1 /* SPDX-License-Identifier: GPL-2.0 */
2 #ifndef _ASM_POWERPC_BOOK3S_64_HASH_64K_H
3 #define _ASM_POWERPC_BOOK3S_64_HASH_64K_H
4 
5 #define H_PTE_INDEX_SIZE   8  // size: 8B <<  8 = 2KB, maps 2^8  x 64KB = 16MB
6 #define H_PMD_INDEX_SIZE  10  // size: 8B << 10 = 8KB, maps 2^10 x 16MB = 16GB
7 #define H_PUD_INDEX_SIZE  10  // size: 8B << 10 = 8KB, maps 2^10 x 16GB = 16TB
8 #define H_PGD_INDEX_SIZE   8  // size: 8B <<  8 = 2KB, maps 2^8  x 16TB =  4PB
9 
10 /*
11  * If we store section details in page->flags we can't increase the MAX_PHYSMEM_BITS
12  * if we increase SECTIONS_WIDTH we will not store node details in page->flags and
13  * page_to_nid does a page->section->node lookup
14  * Hence only increase for VMEMMAP. Further depending on SPARSEMEM_EXTREME reduce
15  * memory requirements with large number of sections.
16  * 51 bits is the max physical real address on POWER9
17  */
18 #if defined(CONFIG_SPARSEMEM_VMEMMAP) && defined(CONFIG_SPARSEMEM_EXTREME)
19 #define H_MAX_PHYSMEM_BITS	51
20 #else
21 #define H_MAX_PHYSMEM_BITS	46
22 #endif
23 
24 /*
25  * Each context is 512TB size. SLB miss for first context/default context
26  * is handled in the hotpath.
27  */
28 #define MAX_EA_BITS_PER_CONTEXT		49
29 #define REGION_SHIFT		MAX_EA_BITS_PER_CONTEXT
30 
31 /*
32  * We use one context for each MAP area.
33  */
34 #define H_KERN_MAP_SIZE		(1UL << MAX_EA_BITS_PER_CONTEXT)
35 
36 /*
37  * Define the address range of the kernel non-linear virtual area
38  * 2PB
39  */
40 #define H_KERN_VIRT_START	ASM_CONST(0xc008000000000000)
41 
42 /*
43  * 64k aligned address free up few of the lower bits of RPN for us
44  * We steal that here. For more deatils look at pte_pfn/pfn_pte()
45  */
46 #define H_PAGE_COMBO	_RPAGE_RPN0 /* this is a combo 4k page */
47 #define H_PAGE_4K_PFN	_RPAGE_RPN1 /* PFN is for a single 4k page */
48 #define H_PAGE_BUSY	_RPAGE_RSV1     /* software: PTE & hash are busy */
49 #define H_PAGE_HASHPTE	_RPAGE_RPN43	/* PTE has associated HPTE */
50 
51 /* memory key bits. */
52 #define H_PTE_PKEY_BIT4		_RPAGE_PKEY_BIT4
53 #define H_PTE_PKEY_BIT3		_RPAGE_PKEY_BIT3
54 #define H_PTE_PKEY_BIT2		_RPAGE_PKEY_BIT2
55 #define H_PTE_PKEY_BIT1		_RPAGE_PKEY_BIT1
56 #define H_PTE_PKEY_BIT0		_RPAGE_PKEY_BIT0
57 
58 /*
59  * We need to differentiate between explicit huge page and THP huge
60  * page, since THP huge page also need to track real subpage details
61  */
62 #define H_PAGE_THP_HUGE  H_PAGE_4K_PFN
63 
64 /* PTE flags to conserve for HPTE identification */
65 #define _PAGE_HPTEFLAGS (H_PAGE_BUSY | H_PAGE_HASHPTE | H_PAGE_COMBO)
66 /*
67  * We use a 2K PTE page fragment and another 2K for storing
68  * real_pte_t hash index
69  * 8 bytes per each pte entry and another 8 bytes for storing
70  * slot details.
71  */
72 #define H_PTE_FRAG_SIZE_SHIFT  (H_PTE_INDEX_SIZE + 3 + 1)
73 #define H_PTE_FRAG_NR	(PAGE_SIZE >> H_PTE_FRAG_SIZE_SHIFT)
74 
75 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLB_PAGE)
76 #define H_PMD_FRAG_SIZE_SHIFT  (H_PMD_INDEX_SIZE + 3 + 1)
77 #else
78 #define H_PMD_FRAG_SIZE_SHIFT  (H_PMD_INDEX_SIZE + 3)
79 #endif
80 #define H_PMD_FRAG_NR	(PAGE_SIZE >> H_PMD_FRAG_SIZE_SHIFT)
81 
82 #ifndef __ASSEMBLY__
83 #include <asm/errno.h>
84 
85 /*
86  * With 64K pages on hash table, we have a special PTE format that
87  * uses a second "half" of the page table to encode sub-page information
88  * in order to deal with 64K made of 4K HW pages. Thus we override the
89  * generic accessors and iterators here
90  */
91 #define __real_pte __real_pte
__real_pte(pte_t pte,pte_t * ptep,int offset)92 static inline real_pte_t __real_pte(pte_t pte, pte_t *ptep, int offset)
93 {
94 	real_pte_t rpte;
95 	unsigned long *hidxp;
96 
97 	rpte.pte = pte;
98 
99 	/*
100 	 * Ensure that we do not read the hidx before we read the PTE. Because
101 	 * the writer side is expected to finish writing the hidx first followed
102 	 * by the PTE, by using smp_wmb(). pte_set_hash_slot() ensures that.
103 	 */
104 	smp_rmb();
105 
106 	hidxp = (unsigned long *)(ptep + offset);
107 	rpte.hidx = *hidxp;
108 	return rpte;
109 }
110 
111 /*
112  * shift the hidx representation by one-modulo-0xf; i.e hidx 0 is respresented
113  * as 1, 1 as 2,... , and 0xf as 0.  This convention lets us represent a
114  * invalid hidx 0xf with a 0x0 bit value. PTEs are anyway zero'd when
115  * allocated. We dont have to zero them gain; thus save on the initialization.
116  */
117 #define HIDX_UNSHIFT_BY_ONE(x) ((x + 0xfUL) & 0xfUL) /* shift backward by one */
118 #define HIDX_SHIFT_BY_ONE(x) ((x + 0x1UL) & 0xfUL)   /* shift forward by one */
119 #define HIDX_BITS(x, index)  (x << (index << 2))
120 #define BITS_TO_HIDX(x, index)  ((x >> (index << 2)) & 0xfUL)
121 #define INVALID_RPTE_HIDX  0x0UL
122 
__rpte_to_hidx(real_pte_t rpte,unsigned long index)123 static inline unsigned long __rpte_to_hidx(real_pte_t rpte, unsigned long index)
124 {
125 	return HIDX_UNSHIFT_BY_ONE(BITS_TO_HIDX(rpte.hidx, index));
126 }
127 
128 /*
129  * Commit the hidx and return PTE bits that needs to be modified. The caller is
130  * expected to modify the PTE bits accordingly and commit the PTE to memory.
131  */
pte_set_hidx(pte_t * ptep,real_pte_t rpte,unsigned int subpg_index,unsigned long hidx,int offset)132 static inline unsigned long pte_set_hidx(pte_t *ptep, real_pte_t rpte,
133 					 unsigned int subpg_index,
134 					 unsigned long hidx, int offset)
135 {
136 	unsigned long *hidxp = (unsigned long *)(ptep + offset);
137 
138 	rpte.hidx &= ~HIDX_BITS(0xfUL, subpg_index);
139 	*hidxp = rpte.hidx  | HIDX_BITS(HIDX_SHIFT_BY_ONE(hidx), subpg_index);
140 
141 	/*
142 	 * Anyone reading PTE must ensure hidx bits are read after reading the
143 	 * PTE by using the read-side barrier smp_rmb(). __real_pte() can be
144 	 * used for that.
145 	 */
146 	smp_wmb();
147 
148 	/* No PTE bits to be modified, return 0x0UL */
149 	return 0x0UL;
150 }
151 
152 #define __rpte_to_pte(r)	((r).pte)
153 extern bool __rpte_sub_valid(real_pte_t rpte, unsigned long index);
154 /*
155  * Trick: we set __end to va + 64k, which happens works for
156  * a 16M page as well as we want only one iteration
157  */
158 #define pte_iterate_hashed_subpages(rpte, psize, vpn, index, shift)	\
159 	do {								\
160 		unsigned long __end = vpn + (1UL << (PAGE_SHIFT - VPN_SHIFT));	\
161 		unsigned __split = (psize == MMU_PAGE_4K ||		\
162 				    psize == MMU_PAGE_64K_AP);		\
163 		shift = mmu_psize_defs[psize].shift;			\
164 		for (index = 0; vpn < __end; index++,			\
165 			     vpn += (1L << (shift - VPN_SHIFT))) {	\
166 		if (!__split || __rpte_sub_valid(rpte, index))
167 
168 #define pte_iterate_hashed_end()  } } while(0)
169 
170 #define pte_pagesize_index(mm, addr, pte)	\
171 	(((pte) & H_PAGE_COMBO)? MMU_PAGE_4K: MMU_PAGE_64K)
172 
173 extern int remap_pfn_range(struct vm_area_struct *, unsigned long addr,
174 			   unsigned long pfn, unsigned long size, pgprot_t);
hash__remap_4k_pfn(struct vm_area_struct * vma,unsigned long addr,unsigned long pfn,pgprot_t prot)175 static inline int hash__remap_4k_pfn(struct vm_area_struct *vma, unsigned long addr,
176 				 unsigned long pfn, pgprot_t prot)
177 {
178 	if (pfn > (PTE_RPN_MASK >> PAGE_SHIFT)) {
179 		WARN(1, "remap_4k_pfn called with wrong pfn value\n");
180 		return -EINVAL;
181 	}
182 	return remap_pfn_range(vma, addr, pfn, PAGE_SIZE,
183 			       __pgprot(pgprot_val(prot) | H_PAGE_4K_PFN));
184 }
185 
186 #define H_PTE_TABLE_SIZE	PTE_FRAG_SIZE
187 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined (CONFIG_HUGETLB_PAGE)
188 #define H_PMD_TABLE_SIZE	((sizeof(pmd_t) << PMD_INDEX_SIZE) + \
189 				 (sizeof(unsigned long) << PMD_INDEX_SIZE))
190 #else
191 #define H_PMD_TABLE_SIZE	(sizeof(pmd_t) << PMD_INDEX_SIZE)
192 #endif
193 #ifdef CONFIG_HUGETLB_PAGE
194 #define H_PUD_TABLE_SIZE	((sizeof(pud_t) << PUD_INDEX_SIZE) +	\
195 				 (sizeof(unsigned long) << PUD_INDEX_SIZE))
196 #else
197 #define H_PUD_TABLE_SIZE	(sizeof(pud_t) << PUD_INDEX_SIZE)
198 #endif
199 #define H_PGD_TABLE_SIZE	(sizeof(pgd_t) << PGD_INDEX_SIZE)
200 
201 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
get_hpte_slot_array(pmd_t * pmdp)202 static inline char *get_hpte_slot_array(pmd_t *pmdp)
203 {
204 	/*
205 	 * The hpte hindex is stored in the pgtable whose address is in the
206 	 * second half of the PMD
207 	 *
208 	 * Order this load with the test for pmd_trans_huge in the caller
209 	 */
210 	smp_rmb();
211 	return *(char **)(pmdp + PTRS_PER_PMD);
212 
213 
214 }
215 /*
216  * The linux hugepage PMD now include the pmd entries followed by the address
217  * to the stashed pgtable_t. The stashed pgtable_t contains the hpte bits.
218  * [ 000 | 1 bit secondary | 3 bit hidx | 1 bit valid]. We use one byte per
219  * each HPTE entry. With 16MB hugepage and 64K HPTE we need 256 entries and
220  * with 4K HPTE we need 4096 entries. Both will fit in a 4K pgtable_t.
221  *
222  * The top three bits are intentionally left as zero. This memory location
223  * are also used as normal page PTE pointers. So if we have any pointers
224  * left around while we collapse a hugepage, we need to make sure
225  * _PAGE_PRESENT bit of that is zero when we look at them
226  */
hpte_valid(unsigned char * hpte_slot_array,int index)227 static inline unsigned int hpte_valid(unsigned char *hpte_slot_array, int index)
228 {
229 	return hpte_slot_array[index] & 0x1;
230 }
231 
hpte_hash_index(unsigned char * hpte_slot_array,int index)232 static inline unsigned int hpte_hash_index(unsigned char *hpte_slot_array,
233 					   int index)
234 {
235 	return hpte_slot_array[index] >> 1;
236 }
237 
mark_hpte_slot_valid(unsigned char * hpte_slot_array,unsigned int index,unsigned int hidx)238 static inline void mark_hpte_slot_valid(unsigned char *hpte_slot_array,
239 					unsigned int index, unsigned int hidx)
240 {
241 	hpte_slot_array[index] = (hidx << 1) | 0x1;
242 }
243 
244 /*
245  *
246  * For core kernel code by design pmd_trans_huge is never run on any hugetlbfs
247  * page. The hugetlbfs page table walking and mangling paths are totally
248  * separated form the core VM paths and they're differentiated by
249  *  VM_HUGETLB being set on vm_flags well before any pmd_trans_huge could run.
250  *
251  * pmd_trans_huge() is defined as false at build time if
252  * CONFIG_TRANSPARENT_HUGEPAGE=n to optimize away code blocks at build
253  * time in such case.
254  *
255  * For ppc64 we need to differntiate from explicit hugepages from THP, because
256  * for THP we also track the subpage details at the pmd level. We don't do
257  * that for explicit huge pages.
258  *
259  */
hash__pmd_trans_huge(pmd_t pmd)260 static inline int hash__pmd_trans_huge(pmd_t pmd)
261 {
262 	return !!((pmd_val(pmd) & (_PAGE_PTE | H_PAGE_THP_HUGE | _PAGE_DEVMAP)) ==
263 		  (_PAGE_PTE | H_PAGE_THP_HUGE));
264 }
265 
hash__pmd_same(pmd_t pmd_a,pmd_t pmd_b)266 static inline int hash__pmd_same(pmd_t pmd_a, pmd_t pmd_b)
267 {
268 	return (((pmd_raw(pmd_a) ^ pmd_raw(pmd_b)) & ~cpu_to_be64(_PAGE_HPTEFLAGS)) == 0);
269 }
270 
hash__pmd_mkhuge(pmd_t pmd)271 static inline pmd_t hash__pmd_mkhuge(pmd_t pmd)
272 {
273 	return __pmd(pmd_val(pmd) | (_PAGE_PTE | H_PAGE_THP_HUGE));
274 }
275 
276 extern unsigned long hash__pmd_hugepage_update(struct mm_struct *mm,
277 					   unsigned long addr, pmd_t *pmdp,
278 					   unsigned long clr, unsigned long set);
279 extern pmd_t hash__pmdp_collapse_flush(struct vm_area_struct *vma,
280 				   unsigned long address, pmd_t *pmdp);
281 extern void hash__pgtable_trans_huge_deposit(struct mm_struct *mm, pmd_t *pmdp,
282 					 pgtable_t pgtable);
283 extern pgtable_t hash__pgtable_trans_huge_withdraw(struct mm_struct *mm, pmd_t *pmdp);
284 extern pmd_t hash__pmdp_huge_get_and_clear(struct mm_struct *mm,
285 				       unsigned long addr, pmd_t *pmdp);
286 extern int hash__has_transparent_hugepage(void);
287 #endif /*  CONFIG_TRANSPARENT_HUGEPAGE */
288 
hash__pmd_mkdevmap(pmd_t pmd)289 static inline pmd_t hash__pmd_mkdevmap(pmd_t pmd)
290 {
291 	return __pmd(pmd_val(pmd) | (_PAGE_PTE | H_PAGE_THP_HUGE | _PAGE_DEVMAP));
292 }
293 
294 #endif	/* __ASSEMBLY__ */
295 
296 #endif /* _ASM_POWERPC_BOOK3S_64_HASH_64K_H */
297