xref: /rk3399_ARM-atf/lib/gpt_rme/gpt_rme.c (revision 09a4bcb81bc37def40ecb638b121e00e05bcbcbd)
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_NSO         U(0xD)
268  *   GPT_GPI_ANY         U(0xF)
269  *
270  * Parameters
271  *   gpi		GPI to check for validity.
272  *
273  * Return
274  *   true for a valid GPI, false for an invalid one.
275  */
276 static bool is_gpi_valid(unsigned int gpi)
277 {
278 	switch (gpi) {
279 	case GPT_GPI_NO_ACCESS:
280 	case GPT_GPI_SECURE:
281 	case GPT_GPI_NS:
282 	case GPT_GPI_ROOT:
283 	case GPT_GPI_REALM:
284 	case GPT_GPI_ANY:
285 		return true;
286 	case GPT_GPI_NSO:
287 		return is_feat_rme_gpc2_present();
288 	default:
289 		return false;
290 	}
291 }
292 
293 /*
294  * This function checks to see if two PAS regions overlap.
295  *
296  * Parameters
297  *   base_1: base address of first PAS
298  *   size_1: size of first PAS
299  *   base_2: base address of second PAS
300  *   size_2: size of second PAS
301  *
302  * Return
303  *   True if PAS regions overlap, false if they do not.
304  */
305 static bool check_pas_overlap(uintptr_t base_1, size_t size_1,
306 			      uintptr_t base_2, size_t size_2)
307 {
308 	if (((base_1 + size_1) > base_2) && ((base_2 + size_2) > base_1)) {
309 		return true;
310 	}
311 	return false;
312 }
313 
314 /*
315  * This helper function checks to see if a PAS region from index 0 to
316  * (pas_idx - 1) occupies the L0 region at index l0_idx in the L0 table.
317  *
318  * Parameters
319  *   l0_idx:      Index of the L0 entry to check
320  *   pas_regions: PAS region array
321  *   pas_idx:     Upper bound of the PAS array index.
322  *
323  * Return
324  *   True if a PAS region occupies the L0 region in question, false if not.
325  */
326 static bool does_previous_pas_exist_here(unsigned int l0_idx,
327 					 pas_region_t *pas_regions,
328 					 unsigned int pas_idx)
329 {
330 	/* Iterate over PAS regions up to pas_idx */
331 	for (unsigned int i = 0U; i < pas_idx; i++) {
332 		if (check_pas_overlap((GPT_L0GPTSZ_ACTUAL_SIZE * l0_idx),
333 		    GPT_L0GPTSZ_ACTUAL_SIZE,
334 		    pas_regions[i].base_pa, pas_regions[i].size)) {
335 			return true;
336 		}
337 	}
338 	return false;
339 }
340 
341 /*
342  * This function iterates over all of the PAS regions and checks them to ensure
343  * proper alignment of base and size, that the GPI is valid, and that no regions
344  * overlap. As a part of the overlap checks, this function checks existing L0
345  * mappings against the new PAS regions in the event that gpt_init_pas_l1_tables
346  * is called multiple times to place L1 tables in different areas of memory. It
347  * also counts the number of L1 tables needed and returns it on success.
348  *
349  * Parameters
350  *   *pas_regions	Pointer to array of PAS region structures.
351  *   pas_region_cnt	Total number of PAS regions in the array.
352  *
353  * Return
354  *   Negative Linux error code in the event of a failure, number of L1 regions
355  *   required when successful.
356  */
357 static int validate_pas_mappings(pas_region_t *pas_regions,
358 				 unsigned int pas_region_cnt)
359 {
360 	unsigned int idx;
361 	unsigned int l1_cnt = 0U;
362 	unsigned int pas_l1_cnt;
363 	uint64_t *l0_desc = (uint64_t *)gpt_config.plat_gpt_l0_base;
364 
365 	assert(pas_regions != NULL);
366 	assert(pas_region_cnt != 0U);
367 
368 	for (idx = 0U; idx < pas_region_cnt; idx++) {
369 		/* Check for arithmetic overflow in region */
370 		if ((ULONG_MAX - pas_regions[idx].base_pa) <
371 		    pas_regions[idx].size) {
372 			ERROR("GPT: Address overflow in PAS[%u]!\n", idx);
373 			return -EOVERFLOW;
374 		}
375 
376 		/* Initial checks for PAS validity */
377 		if (((pas_regions[idx].base_pa + pas_regions[idx].size) >
378 		    GPT_PPS_ACTUAL_SIZE(gpt_config.t)) ||
379 		    !is_gpi_valid(GPT_PAS_ATTR_GPI(pas_regions[idx].attrs))) {
380 			ERROR("GPT: PAS[%u] is invalid!\n", idx);
381 			return -EFAULT;
382 		}
383 
384 		/*
385 		 * Make sure this PAS does not overlap with another one. We
386 		 * start from idx + 1 instead of 0 since prior PAS mappings will
387 		 * have already checked themselves against this one.
388 		 */
389 		for (unsigned int i = idx + 1U; i < pas_region_cnt; i++) {
390 			if (check_pas_overlap(pas_regions[idx].base_pa,
391 			    pas_regions[idx].size,
392 			    pas_regions[i].base_pa,
393 			    pas_regions[i].size)) {
394 				ERROR("GPT: PAS[%u] overlaps with PAS[%u]\n",
395 					i, idx);
396 				return -EFAULT;
397 			}
398 		}
399 
400 		/*
401 		 * Since this function can be called multiple times with
402 		 * separate L1 tables we need to check the existing L0 mapping
403 		 * to see if this PAS would fall into one that has already been
404 		 * initialized.
405 		 */
406 		for (unsigned int i =
407 			(unsigned int)GPT_L0_IDX(pas_regions[idx].base_pa);
408 			i <= GPT_L0_IDX(pas_regions[idx].base_pa +
409 					pas_regions[idx].size - 1UL);
410 			i++) {
411 			if ((GPT_L0_TYPE(l0_desc[i]) == GPT_L0_TYPE_BLK_DESC) &&
412 			    (GPT_L0_BLKD_GPI(l0_desc[i]) == GPT_GPI_ANY)) {
413 				/* This descriptor is unused so continue */
414 				continue;
415 			}
416 
417 			/*
418 			 * This descriptor has been initialized in a previous
419 			 * call to this function so cannot be initialized again.
420 			 */
421 			ERROR("GPT: PAS[%u] overlaps with previous L0[%u]!\n",
422 			      idx, i);
423 			return -EFAULT;
424 		}
425 
426 		/* Check for block mapping (L0) type */
427 		if (GPT_PAS_ATTR_MAP_TYPE(pas_regions[idx].attrs) ==
428 		    GPT_PAS_ATTR_MAP_TYPE_BLOCK) {
429 			/* Make sure base and size are block-aligned */
430 			if (!GPT_IS_L0_ALIGNED(pas_regions[idx].base_pa) ||
431 			    !GPT_IS_L0_ALIGNED(pas_regions[idx].size)) {
432 				ERROR("GPT: PAS[%u] is not block-aligned!\n",
433 				      idx);
434 				return -EFAULT;
435 			}
436 
437 			continue;
438 		}
439 
440 		/* Check for granule mapping (L1) type */
441 		if (GPT_PAS_ATTR_MAP_TYPE(pas_regions[idx].attrs) ==
442 		    GPT_PAS_ATTR_MAP_TYPE_GRANULE) {
443 			/* Make sure base and size are granule-aligned */
444 			if (!GPT_IS_L1_ALIGNED(gpt_config.p, pas_regions[idx].base_pa) ||
445 			    !GPT_IS_L1_ALIGNED(gpt_config.p, pas_regions[idx].size)) {
446 				ERROR("GPT: PAS[%u] is not granule-aligned!\n",
447 				      idx);
448 				return -EFAULT;
449 			}
450 
451 			/* Find how many L1 tables this PAS occupies */
452 			pas_l1_cnt = (GPT_L0_IDX(pas_regions[idx].base_pa +
453 				     pas_regions[idx].size - 1UL) -
454 				     GPT_L0_IDX(pas_regions[idx].base_pa) + 1U);
455 
456 			/*
457 			 * This creates a situation where, if multiple PAS
458 			 * regions occupy the same table descriptor, we can get
459 			 * an artificially high total L1 table count. The way we
460 			 * handle this is by checking each PAS against those
461 			 * before it in the array, and if they both occupy the
462 			 * same PAS we subtract from pas_l1_cnt and only the
463 			 * first PAS in the array gets to count it.
464 			 */
465 
466 			/*
467 			 * If L1 count is greater than 1 we know the start and
468 			 * end PAs are in different L0 regions so we must check
469 			 * both for overlap against other PAS.
470 			 */
471 			if (pas_l1_cnt > 1) {
472 				if (does_previous_pas_exist_here(
473 				    GPT_L0_IDX(pas_regions[idx].base_pa +
474 				    pas_regions[idx].size - 1UL),
475 				    pas_regions, idx)) {
476 					pas_l1_cnt--;
477 				}
478 			}
479 
480 			if (does_previous_pas_exist_here(
481 			    GPT_L0_IDX(pas_regions[idx].base_pa),
482 			    pas_regions, idx)) {
483 				pas_l1_cnt--;
484 			}
485 
486 			l1_cnt += pas_l1_cnt;
487 			continue;
488 		}
489 
490 		/* If execution reaches this point, mapping type is invalid */
491 		ERROR("GPT: PAS[%u] has invalid mapping type 0x%x.\n", idx,
492 		      GPT_PAS_ATTR_MAP_TYPE(pas_regions[idx].attrs));
493 		return -EINVAL;
494 	}
495 
496 	return l1_cnt;
497 }
498 
499 /*
500  * This function validates L0 initialization parameters.
501  *
502  * Parameters
503  *   l0_mem_base	Base address of memory used for L0 table.
504  *   l0_mem_size	Size of memory available for L0 table.
505  *
506  * Return
507  *   Negative Linux error code in the event of a failure, 0 for success.
508  */
509 static int validate_l0_params(gpccr_pps_e pps, uintptr_t l0_mem_base,
510 				size_t l0_mem_size)
511 {
512 	size_t l0_alignment;
513 
514 	/*
515 	 * Make sure PPS is valid and then store it since macros need this value
516 	 * to work.
517 	 */
518 	if (pps > GPT_PPS_MAX) {
519 		ERROR("GPT: Invalid PPS: 0x%x\n", pps);
520 		return -EINVAL;
521 	}
522 	gpt_config.pps = pps;
523 	gpt_config.t = gpt_t_lookup[pps];
524 
525 	/* Alignment must be the greater of 4KB or L0 table size */
526 	l0_alignment = SZ_4K;
527 	if (l0_alignment < GPT_L0_TABLE_SIZE(gpt_config.t)) {
528 		l0_alignment = GPT_L0_TABLE_SIZE(gpt_config.t);
529 	}
530 
531 	/* Check base address */
532 	if ((l0_mem_base == 0UL) ||
533 	   ((l0_mem_base & (l0_alignment - 1UL)) != 0UL)) {
534 		ERROR("GPT: Invalid L0 base address: 0x%lx\n", l0_mem_base);
535 		return -EFAULT;
536 	}
537 
538 	/* Check memory size for L0 table */
539 	if (l0_mem_size < GPT_L0_TABLE_SIZE(gpt_config.t)) {
540 		ERROR("GPT: Inadequate L0 memory\n");
541 		ERROR("      Expected 0x%lx bytes, got 0x%lx\n",
542 				GPT_L0_TABLE_SIZE(gpt_config.t), l0_mem_size);
543 		return -ENOMEM;
544 	}
545 
546 	return 0;
547 }
548 
549 /*
550  * In the event that L1 tables are needed, this function validates
551  * the L1 table generation parameters.
552  *
553  * Parameters
554  *   l1_mem_base	Base address of memory used for L1 table allocation.
555  *   l1_mem_size	Total size of memory available for L1 tables.
556  *   l1_gpt_cnt		Number of L1 tables needed.
557  *
558  * Return
559  *   Negative Linux error code in the event of a failure, 0 for success.
560  */
561 static int validate_l1_params(uintptr_t l1_mem_base, size_t l1_mem_size,
562 				unsigned int l1_gpt_cnt)
563 {
564 	size_t l1_gpt_mem_sz;
565 
566 	/* Check if the granularity is supported */
567 	if (!xlat_arch_is_granule_size_supported(
568 	    GPT_PGS_ACTUAL_SIZE(gpt_config.p))) {
569 		return -EPERM;
570 	}
571 
572 	/* Make sure L1 tables are aligned to their size */
573 	if ((l1_mem_base & (GPT_L1_TABLE_SIZE(gpt_config.p) - 1UL)) != 0UL) {
574 		ERROR("GPT: Unaligned L1 GPT base address: 0x%"PRIxPTR"\n",
575 		      l1_mem_base);
576 		return -EFAULT;
577 	}
578 
579 	/* Get total memory needed for L1 tables */
580 	l1_gpt_mem_sz = l1_gpt_cnt * GPT_L1_TABLE_SIZE(gpt_config.p);
581 
582 	/* Check for overflow */
583 	if ((l1_gpt_mem_sz / GPT_L1_TABLE_SIZE(gpt_config.p)) != l1_gpt_cnt) {
584 		ERROR("GPT: Overflow calculating L1 memory size\n");
585 		return -ENOMEM;
586 	}
587 
588 	/* Make sure enough space was supplied */
589 	if (l1_mem_size < l1_gpt_mem_sz) {
590 		ERROR("%sL1 GPTs%s", (const char *)"GPT: Inadequate ",
591 			(const char *)" memory\n");
592 		ERROR("      Expected 0x%lx bytes, got 0x%lx\n",
593 			l1_gpt_mem_sz, l1_mem_size);
594 		return -ENOMEM;
595 	}
596 
597 	VERBOSE("GPT: Requested 0x%lx bytes for L1 GPTs\n", l1_gpt_mem_sz);
598 	return 0;
599 }
600 
601 /*
602  * This function initializes L0 block descriptors (regions that cannot be
603  * transitioned at the granule level) according to the provided PAS.
604  *
605  * Parameters
606  *   *pas		Pointer to the structure defining the PAS region to
607  *			initialize.
608  */
609 static void generate_l0_blk_desc(pas_region_t *pas)
610 {
611 	uint64_t gpt_desc;
612 	unsigned long idx, end_idx;
613 	uint64_t *l0_gpt_arr;
614 
615 	assert(gpt_config.plat_gpt_l0_base != 0UL);
616 	assert(pas != NULL);
617 
618 	/*
619 	 * Checking of PAS parameters has already been done in
620 	 * validate_pas_mappings so no need to check the same things again.
621 	 */
622 
623 	l0_gpt_arr = (uint64_t *)gpt_config.plat_gpt_l0_base;
624 
625 	/* Create the GPT Block descriptor for this PAS region */
626 	gpt_desc = GPT_L0_BLK_DESC(GPT_PAS_ATTR_GPI(pas->attrs));
627 
628 	/* Start index of this region in L0 GPTs */
629 	idx = GPT_L0_IDX(pas->base_pa);
630 
631 	/*
632 	 * Determine number of L0 GPT descriptors covered by
633 	 * this PAS region and use the count to populate these
634 	 * descriptors.
635 	 */
636 	end_idx = GPT_L0_IDX(pas->base_pa + pas->size);
637 
638 	/* Generate the needed block descriptors */
639 	for (; idx < end_idx; idx++) {
640 		l0_gpt_arr[idx] = gpt_desc;
641 		VERBOSE("GPT: L0 entry (BLOCK) index %lu [%p]: GPI = 0x%"PRIx64" (0x%"PRIx64")\n",
642 			idx, &l0_gpt_arr[idx],
643 			(gpt_desc >> GPT_L0_BLK_DESC_GPI_SHIFT) &
644 			GPT_L0_BLK_DESC_GPI_MASK, l0_gpt_arr[idx]);
645 	}
646 }
647 
648 /*
649  * Helper function to determine if the end physical address lies in the same L0
650  * region as the current physical address. If true, the end physical address is
651  * returned else, the start address of the next region is returned.
652  *
653  * Parameters
654  *   cur_pa		Physical address of the current PA in the loop through
655  *			the range.
656  *   end_pa		Physical address of the end PA in a PAS range.
657  *
658  * Return
659  *   The PA of the end of the current range.
660  */
661 static uintptr_t get_l1_end_pa(uintptr_t cur_pa, uintptr_t end_pa)
662 {
663 	uintptr_t cur_idx;
664 	uintptr_t end_idx;
665 
666 	cur_idx = GPT_L0_IDX(cur_pa);
667 	end_idx = GPT_L0_IDX(end_pa);
668 
669 	assert(cur_idx <= end_idx);
670 
671 	if (cur_idx == end_idx) {
672 		return end_pa;
673 	}
674 
675 	return (cur_idx + 1UL) << GPT_L0_IDX_SHIFT;
676 }
677 
678 /*
679  * Helper function to fill out GPI entries from 'first' granule address of
680  * the specified 'length' in a single L1 table with 'l1_desc' Contiguous
681  * descriptor.
682  *
683  * Parameters
684  *   l1			Pointer to L1 table to fill out
685  *   first		Address of first granule in range
686  *   length		Length of the range in bytes
687  *   gpi		GPI set this range to
688  *
689  * Return
690  *   Address of next granule in range.
691  */
692 __unused static uintptr_t fill_l1_cont_desc(uint64_t *l1, uintptr_t first,
693 					    size_t length, unsigned int gpi)
694 {
695 	/*
696 	 * Look up table for contiguous blocks and descriptors.
697 	 * Entries should be defined in descending block sizes:
698 	 * 512MB, 32MB and 2MB.
699 	 */
700 	static const gpt_fill_lookup_t gpt_fill_lookup[] = {
701 #if (RME_GPT_MAX_BLOCK == 512)
702 		{ SZ_512M, GPT_L1_CONT_DESC_512MB },
703 #endif
704 #if (RME_GPT_MAX_BLOCK >= 32)
705 		{ SZ_32M, GPT_L1_CONT_DESC_32MB },
706 #endif
707 #if (RME_GPT_MAX_BLOCK != 0)
708 		{ SZ_2M, GPT_L1_CONT_DESC_2MB }
709 #endif
710 	};
711 
712 	/*
713 	 * Iterate through all block sizes (512MB, 32MB and 2MB)
714 	 * starting with maximum supported.
715 	 */
716 	for (unsigned long i = 0UL; i < ARRAY_SIZE(gpt_fill_lookup); i++) {
717 		/* Calculate index */
718 		unsigned long idx = GPT_L1_INDEX(first);
719 
720 		/* Contiguous block size */
721 		size_t cont_size = gpt_fill_lookup[i].size;
722 
723 		if (GPT_REGION_IS_CONT(length, first, cont_size)) {
724 
725 			/* Generate Contiguous descriptor */
726 			uint64_t l1_desc = GPT_L1_GPI_CONT_DESC(gpi,
727 						gpt_fill_lookup[i].desc);
728 
729 			/* Number of 128-bit L1 entries in block */
730 			unsigned int cnt;
731 
732 			switch (cont_size) {
733 			case SZ_512M:
734 				cnt = L1_QWORDS_512MB;
735 				break;
736 			case SZ_32M:
737 				cnt = L1_QWORDS_32MB;
738 				break;
739 			default:			/* SZ_2MB */
740 				cnt = L1_QWORDS_2MB;
741 			}
742 
743 			VERBOSE("GPT: Contiguous descriptor 0x%"PRIxPTR" %luMB\n",
744 				first, cont_size / SZ_1M);
745 
746 			/* Fill Contiguous descriptors */
747 			fill_desc(&l1[idx], l1_desc, cnt);
748 			return (first + cont_size);
749 		}
750 	}
751 
752 	return first;
753 }
754 
755 /* Build Granules descriptor with the same 'gpi' for every GPI entry */
756 static uint64_t build_l1_desc(unsigned int gpi)
757 {
758 	uint64_t l1_desc = (uint64_t)gpi | ((uint64_t)gpi << 4);
759 
760 	l1_desc |= (l1_desc << 8);
761 	l1_desc |= (l1_desc << 16);
762 	return (l1_desc | (l1_desc << 32));
763 }
764 
765 /*
766  * Helper function to fill out GPI entries from 'first' to 'last' granule
767  * address in a single L1 table with 'l1_desc' Granules descriptor.
768  *
769  * Parameters
770  *   l1			Pointer to L1 table to fill out
771  *   first		Address of first granule in range
772  *   last		Address of last granule in range (inclusive)
773  *   gpi		GPI set this range to
774  *
775  * Return
776  *   Address of next granule in range.
777  */
778 static uintptr_t fill_l1_gran_desc(uint64_t *l1, uintptr_t first,
779 				   uintptr_t last, unsigned int gpi)
780 {
781 	uint64_t gpi_mask;
782 	unsigned long i;
783 
784 	/* Generate Granules descriptor */
785 	uint64_t l1_desc = build_l1_desc(gpi);
786 
787 	/* Shift the mask if we're starting in the middle of an L1 entry */
788 	gpi_mask = ULONG_MAX << (GPT_L1_GPI_IDX(gpt_config.p, first) << 2);
789 
790 	/* Fill out each L1 entry for this region */
791 	for (i = GPT_L1_INDEX(first); i <= GPT_L1_INDEX(last); i++) {
792 
793 		/* Account for stopping in the middle of an L1 entry */
794 		if (i == GPT_L1_INDEX(last)) {
795 			gpi_mask &= (gpi_mask >> ((15U -
796 				    GPT_L1_GPI_IDX(gpt_config.p, last)) << 2));
797 		}
798 
799 		assert((l1[i] & gpi_mask) == (GPT_L1_ANY_DESC & gpi_mask));
800 
801 		/* Write GPI values */
802 		l1[i] = (l1[i] & ~gpi_mask) | (l1_desc & gpi_mask);
803 
804 		/* Reset mask */
805 		gpi_mask = ULONG_MAX;
806 	}
807 
808 	return last + GPT_PGS_ACTUAL_SIZE(gpt_config.p);
809 }
810 
811 /*
812  * Helper function to fill out GPI entries in a single L1 table.
813  * This function fills out an entire L1 table with either Granules or Contiguous
814  * (RME_GPT_MAX_BLOCK != 0) descriptors depending on region length and alignment.
815  * Note. If RME_GPT_MAX_BLOCK == 0, then the L1 tables are filled with regular
816  * Granules descriptors.
817  *
818  * Parameters
819  *   l1			Pointer to L1 table to fill out
820  *   first		Address of first granule in range
821  *   last		Address of last granule in range (inclusive)
822  *   gpi		GPI set this range to
823  */
824 static void fill_l1_tbl(uint64_t *l1, uintptr_t first, uintptr_t last,
825 			unsigned int gpi)
826 {
827 	assert(l1 != NULL);
828 	assert(first <= last);
829 	assert((first & (GPT_PGS_ACTUAL_SIZE(gpt_config.p) - 1UL)) == 0UL);
830 	assert((last & (GPT_PGS_ACTUAL_SIZE(gpt_config.p) - 1UL)) == 0UL);
831 	assert(GPT_L0_IDX(first) == GPT_L0_IDX(last));
832 
833 #if (RME_GPT_MAX_BLOCK != 0)
834 	while (first <= last) {
835 		/* Region length */
836 		size_t length = last - first + GPT_PGS_ACTUAL_SIZE(gpt_config.p);
837 
838 		if (length < SZ_2M) {
839 			/*
840 			 * Fill with Granule descriptors in case of
841 			 * region length < 2MB.
842 			 */
843 			first = fill_l1_gran_desc(l1, first, last, gpi);
844 
845 		} else if ((first & (SZ_2M - UL(1))) == UL(0)) {
846 			/*
847 			 * For region length >= 2MB and at least 2MB aligned
848 			 * call to fill_l1_cont_desc will iterate through
849 			 * all block sizes (512MB, 32MB and 2MB) supported and
850 			 * fill corresponding Contiguous descriptors.
851 			 */
852 			first = fill_l1_cont_desc(l1, first, length, gpi);
853 		} else {
854 			/*
855 			 * For not aligned region >= 2MB fill with Granules
856 			 * descriptors up to the next 2MB aligned address.
857 			 */
858 			uintptr_t new_last = ALIGN_2MB(first + SZ_2M) -
859 					GPT_PGS_ACTUAL_SIZE(gpt_config.p);
860 
861 			first = fill_l1_gran_desc(l1, first, new_last, gpi);
862 		}
863 	}
864 #else
865 	/* Fill with Granule descriptors */
866 	first = fill_l1_gran_desc(l1, first, last, gpi);
867 #endif
868 	assert(first == (last + GPT_PGS_ACTUAL_SIZE(gpt_config.p)));
869 }
870 
871 /*
872  * This function finds the next available unused L1 table and initializes all
873  * granules descriptor entries to GPI_ANY. This ensures that there are no chunks
874  * of GPI_NO_ACCESS (0b0000) memory floating around in the system in the
875  * event that a PAS region stops midway through an L1 table, thus guaranteeing
876  * that all memory not explicitly assigned is GPI_ANY. This function does not
877  * check for overflow conditions, that should be done by the caller.
878  *
879  * Return
880  *   Pointer to the next available L1 table.
881  */
882 static uint64_t *get_new_l1_tbl(void)
883 {
884 	/* Retrieve the next L1 table */
885 	uint64_t *l1 = (uint64_t *)gpt_l1_tbl;
886 
887 	/* Increment L1 GPT address */
888 	gpt_l1_tbl += GPT_L1_TABLE_SIZE(gpt_config.p);
889 
890 	/* Initialize all GPIs to GPT_GPI_ANY */
891 	for (unsigned int i = 0U; i < GPT_L1_ENTRY_COUNT(gpt_config.p); i++) {
892 		l1[i] = GPT_L1_ANY_DESC;
893 	}
894 
895 	return l1;
896 }
897 
898 /*
899  * When L1 tables are needed, this function creates the necessary L0 table
900  * descriptors and fills out the L1 table entries according to the supplied
901  * PAS range.
902  *
903  * Parameters
904  *   *pas		Pointer to the structure defining the PAS region.
905  */
906 static void generate_l0_tbl_desc(pas_region_t *pas)
907 {
908 	uintptr_t end_pa;
909 	uintptr_t cur_pa;
910 	uintptr_t last_gran_pa;
911 	uint64_t *l0_gpt_base;
912 	uint64_t *l1_gpt_arr;
913 	unsigned int l0_idx, gpi;
914 
915 	assert(gpt_config.plat_gpt_l0_base != 0UL);
916 	assert(pas != NULL);
917 
918 	/*
919 	 * Checking of PAS parameters has already been done in
920 	 * validate_pas_mappings so no need to check the same things again.
921 	 */
922 	end_pa = pas->base_pa + pas->size;
923 	l0_gpt_base = (uint64_t *)gpt_config.plat_gpt_l0_base;
924 
925 	/* We start working from the granule at base PA */
926 	cur_pa = pas->base_pa;
927 
928 	/* Get GPI */
929 	gpi = GPT_PAS_ATTR_GPI(pas->attrs);
930 
931 	/* Iterate over each L0 region in this memory range */
932 	for (l0_idx = (unsigned int)GPT_L0_IDX(pas->base_pa);
933 	     l0_idx <= (unsigned int)GPT_L0_IDX(end_pa - 1UL);
934 	     l0_idx++) {
935 		/*
936 		 * See if the L0 entry is already a table descriptor or if we
937 		 * need to create one.
938 		 */
939 		if (GPT_L0_TYPE(l0_gpt_base[l0_idx]) == GPT_L0_TYPE_TBL_DESC) {
940 			/* Get the L1 array from the L0 entry */
941 			l1_gpt_arr = GPT_L0_TBLD_ADDR(l0_gpt_base[l0_idx]);
942 		} else {
943 			/* Get a new L1 table from the L1 memory space */
944 			l1_gpt_arr = get_new_l1_tbl();
945 
946 			/* Fill out the L0 descriptor and flush it */
947 			l0_gpt_base[l0_idx] = GPT_L0_TBL_DESC(l1_gpt_arr);
948 		}
949 
950 		VERBOSE("GPT: L0 entry (TABLE) index %u [%p] ==> L1 Addr %p (0x%"PRIx64")\n",
951 			l0_idx, &l0_gpt_base[l0_idx], l1_gpt_arr, l0_gpt_base[l0_idx]);
952 
953 		/*
954 		 * Determine the PA of the last granule in this L0 descriptor.
955 		 */
956 		last_gran_pa = get_l1_end_pa(cur_pa, end_pa) -
957 			       GPT_PGS_ACTUAL_SIZE(gpt_config.p);
958 
959 		/*
960 		 * Fill up L1 GPT entries between these two addresses. This
961 		 * function needs the addresses of the first granule and last
962 		 * granule in the range.
963 		 */
964 		fill_l1_tbl(l1_gpt_arr, cur_pa, last_gran_pa, gpi);
965 
966 		/* Advance cur_pa to first granule in next L0 region */
967 		cur_pa = get_l1_end_pa(cur_pa, end_pa);
968 	}
969 }
970 
971 /*
972  * This function flushes a range of L0 descriptors used by a given PAS region
973  * array. There is a chance that some unmodified L0 descriptors would be flushed
974  * in the case that there are "holes" in an array of PAS regions but overall
975  * this should be faster than individually flushing each modified L0 descriptor
976  * as they are created.
977  *
978  * Parameters
979  *   *pas		Pointer to an array of PAS regions.
980  *   pas_count		Number of entries in the PAS array.
981  */
982 static void flush_l0_for_pas_array(pas_region_t *pas, unsigned int pas_count)
983 {
984 	unsigned long idx;
985 	unsigned long start_idx;
986 	unsigned long end_idx;
987 	uint64_t *l0 = (uint64_t *)gpt_config.plat_gpt_l0_base;
988 
989 	assert(pas != NULL);
990 	assert(pas_count != 0U);
991 
992 	/* Initial start and end values */
993 	start_idx = GPT_L0_IDX(pas[0].base_pa);
994 	end_idx = GPT_L0_IDX(pas[0].base_pa + pas[0].size - 1UL);
995 
996 	/* Find lowest and highest L0 indices used in this PAS array */
997 	for (idx = 1UL; idx < pas_count; idx++) {
998 		if (GPT_L0_IDX(pas[idx].base_pa) < start_idx) {
999 			start_idx = GPT_L0_IDX(pas[idx].base_pa);
1000 		}
1001 		if (GPT_L0_IDX(pas[idx].base_pa + pas[idx].size - 1UL) > end_idx) {
1002 			end_idx = GPT_L0_IDX(pas[idx].base_pa + pas[idx].size - 1UL);
1003 		}
1004 	}
1005 
1006 	/*
1007 	 * Flush all covered L0 descriptors, add 1 because we need to include
1008 	 * the end index value.
1009 	 */
1010 	flush_dcache_range((uintptr_t)&l0[start_idx],
1011 			   ((end_idx + 1UL) - start_idx) * sizeof(uint64_t));
1012 }
1013 
1014 /*
1015  * Public API to enable granule protection checks once the tables have all been
1016  * initialized. This function is called at first initialization and then again
1017  * later during warm boots of CPU cores.
1018  *
1019  * Return
1020  *   Negative Linux error code in the event of a failure, 0 for success.
1021  */
1022 int gpt_enable(void)
1023 {
1024 	u_register_t gpccr_el3;
1025 
1026 	/*
1027 	 * Granule tables must be initialised before enabling
1028 	 * granule protection.
1029 	 */
1030 	if (gpt_config.plat_gpt_l0_base == 0UL) {
1031 		ERROR("GPT: Tables have not been initialized!\n");
1032 		return -EPERM;
1033 	}
1034 
1035 	/* Write the base address of the L0 tables into GPTBR */
1036 	write_gptbr_el3(((gpt_config.plat_gpt_l0_base >> GPTBR_BADDR_VAL_SHIFT)
1037 			>> GPTBR_BADDR_SHIFT) & GPTBR_BADDR_MASK);
1038 
1039 	/* GPCCR_EL3.PPS */
1040 	gpccr_el3 = SET_GPCCR_PPS(gpt_config.pps);
1041 
1042 	/* GPCCR_EL3.PGS */
1043 	gpccr_el3 |= SET_GPCCR_PGS(gpt_config.pgs);
1044 
1045 	/*
1046 	 * Since EL3 maps the L1 region as Inner shareable, use the same
1047 	 * shareability attribute for GPC as well so that
1048 	 * GPC fetches are visible to PEs
1049 	 */
1050 	gpccr_el3 |= SET_GPCCR_SH(GPCCR_SH_IS);
1051 
1052 	/* Outer and Inner cacheability set to Normal memory, WB, RA, WA */
1053 	gpccr_el3 |= SET_GPCCR_ORGN(GPCCR_ORGN_WB_RA_WA);
1054 	gpccr_el3 |= SET_GPCCR_IRGN(GPCCR_IRGN_WB_RA_WA);
1055 
1056 	/* Prepopulate GPCCR_EL3 but don't enable GPC yet */
1057 	write_gpccr_el3(gpccr_el3);
1058 	isb();
1059 
1060 	/* Invalidate any stale TLB entries and any cached register fields */
1061 	tlbipaallos();
1062 	dsb();
1063 	isb();
1064 
1065 	/* Enable GPT */
1066 	gpccr_el3 |= GPCCR_GPC_BIT;
1067 
1068 	/* Enable NSO encoding if FEAT_RME_GPC2 is supported. */
1069 	if (is_feat_rme_gpc2_present()) {
1070 		gpccr_el3 |= GPCCR_NSO_BIT;
1071 	}
1072 
1073 	/* TODO: Configure GPCCR_EL3_GPCP for Fault control */
1074 	write_gpccr_el3(gpccr_el3);
1075 	isb();
1076 	tlbipaallos();
1077 	dsb();
1078 	isb();
1079 
1080 	return 0;
1081 }
1082 
1083 /*
1084  * Public API to disable granule protection checks.
1085  */
1086 void gpt_disable(void)
1087 {
1088 	u_register_t gpccr_el3 = read_gpccr_el3();
1089 
1090 	write_gpccr_el3(gpccr_el3 & ~GPCCR_GPC_BIT);
1091 	dsbsy();
1092 	isb();
1093 }
1094 
1095 /*
1096  * Public API that initializes the entire protected space to GPT_GPI_ANY using
1097  * the L0 tables (block descriptors). Ideally, this function is invoked prior
1098  * to DDR discovery and initialization. The MMU must be initialized before
1099  * calling this function.
1100  *
1101  * Parameters
1102  *   pps		PPS value to use for table generation
1103  *   l0_mem_base	Base address of L0 tables in memory.
1104  *   l0_mem_size	Total size of memory available for L0 tables.
1105  *
1106  * Return
1107  *   Negative Linux error code in the event of a failure, 0 for success.
1108  */
1109 int gpt_init_l0_tables(gpccr_pps_e pps, uintptr_t l0_mem_base,
1110 		       size_t l0_mem_size)
1111 {
1112 	uint64_t gpt_desc;
1113 	int ret;
1114 
1115 	/* Ensure that MMU and Data caches are enabled */
1116 	assert((read_sctlr_el3() & SCTLR_C_BIT) != 0UL);
1117 
1118 	/* Validate other parameters */
1119 	ret = validate_l0_params(pps, l0_mem_base, l0_mem_size);
1120 	if (ret != 0) {
1121 		return ret;
1122 	}
1123 
1124 	/* Create the descriptor to initialize L0 entries with */
1125 	gpt_desc = GPT_L0_BLK_DESC(GPT_GPI_ANY);
1126 
1127 	/* Iterate through all L0 entries */
1128 	for (unsigned int i = 0U; i < GPT_L0_REGION_COUNT(gpt_config.t); i++) {
1129 		((uint64_t *)l0_mem_base)[i] = gpt_desc;
1130 	}
1131 
1132 	/* Flush updated L0 table to memory */
1133 	flush_dcache_range((uintptr_t)l0_mem_base, GPT_L0_TABLE_SIZE(gpt_config.t));
1134 
1135 	/* Stash the L0 base address once initial setup is complete */
1136 	gpt_config.plat_gpt_l0_base = l0_mem_base;
1137 
1138 	return 0;
1139 }
1140 
1141 /*
1142  * Public API that carves out PAS regions from the L0 tables and builds any L1
1143  * tables that are needed. This function ideally is run after DDR discovery and
1144  * initialization. The L0 tables must have already been initialized to GPI_ANY
1145  * when this function is called.
1146  *
1147  * This function can be called multiple times with different L1 memory ranges
1148  * and PAS regions if it is desirable to place L1 tables in different locations
1149  * in memory. (ex: you have multiple DDR banks and want to place the L1 tables
1150  * in the DDR bank that they control).
1151  *
1152  * Parameters
1153  *   pgs		PGS value to use for table generation.
1154  *   l1_mem_base	Base address of memory used for L1 tables.
1155  *   l1_mem_size	Total size of memory available for L1 tables.
1156  *   *pas_regions	Pointer to PAS regions structure array.
1157  *   pas_count		Total number of PAS regions.
1158  *
1159  * Return
1160  *   Negative Linux error code in the event of a failure, 0 for success.
1161  */
1162 int gpt_init_pas_l1_tables(gpccr_pgs_e pgs, uintptr_t l1_mem_base,
1163 			   size_t l1_mem_size, pas_region_t *pas_regions,
1164 			   unsigned int pas_count)
1165 {
1166 	int l1_gpt_cnt, ret;
1167 
1168 	/* Ensure that MMU and Data caches are enabled */
1169 	assert((read_sctlr_el3() & SCTLR_C_BIT) != 0UL);
1170 
1171 	/* PGS is needed for validate_pas_mappings so check it now */
1172 	if (pgs > GPT_PGS_MAX) {
1173 		ERROR("GPT: Invalid PGS: 0x%x\n", pgs);
1174 		return -EINVAL;
1175 	}
1176 	gpt_config.pgs = pgs;
1177 	gpt_config.p = gpt_p_lookup[pgs];
1178 
1179 	/* Make sure L0 tables have been initialized */
1180 	if (gpt_config.plat_gpt_l0_base == 0UL) {
1181 		ERROR("GPT: L0 tables must be initialized first!\n");
1182 		return -EPERM;
1183 	}
1184 
1185 	/* Check if L1 GPTs are required and how many */
1186 	l1_gpt_cnt = validate_pas_mappings(pas_regions, pas_count);
1187 	if (l1_gpt_cnt < 0) {
1188 		return l1_gpt_cnt;
1189 	}
1190 
1191 	VERBOSE("GPT: %i L1 GPTs requested\n", l1_gpt_cnt);
1192 
1193 	/* If L1 tables are needed then validate the L1 parameters */
1194 	if (l1_gpt_cnt > 0) {
1195 		ret = validate_l1_params(l1_mem_base, l1_mem_size,
1196 					(unsigned int)l1_gpt_cnt);
1197 		if (ret != 0) {
1198 			return ret;
1199 		}
1200 
1201 		/* Set up parameters for L1 table generation */
1202 		gpt_l1_tbl = l1_mem_base;
1203 	}
1204 
1205 	/* Number of L1 entries in 2MB depends on GPCCR_EL3.PGS value */
1206 	gpt_l1_cnt_2mb = (unsigned int)GPT_L1_ENTRY_COUNT_2MB(gpt_config.p);
1207 
1208 	/* Mask for the L1 index field */
1209 	gpt_l1_index_mask = GPT_L1_IDX_MASK(gpt_config.p);
1210 
1211 	INFO("GPT: Boot Configuration\n");
1212 	INFO("  PPS/T:     0x%x/%u\n", gpt_config.pps, gpt_config.t);
1213 	INFO("  PGS/P:     0x%x/%u\n", gpt_config.pgs, gpt_config.p);
1214 	INFO("  L0GPTSZ/S: 0x%x/%u\n", GPT_L0GPTSZ, GPT_S_VAL);
1215 	INFO("  PAS count: %u\n", pas_count);
1216 	INFO("  L0 base:   0x%"PRIxPTR"\n", gpt_config.plat_gpt_l0_base);
1217 
1218 	/* Generate the tables in memory */
1219 	for (unsigned int idx = 0U; idx < pas_count; idx++) {
1220 		VERBOSE("GPT: PAS[%u]: base 0x%"PRIxPTR"\tsize 0x%lx\tGPI 0x%x\ttype 0x%x\n",
1221 			idx, pas_regions[idx].base_pa, pas_regions[idx].size,
1222 			GPT_PAS_ATTR_GPI(pas_regions[idx].attrs),
1223 			GPT_PAS_ATTR_MAP_TYPE(pas_regions[idx].attrs));
1224 
1225 		/* Check if a block or table descriptor is required */
1226 		if (GPT_PAS_ATTR_MAP_TYPE(pas_regions[idx].attrs) ==
1227 		    GPT_PAS_ATTR_MAP_TYPE_BLOCK) {
1228 			generate_l0_blk_desc(&pas_regions[idx]);
1229 
1230 		} else {
1231 			generate_l0_tbl_desc(&pas_regions[idx]);
1232 		}
1233 	}
1234 
1235 	/* Flush modified L0 tables */
1236 	flush_l0_for_pas_array(pas_regions, pas_count);
1237 
1238 	/* Flush L1 tables if needed */
1239 	if (l1_gpt_cnt > 0) {
1240 		flush_dcache_range(l1_mem_base,
1241 				   GPT_L1_TABLE_SIZE(gpt_config.p) *
1242 				   (size_t)l1_gpt_cnt);
1243 	}
1244 
1245 	/* Make sure that all the entries are written to the memory */
1246 	dsbishst();
1247 	tlbipaallos();
1248 	dsb();
1249 	isb();
1250 
1251 	return 0;
1252 }
1253 
1254 /*
1255  * Public API to initialize the runtime gpt_config structure based on the values
1256  * present in the GPTBR_EL3 and GPCCR_EL3 registers. GPT initialization
1257  * typically happens in a bootloader stage prior to setting up the EL3 runtime
1258  * environment for the granule transition service so this function detects the
1259  * initialization from a previous stage. Granule protection checks must be
1260  * enabled already or this function will return an error.
1261  *
1262  * Parameters
1263  *   l1_bitlocks_base	Base address of memory for L1 tables bitlocks.
1264  *   l1_bitlocks_size	Total size of memory available for L1 tables bitlocks.
1265  *
1266  * Return
1267  *   Negative Linux error code in the event of a failure, 0 for success.
1268  */
1269 int gpt_runtime_init(uintptr_t l1_bitlocks_base, size_t l1_bitlocks_size)
1270 {
1271 	u_register_t reg;
1272 	__unused size_t locks_size;
1273 
1274 	/* Ensure that MMU and Data caches are enabled */
1275 	assert((read_sctlr_el3() & SCTLR_C_BIT) != 0UL);
1276 
1277 	/* Ensure GPC are already enabled */
1278 	if ((read_gpccr_el3() & GPCCR_GPC_BIT) == 0UL) {
1279 		ERROR("GPT: Granule protection checks are not enabled!\n");
1280 		return -EPERM;
1281 	}
1282 
1283 	/*
1284 	 * Read the L0 table address from GPTBR, we don't need the L1 base
1285 	 * address since those are included in the L0 tables as needed.
1286 	 */
1287 	reg = read_gptbr_el3();
1288 	gpt_config.plat_gpt_l0_base = ((reg >> GPTBR_BADDR_SHIFT) &
1289 				      GPTBR_BADDR_MASK) <<
1290 				      GPTBR_BADDR_VAL_SHIFT;
1291 
1292 	/* Read GPCCR to get PGS and PPS values */
1293 	reg = read_gpccr_el3();
1294 	gpt_config.pps = (reg >> GPCCR_PPS_SHIFT) & GPCCR_PPS_MASK;
1295 	gpt_config.t = gpt_t_lookup[gpt_config.pps];
1296 	gpt_config.pgs = (reg >> GPCCR_PGS_SHIFT) & GPCCR_PGS_MASK;
1297 	gpt_config.p = gpt_p_lookup[gpt_config.pgs];
1298 
1299 	/* Number of L1 entries in 2MB depends on GPCCR_EL3.PGS value */
1300 	gpt_l1_cnt_2mb = (unsigned int)GPT_L1_ENTRY_COUNT_2MB(gpt_config.p);
1301 
1302 	/* Mask for the L1 index field */
1303 	gpt_l1_index_mask = GPT_L1_IDX_MASK(gpt_config.p);
1304 
1305 #if (RME_GPT_BITLOCK_BLOCK != 0)
1306 	/*
1307 	 * Size of GPT bitlocks in bytes for the protected address space
1308 	 * with RME_GPT_BITLOCK_BLOCK * 512MB per bitlock.
1309 	 */
1310 	locks_size = GPT_PPS_ACTUAL_SIZE(gpt_config.t) /
1311 			(RME_GPT_BITLOCK_BLOCK * SZ_512M * 8U);
1312 	/*
1313 	 * If protected space size is less than the size covered
1314 	 * by 'bitlock' structure, check for a single bitlock.
1315 	 */
1316 	if (locks_size < LOCK_SIZE) {
1317 		locks_size = LOCK_SIZE;
1318 	/* Check bitlocks array size */
1319 	} else if (locks_size > l1_bitlocks_size) {
1320 		ERROR("GPT: Inadequate GPT bitlocks memory\n");
1321 		ERROR("      Expected 0x%lx bytes, got 0x%lx\n",
1322 			locks_size, l1_bitlocks_size);
1323 		return -ENOMEM;
1324 	}
1325 
1326 	gpt_bitlock = (bitlock_t *)l1_bitlocks_base;
1327 
1328 	/* Initialise GPT bitlocks */
1329 	(void)memset((void *)gpt_bitlock, 0, locks_size);
1330 
1331 	/* Flush GPT bitlocks to memory */
1332 	flush_dcache_range((uintptr_t)gpt_bitlock, locks_size);
1333 #endif /* RME_GPT_BITLOCK_BLOCK */
1334 
1335 	VERBOSE("GPT: Runtime Configuration\n");
1336 	VERBOSE("  PPS/T:     0x%x/%u\n", gpt_config.pps, gpt_config.t);
1337 	VERBOSE("  PGS/P:     0x%x/%u\n", gpt_config.pgs, gpt_config.p);
1338 	VERBOSE("  L0GPTSZ/S: 0x%x/%u\n", GPT_L0GPTSZ, GPT_S_VAL);
1339 	VERBOSE("  L0 base:   0x%"PRIxPTR"\n", gpt_config.plat_gpt_l0_base);
1340 #if (RME_GPT_BITLOCK_BLOCK != 0)
1341 	VERBOSE("  Bitlocks:  0x%"PRIxPTR"/0x%lx\n", (uintptr_t)gpt_bitlock,
1342 					locks_size);
1343 #endif
1344 	return 0;
1345 }
1346 
1347 /*
1348  * A helper to write the value (target_pas << gpi_shift) to the index of
1349  * the gpt_l1_addr.
1350  */
1351 static inline void write_gpt(uint64_t *gpt_l1_desc, uint64_t *gpt_l1_addr,
1352 			     unsigned int gpi_shift, unsigned int idx,
1353 			     unsigned int target_pas)
1354 {
1355 	*gpt_l1_desc &= ~(GPT_L1_GRAN_DESC_GPI_MASK << gpi_shift);
1356 	*gpt_l1_desc |= ((uint64_t)target_pas << gpi_shift);
1357 	gpt_l1_addr[idx] = *gpt_l1_desc;
1358 
1359 	dsboshst();
1360 }
1361 
1362 /*
1363  * Helper to retrieve the gpt_l1_* information from the base address
1364  * returned in gpi_info.
1365  */
1366 static int get_gpi_params(uint64_t base, gpi_info_t *gpi_info)
1367 {
1368 	uint64_t gpt_l0_desc, *gpt_l0_base;
1369 	__unused unsigned int block_idx;
1370 
1371 	gpt_l0_base = (uint64_t *)gpt_config.plat_gpt_l0_base;
1372 	gpt_l0_desc = gpt_l0_base[GPT_L0_IDX(base)];
1373 	if (GPT_L0_TYPE(gpt_l0_desc) != GPT_L0_TYPE_TBL_DESC) {
1374 		VERBOSE("GPT: Granule is not covered by a table descriptor!\n");
1375 		VERBOSE("      Base=0x%"PRIx64"\n", base);
1376 		return -EINVAL;
1377 	}
1378 
1379 	/* Get the table index and GPI shift from PA */
1380 	gpi_info->gpt_l1_addr = GPT_L0_TBLD_ADDR(gpt_l0_desc);
1381 	gpi_info->idx = (unsigned int)GPT_L1_INDEX(base);
1382 	gpi_info->gpi_shift = GPT_L1_GPI_IDX(gpt_config.p, base) << 2;
1383 
1384 #if (RME_GPT_BITLOCK_BLOCK != 0)
1385 	/* Block index */
1386 	block_idx = (unsigned int)(base / (RME_GPT_BITLOCK_BLOCK * SZ_512M));
1387 
1388 	/* Bitlock address and mask */
1389 	gpi_info->lock = &gpt_bitlock[block_idx / LOCK_BITS];
1390 	gpi_info->mask = 1U << (block_idx & (LOCK_BITS - 1U));
1391 #endif
1392 	return 0;
1393 }
1394 
1395 /*
1396  * Helper to retrieve the gpt_l1_desc and GPI information from gpi_info.
1397  * This function is called with bitlock or spinlock acquired.
1398  */
1399 static void read_gpi(gpi_info_t *gpi_info)
1400 {
1401 	gpi_info->gpt_l1_desc = (gpi_info->gpt_l1_addr)[gpi_info->idx];
1402 
1403 	if ((gpi_info->gpt_l1_desc & GPT_L1_TYPE_CONT_DESC_MASK) ==
1404 				 GPT_L1_TYPE_CONT_DESC) {
1405 		/* Read GPI from Contiguous descriptor */
1406 		gpi_info->gpi = (unsigned int)GPT_L1_CONT_GPI(gpi_info->gpt_l1_desc);
1407 	} else {
1408 		/* Read GPI from Granules descriptor */
1409 		gpi_info->gpi = (unsigned int)((gpi_info->gpt_l1_desc >> gpi_info->gpi_shift) &
1410 						GPT_L1_GRAN_DESC_GPI_MASK);
1411 	}
1412 }
1413 
1414 static void flush_page_to_popa(uintptr_t addr)
1415 {
1416 	size_t size = GPT_PGS_ACTUAL_SIZE(gpt_config.p);
1417 
1418 	if (is_feat_mte2_supported()) {
1419 		flush_dcache_to_popa_range_mte2(addr, size);
1420 	} else {
1421 		flush_dcache_to_popa_range(addr, size);
1422 	}
1423 }
1424 
1425 /*
1426  * Helper function to check if all L1 entries in 2MB block have
1427  * the same Granules descriptor value.
1428  *
1429  * Parameters
1430  *   base		Base address of the region to be checked
1431  *   gpi_info		Pointer to 'gpt_config_t' structure
1432  *   l1_desc		GPT Granules descriptor with all entries
1433  *			set to the same GPI.
1434  *
1435  * Return
1436  *   true if L1 all entries have the same descriptor value, false otherwise.
1437  */
1438 __unused static bool check_fuse_2mb(uint64_t base, const gpi_info_t *gpi_info,
1439 					uint64_t l1_desc)
1440 {
1441 	/* Last L1 entry index in 2MB block */
1442 	unsigned int long idx = GPT_L1_INDEX(ALIGN_2MB(base)) +
1443 						gpt_l1_cnt_2mb - 1UL;
1444 
1445 	/* Number of L1 entries in 2MB block */
1446 	unsigned int cnt = gpt_l1_cnt_2mb;
1447 
1448 	/*
1449 	 * Start check from the last L1 entry and continue until the first
1450 	 * non-matching to the passed Granules descriptor value is found.
1451 	 */
1452 	while (cnt-- != 0U) {
1453 		if (gpi_info->gpt_l1_addr[idx--] != l1_desc) {
1454 			/* Non-matching L1 entry found */
1455 			return false;
1456 		}
1457 	}
1458 
1459 	return true;
1460 }
1461 
1462 __unused static void fuse_2mb(uint64_t base, const gpi_info_t *gpi_info,
1463 				uint64_t l1_desc)
1464 {
1465 	/* L1 entry index of the start of 2MB block */
1466 	unsigned long idx_2 = GPT_L1_INDEX(ALIGN_2MB(base));
1467 
1468 	/* 2MB Contiguous descriptor */
1469 	uint64_t l1_cont_desc = GPT_L1_CONT_DESC(l1_desc, 2MB);
1470 
1471 	VERBOSE("GPT: %s(0x%"PRIxPTR" 0x%"PRIx64")\n", __func__, base, l1_desc);
1472 
1473 	fill_desc(&gpi_info->gpt_l1_addr[idx_2], l1_cont_desc, L1_QWORDS_2MB);
1474 }
1475 
1476 /*
1477  * Helper function to check if all 1st L1 entries of 2MB blocks
1478  * in 32MB have the same 2MB Contiguous descriptor value.
1479  *
1480  * Parameters
1481  *   base		Base address of the region to be checked
1482  *   gpi_info		Pointer to 'gpt_config_t' structure
1483  *   l1_desc		GPT Granules descriptor.
1484  *
1485  * Return
1486  *   true if all L1 entries have the same descriptor value, false otherwise.
1487  */
1488 __unused static bool check_fuse_32mb(uint64_t base, const gpi_info_t *gpi_info,
1489 					uint64_t l1_desc)
1490 {
1491 	/* The 1st L1 entry index of the last 2MB block in 32MB */
1492 	unsigned long idx = GPT_L1_INDEX(ALIGN_32MB(base)) +
1493 					(15UL * gpt_l1_cnt_2mb);
1494 
1495 	/* 2MB Contiguous descriptor */
1496 	uint64_t l1_cont_desc = GPT_L1_CONT_DESC(l1_desc, 2MB);
1497 
1498 	/* Number of 2MB blocks in 32MB */
1499 	unsigned int cnt = 16U;
1500 
1501 	/* Set the first L1 entry to 2MB Contiguous descriptor */
1502 	gpi_info->gpt_l1_addr[GPT_L1_INDEX(ALIGN_2MB(base))] = l1_cont_desc;
1503 
1504 	/*
1505 	 * Start check from the 1st L1 entry of the last 2MB block and
1506 	 * continue until the first non-matching to 2MB Contiguous descriptor
1507 	 * value is found.
1508 	 */
1509 	while (cnt-- != 0U) {
1510 		if (gpi_info->gpt_l1_addr[idx] != l1_cont_desc) {
1511 			/* Non-matching L1 entry found */
1512 			return false;
1513 		}
1514 		idx -= gpt_l1_cnt_2mb;
1515 	}
1516 
1517 	return true;
1518 }
1519 
1520 __unused static void fuse_32mb(uint64_t base, const gpi_info_t *gpi_info,
1521 				uint64_t l1_desc)
1522 {
1523 	/* L1 entry index of the start of 32MB block */
1524 	unsigned long idx_32 = GPT_L1_INDEX(ALIGN_32MB(base));
1525 
1526 	/* 32MB Contiguous descriptor */
1527 	uint64_t l1_cont_desc = GPT_L1_CONT_DESC(l1_desc, 32MB);
1528 
1529 	VERBOSE("GPT: %s(0x%"PRIxPTR" 0x%"PRIx64")\n", __func__, base, l1_desc);
1530 
1531 	fill_desc(&gpi_info->gpt_l1_addr[idx_32], l1_cont_desc, L1_QWORDS_32MB);
1532 }
1533 
1534 /*
1535  * Helper function to check if all 1st L1 entries of 32MB blocks
1536  * in 512MB have the same 32MB Contiguous descriptor value.
1537  *
1538  * Parameters
1539  *   base		Base address of the region to be checked
1540  *   gpi_info		Pointer to 'gpt_config_t' structure
1541  *   l1_desc		GPT Granules descriptor.
1542  *
1543  * Return
1544  *   true if all L1 entries have the same descriptor value, false otherwise.
1545  */
1546 __unused static bool check_fuse_512mb(uint64_t base, const gpi_info_t *gpi_info,
1547 					uint64_t l1_desc)
1548 {
1549 	/* The 1st L1 entry index of the last 32MB block in 512MB */
1550 	unsigned long idx = GPT_L1_INDEX(ALIGN_512MB(base)) +
1551 					(15UL * 16UL * gpt_l1_cnt_2mb);
1552 
1553 	/* 32MB Contiguous descriptor */
1554 	uint64_t l1_cont_desc = GPT_L1_CONT_DESC(l1_desc, 32MB);
1555 
1556 	/* Number of 32MB blocks in 512MB */
1557 	unsigned int cnt = 16U;
1558 
1559 	/* Set the first L1 entry to 2MB Contiguous descriptor */
1560 	gpi_info->gpt_l1_addr[GPT_L1_INDEX(ALIGN_32MB(base))] = l1_cont_desc;
1561 
1562 	/*
1563 	 * Start check from the 1st L1 entry of the last 32MB block and
1564 	 * continue until the first non-matching to 32MB Contiguous descriptor
1565 	 * value is found.
1566 	 */
1567 	while (cnt-- != 0U) {
1568 		if (gpi_info->gpt_l1_addr[idx] != l1_cont_desc) {
1569 			/* Non-matching L1 entry found */
1570 			return false;
1571 		}
1572 		idx -= 16UL * gpt_l1_cnt_2mb;
1573 	}
1574 
1575 	return true;
1576 }
1577 
1578 __unused static void fuse_512mb(uint64_t base, const gpi_info_t *gpi_info,
1579 				uint64_t l1_desc)
1580 {
1581 	/* L1 entry index of the start of 512MB block */
1582 	unsigned long idx_512 = GPT_L1_INDEX(ALIGN_512MB(base));
1583 
1584 	/* 512MB Contiguous descriptor */
1585 	uint64_t l1_cont_desc = GPT_L1_CONT_DESC(l1_desc, 512MB);
1586 
1587 	VERBOSE("GPT: %s(0x%"PRIxPTR" 0x%"PRIx64")\n", __func__, base, l1_desc);
1588 
1589 	fill_desc(&gpi_info->gpt_l1_addr[idx_512], l1_cont_desc, L1_QWORDS_512MB);
1590 }
1591 
1592 /*
1593  * Helper function to convert GPI entries in a single L1 table
1594  * from Granules to Contiguous descriptor.
1595  *
1596  * Parameters
1597  *   base		Base address of the region to be written
1598  *   gpi_info		Pointer to 'gpt_config_t' structure
1599  *   l1_desc		GPT Granules descriptor with all entries
1600  *			set to the same GPI.
1601  */
1602 __unused static void fuse_block(uint64_t base, const gpi_info_t *gpi_info,
1603 				uint64_t l1_desc)
1604 {
1605 	/* Start with check for 2MB block */
1606 	if (!check_fuse_2mb(base, gpi_info, l1_desc)) {
1607 		/* Check for 2MB fusing failed */
1608 		return;
1609 	}
1610 
1611 #if (RME_GPT_MAX_BLOCK == 2)
1612 	fuse_2mb(base, gpi_info, l1_desc);
1613 #else
1614 	/* Check for 32MB block */
1615 	if (!check_fuse_32mb(base, gpi_info, l1_desc)) {
1616 		/* Check for 32MB fusing failed, fuse to 2MB */
1617 		fuse_2mb(base, gpi_info, l1_desc);
1618 		return;
1619 	}
1620 
1621 #if (RME_GPT_MAX_BLOCK == 32)
1622 	fuse_32mb(base, gpi_info, l1_desc);
1623 #else
1624 	/* Check for 512MB block */
1625 	if (!check_fuse_512mb(base, gpi_info, l1_desc)) {
1626 		/* Check for 512MB fusing failed, fuse to 32MB */
1627 		fuse_32mb(base, gpi_info, l1_desc);
1628 		return;
1629 	}
1630 
1631 	/* Fuse to 512MB */
1632 	fuse_512mb(base, gpi_info, l1_desc);
1633 
1634 #endif	/* RME_GPT_MAX_BLOCK == 32 */
1635 #endif	/* RME_GPT_MAX_BLOCK == 2 */
1636 }
1637 
1638 /*
1639  * Helper function to convert GPI entries in a single L1 table
1640  * from Contiguous to Granules descriptor. This function updates
1641  * descriptor to Granules in passed 'gpt_config_t' structure as
1642  * the result of shuttering.
1643  *
1644  * Parameters
1645  *   base		Base address of the region to be written
1646  *   gpi_info		Pointer to 'gpt_config_t' structure
1647  *   l1_desc		GPT Granules descriptor set this range to.
1648  */
1649 __unused static void shatter_block(uint64_t base, gpi_info_t *gpi_info,
1650 				   uint64_t l1_desc)
1651 {
1652 	/* Look-up table for 2MB, 32MB and 512MB locks shattering */
1653 	static const gpt_shatter_func gpt_shatter_lookup[] = {
1654 		shatter_2mb,
1655 		shatter_32mb,
1656 		shatter_512mb
1657 	};
1658 
1659 	/* Look-up table for invalidation TLBs for 2MB, 32MB and 512MB blocks */
1660 	static const gpt_tlbi_lookup_t tlbi_lookup[] = {
1661 		{ tlbirpalos_2m, ~(SZ_2M - 1UL) },
1662 		{ tlbirpalos_32m, ~(SZ_32M - 1UL) },
1663 		{ tlbirpalos_512m, ~(SZ_512M - 1UL) }
1664 	};
1665 
1666 	/* Get shattering level from Contig field of Contiguous descriptor */
1667 	unsigned long level = GPT_L1_CONT_CONTIG(gpi_info->gpt_l1_desc) - 1UL;
1668 
1669 	/* Shatter contiguous block */
1670 	gpt_shatter_lookup[level](base, gpi_info, l1_desc);
1671 
1672 	tlbi_lookup[level].function(base & tlbi_lookup[level].mask);
1673 	dsbosh();
1674 
1675 	/*
1676 	 * Update 'gpt_config_t' structure's descriptor to Granules to reflect
1677 	 * the shattered GPI back to caller.
1678 	 */
1679 	gpi_info->gpt_l1_desc = l1_desc;
1680 }
1681 
1682 /*
1683  * This function is the granule transition delegate service. When a granule
1684  * transition request occurs it is routed to this function to have the request,
1685  * if valid, fulfilled following A1.1.1 Delegate of RME supplement.
1686  *
1687  * TODO: implement support for transitioning multiple granules at once.
1688  *
1689  * Parameters
1690  *   base		Base address of the region to transition, must be
1691  *			aligned to granule size.
1692  *   size		Size of region to transition, must be aligned to granule
1693  *			size.
1694  *   src_sec_state	Security state of the caller.
1695  *
1696  * Return
1697  *   Negative Linux error code in the event of a failure, 0 for success.
1698  */
1699 int gpt_delegate_pas(uint64_t base, size_t size, unsigned int src_sec_state)
1700 {
1701 	gpi_info_t gpi_info;
1702 	uint64_t nse, __unused l1_desc;
1703 	unsigned int target_pas;
1704 	int res;
1705 
1706 	/* Ensure that the tables have been set up before taking requests */
1707 	assert(gpt_config.plat_gpt_l0_base != 0UL);
1708 
1709 	/* Ensure that caches are enabled */
1710 	assert((read_sctlr_el3() & SCTLR_C_BIT) != 0UL);
1711 
1712 	/* See if this is a single or a range of granule transition */
1713 	if (size != GPT_PGS_ACTUAL_SIZE(gpt_config.p)) {
1714 		return -EINVAL;
1715 	}
1716 
1717 	/* Check that base and size are valid */
1718 	if ((ULONG_MAX - base) < size) {
1719 		VERBOSE("GPT: Transition request address overflow!\n");
1720 		VERBOSE("      Base=0x%"PRIx64"\n", base);
1721 		VERBOSE("      Size=0x%lx\n", size);
1722 		return -EINVAL;
1723 	}
1724 
1725 	/* Make sure base and size are valid */
1726 	if (((base & (GPT_PGS_ACTUAL_SIZE(gpt_config.p) - 1UL)) != 0UL) ||
1727 	    ((size & (GPT_PGS_ACTUAL_SIZE(gpt_config.p) - 1UL)) != 0UL) ||
1728 	    (size == 0UL) ||
1729 	    ((base + size) >= GPT_PPS_ACTUAL_SIZE(gpt_config.t))) {
1730 		VERBOSE("GPT: Invalid granule transition address range!\n");
1731 		VERBOSE("      Base=0x%"PRIx64"\n", base);
1732 		VERBOSE("      Size=0x%lx\n", size);
1733 		return -EINVAL;
1734 	}
1735 
1736 	/* Delegate request can only come from REALM or SECURE */
1737 	if ((src_sec_state != SMC_FROM_REALM) &&
1738 	    (src_sec_state != SMC_FROM_SECURE)) {
1739 		VERBOSE("GPT: Invalid caller security state 0x%x\n",
1740 							src_sec_state);
1741 		return -EINVAL;
1742 	}
1743 
1744 	if (src_sec_state == SMC_FROM_REALM) {
1745 		target_pas = GPT_GPI_REALM;
1746 		nse = (uint64_t)GPT_NSE_REALM << GPT_NSE_SHIFT;
1747 		l1_desc = GPT_L1_REALM_DESC;
1748 	} else {
1749 		target_pas = GPT_GPI_SECURE;
1750 		nse = (uint64_t)GPT_NSE_SECURE << GPT_NSE_SHIFT;
1751 		l1_desc = GPT_L1_SECURE_DESC;
1752 	}
1753 
1754 	res = get_gpi_params(base, &gpi_info);
1755 	if (res != 0) {
1756 		return res;
1757 	}
1758 
1759 	/*
1760 	 * Access to GPT is controlled by a lock to ensure that no more
1761 	 * than one CPU is allowed to make changes at any given time.
1762 	 */
1763 	GPT_LOCK;
1764 	read_gpi(&gpi_info);
1765 
1766 	/* Check that the current address is in NS state */
1767 	if (gpi_info.gpi != GPT_GPI_NS) {
1768 		VERBOSE("GPT: Only Granule in NS state can be delegated.\n");
1769 		VERBOSE("      Caller: %u, Current GPI: %u\n", src_sec_state,
1770 			gpi_info.gpi);
1771 		GPT_UNLOCK;
1772 		return -EPERM;
1773 	}
1774 
1775 #if (RME_GPT_MAX_BLOCK != 0)
1776 	/* Check for Contiguous descriptor */
1777 	if ((gpi_info.gpt_l1_desc & GPT_L1_TYPE_CONT_DESC_MASK) ==
1778 					GPT_L1_TYPE_CONT_DESC) {
1779 		shatter_block(base, &gpi_info, GPT_L1_NS_DESC);
1780 	}
1781 #endif
1782 	/*
1783 	 * In order to maintain mutual distrust between Realm and Secure
1784 	 * states, remove any data speculatively fetched into the target
1785 	 * physical address space.
1786 	 * Issue DC CIPAPA or DC_CIGDPAPA on implementations with FEAT_MTE2.
1787 	 */
1788 	flush_page_to_popa(base | nse);
1789 
1790 	write_gpt(&gpi_info.gpt_l1_desc, gpi_info.gpt_l1_addr,
1791 		  gpi_info.gpi_shift, gpi_info.idx, target_pas);
1792 
1793 	/* Ensure that all agents observe the new configuration */
1794 	tlbi_page_dsbosh(base);
1795 
1796 	nse = (uint64_t)GPT_NSE_NS << GPT_NSE_SHIFT;
1797 
1798 	/* Ensure that the scrubbed data have made it past the PoPA */
1799 	flush_page_to_popa(base | nse);
1800 
1801 #if (RME_GPT_MAX_BLOCK != 0)
1802 	if (gpi_info.gpt_l1_desc == l1_desc) {
1803 		/* Try to fuse */
1804 		fuse_block(base, &gpi_info, l1_desc);
1805 	}
1806 #endif
1807 
1808 	/* Unlock the lock to GPT */
1809 	GPT_UNLOCK;
1810 
1811 	/*
1812 	 * The isb() will be done as part of context
1813 	 * synchronization when returning to lower EL.
1814 	 */
1815 	VERBOSE("GPT: Granule 0x%"PRIx64" GPI 0x%x->0x%x\n",
1816 		base, gpi_info.gpi, target_pas);
1817 
1818 	return 0;
1819 }
1820 
1821 /*
1822  * This function is the granule transition undelegate service. When a granule
1823  * transition request occurs it is routed to this function where the request is
1824  * validated then fulfilled if possible.
1825  *
1826  * TODO: implement support for transitioning multiple granules at once.
1827  *
1828  * Parameters
1829  *   base		Base address of the region to transition, must be
1830  *			aligned to granule size.
1831  *   size		Size of region to transition, must be aligned to granule
1832  *			size.
1833  *   src_sec_state	Security state of the caller.
1834  *
1835  * Return
1836  *    Negative Linux error code in the event of a failure, 0 for success.
1837  */
1838 int gpt_undelegate_pas(uint64_t base, size_t size, unsigned int src_sec_state)
1839 {
1840 	gpi_info_t gpi_info;
1841 	uint64_t nse, __unused l1_desc;
1842 	int res;
1843 
1844 	/* Ensure that the tables have been set up before taking requests */
1845 	assert(gpt_config.plat_gpt_l0_base != 0UL);
1846 
1847 	/* Ensure that MMU and caches are enabled */
1848 	assert((read_sctlr_el3() & SCTLR_C_BIT) != 0UL);
1849 
1850 	/* See if this is a single or a range of granule transition */
1851 	if (size != GPT_PGS_ACTUAL_SIZE(gpt_config.p)) {
1852 		return -EINVAL;
1853 	}
1854 
1855 	/* Check that base and size are valid */
1856 	if ((ULONG_MAX - base) < size) {
1857 		VERBOSE("GPT: Transition request address overflow!\n");
1858 		VERBOSE("      Base=0x%"PRIx64"\n", base);
1859 		VERBOSE("      Size=0x%lx\n", size);
1860 		return -EINVAL;
1861 	}
1862 
1863 	/* Make sure base and size are valid */
1864 	if (((base & (GPT_PGS_ACTUAL_SIZE(gpt_config.p) - 1UL)) != 0UL) ||
1865 	    ((size & (GPT_PGS_ACTUAL_SIZE(gpt_config.p) - 1UL)) != 0UL) ||
1866 	    (size == 0UL) ||
1867 	    ((base + size) >= GPT_PPS_ACTUAL_SIZE(gpt_config.t))) {
1868 		VERBOSE("GPT: Invalid granule transition address range!\n");
1869 		VERBOSE("      Base=0x%"PRIx64"\n", base);
1870 		VERBOSE("      Size=0x%lx\n", size);
1871 		return -EINVAL;
1872 	}
1873 
1874 	res = get_gpi_params(base, &gpi_info);
1875 	if (res != 0) {
1876 		return res;
1877 	}
1878 
1879 	/*
1880 	 * Access to GPT is controlled by a lock to ensure that no more
1881 	 * than one CPU is allowed to make changes at any given time.
1882 	 */
1883 	GPT_LOCK;
1884 	read_gpi(&gpi_info);
1885 
1886 	/* Check that the current address is in the delegated state */
1887 	if ((src_sec_state == SMC_FROM_REALM) &&
1888 		(gpi_info.gpi == GPT_GPI_REALM)) {
1889 		l1_desc = GPT_L1_REALM_DESC;
1890 		nse = (uint64_t)GPT_NSE_REALM << GPT_NSE_SHIFT;
1891 	} else if ((src_sec_state == SMC_FROM_SECURE) &&
1892 		(gpi_info.gpi == GPT_GPI_SECURE)) {
1893 		l1_desc = GPT_L1_SECURE_DESC;
1894 		nse = (uint64_t)GPT_NSE_SECURE << GPT_NSE_SHIFT;
1895 	} else {
1896 		VERBOSE("GPT: Only Granule in REALM or SECURE state can be undelegated\n");
1897 		VERBOSE("      Caller: %u Current GPI: %u\n", src_sec_state,
1898 			gpi_info.gpi);
1899 		GPT_UNLOCK;
1900 		return -EPERM;
1901 	}
1902 
1903 #if (RME_GPT_MAX_BLOCK != 0)
1904 	/* Check for Contiguous descriptor */
1905 	if ((gpi_info.gpt_l1_desc & GPT_L1_TYPE_CONT_DESC_MASK) ==
1906 					GPT_L1_TYPE_CONT_DESC) {
1907 		shatter_block(base, &gpi_info, l1_desc);
1908 	}
1909 #endif
1910 	/*
1911 	 * In order to maintain mutual distrust between Realm and Secure
1912 	 * states, remove access now, in order to guarantee that writes
1913 	 * to the currently-accessible physical address space will not
1914 	 * later become observable.
1915 	 */
1916 	write_gpt(&gpi_info.gpt_l1_desc, gpi_info.gpt_l1_addr,
1917 		  gpi_info.gpi_shift, gpi_info.idx, GPT_GPI_NO_ACCESS);
1918 
1919 	/* Ensure that all agents observe the new NO_ACCESS configuration */
1920 	tlbi_page_dsbosh(base);
1921 
1922 	/* Ensure that the scrubbed data have made it past the PoPA */
1923 	flush_page_to_popa(base | nse);
1924 
1925 	/*
1926 	 * Remove any data loaded speculatively in NS space from before
1927 	 * the scrubbing.
1928 	 */
1929 	nse = (uint64_t)GPT_NSE_NS << GPT_NSE_SHIFT;
1930 
1931 	flush_page_to_popa(base | nse);
1932 
1933 	/* Clear existing GPI encoding and transition granule */
1934 	write_gpt(&gpi_info.gpt_l1_desc, gpi_info.gpt_l1_addr,
1935 		  gpi_info.gpi_shift, gpi_info.idx, GPT_GPI_NS);
1936 
1937 	/* Ensure that all agents observe the new NS configuration */
1938 	tlbi_page_dsbosh(base);
1939 
1940 #if (RME_GPT_MAX_BLOCK != 0)
1941 	if (gpi_info.gpt_l1_desc == GPT_L1_NS_DESC) {
1942 		/* Try to fuse */
1943 		fuse_block(base, &gpi_info, GPT_L1_NS_DESC);
1944 	}
1945 #endif
1946 	/* Unlock the lock to GPT */
1947 	GPT_UNLOCK;
1948 
1949 	/*
1950 	 * The isb() will be done as part of context
1951 	 * synchronization when returning to lower EL.
1952 	 */
1953 	VERBOSE("GPT: Granule 0x%"PRIx64" GPI 0x%x->0x%x\n",
1954 		base, gpi_info.gpi, GPT_GPI_NS);
1955 
1956 	return 0;
1957 }
1958