1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD 3 * 4 * Copyright (C) 2010 Andreas Tobler 5 * All rights reserved. 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 1. Redistributions of source code must retain the above copyright 11 * notice, this list of conditions and the following disclaimer. 12 * 2. Redistributions in binary form must reproduce the above copyright 13 * notice, this list of conditions and the following disclaimer in the 14 * documentation and/or other materials provided with the distribution. 15 * 16 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR 17 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 18 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 19 * IN NO EVENT SHALL TOOLS GMBH BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 20 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, 21 * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; 22 * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, 23 * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR 24 * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF 25 * ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 26 */ 27 28 #include <sys/cdefs.h> 29 __FBSDID("$FreeBSD$"); 30 31 #include <sys/param.h> 32 #include <sys/kernel.h> 33 #include <sys/ktr.h> 34 #include <sys/lock.h> 35 #include <sys/rmlock.h> 36 #include <sys/mutex.h> 37 #include <sys/proc.h> 38 #include <sys/sysctl.h> 39 #include <sys/systm.h> 40 #include <sys/vmmeter.h> 41 42 #include <dev/ofw/openfirm.h> 43 #include <machine/ofw_machdep.h> 44 45 #include <vm/vm.h> 46 #include <vm/vm_param.h> 47 #include <vm/vm_kern.h> 48 #include <vm/vm_page.h> 49 #include <vm/vm_map.h> 50 #include <vm/vm_object.h> 51 #include <vm/vm_extern.h> 52 #include <vm/vm_pageout.h> 53 #include <vm/uma.h> 54 55 #include <powerpc/aim/mmu_oea64.h> 56 57 #include "mmu_if.h" 58 #include "moea64_if.h" 59 60 #include "phyp-hvcall.h" 61 62 #define MMU_PHYP_DEBUG 0 63 #define MMU_PHYP_ID "mmu_phyp: " 64 #if MMU_PHYP_DEBUG 65 #define dprintf(fmt, ...) printf(fmt, ## __VA_ARGS__) 66 #define dprintf0(fmt, ...) dprintf(MMU_PHYP_ID fmt, ## __VA_ARGS__) 67 #else 68 #define dprintf(fmt, args...) do { ; } while(0) 69 #define dprintf0(fmt, args...) do { ; } while(0) 70 #endif 71 72 static struct rmlock mphyp_eviction_lock; 73 74 /* 75 * Kernel MMU interface 76 */ 77 78 static void mphyp_bootstrap(mmu_t mmup, vm_offset_t kernelstart, 79 vm_offset_t kernelend); 80 static void mphyp_cpu_bootstrap(mmu_t mmup, int ap); 81 static int64_t mphyp_pte_synch(mmu_t, struct pvo_entry *pvo); 82 static int64_t mphyp_pte_clear(mmu_t, struct pvo_entry *pvo, uint64_t ptebit); 83 static int64_t mphyp_pte_unset(mmu_t, struct pvo_entry *pvo); 84 static int mphyp_pte_insert(mmu_t, struct pvo_entry *pvo); 85 86 static mmu_method_t mphyp_methods[] = { 87 MMUMETHOD(mmu_bootstrap, mphyp_bootstrap), 88 MMUMETHOD(mmu_cpu_bootstrap, mphyp_cpu_bootstrap), 89 90 MMUMETHOD(moea64_pte_synch, mphyp_pte_synch), 91 MMUMETHOD(moea64_pte_clear, mphyp_pte_clear), 92 MMUMETHOD(moea64_pte_unset, mphyp_pte_unset), 93 MMUMETHOD(moea64_pte_insert, mphyp_pte_insert), 94 95 /* XXX: pmap_copy_page, pmap_init_page with H_PAGE_INIT */ 96 97 { 0, 0 } 98 }; 99 100 MMU_DEF_INHERIT(pseries_mmu, "mmu_phyp", mphyp_methods, 0, oea64_mmu); 101 102 static int brokenkvm = 0; 103 104 static void 105 print_kvm_bug_warning(void *data) 106 { 107 108 if (brokenkvm) 109 printf("WARNING: Running on a broken hypervisor that does " 110 "not support mandatory H_CLEAR_MOD and H_CLEAR_REF " 111 "hypercalls. Performance will be suboptimal.\n"); 112 } 113 114 SYSINIT(kvmbugwarn1, SI_SUB_COPYRIGHT, SI_ORDER_THIRD + 1, 115 print_kvm_bug_warning, NULL); 116 SYSINIT(kvmbugwarn2, SI_SUB_LAST, SI_ORDER_THIRD + 1, print_kvm_bug_warning, 117 NULL); 118 119 static void 120 mphyp_bootstrap(mmu_t mmup, vm_offset_t kernelstart, vm_offset_t kernelend) 121 { 122 uint64_t final_pteg_count = 0; 123 char buf[8]; 124 uint32_t prop[2]; 125 uint32_t nptlp, shift = 0, slb_encoding = 0; 126 uint32_t lp_size, lp_encoding; 127 struct lpte old; 128 uint64_t vsid; 129 phandle_t dev, node, root; 130 int idx, len, res; 131 132 rm_init(&mphyp_eviction_lock, "pte eviction"); 133 134 moea64_early_bootstrap(mmup, kernelstart, kernelend); 135 136 root = OF_peer(0); 137 138 dev = OF_child(root); 139 while (dev != 0) { 140 res = OF_getprop(dev, "name", buf, sizeof(buf)); 141 if (res > 0 && strcmp(buf, "cpus") == 0) 142 break; 143 dev = OF_peer(dev); 144 } 145 146 node = OF_child(dev); 147 148 while (node != 0) { 149 res = OF_getprop(node, "device_type", buf, sizeof(buf)); 150 if (res > 0 && strcmp(buf, "cpu") == 0) 151 break; 152 node = OF_peer(node); 153 } 154 155 res = OF_getencprop(node, "ibm,pft-size", prop, sizeof(prop)); 156 if (res <= 0) 157 panic("mmu_phyp: unknown PFT size"); 158 final_pteg_count = 1 << prop[1]; 159 res = OF_getencprop(node, "ibm,slb-size", prop, sizeof(prop[0])); 160 if (res > 0) 161 n_slbs = prop[0]; 162 dprintf0("slb-size=%i\n", n_slbs); 163 164 moea64_pteg_count = final_pteg_count / sizeof(struct lpteg); 165 166 /* Clear any old page table entries */ 167 for (idx = 0; idx < moea64_pteg_count*8; idx++) { 168 phyp_pft_hcall(H_READ, 0, idx, 0, 0, &old.pte_hi, 169 &old.pte_lo, &old.pte_lo); 170 vsid = (old.pte_hi << (ADDR_API_SHFT64 - ADDR_PIDX_SHFT)) >> 28; 171 if (vsid == VSID_VRMA || vsid == 0 /* Older VRMA */) 172 continue; 173 174 if (old.pte_hi & LPTE_VALID) 175 phyp_hcall(H_REMOVE, 0, idx, 0); 176 } 177 178 /* 179 * Scan the large page size property for PAPR compatible machines. 180 * See PAPR D.5 Changes to Section 5.1.4, 'CPU Node Properties' 181 * for the encoding of the property. 182 */ 183 184 len = OF_getproplen(node, "ibm,segment-page-sizes"); 185 if (len > 0) { 186 /* 187 * We have to use a variable length array on the stack 188 * since we have very limited stack space. 189 */ 190 pcell_t arr[len/sizeof(cell_t)]; 191 res = OF_getencprop(node, "ibm,segment-page-sizes", arr, 192 sizeof(arr)); 193 len /= 4; 194 idx = 0; 195 while (len > 0) { 196 shift = arr[idx]; 197 slb_encoding = arr[idx + 1]; 198 nptlp = arr[idx + 2]; 199 200 dprintf0("Segment Page Size: " 201 "%uKB, slb_enc=0x%X: {size, encoding}[%u] =", 202 shift > 10? 1 << (shift-10) : 0, 203 slb_encoding, nptlp); 204 205 idx += 3; 206 len -= 3; 207 while (len > 0 && nptlp) { 208 lp_size = arr[idx]; 209 lp_encoding = arr[idx+1]; 210 211 dprintf(" {%uKB, 0x%X}", 212 lp_size > 10? 1 << (lp_size-10) : 0, 213 lp_encoding); 214 215 if (slb_encoding == SLBV_L && lp_encoding == 0) 216 break; 217 218 idx += 2; 219 len -= 2; 220 nptlp--; 221 } 222 dprintf("\n"); 223 if (nptlp && slb_encoding == SLBV_L && lp_encoding == 0) 224 break; 225 } 226 227 if (len > 0) { 228 moea64_large_page_shift = shift; 229 moea64_large_page_size = 1ULL << lp_size; 230 moea64_large_page_mask = moea64_large_page_size - 1; 231 hw_direct_map = 1; 232 printf(MMU_PHYP_ID 233 "Support for hugepages of %uKB detected\n", 234 moea64_large_page_shift > 10? 235 1 << (moea64_large_page_shift-10) : 0); 236 } else { 237 moea64_large_page_size = 0; 238 moea64_large_page_shift = 0; 239 moea64_large_page_mask = 0; 240 hw_direct_map = 0; 241 printf(MMU_PHYP_ID 242 "Support for hugepages not found\n"); 243 } 244 } 245 246 moea64_mid_bootstrap(mmup, kernelstart, kernelend); 247 moea64_late_bootstrap(mmup, kernelstart, kernelend); 248 249 /* Test for broken versions of KVM that don't conform to the spec */ 250 if (phyp_hcall(H_CLEAR_MOD, 0, 0) == H_FUNCTION) 251 brokenkvm = 1; 252 } 253 254 static void 255 mphyp_cpu_bootstrap(mmu_t mmup, int ap) 256 { 257 struct slb *slb = PCPU_GET(aim.slb); 258 register_t seg0; 259 int i; 260 261 /* 262 * Install kernel SLB entries 263 */ 264 265 __asm __volatile ("slbia"); 266 __asm __volatile ("slbmfee %0,%1; slbie %0;" : "=r"(seg0) : "r"(0)); 267 for (i = 0; i < 64; i++) { 268 if (!(slb[i].slbe & SLBE_VALID)) 269 continue; 270 271 __asm __volatile ("slbmte %0, %1" :: 272 "r"(slb[i].slbv), "r"(slb[i].slbe)); 273 } 274 } 275 276 static int64_t 277 mphyp_pte_synch(mmu_t mmu, struct pvo_entry *pvo) 278 { 279 struct lpte pte; 280 uint64_t junk; 281 282 __asm __volatile("ptesync"); 283 phyp_pft_hcall(H_READ, 0, pvo->pvo_pte.slot, 0, 0, &pte.pte_hi, 284 &pte.pte_lo, &junk); 285 if ((pte.pte_hi & LPTE_AVPN_MASK) != 286 ((pvo->pvo_vpn >> (ADDR_API_SHFT64 - ADDR_PIDX_SHFT)) & 287 LPTE_AVPN_MASK)) 288 return (-1); 289 if (!(pte.pte_hi & LPTE_VALID)) 290 return (-1); 291 292 return (pte.pte_lo & (LPTE_CHG | LPTE_REF)); 293 } 294 295 static int64_t 296 mphyp_pte_clear(mmu_t mmu, struct pvo_entry *pvo, uint64_t ptebit) 297 { 298 struct rm_priotracker track; 299 int64_t refchg; 300 uint64_t ptelo, junk; 301 int err; 302 303 /* 304 * This involves two steps (synch and clear) so we need the entry 305 * not to change in the middle. We are protected against deliberate 306 * unset by virtue of holding the pmap lock. Protection against 307 * incidental unset (page table eviction) comes from holding the 308 * shared eviction lock. 309 */ 310 PMAP_LOCK_ASSERT(pvo->pvo_pmap, MA_OWNED); 311 rm_rlock(&mphyp_eviction_lock, &track); 312 313 refchg = mphyp_pte_synch(mmu, pvo); 314 if (refchg < 0) { 315 rm_runlock(&mphyp_eviction_lock, &track); 316 return (refchg); 317 } 318 319 if (brokenkvm) { 320 /* 321 * No way to clear either bit, which is total madness. 322 * Pessimistically claim that, once modified, it stays so 323 * forever and that it is never referenced. 324 */ 325 rm_runlock(&mphyp_eviction_lock, &track); 326 return (refchg & ~LPTE_REF); 327 } 328 329 if (ptebit & LPTE_CHG) { 330 err = phyp_pft_hcall(H_CLEAR_MOD, 0, pvo->pvo_pte.slot, 0, 0, 331 &ptelo, &junk, &junk); 332 KASSERT(err == H_SUCCESS, 333 ("Error clearing page change bit: %d", err)); 334 refchg |= (ptelo & LPTE_CHG); 335 } 336 if (ptebit & LPTE_REF) { 337 err = phyp_pft_hcall(H_CLEAR_REF, 0, pvo->pvo_pte.slot, 0, 0, 338 &ptelo, &junk, &junk); 339 KASSERT(err == H_SUCCESS, 340 ("Error clearing page reference bit: %d", err)); 341 refchg |= (ptelo & LPTE_REF); 342 } 343 344 rm_runlock(&mphyp_eviction_lock, &track); 345 346 return (refchg); 347 } 348 349 static int64_t 350 mphyp_pte_unset(mmu_t mmu, struct pvo_entry *pvo) 351 { 352 struct lpte pte; 353 uint64_t junk; 354 int err; 355 356 PMAP_LOCK_ASSERT(pvo->pvo_pmap, MA_OWNED); 357 358 moea64_pte_from_pvo(pvo, &pte); 359 360 err = phyp_pft_hcall(H_REMOVE, H_AVPN, pvo->pvo_pte.slot, 361 pte.pte_hi & LPTE_AVPN_MASK, 0, &pte.pte_hi, &pte.pte_lo, 362 &junk); 363 KASSERT(err == H_SUCCESS || err == H_NOT_FOUND, 364 ("Error removing page: %d", err)); 365 366 if (err == H_NOT_FOUND) { 367 STAT_MOEA64(moea64_pte_overflow--); 368 return (-1); 369 } 370 371 return (pte.pte_lo & (LPTE_REF | LPTE_CHG)); 372 } 373 374 static uintptr_t 375 mphyp_pte_spillable_ident(uintptr_t ptegbase, struct lpte *to_evict) 376 { 377 uint64_t slot, junk, k; 378 struct lpte pt; 379 int i, j; 380 381 /* Start at a random slot */ 382 i = mftb() % 8; 383 k = -1; 384 for (j = 0; j < 8; j++) { 385 slot = ptegbase + (i + j) % 8; 386 phyp_pft_hcall(H_READ, 0, slot, 0, 0, &pt.pte_hi, 387 &pt.pte_lo, &junk); 388 389 if (pt.pte_hi & LPTE_WIRED) 390 continue; 391 392 /* This is a candidate, so remember it */ 393 k = slot; 394 395 /* Try to get a page that has not been used lately */ 396 if (!(pt.pte_hi & LPTE_VALID) || !(pt.pte_lo & LPTE_REF)) { 397 memcpy(to_evict, &pt, sizeof(struct lpte)); 398 return (k); 399 } 400 } 401 402 if (k == -1) 403 return (k); 404 405 phyp_pft_hcall(H_READ, 0, k, 0, 0, &to_evict->pte_hi, 406 &to_evict->pte_lo, &junk); 407 return (k); 408 } 409 410 static int 411 mphyp_pte_insert(mmu_t mmu, struct pvo_entry *pvo) 412 { 413 struct rm_priotracker track; 414 int64_t result; 415 struct lpte evicted, pte; 416 uint64_t index, junk, lastptelo; 417 418 PMAP_LOCK_ASSERT(pvo->pvo_pmap, MA_OWNED); 419 420 /* Initialize PTE */ 421 moea64_pte_from_pvo(pvo, &pte); 422 evicted.pte_hi = 0; 423 424 /* Make sure further insertion is locked out during evictions */ 425 rm_rlock(&mphyp_eviction_lock, &track); 426 427 /* 428 * First try primary hash. 429 */ 430 pvo->pvo_pte.slot &= ~7UL; /* Base slot address */ 431 result = phyp_pft_hcall(H_ENTER, 0, pvo->pvo_pte.slot, pte.pte_hi, 432 pte.pte_lo, &index, &evicted.pte_lo, &junk); 433 if (result == H_SUCCESS) { 434 rm_runlock(&mphyp_eviction_lock, &track); 435 pvo->pvo_pte.slot = index; 436 return (0); 437 } 438 KASSERT(result == H_PTEG_FULL, ("Page insertion error: %ld " 439 "(ptegidx: %#zx/%#lx, PTE %#lx/%#lx", result, pvo->pvo_pte.slot, 440 moea64_pteg_count, pte.pte_hi, pte.pte_lo)); 441 442 /* 443 * Next try secondary hash. 444 */ 445 pvo->pvo_vaddr ^= PVO_HID; 446 pte.pte_hi ^= LPTE_HID; 447 pvo->pvo_pte.slot ^= (moea64_pteg_mask << 3); 448 449 result = phyp_pft_hcall(H_ENTER, 0, pvo->pvo_pte.slot, 450 pte.pte_hi, pte.pte_lo, &index, &evicted.pte_lo, &junk); 451 if (result == H_SUCCESS) { 452 rm_runlock(&mphyp_eviction_lock, &track); 453 pvo->pvo_pte.slot = index; 454 return (0); 455 } 456 KASSERT(result == H_PTEG_FULL, ("Secondary page insertion error: %ld", 457 result)); 458 459 /* 460 * Out of luck. Find a PTE to sacrifice. 461 */ 462 463 /* Lock out all insertions for a bit */ 464 rm_runlock(&mphyp_eviction_lock, &track); 465 rm_wlock(&mphyp_eviction_lock); 466 467 index = mphyp_pte_spillable_ident(pvo->pvo_pte.slot, &evicted); 468 if (index == -1L) { 469 /* Try other hash table? */ 470 pvo->pvo_vaddr ^= PVO_HID; 471 pte.pte_hi ^= LPTE_HID; 472 pvo->pvo_pte.slot ^= (moea64_pteg_mask << 3); 473 index = mphyp_pte_spillable_ident(pvo->pvo_pte.slot, &evicted); 474 } 475 476 if (index == -1L) { 477 /* No freeable slots in either PTEG? We're hosed. */ 478 rm_wunlock(&mphyp_eviction_lock); 479 panic("mphyp_pte_insert: overflow"); 480 return (-1); 481 } 482 483 /* Victim acquired: update page before waving goodbye */ 484 if (evicted.pte_hi & LPTE_VALID) { 485 result = phyp_pft_hcall(H_REMOVE, H_AVPN, index, 486 evicted.pte_hi & LPTE_AVPN_MASK, 0, &junk, &lastptelo, 487 &junk); 488 STAT_MOEA64(moea64_pte_overflow++); 489 KASSERT(result == H_SUCCESS || result == H_NOT_FOUND, 490 ("Error evicting page: %d", (int)result)); 491 } 492 493 /* 494 * Set the new PTE. 495 */ 496 result = phyp_pft_hcall(H_ENTER, H_EXACT, index, pte.pte_hi, 497 pte.pte_lo, &index, &evicted.pte_lo, &junk); 498 rm_wunlock(&mphyp_eviction_lock); /* All clear */ 499 500 pvo->pvo_pte.slot = index; 501 if (result == H_SUCCESS) 502 return (0); 503 504 panic("Page replacement error: %ld", result); 505 return (result); 506 } 507 508