xref: /rk3399_ARM-atf/lib/gpt_rme/gpt_rme.c (revision 52e486f6a6192bd18d36cdcbc35c59092eefc810)
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_SECURE      U(0x8)
264  *   GPT_GPI_NS          U(0x9)
265  *   GPT_GPI_ROOT        U(0xA)
266  *   GPT_GPI_REALM       U(0xB)
267  *   GPT_GPI_ANY         U(0xF)
268  *
269  * Parameters
270  *   gpi		GPI to check for validity.
271  *
272  * Return
273  *   true for a valid GPI, false for an invalid one.
274  */
275 static bool is_gpi_valid(unsigned int gpi)
276 {
277 	if ((gpi == GPT_GPI_NO_ACCESS) || (gpi == GPT_GPI_ANY) ||
278 	    ((gpi >= GPT_GPI_SECURE) && (gpi <= GPT_GPI_REALM))) {
279 		return true;
280 	}
281 	return false;
282 }
283 
284 /*
285  * This function checks to see if two PAS regions overlap.
286  *
287  * Parameters
288  *   base_1: base address of first PAS
289  *   size_1: size of first PAS
290  *   base_2: base address of second PAS
291  *   size_2: size of second PAS
292  *
293  * Return
294  *   True if PAS regions overlap, false if they do not.
295  */
296 static bool check_pas_overlap(uintptr_t base_1, size_t size_1,
297 			      uintptr_t base_2, size_t size_2)
298 {
299 	if (((base_1 + size_1) > base_2) && ((base_2 + size_2) > base_1)) {
300 		return true;
301 	}
302 	return false;
303 }
304 
305 /*
306  * This helper function checks to see if a PAS region from index 0 to
307  * (pas_idx - 1) occupies the L0 region at index l0_idx in the L0 table.
308  *
309  * Parameters
310  *   l0_idx:      Index of the L0 entry to check
311  *   pas_regions: PAS region array
312  *   pas_idx:     Upper bound of the PAS array index.
313  *
314  * Return
315  *   True if a PAS region occupies the L0 region in question, false if not.
316  */
317 static bool does_previous_pas_exist_here(unsigned int l0_idx,
318 					 pas_region_t *pas_regions,
319 					 unsigned int pas_idx)
320 {
321 	/* Iterate over PAS regions up to pas_idx */
322 	for (unsigned int i = 0U; i < pas_idx; i++) {
323 		if (check_pas_overlap((GPT_L0GPTSZ_ACTUAL_SIZE * l0_idx),
324 		    GPT_L0GPTSZ_ACTUAL_SIZE,
325 		    pas_regions[i].base_pa, pas_regions[i].size)) {
326 			return true;
327 		}
328 	}
329 	return false;
330 }
331 
332 /*
333  * This function iterates over all of the PAS regions and checks them to ensure
334  * proper alignment of base and size, that the GPI is valid, and that no regions
335  * overlap. As a part of the overlap checks, this function checks existing L0
336  * mappings against the new PAS regions in the event that gpt_init_pas_l1_tables
337  * is called multiple times to place L1 tables in different areas of memory. It
338  * also counts the number of L1 tables needed and returns it on success.
339  *
340  * Parameters
341  *   *pas_regions	Pointer to array of PAS region structures.
342  *   pas_region_cnt	Total number of PAS regions in the array.
343  *
344  * Return
345  *   Negative Linux error code in the event of a failure, number of L1 regions
346  *   required when successful.
347  */
348 static int validate_pas_mappings(pas_region_t *pas_regions,
349 				 unsigned int pas_region_cnt)
350 {
351 	unsigned int idx;
352 	unsigned int l1_cnt = 0U;
353 	unsigned int pas_l1_cnt;
354 	uint64_t *l0_desc = (uint64_t *)gpt_config.plat_gpt_l0_base;
355 
356 	assert(pas_regions != NULL);
357 	assert(pas_region_cnt != 0U);
358 
359 	for (idx = 0U; idx < pas_region_cnt; idx++) {
360 		/* Check for arithmetic overflow in region */
361 		if ((ULONG_MAX - pas_regions[idx].base_pa) <
362 		    pas_regions[idx].size) {
363 			ERROR("GPT: Address overflow in PAS[%u]!\n", idx);
364 			return -EOVERFLOW;
365 		}
366 
367 		/* Initial checks for PAS validity */
368 		if (((pas_regions[idx].base_pa + pas_regions[idx].size) >
369 		    GPT_PPS_ACTUAL_SIZE(gpt_config.t)) ||
370 		    !is_gpi_valid(GPT_PAS_ATTR_GPI(pas_regions[idx].attrs))) {
371 			ERROR("GPT: PAS[%u] is invalid!\n", idx);
372 			return -EFAULT;
373 		}
374 
375 		/*
376 		 * Make sure this PAS does not overlap with another one. We
377 		 * start from idx + 1 instead of 0 since prior PAS mappings will
378 		 * have already checked themselves against this one.
379 		 */
380 		for (unsigned int i = idx + 1U; i < pas_region_cnt; i++) {
381 			if (check_pas_overlap(pas_regions[idx].base_pa,
382 			    pas_regions[idx].size,
383 			    pas_regions[i].base_pa,
384 			    pas_regions[i].size)) {
385 				ERROR("GPT: PAS[%u] overlaps with PAS[%u]\n",
386 					i, idx);
387 				return -EFAULT;
388 			}
389 		}
390 
391 		/*
392 		 * Since this function can be called multiple times with
393 		 * separate L1 tables we need to check the existing L0 mapping
394 		 * to see if this PAS would fall into one that has already been
395 		 * initialized.
396 		 */
397 		for (unsigned int i =
398 			(unsigned int)GPT_L0_IDX(pas_regions[idx].base_pa);
399 			i <= GPT_L0_IDX(pas_regions[idx].base_pa +
400 					pas_regions[idx].size - 1UL);
401 			i++) {
402 			if ((GPT_L0_TYPE(l0_desc[i]) == GPT_L0_TYPE_BLK_DESC) &&
403 			    (GPT_L0_BLKD_GPI(l0_desc[i]) == GPT_GPI_ANY)) {
404 				/* This descriptor is unused so continue */
405 				continue;
406 			}
407 
408 			/*
409 			 * This descriptor has been initialized in a previous
410 			 * call to this function so cannot be initialized again.
411 			 */
412 			ERROR("GPT: PAS[%u] overlaps with previous L0[%u]!\n",
413 			      idx, i);
414 			return -EFAULT;
415 		}
416 
417 		/* Check for block mapping (L0) type */
418 		if (GPT_PAS_ATTR_MAP_TYPE(pas_regions[idx].attrs) ==
419 		    GPT_PAS_ATTR_MAP_TYPE_BLOCK) {
420 			/* Make sure base and size are block-aligned */
421 			if (!GPT_IS_L0_ALIGNED(pas_regions[idx].base_pa) ||
422 			    !GPT_IS_L0_ALIGNED(pas_regions[idx].size)) {
423 				ERROR("GPT: PAS[%u] is not block-aligned!\n",
424 				      idx);
425 				return -EFAULT;
426 			}
427 
428 			continue;
429 		}
430 
431 		/* Check for granule mapping (L1) type */
432 		if (GPT_PAS_ATTR_MAP_TYPE(pas_regions[idx].attrs) ==
433 		    GPT_PAS_ATTR_MAP_TYPE_GRANULE) {
434 			/* Make sure base and size are granule-aligned */
435 			if (!GPT_IS_L1_ALIGNED(gpt_config.p, pas_regions[idx].base_pa) ||
436 			    !GPT_IS_L1_ALIGNED(gpt_config.p, pas_regions[idx].size)) {
437 				ERROR("GPT: PAS[%u] is not granule-aligned!\n",
438 				      idx);
439 				return -EFAULT;
440 			}
441 
442 			/* Find how many L1 tables this PAS occupies */
443 			pas_l1_cnt = (GPT_L0_IDX(pas_regions[idx].base_pa +
444 				     pas_regions[idx].size - 1UL) -
445 				     GPT_L0_IDX(pas_regions[idx].base_pa) + 1U);
446 
447 			/*
448 			 * This creates a situation where, if multiple PAS
449 			 * regions occupy the same table descriptor, we can get
450 			 * an artificially high total L1 table count. The way we
451 			 * handle this is by checking each PAS against those
452 			 * before it in the array, and if they both occupy the
453 			 * same PAS we subtract from pas_l1_cnt and only the
454 			 * first PAS in the array gets to count it.
455 			 */
456 
457 			/*
458 			 * If L1 count is greater than 1 we know the start and
459 			 * end PAs are in different L0 regions so we must check
460 			 * both for overlap against other PAS.
461 			 */
462 			if (pas_l1_cnt > 1) {
463 				if (does_previous_pas_exist_here(
464 				    GPT_L0_IDX(pas_regions[idx].base_pa +
465 				    pas_regions[idx].size - 1UL),
466 				    pas_regions, idx)) {
467 					pas_l1_cnt--;
468 				}
469 			}
470 
471 			if (does_previous_pas_exist_here(
472 			    GPT_L0_IDX(pas_regions[idx].base_pa),
473 			    pas_regions, idx)) {
474 				pas_l1_cnt--;
475 			}
476 
477 			l1_cnt += pas_l1_cnt;
478 			continue;
479 		}
480 
481 		/* If execution reaches this point, mapping type is invalid */
482 		ERROR("GPT: PAS[%u] has invalid mapping type 0x%x.\n", idx,
483 		      GPT_PAS_ATTR_MAP_TYPE(pas_regions[idx].attrs));
484 		return -EINVAL;
485 	}
486 
487 	return l1_cnt;
488 }
489 
490 /*
491  * This function validates L0 initialization parameters.
492  *
493  * Parameters
494  *   l0_mem_base	Base address of memory used for L0 table.
495  *   l0_mem_size	Size of memory available for L0 table.
496  *
497  * Return
498  *   Negative Linux error code in the event of a failure, 0 for success.
499  */
500 static int validate_l0_params(gpccr_pps_e pps, uintptr_t l0_mem_base,
501 				size_t l0_mem_size)
502 {
503 	size_t l0_alignment;
504 
505 	/*
506 	 * Make sure PPS is valid and then store it since macros need this value
507 	 * to work.
508 	 */
509 	if (pps > GPT_PPS_MAX) {
510 		ERROR("GPT: Invalid PPS: 0x%x\n", pps);
511 		return -EINVAL;
512 	}
513 	gpt_config.pps = pps;
514 	gpt_config.t = gpt_t_lookup[pps];
515 
516 	/* Alignment must be the greater of 4KB or L0 table size */
517 	l0_alignment = SZ_4K;
518 	if (l0_alignment < GPT_L0_TABLE_SIZE(gpt_config.t)) {
519 		l0_alignment = GPT_L0_TABLE_SIZE(gpt_config.t);
520 	}
521 
522 	/* Check base address */
523 	if ((l0_mem_base == 0UL) ||
524 	   ((l0_mem_base & (l0_alignment - 1UL)) != 0UL)) {
525 		ERROR("GPT: Invalid L0 base address: 0x%lx\n", l0_mem_base);
526 		return -EFAULT;
527 	}
528 
529 	/* Check memory size for L0 table */
530 	if (l0_mem_size < GPT_L0_TABLE_SIZE(gpt_config.t)) {
531 		ERROR("GPT: Inadequate L0 memory\n");
532 		ERROR("      Expected 0x%lx bytes, got 0x%lx\n",
533 				GPT_L0_TABLE_SIZE(gpt_config.t), l0_mem_size);
534 		return -ENOMEM;
535 	}
536 
537 	return 0;
538 }
539 
540 /*
541  * In the event that L1 tables are needed, this function validates
542  * the L1 table generation parameters.
543  *
544  * Parameters
545  *   l1_mem_base	Base address of memory used for L1 table allocation.
546  *   l1_mem_size	Total size of memory available for L1 tables.
547  *   l1_gpt_cnt		Number of L1 tables needed.
548  *
549  * Return
550  *   Negative Linux error code in the event of a failure, 0 for success.
551  */
552 static int validate_l1_params(uintptr_t l1_mem_base, size_t l1_mem_size,
553 				unsigned int l1_gpt_cnt)
554 {
555 	size_t l1_gpt_mem_sz;
556 
557 	/* Check if the granularity is supported */
558 	if (!xlat_arch_is_granule_size_supported(
559 	    GPT_PGS_ACTUAL_SIZE(gpt_config.p))) {
560 		return -EPERM;
561 	}
562 
563 	/* Make sure L1 tables are aligned to their size */
564 	if ((l1_mem_base & (GPT_L1_TABLE_SIZE(gpt_config.p) - 1UL)) != 0UL) {
565 		ERROR("GPT: Unaligned L1 GPT base address: 0x%"PRIxPTR"\n",
566 		      l1_mem_base);
567 		return -EFAULT;
568 	}
569 
570 	/* Get total memory needed for L1 tables */
571 	l1_gpt_mem_sz = l1_gpt_cnt * GPT_L1_TABLE_SIZE(gpt_config.p);
572 
573 	/* Check for overflow */
574 	if ((l1_gpt_mem_sz / GPT_L1_TABLE_SIZE(gpt_config.p)) != l1_gpt_cnt) {
575 		ERROR("GPT: Overflow calculating L1 memory size\n");
576 		return -ENOMEM;
577 	}
578 
579 	/* Make sure enough space was supplied */
580 	if (l1_mem_size < l1_gpt_mem_sz) {
581 		ERROR("%sL1 GPTs%s", (const char *)"GPT: Inadequate ",
582 			(const char *)" memory\n");
583 		ERROR("      Expected 0x%lx bytes, got 0x%lx\n",
584 			l1_gpt_mem_sz, l1_mem_size);
585 		return -ENOMEM;
586 	}
587 
588 	VERBOSE("GPT: Requested 0x%lx bytes for L1 GPTs\n", l1_gpt_mem_sz);
589 	return 0;
590 }
591 
592 /*
593  * This function initializes L0 block descriptors (regions that cannot be
594  * transitioned at the granule level) according to the provided PAS.
595  *
596  * Parameters
597  *   *pas		Pointer to the structure defining the PAS region to
598  *			initialize.
599  */
600 static void generate_l0_blk_desc(pas_region_t *pas)
601 {
602 	uint64_t gpt_desc;
603 	unsigned long idx, end_idx;
604 	uint64_t *l0_gpt_arr;
605 
606 	assert(gpt_config.plat_gpt_l0_base != 0UL);
607 	assert(pas != NULL);
608 
609 	/*
610 	 * Checking of PAS parameters has already been done in
611 	 * validate_pas_mappings so no need to check the same things again.
612 	 */
613 
614 	l0_gpt_arr = (uint64_t *)gpt_config.plat_gpt_l0_base;
615 
616 	/* Create the GPT Block descriptor for this PAS region */
617 	gpt_desc = GPT_L0_BLK_DESC(GPT_PAS_ATTR_GPI(pas->attrs));
618 
619 	/* Start index of this region in L0 GPTs */
620 	idx = GPT_L0_IDX(pas->base_pa);
621 
622 	/*
623 	 * Determine number of L0 GPT descriptors covered by
624 	 * this PAS region and use the count to populate these
625 	 * descriptors.
626 	 */
627 	end_idx = GPT_L0_IDX(pas->base_pa + pas->size);
628 
629 	/* Generate the needed block descriptors */
630 	for (; idx < end_idx; idx++) {
631 		l0_gpt_arr[idx] = gpt_desc;
632 		VERBOSE("GPT: L0 entry (BLOCK) index %lu [%p]: GPI = 0x%"PRIx64" (0x%"PRIx64")\n",
633 			idx, &l0_gpt_arr[idx],
634 			(gpt_desc >> GPT_L0_BLK_DESC_GPI_SHIFT) &
635 			GPT_L0_BLK_DESC_GPI_MASK, l0_gpt_arr[idx]);
636 	}
637 }
638 
639 /*
640  * Helper function to determine if the end physical address lies in the same L0
641  * region as the current physical address. If true, the end physical address is
642  * returned else, the start address of the next region is returned.
643  *
644  * Parameters
645  *   cur_pa		Physical address of the current PA in the loop through
646  *			the range.
647  *   end_pa		Physical address of the end PA in a PAS range.
648  *
649  * Return
650  *   The PA of the end of the current range.
651  */
652 static uintptr_t get_l1_end_pa(uintptr_t cur_pa, uintptr_t end_pa)
653 {
654 	uintptr_t cur_idx;
655 	uintptr_t end_idx;
656 
657 	cur_idx = GPT_L0_IDX(cur_pa);
658 	end_idx = GPT_L0_IDX(end_pa);
659 
660 	assert(cur_idx <= end_idx);
661 
662 	if (cur_idx == end_idx) {
663 		return end_pa;
664 	}
665 
666 	return (cur_idx + 1UL) << GPT_L0_IDX_SHIFT;
667 }
668 
669 /*
670  * Helper function to fill out GPI entries from 'first' granule address of
671  * the specified 'length' in a single L1 table with 'l1_desc' Contiguous
672  * descriptor.
673  *
674  * Parameters
675  *   l1			Pointer to L1 table to fill out
676  *   first		Address of first granule in range
677  *   length		Length of the range in bytes
678  *   gpi		GPI set this range to
679  *
680  * Return
681  *   Address of next granule in range.
682  */
683 __unused static uintptr_t fill_l1_cont_desc(uint64_t *l1, uintptr_t first,
684 					    size_t length, unsigned int gpi)
685 {
686 	/*
687 	 * Look up table for contiguous blocks and descriptors.
688 	 * Entries should be defined in descending block sizes:
689 	 * 512MB, 32MB and 2MB.
690 	 */
691 	static const gpt_fill_lookup_t gpt_fill_lookup[] = {
692 #if (RME_GPT_MAX_BLOCK == 512)
693 		{ SZ_512M, GPT_L1_CONT_DESC_512MB },
694 #endif
695 #if (RME_GPT_MAX_BLOCK >= 32)
696 		{ SZ_32M, GPT_L1_CONT_DESC_32MB },
697 #endif
698 #if (RME_GPT_MAX_BLOCK != 0)
699 		{ SZ_2M, GPT_L1_CONT_DESC_2MB }
700 #endif
701 	};
702 
703 	/*
704 	 * Iterate through all block sizes (512MB, 32MB and 2MB)
705 	 * starting with maximum supported.
706 	 */
707 	for (unsigned long i = 0UL; i < ARRAY_SIZE(gpt_fill_lookup); i++) {
708 		/* Calculate index */
709 		unsigned long idx = GPT_L1_INDEX(first);
710 
711 		/* Contiguous block size */
712 		size_t cont_size = gpt_fill_lookup[i].size;
713 
714 		if (GPT_REGION_IS_CONT(length, first, cont_size)) {
715 
716 			/* Generate Contiguous descriptor */
717 			uint64_t l1_desc = GPT_L1_GPI_CONT_DESC(gpi,
718 						gpt_fill_lookup[i].desc);
719 
720 			/* Number of 128-bit L1 entries in block */
721 			unsigned int cnt;
722 
723 			switch (cont_size) {
724 			case SZ_512M:
725 				cnt = L1_QWORDS_512MB;
726 				break;
727 			case SZ_32M:
728 				cnt = L1_QWORDS_32MB;
729 				break;
730 			default:			/* SZ_2MB */
731 				cnt = L1_QWORDS_2MB;
732 			}
733 
734 			VERBOSE("GPT: Contiguous descriptor 0x%"PRIxPTR" %luMB\n",
735 				first, cont_size / SZ_1M);
736 
737 			/* Fill Contiguous descriptors */
738 			fill_desc(&l1[idx], l1_desc, cnt);
739 			return (first + cont_size);
740 		}
741 	}
742 
743 	return first;
744 }
745 
746 /* Build Granules descriptor with the same 'gpi' for every GPI entry */
747 static uint64_t build_l1_desc(unsigned int gpi)
748 {
749 	uint64_t l1_desc = (uint64_t)gpi | ((uint64_t)gpi << 4);
750 
751 	l1_desc |= (l1_desc << 8);
752 	l1_desc |= (l1_desc << 16);
753 	return (l1_desc | (l1_desc << 32));
754 }
755 
756 /*
757  * Helper function to fill out GPI entries from 'first' to 'last' granule
758  * address in a single L1 table with 'l1_desc' Granules descriptor.
759  *
760  * Parameters
761  *   l1			Pointer to L1 table to fill out
762  *   first		Address of first granule in range
763  *   last		Address of last granule in range (inclusive)
764  *   gpi		GPI set this range to
765  *
766  * Return
767  *   Address of next granule in range.
768  */
769 static uintptr_t fill_l1_gran_desc(uint64_t *l1, uintptr_t first,
770 				   uintptr_t last, unsigned int gpi)
771 {
772 	uint64_t gpi_mask;
773 	unsigned long i;
774 
775 	/* Generate Granules descriptor */
776 	uint64_t l1_desc = build_l1_desc(gpi);
777 
778 	/* Shift the mask if we're starting in the middle of an L1 entry */
779 	gpi_mask = ULONG_MAX << (GPT_L1_GPI_IDX(gpt_config.p, first) << 2);
780 
781 	/* Fill out each L1 entry for this region */
782 	for (i = GPT_L1_INDEX(first); i <= GPT_L1_INDEX(last); i++) {
783 
784 		/* Account for stopping in the middle of an L1 entry */
785 		if (i == GPT_L1_INDEX(last)) {
786 			gpi_mask &= (gpi_mask >> ((15U -
787 				    GPT_L1_GPI_IDX(gpt_config.p, last)) << 2));
788 		}
789 
790 		assert((l1[i] & gpi_mask) == (GPT_L1_ANY_DESC & gpi_mask));
791 
792 		/* Write GPI values */
793 		l1[i] = (l1[i] & ~gpi_mask) | (l1_desc & gpi_mask);
794 
795 		/* Reset mask */
796 		gpi_mask = ULONG_MAX;
797 	}
798 
799 	return last + GPT_PGS_ACTUAL_SIZE(gpt_config.p);
800 }
801 
802 /*
803  * Helper function to fill out GPI entries in a single L1 table.
804  * This function fills out an entire L1 table with either Granules or Contiguous
805  * (RME_GPT_MAX_BLOCK != 0) descriptors depending on region length and alignment.
806  * Note. If RME_GPT_MAX_BLOCK == 0, then the L1 tables are filled with regular
807  * Granules descriptors.
808  *
809  * Parameters
810  *   l1			Pointer to L1 table to fill out
811  *   first		Address of first granule in range
812  *   last		Address of last granule in range (inclusive)
813  *   gpi		GPI set this range to
814  */
815 static void fill_l1_tbl(uint64_t *l1, uintptr_t first, uintptr_t last,
816 			unsigned int gpi)
817 {
818 	assert(l1 != NULL);
819 	assert(first <= last);
820 	assert((first & (GPT_PGS_ACTUAL_SIZE(gpt_config.p) - 1UL)) == 0UL);
821 	assert((last & (GPT_PGS_ACTUAL_SIZE(gpt_config.p) - 1UL)) == 0UL);
822 	assert(GPT_L0_IDX(first) == GPT_L0_IDX(last));
823 
824 #if (RME_GPT_MAX_BLOCK != 0)
825 	while (first <= last) {
826 		/* Region length */
827 		size_t length = last - first + GPT_PGS_ACTUAL_SIZE(gpt_config.p);
828 
829 		if (length < SZ_2M) {
830 			/*
831 			 * Fill with Granule descriptors in case of
832 			 * region length < 2MB.
833 			 */
834 			first = fill_l1_gran_desc(l1, first, last, gpi);
835 
836 		} else if ((first & (SZ_2M - UL(1))) == UL(0)) {
837 			/*
838 			 * For region length >= 2MB and at least 2MB aligned
839 			 * call to fill_l1_cont_desc will iterate through
840 			 * all block sizes (512MB, 32MB and 2MB) supported and
841 			 * fill corresponding Contiguous descriptors.
842 			 */
843 			first = fill_l1_cont_desc(l1, first, length, gpi);
844 		} else {
845 			/*
846 			 * For not aligned region >= 2MB fill with Granules
847 			 * descriptors up to the next 2MB aligned address.
848 			 */
849 			uintptr_t new_last = ALIGN_2MB(first + SZ_2M) -
850 					GPT_PGS_ACTUAL_SIZE(gpt_config.p);
851 
852 			first = fill_l1_gran_desc(l1, first, new_last, gpi);
853 		}
854 	}
855 #else
856 	/* Fill with Granule descriptors */
857 	first = fill_l1_gran_desc(l1, first, last, gpi);
858 #endif
859 	assert(first == (last + GPT_PGS_ACTUAL_SIZE(gpt_config.p)));
860 }
861 
862 /*
863  * This function finds the next available unused L1 table and initializes all
864  * granules descriptor entries to GPI_ANY. This ensures that there are no chunks
865  * of GPI_NO_ACCESS (0b0000) memory floating around in the system in the
866  * event that a PAS region stops midway through an L1 table, thus guaranteeing
867  * that all memory not explicitly assigned is GPI_ANY. This function does not
868  * check for overflow conditions, that should be done by the caller.
869  *
870  * Return
871  *   Pointer to the next available L1 table.
872  */
873 static uint64_t *get_new_l1_tbl(void)
874 {
875 	/* Retrieve the next L1 table */
876 	uint64_t *l1 = (uint64_t *)gpt_l1_tbl;
877 
878 	/* Increment L1 GPT address */
879 	gpt_l1_tbl += GPT_L1_TABLE_SIZE(gpt_config.p);
880 
881 	/* Initialize all GPIs to GPT_GPI_ANY */
882 	for (unsigned int i = 0U; i < GPT_L1_ENTRY_COUNT(gpt_config.p); i++) {
883 		l1[i] = GPT_L1_ANY_DESC;
884 	}
885 
886 	return l1;
887 }
888 
889 /*
890  * When L1 tables are needed, this function creates the necessary L0 table
891  * descriptors and fills out the L1 table entries according to the supplied
892  * PAS range.
893  *
894  * Parameters
895  *   *pas		Pointer to the structure defining the PAS region.
896  */
897 static void generate_l0_tbl_desc(pas_region_t *pas)
898 {
899 	uintptr_t end_pa;
900 	uintptr_t cur_pa;
901 	uintptr_t last_gran_pa;
902 	uint64_t *l0_gpt_base;
903 	uint64_t *l1_gpt_arr;
904 	unsigned int l0_idx, gpi;
905 
906 	assert(gpt_config.plat_gpt_l0_base != 0UL);
907 	assert(pas != NULL);
908 
909 	/*
910 	 * Checking of PAS parameters has already been done in
911 	 * validate_pas_mappings so no need to check the same things again.
912 	 */
913 	end_pa = pas->base_pa + pas->size;
914 	l0_gpt_base = (uint64_t *)gpt_config.plat_gpt_l0_base;
915 
916 	/* We start working from the granule at base PA */
917 	cur_pa = pas->base_pa;
918 
919 	/* Get GPI */
920 	gpi = GPT_PAS_ATTR_GPI(pas->attrs);
921 
922 	/* Iterate over each L0 region in this memory range */
923 	for (l0_idx = (unsigned int)GPT_L0_IDX(pas->base_pa);
924 	     l0_idx <= (unsigned int)GPT_L0_IDX(end_pa - 1UL);
925 	     l0_idx++) {
926 		/*
927 		 * See if the L0 entry is already a table descriptor or if we
928 		 * need to create one.
929 		 */
930 		if (GPT_L0_TYPE(l0_gpt_base[l0_idx]) == GPT_L0_TYPE_TBL_DESC) {
931 			/* Get the L1 array from the L0 entry */
932 			l1_gpt_arr = GPT_L0_TBLD_ADDR(l0_gpt_base[l0_idx]);
933 		} else {
934 			/* Get a new L1 table from the L1 memory space */
935 			l1_gpt_arr = get_new_l1_tbl();
936 
937 			/* Fill out the L0 descriptor and flush it */
938 			l0_gpt_base[l0_idx] = GPT_L0_TBL_DESC(l1_gpt_arr);
939 		}
940 
941 		VERBOSE("GPT: L0 entry (TABLE) index %u [%p] ==> L1 Addr %p (0x%"PRIx64")\n",
942 			l0_idx, &l0_gpt_base[l0_idx], l1_gpt_arr, l0_gpt_base[l0_idx]);
943 
944 		/*
945 		 * Determine the PA of the last granule in this L0 descriptor.
946 		 */
947 		last_gran_pa = get_l1_end_pa(cur_pa, end_pa) -
948 			       GPT_PGS_ACTUAL_SIZE(gpt_config.p);
949 
950 		/*
951 		 * Fill up L1 GPT entries between these two addresses. This
952 		 * function needs the addresses of the first granule and last
953 		 * granule in the range.
954 		 */
955 		fill_l1_tbl(l1_gpt_arr, cur_pa, last_gran_pa, gpi);
956 
957 		/* Advance cur_pa to first granule in next L0 region */
958 		cur_pa = get_l1_end_pa(cur_pa, end_pa);
959 	}
960 }
961 
962 /*
963  * This function flushes a range of L0 descriptors used by a given PAS region
964  * array. There is a chance that some unmodified L0 descriptors would be flushed
965  * in the case that there are "holes" in an array of PAS regions but overall
966  * this should be faster than individually flushing each modified L0 descriptor
967  * as they are created.
968  *
969  * Parameters
970  *   *pas		Pointer to an array of PAS regions.
971  *   pas_count		Number of entries in the PAS array.
972  */
973 static void flush_l0_for_pas_array(pas_region_t *pas, unsigned int pas_count)
974 {
975 	unsigned long idx;
976 	unsigned long start_idx;
977 	unsigned long end_idx;
978 	uint64_t *l0 = (uint64_t *)gpt_config.plat_gpt_l0_base;
979 
980 	assert(pas != NULL);
981 	assert(pas_count != 0U);
982 
983 	/* Initial start and end values */
984 	start_idx = GPT_L0_IDX(pas[0].base_pa);
985 	end_idx = GPT_L0_IDX(pas[0].base_pa + pas[0].size - 1UL);
986 
987 	/* Find lowest and highest L0 indices used in this PAS array */
988 	for (idx = 1UL; idx < pas_count; idx++) {
989 		if (GPT_L0_IDX(pas[idx].base_pa) < start_idx) {
990 			start_idx = GPT_L0_IDX(pas[idx].base_pa);
991 		}
992 		if (GPT_L0_IDX(pas[idx].base_pa + pas[idx].size - 1UL) > end_idx) {
993 			end_idx = GPT_L0_IDX(pas[idx].base_pa + pas[idx].size - 1UL);
994 		}
995 	}
996 
997 	/*
998 	 * Flush all covered L0 descriptors, add 1 because we need to include
999 	 * the end index value.
1000 	 */
1001 	flush_dcache_range((uintptr_t)&l0[start_idx],
1002 			   ((end_idx + 1UL) - start_idx) * sizeof(uint64_t));
1003 }
1004 
1005 /*
1006  * Public API to enable granule protection checks once the tables have all been
1007  * initialized. This function is called at first initialization and then again
1008  * later during warm boots of CPU cores.
1009  *
1010  * Return
1011  *   Negative Linux error code in the event of a failure, 0 for success.
1012  */
1013 int gpt_enable(void)
1014 {
1015 	u_register_t gpccr_el3;
1016 
1017 	/*
1018 	 * Granule tables must be initialised before enabling
1019 	 * granule protection.
1020 	 */
1021 	if (gpt_config.plat_gpt_l0_base == 0UL) {
1022 		ERROR("GPT: Tables have not been initialized!\n");
1023 		return -EPERM;
1024 	}
1025 
1026 	/* Write the base address of the L0 tables into GPTBR */
1027 	write_gptbr_el3(((gpt_config.plat_gpt_l0_base >> GPTBR_BADDR_VAL_SHIFT)
1028 			>> GPTBR_BADDR_SHIFT) & GPTBR_BADDR_MASK);
1029 
1030 	/* GPCCR_EL3.PPS */
1031 	gpccr_el3 = SET_GPCCR_PPS(gpt_config.pps);
1032 
1033 	/* GPCCR_EL3.PGS */
1034 	gpccr_el3 |= SET_GPCCR_PGS(gpt_config.pgs);
1035 
1036 	/*
1037 	 * Since EL3 maps the L1 region as Inner shareable, use the same
1038 	 * shareability attribute for GPC as well so that
1039 	 * GPC fetches are visible to PEs
1040 	 */
1041 	gpccr_el3 |= SET_GPCCR_SH(GPCCR_SH_IS);
1042 
1043 	/* Outer and Inner cacheability set to Normal memory, WB, RA, WA */
1044 	gpccr_el3 |= SET_GPCCR_ORGN(GPCCR_ORGN_WB_RA_WA);
1045 	gpccr_el3 |= SET_GPCCR_IRGN(GPCCR_IRGN_WB_RA_WA);
1046 
1047 	/* Prepopulate GPCCR_EL3 but don't enable GPC yet */
1048 	write_gpccr_el3(gpccr_el3);
1049 	isb();
1050 
1051 	/* Invalidate any stale TLB entries and any cached register fields */
1052 	tlbipaallos();
1053 	dsb();
1054 	isb();
1055 
1056 	/* Enable GPT */
1057 	gpccr_el3 |= GPCCR_GPC_BIT;
1058 
1059 	/* TODO: Configure GPCCR_EL3_GPCP for Fault control */
1060 	write_gpccr_el3(gpccr_el3);
1061 	isb();
1062 	tlbipaallos();
1063 	dsb();
1064 	isb();
1065 
1066 	return 0;
1067 }
1068 
1069 /*
1070  * Public API to disable granule protection checks.
1071  */
1072 void gpt_disable(void)
1073 {
1074 	u_register_t gpccr_el3 = read_gpccr_el3();
1075 
1076 	write_gpccr_el3(gpccr_el3 & ~GPCCR_GPC_BIT);
1077 	dsbsy();
1078 	isb();
1079 }
1080 
1081 /*
1082  * Public API that initializes the entire protected space to GPT_GPI_ANY using
1083  * the L0 tables (block descriptors). Ideally, this function is invoked prior
1084  * to DDR discovery and initialization. The MMU must be initialized before
1085  * calling this function.
1086  *
1087  * Parameters
1088  *   pps		PPS value to use for table generation
1089  *   l0_mem_base	Base address of L0 tables in memory.
1090  *   l0_mem_size	Total size of memory available for L0 tables.
1091  *
1092  * Return
1093  *   Negative Linux error code in the event of a failure, 0 for success.
1094  */
1095 int gpt_init_l0_tables(gpccr_pps_e pps, uintptr_t l0_mem_base,
1096 		       size_t l0_mem_size)
1097 {
1098 	uint64_t gpt_desc;
1099 	int ret;
1100 
1101 	/* Ensure that MMU and Data caches are enabled */
1102 	assert((read_sctlr_el3() & SCTLR_C_BIT) != 0UL);
1103 
1104 	/* Validate other parameters */
1105 	ret = validate_l0_params(pps, l0_mem_base, l0_mem_size);
1106 	if (ret != 0) {
1107 		return ret;
1108 	}
1109 
1110 	/* Create the descriptor to initialize L0 entries with */
1111 	gpt_desc = GPT_L0_BLK_DESC(GPT_GPI_ANY);
1112 
1113 	/* Iterate through all L0 entries */
1114 	for (unsigned int i = 0U; i < GPT_L0_REGION_COUNT(gpt_config.t); i++) {
1115 		((uint64_t *)l0_mem_base)[i] = gpt_desc;
1116 	}
1117 
1118 	/* Flush updated L0 table to memory */
1119 	flush_dcache_range((uintptr_t)l0_mem_base, GPT_L0_TABLE_SIZE(gpt_config.t));
1120 
1121 	/* Stash the L0 base address once initial setup is complete */
1122 	gpt_config.plat_gpt_l0_base = l0_mem_base;
1123 
1124 	return 0;
1125 }
1126 
1127 /*
1128  * Public API that carves out PAS regions from the L0 tables and builds any L1
1129  * tables that are needed. This function ideally is run after DDR discovery and
1130  * initialization. The L0 tables must have already been initialized to GPI_ANY
1131  * when this function is called.
1132  *
1133  * This function can be called multiple times with different L1 memory ranges
1134  * and PAS regions if it is desirable to place L1 tables in different locations
1135  * in memory. (ex: you have multiple DDR banks and want to place the L1 tables
1136  * in the DDR bank that they control).
1137  *
1138  * Parameters
1139  *   pgs		PGS value to use for table generation.
1140  *   l1_mem_base	Base address of memory used for L1 tables.
1141  *   l1_mem_size	Total size of memory available for L1 tables.
1142  *   *pas_regions	Pointer to PAS regions structure array.
1143  *   pas_count		Total number of PAS regions.
1144  *
1145  * Return
1146  *   Negative Linux error code in the event of a failure, 0 for success.
1147  */
1148 int gpt_init_pas_l1_tables(gpccr_pgs_e pgs, uintptr_t l1_mem_base,
1149 			   size_t l1_mem_size, pas_region_t *pas_regions,
1150 			   unsigned int pas_count)
1151 {
1152 	int l1_gpt_cnt, ret;
1153 
1154 	/* Ensure that MMU and Data caches are enabled */
1155 	assert((read_sctlr_el3() & SCTLR_C_BIT) != 0UL);
1156 
1157 	/* PGS is needed for validate_pas_mappings so check it now */
1158 	if (pgs > GPT_PGS_MAX) {
1159 		ERROR("GPT: Invalid PGS: 0x%x\n", pgs);
1160 		return -EINVAL;
1161 	}
1162 	gpt_config.pgs = pgs;
1163 	gpt_config.p = gpt_p_lookup[pgs];
1164 
1165 	/* Make sure L0 tables have been initialized */
1166 	if (gpt_config.plat_gpt_l0_base == 0UL) {
1167 		ERROR("GPT: L0 tables must be initialized first!\n");
1168 		return -EPERM;
1169 	}
1170 
1171 	/* Check if L1 GPTs are required and how many */
1172 	l1_gpt_cnt = validate_pas_mappings(pas_regions, pas_count);
1173 	if (l1_gpt_cnt < 0) {
1174 		return l1_gpt_cnt;
1175 	}
1176 
1177 	VERBOSE("GPT: %i L1 GPTs requested\n", l1_gpt_cnt);
1178 
1179 	/* If L1 tables are needed then validate the L1 parameters */
1180 	if (l1_gpt_cnt > 0) {
1181 		ret = validate_l1_params(l1_mem_base, l1_mem_size,
1182 					(unsigned int)l1_gpt_cnt);
1183 		if (ret != 0) {
1184 			return ret;
1185 		}
1186 
1187 		/* Set up parameters for L1 table generation */
1188 		gpt_l1_tbl = l1_mem_base;
1189 	}
1190 
1191 	/* Number of L1 entries in 2MB depends on GPCCR_EL3.PGS value */
1192 	gpt_l1_cnt_2mb = (unsigned int)GPT_L1_ENTRY_COUNT_2MB(gpt_config.p);
1193 
1194 	/* Mask for the L1 index field */
1195 	gpt_l1_index_mask = GPT_L1_IDX_MASK(gpt_config.p);
1196 
1197 	INFO("GPT: Boot Configuration\n");
1198 	INFO("  PPS/T:     0x%x/%u\n", gpt_config.pps, gpt_config.t);
1199 	INFO("  PGS/P:     0x%x/%u\n", gpt_config.pgs, gpt_config.p);
1200 	INFO("  L0GPTSZ/S: 0x%x/%u\n", GPT_L0GPTSZ, GPT_S_VAL);
1201 	INFO("  PAS count: %u\n", pas_count);
1202 	INFO("  L0 base:   0x%"PRIxPTR"\n", gpt_config.plat_gpt_l0_base);
1203 
1204 	/* Generate the tables in memory */
1205 	for (unsigned int idx = 0U; idx < pas_count; idx++) {
1206 		VERBOSE("GPT: PAS[%u]: base 0x%"PRIxPTR"\tsize 0x%lx\tGPI 0x%x\ttype 0x%x\n",
1207 			idx, pas_regions[idx].base_pa, pas_regions[idx].size,
1208 			GPT_PAS_ATTR_GPI(pas_regions[idx].attrs),
1209 			GPT_PAS_ATTR_MAP_TYPE(pas_regions[idx].attrs));
1210 
1211 		/* Check if a block or table descriptor is required */
1212 		if (GPT_PAS_ATTR_MAP_TYPE(pas_regions[idx].attrs) ==
1213 		    GPT_PAS_ATTR_MAP_TYPE_BLOCK) {
1214 			generate_l0_blk_desc(&pas_regions[idx]);
1215 
1216 		} else {
1217 			generate_l0_tbl_desc(&pas_regions[idx]);
1218 		}
1219 	}
1220 
1221 	/* Flush modified L0 tables */
1222 	flush_l0_for_pas_array(pas_regions, pas_count);
1223 
1224 	/* Flush L1 tables if needed */
1225 	if (l1_gpt_cnt > 0) {
1226 		flush_dcache_range(l1_mem_base,
1227 				   GPT_L1_TABLE_SIZE(gpt_config.p) *
1228 				   (size_t)l1_gpt_cnt);
1229 	}
1230 
1231 	/* Make sure that all the entries are written to the memory */
1232 	dsbishst();
1233 	tlbipaallos();
1234 	dsb();
1235 	isb();
1236 
1237 	return 0;
1238 }
1239 
1240 /*
1241  * Public API to initialize the runtime gpt_config structure based on the values
1242  * present in the GPTBR_EL3 and GPCCR_EL3 registers. GPT initialization
1243  * typically happens in a bootloader stage prior to setting up the EL3 runtime
1244  * environment for the granule transition service so this function detects the
1245  * initialization from a previous stage. Granule protection checks must be
1246  * enabled already or this function will return an error.
1247  *
1248  * Parameters
1249  *   l1_bitlocks_base	Base address of memory for L1 tables bitlocks.
1250  *   l1_bitlocks_size	Total size of memory available for L1 tables bitlocks.
1251  *
1252  * Return
1253  *   Negative Linux error code in the event of a failure, 0 for success.
1254  */
1255 int gpt_runtime_init(uintptr_t l1_bitlocks_base, size_t l1_bitlocks_size)
1256 {
1257 	u_register_t reg;
1258 	__unused size_t locks_size;
1259 
1260 	/* Ensure that MMU and Data caches are enabled */
1261 	assert((read_sctlr_el3() & SCTLR_C_BIT) != 0UL);
1262 
1263 	/* Ensure GPC are already enabled */
1264 	if ((read_gpccr_el3() & GPCCR_GPC_BIT) == 0UL) {
1265 		ERROR("GPT: Granule protection checks are not enabled!\n");
1266 		return -EPERM;
1267 	}
1268 
1269 	/*
1270 	 * Read the L0 table address from GPTBR, we don't need the L1 base
1271 	 * address since those are included in the L0 tables as needed.
1272 	 */
1273 	reg = read_gptbr_el3();
1274 	gpt_config.plat_gpt_l0_base = ((reg >> GPTBR_BADDR_SHIFT) &
1275 				      GPTBR_BADDR_MASK) <<
1276 				      GPTBR_BADDR_VAL_SHIFT;
1277 
1278 	/* Read GPCCR to get PGS and PPS values */
1279 	reg = read_gpccr_el3();
1280 	gpt_config.pps = (reg >> GPCCR_PPS_SHIFT) & GPCCR_PPS_MASK;
1281 	gpt_config.t = gpt_t_lookup[gpt_config.pps];
1282 	gpt_config.pgs = (reg >> GPCCR_PGS_SHIFT) & GPCCR_PGS_MASK;
1283 	gpt_config.p = gpt_p_lookup[gpt_config.pgs];
1284 
1285 	/* Number of L1 entries in 2MB depends on GPCCR_EL3.PGS value */
1286 	gpt_l1_cnt_2mb = (unsigned int)GPT_L1_ENTRY_COUNT_2MB(gpt_config.p);
1287 
1288 	/* Mask for the L1 index field */
1289 	gpt_l1_index_mask = GPT_L1_IDX_MASK(gpt_config.p);
1290 
1291 #if (RME_GPT_BITLOCK_BLOCK != 0)
1292 	/*
1293 	 * Size of GPT bitlocks in bytes for the protected address space
1294 	 * with RME_GPT_BITLOCK_BLOCK * 512MB per bitlock.
1295 	 */
1296 	locks_size = GPT_PPS_ACTUAL_SIZE(gpt_config.t) /
1297 			(RME_GPT_BITLOCK_BLOCK * SZ_512M * 8U);
1298 	/*
1299 	 * If protected space size is less than the size covered
1300 	 * by 'bitlock' structure, check for a single bitlock.
1301 	 */
1302 	if (locks_size < LOCK_SIZE) {
1303 		locks_size = LOCK_SIZE;
1304 	/* Check bitlocks array size */
1305 	} else if (locks_size > l1_bitlocks_size) {
1306 		ERROR("GPT: Inadequate GPT bitlocks memory\n");
1307 		ERROR("      Expected 0x%lx bytes, got 0x%lx\n",
1308 			locks_size, l1_bitlocks_size);
1309 		return -ENOMEM;
1310 	}
1311 
1312 	gpt_bitlock = (bitlock_t *)l1_bitlocks_base;
1313 
1314 	/* Initialise GPT bitlocks */
1315 	(void)memset((void *)gpt_bitlock, 0, locks_size);
1316 
1317 	/* Flush GPT bitlocks to memory */
1318 	flush_dcache_range((uintptr_t)gpt_bitlock, locks_size);
1319 #endif /* RME_GPT_BITLOCK_BLOCK */
1320 
1321 	VERBOSE("GPT: Runtime Configuration\n");
1322 	VERBOSE("  PPS/T:     0x%x/%u\n", gpt_config.pps, gpt_config.t);
1323 	VERBOSE("  PGS/P:     0x%x/%u\n", gpt_config.pgs, gpt_config.p);
1324 	VERBOSE("  L0GPTSZ/S: 0x%x/%u\n", GPT_L0GPTSZ, GPT_S_VAL);
1325 	VERBOSE("  L0 base:   0x%"PRIxPTR"\n", gpt_config.plat_gpt_l0_base);
1326 #if (RME_GPT_BITLOCK_BLOCK != 0)
1327 	VERBOSE("  Bitlocks:  0x%"PRIxPTR"/0x%lx\n", (uintptr_t)gpt_bitlock,
1328 					locks_size);
1329 #endif
1330 	return 0;
1331 }
1332 
1333 /*
1334  * A helper to write the value (target_pas << gpi_shift) to the index of
1335  * the gpt_l1_addr.
1336  */
1337 static inline void write_gpt(uint64_t *gpt_l1_desc, uint64_t *gpt_l1_addr,
1338 			     unsigned int gpi_shift, unsigned int idx,
1339 			     unsigned int target_pas)
1340 {
1341 	*gpt_l1_desc &= ~(GPT_L1_GRAN_DESC_GPI_MASK << gpi_shift);
1342 	*gpt_l1_desc |= ((uint64_t)target_pas << gpi_shift);
1343 	gpt_l1_addr[idx] = *gpt_l1_desc;
1344 
1345 	dsboshst();
1346 }
1347 
1348 /*
1349  * Helper to retrieve the gpt_l1_* information from the base address
1350  * returned in gpi_info.
1351  */
1352 static int get_gpi_params(uint64_t base, gpi_info_t *gpi_info)
1353 {
1354 	uint64_t gpt_l0_desc, *gpt_l0_base;
1355 	__unused unsigned int block_idx;
1356 
1357 	gpt_l0_base = (uint64_t *)gpt_config.plat_gpt_l0_base;
1358 	gpt_l0_desc = gpt_l0_base[GPT_L0_IDX(base)];
1359 	if (GPT_L0_TYPE(gpt_l0_desc) != GPT_L0_TYPE_TBL_DESC) {
1360 		VERBOSE("GPT: Granule is not covered by a table descriptor!\n");
1361 		VERBOSE("      Base=0x%"PRIx64"\n", base);
1362 		return -EINVAL;
1363 	}
1364 
1365 	/* Get the table index and GPI shift from PA */
1366 	gpi_info->gpt_l1_addr = GPT_L0_TBLD_ADDR(gpt_l0_desc);
1367 	gpi_info->idx = (unsigned int)GPT_L1_INDEX(base);
1368 	gpi_info->gpi_shift = GPT_L1_GPI_IDX(gpt_config.p, base) << 2;
1369 
1370 #if (RME_GPT_BITLOCK_BLOCK != 0)
1371 	/* Block index */
1372 	block_idx = (unsigned int)(base / (RME_GPT_BITLOCK_BLOCK * SZ_512M));
1373 
1374 	/* Bitlock address and mask */
1375 	gpi_info->lock = &gpt_bitlock[block_idx / LOCK_BITS];
1376 	gpi_info->mask = 1U << (block_idx & (LOCK_BITS - 1U));
1377 #endif
1378 	return 0;
1379 }
1380 
1381 /*
1382  * Helper to retrieve the gpt_l1_desc and GPI information from gpi_info.
1383  * This function is called with bitlock or spinlock acquired.
1384  */
1385 static void read_gpi(gpi_info_t *gpi_info)
1386 {
1387 	gpi_info->gpt_l1_desc = (gpi_info->gpt_l1_addr)[gpi_info->idx];
1388 
1389 	if ((gpi_info->gpt_l1_desc & GPT_L1_TYPE_CONT_DESC_MASK) ==
1390 				 GPT_L1_TYPE_CONT_DESC) {
1391 		/* Read GPI from Contiguous descriptor */
1392 		gpi_info->gpi = (unsigned int)GPT_L1_CONT_GPI(gpi_info->gpt_l1_desc);
1393 	} else {
1394 		/* Read GPI from Granules descriptor */
1395 		gpi_info->gpi = (unsigned int)((gpi_info->gpt_l1_desc >> gpi_info->gpi_shift) &
1396 						GPT_L1_GRAN_DESC_GPI_MASK);
1397 	}
1398 }
1399 
1400 static void flush_page_to_popa(uintptr_t addr)
1401 {
1402 	size_t size = GPT_PGS_ACTUAL_SIZE(gpt_config.p);
1403 
1404 	if (is_feat_mte2_supported()) {
1405 		flush_dcache_to_popa_range_mte2(addr, size);
1406 	} else {
1407 		flush_dcache_to_popa_range(addr, size);
1408 	}
1409 }
1410 
1411 /*
1412  * Helper function to check if all L1 entries in 2MB block have
1413  * the same Granules descriptor value.
1414  *
1415  * Parameters
1416  *   base		Base address of the region to be checked
1417  *   gpi_info		Pointer to 'gpt_config_t' structure
1418  *   l1_desc		GPT Granules descriptor with all entries
1419  *			set to the same GPI.
1420  *
1421  * Return
1422  *   true if L1 all entries have the same descriptor value, false otherwise.
1423  */
1424 __unused static bool check_fuse_2mb(uint64_t base, const gpi_info_t *gpi_info,
1425 					uint64_t l1_desc)
1426 {
1427 	/* Last L1 entry index in 2MB block */
1428 	unsigned int long idx = GPT_L1_INDEX(ALIGN_2MB(base)) +
1429 						gpt_l1_cnt_2mb - 1UL;
1430 
1431 	/* Number of L1 entries in 2MB block */
1432 	unsigned int cnt = gpt_l1_cnt_2mb;
1433 
1434 	/*
1435 	 * Start check from the last L1 entry and continue until the first
1436 	 * non-matching to the passed Granules descriptor value is found.
1437 	 */
1438 	while (cnt-- != 0U) {
1439 		if (gpi_info->gpt_l1_addr[idx--] != l1_desc) {
1440 			/* Non-matching L1 entry found */
1441 			return false;
1442 		}
1443 	}
1444 
1445 	return true;
1446 }
1447 
1448 __unused static void fuse_2mb(uint64_t base, const gpi_info_t *gpi_info,
1449 				uint64_t l1_desc)
1450 {
1451 	/* L1 entry index of the start of 2MB block */
1452 	unsigned long idx_2 = GPT_L1_INDEX(ALIGN_2MB(base));
1453 
1454 	/* 2MB Contiguous descriptor */
1455 	uint64_t l1_cont_desc = GPT_L1_CONT_DESC(l1_desc, 2MB);
1456 
1457 	VERBOSE("GPT: %s(0x%"PRIxPTR" 0x%"PRIx64")\n", __func__, base, l1_desc);
1458 
1459 	fill_desc(&gpi_info->gpt_l1_addr[idx_2], l1_cont_desc, L1_QWORDS_2MB);
1460 }
1461 
1462 /*
1463  * Helper function to check if all 1st L1 entries of 2MB blocks
1464  * in 32MB have the same 2MB Contiguous descriptor value.
1465  *
1466  * Parameters
1467  *   base		Base address of the region to be checked
1468  *   gpi_info		Pointer to 'gpt_config_t' structure
1469  *   l1_desc		GPT Granules descriptor.
1470  *
1471  * Return
1472  *   true if all L1 entries have the same descriptor value, false otherwise.
1473  */
1474 __unused static bool check_fuse_32mb(uint64_t base, const gpi_info_t *gpi_info,
1475 					uint64_t l1_desc)
1476 {
1477 	/* The 1st L1 entry index of the last 2MB block in 32MB */
1478 	unsigned long idx = GPT_L1_INDEX(ALIGN_32MB(base)) +
1479 					(15UL * gpt_l1_cnt_2mb);
1480 
1481 	/* 2MB Contiguous descriptor */
1482 	uint64_t l1_cont_desc = GPT_L1_CONT_DESC(l1_desc, 2MB);
1483 
1484 	/* Number of 2MB blocks in 32MB */
1485 	unsigned int cnt = 16U;
1486 
1487 	/* Set the first L1 entry to 2MB Contiguous descriptor */
1488 	gpi_info->gpt_l1_addr[GPT_L1_INDEX(ALIGN_2MB(base))] = l1_cont_desc;
1489 
1490 	/*
1491 	 * Start check from the 1st L1 entry of the last 2MB block and
1492 	 * continue until the first non-matching to 2MB Contiguous descriptor
1493 	 * value is found.
1494 	 */
1495 	while (cnt-- != 0U) {
1496 		if (gpi_info->gpt_l1_addr[idx] != l1_cont_desc) {
1497 			/* Non-matching L1 entry found */
1498 			return false;
1499 		}
1500 		idx -= gpt_l1_cnt_2mb;
1501 	}
1502 
1503 	return true;
1504 }
1505 
1506 __unused static void fuse_32mb(uint64_t base, const gpi_info_t *gpi_info,
1507 				uint64_t l1_desc)
1508 {
1509 	/* L1 entry index of the start of 32MB block */
1510 	unsigned long idx_32 = GPT_L1_INDEX(ALIGN_32MB(base));
1511 
1512 	/* 32MB Contiguous descriptor */
1513 	uint64_t l1_cont_desc = GPT_L1_CONT_DESC(l1_desc, 32MB);
1514 
1515 	VERBOSE("GPT: %s(0x%"PRIxPTR" 0x%"PRIx64")\n", __func__, base, l1_desc);
1516 
1517 	fill_desc(&gpi_info->gpt_l1_addr[idx_32], l1_cont_desc, L1_QWORDS_32MB);
1518 }
1519 
1520 /*
1521  * Helper function to check if all 1st L1 entries of 32MB blocks
1522  * in 512MB have the same 32MB Contiguous descriptor value.
1523  *
1524  * Parameters
1525  *   base		Base address of the region to be checked
1526  *   gpi_info		Pointer to 'gpt_config_t' structure
1527  *   l1_desc		GPT Granules descriptor.
1528  *
1529  * Return
1530  *   true if all L1 entries have the same descriptor value, false otherwise.
1531  */
1532 __unused static bool check_fuse_512mb(uint64_t base, const gpi_info_t *gpi_info,
1533 					uint64_t l1_desc)
1534 {
1535 	/* The 1st L1 entry index of the last 32MB block in 512MB */
1536 	unsigned long idx = GPT_L1_INDEX(ALIGN_512MB(base)) +
1537 					(15UL * 16UL * gpt_l1_cnt_2mb);
1538 
1539 	/* 32MB Contiguous descriptor */
1540 	uint64_t l1_cont_desc = GPT_L1_CONT_DESC(l1_desc, 32MB);
1541 
1542 	/* Number of 32MB blocks in 512MB */
1543 	unsigned int cnt = 16U;
1544 
1545 	/* Set the first L1 entry to 2MB Contiguous descriptor */
1546 	gpi_info->gpt_l1_addr[GPT_L1_INDEX(ALIGN_32MB(base))] = l1_cont_desc;
1547 
1548 	/*
1549 	 * Start check from the 1st L1 entry of the last 32MB block and
1550 	 * continue until the first non-matching to 32MB Contiguous descriptor
1551 	 * value is found.
1552 	 */
1553 	while (cnt-- != 0U) {
1554 		if (gpi_info->gpt_l1_addr[idx] != l1_cont_desc) {
1555 			/* Non-matching L1 entry found */
1556 			return false;
1557 		}
1558 		idx -= 16UL * gpt_l1_cnt_2mb;
1559 	}
1560 
1561 	return true;
1562 }
1563 
1564 __unused static void fuse_512mb(uint64_t base, const gpi_info_t *gpi_info,
1565 				uint64_t l1_desc)
1566 {
1567 	/* L1 entry index of the start of 512MB block */
1568 	unsigned long idx_512 = GPT_L1_INDEX(ALIGN_512MB(base));
1569 
1570 	/* 512MB Contiguous descriptor */
1571 	uint64_t l1_cont_desc = GPT_L1_CONT_DESC(l1_desc, 512MB);
1572 
1573 	VERBOSE("GPT: %s(0x%"PRIxPTR" 0x%"PRIx64")\n", __func__, base, l1_desc);
1574 
1575 	fill_desc(&gpi_info->gpt_l1_addr[idx_512], l1_cont_desc, L1_QWORDS_512MB);
1576 }
1577 
1578 /*
1579  * Helper function to convert GPI entries in a single L1 table
1580  * from Granules to Contiguous descriptor.
1581  *
1582  * Parameters
1583  *   base		Base address of the region to be written
1584  *   gpi_info		Pointer to 'gpt_config_t' structure
1585  *   l1_desc		GPT Granules descriptor with all entries
1586  *			set to the same GPI.
1587  */
1588 __unused static void fuse_block(uint64_t base, const gpi_info_t *gpi_info,
1589 				uint64_t l1_desc)
1590 {
1591 	/* Start with check for 2MB block */
1592 	if (!check_fuse_2mb(base, gpi_info, l1_desc)) {
1593 		/* Check for 2MB fusing failed */
1594 		return;
1595 	}
1596 
1597 #if (RME_GPT_MAX_BLOCK == 2)
1598 	fuse_2mb(base, gpi_info, l1_desc);
1599 #else
1600 	/* Check for 32MB block */
1601 	if (!check_fuse_32mb(base, gpi_info, l1_desc)) {
1602 		/* Check for 32MB fusing failed, fuse to 2MB */
1603 		fuse_2mb(base, gpi_info, l1_desc);
1604 		return;
1605 	}
1606 
1607 #if (RME_GPT_MAX_BLOCK == 32)
1608 	fuse_32mb(base, gpi_info, l1_desc);
1609 #else
1610 	/* Check for 512MB block */
1611 	if (!check_fuse_512mb(base, gpi_info, l1_desc)) {
1612 		/* Check for 512MB fusing failed, fuse to 32MB */
1613 		fuse_32mb(base, gpi_info, l1_desc);
1614 		return;
1615 	}
1616 
1617 	/* Fuse to 512MB */
1618 	fuse_512mb(base, gpi_info, l1_desc);
1619 
1620 #endif	/* RME_GPT_MAX_BLOCK == 32 */
1621 #endif	/* RME_GPT_MAX_BLOCK == 2 */
1622 }
1623 
1624 /*
1625  * Helper function to convert GPI entries in a single L1 table
1626  * from Contiguous to Granules descriptor. This function updates
1627  * descriptor to Granules in passed 'gpt_config_t' structure as
1628  * the result of shuttering.
1629  *
1630  * Parameters
1631  *   base		Base address of the region to be written
1632  *   gpi_info		Pointer to 'gpt_config_t' structure
1633  *   l1_desc		GPT Granules descriptor set this range to.
1634  */
1635 __unused static void shatter_block(uint64_t base, gpi_info_t *gpi_info,
1636 				   uint64_t l1_desc)
1637 {
1638 	/* Look-up table for 2MB, 32MB and 512MB locks shattering */
1639 	static const gpt_shatter_func gpt_shatter_lookup[] = {
1640 		shatter_2mb,
1641 		shatter_32mb,
1642 		shatter_512mb
1643 	};
1644 
1645 	/* Look-up table for invalidation TLBs for 2MB, 32MB and 512MB blocks */
1646 	static const gpt_tlbi_lookup_t tlbi_lookup[] = {
1647 		{ tlbirpalos_2m, ~(SZ_2M - 1UL) },
1648 		{ tlbirpalos_32m, ~(SZ_32M - 1UL) },
1649 		{ tlbirpalos_512m, ~(SZ_512M - 1UL) }
1650 	};
1651 
1652 	/* Get shattering level from Contig field of Contiguous descriptor */
1653 	unsigned long level = GPT_L1_CONT_CONTIG(gpi_info->gpt_l1_desc) - 1UL;
1654 
1655 	/* Shatter contiguous block */
1656 	gpt_shatter_lookup[level](base, gpi_info, l1_desc);
1657 
1658 	tlbi_lookup[level].function(base & tlbi_lookup[level].mask);
1659 	dsbosh();
1660 
1661 	/*
1662 	 * Update 'gpt_config_t' structure's descriptor to Granules to reflect
1663 	 * the shattered GPI back to caller.
1664 	 */
1665 	gpi_info->gpt_l1_desc = l1_desc;
1666 }
1667 
1668 /*
1669  * This function is the granule transition delegate service. When a granule
1670  * transition request occurs it is routed to this function to have the request,
1671  * if valid, fulfilled following A1.1.1 Delegate of RME supplement.
1672  *
1673  * TODO: implement support for transitioning multiple granules at once.
1674  *
1675  * Parameters
1676  *   base		Base address of the region to transition, must be
1677  *			aligned to granule size.
1678  *   size		Size of region to transition, must be aligned to granule
1679  *			size.
1680  *   src_sec_state	Security state of the caller.
1681  *
1682  * Return
1683  *   Negative Linux error code in the event of a failure, 0 for success.
1684  */
1685 int gpt_delegate_pas(uint64_t base, size_t size, unsigned int src_sec_state)
1686 {
1687 	gpi_info_t gpi_info;
1688 	uint64_t nse, __unused l1_desc;
1689 	unsigned int target_pas;
1690 	int res;
1691 
1692 	/* Ensure that the tables have been set up before taking requests */
1693 	assert(gpt_config.plat_gpt_l0_base != 0UL);
1694 
1695 	/* Ensure that caches are enabled */
1696 	assert((read_sctlr_el3() & SCTLR_C_BIT) != 0UL);
1697 
1698 	/* See if this is a single or a range of granule transition */
1699 	if (size != GPT_PGS_ACTUAL_SIZE(gpt_config.p)) {
1700 		return -EINVAL;
1701 	}
1702 
1703 	/* Check that base and size are valid */
1704 	if ((ULONG_MAX - base) < size) {
1705 		VERBOSE("GPT: Transition request address overflow!\n");
1706 		VERBOSE("      Base=0x%"PRIx64"\n", base);
1707 		VERBOSE("      Size=0x%lx\n", size);
1708 		return -EINVAL;
1709 	}
1710 
1711 	/* Make sure base and size are valid */
1712 	if (((base & (GPT_PGS_ACTUAL_SIZE(gpt_config.p) - 1UL)) != 0UL) ||
1713 	    ((size & (GPT_PGS_ACTUAL_SIZE(gpt_config.p) - 1UL)) != 0UL) ||
1714 	    (size == 0UL) ||
1715 	    ((base + size) >= GPT_PPS_ACTUAL_SIZE(gpt_config.t))) {
1716 		VERBOSE("GPT: Invalid granule transition address range!\n");
1717 		VERBOSE("      Base=0x%"PRIx64"\n", base);
1718 		VERBOSE("      Size=0x%lx\n", size);
1719 		return -EINVAL;
1720 	}
1721 
1722 	/* Delegate request can only come from REALM or SECURE */
1723 	if ((src_sec_state != SMC_FROM_REALM) &&
1724 	    (src_sec_state != SMC_FROM_SECURE)) {
1725 		VERBOSE("GPT: Invalid caller security state 0x%x\n",
1726 							src_sec_state);
1727 		return -EINVAL;
1728 	}
1729 
1730 	if (src_sec_state == SMC_FROM_REALM) {
1731 		target_pas = GPT_GPI_REALM;
1732 		nse = (uint64_t)GPT_NSE_REALM << GPT_NSE_SHIFT;
1733 		l1_desc = GPT_L1_REALM_DESC;
1734 	} else {
1735 		target_pas = GPT_GPI_SECURE;
1736 		nse = (uint64_t)GPT_NSE_SECURE << GPT_NSE_SHIFT;
1737 		l1_desc = GPT_L1_SECURE_DESC;
1738 	}
1739 
1740 	res = get_gpi_params(base, &gpi_info);
1741 	if (res != 0) {
1742 		return res;
1743 	}
1744 
1745 	/*
1746 	 * Access to GPT is controlled by a lock to ensure that no more
1747 	 * than one CPU is allowed to make changes at any given time.
1748 	 */
1749 	GPT_LOCK;
1750 	read_gpi(&gpi_info);
1751 
1752 	/* Check that the current address is in NS state */
1753 	if (gpi_info.gpi != GPT_GPI_NS) {
1754 		VERBOSE("GPT: Only Granule in NS state can be delegated.\n");
1755 		VERBOSE("      Caller: %u, Current GPI: %u\n", src_sec_state,
1756 			gpi_info.gpi);
1757 		GPT_UNLOCK;
1758 		return -EPERM;
1759 	}
1760 
1761 #if (RME_GPT_MAX_BLOCK != 0)
1762 	/* Check for Contiguous descriptor */
1763 	if ((gpi_info.gpt_l1_desc & GPT_L1_TYPE_CONT_DESC_MASK) ==
1764 					GPT_L1_TYPE_CONT_DESC) {
1765 		shatter_block(base, &gpi_info, GPT_L1_NS_DESC);
1766 	}
1767 #endif
1768 	/*
1769 	 * In order to maintain mutual distrust between Realm and Secure
1770 	 * states, remove any data speculatively fetched into the target
1771 	 * physical address space.
1772 	 * Issue DC CIPAPA or DC_CIGDPAPA on implementations with FEAT_MTE2.
1773 	 */
1774 	flush_page_to_popa(base | nse);
1775 
1776 	write_gpt(&gpi_info.gpt_l1_desc, gpi_info.gpt_l1_addr,
1777 		  gpi_info.gpi_shift, gpi_info.idx, target_pas);
1778 
1779 	/* Ensure that all agents observe the new configuration */
1780 	tlbi_page_dsbosh(base);
1781 
1782 	nse = (uint64_t)GPT_NSE_NS << GPT_NSE_SHIFT;
1783 
1784 	/* Ensure that the scrubbed data have made it past the PoPA */
1785 	flush_page_to_popa(base | nse);
1786 
1787 #if (RME_GPT_MAX_BLOCK != 0)
1788 	if (gpi_info.gpt_l1_desc == l1_desc) {
1789 		/* Try to fuse */
1790 		fuse_block(base, &gpi_info, l1_desc);
1791 	}
1792 #endif
1793 
1794 	/* Unlock the lock to GPT */
1795 	GPT_UNLOCK;
1796 
1797 	/*
1798 	 * The isb() will be done as part of context
1799 	 * synchronization when returning to lower EL.
1800 	 */
1801 	VERBOSE("GPT: Granule 0x%"PRIx64" GPI 0x%x->0x%x\n",
1802 		base, gpi_info.gpi, target_pas);
1803 
1804 	return 0;
1805 }
1806 
1807 /*
1808  * This function is the granule transition undelegate service. When a granule
1809  * transition request occurs it is routed to this function where the request is
1810  * validated then fulfilled if possible.
1811  *
1812  * TODO: implement support for transitioning multiple granules at once.
1813  *
1814  * Parameters
1815  *   base		Base address of the region to transition, must be
1816  *			aligned to granule size.
1817  *   size		Size of region to transition, must be aligned to granule
1818  *			size.
1819  *   src_sec_state	Security state of the caller.
1820  *
1821  * Return
1822  *    Negative Linux error code in the event of a failure, 0 for success.
1823  */
1824 int gpt_undelegate_pas(uint64_t base, size_t size, unsigned int src_sec_state)
1825 {
1826 	gpi_info_t gpi_info;
1827 	uint64_t nse, __unused l1_desc;
1828 	int res;
1829 
1830 	/* Ensure that the tables have been set up before taking requests */
1831 	assert(gpt_config.plat_gpt_l0_base != 0UL);
1832 
1833 	/* Ensure that MMU and caches are enabled */
1834 	assert((read_sctlr_el3() & SCTLR_C_BIT) != 0UL);
1835 
1836 	/* See if this is a single or a range of granule transition */
1837 	if (size != GPT_PGS_ACTUAL_SIZE(gpt_config.p)) {
1838 		return -EINVAL;
1839 	}
1840 
1841 	/* Check that base and size are valid */
1842 	if ((ULONG_MAX - base) < size) {
1843 		VERBOSE("GPT: Transition request address overflow!\n");
1844 		VERBOSE("      Base=0x%"PRIx64"\n", base);
1845 		VERBOSE("      Size=0x%lx\n", size);
1846 		return -EINVAL;
1847 	}
1848 
1849 	/* Make sure base and size are valid */
1850 	if (((base & (GPT_PGS_ACTUAL_SIZE(gpt_config.p) - 1UL)) != 0UL) ||
1851 	    ((size & (GPT_PGS_ACTUAL_SIZE(gpt_config.p) - 1UL)) != 0UL) ||
1852 	    (size == 0UL) ||
1853 	    ((base + size) >= GPT_PPS_ACTUAL_SIZE(gpt_config.t))) {
1854 		VERBOSE("GPT: Invalid granule transition address range!\n");
1855 		VERBOSE("      Base=0x%"PRIx64"\n", base);
1856 		VERBOSE("      Size=0x%lx\n", size);
1857 		return -EINVAL;
1858 	}
1859 
1860 	res = get_gpi_params(base, &gpi_info);
1861 	if (res != 0) {
1862 		return res;
1863 	}
1864 
1865 	/*
1866 	 * Access to GPT is controlled by a lock to ensure that no more
1867 	 * than one CPU is allowed to make changes at any given time.
1868 	 */
1869 	GPT_LOCK;
1870 	read_gpi(&gpi_info);
1871 
1872 	/* Check that the current address is in the delegated state */
1873 	if ((src_sec_state == SMC_FROM_REALM) &&
1874 		(gpi_info.gpi == GPT_GPI_REALM)) {
1875 		l1_desc = GPT_L1_REALM_DESC;
1876 		nse = (uint64_t)GPT_NSE_REALM << GPT_NSE_SHIFT;
1877 	} else if ((src_sec_state == SMC_FROM_SECURE) &&
1878 		(gpi_info.gpi == GPT_GPI_SECURE)) {
1879 		l1_desc = GPT_L1_SECURE_DESC;
1880 		nse = (uint64_t)GPT_NSE_SECURE << GPT_NSE_SHIFT;
1881 	} else {
1882 		VERBOSE("GPT: Only Granule in REALM or SECURE state can be undelegated\n");
1883 		VERBOSE("      Caller: %u Current GPI: %u\n", src_sec_state,
1884 			gpi_info.gpi);
1885 		GPT_UNLOCK;
1886 		return -EPERM;
1887 	}
1888 
1889 #if (RME_GPT_MAX_BLOCK != 0)
1890 	/* Check for Contiguous descriptor */
1891 	if ((gpi_info.gpt_l1_desc & GPT_L1_TYPE_CONT_DESC_MASK) ==
1892 					GPT_L1_TYPE_CONT_DESC) {
1893 		shatter_block(base, &gpi_info, l1_desc);
1894 	}
1895 #endif
1896 	/*
1897 	 * In order to maintain mutual distrust between Realm and Secure
1898 	 * states, remove access now, in order to guarantee that writes
1899 	 * to the currently-accessible physical address space will not
1900 	 * later become observable.
1901 	 */
1902 	write_gpt(&gpi_info.gpt_l1_desc, gpi_info.gpt_l1_addr,
1903 		  gpi_info.gpi_shift, gpi_info.idx, GPT_GPI_NO_ACCESS);
1904 
1905 	/* Ensure that all agents observe the new NO_ACCESS configuration */
1906 	tlbi_page_dsbosh(base);
1907 
1908 	/* Ensure that the scrubbed data have made it past the PoPA */
1909 	flush_page_to_popa(base | nse);
1910 
1911 	/*
1912 	 * Remove any data loaded speculatively in NS space from before
1913 	 * the scrubbing.
1914 	 */
1915 	nse = (uint64_t)GPT_NSE_NS << GPT_NSE_SHIFT;
1916 
1917 	flush_page_to_popa(base | nse);
1918 
1919 	/* Clear existing GPI encoding and transition granule */
1920 	write_gpt(&gpi_info.gpt_l1_desc, gpi_info.gpt_l1_addr,
1921 		  gpi_info.gpi_shift, gpi_info.idx, GPT_GPI_NS);
1922 
1923 	/* Ensure that all agents observe the new NS configuration */
1924 	tlbi_page_dsbosh(base);
1925 
1926 #if (RME_GPT_MAX_BLOCK != 0)
1927 	if (gpi_info.gpt_l1_desc == GPT_L1_NS_DESC) {
1928 		/* Try to fuse */
1929 		fuse_block(base, &gpi_info, GPT_L1_NS_DESC);
1930 	}
1931 #endif
1932 	/* Unlock the lock to GPT */
1933 	GPT_UNLOCK;
1934 
1935 	/*
1936 	 * The isb() will be done as part of context
1937 	 * synchronization when returning to lower EL.
1938 	 */
1939 	VERBOSE("GPT: Granule 0x%"PRIx64" GPI 0x%x->0x%x\n",
1940 		base, gpi_info.gpi, GPT_GPI_NS);
1941 
1942 	return 0;
1943 }
1944