Add these simple debugging aids to the GNOME GitLab CI test job.
They've been needed before [1] so add them permanently.
[1] 57983b9fc2
Create block special devices needed by test_BlockSpecial in GitLab
CI jobs (!59)
"Contents of /proc/partitions inside the Docker image when this
test CI job failed:
...
And the listing of /dev/:
"
Closed!113 - Fix occasional GitLab CI test jobs failures on
BlockSpecial unit tests
STAT_FORMATTED is only used inside snap_to_mebibyte() to suppress
enforcement that partition boundaries must not overlay the MBR or EBRs
when merely formatting existing partitions. However since commit [1],
snap_to_mebibyte() is only called inside the dialogs composing Create
New, Copy / Paste into New and Resize / Move operations and never when
composing a Format operation or any other operation which doesn't change
partition boundaries. Therefore remove STAT_FORMATTED.
[1] 7c94b7d920
Snap partition boundaries before dialogs update FS usage (#48)
The available number of sectors 'total_length' is calculated as one
sector too few. However this doesn't matter because when composing a
new partition in Dialog_Partition_New::Get_New_Partition() the dialog
fits the end of the partition to the end of the unallocated partition
'new_partition->sector_end' in which it is being created, not the
available space. Missed in earlier commit:
4f84cff781
cleanups
Anyway correct the calculation.
As a partition's position is completely defined by it's starting and
ending sector, it is aligned if both boundaries are aligned. The size
of a partition just depends on the starting and ending sector.
Therefore it is never necessary to also check if the size of the
partition is also an exact multiple of MiB. Remove this unnecessary
portion of the alignment check when resizing/moving a partition (or
copy/pasting into a new partition).
Now Win_GParted::m_display_device.partitions is an identical copy of
Win_GParted::display_partitions with the same lifetime. That's wasteful
and pointless. Therefore remove the later and use the former in it's
place.
Closes#227 - Unable to allocate 1 MiB between partitions when moving to
the right
The Create New and Paste dialogs also create partitions and have to
honour currently composed partitions while doing so. Therefore they
must have a Device object containing the currently composed partition
layout for passing into snap_to_alignment() and below. So copy the
current Device object when refreshing the visual at the same time
visual_partitions is generated and use in all 3 dialogs which compose
new partitions.
Note that Create New and Paste aren't subject to the same bug as Resize/
Move was because the code in snap_to_mebibyte() [1] checked the
partition object being composed has status STAT_REAL. This is true for
partition objects created by the Resize/Move dialog, but not true for
the Create New and Paste dialogs which set status to STAT_NEW and
STAT_COPY respectively instead.
[1] Dialog_Base_Partition::snap_to_mebibyte() lines 418 to 438
https://gitlab.gnome.org/GNOME/gparted/-/blob/GPARTED_1_5_0/src/Dialog_Base_Partition.cc#L418Closes#227 - Unable to allocate 1 MiB between partitions when moving to
the right
Start with 2 partitions next to each other, containing file systems that
GParted can move and resize.
EG:
|[#1 ext2 ][#2 swap ] |
Move the start of partition #2 to the right. Then attempt to move the
end of partition #1 to the right to meet it.
EG:
|[#1 ext2 ][#2 swap ] |
The Resize/Move dialog will allow the free space following to be set to
0 so partition #1 is again adjacent to partition #2, but after closing
the dialog a forced 1 MiB gap is added, shrinking the composed size of
partition #1 by that 1 MiB.
If instead the first operation to shrink and move partition #2 is
applied, then partition #1 can be successfully resize right up to
partition #2 without a 1 MiB gap.
Relevant call sequence:
Win_GParted::activate_resize()
Dialog_Partition_Resize_Move::Dialog_Partition_Resize_Move()
Dialog_Base_Partition::Get_New_Partition()
prepare_new_partition()
snap_to_alignment()
snap_to_mebibyte()
prepare_new_partition() created a new partition object to correctly
represent the resized/moved partition #1. However this code in
snap_to_mebibyte() [1] determined that the new location for partition #1
overlapped with where partition #2 currently is on disk, not where
partition #2 will be after the previous operation is applied, therefore
it forced a 1 MiB decrease in partition #1's size creating the gap.
This is because snap_to_mebibyte() is working with the on disk view of
the partitions obtained from devices[current_device] passed into the
Dialog_Partition_Resize_Move() constructor. Hence why applying the
operations one at a time doesn't suffer from the forced 1 MiB gap.
Fix this by creating a copy of the current device object, replacing the
on disk partition layout with the composed partition layout as displayed
in the UI. Then pass this into the Dialog_Partition_Resize_Move
constructor.
[1] Dialog_Base_Partition::snap_to_mebibyte() lines 418 to 438
https://gitlab.gnome.org/GNOME/gparted/-/blob/GPARTED_1_5_0/src/Dialog_Base_Partition.cc#L418Closes#227 - Unable to allocate 1 MiB between partitions when moving to
the right
More recent g++ versions produce these warnings:
test_PasswordRAMStore.cc: In member function ‘virtual void GParted::PasswordRAMStoreTest_TotalErasure_Test::TestBody()’:
test_PasswordRAMStore.cc:61:32: warning: ‘ ’ directive output truncated writing 20 bytes into a region of size 10 [-Wformat-truncation=]
snprintf( buf, sizeof( buf ), "password%03u ", i );
^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
test_PasswordRAMStore.cc:61:10: note: ‘snprintf’ output 32 bytes into a destination of size 21
snprintf( buf, sizeof( buf ), "password%03u ", i );
~~~~~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
snprintf() [1] truncates the printed string to the specified size so
didn't overflow the buffer. However clear the warning by making the
formatted string always exactly 20 characters long, followed by the
terminating NUL character to exactly fill the buffer.
[1] print3(f) - Linux manual page
https://man7.org/linux/man-pages/man3/snprintf.3.html
"The functions snprintf() and vsnprintf() write at most size bytes
(including the terminating null byte ('\0')) to str.
"
Using Automake variable EXTRA_DIST [1] to list the GParted header files
seems overly general. Instead use noinst_HEADERS [2] as it better
describes GParted header files. Header files which need to be
distributed in the archive, but not part of an installed library so not
to be installed below /usr/include.
[1] GNU Automake manual, 14.1 Basics of Distribution
https://www.gnu.org/software/automake/manual/html_node/Basics-of-Distribution.html
"..., it is still common to have files to be distributed which
are not found by the automatic rules. You should listed these
files in the EXTRA_DIST variable. You can mention files in
subdirectories in EXTRA_DIST.
"
[2] GNU Automake manual, 9.2 Header files
https://www.gnu.org/software/automake/manual/html_node/Headers.html
"Usually, only header files that accompany installed libraries
need to be installed. Headers used by programs or convenience
libraries are not installed. The noinst_HEADERS variable can be
used for such headers. However, when the header belongs to a
single convenience library or program, we recommend listing it
in the program's or library's _SOURCES variable (see Defining
program sources) instead of in noinst_HEADERS. This is clearer
for the Makefile.am reader. noinst_HEADERS would be the right
variable to use in a directory containing only headers and no
associated library or program.
All header files must be listed somewhere; in a _SOURCES
variable or in a _HEADERS variable. Missing ones will not
appear in the distribution.
"
The code in erase_filesystem_signatures() used libparted
ped_device_write() which allowed any sector in the whole disk device to
be written. The code only depended on calculations of somewhat
complicated zero offset ranges and the start partition offset to ensure
that it didn't zero sectors outside the target partition. The code
doesn't overwrite partition boundaries, but there have been updates and
bug fixes to the calculation code. To improve the safety create a
libparted geometry representing the partition, or whole disk device,
to be cleared and use ped_geometry_write() so that libparted enforces
writes are only within the partition boundary being erased.
Deliberately breaking erase_filesystem_signatures() code so that it
tries to write past the end of the partition produces this dialog:
Libparted Error
(-) Attempt to write sectors 1024000-1024007 outside of
partition on /dev/sdb.
[ Cancel ] [ Ignore ]
And trying to write before the start of the partition produces this
dialog:
Libparted Bug
(-) Assert (offset >= 0) at cs/geom.c:375 in function
ped_geometry_write() failed.
[ No ]
Followed by GParted aborting and producing a core dump. Not ideal from
libparted, but it does prevent GParted writing outside the partition
boundaries and only occurs in the case of a bug in
erase_filesystem_signatures() which is exercised on every Create and
Format Partition operation and now also unit tested. So not something
we will let through to the users.
Since the previous commit "Also erase all Promise FastTrack RAID
signatures" the previous failing IntelSoftwareRAIDUnaligned test now
passes along with the new PromiseFastTrackRaid* tests.
$ ./test_EraseFileSystemSignatures
Running main() from test_EraseFileSystemSignatures.cc
DISPLAY=":0.0"
[==========] Running 4 tests from 1 test case.
[----------] Global test environment set-up.
[----------] 4 tests from EraseFileSystemSignaturesTest
[ RUN ] EraseFileSystemSignaturesTest.IntelSoftwareRAIDAligned
[ OK ] EraseFileSystemSignaturesTest.IntelSoftwareRAIDAligned (158 ms)
[ RUN ] EraseFileSystemSignaturesTest.IntelSoftwareRAIDUnaligned
[ OK ] EraseFileSystemSignaturesTest.IntelSoftwareRAIDUnaligned (81 ms)
[ RUN ] EraseFileSystemSignaturesTest.PromiseFastTrackRAIDAligned
[ OK ] EraseFileSystemSignaturesTest.PromiseFastTrackRAIDAligned (74 ms)
[ RUN ] EraseFileSystemSignaturesTest.PromiseFastTrackRAIDUnaligned
[ OK ] EraseFileSystemSignaturesTest.PromiseFastTrackRAIDUnaligned (74 ms)
[----------] 4 tests from EraseFileSystemSignaturesTest (387 ms total)
[----------] Global test environment tear-down
[==========] 4 tests from 1 test case ran. (387 ms total)
[ PASSED ] 4 tests.
Closes#220 - Format to Cleared not clearing "pdc" ataraid signature
User reported that GParted didn't clear a pdc (Promise FastTrack) RAID
signature [1]. Reproduce this issue by creating a 16 MiB - 512 byte
test image with Promise FastTrack RAID signatures at all recognised
offsets [2].
$ python << 'EOF'
signature = b'Promise Technology, Inc.'
import os
fd = os.open('/tmp/test.img', os.O_CREAT|os.O_WRONLY)
os.ftruncate(fd, 16*1024*1024 - 512)
for offset in [63, 255, 256, 16, 399, 591, 675, 735, 911, 974, 991, 951, 3087]:
os.lseek(fd, -(offset*512), os.SEEK_END)
os.write(fd, signature)
os.close(fd)
EOF
Then use GParted Format to > Cleared.
$ sudo ./gpartedbin /tmp/test.img
Afterwards blkid, and therefore GParted, still recognises this as a
Promise FastTrack RAID member.
$ blkid /tmp/test.img
/tmp/test.img: TYPE="promise_fasttrack_raid_member"
This is because the test image still contains multiple signatures.
$ hexdump -C /tmp/test.img | grep Promise
00e7e000 50 72 6f 6d 69 73 65 20 54 65 63 ... |Promise Technolo|
00fce000 50 72 6f 6d 69 73 65 20 54 65 63 ... |Promise Technolo|
00fdfe00 50 72 6f 6d 69 73 65 20 54 65 63 ... |Promise Technolo|
00ff8000 50 72 6f 6d 69 73 65 20 54 65 63 ... |Promise Technolo|
Used a test image not an exact multiple of MiBs because drives generally
aren't an exact MiB multiple in size either and as the clearing of ZFS
labels L2 and L3 by writes of zeros at the end of the drive is rounded
to 256 KiBs there will be sectors after that not zeroed where other
Promise signatures remain. The above signatures map back to these
sectors before the end:
16*1024*1024 - 512 = 16776704
512b sectors KiB
(0x00e7e000 - 16776704) / 512 = -3087 -1543.5
(0x00fce000 - 16776704) / 512 = -399 -199.5
(0x00fdfe00 - 16776704) / 512 = -256 -128
(0x00ff8000 - 16776704) / 512 = -63 -31.5
Promise FastTrack RAID signatures are always at multiples 512-byte
sectors (code uses left shift 9 to convert from sectors to byte offset)
[2].
Fix this by:
1. Replace existing zeroing of 3 ranges relative to the end of the
device to be a single range covering the ZFS labels L2 and L3 to the
end of the drive. This will also clear the SWRaid 0.90 & 1.0
super blocks, the Nilfs2 secondary super block, the Intel Software
RAID signature found not zeroed in the unaligned unit test case and
the above Promise FastTrack RAID signatures at -199.5 KiB and later.
2. Add zeroing of the final Promise FastTrack RAID signature at sector
-3087.
Performed a review of all the other ATARAID super blocks detected by
blkid (files *_raid.c) [3] and they are all located within the last 11
sectors so will be zeroed by case 1. above.
[1] GParted forum thread: How to remove a ataraid partition ?
http://gparted-forum.surf4.info/viewtopic.php?id=18104
[2] blkid from util-linux promise_raid.c:probe_pdcraid()
https://git.kernel.org/pub/scm/utils/util-linux/util-linux.git/tree/libblkid/src/superblocks/promise_raid.c?h=v2.38.1#n27
[3] blkid RAID member detection (files *_raid.c)
https://git.kernel.org/pub/scm/utils/util-linux/util-linux.git/tree/libblkid/src/superblocks/?h=v2.38.1Closes#220 - Format to Cleared not clearing "pdc" ataraid signature
Each test in test_EraseFileSystemSignatures is taking just over 10
seconds to run in the Alpine Linux CI image:
[ RUN ] EraseFileSystemSignaturesTest.IntelSoftwareRAIDAligned
[ OK ] EraseFileSystemSignaturesTest.IntelSoftwareRAIDAligned (10045 ms)
[ RUN ] EraseFileSystemSignaturesTest.IntelSoftwareRAIDUnaligned
...
[ FAILED ] EraseFileSystemSignaturesTest.IntelSoftwareRAIDUnaligned (10048 ms)
[----------] 2 tests from EraseFileSystemSignaturesTest (20093 ms total)
[----------] Global test environment tear-down
[==========] 2 tests from 1 test case ran. (20093 ms total)
This is because the udevadm command is not found and so settle_device()
waits for 10 seconds in this call chain:
erase_filesystem_signatures()
settle_device(SETTLE_DEVICE_APPLY_MAX_WAIT_SECONDS)
sleep(10)
Install udevadm command into the Alpine Linux CI job docker image to fix
this. Now it's on a par with the time taken in the other distro CI test
jobs:
[ RUN ] EraseFileSystemSignaturesTest.IntelSoftwareRAIDAligned
[ OK ] EraseFileSystemSignaturesTest.IntelSoftwareRAIDAligned (417 ms)
[ RUN ] EraseFileSystemSignaturesTest.IntelSoftwareRAIDUnaligned
...
[ FAILED ] EraseFileSystemSignaturesTest.IntelSoftwareRAIDUnaligned (165 ms)
[----------] 2 tests from EraseFileSystemSignaturesTest (582 ms total)
[----------] Global test environment tear-down
[==========] 2 tests from 1 test case ran. (582 ms total)
Closes#220 - Format to Cleared not clearing "pdc" ataraid signature
Move common testing code which doesn't need linking with GParted objects
into the common module. Move the remaining common code used to print
GParted objects using the insertion operator (operator<<) into the
insertion_operators module. Split the common code like this so that the
operator<<(std::ostream&, const OperationDetail&) function is not
included in test_PipeCapture and it is not forced to link with all the
non-UI related GParted objects.
The Automake manual provides guidance that when a header belongs to a
single program it is recommended to be listed in the program's _SOURCES
variable and for a directory only containing header files listing them
in the noinst_HEADERS variable is the right variable to use [1].
However the guidance doesn't cover this case for common.h and
insertion_operators.h; header files in a directory with other files and
used by multiple programs. So just because we have gparted_core_OBJECTS
(normal Makefile, not Automake special variable) listing objects to link
with, choose to use noinst_HEADERS Automake variable to list needed
headers.
[1] GNU Automake manual, 9.2 Header files
https://www.gnu.org/software/automake/manual/html_node/Headers.html
"Usually, only header files that accompany installed libraries
need to be installed. Headers used by programs or convenience
libraries are not installed. The noinst_HEADERS variable can be
used for such headers. However, when the header belongs to a
single convenience library or program, we recommend listing it
in the program's or library's _SOURCES variable (see Defining
program sources) instead of in noinst_HEADERS. This is clearer
for the Makefile.am reader. noinst_HEADERS would be the right
variable to use in a directory containing only headers and no
associated library or program.
All header files must be listed somewhere; in a _SOURCES
variable or in a _HEADERS variable. Missing ones will not
appear in the distribution.
"
Closes#220 - Format to Cleared not clearing "pdc" ataraid signature
Initially just testing erasing of Intel Software RAID signatures.
Chosen because it was expected to work, but turned out not to be true in
all cases.
The code needs to initialise GParted_Core::mainthread, construct
Gtk::Main() and execute xvfb-run because of this call chain:
GParted_Core::erase_filesystem_signatures()
GParted_Core::settle_device()
Utils::execute_command ("udevadm settle ...")
status.foreground = (Glib::Thread::self() == GParted_Core::mainthread)
Gtk::Main::run()
This was also needed when testing file system interface classes as
discussed in commits [1][2].
The test fails like this:
$ ./test_EraseFileSystemSignatures
...
[ RUN ] EraseFileSystemSignaturesTest.IntelSoftwareRAIDAligned
[ OK ] EraseFileSystemSignaturesTest.IntelSoftwareRAIDAligned (155 ms)
[ RUN ] EraseFileSystemSignaturesTest.IntelSoftwareRAIDUnaligned
test_EraseFileSystemSignatures.cc:286: Failure
Failed
image_contains_all_zeros(): First non-zero bytes:
0x00001A00 "Intel Raid ISM C" 49 6E 74 65 6C 20 52 61 69 64 20 49 53 4D 20 43
test_EraseFileSystemSignatures.cc:320: Failure
Value of: image_contains_all_zeros()
Actual: false
Expected: true
[ FAILED ] EraseFileSystemSignaturesTest.IntelSoftwareRAIDUnaligned (92 ms)
Manually write the same test image:
$ python << 'EOF'
signature = b'Intel Raid ISM Cfg Sig. '
import os
fd = os.open('/tmp/test.img', os.O_CREAT|os.O_WRONLY)
os.ftruncate(fd, 16*1024*1024 - 512)
os.lseek(fd, -(2*512), os.SEEK_END)
os.write(fd, signature)
os.close(fd)
EOF
Run gpartedbin /tmp/test.img and Format to > Cleared. GParted continues
to display the the image file as containing an ataraid signature.
$ blkid /tmp/test.img
/tmp/test.img: TYPE="isw_raid_member"
$ hexdump -C /tmp/test.img
00000000 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 |................|
*
00fffa00 49 6e 74 65 6c 20 52 61 69 64 20 49 53 4d 20 43 |Intel Raid ISM C|
00fffa10 66 67 20 53 69 67 2e 20 00 00 00 00 00 00 00 00 |fg Sig. ........|
00fffa20 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 |................|
*
00fffe00
This signature is not being cleared when the device/partition/image size
is 512 bytes smaller than a whole MiB because the last 3.5 KiB is left
unwritten. This is because the last block of zeros written is 8 KiB
aligned to 4 KiB at the end of the device.
[1] a97c23c57c
Add initial create ext2 only FileSystem interface class test (!49)
[2] 8db9a83b39
Run test program under xvfb-run to satisfy need for an X11 display (!49)
Closes#220 - Format to Cleared not clearing "pdc" ataraid signature
As documented in the previous commit xfsprogs >= 5.19.0 refuses to
create an XFS file system smaller than 300 MiB.
$ truncate -s $((300*1024*1024-1)) test.img
$ ls -l test.img
-rw-r--r-- 1 auser auser 314572799 Dec 21 11:01 test.img
$ mkfs.xfs -V
mkfs.xfs version 6.0.0
$ mkfs.xfs test.img
Filesystem must be larger than 300MB.
...
$ echo $?
1
Successfully create an XFS file system at minimum size of 300 MiB.
$ truncate -s $((300*1024*1024)) test.img
$ ls -l test.img
-rw-r--r-- 1 auser auser 314572800 Dec 21 11:05 test.img
$ mkfs.xfs test.img
...
$ echo $?
0
$ blkid test.img
test.img: UUID="..." BLOCK_SIZE="512" TYPE="xfs"
Increase the GParted minimum XFS size to 300 MiB. For simplicity and
because the XFS developers said of smaller XFS file systems [1]:
"are known to have performance and redundancy problems that are not
present on the volume sizes that XFS is best at handling"
regardless of the version of mkfs.xfs used to create that XFS then apply
to all versions of xfsprogs.
[1] https://git.kernel.org/pub/scm/fs/xfs/xfsprogs-dev.git/commit/?id=6e0ed3d19c54603f0f7d628ea04b550151d8a262
mkfs: stop allowing tiny filesystems
Closes#217 - GitLab CI test job failing with new mkfs.xfs error
"Filesystem must be larger than 300MB."
From 27-Nov-2022 the alpine_test GitLab CI job started failing,
reporting errors creating XFS file systems in the
test_SupportedFileSystems unit test like this:
[ RUN ] My/SupportedFileSystemsTest.Create/xfs
test_SupportedFileSystems.cc:501: Failure
Value of: m_fs_object->create(m_partition, m_operation_detail)
Actual: false
Expected: true
Operation details:
mkfs.xfs -f -L '' '/builds/GNOME/gparted/tests/test_SupportedFileSystems.img' 00:00:00 (ERROR)
Filesystem must be larger than 300MB.
...
This is because Docker image "alpine:latest" has updated to Alpine Linux
3.17 which includes xfsprogs 6.0.0 which includes this change (first
released in xfsprogs 5.19.0):
https://git.kernel.org/pub/scm/fs/xfs/xfsprogs-dev.git/commit/?id=6e0ed3d19c54603f0f7d628ea04b550151d8a262
mkfs: stop allowing tiny filesystems
Refuse to format a filesystem that are "too small", because these
configurations are known to have performance and redundancy problems
that are not present on the volume sizes that XFS is best at
handling.
Specifically, this means that we won't allow logs smaller than 64MB,
we won't allow single-AG filesystems, and we won't allow volumes
smaller than 300MB.
Increase the default unit test file system image size from 256 MiB to
256+64 = 320 MiB to avoid this error.
Closes#217 - GitLab CI test job failing with new mkfs.xfs error
"Filesystem must be larger than 300MB."
GParted's check operation is a check and if possible repair. For most
file system types GParted already requests that the file system is
repaired. fsck.exfat -y flag has been available since the first release
of exfatprogs 1.0.1 [1] so unconditionally add this.
[1] exfatprogs 1.0.1 fsck/fsck.c:main() case 'y':
https://github.com/exfatprogs/exfatprogs/blob/1.0.1/fsck/fsck.c#L1231!109 - Enable repair when checking exfat file systems
The code was overly complicated in how it converted to the 32-bit little
endian on-disk representation of the Hidden Sectors field. It did:
1. Formatted the partition start sector as a hexadecimal string.
2. Padded it to 8 digits.
3. Reversed pairs of digits.
4. Converted pairs of hexadecimal digits to bytes of binary data.
5. Wrote the 4 bytes of binary data to the Hidden Sectors field.
There is no need for all this string manipulation to convert to a 32-bit
little endian value. Just do this:
1. Truncate (signed 64-bit) partition start sector to 32-bit.
2. Convert from host native to little endian.
3. Write as 4 bytes of binary data to the Hidden Sectors field.
The code also ignores write errors. ofstream.write() only copies the
data into an in process buffer [1] and the data is not passed to the OS
to write to the open file handle until ofstream.close() [2] is called.
However the status of close() was not checked so a failure of the OS to
perform the write would go unreported.
In the case of a failure providing the user with a command line to set
the Hidden Sectors field is excessive. Updating the Hidden Sectors is
no more or less likely to fail than for any other storage manipulation
action. For example GParted doesn't provide command line instructions
to update a partition size if a libparted call fails. Therefore remove.
Rewrite the code to resolve the above issues and lay it out using
if-operation-fails-return-early pattern.
[1] std::ostream::write()
"... it inserts characters into associated stream_buffer object as
if calling its member function sputc until n characters have been
written or until an insertion fails ..."
https://cplusplus.com/reference/ostream/ostream/write/
[2] std::ofstream::close()
"Any pending output sequence is written to the file."
https://cplusplus.com/reference/fstream/ofstream/close/Closes#164 - GParted crashes copying NTFS partition to starting beyond
2TiB
Create test setup using a 4 TiB loop device:
# truncate -s 4T /tmp/disk.img
# losetup -f --show /tmp/disk.img
/dev/loop0
Create 2 x 1 TiB partitions. First at offset 1 MiB, second at offset
2 TiB:
# sgdisk --new 1:2048:2147485696 --typecode 1:0700 /dev/loop0
# sgdisk --new 2:4294967296:6442450944 --typecode 2:0700 /dev/loop0
# partprobe /dev/loop0
Create NTFS file system in the first partition:
# mkntfs -Q /dev/loop0p1
Then use GParted to copy the first NTFS partition into the second
partition. GParted crashes:
# gpartedbin /dev/loop0
...
(gpartedbin:14660): glibmm-ERROR **: 20:39:01.191:
unhandled exception (type std::exception) in signal handler:
what: basic_string::_M_replace_aux
Trace/breakpoint trap
# echo $?
133
Overview of what is happening is that GParted_Core::update_bootsector()
is attempting to set the Hidden Sectors [1] field in the NTFS Partition
Boot Sector (PBS) to the start sector of the newly copied /dev/loop0p2
partition. But the sector number is greater than will fit in a 32-bit
unsigned integer, which the code doesn't handle.
Specifically the code prints the sector number as a hexadecimal number
into string 'hex'. As the target partition starts at exactly 2 TiB then
hex="100000000" (9 hexadecimal digits long). Next:
hex.insert(0, 8 - hex.length(), '0');
is meant to pad the beginning of the 'hex' string with '0's to make the
string 8 character long. But the string is already 9 character long so
8 - 9 is -1 which as unsigned integral type size_t [2] is 2^64-1. So
insert() is trying to insert 18446744073709551615 '0's at the start of
the 'hex' string! Hence the crash.
mkntfs refuses to accept an explicit partition start sector of 2^32 or
larger:
# mkntfs -Q --partition-start 4294967296 /dev/loop0p2
Invalid partition start sector. Maximum is 4294967295 (2^32-1).
# echo $?
1
When mkntfs can't determine the drive geometry and partition offset, as
is the case on loop devices or the partition start sector is 2^32 or
larger, then mkntfs writes zero into Hidden Sectors:
# mkntfs -Q /dev/loop0p1
The partition start sector was not specified for /dev/loop0p1 and it could not be obtained automatically. It has been set to 0.
...
To boot from a device, Windows needs the 'partition start sector', ...
Windows will not be able to boot from this device.
Creating NTFS volume structures.
mkntfs completed successfully. Have a nice day.
# echo $?
0
# hexdump -C /dev/loop0p1 | head -2
00000000 eb 52 90 4e 54 46 53 20 20 20 20 00 02 08 00 00 |.R.NTFS .....|
00000010 00 00 00 00 00 f8 00 00 00 00 00 00 00 00 00 00 |................|
^^ ^^ ^^ ^^
Hidden Sectors value at offset 0x1C in the NTFS Partition Boot Sector.
So mkntfs is warning, writing the Hidden Sectors as zero and reporting
success. Fix GParted in an equivalent way when it is updating the
Hidden Sectors for a moved or copied NTFS which starts at sector 2^32
and beyond.
After this fix the operational details for the same copy operation are:
Copy /dev/loop0p1 to /dev/loop0p2
* calibrate /dev/loop0p1 (SUCCESS)
* calibrate /dev/loop0p2 (SUCCESS)
* set partition type on /dev/loop0p2 (SUCCESS)
* copy file system from /dev/loop0p1 to /dev/loop0p2 (SUCCESS)
* update boot sector of ntfs file system on /dev/loop0p2 (WARNING)
Partition start (4294967296) is beyond sector 4294967295 (2^32-1).
Windows will not be able to boot from this file system.
* check file system on /dev/loop0p2 for errors and if p... (SUCCESS)
[1] NTFS, Partition Boot Sector (PBS)
"
Byte Field Field name Purpose
offset length
0x1C 5 bytes Hidden Sectors The number of sectors preceding
the partition.
"
https://en.wikipedia.org/wiki/NTFS#Partition_Boot_Sector_(PBS)
[2] std::string::insert
"fill (5) string& insert(size_t pos, size_t n, char c);
Insert into string
Inserts additional characters into the string right before the
character indicated by pos (or p):
(5) fill
Insert n consecutive copies of character c.
"
https://cplusplus.com/reference/string/string/insert/Closes#164 - GParted crashes copying NTFS partition to starting beyond
2TiB
Mike Fleetwood