1 // SPDX-License-Identifier: GPL-2.0-only
4 * Copyright 2010 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
5 * Copyright 2011 David Gibson, IBM Corporation <dwg@au1.ibm.com>
6 * Copyright 2016 Alexey Kardashevskiy, IBM Corporation <aik@au1.ibm.com>
9 #include <linux/types.h>
10 #include <linux/string.h>
11 #include <linux/kvm.h>
12 #include <linux/kvm_host.h>
13 #include <linux/highmem.h>
14 #include <linux/gfp.h>
15 #include <linux/slab.h>
16 #include <linux/hugetlb.h>
17 #include <linux/list.h>
18 #include <linux/stringify.h>
20 #include <asm/kvm_ppc.h>
21 #include <asm/kvm_book3s.h>
22 #include <asm/book3s/64/mmu-hash.h>
23 #include <asm/mmu_context.h>
24 #include <asm/hvcall.h>
25 #include <asm/synch.h>
26 #include <asm/ppc-opcode.h>
28 #include <asm/iommu.h>
30 #include <asm/pte-walk.h>
34 #define WARN_ON_ONCE_RM(condition) ({ \
35 static bool __section(".data.unlikely") __warned; \
36 int __ret_warn_once = !!(condition); \
38 if (unlikely(__ret_warn_once && !__warned)) { \
40 pr_err("WARN_ON_ONCE_RM: (%s) at %s:%u\n", \
41 __stringify(condition), \
42 __func__, __LINE__); \
45 unlikely(__ret_warn_once); \
50 #define WARN_ON_ONCE_RM(condition) ({ \
51 int __ret_warn_on = !!(condition); \
52 unlikely(__ret_warn_on); \
58 * Finds a TCE table descriptor by LIOBN.
60 * WARNING: This will be called in real or virtual mode on HV KVM and virtual
63 struct kvmppc_spapr_tce_table *kvmppc_find_table(struct kvm *kvm,
66 struct kvmppc_spapr_tce_table *stt;
68 list_for_each_entry_lockless(stt, &kvm->arch.spapr_tce_tables, list)
69 if (stt->liobn == liobn)
74 EXPORT_SYMBOL_GPL(kvmppc_find_table);
76 #ifdef CONFIG_KVM_BOOK3S_HV_POSSIBLE
77 static long kvmppc_rm_tce_to_ua(struct kvm *kvm,
78 unsigned long tce, unsigned long *ua)
80 unsigned long gfn = tce >> PAGE_SHIFT;
81 struct kvm_memory_slot *memslot;
83 memslot = __gfn_to_memslot(kvm_memslots_raw(kvm), gfn);
87 *ua = __gfn_to_hva_memslot(memslot, gfn) |
88 (tce & ~(PAGE_MASK | TCE_PCI_READ | TCE_PCI_WRITE));
94 * Validates TCE address.
95 * At the moment flags and page mask are validated.
96 * As the host kernel does not access those addresses (just puts them
97 * to the table and user space is supposed to process them), we can skip
98 * checking other things (such as TCE is a guest RAM address or the page
99 * was actually allocated).
101 static long kvmppc_rm_tce_validate(struct kvmppc_spapr_tce_table *stt,
104 unsigned long gpa = tce & ~(TCE_PCI_READ | TCE_PCI_WRITE);
105 enum dma_data_direction dir = iommu_tce_direction(tce);
106 struct kvmppc_spapr_tce_iommu_table *stit;
107 unsigned long ua = 0;
109 /* Allow userspace to poison TCE table */
113 if (iommu_tce_check_gpa(stt->page_shift, gpa))
116 if (kvmppc_rm_tce_to_ua(stt->kvm, tce, &ua))
119 list_for_each_entry_lockless(stit, &stt->iommu_tables, next) {
120 unsigned long hpa = 0;
121 struct mm_iommu_table_group_mem_t *mem;
122 long shift = stit->tbl->it_page_shift;
124 mem = mm_iommu_lookup_rm(stt->kvm->mm, ua, 1ULL << shift);
128 if (mm_iommu_ua_to_hpa_rm(mem, ua, shift, &hpa))
135 /* Note on the use of page_address() in real mode,
137 * It is safe to use page_address() in real mode on ppc64 because
138 * page_address() is always defined as lowmem_page_address()
139 * which returns __va(PFN_PHYS(page_to_pfn(page))) which is arithmetic
140 * operation and does not access page struct.
142 * Theoretically page_address() could be defined different
143 * but either WANT_PAGE_VIRTUAL or HASHED_PAGE_VIRTUAL
144 * would have to be enabled.
145 * WANT_PAGE_VIRTUAL is never enabled on ppc32/ppc64,
146 * HASHED_PAGE_VIRTUAL could be enabled for ppc32 only and only
147 * if CONFIG_HIGHMEM is defined. As CONFIG_SPARSEMEM_VMEMMAP
148 * is not expected to be enabled on ppc32, page_address()
149 * is safe for ppc32 as well.
151 * WARNING: This will be called in real-mode on HV KVM and virtual
154 static u64 *kvmppc_page_address(struct page *page)
156 #if defined(HASHED_PAGE_VIRTUAL) || defined(WANT_PAGE_VIRTUAL)
157 #error TODO: fix to avoid page_address() here
159 return (u64 *) page_address(page);
163 * Handles TCE requests for emulated devices.
164 * Puts guest TCE values to the table and expects user space to convert them.
165 * Cannot fail so kvmppc_rm_tce_validate must be called before it.
167 static void kvmppc_rm_tce_put(struct kvmppc_spapr_tce_table *stt,
168 unsigned long idx, unsigned long tce)
174 page = stt->pages[idx / TCES_PER_PAGE];
176 * kvmppc_rm_ioba_validate() allows pages not be allocated if TCE is
177 * being cleared, otherwise it returns H_TOO_HARD and we skip this.
180 WARN_ON_ONCE_RM(tce != 0);
183 tbl = kvmppc_page_address(page);
185 tbl[idx % TCES_PER_PAGE] = tce;
189 * TCEs pages are allocated in kvmppc_rm_tce_put() which won't be able to do so
191 * Check if kvmppc_rm_tce_put() can succeed in real mode, i.e. a TCEs page is
192 * allocated or not required (when clearing a tce entry).
194 static long kvmppc_rm_ioba_validate(struct kvmppc_spapr_tce_table *stt,
195 unsigned long ioba, unsigned long npages, bool clearing)
197 unsigned long i, idx, sttpage, sttpages;
198 unsigned long ret = kvmppc_ioba_validate(stt, ioba, npages);
203 * clearing==true says kvmppc_rm_tce_put won't be allocating pages
209 idx = (ioba >> stt->page_shift) - stt->offset;
210 sttpage = idx / TCES_PER_PAGE;
211 sttpages = ALIGN(idx % TCES_PER_PAGE + npages, TCES_PER_PAGE) /
213 for (i = sttpage; i < sttpage + sttpages; ++i)
220 static long iommu_tce_xchg_no_kill_rm(struct mm_struct *mm,
221 struct iommu_table *tbl,
222 unsigned long entry, unsigned long *hpa,
223 enum dma_data_direction *direction)
227 ret = tbl->it_ops->xchg_no_kill(tbl, entry, hpa, direction, true);
229 if (!ret && ((*direction == DMA_FROM_DEVICE) ||
230 (*direction == DMA_BIDIRECTIONAL))) {
231 __be64 *pua = IOMMU_TABLE_USERSPACE_ENTRY_RO(tbl, entry);
233 * kvmppc_rm_tce_iommu_do_map() updates the UA cache after
234 * calling this so we still get here a valid UA.
237 mm_iommu_ua_mark_dirty_rm(mm, be64_to_cpu(*pua));
243 static void iommu_tce_kill_rm(struct iommu_table *tbl,
244 unsigned long entry, unsigned long pages)
246 if (tbl->it_ops->tce_kill)
247 tbl->it_ops->tce_kill(tbl, entry, pages, true);
250 static void kvmppc_rm_clear_tce(struct kvm *kvm, struct kvmppc_spapr_tce_table *stt,
251 struct iommu_table *tbl, unsigned long entry)
254 unsigned long subpages = 1ULL << (stt->page_shift - tbl->it_page_shift);
255 unsigned long io_entry = entry << (stt->page_shift - tbl->it_page_shift);
257 for (i = 0; i < subpages; ++i) {
258 unsigned long hpa = 0;
259 enum dma_data_direction dir = DMA_NONE;
261 iommu_tce_xchg_no_kill_rm(kvm->mm, tbl, io_entry + i, &hpa, &dir);
265 static long kvmppc_rm_tce_iommu_mapped_dec(struct kvm *kvm,
266 struct iommu_table *tbl, unsigned long entry)
268 struct mm_iommu_table_group_mem_t *mem = NULL;
269 const unsigned long pgsize = 1ULL << tbl->it_page_shift;
270 __be64 *pua = IOMMU_TABLE_USERSPACE_ENTRY_RO(tbl, entry);
273 /* it_userspace allocation might be delayed */
276 mem = mm_iommu_lookup_rm(kvm->mm, be64_to_cpu(*pua), pgsize);
280 mm_iommu_mapped_dec(mem);
282 *pua = cpu_to_be64(0);
287 static long kvmppc_rm_tce_iommu_do_unmap(struct kvm *kvm,
288 struct iommu_table *tbl, unsigned long entry)
290 enum dma_data_direction dir = DMA_NONE;
291 unsigned long hpa = 0;
294 if (iommu_tce_xchg_no_kill_rm(kvm->mm, tbl, entry, &hpa, &dir))
296 * real mode xchg can fail if struct page crosses
304 ret = kvmppc_rm_tce_iommu_mapped_dec(kvm, tbl, entry);
306 iommu_tce_xchg_no_kill_rm(kvm->mm, tbl, entry, &hpa, &dir);
311 static long kvmppc_rm_tce_iommu_unmap(struct kvm *kvm,
312 struct kvmppc_spapr_tce_table *stt, struct iommu_table *tbl,
315 unsigned long i, ret = H_SUCCESS;
316 unsigned long subpages = 1ULL << (stt->page_shift - tbl->it_page_shift);
317 unsigned long io_entry = entry * subpages;
319 for (i = 0; i < subpages; ++i) {
320 ret = kvmppc_rm_tce_iommu_do_unmap(kvm, tbl, io_entry + i);
321 if (ret != H_SUCCESS)
325 iommu_tce_kill_rm(tbl, io_entry, subpages);
330 static long kvmppc_rm_tce_iommu_do_map(struct kvm *kvm, struct iommu_table *tbl,
331 unsigned long entry, unsigned long ua,
332 enum dma_data_direction dir)
335 unsigned long hpa = 0;
336 __be64 *pua = IOMMU_TABLE_USERSPACE_ENTRY_RO(tbl, entry);
337 struct mm_iommu_table_group_mem_t *mem;
340 /* it_userspace allocation might be delayed */
343 mem = mm_iommu_lookup_rm(kvm->mm, ua, 1ULL << tbl->it_page_shift);
347 if (WARN_ON_ONCE_RM(mm_iommu_ua_to_hpa_rm(mem, ua, tbl->it_page_shift,
351 if (WARN_ON_ONCE_RM(mm_iommu_mapped_inc(mem)))
354 ret = iommu_tce_xchg_no_kill_rm(kvm->mm, tbl, entry, &hpa, &dir);
356 mm_iommu_mapped_dec(mem);
358 * real mode xchg can fail if struct page crosses
365 kvmppc_rm_tce_iommu_mapped_dec(kvm, tbl, entry);
367 *pua = cpu_to_be64(ua);
372 static long kvmppc_rm_tce_iommu_map(struct kvm *kvm,
373 struct kvmppc_spapr_tce_table *stt, struct iommu_table *tbl,
374 unsigned long entry, unsigned long ua,
375 enum dma_data_direction dir)
377 unsigned long i, pgoff, ret = H_SUCCESS;
378 unsigned long subpages = 1ULL << (stt->page_shift - tbl->it_page_shift);
379 unsigned long io_entry = entry * subpages;
381 for (i = 0, pgoff = 0; i < subpages;
382 ++i, pgoff += IOMMU_PAGE_SIZE(tbl)) {
384 ret = kvmppc_rm_tce_iommu_do_map(kvm, tbl,
385 io_entry + i, ua + pgoff, dir);
386 if (ret != H_SUCCESS)
390 iommu_tce_kill_rm(tbl, io_entry, subpages);
395 long kvmppc_rm_h_put_tce(struct kvm_vcpu *vcpu, unsigned long liobn,
396 unsigned long ioba, unsigned long tce)
398 struct kvmppc_spapr_tce_table *stt;
400 struct kvmppc_spapr_tce_iommu_table *stit;
401 unsigned long entry, ua = 0;
402 enum dma_data_direction dir;
404 /* udbg_printf("H_PUT_TCE(): liobn=0x%lx ioba=0x%lx, tce=0x%lx\n", */
405 /* liobn, ioba, tce); */
407 stt = kvmppc_find_table(vcpu->kvm, liobn);
411 ret = kvmppc_rm_ioba_validate(stt, ioba, 1, tce == 0);
412 if (ret != H_SUCCESS)
415 ret = kvmppc_rm_tce_validate(stt, tce);
416 if (ret != H_SUCCESS)
419 dir = iommu_tce_direction(tce);
420 if ((dir != DMA_NONE) && kvmppc_rm_tce_to_ua(vcpu->kvm, tce, &ua))
423 entry = ioba >> stt->page_shift;
425 list_for_each_entry_lockless(stit, &stt->iommu_tables, next) {
427 ret = kvmppc_rm_tce_iommu_unmap(vcpu->kvm, stt,
430 ret = kvmppc_rm_tce_iommu_map(vcpu->kvm, stt,
431 stit->tbl, entry, ua, dir);
433 if (ret != H_SUCCESS) {
434 kvmppc_rm_clear_tce(vcpu->kvm, stt, stit->tbl, entry);
439 kvmppc_rm_tce_put(stt, entry, tce);
444 static long kvmppc_rm_ua_to_hpa(struct kvm_vcpu *vcpu, unsigned long mmu_seq,
445 unsigned long ua, unsigned long *phpa)
451 * Called in real mode with MSR_EE = 0. We are safe here.
452 * It is ok to do the lookup with arch.pgdir here, because
453 * we are doing this on secondary cpus and current task there
454 * is not the hypervisor. Also this is safe against THP in the
455 * host, because an IPI to primary thread will wait for the secondary
456 * to exit which will agains result in the below page table walk
459 /* an rmap lock won't make it safe. because that just ensure hash
460 * page table entries are removed with rmap lock held. After that
461 * mmu notifier returns and we go ahead and removing ptes from Qemu page table.
463 ptep = find_kvm_host_pte(vcpu->kvm, mmu_seq, ua, &shift);
467 pte = READ_ONCE(*ptep);
468 if (!pte_present(pte))
474 /* Avoid handling anything potentially complicated in realmode */
475 if (shift > PAGE_SHIFT)
481 *phpa = (pte_pfn(pte) << PAGE_SHIFT) | (ua & ((1ULL << shift) - 1)) |
487 long kvmppc_rm_h_put_tce_indirect(struct kvm_vcpu *vcpu,
488 unsigned long liobn, unsigned long ioba,
489 unsigned long tce_list, unsigned long npages)
491 struct kvm *kvm = vcpu->kvm;
492 struct kvmppc_spapr_tce_table *stt;
493 long i, ret = H_SUCCESS;
494 unsigned long tces, entry, ua = 0;
495 unsigned long mmu_seq;
497 struct kvmppc_spapr_tce_iommu_table *stit;
500 * used to check for invalidations in progress
502 mmu_seq = kvm->mmu_notifier_seq;
505 stt = kvmppc_find_table(vcpu->kvm, liobn);
509 entry = ioba >> stt->page_shift;
511 * The spec says that the maximum size of the list is 512 TCEs
512 * so the whole table addressed resides in 4K page
517 if (tce_list & (SZ_4K - 1))
520 ret = kvmppc_rm_ioba_validate(stt, ioba, npages, false);
521 if (ret != H_SUCCESS)
524 if (mm_iommu_preregistered(vcpu->kvm->mm)) {
526 * We get here if guest memory was pre-registered which
527 * is normally VFIO case and gpa->hpa translation does not
530 struct mm_iommu_table_group_mem_t *mem;
532 if (kvmppc_rm_tce_to_ua(vcpu->kvm, tce_list, &ua))
535 mem = mm_iommu_lookup_rm(vcpu->kvm->mm, ua, IOMMU_PAGE_SIZE_4K);
537 prereg = mm_iommu_ua_to_hpa_rm(mem, ua,
538 IOMMU_PAGE_SHIFT_4K, &tces) == 0;
543 * This is usually a case of a guest with emulated devices only
544 * when TCE list is not in preregistered memory.
545 * We do not require memory to be preregistered in this case
546 * so lock rmap and do __find_linux_pte_or_hugepte().
548 if (kvmppc_rm_tce_to_ua(vcpu->kvm, tce_list, &ua))
551 arch_spin_lock(&kvm->mmu_lock.rlock.raw_lock);
552 if (kvmppc_rm_ua_to_hpa(vcpu, mmu_seq, ua, &tces)) {
558 for (i = 0; i < npages; ++i) {
559 unsigned long tce = be64_to_cpu(((u64 *)tces)[i]);
561 ret = kvmppc_rm_tce_validate(stt, tce);
562 if (ret != H_SUCCESS)
566 for (i = 0; i < npages; ++i) {
567 unsigned long tce = be64_to_cpu(((u64 *)tces)[i]);
570 if (kvmppc_rm_tce_to_ua(vcpu->kvm, tce, &ua)) {
575 list_for_each_entry_lockless(stit, &stt->iommu_tables, next) {
576 ret = kvmppc_rm_tce_iommu_map(vcpu->kvm, stt,
577 stit->tbl, entry + i, ua,
578 iommu_tce_direction(tce));
580 if (ret != H_SUCCESS) {
581 kvmppc_rm_clear_tce(vcpu->kvm, stt, stit->tbl,
587 kvmppc_rm_tce_put(stt, entry + i, tce);
592 arch_spin_unlock(&kvm->mmu_lock.rlock.raw_lock);
596 long kvmppc_rm_h_stuff_tce(struct kvm_vcpu *vcpu,
597 unsigned long liobn, unsigned long ioba,
598 unsigned long tce_value, unsigned long npages)
600 struct kvmppc_spapr_tce_table *stt;
602 struct kvmppc_spapr_tce_iommu_table *stit;
604 stt = kvmppc_find_table(vcpu->kvm, liobn);
608 ret = kvmppc_rm_ioba_validate(stt, ioba, npages, tce_value == 0);
609 if (ret != H_SUCCESS)
612 /* Check permission bits only to allow userspace poison TCE for debug */
613 if (tce_value & (TCE_PCI_WRITE | TCE_PCI_READ))
616 list_for_each_entry_lockless(stit, &stt->iommu_tables, next) {
617 unsigned long entry = ioba >> stt->page_shift;
619 for (i = 0; i < npages; ++i) {
620 ret = kvmppc_rm_tce_iommu_unmap(vcpu->kvm, stt,
621 stit->tbl, entry + i);
623 if (ret == H_SUCCESS)
626 if (ret == H_TOO_HARD)
630 kvmppc_rm_clear_tce(vcpu->kvm, stt, stit->tbl, entry + i);
634 for (i = 0; i < npages; ++i, ioba += (1ULL << stt->page_shift))
635 kvmppc_rm_tce_put(stt, ioba >> stt->page_shift, tce_value);
640 /* This can be called in either virtual mode or real mode */
641 long kvmppc_h_get_tce(struct kvm_vcpu *vcpu, unsigned long liobn,
644 struct kvmppc_spapr_tce_table *stt;
650 stt = kvmppc_find_table(vcpu->kvm, liobn);
654 ret = kvmppc_ioba_validate(stt, ioba, 1);
655 if (ret != H_SUCCESS)
658 idx = (ioba >> stt->page_shift) - stt->offset;
659 page = stt->pages[idx / TCES_PER_PAGE];
661 vcpu->arch.regs.gpr[4] = 0;
664 tbl = (u64 *)page_address(page);
666 vcpu->arch.regs.gpr[4] = tbl[idx % TCES_PER_PAGE];
670 EXPORT_SYMBOL_GPL(kvmppc_h_get_tce);
672 #endif /* KVM_BOOK3S_HV_POSSIBLE */