Skip to contentSkip to conversation

i386 Vmi memory hole

RFC PATCH In review Zachary Amsden 3 participants 13 messages Mar 13, 2006 – Mar 15, 2006 exec and binfmt

lore
Share & tools
lore
#0 lore raw

RFC PATCH 7/24 i386 Vmi memory hole

Create a configurable hole in the linear address space at the top of memory. A more advanced interface is needed to negotiate how much space the hypervisor is allowed to steal, but in the end, it seems most likely that a fixed constant size will be chosen for the compiled kernel, potentially propagated to an information page used by paravirtual initialization to determine interface compatibility.

arch/i386/Kconfig +9 -0
Index: linux-2.6.16-rc3/arch/i386/Kconfig
===================================================================
@@ -201,6 +201,15 @@ config VMI_DEBUG
 
 endmenu
 
+config MEMORY_HOLE
+	int "Create hole at top of memory (0-256 MB)"
+	range 0 256
+	default "64" if X86_VMI
+	default "0" if !X86_VMI
+	help
+	   Useful for creating a hole in the top of memory when running
+	   inside of a virtual machine monitor.
+
 config ACPI_SRAT
 	bool
 	default y
include/asm-i386/fixmap.h +1 -1
Index: linux-2.6.16-rc3/include/asm-i386/fixmap.h
===================================================================
@@ -20,7 +20,7 @@
  * Leave one empty page between vmalloc'ed areas and
  * the start of the fixmap.
  */
-#define __FIXADDR_TOP	0xfffff000
+#define __FIXADDR_TOP	0xfffff000-(CONFIG_MEMORY_HOLE << 20)
 
 #ifndef __ASSEMBLY__
 #include <linux/kernel.h>
#1 lore raw
Quoted 15 lines · * Zachary Amsden ([email protected]) wrote:

Create a configurable hole in the linear address space at the top of memory. A more advanced interface is needed to negotiate how much space the hypervisor is allowed to steal, but in the end, it seems most likely that a fixed constant size will be chosen for the compiled kernel, potentially propagated to an information page used by paravirtual initialization to determine interface compatibility.

Index: linux-2.6.16-rc3/arch/i386/Kconfig
===================================================================

arch/i386/Kconfig config VMI_DEBUG · line 208

Quoted 8 lines
arch/i386/Kconfig +5 -0
@@ -201,6 +201,15 @@ config VMI_DEBUG
 
 endmenu
 
+config MEMORY_HOLE
+	int "Create hole at top of memory (0-256 MB)"
+	range 0 256
+	default "64" if X86_VMI
+	default "0" if !X86_VMI

Deja-vu ;-) And still works in context of Xen, but we've just let the subarch define the __FIXADDR_TOP. Having it be dynamic could be interesting.

#2 lore raw
Quoted 20 lines · Chris Wright
Quoted 18 lines · * Zachary Amsden ([email protected]) wrote:

Create a configurable hole in the linear address space at the top of memory. A more advanced interface is needed to negotiate how much space the hypervisor is allowed to steal, but in the end, it seems most likely that a fixed constant size will be chosen for the compiled kernel, potentially propagated to an information page used by paravirtual initialization to determine interface compatibility.

Index: linux-2.6.16-rc3/arch/i386/Kconfig
===================================================================

arch/i386/Kconfig +0 -0
@@ -201,6 +201,15 @@ config VMI_DEBUG
 
 endmenu
 
Quoted 11 lines

arch/i386/Kconfig config VMI_DEBUG · line 209

Quoted 6 lines
arch/i386/Kconfig +5 -0
+config MEMORY_HOLE
+	int "Create hole at top of memory (0-256 MB)"
+	range 0 256
+	default "64" if X86_VMI
+	default "0" if !X86_VMI
 

Deja-vu ;-) And still works in context of Xen, but we've just let the subarch define the __FIXADDR_TOP. Having it be dynamic could be interesting.

Here's dynamic. I hope it still applies.

Allow creation of an compile time hole at the top of linear address space.

Extended to allow a dynamic hole in linear address space, 7/2005. This required some serious hacking to get everything perfect, but the end result appears to function quite nicely. Everyone can now share the appreciation of pseudo-undocumented ELF OS fields, which means core dumps, debuggers and even broken or obsolete linkers may continue to work.

arch/i386/Kconfig +14 -0
Index: linux-2.6.13/arch/i386/Kconfig
===================================================================
@@ -127,6 +127,20 @@
 
 endchoice
 
+config RELOCATABLE_FIXMAP
+	bool "Allow the fixmap to be placed dynamically at runtime"
+	depends on EXPERIMENTAL
+	help
+	  Crazy hackers only.
+
+config MEMORY_HOLE
+	int "Create hole at top of memory (0-512 MB)"
+	range 0 512
+	default "0"
+	help
+	  Useful for creating a hole in the top of memory when running
+	  inside of a virtual machine monitor.
+
 config ACPI_SRAT
 	bool
 	default y
arch/i386/kernel/sysenter.c +76 -8

98 lines not on this page · raw message

arch/i386/kernel/asm-offsets.c +4 -0
Index: linux-2.6.13/arch/i386/kernel/asm-offsets.c
===================================================================
@@ -68,5 +68,9 @@
 		 sizeof(struct tss_struct));
 
 	DEFINE(PAGE_SIZE_asm, PAGE_SIZE);
+#ifdef CONFIG_RELOCATABLE_FIXMAP
+	DEFINE(VSYSCALL_BASE, 0);
+#else
 	DEFINE(VSYSCALL_BASE, __fix_to_virt(FIX_VSYSCALL));
+#endif
 }
arch/i386/kernel/signal.c +4 -2
Index: linux-2.6.13/arch/i386/kernel/signal.c
===================================================================
@@ -345,6 +345,8 @@
    See vsyscall-sigreturn.S.  */
 extern void __user __kernel_sigreturn;
 extern void __user __kernel_rt_sigreturn;
+#define kernel_sigreturn  (VSYSCALL_RELOCATION + (void __user *)&__kernel_sigreturn)
+#define kernel_rt_sigreturn  (VSYSCALL_RELOCATION + (void __user *)&__kernel_rt_sigreturn)
 
 static int setup_frame(int sig, struct k_sigaction *ka,
 		       sigset_t *set, struct pt_regs * regs)
@@ -380,7 +382,7 @@
 			goto give_sigsegv;
 	}
 
-	restorer = &__kernel_sigreturn;
+	restorer = kernel_sigreturn;
 	if (ka->sa.sa_flags & SA_RESTORER)
 		restorer = ka->sa.sa_restorer;
 
@@ -476,7 +478,7 @@
 		goto give_sigsegv;
 
 	/* Set up to return from userspace.  */
-	restorer = &__kernel_rt_sigreturn;
+	restorer = kernel_rt_sigreturn;
 	if (ka->sa.sa_flags & SA_RESTORER)
 		restorer = ka->sa.sa_restorer;
 	err |= __put_user(restorer, &frame->pretcode);
arch/i386/kernel/entry.S +4 -0
Index: linux-2.6.13/arch/i386/kernel/entry.S
===================================================================
@@ -200,7 +200,11 @@
 	pushl %ebp
 	pushfl
 	pushl $(__USER_CS)
+#ifdef CONFIG_RELOCATABLE_FIXMAP
+	pushl %ss:SYSENTER_RETURN_ADDR
+#else
 	pushl $SYSENTER_RETURN
+#endif
 
 /*
  * Load the potential sixth argument from user stack.
arch/i386/mm/init.c +10 -0
Index: linux-2.6.13/arch/i386/mm/init.c
===================================================================
@@ -42,6 +42,10 @@
 
 unsigned int __VMALLOC_RESERVE = 128 << 20;
 
+#ifdef CONFIG_RELOCATABLE_FIXMAP
+unsigned long __FIXADDR_TOP = 0;
+#endif
+
 DEFINE_PER_CPU(struct mmu_gather, mmu_gathers);
 unsigned long highstart_pfn, highend_pfn;
 
@@ -478,6 +482,12 @@
 		printk("NX (Execute Disable) protection: active\n");
 #endif
 
+#ifdef CONFIG_RELOCATABLE_FIXMAP
+	if (!__FIXADDR_TOP) 
+		__FIXADDR_TOP =  0xfffff000UL-(CONFIG_MEMORY_HOLE << 20);
+	printk(KERN_INFO "Fixmap top relocated to %lxh\n", __FIXADDR_TOP);
+#endif
+
 	pagetable_init();
 
 	load_cr3(swapper_pg_dir);
include/asm-i386/fixmap.h +7 -1
Index: linux-2.6.13/include/asm-i386/fixmap.h
===================================================================
@@ -20,7 +20,13 @@
  * Leave one empty page between vmalloc'ed areas and
  * the start of the fixmap.
  */
-#define __FIXADDR_TOP	0xfffff000
+#ifdef CONFIG_RELOCATABLE_FIXMAP
+extern unsigned long __FIXADDR_TOP;
+#define VSYSCALL_RELOCATION __fix_to_virt(FIX_VSYSCALL)
+#else
+#define __FIXADDR_TOP	(0xfffff000-(CONFIG_MEMORY_HOLE << 20))
+#define VSYSCALL_RELOCATION 0
+#endif
 
 #ifndef __ASSEMBLY__
 #include <linux/kernel.h>
include/asm-i386/elf.h +1 -1
Index: linux-2.6.13/include/asm-i386/elf.h
===================================================================
@@ -129,7 +129,7 @@
 
 #define VSYSCALL_BASE	(__fix_to_virt(FIX_VSYSCALL))
 #define VSYSCALL_EHDR	((const struct elfhdr *) VSYSCALL_BASE)
-#define VSYSCALL_ENTRY	((unsigned long) &__kernel_vsyscall)
+#define VSYSCALL_ENTRY	((unsigned long) (VSYSCALL_RELOCATION+&__kernel_vsyscall))
 extern void __kernel_vsyscall;
 
 #define ARCH_DLINFO						\
include/linux/elf.h +3 -0
Index: linux-2.6.13/include/linux/elf.h
===================================================================
@@ -138,6 +138,9 @@
 #define DT_DEBUG	21
 #define DT_TEXTREL	22
 #define DT_JMPREL	23
+#define DT_VERSYM	0x6ffffff0
+#define DT_VERDEF	0x6ffffffc
+#define DT_VERNEED	0x6ffffffe
 #define DT_LOPROC	0x70000000
 #define DT_HIPROC	0x7fffffff
 
#3 lore raw
Quoted 7 lines · * Zachary Amsden ([email protected]) wrote:

Allow creation of an compile time hole at the top of linear address space.

Extended to allow a dynamic hole in linear address space, 7/2005. This required some serious hacking to get everything perfect, but the end result appears to function quite nicely. Everyone can now share the appreciation of pseudo-undocumented ELF OS fields, which means core dumps, debuggers and even broken or obsolete linkers may continue to work.

Thanks. Gerd did something similar (although I believe it's simpler, don't recall the relocation magic) for Xen. Either way, it's useful from Xen perspective.

thanks,
-chris

#4 lore raw
Quoted 11 lines · Chris Wright
Quoted 8 lines · * Zachary Amsden ([email protected]) wrote:

Allow creation of an compile time hole at the top of linear address space.

Extended to allow a dynamic hole in linear address space, 7/2005. This required some serious hacking to get everything perfect, but the end result appears to function quite nicely. Everyone can now share the appreciation of pseudo-undocumented ELF OS fields, which means core dumps, debuggers and even broken or obsolete linkers may continue to work.

Quoted 4 lines

Thanks. Gerd did something similar (although I believe it's simpler, don't recall the relocation magic) for Xen. Either way, it's useful from Xen perspective.

I believe Xen disables sysenter. The complications in my patch come from the fact that the vsyscall page has to be relocated dynamically, requiring, basically run time linking on the page and some tweaks to get sysenter to work. If you don't use vsyscall (say, non-TLS glibc), then you don't need that complexity. But I think it might be needed now, even for Xen.

Zach

#5 lore raw
Quoted 13 lines · * Zachary Amsden ([email protected]) wrote:
Quoted 12 lines · Chris Wright
Quoted 9 lines · * Zachary Amsden ([email protected]) wrote:

Allow creation of an compile time hole at the top of linear address space.

Extended to allow a dynamic hole in linear address space, 7/2005. This required some serious hacking to get everything perfect, but the end result
appears to function quite nicely. Everyone can now share the appreciation of pseudo-undocumented ELF OS fields, which means core dumps, debuggers and even broken or obsolete linkers may continue to work.

Quoted 5 lines
Quoted 3 lines

Thanks. Gerd did something similar (although I believe it's simpler, don't recall the relocation magic) for Xen. Either way, it's useful from Xen perspective.

I believe Xen disables sysenter.

Yes, so vsyscall page has int80 implementation.

Quoted 6 lines

The complications in my patch come
from the fact that the vsyscall page has to be relocated dynamically, requiring, basically run time linking on the page and some tweaks to get sysenter to work. If you don't use vsyscall (say, non-TLS glibc), then you don't need that complexity. But I think it might be needed now, even for Xen.

I believe both Xen and execshield move vsyscall out of fixmap, and then map into userspace as normal vma.

thanks,
-chris

#6 lore raw
Quoted 9 lines
Quoted 6 lines

The complications in my patch come
from the fact that the vsyscall page has to be relocated dynamically, requiring, basically run time linking on the page and some tweaks to get sysenter to work. If you don't use vsyscall (say, non-TLS glibc), then you don't need that complexity. But I think it might be needed now, even for Xen.

I believe both Xen and execshield move vsyscall out of fixmap, and then map into userspace as normal vma.

Yep, my patch (attached below for reference) moves the vsyscall page into user address space, just below PAGE_OFFSET. Works basically the same way the vsyscall page is mapped in the ia32 emulation of the x86_64
architecture. Address stays fixed, thus the relocation magic isn't needed.

Once the vsyscall page is moved out of fixmap it's easy to make fixmap movable and thus have a runtime-resizable address space hole at the top of address space. Patch is attached too, although that one is more proof-of-concept, it doesn't make much sense as-is. It has a kernel command line option to specify the top of address space so you can play around with it ...

Both patches are against -rc3 and most likely still apply just fine, havn't tested that though.

cheers,

Gerd

Signature · 3 lines
Gerd 'just married' Hoffmann <[email protected]>
I'm the hacker formerly known as Gerd Knorr.
http://www.suse.de/~kraxel/just-married.jpeg
arch/i386/kernel/asm-offsets.c +1 -1
Index: vanilla-2.6.16-rc3/arch/i386/kernel/asm-offsets.c
===================================================================
@@ -68,5 +68,5 @@ void foo(void)
 		 sizeof(struct tss_struct));
 
 	DEFINE(PAGE_SIZE_asm, PAGE_SIZE);
-	DEFINE(VSYSCALL_BASE, __fix_to_virt(FIX_VSYSCALL));
+	DEFINE(VSYSCALL_BASE, (PAGE_OFFSET - 2*PAGE_SIZE));
 }
arch/i386/kernel/sysenter.c +71 -5

100 lines not on this page · raw message

include/asm-i386/a.out.h +1 -1
Index: vanilla-2.6.16-rc3/include/asm-i386/a.out.h
===================================================================
@@ -19,7 +19,7 @@ struct exec
 
 #ifdef __KERNEL__
 
-#define STACK_TOP	TASK_SIZE
+#define STACK_TOP	(TASK_SIZE - 3*PAGE_SIZE)
 
 #endif
 
include/asm-i386/elf.h +6 -1
Index: vanilla-2.6.16-rc3/include/asm-i386/elf.h
===================================================================
@@ -129,11 +129,16 @@ extern int dump_task_extended_fpu (struc
 #define ELF_CORE_COPY_FPREGS(tsk, elf_fpregs) dump_task_fpu(tsk, elf_fpregs)
 #define ELF_CORE_COPY_XFPREGS(tsk, elf_xfpregs) dump_task_extended_fpu(tsk, elf_xfpregs)
 
-#define VSYSCALL_BASE	(__fix_to_virt(FIX_VSYSCALL))
+#define VSYSCALL_BASE	(PAGE_OFFSET - 2*PAGE_SIZE)
 #define VSYSCALL_EHDR	((const struct elfhdr *) VSYSCALL_BASE)
 #define VSYSCALL_ENTRY	((unsigned long) &__kernel_vsyscall)
 extern void __kernel_vsyscall;
 
+#define ARCH_HAS_SETUP_ADDITIONAL_PAGES
+struct linux_binprm;
+extern int arch_setup_additional_pages(struct linux_binprm *bprm,
+                                       int executable_stack);
+
 #define ARCH_DLINFO						\
 do {								\
 		NEW_AUX_ENT(AT_SYSINFO,	VSYSCALL_ENTRY);	\
include/asm-i386/fixmap.h +0 -9
Index: vanilla-2.6.16-rc3/include/asm-i386/fixmap.h
===================================================================
@@ -52,7 +52,6 @@
  */
 enum fixed_addresses {
 	FIX_HOLE,
-	FIX_VSYSCALL,
 #ifdef CONFIG_X86_LOCAL_APIC
 	FIX_APIC_BASE,	/* local (CPU) APIC) -- required for SMP or not */
 #endif
@@ -116,14 +115,6 @@ extern void __set_fixmap (enum fixed_add
 #define __fix_to_virt(x)	(FIXADDR_TOP - ((x) << PAGE_SHIFT))
 #define __virt_to_fix(x)	((FIXADDR_TOP - ((x)&PAGE_MASK)) >> PAGE_SHIFT)
 
-/*
- * This is the range that is readable by user mode, and things
- * acting like user mode such as get_user_pages.
- */
-#define FIXADDR_USER_START	(__fix_to_virt(FIX_VSYSCALL))
-#define FIXADDR_USER_END	(FIXADDR_USER_START + PAGE_SIZE)
-
-
 extern void __this_fixmap_does_not_exist(void);
 
 /*
include/asm-i386/page.h +2 -0
Index: vanilla-2.6.16-rc3/include/asm-i386/page.h
===================================================================
@@ -139,6 +139,8 @@ extern int page_is_ram(unsigned long pag
 	((current->personality & READ_IMPLIES_EXEC) ? VM_EXEC : 0 ) | \
 		 VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC)
 
+#define __HAVE_ARCH_GATE_AREA 1
+
 #endif /* __KERNEL__ */
 
 #include <asm-generic/page.h>
include/asm-i386/processor.h +1 -1
Index: vanilla-2.6.16-rc3/include/asm-i386/processor.h
===================================================================
@@ -318,7 +318,7 @@ extern int bootloader_type;
 /*
  * User space process size: 3GB (default).
  */
-#define TASK_SIZE	(PAGE_OFFSET)
+#define TASK_SIZE	(PAGE_OFFSET - 3*PAGE_SIZE)
 
 /* This decides where the kernel will search for a free chunk of vm
  * space during mmap's.
arch/i386/kernel/setup.c +6 -0
Index: vanilla-2.6.16-rc3/arch/i386/kernel/setup.c
===================================================================
@@ -922,6 +922,12 @@ static void __init parse_cmdline_early (
 		else if (!memcmp(from, "vmalloc=", 8))
 			__VMALLOC_RESERVE = memparse(from+8, &from);
 
+		/*
+		 * fixmap=addr
+		 */
+		else if (!memcmp(from, "fixmap=", 7))
+			set_fixaddr_top(simple_strtoul(from+7, NULL, 16));
+
 	next_char:
 		c = *(from++);
 		if (!c)
arch/i386/mm/init.c +36 -0
Index: vanilla-2.6.16-rc3/arch/i386/mm/init.c
===================================================================
@@ -628,6 +628,42 @@ void __init mem_init(void)
 		(unsigned long) (totalhigh_pages << (PAGE_SHIFT-10))
 	       );
 
+#if 1 /* double-sanity-check paranoia */
+	printk("virtual kernel memory layout:\n"
+	       "    fixmap  : 0x%08lx - 0x%08lx   (%4ld kB)\n"
+	       "    pkmap   : 0x%08lx - 0x%08lx   (%4ld kB)\n"
+	       "    vmalloc : 0x%08lx - 0x%08lx   (%4ld MB)\n"
+	       "    lowmem  : 0x%08lx - 0x%08lx   (%4ld MB)\n"
+	       "      .init : 0x%08lx - 0x%08lx   (%4ld kB)\n"
+	       "      .data : 0x%08lx - 0x%08lx   (%4ld kB)\n"
+	       "      .text : 0x%08lx - 0x%08lx   (%4ld kB)\n",
+	       FIXADDR_START, FIXADDR_TOP,
+	       (FIXADDR_TOP - FIXADDR_START) >> 10,
+
+	       PKMAP_BASE, PKMAP_BASE+LAST_PKMAP*PAGE_SIZE,
+	       (LAST_PKMAP*PAGE_SIZE) >> 10,
+
+	       VMALLOC_START, VMALLOC_END,
+	       (VMALLOC_END - VMALLOC_START) >> 20,
+
+	       (unsigned long)__va(0), (unsigned long)high_memory,
+	       ((unsigned long)high_memory - (unsigned long)__va(0)) >> 20,
+
+	       (unsigned long)&__init_begin, (unsigned long)&__init_end,
+	       ((unsigned long)&__init_end - (unsigned long)&__init_begin) >> 10,
+
+	       (unsigned long)&_etext, (unsigned long)&_edata,
+	       ((unsigned long)&_edata - (unsigned long)&_etext) >> 10,
+
+	       (unsigned long)&_text, (unsigned long)&_etext,
+	       ((unsigned long)&_etext - (unsigned long)&_text) >> 10);
+
+	BUG_ON(PKMAP_BASE+LAST_PKMAP*PAGE_SIZE > FIXADDR_START);
+	BUG_ON(VMALLOC_END                     > PKMAP_BASE);
+	BUG_ON(VMALLOC_START                   > VMALLOC_END);
+	BUG_ON((unsigned long)high_memory      > VMALLOC_START);
+#endif /* double-sanity-check paranoia */
+
 #ifdef CONFIG_X86_PAE
 	if (!cpu_has_pae)
 		panic("cannot execute a PAE-enabled kernel on a PAE-less CPU!");
arch/i386/mm/pgtable.c +13 -0
Index: vanilla-2.6.16-rc3/arch/i386/mm/pgtable.c
===================================================================
@@ -13,6 +13,7 @@
 #include <linux/slab.h>
 #include <linux/pagemap.h>
 #include <linux/spinlock.h>
+#include <linux/module.h>
 
 #include <asm/system.h>
 #include <asm/pgtable.h>
@@ -138,6 +139,10 @@ void set_pmd_pfn(unsigned long vaddr, un
 	__flush_tlb_one(vaddr);
 }
 
+static int fixmaps = 0;
+unsigned long __FIXADDR_TOP = 0xfffff000;
+EXPORT_SYMBOL(__FIXADDR_TOP);
+
 void __set_fixmap (enum fixed_addresses idx, unsigned long phys, pgprot_t flags)
 {
 	unsigned long address = __fix_to_virt(idx);
@@ -147,6 +152,14 @@ void __set_fixmap (enum fixed_addresses 
 		return;
 	}
 	set_pte_pfn(address, phys >> PAGE_SHIFT, flags);
+	fixmaps++;
+}
+
+void set_fixaddr_top(unsigned long top)
+{
+	BUG_ON(fixmaps > 0);
+	printk("%s: addr=0x%lx\n", __FUNCTION__, top);
+	__FIXADDR_TOP = top - PAGE_SIZE;
 }
 
 pte_t *pte_alloc_one_kernel(struct mm_struct *mm, unsigned long address)
include/asm-i386/fixmap.h +2 -1
Index: vanilla-2.6.16-rc3/include/asm-i386/fixmap.h
===================================================================
@@ -20,7 +20,7 @@
  * Leave one empty page between vmalloc'ed areas and
  * the start of the fixmap.
  */
-#define __FIXADDR_TOP	0xfffff000
+extern unsigned long __FIXADDR_TOP;
 
 #ifndef __ASSEMBLY__
 #include <linux/kernel.h>
@@ -93,6 +93,7 @@ enum fixed_addresses {
 
 extern void __set_fixmap (enum fixed_addresses idx,
 					unsigned long phys, pgprot_t flags);
+extern void set_fixaddr_top(unsigned long top);
 
 #define set_fixmap(idx, phys) \
 		__set_fixmap(idx, phys, PAGE_KERNEL)
include/asm-i386/page.h +1 -1
Index: vanilla-2.6.16-rc3/include/asm-i386/page.h
===================================================================
@@ -121,7 +121,7 @@ extern int page_is_ram(unsigned long pag
 
 #define PAGE_OFFSET		((unsigned long)__PAGE_OFFSET)
 #define VMALLOC_RESERVE		((unsigned long)__VMALLOC_RESERVE)
-#define MAXMEM			(-__PAGE_OFFSET-__VMALLOC_RESERVE)
+#define MAXMEM			(__FIXADDR_TOP-__PAGE_OFFSET-__VMALLOC_RESERVE)
 #define __pa(x)			((unsigned long)(x)-PAGE_OFFSET)
 #define __va(x)			((void *)((unsigned long)(x)+PAGE_OFFSET))
 #define pfn_to_kaddr(pfn)      __va((pfn) << PAGE_SHIFT)
#7 lore raw
Quoted 18 lines · Gerd Hoffmann
Quoted 5 lines
Quoted 2 lines

The complications in my patch come from the fact that the vsyscall page has to be relocated dynamically, requiring, basically run time linking on the page and some tweaks to get sysenter to work. If you don't use vsyscall (say, non-TLS glibc), then you don't need that complexity. But I think it might be needed now, even for Xen.

I believe both Xen and execshield move vsyscall out of fixmap, and then map into userspace as normal vma.

Yep, my patch (attached below for reference) moves the vsyscall page into user address space, just below PAGE_OFFSET. Works basically the same way the vsyscall page is mapped in the ia32 emulation of the x86_64
architecture. Address stays fixed, thus the relocation magic isn't needed.

Once the vsyscall page is moved out of fixmap it's easy to make fixmap movable and thus have a runtime-resizable address space hole at the top of address space. Patch is attached too, although that one is more proof-of-concept, it doesn't make much sense as-is. It has a kernel command line option to specify the top of address space so you can play around with it ...

Quoted 3 lines

Both patches are against -rc3 and most likely still apply just fine, havn't tested that though.

Your patch looks a lot cleaner and less hackish than mine. But I wonder if it still works with kernels that support the sysenter method of calling into the kernel. Look at the following code:

ENTRY(sysenter_entry)
movl TSS_sysenter_esp0(%esp),%esp
sysenter_past_esp:

       STI
       pushl $(__USER_DS)
       pushl %ebp
       pushfl
       pushl $(__USER_CS)
       pushl $SYSENTER_RETURN

SYSENTER_RETURN is a link time constant that is defined based on the location of the vsyscall page. If the vsyscall page can move, this can not be a constant. The reason is, this "fake" exception frame is used to return back to the EIP of the call site, and sysenter does not record the EIP of the call site.

Zach

#8 lore raw
Quoted 10 lines · * Zachary Amsden ([email protected]) wrote:

ENTRY(sysenter_entry)
movl TSS_sysenter_esp0(%esp),%esp
sysenter_past_esp:

       STI
       pushl $(__USER_DS)
       pushl %ebp
       pushfl
       pushl $(__USER_CS)
       pushl $SYSENTER_RETURN
Quoted 5 lines

SYSENTER_RETURN is a link time constant that is defined based on the location of the vsyscall page. If the vsyscall page can move, this can not be a constant. The reason is, this "fake" exception frame is used to return back to the EIP of the call site, and sysenter does not record the EIP of the call site.

It's only real issue for something like execshield. For this it's easy to do the fixed math since it's still at fixed address.

+ DEFINE(VSYSCALL_BASE, (PAGE_OFFSET - 2*PAGE_SIZE));

But execshield has to make SYSENTER_RETURN context sensitive to current since the vdso is mapped at random location.

thanks,
-chris

#9 lore raw
Quoted 12 lines · Chris Wright
Quoted 10 lines · * Zachary Amsden ([email protected]) wrote:

ENTRY(sysenter_entry)
movl TSS_sysenter_esp0(%esp),%esp
sysenter_past_esp:

       STI
       pushl $(__USER_DS)
       pushl %ebp
       pushfl
       pushl $(__USER_CS)
       pushl $SYSENTER_RETURN
Quoted 8 lines
Quoted 2 lines

SYSENTER_RETURN is a link time constant that is defined based on the location of the vsyscall page. If the vsyscall page can move, this can not be a constant. The reason is, this "fake" exception frame is used to return back to the EIP of the call site, and sysenter does not record the EIP of the call site.

It's only real issue for something like execshield. For this it's easy to do the fixed math since it's still at fixed address.

+ DEFINE(VSYSCALL_BASE, (PAGE_OFFSET - 2*PAGE_SIZE));

Ok, I'm confused. What fixed math? The return EIP that is pushed here is used when sysenter is active and you have to IRET back to userspace. If that EIP is dynamically relocatable, you can't do fixed math unless you patch the pushl site dynamically. Notable reasons for returning via IRET on this fake exception frame were (until my recent submission) IOPL changes, but I believe there were more. I will have to inspect the source to determine if that is still the case.

Zach

#10 lore raw
Quoted 3 lines · * Zachary Amsden ([email protected]) wrote:
Quoted 1 line

+ DEFINE(VSYSCALL_BASE, (PAGE_OFFSET - 2*PAGE_SIZE));

Ok, I'm confused. What fixed math?

Sorry, bad choice of words. From above, the VYSYCALL_BASE is known at compile time (in asm-offsets.h). So the SYSENTER_RETURN is still fixed addr. For execshield it's truly dynamic, so you get something like this instead of the constant SYSENTER_RETURN:

- pushl $SYSENTER_RETURN
+ pushl (TI_sysenter_return-THREAD_SIZE+8+4*4)(%esp)

thanks,
-chris

#11 lore raw
In reply to Zachary Amsden
Quoted 5 lines

pushl $SYSENTER_RETURN

SYSENTER_RETURN is a link time constant that is defined based on the location of the vsyscall page. If the vsyscall page can move, this can not be a constant.

The vsyscall page is at PAGE_OFFSET - 2*PAGE_SIZE. It doesn't move. At least not at runtime. At compile time it can change with the new VMSPLIT config options, but that isn't a problem ;)

cheers,

Gerd

Signature · 3 lines
Gerd 'just married' Hoffmann <[email protected]>
I'm the hacker formerly known as Gerd Knorr.
http://www.suse.de/~kraxel/just-married.jpeg
#12 lore raw
Quoted 11 lines · Gerd Hoffmann
Quoted 6 lines

pushl $SYSENTER_RETURN

SYSENTER_RETURN is a link time constant that is defined based on the location of the vsyscall page. If the vsyscall page can move, this can not be a constant.

The vsyscall page is at PAGE_OFFSET - 2*PAGE_SIZE. It doesn't move. At least not at runtime. At compile time it can change with the new VMSPLIT config options, but that isn't a problem ;)

Okay, I get it now. Thanks for the explanation. This certainly does simplify the problem.

Zach