qubes-linux-kernel/patches.suse/mm-devzero-optimisation.patch

From: Nick Piggin <npiggin@suse.de>
Subject: mm: /dev/zero optimisation
References: bnc#430738
Patch-mainline: no (could be submit)

Patch for removal of ZERO_PAGE from main VM paths also removed the
/dev/zero optimisation to map directly from ZERO_PAGE when doing
mmap() and also the interesting read(2) "hack" where the MMU was
used to make zero-filling the target buffer zero-copy.

Some benchmarks have run into issues with this. Customers sometimes
use these benchmarks to qualify and test systems, so even if the
benchmarks themselves are "stupid", it saves some trouble to retain
this optimisation for them. Also, while I don't think it was established
that there is a "real" workload where this helps, but it can't be proven
that one does not exist.

Signed-off-by: Nick Piggin <npiggin@suse.de>
---
 drivers/char/mem.c |  105 ++++++++++++++++++++++++++++++++++++++++++++++++++++-
 include/linux/mm.h |    2 +
 mm/memory.c        |   87 +++++++++++++++++++++++++++++++++++++++++++
 3 files changed, 193 insertions(+), 1 deletion(-)

--- a/drivers/char/mem.c
+++ b/drivers/char/mem.c
@@ -639,6 +639,100 @@ static ssize_t splice_write_null(struct
 	return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_null);
 }
 
+#if 1 //ndef CONFIG_XEN
+/*
+ * For fun, we are using the MMU for this.
+ */
+static inline size_t read_zero_pagealigned(char __user * buf, size_t size)
+{
+	struct mm_struct *mm;
+	struct vm_area_struct * vma;
+	unsigned long addr=(unsigned long)buf;
+
+	mm = current->mm;
+	/* Oops, this was forgotten before. -ben */
+	down_read(&mm->mmap_sem);
+
+	/* For private mappings, just map in zero pages. */
+	for (vma = find_vma(mm, addr); vma; vma = vma->vm_next) {
+		unsigned long count;
+
+		if (vma->vm_start > addr || (vma->vm_flags & VM_WRITE) == 0)
+			goto out_up;
+		if (vma->vm_flags & (VM_SHARED | VM_HUGETLB))
+			break;
+		count = vma->vm_end - addr;
+		if (count > size)
+			count = size;
+
+		zap_page_range(vma, addr, count, NULL);
+		if (zeromap_page_range(vma, addr, count, PAGE_COPY))
+			break;
+
+		size -= count;
+		buf += count;
+		addr += count;
+		if (size == 0)
+			goto out_up;
+	}
+
+	up_read(&mm->mmap_sem);
+
+	/* The shared case is hard. Let's do the conventional zeroing. */
+	do {
+		unsigned long unwritten = clear_user(buf, PAGE_SIZE);
+		if (unwritten)
+			return size + unwritten - PAGE_SIZE;
+		cond_resched();
+		buf += PAGE_SIZE;
+		size -= PAGE_SIZE;
+	} while (size);
+
+	return size;
+out_up:
+	up_read(&mm->mmap_sem);
+	return size;
+}
+
+static ssize_t read_zero(struct file *file, char __user *buf,
+			 size_t count, loff_t *ppos)
+{
+	unsigned long left, unwritten, written = 0;
+
+	if (!count)
+		return 0;
+
+	if (!access_ok(VERIFY_WRITE, buf, count))
+		return -EFAULT;
+
+	left = count;
+
+	/* do we want to be clever? Arbitrary cut-off */
+	if (count >= PAGE_SIZE*4) {
+		unsigned long partial;
+
+		/* How much left of the page? */
+		partial = (PAGE_SIZE-1) & -(unsigned long) buf;
+		unwritten = clear_user(buf, partial);
+		written = partial - unwritten;
+		if (unwritten)
+			goto out;
+		left -= partial;
+		buf += partial;
+		unwritten = read_zero_pagealigned(buf, left & PAGE_MASK);
+		written += (left & PAGE_MASK) - unwritten;
+		if (unwritten)
+			goto out;
+		buf += left & PAGE_MASK;
+		left &= ~PAGE_MASK;
+	}
+	unwritten = clear_user(buf, left);
+	written += left - unwritten;
+out:
+	return written ? written : -EFAULT;
+}
+
+#else /* CONFIG_XEN */
 static ssize_t read_zero(struct file *file, char __user *buf,
 			 size_t count, loff_t *ppos)
 {
@@ -669,15 +763,24 @@ static ssize_t read_zero(struct file *fi
 	}
 	return written ? written : -EFAULT;
 }
+#endif /* CONFIG_XEN */
 
 static int mmap_zero(struct file *file, struct vm_area_struct *vma)
 {
+	int err = 0;
+
 #ifndef CONFIG_MMU
 	return -ENOSYS;
 #endif
+
 	if (vma->vm_flags & VM_SHARED)
 		return shmem_zero_setup(vma);
-	return 0;
+#if 1 //ndef CONFIG_XEN
+	err = zeromap_page_range(vma, vma->vm_start,
+			vma->vm_end - vma->vm_start, vma->vm_page_prot);
+	BUG_ON(err == -EEXIST);
+#endif
+	return err;
 }
 
 static ssize_t write_full(struct file *file, const char __user *buf,
--- a/include/linux/mm.h
+++ b/include/linux/mm.h
@@ -799,6 +799,8 @@ void free_pgd_range(struct mmu_gather *t
 		unsigned long end, unsigned long floor, unsigned long ceiling);
 int copy_page_range(struct mm_struct *dst, struct mm_struct *src,
 			struct vm_area_struct *vma);
+int zeromap_page_range(struct vm_area_struct *vma, unsigned long from,
+			unsigned long size, pgprot_t prot);
 void unmap_mapping_range(struct address_space *mapping,
 		loff_t const holebegin, loff_t const holelen, int even_cows);
 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -1590,6 +1590,93 @@ struct page *get_dump_page(unsigned long
 }
 #endif /* CONFIG_ELF_CORE */
 
+static int zeromap_pte_range(struct mm_struct *mm, pmd_t *pmd,
+			unsigned long addr, unsigned long end, pgprot_t prot)
+{
+	pte_t *pte;
+	spinlock_t *ptl;
+	int err = 0;
+
+	pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
+	if (!pte)
+		return -EAGAIN;
+	arch_enter_lazy_mmu_mode();
+	do {
+		pte_t zero_pte;
+
+		if (unlikely(!pte_none(*pte))) {
+			err = -EEXIST;
+			pte++;
+			break;
+		}
+		zero_pte = pte_mkspecial(pfn_pte(my_zero_pfn(addr), prot));
+		zero_pte = pte_wrprotect(zero_pte);
+		set_pte_at(mm, addr, pte, zero_pte);
+	} while (pte++, addr += PAGE_SIZE, addr != end);
+	arch_leave_lazy_mmu_mode();
+	pte_unmap_unlock(pte - 1, ptl);
+	return err;
+}
+
+static inline int zeromap_pmd_range(struct mm_struct *mm, pud_t *pud,
+			unsigned long addr, unsigned long end, pgprot_t prot)
+{
+	pmd_t *pmd;
+	unsigned long next;
+	int err;
+
+	pmd = pmd_alloc(mm, pud, addr);
+	if (!pmd)
+		return -EAGAIN;
+	do {
+		next = pmd_addr_end(addr, end);
+		err = zeromap_pte_range(mm, pmd, addr, next, prot);
+		if (err)
+			break;
+	} while (pmd++, addr = next, addr != end);
+	return err;
+}
+
+static inline int zeromap_pud_range(struct mm_struct *mm, pgd_t *pgd,
+			unsigned long addr, unsigned long end, pgprot_t prot)
+{
+	pud_t *pud;
+	unsigned long next;
+	int err;
+
+	pud = pud_alloc(mm, pgd, addr);
+	if (!pud)
+		return -EAGAIN;
+	do {
+		next = pud_addr_end(addr, end);
+		err = zeromap_pmd_range(mm, pud, addr, next, prot);
+		if (err)
+			break;
+	} while (pud++, addr = next, addr != end);
+	return err;
+}
+
+int zeromap_page_range(struct vm_area_struct *vma,
+			unsigned long addr, unsigned long size, pgprot_t prot)
+{
+	pgd_t *pgd;
+	unsigned long next;
+	unsigned long end = addr + size;
+	struct mm_struct *mm = vma->vm_mm;
+	int err;
+
+	BUG_ON(addr >= end);
+	pgd = pgd_offset(mm, addr);
+	flush_cache_range(vma, addr, end);
+	do {
+		next = pgd_addr_end(addr, end);
+		err = zeromap_pud_range(mm, pgd, addr, next, prot);
+		if (err)
+			break;
+	} while (pgd++, addr = next, addr != end);
+	return err;
+}
+
 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
 			spinlock_t **ptl)
 {