Paulo Alcantara [Sun, 8 Apr 2012 19:33:16 +0000 (16:33 -0300)]
PXELINUX: do not load 32-bit environment twice
Calling load_env32() once, it should never return. If it does, then jump
to kaboom.
Signed-off-by: Paulo Alcantara <pcacjr@zytor.com>
H. Peter Anvin [Tue, 27 Mar 2012 05:51:09 +0000 (22:51 -0700)]
libinstaller: Avoid using <linux/ext2_fs.h>
Don't use <linux/ext2_fs.h> if we can avoid it.
The ioctl constants have been globalized and moved to <linux/fs.h>.
Use a private copy of ext2_fs.h from e2fsprogs with the ioctl
constants removed for the data structures.
Do at least attempt backward compatibility for old kernel headers, but
no real hope of proper operation there...
[ Doing this as a cherry-pick to unblock this branch ]
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
Resolved Conflicts:
libinstaller/linuxioctl.h
H. Peter Anvin [Tue, 27 Mar 2012 21:47:04 +0000 (14:47 -0700)]
Makefile.private: remove obsolete preupload rule
Remove obsolete preupload rule, already deleted on the master branch.
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
H. Peter Anvin [Tue, 27 Mar 2012 21:44:23 +0000 (14:44 -0700)]
version: Bump date to 2012
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
H. Peter Anvin [Tue, 27 Mar 2012 21:42:51 +0000 (14:42 -0700)]
kernel: Support multiple initrds
Support loading multiple initrd images separated by commas. It still
doesn't correctly handle a secondary initrd= argument overriding the
first; not clear if that really was a good idea or if they would be
considered additive arguments.
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
H. Peter Anvin [Tue, 27 Mar 2012 21:41:55 +0000 (14:41 -0700)]
Revert "ldlinux: Disable VGA graphics when loading a kernel"
This reverts commit
87320b8de8f331b9c1810e98e24426e70b0ae29f.
That commit would change the mode before the kernel was actually
loaded, but would leave the console nonfunctional if it was vesacon.
We may need to do something similar eventually, but that one is not
correct.
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
H. Peter Anvin [Tue, 27 Mar 2012 21:39:46 +0000 (14:39 -0700)]
menu: Make "menu clear" the default
Make "menu clear" the default... it was sort of implicitly so at least
for vesamenu when using the old system, since we would end up zapping
the mode when any output happened, but that no longer is the case.
Also move the cursor to the top of the screen.
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
Matt Fleming [Tue, 13 Mar 2012 12:23:56 +0000 (12:23 +0000)]
core: Abstract search_config() into search_dirs()
The core currently abuses search_config() in order to locate
ldlinux.c32. The problem with this is that search_config() updates
ConfigName with the realpath of the file on every invocation. If we've
previously set ConfigName search_config() will overwrite this value.
Signed-off-by: Matt Fleming <matt.fleming@intel.com>
Matt Fleming [Mon, 12 Mar 2012 19:40:33 +0000 (19:40 +0000)]
mk: Don't link against compiler's version of libgcc
Syslinux provides its own copy of libgcc functions in
com32/lib/libgcc, so we don't need to link against the version
provided by the compiler. What's worse, the behaviour is different
between the two versions.
This caused problems, for example, when parsing the TOTALTIMEOUT
config directive in com32/menu because of the following calculation,
totaltimeout = (atoll(skipspace(p + 13)) * CLK_TCK + 9) / 10;
which invokes __divdi3() as provided by libgcc. The results returned
by the Syslinux and compiler's copy of libgcc differ, leading to
differences in behaviour.
Signed-off-by: Matt Fleming <matt.fleming@intel.com>
Matt Fleming [Fri, 9 Mar 2012 12:49:50 +0000 (12:49 +0000)]
ldlinux: Don't alloc initramfs unless initrd= is specified
It only makes sense to allocate memory for 'initramfs' if an initrd=
argument was on the command line, otherwise we're wasting memory.
Furthermore needlessly allocating 'initramfs' actually caused old
kernels, such as memtest86, to fail to run because of the following
check in syslinux_boot_kernel(),
if (initramfs && hdr.version < 0x0200)
goto bail; /* initrd/initramfs not supported */
Move the allocation after we've checked for and found an initrd=
argument.
Signed-off-by: Matt Fleming <matt.fleming@intel.com>
Matt Fleming [Fri, 9 Mar 2012 10:53:52 +0000 (10:53 +0000)]
ldlinux: Disable VGA graphics when loading a kernel
The state of the graphics hardware is unknown when we're about to load
a new kernel. It could be running in VGA mode, so let's disable VGA
graphics before passing control to the kernel.
The old assembly version of this code made extensive used of
writestr(), which always calls vgaclearmode(). In the C version we're
using printf() and so need to call vgaclearmode() directly. Doing it
directly has the added advantage of documenting exactly where we need
to reset the graphics hardware.
Signed-off-by: Matt Fleming <matt.fleming@intel.com>
Matt Fleming [Tue, 13 Mar 2012 09:39:14 +0000 (09:39 +0000)]
ldlinux: Make the default execute type KT_KERNEL
Default to trying to boot a linux kernel image, which was the
behaviour of the old asm image file loader.
Note that we don't need to go through __intcall() to load a kernel as
the dynamic linker will fixup any references to the kernel load
functions for us at runtime. Furthermore, modules invoking execute()
could be located anywhere in the 32-bit address space and may not be
able to represent their pointer arguments with SEG() and OFFS(), so
using the comboot API may not even work.
Signed-off-by: Matt Fleming <matt.fleming@intel.com>
Matt Fleming [Wed, 7 Mar 2012 15:18:51 +0000 (15:18 +0000)]
ldlinux: Avoid initialised data memory corruption
We can't realloc() 'PATH' because realloc() may just extend the
malloc'd region if the adjacent region is free, as opposed to
allocating a new region and then copying the data. This behaviour is
fine in most circumstances but not with initialised string data, such
as 'PATH'. The reason is that other string data pointers may point to
characters in 'PATH' and if we modify it after realloc()'ing, we'll
appear to corrupt unrelated string data.
For example, the string "/" is used in chdir() and the address of that
string is the last "/" in 'PATH'. If we realloc() and then append
"foo" to 'PATH' the string pointer in chdir() will now point to "/foo".
Initialise 'PATH' at runtime using malloc() and free() to avoid
corrupting string data.
Signed-off-by: Matt Fleming <matt.fleming@intel.com>
Matt Fleming [Tue, 13 Mar 2012 10:06:59 +0000 (10:06 +0000)]
ldlinux: chdir to the APPEND str when parsing a CONFIG directive
If an APPEND directive was specified for a CONFIG label then we need
to chdir to that directory after parsing the new config file.
Signed-off-by: Matt Fleming <matt.fleming@intel.com>
Matt Fleming [Tue, 13 Mar 2012 09:58:26 +0000 (09:58 +0000)]
com32: Delete execute.c and use the one from ldlinux
ldlinux provides us with an execute.c and all the functions that we
need, there's no point in duplicating the code in com32/menu.
Also, the copy in ldlinux doesn't use __com32.cs_bounce. Using
cs_bounce no longer works with dynamic ELF modules because we don't
have such tight control over the address space. Instead, memory must
be allocated dynamically or reserved as part of the object file.
Signed-off-by: Matt Fleming <matt.fleming@intel.com>
Matt Fleming [Thu, 1 Mar 2012 17:23:57 +0000 (17:23 +0000)]
elflink: Don't use strcmp on a non-NUL terminated string
The syslinux implementation of strcmp() only checks for NUL in the
first argument and will use any NULs in the second argument for
comparison with the corresponding character in the first string. This
caused problems because at the time strcmp() is called 'path' isn't
NUL-terminated. Since we're only checking the first character of
'path' just code it explicitly.
This bug caused some modules not to load when resolving module
dependencies via the ELF DT_NEEDED field because the "." in the PATH
lookup code wouldn't be expanded to the current working directory. For
example, instead of expanding to "/foo/bar" if /foo was the cwd, the
path would be ".bar".
Signed-off-by: Matt Fleming <matt.fleming@intel.com>
Matt Fleming [Tue, 13 Mar 2012 09:37:57 +0000 (09:37 +0000)]
ldlinux: Loading a config file should cause re-initialisation
There are a number of initialisation steps that need to be performed
*every* time a config file is loaded. Reload ldlinux.c32 so that we
can re-initialise the environment whenever a new config file is
loaded. This involves unloading all the modules that have been loaded
since ldlinux.c32. Luckily the list of loaded modules is sorted by
load order, which means it's trivial to "pop" them from the front of
the list.
Signed-off-by: Matt Fleming <matt.fleming@intel.com>
Matt Fleming [Thu, 22 Mar 2012 15:31:43 +0000 (15:31 +0000)]
elf: Support __constructor and __destructor
The old way of specifying functions that need to be run before/after a
main function has never worked for ELF modules. Instead, the only way
to get similiar behaviour was by using the MODULE_INIT() and
MODULE_EXIT() macros, but no one seems to have bothered converting the
old __constructor users over. Anyway, the old way is superior because
it allows multiple functions be specified. Delete the MODULE_* macros
as there's only one user of them in the entire tree.
We can also get rid of the UNKNOWN_MODULE constant because now a
module doesn't need a __module_init_ptr symbol to be classed as a
library - if a module isn't an executable, it's a library, there's no
such thing as an unknown type.
It's no longer necessary to explicitly call
__syslinux_get_ipappend_strings() from ldlinux.c because the
__constructor tag on the version in com32/lib/syslinux will take care
of invoking it before ldlinux.c32's main() is executed.
Also, since we've refactored unload_module() to now run a module's
destructors let's minimise the number of callers by deleting
load_library() and unload_library(), which were only called by the
test module in com32/elflink.
Signed-off-by: Matt Fleming <matt.fleming@intel.com>
Matt Fleming [Tue, 20 Mar 2012 16:27:17 +0000 (16:27 +0000)]
core: Shrink core/ldlinux.sys to be under 64K bytes
It would seem that running from a btrfs file system has been broken
for some time. The reason is that core/ldlinux.sys has grown
substantially, resulting in it weighing in at a whopping 88K bytes and
no longer fitting within the first 64K segment of a btrfs file system.
This huge size stems from the fact that the core links with
libcom32min.a because it requires various symbols contained within
that archive but it includes and exports *all* the symbols from
libcom32min.a. The reasoning behind this design decision is that when
a module is loaded any undefined symbols that are defined in
libcom32min.a can be resolved at runtime by the core, without the
module also needing to statically link against libcom32min.a.
Unfortunately doing this has increased the size of the core beyond
acceptable limits, such that booting from a btrfs file system no
longer works.
This commit links ldlinux.c32 against libcom32min.a so that now
ldlinux.c32 exports all the symbols contained within the
archive. Since we will always load ldlinux.c32 before any other
modules, any undefined symbols will now be resolved by ldlinux.c32
instead of the core. ldlinux.c32 isn't subject to same size
constraints as the core, e.g. fitting within 64K.
Here are the sizes in bytes before and after this commit,
Before:
33806 com32/elflink/ldlinux/ldlinux.c32
87749 core/ldlinux.sys
After:
393871 com32/elflink/ldlinux/ldlinux.c32
45516 core/ldlinux.sys
Signed-off-by: Matt Fleming <matt.fleming@intel.com>
Matt Fleming [Thu, 1 Mar 2012 11:54:13 +0000 (11:54 +0000)]
ldlinux: Fix detection of function keys
__get_key() directly calls getchar() but doesn't properly detect when
a function key such as F1, F2, etc has been entered. Despite the
previous commit to cli.c which read,
ldlinux: Accept commands from the serial console
To mimic the old (pre-elflink) command-line interface behaviour let's
use getchar() instead of reading from stdin. This way, if the user
types a command on the serial console it will actually be executed.
I can't see how get_key() fails to detect input on the serial
console. So let's revert to calling get_key() because it actually
handles the detection of function keys.
Signed-off-by: Matt Fleming <matt.fleming@intel.com>
Matt Fleming [Tue, 28 Feb 2012 21:06:38 +0000 (21:06 +0000)]
core: Check for func key in getchar()
We need to return the value stored in ah after reading from the
keyboard with __intcall(0x16, ...) otherwise we won't properly detect
function keys such as F1, F2, etc.
Signed-off-by: Matt Fleming <matt.fleming@intel.com>
Matt Fleming [Thu, 16 Feb 2012 12:12:50 +0000 (12:12 +0000)]
core: Delete the aux segment
We don't need to use a special aux segment because we can represent
'fontbuf' with OFFS() and SEG(). We're guaranteed to be able to break
the 32-bit pointer up with these macros because we control where in
the address space the core is loaded. (Not all 32-bit pointers can be
represented with OFFS() and SEG(), only addresses in the range
0..0xFFFFF.)
This fixes the breakage that was introduced in commit
14531c47bc95
("core: Delete code that is duplicated in ldlinux").
This allows the default font to be displayed. Previously junk was
being returned in the COMBOOT API call to query the userfont, leading
the caller to believe that a user font was installed even when it
wasn't.
Signed-off-by: Matt Fleming <matt.fleming@intel.com>
Matt Fleming [Tue, 14 Feb 2012 13:34:00 +0000 (13:34 +0000)]
core: Only adjust screen if we're in text mode
A bug was introduced in commit
9f51b69d ("core: Reimplement lots asm
code in C") when converting the old asm comapi_usingvga function into
C. We only want to adjust the screen if we're not using VGA.
Signed-off-by: Matt Fleming <matt.fleming@intel.com>
Matt Fleming [Mon, 13 Feb 2012 14:27:36 +0000 (14:27 +0000)]
com32/menu: vesamenu.c32 and menu.c32 require multiple obj files
The standard %.c32 rule in mk/elf.mk doesn't work for vesamenu.c32 and
menu.c32 because they need multiple object files to be linked. Replace
the *.elf targets with *.c32 because .c32 *IS* an ELF file now.
Signed-off-by: Matt Fleming <matt.fleming@intel.com>
Matt Fleming [Thu, 9 Feb 2012 10:15:38 +0000 (10:15 +0000)]
ldlinux: Parse kernel type for labels
We need to parse the kernel type for labels aswell as things entered
on the cmdline, instead of always passing KT_KERNEL or KT_NONE to
execute(). Move the logic into a new helper function.
This fixes a bug where an incorrect kernel type would be passed to
execute() if anything other than a linux kernel (such as a .bin) was
specified in a LABEL's KERNEL argument, which resulted in the file not
being executed.
Signed-off-by: Matt Fleming <matt.fleming@intel.com>
Matt Fleming [Thu, 15 Dec 2011 20:47:00 +0000 (20:47 +0000)]
ldlinux: Initialise 'p' before using it.
Signed-off-by: Matt Fleming <matt.fleming@intel.com>
H. Peter Anvin [Thu, 15 Dec 2011 18:47:41 +0000 (10:47 -0800)]
Revert "core: remove redundant zalloc() definition"
This reverts commit
46b993449264a2332427a165b4c77e400a053178.
Build failure was due to a non-clean tree.
Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
H. Peter Anvin [Wed, 14 Dec 2011 22:00:21 +0000 (14:00 -0800)]
core: remove redundant zalloc() definition
Remove a redundant definition of zalloc() which causes a compile
failure.
Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
Matt Fleming [Fri, 25 Nov 2011 16:22:44 +0000 (16:22 +0000)]
ldlinux: Accept commands from the serial console
To mimic the old (pre-elflink) command-line interface behaviour let's
use getchar() instead of reading from stdin. This way, if the user
types a command on the serial console it will actually be executed.
Signed-off-by: Matt Fleming <matt.fleming@intel.com>
Matt Fleming [Fri, 25 Nov 2011 15:54:48 +0000 (15:54 +0000)]
ldlinux: Add eprintf() to print to VGA and serial
printf() is used heavily in the ldlinux code but that only displays
things on the VGA console, not on serial. Introduce eprintf(), which
gives us the benefit of format strings but will also print to both VGA
and serial.
Signed-off-by: Matt Fleming <matt.fleming@intel.com>
Matt Fleming [Fri, 25 Nov 2011 15:36:55 +0000 (15:36 +0000)]
ldlinux: Don't return previous command if user pressenter enter
If the user presses enter at the command prompt then simply print the
command prompt again on a newline - don't re-execute the previous
command.
Signed-off-by: Matt Fleming <matt.fleming@intel.com>
Matt Fleming [Fri, 25 Nov 2011 15:09:11 +0000 (15:09 +0000)]
elflink: Include '\n' when printing error messages
We need to print a newline character at the end of our error messages
otherwise the syslinux command prompt may overwrite them, making them
impossible to see.
Signed-off-by: Matt Fleming <matt.fleming@intel.com>
Matt Fleming [Wed, 9 Nov 2011 13:19:54 +0000 (13:19 +0000)]
ldlinux: Make tab-completion output same as pre-elflink
To maintain compatability with the old pre-elflink command line code,
the list of potential labels should all be printed on one line, not
one per line.
Signed-off-by: Matt Fleming <matt.fleming@intel.com>
Matt Fleming [Fri, 4 Nov 2011 15:02:14 +0000 (15:02 +0000)]
core: Split core console code into separate libs
When the console code was written in asm ldlinux.asm, isolinux.asm and
pxelinux.asm simply included the correct console I/O files (rawcon.inc
and plaincon.inc) but now that these files are implemented in C we
need another way to link against the correct functions.
Create separate libraries for linking ldlinux, isolinux and pxelinux
against.
Signed-off-by: Matt Fleming <matt.fleming@intel.com>
Matt Fleming [Fri, 4 Nov 2011 14:25:20 +0000 (14:25 +0000)]
pxelinux: Call load_env32()
Execute the elflink environment.
Signed-off-by: Matt Fleming <matt.fleming@intel.com>
Matt Fleming [Tue, 24 May 2011 09:45:51 +0000 (10:45 +0100)]
core: Delete code that is duplicated in ldlinux
Lots of code that used to be implemented in the core is now
implemented in ldlinux, but code from core was never deleted. Purge
this code.
Also, move all com32 loading to ldlinux since we need to be able to do
various command line things (which are no longer available in the
core).
Signed-off-by: Matt Fleming <matt.fleming@intel.com>
Matt Fleming [Wed, 19 Oct 2011 09:02:42 +0000 (10:02 +0100)]
core: Reimplement writestr in C
Implement writestr in C instead of asm and move writestr.inc into
modules/ because it's required by modules/ver.asm.
Signed-off-by: Matt Fleming <matt.fleming@intel.com>
Matt Fleming [Tue, 18 Oct 2011 12:13:06 +0000 (13:13 +0100)]
core: Reimplement lots asm code in C
There is an awful lot of code currently implemented in assembly when
it could just as easily be implemented in C. Having it in C makes it
much easier to share code between the BIOS and forthcoming EFI
firmware backend. The following code fragments have been rewritten,
- timer initialisation
- adjust_screen()
- check_esapes() and mem_init()
- conio.inc
- plaincon.inc
- cleanup.inc
- serirq.inc
- font.inc
- graphics
- writehex
Signed-off-by: Matt Fleming <matt.fleming@intel.com>
Matt Fleming [Wed, 18 May 2011 14:55:21 +0000 (15:55 +0100)]
com32: Add unhexchar
Add a C implementation of unhexchar() for use in C code. This
implementation should be an exact duplicate of the old asm version.
Signed-off-by: Matt Fleming <matt.fleming@linux.intel.com>
H. Peter Anvin [Thu, 28 Jul 2011 21:02:58 +0000 (14:02 -0700)]
Merge remote-tracking branch 'mfleming/for-hpa/elflink/ldlinux' into elflink
Matt Fleming [Tue, 12 Jul 2011 15:21:37 +0000 (16:21 +0100)]
pxelinux: open_file() returns a non-negative handle
The usage of open_file() is wrong in core/fs/pxe/pxe.c. Any
non-negative return value indicates success, not just a return value
of zero.
This bug was introduced in commit
ba4fefa9b52b "core: change
load_config() to open_config()". The bug causes pxelinux to fail to
open configuration files.
Signed-off-by: Matt Fleming <matt.fleming@linux.intel.com>
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
Matt Fleming [Tue, 12 Jul 2011 15:12:48 +0000 (16:12 +0100)]
pxelinux: open_file() returns a non-negative handle
The usage of open_file() is wrong in core/fs/pxe/pxe.c. Any
non-negative return value indicates success, not just a return value
of zero.
This bug was introduced in commit
ba4fefa9b52b "core: change
load_config() to open_config()". The bug causes pxelinux to fail to
open configuration files.
Signed-off-by: Matt Fleming <matt.fleming@linux.intel.com>
Matt Fleming [Tue, 12 Jul 2011 10:52:05 +0000 (11:52 +0100)]
core: Fix ldlinux.c32 failing to load when not installed in "/"
Currently, if syslinux is installed in a directory other than "/" we
will fail to load ldlinux.c32.
Because we know where we were installed the simplest fix is to chdir()
to the installation directory before attempting to load
ldlinux.c32. This requires us to add "." to PATH so that we look in
the current working directory when loading ELF modules.
Changing to the install directory needs to handled differently for
different file systems, which is the reason behind the new file system
operation, .chdir_start(). Disk-based Syslinux variants should use
generic_chdir_start() to chdir() to CurrentDirName, which is the
installation path. By using this new fs operation, we can load
ldlinux.c32 when we expand the "." in PATH, without having to search
the entire filesystem for it. However, iso9660 file systems still
require us to search some directories because it has no notion of an
installation directory.
Previously, changing into the install directory was handled when
opening the config file, but because the config file parser is now
part of the ldlinux.c32 ELF module, it needs to be done much earlier.
Reported-by: H. Peter Anvin <hpa@zytor.com>
Signed-off-by: Matt Fleming <matt.fleming@linux.intel.com>
Matt Fleming [Tue, 7 Jun 2011 19:12:36 +0000 (20:12 +0100)]
Merge branch 'DT_NEEDED-module-deps' into for-hpa/elflink/ldlinux
Matt Fleming [Tue, 7 Jun 2011 19:12:17 +0000 (20:12 +0100)]
Merge branch 'PATH-based-search' into for-hpa/elflink/ldlinux
Conflicts:
com32/elflink/ldlinux/readconfig.c
Matt Fleming [Tue, 7 Jun 2011 19:10:59 +0000 (20:10 +0100)]
Merge branch 'fs-config' into for-hpa/elflink/ldlinux
Conflicts:
com32/elflink/ldlinux/readconfig.c
Matt Fleming [Tue, 7 Jun 2011 19:06:01 +0000 (20:06 +0100)]
Merge branch 'pxelinux' into for-hpa/elflink/ldlinux
Conflicts:
core/pxelinux.asm
Matt Fleming [Fri, 27 May 2011 16:57:32 +0000 (17:57 +0100)]
ldlinux: Support "kbdmap"
Add support to the ldlinux config file parser for the "kbdmap"
directive.
Signed-off-by: Matt Fleming <matt.fleming@linux.intel.com>
Matt Fleming [Fri, 27 May 2011 16:57:05 +0000 (17:57 +0100)]
ldlinux: Support "font"
Add support for parsing the "font" config directive to ldlinux.
Signed-off-by: Matt Fleming <matt.fleming@linux.intel.com>
Matt Fleming [Fri, 27 May 2011 16:49:39 +0000 (17:49 +0100)]
ldlinux: Parse "display" in config files
Teach ldlinux to parse the "display" directive.
Signed-off-by: Matt Fleming <matt.fleming@linux.intel.com>
Matt Fleming [Fri, 27 May 2011 16:10:32 +0000 (17:10 +0100)]
ldlinux: Add support for parsing "serial" config directive
Signed-off-by: Matt Fleming <matt.fleming@linux.intel.com>
Matt Fleming [Mon, 6 Jun 2011 13:12:45 +0000 (14:12 +0100)]
ldlinux: Use DT_NEEDED for module dependencies
The ELF file format uses DT_NEEDED entries in the dynamic section to
name any shared library dependencies. Instead of rolling our own
dependency logic via the modules.dep file and elf_gen_dep.sh script
use the DT_NEEDED entries.
Signed-off-by: Matt Fleming <matt.fleming@linux.intel.com>
Matt Fleming [Mon, 6 Jun 2011 21:45:33 +0000 (22:45 +0100)]
ldlinux: Use open_config() to open config files
Instead of duplicating the logic for finding config files on different
filesystems in both ldlinux/readconfig.c and core/fs/* use the
open_config() wrapper that calls into the fs-specific open_config().
Signed-off-by: Matt Fleming <matt.fleming@linux.intel.com>
Matt Fleming [Mon, 6 Jun 2011 21:17:03 +0000 (22:17 +0100)]
core: Return a file descriptor from open_config()
Wrap open_config() the same way that open() wraps open_file(). The
only user of open_config() requires access to a file descriptor so it
makes sense to return a file descriptor.
The file handle implementation is a historic piece of code and this
patch tries to hide it as they will likely be removed at a future
point in time. Furthermore, the file handle code is very core-specific
and should not be exposed to any callers of open_config().
Signed-off-by: Matt Fleming <matt.fleming@linux.intel.com>
Matt Fleming [Mon, 6 Jun 2011 17:19:02 +0000 (18:19 +0100)]
core: open_file() returns a non-negative handle id
We currently only take the success path if open_file() returns 0 but
open_file() returns a handle ID which is a non-negative number.
Signed-off-by: Matt Fleming <matt.fleming@linux.intel.com>
Matt Fleming [Mon, 6 Jun 2011 10:09:42 +0000 (11:09 +0100)]
ldlinux: PATH-based module lookup
Add support for specifying directories to search when loading
modules. A new config directive, "PATH", instructs the module loading
code to search the directories listed in a colon-separated list when
loading a module.
Signed-off-by: Matt Fleming <matt.fleming@linux.intel.com>
H. Peter Anvin [Sat, 28 May 2011 00:45:59 +0000 (17:45 -0700)]
core: change load_config() to open_config()
Change load_config() to open_config(), which is a method that works
just like open_file(). This we can use to get the original
configuration file.
Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
H. Peter Anvin [Sat, 28 May 2011 00:45:22 +0000 (17:45 -0700)]
core: fix warnings in core/mem/init.c
Fix warnings (that block -Werror) in core/mem/init.c
Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
H. Peter Anvin [Sat, 28 May 2011 00:44:48 +0000 (17:44 -0700)]
core.h: add a few include files needed pretty much everywhere
Some include files we might as well centralize...
Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
Matt Fleming [Fri, 27 May 2011 16:09:27 +0000 (17:09 +0100)]
core: Load the 32-bit environment at startup
Do the same as the other Syslinux variants and initialise the 32-bit
environment when we boot.
Signed-off-by: Matt Fleming <matt.fleming@linux.intel.com>
Matt Fleming [Wed, 25 May 2011 22:29:05 +0000 (23:29 +0100)]
ldlinux: Add support for parsing pxeretry option
Signed-off-by: Matt Fleming <matt.fleming@linux.intel.com>
Matt Fleming [Wed, 25 May 2011 19:26:23 +0000 (20:26 +0100)]
ldlinux: Initialise ipappend strings
Previouly __syslinux_get_ipappend_strings() was called as a
__constructor. However, as constructors are executed before we load
any ELF modules we need to explicitly call this function to initialise
the ipappend string.
Signed-off-by: Matt Fleming <matt.fleming@linux.intel.com>
Matt Fleming [Wed, 25 May 2011 14:35:53 +0000 (15:35 +0100)]
ldlinux: Fix "prompt" config option parsing
Previously, we were looking for the value of the prompt option 8
characters after the first 'p', clearly since "prompt" is only 6
characters we should be searching after the sixth character.
I noticed this bug when specifying "prompt 1" in my config file and
Syslinux still tried to boot the default kernel.
Signed-off-by: Matt Fleming <matt.fleming@linux.intel.com>
Matt Fleming [Wed, 25 May 2011 07:45:37 +0000 (08:45 +0100)]
ldlinux: Remove __constructor attribute
The __constructor tag only makes sense in the core as the constructor
functions are executed from load_env32(), before any modules have been
loaded. Therefore, applying __constructor to code inside ldlinux is
pointless as unless the functions are called explicitly, they will
never be executed.
Signed-off-by: Matt Fleming <matt.fleming@linux.intel.com>
Matt Fleming [Wed, 25 May 2011 07:39:47 +0000 (08:39 +0100)]
ldlinux: Add support for Auxillary Data Vector
Move all the code for the ADV into ldlinux so that it doesn't have any
dependencies on other modules.
We also need a way to initialize the ADV from ldlinux, so add another
vector to the comboot API.
Signed-off-by: Matt Fleming <matt.fleming@linux.intel.com>
Matt Fleming [Tue, 24 May 2011 10:54:10 +0000 (11:54 +0100)]
ldlinux: Add localboot support
Add a few tweaks to enable ldlinux to localboot a hard disk. Most of
the support was already there, there were just a few bug fixes in
execute().
Signed-off-by: Matt Fleming <matt.fleming@linux.intel.com>
Matt Fleming [Tue, 24 May 2011 10:37:49 +0000 (11:37 +0100)]
ldlinux: Search in isolinux/ directory for config files
If we're running from an ISO image we should be looking in the
isolinux/ directory for isolinux.cfg, not in the top-level directory.
Signed-off-by: Matt Fleming <matt.fleming@linux.intel.com>
Matt Fleming [Fri, 29 Apr 2011 10:20:11 +0000 (11:20 +0100)]
hdt: libmenu.a no longer exists
Don't link against libmenu.a, it no longer exists as it was removed in
commit
74518b8b691c ("elflink: Make ELF the default object format").
This fixes the following build error,
make[2]: *** No rule to make target `../cmenu/libmenu/libmenu.a', needed by `hdt.c32'. Stop.
Signed-off-by: Matt Fleming <matt.fleming@linux.intel.com>
H. Peter Anvin [Wed, 27 Apr 2011 21:27:35 +0000 (14:27 -0700)]
load_env32 should not include menu.h
Remove stay #include that does not resolve.
Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
Matt Fleming [Fri, 15 Apr 2011 20:35:48 +0000 (21:35 +0100)]
elf_gen_dep.sh: Prioritise symbols exported by core
It's possible for multiple modules to export the same symbol. We
always favour the symbols exported from core/ because this means that
modules will have less runtime dependencies.
For example, some symbols required by ldlinux.c32 are exported by both
core/ and libmenu.c32. Because we don't want ldlinux.c32 to have any
dependencies, we need to make sure we resolve to the symbol exported
by core/.
Note that now the order in which we extract global symbols from ELF
files in elf_gen_dep.sh is important, because this dictates our order
of preference and hence core/isolinux.elf and core/pxelinux.elf need
to come first.
Signed-off-by: Matt Fleming <matt.fleming@linux.intel.com>
Matt Fleming [Fri, 15 Apr 2011 20:27:37 +0000 (21:27 +0100)]
elf_gen_dep.sh: Don't calculate dependencies for whitespace symbols
Don't try to resolve dependencies for a symbol whose name is purely
made up of whitespace.
Signed-off-by: Matt Fleming <matt.fleming@linux.intel.com>
Matt Fleming [Fri, 15 Apr 2011 20:10:55 +0000 (21:10 +0100)]
elflink: Don't require every module to have init/exit functions
Don't complain or refuse to load a module if it doesn't contain an
init or exit function, as many of the init/exit functions are in fact
empty.
Signed-off-by: Matt Fleming <matt.fleming@linux.intel.com>
Matt Fleming [Fri, 15 Apr 2011 20:03:51 +0000 (21:03 +0100)]
elflink: Don't compile in debug output by default
Turn off ELF_DEBUG, we really don't need to see the "[ELF] MODULE
UNLOADED" message everytime we finish executing an ELF module.
Signed-off-by: Matt Fleming <matt.fleming@linux.intel.com>
Matt Fleming [Tue, 26 Apr 2011 09:02:13 +0000 (10:02 +0100)]
elflink: ldlinux should not have any dependencies
ldlinux currently requires the get_key() symbol, which means it has a
dependency on libutil_com.c32. ldlinux.c32 really should be a
standalone module so let's move get_key.c into ldlinux/.
However, moving get_key.c creates a new problem - native linux
applications such as com32/samples/keytest.c require the get_key()
symbol which is currently exported by libutil_lnx.a. To fix this, we
create a new native linux library that exports any required symbols
from ldlinux. With this change we need to update com32/Makefile so
that we build com32/elfink/ldlinux before com32/samples, and seeing as
ldlinux now has no dependencies, we may as well move it to the front
of $SUBDIRS.
Also, update elf_gen_dep.sh to skip ldlinux.c32 and not search for
dependencies for any of its unresolved symbols. This way, if anyone
inadvertently introduces a dependency that cannot be resolved by the
core, ldlinux.c32 will fail to load at runtime (which guarantees that
the newly introduced dependency won't go unnoticed).
Signed-off-by: Matt Fleming <matt.fleming@linux.intel.com>
Matt Fleming [Fri, 15 Apr 2011 10:48:37 +0000 (11:48 +0100)]
exit.c: Truncate exit status to uint8_t
The valid range for an exit status is 0 - 255, so we need to truncate
the value passed to _exit().
I noticed this when a module was doing _exit(-1), and ended up calling
longjmp(.., 0xffffffff + 1)
which meant that setjmp() in spawn_load() returned 0. Obviously, we
wanted the setjmp() to return 256 (0xff + 1), because the code in
spawn_load() handles the return value like so,
ret_val = setjmp(module->u.x.process_exit);
if (ret_val)
ret_val--; /* Valid range is 0-255 */
else if (!module->main_func)
ret_val = -1;
else
exit((module->main_func)(argc, argv)); /* Actually run! */
There actually is code in spawn_load() to properly truncate 'ret_val',
but it is applied too late. The truncation needs to happen when we
pass the exit status to longjmp().
Suggested-by: H. Peter Anvin <hpa@zytor.com>
Signed-off-by: Matt Fleming <matt.fleming@linux.intel.com>
Matt Fleming [Tue, 26 Apr 2011 08:59:53 +0000 (09:59 +0100)]
elflink: Make ELF the default object format
com32/elflink/modules was originally created to house ELF modules and
keep them separate from the COM32 modules as the elflink branch was
being developed. However, this has inadvertently created a maintenance
nightmare because code was copied from elsewhere in the tree into
com32/elflink/modules, resulting in duplication. Bug fixes have been
going into the original code but have not been merged onto the elflink
branch, leaving the duplicate code in com32/elflink/modules buggy.
So let's delete this directory. There really is no reason to keep ELF
and COM32 modules separate because there's no reason to need both
COM32 and ELF modules to coexist. ELF is a far superior object file
format and all modules are not emitted as ELF objects.
Now that we're outputting ELF modules we can use dynamic memory
instead of the cs_bounce bounce buffer.
This commit requires a certain amount of shuffling for some
files. quicksort.c isn't a module and belongs as part of the util
library. cli.h belongs in com32/include so that other modules can make
use of the cli code in ldlinux.c32.
All libraries are now ELF shared libraries which are only loaded to
fixup unresolved symbols for executable modules and renamed from *.a
to *.c32. This reduces the runtime memory footprint because libraries
are only loaded when needed and because every executable no longer
gets its own copy of code/data (as it would if linking with a static
library). Also, remove MINLIBOBJS from libcom32.c32 because it is
already part of libcom32min.a and we don't want to have any duplicate
symbols between the core (which links with libcom32min.a) and
libcom32.c32.
Welcome to the New World Order of ELF modules!
Signed-off-by: Matt Fleming <matt.fleming@linux.intel.com>
Matt Fleming [Wed, 13 Apr 2011 23:59:12 +0000 (00:59 +0100)]
elflink: Delete core-elf.h
This header file really just duplicates code that is available in
other header files. It has no reason to exist.
Signed-off-by: Matt Fleming <matt.fleming@linux.intel.com>
Matt Fleming [Wed, 13 Apr 2011 23:52:23 +0000 (00:52 +0100)]
com32: Move menu.h
menu.h is required by the menu code in com32/menu and also by
ldlinux. So move it to a more neutral place.
Signed-off-by: Matt Fleming <matt.fleming@linux.intel.com>
Matt Fleming [Wed, 13 Apr 2011 23:50:08 +0000 (00:50 +0100)]
com32: We only need one refstr.h
Delete the duplicate refstr.h and move the remaining copy into
com32/include so that it can be included by both ldlinux and the menu
code.
Signed-off-by: Matt Fleming <matt.fleming@linux.intel.com>
Matt Fleming [Wed, 13 Apr 2011 22:00:16 +0000 (23:00 +0100)]
cmenu: Rename menu.h to cmenu.h
In preparation for moving com32/menu/menu.h to com32/include/ let's
rename com32/cmenu/menu.h. This stops us having two header files named
menu.h which would mean that we'd have to be super careful with our
include paths to ensure we included the correct header.
Signed-off-by: Matt Fleming <matt.fleming@linux.intel.com>
Matt Fleming [Tue, 12 Apr 2011 21:28:57 +0000 (22:28 +0100)]
com32: Remove klibc.c32
We don't need this anymore as it includes DYNOBJS but libcom32.c32
also includes those objects and we can now dynamically load
libcom32.c32 at runtime.
Signed-off-by: Matt Fleming <matt.fleming@linux.intel.com>
Matt Fleming [Wed, 13 Apr 2011 11:54:49 +0000 (12:54 +0100)]
com32: Put deleted object files back into the Makefile
For some reason that I can't understand, some object files were
removed from this Makefile in the following commit,
|commit
0cb6490fa96f752c01bbb9add2c9ca6dbf99ece2
|Author: Feng Tang <feng.tang@intel.com>
|Date: Thu Jun 3 10:48:22 2010 +0800
|
| elflink: start merge with pathbased branch
|
| modify these files to make compile pass
|
| modified: com32/MCONFIG
| modified: com32/Makefile
| modified: com32/include/klibc/compiler.h
| modified: com32/include/sys/elfcommon.h
| modified: com32/lib/Makefile
| modified: com32/lib/free.c
| modified: com32/lib/malloc.c
However, with this patch things continue to compile correctly so I'm
unsure why they were ever removed. Furthermore, pieces of code in
com32/ refer to symbols defined in these these object files and so
they are required in order to load modules.
These missing object files were discovered when loading libcom32.c32,
which wouldn't load because it was complaining about unresolved
symbols.
The removal of core/strncasecmp.c needs some explanation. There are
currently two identical copies of strncasecmp.c, one in com32/lib and
one in core/, and while the core copy is part of core/libcom32.a, the
copy in com32/lib isn't compiled. Now, because there are no references
to strncasecmp within core/ the symbol isn't pulled in from libcom32.a
and exported as a global symbol by the core. So, if ELF modules have
references to strncasecmp they cannot be resolved at runtime.
To fix this I've included strncasecmp.o in libcom32min.a which
isolinux.elf, ldlinux.elf and pxelinux.elf link against with
--whole-archive, which means that even if there are no references to
strncasecmp within the core, that symbol is still exported.
Signed-off-by: Matt Fleming <matt.fleming@linux.intel.com>
Matt Fleming [Fri, 8 Apr 2011 12:21:51 +0000 (13:21 +0100)]
elflink: Move more code from core/ into ldlinux
ldlinux now contains all the code necessary to load and execute
modules, none is contained in the core.
This change also allows us to change the spawn_load() prototype and to
push the job of processing arguments to executable functions (e.g. the
contents of char *argv[] as passed to a module's main function) into
ldlinux/execute.c instead of doing it in spawn_load(). Moving it into
ldlinux/ makes sense because the only core user of spawn_load() is
load_env32() and we don't require any sort of argument processing in
that path.
Signed-off-by: Matt Fleming <matt.fleming@linux.intel.com>
Matt Fleming [Thu, 7 Apr 2011 20:02:02 +0000 (21:02 +0100)]
elflink: Remove MODULE_MAIN() from executable ELF modules
To make it easier to move existing COM32 modules over to ELF modules
without having to modify them we need to search for and run their
main() functions when they're loaded instead of requiring their
initial function to be labeled with MODULE_MAIN().
Currently, we require all executable ELF modules to specify their
initial function with the use of the MODULE_MAIN() macro and not via
the traditional method of naming it main(). However, there are weird
restrictions on what functions can be passed to MODULE_MAIN(), for
instance, they must be declared static.
This patch makes life much simpler and allows executable ELF modules
to be loaded without the MODULE_MAIN() wrapper, but rather by naming
their initial function main(). All the modules in com32/modules can
now be run as ELF modules without any modifications.
Signed-off-by: Matt Fleming <matt.fleming@linux.intel.com>
Matt Fleming [Fri, 8 Apr 2011 12:52:30 +0000 (13:52 +0100)]
com32: console is already initialised when we enter main()
There's no need to initialise the console when we enter main() as
ldlinux.c32 will already have initialised it for us. Leaving the calls
to openconsole() causes the console settings to be reset whenever we
run a module.
Signed-off-by: Matt Fleming <matt.fleming@linux.intel.com>
Matt Fleming [Tue, 12 Apr 2011 21:24:36 +0000 (22:24 +0100)]
exit: Delete exit.S and references to __exit_handler
We stopped using exit.S (and thus __exit_handler) in
71d5d1ba646e
("elflink: change the exit.c, remove the exit.S to make process exit
working") when we moved to using the C version of _exit(). If we don't
remove the symbol __exit_handler we'll encounter undefined symbol
errors when moving libcom32.a over to an ELF shared library.
Signed-off-by: Matt Fleming <matt.fleming@linux.intel.com>
Matt Fleming [Tue, 12 Apr 2011 21:23:55 +0000 (22:23 +0100)]
mk: LDFLAGS use elf_i386 emulation for ld
elf32_i386 isn't a valid ld emulation mode. Presumably no one is using
$(LDFLAGS) set in this file.
Signed-off-by: Matt Fleming <matt.fleming@linux.intel.com>
Matt Fleming [Thu, 7 Apr 2011 20:00:31 +0000 (21:00 +0100)]
mk: Add rule for emitting .c32 ELF files
We currently open code the Makefile rules for outputting .c32 files,
so provide a pattern rule to make things easier.
There are no users of this rule yet, but we'll make use of it when we
switch all modules over to ELF.
Signed-off-by: Matt Fleming <matt.fleming@linux.intel.com>
Matt Fleming [Thu, 7 Apr 2011 17:12:51 +0000 (18:12 +0100)]
elflink: Move elf MCONFIG to mk/elf.mk
Catch up with recent changes to the Makefile fragments in the MCONFIG
files.
Signed-off-by: Matt Fleming <matt.fleming@linux.intel.com>
Matt Fleming [Tue, 26 Apr 2011 08:51:32 +0000 (09:51 +0100)]
Merge branch 'for-hpa/elflink/core' into elflink
Matt Fleming [Tue, 26 Apr 2011 08:51:14 +0000 (09:51 +0100)]
Merge branch 'for-hpa/elflink/ldlinux' into elflink
Matt Fleming [Tue, 26 Apr 2011 08:46:34 +0000 (09:46 +0100)]
Merge remote-tracking branch 'zytor/master' into elflink
We need the recent Makefile filename changes to be merged into the
elflink branch because it will make things simpler when converting all
modules to ELF format.
Conflicts:
com32/Makefile
com32/modules/Makefile
version
H. Peter Anvin [Tue, 26 Apr 2011 08:30:29 +0000 (01:30 -0700)]
hdt: add back -I$(com32)
hdt picks up include files from outside the include directory and so
needs an explicit -I$(com32). This probably should be fixed.
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
H. Peter Anvin [Tue, 26 Apr 2011 08:29:28 +0000 (01:29 -0700)]
sysdump: add $(LIBS) to the link rule
$(C_LIBS) doesn't include $(LIBS) so we need to add it explicitly.
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
H. Peter Anvin [Tue, 26 Apr 2011 08:18:18 +0000 (01:18 -0700)]
Merge remote-tracking branch 'mfleming/for-hpa/makefile-cleanup'
Resolved Conflicts:
com32/hdt/Makefile
com32/sysdump/Makefile
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
H. Peter Anvin [Tue, 26 Apr 2011 03:18:07 +0000 (20:18 -0700)]
isolinux: remove broken and obsolete open_file_t
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
H. Peter Anvin [Tue, 26 Apr 2011 03:16:07 +0000 (20:16 -0700)]
isolinux: remove .img file support
Remove disk image support from ISOLINUX. Very few BIOSes support this
call correctly, and it is likely to have been broken for a very long
time anyway.
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
H. Peter Anvin [Tue, 26 Apr 2011 00:15:40 +0000 (17:15 -0700)]
com32: inet_ntoa() takes struct in_addr
The standard definition for inet_ntoa() is to take struct in_addr, and
not doing that causes a conflict on the lwip branch.
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
H. Peter Anvin [Tue, 26 Apr 2011 00:00:22 +0000 (17:00 -0700)]
hdt: remove set-but-unused variable
Remove a set but unused variable in HDT, which prevents building with
gcc 4.6. This is probably a bug; presumably the active flag should be
presented to the user, too.
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
H. Peter Anvin [Mon, 25 Apr 2011 22:24:33 +0000 (15:24 -0700)]
Merge remote-tracking branch 'sherbszt/gfxboot32'