iPXE adds Simple File System Protocol to loaded image handle, as side
effect it also adds Block IO protocol (according to comments, to work
around some bugs in EDK2). GRUB assumes that every device with Block IO
is disk and skips network initialization entirely. But iPXE Block IO
implementation is just a stub which always fails for every operation
so cannot be used. Attempt to detect and skip such devices.
We are using media ID which iPXE sets to "iPXE" and block IO size in
hope that no real device would announce 1B block ...
Closes: 50518
UEFI 2.6 9.3.6.4 File Path Media Device Path says that Path Name is
"A NULL-terminated Path string including directory and file names".
Strip final NULL from Path Name in each File Path node when constructing
full path. To be on safe side, strip all of them.
Fixes failure chainloading grub from grub, when loaded grub truncates
image path and does not find its grub.cfg.
https://bugzilla.opensuse.org/show_bug.cgi?id=1026344
This was triggered by commit ce95549cc54b5d6f494608a7c390dba3aab4fba7;
before it we built Path Name without trailing NULL, and apparently all
other bootloaders use single File Path node, thus not exposing this bug.
1. Do not assume block list and fragment are mutually exclusive. Squash
can pack file tail as fragment (unless -no-fragments is specified); so
check read offset and read either from block list or from fragments as
appropriate.
2. Support sparse files with zero blocks.
3. Fix fragment read - frag.offset is absolute fragment position,
not offset relative to ino.chunk.
Reported and tested by Carlo Caione <carlo@endlessm.com>
In case of GRUB we put remapper after domain pages and not at 0x0.
In this case we use max_addr to put remapper. Unfortunately we increment
max_addr as well in this case resulting in virt mapping mapping page
at old max_addr and trying to boot using new max_addr.
Closes 46014.
Unified Extensible Firmware Interface Specification, Version 2.6,
section 2.3.4, x64 Platforms, boot services, says among others:
The stack must be 16-byte aligned. So, do it. Otherwise OS may
boot only by chance as it happens right now.
Signed-off-by: Daniel Kiper <daniel.kiper@oracle.com>
According to EABI only STT_FUNC has convention of lowest bit indicating
execution mode. R_THM_{JUMP,CALL}* relocations are assumed to be pointing
to thumb mode unless they use STT_FUNC.
If ascent is bigger than height - 2, then we draw over character box but then
to clear cursor we only draw over character box. So trim ascent if necessarry.
These entries have placeholder for device name and so are useless for our
purpose. grub failed with something like
grub-install: error: failed to get canonical path of `systemd-1'.
When we see autofs entry, record it (to keep parent-child relationship) but
continue to look for real mount. If it is found, we process it as usual. If
only autofs entry exists, attempt to trigger mount by opening mount point
and retry. Mount point itself is then kept open to avoid timeout.
Recent systemd is by default using automount for /boot/efi so this should
become more popular problem on EFI systems.
Closes: 49942
The path returned by grub_efi_net_config has already been stripped for the
directory part extracted from cached bootp packet. We should just return the
result to avoild it be stripped again.
It fixed the problem that grub.efi as NBP image always looking for grub.cfg and
platform directory in upper folder rather than current one it gets loaded while
$prefix is empty. The behavior is inconsistent with other architecture and how
we would expect empty $prefix going to be in general.
The only exception to the general rule of empty $prefix is that when loaded
from platform directory itself, the platform part is stripped thus upper folder
is used for looking up files. It meets the case for how grub-mknetdir lay out
the files under tftp root directory, but also hide away this issue to be
identified as it appears to be just works.
Also fix possible memory leak by moving grub_efi_get_filename() call after
grub_efi_net_config().
EFI File Path Media Device Path is defined as NULL terminated string;
but chainloader built file paths without final NULL. This caused error
with Secure Boot and Linux Foundation PreLoader on Acer with InsydeH20 BIOS.
Apparently firmware failed verification with EFI_INVALID_PARAMETER which is
considered fatal error by PreLoader.
Reported and tested by Giovanni Santini <itachi.sama.amaterasu@gmail.com>
get_card_packet() from ofnet.c allocates a netbuff based on the device's MTU:
nb = grub_netbuff_alloc (dev->mtu + 64 + 2);
In the case when the MTU is large, and the received packet is
relatively small, this leads to allocation of significantly more memory,
than it's required. An example could be transmission of TFTP packets
with 0x400 blksize via a network card with 0x10000 MTU.
This patch implements a per-card receive buffer in a way similar to efinet.c,
and makes get_card_packet() allocate a netbuff of the received data size.
Signed-off-by: Stanislav Kholmanskikh <stanislav.kholmanskikh@oracle.com>
Reviewed-by: Daniel Kiper <daniel.kiper@oracle.com>
In the current code search_net_devices() uses the "alloc-mem" command
from the IEEE1275 User Interface for allocation of the transmit buffer
for the case when GRUB_IEEE1275_FLAG_VIRT_TO_REAL_BROKEN is set.
I don't have hardware where this flag is set to verify if this
workaround is still needed. However, further changes to ofnet will
require to execute this workaround one more time. Therefore, to
avoid possible duplication of code I'm moving this piece of
code into a function.
Signed-off-by: Stanislav Kholmanskikh <stanislav.kholmanskikh@oracle.com>
Reviewed-by: Daniel Kiper <daniel.kiper@oracle.com>
We only support FDT files with EFI on arm and arm64 systems, not
on x86. So move the helper that finds a prepopulated FDT UUID
into its own file and only build it for architectures where it
also gets called.
Signed-off-by: Alexander Graf <agraf@suse.de>
Reviewed-by: Daniel Kiper <daniel.kiper@oracle.com>
Network boot autoconfiguration sets default server to next server IP
(siaddr) from BOOTP/DHCP reply, but manual configuration using net_bootp
exports only server name. Unfortunately semantic of server name is not
clearly defined. BOOTP RFC 951 defines it only for client request, and
DHCP RFC 1541 only mentions it, without any implied usage. It looks like
this field is mostly empty in server replies.
Export next server IP as net_<interface>_next_server variable. This allows
grub configuration script to set $root/$prefix based on information obtained
by net_bootp.
Reported and tested by: Nikunj A Dadhania <nikunj@linux.vnet.ibm.com>
Cc: nikunj@linux.vnet.ibm.com
v2: change variable name to net_<interface>_next_server as discussed on the list
Searching for a device tree that EFI passes to us via configuration tables
is nothing architecture specific. Move it into generic code.
Signed-off-by: Alexander Graf <agraf@suse.de>
Reviewed-by: Daniel Kiper <daniel.kiper@oracle.com>
grub_util_get_dm_abstraction() does a string comparison of insufficient
length. When using a UUID such as "CRYPT-PLAIN-sda6_crypt", the function
returns GRUB_DEV_ABSTRACTION_LUKS.
This results in the error:
./grub-probe: error: disk `cryptouuid/sda6_crypt' not found.
This appears to be a copy/paste error introduced in:
a10e7a5a89
The bug was (apparently) latent until revealed by:
3bca85b418
Signed-off-by: Corey Hickey <bugfood-c@fatooh.org>
Modern pvops linux kernels support a p2m list not covered by the
kernel mapping. This capability is flagged by an elf-note specifying
the virtual address the kernel is expecting the p2m list to be mapped
to.
In case the elf-note is set by the kernel don't place the p2m list
into the kernel mapping, but map it to the given address. This will
allow to support domains with larger memory, as the kernel mapping is
limited to 2GB and a domain with huge memory in the TB range will have
a p2m list larger than this.
Signed-off-by: Juergen Gross <jgross@suse.com>
Reviewed-by: Daniel Kiper <daniel.kiper@oracle.com>
Modify the page table construction to allow multiple virtual regions
to be mapped. This is done as preparation for removing the p2m list
from the initial kernel mapping in order to support huge pv domains.
This allows a cleaner approach for mapping the relocator page by
using this capability.
The interface to the assembler level of the relocator has to be changed
in order to be able to process multiple page table areas.
Signed-off-by: Juergen Gross <jgross@suse.com>
Reviewed-by: Daniel Kiper <daniel.kiper@oracle.com>
Modern pvops linux kernels support an initrd not covered by the initial
mapping. This capability is flagged by an elf-note.
In case the elf-note is set by the kernel don't place the initrd into
the initial mapping. This will allow to load larger initrds and/or
support domains with larger memory, as the initial mapping is limited
to 2GB and it is containing the p2m list.
Signed-off-by: Juergen Gross <jgross@suse.com>
Reviewed-by: Daniel Kiper <daniel.kiper@oracle.com>
Do the allocation of page tables in a separate function. This will
allow to do the allocation at different times of the boot preparations
depending on the features the kernel is supporting.
Signed-off-by: Juergen Gross <jgross@suse.com>
Reviewed-by: Daniel Kiper <daniel.kiper@oracle.com>
Do the allocation of special pages (start info, console and xenbus
ring buffers) in a separate function. This will allow to do the
allocation at different times of the boot preparations depending on
the features the kernel is supporting.
Signed-off-by: Juergen Gross <jgross@suse.com>
Reviewed-by: Daniel Kiper <daniel.kiper@oracle.com>
Do the p2m list allocation of the to be loaded kernel in a separate
function. This will allow doing the p2m list allocation at different
times of the boot preparations depending on the features the kernel
is supporting.
While at this remove superfluous setting of first_p2m_pfn and
nr_p2m_frames as those are needed only in case of the p2m list not
being mapped by the initial kernel mapping.
Signed-off-by: Juergen Gross <jgross@suse.com>
Reviewed-by: Daniel Kiper <daniel.kiper@oracle.com>
Various features and parameters of a pv-kernel are specified via
elf notes in the kernel image. Those notes are part of the interface
between the Xen hypervisor and the kernel.
Instead of using num,bers in the code when interpreting the elf notes
make use of the header supplied by Xen for that purpose.
Signed-off-by: Juergen Gross <jgross@suse.com>
Reviewed-by: Daniel Kiper <daniel.kiper@oracle.com>
The loader for xen paravirtualized environment is using lots of global
variables. Reduce the number by making them either local or by putting
them into a single state structure.
Signed-off-by: Juergen Gross <jgross@suse.com>
Reviewed-by: Daniel Kiper <daniel.kiper@oracle.com>
When loading a Xen pv-kernel avoid memory leaks in case of errors.
Signed-off-by: Juergen Gross <jgross@suse.com>
Reviewed-by: Daniel Kiper <daniel.kiper@oracle.com>
The loader for xen paravirtualized environment isn't callable multiple
times as it won't free any memory in case of failure.
Call grub_relocator_unload() as other modules do it before allocating
a new relocator or when unloading the module.
Signed-off-by: Juergen Gross <jgross@suse.com>
Reviewed-by: Daniel Kiper <daniel.kiper@oracle.com>
Currently multiboot2 protocol loads image exactly at address specified in
ELF or multiboot2 header. This solution works quite well on legacy BIOS
platforms. It is possible because memory regions are placed at predictable
addresses (though I was not able to find any spec which says that it is
strong requirement, so, it looks that it is just a goodwill of hardware
designers). However, EFI platforms are more volatile. Even if required
memory regions live at specific addresses then they are sometimes simply
not free (e.g. used by boot/runtime services on Dell PowerEdge R820 and
OVMF). This means that you are not able to just set up final image
destination on build time. You have to provide method to relocate image
contents to real load address which is usually different than load address
specified in ELF and multiboot2 headers.
This patch provides all needed machinery to do self relocation in image code.
First of all GRUB2 reads min_addr (min. load addr), max_addr (max. load addr),
align (required image alignment), preference (it says which memory regions are
preferred by image, e.g. none, low, high) from multiboot_header_tag_relocatable
header tag contained in binary (at this stage load addresses from multiboot2
and/or ELF headers are ignored). Later loader tries to fulfill request (not only
that one) and if it succeeds then it informs image about real load address via
multiboot_tag_load_base_addr tag. At this stage GRUB2 role is finished. Starting
from now executable must cope with relocations itself using whole static and
dynamic knowledge provided by boot loader.
This patch does not provide functionality which could do relocations using
ELF relocation data. However, I was asked by Konrad Rzeszutek Wilk and Vladimir
'phcoder' Serbinenko to investigate that thing. It looks that relevant machinery
could be added to existing code (including this patch) without huge effort.
Additionally, ELF relocation could live in parallel with self relocation provided
by this patch. However, during research I realized that first of all we should
establish the details how ELF relocatable image should look like and how it should
be build. At least to build proper test/example files.
So, this patch just provides support for self relocatable images. If ELF file
with relocs is loaded then GRUB2 complains loudly and ignores it. Support for
such files will be added later.
This patch was tested with Xen image which uses that functionality. However, this Xen
feature is still under development and new patchset will be released in about 2-3 weeks.
Signed-off-by: Daniel Kiper <daniel.kiper@oracle.com>
Reviewed-by: Vladimir Serbinenko <phcoder@gmail.com>
If image requested EFI boot services then skip multiboot2 memory maps.
Main reason for not providing maps is because they will likely be
invalid. We do a few allocations after filling them, e.g. for relocator
needs. Usually we do not care as we would have finished boot services.
If we keep boot services then it is easier/safer to not provide maps.
However, if image needs memory maps and they are not provided by bootloader
then it should get itself just before ExitBootServices() call.
Signed-off-by: Daniel Kiper <daniel.kiper@oracle.com>
Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Reviewed-by: Vladimir Serbinenko <phcoder@gmail.com>
Add tags used to pass ImageHandle to loaded image if requested.
It is used by at least ExitBootServices() function.
Signed-off-by: Daniel Kiper <daniel.kiper@oracle.com>
Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Reviewed-by: Vladimir Serbinenko <phcoder@gmail.com>
Add grub_relocator64_efi relocator. It will be used on EFI 64-bit platforms
when multiboot2 compatible image requests MULTIBOOT_TAG_TYPE_EFI_BS. Relocator
will set lower parts of %rax and %rbx accordingly to multiboot2 specification.
On the other hand processor mode, just before jumping into loaded image, will
be set accordingly to Unified Extensible Firmware Interface Specification,
Version 2.4 Errata B, section 2.3.4, x64 Platforms, boot services. This way
loaded image will be able to use EFI boot services without any issues.
Signed-off-by: Daniel Kiper <daniel.kiper@oracle.com>
Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Reviewed-by: Vladimir Serbinenko <phcoder@gmail.com>
limit_time underflows when current time is less than 90000ms.
This causes packet fragments received during this time, i.e.,
till 90000ms pass since timer init, to be rejected.
Hence, set it to 0 if its less than 90000.
Signed-off-by: Sakar Arora <Sakar.Arora@nxp.com>
Writing the primary GPT before the backup may lead to a confusing
situation: booting a freshly updated system could consistently fail and
next boot will fall back to the old system if writing the primary works
but writing the backup fails. If the backup is written first and fails
the primary is left in the old state so the next boot will re-try and
possibly fail in the exact same way. Making that repeatable should make
it easier for users to identify the error.
Additionally if the firmware and OS disagree on the disk size, making
the backup inaccessible to GRUB, then just skip writing the backup.
When this happens the automatic call to `coreos-setgoodroot` after boot
will take care of repairing the backup.
The firmware and the OS may disagree on the disk configuration and size.
Although such a setup should be avoided users are unlikely to know about
the problem, assuming everything behaves like the OS. Tolerate this as
best we can and trust the reported on-disk location over the firmware
when looking for the backup GPT. If the location is inaccessible report
the error as best we can and move on.
I personally think this reads easier. Also has the side effect of
directly comparing the primary and backup tables instead of presuming
they are equal if the crc32 matches.
This ensures all code modifying GPT data include the same sanity check
that repair does. If revalidation fails the status flags are left in the
appropriate state.
The header was being relocated without checking the new location is
actually safe. If the BIOS thinks the disk is smaller than the OS then
repair may relocate the header into allocated space, failing the final
validation check. So only move it if the disk has grown.
Additionally, if the backup is valid then we can assume its current
location is good enough and leave it as-is.
Use the new status function which checks *_HEADER_VALID and
*_ENTRIES_VALID bits together. It doesn't make sense for the header and
entries bits to mismatch so don't allow for it.
Portions of the code attempted to handle the fact that GPT entries on
disk may be larger than the currently defined struct while others
assumed the data could be indexed by the struct size directly. This
never came up because no utility uses a size larger than 128 bytes but
for the sake of safety we need to do this by the spec.
GPT_BOTH_VALID is 4 bits so simple a boolean check is not sufficient.
This broken condition allowed gptprio to trust bogus disk locations in
headers that were marked invalid causing arbitrary disk corruption.
We may get more than one response before exiting out of loop in
grub_net_dns_lookup, but buffer was allocated for the first response only,
so storing answers from subsequent replies wrote past allocated size.
We never really use more than the very first address during lookup so there
is little point in collecting all of them. Just quit early if we already have
some reply.
Code needs serious redesign to actually collect multiple answers
and select the best fit according to requested type (IPv4 or IPv6).
Reported and tested by Michael Chang <mchang@suse.com>
XFS V5 stores UUID in metadata and compares them with superblock UUID.
To allow changing of user-visible UUID it stores original value in new
superblock field (meta_uuid) and sets incompatible flag to indicate that
new field must be used to verify metadata. Our driver currently does not
check metadata UUID so simply accept such filesystem.
Reported-By: Marcos Mello <marcosfrm@outlook.com>
Reviewd by Jan Kara <jack@suse.cz>
GRUB assumes that no disk is ever larger than 1EiB and rejects
reads/writes to such locations. Unfortunately this is not conveyed in
the usual way with the special GRUB_DISK_SIZE_UNKNOWN value.
send client arch in bootp requests, for now BIOS and x64/aarch64 EFI is
supported.
fix a bug introduced in 4d5d7be005 where
user class was encoded improperly, although this didn't seem to have any
detrimental effects.
properly insert an option terminator.
If we open new connection, we need to reset stall indication, otherwise
nothing will ever be polled (low level code rely on this field being
zero when establishing connection).
Depending on the OS/libc, device macros are defined in different
headers. This change ensures we include the right one.
sys/types.h - BSD
sys/mkdev.h - Sun
sys/sysmacros.h - glibc (Linux)
glibc currently pulls sys/sysmacros.h into sys/types.h, but this may
change in a future release.
https://sourceware.org/ml/libc-alpha/2015-11/msg00253.html
At least the apache sever is very unhappy with that extra null line and will
take more than ten seconds in responding to each range request, which slows
down a lot the entire http file transfer process or even time out.
Some DHCP servers (such as dnsmasq) tokenise parameters with commas, making
it impossible to pass boot files with commas in them. Allow using a semicolon
to separate the protocol from host if a comma wasn't found.
The current logic in the DNS resolution code allocates an address buffer
based on the number of addresses in the response packet. If we receive
multiple response packets in response to a single query packet, this means
that we will reallocate a new buffer large enough for only the addresses in
that specific packet, discarding any previous results in the process. Worse,
we still keep track of the *total* number of addresses resolved in response
to this query, not merely the number in the packet being currently processed.
Use realloc() rather than malloc() to avoid overwriting the existing data,
and allocate a buffer large enough for the total set of addresses rather
than merely the number in this specific response.
Rework TPM measurements to use fewer PCRs. After discussion with upstream,
it's preferable to avoid using so many PCRs. Instead, measure into PCRs 8
and 9 but use a prefix in the event log to indicate which subsystem carried
out the measurements.
Historically this variable hold previous value of filename that
had to be freed if allocated previously. Currently this branch
is entered only if filename was not allocated previously so it
became redundant. It did not cause real problems because grub_free
was not called, but code is confusing and causes compilation error
in some cases.
Returned from the OpenProtocol operation, the grub_efi_block_io_media
structure contains the io_align field, specifying the minimum alignment
required for buffers used in any data transfers with the device.
Make grub_efidisk_readwrite() allocate a temporary buffer, aligned to
this boundary, if the buffer passed to it does not already meet the
requirements.
Also sanity check the io_align field in grub_efidisk_open() for
power-of-two-ness and bail if invalid.
Map EFI_NO_MEDIA to GRUB_ERR_OUT_OF_RANGE that is ignored by diskfilter. This
actually matches pretty close (we obviously attempt to read outside of media)
and avoids adding more error codes.
This affects only internally initiated scans. If read/write from removable is
explicitly requested, we still return an error and text explanation is more
clear for user than generic error.
Reported and tested by Andreas Loew <Andreas.Loew@gmx.net>
ipv6 routing in grub2 is broken, we cannot talk to anything outside our local
network or anything that doesn't route in our global namespace. This patch
fixes this by doing a couple of things
1) Read the router information off of the router advertisement. If we have a
router lifetime we need to take the source address and create a route from it.
2) Changes the routing stuff slightly to allow you to specify a gateway _and_ an
interface. Since the router advertisements come in on the link local address we
need to associate it with the global address on the card. So when we are
processing the router advertisement, either use the SLAAC interface we create
and add the route to that interface, or loop through the global addresses we
currently have on our interface and associate it with one of those addresses.
We need to have a special case here for the default route so that it gets used,
we do this by setting the masksize to 0 to mean it encompasses all networks.
The routing code will automatically select the best route so if there is a
closer match we will use that.
With this patch I can now talk to ipv6 addresses outside of my local network.
Thanks,
Signed-off-by: Josef Bacik <jbacik@fb.com>
Prevent buffer over-read in grub_machine_mmap_iterate. This was
causing phys_base from being calculated properly. This then
caused the wrong value to be placed in ramdisk_image within
struct linux_hdrs. Which prevented the ramdisk from loading on
boot.
Newer SPARC systems contain more than 8 available memory entries.
For example on a T5-8 with 2TB of memory, the memory layout could
look like this:
T5-8 Memory
reg 00000000 30000000 0000003f b0000000
00000800 00000000 00000040 00000000
00001000 00000000 00000040 00000000
00001800 00000000 00000040 00000000
00002000 00000000 00000040 00000000
00002800 00000000 00000040 00000000
00003000 00000000 00000040 00000000
00003800 00000000 00000040 00000000
available 00003800 00000000 0000003f ffcae000
00003000 00000000 00000040 00000000
00002800 00000000 00000040 00000000
00002000 00000000 00000040 00000000
00001800 00000000 00000040 00000000
00001000 00000000 00000040 00000000
00000800 00000000 00000040 00000000
00000000 70000000 0000003f 70000000
00000000 6eef8000 00000000 00002000
00000000 30400000 00000000 3eaf6000
name memory
Signed-off-by: Eric Snowberg <eric.snowberg@oracle.com>
When running grub in a VGA console of a KVM pseries guest on PowerPC,
you can see the cursor sweeping over the whole line when entering a
character in editor mode. This is visible because grub always refreshes
the whole line when entering a character in editor mode, and drawing
characters is quite a slow operation with the firmware used for the
powerpc pseries guests (SLOF).
To avoid this ugliness, the cursor should be disabled when refreshing
the screen contents during update_screen().
Signed-off-by: Thomas Huth <thuth@redhat.com>
lexer calls yylex_fatal on fatal internal errors. yylex_fatal itself is
declared as noreturn and calls exit. Returning from noreturn function has
unpredictable consequences.
all_video module does not have any code or data and exists solely for
.moddeps section to pull in dependencies. This makes all symbols unneeded.
While in current binutils (last released version as of this commit is 2.26)
``strip --strip-unneeded'' unintentionally adds section symbols for each
existing section, this behavior was considered a bug and changed in commit
14f2c699ddca1e2f706342dffc59a6c7e23e844c to completely strip symbol table
in this case.
Older binutils (verified with 2.17) and some other toolchains (at least
elftoolchain r3223M), both used in FreeBSD, remove symbol table in all_video
as well.
Relax run-time check and do not return error for modules without symbol table.
Add additional checks to module verifier to make sure such modules
a) have non-empty .moddeps section. Without either externally visible symbols
or .moddeps modules are completely useless and should not be built.
b) do not have any relocations.
Closes: 46986
v2: add run-time check for empty symbol table if relocations are present as
suggested by Vladimir.
-Wunreachable-code has been a no-op since GCC 4.5; GRUB hasn't been
compiled with it since 2012; and GCC 6 produces "error:
'-Wunreachable-code' is not an option that controls warnings" for these.
Fixes Debian bug #812047.
