2 * PowerPC memory management structures
4 * Dave Engebretsen & Mike Corrigan <{engebret|mikejc}@us.ibm.com>
7 * This program is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU General Public License
9 * as published by the Free Software Foundation; either version
10 * 2 of the License, or (at your option) any later version.
16 #include <linux/config.h>
17 #include <asm/asm-compat.h>
24 #define STE_ESID_V 0x80
25 #define STE_ESID_KS 0x20
26 #define STE_ESID_KP 0x10
27 #define STE_ESID_N 0x08
29 #define STE_VSID_SHIFT 12
31 /* Location of cpu0's segment table */
32 #define STAB0_PAGE 0x6
33 #define STAB0_PHYS_ADDR (STAB0_PAGE<<12)
36 extern char initial_stab[];
37 #endif /* ! __ASSEMBLY */
43 #define SLB_NUM_BOLTED 3
44 #define SLB_CACHE_ENTRIES 8
46 /* Bits in the SLB ESID word */
47 #define SLB_ESID_V ASM_CONST(0x0000000008000000) /* valid */
49 /* Bits in the SLB VSID word */
50 #define SLB_VSID_SHIFT 12
51 #define SLB_VSID_B ASM_CONST(0xc000000000000000)
52 #define SLB_VSID_B_256M ASM_CONST(0x0000000000000000)
53 #define SLB_VSID_B_1T ASM_CONST(0x4000000000000000)
54 #define SLB_VSID_KS ASM_CONST(0x0000000000000800)
55 #define SLB_VSID_KP ASM_CONST(0x0000000000000400)
56 #define SLB_VSID_N ASM_CONST(0x0000000000000200) /* no-execute */
57 #define SLB_VSID_L ASM_CONST(0x0000000000000100)
58 #define SLB_VSID_C ASM_CONST(0x0000000000000080) /* class */
59 #define SLB_VSID_LP ASM_CONST(0x0000000000000030)
60 #define SLB_VSID_LP_00 ASM_CONST(0x0000000000000000)
61 #define SLB_VSID_LP_01 ASM_CONST(0x0000000000000010)
62 #define SLB_VSID_LP_10 ASM_CONST(0x0000000000000020)
63 #define SLB_VSID_LP_11 ASM_CONST(0x0000000000000030)
64 #define SLB_VSID_LLP (SLB_VSID_L|SLB_VSID_LP)
66 #define SLB_VSID_KERNEL (SLB_VSID_KP)
67 #define SLB_VSID_USER (SLB_VSID_KP|SLB_VSID_KS|SLB_VSID_C)
69 #define SLBIE_C (0x08000000)
75 #define HPTES_PER_GROUP 8
77 #define HPTE_V_AVPN_SHIFT 7
78 #define HPTE_V_AVPN ASM_CONST(0xffffffffffffff80)
79 #define HPTE_V_AVPN_VAL(x) (((x) & HPTE_V_AVPN) >> HPTE_V_AVPN_SHIFT)
80 #define HPTE_V_COMPARE(x,y) (!(((x) ^ (y)) & HPTE_V_AVPN))
81 #define HPTE_V_BOLTED ASM_CONST(0x0000000000000010)
82 #define HPTE_V_LOCK ASM_CONST(0x0000000000000008)
83 #define HPTE_V_LARGE ASM_CONST(0x0000000000000004)
84 #define HPTE_V_SECONDARY ASM_CONST(0x0000000000000002)
85 #define HPTE_V_VALID ASM_CONST(0x0000000000000001)
87 #define HPTE_R_PP0 ASM_CONST(0x8000000000000000)
88 #define HPTE_R_TS ASM_CONST(0x4000000000000000)
89 #define HPTE_R_RPN_SHIFT 12
90 #define HPTE_R_RPN ASM_CONST(0x3ffffffffffff000)
91 #define HPTE_R_FLAGS ASM_CONST(0x00000000000003ff)
92 #define HPTE_R_PP ASM_CONST(0x0000000000000003)
93 #define HPTE_R_N ASM_CONST(0x0000000000000004)
95 /* Values for PP (assumes Ks=0, Kp=1) */
96 /* pp0 will always be 0 for linux */
97 #define PP_RWXX 0 /* Supervisor read/write, User none */
98 #define PP_RWRX 1 /* Supervisor read/write, User read */
99 #define PP_RWRW 2 /* Supervisor read/write, User read/write */
100 #define PP_RXRX 3 /* Supervisor read, User read */
109 extern hpte_t *htab_address;
110 extern unsigned long htab_hash_mask;
113 * Page size definition
115 * shift : is the "PAGE_SHIFT" value for that page size
116 * sllp : is a bit mask with the value of SLB L || LP to be or'ed
117 * directly to a slbmte "vsid" value
118 * penc : is the HPTE encoding mask for the "LP" field:
123 unsigned int shift; /* number of bits */
124 unsigned int penc; /* HPTE encoding */
125 unsigned int tlbiel; /* tlbiel supported for that page size */
126 unsigned long avpnm; /* bits to mask out in AVPN in the HPTE */
127 unsigned long sllp; /* SLB L||LP (exact mask to use in slbmte) */
130 #endif /* __ASSEMBLY__ */
133 * The kernel use the constants below to index in the page sizes array.
134 * The use of fixed constants for this purpose is better for performances
135 * of the low level hash refill handlers.
137 * A non supported page size has a "shift" field set to 0
139 * Any new page size being implemented can get a new entry in here. Whether
140 * the kernel will use it or not is a different matter though. The actual page
141 * size used by hugetlbfs is not defined here and may be made variable
144 #define MMU_PAGE_4K 0 /* 4K */
145 #define MMU_PAGE_64K 1 /* 64K */
146 #define MMU_PAGE_64K_AP 2 /* 64K Admixed (in a 4K segment) */
147 #define MMU_PAGE_1M 3 /* 1M */
148 #define MMU_PAGE_16M 4 /* 16M */
149 #define MMU_PAGE_16G 5 /* 16G */
150 #define MMU_PAGE_COUNT 6
155 * The current system page sizes
157 extern struct mmu_psize_def mmu_psize_defs[MMU_PAGE_COUNT];
158 extern int mmu_linear_psize;
159 extern int mmu_virtual_psize;
161 #ifdef CONFIG_HUGETLB_PAGE
163 * The page size index of the huge pages for use by hugetlbfs
165 extern int mmu_huge_psize;
167 #endif /* CONFIG_HUGETLB_PAGE */
170 * This function sets the AVPN and L fields of the HPTE appropriately
173 static inline unsigned long hpte_encode_v(unsigned long va, int psize)
176 v = (va >> 23) & ~(mmu_psize_defs[psize].avpnm);
177 v <<= HPTE_V_AVPN_SHIFT;
178 if (psize != MMU_PAGE_4K)
184 * This function sets the ARPN, and LP fields of the HPTE appropriately
185 * for the page size. We assume the pa is already "clean" that is properly
186 * aligned for the requested page size
188 static inline unsigned long hpte_encode_r(unsigned long pa, int psize)
192 /* A 4K page needs no special encoding */
193 if (psize == MMU_PAGE_4K)
194 return pa & HPTE_R_RPN;
196 unsigned int penc = mmu_psize_defs[psize].penc;
197 unsigned int shift = mmu_psize_defs[psize].shift;
198 return (pa & ~((1ul << shift) - 1)) | (penc << 12);
204 * This hashes a virtual address for a 256Mb segment only for now
207 static inline unsigned long hpt_hash(unsigned long va, unsigned int shift)
209 return ((va >> 28) & 0x7fffffffffUL) ^ ((va & 0x0fffffffUL) >> shift);
212 extern int __hash_page_4K(unsigned long ea, unsigned long access,
213 unsigned long vsid, pte_t *ptep, unsigned long trap,
215 extern int __hash_page_64K(unsigned long ea, unsigned long access,
216 unsigned long vsid, pte_t *ptep, unsigned long trap,
219 extern int hash_huge_page(struct mm_struct *mm, unsigned long access,
220 unsigned long ea, unsigned long vsid, int local);
222 extern void htab_finish_init(void);
223 extern int htab_bolt_mapping(unsigned long vstart, unsigned long vend,
224 unsigned long pstart, unsigned long mode,
227 extern void htab_initialize(void);
228 extern void htab_initialize_secondary(void);
229 extern void hpte_init_native(void);
230 extern void hpte_init_lpar(void);
231 extern void hpte_init_iSeries(void);
232 extern void mm_init_ppc64(void);
234 extern long pSeries_lpar_hpte_insert(unsigned long hpte_group,
235 unsigned long va, unsigned long prpn,
236 unsigned long rflags,
237 unsigned long vflags, int psize);
239 extern long native_hpte_insert(unsigned long hpte_group,
240 unsigned long va, unsigned long prpn,
241 unsigned long rflags,
242 unsigned long vflags, int psize);
244 extern long iSeries_hpte_insert(unsigned long hpte_group,
245 unsigned long va, unsigned long prpn,
246 unsigned long rflags,
247 unsigned long vflags, int psize);
249 extern void stabs_alloc(void);
250 extern void slb_initialize(void);
251 extern void stab_initialize(unsigned long stab);
253 #endif /* __ASSEMBLY__ */
258 * We first generate a 36-bit "proto-VSID". For kernel addresses this
259 * is equal to the ESID, for user addresses it is:
260 * (context << 15) | (esid & 0x7fff)
262 * The two forms are distinguishable because the top bit is 0 for user
263 * addresses, whereas the top two bits are 1 for kernel addresses.
264 * Proto-VSIDs with the top two bits equal to 0b10 are reserved for
267 * The proto-VSIDs are then scrambled into real VSIDs with the
268 * multiplicative hash:
270 * VSID = (proto-VSID * VSID_MULTIPLIER) % VSID_MODULUS
271 * where VSID_MULTIPLIER = 268435399 = 0xFFFFFC7
272 * VSID_MODULUS = 2^36-1 = 0xFFFFFFFFF
274 * This scramble is only well defined for proto-VSIDs below
275 * 0xFFFFFFFFF, so both proto-VSID and actual VSID 0xFFFFFFFFF are
276 * reserved. VSID_MULTIPLIER is prime, so in particular it is
277 * co-prime to VSID_MODULUS, making this a 1:1 scrambling function.
278 * Because the modulus is 2^n-1 we can compute it efficiently without
279 * a divide or extra multiply (see below).
281 * This scheme has several advantages over older methods:
283 * - We have VSIDs allocated for every kernel address
284 * (i.e. everything above 0xC000000000000000), except the very top
285 * segment, which simplifies several things.
287 * - We allow for 15 significant bits of ESID and 20 bits of
288 * context for user addresses. i.e. 8T (43 bits) of address space for
289 * up to 1M contexts (although the page table structure and context
290 * allocation will need changes to take advantage of this).
292 * - The scramble function gives robust scattering in the hash
293 * table (at least based on some initial results). The previous
294 * method was more susceptible to pathological cases giving excessive
298 * WARNING - If you change these you must make sure the asm
299 * implementations in slb_allocate (slb_low.S), do_stab_bolted
300 * (head.S) and ASM_VSID_SCRAMBLE (below) are changed accordingly.
302 * You'll also need to change the precomputed VSID values in head.S
303 * which are used by the iSeries firmware.
306 #define VSID_MULTIPLIER ASM_CONST(200730139) /* 28-bit prime */
308 #define VSID_MODULUS ((1UL<<VSID_BITS)-1)
310 #define CONTEXT_BITS 19
311 #define USER_ESID_BITS 16
313 #define USER_VSID_RANGE (1UL << (USER_ESID_BITS + SID_SHIFT))
316 * This macro generates asm code to compute the VSID scramble
317 * function. Used in slb_allocate() and do_stab_bolted. The function
318 * computed is: (protovsid*VSID_MULTIPLIER) % VSID_MODULUS
320 * rt = register continaing the proto-VSID and into which the
321 * VSID will be stored
322 * rx = scratch register (clobbered)
324 * - rt and rx must be different registers
325 * - The answer will end up in the low 36 bits of rt. The higher
326 * bits may contain other garbage, so you may need to mask the
329 #define ASM_VSID_SCRAMBLE(rt, rx) \
330 lis rx,VSID_MULTIPLIER@h; \
331 ori rx,rx,VSID_MULTIPLIER@l; \
332 mulld rt,rt,rx; /* rt = rt * MULTIPLIER */ \
334 srdi rx,rt,VSID_BITS; \
335 clrldi rt,rt,(64-VSID_BITS); \
336 add rt,rt,rx; /* add high and low bits */ \
337 /* Now, r3 == VSID (mod 2^36-1), and lies between 0 and \
338 * 2^36-1+2^28-1. That in particular means that if r3 >= \
339 * 2^36-1, then r3+1 has the 2^36 bit set. So, if r3+1 has \
340 * the bit clear, r3 already has the answer we want, if it \
341 * doesn't, the answer is the low 36 bits of r3+1. So in all \
342 * cases the answer is the low 36 bits of (r3 + ((r3+1) >> 36))*/\
344 srdi rx,rx,VSID_BITS; /* extract 2^36 bit */ \
350 typedef unsigned long mm_context_id_t;
354 #ifdef CONFIG_HUGETLB_PAGE
355 u16 low_htlb_areas, high_htlb_areas;
360 static inline unsigned long vsid_scramble(unsigned long protovsid)
363 /* The code below is equivalent to this function for arguments
364 * < 2^VSID_BITS, which is all this should ever be called
365 * with. However gcc is not clever enough to compute the
366 * modulus (2^n-1) without a second multiply. */
367 return ((protovsid * VSID_MULTIPLIER) % VSID_MODULUS);
371 x = protovsid * VSID_MULTIPLIER;
372 x = (x >> VSID_BITS) + (x & VSID_MODULUS);
373 return (x + ((x+1) >> VSID_BITS)) & VSID_MODULUS;
377 /* This is only valid for addresses >= KERNELBASE */
378 static inline unsigned long get_kernel_vsid(unsigned long ea)
380 return vsid_scramble(ea >> SID_SHIFT);
383 /* This is only valid for user addresses (which are below 2^41) */
384 static inline unsigned long get_vsid(unsigned long context, unsigned long ea)
386 return vsid_scramble((context << USER_ESID_BITS)
387 | (ea >> SID_SHIFT));
390 #define VSID_SCRAMBLE(pvsid) (((pvsid) * VSID_MULTIPLIER) % VSID_MODULUS)
391 #define KERNEL_VSID(ea) VSID_SCRAMBLE(GET_ESID(ea))
393 #endif /* __ASSEMBLY */
395 #endif /* _PPC64_MMU_H_ */