[PATCH mm,net-next 1/3] mm: Refactor insert_page to prepare for batched-lock insert.

[PATCH mm,net-next 1/3] mm: Refactor insert_page to prepare for batched-lock insert.

[Arjun Roy]

Add helper methods for vm_insert_page()/insert_page() to prepare for
vm_insert_pages(), which batch-inserts pages to reduce spinlock
operations when inserting multiple consecutive pages into the user
page table.

The intention of this patch-set is to reduce atomic ops for
tcp zerocopy receives, which normally hits the same spinlock multiple
times consecutively.

Signed-off-by: Arjun Roy <arjunroy.kdev@gmail.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Soheil Hassas Yeganeh <soheil@google.com>

---
 mm/memory.c | 41 ++++++++++++++++++++++++-----------------
 1 file changed, 24 insertions(+), 17 deletions(-)

diff --git a/mm/memory.c b/mm/memory.c
index 7c6e19bdc428..7e23a9275386 100644
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -1693,6 +1693,27 @@ pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
 	return pte_alloc_map_lock(mm, pmd, addr, ptl);
 }
 
+static int validate_page_before_insert(struct page *page)
+{
+	if (PageAnon(page) || PageSlab(page))
+		return -EINVAL;
+	flush_dcache_page(page);
+	return 0;
+}
+
+static int insert_page_into_pte_locked(struct mm_struct *mm, pte_t *pte,
+			unsigned long addr, struct page *page, pgprot_t prot)
+{
+	if (!pte_none(*pte))
+		return -EBUSY;
+	/* Ok, finally just insert the thing.. */
+	get_page(page);
+	inc_mm_counter_fast(mm, mm_counter_file(page));
+	page_add_file_rmap(page, false);
+	set_pte_at(mm, addr, pte, mk_pte(page, prot));
+	return 0;
+}
+
 /*
  * This is the old fallback for page remapping.
  *
@@ -1708,28 +1729,14 @@ static int insert_page(struct vm_area_struct *vma, unsigned long addr,
 	pte_t *pte;
 	spinlock_t *ptl;
 
-	retval = -EINVAL;
-	if (PageAnon(page) || PageSlab(page))
+	retval = validate_page_before_insert(page);
+	if (retval)
 		goto out;
 	retval = -ENOMEM;
-	flush_dcache_page(page);
 	pte = get_locked_pte(mm, addr, &ptl);
 	if (!pte)
 		goto out;
-	retval = -EBUSY;
-	if (!pte_none(*pte))
-		goto out_unlock;
-
-	/* Ok, finally just insert the thing.. */
-	get_page(page);
-	inc_mm_counter_fast(mm, mm_counter_file(page));
-	page_add_file_rmap(page, false);
-	set_pte_at(mm, addr, pte, mk_pte(page, prot));
-
-	retval = 0;
-	pte_unmap_unlock(pte, ptl);
-	return retval;
-out_unlock:
+	retval = insert_page_into_pte_locked(mm, pte, addr, page, prot);
 	pte_unmap_unlock(pte, ptl);
 out:
 	return retval;
-- 
2.25.0.rc1.283.g88dfdc4193-goog

[PATCH mm,net-next 2/3] mm: Add vm_insert_pages().

[Arjun Roy]

Add the ability to insert multiple pages at once to a user VM with
lower PTE spinlock operations.

The intention of this patch-set is to reduce atomic ops for
tcp zerocopy receives, which normally hits the same spinlock multiple
times consecutively.

Signed-off-by: Arjun Roy <arjunroy.kdev@gmail.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Soheil Hassas Yeganeh <soheil@google.com>

---
 include/linux/mm.h |   2 +
 mm/memory.c        | 111 ++++++++++++++++++++++++++++++++++++++++++++-
 2 files changed, 111 insertions(+), 2 deletions(-)

diff --git a/include/linux/mm.h b/include/linux/mm.h
index 29c928ba6b42..a3ac40fbe8fd 100644
--- a/include/linux/mm.h
+++ b/include/linux/mm.h
@@ -2563,6 +2563,8 @@ struct vm_area_struct *find_extend_vma(struct mm_struct *, unsigned long addr);
 int remap_pfn_range(struct vm_area_struct *, unsigned long addr,
 			unsigned long pfn, unsigned long size, pgprot_t);
 int vm_insert_page(struct vm_area_struct *, unsigned long addr, struct page *);
+int vm_insert_pages(struct vm_area_struct *vma, unsigned long addr,
+			struct page **pages, unsigned long *num);
 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
 			unsigned long pfn);
 int vm_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
diff --git a/mm/memory.c b/mm/memory.c
index 7e23a9275386..1655c5feaf32 100644
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -1670,8 +1670,7 @@ int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
 }
 EXPORT_SYMBOL_GPL(zap_vma_ptes);
 
-pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
-			spinlock_t **ptl)
+static pmd_t *walk_to_pmd(struct mm_struct *mm, unsigned long addr)
 {
 	pgd_t *pgd;
 	p4d_t *p4d;
@@ -1690,6 +1689,16 @@ pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
 		return NULL;
 
 	VM_BUG_ON(pmd_trans_huge(*pmd));
+	return pmd;
+}
+
+pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
+			spinlock_t **ptl)
+{
+	pmd_t *pmd = walk_to_pmd(mm, addr);
+
+	if (!pmd)
+		return NULL;
 	return pte_alloc_map_lock(mm, pmd, addr, ptl);
 }
 
@@ -1714,6 +1723,15 @@ static int insert_page_into_pte_locked(struct mm_struct *mm, pte_t *pte,
 	return 0;
 }
 
+static int insert_page_in_batch_locked(struct mm_struct *mm, pmd_t *pmd,
+			unsigned long addr, struct page *page, pgprot_t prot)
+{
+	const int err = validate_page_before_insert(page);
+
+	return err ? err : insert_page_into_pte_locked(
+		mm, pte_offset_map(pmd, addr), addr, page, prot);
+}
+
 /*
  * This is the old fallback for page remapping.
  *
@@ -1742,6 +1760,95 @@ static int insert_page(struct vm_area_struct *vma, unsigned long addr,
 	return retval;
 }
 
+/* insert_pages() amortizes the cost of spinlock operations
+ * when inserting pages in a loop.
+ */
+static int insert_pages(struct vm_area_struct *vma, unsigned long addr,
+			struct page **pages, unsigned long *num, pgprot_t prot)
+{
+	pmd_t *pmd = NULL;
+	spinlock_t *pte_lock = NULL;
+	struct mm_struct *const mm = vma->vm_mm;
+	unsigned long curr_page_idx = 0;
+	unsigned long remaining_pages_total = *num;
+	unsigned long pages_to_write_in_pmd;
+	int ret;
+more:
+	ret = -EFAULT;
+	pmd = walk_to_pmd(mm, addr);
+	if (!pmd)
+		goto out;
+
+	pages_to_write_in_pmd = min_t(unsigned long,
+		remaining_pages_total, PTRS_PER_PTE - pte_index(addr));
+
+	/* Allocate the PTE if necessary; takes PMD lock once only. */
+	ret = -ENOMEM;
+	if (pte_alloc(mm, pmd, addr))
+		goto out;
+	pte_lock = pte_lockptr(mm, pmd);
+
+	while (pages_to_write_in_pmd) {
+		int pte_idx = 0;
+		const int batch_size = min_t(int, pages_to_write_in_pmd, 8);
+
+		spin_lock(pte_lock);
+		for (; pte_idx < batch_size; ++pte_idx) {
+			int err = insert_page_in_batch_locked(mm, pmd,
+				addr, pages[curr_page_idx], prot);
+			if (unlikely(err)) {
+				spin_unlock(pte_lock);
+				ret = err;
+				remaining_pages_total -= pte_idx;
+				goto out;
+			}
+			addr += PAGE_SIZE;
+			++curr_page_idx;
+		}
+		spin_unlock(pte_lock);
+		pages_to_write_in_pmd -= batch_size;
+		remaining_pages_total -= batch_size;
+	}
+	if (remaining_pages_total)
+		goto more;
+	ret = 0;
+out:
+	*num = remaining_pages_total;
+	return ret;
+}
+
+/**
+ * vm_insert_pages - insert multiple pages into user vma, batching the pmd lock.
+ * @vma: user vma to map to
+ * @addr: target start user address of these pages
+ * @pages: source kernel pages
+ * @num: in: number of pages to map. out: number of pages that were *not*
+ * mapped. (0 means all pages were successfully mapped).
+ *
+ * Preferred over vm_insert_page() when inserting multiple pages.
+ *
+ * In case of error, we may have mapped a subset of the provided
+ * pages. It is the caller's responsibility to account for this case.
+ *
+ * The same restrictions apply as in vm_insert_page().
+ */
+int vm_insert_pages(struct vm_area_struct *vma, unsigned long addr,
+			struct page **pages, unsigned long *num)
+{
+	const unsigned long end_addr = addr + (*num * PAGE_SIZE) - 1;
+
+	if (addr < vma->vm_start || end_addr >= vma->vm_end)
+		return -EFAULT;
+	if (!(vma->vm_flags & VM_MIXEDMAP)) {
+		BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
+		BUG_ON(vma->vm_flags & VM_PFNMAP);
+		vma->vm_flags |= VM_MIXEDMAP;
+	}
+	/* Defer page refcount checking till we're about to map that page. */
+	return insert_pages(vma, addr, pages, num, vma->vm_page_prot);
+}
+EXPORT_SYMBOL(vm_insert_pages);
+
 /**
  * vm_insert_page - insert single page into user vma
  * @vma: user vma to map to
-- 
2.25.0.rc1.283.g88dfdc4193-goog

[PATCH mm,net-next 3/3] net-zerocopy: Use vm_insert_pages() for tcp rcv zerocopy.

[Arjun Roy]

Use vm_insert_pages() for tcp receive zerocopy. Spin lock cycles
(as reported by perf) drop from a couple of percentage points
to a fraction of a percent. This results in a roughly 6% increase in
efficiency, measured roughly as zerocopy receive count divided by CPU
utilization.

The intention of this patch-set is to reduce atomic ops for
tcp zerocopy receives, which normally hits the same spinlock multiple
times consecutively.

Signed-off-by: Arjun Roy <arjunroy.kdev@gmail.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Soheil Hassas Yeganeh <soheil@google.com>

---
 net/ipv4/tcp.c | 67 +++++++++++++++++++++++++++++++++++++++++++++-----
 1 file changed, 61 insertions(+), 6 deletions(-)

diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
index 34490d972758..52f96c3ceab3 100644
--- a/net/ipv4/tcp.c
+++ b/net/ipv4/tcp.c
@@ -1861,14 +1861,48 @@ int tcp_mmap(struct file *file, struct socket *sock,
 }
 EXPORT_SYMBOL(tcp_mmap);
 
+static int tcp_zerocopy_vm_insert_batch(struct vm_area_struct *vma,
+					struct page **pages,
+					unsigned long pages_to_map,
+					unsigned long *insert_addr,
+					u32 *length_with_pending,
+					u32 *seq,
+					struct tcp_zerocopy_receive *zc)
+{
+	unsigned long pages_remaining = pages_to_map;
+	int bytes_mapped;
+	int ret;
+
+	ret = vm_insert_pages(vma, *insert_addr, pages, &pages_remaining);
+	bytes_mapped = PAGE_SIZE * (pages_to_map - pages_remaining);
+	/* Even if vm_insert_pages fails, it may have partially succeeded in
+	 * mapping (some but not all of the pages).
+	 */
+	*seq += bytes_mapped;
+	*insert_addr += bytes_mapped;
+	if (ret) {
+		/* But if vm_insert_pages did fail, we have to unroll some state
+		 * we speculatively touched before.
+		 */
+		const int bytes_not_mapped = PAGE_SIZE * pages_remaining;
+		*length_with_pending -= bytes_not_mapped;
+		zc->recv_skip_hint += bytes_not_mapped;
+	}
+	return ret;
+}
+
 static int tcp_zerocopy_receive(struct sock *sk,
 				struct tcp_zerocopy_receive *zc)
 {
 	unsigned long address = (unsigned long)zc->address;
 	u32 length = 0, seq, offset, zap_len;
+	#define PAGE_BATCH_SIZE 8
+	struct page *pages[PAGE_BATCH_SIZE];
 	const skb_frag_t *frags = NULL;
 	struct vm_area_struct *vma;
 	struct sk_buff *skb = NULL;
+	unsigned long pg_idx = 0;
+	unsigned long curr_addr;
 	struct tcp_sock *tp;
 	int inq;
 	int ret;
@@ -1901,15 +1935,26 @@ static int tcp_zerocopy_receive(struct sock *sk,
 		zc->recv_skip_hint = zc->length;
 	}
 	ret = 0;
+	curr_addr = address;
 	while (length + PAGE_SIZE <= zc->length) {
 		if (zc->recv_skip_hint < PAGE_SIZE) {
+			/* If we're here, finish the current batch. */
+			if (pg_idx) {
+				ret = tcp_zerocopy_vm_insert_batch(vma, pages,
+								   pg_idx,
+								   &curr_addr,
+								   &length,
+								   &seq, zc);
+				if (ret)
+					goto out;
+				pg_idx = 0;
+			}
 			if (skb) {
 				skb = skb->next;
 				offset = seq - TCP_SKB_CB(skb)->seq;
 			} else {
 				skb = tcp_recv_skb(sk, seq, &offset);
 			}
-
 			zc->recv_skip_hint = skb->len - offset;
 			offset -= skb_headlen(skb);
 			if ((int)offset < 0 || skb_has_frag_list(skb))
@@ -1933,14 +1978,24 @@ static int tcp_zerocopy_receive(struct sock *sk,
 			zc->recv_skip_hint -= remaining;
 			break;
 		}
-		ret = vm_insert_page(vma, address + length,
-				     skb_frag_page(frags));
-		if (ret)
-			break;
+		pages[pg_idx] = skb_frag_page(frags);
+		pg_idx++;
 		length += PAGE_SIZE;
-		seq += PAGE_SIZE;
 		zc->recv_skip_hint -= PAGE_SIZE;
 		frags++;
+		if (pg_idx == PAGE_BATCH_SIZE) {
+			ret = tcp_zerocopy_vm_insert_batch(vma, pages, pg_idx,
+							   &curr_addr, &length,
+							   &seq, zc);
+			if (ret)
+				goto out;
+			pg_idx = 0;
+		}
+	}
+	if (pg_idx) {
+		ret = tcp_zerocopy_vm_insert_batch(vma, pages, pg_idx,
+						   &curr_addr, &length, &seq,
+						   zc);
 	}
 out:
 	up_read(&current->mm->mmap_sem);
-- 
2.25.0.rc1.283.g88dfdc4193-goog