[PATCH kernel v4] KVM: PPC: Allocate guest TCEs on demand too

[Alexey Kardashevskiy <aik@ozlabs.ru>]

We already allocate hardware TCE tables in multiple levels and skip
intermediate levels when we can, now it is a turn of the KVM TCE tables.
Thankfully these are allocated already in 2 levels.

This moves the table's last level allocation from the creating helper to
kvmppc_tce_put() and kvm_spapr_tce_fault(). Since such allocation cannot
be done in real mode, this creates a virtual mode version of
kvmppc_tce_put() which handles allocations.

This adds kvmppc_rm_ioba_validate() to do an additional test if
the consequent kvmppc_tce_put() needs a page which has not been allocated;
if this is the case, we bail out to virtual mode handlers.

The allocations are protected by a new mutex as kvm->lock is not suitable
for the task because the fault handler is called with the mmap_sem held
but kvmhv_setup_mmu() locks kvm->lock and mmap_sem in the reverse order.

Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
---
Changes:
v4:
* redid locking
* added a separate virtual mode's kvmppc_tce_put to avoid alloc_page()
in otherwise realmode-only code

v3:
* fixed alignments in kvmppc_rm_ioba_validate

v2:
* added kvm mutex around alloc_page to prevent races; in both place we
test the pointer, if NULL, then take a lock and check again so on a fast
path we do not take a lock at all

---
For NVLink2 pass through guests with 128TiB DMA windows and very fragmented
system RAM the difference is gigabytes of RAM.

Fun fact: after this patch the only remaining real+virtual function is
kvmppc_find_table().
---
 arch/powerpc/include/asm/kvm_host.h |  3 ++
 arch/powerpc/include/asm/kvm_ppc.h  |  2 -
 arch/powerpc/kvm/book3s_64_vio.c    | 72 +++++++++++++++++++++++------
 arch/powerpc/kvm/book3s_64_vio_hv.c | 66 +++++++++++++++++++-------
 4 files changed, 110 insertions(+), 33 deletions(-)

diff --git a/arch/powerpc/include/asm/kvm_host.h b/arch/powerpc/include/asm/kvm_host.h
index e6b5bb012ccb..07cefa53222e 100644
--- a/arch/powerpc/include/asm/kvm_host.h
+++ b/arch/powerpc/include/asm/kvm_host.h
@@ -201,6 +201,8 @@ struct kvmppc_spapr_tce_iommu_table {
 	struct kref kref;
 };
 
+#define TCES_PER_PAGE	(PAGE_SIZE / sizeof(u64))
+
 struct kvmppc_spapr_tce_table {
 	struct list_head list;
 	struct kvm *kvm;
@@ -210,6 +212,7 @@ struct kvmppc_spapr_tce_table {
 	u64 offset;		/* in pages */
 	u64 size;		/* window size in pages */
 	struct list_head iommu_tables;
+	struct mutex alloc_lock;
 	struct page *pages[0];
 };
 
diff --git a/arch/powerpc/include/asm/kvm_ppc.h b/arch/powerpc/include/asm/kvm_ppc.h
index e774a386980a..bde566dfd060 100644
--- a/arch/powerpc/include/asm/kvm_ppc.h
+++ b/arch/powerpc/include/asm/kvm_ppc.h
@@ -197,8 +197,6 @@ extern struct kvmppc_spapr_tce_table *kvmppc_find_table(
 		(iommu_tce_check_ioba((stt)->page_shift, (stt)->offset, \
 				(stt)->size, (ioba), (npages)) ?        \
 				H_PARAMETER : H_SUCCESS)
-extern void kvmppc_tce_put(struct kvmppc_spapr_tce_table *tt,
-		unsigned long idx, unsigned long tce);
 extern long kvmppc_h_put_tce(struct kvm_vcpu *vcpu, unsigned long liobn,
 			     unsigned long ioba, unsigned long tce);
 extern long kvmppc_h_put_tce_indirect(struct kvm_vcpu *vcpu,
diff --git a/arch/powerpc/kvm/book3s_64_vio.c b/arch/powerpc/kvm/book3s_64_vio.c
index 5e3b62c491f8..66270e07449a 100644
--- a/arch/powerpc/kvm/book3s_64_vio.c
+++ b/arch/powerpc/kvm/book3s_64_vio.c
@@ -228,11 +228,33 @@ static void release_spapr_tce_table(struct rcu_head *head)
 	unsigned long i, npages = kvmppc_tce_pages(stt->size);
 
 	for (i = 0; i < npages; i++)
-		__free_page(stt->pages[i]);
+		if (stt->pages[i])
+			__free_page(stt->pages[i]);
 
 	kfree(stt);
 }
 
+static struct page *kvm_spapr_get_tce_page(struct kvmppc_spapr_tce_table *stt,
+		unsigned long sttpage)
+{
+	struct page *page = stt->pages[sttpage];
+
+	if (page)
+		return page;
+
+	mutex_lock(&stt->alloc_lock);
+	page = stt->pages[sttpage];
+	if (!page) {
+		page = alloc_page(GFP_KERNEL | __GFP_ZERO);
+		WARN_ON_ONCE(!page);
+		if (page)
+			stt->pages[sttpage] = page;
+	}
+	mutex_unlock(&stt->alloc_lock);
+
+	return page;
+}
+
 static vm_fault_t kvm_spapr_tce_fault(struct vm_fault *vmf)
 {
 	struct kvmppc_spapr_tce_table *stt = vmf->vma->vm_file->private_data;
@@ -241,7 +263,10 @@ static vm_fault_t kvm_spapr_tce_fault(struct vm_fault *vmf)
 	if (vmf->pgoff >= kvmppc_tce_pages(stt->size))
 		return VM_FAULT_SIGBUS;
 
-	page = stt->pages[vmf->pgoff];
+	page = kvm_spapr_get_tce_page(stt, vmf->pgoff);
+	if (!page)
+		return VM_FAULT_OOM;
+
 	get_page(page);
 	vmf->page = page;
 	return 0;
@@ -296,7 +321,6 @@ long kvm_vm_ioctl_create_spapr_tce(struct kvm *kvm,
 	struct kvmppc_spapr_tce_table *siter;
 	unsigned long npages, size = args->size;
 	int ret = -ENOMEM;
-	int i;
 
 	if (!args->size || args->page_shift < 12 || args->page_shift > 34 ||
 		(args->offset + args->size > (ULLONG_MAX >> args->page_shift)))
@@ -318,14 +342,9 @@ long kvm_vm_ioctl_create_spapr_tce(struct kvm *kvm,
 	stt->offset = args->offset;
 	stt->size = size;
 	stt->kvm = kvm;
+	mutex_init(&stt->alloc_lock);
 	INIT_LIST_HEAD_RCU(&stt->iommu_tables);
 
-	for (i = 0; i < npages; i++) {
-		stt->pages[i] = alloc_page(GFP_KERNEL | __GFP_ZERO);
-		if (!stt->pages[i])
-			goto fail;
-	}
-
 	mutex_lock(&kvm->lock);
 
 	/* Check this LIOBN hasn't been previously allocated */
@@ -352,11 +371,6 @@ long kvm_vm_ioctl_create_spapr_tce(struct kvm *kvm,
 	if (ret >= 0)
 		return ret;
 
- fail:
-	for (i = 0; i < npages; i++)
-		if (stt->pages[i])
-			__free_page(stt->pages[i]);
-
 	kfree(stt);
  fail_acct:
 	kvmppc_account_memlimit(kvmppc_stt_pages(npages), false);
@@ -413,6 +427,36 @@ static long kvmppc_tce_validate(struct kvmppc_spapr_tce_table *stt,
 	return H_SUCCESS;
 }
 
+/*
+ * Handles TCE requests for emulated devices.
+ * Puts guest TCE values to the table and expects user space to convert them.
+ * Cannot fail so kvmppc_tce_validate must be called before it.
+ */
+static void kvmppc_tce_put(struct kvmppc_spapr_tce_table *stt,
+		unsigned long idx, unsigned long tce)
+{
+	struct page *page;
+	u64 *tbl;
+	unsigned long sttpage;
+
+	idx -= stt->offset;
+	sttpage = idx / TCES_PER_PAGE;
+	page = stt->pages[sttpage];
+
+	if (!page) {
+		/* We allow any TCE, not just with read|write permissions */
+		if (!tce)
+			return;
+
+		page = kvm_spapr_get_tce_page(stt, sttpage);
+		if (!page)
+			return;
+	}
+	tbl = page_to_virt(page);
+
+	tbl[idx % TCES_PER_PAGE] = tce;
+}
+
 static void kvmppc_clear_tce(struct mm_struct *mm, struct iommu_table *tbl,
 		unsigned long entry)
 {
diff --git a/arch/powerpc/kvm/book3s_64_vio_hv.c b/arch/powerpc/kvm/book3s_64_vio_hv.c
index 31e59748832a..484b47fa3960 100644
--- a/arch/powerpc/kvm/book3s_64_vio_hv.c
+++ b/arch/powerpc/kvm/book3s_64_vio_hv.c
@@ -66,8 +66,6 @@
 
 #endif
 
-#define TCES_PER_PAGE	(PAGE_SIZE / sizeof(u64))
-
 /*
  * Finds a TCE table descriptor by LIOBN.
  *
@@ -148,7 +146,6 @@ static long kvmppc_rm_tce_validate(struct kvmppc_spapr_tce_table *stt,
 
 	return H_SUCCESS;
 }
-#endif /* CONFIG_KVM_BOOK3S_HV_POSSIBLE */
 
 /* Note on the use of page_address() in real mode,
  *
@@ -180,13 +177,9 @@ static u64 *kvmppc_page_address(struct page *page)
 /*
  * Handles TCE requests for emulated devices.
  * Puts guest TCE values to the table and expects user space to convert them.
- * Called in both real and virtual modes.
- * Cannot fail so kvmppc_tce_validate must be called before it.
- *
- * WARNING: This will be called in real-mode on HV KVM and virtual
- *          mode on PR KVM
+ * Cannot fail so kvmppc_rm_tce_validate must be called before it.
  */
-void kvmppc_tce_put(struct kvmppc_spapr_tce_table *stt,
+static void kvmppc_rm_tce_put(struct kvmppc_spapr_tce_table *stt,
 		unsigned long idx, unsigned long tce)
 {
 	struct page *page;
@@ -194,13 +187,48 @@ void kvmppc_tce_put(struct kvmppc_spapr_tce_table *stt,
 
 	idx -= stt->offset;
 	page = stt->pages[idx / TCES_PER_PAGE];
+	/*
+	 * page must not be NULL in real mode,
+	 * kvmppc_rm_ioba_validate() must have taken care of this.
+	 */
+	WARN_ON_ONCE_RM(!page);
 	tbl = kvmppc_page_address(page);
 
 	tbl[idx % TCES_PER_PAGE] = tce;
 }
-EXPORT_SYMBOL_GPL(kvmppc_tce_put);
 
-#ifdef CONFIG_KVM_BOOK3S_HV_POSSIBLE
+/*
+ * TCEs pages are allocated in kvmppc_rm_tce_put() which won't be able to do so
+ * in real mode.
+ * Check if kvmppc_rm_tce_put() can succeed in real mode, i.e. a TCEs page is
+ * allocated or not required (when clearing a tce entry).
+ */
+static long kvmppc_rm_ioba_validate(struct kvmppc_spapr_tce_table *stt,
+		unsigned long ioba, unsigned long npages, bool clearing)
+{
+	unsigned long i, idx, sttpage, sttpages;
+	unsigned long ret = kvmppc_ioba_validate(stt, ioba, npages);
+
+	if (ret)
+		return ret;
+	/*
+	 * clearing=true says kvmppc_rm_tce_put won't be allocating pages
+	 * for empty tces.
+	 */
+	if (clearing)
+		return H_SUCCESS;
+
+	idx = (ioba >> stt->page_shift) - stt->offset;
+	sttpage = idx / TCES_PER_PAGE;
+	sttpages = _ALIGN_UP(idx % TCES_PER_PAGE + npages, TCES_PER_PAGE) /
+			TCES_PER_PAGE;
+	for (i = sttpage; i < sttpage + sttpages; ++i)
+		if (!stt->pages[i])
+			return H_TOO_HARD;
+
+	return H_SUCCESS;
+}
+
 static long iommu_tce_xchg_rm(struct mm_struct *mm, struct iommu_table *tbl,
 		unsigned long entry, unsigned long *hpa,
 		enum dma_data_direction *direction)
@@ -378,7 +406,7 @@ long kvmppc_rm_h_put_tce(struct kvm_vcpu *vcpu, unsigned long liobn,
 	if (!stt)
 		return H_TOO_HARD;
 
-	ret = kvmppc_ioba_validate(stt, ioba, 1);
+	ret = kvmppc_rm_ioba_validate(stt, ioba, 1, tce = 0);
 	if (ret != H_SUCCESS)
 		return ret;
 
@@ -406,7 +434,7 @@ long kvmppc_rm_h_put_tce(struct kvm_vcpu *vcpu, unsigned long liobn,
 		}
 	}
 
-	kvmppc_tce_put(stt, entry, tce);
+	kvmppc_rm_tce_put(stt, entry, tce);
 
 	return H_SUCCESS;
 }
@@ -477,7 +505,7 @@ long kvmppc_rm_h_put_tce_indirect(struct kvm_vcpu *vcpu,
 	if (tce_list & (SZ_4K - 1))
 		return H_PARAMETER;
 
-	ret = kvmppc_ioba_validate(stt, ioba, npages);
+	ret = kvmppc_rm_ioba_validate(stt, ioba, npages, false);
 	if (ret != H_SUCCESS)
 		return ret;
 
@@ -554,7 +582,7 @@ long kvmppc_rm_h_put_tce_indirect(struct kvm_vcpu *vcpu,
 			}
 		}
 
-		kvmppc_tce_put(stt, entry + i, tce);
+		kvmppc_rm_tce_put(stt, entry + i, tce);
 	}
 
 unlock_exit:
@@ -580,7 +608,7 @@ long kvmppc_rm_h_stuff_tce(struct kvm_vcpu *vcpu,
 	if (!stt)
 		return H_TOO_HARD;
 
-	ret = kvmppc_ioba_validate(stt, ioba, npages);
+	ret = kvmppc_rm_ioba_validate(stt, ioba, npages, tce_value = 0);
 	if (ret != H_SUCCESS)
 		return ret;
 
@@ -607,7 +635,7 @@ long kvmppc_rm_h_stuff_tce(struct kvm_vcpu *vcpu,
 	}
 
 	for (i = 0; i < npages; ++i, ioba += (1ULL << stt->page_shift))
-		kvmppc_tce_put(stt, ioba >> stt->page_shift, tce_value);
+		kvmppc_rm_tce_put(stt, ioba >> stt->page_shift, tce_value);
 
 	return H_SUCCESS;
 }
@@ -632,6 +660,10 @@ long kvmppc_h_get_tce(struct kvm_vcpu *vcpu, unsigned long liobn,
 
 	idx = (ioba >> stt->page_shift) - stt->offset;
 	page = stt->pages[idx / TCES_PER_PAGE];
+	if (!page) {
+		vcpu->arch.regs.gpr[4] = 0;
+		return H_SUCCESS;
+	}
 	tbl = (u64 *)page_address(page);
 
 	vcpu->arch.regs.gpr[4] = tbl[idx % TCES_PER_PAGE];
-- 
2.17.1