Mitchel Humpherys | 86a560e | 2015-09-30 14:23:58 -0700 | [diff] [blame] | 1 | /* Copyright (c) 2016, The Linux Foundation. All rights reserved. |
| 2 | * |
| 3 | * This program is free software; you can redistribute it and/or modify |
| 4 | * it under the terms of the GNU General Public License version 2 and |
| 5 | * only version 2 as published by the Free Software Foundation. |
| 6 | * |
| 7 | * This program is distributed in the hope that it will be useful, |
| 8 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 9 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 10 | * GNU General Public License for more details. |
| 11 | */ |
| 12 | |
| 13 | #define pr_fmt(fmt) "io-pgtable-fast: " fmt |
| 14 | |
| 15 | #include <linux/iommu.h> |
| 16 | #include <linux/kernel.h> |
| 17 | #include <linux/scatterlist.h> |
| 18 | #include <linux/sizes.h> |
| 19 | #include <linux/slab.h> |
| 20 | #include <linux/types.h> |
| 21 | #include <linux/io-pgtable-fast.h> |
| 22 | #include <asm/cacheflush.h> |
| 23 | |
| 24 | #include "io-pgtable.h" |
| 25 | |
| 26 | #define AV8L_FAST_MAX_ADDR_BITS 48 |
| 27 | |
| 28 | /* Struct accessors */ |
| 29 | #define iof_pgtable_to_data(x) \ |
| 30 | container_of((x), struct av8l_fast_io_pgtable, iop) |
| 31 | |
| 32 | #define iof_pgtable_ops_to_pgtable(x) \ |
| 33 | container_of((x), struct io_pgtable, ops) |
| 34 | |
| 35 | #define iof_pgtable_ops_to_data(x) \ |
| 36 | iof_pgtable_to_data(iof_pgtable_ops_to_pgtable(x)) |
| 37 | |
| 38 | struct av8l_fast_io_pgtable { |
| 39 | struct io_pgtable iop; |
| 40 | av8l_fast_iopte *pgd; |
| 41 | av8l_fast_iopte *puds[4]; |
| 42 | av8l_fast_iopte *pmds; |
| 43 | struct page **pages; /* page table memory */ |
| 44 | }; |
| 45 | |
| 46 | /* Page table bits */ |
| 47 | #define AV8L_FAST_PTE_TYPE_SHIFT 0 |
| 48 | #define AV8L_FAST_PTE_TYPE_MASK 0x3 |
| 49 | |
| 50 | #define AV8L_FAST_PTE_TYPE_BLOCK 1 |
| 51 | #define AV8L_FAST_PTE_TYPE_TABLE 3 |
| 52 | #define AV8L_FAST_PTE_TYPE_PAGE 3 |
| 53 | |
| 54 | #define AV8L_FAST_PTE_NSTABLE (((av8l_fast_iopte)1) << 63) |
| 55 | #define AV8L_FAST_PTE_XN (((av8l_fast_iopte)3) << 53) |
| 56 | #define AV8L_FAST_PTE_AF (((av8l_fast_iopte)1) << 10) |
| 57 | #define AV8L_FAST_PTE_SH_NS (((av8l_fast_iopte)0) << 8) |
| 58 | #define AV8L_FAST_PTE_SH_OS (((av8l_fast_iopte)2) << 8) |
| 59 | #define AV8L_FAST_PTE_SH_IS (((av8l_fast_iopte)3) << 8) |
| 60 | #define AV8L_FAST_PTE_NS (((av8l_fast_iopte)1) << 5) |
| 61 | #define AV8L_FAST_PTE_VALID (((av8l_fast_iopte)1) << 0) |
| 62 | |
| 63 | #define AV8L_FAST_PTE_ATTR_LO_MASK (((av8l_fast_iopte)0x3ff) << 2) |
| 64 | /* Ignore the contiguous bit for block splitting */ |
| 65 | #define AV8L_FAST_PTE_ATTR_HI_MASK (((av8l_fast_iopte)6) << 52) |
| 66 | #define AV8L_FAST_PTE_ATTR_MASK (AV8L_FAST_PTE_ATTR_LO_MASK | \ |
| 67 | AV8L_FAST_PTE_ATTR_HI_MASK) |
| 68 | #define AV8L_FAST_PTE_ADDR_MASK ((av8l_fast_iopte)0xfffffffff000) |
| 69 | |
| 70 | |
| 71 | /* Stage-1 PTE */ |
| 72 | #define AV8L_FAST_PTE_AP_PRIV_RW (((av8l_fast_iopte)0) << 6) |
| 73 | #define AV8L_FAST_PTE_AP_RW (((av8l_fast_iopte)1) << 6) |
| 74 | #define AV8L_FAST_PTE_AP_PRIV_RO (((av8l_fast_iopte)2) << 6) |
| 75 | #define AV8L_FAST_PTE_AP_RO (((av8l_fast_iopte)3) << 6) |
| 76 | #define AV8L_FAST_PTE_ATTRINDX_SHIFT 2 |
| 77 | #define AV8L_FAST_PTE_nG (((av8l_fast_iopte)1) << 11) |
| 78 | |
| 79 | /* Stage-2 PTE */ |
| 80 | #define AV8L_FAST_PTE_HAP_FAULT (((av8l_fast_iopte)0) << 6) |
| 81 | #define AV8L_FAST_PTE_HAP_READ (((av8l_fast_iopte)1) << 6) |
| 82 | #define AV8L_FAST_PTE_HAP_WRITE (((av8l_fast_iopte)2) << 6) |
| 83 | #define AV8L_FAST_PTE_MEMATTR_OIWB (((av8l_fast_iopte)0xf) << 2) |
| 84 | #define AV8L_FAST_PTE_MEMATTR_NC (((av8l_fast_iopte)0x5) << 2) |
| 85 | #define AV8L_FAST_PTE_MEMATTR_DEV (((av8l_fast_iopte)0x1) << 2) |
| 86 | |
| 87 | /* Register bits */ |
| 88 | #define ARM_32_LPAE_TCR_EAE (1 << 31) |
| 89 | #define ARM_64_LPAE_S2_TCR_RES1 (1 << 31) |
| 90 | |
| 91 | #define AV8L_FAST_TCR_TG0_4K (0 << 14) |
| 92 | #define AV8L_FAST_TCR_TG0_64K (1 << 14) |
| 93 | #define AV8L_FAST_TCR_TG0_16K (2 << 14) |
| 94 | |
| 95 | #define AV8L_FAST_TCR_SH0_SHIFT 12 |
| 96 | #define AV8L_FAST_TCR_SH0_MASK 0x3 |
| 97 | #define AV8L_FAST_TCR_SH_NS 0 |
| 98 | #define AV8L_FAST_TCR_SH_OS 2 |
| 99 | #define AV8L_FAST_TCR_SH_IS 3 |
| 100 | |
| 101 | #define AV8L_FAST_TCR_ORGN0_SHIFT 10 |
| 102 | #define AV8L_FAST_TCR_IRGN0_SHIFT 8 |
| 103 | #define AV8L_FAST_TCR_RGN_MASK 0x3 |
| 104 | #define AV8L_FAST_TCR_RGN_NC 0 |
| 105 | #define AV8L_FAST_TCR_RGN_WBWA 1 |
| 106 | #define AV8L_FAST_TCR_RGN_WT 2 |
| 107 | #define AV8L_FAST_TCR_RGN_WB 3 |
| 108 | |
| 109 | #define AV8L_FAST_TCR_SL0_SHIFT 6 |
| 110 | #define AV8L_FAST_TCR_SL0_MASK 0x3 |
| 111 | |
| 112 | #define AV8L_FAST_TCR_T0SZ_SHIFT 0 |
| 113 | #define AV8L_FAST_TCR_SZ_MASK 0xf |
| 114 | |
| 115 | #define AV8L_FAST_TCR_PS_SHIFT 16 |
| 116 | #define AV8L_FAST_TCR_PS_MASK 0x7 |
| 117 | |
| 118 | #define AV8L_FAST_TCR_IPS_SHIFT 32 |
| 119 | #define AV8L_FAST_TCR_IPS_MASK 0x7 |
| 120 | |
| 121 | #define AV8L_FAST_TCR_PS_32_BIT 0x0ULL |
| 122 | #define AV8L_FAST_TCR_PS_36_BIT 0x1ULL |
| 123 | #define AV8L_FAST_TCR_PS_40_BIT 0x2ULL |
| 124 | #define AV8L_FAST_TCR_PS_42_BIT 0x3ULL |
| 125 | #define AV8L_FAST_TCR_PS_44_BIT 0x4ULL |
| 126 | #define AV8L_FAST_TCR_PS_48_BIT 0x5ULL |
| 127 | |
| 128 | #define AV8L_FAST_TCR_EPD1_SHIFT 23 |
| 129 | #define AV8L_FAST_TCR_EPD1_FAULT 1 |
| 130 | |
| 131 | #define AV8L_FAST_MAIR_ATTR_SHIFT(n) ((n) << 3) |
| 132 | #define AV8L_FAST_MAIR_ATTR_MASK 0xff |
| 133 | #define AV8L_FAST_MAIR_ATTR_DEVICE 0x04 |
| 134 | #define AV8L_FAST_MAIR_ATTR_NC 0x44 |
| 135 | #define AV8L_FAST_MAIR_ATTR_WBRWA 0xff |
Patrick Daly | bf76227 | 2016-11-03 16:49:44 -0700 | [diff] [blame] | 136 | #define AV8L_FAST_MAIR_ATTR_UPSTREAM 0xf4 |
Mitchel Humpherys | 86a560e | 2015-09-30 14:23:58 -0700 | [diff] [blame] | 137 | #define AV8L_FAST_MAIR_ATTR_IDX_NC 0 |
| 138 | #define AV8L_FAST_MAIR_ATTR_IDX_CACHE 1 |
| 139 | #define AV8L_FAST_MAIR_ATTR_IDX_DEV 2 |
Patrick Daly | bf76227 | 2016-11-03 16:49:44 -0700 | [diff] [blame] | 140 | #define AV8L_FAST_MAIR_ATTR_IDX_UPSTREAM 3 |
Mitchel Humpherys | 86a560e | 2015-09-30 14:23:58 -0700 | [diff] [blame] | 141 | |
| 142 | #define AV8L_FAST_PAGE_SHIFT 12 |
| 143 | |
| 144 | |
Mitchel Humpherys | 5c704e0 | 2015-12-21 15:06:34 -0800 | [diff] [blame] | 145 | #ifdef CONFIG_IOMMU_IO_PGTABLE_FAST_PROVE_TLB |
| 146 | |
| 147 | #include <asm/cacheflush.h> |
| 148 | #include <linux/notifier.h> |
| 149 | |
| 150 | static ATOMIC_NOTIFIER_HEAD(av8l_notifier_list); |
| 151 | |
| 152 | void av8l_register_notify(struct notifier_block *nb) |
| 153 | { |
| 154 | atomic_notifier_chain_register(&av8l_notifier_list, nb); |
| 155 | } |
| 156 | EXPORT_SYMBOL(av8l_register_notify); |
| 157 | |
| 158 | static void __av8l_check_for_stale_tlb(av8l_fast_iopte *ptep) |
| 159 | { |
| 160 | if (unlikely(*ptep)) { |
| 161 | atomic_notifier_call_chain( |
| 162 | &av8l_notifier_list, MAPPED_OVER_STALE_TLB, |
| 163 | (void *) ptep); |
| 164 | pr_err("Tried to map over a non-vacant pte: 0x%llx @ %p\n", |
| 165 | *ptep, ptep); |
| 166 | pr_err("Nearby memory:\n"); |
| 167 | print_hex_dump(KERN_ERR, "pgtbl: ", DUMP_PREFIX_ADDRESS, |
| 168 | 32, 8, ptep - 16, 32 * sizeof(*ptep), false); |
| 169 | } |
| 170 | } |
| 171 | |
| 172 | void av8l_fast_clear_stale_ptes(av8l_fast_iopte *pmds, bool skip_sync) |
| 173 | { |
| 174 | int i; |
| 175 | av8l_fast_iopte *pmdp = pmds; |
| 176 | |
| 177 | for (i = 0; i < ((SZ_1G * 4UL) >> AV8L_FAST_PAGE_SHIFT); ++i) { |
| 178 | if (!(*pmdp & AV8L_FAST_PTE_VALID)) { |
| 179 | *pmdp = 0; |
| 180 | if (!skip_sync) |
| 181 | dmac_clean_range(pmdp, pmdp + 1); |
| 182 | } |
| 183 | pmdp++; |
| 184 | } |
| 185 | } |
| 186 | #else |
| 187 | static void __av8l_check_for_stale_tlb(av8l_fast_iopte *ptep) |
| 188 | { |
| 189 | } |
| 190 | #endif |
| 191 | |
Mitchel Humpherys | 86a560e | 2015-09-30 14:23:58 -0700 | [diff] [blame] | 192 | /* caller must take care of cache maintenance on *ptep */ |
| 193 | int av8l_fast_map_public(av8l_fast_iopte *ptep, phys_addr_t paddr, size_t size, |
| 194 | int prot) |
| 195 | { |
| 196 | int i, nptes = size >> AV8L_FAST_PAGE_SHIFT; |
| 197 | av8l_fast_iopte pte = AV8L_FAST_PTE_XN |
| 198 | | AV8L_FAST_PTE_TYPE_PAGE |
| 199 | | AV8L_FAST_PTE_AF |
| 200 | | AV8L_FAST_PTE_nG |
| 201 | | AV8L_FAST_PTE_SH_IS; |
| 202 | |
| 203 | if (prot & IOMMU_MMIO) |
| 204 | pte |= (AV8L_FAST_MAIR_ATTR_IDX_DEV |
| 205 | << AV8L_FAST_PTE_ATTRINDX_SHIFT); |
| 206 | else if (prot & IOMMU_CACHE) |
| 207 | pte |= (AV8L_FAST_MAIR_ATTR_IDX_CACHE |
| 208 | << AV8L_FAST_PTE_ATTRINDX_SHIFT); |
Patrick Daly | bf76227 | 2016-11-03 16:49:44 -0700 | [diff] [blame] | 209 | else if (prot & IOMMU_USE_UPSTREAM_HINT) |
| 210 | pte |= (AV8L_FAST_MAIR_ATTR_IDX_UPSTREAM |
| 211 | << AV8L_FAST_PTE_ATTRINDX_SHIFT); |
Mitchel Humpherys | 86a560e | 2015-09-30 14:23:58 -0700 | [diff] [blame] | 212 | |
| 213 | if (!(prot & IOMMU_WRITE)) |
| 214 | pte |= AV8L_FAST_PTE_AP_RO; |
| 215 | else |
| 216 | pte |= AV8L_FAST_PTE_AP_RW; |
| 217 | |
| 218 | paddr &= AV8L_FAST_PTE_ADDR_MASK; |
Mitchel Humpherys | 5c704e0 | 2015-12-21 15:06:34 -0800 | [diff] [blame] | 219 | for (i = 0; i < nptes; i++, paddr += SZ_4K) { |
| 220 | __av8l_check_for_stale_tlb(ptep + i); |
Mitchel Humpherys | 86a560e | 2015-09-30 14:23:58 -0700 | [diff] [blame] | 221 | *(ptep + i) = pte | paddr; |
Mitchel Humpherys | 5c704e0 | 2015-12-21 15:06:34 -0800 | [diff] [blame] | 222 | } |
Mitchel Humpherys | 86a560e | 2015-09-30 14:23:58 -0700 | [diff] [blame] | 223 | |
| 224 | return 0; |
| 225 | } |
| 226 | |
| 227 | static int av8l_fast_map(struct io_pgtable_ops *ops, unsigned long iova, |
| 228 | phys_addr_t paddr, size_t size, int prot) |
| 229 | { |
| 230 | struct av8l_fast_io_pgtable *data = iof_pgtable_ops_to_data(ops); |
| 231 | av8l_fast_iopte *ptep = iopte_pmd_offset(data->pmds, iova); |
| 232 | unsigned long nptes = size >> AV8L_FAST_PAGE_SHIFT; |
| 233 | |
| 234 | av8l_fast_map_public(ptep, paddr, size, prot); |
| 235 | dmac_clean_range(ptep, ptep + nptes); |
| 236 | |
| 237 | return 0; |
| 238 | } |
| 239 | |
Mitchel Humpherys | 5c704e0 | 2015-12-21 15:06:34 -0800 | [diff] [blame] | 240 | static void __av8l_fast_unmap(av8l_fast_iopte *ptep, size_t size, |
| 241 | bool need_stale_tlb_tracking) |
| 242 | { |
| 243 | unsigned long nptes = size >> AV8L_FAST_PAGE_SHIFT; |
| 244 | int val = need_stale_tlb_tracking |
| 245 | ? AV8L_FAST_PTE_UNMAPPED_NEED_TLBI |
| 246 | : 0; |
| 247 | |
| 248 | memset(ptep, val, sizeof(*ptep) * nptes); |
| 249 | } |
| 250 | |
Mitchel Humpherys | 86a560e | 2015-09-30 14:23:58 -0700 | [diff] [blame] | 251 | /* caller must take care of cache maintenance on *ptep */ |
| 252 | void av8l_fast_unmap_public(av8l_fast_iopte *ptep, size_t size) |
| 253 | { |
Mitchel Humpherys | 5c704e0 | 2015-12-21 15:06:34 -0800 | [diff] [blame] | 254 | __av8l_fast_unmap(ptep, size, true); |
Mitchel Humpherys | 86a560e | 2015-09-30 14:23:58 -0700 | [diff] [blame] | 255 | } |
| 256 | |
| 257 | /* upper layer must take care of TLB invalidation */ |
| 258 | static size_t av8l_fast_unmap(struct io_pgtable_ops *ops, unsigned long iova, |
| 259 | size_t size) |
| 260 | { |
| 261 | struct av8l_fast_io_pgtable *data = iof_pgtable_ops_to_data(ops); |
| 262 | av8l_fast_iopte *ptep = iopte_pmd_offset(data->pmds, iova); |
| 263 | unsigned long nptes = size >> AV8L_FAST_PAGE_SHIFT; |
| 264 | |
Mitchel Humpherys | 5c704e0 | 2015-12-21 15:06:34 -0800 | [diff] [blame] | 265 | __av8l_fast_unmap(ptep, size, false); |
Mitchel Humpherys | 86a560e | 2015-09-30 14:23:58 -0700 | [diff] [blame] | 266 | dmac_clean_range(ptep, ptep + nptes); |
| 267 | |
| 268 | return size; |
| 269 | } |
| 270 | |
| 271 | static phys_addr_t av8l_fast_iova_to_phys(struct io_pgtable_ops *ops, |
| 272 | unsigned long iova) |
| 273 | { |
| 274 | struct av8l_fast_io_pgtable *data = iof_pgtable_ops_to_data(ops); |
| 275 | av8l_fast_iopte pte, *pgdp, *pudp, *pmdp; |
| 276 | phys_addr_t phys; |
| 277 | const unsigned long pts = AV8L_FAST_PTE_TYPE_SHIFT; |
| 278 | const unsigned long ptm = AV8L_FAST_PTE_TYPE_MASK; |
| 279 | const unsigned long ptt = AV8L_FAST_PTE_TYPE_TABLE; |
| 280 | const unsigned long ptp = AV8L_FAST_PTE_TYPE_PAGE; |
| 281 | const av8l_fast_iopte am = AV8L_FAST_PTE_ADDR_MASK; |
| 282 | |
| 283 | /* TODO: clean up some of these magic numbers... */ |
| 284 | |
| 285 | pgdp = (av8l_fast_iopte *) |
| 286 | (((unsigned long)data->pgd) | ((iova & 0x7fc0000000) >> 27)); |
| 287 | pte = *pgdp; |
| 288 | if (((pte >> pts) & ptm) != ptt) |
| 289 | return 0; |
| 290 | pudp = phys_to_virt((pte & am) | ((iova & 0x3fe00000) >> 18)); |
| 291 | |
| 292 | pte = *pudp; |
| 293 | if (((pte >> pts) & ptm) != ptt) |
| 294 | return 0; |
| 295 | pmdp = phys_to_virt((pte & am) | ((iova & 0x1ff000) >> 9)); |
| 296 | |
| 297 | pte = *pmdp; |
| 298 | if (((pte >> pts) & ptm) != ptp) |
| 299 | return 0; |
| 300 | phys = pte & am; |
| 301 | |
| 302 | return phys | (iova & 0xfff); |
| 303 | } |
| 304 | |
| 305 | static int av8l_fast_map_sg(struct io_pgtable_ops *ops, unsigned long iova, |
| 306 | struct scatterlist *sg, unsigned int nents, |
| 307 | int prot, size_t *size) |
| 308 | { |
| 309 | return -ENODEV; |
| 310 | } |
| 311 | |
| 312 | static struct av8l_fast_io_pgtable * |
| 313 | av8l_fast_alloc_pgtable_data(struct io_pgtable_cfg *cfg) |
| 314 | { |
| 315 | struct av8l_fast_io_pgtable *data; |
| 316 | |
| 317 | data = kmalloc(sizeof(*data), GFP_KERNEL); |
| 318 | if (!data) |
| 319 | return NULL; |
| 320 | |
| 321 | data->iop.ops = (struct io_pgtable_ops) { |
| 322 | .map = av8l_fast_map, |
| 323 | .map_sg = av8l_fast_map_sg, |
| 324 | .unmap = av8l_fast_unmap, |
| 325 | .iova_to_phys = av8l_fast_iova_to_phys, |
| 326 | }; |
| 327 | |
| 328 | return data; |
| 329 | } |
| 330 | |
| 331 | /* |
| 332 | * We need 1 page for the pgd, 4 pages for puds (1GB VA per pud page) and |
| 333 | * 2048 pages for pmds (each pud page contains 512 table entries, each |
| 334 | * pointing to a pmd). |
| 335 | */ |
| 336 | #define NUM_PGD_PAGES 1 |
| 337 | #define NUM_PUD_PAGES 4 |
| 338 | #define NUM_PMD_PAGES 2048 |
| 339 | #define NUM_PGTBL_PAGES (NUM_PGD_PAGES + NUM_PUD_PAGES + NUM_PMD_PAGES) |
| 340 | |
| 341 | static int |
| 342 | av8l_fast_prepopulate_pgtables(struct av8l_fast_io_pgtable *data, |
| 343 | struct io_pgtable_cfg *cfg, void *cookie) |
| 344 | { |
| 345 | int i, j, pg = 0; |
| 346 | struct page **pages, *page; |
| 347 | |
| 348 | pages = kmalloc(sizeof(*pages) * NUM_PGTBL_PAGES, GFP_KERNEL); |
| 349 | if (!pages) |
| 350 | return -ENOMEM; |
| 351 | |
| 352 | page = alloc_page(GFP_KERNEL | __GFP_ZERO); |
| 353 | if (!page) |
| 354 | goto err_free_pages_arr; |
| 355 | pages[pg++] = page; |
| 356 | data->pgd = page_address(page); |
| 357 | |
| 358 | /* |
| 359 | * We need 2048 entries at level 2 to map 4GB of VA space. A page |
| 360 | * can hold 512 entries, so we need 4 pages. |
| 361 | */ |
| 362 | for (i = 0; i < 4; ++i) { |
| 363 | av8l_fast_iopte pte, *ptep; |
| 364 | |
| 365 | page = alloc_page(GFP_KERNEL | __GFP_ZERO); |
| 366 | if (!page) |
| 367 | goto err_free_pages; |
| 368 | pages[pg++] = page; |
| 369 | data->puds[i] = page_address(page); |
| 370 | pte = page_to_phys(page) | AV8L_FAST_PTE_TYPE_TABLE; |
| 371 | ptep = ((av8l_fast_iopte *)data->pgd) + i; |
| 372 | *ptep = pte; |
| 373 | } |
| 374 | dmac_clean_range(data->pgd, data->pgd + 4); |
| 375 | |
| 376 | /* |
| 377 | * We have 4 puds, each of which can point to 512 pmds, so we'll |
| 378 | * have 2048 pmds, each of which can hold 512 ptes, for a grand |
| 379 | * total of 2048*512=1048576 PTEs. |
| 380 | */ |
| 381 | for (i = 0; i < 4; ++i) { |
| 382 | for (j = 0; j < 512; ++j) { |
| 383 | av8l_fast_iopte pte, *pudp; |
| 384 | |
| 385 | page = alloc_page(GFP_KERNEL | __GFP_ZERO); |
| 386 | if (!page) |
| 387 | goto err_free_pages; |
| 388 | pages[pg++] = page; |
| 389 | pte = page_to_phys(page) | AV8L_FAST_PTE_TYPE_TABLE; |
| 390 | pudp = data->puds[i] + j; |
| 391 | *pudp = pte; |
| 392 | } |
| 393 | dmac_clean_range(data->puds[i], data->puds[i] + 512); |
| 394 | } |
| 395 | |
| 396 | if (WARN_ON(pg != NUM_PGTBL_PAGES)) |
| 397 | goto err_free_pages; |
| 398 | |
| 399 | /* |
| 400 | * We map the pmds into a virtually contiguous space so that we |
| 401 | * don't have to traverse the first two levels of the page tables |
| 402 | * to find the appropriate pud. Instead, it will be a simple |
| 403 | * offset from the virtual base of the pmds. |
| 404 | */ |
| 405 | data->pmds = vmap(&pages[NUM_PGD_PAGES + NUM_PUD_PAGES], NUM_PMD_PAGES, |
| 406 | VM_IOREMAP, PAGE_KERNEL); |
| 407 | if (!data->pmds) |
| 408 | goto err_free_pages; |
| 409 | |
| 410 | data->pages = pages; |
| 411 | return 0; |
| 412 | |
| 413 | err_free_pages: |
| 414 | for (i = 0; i < pg; ++i) |
| 415 | __free_page(pages[i]); |
| 416 | err_free_pages_arr: |
| 417 | kfree(pages); |
| 418 | return -ENOMEM; |
| 419 | } |
| 420 | |
| 421 | static struct io_pgtable * |
| 422 | av8l_fast_alloc_pgtable(struct io_pgtable_cfg *cfg, void *cookie) |
| 423 | { |
| 424 | u64 reg; |
| 425 | struct av8l_fast_io_pgtable *data = |
| 426 | av8l_fast_alloc_pgtable_data(cfg); |
| 427 | |
Shiraz Hashim | 06f3121 | 2016-07-04 15:05:14 +0530 | [diff] [blame] | 428 | if (!data) |
| 429 | return NULL; |
| 430 | |
Mitchel Humpherys | 86a560e | 2015-09-30 14:23:58 -0700 | [diff] [blame] | 431 | /* restrict according to the fast map requirements */ |
| 432 | cfg->ias = 32; |
| 433 | cfg->pgsize_bitmap = SZ_4K; |
| 434 | |
| 435 | /* TCR */ |
Liam Mark | 5649c82 | 2016-12-19 14:35:08 -0800 | [diff] [blame] | 436 | if (cfg->quirks & IO_PGTABLE_QUIRK_QCOM_USE_UPSTREAM_HINT) |
Patrick Daly | ce6786f | 2016-11-09 14:19:23 -0800 | [diff] [blame] | 437 | reg = (AV8L_FAST_TCR_SH_OS << AV8L_FAST_TCR_SH0_SHIFT) | |
| 438 | (AV8L_FAST_TCR_RGN_NC << AV8L_FAST_TCR_IRGN0_SHIFT) | |
| 439 | (AV8L_FAST_TCR_RGN_WBWA << AV8L_FAST_TCR_ORGN0_SHIFT); |
| 440 | else |
| 441 | reg = (AV8L_FAST_TCR_SH_IS << AV8L_FAST_TCR_SH0_SHIFT) | |
| 442 | (AV8L_FAST_TCR_RGN_NC << AV8L_FAST_TCR_IRGN0_SHIFT) | |
| 443 | (AV8L_FAST_TCR_RGN_NC << AV8L_FAST_TCR_ORGN0_SHIFT); |
Mitchel Humpherys | 86a560e | 2015-09-30 14:23:58 -0700 | [diff] [blame] | 444 | |
| 445 | reg |= AV8L_FAST_TCR_TG0_4K; |
| 446 | |
| 447 | switch (cfg->oas) { |
| 448 | case 32: |
| 449 | reg |= (AV8L_FAST_TCR_PS_32_BIT << AV8L_FAST_TCR_IPS_SHIFT); |
| 450 | break; |
| 451 | case 36: |
| 452 | reg |= (AV8L_FAST_TCR_PS_36_BIT << AV8L_FAST_TCR_IPS_SHIFT); |
| 453 | break; |
| 454 | case 40: |
| 455 | reg |= (AV8L_FAST_TCR_PS_40_BIT << AV8L_FAST_TCR_IPS_SHIFT); |
| 456 | break; |
| 457 | case 42: |
| 458 | reg |= (AV8L_FAST_TCR_PS_42_BIT << AV8L_FAST_TCR_IPS_SHIFT); |
| 459 | break; |
| 460 | case 44: |
| 461 | reg |= (AV8L_FAST_TCR_PS_44_BIT << AV8L_FAST_TCR_IPS_SHIFT); |
| 462 | break; |
| 463 | case 48: |
| 464 | reg |= (AV8L_FAST_TCR_PS_48_BIT << AV8L_FAST_TCR_IPS_SHIFT); |
| 465 | break; |
| 466 | default: |
| 467 | goto out_free_data; |
| 468 | } |
| 469 | |
| 470 | reg |= (64ULL - cfg->ias) << AV8L_FAST_TCR_T0SZ_SHIFT; |
| 471 | reg |= AV8L_FAST_TCR_EPD1_FAULT << AV8L_FAST_TCR_EPD1_SHIFT; |
| 472 | cfg->av8l_fast_cfg.tcr = reg; |
| 473 | |
| 474 | /* MAIRs */ |
| 475 | reg = (AV8L_FAST_MAIR_ATTR_NC |
| 476 | << AV8L_FAST_MAIR_ATTR_SHIFT(AV8L_FAST_MAIR_ATTR_IDX_NC)) | |
| 477 | (AV8L_FAST_MAIR_ATTR_WBRWA |
| 478 | << AV8L_FAST_MAIR_ATTR_SHIFT(AV8L_FAST_MAIR_ATTR_IDX_CACHE)) | |
| 479 | (AV8L_FAST_MAIR_ATTR_DEVICE |
Patrick Daly | bf76227 | 2016-11-03 16:49:44 -0700 | [diff] [blame] | 480 | << AV8L_FAST_MAIR_ATTR_SHIFT(AV8L_FAST_MAIR_ATTR_IDX_DEV)) | |
| 481 | (AV8L_FAST_MAIR_ATTR_UPSTREAM |
| 482 | << AV8L_FAST_MAIR_ATTR_SHIFT(AV8L_FAST_MAIR_ATTR_IDX_UPSTREAM)); |
Mitchel Humpherys | 86a560e | 2015-09-30 14:23:58 -0700 | [diff] [blame] | 483 | |
| 484 | cfg->av8l_fast_cfg.mair[0] = reg; |
| 485 | cfg->av8l_fast_cfg.mair[1] = 0; |
| 486 | |
| 487 | /* Allocate all page table memory! */ |
| 488 | if (av8l_fast_prepopulate_pgtables(data, cfg, cookie)) |
| 489 | goto out_free_data; |
| 490 | |
| 491 | cfg->av8l_fast_cfg.pmds = data->pmds; |
| 492 | |
| 493 | /* TTBRs */ |
| 494 | cfg->av8l_fast_cfg.ttbr[0] = virt_to_phys(data->pgd); |
| 495 | cfg->av8l_fast_cfg.ttbr[1] = 0; |
| 496 | return &data->iop; |
| 497 | |
| 498 | out_free_data: |
| 499 | kfree(data); |
| 500 | return NULL; |
| 501 | } |
| 502 | |
| 503 | static void av8l_fast_free_pgtable(struct io_pgtable *iop) |
| 504 | { |
| 505 | int i; |
| 506 | struct av8l_fast_io_pgtable *data = iof_pgtable_to_data(iop); |
| 507 | |
| 508 | vunmap(data->pmds); |
| 509 | for (i = 0; i < NUM_PGTBL_PAGES; ++i) |
| 510 | __free_page(data->pages[i]); |
| 511 | kfree(data->pages); |
| 512 | kfree(data); |
| 513 | } |
| 514 | |
| 515 | struct io_pgtable_init_fns io_pgtable_av8l_fast_init_fns = { |
| 516 | .alloc = av8l_fast_alloc_pgtable, |
| 517 | .free = av8l_fast_free_pgtable, |
| 518 | }; |
| 519 | |
| 520 | |
| 521 | #ifdef CONFIG_IOMMU_IO_PGTABLE_FAST_SELFTEST |
| 522 | |
| 523 | #include <linux/dma-contiguous.h> |
| 524 | |
| 525 | static struct io_pgtable_cfg *cfg_cookie; |
| 526 | |
| 527 | static void dummy_tlb_flush_all(void *cookie) |
| 528 | { |
| 529 | WARN_ON(cookie != cfg_cookie); |
| 530 | } |
| 531 | |
| 532 | static void dummy_tlb_add_flush(unsigned long iova, size_t size, size_t granule, |
| 533 | bool leaf, void *cookie) |
| 534 | { |
| 535 | WARN_ON(cookie != cfg_cookie); |
| 536 | WARN_ON(!(size & cfg_cookie->pgsize_bitmap)); |
| 537 | } |
| 538 | |
| 539 | static void dummy_tlb_sync(void *cookie) |
| 540 | { |
| 541 | WARN_ON(cookie != cfg_cookie); |
| 542 | } |
| 543 | |
| 544 | static struct iommu_gather_ops dummy_tlb_ops __initdata = { |
| 545 | .tlb_flush_all = dummy_tlb_flush_all, |
| 546 | .tlb_add_flush = dummy_tlb_add_flush, |
| 547 | .tlb_sync = dummy_tlb_sync, |
| 548 | }; |
| 549 | |
| 550 | /* |
| 551 | * Returns true if the iova range is successfully mapped to the contiguous |
| 552 | * phys range in ops. |
| 553 | */ |
| 554 | static bool av8l_fast_range_has_specific_mapping(struct io_pgtable_ops *ops, |
| 555 | const unsigned long iova_start, |
| 556 | const phys_addr_t phys_start, |
| 557 | const size_t size) |
| 558 | { |
| 559 | unsigned long iova = iova_start; |
| 560 | phys_addr_t phys = phys_start; |
| 561 | |
| 562 | while (iova < (iova_start + size)) { |
| 563 | /* + 42 just to make sure offsetting is working */ |
| 564 | if (ops->iova_to_phys(ops, iova + 42) != (phys + 42)) |
| 565 | return false; |
| 566 | iova += SZ_4K; |
| 567 | phys += SZ_4K; |
| 568 | } |
| 569 | return true; |
| 570 | } |
| 571 | |
| 572 | static int __init av8l_fast_positive_testing(void) |
| 573 | { |
| 574 | int failed = 0; |
| 575 | unsigned long iova; |
| 576 | struct io_pgtable_ops *ops; |
| 577 | struct io_pgtable_cfg cfg; |
| 578 | struct av8l_fast_io_pgtable *data; |
| 579 | av8l_fast_iopte *pmds; |
| 580 | |
| 581 | cfg = (struct io_pgtable_cfg) { |
| 582 | .tlb = &dummy_tlb_ops, |
| 583 | .ias = 32, |
| 584 | .oas = 32, |
| 585 | .pgsize_bitmap = SZ_4K, |
| 586 | }; |
| 587 | |
| 588 | cfg_cookie = &cfg; |
| 589 | ops = alloc_io_pgtable_ops(ARM_V8L_FAST, &cfg, &cfg); |
| 590 | |
| 591 | if (WARN_ON(!ops)) |
| 592 | return 1; |
| 593 | |
| 594 | data = iof_pgtable_ops_to_data(ops); |
| 595 | pmds = data->pmds; |
| 596 | |
| 597 | /* map the entire 4GB VA space with 4K map calls */ |
| 598 | for (iova = 0; iova < SZ_1G * 4UL; iova += SZ_4K) { |
| 599 | if (WARN_ON(ops->map(ops, iova, iova, SZ_4K, IOMMU_READ))) { |
| 600 | failed++; |
| 601 | continue; |
| 602 | } |
| 603 | } |
| 604 | |
| 605 | if (WARN_ON(!av8l_fast_range_has_specific_mapping(ops, 0, 0, |
| 606 | SZ_1G * 4UL))) |
| 607 | failed++; |
| 608 | |
| 609 | /* unmap it all */ |
| 610 | for (iova = 0; iova < SZ_1G * 4UL; iova += SZ_4K) { |
| 611 | if (WARN_ON(ops->unmap(ops, iova, SZ_4K) != SZ_4K)) |
| 612 | failed++; |
| 613 | } |
| 614 | |
Mitchel Humpherys | 5c704e0 | 2015-12-21 15:06:34 -0800 | [diff] [blame] | 615 | /* sweep up TLB proving PTEs */ |
| 616 | av8l_fast_clear_stale_ptes(pmds, false); |
| 617 | |
Mitchel Humpherys | 86a560e | 2015-09-30 14:23:58 -0700 | [diff] [blame] | 618 | /* map the entire 4GB VA space with 8K map calls */ |
| 619 | for (iova = 0; iova < SZ_1G * 4UL; iova += SZ_8K) { |
| 620 | if (WARN_ON(ops->map(ops, iova, iova, SZ_8K, IOMMU_READ))) { |
| 621 | failed++; |
| 622 | continue; |
| 623 | } |
| 624 | } |
| 625 | |
| 626 | if (WARN_ON(!av8l_fast_range_has_specific_mapping(ops, 0, 0, |
| 627 | SZ_1G * 4UL))) |
| 628 | failed++; |
| 629 | |
| 630 | /* unmap it all with 8K unmap calls */ |
| 631 | for (iova = 0; iova < SZ_1G * 4UL; iova += SZ_8K) { |
| 632 | if (WARN_ON(ops->unmap(ops, iova, SZ_8K) != SZ_8K)) |
| 633 | failed++; |
| 634 | } |
| 635 | |
Mitchel Humpherys | 5c704e0 | 2015-12-21 15:06:34 -0800 | [diff] [blame] | 636 | /* sweep up TLB proving PTEs */ |
| 637 | av8l_fast_clear_stale_ptes(pmds, false); |
| 638 | |
Mitchel Humpherys | 86a560e | 2015-09-30 14:23:58 -0700 | [diff] [blame] | 639 | /* map the entire 4GB VA space with 16K map calls */ |
| 640 | for (iova = 0; iova < SZ_1G * 4UL; iova += SZ_16K) { |
| 641 | if (WARN_ON(ops->map(ops, iova, iova, SZ_16K, IOMMU_READ))) { |
| 642 | failed++; |
| 643 | continue; |
| 644 | } |
| 645 | } |
| 646 | |
| 647 | if (WARN_ON(!av8l_fast_range_has_specific_mapping(ops, 0, 0, |
| 648 | SZ_1G * 4UL))) |
| 649 | failed++; |
| 650 | |
| 651 | /* unmap it all */ |
| 652 | for (iova = 0; iova < SZ_1G * 4UL; iova += SZ_16K) { |
| 653 | if (WARN_ON(ops->unmap(ops, iova, SZ_16K) != SZ_16K)) |
| 654 | failed++; |
| 655 | } |
| 656 | |
Mitchel Humpherys | 5c704e0 | 2015-12-21 15:06:34 -0800 | [diff] [blame] | 657 | /* sweep up TLB proving PTEs */ |
| 658 | av8l_fast_clear_stale_ptes(pmds, false); |
| 659 | |
Mitchel Humpherys | 86a560e | 2015-09-30 14:23:58 -0700 | [diff] [blame] | 660 | /* map the entire 4GB VA space with 64K map calls */ |
| 661 | for (iova = 0; iova < SZ_1G * 4UL; iova += SZ_64K) { |
| 662 | if (WARN_ON(ops->map(ops, iova, iova, SZ_64K, IOMMU_READ))) { |
| 663 | failed++; |
| 664 | continue; |
| 665 | } |
| 666 | } |
| 667 | |
| 668 | if (WARN_ON(!av8l_fast_range_has_specific_mapping(ops, 0, 0, |
| 669 | SZ_1G * 4UL))) |
| 670 | failed++; |
| 671 | |
| 672 | /* unmap it all at once */ |
| 673 | if (WARN_ON(ops->unmap(ops, 0, SZ_1G * 4UL) != SZ_1G * 4UL)) |
| 674 | failed++; |
| 675 | |
| 676 | free_io_pgtable_ops(ops); |
| 677 | return failed; |
| 678 | } |
| 679 | |
| 680 | static int __init av8l_fast_do_selftests(void) |
| 681 | { |
| 682 | int failed = 0; |
| 683 | |
| 684 | failed += av8l_fast_positive_testing(); |
| 685 | |
| 686 | pr_err("selftest: completed with %d failures\n", failed); |
| 687 | |
| 688 | return 0; |
| 689 | } |
| 690 | subsys_initcall(av8l_fast_do_selftests); |
| 691 | #endif |