From d6beea7f9ea1aa2ae5abca7fccb252767820aa13 Mon Sep 17 00:00:00 2001

From: Lynn Boger <laboger@linux.vnet.ibm.com>

Date: Tue, 26 Jul 2016 08:51:10 -0500

Subject: [PATCH] cmd/link: split large elf text sections for ppc64x

Some applications built with Go on ppc64x with

external linking can fail to link with relocation

truncation errors, due to the way that the golang

compiler generates a single go.o file containing

a single large text section to send to the GNU

linker.  If the size of the single text section is

greater than 2^26, this can lead to link errors

due to 24 bit offset field in the bl (call)

instruction.

This fix solves the problem by splitting into

multiple text sections when this limit is reached.

When this is done then the GNU linker can fix the

long calls and insert jump tables where needed.

---

 src/cmd/link/internal/ld/data.go   | 52 +++++++++++++++++++++++++++++--

 src/cmd/link/internal/ld/elf.go    | 60 ++++++++++++++++++++++++++++++++---

 src/cmd/link/internal/ld/lib.go    | 20 ++++++++++++

 src/cmd/link/internal/ld/symtab.go | 64 ++++++++++++++++++++++++++++++++++++++

 src/cmd/link/internal/ppc64/asm.go | 12 ++++---

 src/runtime/symtab.go              | 34 +++++++++++++++-----

 src/runtime/type.go                | 16 +++++++++-

 7 files changed, 237 insertions(+), 21 deletions(-)

diff --git a/src/cmd/link/internal/ld/data.go b/src/cmd/link/internal/ld/data.go

index 57a0dad..58ce18c 100644

--- a/src/cmd/link/internal/ld/data.go

+++ b/src/cmd/link/internal/ld/data.go

@@ -527,7 +527,15 @@ func relocsym(s *LSym) {

 			o = Symaddr(r.Sym) + r.Add - int64(r.Sym.Sect.Vaddr)

 		case obj.R_ADDROFF:

-			o = Symaddr(r.Sym) - int64(r.Sym.Sect.Vaddr) + r.Add

+

+			// The method offset tables using this relocation expect the offset to be relative

+			// to the start of the first text section, even if there are multiple.

+

+			if Linkmode == LinkExternal && r.Sym.Sect.Name == ".text" && r.Sym.Sect.Vaddr != Segtext.Vaddr {

+				o = Symaddr(r.Sym) - int64(Segtext.Vaddr) + r.Add

+			} else {

+				o = Symaddr(r.Sym) - int64(r.Sym.Sect.Vaddr) + r.Add

+			}

 			// r->sym can be null when CALL $(constant) is transformed from absolute PC to relative PC call.

 		case obj.R_CALL, obj.R_GOTPCREL, obj.R_PCREL:

@@ -1926,6 +1934,7 @@ func textaddress() {

 	}

 	va := uint64(INITTEXT)

 	sect.Vaddr = va

+	n := 1

 	for _, sym := range Ctxt.Textp {

 		sym.Sect = sect

 		if sym.Type&obj.SSUB != 0 {

@@ -1948,9 +1957,30 @@ func textaddress() {

 		} else {

 			va += uint64(sym.Size)

 		}

+		// On ppc64x a text section should not be larger than 2^26 bytes due to the size of

+		// call target offset field in the bl instruction.  Splitting into text

+		// sections smaller than this limit allows the GNU linker to modify the long calls

+		// appropriately.  The limit allows for the space for linker tables.

+

+		// Only break at outermost syms.

+

+		if sym.Outer == nil && Iself && Linkmode == LinkExternal && SysArch.InFamily(sys.PPC64) && va-sect.Vaddr > uint64(0x1c00000) {

+

+			// Set the length for the previous text section

+			sect.Length = va - sect.Vaddr

+

+			// Create new section, set the starting Vaddr

+			sect = addsection(&Segtext, ".text", 05)

+			sect.Vaddr = va

+

+			// Create a symbol for the start and end of the secondary text section

+			Linklookup(Ctxt, fmt.Sprintf("runtime.text.%d", n), 0).Sect = sect

+			n++

+		}

 	}

 	sect.Length = va - sect.Vaddr

+	Linklookup(Ctxt, "runtime.etext", 0).Sect = sect

 }

 // assign addresses

@@ -2057,11 +2087,27 @@ func address() {

 	Segdwarf.Filelen = va - Segdwarf.Vaddr

 	text := Segtext.Sect

+	lasttext := text

 	var rodata *Section

 	if Segrodata.Sect != nil {

 		rodata = Segrodata.Sect

 	} else {

-		rodata = text.Next

+		// Could be multiple .text sections

+		n := 1

+		for sect := Segtext.Sect.Next; sect != nil; sect = sect.Next {

+			if sect.Name != ".text" {

+				break

+			}

+			lasttext = sect

+			symname := fmt.Sprintf("runtime.text.%d", n)

+			xdefine(symname, obj.STEXT, int64(sect.Vaddr))

+			n++

+		}

+

+		rodata = lasttext.Next

+		if rodata != nil && rodata.Name != ".rodata" {

+			Diag("Unexpected section order in text segment")

+		}

 	}

 	var relrodata *Section

 	typelink := rodata.Next

@@ -2108,7 +2154,7 @@ func address() {

 	}

 	xdefine("runtime.text", obj.STEXT, int64(text.Vaddr))

-	xdefine("runtime.etext", obj.STEXT, int64(text.Vaddr+text.Length))

+	xdefine("runtime.etext", obj.STEXT, int64(lasttext.Vaddr+lasttext.Length))

 	if HEADTYPE == obj.Hwindows {

 		xdefine(".text", obj.STEXT, int64(text.Vaddr))

 	}

diff --git a/src/cmd/link/internal/ld/elf.go b/src/cmd/link/internal/ld/elf.go

index 39d3609..ecb00c2 100644

--- a/src/cmd/link/internal/ld/elf.go

+++ b/src/cmd/link/internal/ld/elf.go

@@ -1623,6 +1623,25 @@ func elfshname(name string) *ElfShdr {

 	return nil

 }

+// Create an ElfShdr for the section with name.

+// A new one is created even if one already exists with

+// the same name.

+func elfshnamedup(name string) *ElfShdr {

+	var off int

+	var sh *ElfShdr

+

+	for i := 0; i < nelfstr; i++ {

+		if name == elfstr[i].s {

+			off = elfstr[i].off

+			sh = newElfShdr(int64(off))

+			return sh

+		}

+	}

+	Diag("cannot find elf name %s", name)

+	errorexit()

+	return nil

+}

+

 func elfshalloc(sect *Section) *ElfShdr {

 	sh := elfshname(sect.Name)

 	sect.Elfsect = sh

@@ -1630,7 +1649,17 @@ func elfshalloc(sect *Section) *ElfShdr {

 }

 func elfshbits(sect *Section) *ElfShdr {

-	sh := elfshalloc(sect)

+	var sh *ElfShdr

+

+	if sect.Name == ".text" {

+		if sect.Elfsect == nil {

+			sect.Elfsect = elfshnamedup(sect.Name)

+		}

+		sh = sect.Elfsect

+	} else {

+		sh = elfshalloc(sect)

+	}

+

 	// If this section has already been set up as a note, we assume type_ and

 	// flags are already correct, but the other fields still need filling in.

 	if sh.type_ == SHT_NOTE {

@@ -1706,6 +1735,15 @@ func elfshreloc(sect *Section) *ElfShdr {

 	}

 	sh := elfshname(elfRelType + sect.Name)

+

+	// There could be multiple text sections but each needs

+	// its own .rela.text.

+	if sect.Name == ".text" {

+		if sh.info != 0 && sh.info != uint32(sect.Elfsect.shnum) {

+			sh = elfshnamedup(elfRelType + sect.Name)

+		}

+	}

+

 	sh.type_ = uint32(typ)

 	sh.entsize = uint64(SysArch.RegSize) * 2

 	if typ == SHT_RELA {

@@ -1776,9 +1814,12 @@ func Elfemitreloc() {

 		Cput(0)

 	}

-	elfrelocsect(Segtext.Sect, Ctxt.Textp)

-	for sect := Segtext.Sect.Next; sect != nil; sect = sect.Next {

-		elfrelocsect(sect, datap)

+	for sect := Segtext.Sect; sect != nil; sect = sect.Next {

+		if sect.Name == ".text" {

+			elfrelocsect(sect, Ctxt.Textp)

+		} else {

+			elfrelocsect(sect, datap)

+		}

 	}

 	for sect := Segrodata.Sect; sect != nil; sect = sect.Next {

 		elfrelocsect(sect, datap)

@@ -2109,7 +2150,16 @@ func Asmbelfsetup() {

 	elfshname("")

 	for sect := Segtext.Sect; sect != nil; sect = sect.Next {

-		elfshalloc(sect)

+

+		// There could be multiple .text sections. Instead check the Elfsect

+		// field to determine if already has an ElfShdr and if not, create one.

+		if sect.Name == ".text" {

+			if sect.Elfsect == nil {

+				sect.Elfsect = elfshnamedup(sect.Name)

+			}

+		} else {

+			elfshalloc(sect)

+		}

 	}

 	for sect := Segrodata.Sect; sect != nil; sect = sect.Next {

 		elfshalloc(sect)

diff --git a/src/cmd/link/internal/ld/lib.go b/src/cmd/link/internal/ld/lib.go

index 14f4fa9..709e7ca 100644

--- a/src/cmd/link/internal/ld/lib.go

+++ b/src/cmd/link/internal/ld/lib.go

@@ -1956,6 +1956,26 @@ func genasmsym(put func(*LSym, string, int, int64, int64, int, *LSym)) {

 	if s.Type == obj.STEXT {

 		put(s, s.Name, 'T', s.Value, s.Size, int(s.Version), nil)

 	}

+	n := 0

+

+	// Generate base addresses for all text sections if there are multiple

+	for sect := Segtext.Sect; sect != nil; sect = sect.Next {

+		if n == 0 {

+			n++

+			continue

+		}

+		if sect.Name != ".text" {

+			break

+		}

+		s = Linkrlookup(Ctxt, fmt.Sprintf("runtime.text.%d", n), 0)

+		if s == nil {

+			break

+		}

+		if s.Type == obj.STEXT {

+			put(s, s.Name, 'T', s.Value, s.Size, int(s.Version), nil)

+		}

+		n++

+	}

 	s = Linklookup(Ctxt, "runtime.etext", 0)

 	if s.Type == obj.STEXT {

 		put(s, s.Name, 'T', s.Value, s.Size, int(s.Version), nil)

diff --git a/src/cmd/link/internal/ld/symtab.go b/src/cmd/link/internal/ld/symtab.go

index 06d7792..80eb33d 100644

--- a/src/cmd/link/internal/ld/symtab.go

+++ b/src/cmd/link/internal/ld/symtab.go

@@ -315,6 +315,62 @@ func (libs byPkg) Swap(a, b int) {

 	libs[a], libs[b] = libs[b], libs[a]

 }

+// Create a table with information on the text sections.

+

+func textsectionmap() uint32 {

+

+	t := Linklookup(Ctxt, "runtime.textsectionmap", 0)

+	t.Type = obj.SRODATA

+	t.Attr |= AttrReachable

+	nsections := int64(0)

+

+	for sect := Segtext.Sect; sect != nil; sect = sect.Next {

+		if sect.Name == ".text" {

+			nsections++

+		} else {

+			break

+		}

+	}

+	Symgrow(Ctxt, t, nsections*3*8)

+

+	off := int64(0)

+	n := 0

+

+	// The vaddr for each text section is the difference between the section's

+	// Vaddr and the Vaddr for the first text section as determined at compile

+	// time.

+

+	// The symbol name for the start address of the first text section is

+	// runtime.text.  Additional text sections are named runtime.text.n where n is the

+	// order of creation starting with 1. These symbols provide the section's

+	// start address after relocation by the linker.

+

+	textbase := Segtext.Sect.Vaddr

+	for sect := Segtext.Sect; sect != nil; sect = sect.Next {

+		if sect.Name != ".text" {

+			break

+		}

+		off = setuintxx(Ctxt, t, off, sect.Vaddr-textbase, int64(SysArch.IntSize))

+		off = setuintxx(Ctxt, t, off, sect.Length, int64(SysArch.IntSize))

+		if n == 0 {

+			s := Linkrlookup(Ctxt, "runtime.text", 0)

+			if s == nil {

+				Diag("Unable to find symbol runtime.text\n")

+			}

+			off = setaddr(Ctxt, t, off, s)

+

+		} else {

+			s := Linklookup(Ctxt, fmt.Sprintf("runtime.text.%d", n), 0)

+			if s == nil {

+				Diag("Unable to find symbol runtime.text.%d\n", n)

+			}

+			off = setaddr(Ctxt, t, off, s)

+		}

+		n++

+	}

+	return uint32(n)

+}

+

 func symtab() {

 	dosymtype()

@@ -489,6 +545,8 @@ func symtab() {

 		adduint(Ctxt, abihashgostr, uint64(hashsym.Size))

 	}

+	nsections := textsectionmap()

+

 	// Information about the layout of the executable image for the

 	// runtime to use. Any changes here must be matched by changes to

 	// the definition of moduledata in runtime/symtab.go.

@@ -527,6 +585,12 @@ func symtab() {

 	Addaddr(Ctxt, moduledata, Linklookup(Ctxt, "runtime.gcbss", 0))

 	Addaddr(Ctxt, moduledata, Linklookup(Ctxt, "runtime.types", 0))

 	Addaddr(Ctxt, moduledata, Linklookup(Ctxt, "runtime.etypes", 0))

+

+	// text section information

+	Addaddr(Ctxt, moduledata, Linklookup(Ctxt, "runtime.textsectionmap", 0))

+	adduint(Ctxt, moduledata, uint64(nsections))

+	adduint(Ctxt, moduledata, uint64(nsections))

+

 	// The typelinks slice

 	Addaddr(Ctxt, moduledata, Linklookup(Ctxt, "runtime.typelink", 0))

 	adduint(Ctxt, moduledata, uint64(ntypelinks))

diff --git a/src/cmd/link/internal/ppc64/asm.go b/src/cmd/link/internal/ppc64/asm.go

index bd2e23f..ab59fa8 100644

--- a/src/cmd/link/internal/ppc64/asm.go

+++ b/src/cmd/link/internal/ppc64/asm.go

@@ -813,12 +813,14 @@ func asmb() {

 		ld.Asmbelfsetup()

 	}

-	sect := ld.Segtext.Sect

-	ld.Cseek(int64(sect.Vaddr - ld.Segtext.Vaddr + ld.Segtext.Fileoff))

-	ld.Codeblk(int64(sect.Vaddr), int64(sect.Length))

-	for sect = sect.Next; sect != nil; sect = sect.Next {

+	for sect := ld.Segtext.Sect; sect != nil; sect = sect.Next {

 		ld.Cseek(int64(sect.Vaddr - ld.Segtext.Vaddr + ld.Segtext.Fileoff))

-		ld.Datblk(int64(sect.Vaddr), int64(sect.Length))

+		// Might have multiple text sections

+		if sect.Name == ".text" {

+			ld.Codeblk(int64(sect.Vaddr), int64(sect.Length))

+		} else {

+			ld.Datblk(int64(sect.Vaddr), int64(sect.Length))

+		}

 	}

 	if ld.Segrodata.Filelen > 0 {

diff --git a/src/runtime/symtab.go b/src/runtime/symtab.go

index 4f6fae2..23e80ce 100644

--- a/src/runtime/symtab.go

+++ b/src/runtime/symtab.go

@@ -195,8 +195,9 @@ type moduledata struct {

 	end, gcdata, gcbss    uintptr

 	types, etypes         uintptr

-	typelinks []int32 // offsets from types

-	itablinks []*itab

+	textsectmap []textsect

+	typelinks   []int32 // offsets from types

+	itablinks   []*itab

 	modulename   string

 	modulehashes []modulehash

@@ -226,6 +227,14 @@ type functab struct {

 	funcoff uintptr

 }

+// Mapping information for secondary text sections

+

+type textsect struct {

+	vaddr    uint64  // prelinked section vaddr

+	length   uint64  // section length

+	baseaddr uintptr // relocated section address

+}

+

 const minfunc = 16                 // minimum function size

 const pcbucketsize = 256 * minfunc // size of bucket in the pc->func lookup table

@@ -368,12 +377,23 @@ func findfunc(pc uintptr) *_func {

 	ffb := (*findfuncbucket)(add(unsafe.Pointer(datap.findfunctab), b*unsafe.Sizeof(findfuncbucket{})))

 	idx := ffb.idx + uint32(ffb.subbuckets[i])

 	if pc < datap.ftab[idx].entry {

-		throw("findfunc: bad findfunctab entry")

-	}

-	// linear search to find func with pc >= entry.

-	for datap.ftab[idx+1].entry <= pc {

-		idx++

+		// If there are multiple text sections then the buckets for the secondary

+		// text sections will be off because the addresses in those text sections

+		// were relocated to higher addresses.  Search back to find it.

+

+		for datap.ftab[idx].entry > pc && idx > 0 {

+			idx--

+		}

+		if idx == 0 {

+			throw("findfunc: bad findfunctab entry idx")

+		}

+	} else {

+

+		// linear search to find func with pc >= entry.

+		for datap.ftab[idx+1].entry <= pc {

+			idx++

+		}

 	}

 	return (*_func)(unsafe.Pointer(&datap.pclntable[datap.ftab[idx].funcoff]))

 }

diff --git a/src/runtime/type.go b/src/runtime/type.go

index 5ef11a4..f641adc 100644

--- a/src/runtime/type.go

+++ b/src/runtime/type.go

@@ -257,7 +257,21 @@ func (t *_type) textOff(off textOff) unsafe.Pointer {

 		}

 		return res

 	}

-	res := md.text + uintptr(off)

+	res := uintptr(0)

+

+	// Find the text section range that contains the offset to determine the section's base

+	// address.  In cases where there are multiple text sections, the base address might be

+	// relocated by the linker.

+

+	for i := 0; i < len(md.textsectmap); i++ {

+		sectaddr := md.textsectmap[i].vaddr

+		sectlen := md.textsectmap[i].length

+		if uint64(off) >= sectaddr && uint64(off) <= sectaddr+sectlen {

+			res = md.textsectmap[i].baseaddr + uintptr(off) - uintptr(md.textsectmap[i].vaddr)

+			break

+		}

+	}

+

 	if res > md.etext {

 		println("runtime: textOff", hex(off), "out of range", hex(md.text), "-", hex(md.etext))

 		throw("runtime: text offset out of range")