Vineet Gupta | f1f3347 | 2013-01-18 15:12:19 +0530 | [diff] [blame] | 1 | /* |
| 2 | * TLB Management (flush/create/diagnostics) for ARC700 |
| 3 | * |
| 4 | * Copyright (C) 2004, 2007-2010, 2011-2012 Synopsys, Inc. (www.synopsys.com) |
| 5 | * |
| 6 | * This program is free software; you can redistribute it and/or modify |
| 7 | * it under the terms of the GNU General Public License version 2 as |
| 8 | * published by the Free Software Foundation. |
Vineet Gupta | d79e678 | 2013-01-18 15:12:20 +0530 | [diff] [blame] | 9 | * |
| 10 | * vineetg: Aug 2011 |
| 11 | * -Reintroduce duplicate PD fixup - some customer chips still have the issue |
| 12 | * |
| 13 | * vineetg: May 2011 |
| 14 | * -No need to flush_cache_page( ) for each call to update_mmu_cache() |
| 15 | * some of the LMBench tests improved amazingly |
| 16 | * = page-fault thrice as fast (75 usec to 28 usec) |
| 17 | * = mmap twice as fast (9.6 msec to 4.6 msec), |
| 18 | * = fork (5.3 msec to 3.7 msec) |
| 19 | * |
| 20 | * vineetg: April 2011 : |
| 21 | * -MMU v3: PD{0,1} bits layout changed: They don't overlap anymore, |
| 22 | * helps avoid a shift when preparing PD0 from PTE |
| 23 | * |
| 24 | * vineetg: April 2011 : Preparing for MMU V3 |
| 25 | * -MMU v2/v3 BCRs decoded differently |
| 26 | * -Remove TLB_SIZE hardcoding as it's variable now: 256 or 512 |
| 27 | * -tlb_entry_erase( ) can be void |
| 28 | * -local_flush_tlb_range( ): |
| 29 | * = need not "ceil" @end |
| 30 | * = walks MMU only if range spans < 32 entries, as opposed to 256 |
| 31 | * |
| 32 | * Vineetg: Sept 10th 2008 |
| 33 | * -Changes related to MMU v2 (Rel 4.8) |
| 34 | * |
| 35 | * Vineetg: Aug 29th 2008 |
| 36 | * -In TLB Flush operations (Metal Fix MMU) there is a explict command to |
| 37 | * flush Micro-TLBS. If TLB Index Reg is invalid prior to TLBIVUTLB cmd, |
| 38 | * it fails. Thus need to load it with ANY valid value before invoking |
| 39 | * TLBIVUTLB cmd |
| 40 | * |
| 41 | * Vineetg: Aug 21th 2008: |
| 42 | * -Reduced the duration of IRQ lockouts in TLB Flush routines |
| 43 | * -Multiple copies of TLB erase code seperated into a "single" function |
| 44 | * -In TLB Flush routines, interrupt disabling moved UP to retrieve ASID |
| 45 | * in interrupt-safe region. |
| 46 | * |
| 47 | * Vineetg: April 23rd Bug #93131 |
| 48 | * Problem: tlb_flush_kernel_range() doesnt do anything if the range to |
| 49 | * flush is more than the size of TLB itself. |
| 50 | * |
| 51 | * Rahul Trivedi : Codito Technologies 2004 |
Vineet Gupta | f1f3347 | 2013-01-18 15:12:19 +0530 | [diff] [blame] | 52 | */ |
| 53 | |
| 54 | #include <linux/module.h> |
| 55 | #include <asm/arcregs.h> |
Vineet Gupta | d79e678 | 2013-01-18 15:12:20 +0530 | [diff] [blame] | 56 | #include <asm/setup.h> |
Vineet Gupta | f1f3347 | 2013-01-18 15:12:19 +0530 | [diff] [blame] | 57 | #include <asm/mmu_context.h> |
Vineet Gupta | da1677b | 2013-05-14 13:28:17 +0530 | [diff] [blame] | 58 | #include <asm/mmu.h> |
Vineet Gupta | f1f3347 | 2013-01-18 15:12:19 +0530 | [diff] [blame] | 59 | |
Vineet Gupta | d79e678 | 2013-01-18 15:12:20 +0530 | [diff] [blame] | 60 | /* Need for ARC MMU v2 |
| 61 | * |
| 62 | * ARC700 MMU-v1 had a Joint-TLB for Code and Data and is 2 way set-assoc. |
| 63 | * For a memcpy operation with 3 players (src/dst/code) such that all 3 pages |
| 64 | * map into same set, there would be contention for the 2 ways causing severe |
| 65 | * Thrashing. |
| 66 | * |
| 67 | * Although J-TLB is 2 way set assoc, ARC700 caches J-TLB into uTLBS which has |
| 68 | * much higher associativity. u-D-TLB is 8 ways, u-I-TLB is 4 ways. |
| 69 | * Given this, the thrasing problem should never happen because once the 3 |
| 70 | * J-TLB entries are created (even though 3rd will knock out one of the prev |
| 71 | * two), the u-D-TLB and u-I-TLB will have what is required to accomplish memcpy |
| 72 | * |
| 73 | * Yet we still see the Thrashing because a J-TLB Write cause flush of u-TLBs. |
| 74 | * This is a simple design for keeping them in sync. So what do we do? |
| 75 | * The solution which James came up was pretty neat. It utilised the assoc |
| 76 | * of uTLBs by not invalidating always but only when absolutely necessary. |
| 77 | * |
| 78 | * - Existing TLB commands work as before |
| 79 | * - New command (TLBWriteNI) for TLB write without clearing uTLBs |
| 80 | * - New command (TLBIVUTLB) to invalidate uTLBs. |
| 81 | * |
| 82 | * The uTLBs need only be invalidated when pages are being removed from the |
| 83 | * OS page table. If a 'victim' TLB entry is being overwritten in the main TLB |
| 84 | * as a result of a miss, the removed entry is still allowed to exist in the |
| 85 | * uTLBs as it is still valid and present in the OS page table. This allows the |
| 86 | * full associativity of the uTLBs to hide the limited associativity of the main |
| 87 | * TLB. |
| 88 | * |
| 89 | * During a miss handler, the new "TLBWriteNI" command is used to load |
| 90 | * entries without clearing the uTLBs. |
| 91 | * |
| 92 | * When the OS page table is updated, TLB entries that may be associated with a |
| 93 | * removed page are removed (flushed) from the TLB using TLBWrite. In this |
| 94 | * circumstance, the uTLBs must also be cleared. This is done by using the |
| 95 | * existing TLBWrite command. An explicit IVUTLB is also required for those |
| 96 | * corner cases when TLBWrite was not executed at all because the corresp |
| 97 | * J-TLB entry got evicted/replaced. |
| 98 | */ |
| 99 | |
Vineet Gupta | da1677b | 2013-05-14 13:28:17 +0530 | [diff] [blame] | 100 | |
Vineet Gupta | f1f3347 | 2013-01-18 15:12:19 +0530 | [diff] [blame] | 101 | /* A copy of the ASID from the PID reg is kept in asid_cache */ |
| 102 | int asid_cache = FIRST_ASID; |
| 103 | |
| 104 | /* ASID to mm struct mapping. We have one extra entry corresponding to |
| 105 | * NO_ASID to save us a compare when clearing the mm entry for old asid |
| 106 | * see get_new_mmu_context (asm-arc/mmu_context.h) |
| 107 | */ |
| 108 | struct mm_struct *asid_mm_map[NUM_ASID + 1]; |
Vineet Gupta | cc562d2 | 2013-01-18 15:12:19 +0530 | [diff] [blame] | 109 | |
Vineet Gupta | d79e678 | 2013-01-18 15:12:20 +0530 | [diff] [blame] | 110 | /* |
| 111 | * Utility Routine to erase a J-TLB entry |
| 112 | * The procedure is to look it up in the MMU. If found, ERASE it by |
| 113 | * issuing a TlbWrite CMD with PD0 = PD1 = 0 |
| 114 | */ |
| 115 | |
| 116 | static void __tlb_entry_erase(void) |
| 117 | { |
| 118 | write_aux_reg(ARC_REG_TLBPD1, 0); |
| 119 | write_aux_reg(ARC_REG_TLBPD0, 0); |
| 120 | write_aux_reg(ARC_REG_TLBCOMMAND, TLBWrite); |
| 121 | } |
| 122 | |
| 123 | static void tlb_entry_erase(unsigned int vaddr_n_asid) |
| 124 | { |
| 125 | unsigned int idx; |
| 126 | |
| 127 | /* Locate the TLB entry for this vaddr + ASID */ |
| 128 | write_aux_reg(ARC_REG_TLBPD0, vaddr_n_asid); |
| 129 | write_aux_reg(ARC_REG_TLBCOMMAND, TLBProbe); |
| 130 | idx = read_aux_reg(ARC_REG_TLBINDEX); |
| 131 | |
| 132 | /* No error means entry found, zero it out */ |
| 133 | if (likely(!(idx & TLB_LKUP_ERR))) { |
| 134 | __tlb_entry_erase(); |
| 135 | } else { /* Some sort of Error */ |
| 136 | |
| 137 | /* Duplicate entry error */ |
| 138 | if (idx & 0x1) { |
| 139 | /* TODO we need to handle this case too */ |
| 140 | pr_emerg("unhandled Duplicate flush for %x\n", |
| 141 | vaddr_n_asid); |
| 142 | } |
| 143 | /* else entry not found so nothing to do */ |
| 144 | } |
| 145 | } |
| 146 | |
| 147 | /**************************************************************************** |
| 148 | * ARC700 MMU caches recently used J-TLB entries (RAM) as uTLBs (FLOPs) |
| 149 | * |
| 150 | * New IVUTLB cmd in MMU v2 explictly invalidates the uTLB |
| 151 | * |
| 152 | * utlb_invalidate ( ) |
| 153 | * -For v2 MMU calls Flush uTLB Cmd |
| 154 | * -For v1 MMU does nothing (except for Metal Fix v1 MMU) |
| 155 | * This is because in v1 TLBWrite itself invalidate uTLBs |
| 156 | ***************************************************************************/ |
| 157 | |
| 158 | static void utlb_invalidate(void) |
| 159 | { |
| 160 | #if (CONFIG_ARC_MMU_VER >= 2) |
| 161 | |
| 162 | #if (CONFIG_ARC_MMU_VER < 3) |
| 163 | /* MMU v2 introduced the uTLB Flush command. |
| 164 | * There was however an obscure hardware bug, where uTLB flush would |
| 165 | * fail when a prior probe for J-TLB (both totally unrelated) would |
| 166 | * return lkup err - because the entry didnt exist in MMU. |
| 167 | * The Workround was to set Index reg with some valid value, prior to |
| 168 | * flush. This was fixed in MMU v3 hence not needed any more |
| 169 | */ |
| 170 | unsigned int idx; |
| 171 | |
| 172 | /* make sure INDEX Reg is valid */ |
| 173 | idx = read_aux_reg(ARC_REG_TLBINDEX); |
| 174 | |
| 175 | /* If not write some dummy val */ |
| 176 | if (unlikely(idx & TLB_LKUP_ERR)) |
| 177 | write_aux_reg(ARC_REG_TLBINDEX, 0xa); |
| 178 | #endif |
| 179 | |
| 180 | write_aux_reg(ARC_REG_TLBCOMMAND, TLBIVUTLB); |
| 181 | #endif |
| 182 | |
| 183 | } |
| 184 | |
| 185 | /* |
| 186 | * Un-conditionally (without lookup) erase the entire MMU contents |
| 187 | */ |
| 188 | |
| 189 | noinline void local_flush_tlb_all(void) |
| 190 | { |
| 191 | unsigned long flags; |
| 192 | unsigned int entry; |
| 193 | struct cpuinfo_arc_mmu *mmu = &cpuinfo_arc700[smp_processor_id()].mmu; |
| 194 | |
| 195 | local_irq_save(flags); |
| 196 | |
| 197 | /* Load PD0 and PD1 with template for a Blank Entry */ |
| 198 | write_aux_reg(ARC_REG_TLBPD1, 0); |
| 199 | write_aux_reg(ARC_REG_TLBPD0, 0); |
| 200 | |
| 201 | for (entry = 0; entry < mmu->num_tlb; entry++) { |
| 202 | /* write this entry to the TLB */ |
| 203 | write_aux_reg(ARC_REG_TLBINDEX, entry); |
| 204 | write_aux_reg(ARC_REG_TLBCOMMAND, TLBWrite); |
| 205 | } |
| 206 | |
| 207 | utlb_invalidate(); |
| 208 | |
| 209 | local_irq_restore(flags); |
| 210 | } |
| 211 | |
| 212 | /* |
| 213 | * Flush the entrie MM for userland. The fastest way is to move to Next ASID |
| 214 | */ |
| 215 | noinline void local_flush_tlb_mm(struct mm_struct *mm) |
| 216 | { |
| 217 | /* |
| 218 | * Small optimisation courtesy IA64 |
| 219 | * flush_mm called during fork,exit,munmap etc, multiple times as well. |
| 220 | * Only for fork( ) do we need to move parent to a new MMU ctxt, |
| 221 | * all other cases are NOPs, hence this check. |
| 222 | */ |
| 223 | if (atomic_read(&mm->mm_users) == 0) |
| 224 | return; |
| 225 | |
| 226 | /* |
| 227 | * Workaround for Android weirdism: |
| 228 | * A binder VMA could end up in a task such that vma->mm != tsk->mm |
| 229 | * old code would cause h/w - s/w ASID to get out of sync |
| 230 | */ |
| 231 | if (current->mm != mm) |
| 232 | destroy_context(mm); |
| 233 | else |
| 234 | get_new_mmu_context(mm); |
| 235 | } |
| 236 | |
| 237 | /* |
| 238 | * Flush a Range of TLB entries for userland. |
| 239 | * @start is inclusive, while @end is exclusive |
| 240 | * Difference between this and Kernel Range Flush is |
| 241 | * -Here the fastest way (if range is too large) is to move to next ASID |
| 242 | * without doing any explicit Shootdown |
| 243 | * -In case of kernel Flush, entry has to be shot down explictly |
| 244 | */ |
| 245 | void local_flush_tlb_range(struct vm_area_struct *vma, unsigned long start, |
| 246 | unsigned long end) |
| 247 | { |
| 248 | unsigned long flags; |
| 249 | unsigned int asid; |
| 250 | |
| 251 | /* If range @start to @end is more than 32 TLB entries deep, |
| 252 | * its better to move to a new ASID rather than searching for |
| 253 | * individual entries and then shooting them down |
| 254 | * |
| 255 | * The calc above is rough, doesn't account for unaligned parts, |
| 256 | * since this is heuristics based anyways |
| 257 | */ |
| 258 | if (unlikely((end - start) >= PAGE_SIZE * 32)) { |
| 259 | local_flush_tlb_mm(vma->vm_mm); |
| 260 | return; |
| 261 | } |
| 262 | |
| 263 | /* |
| 264 | * @start moved to page start: this alone suffices for checking |
| 265 | * loop end condition below, w/o need for aligning @end to end |
| 266 | * e.g. 2000 to 4001 will anyhow loop twice |
| 267 | */ |
| 268 | start &= PAGE_MASK; |
| 269 | |
| 270 | local_irq_save(flags); |
| 271 | asid = vma->vm_mm->context.asid; |
| 272 | |
| 273 | if (asid != NO_ASID) { |
| 274 | while (start < end) { |
| 275 | tlb_entry_erase(start | (asid & 0xff)); |
| 276 | start += PAGE_SIZE; |
| 277 | } |
| 278 | } |
| 279 | |
| 280 | utlb_invalidate(); |
| 281 | |
| 282 | local_irq_restore(flags); |
| 283 | } |
| 284 | |
| 285 | /* Flush the kernel TLB entries - vmalloc/modules (Global from MMU perspective) |
| 286 | * @start, @end interpreted as kvaddr |
| 287 | * Interestingly, shared TLB entries can also be flushed using just |
| 288 | * @start,@end alone (interpreted as user vaddr), although technically SASID |
| 289 | * is also needed. However our smart TLbProbe lookup takes care of that. |
| 290 | */ |
| 291 | void local_flush_tlb_kernel_range(unsigned long start, unsigned long end) |
| 292 | { |
| 293 | unsigned long flags; |
| 294 | |
| 295 | /* exactly same as above, except for TLB entry not taking ASID */ |
| 296 | |
| 297 | if (unlikely((end - start) >= PAGE_SIZE * 32)) { |
| 298 | local_flush_tlb_all(); |
| 299 | return; |
| 300 | } |
| 301 | |
| 302 | start &= PAGE_MASK; |
| 303 | |
| 304 | local_irq_save(flags); |
| 305 | while (start < end) { |
| 306 | tlb_entry_erase(start); |
| 307 | start += PAGE_SIZE; |
| 308 | } |
| 309 | |
| 310 | utlb_invalidate(); |
| 311 | |
| 312 | local_irq_restore(flags); |
| 313 | } |
| 314 | |
| 315 | /* |
| 316 | * Delete TLB entry in MMU for a given page (??? address) |
| 317 | * NOTE One TLB entry contains translation for single PAGE |
| 318 | */ |
| 319 | |
| 320 | void local_flush_tlb_page(struct vm_area_struct *vma, unsigned long page) |
| 321 | { |
| 322 | unsigned long flags; |
| 323 | |
| 324 | /* Note that it is critical that interrupts are DISABLED between |
| 325 | * checking the ASID and using it flush the TLB entry |
| 326 | */ |
| 327 | local_irq_save(flags); |
| 328 | |
| 329 | if (vma->vm_mm->context.asid != NO_ASID) { |
| 330 | tlb_entry_erase((page & PAGE_MASK) | |
| 331 | (vma->vm_mm->context.asid & 0xff)); |
| 332 | utlb_invalidate(); |
| 333 | } |
| 334 | |
| 335 | local_irq_restore(flags); |
| 336 | } |
Vineet Gupta | cc562d2 | 2013-01-18 15:12:19 +0530 | [diff] [blame] | 337 | |
| 338 | /* |
| 339 | * Routine to create a TLB entry |
| 340 | */ |
| 341 | void create_tlb(struct vm_area_struct *vma, unsigned long address, pte_t *ptep) |
| 342 | { |
| 343 | unsigned long flags; |
Vineet Gupta | 64b703e | 2013-06-17 18:12:13 +0530 | [diff] [blame^] | 344 | unsigned int idx, asid_or_sasid, rwx; |
Vineet Gupta | cc562d2 | 2013-01-18 15:12:19 +0530 | [diff] [blame] | 345 | unsigned long pd0_flags; |
| 346 | |
| 347 | /* |
| 348 | * create_tlb() assumes that current->mm == vma->mm, since |
| 349 | * -it ASID for TLB entry is fetched from MMU ASID reg (valid for curr) |
| 350 | * -completes the lazy write to SASID reg (again valid for curr tsk) |
| 351 | * |
| 352 | * Removing the assumption involves |
| 353 | * -Using vma->mm->context{ASID,SASID}, as opposed to MMU reg. |
| 354 | * -Fix the TLB paranoid debug code to not trigger false negatives. |
| 355 | * -More importantly it makes this handler inconsistent with fast-path |
| 356 | * TLB Refill handler which always deals with "current" |
| 357 | * |
| 358 | * Lets see the use cases when current->mm != vma->mm and we land here |
| 359 | * 1. execve->copy_strings()->__get_user_pages->handle_mm_fault |
| 360 | * Here VM wants to pre-install a TLB entry for user stack while |
| 361 | * current->mm still points to pre-execve mm (hence the condition). |
| 362 | * However the stack vaddr is soon relocated (randomization) and |
| 363 | * move_page_tables() tries to undo that TLB entry. |
| 364 | * Thus not creating TLB entry is not any worse. |
| 365 | * |
| 366 | * 2. ptrace(POKETEXT) causes a CoW - debugger(current) inserting a |
| 367 | * breakpoint in debugged task. Not creating a TLB now is not |
| 368 | * performance critical. |
| 369 | * |
| 370 | * Both the cases above are not good enough for code churn. |
| 371 | */ |
| 372 | if (current->active_mm != vma->vm_mm) |
| 373 | return; |
| 374 | |
| 375 | local_irq_save(flags); |
| 376 | |
| 377 | tlb_paranoid_check(vma->vm_mm->context.asid, address); |
| 378 | |
| 379 | address &= PAGE_MASK; |
| 380 | |
| 381 | /* update this PTE credentials */ |
| 382 | pte_val(*ptep) |= (_PAGE_PRESENT | _PAGE_ACCESSED); |
| 383 | |
| 384 | /* Create HW TLB entry Flags (in PD0) from PTE Flags */ |
| 385 | #if (CONFIG_ARC_MMU_VER <= 2) |
| 386 | pd0_flags = ((pte_val(*ptep) & PTE_BITS_IN_PD0) >> 1); |
| 387 | #else |
| 388 | pd0_flags = ((pte_val(*ptep) & PTE_BITS_IN_PD0)); |
| 389 | #endif |
| 390 | |
| 391 | /* ASID for this task */ |
| 392 | asid_or_sasid = read_aux_reg(ARC_REG_PID) & 0xff; |
| 393 | |
| 394 | write_aux_reg(ARC_REG_TLBPD0, address | pd0_flags | asid_or_sasid); |
| 395 | |
Vineet Gupta | 64b703e | 2013-06-17 18:12:13 +0530 | [diff] [blame^] | 396 | /* |
| 397 | * ARC MMU provides fully orthogonal access bits for K/U mode, |
| 398 | * however Linux only saves 1 set to save PTE real-estate |
| 399 | * Here we convert 3 PTE bits into 6 MMU bits: |
| 400 | * -Kernel only entries have Kr Kw Kx 0 0 0 |
| 401 | * -User entries have mirrored K and U bits |
| 402 | */ |
| 403 | rwx = pte_val(*ptep) & PTE_BITS_RWX; |
| 404 | |
| 405 | if (pte_val(*ptep) & _PAGE_GLOBAL) |
| 406 | rwx <<= 3; /* r w x => Kr Kw Kx 0 0 0 */ |
| 407 | else |
| 408 | rwx |= (rwx << 3); /* r w x => Kr Kw Kx Ur Uw Ux */ |
| 409 | |
Vineet Gupta | cc562d2 | 2013-01-18 15:12:19 +0530 | [diff] [blame] | 410 | /* Load remaining info in PD1 (Page Frame Addr and Kx/Kw/Kr Flags) */ |
Vineet Gupta | 64b703e | 2013-06-17 18:12:13 +0530 | [diff] [blame^] | 411 | write_aux_reg(ARC_REG_TLBPD1, |
| 412 | rwx | (pte_val(*ptep) & PTE_BITS_NON_RWX_IN_PD1)); |
Vineet Gupta | cc562d2 | 2013-01-18 15:12:19 +0530 | [diff] [blame] | 413 | |
| 414 | /* First verify if entry for this vaddr+ASID already exists */ |
| 415 | write_aux_reg(ARC_REG_TLBCOMMAND, TLBProbe); |
| 416 | idx = read_aux_reg(ARC_REG_TLBINDEX); |
| 417 | |
| 418 | /* |
| 419 | * If Not already present get a free slot from MMU. |
| 420 | * Otherwise, Probe would have located the entry and set INDEX Reg |
| 421 | * with existing location. This will cause Write CMD to over-write |
| 422 | * existing entry with new PD0 and PD1 |
| 423 | */ |
| 424 | if (likely(idx & TLB_LKUP_ERR)) |
| 425 | write_aux_reg(ARC_REG_TLBCOMMAND, TLBGetIndex); |
| 426 | |
| 427 | /* |
| 428 | * Commit the Entry to MMU |
| 429 | * It doesnt sound safe to use the TLBWriteNI cmd here |
| 430 | * which doesn't flush uTLBs. I'd rather be safe than sorry. |
| 431 | */ |
| 432 | write_aux_reg(ARC_REG_TLBCOMMAND, TLBWrite); |
| 433 | |
| 434 | local_irq_restore(flags); |
| 435 | } |
| 436 | |
Vineet Gupta | eacd0e95 | 2013-04-16 14:10:48 +0530 | [diff] [blame] | 437 | /* |
| 438 | * Called at the end of pagefault, for a userspace mapped page |
| 439 | * -pre-install the corresponding TLB entry into MMU |
Vineet Gupta | 4102b53 | 2013-05-09 21:54:51 +0530 | [diff] [blame] | 440 | * -Finalize the delayed D-cache flush of kernel mapping of page due to |
| 441 | * flush_dcache_page(), copy_user_page() |
| 442 | * |
| 443 | * Note that flush (when done) involves both WBACK - so physical page is |
| 444 | * in sync as well as INV - so any non-congruent aliases don't remain |
Vineet Gupta | cc562d2 | 2013-01-18 15:12:19 +0530 | [diff] [blame] | 445 | */ |
Vineet Gupta | 24603fd | 2013-04-11 18:36:35 +0530 | [diff] [blame] | 446 | void update_mmu_cache(struct vm_area_struct *vma, unsigned long vaddr_unaligned, |
Vineet Gupta | cc562d2 | 2013-01-18 15:12:19 +0530 | [diff] [blame] | 447 | pte_t *ptep) |
| 448 | { |
Vineet Gupta | 24603fd | 2013-04-11 18:36:35 +0530 | [diff] [blame] | 449 | unsigned long vaddr = vaddr_unaligned & PAGE_MASK; |
Vineet Gupta | 4102b53 | 2013-05-09 21:54:51 +0530 | [diff] [blame] | 450 | unsigned long paddr = pte_val(*ptep) & PAGE_MASK; |
Vineet Gupta | 29b93c6 | 2013-05-19 15:51:03 +0530 | [diff] [blame] | 451 | struct page *page = pfn_to_page(pte_pfn(*ptep)); |
Vineet Gupta | cc562d2 | 2013-01-18 15:12:19 +0530 | [diff] [blame] | 452 | |
Vineet Gupta | 24603fd | 2013-04-11 18:36:35 +0530 | [diff] [blame] | 453 | create_tlb(vma, vaddr, ptep); |
| 454 | |
Vineet Gupta | 29b93c6 | 2013-05-19 15:51:03 +0530 | [diff] [blame] | 455 | if (page == ZERO_PAGE(0)) { |
| 456 | return; |
| 457 | } |
| 458 | |
Vineet Gupta | 4102b53 | 2013-05-09 21:54:51 +0530 | [diff] [blame] | 459 | /* |
| 460 | * Exec page : Independent of aliasing/page-color considerations, |
| 461 | * since icache doesn't snoop dcache on ARC, any dirty |
| 462 | * K-mapping of a code page needs to be wback+inv so that |
| 463 | * icache fetch by userspace sees code correctly. |
| 464 | * !EXEC page: If K-mapping is NOT congruent to U-mapping, flush it |
| 465 | * so userspace sees the right data. |
| 466 | * (Avoids the flush for Non-exec + congruent mapping case) |
| 467 | */ |
Vineet Gupta | 3e87974 | 2013-05-22 18:38:10 +0530 | [diff] [blame] | 468 | if ((vma->vm_flags & VM_EXEC) || |
| 469 | addr_not_cache_congruent(paddr, vaddr)) { |
Vineet Gupta | eacd0e95 | 2013-04-16 14:10:48 +0530 | [diff] [blame] | 470 | |
Vineet Gupta | 2ed21da | 2013-05-13 17:23:58 +0530 | [diff] [blame] | 471 | int dirty = !test_and_set_bit(PG_dc_clean, &page->flags); |
Vineet Gupta | eacd0e95 | 2013-04-16 14:10:48 +0530 | [diff] [blame] | 472 | if (dirty) { |
Vineet Gupta | 4102b53 | 2013-05-09 21:54:51 +0530 | [diff] [blame] | 473 | /* wback + inv dcache lines */ |
Vineet Gupta | 6ec18a8 | 2013-05-09 15:10:18 +0530 | [diff] [blame] | 474 | __flush_dcache_page(paddr, paddr); |
Vineet Gupta | 4102b53 | 2013-05-09 21:54:51 +0530 | [diff] [blame] | 475 | |
| 476 | /* invalidate any existing icache lines */ |
| 477 | if (vma->vm_flags & VM_EXEC) |
| 478 | __inv_icache_page(paddr, vaddr); |
Vineet Gupta | eacd0e95 | 2013-04-16 14:10:48 +0530 | [diff] [blame] | 479 | } |
Vineet Gupta | 24603fd | 2013-04-11 18:36:35 +0530 | [diff] [blame] | 480 | } |
Vineet Gupta | cc562d2 | 2013-01-18 15:12:19 +0530 | [diff] [blame] | 481 | } |
| 482 | |
| 483 | /* Read the Cache Build Confuration Registers, Decode them and save into |
| 484 | * the cpuinfo structure for later use. |
| 485 | * No Validation is done here, simply read/convert the BCRs |
| 486 | */ |
Paul Gortmaker | ce75995 | 2013-06-24 15:30:15 -0400 | [diff] [blame] | 487 | void read_decode_mmu_bcr(void) |
Vineet Gupta | cc562d2 | 2013-01-18 15:12:19 +0530 | [diff] [blame] | 488 | { |
Vineet Gupta | cc562d2 | 2013-01-18 15:12:19 +0530 | [diff] [blame] | 489 | struct cpuinfo_arc_mmu *mmu = &cpuinfo_arc700[smp_processor_id()].mmu; |
Vineet Gupta | da1677b | 2013-05-14 13:28:17 +0530 | [diff] [blame] | 490 | unsigned int tmp; |
| 491 | struct bcr_mmu_1_2 { |
| 492 | #ifdef CONFIG_CPU_BIG_ENDIAN |
| 493 | unsigned int ver:8, ways:4, sets:4, u_itlb:8, u_dtlb:8; |
| 494 | #else |
| 495 | unsigned int u_dtlb:8, u_itlb:8, sets:4, ways:4, ver:8; |
| 496 | #endif |
| 497 | } *mmu2; |
| 498 | |
| 499 | struct bcr_mmu_3 { |
| 500 | #ifdef CONFIG_CPU_BIG_ENDIAN |
| 501 | unsigned int ver:8, ways:4, sets:4, osm:1, reserv:3, pg_sz:4, |
| 502 | u_itlb:4, u_dtlb:4; |
| 503 | #else |
| 504 | unsigned int u_dtlb:4, u_itlb:4, pg_sz:4, reserv:3, osm:1, sets:4, |
| 505 | ways:4, ver:8; |
| 506 | #endif |
| 507 | } *mmu3; |
Vineet Gupta | cc562d2 | 2013-01-18 15:12:19 +0530 | [diff] [blame] | 508 | |
| 509 | tmp = read_aux_reg(ARC_REG_MMU_BCR); |
| 510 | mmu->ver = (tmp >> 24); |
| 511 | |
| 512 | if (mmu->ver <= 2) { |
| 513 | mmu2 = (struct bcr_mmu_1_2 *)&tmp; |
| 514 | mmu->pg_sz = PAGE_SIZE; |
| 515 | mmu->sets = 1 << mmu2->sets; |
| 516 | mmu->ways = 1 << mmu2->ways; |
| 517 | mmu->u_dtlb = mmu2->u_dtlb; |
| 518 | mmu->u_itlb = mmu2->u_itlb; |
| 519 | } else { |
| 520 | mmu3 = (struct bcr_mmu_3 *)&tmp; |
| 521 | mmu->pg_sz = 512 << mmu3->pg_sz; |
| 522 | mmu->sets = 1 << mmu3->sets; |
| 523 | mmu->ways = 1 << mmu3->ways; |
| 524 | mmu->u_dtlb = mmu3->u_dtlb; |
| 525 | mmu->u_itlb = mmu3->u_itlb; |
| 526 | } |
| 527 | |
| 528 | mmu->num_tlb = mmu->sets * mmu->ways; |
| 529 | } |
| 530 | |
Vineet Gupta | af61742 | 2013-01-18 15:12:24 +0530 | [diff] [blame] | 531 | char *arc_mmu_mumbojumbo(int cpu_id, char *buf, int len) |
| 532 | { |
| 533 | int n = 0; |
Noam Camus | e3edeb6 | 2013-02-26 09:22:46 +0200 | [diff] [blame] | 534 | struct cpuinfo_arc_mmu *p_mmu = &cpuinfo_arc700[cpu_id].mmu; |
Vineet Gupta | af61742 | 2013-01-18 15:12:24 +0530 | [diff] [blame] | 535 | |
| 536 | n += scnprintf(buf + n, len - n, "ARC700 MMU [v%x]\t: %dk PAGE, ", |
| 537 | p_mmu->ver, TO_KB(p_mmu->pg_sz)); |
| 538 | |
| 539 | n += scnprintf(buf + n, len - n, |
| 540 | "J-TLB %d (%dx%d), uDTLB %d, uITLB %d, %s\n", |
| 541 | p_mmu->num_tlb, p_mmu->sets, p_mmu->ways, |
| 542 | p_mmu->u_dtlb, p_mmu->u_itlb, |
Vineet Gupta | 8235703 | 2013-06-01 12:55:42 +0530 | [diff] [blame] | 543 | IS_ENABLED(CONFIG_ARC_MMU_SASID) ? "SASID" : ""); |
Vineet Gupta | af61742 | 2013-01-18 15:12:24 +0530 | [diff] [blame] | 544 | |
| 545 | return buf; |
| 546 | } |
| 547 | |
Paul Gortmaker | ce75995 | 2013-06-24 15:30:15 -0400 | [diff] [blame] | 548 | void arc_mmu_init(void) |
Vineet Gupta | cc562d2 | 2013-01-18 15:12:19 +0530 | [diff] [blame] | 549 | { |
Vineet Gupta | af61742 | 2013-01-18 15:12:24 +0530 | [diff] [blame] | 550 | char str[256]; |
| 551 | struct cpuinfo_arc_mmu *mmu = &cpuinfo_arc700[smp_processor_id()].mmu; |
| 552 | |
| 553 | printk(arc_mmu_mumbojumbo(0, str, sizeof(str))); |
| 554 | |
| 555 | /* For efficiency sake, kernel is compile time built for a MMU ver |
| 556 | * This must match the hardware it is running on. |
| 557 | * Linux built for MMU V2, if run on MMU V1 will break down because V1 |
| 558 | * hardware doesn't understand cmds such as WriteNI, or IVUTLB |
| 559 | * On the other hand, Linux built for V1 if run on MMU V2 will do |
| 560 | * un-needed workarounds to prevent memcpy thrashing. |
| 561 | * Similarly MMU V3 has new features which won't work on older MMU |
| 562 | */ |
| 563 | if (mmu->ver != CONFIG_ARC_MMU_VER) { |
| 564 | panic("MMU ver %d doesn't match kernel built for %d...\n", |
| 565 | mmu->ver, CONFIG_ARC_MMU_VER); |
| 566 | } |
| 567 | |
| 568 | if (mmu->pg_sz != PAGE_SIZE) |
| 569 | panic("MMU pg size != PAGE_SIZE (%luk)\n", TO_KB(PAGE_SIZE)); |
| 570 | |
Vineet Gupta | cc562d2 | 2013-01-18 15:12:19 +0530 | [diff] [blame] | 571 | /* |
| 572 | * ASID mgmt data structures are compile time init |
| 573 | * asid_cache = FIRST_ASID and asid_mm_map[] all zeroes |
| 574 | */ |
| 575 | |
| 576 | local_flush_tlb_all(); |
| 577 | |
| 578 | /* Enable the MMU */ |
| 579 | write_aux_reg(ARC_REG_PID, MMU_ENABLE); |
Vineet Gupta | 41195d2 | 2013-01-18 15:12:23 +0530 | [diff] [blame] | 580 | |
| 581 | /* In smp we use this reg for interrupt 1 scratch */ |
| 582 | #ifndef CONFIG_SMP |
| 583 | /* swapper_pg_dir is the pgd for the kernel, used by vmalloc */ |
| 584 | write_aux_reg(ARC_REG_SCRATCH_DATA0, swapper_pg_dir); |
| 585 | #endif |
Vineet Gupta | cc562d2 | 2013-01-18 15:12:19 +0530 | [diff] [blame] | 586 | } |
| 587 | |
| 588 | /* |
| 589 | * TLB Programmer's Model uses Linear Indexes: 0 to {255, 511} for 128 x {2,4} |
| 590 | * The mapping is Column-first. |
| 591 | * --------------------- ----------- |
| 592 | * |way0|way1|way2|way3| |way0|way1| |
| 593 | * --------------------- ----------- |
| 594 | * [set0] | 0 | 1 | 2 | 3 | | 0 | 1 | |
| 595 | * [set1] | 4 | 5 | 6 | 7 | | 2 | 3 | |
| 596 | * ~ ~ ~ ~ |
| 597 | * [set127] | 508| 509| 510| 511| | 254| 255| |
| 598 | * --------------------- ----------- |
| 599 | * For normal operations we don't(must not) care how above works since |
| 600 | * MMU cmd getIndex(vaddr) abstracts that out. |
| 601 | * However for walking WAYS of a SET, we need to know this |
| 602 | */ |
| 603 | #define SET_WAY_TO_IDX(mmu, set, way) ((set) * mmu->ways + (way)) |
| 604 | |
| 605 | /* Handling of Duplicate PD (TLB entry) in MMU. |
| 606 | * -Could be due to buggy customer tapeouts or obscure kernel bugs |
| 607 | * -MMU complaints not at the time of duplicate PD installation, but at the |
| 608 | * time of lookup matching multiple ways. |
| 609 | * -Ideally these should never happen - but if they do - workaround by deleting |
| 610 | * the duplicate one. |
| 611 | * -Knob to be verbose abt it.(TODO: hook them up to debugfs) |
| 612 | */ |
| 613 | volatile int dup_pd_verbose = 1;/* Be slient abt it or complain (default) */ |
| 614 | |
| 615 | void do_tlb_overlap_fault(unsigned long cause, unsigned long address, |
| 616 | struct pt_regs *regs) |
| 617 | { |
| 618 | int set, way, n; |
| 619 | unsigned int pd0[4], pd1[4]; /* assume max 4 ways */ |
| 620 | unsigned long flags, is_valid; |
| 621 | struct cpuinfo_arc_mmu *mmu = &cpuinfo_arc700[smp_processor_id()].mmu; |
| 622 | |
| 623 | local_irq_save(flags); |
| 624 | |
| 625 | /* re-enable the MMU */ |
| 626 | write_aux_reg(ARC_REG_PID, MMU_ENABLE | read_aux_reg(ARC_REG_PID)); |
| 627 | |
| 628 | /* loop thru all sets of TLB */ |
| 629 | for (set = 0; set < mmu->sets; set++) { |
| 630 | |
| 631 | /* read out all the ways of current set */ |
| 632 | for (way = 0, is_valid = 0; way < mmu->ways; way++) { |
| 633 | write_aux_reg(ARC_REG_TLBINDEX, |
| 634 | SET_WAY_TO_IDX(mmu, set, way)); |
| 635 | write_aux_reg(ARC_REG_TLBCOMMAND, TLBRead); |
| 636 | pd0[way] = read_aux_reg(ARC_REG_TLBPD0); |
| 637 | pd1[way] = read_aux_reg(ARC_REG_TLBPD1); |
| 638 | is_valid |= pd0[way] & _PAGE_PRESENT; |
| 639 | } |
| 640 | |
| 641 | /* If all the WAYS in SET are empty, skip to next SET */ |
| 642 | if (!is_valid) |
| 643 | continue; |
| 644 | |
| 645 | /* Scan the set for duplicate ways: needs a nested loop */ |
| 646 | for (way = 0; way < mmu->ways; way++) { |
| 647 | if (!pd0[way]) |
| 648 | continue; |
| 649 | |
| 650 | for (n = way + 1; n < mmu->ways; n++) { |
| 651 | if ((pd0[way] & PAGE_MASK) == |
| 652 | (pd0[n] & PAGE_MASK)) { |
| 653 | |
| 654 | if (dup_pd_verbose) { |
| 655 | pr_info("Duplicate PD's @" |
| 656 | "[%d:%d]/[%d:%d]\n", |
| 657 | set, way, set, n); |
| 658 | pr_info("TLBPD0[%u]: %08x\n", |
| 659 | way, pd0[way]); |
| 660 | } |
| 661 | |
| 662 | /* |
| 663 | * clear entry @way and not @n. This is |
| 664 | * critical to our optimised loop |
| 665 | */ |
| 666 | pd0[way] = pd1[way] = 0; |
| 667 | write_aux_reg(ARC_REG_TLBINDEX, |
| 668 | SET_WAY_TO_IDX(mmu, set, way)); |
| 669 | __tlb_entry_erase(); |
| 670 | } |
| 671 | } |
| 672 | } |
| 673 | } |
| 674 | |
| 675 | local_irq_restore(flags); |
| 676 | } |
| 677 | |
| 678 | /*********************************************************************** |
| 679 | * Diagnostic Routines |
| 680 | * -Called from Low Level TLB Hanlders if things don;t look good |
| 681 | **********************************************************************/ |
| 682 | |
| 683 | #ifdef CONFIG_ARC_DBG_TLB_PARANOIA |
| 684 | |
| 685 | /* |
| 686 | * Low Level ASM TLB handler calls this if it finds that HW and SW ASIDS |
| 687 | * don't match |
| 688 | */ |
| 689 | void print_asid_mismatch(int is_fast_path) |
| 690 | { |
| 691 | int pid_sw, pid_hw; |
| 692 | pid_sw = current->active_mm->context.asid; |
| 693 | pid_hw = read_aux_reg(ARC_REG_PID) & 0xff; |
| 694 | |
| 695 | pr_emerg("ASID Mismatch in %s Path Handler: sw-pid=0x%x hw-pid=0x%x\n", |
| 696 | is_fast_path ? "Fast" : "Slow", pid_sw, pid_hw); |
| 697 | |
| 698 | __asm__ __volatile__("flag 1"); |
| 699 | } |
| 700 | |
| 701 | void tlb_paranoid_check(unsigned int pid_sw, unsigned long addr) |
| 702 | { |
| 703 | unsigned int pid_hw; |
| 704 | |
| 705 | pid_hw = read_aux_reg(ARC_REG_PID) & 0xff; |
| 706 | |
| 707 | if (addr < 0x70000000 && ((pid_hw != pid_sw) || (pid_sw == NO_ASID))) |
| 708 | print_asid_mismatch(0); |
| 709 | } |
| 710 | #endif |