libhugetlbfs HOWTO
==================

Author: David Gibson <dwg@au1.ibm.com>, Adam Litke <agl@us.ibm.com>, and others
Last updated: Tuesday, May 13th, 2008

Introduction
============

In Linux(TM), access to hugepages is provided through a virtual file
system, "hugetlbfs".  The libhugetlbfs library interface works with
hugetlbfs to provide more convenient specific application-level
services.  In particular libhugetlbfs has three main functions:

	* library functions
libhugetlbfs provides functions that allow an applications to
explicitly allocate and use hugepages more easily they could by
directly accessing the hugetblfs filesystem

	* hugepage malloc()
libhugetlbfs can be used to make an existing application use hugepages
for all its malloc() calls.  This works on an existing (dynamically
linked) application binary without modification.

	* hugepage text/data/BSS
libhugetlbfs, in conjunction with included special linker scripts can
be used to make an application which will store its executable text,
its initialized data or BSS, or all of the above in hugepages.  This
requires relinking an application, but does not require source-level
modifications.

This HOWTO explains how to use the libhugetlbfs library.  It is for
application developers or system administrators who wish to use any of
the above functions.

The libhugetlbfs library is a focal point to simplify and standardise
the use of the kernel API.

Prerequisites
=============

Hardware prerequisites
----------------------

You will need a CPU with some sort of hugepage support, which is
handled by your kernel.  The covers recent x86, AMD64 and 64-bit
PowerPC(R) (POWER4, PPC970 and later) CPUs.

Currently, only x86, AMD64 and PowerPC are fully supported by
libhugetlbfs. IA64 and Sparc64 have a working malloc, and SH64
should also but it has not been tested. IA64, Sparc64, and SH64
do not support segment remapping at this time.

Kernel prerequisites
--------------------

To use all the features of libhugetlbfs you will need a 2.6.16 or
later kernel.  Many things will work with earlier kernels, but they
have important bugs and missing features.  The later sections of the
HOWTO assume a 2.6.16 or later kernel.  The kernel must also have
hugepages enabled, that is to say the CONFIG_HUGETLB_PAGE and
CONFIG_HUGETLBFS options must be switched on.

To check if hugetlbfs is enabled, use one of the following methods:

  * (Preferred) Use "grep hugetlbfs /proc/filesystems" to see if
    hugetlbfs is a supported file system.
  * On kernels which support /proc/config.gz (for example SLES10
    kernels), you can search for the CONFIG_HUGETLB_PAGE and
    CONFIG_HUGETLBFS options in /proc/config.gz
  * Finally, attempt to mount hugetlbfs. If it works, the required
    hugepage support is enabled.

Any kernel which meets the above test (even old ones) should support
at least basic libhugetlbfs functions, although old kernels may have
serious bugs.

The MAP_PRIVATE flag instructs the kernel to return a memory area that
is private to the requesting process.  To use MAP_PRIVATE mappings,
libhugetlbfs's automatic malloc() (morecore) feature, or the hugepage
text, data, or BSS features, you will need a kernel with hugepage
Copy-on-Write (CoW) support.  The 2.6.16 kernel has this.

PowerPC note: The malloc()/morecore features will generate warnings if
used on PowerPC chips with a kernel where hugepage mappings don't
respect the mmap() hint address (the "hint address" is the first
parameter to mmap(), when MAP_FIXED is not specified; the kernel is
not required to mmap() at this address, but should do so when
possible).  2.6.16 and later kernels do honor the hint address.
Hugepage malloc()/morecore should still work without this patch, but
the size of the hugepage heap will be limited (to around 256M for
32-bit and 1TB for 64-bit).

Toolchain prerequisites
-----------------------

The library uses a number of GNU specific features, so you will need
to use both gcc and GNU binutils.  For PowerPC and AMD64 systems you
will need a "biarch" compiler, which can build both 32-bit and 64-bit
binaries.

Configuration prerequisites
---------------------------

In kernels before 2.6.24, hugepages must be allocated at boot-time via
the hugepages= command-line parameter or at run-time via the
/proc/sys/vm/nr_hugepages sysctl. If memory is restricted on the system,
boot-time allocation is recommended. Hugepages so allocated will be in
the static hugepage pool.

In kernels starting with 2.6.24, the hugepage pool can grown on-demand.
If this feature should be used, /proc/sys/vm/nr_overcommit_hugepages
should be set to the maximum size of the hugepage pool. No hugepages
need to be allocated via /proc/sys/vm/nr_hugepages or hugepages= in this
case. Hugepages so allocated will be in the dynamic hugepage pool.

For the running of the libhugetlbfs testsuite (see below), allocating 20
static hugepages is recommended. Due to memory restrictions, the number
of hugepages requested may not be allocated if the allocation is
attempted at run-time. Users should verify the actual number of
hugepages allocated by either

       cat /proc/sys/vm/nr_hugepages

or

       grep HugePages_Free /proc/meminfo

With 20 hugepages allocated, most tests should succeed. However, with
smaller hugepages sizes, many more hugepages may be necessary.

To use libhugetlbfs features, as well as to run the testsuite, hugetlbfs
must be mounted:

       mkdir -p /mnt/hugetlbfs
       mount -t hugetlbfs none /mnt/hugetlbfs

If hugepages should be available to non-root users, the permissions on
the mountpoint need to be set appropriately.

Installation
============

1. Type "make" to build the library

This will create "obj32" and/or "obj64" under the top level
libhugetlbfs directory, and build, respectively, 32-bit and 64-bit
shared and static versions (as applicable) of the library into each
directory.  This will also build (but not run) the testsuite.

On i386 systems, only the 32-bit library will be built.  On PowerPC
and AMD64 systems, both 32-bit and 64-bit versions will be built (the
32-bit AMD64 version is identical to the i386 version).

2. Run the testsuite with "make check"

Running the testsuite is a good idea to ensure that the library is
working properly, and is quite quick (under 3 minutes on a 2GHz Apple
G5).  "make func" will run the just the functionality tests, rather
than stress tests (a subset of "make check") which is much quicker.
The testsuite contains tests both for the library's features and for
the underlying kernel hugepage functionality.

NOTE: The testsuite must be run as the root user.

WARNING: The testsuite contains testcases explicitly designed to test
for a number of hugepage related kernel bugs uncovered during the
library's development.  Some of these testcases WILL CRASH HARD a
kernel without the relevant fixes.  2.6.16 contains all such fixes for
all testcases included as of this writing.

3. (Optional) Install to system paths with "make install"

This will install the library images to the system lib/lib32/lib64
as appropriate.  By default it will install under /usr/local.  To put
it somewhere else use PREFIX=/path/to/install on the make command
line.  For example:
	make install PREFIX=/opt/hugetlbfs
Will install under /opt/hugetlbfs.

"make install" will also install the linker scripts and wrapper for ld
used for hugepage test/data/BSS (see below for details).

Alternatively, you can use the library from the directory in which it
was built, using the LD_LIBRARY_PATH environment variable.

Usage
=====

Using hugepages for malloc() (morecore)
---------------------------------------

This feature allows an existing (dynamically linked) binary executable
to use hugepages for all its malloc() calls.  To run a program using
the automatic hugepage malloc() feature, you must set several
environment variables:

1. Set LD_PRELOAD=libhugetlbfs.so
  This tells the dynamic linker to load the libhugetlbfs shared
  library, even though the program wasn't originally linked against it.

  Note: If the program is linked against libhugetlbfs, preloading the
        library may lead to application crashes. You should skip this
        step in that case.

2. Set LD_LIBRARY_PATH to the directory containing libhugetlbfs.so
  This is only necessary if you haven't installed libhugetlbfs.so to a
  system default path.  If you set LD_LIBRARY_PATH, make sure the
  directory referenced contains the right version of the library
  (32-bit or 64-bit) as appropriate to the binary you want to run.

3. Set HUGETLB_MORECORE=yes
  This enables the hugepage malloc() feature, instructing libhugetlbfs
  to override libc's normal morecore() function with a hugepage
  version and use it for malloc().  From this point all malloc()s
  should come from hugepage memory until it runs out.

Usually it's preferable to set these environment variables on the
command line of the program you wish to run, rather than using
"export", because you'll only want to enable the hugepage malloc() for
particular programs, not everything.

Examples:

If you've installed libhugetlbfs in the default place (under
/usr/local) which is in the system library search path use:
  $ LD_PRELOAD=libhugetlbfs.so HUGETLB_MORECORE=yes <your app command line>

If you have built libhugetlbfs in ~/libhugetlbfs and haven't installed
it yet, the following would work for a 64-bit program:

  $ LD_PRELOAD=libhugetlbfs.so LD_LIBRARY_PATH=~/libhugetlbfs/obj64 \
	HUGETLB_MORECORE=yes <your app command line>

Under some circumstances, you might want to specify the address where
the hugepage heap is located.  You can do this by setting the
HUGETLB_MORECORE_HEAPBASE environment variable to the heap address in
hexadecimal.  (NOTE: this will not work on PowerPC systems with old kernels
which don't respect the hugepage hint address; see Kernel Prerequisites
above).

By default, the hugepage heap begins at roughly the same place a
normal page heap would, rounded up by an amount determined by your
platform.  For 32-bit PowerPC binaries the normal page heap address is
rounded-up to a multiple of 256MB (that is, putting it in the next MMU
segment); for 64-bit PowerPC binaries the address is rounded-up to a
multiple of 1TB.  On all other platforms the address is rounded-up to
the size of a hugepage.

By default, the hugepage heap will be prefaulted by libhugetlbfs to
guarantee enough hugepages exist and are reserved for the application
(if this was not done, applications could receive a SIGKILL signal if
hugepages needed for the heap are used by another application before
they are faulted in). This leads to local-node allocations when no
memory policy is in place for hugepages. Therefore, it is recommended to
use

  $ numactl --interleave=all <your app command line>

to regain some of the performance impact of local-node allocations on
large NUMA systems. This can still result in poor performance for those
applications which carefully place their threads on particular nodes
(such as by using OpenMP). In that case, thread-local allocation is
preferred so users should select a memory policy that corresponds to
the run-time behavior of the process' CPU usage. Users can specify
HUGETLB_NO_PREFAULT to prevent the prefaulting of hugepages and instead
rely on run-time faulting of hugepages.  NOTE: specifying
HUGETLB_NO_PREFAULT on a system where hugepages are available to and
used by many process can result in some applications receving SIGKILL,
so its use is not recommended in high-availability or production
environments.

By default, the hugepage heap does not shrink.  To enable hugepage heap
shrinking, set HUGETLB_MORECORE_SHRINK=yes.  NB: We have been seeing some
unexpected behavior from glibc's malloc when this is enabled.

Using hugepage text, data, or BSS
---------------------------------

To use the hugepage text, data, or BSS segments feature, you need to
specially link your application.

	Linking the application:
	------------------------

To link an application for hugepages, you should use the the
ld.hugetlbfs script included with libhugetlbfs in place of your normal
linker.  Without any special options this will simply invoke GNU ld
with the same parameters.  

To link a program for hugepages, one of the following options should
be specified:

	--hugetlbfs-link=B

will link the application to store BSS data (only) into hugepages

	--hugetlbfs-link=BDT

will link the application to store text, initialized data and BSS data
into hugepages.

These are the only supported linking options.

The ld.hugetlbfs script will invoke the system linker with all the
necessary options to link for hugepages, in particular selecting the
right linker script.

If you installed ld.hugetlbfs using "make install", or if you run it
from the place where you built libhugetlbfs, it should automatically
be able to find the libhugetlbfs linker scripts.  Otherwise you may
need to explicitly instruct it where to find the scripts with the
option:
	--hugetlbfs-script-path=/path/to/scripts
(The linker scripts are in the ldscripts/ subdirectory of the
libhugetlbfs source tree).

	Linking via gcc:
	----------------

In many cases it's normal to link an application by invoking gcc,
which will then invoke the linker with appropriate options, rather
than invoking ld directly.  In such cases it's usually best to
convince gcc to invoke the ld.hugetlbfs script instead of the system
linker, rather than modifying your build procedure to invoke the
ld.hugetlbfs directly; the compilers may often add special libraries
or other linker options which can be fiddly to reproduce by hand.
To make this easier, 'make install' will install ld.hugetlbfs into 
$PREFIX/share/libhugetlbfs and create an 'ld' symlink to it.

Then with gcc, you invoke it as a linker with two options:

	-B $PREFIX/share/libhugetlbfs

This option tells gcc to look in a non-standard location for the
linker, thus finding our script rather than the normal linker. This
can optionally be set in the CFLAGS environment variable.

	-Wl,--hugetlbfs-link=B
OR	-Wl,--hugetlbfs-link=BDT

This option instructs gcc to pass the --hugetblfs-link option down to
the linker, thus invoking the special behaviour of the ld.hugetblfs
script. This can optionally be set in the LDFLAGS environment variable.

If you use a compiler other than gcc, you will need to consult its
documentation to see how to convince it to invoke ld.hugetlbfs in
place of the system linker.

	Running the application:
	------------------------

The specially-linked application needs the libhugetlbfs library, so
you might need to set the LD_LIBRARY_PATH environment variable so the
application can locate libhugetlbfs.so.  Other than that, after you
link the application with the correct script, it should only be
necessary to run it normally to activate the text, data, or BSS
hugepage feature.  Upon initialization, libhugetlbfs will detect the
special flags placed in the application's ELF header by the linker,
and remap the requested program segments into hugepages.

	Environment variables:
	----------------------

There are a number of private environment variables which can affect
libhugetlbfs:
	HUGETLB_ELFMAP
		If equal to "no", segment remapping is disabled;
		otherwise, it is enabled (default)
	HUGETLB_MINIMAL_COPY
		If equal to "no", the entire segment will be copied;
		otherwise, only the necessary parts will be, which can
		be much more efficient (default)

	HUGETLB_FORCE_ELFMAP
		Explained in "Partial segment remapping"

	HUGETLB_MORECORE
	HUGETLB_MORECORE_HEAPBASE
	HUGETLB_NO_PREFAULT
		Explained in "Using hugepages for malloc()
		(morecore)"

	HUGETLB_VERBOSE
		Specify the verbosity level of debugging output from 1
		to 99 (default is 1)
	HUGETLB_PATH
		Specify the path to the hugetlbfs mount point
	HUGETLB_SHARE
		Explained in "Sharing remapped segments"
	HUGETLB_DEBUG
		Set to 1 if an application segfaults. Gives very detailed output
		and runs extra diagnostics.

	Sharing remapped segments:
	--------------------------

By default, when libhugetlbfs uses anonymous, unlinked hugetlbfs files
to store remapped program segment data.  This means that if the same
program is started multiple times using hugepage segments, multiple
huge pages will be used to store the same program data.

The reduce this wastage, libugetlbfs can be instructed to allow
sharing segments between multiple invocations of a program.  To do
this, you must set the HUGETLB_SHARE variable must be set for all the
processes in question.  This variable has two possible values:
	anything but 1: the default, indicates no segments should be shared
	1: indicates that read-only segments (i.e. the program text,
in most cases) should be shared, read-write segments (data and bss)
will not be shared.

If the HUGETLB_MINIMAL_COPY variable is set for any program using
shared segments, it must be set to the same value for all invocations
of that program.

Segment sharing is implemented by creating persistent files in a
hugetlbfs containing the necessary segment data.  By default, these
files are stored in a subdirectory of the first located hugetlbfs
filesystem, named 'elflink-uid-XXX' where XXX is the uid of the
process using sharing.  This directory must be owned by the uid in
question, and have mode 0700.  If it doesn't exist, libhugetlbfs will
create it automatically.  This means that (by default) separate
invocations of the same program by different users will not share huge
pages.

The location for storing the hugetlbfs page files can be changed by
setting the HUGETLB_SHARE_PATH environment variable.  If set, this
variable must contain the path of an accessible, already created
directory located in a hugetlbfs filesystem.  The owner and mode of
this directory are not checked, so this method can be used to allow
processes of multiple uids to share huge pages.  IMPORTANT SECURITY
NOTE: any process sharing hugepages can insert arbitrary executable
code into any other process sharing hugepages in the same directory.
Therefore, when using HUGETLB_SHARE_PATH, the directory created *must*
allow access only to a set of uids who are mutually trusted.

The files created in hugetlbfs for sharing are persistent, and must be
manually deleted to free the hugepages in question.  Future versions
of libhugetlbfs should include tools and scripts to automate this
cleanup.

	Partial segment remapping
	-------------------------

libhugetlbfs has limited support for remapping a normal, non-relinked
binary's data, text and BSS into hugepages. To enable this feature,
HUGETLB_FORCE_ELFMAP must be set to "yes".

Partial segment remapping is not guaranteed to work. Most importantly, a
binary's segments must be large enough even when not relinked by
libhugetlbfs:

	architecture	address		minimum segment size
	------------	-------		--------------------
	i386, x86_64	all		hugepage size
	ppc32		all		256M
	ppc64		0-4G		256M
	ppc64		4G-1T		1020G
	ppc64		1T+		1T

The raw size, though, is not sufficient to indicate if the code will
succeed, due to alignment. Since the binary is not relinked, however,
this is relatively straightforward to 'test and see'.

NOTE: You must use LD_PRELOAD to load libhugetlbfs.so when using
partial remapping.


Examples
========

Example 1:  Application Developer
---------------------------------

To have a program use hugepages, complete the following steps:

1. Make sure you are working with kernel 2.6.16 or greater.

2. Modify the build procedure so your application is linked against
libhugetlbfs.

For the remapping, you link against the library with the appropriate
linker script (if necessary or desired).  Linking against the library
should result in transparent usage of hugepages.

Example 2:  End Users and System Administrators
-----------------------------------------------

To have an application use libhugetlbfs, complete the following steps:

1. Make sure you are using kernel 2.6.16.

2. Make sure the library is in the path, which you can set with the
LD_LIBRARY_PATH environment variable. You might need to set other
environment variables, including LD_PRELOAD as described above.


Troubleshooting
===============

The library has a certain amount of debugging code built in, which can
be controlled with the environment variable HUGETLB_VERBOSE.  By
default the debug level is "1" which means the library will only print
relatively serious error messages.  Setting HUGETLB_VERBOSE=2 or
higher will enable more debug messages (at present 2 is the highest
debug level, but that may change).  Setting HUGETLB_VERBOSE=0 will
silence the library completely, even in the case of errors - the only
exception is in cases where the library has to abort(), which can
happen if something goes wrong in the middle of unmapping and
remapping segments for the text/data/bss feature.

If an application fails to run, set the environment variable HUGETLB_DEBUG
to 1. This causes additional diagnostics to be run. This information should
be included when sending bug reports to the libhugetlbfs team.

Trademarks
==========

This work represents the view of the author and does not necessarily
represent the view of IBM.

PowerPC is a registered trademark of International Business Machines
Corporation in the United States, other countries, or both.  Linux is
a trademark of Linus Torvalds in the United States, other countries,
or both.