linux 2.6.18: COMPAT_VDSO

This adds support for CONFIG_COMPAT_VDSO. As this will certainly raise questions, I left in the code needed for an alternative approach (which requires mode C code, but less build script changes).

Quoted 1 line

Quoted 1 line

Keir Fraser <[email protected]> 05.03.07 14:42 >>>

Quoted 3 lines · Jan Beulich

This adds support for CONFIG_COMPAT_VDSO. As this will certainly raise questions, I left in the code needed for an alternative approach (which requires mode C code, but less build script changes).

I thought COMPAT_VDSO meant that the VDSO had to be mapped at a fixed location at top of memory, which is impossible when running on Xen?

It's gonna be really old libc-s that need an absolutely invariable address. But
you certainly recall that in 2.6.16 (and probably earlier) we had a hack to map
it in right below PAGE_OFFSET, which worked too for not really new but also not really old libc-s. It is those intermediate libc (supporting AT_SYSINFO but
not AT_SYSINFO_EHDR) versions that the option attempts to address for Xen. I happen to have a SuSE 9.0 system still in active use that doesn't work without this option, and I think glibc 2.3.2 shouldn't be considered entirely obsolete, yet.

Quoted 1 line

Quoted 1 line

Keir Fraser <[email protected]> 05.03.07 16:55 >>>

Quoted 11 lines · Jan Beulich

It's gonna be really old libc-s that need an absolutely invariable address. But
you certainly recall that in 2.6.16 (and probably earlier) we had a hack to map
it in right below PAGE_OFFSET, which worked too for not really new but also not really old libc-s. It is those intermediate libc (supporting AT_SYSINFO but
not AT_SYSINFO_EHDR) versions that the option attempts to address for Xen. I happen to have a SuSE 9.0 system still in active use that doesn't work without this option, and I think glibc 2.3.2 shouldn't be considered entirely obsolete, yet.

The build-system approach looks simplest (certainly smaller than the C code) but is impenetrable to read. I think we'd take that approach if you add some comments to explain what's going on in the makefile runes.

Quoted 3 lines

The build-system approach looks simplest (certainly smaller than the C code) but is impenetrable to read. I think we'd take that approach if you add some comments to explain what's going on in the makefile runes.

I didn't follow up on this so far because kernel mainline seems to want the fix
for paravirt guests, too, but seems more favorable of the C approach. I haven't
seen anything new on this during the last two days anymore, so would want to re-inquire whether you wouldn't then change your mind on which approach to take.

I didn't follow up on this so far because kernel mainline seems to want the fix for paravirt guests, too, but seems more favorable of the C approach. I haven't seen anything new on this during the last two days anymore, so would want to re-inquire whether you wouldn't then change your mind on which approach to take.

This adds support for CONFIG_COMPAT_VDSO. As this will certainly raise questions, I left in the code needed for an alternative approach (which requires mode C code, but less build script changes).

Signed-off-by Jan Beulich <[email protected]>

Index: head-2007-02-27/arch/i386/Kconfig
===================================================================

arch/i386/Kconfig +0 -1

@@ -819,7 +819,6 @@ config HOTPLUG_CPU
 
 config COMPAT_VDSO
 	bool "Compat VDSO support"
-	depends on !X86_XEN
 	default y
 	help
 	  Map the VDSO to the predictable old-style address too.

Index: head-2007-02-27/arch/i386/kernel/Makefile
===================================================================

arch/i386/kernel/Makefile +46 -0

@@ -74,6 +74,52 @@ $(obj)/vsyscall-%.so: $(src)/vsyscall.ld
 		      $(obj)/vsyscall-%.o $(obj)/$(vsyscall_note) FORCE
 	$(call if_changed,syscall)
 
+ifeq ($(CONFIG_XEN)$(CONFIG_COMPAT_VDSO),yy)
+
+# vsyscall.o also contains the vsyscall DSO relocation info as __initdata.
+# We must build both alternative images before we can assemble it.
+# Note: kbuild does not track this dependency due to usage of .include
+$(obj)/vsyscall.o: $(obj)/vsyscall-int80.rel $(obj)/vsyscall-sysenter.rel
+targets += $(foreach F,int80 sysenter,vsyscall-$F.so.alt vsyscall-$F.rel)
+targets += vsyscall.lds.alt
+
+# The alternative DSO images are built using an alternate base address.
+quiet_cmd_syscall_alt = REBASE  $@
+      cmd_syscall_alt = sed 's,^\([[:space:]]*\.[[:space:]]*=[[:space:]]*\),\1-0x55AA0000 + ,' $< >$@
+
+quiet_cmd_syscall_rel = COMPARE $@
+      cmd_syscall_rel = set -e; \
+			cmp -l $(basename $<) $< \
+			| { read off1 old1 new1; \
+			    read off2 old2 new2; \
+			    test $$(expr $$off1 + 1) = $$off2; \
+			    echo " .long $$(expr $$off1 - 3)"; \
+			    while read off1 old3 new3; do \
+				read off2 old4 new4; \
+				test $$(expr $$off1 + 1) = $$off2; \
+				test $$old1 = $$old3 -a $$new1 = $$new3; \
+				test $$old2 = $$old4 -a $$new2 = $$new4; \
+				echo " .long $$(expr $$off1 - 3)"; \
+			    done; \
+			  } >$@
+
+SYSCFLAGS_vsyscall-sysenter.so.alt = $(vsyscall-flags)
+SYSCFLAGS_vsyscall-int80.so.alt    = $(vsyscall-flags)
+
+$(obj)/vsyscall.lds.alt: $(obj)/vsyscall.lds FORCE
+	$(call if_changed,syscall_alt)
+
+$(obj)/vsyscall-int80.so.alt $(obj)/vsyscall-sysenter.so.alt: \
+$(obj)/vsyscall-%.so.alt: $(obj)/vsyscall.lds.alt \
+		      $(obj)/vsyscall-%.o $(obj)/$(vsyscall_note) FORCE
+	$(call if_changed,syscall)
+
+$(obj)/vsyscall-int80.rel $(obj)/vsyscall-sysenter.rel: \
+$(obj)/vsyscall-%.rel: $(obj)/vsyscall-%.so.alt FORCE
+	$(call if_changed,syscall_rel)
+
+endif
+
 # We also create a special relocatable object that should mirror the symbol
 # table and layout of the linked DSO.  With ld -R we can then refer to
 # these symbols in the kernel code rather than hand-coded addresses.

Index: head-2007-02-27/arch/i386/kernel/sysenter.c
===================================================================

arch/i386/kernel/sysenter.c +138 -0

@@ -25,6 +25,8 @@
 
 #ifdef CONFIG_XEN
 #include <xen/interface/callback.h>
+extern const unsigned long vdso_rel_int80_start[], vdso_rel_int80_end[];
+extern const unsigned long vdso_rel_sysenter_start[], vdso_rel_sysenter_end[];
 #endif
 
 /*

@@ -66,6 +68,142 @@ void enable_sep_cpu(void)
 #endif
 }
 
+#if defined(CONFIG_XEN) && defined(CONFIG_COMPAT_VDSO)
+static void __init relocate_vdso(Elf32_Ehdr *ehdr, unsigned long old_base, unsigned long new_base,
+                                 const unsigned long *reloc_start, const unsigned long *reloc_end)
+{
+#if 1
+	const unsigned long *reloc;
+
+	for (reloc = reloc_start; reloc < reloc_end; ++reloc) {
+		unsigned long *ptr = (void *)((unsigned long)ehdr + *reloc);
+
+		*ptr += new_base - old_base;
+	}
+#else
+	unsigned i, ndynsym = 0, szdynsym = 0;
+	unsigned long dynsym = 0;
+
+	BUG_ON(ehdr->e_ident[EI_MAG0] != ELFMAG0);
+	BUG_ON(ehdr->e_ident[EI_MAG1] != ELFMAG1);
+	BUG_ON(ehdr->e_ident[EI_MAG2] != ELFMAG2);
+	BUG_ON(ehdr->e_ident[EI_MAG3] != ELFMAG3);
+	BUG_ON(ehdr->e_ident[EI_CLASS] != ELFCLASS32);
+	BUG_ON(ehdr->e_ident[EI_DATA] != ELFDATA2LSB);
+	BUG_ON(ehdr->e_ehsize < sizeof(*ehdr));
+	ehdr->e_entry += new_base - old_base;
+	BUG_ON(ehdr->e_phentsize < sizeof(Elf32_Phdr));
+	for (i = 0; i < ehdr->e_phnum; ++i) {
+		Elf32_Phdr *phdr = (void *)((unsigned long)ehdr + ehdr->e_phoff + i * ehdr->e_phentsize);
+
+		phdr->p_vaddr += new_base - old_base;
+		switch(phdr->p_type) {
+		case PT_LOAD:
+		case PT_NOTE:
+			break;
+		case PT_DYNAMIC: {
+				Elf32_Dyn *dyn = (void *)(phdr->p_vaddr - new_base + (unsigned long)ehdr);
+				unsigned j;
+
+				for(j = 0; dyn[j].d_tag != DT_NULL; ++j) {
+					switch(dyn[j].d_tag) {
+					case DT_HASH:
+					case DT_STRTAB:
+					case DT_SYMTAB:
+					case 0x6ffffff0: /* DT_VERSYM */
+					case 0x6ffffffc: /* DT_VERDEF */
+						break;
+					case DT_SONAME:
+					case DT_STRSZ:
+					case 0x6ffffffd: /* DT_VERDEFNUM */
+						continue;
+					case DT_SYMENT:
+						szdynsym = dyn[j].d_un.d_val;
+						continue;
+					default:
+						if (dyn[j].d_tag >= 0x60000000 /* OLD_DT_LOOS */
+						    || dyn[j].d_tag < 31 /* DT_ENCODING */
+						    || !(dyn[j].d_tag & 1)) {
+							printk(KERN_WARNING "vDSO dynamic info %u has unsupported tag
 %08X\n", j, dyn[j].d_tag);
+							WARN_ON(1);
+							continue;
+						}
+						break;
+					}
+					dyn[j].d_un.d_ptr += new_base - old_base;
+					switch(dyn[j].d_tag) {
+					case DT_HASH:
+						ndynsym = ((Elf32_Word *)dyn[j].d_un.d_ptr)[1];
+						break;
+					case DT_SYMTAB:
+						dynsym = dyn[j].d_un.d_ptr;
+						break;
+					}
+				}
+			}
+			break;
+		case PT_GNU_EH_FRAME:
+			/* XXX */
+			break;
+		default:
+			printk(KERN_WARNING "vDSO program header %u has unsupported type %08X\n", i, phdr->p_type);
+			WARN_ON(1);
+			break;
+		}
+	}
+	BUG_ON(ehdr->e_shentsize < sizeof(Elf32_Shdr));
+	BUG_ON(ehdr->e_shnum >= SHN_LORESERVE);
+	for (i = 1; i < ehdr->e_shnum; ++i) {
+		Elf32_Shdr *shdr = (void *)((unsigned long)ehdr + ehdr->e_shoff + i * ehdr->e_shentsize);
+
+		if (!(shdr->sh_flags & SHF_ALLOC))
+			continue;
+		shdr->sh_addr += new_base - old_base;
+		switch(shdr->sh_type) {
+		case SHT_DYNAMIC:
+		case SHT_HASH:
+		case SHT_NOBITS:
+		case SHT_NOTE:
+		case SHT_PROGBITS:
+		case SHT_STRTAB:
+		case 0x6ffffffd: /* SHT_GNU_verdef */
+		case 0x6fffffff: /* SHT_GNU_versym */
+			break;
+		case SHT_DYNSYM:
+			BUG_ON(shdr->sh_entsize < sizeof(Elf32_Sym));
+			if (!szdynsym)
+				szdynsym = shdr->sh_entsize;
+			else
+				WARN_ON(szdynsym != shdr->sh_entsize);
+			if (!ndynsym)
+				ndynsym = shdr->sh_size / szdynsym;
+			else
+				WARN_ON(ndynsym != shdr->sh_size / szdynsym);
+			if (!dynsym)
+				dynsym = shdr->sh_addr;
+			else
+				WARN_ON(dynsym != shdr->sh_addr);
+			break;
+		default:
+			printk(KERN_WARNING "vDSO section %u has unsupported type %08X\n", i, shdr->sh_type);
+			WARN_ON(shdr->sh_size);
+			break;
+		}
+	}
+	dynsym += (unsigned long)ehdr - new_base;
+	for(i = 1; i < ndynsym; ++i) {
+		Elf32_Sym *sym = (void *)(dynsym + i * szdynsym);
+
+		if (sym->st_shndx == SHN_ABS)
+			continue;
+		sym->st_value += new_base - old_base;
+	}
+#endif
+}
+#else
+#define relocate_vdso(ehdr, old, new, start, end) ((void)0)
+#endif
+
 /*
  * These symbols are defined by vsyscall.o to mark the bounds

* of the ELF DSO images included therein.

arch/i386/kernel/sysenter.c +4 -0

@@ -104,12 +245,16 @@ int __init sysenter_setup(void)
 		memcpy(syscall_page,
 		       &vsyscall_int80_start,
 		       &vsyscall_int80_end - &vsyscall_int80_start);
+		relocate_vdso(syscall_page, VDSO_PRELINK, __fix_to_virt(FIX_VDSO),
+		              vdso_rel_int80_start, vdso_rel_int80_end);
 		return 0;
 	}
 
 	memcpy(syscall_page,
 	       &vsyscall_sysenter_start,
 	       &vsyscall_sysenter_end - &vsyscall_sysenter_start);
+	relocate_vdso(syscall_page, VDSO_PRELINK, __fix_to_virt(FIX_VDSO),
+	              vdso_rel_sysenter_start, vdso_rel_sysenter_end);
 
 	return 0;
 }

Index: head-2007-02-27/arch/i386/kernel/vsyscall.S
===================================================================

arch/i386/kernel/vsyscall.S +16 -0

@@ -12,4 +12,20 @@ vsyscall_sysenter_start:
 	.incbin "arch/i386/kernel/vsyscall-sysenter.so"
 vsyscall_sysenter_end:
 
+#if defined(CONFIG_XEN) && defined(CONFIG_COMPAT_VDSO)
+
+	.align 4
+
+	.globl vdso_rel_int80_start, vdso_rel_int80_end
+vdso_rel_int80_start:
+	.include "arch/i386/kernel/vsyscall-int80.rel"
+vdso_rel_int80_end:
+
+	.globl vdso_rel_sysenter_start, vdso_rel_sysenter_end
+vdso_rel_sysenter_start:
+	.include "arch/i386/kernel/vsyscall-sysenter.rel"
+vdso_rel_sysenter_end:
+
+#endif
+
 __FINIT

Index: head-2007-02-27/include/asm-i386/elf.h
===================================================================

include/asm-i386/elf.h +5 -1

@@ -137,7 +137,11 @@ extern int dump_task_extended_fpu (struc
 
 #ifdef CONFIG_COMPAT_VDSO
 # define VDSO_COMPAT_BASE	VDSO_HIGH_BASE
-# define VDSO_PRELINK		VDSO_HIGH_BASE
+# ifndef CONFIG_XEN
+#  define VDSO_PRELINK		VDSO_HIGH_BASE
+# else
+#  define VDSO_PRELINK		(0UL - FIX_VDSO * PAGE_SIZE)
+# endif
 #else
 # define VDSO_COMPAT_BASE	VDSO_BASE
 # define VDSO_PRELINK		0

_______________________________________________
Xen-devel mailing list
[email protected]
http://lists.xensource.com/xen-devel

Ingo just raised this as an issue for paravirt_ops as well. I don't
quite follow what's going on there. My understanding is that there are
some old versions of glibc (which were unreleased CVS snapshots shipped by some vendors) which don't use the vdso's ELF header, but instead have their own canned one which built into the library itself.