You cannot select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
ELFv2-ABI/specification/ch_4.xml

1128 lines
52 KiB
XML

This file contains ambiguous Unicode characters!

This file contains ambiguous Unicode characters that may be confused with others in your current locale. If your use case is intentional and legitimate, you can safely ignore this warning. Use the Escape button to highlight these characters.

<!--
Copyright (c) 2016 OpenPOWER Foundation
Licensed under the GNU Free Documentation License, Version 1.3;
with no Invariants Sections, with no Front-Cover Texts,
and with no Back-Cover Texts (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.gnu.org/licenses/fdl-1.3.txt
-->
<chapter xmlns="http://docbook.org/ns/docbook"
xmlns:xl="http://www.w3.org/1999/xlink" version="5.0" xml:lang="en">
<title>Program Loading and Dynamic Linking</title>
<section xml:id="dbdoclet.50655242___RefHeading___Toc377640650">
<title>Program Loading</title>
<para>A number of criteria constrain the mapping of an executable file or
shared object file to virtual memory segments. During mapping, the
operating system may use delayed physical reads to improve performance,
which necessitates that file offsets and virtual addresses are congruent,
modulo the page size.</para>
<para>Page size must be less than or equal to the operating system
implemented congruency. This ABI defines 64 KB congruency as the minimum
allowable. To maintain interoperability between operating system
implementations, 64 KB congruency is recommended.</para>
<note>
<para>There is historical precedence for 64 KB congruency in that
there is synergy with the Power Architecture instruction set whereby
low and high adjusted relocations can be easily performed using addi or
addis instructions.</para>
</note>
<para>The value of the p_align member of the program header struct must be
0x10000 or a larger power of 2. If a larger congruency size is used for
large pages, p_align should match the congruency value.</para>
<para>The following program header information illustrates an application
that is mapped with a base address of 0x10000000:</para>
<table frame="all" pgwide="1" xml:id="dbdoclet.50655242_44623">
<title>Program Header Example</title>
<tgroup cols="3">
<colspec colname="c1" colwidth="33*" align="center" />
<colspec colname="c2" colwidth="33*" align="center" />
<colspec colname="c3" colwidth="33*" align="center" />
<thead>
<row>
<entry>
<para>
<emphasis role="bold">Header Member</emphasis>
</para>
</entry>
<entry>
<para>
<emphasis role="bold">Text Segment</emphasis>
</para>
</entry>
<entry>
<para>
<emphasis role="bold">Data Segment</emphasis>
</para>
</entry>
</row>
</thead>
<tbody>
<row>
<entry>
<para>p_type</para>
</entry>
<entry>
<para>PT_LOAD</para>
</entry>
<entry>
<para>PT_LOAD</para>
</entry>
</row>
<row>
<entry>
<para>p_offset</para>
</entry>
<entry>
<para>0x000000</para>
</entry>
<entry>
<para>0x000af0</para>
</entry>
</row>
<row>
<entry>
<para>p_vaddr</para>
</entry>
<entry>
<para>0x10000000</para>
</entry>
<entry>
<para>0x10010af0</para>
</entry>
</row>
<row>
<entry>
<para>p_paddr</para>
</entry>
<entry>
<para>0x10000000</para>
</entry>
<entry>
<para>0x10010af0</para>
</entry>
</row>
<row>
<entry>
<para>p_filesz</para>
</entry>
<entry>
<para>0x00af0</para>
</entry>
<entry>
<para>0x00124</para>
</entry>
</row>
<row>
<entry>
<para>p_memsz</para>
</entry>
<entry>
<para>0x00af0</para>
</entry>
<entry>
<para>0x00128</para>
</entry>
</row>
<row>
<entry>
<para>p_flags</para>
</entry>
<entry>
<para>R-E</para>
</entry>
<entry>
<para>RW-</para>
</entry>
</row>
<row>
<entry>
<para>p_align</para>
</entry>
<entry>
<para>0x10000</para>
</entry>
<entry>
<para>0x10000</para>
</entry>
</row>
</tbody>
</tgroup>
</table>
<note>
<para>For the PT_LOAD entry describing the data segment, the
p_memsz may be greater than the p_filesz. The difference is the size of
the .bss section. On implementations that use virtual memory file
mapping, only the portion of the file between the .data p_offset
(rounded down to the nearest page) to p_offset + p_filesz (rounded up
to the next page size) is included. If the distance between p_offset +
p_filesz and p_offset + p_memsz crosses a page boundary, then
additional memory must be allocated out of anonymous memory to include
data through p_vaddr + p_memsz.</para>
</note>
<para>
<xref linkend="dbdoclet.50655242_45730" /> demonstrates a typical mapping of
file to memory segments.</para>
<table frame="all" pgwide="1" xml:id="dbdoclet.50655242_45730">
<title>Memory Segment Mappings</title>
<tgroup cols="3">
<colspec colname="c1" colwidth="33*" align="center" />
<colspec colname="c2" colwidth="33*" align="center" />
<colspec colname="c3" colwidth="33*" align="center" />
<thead>
<row>
<entry>
<para>
<emphasis role="bold">File</emphasis>
</para>
</entry>
<entry>
<para>
<emphasis role="bold">Section</emphasis>
</para>
</entry>
<entry>
<para>
<emphasis role="bold">Virtual Address</emphasis>
</para>
</entry>
</row>
</thead>
<tbody>
<row>
<entry>
<para>0x0</para>
</entry>
<entry>
<para>header</para>
</entry>
<entry>
<para>0x10000000</para>
</entry>
</row>
<row>
<entry>
<para>0x100</para>
</entry>
<entry>
<para>.text</para>
</entry>
<entry>
<para>0x10000100</para>
</entry>
</row>
<row>
<entry>
<para>0xaf0</para>
</entry>
<entry>
<para>.data</para>
</entry>
<entry>
<para>0x10010af0</para>
</entry>
</row>
<row>
<entry>
<para>Not applicable. Zero-initialized data is not stored in the
file.</para>
</entry>
<entry>
<para>.bss</para>
</entry>
<entry>
<para>0x10010c14</para>
</entry>
</row>
<row>
<entry>
<para>Not stored in the file.</para>
</entry>
<entry>
<para>End of sections</para>
</entry>
<entry>
<para>0x10010c18</para>
</entry>
</row>
</tbody>
</tgroup>
</table>
<para>Operating systems typically enforce memory permission on a per-page
granularity. This ABI maintains that the memory permissions are consistent
across each memory segment when a file image is mapped to a process memory
segment. The text segment and data segment require differing memory
permissions. To maintain congruency of file offset to virtual address
modulo the page size, the system maps the file region holding the
overlapped text and data twice at different virtual addresses for each
segment (see
<xref linkend="dbdoclet.50655242_42741" />).</para>
<para>To increase the security attributes of this ABI, the text and certain
sections of the data segment (such as the .rodata section) may be protected
as read only after the pages are mapped and relocations are resolved. See
<xref linkend="dbdoclet.50655242_93623" /> for more information.</para>
<figure pgwide="1" xml:id="dbdoclet.50655242_42741">
<title>File Image to Process Memory Image Mapping</title>
<mediaobject>
<imageobject>
<imagedata fileref="figures/fig4-1.png" format="PNG"
scalefit="1" width="100%" />
</imageobject>
</mediaobject>
</figure>
<para>As a result of this mapping, there can be up to four pages of impure
text or data in the virtual memory segments for the application as
described in the following list:</para>
<orderedlist>
<listitem>
<para>ELF header information, program headers, and other information will
precede the .text section and reside at the beginning of the text segment.</para>
</listitem>
<listitem>
<para>The last memory page of the text segment can contain a copy of
the partial, first file-image data page as an artifact of page faulting
the last file-image text page from the file image to the text segment
while maintaining the required offsets as shown in
<xref linkend="dbdoclet.50655242_42741" />.</para>
</listitem>
<listitem>
<para>Likewise, the first memory page of the data segment may
contain a copy of the partial, last file-image text page as an artifact
of page faulting the first file-image data page from the file image to
the data segment while maintaining the required offsets.</para>
</listitem>
<listitem>
<para>The last faulted data-segment memory page may contain residual
data from the last file-image data page that is not part of the actual
file image. The system is required to zero this residual memory after
that page is mapped to the data segment. If the application requires
static data, the remainder of this page is used for that purpose. If
the static data requirements exceed the remnant left in the last
faulted memory page, additional pages shall be mapped from anonymous
memory and zeroed.</para>
</listitem>
</orderedlist>
<note>
<para>The handling of the contents of the first three
pages is undefined by this ABI. They are unused by the
executable program once started.</para>
</note>
<section xml:id="dbdoclet.50655242_10560">
<title>Addressing Models</title>
<para>When mapping an executable file or shared object file to memory,
the system can use the following addressing models. Each application is
allocated its own virtual address space.</para>
<itemizedlist>
<listitem>
<para>Traditionally, executable files are mapped to virtual memory
using an absolute addressing model, where the mapping of the sections to
segments uses the section p_vaddr specified by the ELF header directly
as an absolute address.
</para>
</listitem>
<listitem>
<para>The position-independent code (PIC) addressing model allows the
file image text of an executable file or shared object file to be
loaded into the virtual address space of a process at an arbitrary
starting address chosen by the kernel loader or program interpreter
(dynamic linker).</para>
</listitem>
</itemizedlist>
<note>
<itemizedlist>
<listitem>
<para>Shared objects need to use the PIC addressing model
so that all references to global variables go through the
Global Offset Table.</para>
</listitem>
<listitem>
<para>Position-independent executables should use the PIC
addressing model.</para>
</listitem>
</itemizedlist>
</note>
</section>
<section xml:id="dbdoclet.50655242_82123">
<title>Process Initialization</title>
<para>To provide a standard environment for application programs, the
exec system call creates an initial program machine state. That state
includes the use of registers, the layout of the stack frame, and
argument passing. For example, a C program might typically issue the
following declaration to begin executing at the local entry point of a
function named main:</para>
<programlisting>extern int main (int argc, char *argv[ ], char *envp[ ], void *auxv[ ]);
int main(int argc, char *argv[ ], char *envp[ ], ElfW(auxv_t) *auxvec)</programlisting>
<para>where:</para>
<itemizedlist mark="none">
<listitem>
<para>argc is a nonnegative argument count.</para>
</listitem>
<listitem>
<para> argv is an array of argument strings.
It is terminated by a NULL pointer, argv[argc] == 0.</para>
</listitem>
<listitem>
<para>envp is an array of environment strings. It is also
terminated by a NULL pointer.</para>
</listitem>
<listitem>
<para>auxv is an array of structures that contain the auxiliary
vector. It is terminated by a structure entry with an a_type of
AT_NULL. For more information, see
<xref linkend="dbdoclet.50655242_98651" />.</para>
</listitem>
</itemizedlist>
<para>This section explains how to implement the call to main or to the
entry point.</para>
<section xml:id="dbdoclet.50655242___RefHeading___Toc377640653">
<title xml:id="dbdoclet.50655242_PROC-REG">Registers</title>
<para>The contents of most registers are
<emphasis>not</emphasis> specified when a process is first entered from an
exec system call. A program should not expect the operating system to set
all registers to 0. If a register other than those listed in
<xref linkend="dbdoclet.50655242_74550" /> must have a specific value, the
program must set it to that value during process initialization.</para>
<para>The contents of the following registers
<emphasis>are</emphasis> specified:</para>
<para> </para>
<table frame="all" pgwide="1" xml:id="dbdoclet.50655242_74550">
<title>Registers Specified during Process Initialization</title>
<tgroup cols="2">
<colspec colname="c1" colwidth="15*" align="center"/>
<colspec colname="c2" colwidth="85*" />
<thead>
<row>
<entry>
<para>
<emphasis role="bold">Register</emphasis>
</para>
</entry>
<entry align="center">
<para>
<emphasis role="bold">Description</emphasis>
</para>
</entry>
</row>
</thead>
<tbody>
<row>
<entry>
<para>r1</para>
</entry>
<entry>
<para>The initial stack pointer, aligned to a quadword
boundary.</para>
</entry>
</row>
<row>
<entry>
<para>r2</para>
</entry>
<entry>
<para>Undefined.</para>
</entry>
</row>
<row>
<entry>
<para>r3</para>
</entry>
<entry>
<para>Contains argc, the nonnegative argument count.</para>
</entry>
</row>
<row>
<entry>
<para>r4</para>
</entry>
<entry>
<para>Contains argv, a pointer to the array of argument
pointers in the stack. The array is immediately followed by a
NULL pointer. If there are no arguments, r4 points to a NULL
pointer.</para>
</entry>
</row>
<row>
<entry>
<para>r5</para>
</entry>
<entry>
<para>Contains envp, a pointer to the array of environment
pointers in the stack. The array is immediately followed by a
NULL pointer. If no environment exists, r5 points to a NULL
pointer.</para>
</entry>
</row>
<row>
<entry>
<para>r6</para>
</entry>
<entry>
<para>Contains a pointer to the auxiliary vector. The auxiliary
vector shall have at least one member, a terminating entry with
an a_type of AT_NULL (see
<xref linkend="dbdoclet.50655242_98651" />).</para>
</entry>
</row>
<row>
<entry>
<para>r7</para>
</entry>
<entry>
<para>Contains a termination function pointer. If r7 contains a
nonzero value, the value represents a function pointer that the
application should register with atexit. If r7 contains zero,
no action is required.</para>
</entry>
</row>
<row>
<entry>
<para>r12</para>
</entry>
<entry>
<para>Contains the address of the global entry point of the
first function being invoked, which represents the start
address of the executable specified in the exec call.</para>
</entry>
</row>
<row>
<entry>
<para>FPSCR</para>
</entry>
<entry>
<para>Contains 0, specifying “round to nearest” mode for both
binary and decimal rounding modes, IEEE Mode, and the disabling
of floating-point exceptions.</para>
</entry>
</row>
<row>
<entry>
<para>VSCR</para>
</entry>
<entry>
<para>Vector Status and Control Register. Contains 0,
specifying vector Java/IEEE mode and that no saturation has
occurred.</para>
</entry>
</row>
</tbody>
</tgroup>
</table>
<para>The run-time that gets control from _start is responsible for:</para>
<itemizedlist spacing="compact">
<listitem>
<para>Creating the first stack frame</para>
</listitem>
<listitem>
<para>Initializing the first stack frame's back chain pointer to
NULL</para>
</listitem>
<listitem>
<para>Allocating and initializing TLS storage</para>
</listitem>
<listitem>
<para>Initializing the thread control block (TCB) and dynamic thread
vector (DTV)</para>
</listitem>
<listitem>
<para>Initializing any __thread variables</para>
</listitem>
<listitem>
<para>Setting R13 for the initial process thread.</para>
</listitem>
</itemizedlist>
<para>This initialization must be completed before any library
initialization codes are run and before control is transferred to the
main program (main( )).</para>
</section>
<section xml:id="dbdoclet.50655242_83727">
<title>Process Stack</title>
<para>Although every process has a stack, no fixed stack address is
defined by the system. In addition, a program's stack address can change
from one system to another. It can even change from one process
invocation to another. Thus, the process initialization code must use the
stack address in general-purpose register r1. Data in the stack segment
at addresses below the stack pointer contain undefined values.</para>
</section>
<section xml:id="dbdoclet.50655242_98651">
<title>Auxiliary Vector</title>
<para>The argument and environment vectors transmit information from one
application program to another. However, the auxiliary vector conveys
information from the operating system to the program. This vector is an
array of structures, defined as follows:</para>
<programlisting>typedef struct
{
long a_type;
union
{
long a_val;
void *a_ptr;
void (*a_fcn)( );
} a_un;
} auxv_t;
Name Value a_un field Comment
AT_NULL 0 ignored /* End of vector */
AT_PHDR 3 a_ptr /* Program headers for program */
AT_PHENT 4 a_val /* Size of program header entry */
AT_PHNUM 5 a_val /* Number of program headers */
AT_PAGESZ 6 a_val /* System page size */
AT_BASE 7 a_ptr /* Base address of interpreter */
AT_FLAGS 8 a_val /* Flags */
AT_ENTRY 9 a_ptr /* Entry point of program */
AT_UID 11 /* Real user ID (uid) */
AT_EUID 12 /* Effective user ID (euid) */
AT_GID 13 /* Real group ID (gid) */
AT_EGID 14 /* Effective group ID (egid) */
AT_PLATFORM 15 a_ptr /* String identifying platform. */
AT_HWCAP 16 a_val /* Machine-dependent hints about
processor capabilities. */
AT_CLKTCK 17 /* Frequency of times( ), always 100 */
AT_DCACHEBSIZE 19 a_val /* Data cache block size */
AT_ICACHEBSIZE 20 a_val /* Instruction cache block size */
AT_UCACHEBSIZE 21 a_val /* Unified cache block size */
AT_IGNOREPPC 22 /* Ignore this entry! */
AT_SECURE 23 /* Boolean, was exec authorized to use
setuid or setgid */
AT_BASE_PLATFORM 24 a_ptr /* String identifying real platforms */
AT_RANDOM 25 /* Address of 16 random bytes */
AT_HWCAP2 26 a_val /* More machine-dependent hints about
processor capabilities. */
AT_EXECFN 31 /* File name of executable */
AT_SYSINFO_EHDR 33 /* In many architectures, the kernel
provides a virtual dynamic shared
object (VDSO) that contains a function
callable from the user state.
AT_SYSINFO_EHDR is the address of the
VDSO header that is used by the
dynamic linker to resolve function
symbols with the VDSO. */
AT_L1I_CACHESIZE 40 /* Cache sizes and geometries. */
AT_L1I_CACHEGEOMETRY 41
AT_L1D_CACHESIZE 42
AT_L1D_CACHEGEOMETRY 43
AT_L2_CACHESIZE 44
AT_L2_CACHEGEOMETRY 45
AT_L3_CACHESIZE 46
AT_L3_CACHEGEOMETRY 47
</programlisting>
<para>AT_NULL</para>
<para>The auxiliary vector has no fixed length; instead an entry of this
type denotes the end of the vector. The corresponding value of a_un is
undefined.</para>
<para>AT_PHDR</para>
<para>Under some conditions, the system creates the memory image of the
application program before passing control to an interpreter program.
When this happens, the a_ptr member of the AT_PHDR entry tells the
interpreter where to find the program header table in the memory image.
If the AT_PHDR entry is present, entries of types AT_PHENT, AT_PHNUM, and
AT_ENTRY must also be present. See the Program Header section in Chapter
5 of the
<citetitle>System V ABI</citetitle> for more information about the program
header table.</para>
<para>AT_PHENT</para>
<para>The a_val member of this entry holds the size, in bytes, of one
entry in the program header table to which the AT_PHDR entry
points.</para>
<para>AT_PHNUM</para>
<para>The a_val member of this entry holds the number of entries in the
program header table to which the AT_PHDR entry points.</para>
<para>AT_PAGESZ</para>
<para>If present, this entry's a_val member gives the system page size in
bytes. The same information is also available through the sysconf system
call.</para>
<para>AT_BASE</para>
<para>The a_ptr member of this entry holds the base address at which the
interpreter program was loaded into memory. See the Program Header
section in Chapter 5 of the
<citetitle>System V ABI</citetitle> for more information about the base
address.</para>
<para>AT_FLAGS</para>
<para>If present, the a_val member of this entry holds 1-bit flags. Bits
with undefined semantics are set to zero. Other auxiliary vector types
are reserved. No flags are currently defined for AT_FLAGS on the 64-bit
OpenPOWER ABI Architecture.</para>
<para>AT_ENTRY</para>
<para>The a_ptr member of this entry holds the entry point of the
application program to which the interpreter program should transfer
control.</para>
<para>AT_DCACHEBSIZE</para>
<para>The a_val member of this entry gives the data cache block size for
processors on the system on which this program is running. If the
processors have unified caches, AT_DCACHEBSIZE is the same as
AT_UCACHEBSIZE.</para>
<para>AT_ICACHEBSIZE</para>
<para>The a_val member of this entry gives the instruction cache block
size for processors on the system on which this program is running. If
the processors have unified caches, AT_ICACHEBSIZE is the same as
AT_UCACHEBSIZE.</para>
<para>AT_UCACHEBSIZE</para>
<para>The a_val member of this entry is zero if the processors on the
system on which this program is running do not have a unified instruction
and data cache. Otherwise, it gives the cache block size.</para>
<para>AT_PLATFORM</para>
<para>The a_ptr member is the address of the platform name string. For
virtualized systems, this may be different (that is, an older platform)
than the physical machine running this environment.</para>
<para>AT_BASE_PLATFORM</para>
<para>The a_ptr member is the address of the platform name string for the
physical machine. For virtualized systems, this will be the platform name
of the real hardware.</para>
<para>AT_HWCAP</para>
<para>The a_val member of this entry is a bit map of hardware
capabilities. Some bit mask values include:</para>
<programlisting>PPC_FEATURE_32 0x80000000 /* Always set for powerpc64 */
PPC_FEATURE_64 0x40000000 /* Always set for powerpc64 */
PPC_FEATURE_HAS_ALTIVEC 0x10000000
PPC_FEATURE_HAS_FPU 0x08000000
PPC_FEATURE_HAS_MMU 0x04000000
PPC_FEATURE_UNIFIED_CACHE 0x01000000
PPC_FEATURE_NO_TB 0x00100000 /* 601/403gx have no timebase */
PPC_FEATURE_POWER4 0x00080000 /* POWER4 ISA 2.00 */
PPC_FEATURE_POWER5 0x00040000 /* POWER5 ISA 2.02 */
PPC_FEATURE_POWER5_PLUS 0x00020000 /* POWER5+ ISA 2.03 */
PPC_FEATURE_CELL_BE 0x00010000 /* CELL Broadband Engine */
PPC_FEATURE_BOOKE 0x00008000 /* ISA Category Embedded */
PPC_FEATURE_SMT 0x00004000 /* Simultaneous Multi-Threading */
PPC_FEATURE_ICACHE_SNOOP 0x00002000
PPC_FEATURE_ARCH_2_05 0x00001000 /* ISA 2.05 */
PPC_FEATURE_PA6T 0x00000800 /* PA Semi 6T Core */
PPC_FEATURE_HAS_DFP 0x00000400 /* Decimal FP Unit */
PPC_FEATURE_POWER6_EXT 0x00000200 /* P6 + mffgpr/mftgpr */
PPC_FEATURE_ARCH_2_06 0x00000100 /* ISA 2.06 */
PPC_FEATURE_HAS_VSX 0x00000080 /* P7 Vector Extension. */
PPC_FEATURE_PSERIES_PERFMON_COMPAT 0x00000040
PPC_FEATURE_TRUE_LE 0x00000002
PPC_FEATURE_PPC_LE 0x00000001</programlisting>
<para revisionflag="added">Bit 0x00000004 is reserved for kernel use.
</para>
<para>AT_HWCAP2</para>
<para>The a_val member of this entry is a bit map of hardware
capabilities. Some bit mask values include:</para>
<programlisting revisionflag="changed">PPC_FEATURE2_ARCH_2_07 0x80000000 /* ISA 2.07 */
PPC_FEATURE2_HAS_HTM 0x40000000 /* Hardware Transactional Memory */
PPC_FEATURE2_HAS_DSCR 0x20000000 /* Data Stream Control Register */
PPC_FEATURE2_HAS_EBB 0x10000000 /* Event Base Branching */
PPC_FEATURE2_HAS_ISEL 0x08000000 /* Integer Select */
PPC_FEATURE2_HAS_TAR 0x04000000 /* Target Address Register */
PPC_FEATURE2_HAS_VCRYPTO 0x02000000 /* The processor implements the
Vector.AES category */
PPC_FEATURE2_HTM_NOSC 0x01000000
PPC_FEATURE2_ARCH_3_00 0x00800000 /* ISA 3.0 */
PPC_FEATURE2_HAS_IEEE128 0x00400000 /* VSX IEEE Binary Float 128-bit */
PPC_FEATURE2_DARN 0x00200000 /* darn instruction */
PPC_FEATURE2_SCV 0x00100000 /* scv syscall */
PPC_FEATURE2_HTM_NO_SUSPEND 0x00080000 /* TM without suspended state */
PPC_FEATURE2_ARCH_3_1 0x00040000 /* ISA 3.1 */
PPC_FEATURE2_MMA 0x00020000 /* Matrix Multiply Accumulate */</programlisting>
<para>When a process starts to execute, its stack holds the arguments,
environment, and auxiliary vector received from the exec call. The system
makes no guarantees about the relative arrangement of argument strings,
environment strings, and the auxiliary information, which appear in no
defined or predictable order. Further, the system may allocate memory
after the null auxiliary vector entry and before the beginning of the
information block.</para>
<para>AT_L1I_CACHESIZE</para>
<para>The size of the level-1 instruction cache, in bytes.</para>
<para>AT_L1I_CACHEGEOMETRY</para>
<para>The geometry of the level-1 instruction cache. The low-order
sixteen bits contain the cache associativity as a value N, where
N = 1 represents a direct-mapped cache, N = 0xffff represents a
fully associative cache, and any other N represents an N-way
set-associative cache. The next higher-order sixteen bits contain
the size of the cache line in bytes. Note that the cache line
size is not necessarily the same as the cache block size.</para>
<para>AT_L1D_CACHESIZE</para>
<para>The size of the level-1 data cache, in bytes.</para>
<para>AT_L1D_CACHEGEOMETRY</para>
<para>The geometry of the level-1 data cache, defined in the same
manner as for AT_L1I_CACHEGEOMETRY.</para>
<para>AT_L2_CACHESIZE</para>
<para>The size of the level-2 cache, in bytes.</para>
<para>AT_L2_CACHEGEOMETRY</para>
<para>The geometry of the level-2 cache, defined in the same
manner as for AT_L1I_CACHEGEOMETRY.</para>
<para>AT_L3_CACHESIZE</para>
<para>The size of the level-3 cache, in bytes.</para>
<para>AT_L3_CACHEGEOMETRY</para>
<para>The geometry of the level-3 cache, defined in the same
manner as for AT_L1I_CACHEGEOMETRY.</para>
</section>
</section>
</section>
<section>
<title>Dynamic Linking</title>
<section xml:id="dbdoclet.50655242___RefHeading___Toc377640656">
<title>Program Interpreter</title>
<para>For dynamic linking, the standard program interpreter is
/lib/ld64.so.2. It may be located in different places on different
distributions.</para>
</section>
<section>
<title>Dynamic Section</title>
<para><anchor xml:id="dbdoclet.50655242_page119" xreflabel="" />The dynamic
section provides information used by the dynamic linker to manage
dynamically loaded shared objects, including relocation, initialization,
and termination when loaded or unloaded, resolving dependencies on other
shared objects, resolving references to symbols in the shared object, and
supporting debugging. The following dynamic tags are relevant to this
processor-specific ABI:</para>
<para>DT_PLTGOT</para>
<para>The d_ptr member of this dynamic tag points to the first byte of
the PLT.</para>
<para>DT_JMPREL</para>
<para>The d_ptr member of this dynamic tag points to the first byte of
the table of relocation entries, which have a one-to-one correspondence
with PLT entries. Any executable or shared object with a PLT must have
DT_JMPREL. A shared object containing only data will not have a PLT and
thus will not have DT_JMPREL.</para>
<para>DT_PPC64_GLINK (DT_LOPROC + 0)</para>
<para>The d_ptr member of this dynamic tag points to 32 bytes before the
.glink lazy link symbol resolver stubs that are described in
<xref linkend="dbdoclet.50655242_82622" />.</para>
<para>DT_PPC64_OPT (DT_LOPROC + 3)</para>
<para>The d_val member of this dynamic tag specifies whether various
optimizations are possible. The low bit will be set to indicate that an
optimized __tls_get_addr call stub is used. The next most-significant bit
will be set if multiple TOCs are present.</para>
</section>
<section xml:id="dbdoclet.50655242_47739">
<title>Global Offset Table</title>
<para>To support position-independent code, a Global Offset Table (GOT)
shall be constructed by the link editor in the data segment when linking
code that contains any of the various R_PPC64_GOT* relocations or when
linking code that references the .TOC. address. The GOT consists of an
8-byte header that contains the TOC base (the first TOC base when
multiple TOCs are present), followed by an array of 8-byte addresses.
The link editor shall emit dynamic relocations as appropriate for each
entry in the GOT. At runtime, the dynamic linker will apply these
relocations after the addresses of all memory segments are known (and
thus the addresses of all symbols). While the GOT may be appear to be an
array of absolute addresses, this ABI does not preclude the GOT
containing nonaddress entries and specifies the presence of nonaddress
tls_index entries.</para>
<para>Absolute addresses are generated for all GOT relocations by the
dynamic linker before giving control to general application code.
(However, IFUNC resolution functions may be invoked before relocation is
completed, limiting the use of global variables by such functions.) The
dynamic linker is free to choose different memory segment addresses for
the executable or shared objects in a different process image. After the
initial mapping of the process image by the dynamic linker, memory
segments reside at fixed addresses for the life of a process.</para>
<para>The symbol .TOC. may be used to access the GOT or in TOC-relative
addressing to other data constructs, such as the procedure linkage table.
The symbol may be offset by 0x8000 bytes, or another offset, from the
start of the .got section. This offset allows the use of the full (64 KB)
signed range of 16-bit displacement fields by using both positive and
negative subscripts into the array of addresses, or a larger offset to
afford addressing using references within ±2 GB with 32-bit
displacements. The 32-bit displacements are constructed by using the
addis instruction to provide a first high-order 16-bit portion of a
32-bit displacement in conjunction with an instruction to supply a
low-order 16-bit portion of a 32-bit displacement.</para>
<para>In PIC code<phrase revisionflag="added"> that uses the
TOC</phrase>, the TOC pointer r2 points to the TOC base, enabling
easy reference. For static nonrelocatable modules, the GOT address is
fixed and can be directly used by code.</para>
<para revisionflag="deleted">All functions except leaf routines must
load the value of the TOC base into the TOC register r2.</para>
<para revisionflag="added">
Code may access GOT entries directly using PC-relative addressing,
where available.
</para>
</section>
<section>
<title>Function Addresses</title>
<para>The following requirements concern function addresses.</para>
<para>When referencing a function address, consider the following
requirements:</para>
<itemizedlist>
<listitem>
<para>Intraobject executable or shared object function address
references may be resolved by the dynamic linker to the absolute
virtual address of the symbol.</para>
</listitem>
<listitem>
<para>Function address references from within the executable file
to a function defined in a shared object file are resolved by the
link editor to the .text section address of the PLT call stub for
that functionwithin the executable file.
</para>
</listitem>
<listitem>
<para>In a static module, when a function pointer reference is made
to a function provided by a dynamically loaded shared module, the
function may be resolved to the address of a PLT stub. If this
resolution is made, all function pointer references must be made
through the same PLT stub in the static module to ensure correct
intraobject comparisons for function addresses.</para>
</listitem>
<listitem>
<para>A function address of a nested function
<emphasis>may</emphasis> also be resolved to the address of a
trampoline used to call it.</para>
</listitem>
</itemizedlist>
<para>When comparing function addresses, consider the following
requirements:</para>
<itemizedlist>
<listitem>
<para>The address of a function shall compare to the same value in
executables and shared objects.</para>
</listitem>
<listitem>
<para>For intraobject comparisons of function addresses within the
executable or shared object, the link editor may directly compare the
absolute virtual addresses.</para>
</listitem>
<listitem>
<para>For a function address comparison where an executable
references a function defined in a a shared object, the link
editor will place the address of a .text section PLT call stub
for that function in the corresponding dynamic symbol table
entry's st_value field (see <xref linkend="dbdoclet.50655241_95185" />).</para>
</listitem>
<listitem>
<para>When the dynamic linker loads shared objects associated with an
executable and resolves any GOT entry relocations into absolute
addresses, it will search the dynamic symbol table of the executable
for each symbol that needs to be resolved.</para>
</listitem>
<listitem>
<para>If it finds the symbol and the st_value of the symbol table
entry is nonzero, it shall use the address indicated in the st_value
entry as the symbols address. If the dynamic linker does not find
the symbol in the executables dynamic symbol table or the entrys
st_value member is zero, the dynamic linker may consider the symbol
as undefined in the executable file.</para>
</listitem>
</itemizedlist>
</section>
<section xml:id="dbdoclet.50655242_20388">
<title>Procedure Linkage Table</title>
<para>When the link editor builds an executable file or shared object
file, it does not know the absolute address of undefined function calls.
Therefore, it cannot generate code to directly transfer execution to
another shared object or executable. For each execution transfer to an
undefined function call in the file image, the link editor places a
relocation against an entry in the Procedure Linkage Table (PLT) of the
executable or shared object that corresponds to that function
call.</para>
<para>Additionally, for all nonstatic functions with standard (nonhidden)
visibility in a shared object, the link editor invokes the function
through the PLT, even if the shared object defines the function. The same
is not true for executables.</para>
<para>The link editor knows the number of functions invoked through the
PLT, and it reserves space for an appropriately sized .plt section. The
.plt section is located in the section following the .got. It consists of
an array of addresses and is initialized by the module loader. There will
also be an array of R_PPC_JMP_SLOT relocations in .rela.plt, with a
one-to-one correspondence between elements of each array. Each
R_PPC_JMP_SLOT relocation will have r_offset pointing at the .plt word it
relocates.</para>
<para>A unique PLT is constructed by the static linker for each static
module (that is, the main executable) and each dynamic shared object. The
PLT is located in the data segment of the process image at object load
time by the dynamic linker using the information about the .plt section
stored in the file image. The individual PLT entries are populated by the
dynamic linker using one of the following binding methods. Execution can
then be redirected to a dependent shared object or executable.</para>
<section>
<title>Lazy Binding</title>
<para>The lazy binding method is the default. It delays the resolution of
a PLT entry to an absolute address until the function call is made the
first time. The benefit of this method is that the application does not
pay the resolution cost until the first time it needs to call the
function, if at all.</para>
<para>To implement lazy binding, the dynamic loader points each PLT entry
to a lazy resolution stub at load time. After the function call is made
the first time, this lazy resolution stub gets control, resolves the
symbol, and updates the PLT entry to hold the final value to be used for
future calls.</para>
</section>
<section xml:id="dbdoclet.50655242_93623">
<title>Immediate Binding</title>
<para>The immediate binding method resolves the absolute addresses of all
PLT entries in the executable and dependent shared objects at load time,
before passing execution control to the application. The environment
variable LD_BIND_NOW may be set to a nonnull value to signal the dynamic
linker that immediate binding is requested at load time, before control
is given to the application.</para>
<para>For some performance-sensitive situations, it may be better to pay
the resolution cost to populate the PLT entries up front rather than
during execution.</para>
</section>
<section xml:id="dbdoclet.50655242_82622">
<title>Procedure Linkage Table</title>
<para>For every call site that needs to use the PLT, the link editor
constructs a call stub in the .text section and resolves the call site to
use that call stub. The call stub transfers control to the address
indicated in the PLT entry. These call stubs need not be adjacent to one
another or unique. They can be scattered throughout the text segment so
that they can be reached with a branch and link instruction.</para>
<para>Depending on relocation information at the call site, the stub
provides one of the following properties:</para>
<orderedlist>
<listitem>
<para>The caller has set up r2 to hold the TOC pointer and expects
the PLT call stub to save that value to the TOC save stack slot. This
is the default.</para>
</listitem>
<listitem>
<para>The caller has set up r2 to hold the TOC pointer and has
already saved that value to the TOC save stack slot itself. This is
indicated by the presence of a R_PPC64_TOCSAVE relocation on the nop
following the call.</para>
</listitem>
</orderedlist>
<programlisting>tocsaveloc:
nop
...
bl target
.reloc ., R_PPC64_TOCSAVE, tocsaveloc
nop</programlisting>
<orderedlist continuation="continues">
<listitem>
<para>The caller has not set up r2 to hold the TOC pointer. This
is indicated by use of a R_PPC64_REL24_NOTOC relocation (instead of
R_PPC64_REL24) on the call instruction.</para>
</listitem>
</orderedlist>
<para>In any scenario, the PLT call stub must transfer control to the
function whose address is provided in the associated PLT entry. This
address is treated as a global entry point for ABI purposes. This means
that the PLT call stub loads the address into r12 before transferring
control.</para>
<para>Although the details of the call stub implementation are left to
the link editor, some examples are provided. In those examples, func@plt
is used to denote the address of the PLT entry for func; func@plt@toc
denotes the offset of that address relative to the TOC pointer; and the
@ha and @l variants denote the high-adjusted and low parts of these
values as usual. Because the link editor synthesizes the PLT call stubs
directly, it can determine all these values as immediate constants. The
assembler is not required to support those notations.</para>
<para>A possible implementation for case 1 looks as follows (if
func@plt@toc is less than 32 KB, the call stub may be simplified to omit
the addis):</para>
<programlisting>std r2,24(r1)
addis r12,r2,func@plt@toc@ha
ld r12,func@plt@toc@l(r12)
mtctr r12
bctr</programlisting>
<para>For case 2, the same implementation as for case 1 may be used,
except that the first instruction “std r2,24(r1)” is omitted:</para>
<programlisting>addis r12,r2,func@plt@toc@ha
ld r12,func@plt@toc@l(r12)
mtctr r12
bctr</programlisting>
<para><anchor xml:id="dbdoclet.50655242___DdeLink__61883_1749258592"
xreflabel="" />A possible implementation for case 3 looks as
follows:</para>
<programlisting> mflr r0
bcl 20,31,1f
1: mflr r2
mtlr r0
addis r2,r2,(.TOC.-1b)@ha
addi r2,r2,(.TOC.-1b)@l
addis r12,r2,func@plt@toc@ha
ld r12,func@plt@toc@l(r12)
mtctr r12
bctr</programlisting>
<para>When generating non-PIC code for the small or medium code model, a
simpler variant may alternatively be used for cases 2 or 3:</para>
<programlisting>lis r12,func@plt@ha
ld r12,func@plt@l(r12)
mtctr r12
bctr</programlisting>
<para revisionflag="added">
When PC-relative addressing is available, another simpler variant
may alternatively be used for cases 2 or 3:
</para>
<programlisting revisionflag="added">pld r12, func@plt@pcrel
mtctr r12
bctr</programlisting>
<para>To support lazy binding, the link editor also provides a set of
symbol resolver stubs, one for each PLT entry. Each resolver stub
consists of a single instruction, which is usually a branch to a common
resolver entry point or a nop. The resolver stubs are placed in the
.glink section, which is merged into the .text section of the final
executable or dynamic object. The address of the resolver stubs is
communicated to the dynamic loader through the DT_PPC64_GLINK dynamic
section entry. The address of the symbol resolver stub associated with
PLT entry N is determined by adding 4xN + 32 to the d_ptr field of the
DT_PPC64_GLINK entry. When using lazy binding, the dynamic linker
initializes each PLT entry at load time to that address.</para>
<para>The resolver stubs provided by the link editor must call into the
main resolver routine provided by the dynamic linker. This resolver
routine must be called with r0 set to the index of the PLT entry to be
resolved, r11 set to the identifier of the current dynamic object, and
r12 set to the resolver entry point address (as usual when calling a
global entry point). The resolver entry point address and the dynamic
object identifier are installed at load time by the dynamic linker into
the two doublewords immediately preceding the array of PLT entries,
allowing the resolver stubs to retrieve these values from there. These
two doublewords are considered part of the .plt section; the DT_PLTGOT
dynamic section entry points to the first of those words.</para>
<para>Beyond the above requirements, the implementation of the .glink
resolver stubs is up to the link editor. The following shows an example
implementation:</para>
<programlisting> # ABI note: At entry to the resolver stub:
# - r12 holds the address of the res_N stub for the target routine
# - all argument registers hold arguments for the target routine
PLTresolve:
# Determine addressability. This sequence works for both PIC
# and non-PIC code and does not rely on presence of the TOC pointer.
mflr r0
bcl 20,31,1f
1: mflr r11
mtlr r0
# Compute .plt section index from entry point address in r12
# .plt section index is placed into r0 as argument to the resolver
sub <phrase revisionflag="changed">r12</phrase>,r12,r11
subi <phrase revisionflag="changed">r12,r12</phrase>,res_0-1b
srdi r0,<phrase revisionflag="changed">r12</phrase>,2
# Load address of the first byte of the PLT
ld r12,PLToffset-1b(r11)
add r11,r12,r11
# Load resolver address and DSO identifier from the
# first two doublewords of the PLT
ld r12,0(r11)
ld r11,8(r11)
# Branch to resolver
mtctr r12
bctr
# ABI note: At entry to the resolver:
# - r12 holds the resolver address
# - r11 holds the DSO identifier
# - r0 holds the PLT index of the target routine
# - all argument registers hold arguments for the target routine
# Constant pool holding offset to the PLT
# Note that there is no actual symbol PLT; the link editor
# synthesizes this value when creating the .glink section
PLToffset:
.quad PLT-.
# A table of branches, one for each PLT entry
# The idea is that the PLT call stub loads r12 with these
# addresses, so (r12 - res_0) gives the PLT index × 4.
res_0: b PLTresolve
res_1: b PLTresolve
...</programlisting>
<para>After resolution, the value of a PLT entry in the PLT is the
address of the functions global entry point, unless the resolver
can determine that a module-local call occurs with a shared TOC value
wherein the TOC is shared between the caller and the callee.</para>
</section>
</section>
</section>
</chapter>