Skip to contentSkip to conversation

mm/kvm/vfio/ppc64: Migrate compound pages out of CMA region

PATCH Superseded Aneesh Kumar K.V 3 participants 15 messages Jan 14, 2019 – Feb 27, 2019

Subsystems (6)

v7 of 8 Superseded Latest is v8 →

lore
Share & tools
lore
#0 lore raw

v7 cover mm/kvm/vfio/ppc64: Migrate compound pages out of CMA region

ppc64 use CMA area for the allocation of guest page table (hash page table). We won't be able to start guest if we fail to allocate hash page table. We have observed hash table allocation failure because we failed to migrate pages out of CMA region because they were pinned. This happen when we are using VFIO. VFIO on ppc64 pins the entire guest RAM. If the guest RAM pages get allocated out of CMA region, we won't be able to migrate those pages. The pages are also pinned for the lifetime of the guest.

Currently we support migration of non-compound pages. With THP and with the addition of
hugetlb migration we can end up allocating compound pages from CMA region. This
patch series add support for migrating compound pages.

Changes from V6:
* use get_user_pages_longterm instead of get_user_pages_cma_migrate()

Changes from V5:
* Add PF_MEMALLOC_NOCMA
* remote __GFP_THISNODE when allocating target page for migration

Changes from V4:
* use __GFP_NOWARN when allocating pages to avoid page allocation failure warnings.

Changes from V3:
* Move the hugetlb check before transhuge check
* Use compound head page when isolating hugetlb page

Aneesh Kumar K.V (4):
mm/cma: Add PF flag to force non cma alloc
mm: Update get_user_pages_longterm to migrate pages allocated from CMA
region
powerpc/mm/iommu: Allow migration of cma allocated pages during
mm_iommu_do_alloc
powerpc/mm/iommu: Allow large IOMMU page size only for hugetlb backing

 arch/powerpc/mm/mmu_context_iommu.c | 146 ++++++--------------
 include/linux/hugetlb.h             |   2 +
 include/linux/mm.h                  |   3 +-
 include/linux/sched.h               |   1 +
 include/linux/sched/mm.h            |  48 +++++--
 mm/gup.c                            | 200 ++++++++++++++++++++++++----
 mm/hugetlb.c                        |   4 +-
 7 files changed, 267 insertions(+), 137 deletions(-)
Signature · 1 line
2.20.1
#1 lore raw

PATCH v7 1/4 mm/cma: Add PF flag to force non cma alloc

This patch adds PF_MEMALLOC_NOCMA which make sure any allocation in that context is marked non-movable and hence cannot be satisfied by CMA region.

This is useful with get_user_pages_longterm where we want to take a page pin by migrating pages from CMA region. Marking the section PF_MEMALLOC_NOCMA ensures that we avoid unnecessary page migration later.

---
 include/linux/sched.h    |  1 +
 include/linux/sched/mm.h | 48 +++++++++++++++++++++++++++++++++-------
 2 files changed, 41 insertions(+), 8 deletions(-)
include/linux/sched.h +1 -0
index 224666226e87..687b0edf97b2 100644
@@ -1406,6 +1406,7 @@ extern struct pid *cad_pid;
 #define PF_RANDOMIZE		0x00400000	/* Randomize virtual address space */
 #define PF_SWAPWRITE		0x00800000	/* Allowed to write to swap */
 #define PF_MEMSTALL		0x01000000	/* Stalled due to lack of memory */
+#define PF_MEMALLOC_NOCMA	0x02000000	/* All allocation request will have _GFP_MOVABLE cleared */
 #define PF_NO_SETAFFINITY	0x04000000	/* Userland is not allowed to meddle with cpus_allowed */
 #define PF_MCE_EARLY		0x08000000      /* Early kill for mce process policy */
 #define PF_MUTEX_TESTER		0x20000000	/* Thread belongs to the rt mutex tester */
include/linux/sched/mm.h +40 -8

63 lines not on this page · raw message

Signature · 1 line
2.20.1
maintainer
#2 lore raw

Re: [PATCH V7 1/4] mm/cma: Add PF flag to force non cma alloc

Quoted 12 lines · On Mon, 14 Jan 2019 15:24:33 +0530 "Aneesh Kumar K.V"…

This patch adds PF_MEMALLOC_NOCMA which make sure any allocation in that context is marked non-movable and hence cannot be satisfied by CMA region.

This is useful with get_user_pages_longterm where we want to take a page pin by migrating pages from CMA region. Marking the section PF_MEMALLOC_NOCMA ensures that we avoid unnecessary page migration later.

...

include/linux/sched.h extern struct pid *cad_pid; · line 1412

Quoted 7 lines
include/linux/sched.h +1 -0
@@ -1406,6 +1406,7 @@ extern struct pid *cad_pid;
 #define PF_RANDOMIZE		0x00400000	/* Randomize virtual address space */
 #define PF_SWAPWRITE		0x00800000	/* Allowed to write to swap */
 #define PF_MEMSTALL		0x01000000	/* Stalled due to lack of memory */
+#define PF_MEMALLOC_NOCMA	0x02000000	/* All allocation request will have _GFP_MOVABLE cleared */
 #define PF_NO_SETAFFINITY	0x04000000	/* Userland is not allowed to meddle with cpus_allowed */
 #define PF_MCE_EARLY		0x08000000      /* Early kill for mce process policy */
 #define PF_MUTEX_TESTER		0x20000000	/* Thread belongs to the rt mutex tester */

This flag has been taken by PF_UMH so I moved it to 0x10000000.

And we have now run out of PF_ flags.

#3 lore raw

PATCH v7 2/4 mm: Update get_user_pages_longterm to migrate pages allocated from CMA region

In reply to Aneesh Kumar K.V

This patch updates get_user_pages_longterm to migrate pages allocated out of CMA region. This makes sure that we don't keep non-movable pages (due to page reference count) in the CMA area.

This will be used by ppc64 in a later patch to avoid pinning pages in the CMA region. ppc64 uses CMA region for allocation of the hardware page table (hash page table) and not able to migrate pages out of CMA region results in page table allocation failures.

One case where we hit this easy is when a guest using a VFIO passthrough device. VFIO locks all the guest's memory and if the guest memory is backed by CMA region, it becomes unmovable resulting in fragmenting the CMA and possibly preventing other guests from allocation a large enough hash page table.

NOTE: We allocate the new page without using __GFP_THISNODE

---
 include/linux/hugetlb.h |   2 +
 include/linux/mm.h      |   3 +-
 mm/gup.c                | 200 +++++++++++++++++++++++++++++++++++-----
 mm/hugetlb.c            |   4 +-
 4 files changed, 182 insertions(+), 27 deletions(-)
include/linux/hugetlb.h +2 -0
index 087fd5f48c91..1eed0cdaec0e 100644
@@ -371,6 +371,8 @@ struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
 				nodemask_t *nmask);
 struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
 				unsigned long address);
+struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
+				     int nid, nodemask_t *nmask);
 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
 			pgoff_t idx);
 
include/linux/mm.h +2 -1
index 80bb6408fe73..20ec56f8e2bb 100644
@@ -1536,7 +1536,8 @@ long get_user_pages_locked(unsigned long start, unsigned long nr_pages,
 		    unsigned int gup_flags, struct page **pages, int *locked);
 long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
 		    struct page **pages, unsigned int gup_flags);
-#ifdef CONFIG_FS_DAX
+
+#if defined(CONFIG_FS_DAX) || defined(CONFIG_CMA)
 long get_user_pages_longterm(unsigned long start, unsigned long nr_pages,
 			    unsigned int gup_flags, struct page **pages,
 			    struct vm_area_struct **vmas);
mm/gup.c +176 -24

231 lines not on this page · raw message

mm/hugetlb.c +2 -2
index df2e7dd5ff17..913862771808 100644
@@ -1586,8 +1586,8 @@ static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
 	return page;
 }
 
-static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
-		int nid, nodemask_t *nmask)
+struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
+				     int nid, nodemask_t *nmask)
 {
 	struct page *page;
 
Signature · 1 line
2.20.1
#4 lore raw

PATCH v7 2/4 mm: Update get_user_pages_longterm to migrate pages allocated from CMA region

In reply to Aneesh Kumar K.V

This patch updates get_user_pages_longterm to migrate pages allocated out of CMA region. This makes sure that we don't keep non-movable pages (due to page reference count) in the CMA area.

This will be used by ppc64 in a later patch to avoid pinning pages in the CMA region. ppc64 uses CMA region for allocation of the hardware page table (hash page table) and not able to migrate pages out of CMA region results in page table allocation failures.

One case where we hit this easy is when a guest using a VFIO passthrough device. VFIO locks all the guest's memory and if the guest memory is backed by CMA region, it becomes unmovable resulting in fragmenting the CMA and possibly preventing other guests from allocation a large enough hash page table.

NOTE: We allocate the new page without using __GFP_THISNODE

---
 include/linux/hugetlb.h |   2 +
 include/linux/mm.h      |   3 +-
 mm/gup.c                | 200 +++++++++++++++++++++++++++++++++++-----
 mm/hugetlb.c            |   4 +-
 4 files changed, 182 insertions(+), 27 deletions(-)
include/linux/hugetlb.h +2 -0
index 087fd5f48c91..1eed0cdaec0e 100644
@@ -371,6 +371,8 @@ struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
 				nodemask_t *nmask);
 struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
 				unsigned long address);
+struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
+				     int nid, nodemask_t *nmask);
 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
 			pgoff_t idx);
 
include/linux/mm.h +2 -1
index 80bb6408fe73..20ec56f8e2bb 100644
@@ -1536,7 +1536,8 @@ long get_user_pages_locked(unsigned long start, unsigned long nr_pages,
 		    unsigned int gup_flags, struct page **pages, int *locked);
 long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
 		    struct page **pages, unsigned int gup_flags);
-#ifdef CONFIG_FS_DAX
+
+#if defined(CONFIG_FS_DAX) || defined(CONFIG_CMA)
 long get_user_pages_longterm(unsigned long start, unsigned long nr_pages,
 			    unsigned int gup_flags, struct page **pages,
 			    struct vm_area_struct **vmas);
mm/gup.c +176 -24

231 lines not on this page · raw message

mm/hugetlb.c +2 -2
index df2e7dd5ff17..913862771808 100644
@@ -1586,8 +1586,8 @@ static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
 	return page;
 }
 
-static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
-		int nid, nodemask_t *nmask)
+struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
+				     int nid, nodemask_t *nmask)
 {
 	struct page *page;
 
Signature · 1 line
2.20.1
#5 lore raw

PATCH v7 3/4 powerpc/mm/iommu: Allow migration of cma allocated pages during mm_iommu_do_alloc

In reply to Aneesh Kumar K.V

The current code doesn't do page migration if the page allocated is a compound page. With HugeTLB migration support, we can end up allocating hugetlb pages from CMA region. Also, THP pages can be allocated from CMA region. This patch updates the code to handle compound pages correctly. The patch also switches to a single get_user_pages with the right count, instead of doing one get_user_pages per page. That avoids reading page table multiple times.

Since these page reference updates are long term pin, switch to get_user_pages_longterm. That makes sure we fail correctly if the guest RAM is backed by DAX pages.

The patch also converts the hpas member of mm_iommu_table_group_mem_t to a union. We use the same storage location to store pointers to struct page. We cannot update all the code path use struct page *, because we access hpas in real mode and we can't do that struct page * to pfn conversion in real mode.

---
 arch/powerpc/mm/mmu_context_iommu.c | 126 +++++++++-------------------
 1 file changed, 39 insertions(+), 87 deletions(-)
arch/powerpc/mm/mmu_context_iommu.c +39 -87
index a712a650a8b6..f11a2f15071f 100644
@@ -21,6 +21,7 @@
 #include <linux/sizes.h>
 #include <asm/mmu_context.h>
 #include <asm/pte-walk.h>
+#include <linux/mm_inline.h>
 
 static DEFINE_MUTEX(mem_list_mutex);
 
@@ -34,8 +35,18 @@ struct mm_iommu_table_group_mem_t {
 	atomic64_t mapped;
 	unsigned int pageshift;
 	u64 ua;			/* userspace address */
-	u64 entries;		/* number of entries in hpas[] */
-	u64 *hpas;		/* vmalloc'ed */
+	u64 entries;		/* number of entries in hpas/hpages[] */
+	/*
+	 * in mm_iommu_get we temporarily use this to store
+	 * struct page address.
+	 *
+	 * We need to convert ua to hpa in real mode. Make it
+	 * simpler by storing physical address.
+	 */
+	union {
+		struct page **hpages;	/* vmalloc'ed */
+		phys_addr_t *hpas;
+	};
 #define MM_IOMMU_TABLE_INVALID_HPA	((uint64_t)-1)
 	u64 dev_hpa;		/* Device memory base address */
 };
@@ -80,64 +91,15 @@ bool mm_iommu_preregistered(struct mm_struct *mm)
 }
 EXPORT_SYMBOL_GPL(mm_iommu_preregistered);
 
-/*
- * Taken from alloc_migrate_target with changes to remove CMA allocations
- */
-struct page *new_iommu_non_cma_page(struct page *page, unsigned long private)
-{
-	gfp_t gfp_mask = GFP_USER;
-	struct page *new_page;
-
-	if (PageCompound(page))
-		return NULL;
-
-	if (PageHighMem(page))
-		gfp_mask |= __GFP_HIGHMEM;
-
-	/*
-	 * We don't want the allocation to force an OOM if possibe
-	 */
-	new_page = alloc_page(gfp_mask | __GFP_NORETRY | __GFP_NOWARN);
-	return new_page;
-}
-
-static int mm_iommu_move_page_from_cma(struct page *page)
-{
-	int ret = 0;
-	LIST_HEAD(cma_migrate_pages);
-
-	/* Ignore huge pages for now */
-	if (PageCompound(page))
-		return -EBUSY;
-
-	lru_add_drain();
-	ret = isolate_lru_page(page);
-	if (ret)
-		return ret;
-
-	list_add(&page->lru, &cma_migrate_pages);
-	put_page(page); /* Drop the gup reference */
-
-	ret = migrate_pages(&cma_migrate_pages, new_iommu_non_cma_page,
-				NULL, 0, MIGRATE_SYNC, MR_CONTIG_RANGE);
-	if (ret) {
-		if (!list_empty(&cma_migrate_pages))
-			putback_movable_pages(&cma_migrate_pages);
-	}
-
-	return 0;
-}
-
 static long mm_iommu_do_alloc(struct mm_struct *mm, unsigned long ua,
-		unsigned long entries, unsigned long dev_hpa,
-		struct mm_iommu_table_group_mem_t **pmem)
+			      unsigned long entries, unsigned long dev_hpa,
+			      struct mm_iommu_table_group_mem_t **pmem)
 {
 	struct mm_iommu_table_group_mem_t *mem;
-	long i, j, ret = 0, locked_entries = 0;
+	long i, ret = 0, locked_entries = 0;
 	unsigned int pageshift;
 	unsigned long flags;
 	unsigned long cur_ua;
-	struct page *page = NULL;
 
 	mutex_lock(&mem_list_mutex);
 
@@ -187,41 +149,27 @@ static long mm_iommu_do_alloc(struct mm_struct *mm, unsigned long ua,
 		goto unlock_exit;
 	}
 
+	down_read(&mm->mmap_sem);
+	ret = get_user_pages_longterm(ua, entries, FOLL_WRITE, mem->hpages, NULL);
+	up_read(&mm->mmap_sem);
+	if (ret != entries) {
+		/* free the reference taken */
+		for (i = 0; i < ret; i++)
+			put_page(mem->hpages[i]);
+
+		vfree(mem->hpas);
+		kfree(mem);
+		ret = -EFAULT;
+		goto unlock_exit;
+	} else {
+		ret = 0;
+	}
+
+	pageshift = PAGE_SHIFT;
 	for (i = 0; i < entries; ++i) {
+		struct page *page = mem->hpages[i];
+
 		cur_ua = ua + (i << PAGE_SHIFT);
-		if (1 != get_user_pages_fast(cur_ua,
-					1/* pages */, 1/* iswrite */, &page)) {
-			ret = -EFAULT;
-			for (j = 0; j < i; ++j)
-				put_page(pfn_to_page(mem->hpas[j] >>
-						PAGE_SHIFT));
-			vfree(mem->hpas);
-			kfree(mem);
-			goto unlock_exit;
-		}
-		/*
-		 * If we get a page from the CMA zone, since we are going to
-		 * be pinning these entries, we might as well move them out
-		 * of the CMA zone if possible. NOTE: faulting in + migration
-		 * can be expensive. Batching can be considered later
-		 */
-		if (is_migrate_cma_page(page)) {
-			if (mm_iommu_move_page_from_cma(page))
-				goto populate;
-			if (1 != get_user_pages_fast(cur_ua,
-						1/* pages */, 1/* iswrite */,
-						&page)) {
-				ret = -EFAULT;
-				for (j = 0; j < i; ++j)
-					put_page(pfn_to_page(mem->hpas[j] >>
-								PAGE_SHIFT));
-				vfree(mem->hpas);
-				kfree(mem);
-				goto unlock_exit;
-			}
-		}
-populate:
-		pageshift = PAGE_SHIFT;
 		if (mem->pageshift > PAGE_SHIFT && PageCompound(page)) {
 			pte_t *pte;
 			struct page *head = compound_head(page);
@@ -239,6 +187,10 @@ static long mm_iommu_do_alloc(struct mm_struct *mm, unsigned long ua,
 			local_irq_restore(flags);
 		}
 		mem->pageshift = min(mem->pageshift, pageshift);
+		/*
+		 * We don't need struct page reference any more, switch
+		 * to physical address.
+		 */
 		mem->hpas[i] = page_to_pfn(page) << PAGE_SHIFT;
 	}
 
Signature · 1 line
2.20.1
reviewer Reviewed-by on v7 3/4
#6 lore raw

Re: [PATCH V7 3/4] powerpc/mm/iommu: Allow migration of cma allocated pages during mm_iommu_do_alloc

Quoted 6 lines · Aneesh Kumar K.V

The current code doesn't do page migration if the page allocated is a compound page. With HugeTLB migration support, we can end up allocating hugetlb pages from CMA region. Also, THP pages can be allocated from CMA region. This patch updates the code to handle compound pages correctly. The patch also switches to a single get_user_pages with the right count, instead of doing one get_user_pages per page. That avoids reading page table multiple times.

It's not very obvious from the above description that the migration logic is now being done by get_user_pages_longterm(), it just looks like it's all being deleted in this patch. Would be good to mention that.

Quoted 3 lines

Since these page reference updates are long term pin, switch to get_user_pages_longterm. That makes sure we fail correctly if the guest RAM is backed by DAX pages.

Can you explain that in more detail?

Quoted 4 lines

The patch also converts the hpas member of mm_iommu_table_group_mem_t to a union. We use the same storage location to store pointers to struct page. We cannot update all the code path use struct page *, because we access hpas in real mode and we can't do that struct page * to pfn conversion in real mode.

That's a pain, it's asking for bugs mixing two different values in the same array. But I guess it's the least worst option.

It sounds like that's a separate change you could do in a separate patch. But it's not, because it's tied to the fact that we're doing a single GUP call.

Quoted 84 lines
arch/powerpc/mm/mmu_context_iommu.c +13 -49
index a712a650a8b6..f11a2f15071f 100644
@@ -21,6 +21,7 @@
 #include <linux/sizes.h>
 #include <asm/mmu_context.h>
 #include <asm/pte-walk.h>
+#include <linux/mm_inline.h>
 
 static DEFINE_MUTEX(mem_list_mutex);
 
@@ -34,8 +35,18 @@ struct mm_iommu_table_group_mem_t {
 	atomic64_t mapped;
 	unsigned int pageshift;
 	u64 ua;			/* userspace address */
-	u64 entries;		/* number of entries in hpas[] */
-	u64 *hpas;		/* vmalloc'ed */
+	u64 entries;		/* number of entries in hpas/hpages[] */
+	/*
+	 * in mm_iommu_get we temporarily use this to store
+	 * struct page address.
+	 *
+	 * We need to convert ua to hpa in real mode. Make it
+	 * simpler by storing physical address.
+	 */
+	union {
+		struct page **hpages;	/* vmalloc'ed */
+		phys_addr_t *hpas;
+	};
 #define MM_IOMMU_TABLE_INVALID_HPA	((uint64_t)-1)
 	u64 dev_hpa;		/* Device memory base address */
 };
@@ -80,64 +91,15 @@ bool mm_iommu_preregistered(struct mm_struct *mm)
 }
 EXPORT_SYMBOL_GPL(mm_iommu_preregistered);
 
-/*
- * Taken from alloc_migrate_target with changes to remove CMA allocations
- */
-struct page *new_iommu_non_cma_page(struct page *page, unsigned long private)
-{
-	gfp_t gfp_mask = GFP_USER;
-	struct page *new_page;
-
-	if (PageCompound(page))
-		return NULL;
-
-	if (PageHighMem(page))
-		gfp_mask |= __GFP_HIGHMEM;
-
-	/*
-	 * We don't want the allocation to force an OOM if possibe
-	 */
-	new_page = alloc_page(gfp_mask | __GFP_NORETRY | __GFP_NOWARN);
-	return new_page;
-}
-
-static int mm_iommu_move_page_from_cma(struct page *page)
-{
-	int ret = 0;
-	LIST_HEAD(cma_migrate_pages);
-
-	/* Ignore huge pages for now */
-	if (PageCompound(page))
-		return -EBUSY;
-
-	lru_add_drain();
-	ret = isolate_lru_page(page);
-	if (ret)
-		return ret;
-
-	list_add(&page->lru, &cma_migrate_pages);
-	put_page(page); /* Drop the gup reference */
-
-	ret = migrate_pages(&cma_migrate_pages, new_iommu_non_cma_page,
-				NULL, 0, MIGRATE_SYNC, MR_CONTIG_RANGE);
-	if (ret) {
-		if (!list_empty(&cma_migrate_pages))
-			putback_movable_pages(&cma_migrate_pages);
-	}
-
-	return 0;
-}

arch/powerpc/mm/mmu_context_iommu.c bool mm_iommu_preregistered(struct mm_struct *mm) · line 99

Quoted 10 lines
arch/powerpc/mm/mmu_context_iommu.c +3 -4
-
 static long mm_iommu_do_alloc(struct mm_struct *mm, unsigned long ua,
-		unsigned long entries, unsigned long dev_hpa,
-		struct mm_iommu_table_group_mem_t **pmem)
+			      unsigned long entries, unsigned long dev_hpa,
+			      struct mm_iommu_table_group_mem_t **pmem)
 {
 	struct mm_iommu_table_group_mem_t *mem;
-	long i, j, ret = 0, locked_entries = 0;
+	long i, ret = 0, locked_entries = 0;

I'd prefer we didn't initialise ret here.

Quoted 15 lines
arch/powerpc/mm/mmu_context_iommu.c +4 -1
 	unsigned int pageshift;
 	unsigned long flags;
 	unsigned long cur_ua;
-	struct page *page = NULL;
 
 	mutex_lock(&mem_list_mutex);
 
@@ -187,41 +149,27 @@ static long mm_iommu_do_alloc(struct mm_struct *mm, unsigned long ua,
 		goto unlock_exit;
 	}
 
+	down_read(&mm->mmap_sem);
+	ret = get_user_pages_longterm(ua, entries, FOLL_WRITE, mem->hpages, NULL);
+	up_read(&mm->mmap_sem);
+	if (ret != entries) {

arch/powerpc/mm/mmu_context_iommu.c static long mm_iommu_do_alloc(struct mm_struct *mm, unsigned long ua, · line 165

Quoted 10 lines
arch/powerpc/mm/mmu_context_iommu.c +10 -0
+		/* free the reference taken */
+		for (i = 0; i < ret; i++)
+			put_page(mem->hpages[i]);
+
+		vfree(mem->hpas);
+		kfree(mem);
+		ret = -EFAULT;
+		goto unlock_exit;
+	} else {
+		ret = 0;

Or here.

Instead it should be set to 0 at good_exit.

Quoted 44 lines
arch/powerpc/mm/mmu_context_iommu.c +5 -33
+	}
+
+	pageshift = PAGE_SHIFT;
 	for (i = 0; i < entries; ++i) {
+		struct page *page = mem->hpages[i];
+
 		cur_ua = ua + (i << PAGE_SHIFT);
-		if (1 != get_user_pages_fast(cur_ua,
-					1/* pages */, 1/* iswrite */, &page)) {
-			ret = -EFAULT;
-			for (j = 0; j < i; ++j)
-				put_page(pfn_to_page(mem->hpas[j] >>
-						PAGE_SHIFT));
-			vfree(mem->hpas);
-			kfree(mem);
-			goto unlock_exit;
-		}
-		/*
-		 * If we get a page from the CMA zone, since we are going to
-		 * be pinning these entries, we might as well move them out
-		 * of the CMA zone if possible. NOTE: faulting in + migration
-		 * can be expensive. Batching can be considered later
-		 */
-		if (is_migrate_cma_page(page)) {
-			if (mm_iommu_move_page_from_cma(page))
-				goto populate;
-			if (1 != get_user_pages_fast(cur_ua,
-						1/* pages */, 1/* iswrite */,
-						&page)) {
-				ret = -EFAULT;
-				for (j = 0; j < i; ++j)
-					put_page(pfn_to_page(mem->hpas[j] >>
-								PAGE_SHIFT));
-				vfree(mem->hpas);
-				kfree(mem);
-				goto unlock_exit;
-			}
-		}
-populate:
-		pageshift = PAGE_SHIFT;
 		if (mem->pageshift > PAGE_SHIFT && PageCompound(page)) {
 			pte_t *pte;
 			struct page *head = compound_head(page);

arch/powerpc/mm/mmu_context_iommu.c static long mm_iommu_do_alloc(struct mm_struct *mm, unsigned long ua, · line 195

Quoted 9 lines
arch/powerpc/mm/mmu_context_iommu.c +4 -0
@@ -239,6 +187,10 @@ static long mm_iommu_do_alloc(struct mm_struct *mm, unsigned long ua,
 			local_irq_restore(flags);
 		}
 		mem->pageshift = min(mem->pageshift, pageshift);
+		/*
+		 * We don't need struct page reference any more, switch
+		 * to physical address.
+		 */
 		mem->hpas[i] = page_to_pfn(page) << PAGE_SHIFT;
 	}

I'm not any sort of expert on this code, but I don't see anything wrong.

cheers

#7 lore raw

Re: [PATCH V7 3/4] powerpc/mm/iommu: Allow migration of cma allocated pages during mm_iommu_do_alloc

Quoted 13 lines · Michael Ellerman
Quoted 6 lines · Aneesh Kumar K.V

The current code doesn't do page migration if the page allocated is a compound page. With HugeTLB migration support, we can end up allocating hugetlb pages from CMA region. Also, THP pages can be allocated from CMA region. This patch updates the code to handle compound pages correctly. The patch also switches to a single get_user_pages with the right count, instead of doing one get_user_pages per page. That avoids reading page table multiple times.

It's not very obvious from the above description that the migration logic is now being done by get_user_pages_longterm(), it just looks like it's all being deleted in this patch. Would be good to mention that.

Quoted 5 lines
Quoted 3 lines

Since these page reference updates are long term pin, switch to get_user_pages_longterm. That makes sure we fail correctly if the guest RAM is backed by DAX pages.

Can you explain that in more detail?

DAX pages lifetime is dictated by file system rules and as such, we need to make sure that we free these pages on operations like truncate and punch hole. If we have long term pin on these pages, which are mostly return to userspace with elevated page count, the entity holding the long term pin may not be aware of the fact that file got truncated and the file system blocks possibly got reused. That can result in corruption.

Work is going on to solve this issue by either making operations like truncate wait or to make these elevated reference counted page/file system blocks not to be released back to the file system free list.

Till then we prevent long term pin on DAX pages.

Now that we have an API for long term pin, we should ideally be using that in the vfio code.

Quoted 6 lines
Quoted 5 lines

The patch also converts the hpas member of mm_iommu_table_group_mem_t to a union. We use the same storage location to store pointers to struct page. We cannot update all the code path use struct page *, because we access hpas in real mode and we can't do that struct page * to pfn conversion in real mode.

Quoted 6 lines

That's a pain, it's asking for bugs mixing two different values in the same array. But I guess it's the least worst option.

It sounds like that's a separate change you could do in a separate patch. But it's not, because it's tied to the fact that we're doing a single GUP call.

-aneesh

#8 lore raw

PATCH v7 4/4 powerpc/mm/iommu: Allow large IOMMU page size only for hugetlb backing

In reply to Aneesh Kumar K.V

THP pages can get split during different code paths. An incremented reference count does imply we will not split the compound page. But the pmd entry can be converted to level 4 pte entries. Keep the code simpler by allowing large IOMMU page size only if the guest ram is backed by hugetlb pages.

---
 arch/powerpc/mm/mmu_context_iommu.c | 24 +++++++-----------------
 1 file changed, 7 insertions(+), 17 deletions(-)
arch/powerpc/mm/mmu_context_iommu.c +7 -17
index f11a2f15071f..74d43f522dc2 100644
@@ -98,8 +98,6 @@ static long mm_iommu_do_alloc(struct mm_struct *mm, unsigned long ua,
 	struct mm_iommu_table_group_mem_t *mem;
 	long i, ret = 0, locked_entries = 0;
 	unsigned int pageshift;
-	unsigned long flags;
-	unsigned long cur_ua;
 
 	mutex_lock(&mem_list_mutex);
 
@@ -169,22 +167,14 @@ static long mm_iommu_do_alloc(struct mm_struct *mm, unsigned long ua,
 	for (i = 0; i < entries; ++i) {
 		struct page *page = mem->hpages[i];
 
-		cur_ua = ua + (i << PAGE_SHIFT);
-		if (mem->pageshift > PAGE_SHIFT && PageCompound(page)) {
-			pte_t *pte;
+		/*
+		 * Allow to use larger than 64k IOMMU pages. Only do that
+		 * if we are backed by hugetlb.
+		 */
+		if ((mem->pageshift > PAGE_SHIFT) && PageHuge(page)) {
 			struct page *head = compound_head(page);
-			unsigned int compshift = compound_order(head);
-			unsigned int pteshift;
-
-			local_irq_save(flags); /* disables as well */
-			pte = find_linux_pte(mm->pgd, cur_ua, NULL, &pteshift);
-
-			/* Double check it is still the same pinned page */
-			if (pte && pte_page(*pte) == head &&
-			    pteshift == compshift + PAGE_SHIFT)
-				pageshift = max_t(unsigned int, pteshift,
-						PAGE_SHIFT);
-			local_irq_restore(flags);
+
+			pageshift = compound_order(head) + PAGE_SHIFT;
 		}
 		mem->pageshift = min(mem->pageshift, pageshift);
 		/*
Signature · 1 line
2.20.1
#9 lore raw

PATCH v7 5/5 testing

In reply to Aneesh Kumar K.V
---
 mm/gup.c | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)
mm/gup.c +4 -2
index 6e8152594e83..91849c39931a 100644
@@ -1226,7 +1226,7 @@ static long check_and_migrate_cma_pages(unsigned long start, long nr_pages,
 		 * be pinning these entries, we might as well move them out
 		 * of the CMA zone if possible.
 		 */
-		if (is_migrate_cma_page(pages[i])) {
+		if (true || is_migrate_cma_page(pages[i])) {
 
 			struct page *head = compound_head(pages[i]);
 
@@ -1256,6 +1256,7 @@ static long check_and_migrate_cma_pages(unsigned long start, long nr_pages,
 		for (i = 0; i < nr_pages; i++)
 			put_page(pages[i]);
 
+		pr_emerg("migrating nr_pages");
 		if (migrate_pages(&cma_page_list, new_non_cma_page,
 				  NULL, 0, MIGRATE_SYNC, MR_CONTIG_RANGE)) {
 			/*
@@ -1274,10 +1275,11 @@ static long check_and_migrate_cma_pages(unsigned long start, long nr_pages,
 		nr_pages = get_user_pages(start, nr_pages, gup_flags, pages, vmas);
 		if ((nr_pages > 0) && migrate_allow) {
 			drain_allow = true;
-			goto check_again;
+			//goto check_again;
 		}
 	}
 
+	pr_emerg("Returning with %ld\n", nr_pages);
 	return nr_pages;
 }
 #else
Signature · 1 line
2.20.1
#10 lore raw

Re: [PATCH V7 5/5] testing

You can ignore this one.

Quoted 39 lines · Aneesh Kumar K.V
---
  mm/gup.c | 6 ++++--
  1 file changed, 4 insertions(+), 2 deletions(-)
mm/gup.c +4 -2
index 6e8152594e83..91849c39931a 100644
@@ -1226,7 +1226,7 @@ static long check_and_migrate_cma_pages(unsigned long start, long nr_pages,
  		 * be pinning these entries, we might as well move them out
  		 * of the CMA zone if possible.
  		 */
-		if (is_migrate_cma_page(pages[i])) {
+		if (true || is_migrate_cma_page(pages[i])) {
 
  			struct page *head = compound_head(pages[i]);
 
@@ -1256,6 +1256,7 @@ static long check_and_migrate_cma_pages(unsigned long start, long nr_pages,
  		for (i = 0; i < nr_pages; i++)
  			put_page(pages[i]);
 
+		pr_emerg("migrating nr_pages");
  		if (migrate_pages(&cma_page_list, new_non_cma_page,
  				  NULL, 0, MIGRATE_SYNC, MR_CONTIG_RANGE)) {
  			/*
@@ -1274,10 +1275,11 @@ static long check_and_migrate_cma_pages(unsigned long start, long nr_pages,
  		nr_pages = get_user_pages(start, nr_pages, gup_flags, pages, vmas);
  		if ((nr_pages > 0) && migrate_allow) {
  			drain_allow = true;
-			goto check_again;
+			//goto check_again;
  		}
  	}
 
+	pr_emerg("Returning with %ld\n", nr_pages);
  	return nr_pages;
  }
  #else
maintainer
#11 lore raw
In reply to Aneesh Kumar K.V
Quoted 11 lines · On Mon, 14 Jan 2019 15:24:32 +0530 "Aneesh Kumar K.V"…

ppc64 use CMA area for the allocation of guest page table (hash page table). We won't be able to start guest if we fail to allocate hash page table. We have observed hash table allocation failure because we failed to migrate pages out of CMA region because they were pinned. This happen when we are using VFIO. VFIO on ppc64 pins the entire guest RAM. If the guest RAM pages get allocated out of CMA region, we won't be able to migrate those pages. The pages are also pinned for the lifetime of the guest.

Currently we support migration of non-compound pages. With THP and with the addition of
hugetlb migration we can end up allocating compound pages from CMA region. This
patch series add support for migrating compound pages.

Very little review activity is in evidence. Please identify some appropriate reviewers and ask them to take a look?

maintainer
#12 lore raw
In reply to Aneesh Kumar K.V

[patch 1/4]: OK. I guess. Was this worth consuming our last PF_ flag? [patch 2/4]: unreviewed
[patch 3/4]: unreviewed, mpe still unhappy, I expect?
[patch 4/4]: unreviewed

#13 lore raw
Quoted 1 line · Andrew Morton

[patch 1/4]: OK. I guess. Was this worth consuming our last PF_ flag?

That was done based on request from Andrea and it also helps in avoiding allocating pages from CMA region where we know we are anyway going to migrate them out. So yes, this helps.

Quoted 2 lines

[patch 2/4]: unreviewed
[patch 3/4]: unreviewed, mpe still unhappy, I expect?

I did reply to that email. I guess mpe is ok with that?

Quoted 1 line

[patch 4/4]: unreviewed

-aneesh

#14 lore raw
Quoted 4 lines · Aneesh Kumar K.V
Quoted 1 line · Andrew Morton

[patch 1/4]: OK. I guess. Was this worth consuming our last PF_ flag?

Quoted 8 lines

That was done based on request from Andrea and it also helps in avoiding allocating pages from CMA region where we know we are anyway going to migrate them out. So yes, this helps.

Quoted 2 lines

[patch 2/4]: unreviewed
[patch 3/4]: unreviewed, mpe still unhappy, I expect?

I did reply to that email. I guess mpe is ok with that?

It would be nice to fold your explanation about DAX into the change log, so it's there for people to see.

And I think my comment about initialising ret still stands.

cheers