stage3 still contains a legacy fallback path that transfers control to
the new kernel via LPSW 0 when diag 0x308 is not available.
The current start_kernel() inline asm issues "diag %[code],%[code],0x308"
without enforcing the required even/odd register pairing for the first
operand. With recent distro toolchains (e.g. GCC 15.x as shipped by Fedora
43 and Ubuntu 25.10), register allocation picks an odd register for the
first diag operand in start_kernel(), which triggers a specification
exception. That exception is then caught by the program-check handler
and stage3 silently takes the fallback, booting the kernel via LPSW 0
instead of performing diag308 LOAD_NORMAL_RESET.
All supported environments are expected to provide diag308, and the rest
of the code already relies on it. Drop start_kernel() and the obsolete
LPSW fallback and always reset into the new kernel via the properly
coded diag308() helper. Mark start() as __noreturn.
Reviewed-by: Eduard Shishkin <edward6@linux.ibm.com>
Tested-by: Eduard Shishkin <edward6@linux.ibm.com>
Signed-off-by: Vasily Gorbik <gor@linux.ibm.com>
Signed-off-by: Jan Höppner <hoeppner@linux.ibm.com>
Fix incorrect handling of tape devices leading to inability of
creating dumps on them.
Make the check for tape device go first, to not miss it on irrelevant
errors
Reported-by: Mikhail Zaslonko <zaslonko@linux.ibm.com>
Reviewed-by: Mikhail Zaslonko <zaslonko@linux.ibm.com>
Signed-off-by: Eduard Shishkin <edward6@linux.ibm.com>
Signed-off-by: Jan Höppner <hoeppner@linux.ibm.com>
Fix two issues in boot menu input parsing:
1. ebcdic_strtoul returns unsigned long but the value was stored in an int.
2. ebcdic_strtoul could overflow if @value exceeds ULONG_MAX.
Both problems are easy to trigger by entering an excessively large value
in the boot menu, which can lead to unsigned long overflow and memory
corruption.
Use a checked addition to prevent overflow and change menu_read() return
type to unsigned long.
Suggested-by: Eduard Shishkin <edward6@linux.ibm.com>
Reviewed-by: Eduard Shishkin <edward6@linux.ibm.com>
Signed-off-by: Marc Hartmayer <mhartmay@linux.ibm.com>
Signed-off-by: Jan Höppner <hoeppner@linux.ibm.com>
Fix two off-by-one errors:
+ One caused an out-of-bounds read when the loadparm value was 63
+ The other made a boot entry unselectable.
Currently, these bugs have no practical impact because:
1. The memory area beyond __stage2_params.config[63] is empty.
2. BOOT_MENU_ENTRIES is set to 63, which exceeds the number of boot menu
entries that can be written to disk.
Reviewed-by: Mikhail Zaslonko <zaslonko@linux.ibm.com>
Reviewed-by: Eduard Shishkin <edward6@linux.ibm.com>
Signed-off-by: Marc Hartmayer <mhartmay@linux.ibm.com>
Signed-off-by: Steffen Eiden <seiden@linux.ibm.com>
When a loadparm is presented by the user it's not checked whether it's
too large. Add this missing check to fix an out-of-bounds read.
The problem did not show up because the area behind the buffer is empty.
Reviewed-by: Steffen Eiden <seiden@linux.ibm.com>
Reviewed-by: Eduard Shishkin <edward6@linux.ibm.com>
Signed-off-by: Marc Hartmayer <mhartmay@linux.ibm.com>
Signed-off-by: Steffen Eiden <seiden@linux.ibm.com>
Currently, 'sclp_setup(SCLP_INIT)' fails if no SCLP line-mode console is
available. As a result 'menu_param()' is never called, even though it is
required to retrieve the 'loadparm' value.
However, reading the loadparm via SCLP remains useful even when a SCLP
line-mode console is absent, because this value determines which boot
entry should be selected. Therefore, the boot process should continue by
retrieving the loadparm without requiring an SCLP line-mode console.
It's safe to continue without a SCLP console as 'printf' and
'menu_param' tolerates the absence of a SCLP line-mode console.
Fixes: https://github.com/ibm-s390-linux/s390-tools/issues/196
Reported-by: Thomas Huth <thuth@redhat.com>
Reviewed-by: Steffen Eiden <seiden@linux.ibm.com>
Reviewed-by: Eduard Shishkin <edward6@linux.ibm.com>
Signed-off-by: Marc Hartmayer <mhartmay@linux.ibm.com>
Signed-off-by: Steffen Eiden <seiden@linux.ibm.com>
Output all text input sections into one text section and map all input
sections .bss.* .text.*, .rodata.*, and .data.* as it's done by the
default linker script [1]. In addition, make the linker script easier to
read by replacing the magic value of 32 with 'SIZEOF(.sb.trailer)'.
[1] Check the output of 'ld --verbose'.
Reviewed-by: Mikhail Zaslonko <zaslonko@linux.ibm.com>
Reviewed-by: Stefan Haberland <sth@linux.ibm.com>
Signed-off-by: Marc Hartmayer <mhartmay@linux.ibm.com>
Signed-off-by: Jan Höppner <hoeppner@linux.ibm.com>
Although stage2.head is not directly used in stage2, it is required by
other stages. To ensure it is retained during link-time
optimization (LTO), explicitly mark it as used. This improves code
readability and prevents LTO from mistakenly removing it.
Reviewed-by: Mikhail Zaslonko <zaslonko@linux.ibm.com>
Reviewed-by: Stefan Haberland <sth@linux.ibm.com>
Signed-off-by: Marc Hartmayer <mhartmay@linux.ibm.com>
Signed-off-by: Jan Höppner <hoeppner@linux.ibm.com>
Rename .exec into .elf because they are ELF files, no reason to use a
different file extension.
This change improves the shell's autocompletion of tools such as
'readelf'.
Acked-by: Mikhail Zaslonko <zaslonko@linux.ibm.com>
Reviewed-by: Stefan Haberland <sth@linux.ibm.com>
Signed-off-by: Marc Hartmayer <mhartmay@linux.ibm.com>
Signed-off-by: Jan Höppner <hoeppner@linux.ibm.com>
The definition NO_PIE_CFLAGS has only two users and specifies only a
single compiler flag. There is no point having this definition anymore.
Remove the definition and specify -fno-pie directly for the two users.
Reviewed-by: Marc Hartmayer <mhartmay@linux.ibm.com>
Signed-off-by: Jan Höppner <hoeppner@linux.ibm.com>
Commit c5a91199e3 ("zipl: Always build and link without PIE.")
introduced -fno-pie (for compilation) and -no-pie (for linking) for
zipl. At the time the linker was still used directly before commit
5e46632767 ("zipl: Use the compiler for linking instead of ld")
eventually switched to calling the compiler for the linking step.
During that adaption -static was introduced to the linker flags. -no-pie
was carried over as well. However -static implies -no-pie and it is
therefore not required. For GCC see also man 1 gcc (-static) [1]. Whilst
not explicitly documented, Clang shows the same behaviour. Clang also
complains when -static and -no-pie are specified in the linker step at
the same time with the following warning:
clang: warning: argument unused during compilation: '-no-pie' [-Wunused-command-line-argument]
Since -no-pie is not required, remove it and get rid of the warning.
[1] https://gcc.gnu.org/onlinedocs/gcc/Link-Options.html#index-static
Reviewed-by: Marc Hartmayer <mhartmay@linux.ibm.com>
Signed-off-by: Jan Höppner <hoeppner@linux.ibm.com>
The code within 'menu_param()' previously assumed that a read from sclp
will never fail.
If 'sclp_param()' fails then 'endptr' is never initialised and
'loadparm' is compared with 'endptr' which is undefined behvaiour.
If a sclp read fails, an undefined 'endptr' is never accessed, and upon
a failed read, will return a new error code 'SCLP_ERROR' instead of
returning 'NUMBER_FOUND' which is incorrect logic wise.
Remove compare conditions and assignments of 0 in 'value', as 'value'
is initialised with 0 ('DEFAULT_MENU_ENTRY') and cannot be non zero,
only in the case where a number is found and we go to boot.
Logic:
Check if we got a number and boot from it.
If 'PRINT_PROMPT', break out to menu print logic.
If an 'SCLP_ERROR' occurs, print an error message and boot the default
since 'value' is initialised with 'DEFAULT_MENU_ENTRY'.
If 'NOTHING_FOUND', check if the menu is disabled. If disabled, go to
default boot. Otherwise break out to print logic.
Reviewed-by: Marc Hartmayer <mhartmay@linux.ibm.com>
Reviewed-by: Steffen Eiden <seiden@linux.ibm.com>
Signed-off-by: Richie Buturla <richie@linux.ibm.com>
Signed-off-by: Jan Höppner <hoeppner@linux.ibm.com>
The 'temp_area' buffer was not zeroed out for a retry in case of an
invalid input to select a boot menu entry.
Before:
zIPL v2.38.0-build-20250822 interactive boot menu
0. default (1TTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTT...
1. 1TTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTT...
2. 2TTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTT...
3. 3TTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTT...
Note: VM users please use '#cp vi vmsg <input>'
Please choose (default will boot in 3 seconds):111
Error: undefined configuration
Please choose:1
Error: undefined configuration
After:
zIPL v2.38.0-build-20250822 interactive boot menu
0. default (1TTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTT...
1. 1TTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTT...
2. 2TTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTT...
3. 3TTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTT...
Note: VM users please use '#cp vi vmsg <input>'
Please choose (default will boot in 3 seconds):111
Error: undefined configuration
Please choose:1
Booting 1TTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTTT...
Acked-by: Eduard Shishkin <edward6@linux.ibm.com>
Acked-by: Mikhail Zaslonko <zaslonko@linux.ibm.com>
Signed-off-by: Marc Hartmayer <mhartmay@linux.ibm.com>
Signed-off-by: Steffen Eiden <seiden@linux.ibm.com>
Fix a design mistake leading to IPL programs corruption.
Use a matrix of component ranges to locate the added components in the
body of bootmap file instead of an array.
Earlier an array of NR_PROGRAM_COMPONENTS was used, which is incorrect
Now a matrix of (NR_PROGRAM_COMPONENTS X BOOT_MENU_ENTRIES) is used.
Don't duplicate environment block for each menu entry. Instead, reuse
the one that was added when processing the first menu entry.
Tested-by: Alexander Egorenkov <egorenar@linux.ibm.com>
Reported-by: Alexander Egorenkov <egorenar@linux.ibm.com>
Signed-off-by: Eduard Shishkin <edward6@linux.ibm.com>
Signed-off-by: Steffen Eiden <seiden@linux.ibm.com>
Apply previously introduced generic infrastructure to add an
environment block as a "buffer component".
Signed-off-by: Eduard Shishkin <edward6@linux.ibm.com>
Reviewed-by: Stefan Haberland <sth@linux.ibm.com>
Signed-off-by: Steffen Eiden <seiden@linux.ibm.com>
Don't duplicate data of components added via add_component_buffer()
and friends to bootmap file. Instead, reuse data that were previously
added when preparing a program table for the first mirror to create
metadata (block lists, program tables, etc) specific for other mirrors.
Signed-off-by: Eduard Shishkin <edward6@linux.ibm.com>
Reviewed-by: Stefan Haberland <sth@linux.ibm.com>
Signed-off-by: Steffen Eiden <seiden@linux.ibm.com>
If the option '--add-files' is specified, don't duplicate data of
components added via add_component_file() and friends to bootmap
file for each mirror. Instead, reuse the data that were added when
preparing a program table for the first mirror to create metadata
(block lists, program tables, etc) specific for other mirrors.
Signed-off-by: Eduard Shishkin <edward6@linux.ibm.com>
Reviewed-by: Stefan Haberland <sth@linux.ibm.com>
Signed-off-by: Steffen Eiden <seiden@linux.ibm.com>
Drop limitations on identical target parameters of base disks per
logical device;
Fix verbose zipl output to include geometry of each mirror and
component load addresses that would be used when booting from each
mirror.
Signed-off-by: Eduard Shishkin <edward6@linux.ibm.com>
Reviewed-by: Stefan Haberland <sth@linux.ibm.com>
Signed-off-by: Steffen Eiden <seiden@linux.ibm.com>
Replace struct install_set with a new type containing multiple
program tables and program components. Refactor the code accordingly.
When retrieving device INFO, for each mirror complete a respective
structure disk_info in the INFO.BASE array.
When building a bootmap, for each mirror create a separate program
table (or a pair of tables in case of DASD - for CCW-TYPE IPL and
for LD-IPL), using respective components in the arrays INFO.BASE and
BIS.MIRRORS
Make data of program components added via get_component_buffer() and
get_component_file() be duplicated per each mirror.
Make boot record on each mirror point out to a respective program
table in the bootmap (when booting from different mirrors, different
program tables in the bootmap are used).
This patch doesn't make functional changes. However, test cases
comparing boot meta-data dumps of different mirrors may fail (since
boot records on different mirrors now refer different copies of boot
data). This will be fixed by the next patches in the series which
allow boot data to be reused.
Signed-off-by: Eduard Shishkin <edward6@linux.ibm.com>
Reviewed-by: Stefan Haberland <sth@linux.ibm.com>
Signed-off-by: Steffen Eiden <seiden@linux.ibm.com>
Replace struct disk_info with new 2 types.
Old type:
struct disk_info {
A; /* logical device info */
B; /* basedisk info */
};
New types:
struct disk_info {
B; /* basedisk info */
};
and
struct device_info {
A; /* logical device info */
struct disk_info C [MAX_TARGETS]; /* array of base disks */
};
Here A (logical device info) is the following:
dev_t device; /* logical device for bootmap creation */
char *name; /* name of logical device as reetrieved from
"/proc/partitions" */
char *drv_name; /* name of the driver managing the logical
device as retrieved from "/proc/devices",
or evaluated */
int fs_block_size;
Refactor the code respectively, to use only the first element of
the array C, so that this patch represents an equivalent transform.
Signed-off-by: Eduard Shishkin <edward6@linux.ibm.com>
Reviewed-by: Stefan Haberland <sth@linux.ibm.com>
Signed-off-by: Steffen Eiden <seiden@linux.ibm.com>
Fix segmentation fault when trying to process not supported dm-targets
Release allocated memory in error paths
Reviewed-by: Alexander Egorenkov <egorenar@linux.ibm.com>
Signed-off-by: Eduard Shishkin <edward6@linux.ibm.com>
Signed-off-by: Steffen Eiden <seiden@linux.ibm.com>
Fix is_device_mapper() predicate to not base on checking a hardcoded
major number (253), which not always correct, since on some systems
dm-devices have different majors".
Reviewed-by: Stefan Haberland <sth@linux.ibm.com>
Signed-off-by: Eduard Shishkin <edward6@linux.ibm.com>
Signed-off-by: Jan Höppner <hoeppner@linux.ibm.com>
Add support for complex dm- over md-devices.
It includes an important particular use case in distro installation
process, when a linear dm-target is built on the top of software
RAID1:
dasda 94:0 0 20.6G 0 disk
\_dasda1 94:1 0 20G 0 part
\_md127 9:127 0 20G 0 raid1
\_rhel_a46lp05-root00 253:5 0 20G 0 lvm /
dasdb 94:4 0 20.6G 0 disk
\_dasdb1 94:5 0 20G 0 part
\_md127 9:127 0 20G 0 raid1
\_rhel_a46lp05-root00 253:5 0 20G 0 lvm /
\ # zipl_helper.device-mapper /
Expected result:
targetbase=94:0
targettype=CDL
targetgeometry=30051,15,12
targetblocksize=4096
targetoffset=4632
targetbase=94:4
targettype=CDL
targetgeometry=30051,15,12
targetblocksize=4096
targetoffset=4632
Actual result (incorrect):
targetbase=9:127
targettype=SCSI
targetblocksize=4096
targetoffset=256
Reviewed-by: Stefan Haberland <sth@linux.ibm.com>
Signed-off-by: Eduard Shishkin <edward6@linux.ibm.com>
Signed-off-by: Jan Höppner <hoeppner@linux.ibm.com>
Since DASD standalone dumper does not support non-extended s390
multi-volume dumps for years, drop zgetdump support of non-extended s390_mv
DASD dumps and dump-tools:
- Merge dfi_s390mv_ext and dt_s390mv_ext with related dfi_s390mv and
dt_s390mv counterparts.
- Update DFI and DT vectors of supported dump formats and dump-tools
removing dfi_s390mv and dt_s390mv entries and keeping dt_s390mv_ext and
dfi_s390mv_ext ones for s390_mv extended dump and dump-tool accordingly.
- Remove dfi_s390mv.c and dt_s390mv.c sources and update the Makefile.
- Make dfi_s390mv_ext_init() and dfi_s390mv_info() static.
- Remove magic number relevant to non-extended s390_mv DASD dumper.
- Completely drop DASD dumper version 5 case in df_s390_dumper_read().
Note: Since non-extended s390 multi-volume dumps can reside on DASD
partitions only and cannot be produced via DFO interface, we can drop this
dump format entirely.
Signed-off-by: Mikhail Zaslonko <zaslonko@linux.ibm.com>
Reviewed-by: Alexander Egorenkov <egorenar@linux.ibm.com>
Signed-off-by: Jan Höppner <hoeppner@linux.ibm.com>
Since DASD standalone dumper does not support non-extended s390 dump
format for years, drop zgetdump support of non-extended s390 single volume
DASD dump-tool:
- Merge dt_390sv_ext.c and dt_s390sv.c counterparts.
- Update DT vector of supported dump tools removing dt_s390sv entry and
keeping dt_s390sv_ext entry for s390 single volume extended dump tool.
- Remove dt_s390sv.c source and update the Makefile accordingly.
- Remove magic constants relevant to non-extended s390 single volume DASD
dumper.
- Drop s390 single volume DASD dumper version 5 in df_s390_dumper_read().
Note: We still need support of non-extended s390 dump format since such
output dump files can be produced by 'zgetdump -f s390' via dfo_s390. Thus,
both dfi_s390 and dfi_s390_ext DFI vector entries remain in order to
process s390 dumps files as well as s390_ext dumps on the DASD partition.
Signed-off-by: Mikhail Zaslonko <zaslonko@linux.ibm.com>
Reviewed-by: Alexander Egorenkov <egorenar@linux.ibm.com>
Signed-off-by: Jan Höppner <hoeppner@linux.ibm.com>
Drop support of obsolete dump and dump-tool versions (single-volume DASD,
FBA and Tape) in order to simplify zgetdump logic:
- CCW dumpers written in assembler instructions as well as stage2 dumpers
of size less than 0x3000 (STAGE2_DUMPER_SIZE_V1 or STAGE2_DUMPER_SIZE_V2)
haven't been used for years. Remove its traces completely as a cleanup.
Keep the last version (version 5) of non-extended DASD dumper as well as
newer extended DASD dumpers.
- Rename STAGE2_DUMPER_SIZE_V3 and STAGE2_DUMPER_SIZE_ZLIB constants.
- Drop support of non-extended s390 dumps of version < 5. Dump files
of s390 format version 5 can be still produced by zgetdump (dfo_s390).
- Drop excessive dump version checking in df_s390_cpu_info_add() and
df_s390_hdr_add() considering that obsolete s390 dumps of version lower
than 5 no longer supported.
- Use cpu_cnt field in s390 dump header instead of the s390 dump version to
indicate no cpu info available (DFI_CPU_CONTENT_NONE) for dfo_s390.
- Make df_s390_dumper_read() return error code upon unknown dumper
version/magic detection.
Signed-off-by: Mikhail Zaslonko <zaslonko@linux.ibm.com>
Reviewed-by: Alexander Egorenkov <egorenar@linux.ibm.com>
Signed-off-by: Jan Höppner <hoeppner@linux.ibm.com>
Use the directory specified by the shell environment variable TMPDIR
for temporary objects creation and ngdump job simulation. If TMPDIR is
not set, then use "/tmp" for the mentioned purposes;
While running ngdump job in 'dry-run' mode:
. Don't format/mount the target dump device. Instead, create the
bootmap file and the meta-file at the temporary mount point without
mounting anything to it. Thus, the mentioned files to be acrually
created in the "proxy" file system owning the temporary mount point;
. Retrieve base disk info from the read-only dump device and
complete that info with the block size of the proxy file system;
Separate the steps on retrieving/setting file system block size
into a dedicated procedure;
Use definitions instead of hardcoded file names;
Fix a bug in an error path (accessing freed memory);
Make misc_open_simulate() and misc_open_exclusive() static;
Update man pages with the requirements on the system environment
(resources) for ngdump job being executed in dry-run mode;
Provide hints for user (in stderr) in case when ngdump job in dry-run
mode failed due to inappropriate system environment.
Tested-by: Mikhail Zaslonko <zaslonko@linux.ibm.com>
Reviewed-by: Mikhail Zaslonko <zaslonko@linux.ibm.com>
Signed-off-by: Eduard Shishkin <edward6@linux.ibm.com>
Signed-off-by: Jan Höppner <hoeppner@linux.ibm.com>
Introduce misc_open_simulate() to mark individual files as "opened
for write simulation".
Introduce misc_open_device() to open a file either in "usual", or
in "simulation" mode, depending on the passed argument
Tested-by: Mikhail Zaslonko <zaslonko@linux.ibm.com>
Reviewed-by: Mikhail Zaslonko <zaslonko@linux.ibm.com>
Signed-off-by: Eduard Shishkin <edward6@linux.ibm.com>
Signed-off-by: Jan Höppner <hoeppner@linux.ibm.com>
Introduce a write simulation operation, which doesn't write actual
data and just updates the current position in the file.
This allows to emulate block lists for files located on raw devices
(not formatted with a file system). This is used by the next patches
in the series to implement '--dry-run' zipl option for dumps of all
types (not only ngdumps).
Introduce a 'misc file descriptor', which allows to mark individual
files as "opened in a simulation mode".
Whenever bulding a bootmap file, use either real write, or write
simulation depending on the mode set in the 'misc file descriptor'.
Tested-by: Mikhail Zaslonko <zaslonko@linux.ibm.com>
Reviewed-by: Mikhail Zaslonko <zaslonko@linux.ibm.com>
Signed-off-by: Eduard Shishkin <edward6@linux.ibm.com>
Signed-off-by: Jan Höppner <hoeppner@linux.ibm.com>
Update eckd2dump linker script to remove stage2 space constraints and
avoid linker section overlaps when using 'unfortunate' compiler
version & flags combination.
Since Multi-volume dumper size changes, we also update the dump-tool version
number for 'zgetdump' to distinguish.
Signed-off-by: Mikhail Zaslonko <zaslonko@linux.ibm.com>
Fixes: https://github.com/ibm-s390-linux/s390-tools/issues/171
Signed-off-by: Jan Höppner <hoeppner@linux.ibm.com>
