Xen devel - Jan 2007 - [PATCH] linux/i386: allow CONFIG_HIGHPTE on i386 (take 2)

While, as discussed, the performance impact of this option is certainly higher
than on native Linux, the option should not be entirely disallowed if people
want to sacrifice performance for less lowmem pressure.

Signed-off-by: Jan Beulich <jbeulich@novell.com>

Index: sle10-sp1-2007-01-10/arch/i386/Kconfig
==================================================================---
sle10-sp1-2007-01-10.orig/arch/i386/Kconfig	2007-01-10 13:33:54.000000000 +0100
+++ sle10-sp1-2007-01-10/arch/i386/Kconfig	2007-01-09 11:47:18.000000000 +0100
@@ -594,7 +594,7 @@ config HAVE_ARCH_EARLY_PFN_TO_NID
 
 config HIGHPTE
 	bool "Allocate 3rd-level pagetables from highmem"
-	depends on (HIGHMEM4G || HIGHMEM64G) && !X86_XEN
+	depends on HIGHMEM4G || HIGHMEM64G
 	help
 	  The VM uses one page table entry for each page of physical memory.
 	  For systems with a lot of RAM, this can be wasteful of precious
Index: sle10-sp1-2007-01-10/arch/i386/mm/highmem-xen.c
==================================================================---
sle10-sp1-2007-01-10.orig/arch/i386/mm/highmem-xen.c	2007-01-10
13:33:54.000000000 +0100
+++ sle10-sp1-2007-01-10/arch/i386/mm/highmem-xen.c	2007-01-09
12:45:43.000000000 +0100
@@ -129,5 +129,6 @@ struct page *kmap_atomic_to_page(void *p
 EXPORT_SYMBOL(kmap);
 EXPORT_SYMBOL(kunmap);
 EXPORT_SYMBOL(kmap_atomic);
+EXPORT_SYMBOL(kmap_atomic_pte);
 EXPORT_SYMBOL(kunmap_atomic);
 EXPORT_SYMBOL(kmap_atomic_to_page);
Index: sle10-sp1-2007-01-10/arch/i386/mm/pgtable-xen.c
==================================================================---
sle10-sp1-2007-01-10.orig/arch/i386/mm/pgtable-xen.c	2007-01-10
13:33:54.000000000 +0100
+++ sle10-sp1-2007-01-10/arch/i386/mm/pgtable-xen.c	2007-01-10
14:14:47.000000000 +0100
@@ -238,23 +238,41 @@ struct page *pte_alloc_one(struct mm_str
 
 #ifdef CONFIG_HIGHPTE
 	pte = alloc_pages(GFP_KERNEL|__GFP_HIGHMEM|__GFP_REPEAT|__GFP_ZERO, 0);
+	if (pte && PageHighMem(pte)) {
+		struct mmuext_op op;
+
+		kmap_flush_unused();
+		op.cmd = MMUEXT_PIN_L1_TABLE;
+		op.arg1.mfn = pfn_to_mfn(page_to_pfn(pte));
+		BUG_ON(HYPERVISOR_mmuext_op(&op, 1, NULL, DOMID_SELF) < 0);
+	}
 #else
 	pte = alloc_pages(GFP_KERNEL|__GFP_REPEAT|__GFP_ZERO, 0);
+#endif
 	if (pte) {
 		SetPageForeign(pte, pte_free);
 		set_page_count(pte, 1);
 	}
-#endif
 	return pte;
 }
 
 void pte_free(struct page *pte)
 {
-	unsigned long va = (unsigned long)__va(page_to_pfn(pte)<<PAGE_SHIFT);
+	unsigned long pfn = page_to_pfn(pte);
 
-	if (!pte_write(*virt_to_ptep(va)))
-		BUG_ON(HYPERVISOR_update_va_mapping(
-			va, pfn_pte(page_to_pfn(pte), PAGE_KERNEL), 0));
+	if (!PageHighMem(pte)) {
+		unsigned long va = (unsigned long)__va(pfn << PAGE_SHIFT);
+
+		if (!pte_write(*virt_to_ptep(va)))
+			BUG_ON(HYPERVISOR_update_va_mapping(
+			       va, pfn_pte(pfn, PAGE_KERNEL), 0));
+	} else {
+		struct mmuext_op op;
+
+		op.cmd = MMUEXT_UNPIN_TABLE;
+		op.arg1.mfn = pfn_to_mfn(pfn);
+		BUG_ON(HYPERVISOR_mmuext_op(&op, 1, NULL, DOMID_SELF) < 0);
+	}
 
 	ClearPageForeign(pte);
 	set_page_count(pte, 1);
Index: sle10-sp1-2007-01-10/include/asm-i386/mach-xen/asm/pgalloc.h
==================================================================---
sle10-sp1-2007-01-10.orig/include/asm-i386/mach-xen/asm/pgalloc.h	2007-01-10
13:33:54.000000000 +0100
+++ sle10-sp1-2007-01-10/include/asm-i386/mach-xen/asm/pgalloc.h	2007-01-10
14:15:14.000000000 +0100
@@ -42,7 +42,7 @@ extern struct page *pte_alloc_one(struct
 static inline void pte_free_kernel(pte_t *pte)
 {
 	free_page((unsigned long)pte);
-	make_page_writable(pte, XENFEAT_writable_page_tables);
+	make_lowmem_page_writable(pte, XENFEAT_writable_page_tables);
 }
 
 extern void pte_free(struct page *pte);



_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xensource.com
http://lists.xensource.com/xen-devel

On 10/1/07 15:53, "Jan Beulich" <jbeulich@novell.com> wrote:
> While, as discussed, the performance impact of this option is certainly
higher
> than on native Linux, the option should not be entirely disallowed if
people
> want to sacrifice performance for less lowmem pressure.
> 
> Signed-off-by: Jan Beulich <jbeulich@novell.com>
Much of the suckiness will come from the fact that highptes are being pinned
for their lifetime. This means that you don''t avoid emulation traps for
updates during process fork() and exit(), which are big wins. Obviously  the
problem is that kmap_atomic_pte() doesn''t know whether to map the page
writeable or read-only (is it part of a pinned pagetable yet?). We could
steal a page flag, or use page->private, or something like that do give this
info... This patch is fine for now though.

 -- Keir


_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xensource.com
http://lists.xensource.com/xen-devel