1 // SPDX-License-Identifier: GPL-2.0-only
4 * Copyright 2010 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
7 #include <linux/types.h>
8 #include <linux/string.h>
10 #include <linux/kvm_host.h>
11 #include <linux/highmem.h>
12 #include <linux/gfp.h>
13 #include <linux/slab.h>
14 #include <linux/hugetlb.h>
15 #include <linux/vmalloc.h>
16 #include <linux/srcu.h>
17 #include <linux/anon_inodes.h>
18 #include <linux/file.h>
19 #include <linux/debugfs.h>
21 #include <asm/kvm_ppc.h>
22 #include <asm/kvm_book3s.h>
23 #include <asm/book3s/64/mmu-hash.h>
24 #include <asm/hvcall.h>
25 #include <asm/synch.h>
26 #include <asm/ppc-opcode.h>
27 #include <asm/cputable.h>
28 #include <asm/pte-walk.h>
33 //#define DEBUG_RESIZE_HPT 1
35 #ifdef DEBUG_RESIZE_HPT
36 #define resize_hpt_debug(resize, ...) \
38 printk(KERN_DEBUG "RESIZE HPT %p: ", resize); \
39 printk(__VA_ARGS__); \
42 #define resize_hpt_debug(resize, ...) \
46 static long kvmppc_virtmode_do_h_enter(struct kvm *kvm, unsigned long flags,
47 long pte_index, unsigned long pteh,
48 unsigned long ptel, unsigned long *pte_idx_ret);
50 struct kvm_resize_hpt {
51 /* These fields read-only after init */
53 struct work_struct work;
56 /* These fields protected by kvm->arch.mmu_setup_lock */
58 /* Possible values and their usage:
59 * <0 an error occurred during allocation,
60 * -EBUSY allocation is in the progress,
61 * 0 allocation made successfully.
65 /* Private to the work thread, until error != -EBUSY,
66 * then protected by kvm->arch.mmu_setup_lock.
68 struct kvm_hpt_info hpt;
71 int kvmppc_allocate_hpt(struct kvm_hpt_info *info, u32 order)
73 unsigned long hpt = 0;
75 struct page *page = NULL;
76 struct revmap_entry *rev;
79 if ((order < PPC_MIN_HPT_ORDER) || (order > PPC_MAX_HPT_ORDER))
82 page = kvm_alloc_hpt_cma(1ul << (order - PAGE_SHIFT));
84 hpt = (unsigned long)pfn_to_kaddr(page_to_pfn(page));
85 memset((void *)hpt, 0, (1ul << order));
90 hpt = __get_free_pages(GFP_KERNEL|__GFP_ZERO|__GFP_RETRY_MAYFAIL
91 |__GFP_NOWARN, order - PAGE_SHIFT);
96 /* HPTEs are 2**4 bytes long */
97 npte = 1ul << (order - 4);
99 /* Allocate reverse map array */
100 rev = vmalloc(array_size(npte, sizeof(struct revmap_entry)));
103 kvm_free_hpt_cma(page, 1 << (order - PAGE_SHIFT));
105 free_pages(hpt, order - PAGE_SHIFT);
117 void kvmppc_set_hpt(struct kvm *kvm, struct kvm_hpt_info *info)
119 atomic64_set(&kvm->arch.mmio_update, 0);
120 kvm->arch.hpt = *info;
121 kvm->arch.sdr1 = __pa(info->virt) | (info->order - 18);
123 pr_debug("KVM guest htab at %lx (order %ld), LPID %x\n",
124 info->virt, (long)info->order, kvm->arch.lpid);
127 long kvmppc_alloc_reset_hpt(struct kvm *kvm, int order)
130 struct kvm_hpt_info info;
132 mutex_lock(&kvm->arch.mmu_setup_lock);
133 if (kvm->arch.mmu_ready) {
134 kvm->arch.mmu_ready = 0;
135 /* order mmu_ready vs. vcpus_running */
137 if (atomic_read(&kvm->arch.vcpus_running)) {
138 kvm->arch.mmu_ready = 1;
142 if (kvm_is_radix(kvm)) {
143 err = kvmppc_switch_mmu_to_hpt(kvm);
148 if (kvm->arch.hpt.order == order) {
149 /* We already have a suitable HPT */
151 /* Set the entire HPT to 0, i.e. invalid HPTEs */
152 memset((void *)kvm->arch.hpt.virt, 0, 1ul << order);
154 * Reset all the reverse-mapping chains for all memslots
156 kvmppc_rmap_reset(kvm);
161 if (kvm->arch.hpt.virt) {
162 kvmppc_free_hpt(&kvm->arch.hpt);
163 kvmppc_rmap_reset(kvm);
166 err = kvmppc_allocate_hpt(&info, order);
169 kvmppc_set_hpt(kvm, &info);
173 /* Ensure that each vcpu will flush its TLB on next entry. */
174 cpumask_setall(&kvm->arch.need_tlb_flush);
176 mutex_unlock(&kvm->arch.mmu_setup_lock);
180 void kvmppc_free_hpt(struct kvm_hpt_info *info)
185 kvm_free_hpt_cma(virt_to_page(info->virt),
186 1 << (info->order - PAGE_SHIFT));
188 free_pages(info->virt, info->order - PAGE_SHIFT);
193 /* Bits in first HPTE dword for pagesize 4k, 64k or 16M */
194 static inline unsigned long hpte0_pgsize_encoding(unsigned long pgsize)
196 return (pgsize > 0x1000) ? HPTE_V_LARGE : 0;
199 /* Bits in second HPTE dword for pagesize 4k, 64k or 16M */
200 static inline unsigned long hpte1_pgsize_encoding(unsigned long pgsize)
202 return (pgsize == 0x10000) ? 0x1000 : 0;
205 void kvmppc_map_vrma(struct kvm_vcpu *vcpu, struct kvm_memory_slot *memslot,
206 unsigned long porder)
209 unsigned long npages;
210 unsigned long hp_v, hp_r;
211 unsigned long addr, hash;
213 unsigned long hp0, hp1;
214 unsigned long idx_ret;
216 struct kvm *kvm = vcpu->kvm;
218 psize = 1ul << porder;
219 npages = memslot->npages >> (porder - PAGE_SHIFT);
221 /* VRMA can't be > 1TB */
222 if (npages > 1ul << (40 - porder))
223 npages = 1ul << (40 - porder);
224 /* Can't use more than 1 HPTE per HPTEG */
225 if (npages > kvmppc_hpt_mask(&kvm->arch.hpt) + 1)
226 npages = kvmppc_hpt_mask(&kvm->arch.hpt) + 1;
228 hp0 = HPTE_V_1TB_SEG | (VRMA_VSID << (40 - 16)) |
229 HPTE_V_BOLTED | hpte0_pgsize_encoding(psize);
230 hp1 = hpte1_pgsize_encoding(psize) |
231 HPTE_R_R | HPTE_R_C | HPTE_R_M | PP_RWXX;
233 for (i = 0; i < npages; ++i) {
235 /* can't use hpt_hash since va > 64 bits */
236 hash = (i ^ (VRMA_VSID ^ (VRMA_VSID << 25)))
237 & kvmppc_hpt_mask(&kvm->arch.hpt);
239 * We assume that the hash table is empty and no
240 * vcpus are using it at this stage. Since we create
241 * at most one HPTE per HPTEG, we just assume entry 7
242 * is available and use it.
244 hash = (hash << 3) + 7;
245 hp_v = hp0 | ((addr >> 16) & ~0x7fUL);
247 ret = kvmppc_virtmode_do_h_enter(kvm, H_EXACT, hash, hp_v, hp_r,
249 if (ret != H_SUCCESS) {
250 pr_err("KVM: map_vrma at %lx failed, ret=%ld\n",
257 int kvmppc_mmu_hv_init(void)
259 unsigned long nr_lpids;
261 if (!mmu_has_feature(MMU_FTR_LOCKLESS_TLBIE))
264 if (cpu_has_feature(CPU_FTR_HVMODE)) {
265 if (WARN_ON(mfspr(SPRN_LPID) != 0))
267 nr_lpids = 1UL << mmu_lpid_bits;
269 nr_lpids = 1UL << KVM_MAX_NESTED_GUESTS_SHIFT;
272 if (!cpu_has_feature(CPU_FTR_ARCH_300)) {
273 /* POWER7 has 10-bit LPIDs, POWER8 has 12-bit LPIDs */
274 if (cpu_has_feature(CPU_FTR_ARCH_207S))
275 WARN_ON(nr_lpids != 1UL << 12);
277 WARN_ON(nr_lpids != 1UL << 10);
280 * Reserve the last implemented LPID use in partition
281 * switching for POWER7 and POWER8.
286 kvmppc_init_lpid(nr_lpids);
291 static long kvmppc_virtmode_do_h_enter(struct kvm *kvm, unsigned long flags,
292 long pte_index, unsigned long pteh,
293 unsigned long ptel, unsigned long *pte_idx_ret)
298 ret = kvmppc_do_h_enter(kvm, flags, pte_index, pteh, ptel,
299 kvm->mm->pgd, false, pte_idx_ret);
301 if (ret == H_TOO_HARD) {
302 /* this can't happen */
303 pr_err("KVM: Oops, kvmppc_h_enter returned too hard!\n");
304 ret = H_RESOURCE; /* or something */
310 static struct kvmppc_slb *kvmppc_mmu_book3s_hv_find_slbe(struct kvm_vcpu *vcpu,
316 for (i = 0; i < vcpu->arch.slb_nr; i++) {
317 if (!(vcpu->arch.slb[i].orige & SLB_ESID_V))
320 if (vcpu->arch.slb[i].origv & SLB_VSID_B_1T)
325 if (((vcpu->arch.slb[i].orige ^ eaddr) & mask) == 0)
326 return &vcpu->arch.slb[i];
331 static unsigned long kvmppc_mmu_get_real_addr(unsigned long v, unsigned long r,
334 unsigned long ra_mask;
336 ra_mask = kvmppc_actual_pgsz(v, r) - 1;
337 return (r & HPTE_R_RPN & ~ra_mask) | (ea & ra_mask);
340 static int kvmppc_mmu_book3s_64_hv_xlate(struct kvm_vcpu *vcpu, gva_t eaddr,
341 struct kvmppc_pte *gpte, bool data, bool iswrite)
343 struct kvm *kvm = vcpu->kvm;
344 struct kvmppc_slb *slbe;
346 unsigned long pp, key;
347 unsigned long v, orig_v, gr;
350 int virtmode = vcpu->arch.shregs.msr & (data ? MSR_DR : MSR_IR);
352 if (kvm_is_radix(vcpu->kvm))
353 return kvmppc_mmu_radix_xlate(vcpu, eaddr, gpte, data, iswrite);
357 slbe = kvmppc_mmu_book3s_hv_find_slbe(vcpu, eaddr);
362 /* real mode access */
363 slb_v = vcpu->kvm->arch.vrma_slb_v;
367 /* Find the HPTE in the hash table */
368 index = kvmppc_hv_find_lock_hpte(kvm, eaddr, slb_v,
369 HPTE_V_VALID | HPTE_V_ABSENT);
374 hptep = (__be64 *)(kvm->arch.hpt.virt + (index << 4));
375 v = orig_v = be64_to_cpu(hptep[0]) & ~HPTE_V_HVLOCK;
376 if (cpu_has_feature(CPU_FTR_ARCH_300))
377 v = hpte_new_to_old_v(v, be64_to_cpu(hptep[1]));
378 gr = kvm->arch.hpt.rev[index].guest_rpte;
380 unlock_hpte(hptep, orig_v);
384 gpte->vpage = ((v & HPTE_V_AVPN) << 4) | ((eaddr >> 12) & 0xfff);
386 /* Get PP bits and key for permission check */
387 pp = gr & (HPTE_R_PP0 | HPTE_R_PP);
388 key = (vcpu->arch.shregs.msr & MSR_PR) ? SLB_VSID_KP : SLB_VSID_KS;
391 /* Calculate permissions */
392 gpte->may_read = hpte_read_permission(pp, key);
393 gpte->may_write = hpte_write_permission(pp, key);
394 gpte->may_execute = gpte->may_read && !(gr & (HPTE_R_N | HPTE_R_G));
396 /* Storage key permission check for POWER7 */
397 if (data && virtmode) {
398 int amrfield = hpte_get_skey_perm(gr, vcpu->arch.amr);
405 /* Get the guest physical address */
406 gpte->raddr = kvmppc_mmu_get_real_addr(v, gr, eaddr);
411 * Quick test for whether an instruction is a load or a store.
412 * If the instruction is a load or a store, then this will indicate
413 * which it is, at least on server processors. (Embedded processors
414 * have some external PID instructions that don't follow the rule
415 * embodied here.) If the instruction isn't a load or store, then
416 * this doesn't return anything useful.
418 static int instruction_is_store(unsigned int instr)
423 if ((instr & 0xfc000000) == 0x7c000000)
424 mask = 0x100; /* major opcode 31 */
425 return (instr & mask) != 0;
428 int kvmppc_hv_emulate_mmio(struct kvm_vcpu *vcpu,
429 unsigned long gpa, gva_t ea, int is_store)
434 * Fast path - check if the guest physical address corresponds to a
435 * device on the FAST_MMIO_BUS, if so we can avoid loading the
436 * instruction all together, then we can just handle it and return.
441 idx = srcu_read_lock(&vcpu->kvm->srcu);
442 ret = kvm_io_bus_write(vcpu, KVM_FAST_MMIO_BUS, (gpa_t) gpa, 0,
444 srcu_read_unlock(&vcpu->kvm->srcu, idx);
446 kvmppc_set_pc(vcpu, kvmppc_get_pc(vcpu) + 4);
452 * If we fail, we just return to the guest and try executing it again.
454 if (kvmppc_get_last_inst(vcpu, INST_GENERIC, &last_inst) !=
459 * WARNING: We do not know for sure whether the instruction we just
460 * read from memory is the same that caused the fault in the first
461 * place. If the instruction we read is neither an load or a store,
462 * then it can't access memory, so we don't need to worry about
463 * enforcing access permissions. So, assuming it is a load or
464 * store, we just check that its direction (load or store) is
465 * consistent with the original fault, since that's what we
466 * checked the access permissions against. If there is a mismatch
467 * we just return and retry the instruction.
470 if (instruction_is_store(last_inst) != !!is_store)
474 * Emulated accesses are emulated by looking at the hash for
475 * translation once, then performing the access later. The
476 * translation could be invalidated in the meantime in which
477 * point performing the subsequent memory access on the old
478 * physical address could possibly be a security hole for the
479 * guest (but not the host).
481 * This is less of an issue for MMIO stores since they aren't
482 * globally visible. It could be an issue for MMIO loads to
483 * a certain extent but we'll ignore it for now.
486 vcpu->arch.paddr_accessed = gpa;
487 vcpu->arch.vaddr_accessed = ea;
488 return kvmppc_emulate_mmio(vcpu);
491 int kvmppc_book3s_hv_page_fault(struct kvm_vcpu *vcpu,
492 unsigned long ea, unsigned long dsisr)
494 struct kvm *kvm = vcpu->kvm;
495 unsigned long hpte[3], r;
496 unsigned long hnow_v, hnow_r;
498 unsigned long mmu_seq, psize, pte_size;
499 unsigned long gpa_base, gfn_base;
500 unsigned long gpa, gfn, hva, pfn, hpa;
501 struct kvm_memory_slot *memslot;
503 struct revmap_entry *rev;
507 bool writing, write_ok;
509 unsigned long rcbits;
513 if (kvm_is_radix(kvm))
514 return kvmppc_book3s_radix_page_fault(vcpu, ea, dsisr);
517 * Real-mode code has already searched the HPT and found the
518 * entry we're interested in. Lock the entry and check that
519 * it hasn't changed. If it has, just return and re-execute the
522 if (ea != vcpu->arch.pgfault_addr)
525 if (vcpu->arch.pgfault_cache) {
526 mmio_update = atomic64_read(&kvm->arch.mmio_update);
527 if (mmio_update == vcpu->arch.pgfault_cache->mmio_update) {
528 r = vcpu->arch.pgfault_cache->rpte;
529 psize = kvmppc_actual_pgsz(vcpu->arch.pgfault_hpte[0],
531 gpa_base = r & HPTE_R_RPN & ~(psize - 1);
532 gfn_base = gpa_base >> PAGE_SHIFT;
533 gpa = gpa_base | (ea & (psize - 1));
534 return kvmppc_hv_emulate_mmio(vcpu, gpa, ea,
535 dsisr & DSISR_ISSTORE);
538 index = vcpu->arch.pgfault_index;
539 hptep = (__be64 *)(kvm->arch.hpt.virt + (index << 4));
540 rev = &kvm->arch.hpt.rev[index];
542 while (!try_lock_hpte(hptep, HPTE_V_HVLOCK))
544 hpte[0] = be64_to_cpu(hptep[0]) & ~HPTE_V_HVLOCK;
545 hpte[1] = be64_to_cpu(hptep[1]);
546 hpte[2] = r = rev->guest_rpte;
547 unlock_hpte(hptep, hpte[0]);
550 if (cpu_has_feature(CPU_FTR_ARCH_300)) {
551 hpte[0] = hpte_new_to_old_v(hpte[0], hpte[1]);
552 hpte[1] = hpte_new_to_old_r(hpte[1]);
554 if (hpte[0] != vcpu->arch.pgfault_hpte[0] ||
555 hpte[1] != vcpu->arch.pgfault_hpte[1])
558 /* Translate the logical address and get the page */
559 psize = kvmppc_actual_pgsz(hpte[0], r);
560 gpa_base = r & HPTE_R_RPN & ~(psize - 1);
561 gfn_base = gpa_base >> PAGE_SHIFT;
562 gpa = gpa_base | (ea & (psize - 1));
563 gfn = gpa >> PAGE_SHIFT;
564 memslot = gfn_to_memslot(kvm, gfn);
566 trace_kvm_page_fault_enter(vcpu, hpte, memslot, ea, dsisr);
568 /* No memslot means it's an emulated MMIO region */
569 if (!memslot || (memslot->flags & KVM_MEMSLOT_INVALID))
570 return kvmppc_hv_emulate_mmio(vcpu, gpa, ea,
571 dsisr & DSISR_ISSTORE);
574 * This should never happen, because of the slot_is_aligned()
575 * check in kvmppc_do_h_enter().
577 if (gfn_base < memslot->base_gfn)
580 /* used to check for invalidations in progress */
581 mmu_seq = kvm->mmu_notifier_seq;
586 writing = (dsisr & DSISR_ISSTORE) != 0;
587 /* If writing != 0, then the HPTE must allow writing, if we get here */
589 hva = gfn_to_hva_memslot(memslot, gfn);
592 * Do a fast check first, since __gfn_to_pfn_memslot doesn't
593 * do it with !atomic && !async, which is how we call it.
594 * We always ask for write permission since the common case
595 * is that the page is writable.
597 if (get_user_page_fast_only(hva, FOLL_WRITE, &page)) {
600 /* Call KVM generic code to do the slow-path check */
601 pfn = __gfn_to_pfn_memslot(memslot, gfn, false, NULL,
602 writing, &write_ok, NULL);
603 if (is_error_noslot_pfn(pfn))
606 if (pfn_valid(pfn)) {
607 page = pfn_to_page(pfn);
608 if (PageReserved(page))
614 * Read the PTE from the process' radix tree and use that
615 * so we get the shift and attribute bits.
617 spin_lock(&kvm->mmu_lock);
618 ptep = find_kvm_host_pte(kvm, mmu_seq, hva, &shift);
621 pte = READ_ONCE(*ptep);
622 spin_unlock(&kvm->mmu_lock);
624 * If the PTE disappeared temporarily due to a THP
625 * collapse, just return and let the guest try again.
627 if (!pte_present(pte)) {
632 hpa = pte_pfn(pte) << PAGE_SHIFT;
633 pte_size = PAGE_SIZE;
635 pte_size = 1ul << shift;
638 if (psize > pte_size)
640 if (pte_size > psize)
641 hpa |= hva & (pte_size - psize);
643 /* Check WIMG vs. the actual page we're accessing */
644 if (!hpte_cache_flags_ok(r, is_ci)) {
648 * Allow guest to map emulated device memory as
649 * uncacheable, but actually make it cacheable.
651 r = (r & ~(HPTE_R_W|HPTE_R_I|HPTE_R_G)) | HPTE_R_M;
655 * Set the HPTE to point to hpa.
656 * Since the hpa is at PAGE_SIZE granularity, make sure we
657 * don't mask out lower-order bits if psize < PAGE_SIZE.
659 if (psize < PAGE_SIZE)
661 r = (r & HPTE_R_KEY_HI) | (r & ~(HPTE_R_PP0 - psize)) | hpa;
662 if (hpte_is_writable(r) && !write_ok)
663 r = hpte_make_readonly(r);
666 while (!try_lock_hpte(hptep, HPTE_V_HVLOCK))
668 hnow_v = be64_to_cpu(hptep[0]);
669 hnow_r = be64_to_cpu(hptep[1]);
670 if (cpu_has_feature(CPU_FTR_ARCH_300)) {
671 hnow_v = hpte_new_to_old_v(hnow_v, hnow_r);
672 hnow_r = hpte_new_to_old_r(hnow_r);
676 * If the HPT is being resized, don't update the HPTE,
677 * instead let the guest retry after the resize operation is complete.
678 * The synchronization for mmu_ready test vs. set is provided
681 if (!kvm->arch.mmu_ready)
684 if ((hnow_v & ~HPTE_V_HVLOCK) != hpte[0] || hnow_r != hpte[1] ||
685 rev->guest_rpte != hpte[2])
686 /* HPTE has been changed under us; let the guest retry */
688 hpte[0] = (hpte[0] & ~HPTE_V_ABSENT) | HPTE_V_VALID;
690 /* Always put the HPTE in the rmap chain for the page base address */
691 rmap = &memslot->arch.rmap[gfn_base - memslot->base_gfn];
694 /* Check if we might have been invalidated; let the guest retry if so */
696 if (mmu_notifier_retry(vcpu->kvm, mmu_seq)) {
701 /* Only set R/C in real HPTE if set in both *rmap and guest_rpte */
702 rcbits = *rmap >> KVMPPC_RMAP_RC_SHIFT;
703 r &= rcbits | ~(HPTE_R_R | HPTE_R_C);
705 if (be64_to_cpu(hptep[0]) & HPTE_V_VALID) {
706 /* HPTE was previously valid, so we need to invalidate it */
708 hptep[0] |= cpu_to_be64(HPTE_V_ABSENT);
709 kvmppc_invalidate_hpte(kvm, hptep, index);
710 /* don't lose previous R and C bits */
711 r |= be64_to_cpu(hptep[1]) & (HPTE_R_R | HPTE_R_C);
713 kvmppc_add_revmap_chain(kvm, rev, rmap, index, 0);
716 if (cpu_has_feature(CPU_FTR_ARCH_300)) {
717 r = hpte_old_to_new_r(hpte[0], r);
718 hpte[0] = hpte_old_to_new_v(hpte[0]);
720 hptep[1] = cpu_to_be64(r);
722 __unlock_hpte(hptep, hpte[0]);
723 asm volatile("ptesync" : : : "memory");
725 if (page && hpte_is_writable(r))
726 set_page_dirty_lock(page);
729 trace_kvm_page_fault_exit(vcpu, hpte, ret);
736 __unlock_hpte(hptep, be64_to_cpu(hptep[0]));
741 void kvmppc_rmap_reset(struct kvm *kvm)
743 struct kvm_memslots *slots;
744 struct kvm_memory_slot *memslot;
747 srcu_idx = srcu_read_lock(&kvm->srcu);
748 slots = kvm_memslots(kvm);
749 kvm_for_each_memslot(memslot, bkt, slots) {
750 /* Mutual exclusion with kvm_unmap_hva_range etc. */
751 spin_lock(&kvm->mmu_lock);
753 * This assumes it is acceptable to lose reference and
754 * change bits across a reset.
756 memset(memslot->arch.rmap, 0,
757 memslot->npages * sizeof(*memslot->arch.rmap));
758 spin_unlock(&kvm->mmu_lock);
760 srcu_read_unlock(&kvm->srcu, srcu_idx);
763 /* Must be called with both HPTE and rmap locked */
764 static void kvmppc_unmap_hpte(struct kvm *kvm, unsigned long i,
765 struct kvm_memory_slot *memslot,
766 unsigned long *rmapp, unsigned long gfn)
768 __be64 *hptep = (__be64 *) (kvm->arch.hpt.virt + (i << 4));
769 struct revmap_entry *rev = kvm->arch.hpt.rev;
771 unsigned long ptel, psize, rcbits;
775 /* chain is now empty */
776 *rmapp &= ~(KVMPPC_RMAP_PRESENT | KVMPPC_RMAP_INDEX);
778 /* remove i from chain */
782 rev[i].forw = rev[i].back = i;
783 *rmapp = (*rmapp & ~KVMPPC_RMAP_INDEX) | j;
786 /* Now check and modify the HPTE */
787 ptel = rev[i].guest_rpte;
788 psize = kvmppc_actual_pgsz(be64_to_cpu(hptep[0]), ptel);
789 if ((be64_to_cpu(hptep[0]) & HPTE_V_VALID) &&
790 hpte_rpn(ptel, psize) == gfn) {
791 hptep[0] |= cpu_to_be64(HPTE_V_ABSENT);
792 kvmppc_invalidate_hpte(kvm, hptep, i);
793 hptep[1] &= ~cpu_to_be64(HPTE_R_KEY_HI | HPTE_R_KEY_LO);
794 /* Harvest R and C */
795 rcbits = be64_to_cpu(hptep[1]) & (HPTE_R_R | HPTE_R_C);
796 *rmapp |= rcbits << KVMPPC_RMAP_RC_SHIFT;
797 if ((rcbits & HPTE_R_C) && memslot->dirty_bitmap)
798 kvmppc_update_dirty_map(memslot, gfn, psize);
799 if (rcbits & ~rev[i].guest_rpte) {
800 rev[i].guest_rpte = ptel | rcbits;
801 note_hpte_modification(kvm, &rev[i]);
806 static void kvm_unmap_rmapp(struct kvm *kvm, struct kvm_memory_slot *memslot,
811 unsigned long *rmapp;
813 rmapp = &memslot->arch.rmap[gfn - memslot->base_gfn];
816 if (!(*rmapp & KVMPPC_RMAP_PRESENT)) {
822 * To avoid an ABBA deadlock with the HPTE lock bit,
823 * we can't spin on the HPTE lock while holding the
826 i = *rmapp & KVMPPC_RMAP_INDEX;
827 hptep = (__be64 *) (kvm->arch.hpt.virt + (i << 4));
828 if (!try_lock_hpte(hptep, HPTE_V_HVLOCK)) {
829 /* unlock rmap before spinning on the HPTE lock */
831 while (be64_to_cpu(hptep[0]) & HPTE_V_HVLOCK)
836 kvmppc_unmap_hpte(kvm, i, memslot, rmapp, gfn);
838 __unlock_hpte(hptep, be64_to_cpu(hptep[0]));
842 bool kvm_unmap_gfn_range_hv(struct kvm *kvm, struct kvm_gfn_range *range)
846 if (kvm_is_radix(kvm)) {
847 for (gfn = range->start; gfn < range->end; gfn++)
848 kvm_unmap_radix(kvm, range->slot, gfn);
850 for (gfn = range->start; gfn < range->end; gfn++)
851 kvm_unmap_rmapp(kvm, range->slot, gfn);
857 void kvmppc_core_flush_memslot_hv(struct kvm *kvm,
858 struct kvm_memory_slot *memslot)
862 unsigned long *rmapp;
864 gfn = memslot->base_gfn;
865 rmapp = memslot->arch.rmap;
866 if (kvm_is_radix(kvm)) {
867 kvmppc_radix_flush_memslot(kvm, memslot);
871 for (n = memslot->npages; n; --n, ++gfn) {
873 * Testing the present bit without locking is OK because
874 * the memslot has been marked invalid already, and hence
875 * no new HPTEs referencing this page can be created,
876 * thus the present bit can't go from 0 to 1.
878 if (*rmapp & KVMPPC_RMAP_PRESENT)
879 kvm_unmap_rmapp(kvm, memslot, gfn);
884 static bool kvm_age_rmapp(struct kvm *kvm, struct kvm_memory_slot *memslot,
887 struct revmap_entry *rev = kvm->arch.hpt.rev;
888 unsigned long head, i, j;
891 unsigned long *rmapp;
893 rmapp = &memslot->arch.rmap[gfn - memslot->base_gfn];
896 if (*rmapp & KVMPPC_RMAP_REFERENCED) {
897 *rmapp &= ~KVMPPC_RMAP_REFERENCED;
900 if (!(*rmapp & KVMPPC_RMAP_PRESENT)) {
905 i = head = *rmapp & KVMPPC_RMAP_INDEX;
907 hptep = (__be64 *) (kvm->arch.hpt.virt + (i << 4));
910 /* If this HPTE isn't referenced, ignore it */
911 if (!(be64_to_cpu(hptep[1]) & HPTE_R_R))
914 if (!try_lock_hpte(hptep, HPTE_V_HVLOCK)) {
915 /* unlock rmap before spinning on the HPTE lock */
917 while (be64_to_cpu(hptep[0]) & HPTE_V_HVLOCK)
922 /* Now check and modify the HPTE */
923 if ((be64_to_cpu(hptep[0]) & HPTE_V_VALID) &&
924 (be64_to_cpu(hptep[1]) & HPTE_R_R)) {
925 kvmppc_clear_ref_hpte(kvm, hptep, i);
926 if (!(rev[i].guest_rpte & HPTE_R_R)) {
927 rev[i].guest_rpte |= HPTE_R_R;
928 note_hpte_modification(kvm, &rev[i]);
932 __unlock_hpte(hptep, be64_to_cpu(hptep[0]));
933 } while ((i = j) != head);
939 bool kvm_age_gfn_hv(struct kvm *kvm, struct kvm_gfn_range *range)
944 if (kvm_is_radix(kvm)) {
945 for (gfn = range->start; gfn < range->end; gfn++)
946 ret |= kvm_age_radix(kvm, range->slot, gfn);
948 for (gfn = range->start; gfn < range->end; gfn++)
949 ret |= kvm_age_rmapp(kvm, range->slot, gfn);
955 static bool kvm_test_age_rmapp(struct kvm *kvm, struct kvm_memory_slot *memslot,
958 struct revmap_entry *rev = kvm->arch.hpt.rev;
959 unsigned long head, i, j;
962 unsigned long *rmapp;
964 rmapp = &memslot->arch.rmap[gfn - memslot->base_gfn];
965 if (*rmapp & KVMPPC_RMAP_REFERENCED)
969 if (*rmapp & KVMPPC_RMAP_REFERENCED)
972 if (*rmapp & KVMPPC_RMAP_PRESENT) {
973 i = head = *rmapp & KVMPPC_RMAP_INDEX;
975 hp = (unsigned long *)(kvm->arch.hpt.virt + (i << 4));
977 if (be64_to_cpu(hp[1]) & HPTE_R_R)
979 } while ((i = j) != head);
988 bool kvm_test_age_gfn_hv(struct kvm *kvm, struct kvm_gfn_range *range)
990 WARN_ON(range->start + 1 != range->end);
992 if (kvm_is_radix(kvm))
993 return kvm_test_age_radix(kvm, range->slot, range->start);
995 return kvm_test_age_rmapp(kvm, range->slot, range->start);
998 bool kvm_set_spte_gfn_hv(struct kvm *kvm, struct kvm_gfn_range *range)
1000 WARN_ON(range->start + 1 != range->end);
1002 if (kvm_is_radix(kvm))
1003 kvm_unmap_radix(kvm, range->slot, range->start);
1005 kvm_unmap_rmapp(kvm, range->slot, range->start);
1010 static int vcpus_running(struct kvm *kvm)
1012 return atomic_read(&kvm->arch.vcpus_running) != 0;
1016 * Returns the number of system pages that are dirty.
1017 * This can be more than 1 if we find a huge-page HPTE.
1019 static int kvm_test_clear_dirty_npages(struct kvm *kvm, unsigned long *rmapp)
1021 struct revmap_entry *rev = kvm->arch.hpt.rev;
1022 unsigned long head, i, j;
1026 int npages_dirty = 0;
1030 if (!(*rmapp & KVMPPC_RMAP_PRESENT)) {
1032 return npages_dirty;
1035 i = head = *rmapp & KVMPPC_RMAP_INDEX;
1037 unsigned long hptep1;
1038 hptep = (__be64 *) (kvm->arch.hpt.virt + (i << 4));
1042 * Checking the C (changed) bit here is racy since there
1043 * is no guarantee about when the hardware writes it back.
1044 * If the HPTE is not writable then it is stable since the
1045 * page can't be written to, and we would have done a tlbie
1046 * (which forces the hardware to complete any writeback)
1047 * when making the HPTE read-only.
1048 * If vcpus are running then this call is racy anyway
1049 * since the page could get dirtied subsequently, so we
1050 * expect there to be a further call which would pick up
1051 * any delayed C bit writeback.
1052 * Otherwise we need to do the tlbie even if C==0 in
1053 * order to pick up any delayed writeback of C.
1055 hptep1 = be64_to_cpu(hptep[1]);
1056 if (!(hptep1 & HPTE_R_C) &&
1057 (!hpte_is_writable(hptep1) || vcpus_running(kvm)))
1060 if (!try_lock_hpte(hptep, HPTE_V_HVLOCK)) {
1061 /* unlock rmap before spinning on the HPTE lock */
1063 while (hptep[0] & cpu_to_be64(HPTE_V_HVLOCK))
1068 /* Now check and modify the HPTE */
1069 if (!(hptep[0] & cpu_to_be64(HPTE_V_VALID))) {
1070 __unlock_hpte(hptep, be64_to_cpu(hptep[0]));
1074 /* need to make it temporarily absent so C is stable */
1075 hptep[0] |= cpu_to_be64(HPTE_V_ABSENT);
1076 kvmppc_invalidate_hpte(kvm, hptep, i);
1077 v = be64_to_cpu(hptep[0]);
1078 r = be64_to_cpu(hptep[1]);
1080 hptep[1] = cpu_to_be64(r & ~HPTE_R_C);
1081 if (!(rev[i].guest_rpte & HPTE_R_C)) {
1082 rev[i].guest_rpte |= HPTE_R_C;
1083 note_hpte_modification(kvm, &rev[i]);
1085 n = kvmppc_actual_pgsz(v, r);
1086 n = (n + PAGE_SIZE - 1) >> PAGE_SHIFT;
1087 if (n > npages_dirty)
1091 v &= ~HPTE_V_ABSENT;
1093 __unlock_hpte(hptep, v);
1094 } while ((i = j) != head);
1097 return npages_dirty;
1100 void kvmppc_harvest_vpa_dirty(struct kvmppc_vpa *vpa,
1101 struct kvm_memory_slot *memslot,
1106 if (!vpa->dirty || !vpa->pinned_addr)
1108 gfn = vpa->gpa >> PAGE_SHIFT;
1109 if (gfn < memslot->base_gfn ||
1110 gfn >= memslot->base_gfn + memslot->npages)
1115 __set_bit_le(gfn - memslot->base_gfn, map);
1118 long kvmppc_hv_get_dirty_log_hpt(struct kvm *kvm,
1119 struct kvm_memory_slot *memslot, unsigned long *map)
1122 unsigned long *rmapp;
1125 rmapp = memslot->arch.rmap;
1126 for (i = 0; i < memslot->npages; ++i) {
1127 int npages = kvm_test_clear_dirty_npages(kvm, rmapp);
1129 * Note that if npages > 0 then i must be a multiple of npages,
1130 * since we always put huge-page HPTEs in the rmap chain
1131 * corresponding to their page base address.
1134 set_dirty_bits(map, i, npages);
1141 void *kvmppc_pin_guest_page(struct kvm *kvm, unsigned long gpa,
1142 unsigned long *nb_ret)
1144 struct kvm_memory_slot *memslot;
1145 unsigned long gfn = gpa >> PAGE_SHIFT;
1146 struct page *page, *pages[1];
1148 unsigned long hva, offset;
1151 srcu_idx = srcu_read_lock(&kvm->srcu);
1152 memslot = gfn_to_memslot(kvm, gfn);
1153 if (!memslot || (memslot->flags & KVM_MEMSLOT_INVALID))
1155 hva = gfn_to_hva_memslot(memslot, gfn);
1156 npages = get_user_pages_fast(hva, 1, FOLL_WRITE, pages);
1160 srcu_read_unlock(&kvm->srcu, srcu_idx);
1162 offset = gpa & (PAGE_SIZE - 1);
1164 *nb_ret = PAGE_SIZE - offset;
1165 return page_address(page) + offset;
1168 srcu_read_unlock(&kvm->srcu, srcu_idx);
1172 void kvmppc_unpin_guest_page(struct kvm *kvm, void *va, unsigned long gpa,
1175 struct page *page = virt_to_page(va);
1176 struct kvm_memory_slot *memslot;
1185 /* We need to mark this page dirty in the memslot dirty_bitmap, if any */
1186 gfn = gpa >> PAGE_SHIFT;
1187 srcu_idx = srcu_read_lock(&kvm->srcu);
1188 memslot = gfn_to_memslot(kvm, gfn);
1189 if (memslot && memslot->dirty_bitmap)
1190 set_bit_le(gfn - memslot->base_gfn, memslot->dirty_bitmap);
1191 srcu_read_unlock(&kvm->srcu, srcu_idx);
1197 static int resize_hpt_allocate(struct kvm_resize_hpt *resize)
1201 rc = kvmppc_allocate_hpt(&resize->hpt, resize->order);
1205 resize_hpt_debug(resize, "resize_hpt_allocate(): HPT @ 0x%lx\n",
1211 static unsigned long resize_hpt_rehash_hpte(struct kvm_resize_hpt *resize,
1214 struct kvm *kvm = resize->kvm;
1215 struct kvm_hpt_info *old = &kvm->arch.hpt;
1216 struct kvm_hpt_info *new = &resize->hpt;
1217 unsigned long old_hash_mask = (1ULL << (old->order - 7)) - 1;
1218 unsigned long new_hash_mask = (1ULL << (new->order - 7)) - 1;
1219 __be64 *hptep, *new_hptep;
1220 unsigned long vpte, rpte, guest_rpte;
1222 struct revmap_entry *rev;
1223 unsigned long apsize, avpn, pteg, hash;
1224 unsigned long new_idx, new_pteg, replace_vpte;
1227 hptep = (__be64 *)(old->virt + (idx << 4));
1229 /* Guest is stopped, so new HPTEs can't be added or faulted
1230 * in, only unmapped or altered by host actions. So, it's
1231 * safe to check this before we take the HPTE lock */
1232 vpte = be64_to_cpu(hptep[0]);
1233 if (!(vpte & HPTE_V_VALID) && !(vpte & HPTE_V_ABSENT))
1234 return 0; /* nothing to do */
1236 while (!try_lock_hpte(hptep, HPTE_V_HVLOCK))
1239 vpte = be64_to_cpu(hptep[0]);
1242 if (!(vpte & HPTE_V_VALID) && !(vpte & HPTE_V_ABSENT))
1246 if (cpu_has_feature(CPU_FTR_ARCH_300)) {
1247 rpte = be64_to_cpu(hptep[1]);
1248 vpte = hpte_new_to_old_v(vpte, rpte);
1252 rev = &old->rev[idx];
1253 guest_rpte = rev->guest_rpte;
1256 apsize = kvmppc_actual_pgsz(vpte, guest_rpte);
1260 if (vpte & HPTE_V_VALID) {
1261 unsigned long gfn = hpte_rpn(guest_rpte, apsize);
1262 int srcu_idx = srcu_read_lock(&kvm->srcu);
1263 struct kvm_memory_slot *memslot =
1264 __gfn_to_memslot(kvm_memslots(kvm), gfn);
1267 unsigned long *rmapp;
1268 rmapp = &memslot->arch.rmap[gfn - memslot->base_gfn];
1271 kvmppc_unmap_hpte(kvm, idx, memslot, rmapp, gfn);
1275 srcu_read_unlock(&kvm->srcu, srcu_idx);
1278 /* Reload PTE after unmap */
1279 vpte = be64_to_cpu(hptep[0]);
1280 BUG_ON(vpte & HPTE_V_VALID);
1281 BUG_ON(!(vpte & HPTE_V_ABSENT));
1284 if (!(vpte & HPTE_V_BOLTED))
1287 rpte = be64_to_cpu(hptep[1]);
1289 if (cpu_has_feature(CPU_FTR_ARCH_300)) {
1290 vpte = hpte_new_to_old_v(vpte, rpte);
1291 rpte = hpte_new_to_old_r(rpte);
1294 pshift = kvmppc_hpte_base_page_shift(vpte, rpte);
1295 avpn = HPTE_V_AVPN_VAL(vpte) & ~(((1ul << pshift) - 1) >> 23);
1296 pteg = idx / HPTES_PER_GROUP;
1297 if (vpte & HPTE_V_SECONDARY)
1300 if (!(vpte & HPTE_V_1TB_SEG)) {
1301 unsigned long offset, vsid;
1303 /* We only have 28 - 23 bits of offset in avpn */
1304 offset = (avpn & 0x1f) << 23;
1306 /* We can find more bits from the pteg value */
1308 offset |= ((vsid ^ pteg) & old_hash_mask) << pshift;
1310 hash = vsid ^ (offset >> pshift);
1312 unsigned long offset, vsid;
1314 /* We only have 40 - 23 bits of seg_off in avpn */
1315 offset = (avpn & 0x1ffff) << 23;
1318 offset |= ((vsid ^ (vsid << 25) ^ pteg) & old_hash_mask) << pshift;
1320 hash = vsid ^ (vsid << 25) ^ (offset >> pshift);
1323 new_pteg = hash & new_hash_mask;
1324 if (vpte & HPTE_V_SECONDARY)
1325 new_pteg = ~hash & new_hash_mask;
1327 new_idx = new_pteg * HPTES_PER_GROUP + (idx % HPTES_PER_GROUP);
1328 new_hptep = (__be64 *)(new->virt + (new_idx << 4));
1330 replace_vpte = be64_to_cpu(new_hptep[0]);
1331 if (cpu_has_feature(CPU_FTR_ARCH_300)) {
1332 unsigned long replace_rpte = be64_to_cpu(new_hptep[1]);
1333 replace_vpte = hpte_new_to_old_v(replace_vpte, replace_rpte);
1336 if (replace_vpte & (HPTE_V_VALID | HPTE_V_ABSENT)) {
1337 BUG_ON(new->order >= old->order);
1339 if (replace_vpte & HPTE_V_BOLTED) {
1340 if (vpte & HPTE_V_BOLTED)
1341 /* Bolted collision, nothing we can do */
1343 /* Discard the new HPTE */
1347 /* Discard the previous HPTE */
1350 if (cpu_has_feature(CPU_FTR_ARCH_300)) {
1351 rpte = hpte_old_to_new_r(vpte, rpte);
1352 vpte = hpte_old_to_new_v(vpte);
1355 new_hptep[1] = cpu_to_be64(rpte);
1356 new->rev[new_idx].guest_rpte = guest_rpte;
1357 /* No need for a barrier, since new HPT isn't active */
1358 new_hptep[0] = cpu_to_be64(vpte);
1359 unlock_hpte(new_hptep, vpte);
1362 unlock_hpte(hptep, vpte);
1366 static int resize_hpt_rehash(struct kvm_resize_hpt *resize)
1368 struct kvm *kvm = resize->kvm;
1372 for (i = 0; i < kvmppc_hpt_npte(&kvm->arch.hpt); i++) {
1373 rc = resize_hpt_rehash_hpte(resize, i);
1381 static void resize_hpt_pivot(struct kvm_resize_hpt *resize)
1383 struct kvm *kvm = resize->kvm;
1384 struct kvm_hpt_info hpt_tmp;
1386 /* Exchange the pending tables in the resize structure with
1387 * the active tables */
1389 resize_hpt_debug(resize, "resize_hpt_pivot()\n");
1391 spin_lock(&kvm->mmu_lock);
1392 asm volatile("ptesync" : : : "memory");
1394 hpt_tmp = kvm->arch.hpt;
1395 kvmppc_set_hpt(kvm, &resize->hpt);
1396 resize->hpt = hpt_tmp;
1398 spin_unlock(&kvm->mmu_lock);
1400 synchronize_srcu_expedited(&kvm->srcu);
1402 if (cpu_has_feature(CPU_FTR_ARCH_300))
1403 kvmppc_setup_partition_table(kvm);
1405 resize_hpt_debug(resize, "resize_hpt_pivot() done\n");
1408 static void resize_hpt_release(struct kvm *kvm, struct kvm_resize_hpt *resize)
1410 if (WARN_ON(!mutex_is_locked(&kvm->arch.mmu_setup_lock)))
1416 if (resize->error != -EBUSY) {
1417 if (resize->hpt.virt)
1418 kvmppc_free_hpt(&resize->hpt);
1422 if (kvm->arch.resize_hpt == resize)
1423 kvm->arch.resize_hpt = NULL;
1426 static void resize_hpt_prepare_work(struct work_struct *work)
1428 struct kvm_resize_hpt *resize = container_of(work,
1429 struct kvm_resize_hpt,
1431 struct kvm *kvm = resize->kvm;
1434 if (WARN_ON(resize->error != -EBUSY))
1437 mutex_lock(&kvm->arch.mmu_setup_lock);
1439 /* Request is still current? */
1440 if (kvm->arch.resize_hpt == resize) {
1441 /* We may request large allocations here:
1442 * do not sleep with kvm->arch.mmu_setup_lock held for a while.
1444 mutex_unlock(&kvm->arch.mmu_setup_lock);
1446 resize_hpt_debug(resize, "resize_hpt_prepare_work(): order = %d\n",
1449 err = resize_hpt_allocate(resize);
1451 /* We have strict assumption about -EBUSY
1452 * when preparing for HPT resize.
1454 if (WARN_ON(err == -EBUSY))
1457 mutex_lock(&kvm->arch.mmu_setup_lock);
1458 /* It is possible that kvm->arch.resize_hpt != resize
1459 * after we grab kvm->arch.mmu_setup_lock again.
1463 resize->error = err;
1465 if (kvm->arch.resize_hpt != resize)
1466 resize_hpt_release(kvm, resize);
1468 mutex_unlock(&kvm->arch.mmu_setup_lock);
1471 long kvm_vm_ioctl_resize_hpt_prepare(struct kvm *kvm,
1472 struct kvm_ppc_resize_hpt *rhpt)
1474 unsigned long flags = rhpt->flags;
1475 unsigned long shift = rhpt->shift;
1476 struct kvm_resize_hpt *resize;
1479 if (flags != 0 || kvm_is_radix(kvm))
1482 if (shift && ((shift < 18) || (shift > 46)))
1485 mutex_lock(&kvm->arch.mmu_setup_lock);
1487 resize = kvm->arch.resize_hpt;
1490 if (resize->order == shift) {
1491 /* Suitable resize in progress? */
1492 ret = resize->error;
1494 ret = 100; /* estimated time in ms */
1496 resize_hpt_release(kvm, resize);
1501 /* not suitable, cancel it */
1502 resize_hpt_release(kvm, resize);
1507 goto out; /* nothing to do */
1509 /* start new resize */
1511 resize = kzalloc(sizeof(*resize), GFP_KERNEL);
1517 resize->error = -EBUSY;
1518 resize->order = shift;
1520 INIT_WORK(&resize->work, resize_hpt_prepare_work);
1521 kvm->arch.resize_hpt = resize;
1523 schedule_work(&resize->work);
1525 ret = 100; /* estimated time in ms */
1528 mutex_unlock(&kvm->arch.mmu_setup_lock);
1532 static void resize_hpt_boot_vcpu(void *opaque)
1534 /* Nothing to do, just force a KVM exit */
1537 long kvm_vm_ioctl_resize_hpt_commit(struct kvm *kvm,
1538 struct kvm_ppc_resize_hpt *rhpt)
1540 unsigned long flags = rhpt->flags;
1541 unsigned long shift = rhpt->shift;
1542 struct kvm_resize_hpt *resize;
1545 if (flags != 0 || kvm_is_radix(kvm))
1548 if (shift && ((shift < 18) || (shift > 46)))
1551 mutex_lock(&kvm->arch.mmu_setup_lock);
1553 resize = kvm->arch.resize_hpt;
1555 /* This shouldn't be possible */
1557 if (WARN_ON(!kvm->arch.mmu_ready))
1560 /* Stop VCPUs from running while we mess with the HPT */
1561 kvm->arch.mmu_ready = 0;
1564 /* Boot all CPUs out of the guest so they re-read
1566 on_each_cpu(resize_hpt_boot_vcpu, NULL, 1);
1569 if (!resize || (resize->order != shift))
1572 ret = resize->error;
1576 ret = resize_hpt_rehash(resize);
1580 resize_hpt_pivot(resize);
1583 /* Let VCPUs run again */
1584 kvm->arch.mmu_ready = 1;
1587 resize_hpt_release(kvm, resize);
1588 mutex_unlock(&kvm->arch.mmu_setup_lock);
1593 * Functions for reading and writing the hash table via reads and
1594 * writes on a file descriptor.
1596 * Reads return the guest view of the hash table, which has to be
1597 * pieced together from the real hash table and the guest_rpte
1598 * values in the revmap array.
1600 * On writes, each HPTE written is considered in turn, and if it
1601 * is valid, it is written to the HPT as if an H_ENTER with the
1602 * exact flag set was done. When the invalid count is non-zero
1603 * in the header written to the stream, the kernel will make
1604 * sure that that many HPTEs are invalid, and invalidate them
1608 struct kvm_htab_ctx {
1609 unsigned long index;
1610 unsigned long flags;
1615 #define HPTE_SIZE (2 * sizeof(unsigned long))
1618 * Returns 1 if this HPT entry has been modified or has pending
1621 static int hpte_dirty(struct revmap_entry *revp, __be64 *hptp)
1623 unsigned long rcbits_unset;
1625 if (revp->guest_rpte & HPTE_GR_MODIFIED)
1628 /* Also need to consider changes in reference and changed bits */
1629 rcbits_unset = ~revp->guest_rpte & (HPTE_R_R | HPTE_R_C);
1630 if ((be64_to_cpu(hptp[0]) & HPTE_V_VALID) &&
1631 (be64_to_cpu(hptp[1]) & rcbits_unset))
1637 static long record_hpte(unsigned long flags, __be64 *hptp,
1638 unsigned long *hpte, struct revmap_entry *revp,
1639 int want_valid, int first_pass)
1641 unsigned long v, r, hr;
1642 unsigned long rcbits_unset;
1646 /* Unmodified entries are uninteresting except on the first pass */
1647 dirty = hpte_dirty(revp, hptp);
1648 if (!first_pass && !dirty)
1652 if (be64_to_cpu(hptp[0]) & (HPTE_V_VALID | HPTE_V_ABSENT)) {
1654 if ((flags & KVM_GET_HTAB_BOLTED_ONLY) &&
1655 !(be64_to_cpu(hptp[0]) & HPTE_V_BOLTED))
1658 if (valid != want_valid)
1662 if (valid || dirty) {
1663 /* lock the HPTE so it's stable and read it */
1665 while (!try_lock_hpte(hptp, HPTE_V_HVLOCK))
1667 v = be64_to_cpu(hptp[0]);
1668 hr = be64_to_cpu(hptp[1]);
1669 if (cpu_has_feature(CPU_FTR_ARCH_300)) {
1670 v = hpte_new_to_old_v(v, hr);
1671 hr = hpte_new_to_old_r(hr);
1674 /* re-evaluate valid and dirty from synchronized HPTE value */
1675 valid = !!(v & HPTE_V_VALID);
1676 dirty = !!(revp->guest_rpte & HPTE_GR_MODIFIED);
1678 /* Harvest R and C into guest view if necessary */
1679 rcbits_unset = ~revp->guest_rpte & (HPTE_R_R | HPTE_R_C);
1680 if (valid && (rcbits_unset & hr)) {
1681 revp->guest_rpte |= (hr &
1682 (HPTE_R_R | HPTE_R_C)) | HPTE_GR_MODIFIED;
1686 if (v & HPTE_V_ABSENT) {
1687 v &= ~HPTE_V_ABSENT;
1691 if ((flags & KVM_GET_HTAB_BOLTED_ONLY) && !(v & HPTE_V_BOLTED))
1694 r = revp->guest_rpte;
1695 /* only clear modified if this is the right sort of entry */
1696 if (valid == want_valid && dirty) {
1697 r &= ~HPTE_GR_MODIFIED;
1698 revp->guest_rpte = r;
1700 unlock_hpte(hptp, be64_to_cpu(hptp[0]));
1702 if (!(valid == want_valid && (first_pass || dirty)))
1705 hpte[0] = cpu_to_be64(v);
1706 hpte[1] = cpu_to_be64(r);
1710 static ssize_t kvm_htab_read(struct file *file, char __user *buf,
1711 size_t count, loff_t *ppos)
1713 struct kvm_htab_ctx *ctx = file->private_data;
1714 struct kvm *kvm = ctx->kvm;
1715 struct kvm_get_htab_header hdr;
1717 struct revmap_entry *revp;
1718 unsigned long i, nb, nw;
1719 unsigned long __user *lbuf;
1720 struct kvm_get_htab_header __user *hptr;
1721 unsigned long flags;
1723 unsigned long hpte[2];
1725 if (!access_ok(buf, count))
1727 if (kvm_is_radix(kvm))
1730 first_pass = ctx->first_pass;
1734 hptp = (__be64 *)(kvm->arch.hpt.virt + (i * HPTE_SIZE));
1735 revp = kvm->arch.hpt.rev + i;
1736 lbuf = (unsigned long __user *)buf;
1739 while (nb + sizeof(hdr) + HPTE_SIZE < count) {
1740 /* Initialize header */
1741 hptr = (struct kvm_get_htab_header __user *)buf;
1746 lbuf = (unsigned long __user *)(buf + sizeof(hdr));
1748 /* Skip uninteresting entries, i.e. clean on not-first pass */
1750 while (i < kvmppc_hpt_npte(&kvm->arch.hpt) &&
1751 !hpte_dirty(revp, hptp)) {
1759 /* Grab a series of valid entries */
1760 while (i < kvmppc_hpt_npte(&kvm->arch.hpt) &&
1761 hdr.n_valid < 0xffff &&
1762 nb + HPTE_SIZE < count &&
1763 record_hpte(flags, hptp, hpte, revp, 1, first_pass)) {
1764 /* valid entry, write it out */
1766 if (__put_user(hpte[0], lbuf) ||
1767 __put_user(hpte[1], lbuf + 1))
1775 /* Now skip invalid entries while we can */
1776 while (i < kvmppc_hpt_npte(&kvm->arch.hpt) &&
1777 hdr.n_invalid < 0xffff &&
1778 record_hpte(flags, hptp, hpte, revp, 0, first_pass)) {
1779 /* found an invalid entry */
1786 if (hdr.n_valid || hdr.n_invalid) {
1787 /* write back the header */
1788 if (__copy_to_user(hptr, &hdr, sizeof(hdr)))
1791 buf = (char __user *)lbuf;
1796 /* Check if we've wrapped around the hash table */
1797 if (i >= kvmppc_hpt_npte(&kvm->arch.hpt)) {
1799 ctx->first_pass = 0;
1809 static ssize_t kvm_htab_write(struct file *file, const char __user *buf,
1810 size_t count, loff_t *ppos)
1812 struct kvm_htab_ctx *ctx = file->private_data;
1813 struct kvm *kvm = ctx->kvm;
1814 struct kvm_get_htab_header hdr;
1817 unsigned long __user *lbuf;
1819 unsigned long tmp[2];
1825 if (!access_ok(buf, count))
1827 if (kvm_is_radix(kvm))
1830 /* lock out vcpus from running while we're doing this */
1831 mutex_lock(&kvm->arch.mmu_setup_lock);
1832 mmu_ready = kvm->arch.mmu_ready;
1834 kvm->arch.mmu_ready = 0; /* temporarily */
1835 /* order mmu_ready vs. vcpus_running */
1837 if (atomic_read(&kvm->arch.vcpus_running)) {
1838 kvm->arch.mmu_ready = 1;
1839 mutex_unlock(&kvm->arch.mmu_setup_lock);
1845 for (nb = 0; nb + sizeof(hdr) <= count; ) {
1847 if (__copy_from_user(&hdr, buf, sizeof(hdr)))
1851 if (nb + hdr.n_valid * HPTE_SIZE > count)
1859 if (i >= kvmppc_hpt_npte(&kvm->arch.hpt) ||
1860 i + hdr.n_valid + hdr.n_invalid > kvmppc_hpt_npte(&kvm->arch.hpt))
1863 hptp = (__be64 *)(kvm->arch.hpt.virt + (i * HPTE_SIZE));
1864 lbuf = (unsigned long __user *)buf;
1865 for (j = 0; j < hdr.n_valid; ++j) {
1870 if (__get_user(hpte_v, lbuf) ||
1871 __get_user(hpte_r, lbuf + 1))
1873 v = be64_to_cpu(hpte_v);
1874 r = be64_to_cpu(hpte_r);
1876 if (!(v & HPTE_V_VALID))
1878 pshift = kvmppc_hpte_base_page_shift(v, r);
1884 if (be64_to_cpu(hptp[0]) & (HPTE_V_VALID | HPTE_V_ABSENT))
1885 kvmppc_do_h_remove(kvm, 0, i, 0, tmp);
1887 ret = kvmppc_virtmode_do_h_enter(kvm, H_EXACT, i, v, r,
1889 if (ret != H_SUCCESS) {
1890 pr_err("kvm_htab_write ret %ld i=%ld v=%lx "
1891 "r=%lx\n", ret, i, v, r);
1894 if (!mmu_ready && is_vrma_hpte(v)) {
1895 unsigned long senc, lpcr;
1897 senc = slb_pgsize_encoding(1ul << pshift);
1898 kvm->arch.vrma_slb_v = senc | SLB_VSID_B_1T |
1899 (VRMA_VSID << SLB_VSID_SHIFT_1T);
1900 if (!cpu_has_feature(CPU_FTR_ARCH_300)) {
1901 lpcr = senc << (LPCR_VRMASD_SH - 4);
1902 kvmppc_update_lpcr(kvm, lpcr,
1905 kvmppc_setup_partition_table(kvm);
1913 for (j = 0; j < hdr.n_invalid; ++j) {
1914 if (be64_to_cpu(hptp[0]) & (HPTE_V_VALID | HPTE_V_ABSENT))
1915 kvmppc_do_h_remove(kvm, 0, i, 0, tmp);
1923 /* Order HPTE updates vs. mmu_ready */
1925 kvm->arch.mmu_ready = mmu_ready;
1926 mutex_unlock(&kvm->arch.mmu_setup_lock);
1933 static int kvm_htab_release(struct inode *inode, struct file *filp)
1935 struct kvm_htab_ctx *ctx = filp->private_data;
1937 filp->private_data = NULL;
1938 if (!(ctx->flags & KVM_GET_HTAB_WRITE))
1939 atomic_dec(&ctx->kvm->arch.hpte_mod_interest);
1940 kvm_put_kvm(ctx->kvm);
1945 static const struct file_operations kvm_htab_fops = {
1946 .read = kvm_htab_read,
1947 .write = kvm_htab_write,
1948 .llseek = default_llseek,
1949 .release = kvm_htab_release,
1952 int kvm_vm_ioctl_get_htab_fd(struct kvm *kvm, struct kvm_get_htab_fd *ghf)
1955 struct kvm_htab_ctx *ctx;
1958 /* reject flags we don't recognize */
1959 if (ghf->flags & ~(KVM_GET_HTAB_BOLTED_ONLY | KVM_GET_HTAB_WRITE))
1961 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1966 ctx->index = ghf->start_index;
1967 ctx->flags = ghf->flags;
1968 ctx->first_pass = 1;
1970 rwflag = (ghf->flags & KVM_GET_HTAB_WRITE) ? O_WRONLY : O_RDONLY;
1971 ret = anon_inode_getfd("kvm-htab", &kvm_htab_fops, ctx, rwflag | O_CLOEXEC);
1974 kvm_put_kvm_no_destroy(kvm);
1978 if (rwflag == O_RDONLY) {
1979 mutex_lock(&kvm->slots_lock);
1980 atomic_inc(&kvm->arch.hpte_mod_interest);
1981 /* make sure kvmppc_do_h_enter etc. see the increment */
1982 synchronize_srcu_expedited(&kvm->srcu);
1983 mutex_unlock(&kvm->slots_lock);
1989 struct debugfs_htab_state {
1992 unsigned long hpt_index;
1998 static int debugfs_htab_open(struct inode *inode, struct file *file)
2000 struct kvm *kvm = inode->i_private;
2001 struct debugfs_htab_state *p;
2003 p = kzalloc(sizeof(*p), GFP_KERNEL);
2009 mutex_init(&p->mutex);
2010 file->private_data = p;
2012 return nonseekable_open(inode, file);
2015 static int debugfs_htab_release(struct inode *inode, struct file *file)
2017 struct debugfs_htab_state *p = file->private_data;
2019 kvm_put_kvm(p->kvm);
2024 static ssize_t debugfs_htab_read(struct file *file, char __user *buf,
2025 size_t len, loff_t *ppos)
2027 struct debugfs_htab_state *p = file->private_data;
2030 unsigned long v, hr, gr;
2035 if (kvm_is_radix(kvm))
2038 ret = mutex_lock_interruptible(&p->mutex);
2042 if (p->chars_left) {
2046 r = copy_to_user(buf, p->buf + p->buf_index, n);
2061 hptp = (__be64 *)(kvm->arch.hpt.virt + (i * HPTE_SIZE));
2062 for (; len != 0 && i < kvmppc_hpt_npte(&kvm->arch.hpt);
2064 if (!(be64_to_cpu(hptp[0]) & (HPTE_V_VALID | HPTE_V_ABSENT)))
2067 /* lock the HPTE so it's stable and read it */
2069 while (!try_lock_hpte(hptp, HPTE_V_HVLOCK))
2071 v = be64_to_cpu(hptp[0]) & ~HPTE_V_HVLOCK;
2072 hr = be64_to_cpu(hptp[1]);
2073 gr = kvm->arch.hpt.rev[i].guest_rpte;
2074 unlock_hpte(hptp, v);
2077 if (!(v & (HPTE_V_VALID | HPTE_V_ABSENT)))
2080 n = scnprintf(p->buf, sizeof(p->buf),
2081 "%6lx %.16lx %.16lx %.16lx\n",
2086 r = copy_to_user(buf, p->buf, n);
2102 mutex_unlock(&p->mutex);
2106 static ssize_t debugfs_htab_write(struct file *file, const char __user *buf,
2107 size_t len, loff_t *ppos)
2112 static const struct file_operations debugfs_htab_fops = {
2113 .owner = THIS_MODULE,
2114 .open = debugfs_htab_open,
2115 .release = debugfs_htab_release,
2116 .read = debugfs_htab_read,
2117 .write = debugfs_htab_write,
2118 .llseek = generic_file_llseek,
2121 void kvmppc_mmu_debugfs_init(struct kvm *kvm)
2123 debugfs_create_file("htab", 0400, kvm->debugfs_dentry, kvm,
2124 &debugfs_htab_fops);
2127 void kvmppc_mmu_book3s_hv_init(struct kvm_vcpu *vcpu)
2129 struct kvmppc_mmu *mmu = &vcpu->arch.mmu;
2131 vcpu->arch.slb_nr = 32; /* POWER7/POWER8 */
2133 mmu->xlate = kvmppc_mmu_book3s_64_hv_xlate;
2135 vcpu->arch.hflags |= BOOK3S_HFLAG_SLB;