xref: /openbmc/linux/arch/powerpc/mm/nohash/tlb_64e.c (revision a2977c0c)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * Copyright 2008,2009 Ben Herrenschmidt <benh@kernel.crashing.org>
4  *                     IBM Corp.
5  *
6  *  Derived from arch/ppc/mm/init.c:
7  *    Copyright (C) 1995-1996 Gary Thomas (gdt@linuxppc.org)
8  *
9  *  Modifications by Paul Mackerras (PowerMac) (paulus@cs.anu.edu.au)
10  *  and Cort Dougan (PReP) (cort@cs.nmt.edu)
11  *    Copyright (C) 1996 Paul Mackerras
12  *
13  *  Derived from "arch/i386/mm/init.c"
14  *    Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
15  */
16 
17 #include <linux/kernel.h>
18 #include <linux/export.h>
19 #include <linux/mm.h>
20 #include <linux/init.h>
21 #include <linux/pagemap.h>
22 #include <linux/memblock.h>
23 
24 #include <asm/pgalloc.h>
25 #include <asm/tlbflush.h>
26 #include <asm/tlb.h>
27 #include <asm/code-patching.h>
28 #include <asm/cputhreads.h>
29 
30 #include <mm/mmu_decl.h>
31 
32 /* The variables below are currently only used on 64-bit Book3E
33  * though this will probably be made common with other nohash
34  * implementations at some point
35  */
36 static int mmu_pte_psize;	/* Page size used for PTE pages */
37 int mmu_vmemmap_psize;		/* Page size used for the virtual mem map */
38 int book3e_htw_mode;		/* HW tablewalk?  Value is PPC_HTW_* */
39 unsigned long linear_map_top;	/* Top of linear mapping */
40 
41 
42 /*
43  * Number of bytes to add to SPRN_SPRG_TLB_EXFRAME on crit/mcheck/debug
44  * exceptions.  This is used for bolted and e6500 TLB miss handlers which
45  * do not modify this SPRG in the TLB miss code; for other TLB miss handlers,
46  * this is set to zero.
47  */
48 int extlb_level_exc;
49 
50 /*
51  * Handling of virtual linear page tables or indirect TLB entries
52  * flushing when PTE pages are freed
53  */
54 void tlb_flush_pgtable(struct mmu_gather *tlb, unsigned long address)
55 {
56 	int tsize = mmu_psize_defs[mmu_pte_psize].enc;
57 
58 	if (book3e_htw_mode != PPC_HTW_NONE) {
59 		unsigned long start = address & PMD_MASK;
60 		unsigned long end = address + PMD_SIZE;
61 		unsigned long size = 1UL << mmu_psize_defs[mmu_pte_psize].shift;
62 
63 		/* This isn't the most optimal, ideally we would factor out the
64 		 * while preempt & CPU mask mucking around, or even the IPI but
65 		 * it will do for now
66 		 */
67 		while (start < end) {
68 			__flush_tlb_page(tlb->mm, start, tsize, 1);
69 			start += size;
70 		}
71 	} else {
72 		unsigned long rmask = 0xf000000000000000ul;
73 		unsigned long rid = (address & rmask) | 0x1000000000000000ul;
74 		unsigned long vpte = address & ~rmask;
75 
76 		vpte = (vpte >> (PAGE_SHIFT - 3)) & ~0xffful;
77 		vpte |= rid;
78 		__flush_tlb_page(tlb->mm, vpte, tsize, 0);
79 	}
80 }
81 
82 static void __init setup_page_sizes(void)
83 {
84 	unsigned int tlb0cfg;
85 	unsigned int eptcfg;
86 	int psize;
87 
88 #ifdef CONFIG_PPC_E500
89 	unsigned int mmucfg = mfspr(SPRN_MMUCFG);
90 	int fsl_mmu = mmu_has_feature(MMU_FTR_TYPE_FSL_E);
91 
92 	if (fsl_mmu && (mmucfg & MMUCFG_MAVN) == MMUCFG_MAVN_V1) {
93 		unsigned int tlb1cfg = mfspr(SPRN_TLB1CFG);
94 		unsigned int min_pg, max_pg;
95 
96 		min_pg = (tlb1cfg & TLBnCFG_MINSIZE) >> TLBnCFG_MINSIZE_SHIFT;
97 		max_pg = (tlb1cfg & TLBnCFG_MAXSIZE) >> TLBnCFG_MAXSIZE_SHIFT;
98 
99 		for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
100 			struct mmu_psize_def *def;
101 			unsigned int shift;
102 
103 			def = &mmu_psize_defs[psize];
104 			shift = def->shift;
105 
106 			if (shift == 0 || shift & 1)
107 				continue;
108 
109 			/* adjust to be in terms of 4^shift Kb */
110 			shift = (shift - 10) >> 1;
111 
112 			if ((shift >= min_pg) && (shift <= max_pg))
113 				def->flags |= MMU_PAGE_SIZE_DIRECT;
114 		}
115 
116 		goto out;
117 	}
118 
119 	if (fsl_mmu && (mmucfg & MMUCFG_MAVN) == MMUCFG_MAVN_V2) {
120 		u32 tlb1cfg, tlb1ps;
121 
122 		tlb0cfg = mfspr(SPRN_TLB0CFG);
123 		tlb1cfg = mfspr(SPRN_TLB1CFG);
124 		tlb1ps = mfspr(SPRN_TLB1PS);
125 		eptcfg = mfspr(SPRN_EPTCFG);
126 
127 		if ((tlb1cfg & TLBnCFG_IND) && (tlb0cfg & TLBnCFG_PT))
128 			book3e_htw_mode = PPC_HTW_E6500;
129 
130 		/*
131 		 * We expect 4K subpage size and unrestricted indirect size.
132 		 * The lack of a restriction on indirect size is a Freescale
133 		 * extension, indicated by PSn = 0 but SPSn != 0.
134 		 */
135 		if (eptcfg != 2)
136 			book3e_htw_mode = PPC_HTW_NONE;
137 
138 		for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
139 			struct mmu_psize_def *def = &mmu_psize_defs[psize];
140 
141 			if (!def->shift)
142 				continue;
143 
144 			if (tlb1ps & (1U << (def->shift - 10))) {
145 				def->flags |= MMU_PAGE_SIZE_DIRECT;
146 
147 				if (book3e_htw_mode && psize == MMU_PAGE_2M)
148 					def->flags |= MMU_PAGE_SIZE_INDIRECT;
149 			}
150 		}
151 
152 		goto out;
153 	}
154 #endif
155 out:
156 	/* Cleanup array and print summary */
157 	pr_info("MMU: Supported page sizes\n");
158 	for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
159 		struct mmu_psize_def *def = &mmu_psize_defs[psize];
160 		const char *__page_type_names[] = {
161 			"unsupported",
162 			"direct",
163 			"indirect",
164 			"direct & indirect"
165 		};
166 		if (def->flags == 0) {
167 			def->shift = 0;
168 			continue;
169 		}
170 		pr_info("  %8ld KB as %s\n", 1ul << (def->shift - 10),
171 			__page_type_names[def->flags & 0x3]);
172 	}
173 }
174 
175 static void __init setup_mmu_htw(void)
176 {
177 	/*
178 	 * If we want to use HW tablewalk, enable it by patching the TLB miss
179 	 * handlers to branch to the one dedicated to it.
180 	 */
181 
182 	switch (book3e_htw_mode) {
183 #ifdef CONFIG_PPC_E500
184 	case PPC_HTW_E6500:
185 		extlb_level_exc = EX_TLB_SIZE;
186 		patch_exception(0x1c0, exc_data_tlb_miss_e6500_book3e);
187 		patch_exception(0x1e0, exc_instruction_tlb_miss_e6500_book3e);
188 		break;
189 #endif
190 	}
191 	pr_info("MMU: Book3E HW tablewalk %s\n",
192 		book3e_htw_mode != PPC_HTW_NONE ? "enabled" : "not supported");
193 }
194 
195 /*
196  * Early initialization of the MMU TLB code
197  */
198 static void early_init_this_mmu(void)
199 {
200 	unsigned int mas4;
201 
202 	/* Set MAS4 based on page table setting */
203 
204 	mas4 = 0x4 << MAS4_WIMGED_SHIFT;
205 	switch (book3e_htw_mode) {
206 	case PPC_HTW_E6500:
207 		mas4 |= MAS4_INDD;
208 		mas4 |= BOOK3E_PAGESZ_2M << MAS4_TSIZED_SHIFT;
209 		mas4 |= MAS4_TLBSELD(1);
210 		mmu_pte_psize = MMU_PAGE_2M;
211 		break;
212 
213 	case PPC_HTW_NONE:
214 		mas4 |=	BOOK3E_PAGESZ_4K << MAS4_TSIZED_SHIFT;
215 		mmu_pte_psize = mmu_virtual_psize;
216 		break;
217 	}
218 	mtspr(SPRN_MAS4, mas4);
219 
220 #ifdef CONFIG_PPC_E500
221 	if (mmu_has_feature(MMU_FTR_TYPE_FSL_E)) {
222 		unsigned int num_cams;
223 		bool map = true;
224 
225 		/* use a quarter of the TLBCAM for bolted linear map */
226 		num_cams = (mfspr(SPRN_TLB1CFG) & TLBnCFG_N_ENTRY) / 4;
227 
228 		/*
229 		 * Only do the mapping once per core, or else the
230 		 * transient mapping would cause problems.
231 		 */
232 #ifdef CONFIG_SMP
233 		if (hweight32(get_tensr()) > 1)
234 			map = false;
235 #endif
236 
237 		if (map)
238 			linear_map_top = map_mem_in_cams(linear_map_top,
239 							 num_cams, false, true);
240 	}
241 #endif
242 
243 	/* A sync won't hurt us after mucking around with
244 	 * the MMU configuration
245 	 */
246 	mb();
247 }
248 
249 static void __init early_init_mmu_global(void)
250 {
251 	/* XXX This should be decided at runtime based on supported
252 	 * page sizes in the TLB, but for now let's assume 16M is
253 	 * always there and a good fit (which it probably is)
254 	 *
255 	 * Freescale booke only supports 4K pages in TLB0, so use that.
256 	 */
257 	if (mmu_has_feature(MMU_FTR_TYPE_FSL_E))
258 		mmu_vmemmap_psize = MMU_PAGE_4K;
259 	else
260 		mmu_vmemmap_psize = MMU_PAGE_16M;
261 
262 	/* XXX This code only checks for TLB 0 capabilities and doesn't
263 	 *     check what page size combos are supported by the HW. It
264 	 *     also doesn't handle the case where a separate array holds
265 	 *     the IND entries from the array loaded by the PT.
266 	 */
267 	/* Look for supported page sizes */
268 	setup_page_sizes();
269 
270 	/* Look for HW tablewalk support */
271 	setup_mmu_htw();
272 
273 #ifdef CONFIG_PPC_E500
274 	if (mmu_has_feature(MMU_FTR_TYPE_FSL_E)) {
275 		if (book3e_htw_mode == PPC_HTW_NONE) {
276 			extlb_level_exc = EX_TLB_SIZE;
277 			patch_exception(0x1c0, exc_data_tlb_miss_bolted_book3e);
278 			patch_exception(0x1e0,
279 				exc_instruction_tlb_miss_bolted_book3e);
280 		}
281 	}
282 #endif
283 
284 	/* Set the global containing the top of the linear mapping
285 	 * for use by the TLB miss code
286 	 */
287 	linear_map_top = memblock_end_of_DRAM();
288 
289 	ioremap_bot = IOREMAP_BASE;
290 }
291 
292 static void __init early_mmu_set_memory_limit(void)
293 {
294 #ifdef CONFIG_PPC_E500
295 	if (mmu_has_feature(MMU_FTR_TYPE_FSL_E)) {
296 		/*
297 		 * Limit memory so we dont have linear faults.
298 		 * Unlike memblock_set_current_limit, which limits
299 		 * memory available during early boot, this permanently
300 		 * reduces the memory available to Linux.  We need to
301 		 * do this because highmem is not supported on 64-bit.
302 		 */
303 		memblock_enforce_memory_limit(linear_map_top);
304 	}
305 #endif
306 
307 	memblock_set_current_limit(linear_map_top);
308 }
309 
310 /* boot cpu only */
311 void __init early_init_mmu(void)
312 {
313 	early_init_mmu_global();
314 	early_init_this_mmu();
315 	early_mmu_set_memory_limit();
316 }
317 
318 void early_init_mmu_secondary(void)
319 {
320 	early_init_this_mmu();
321 }
322 
323 void setup_initial_memory_limit(phys_addr_t first_memblock_base,
324 				phys_addr_t first_memblock_size)
325 {
326 	/* On non-FSL Embedded 64-bit, we adjust the RMA size to match
327 	 * the bolted TLB entry. We know for now that only 1G
328 	 * entries are supported though that may eventually
329 	 * change.
330 	 *
331 	 * on FSL Embedded 64-bit, usually all RAM is bolted, but with
332 	 * unusual memory sizes it's possible for some RAM to not be mapped
333 	 * (such RAM is not used at all by Linux, since we don't support
334 	 * highmem on 64-bit).  We limit ppc64_rma_size to what would be
335 	 * mappable if this memblock is the only one.  Additional memblocks
336 	 * can only increase, not decrease, the amount that ends up getting
337 	 * mapped.  We still limit max to 1G even if we'll eventually map
338 	 * more.  This is due to what the early init code is set up to do.
339 	 *
340 	 * We crop it to the size of the first MEMBLOCK to
341 	 * avoid going over total available memory just in case...
342 	 */
343 #ifdef CONFIG_PPC_E500
344 	if (early_mmu_has_feature(MMU_FTR_TYPE_FSL_E)) {
345 		unsigned long linear_sz;
346 		unsigned int num_cams;
347 
348 		/* use a quarter of the TLBCAM for bolted linear map */
349 		num_cams = (mfspr(SPRN_TLB1CFG) & TLBnCFG_N_ENTRY) / 4;
350 
351 		linear_sz = map_mem_in_cams(first_memblock_size, num_cams,
352 					    true, true);
353 
354 		ppc64_rma_size = min_t(u64, linear_sz, 0x40000000);
355 	} else
356 #endif
357 		ppc64_rma_size = min_t(u64, first_memblock_size, 0x40000000);
358 
359 	/* Finally limit subsequent allocations */
360 	memblock_set_current_limit(first_memblock_base + ppc64_rma_size);
361 }
362