(&gpi_info->gpt_l1_addr[idx], l1_desc, L1_QWORDS_2MB);
203 }
204 
205 static void shatter_32mb(uintptr_t base, const gpi_info_t *gpi_info,
206 				uint64_t l1_desc)
207 {
208 	unsigned long idx = GPT_L1_INDEX(ALIGN_2MB(base));
209 	const uint64_t *l1_gran = &gpi_info->gpt_l1_addr[idx];
210 	uint64_t l1_cont_desc = GPT_L1_CONT_DESC(l1_desc, 2MB);
211 	uint64_t *l1;
212 
213 	VERBOSE("GPT: %s(0x%"PRIxPTR" 0x%"PRIx64")\n",
214 				__func__, base, l1_desc);
215 
216 	/* Get index corresponding to 32MB aligned address */
217 	idx = GPT_L1_INDEX(ALIGN_32MB(base));
218 	l1 = &gpi_info->gpt_l1_addr[idx];
219 
220 	/* 16 x 2MB blocks in 32MB */
221 	for (unsigned int i = 0U; i < 16U; i++) {
222 		/* Fill with Granules or Contiguous descriptors */
223 		fill_desc(l1, (l1 == l1_gran) ? l1_desc : l1_cont_desc,
224 							L1_QWORDS_2MB);
225 		l1 = (uint64_t *)((uintptr_t)l1 + L1_BYTES_2MB);
226 	}
227 }
228 
229 static void shatter_512mb(uintptr_t base, const gpi_info_t *gpi_info,
230 				uint64_t l1_desc)
231 {
232 	unsigned long idx = GPT_L1_INDEX(ALIGN_32MB(base));
233 	const uint64_t *l1_32mb = &gpi_info->gpt_l1_addr[idx];
234 	uint64_t l1_cont_desc = GPT_L1_CONT_DESC(l1_desc, 32MB);
235 	uint64_t *l1;
236 
237 	VERBOSE("GPT: %s(0x%"PRIxPTR" 0x%"PRIx64")\n",
238 				__func__, base, l1_desc);
239 
240 	/* Get index corresponding to 512MB aligned address */
241 	idx = GPT_L1_INDEX(ALIGN_512MB(base));
242 	l1 = &gpi_info->gpt_l1_addr[idx];
243 
244 	/* 16 x 32MB blocks in 512MB */
245 	for (unsigned int i = 0U; i < 16U; i++) {
246 		if (l1 == l1_32mb) {
247 			/* Shatter this 32MB block */
248 			shatter_32mb(base, gpi_info, l1_desc);
249 		} else {
250 			/* Fill 32MB with Contiguous descriptors */
251 			fill_desc(l1, l1_cont_desc, L1_QWORDS_32MB);
252 		}
253 
254 		l1 = (uint64_t *)((uintptr_t)l1 + L1_BYTES_32MB);
255 	}
256 }
257 
258 /*
259  * This function checks to see if a GPI value is valid.
260  *
261  * These are valid GPI values.
262  *   GPT_GPI_NO_ACCESS   U(0x0)
263  *   GPT_GPI_SA          U(0x4)
264  *   GPT_GPI_NSP	 U(0x5)
265  *   GPT_GPI_SECURE      U(0x8)
266  *   GPT_GPI_NS          U(0x9)
267  *   GPT_GPI_ROOT        U(0xA)
268  *   GPT_GPI_REALM       U(0xB)
269  *   GPT_GPI_NSO         U(0xD)
270  *   GPT_GPI_ANY         U(0xF)
271  *
272  * Parameters
273  *   gpi		GPI to check for validity.
274  *
275  * Return
276  *   true for a valid GPI, false for an invalid one.
277  */
278 static bool is_gpi_valid(unsigned int gpi)
279 {
280 	switch (gpi) {
281 	case GPT_GPI_NO_ACCESS:
282 	case GPT_GPI_SECURE:
283 	case GPT_GPI_NS:
284 	case GPT_GPI_ROOT:
285 	case GPT_GPI_REALM:
286 	case GPT_GPI_ANY:
287 		return true;
288 	case GPT_GPI_NSO:
289 		return is_feat_rme_gpc2_present();
290 	case GPT_GPI_SA:
291 	case GPT_GPI_NSP:
292 		return is_feat_rme_gdi_supported();
293 	default:
294 		return false;
295 	}
296 }
297 
298 /*
299  * This function checks to see if two PAS regions overlap.
300  *
301  * Parameters
302  *   base_1: base address of first PAS
303  *   size_1: size of first PAS
304  *   base_2: base address of second PAS
305  *   size_2: size of second PAS
306  *
307  * Return
308  *   True if PAS regions overlap, false if they do not.
309  */
310 static bool check_pas_overlap(uintptr_t base_1, size_t size_1,
311 			      uintptr_t base_2, size_t size_2)
312 {
313 	if (((base_1 + size_1) > base_2) && ((base_2 + size_2) > base_1)) {
314 		return true;
315 	}
316 	return false;
317 }
318 
319 /*
320  * This helper function checks to see if a PAS region from index 0 to
321  * (pas_idx - 1) occupies the L0 region at index l0_idx in the L0 table.
322  *
323  * Parameters
324  *   l0_idx:      Index of the L0 entry to check
325  *   pas_regions: PAS region array
326  *   pas_idx:     Upper bound of the PAS array index.
327  *
328  * Return
329  *   True if a PAS region occupies the L0 region in question, false if not.
330  */
331 static bool does_previous_pas_exist_here(unsigned int l0_idx,
332 					 pas_region_t *pas_regions,
333 					 unsigned int pas_idx)
334 {
335 	/* Iterate over PAS regions up to pas_idx */
336 	for (unsigned int i = 0U; i < pas_idx; i++) {
337 		if (check_pas_overlap((GPT_L0GPTSZ_ACTUAL_SIZE * l0_idx),
338 		    GPT_L0GPTSZ_ACTUAL_SIZE,
339 		    pas_regions[i].base_pa, pas_regions[i].size)) {
340 			return true;
341 		}
342 	}
343 	return false;
344 }
345 
346 /*
347  * This function iterates over all of the PAS regions and checks them to ensure
348  * proper alignment of base and size, that the GPI is valid, and that no regions
349  * overlap. As a part of the overlap checks, this function checks existing L0
350  * mappings against the new PAS regions in the event that gpt_init_pas_l1_tables
351  * is called multiple times to place L1 tables in different areas of memory. It
352  * also counts the number of L1 tables needed and returns it on success.
353  *
354  * Parameters
355  *   *pas_regions	Pointer to array of PAS region structures.
356  *   pas_region_cnt	Total number of PAS regions in the array.
357  *
358  * Return
359  *   Negative Linux error code in the event of a failure, number of L1 regions
360  *   required when successful.
361  */
362 static int validate_pas_mappings(pas_region_t *pas_regions,
363 				 unsigned int pas_region_cnt)
364 {
365 	unsigned int idx;
366 	unsigned int l1_cnt = 0U;
367 	unsigned int pas_l1_cnt;
368 	uint64_t *l0_desc = (uint64_t *)gpt_config.plat_gpt_l0_base;
369 
370 	assert(pas_regions != NULL);
371 	assert(pas_region_cnt != 0U);
372 
373 	for (idx = 0U; idx < pas_region_cnt; idx++) {
374 		/* Check for arithmetic overflow in region */
375 		if ((ULONG_MAX - pas_regions[idx].base_pa) <
376 		    pas_regions[idx].size) {
377 			ERROR("GPT: Address overflow in PAS[%u]!\n", idx);
378 			return -EOVERFLOW;
379 		}
380 
381 		/* Initial checks for PAS validity */
382 		if (((pas_regions[idx].base_pa + pas_regions[idx].size) >
383 		    GPT_PPS_ACTUAL_SIZE(gpt_config.t)) ||
384 		    !is_gpi_valid(GPT_PAS_ATTR_GPI(pas_regions[idx].attrs))) {
385 			ERROR("GPT: PAS[%u] is invalid!\n", idx);
386 			return -EFAULT;
387 		}
388 
389 		/*
390 		 * Make sure this PAS does not overlap with another one. We
391 		 * start from idx + 1 instead of 0 since prior PAS mappings will
392 		 * have already checked themselves against this one.
393 		 */
394 		for (unsigned int i = idx + 1U; i < pas_region_cnt; i++) {
395 			if (check_pas_overlap(pas_regions[idx].base_pa,
396 			    pas_regions[idx].size,
397 			    pas_regions[i].base_pa,
398 			    pas_regions[i].size)) {
399 				ERROR("GPT: PAS[%u] overlaps with PAS[%u]\n",
400 					i, idx);
401 				return -EFAULT;
402 			}
403 		}
404 
405 		/*
406 		 * Since this function can be called multiple times with
407 		 * separate L1 tables we need to check the existing L0 mapping
408 		 * to see if this PAS would fall into one that has already been
409 		 * initialized.
410 		 */
411 		for (unsigned int i =
412 			(unsigned int)GPT_L0_IDX(pas_regions[idx].base_pa);
413 			i <= GPT_L0_IDX(pas_regions[idx].base_pa +
414 					pas_regions[idx].size - 1UL);
415 			i++) {
416 			if ((GPT_L0_TYPE(l0_desc[i]) == GPT_L0_TYPE_BLK_DESC) &&
417 			    (GPT_L0_BLKD_GPI(l0_desc[i]) == GPT_GPI_ANY)) {
418 				/* This descriptor is unused so continue */
419 				continue;
420 			}
421 
422 			/*
423 			 * This descriptor has been initialized in a previous
424 			 * call to this function so cannot be initialized again.
425 			 */
426 			ERROR("GPT: PAS[%u] overlaps with previous L0[%u]!\n",
427 			      idx, i);
428 			return -EFAULT;
429 		}
430 
431 		/* Check for block mapping (L0) type */
432 		if (GPT_PAS_ATTR_MAP_TYPE(pas_regions[idx].attrs) ==
433 		    GPT_PAS_ATTR_MAP_TYPE_BLOCK) {
434 			/* Make sure base and size are block-aligned */
435 			if (!GPT_IS_L0_ALIGNED(pas_regions[idx].base_pa) ||
436 			    !GPT_IS_L0_ALIGNED(pas_regions[idx].size)) {
437 				ERROR("GPT: PAS[%u] is not block-aligned!\n",
438 				      idx);
439 				return -EFAULT;
440 			}
441 
442 			continue;
443 		}
444 
445 		/* Check for granule mapping (L1) type */
446 		if (GPT_PAS_ATTR_MAP_TYPE(pas_regions[idx].attrs) ==
447 		    GPT_PAS_ATTR_MAP_TYPE_GRANULE) {
448 			/* Make sure base and size are granule-aligned */
449 			if (!GPT_IS_L1_ALIGNED(gpt_config.p, pas_regions[idx].base_pa) ||
450 			    !GPT_IS_L1_ALIGNED(gpt_config.p, pas_regions[idx].size)) {
451 				ERROR("GPT: PAS[%u] is not granule-aligned!\n",
452 				      idx);
453 				return -EFAULT;
454 			}
455 
456 			/* Find how many L1 tables this PAS occupies */
457 			pas_l1_cnt = (GPT_L0_IDX(pas_regions[idx].base_pa +
458 				     pas_regions[idx].size - 1UL) -
459 				     GPT_L0_IDX(pas_regions[idx].base_pa) + 1U);
460 
461 			/*
462 			 * This creates a situation where, if multiple PAS
463 			 * regions occupy the same table descriptor, we can get
464 			 * an artificially high total L1 table count. The way we
465 			 * handle this is by checking each PAS against those
466 			 * before it in the array, and if they both occupy the
467 			 * same PAS we subtract from pas_l1_cnt and only the
468 			 * first PAS in the array gets to count it.
469 			 */
470 
471 			/*
472 			 * If L1 count is greater than 1 we know the start and
473 			 * end PAs are in different L0 regions so we must check
474 			 * both for overlap against other PAS.
475 			 */
476 			if (pas_l1_cnt > 1) {
477 				if (does_previous_pas_exist_here(
478 				    GPT_L0_IDX(pas_regions[idx].base_pa +
479 				    pas_regions[idx].size - 1UL),
480 				    pas_regions, idx)) {
481 					pas_l1_cnt--;
482 				}
483 			}
484 
485 			if (does_previous_pas_exist_here(
486 			    GPT_L0_IDX(pas_regions[idx].base_pa),
487 			    pas_regions, idx)) {
488 				pas_l1_cnt--;
489 			}
490 
491 			l1_cnt += pas_l1_cnt;
492 			continue;
493 		}
494 
495 		/* If execution reaches this point, mapping type is invalid */
496 		ERROR("GPT: PAS[%u] has invalid mapping type 0x%x.\n", idx,
497 		      GPT_PAS_ATTR_MAP_TYPE(pas_regions[idx].attrs));
498 		return -EINVAL;
499 	}
500 
501 	return l1_cnt;
502 }
503 
504 /*
505  * This function validates L0 initialization parameters.
506  *
507  * Parameters
508  *   l0_mem_base	Base address of memory used for L0 table.
509  *   l0_mem_size	Size of memory available for L0 table.
510  *
511  * Return
512  *   Negative Linux error code in the event of a failure, 0 for success.
513  */
514 static int validate_l0_params(gpccr_pps_e pps, uintptr_t l0_mem_base,
515 				size_t l0_mem_size)
516 {
517 	size_t l0_alignment;
518 
519 	/*
520 	 * Make sure PPS is valid and then store it since macros need this value
521 	 * to work.
522 	 */
523 	if (pps > GPT_PPS_MAX) {
524 		ERROR("GPT: Invalid PPS: 0x%x\n", pps);
525 		return -EINVAL;
526 	}
527 	gpt_config.pps = pps;
528 	gpt_config.t = gpt_t_lookup[pps];
529 
530 	/* Alignment must be the greater of 4KB or L0 table size */
531 	l0_alignment = SZ_4K;
532 	if (l0_alignment < GPT_L0_TABLE_SIZE(gpt_config.t)) {
533 		l0_alignment = GPT_L0_TABLE_SIZE(gpt_config.t);
534 	}
535 
536 	/* Check base address */
537 	if ((l0_mem_base == 0UL) ||
538 	   ((l0_mem_base & (l0_alignment - 1UL)) != 0UL)) {
539 		ERROR("GPT: Invalid L0 base address: 0x%lx\n", l0_mem_base);
540 		return -EFAULT;
541 	}
542 
543 	/* Check memory size for L0 table */
544 	if (l0_mem_size < GPT_L0_TABLE_SIZE(gpt_config.t)) {
545 		ERROR("GPT: Inadequate L0 memory\n");
546 		ERROR("      Expected 0x%lx bytes, got 0x%lx\n",
547 				GPT_L0_TABLE_SIZE(gpt_config.t), l0_mem_size);
548 		return -ENOMEM;
549 	}
550 
551 	return 0;
552 }
553 
554 /*
555  * In the event that L1 tables are needed, this function validates
556  * the L1 table generation parameters.
557  *
558  * Parameters
559  *   l1_mem_base	Base address of memory used for L1 table allocation.
560  *   l1_mem_size	Total size of memory available for L1 tables.
561  *   l1_gpt_cnt		Number of L1 tables needed.
562  *
563  * Return
564  *   Negative Linux error code in the event of a failure, 0 for success.
565  */
566 static int validate_l1_params(uintptr_t l1_mem_base, size_t l1_mem_size,
567 				unsigned int l1_gpt_cnt)
568 {
569 	size_t l1_gpt_mem_sz;
570 
571 	/* Check if the granularity is supported */
572 	if (!xlat_arch_is_granule_size_supported(
573 	    GPT_PGS_ACTUAL_SIZE(gpt_config.p))) {
574 		return -EPERM;
575 	}
576 
577 	/* Make sure L1 tables are aligned to their size */
578 	if ((l1_mem_base & (GPT_L1_TABLE_SIZE(gpt_config.p) - 1UL)) != 0UL) {
579 		ERROR("GPT: Unaligned L1 GPT base address: 0x%"PRIxPTR"\n",
580 		      l1_mem_base);
581 		return -EFAULT;
582 	}
583 
584 	/* Get total memory needed for L1 tables */
585 	l1_gpt_mem_sz = l1_gpt_cnt * GPT_L1_TABLE_SIZE(gpt_config.p);
586 
587 	/* Check for overflow */
588 	if ((l1_gpt_mem_sz / GPT_L1_TABLE_SIZE(gpt_config.p)) != l1_gpt_cnt) {
589 		ERROR("GPT: Overflow calculating L1 memory size\n");
590 		return -ENOMEM;
591 	}
592 
593 	/* Make sure enough space was supplied */
594 	if (l1_mem_size < l1_gpt_mem_sz) {
595 		ERROR("%sL1 GPTs%s", (const char *)"GPT: Inadequate ",
596 			(const char *)" memory\n");
597 		ERROR("      Expected 0x%lx bytes, got 0x%lx\n",
598 			l1_gpt_mem_sz, l1_mem_size);
599 		return -ENOMEM;
600 	}
601 
602 	VERBOSE("GPT: Requested 0x%lx bytes for L1 GPTs\n", l1_gpt_mem_sz);
603 	return 0;
604 }
605 
606 /*
607  * This function initializes L0 block descriptors (regions that cannot be
608  * transitioned at the granule level) according to the provided PAS.
609  *
610  * Parameters
611  *   *pas		Pointer to the structure defining the PAS region to
612  *			initialize.
613  */
614 static void generate_l0_blk_desc(pas_region_t *pas)
615 {
616 	uint64_t gpt_desc;
617 	unsigned long idx, end_idx;
618 	uint64_t *l0_gpt_arr;
619 
620 	assert(gpt_config.plat_gpt_l0_base != 0UL);
621 	assert(pas != NULL);
622 
623 	/*
624 	 * Checking of PAS parameters has already been done in
625 	 * validate_pas_mappings so no need to check the same things again.
626 	 */
627 
628 	l0_gpt_arr = (uint64_t *)gpt_config.plat_gpt_l0_base;
629 
630 	/* Create the GPT Block descriptor for this PAS region */
631 	gpt_desc = GPT_L0_BLK_DESC(GPT_PAS_ATTR_GPI(pas->attrs));
632 
633 	/* Start index of this region in L0 GPTs */
634 	idx = GPT_L0_IDX(pas->base_pa);
635 
636 	/*
637 	 * Determine number of L0 GPT descriptors covered by
638 	 * this PAS region and use the count to populate these
639 	 * descriptors.
640 	 */
641 	end_idx = GPT_L0_IDX(pas->base_pa + pas->size);
642 
643 	/* Generate the needed block descriptors */
644 	for (; idx < end_idx; idx++) {
645 		l0_gpt_arr[idx] = gpt_desc;
646 		VERBOSE("GPT: L0 entry (BLOCK) index %lu [%p]: GPI = 0x%"PRIx64" (0x%"PRIx64")\n",
647 			idx, &l0_gpt_arr[idx],
648 			(gpt_desc >> GPT_L0_BLK_DESC_GPI_SHIFT) &
649 			GPT_L0_BLK_DESC_GPI_MASK, l0_gpt_arr[idx]);
650 	}
651 }
652 
653 /*
654  * Helper function to determine if the end physical address lies in the same L0
655  * region as the current physical address. If true, the end physical address is
656  * returned else, the start address of the next region is returned.
657  *
658  * Parameters
659  *   cur_pa		Physical address of the current PA in the loop through
660  *			the range.
661  *   end_pa		Physical address of the end PA in a PAS range.
662  *
663  * Return
664  *   The PA of the end of the current range.
665  */
666 static uintptr_t get_l1_end_pa(uintptr_t cur_pa, uintptr_t end_pa)
667 {
668 	uintptr_t cur_idx;
669 	uintptr_t end_idx;
670 
671 	cur_idx = GPT_L0_IDX(cur_pa);
672 	end_idx = GPT_L0_IDX(end_pa);
673 
674 	assert(cur_idx <= end_idx);
675 
676 	if (cur_idx == end_idx) {
677 		return end_pa;
678 	}
679 
680 	return (cur_idx + 1UL) << GPT_L0_IDX_SHIFT;
681 }
682 
683 /*
684  * Helper function to fill out GPI entries from 'first' granule address of
685  * the specified 'length' in a single L1 table with 'l1_desc' Contiguous
686  * descriptor.
687  *
688  * Parameters
689  *   l1			Pointer to L1 table to fill out
690  *   first		Address of first granule in range
691  *   length		Length of the range in bytes
692  *   gpi		GPI set this range to
693  *
694  * Return
695  *   Address of next granule in range.
696  */
697 __unused static uintptr_t fill_l1_cont_desc(uint64_t *l1, uintptr_t first,
698 					    size_t length, unsigned int gpi)
699 {
700 	/*
701 	 * Look up table for contiguous blocks and descriptors.
702 	 * Entries should be defined in descending block sizes:
703 	 * 512MB, 32MB and 2MB.
704 	 */
705 	static const gpt_fill_lookup_t gpt_fill_lookup[] = {
706 #if (RME_GPT_MAX_BLOCK == 512)
707 		{ SZ_512M, GPT_L1_CONT_DESC_512MB },
708 #endif
709 #if (RME_GPT_MAX_BLOCK >= 32)
710 		{ SZ_32M, GPT_L1_CONT_DESC_32MB },
711 #endif
712 #if (RME_GPT_MAX_BLOCK != 0)
713 		{ SZ_2M, GPT_L1_CONT_DESC_2MB }
714 #endif
715 	};
716 
717 	/*
718 	 * Iterate through all block sizes (512MB, 32MB and 2MB)
719 	 * starting with maximum supported.
720 	 */
721 	for (unsigned long i = 0UL; i < ARRAY_SIZE(gpt_fill_lookup); i++) {
722 		/* Calculate index */
723 		unsigned long idx = GPT_L1_INDEX(first);
724 
725 		/* Contiguous block size */
726 		size_t cont_size = gpt_fill_lookup[i].size;
727 
728 		if (GPT_REGION_IS_CONT(length, first, cont_size)) {
729 
730 			/* Generate Contiguous descriptor */
731 			uint64_t l1_desc = GPT_L1_GPI_CONT_DESC(gpi,
732 						gpt_fill_lookup[i].desc);
733 
734 			/* Number of 128-bit L1 entries in block */
735 			unsigned int cnt;
736 
737 			switch (cont_size) {
738 			case SZ_512M:
739 				cnt = L1_QWORDS_512MB;
740 				break;
741 			case SZ_32M:
742 				cnt = L1_QWORDS_32MB;
743 				break;
744 			default:			/* SZ_2MB */
745 				cnt = L1_QWORDS_2MB;
746 			}
747 
748 			VERBOSE("GPT: Contiguous descriptor 0x%"PRIxPTR" %luMB\n",
749 				first, cont_size / SZ_1M);
750 
751 			/* Fill Contiguous descriptors */
752 			fill_desc(&l1[idx], l1_desc, cnt);
753 			return (first + cont_size);
754 		}
755 	}
756 
757 	return first;
758 }
759 
760 /* Build Granules descriptor with the same 'gpi' for every GPI entry */
761 static uint64_t build_l1_desc(unsigned int gpi)
762 {
763 	uint64_t l1_desc = (uint64_t)gpi | ((uint64_t)gpi << 4);
764 
765 	l1_desc |= (l1_desc << 8);
766 	l1_desc |= (l1_desc << 16);
767 	return (l1_desc | (l1_desc << 32));
768 }
769 
770 /*
771  * Helper function to fill out GPI entries from 'first' to 'last' granule
772  * address in a single L1 table with 'l1_desc' Granules descriptor.
773  *
774  * Parameters
775  *   l1			Pointer to L1 table to fill out
776  *   first		Address of first granule in range
777  *   last		Address of last granule in range (inclusive)
778  *   gpi		GPI set this range to
779  *
780  * Return
781  *   Address of next granule in range.
782  */
783 static uintptr_t fill_l1_gran_desc(uint64_t *l1, uintptr_t first,
784 				   uintptr_t last, unsigned int gpi)
785 {
786 	uint64_t gpi_mask;
787 	unsigned long i;
788 
789 	/* Generate Granules descriptor */
790 	uint64_t l1_desc = build_l1_desc(gpi);
791 
792 	/* Shift the mask if we're starting in the middle of an L1 entry */
793 	gpi_mask = ULONG_MAX << (GPT_L1_GPI_IDX(gpt_config.p, first) << 2);
794 
795 	/* Fill out each L1 entry for this region */
796 	for (i = GPT_L1_INDEX(first); i <= GPT_L1_INDEX(last); i++) {
797 
798 		/* Account for stopping in the middle of an L1 entry */
799 		if (i == GPT_L1_INDEX(last)) {
800 			gpi_mask &= (gpi_mask >> ((15U -
801 				    GPT_L1_GPI_IDX(gpt_config.p, last)) << 2));
802 		}
803 
804 		assert((l1[i] & gpi_mask) == (GPT_L1_ANY_DESC & gpi_mask));
805 
806 		/* Write GPI values */
807 		l1[i] = (l1[i] & ~gpi_mask) | (l1_desc & gpi_mask);
808 
809 		/* Reset mask */
810 		gpi_mask = ULONG_MAX;
811 	}
812 
813 	return last + GPT_PGS_ACTUAL_SIZE(gpt_config.p);
814 }
815 
816 /*
817  * Helper function to fill out GPI entries in a single L1 table.
818  * This function fills out an entire L1 table with either Granules or Contiguous
819  * (RME_GPT_MAX_BLOCK != 0) descriptors depending on region length and alignment.
820  * Note. If RME_GPT_MAX_BLOCK == 0, then the L1 tables are filled with regular
821  * Granules descriptors.
822  *
823  * Parameters
824  *   l1			Pointer to L1 table to fill out
825  *   first		Address of first granule in range
826  *   last		Address of last granule in range (inclusive)
827  *   gpi		GPI set this range to
828  */
829 static void fill_l1_tbl(uint64_t *l1, uintptr_t first, uintptr_t last,
830 			unsigned int gpi)
831 {
832 	assert(l1 != NULL);
833 	assert(first <= last);
834 	assert((first & (GPT_PGS_ACTUAL_SIZE(gpt_config.p) - 1UL)) == 0UL);
835 	assert((last & (GPT_PGS_ACTUAL_SIZE(gpt_config.p) - 1UL)) == 0UL);
836 	assert(GPT_L0_IDX(first) == GPT_L0_IDX(last));
837 
838 #if (RME_GPT_MAX_BLOCK != 0)
839 	while (first <= last) {
840 		/* Region length */
841 		size_t length = last - first + GPT_PGS_ACTUAL_SIZE(gpt_config.p);
842 
843 		if (length < SZ_2M) {
844 			/*
845 			 * Fill with Granule descriptors in case of
846 			 * region length < 2MB.
847 			 */
848 			first = fill_l1_gran_desc(l1, first, last, gpi);
849 
850 		} else if ((first & (SZ_2M - UL(1))) == UL(0)) {
851 			/*
852 			 * For region length >= 2MB and at least 2MB aligned
853 			 * call to fill_l1_cont_desc will iterate through
854 			 * all block sizes (512MB, 32MB and 2MB) supported and
855 			 * fill corresponding Contiguous descriptors.
856 			 */
857 			first = fill_l1_cont_desc(l1, first, length, gpi);
858 		} else {
859 			/*
860 			 * For not aligned region >= 2MB fill with Granules
861 			 * descriptors up to the next 2MB aligned address.
862 			 */
863 			uintptr_t new_last = ALIGN_2MB(first + SZ_2M) -
864 					GPT_PGS_ACTUAL_SIZE(gpt_config.p);
865 
866 			first = fill_l1_gran_desc(l1, first, new_last, gpi);
867 		}
868 	}
869 #else
870 	/* Fill with Granule descriptors */
871 	first = fill_l1_gran_desc(l1, first, last, gpi);
872 #endif
873 	assert(first == (last + GPT_PGS_ACTUAL_SIZE(gpt_config.p)));
874 }
875 
876 /*
877  * This function finds the next available unused L1 table and initializes all
878  * granules descriptor entries to GPI_ANY. This ensures that there are no chunks
879  * of GPI_NO_ACCESS (0b0000) memory floating around in the system in the
880  * event that a PAS region stops midway through an L1 table, thus guaranteeing
881  * that all memory not explicitly assigned is GPI_ANY. This function does not
882  * check for overflow conditions, that should be done by the caller.
883  *
884  * Return
885  *   Pointer to the next available L1 table.
886  */
887 static uint64_t *get_new_l1_tbl(void)
888 {
889 	/* Retrieve the next L1 table */
890 	uint64_t *l1 = (uint64_t *)gpt_l1_tbl;
891 
892 	/* Increment L1 GPT address */
893 	gpt_l1_tbl += GPT_L1_TABLE_SIZE(gpt_config.p);
894 
895 	/* Initialize all GPIs to GPT_GPI_ANY */
896 	for (unsigned int i = 0U; i < GPT_L1_ENTRY_COUNT(gpt_config.p); i++) {
897 		l1[i] = GPT_L1_ANY_DESC;
898 	}
899 
900 	return l1;
901 }
902 
903 /*
904  * When L1 tables are needed, this function creates the necessary L0 table
905  * descriptors and fills out the L1 table entries according to the supplied
906  * PAS range.
907  *
908  * Parameters
909  *   *pas		Pointer to the structure defining the PAS region.
910  */
911 static void generate_l0_tbl_desc(pas_region_t *pas)
912 {
913 	uintptr_t end_pa;
914 	uintptr_t cur_pa;
915 	uintptr_t last_gran_pa;
916 	uint64_t *l0_gpt_base;
917 	uint64_t *l1_gpt_arr;
918 	unsigned int l0_idx, gpi;
919 
920 	assert(gpt_config.plat_gpt_l0_base != 0UL);
921 	assert(pas != NULL);
922 
923 	/*
924 	 * Checking of PAS parameters has already been done in
925 	 * validate_pas_mappings so no need to check the same things again.
926 	 */
927 	end_pa = pas->base_pa + pas->size;
928 	l0_gpt_base = (uint64_t *)gpt_config.plat_gpt_l0_base;
929 
930 	/* We start working from the granule at base PA */
931 	cur_pa = pas->base_pa;
932 
933 	/* Get GPI */
934 	gpi = GPT_PAS_ATTR_GPI(pas->attrs);
935 
936 	/* Iterate over each L0 region in this memory range */
937 	for (l0_idx = (unsigned int)GPT_L0_IDX(pas->base_pa);
938 	     l0_idx <= (unsigned int)GPT_L0_IDX(end_pa - 1UL);
939 	     l0_idx++) {
940 		/*
941 		 * See if the L0 entry is already a table descriptor or if we
942 		 * need to create one.
943 		 */
944 		if (GPT_L0_TYPE(l0_gpt_base[l0_idx]) == GPT_L0_TYPE_TBL_DESC) {
945 			/* Get the L1 array from the L0 entry */
946 			l1_gpt_arr = GPT_L0_TBLD_ADDR(l0_gpt_base[l0_idx]);
947 		} else {
948 			/* Get a new L1 table from the L1 memory space */
949 			l1_gpt_arr = get_new_l1_tbl();
950 
951 			/* Fill out the L0 descriptor and flush it */
952 			l0_gpt_base[l0_idx] = GPT_L0_TBL_DESC(l1_gpt_arr);
953 		}
954 
955 		VERBOSE("GPT: L0 entry (TABLE) index %u [%p] ==> L1 Addr %p (0x%"PRIx64")\n",
956 			l0_idx, &l0_gpt_base[l0_idx], l1_gpt_arr, l0_gpt_base[l0_idx]);
957 
958 		/*
959 		 * Determine the PA of the last granule in this L0 descriptor.
960 		 */
961 		last_gran_pa = get_l1_end_pa(cur_pa, end_pa) -
962 			       GPT_PGS_ACTUAL_SIZE(gpt_config.p);
963 
964 		/*
965 		 * Fill up L1 GPT entries between these two addresses. This
966 		 * function needs the addresses of the first granule and last
967 		 * granule in the range.
968 		 */
969 		fill_l1_tbl(l1_gpt_arr, cur_pa, last_gran_pa, gpi);
970 
971 		/* Advance cur_pa to first granule in next L0 region */
972 		cur_pa = get_l1_end_pa(cur_pa, end_pa);
973 	}
974 }
975 
976 /*
977  * This function flushes a range of L0 descriptors used by a given PAS region
978  * array. There is a chance that some unmodified L0 descriptors would be flushed
979  * in the case that there are "holes" in an array of PAS regions but overall
980  * this should be faster than individually flushing each modified L0 descriptor
981  * as they are created.
982  *
983  * Parameters
984  *   *pas		Pointer to an array of PAS regions.
985  *   pas_count		Number of entries in the PAS array.
986  */
987 static void flush_l0_for_pas_array(pas_region_t *pas, unsigned int pas_count)
988 {
989 	unsigned long idx;
990 	unsigned long start_idx;
991 	unsigned long end_idx;
992 	uint64_t *l0 = (uint64_t *)gpt_config.plat_gpt_l0_base;
993 
994 	assert(pas != NULL);
995 	assert(pas_count != 0U);
996 
997 	/* Initial start and end values */
998 	start_idx = GPT_L0_IDX(pas[0].base_pa);
999 	end_idx = GPT_L0_IDX(pas[0].base_pa + pas[0].size - 1UL);
1000 
1001 	/* Find lowest and highest L0 indices used in this PAS array */
1002 	for (idx = 1UL; idx < pas_count; idx++) {
1003 		if (GPT_L0_IDX(pas[idx].base_pa) < start_idx) {
1004 			start_idx = GPT_L0_IDX(pas[idx].base_pa);
1005 		}
1006 		if (GPT_L0_IDX(pas[idx].base_pa + pas[idx].size - 1UL) > end_idx) {
1007 			end_idx = GPT_L0_IDX(pas[idx].base_pa + pas[idx].size - 1UL);
1008 		}
1009 	}
1010 
1011 	/*
1012 	 * Flush all covered L0 descriptors, add 1 because we need to include
1013 	 * the end index value.
1014 	 */
1015 	flush_dcache_range((uintptr_t)&l0[start_idx],
1016 			   ((end_idx + 1UL) - start_idx) * sizeof(uint64_t));
1017 }
1018 
1019 /*
1020  * Public API to enable granule protection checks once the tables have all been
1021  * initialized. This function is called at first initialization and then again
1022  * later during warm boots of CPU cores.
1023  *
1024  * Return
1025  *   Negative Linux error code in the event of a failure, 0 for success.
1026  */
1027 int gpt_enable(void)
1028 {
1029 	u_register_t gpccr_el3;
1030 
1031 	/*
1032 	 * Granule tables must be initialised before enabling
1033 	 * granule protection.
1034 	 */
1035 	if (gpt_config.plat_gpt_l0_base == 0UL) {
1036 		ERROR("GPT: Tables have not been initialized!\n");
1037 		return -EPERM;
1038 	}
1039 
1040 	/* Write the base address of the L0 tables into GPTBR */
1041 	write_gptbr_el3(((gpt_config.plat_gpt_l0_base >> GPTBR_BADDR_VAL_SHIFT)
1042 			>> GPTBR_BADDR_SHIFT) & GPTBR_BADDR_MASK);
1043 
1044 	/* GPCCR_EL3.PPS */
1045 	gpccr_el3 = SET_GPCCR_PPS(gpt_config.pps);
1046 
1047 	/* GPCCR_EL3.PGS */
1048 	gpccr_el3 |= SET_GPCCR_PGS(gpt_config.pgs);
1049 
1050 	/*
1051 	 * Since EL3 maps the L1 region as Inner shareable, use the same
1052 	 * shareability attribute for GPC as well so that
1053 	 * GPC fetches are visible to PEs
1054 	 */
1055 	gpccr_el3 |= SET_GPCCR_SH(GPCCR_SH_IS);
1056 
1057 	/* Outer and Inner cacheability set to Normal memory, WB, RA, WA */
1058 	gpccr_el3 |= SET_GPCCR_ORGN(GPCCR_ORGN_WB_RA_WA);
1059 	gpccr_el3 |= SET_GPCCR_IRGN(GPCCR_IRGN_WB_RA_WA);
1060 
1061 	/* Enable NSP and SA if FEAT_RME_GDI is implemented */
1062 	if (is_feat_rme_gdi_supported()) {
1063 		gpccr_el3 |= GPCCR_NSP_BIT;
1064 		gpccr_el3 |= GPCCR_SA_BIT;
1065 	}
1066 
1067 	/* Prepopulate GPCCR_EL3 but don't enable GPC yet */
1068 	write_gpccr_el3(gpccr_el3);
1069 	isb();
1070 
1071 	/* Invalidate any stale TLB entries and any cached register fields */
1072 	tlbipaallos();
1073 	dsb();
1074 	isb();
1075 
1076 	/* Enable GPT */
1077 	gpccr_el3 |= GPCCR_GPC_BIT;
1078 
1079 	/* Enable NSO encoding if FEAT_RME_GPC2 is supported. */
1080 	if (is_feat_rme_gpc2_present()) {
1081 		gpccr_el3 |= GPCCR_NSO_BIT;
1082 	}
1083 
1084 	/* TODO: Configure GPCCR_EL3_GPCP for Fault control */
1085 	write_gpccr_el3(gpccr_el3);
1086 	isb();
1087 	tlbipaallos();
1088 	dsb();
1089 	isb();
1090 
1091 	return 0;
1092 }
1093 
1094 /*
1095  * Public API to disable granule protection checks.
1096  */
1097 void gpt_disable(void)
1098 {
1099 	u_register_t gpccr_el3 = read_gpccr_el3();
1100 
1101 	write_gpccr_el3(gpccr_el3 & ~GPCCR_GPC_BIT);
1102 	dsbsy();
1103 	isb();
1104 }
1105 
1106 /*
1107  * Public API that initializes the entire protected space to GPT_GPI_ANY using
1108  * the L0 tables (block descriptors). Ideally, this function is invoked prior
1109  * to DDR discovery and initialization. The MMU must be initialized before
1110  * calling this function.
1111  *
1112  * Parameters
1113  *   pps		PPS value to use for table generation
1114  *   l0_mem_base	Base address of L0 tables in memory.
1115  *   l0_mem_size	Total size of memory available for L0 tables.
1116  *
1117  * Return
1118  *   Negative Linux error code in the event of a failure, 0 for success.
1119  */
1120 int gpt_init_l0_tables(gpccr_pps_e pps, uintptr_t l0_mem_base,
1121 		       size_t l0_mem_size)
1122 {
1123 	uint64_t gpt_desc;
1124 	int ret;
1125 
1126 	/* Ensure that MMU and Data caches are enabled */
1127 	assert((read_sctlr_el3() & SCTLR_C_BIT) != 0UL);
1128 
1129 	/* Validate other parameters */
1130 	ret = validate_l0_params(pps, l0_mem_base, l0_mem_size);
1131 	if (ret != 0) {
1132 		return ret;
1133 	}
1134 
1135 	/* Create the descriptor to initialize L0 entries with */
1136 	gpt_desc = GPT_L0_BLK_DESC(GPT_GPI_ANY);
1137 
1138 	/* Iterate through all L0 entries */
1139 	for (unsigned int i = 0U; i < GPT_L0_REGION_COUNT(gpt_config.t); i++) {
1140 		((uint64_t *)l0_mem_base)[i] = gpt_desc;
1141 	}
1142 
1143 	/* Flush updated L0 table to memory */
1144 	flush_dcache_range((uintptr_t)l0_mem_base, GPT_L0_TABLE_SIZE(gpt_config.t));
1145 
1146 	/* Stash the L0 base address once initial setup is complete */
1147 	gpt_config.plat_gpt_l0_base = l0_mem_base;
1148 
1149 	return 0;
1150 }
1151 
1152 /*
1153  * Public API that carves out PAS regions from the L0 tables and builds any L1
1154  * tables that are needed. This function ideally is run after DDR discovery and
1155  * initialization. The L0 tables must have already been initialized to GPI_ANY
1156  * when this function is called.
1157  *
1158  * This function can be called multiple times with different L1 memory ranges
1159  * and PAS regions if it is desirable to place L1 tables in different locations
1160  * in memory. (ex: you have multiple DDR banks and want to place the L1 tables
1161  * in the DDR bank that they control).
1162  *
1163  * Parameters
1164  *   pgs		PGS value to use for table generation.
1165  *   l1_mem_base	Base address of memory used for L1 tables.
1166  *   l1_mem_size	Total size of memory available for L1 tables.
1167  *   *pas_regions	Pointer to PAS regions structure array.
1168  *   pas_count		Total number of PAS regions.
1169  *
1170  * Return
1171  *   Negative Linux error code in the event of a failure, 0 for success.
1172  */
1173 int gpt_init_pas_l1_tables(gpccr_pgs_e pgs, uintptr_t l1_mem_base,
1174 			   size_t l1_mem_size, pas_region_t *pas_regions,
1175 			   unsigned int pas_count)
1176 {
1177 	int l1_gpt_cnt, ret;
1178 
1179 	/* Ensure that MMU and Data caches are enabled */
1180 	assert((read_sctlr_el3() & SCTLR_C_BIT) != 0UL);
1181 
1182 	/* PGS is needed for validate_pas_mappings so check it now */
1183 	if (pgs > GPT_PGS_MAX) {
1184 		ERROR("GPT: Invalid PGS: 0x%x\n", pgs);
1185 		return -EINVAL;
1186 	}
1187 	gpt_config.pgs = pgs;
1188 	gpt_config.p = gpt_p_lookup[pgs];
1189 
1190 	/* Make sure L0 tables have been initialized */
1191 	if (gpt_config.plat_gpt_l0_base == 0UL) {
1192 		ERROR("GPT: L0 tables must be initialized first!\n");
1193 		return -EPERM;
1194 	}
1195 
1196 	/* Check if L1 GPTs are required and how many */
1197 	l1_gpt_cnt = validate_pas_mappings(pas_regions, pas_count);
1198 	if (l1_gpt_cnt < 0) {
1199 		return l1_gpt_cnt;
1200 	}
1201 
1202 	VERBOSE("GPT: %i L1 GPTs requested\n", l1_gpt_cnt);
1203 
1204 	/* If L1 tables are needed then validate the L1 parameters */
1205 	if (l1_gpt_cnt > 0) {
1206 		ret = validate_l1_params(l1_mem_base, l1_mem_size,
1207 					(unsigned int)l1_gpt_cnt);
1208 		if (ret != 0) {
1209 			return ret;
1210 		}
1211 
1212 		/* Set up parameters for L1 table generation */
1213 		gpt_l1_tbl = l1_mem_base;
1214 	}
1215 
1216 	/* Number of L1 entries in 2MB depends on GPCCR_EL3.PGS value */
1217 	gpt_l1_cnt_2mb = (unsigned int)GPT_L1_ENTRY_COUNT_2MB(gpt_config.p);
1218 
1219 	/* Mask for the L1 index field */
1220 	gpt_l1_index_mask = GPT_L1_IDX_MASK(gpt_config.p);
1221 
1222 	INFO("GPT: Boot Configuration\n");
1223 	INFO("  PPS/T:     0x%x/%u\n", gpt_config.pps, gpt_config.t);
1224 	INFO("  PGS/P:     0x%x/%u\n", gpt_config.pgs, gpt_config.p);
1225 	INFO("  L0GPTSZ/S: 0x%x/%u\n", GPT_L0GPTSZ, GPT_S_VAL);
1226 	INFO("  PAS count: %u\n", pas_count);
1227 	INFO("  L0 base:   0x%"PRIxPTR"\n", gpt_config.plat_gpt_l0_base);
1228 
1229 	/* Generate the tables in memory */
1230 	for (unsigned int idx = 0U; idx < pas_count; idx++) {
1231 		VERBOSE("GPT: PAS[%u]: base 0x%"PRIxPTR"\tsize 0x%lx\tGPI 0x%x\ttype 0x%x\n",
1232 			idx, pas_regions[idx].base_pa, pas_regions[idx].size,
1233 			GPT_PAS_ATTR_GPI(pas_regions[idx].attrs),
1234 			GPT_PAS_ATTR_MAP_TYPE(pas_regions[idx].attrs));
1235 
1236 		/* Check if a block or table descriptor is required */
1237 		if (GPT_PAS_ATTR_MAP_TYPE(pas_regions[idx].attrs) ==
1238 		    GPT_PAS_ATTR_MAP_TYPE_BLOCK) {
1239 			generate_l0_blk_desc(&pas_regions[idx]);
1240 
1241 		} else {
1242 			generate_l0_tbl_desc(&pas_regions[idx]);
1243 		}
1244 	}
1245 
1246 	/* Flush modified L0 tables */
1247 	flush_l0_for_pas_array(pas_regions, pas_count);
1248 
1249 	/* Flush L1 tables if needed */
1250 	if (l1_gpt_cnt > 0) {
1251 		flush_dcache_range(l1_mem_base,
1252 				   GPT_L1_TABLE_SIZE(gpt_config.p) *
1253 				   (size_t)l1_gpt_cnt);
1254 	}
1255 
1256 	/* Make sure that all the entries are written to the memory */
1257 	dsbishst();
1258 	tlbipaallos();
1259 	dsb();
1260 	isb();
1261 
1262 	return 0;
1263 }
1264 
1265 /*
1266  * Public API to initialize the runtime gpt_config structure based on the values
1267  * present in the GPTBR_EL3 and GPCCR_EL3 registers. GPT initialization
1268  * typically happens in a bootloader stage prior to setting up the EL3 runtime
1269  * environment for the granule transition service so this function detects the
1270  * initialization from a previous stage. Granule protection checks must be
1271  * enabled already or this function will return an error.
1272  *
1273  * Parameters
1274  *   l1_bitlocks_base	Base address of memory for L1 tables bitlocks.
1275  *   l1_bitlocks_size	Total size of memory available for L1 tables bitlocks.
1276  *
1277  * Return
1278  *   Negative Linux error code in the event of a failure, 0 for success.
1279  */
1280 int gpt_runtime_init(uintptr_t l1_bitlocks_base, size_t l1_bitlocks_size)
1281 {
1282 	u_register_t reg;
1283 	__unused size_t locks_size;
1284 
1285 	/* Ensure that MMU and Data caches are enabled */
1286 	assert((read_sctlr_el3() & SCTLR_C_BIT) != 0UL);
1287 
1288 	/* Ensure GPC are already enabled */
1289 	if ((read_gpccr_el3() & GPCCR_GPC_BIT) == 0UL) {
1290 		ERROR("GPT: Granule protection checks are not enabled!\n");
1291 		return -EPERM;
1292 	}
1293 
1294 	/*
1295 	 * Read the L0 table address from GPTBR, we don't need the L1 base
1296 	 * address since those are included in the L0 tables as needed.
1297 	 */
1298 	reg = read_gptbr_el3();
1299 	gpt_config.plat_gpt_l0_base = ((reg >> GPTBR_BADDR_SHIFT) &
1300 				      GPTBR_BADDR_MASK) <<
1301 				      GPTBR_BADDR_VAL_SHIFT;
1302 
1303 	/* Read GPCCR to get PGS and PPS values */
1304 	reg = read_gpccr_el3();
1305 	gpt_config.pps = (reg >> GPCCR_PPS_SHIFT) & GPCCR_PPS_MASK;
1306 	gpt_config.t = gpt_t_lookup[gpt_config.pps];
1307 	gpt_config.pgs = (reg >> GPCCR_PGS_SHIFT) & GPCCR_PGS_MASK;
1308 	gpt_config.p = gpt_p_lookup[gpt_config.pgs];
1309 
1310 	/* Number of L1 entries in 2MB depends on GPCCR_EL3.PGS value */
1311 	gpt_l1_cnt_2mb = (unsigned int)GPT_L1_ENTRY_COUNT_2MB(gpt_config.p);
1312 
1313 	/* Mask for the L1 index field */
1314 	gpt_l1_index_mask = GPT_L1_IDX_MASK(gpt_config.p);
1315 
1316 #if (RME_GPT_BITLOCK_BLOCK != 0)
1317 	/*
1318 	 * Size of GPT bitlocks in bytes for the protected address space
1319 	 * with RME_GPT_BITLOCK_BLOCK * 512MB per bitlock.
1320 	 */
1321 	locks_size = GPT_PPS_ACTUAL_SIZE(gpt_config.t) /
1322 			(RME_GPT_BITLOCK_BLOCK * SZ_512M * 8U);
1323 	/*
1324 	 * If protected space size is less than the size covered
1325 	 * by 'bitlock' structure, check for a single bitlock.
1326 	 */
1327 	if (locks_size < LOCK_SIZE) {
1328 		locks_size = LOCK_SIZE;
1329 	/* Check bitlocks array size */
1330 	} else if (locks_size > l1_bitlocks_size) {
1331 		ERROR("GPT: Inadequate GPT bitlocks memory\n");
1332 		ERROR("      Expected 0x%lx bytes, got 0x%lx\n",
1333 			locks_size, l1_bitlocks_size);
1334 		return -ENOMEM;
1335 	}
1336 
1337 	gpt_bitlock = (bitlock_t *)l1_bitlocks_base;
1338 
1339 	/* Initialise GPT bitlocks */
1340 	(void)memset((void *)gpt_bitlock, 0, locks_size);
1341 
1342 	/* Flush GPT bitlocks to memory */
1343 	flush_dcache_range((uintptr_t)gpt_bitlock, locks_size);
1344 #endif /* RME_GPT_BITLOCK_BLOCK */
1345 
1346 	VERBOSE("GPT: Runtime Configuration\n");
1347 	VERBOSE("  PPS/T:     0x%x/%u\n", gpt_config.pps, gpt_config.t);
1348 	VERBOSE("  PGS/P:     0x%x/%u\n", gpt_config.pgs, gpt_config.p);
1349 	VERBOSE("  L0GPTSZ/S: 0x%x/%u\n", GPT_L0GPTSZ, GPT_S_VAL);
1350 	VERBOSE("  L0 base:   0x%"PRIxPTR"\n", gpt_config.plat_gpt_l0_base);
1351 #if (RME_GPT_BITLOCK_BLOCK != 0)
1352 	VERBOSE("  Bitlocks:  0x%"PRIxPTR"/0x%lx\n", (uintptr_t)gpt_bitlock,
1353 					locks_size);
1354 #endif
1355 	return 0;
1356 }
1357 
1358 /*
1359  * A helper to write the value (target_pas << gpi_shift) to the index of
1360  * the gpt_l1_addr.
1361  */
1362 static inline void write_gpt(uint64_t *gpt_l1_desc, uint64_t *gpt_l1_addr,
1363 			     unsigned int gpi_shift, unsigned int idx,
1364 			     unsigned int target_pas)
1365 {
1366 	*gpt_l1_desc &= ~(GPT_L1_GRAN_DESC_GPI_MASK << gpi_shift);
1367 	*gpt_l1_desc |= ((uint64_t)target_pas << gpi_shift);
1368 	gpt_l1_addr[idx] = *gpt_l1_desc;
1369 
1370 	dsboshst();
1371 }
1372 
1373 /*
1374  * Helper to retrieve the gpt_l1_* information from the base address
1375  * returned in gpi_info.
1376  */
1377 static int get_gpi_params(uint64_t base, gpi_info_t *gpi_info)
1378 {
1379 	uint64_t gpt_l0_desc, *gpt_l0_base;
1380 	__unused unsigned int block_idx;
1381 
1382 	gpt_l0_base = (uint64_t *)gpt_config.plat_gpt_l0_base;
1383 	gpt_l0_desc = gpt_l0_base[GPT_L0_IDX(base)];
1384 	if (GPT_L0_TYPE(gpt_l0_desc) != GPT_L0_TYPE_TBL_DESC) {
1385 		VERBOSE("GPT: Granule is not covered by a table descriptor!\n");
1386 		VERBOSE("      Base=0x%"PRIx64"\n", base);
1387 		return -EINVAL;
1388 	}
1389 
1390 	/* Get the table index and GPI shift from PA */
1391 	gpi_info->gpt_l1_addr = GPT_L0_TBLD_ADDR(gpt_l0_desc);
1392 	gpi_info->idx = (unsigned int)GPT_L1_INDEX(base);
1393 	gpi_info->gpi_shift = GPT_L1_GPI_IDX(gpt_config.p, base) << 2;
1394 
1395 #if (RME_GPT_BITLOCK_BLOCK != 0)
1396 	/* Block index */
1397 	block_idx = (unsigned int)(base / (RME_GPT_BITLOCK_BLOCK * SZ_512M));
1398 
1399 	/* Bitlock address and mask */
1400 	gpi_info->lock = &gpt_bitlock[block_idx / LOCK_BITS];
1401 	gpi_info->mask = 1U << (block_idx & (LOCK_BITS - 1U));
1402 #endif
1403 	return 0;
1404 }
1405 
1406 /*
1407  * Helper to retrieve the gpt_l1_desc and GPI information from gpi_info.
1408  * This function is called with bitlock or spinlock acquired.
1409  */
1410 static void read_gpi(gpi_info_t *gpi_info)
1411 {
1412 	gpi_info->gpt_l1_desc = (gpi_info->gpt_l1_addr)[gpi_info->idx];
1413 
1414 	if ((gpi_info->gpt_l1_desc & GPT_L1_TYPE_CONT_DESC_MASK) ==
1415 				 GPT_L1_TYPE_CONT_DESC) {
1416 		/* Read GPI from Contiguous descriptor */
1417 		gpi_info->gpi = (unsigned int)GPT_L1_CONT_GPI(gpi_info->gpt_l1_desc);
1418 	} else {
1419 		/* Read GPI from Granules descriptor */
1420 		gpi_info->gpi = (unsigned int)((gpi_info->gpt_l1_desc >> gpi_info->gpi_shift) &
1421 						GPT_L1_GRAN_DESC_GPI_MASK);
1422 	}
1423 }
1424 
1425 static void flush_page_to_popa(uintptr_t addr)
1426 {
1427 	size_t size = GPT_PGS_ACTUAL_SIZE(gpt_config.p);
1428 
1429 	if (is_feat_mte2_supported()) {
1430 		flush_dcache_to_popa_range_mte2(addr, size);
1431 	} else {
1432 		flush_dcache_to_popa_range(addr, size);
1433 	}
1434 }
1435 
1436 /*
1437  * Helper function to check if all L1 entries in 2MB block have
1438  * the same Granules descriptor value.
1439  *
1440  * Parameters
1441  *   base		Base address of the region to be checked
1442  *   gpi_info		Pointer to 'gpt_config_t' structure
1443  *   l1_desc		GPT Granules descriptor with all entries
1444  *			set to the same GPI.
1445  *
1446  * Return
1447  *   true if L1 all entries have the same descriptor value, false otherwise.
1448  */
1449 __unused static bool check_fuse_2mb(uint64_t base, const gpi_info_t *gpi_info,
1450 					uint64_t l1_desc)
1451 {
1452 	/* Last L1 entry index in 2MB block */
1453 	unsigned int long idx = GPT_L1_INDEX(ALIGN_2MB(base)) +
1454 						gpt_l1_cnt_2mb - 1UL;
1455 
1456 	/* Number of L1 entries in 2MB block */
1457 	unsigned int cnt = gpt_l1_cnt_2mb;
1458 
1459 	/*
1460 	 * Start check from the last L1 entry and continue until the first
1461 	 * non-matching to the passed Granules descriptor value is found.
1462 	 */
1463 	while (cnt-- != 0U) {
1464 		if (gpi_info->gpt_l1_addr[idx--] != l1_desc) {
1465 			/* Non-matching L1 entry found */
1466 			return false;
1467 		}
1468 	}
1469 
1470 	return true;
1471 }
1472 
1473 __unused static void fuse_2mb(uint64_t base, const gpi_info_t *gpi_info,
1474 				uint64_t l1_desc)
1475 {
1476 	/* L1 entry index of the start of 2MB block */
1477 	unsigned long idx_2 = GPT_L1_INDEX(ALIGN_2MB(base));
1478 
1479 	/* 2MB Contiguous descriptor */
1480 	uint64_t l1_cont_desc = GPT_L1_CONT_DESC(l1_desc, 2MB);
1481 
1482 	VERBOSE("GPT: %s(0x%"PRIxPTR" 0x%"PRIx64")\n", __func__, base, l1_desc);
1483 
1484 	fill_desc(&gpi_info->gpt_l1_addr[idx_2], l1_cont_desc, L1_QWORDS_2MB);
1485 }
1486 
1487 /*
1488  * Helper function to check if all 1st L1 entries of 2MB blocks
1489  * in 32MB have the same 2MB Contiguous descriptor value.
1490  *
1491  * Parameters
1492  *   base		Base address of the region to be checked
1493  *   gpi_info		Pointer to 'gpt_config_t' structure
1494  *   l1_desc		GPT Granules descriptor.
1495  *
1496  * Return
1497  *   true if all L1 entries have the same descriptor value, false otherwise.
1498  */
1499 __unused static bool check_fuse_32mb(uint64_t base, const gpi_info_t *gpi_info,
1500 					uint64_t l1_desc)
1501 {
1502 	/* The 1st L1 entry index of the last 2MB block in 32MB */
1503 	unsigned long idx = GPT_L1_INDEX(ALIGN_32MB(base)) +
1504 					(15UL * gpt_l1_cnt_2mb);
1505 
1506 	/* 2MB Contiguous descriptor */
1507 	uint64_t l1_cont_desc = GPT_L1_CONT_DESC(l1_desc, 2MB);
1508 
1509 	/* Number of 2MB blocks in 32MB */
1510 	unsigned int cnt = 16U;
1511 
1512 	/* Set the first L1 entry to 2MB Contiguous descriptor */
1513 	gpi_info->gpt_l1_addr[GPT_L1_INDEX(ALIGN_2MB(base))] = l1_cont_desc;
1514 
1515 	/*
1516 	 * Start check from the 1st L1 entry of the last 2MB block and
1517 	 * continue until the first non-matching to 2MB Contiguous descriptor
1518 	 * value is found.
1519 	 */
1520 	while (cnt-- != 0U) {
1521 		if (gpi_info->gpt_l1_addr[idx] != l1_cont_desc) {
1522 			/* Non-matching L1 entry found */
1523 			return false;
1524 		}
1525 		idx -= gpt_l1_cnt_2mb;
1526 	}
1527 
1528 	return true;
1529 }
1530 
1531 __unused static void fuse_32mb(uint64_t base, const gpi_info_t *gpi_info,
1532 				uint64_t l1_desc)
1533 {
1534 	/* L1 entry index of the start of 32MB block */
1535 	unsigned long idx_32 = GPT_L1_INDEX(ALIGN_32MB(base));
1536 
1537 	/* 32MB Contiguous descriptor */
1538 	uint64_t l1_cont_desc = GPT_L1_CONT_DESC(l1_desc, 32MB);
1539 
1540 	VERBOSE("GPT: %s(0x%"PRIxPTR" 0x%"PRIx64")\n", __func__, base, l1_desc);
1541 
1542 	fill_desc(&gpi_info->gpt_l1_addr[idx_32], l1_cont_desc, L1_QWORDS_32MB);
1543 }
1544 
1545 /*
1546  * Helper function to check if all 1st L1 entries of 32MB blocks
1547  * in 512MB have the same 32MB Contiguous descriptor value.
1548  *
1549  * Parameters
1550  *   base		Base address of the region to be checked
1551  *   gpi_info		Pointer to 'gpt_config_t' structure
1552  *   l1_desc		GPT Granules descriptor.
1553  *
1554  * Return
1555  *   true if all L1 entries have the same descriptor value, false otherwise.
1556  */
1557 __unused static bool check_fuse_512mb(uint64_t base, const gpi_info_t *gpi_info,
1558 					uint64_t l1_desc)
1559 {
1560 	/* The 1st L1 entry index of the last 32MB block in 512MB */
1561 	unsigned long idx = GPT_L1_INDEX(ALIGN_512MB(base)) +
1562 					(15UL * 16UL * gpt_l1_cnt_2mb);
1563 
1564 	/* 32MB Contiguous descriptor */
1565 	uint64_t l1_cont_desc = GPT_L1_CONT_DESC(l1_desc, 32MB);
1566 
1567 	/* Number of 32MB blocks in 512MB */
1568 	unsigned int cnt = 16U;
1569 
1570 	/* Set the first L1 entry to 2MB Contiguous descriptor */
1571 	gpi_info->gpt_l1_addr[GPT_L1_INDEX(ALIGN_32MB(base))] = l1_cont_desc;
1572 
1573 	/*
1574 	 * Start check from the 1st L1 entry of the last 32MB block and
1575 	 * continue until the first non-matching to 32MB Contiguous descriptor
1576 	 * value is found.
1577 	 */
1578 	while (cnt-- != 0U) {
1579 		if (gpi_info->gpt_l1_addr[idx] != l1_cont_desc) {
1580 			/* Non-matching L1 entry found */
1581 			return false;
1582 		}
1583 		idx -= 16UL * gpt_l1_cnt_2mb;
1584 	}
1585 
1586 	return true;
1587 }
1588 
1589 __unused static void fuse_512mb(uint64_t base, const gpi_info_t *gpi_info,
1590 				uint64_t l1_desc)
1591 {
1592 	/* L1 entry index of the start of 512MB block */
1593 	unsigned long idx_512 = GPT_L1_INDEX(ALIGN_512MB(base));
1594 
1595 	/* 512MB Contiguous descriptor */
1596 	uint64_t l1_cont_desc = GPT_L1_CONT_DESC(l1_desc, 512MB);
1597 
1598 	VERBOSE("GPT: %s(0x%"PRIxPTR" 0x%"PRIx64")\n", __func__, base, l1_desc);
1599 
1600 	fill_desc(&gpi_info->gpt_l1_addr[idx_512], l1_cont_desc, L1_QWORDS_512MB);
1601 }
1602 
1603 /*
1604  * Helper function to convert GPI entries in a single L1 table
1605  * from Granules to Contiguous descriptor.
1606  *
1607  * Parameters
1608  *   base		Base address of the region to be written
1609  *   gpi_info		Pointer to 'gpt_config_t' structure
1610  *   l1_desc		GPT Granules descriptor with all entries
1611  *			set to the same GPI.
1612  */
1613 __unused static void fuse_block(uint64_t base, const gpi_info_t *gpi_info,
1614 				uint64_t l1_desc)
1615 {
1616 	/* Start with check for 2MB block */
1617 	if (!check_fuse_2mb(base, gpi_info, l1_desc)) {
1618 		/* Check for 2MB fusing failed */
1619 		return;
1620 	}
1621 
1622 #if (RME_GPT_MAX_BLOCK == 2)
1623 	fuse_2mb(base, gpi_info, l1_desc);
1624 #else
1625 	/* Check for 32MB block */
1626 	if (!check_fuse_32mb(base, gpi_info, l1_desc)) {
1627 		/* Check for 32MB fusing failed, fuse to 2MB */
1628 		fuse_2mb(base, gpi_info, l1_desc);
1629 		return;
1630 	}
1631 
1632 #if (RME_GPT_MAX_BLOCK == 32)
1633 	fuse_32mb(base, gpi_info, l1_desc);
1634 #else
1635 	/* Check for 512MB block */
1636 	if (!check_fuse_512mb(base, gpi_info, l1_desc)) {
1637 		/* Check for 512MB fusing failed, fuse to 32MB */
1638 		fuse_32mb(base, gpi_info, l1_desc);
1639 		return;
1640 	}
1641 
1642 	/* Fuse to 512MB */
1643 	fuse_512mb(base, gpi_info, l1_desc);
1644 
1645 #endif	/* RME_GPT_MAX_BLOCK == 32 */
1646 #endif	/* RME_GPT_MAX_BLOCK == 2 */
1647 }
1648 
1649 /*
1650  * Helper function to convert GPI entries in a single L1 table
1651  * from Contiguous to Granules descriptor. This function updates
1652  * descriptor to Granules in passed 'gpt_config_t' structure as
1653  * the result of shuttering.
1654  *
1655  * Parameters
1656  *   base		Base address of the region to be written
1657  *   gpi_info		Pointer to 'gpt_config_t' structure
1658  *   l1_desc		GPT Granules descriptor set this range to.
1659  */
1660 __unused static void shatter_block(uint64_t base, gpi_info_t *gpi_info,
1661 				   uint64_t l1_desc)
1662 {
1663 	/* Look-up table for 2MB, 32MB and 512MB locks shattering */
1664 	static const gpt_shatter_func gpt_shatter_lookup[] = {
1665 		shatter_2mb,
1666 		shatter_32mb,
1667 		shatter_512mb
1668 	};
1669 
1670 	/* Look-up table for invalidation TLBs for 2MB, 32MB and 512MB blocks */
1671 	static const gpt_tlbi_lookup_t tlbi_lookup[] = {
1672 		{ tlbirpalos_2m, ~(SZ_2M - 1UL) },
1673 		{ tlbirpalos_32m, ~(SZ_32M - 1UL) },
1674 		{ tlbirpalos_512m, ~(SZ_512M - 1UL) }
1675 	};
1676 
1677 	/* Get shattering level from Contig field of Contiguous descriptor */
1678 	unsigned long level = GPT_L1_CONT_CONTIG(gpi_info->gpt_l1_desc) - 1UL;
1679 
1680 	/* Shatter contiguous block */
1681 	gpt_shatter_lookup[level](base, gpi_info, l1_desc);
1682 
1683 	tlbi_lookup[level].function(base & tlbi_lookup[level].mask);
1684 	dsbosh();
1685 
1686 	/*
1687 	 * Update 'gpt_config_t' structure's descriptor to Granules to reflect
1688 	 * the shattered GPI back to caller.
1689 	 */
1690 	gpi_info->gpt_l1_desc = l1_desc;
1691 }
1692 
1693 /*
1694  * This function is the granule transition delegate service. When a granule
1695  * transition request occurs it is routed to this function to have the request,
1696  * if valid, fulfilled following A1.1.1 Delegate of RME supplement.
1697  *
1698  * TODO: implement support for transitioning multiple granules at once.
1699  *
1700  * Parameters
1701  *   base		Base address of the region to transition, must be
1702  *			aligned to granule size.
1703  *   size		Size of region to transition, must be aligned to granule
1704  *			size.
1705  *   src_sec_state	Security state of the caller.
1706  *
1707  * Return
1708  *   Negative Linux error code in the event of a failure, 0 for success.
1709  */
1710 int gpt_delegate_pas(uint64_t base, size_t size, unsigned int src_sec_state)
1711 {
1712 	gpi_info_t gpi_info;
1713 	uint64_t nse, __unused l1_desc;
1714 	unsigned int target_pas;
1715 	int res;
1716 
1717 	/* Ensure that the tables have been set up before taking requests */
1718 	assert(gpt_config.plat_gpt_l0_base != 0UL);
1719 
1720 	/* Ensure that caches are enabled */
1721 	assert((read_sctlr_el3() & SCTLR_C_BIT) != 0UL);
1722 
1723 	/* See if this is a single or a range of granule transition */
1724 	if (size != GPT_PGS_ACTUAL_SIZE(gpt_config.p)) {
1725 		return -EINVAL;
1726 	}
1727 
1728 	/* Check that base and size are valid */
1729 	if ((ULONG_MAX - base) < size) {
1730 		VERBOSE("GPT: Transition request address overflow!\n");
1731 		VERBOSE("      Base=0x%"PRIx64"\n", base);
1732 		VERBOSE("      Size=0x%lx\n", size);
1733 		return -EINVAL;
1734 	}
1735 
1736 	/* Make sure base and size are valid */
1737 	if (((base & (GPT_PGS_ACTUAL_SIZE(gpt_config.p) - 1UL)) != 0UL) ||
1738 	    ((size & (GPT_PGS_ACTUAL_SIZE(gpt_config.p) - 1UL)) != 0UL) ||
1739 	    (size == 0UL) ||
1740 	    ((base + size) >= GPT_PPS_ACTUAL_SIZE(gpt_config.t))) {
1741 		VERBOSE("GPT: Invalid granule transition address range!\n");
1742 		VERBOSE("      Base=0x%"PRIx64"\n", base);
1743 		VERBOSE("      Size=0x%lx\n", size);
1744 		return -EINVAL;
1745 	}
1746 
1747 	/* Delegate request can only come from REALM or SECURE */
1748 	if ((src_sec_state != SMC_FROM_REALM) &&
1749 	    (src_sec_state != SMC_FROM_SECURE)) {
1750 		VERBOSE("GPT: Invalid caller security state 0x%x\n",
1751 							src_sec_state);
1752 		return -EINVAL;
1753 	}
1754 
1755 	if (src_sec_state == SMC_FROM_REALM) {
1756 		target_pas = GPT_GPI_REALM;
1757 		nse = (uint64_t)GPT_NSE_REALM << GPT_NSE_SHIFT;
1758 		l1_desc = GPT_L1_REALM_DESC;
1759 	} else {
1760 		target_pas = GPT_GPI_SECURE;
1761 		nse = (uint64_t)GPT_NSE_SECURE << GPT_NSE_SHIFT;
1762 		l1_desc = GPT_L1_SECURE_DESC;
1763 	}
1764 
1765 	res = get_gpi_params(base, &gpi_info);
1766 	if (res != 0) {
1767 		return res;
1768 	}
1769 
1770 	/*
1771 	 * Access to GPT is controlled by a lock to ensure that no more
1772 	 * than one CPU is allowed to make changes at any given time.
1773 	 */
1774 	GPT_LOCK;
1775 	read_gpi(&gpi_info);
1776 
1777 	/* Check that the current address is in NS state */
1778 	if (gpi_info.gpi != GPT_GPI_NS) {
1779 		VERBOSE("GPT: Only Granule in NS state can be delegated.\n");
1780 		VERBOSE("      Caller: %u, Current GPI: %u\n", src_sec_state,
1781 			gpi_info.gpi);
1782 		GPT_UNLOCK;
1783 		return -EPERM;
1784 	}
1785 
1786 #if (RME_GPT_MAX_BLOCK != 0)
1787 	/* Check for Contiguous descriptor */
1788 	if ((gpi_info.gpt_l1_desc & GPT_L1_TYPE_CONT_DESC_MASK) ==
1789 					GPT_L1_TYPE_CONT_DESC) {
1790 		shatter_block(base, &gpi_info, GPT_L1_NS_DESC);
1791 	}
1792 #endif
1793 	/*
1794 	 * In order to maintain mutual distrust between Realm and Secure
1795 	 * states, remove any data speculatively fetched into the target
1796 	 * physical address space.
1797 	 * Issue DC CIPAPA or DC_CIGDPAPA on implementations with FEAT_MTE2.
1798 	 */
1799 	flush_page_to_popa(base | nse);
1800 
1801 	write_gpt(&gpi_info.gpt_l1_desc, gpi_info.gpt_l1_addr,
1802 		  gpi_info.gpi_shift, gpi_info.idx, target_pas);
1803 
1804 	/* Ensure that all agents observe the new configuration */
1805 	tlbi_page_dsbosh(base);
1806 
1807 	nse = (uint64_t)GPT_NSE_NS << GPT_NSE_SHIFT;
1808 
1809 	/* Ensure that the scrubbed data have made it past the PoPA */
1810 	flush_page_to_popa(base | nse);
1811 
1812 #if (RME_GPT_MAX_BLOCK != 0)
1813 	if (gpi_info.gpt_l1_desc == l1_desc) {
1814 		/* Try to fuse */
1815 		fuse_block(base, &gpi_info, l1_desc);
1816 	}
1817 #endif
1818 
1819 	/* Unlock the lock to GPT */
1820 	GPT_UNLOCK;
1821 
1822 	/*
1823 	 * The isb() will be done as part of context
1824 	 * synchronization when returning to lower EL.
1825 	 */
1826 	VERBOSE("GPT: Granule 0x%"PRIx64" GPI 0x%x->0x%x\n",
1827 		base, gpi_info.gpi, target_pas);
1828 
1829 	return 0;
1830 }
1831 
1832 /*
1833  * This function is the granule transition undelegate service. When a granule
1834  * transition request occurs it is routed to this function where the request is
1835  * validated then fulfilled if possible.
1836  *
1837  * TODO: implement support for transitioning multiple granules at once.
1838  *
1839  * Parameters
1840  *   base		Base address of the region to transition, must be
1841  *			aligned to granule size.
1842  *   size		Size of region to transition, must be aligned to granule
1843  *			size.
1844  *   src_sec_state	Security state of the caller.
1845  *
1846  * Return
1847  *    Negative Linux error code in the event of a failure, 0 for success.
1848  */
1849 int gpt_undelegate_pas(uint64_t base, size_t size, unsigned int src_sec_state)
1850 {
1851 	gpi_info_t gpi_info;
1852 	uint64_t nse, __unused l1_desc;
1853 	int res;
1854 
1855 	/* Ensure that the tables have been set up before taking requests */
1856 	assert(gpt_config.plat_gpt_l0_base != 0UL);
1857 
1858 	/* Ensure that MMU and caches are enabled */
1859 	assert((read_sctlr_el3() & SCTLR_C_BIT) != 0UL);
1860 
1861 	/* See if this is a single or a range of granule transition */
1862 	if (size != GPT_PGS_ACTUAL_SIZE(gpt_config.p)) {
1863 		return -EINVAL;
1864 	}
1865 
1866 	/* Check that base and size are valid */
1867 	if ((ULONG_MAX - base) < size) {
1868 		VERBOSE("GPT: Transition request address overflow!\n");
1869 		VERBOSE("      Base=0x%"PRIx64"\n", base);
1870 		VERBOSE("      Size=0x%lx\n", size);
1871 		return -EINVAL;
1872 	}
1873 
1874 	/* Make sure base and size are valid */
1875 	if (((base & (GPT_PGS_ACTUAL_SIZE(gpt_config.p) - 1UL)) != 0UL) ||
1876 	    ((size & (GPT_PGS_ACTUAL_SIZE(gpt_config.p) - 1UL)) != 0UL) ||
1877 	    (size == 0UL) ||
1878 	    ((base + size) >= GPT_PPS_ACTUAL_SIZE(gpt_config.t))) {
1879 		VERBOSE("GPT: Invalid granule transition address range!\n");
1880 		VERBOSE("      Base=0x%"PRIx64"\n", base);
1881 		VERBOSE("      Size=0x%lx\n", size);
1882 		return -EINVAL;
1883 	}
1884 
1885 	res = get_gpi_params(base, &gpi_info);
1886 	if (res != 0) {
1887 		return res;
1888 	}
1889 
1890 	/*
1891 	 * Access to GPT is controlled by a lock to ensure that no more
1892 	 * than one CPU is allowed to make changes at any given time.
1893 	 */
1894 	GPT_LOCK;
1895 	read_gpi(&gpi_info);
1896 
1897 	/* Check that the current address is in the delegated state */
1898 	if ((src_sec_state == SMC_FROM_REALM) &&
1899 		(gpi_info.gpi == GPT_GPI_REALM)) {
1900 		l1_desc = GPT_L1_REALM_DESC;
1901 		nse = (uint64_t)GPT_NSE_REALM << GPT_NSE_SHIFT;
1902 	} else if ((src_sec_state == SMC_FROM_SECURE) &&
1903 		(gpi_info.gpi == GPT_GPI_SECURE)) {
1904 		l1_desc = GPT_L1_SECURE_DESC;
1905 		nse = (uint64_t)GPT_NSE_SECURE << GPT_NSE_SHIFT;
1906 	} else {
1907 		VERBOSE("GPT: Only Granule in REALM or SECURE state can be undelegated\n");
1908 		VERBOSE("      Caller: %u Current GPI: %u\n", src_sec_state,
1909 			gpi_info.gpi);
1910 		GPT_UNLOCK;
1911 		return -EPERM;
1912 	}
1913 
1914 #if (RME_GPT_MAX_BLOCK != 0)
1915 	/* Check for Contiguous descriptor */
1916 	if ((gpi_info.gpt_l1_desc & GPT_L1_TYPE_CONT_DESC_MASK) ==
1917 					GPT_L1_TYPE_CONT_DESC) {
1918 		shatter_block(base, &gpi_info, l1_desc);
1919 	}
1920 #endif
1921 	/*
1922 	 * In order to maintain mutual distrust between Realm and Secure
1923 	 * states, remove access now, in order to guarantee that writes
1924 	 * to the currently-accessible physical address space will not
1925 	 * later become observable.
1926 	 */
1927 	write_gpt(&gpi_info.gpt_l1_desc, gpi_info.gpt_l1_addr,
1928 		  gpi_info.gpi_shift, gpi_info.idx, GPT_GPI_NO_ACCESS);
1929 
1930 	/* Ensure that all agents observe the new NO_ACCESS configuration */
1931 	tlbi_page_dsbosh(base);
1932 
1933 	/* Ensure that the scrubbed data have made it past the PoPA */
1934 	flush_page_to_popa(base | nse);
1935 
1936 	/*
1937 	 * Remove any data loaded speculatively in NS space from before
1938 	 * the scrubbing.
1939 	 */
1940 	nse = (uint64_t)GPT_NSE_NS << GPT_NSE_SHIFT;
1941 
1942 	flush_page_to_popa(base | nse);
1943 
1944 	/* Clear existing GPI encoding and transition granule */
1945 	write_gpt(&gpi_info.gpt_l1_desc, gpi_info.gpt_l1_addr,
1946 		  gpi_info.gpi_shift, gpi_info.idx, GPT_GPI_NS);
1947 
1948 	/* Ensure that all agents observe the new NS configuration */
1949 	tlbi_page_dsbosh(base);
1950 
1951 #if (RME_GPT_MAX_BLOCK != 0)
1952 	if (gpi_info.gpt_l1_desc == GPT_L1_NS_DESC) {
1953 		/* Try to fuse */
1954 		fuse_block(base, &gpi_info, GPT_L1_NS_DESC);
1955 	}
1956 #endif
1957 	/* Unlock the lock to GPT */
1958 	GPT_UNLOCK;
1959 
1960 	/*
1961 	 * The isb() will be done as part of context
1962 	 * synchronization when returning to lower EL.
1963 	 */
1964 	VERBOSE("GPT: Granule 0x%"PRIx64" GPI 0x%x->0x%x\n",
1965 		base, gpi_info.gpi, GPT_GPI_NS);
1966 
1967 	return 0;
1968 }
1969