1 // SPDX-License-Identifier: GPL-2.0-only
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 iommu_table *tbl,
253 unsigned long hpa = 0;
254 enum dma_data_direction dir = DMA_NONE;
256 iommu_tce_xchg_no_kill_rm(kvm->mm, tbl, entry, &hpa, &dir);
259 static long kvmppc_rm_tce_iommu_mapped_dec(struct kvm *kvm,
260 struct iommu_table *tbl, unsigned long entry)
262 struct mm_iommu_table_group_mem_t *mem = NULL;
263 const unsigned long pgsize = 1ULL << tbl->it_page_shift;
264 __be64 *pua = IOMMU_TABLE_USERSPACE_ENTRY_RO(tbl, entry);
267 /* it_userspace allocation might be delayed */
270 mem = mm_iommu_lookup_rm(kvm->mm, be64_to_cpu(*pua), pgsize);
274 mm_iommu_mapped_dec(mem);
276 *pua = cpu_to_be64(0);
281 static long kvmppc_rm_tce_iommu_do_unmap(struct kvm *kvm,
282 struct iommu_table *tbl, unsigned long entry)
284 enum dma_data_direction dir = DMA_NONE;
285 unsigned long hpa = 0;
288 if (iommu_tce_xchg_no_kill_rm(kvm->mm, tbl, entry, &hpa, &dir))
290 * real mode xchg can fail if struct page crosses
298 ret = kvmppc_rm_tce_iommu_mapped_dec(kvm, tbl, entry);
300 iommu_tce_xchg_no_kill_rm(kvm->mm, tbl, entry, &hpa, &dir);
305 static long kvmppc_rm_tce_iommu_unmap(struct kvm *kvm,
306 struct kvmppc_spapr_tce_table *stt, struct iommu_table *tbl,
309 unsigned long i, ret = H_SUCCESS;
310 unsigned long subpages = 1ULL << (stt->page_shift - tbl->it_page_shift);
311 unsigned long io_entry = entry * subpages;
313 for (i = 0; i < subpages; ++i) {
314 ret = kvmppc_rm_tce_iommu_do_unmap(kvm, tbl, io_entry + i);
315 if (ret != H_SUCCESS)
322 static long kvmppc_rm_tce_iommu_do_map(struct kvm *kvm, struct iommu_table *tbl,
323 unsigned long entry, unsigned long ua,
324 enum dma_data_direction dir)
327 unsigned long hpa = 0;
328 __be64 *pua = IOMMU_TABLE_USERSPACE_ENTRY_RO(tbl, entry);
329 struct mm_iommu_table_group_mem_t *mem;
332 /* it_userspace allocation might be delayed */
335 mem = mm_iommu_lookup_rm(kvm->mm, ua, 1ULL << tbl->it_page_shift);
339 if (WARN_ON_ONCE_RM(mm_iommu_ua_to_hpa_rm(mem, ua, tbl->it_page_shift,
343 if (WARN_ON_ONCE_RM(mm_iommu_mapped_inc(mem)))
346 ret = iommu_tce_xchg_no_kill_rm(kvm->mm, tbl, entry, &hpa, &dir);
348 mm_iommu_mapped_dec(mem);
350 * real mode xchg can fail if struct page crosses
357 kvmppc_rm_tce_iommu_mapped_dec(kvm, tbl, entry);
359 *pua = cpu_to_be64(ua);
364 static long kvmppc_rm_tce_iommu_map(struct kvm *kvm,
365 struct kvmppc_spapr_tce_table *stt, struct iommu_table *tbl,
366 unsigned long entry, unsigned long ua,
367 enum dma_data_direction dir)
369 unsigned long i, pgoff, ret = H_SUCCESS;
370 unsigned long subpages = 1ULL << (stt->page_shift - tbl->it_page_shift);
371 unsigned long io_entry = entry * subpages;
373 for (i = 0, pgoff = 0; i < subpages;
374 ++i, pgoff += IOMMU_PAGE_SIZE(tbl)) {
376 ret = kvmppc_rm_tce_iommu_do_map(kvm, tbl,
377 io_entry + i, ua + pgoff, dir);
378 if (ret != H_SUCCESS)
385 long kvmppc_rm_h_put_tce(struct kvm_vcpu *vcpu, unsigned long liobn,
386 unsigned long ioba, unsigned long tce)
388 struct kvmppc_spapr_tce_table *stt;
390 struct kvmppc_spapr_tce_iommu_table *stit;
391 unsigned long entry, ua = 0;
392 enum dma_data_direction dir;
394 /* udbg_printf("H_PUT_TCE(): liobn=0x%lx ioba=0x%lx, tce=0x%lx\n", */
395 /* liobn, ioba, tce); */
397 stt = kvmppc_find_table(vcpu->kvm, liobn);
401 ret = kvmppc_rm_ioba_validate(stt, ioba, 1, tce == 0);
402 if (ret != H_SUCCESS)
405 ret = kvmppc_rm_tce_validate(stt, tce);
406 if (ret != H_SUCCESS)
409 dir = iommu_tce_direction(tce);
410 if ((dir != DMA_NONE) && kvmppc_rm_tce_to_ua(vcpu->kvm, tce, &ua))
413 entry = ioba >> stt->page_shift;
415 list_for_each_entry_lockless(stit, &stt->iommu_tables, next) {
417 ret = kvmppc_rm_tce_iommu_unmap(vcpu->kvm, stt,
420 ret = kvmppc_rm_tce_iommu_map(vcpu->kvm, stt,
421 stit->tbl, entry, ua, dir);
423 iommu_tce_kill_rm(stit->tbl, entry, 1);
425 if (ret != H_SUCCESS) {
426 kvmppc_rm_clear_tce(vcpu->kvm, stit->tbl, entry);
431 kvmppc_rm_tce_put(stt, entry, tce);
436 static long kvmppc_rm_ua_to_hpa(struct kvm_vcpu *vcpu, unsigned long mmu_seq,
437 unsigned long ua, unsigned long *phpa)
443 * Called in real mode with MSR_EE = 0. We are safe here.
444 * It is ok to do the lookup with arch.pgdir here, because
445 * we are doing this on secondary cpus and current task there
446 * is not the hypervisor. Also this is safe against THP in the
447 * host, because an IPI to primary thread will wait for the secondary
448 * to exit which will agains result in the below page table walk
451 /* an rmap lock won't make it safe. because that just ensure hash
452 * page table entries are removed with rmap lock held. After that
453 * mmu notifier returns and we go ahead and removing ptes from Qemu page table.
455 ptep = find_kvm_host_pte(vcpu->kvm, mmu_seq, ua, &shift);
459 pte = READ_ONCE(*ptep);
460 if (!pte_present(pte))
466 /* Avoid handling anything potentially complicated in realmode */
467 if (shift > PAGE_SHIFT)
473 *phpa = (pte_pfn(pte) << PAGE_SHIFT) | (ua & ((1ULL << shift) - 1)) |
479 long kvmppc_rm_h_put_tce_indirect(struct kvm_vcpu *vcpu,
480 unsigned long liobn, unsigned long ioba,
481 unsigned long tce_list, unsigned long npages)
483 struct kvm *kvm = vcpu->kvm;
484 struct kvmppc_spapr_tce_table *stt;
485 long i, ret = H_SUCCESS;
486 unsigned long tces, entry, ua = 0;
487 unsigned long mmu_seq;
489 struct kvmppc_spapr_tce_iommu_table *stit;
492 * used to check for invalidations in progress
494 mmu_seq = kvm->mmu_notifier_seq;
497 stt = kvmppc_find_table(vcpu->kvm, liobn);
501 entry = ioba >> stt->page_shift;
503 * The spec says that the maximum size of the list is 512 TCEs
504 * so the whole table addressed resides in 4K page
509 if (tce_list & (SZ_4K - 1))
512 ret = kvmppc_rm_ioba_validate(stt, ioba, npages, false);
513 if (ret != H_SUCCESS)
516 if (mm_iommu_preregistered(vcpu->kvm->mm)) {
518 * We get here if guest memory was pre-registered which
519 * is normally VFIO case and gpa->hpa translation does not
522 struct mm_iommu_table_group_mem_t *mem;
524 if (kvmppc_rm_tce_to_ua(vcpu->kvm, tce_list, &ua))
527 mem = mm_iommu_lookup_rm(vcpu->kvm->mm, ua, IOMMU_PAGE_SIZE_4K);
529 prereg = mm_iommu_ua_to_hpa_rm(mem, ua,
530 IOMMU_PAGE_SHIFT_4K, &tces) == 0;
535 * This is usually a case of a guest with emulated devices only
536 * when TCE list is not in preregistered memory.
537 * We do not require memory to be preregistered in this case
538 * so lock rmap and do __find_linux_pte_or_hugepte().
540 if (kvmppc_rm_tce_to_ua(vcpu->kvm, tce_list, &ua))
543 arch_spin_lock(&kvm->mmu_lock.rlock.raw_lock);
544 if (kvmppc_rm_ua_to_hpa(vcpu, mmu_seq, ua, &tces)) {
550 for (i = 0; i < npages; ++i) {
551 unsigned long tce = be64_to_cpu(((u64 *)tces)[i]);
553 ret = kvmppc_rm_tce_validate(stt, tce);
554 if (ret != H_SUCCESS)
558 for (i = 0; i < npages; ++i) {
559 unsigned long tce = be64_to_cpu(((u64 *)tces)[i]);
562 if (kvmppc_rm_tce_to_ua(vcpu->kvm, tce, &ua)) {
564 goto invalidate_exit;
567 list_for_each_entry_lockless(stit, &stt->iommu_tables, next) {
568 ret = kvmppc_rm_tce_iommu_map(vcpu->kvm, stt,
569 stit->tbl, entry + i, ua,
570 iommu_tce_direction(tce));
572 if (ret != H_SUCCESS) {
573 kvmppc_rm_clear_tce(vcpu->kvm, stit->tbl,
575 goto invalidate_exit;
579 kvmppc_rm_tce_put(stt, entry + i, tce);
583 list_for_each_entry_lockless(stit, &stt->iommu_tables, next)
584 iommu_tce_kill_rm(stit->tbl, entry, npages);
588 arch_spin_unlock(&kvm->mmu_lock.rlock.raw_lock);
592 long kvmppc_rm_h_stuff_tce(struct kvm_vcpu *vcpu,
593 unsigned long liobn, unsigned long ioba,
594 unsigned long tce_value, unsigned long npages)
596 struct kvmppc_spapr_tce_table *stt;
598 struct kvmppc_spapr_tce_iommu_table *stit;
600 stt = kvmppc_find_table(vcpu->kvm, liobn);
604 ret = kvmppc_rm_ioba_validate(stt, ioba, npages, tce_value == 0);
605 if (ret != H_SUCCESS)
608 /* Check permission bits only to allow userspace poison TCE for debug */
609 if (tce_value & (TCE_PCI_WRITE | TCE_PCI_READ))
612 list_for_each_entry_lockless(stit, &stt->iommu_tables, next) {
613 unsigned long entry = ioba >> stt->page_shift;
615 for (i = 0; i < npages; ++i) {
616 ret = kvmppc_rm_tce_iommu_unmap(vcpu->kvm, stt,
617 stit->tbl, entry + i);
619 if (ret == H_SUCCESS)
622 if (ret == H_TOO_HARD)
623 goto invalidate_exit;
626 kvmppc_rm_clear_tce(vcpu->kvm, stit->tbl, entry);
630 for (i = 0; i < npages; ++i, ioba += (1ULL << stt->page_shift))
631 kvmppc_rm_tce_put(stt, ioba >> stt->page_shift, tce_value);
634 list_for_each_entry_lockless(stit, &stt->iommu_tables, next)
635 iommu_tce_kill_rm(stit->tbl, ioba >> stt->page_shift, npages);
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 */