When opening the Partition Information dialog, creation of the dialog
object was creating a copy of the partition object to be displayed. If
this was an extended partition it also included recursively constructing
the contained logical partitions too.
Instead, replace the partition object in the Dialog_Partition_Info class
with a reference to it.
NOTE:
In C++ a reference is really just a pointer under the hood. As such,
dereferences of a pointer to an object in the context of needing a
reference to the object doesn't copy the object. It merely initialises
the reference from the pointer.
Specifically, with this prototype:
Dialog_Partition_Info( const Partition & partition );
and the dialog object being constructed in Win_GParted::activate_info():
Dialog_Partition_Info dialog( *selected_partition_ptr );
the partition object is not copy constructed. A reference (pointer) to
it is merely passed to the dialog constructor.
Bug 750168 - Reduce the amount of copying of partition objects
Further ensure that a bug doesn't get introduced with the use of
selected_partition_ptr, by asserting that it points to a current
partition object in the vector of display partitions.
After deliberately breaking the code so that selected_partition_ptr
points to some other partition object, trying to display the Information
dialog causes this crash:
======================
libparted : 2.4
======================
**
ERROR:Win_GParted.cc:989:void GParted::Win_GParted::set_valid_operations(): assertion failed: (valid_display_partition_ptr( selected_partition_ptr ))
Aborted (core dumped)
At this point in the code:
973 void Win_GParted::set_valid_operations()
974 {
...
986 // No partition selected ...
987 if ( ! selected_partition_ptr )
988 return ;
>> 989 g_assert( valid_display_partition_ptr( selected_partition_ptr ) ); // Bug: Not pointing at a valid display partition object
Bug 750168 - Reduce the amount of copying of partition objects
Now that TreeView_Details and DrawingAreaVisualDisk classes store and
pass pointers to partition objects in the Gtk signal callbacks, change
the selected partition into a pointer too.
Bug 750168 - Reduce the amount of copying of partition objects
This stops copying of each displayed partition object into the
DrawingAreaVisualDisk class.
Bug 750168 - Reduce the amount of copying of partition objects
This stops copying of each displayed partition object into the
TreeView_Details class.
It also stops copy constructing lots of partition objects when just
clicking on a partition in the disk graphic. The disk graphic needs to
inform the main GUI and then the partition list which partition has been
selected. The call sequence goes like:
DrawingAreaVisualDisk::on_button_press_event(event)
Win_GParted::on_partition_selected(partition_ptr, src_is_treeview)
TreeView_Detail::set_selected(partition_ptr)
TreeView_Detail::set_selected(rows, partition_ptr,
inside_extended)
Relevant source and highlighted comparison line:
140 bool TreeView_Detail::set_selected( Gtk::TreeModel::Children rows,
141 const Partition * partition_ptr, bool inside_extended )
142 {
143 for ( unsigned int t = 0 ; t < rows .size() ; t++ )
144 {
>> 145 if ( static_cast<Partition>( rows[t][treeview_detail_columns.partition] ) == *partition_ptr )
146 {
147 if ( inside_extended )
148 expand_all() ;
149
150 set_cursor( static_cast<Gtk::TreePath>( rows[ t ] ) ) ;
151 return true ;
152 }
153
154 if ( set_selected( rows[t].children(), partition_ptr, true ) )
155 return true ;
156 }
157
158 return false ;
159 }
Then in this function the partition selected in the disk graphic
(partition_ptr parameter) is compared in turn with each partition object
stored in the Gtk::TreeView model to find the matching one to mark it as
selected. This mere act of accessing the partition object stored in a
row of the Gtk::TreeView model causes it to be copy constructed. So
clicking on the 5th partition in the disk graphic will copy construct
the first 5 partition objects just to do a compare to find the matching
one.
This is because it is not possible to get a reference from a
Gtk:TreeViewProxy in gtkmm. Merely accessing a value in a Gtk::TreeView
model takes a copy of that value.
Subject: get a reference from a Gtk::TreeValueProxy
http://comments.gmane.org/gmane.comp.gnome.gtkmm/2217http://marc.info/?t=104400417500001&r=1&w=4
Bug 750168 - Reduce the amount of copying of partition objects
Change from passing a reference to the selected partition, to passing a
pointer to the selected partition in the signal_partition_selected
callbacks between the disk graphic, partition list and core GUI modules.
This is an enabler for the following patches.
Bug 750168 - Reduce the amount of copying of partition objects
Win_GParted::Refresh_Visual() used a local variable containing a copy of
the vector of partitions in the current device to be displayed. After
visually applying pending operations it loaded copies of each partition
object into the GUI widgets to display the disk graphic and partition
list, DrawingAreaVisualDisk and TreeView_Details classes respectively.
When a partition is selected in the UI, again a partition object is
copied. Also several of the partition dialogs, including the
information dialog, take a copy of the partition object. All these are
copies of the same set of partition objects, those currently being
displayed in the UI.
Move the vector of displayed partitions from a local variable in
Refresh_Visual() to a Win_GParted member variable. This will allow for
the above cases to be changed to used pointers and references to the
same set of partition objects.
The valid lifetime of pointers to elements in this partition object
vector is from one refresh to the next, when the vector is cleared and
repopulated with a new set of partition objects. This is exactly what
is needed as the GUI widgets are reloaded on each refresh, the selected
partition is reset and none of the partition dialog objects exist.
Dialog objects being created and destroyed on each use.
On the other hand some copies of partition objects currently being
displayed, still need to be made because they have lifetimes which need
to last longer than the next call to Refresh_Visual(). Specifically the
source of the copy partition and the partition objects copied into the
in the list of pending operations.
Bug 750168 - Reduce the amount of copying of partition objects
BUF in the copy dialog class, Dialog_Partition_Copy, is use to adjust
limits in 2 cases:
1) Minimum size when copying an XFS file system
Minimum size was set to the used space + 2 * cylinder size (typically
plus ~16 MiB). This commit from 2004-12-20 added it:
a54b52ea33
xfs copy now uses xfsdump and xfsrestore. icw some hacks in the other 2
Issues:
* This is increasing the minimum XFS file system size when copying it,
which doesn't happen in the resize case for other file systems.
* It allows an XFS file system to be created which is smaller than the
minimum size allowed by GParted. Copying an empty XFS file system can
create a new file system as small as 26 MiB. This is smaller than the
minimum GParted allows of 32 MiB because that is the minimum
xfs_repair can handle.
Remove this addition when copying an XFS file system and enforce minimum
file system size.
2) Maximum size when copying a file system into empty space larger than
it's maximum size
Maximum size was set to maximum file system size - cylinder size
(typically minus ~8 MiB). Only applied to FAT16 which has a maximum
file system size set in and can be grown. Added by this commit from
2004-12-15:
10e8f3338d
:get_fs now returns a const reference. in copy and resizedialog
...
* in copy and resizedialog filesystems with MAX set now have a max size of MAX - one cylinder .
Issue:
* This is applying a lower maximum resize when copying the file system
compared to that when creating the file system.
NOTE:
GParted currently allows all file systems to be resize to any size,
regardless of the maximum file system size. This is probably an
oversight, but it does allow libparted to convert FAT16 to FAT32 file
system when resizing.
Remove this lower maximum file system size when copying and resizing,
compared to creating.
Bug 749867 - Some limits are adjusted by arcane cylinder size amount
when copying and resizing in a single operation
This commit from 2010-05-20 removed use of cylinder size increase in the
minimum, and cylinder size decrease in the maximum file system sizes
from the resize/move dialog.
e62a23b5b5
Add partition alignment option to align to MiB (#617409)
This cylinder size limit adjustments were being performed using the
Dialog_Base_Partition::BUF member variable. Now in the
Dialog_Partition_Resize_Move class it is never accessed, and only
unnecessarily set. Move BUF from the common base class into the
Dialog_Partition_Copy class where it is still used.
Bug 749867 - Some limits are adjusted by arcane cylinder size amount
when copying and resizing in a single operation
The function never modifies any member variables so make it a const
member function.
(FileSystem::get_custom_text() is a virtual function so can't be made
static).
Rename a couple of GParted_Core methods for consistency and to better
distinguish get_filesystem() from get_filesystems() which do completely
unrelated things.
get_filesystem() -> detect_filesystem()
recognise_filesystem_signature() -> detect_filesystem_internal()
Also make detect_filesystem() a static member method as it doesn't use
any member variables. Requirement cascades to get_partition_path().
GParted_Core methods:
flush_device()
get_device()
get_disk()
get_device_and_disk()
destroy_device_and_disk()
commit()
commit_to_os()
settle_device()
This group of methods only call libparted API functions and run external
executables. None of them access any GParted_Core member variables.
Make them all static member functions.
These member functions are only used within the GParted_Core class and
only operate on the static member variable FILESYSTEM_MAP.
Make both functions private and also make init_filesystems() static.
The FileSystem objects stored in the FILESYSTEM_MAP are allocated once
using new in init_filesystems() but never deleted.
Valgrind output fragment:
# valgrind --leak-check=full ./gparted
==29314== 353 (72 direct, 281 indirect) bytes in 1 blocks are definitely lost in loss record 6,287 of 6,905
==29314== at 0x4A075FC: operator new(unsigned long) (vg_replace_malloc.c:298)
>> ==29314== by 0x46EDA5: GParted::GParted_Core::init_filesystems() (GParted_Core.cc:106)
==29314== by 0x46EC5F: GParted::GParted_Core::GParted_Core() (GParted_Core.cc:96)
==29314== by 0x4A74F4: GParted::Win_GParted::Win_GParted(std::vector<Glib::ustring, std::allocator<Glib::ustring> > const&) (Win_GParted.cc:51)
==29314== by 0x4D600A: main (main.cc:56)
...
==29314== 161 (72 direct, 89 indirect) bytes in 1 blocks are definitely lost in loss record 6,119 of 6,905
==29314== at 0x4A075FC: operator new(unsigned long) (vg_replace_malloc.c:298)
>> ==29314== by 0x46F50C: GParted::GParted_Core::init_filesystems() (GParted_Core.cc:124)
==29314== by 0x46EC5F: GParted::GParted_Core::GParted_Core() (GParted_Core.cc:96)
==29314== by 0x4A74F4: GParted::Win_GParted::Win_GParted(std::vector<Glib::ustring, std::allocator<Glib::ustring> > const&) (Win_GParted.cc:51)
==29314== by 0x4D600A: main (main.cc:56)
GParted_Core.cc source:
102 void GParted_Core::init_filesystems()
103 {
104 FILESYSTEM_MAP[ FS_UNKNOWN ] = NULL ;
105 FILESYSTEM_MAP[ FS_CLEARED ] = NULL ;
>> 106 FILESYSTEM_MAP[ FS_BTRFS ] = new btrfs() ;
...
>> 124 FILESYSTEM_MAP[ FS_XFS ] = new xfs() ;
125 FILESYSTEM_MAP[ FS_BITLOCKER ] = NULL ;
Fix by deleting all FILESYSTEM_MAP pointers. Note that delete on a NULL
pointer is defined by C++ as a safe do nothing operation.
C++ FAQ / Do I need to check for null before delete p?
https://isocpp.org/wiki/faq/freestore-mgmt#delete-handles-null
Fixing this reduces the valgrind reported definitely lost memory blocks
count from 25 down to 6. 19 FileSystem objects deleted and 19 memory
blocks no longer lost.
Bug 749036 - FileSystem objects are memory leaked in init_filesystems()
Add a partition name entry box to the Create New Partition dialog. The
entry box is greyed out (not sensitive) for partition table types which
don't support partition naming. Currently only supported for GPTs. See
Utils::get_max_partition_name_length() for details.
There was a slightly wider gap between the file system combobox row and
the label entry row when there were only three widgets on the right hand
side of the dialog. This has been removed now that there are four
widgets so that they are all evenly spaced and they line up with the
four widgets on the left hand side.
So far the partition name can be entered and previewed, but isn't yet
applied to the disk.
Bug 746214 - Partition naming enhancements
Adding a partition name entry to the Create New Partition dialog will
need access to these two Device methods: partition_naming_supported()
and get_max_partition_length(). The Set_Data() function already takes
two parameters, only_unformatted and disktype, taken from Device member
variables.
Rather than add two more parameters to the Set_Data() function pass the
Device object instead, replacing the current only_unformatted and
disktype parameters.
Bug 746214 - Partition name enhancements
Rename Gtk::Entry object entry -> filesystem_label_entry in the
Dialog_Partition_New class. This is in preparation for the introduction
of the partition name entry box in the Create New Partition dialog.
Bug 746214 - Partition name enhancements
Previously partition naming had only been implemented for gpt. Make the
code ready to support naming of the other partition table types for
which libparted supports naming. Specifically: amiga, dvh, mac and
pc98 in addition to gpt. Document issues found with some of these
partition table types, which can relatively easily been worked around.
Leave support of naming for partition table types other than gpt
disabled, mostly just to reduce ongoing testing effort, at least until
there is any user demand for it.
Bug 746214 - Partition naming enhancements
resize_move() and move() stopped using the device parameter in this
commit from 2006-07-23:
d663c3c277
removed cylindersize buffering during resize from the filesystems. It is
create() stopped using the device parameter in this commit from 2006-03-19:
ad9f2126e7
fixed issues with copying (see also #335004) cleanups + added FIXME added
For reference most other operation methods had the device parameter
removed in this earlier commit from 2005-12-07:
642f0a145b
from now on each partition has a reference to it's device. make use of new
When the following conditions were met GParted would fail to recognise a
newly created whole disk device file system, and instead show an unknown
file system filling the disk:
1) Disk was previously partitioned and contained at least one partition.
2) Using libparted version 2.0 to 3.0 inclusive.
Initial status:
# blkid | fgrep sdc
# fgrep sdc /proc/partitions
8 32 976762584 sdc
8 33 104857600 sdc1
# parted /dev/sdc
GNU Parted 2.4
Using /dev/sdc
Welcome to GNU Parted! Type 'help' to view a list of commands.
(parted) print
Model: ATA ST1000LM024 HN-M (scsi)
Disk /dev/sdc: 1000GB
Sector size (logical/physical): 512B/4096B
Partition Table: msdos
Number Start End Size Type File system Flags
1 1049kB 107GB 107GB primary
When creating the loop partition table libparted would not inform the
kernel to delete the old partitions. /proc/partitions still contained
the details of the old partitions.
(parted) mktable loop
Warning: The existing disk label on /dev/sdc will be destroyed and
all data on this disk will be lost. Do you want to continue?
Yes/No? Yes
(parted) print
Model: ATA ST1000LM024 HN-M (scsi)
Disk /dev/sdc: 1000GB
Sector size (logical/physical): 512B/4096B
Partition Table: loop
Number Start End Size File system Flags
(parted) quit
# fgrep sdc /proc/partitions
8 32 976762584 sdc
8 33 104857600 sdc1
Creation of the whole disk device file system goes unnoticed by blkid
because the kernel and therefore blkid's cache have stale partition
information.
# mkfs.xfs -f /dev/sdc
# blkid | fgrep sdc
NOTE:
On a Linux Software RAID array, as opposed to a hard disk, blkid does
notice creation of the whole disk device file system. However the
kernel still has old partition details.
This was fixed in libparted 3.1 by commit:
http://git.savannah.gnu.org/cgit/parted.git/commit/?id=f5c909c0cd50ed52a48dae6d35907dc08b137e88
libparted: remove has_partitions check to allow loopback partitions
Fix by deleting old partitions before creating the loop table when
compiled with a broken version of libparted. The GParted UI provides
no feedback while a new partition table is created, and with some
versions of GTK the UI become unresponsive too, so it is important to be
as fast as possible. Evaluated three different methods, deleting 15 and
22 MSDOS partitions on a physical 5400 RPM hard drive using libparted
2.4:
M1) Delete and commit one partition at a time.
Takes up to 24 seconds to delete 15 partitions. With 22 partitions
libparted always reports finding some of the partitions busy and
unable to inform the kernel about the modifications.
Too slow and doesn't work.
M2) Delete all partitions in one go and commit once.
Takes up to 1.4 seconds to delete either 15 or 22 partitions. Never
removes partitions 17 and higher from the kernel.
Doesn't work.
M3) Write GPT table (letting libparted delete any old partitions).
Takes up to 0.8 seconds to delete either 15 or 22 partitions.
Fast and works.
Use method 3 - write a GPT table thus using libparted code to inform the
kernel of the old partition deletions.
Bug 743181 - Add unpartitioned drive read-write support
Older versions of blkid don't correctly distinguish between FAT16 and
FAT32 file systems when overwriting one with the other. This effects
GParted too with these file systems on whole disk devices where only
blkid is used to recognise the contents. See previous fix for why only
blkid is used in this case:
Avoid whole disk FAT being detected as MSDOS partition table
(#743181)
Example:
# blkid -v
blkid from util-linux 2.20.1 (liblkid 2.20.0, 19-Oct-2011)
# mkdosfs -F16 -I /dev/md1
# blkid | fgrep md1
/dev/md1: SEC_TYPE="msdos" UUID="7C23-95D9" TYPE="vfat"
# mkdosfs -F32 -I /dev/md1
# blkid | fgrep md1
/dev/md1: SEC_TYPE="msdos" UUID="7F93-98F4" TYPE="vfat"
So blkid recognised the UUID changed but didn't remove the SEC_TYPE for
the FAT32 file system. See FS_Info::get_fs_type() as it uses this to
distinguish between FAT16 and FAT32. This is a caching update bug in
blkid, because telling blkid not to use the cache gets the right
results:
# blkid -c /dev/null | fgrep md1
/dev/md1: UUID="7F93-98F4" TYPE="vfat"
With testing determined that blkid from util-linux 2.23 and later are
not affected and earlier versions are affected. Mostly recently known
affected distribution is Ubuntu 14.04 LTS with util-linux 2.20.1.
The straight forward fix would be to instruct blkid to not use its cache
with 'blkid -c /dev/null'. But using blkid's cache is needed to prevent
blkid hanging for minutes when trying to access a non-existent floppy
drive when the BIOS is set incorrectly. See commit:
18f863151c
Fix long scan problem when BIOS floppy setting incorrect
Instead, when using an older affected version of blkid and when blkid
cache reports a vfat file system, run blkid again bypassing the cache.
The device is known to exist and contain a vfat file system, just not
whether it is a FAT16 or FAT32 file system, so can't be a non-existent
floppy device and won't hang.
Bug 743181 - Add unpartitioned drive read-write support
When writing "loop" partition table over the top of some whole disk
device file system types GParted continued to show those whole disk
device file systems rather than the virtual unknown partition from the
"loop" partition table.
This affected btrfs, jfs, reiser4 and reiserfs. It occurred because of
several factors:
1) Libparted only zeroed the first and last 9.5 KiB (assuming 512 byte
sectors) of the device before writing a new partition table. See
ped_disk_clobber().
2) These file systems have their super blocks and therefore signatures
after the first 9.5 KiB.
3) Whole disk device file system detection is performed using blkid
before checking for a libparted "loop" partition table. See
GParted_Core::set_devices_thread().
Ref:
libparted 3.2: disk.c:ped_disk_clobber()
http://git.savannah.gnu.org/cgit/parted.git/tree/libparted/disk.c?id=v3.2#n302
Fix by always erasing any possible file system signatures on the device
before creating a new "loop" partition table.
NOTE:
This is typically taking up to 0.5 seconds in my testing on a 5400 RPM
hard drive, during which time the GParted UI is hung and the create
partition table dialog shows the apply button pressed but no other
progress indication.
Bug 743181 - Add unpartitioned drive read-write support
get_device_and_disk() basically calls libparted to get a PedDevice
object representing a disk device and a PedDisk object representing a
partition table. Re-implement get_device_and_disk() using two separate
functions, get_device() and get_disk(), to get one of these objects
each.
No functionality changes with this commit. It enables future commits to
incrementally add support for whole disk devices into GParted without
needing libparted to recognise the contents and create a virtual "loop"
partition table.
Bug 743181 - Add unpartitioned drive read-write support
Need to be able to take different actions in the GParted_Core partition
manipulation methods and in Win_GParted UI methods to deal with
libparted supported partitions or whole disk devices without a partition
table. Add boolean whole_device to the partition object and set
appropriately to allow for this.
Bug 743181 - Add unpartitioned drive read-write support
Function Utils::get_filesystem_kernel_name() returns the name of the
file system as needed for use in the mount command:
mount -t TYPE DEVICE DIR
Needed because the kernel / mount name is 'hfsplus' where as libparted /
GParted, as reported by Utils::get_filesystem_string(), calls it 'hfs+'.
So far just added debugging when mounting a file system to test the
function works.
# ./gartedbin
======================
libparted : 2.1
======================
DEBUG: (hfsplus) # mount -v /dev/sdb5 "/mnt/5"
DEBUG: (nilfs2) # mount -v /dev/sdb1 "/mnt/1"
Bug 742741 - Nilfs2 file system is unusable on RHEL/CentOS 6
For file systems which libparted recognises, when found on the whole
disk device, it reports with partition table "loop" and a partition
covering the whole disk. GParted duly displays this to the user.
For file systems which libparted doesn't recognise it reports
"unrecognised disk label". As of the latest libparted 3.2, these file
system aren't recognised and can't currently be shown when on the whole
disk device:
BitLocker, Crypt LUKS, exFAT, F2FS, LVM2 Physical Volume,
Linux Software RAID, ReFS, Reiser 4
So only when libparted doesn't recognise a file system on the whole disk
device and GParted does, either via blkid or it's internal code, display
this with partition table "none".
Bug 741430 - GParted cannot recognise LVM signature on unpartitioned
drive
Move code which queries the file system label and UUID of a partition
into a separate helper function.
Bug 741430 - GParted cannot recognise LVM signature on unpartitioned
drive
Refactor GParted internal file system signature detection to remove code
duplication. There were 5 separate copies of code to: allocate a
buffer, open, read and close the device, free the buffer and compare the
signature.
Bug 741430 - GParted cannot recognise LVM signature on unpartitioned
drive
Embedded devices (Android) use GPT partition names to identify
partitions, instead of file system labels. Add support for viewing and
changing them.
As partition names are used to provide unique identification they are
never copied when copying the contents of one partition to another.
Note that GNU/Linux uses file system labels, UUIDs or device names for
identification during the boot process and afterwards so while partition
names can be used, they are optional and purely for user information.
Bug 741424 - Add support for GPT partition names
This and the following few commits rename variables, methods, classes,
etc from *label_partition* to *label_filesystem* so that the code also
reflects that it is the label of the file system that is being modified
and to separate it from the name partition operation about to be added.
enum OPERATION_LABEL_PARTITION -> OPERATION_LABEL_FILESYSTEM
Bug 741424 - Add support for GPT partition names
Only recognises ReFS file system. No other actions are supported.
Requires blkid from util-linux >= 2.24.
Bug #738471 - ReFS file system is not recognised
Helper to check whether a recognised file system type is supported by
GParted or not. Supported means there is an implementation class and
will appear in the File System Support dialog.
Make supported_filesystem() a static member function so that it can be
called without a class object so that GParted_Core::GParted_Core()
initialiser isn't called multiple times. This requires FILESYSTEM_MAP
to become a static member variable too.
Bug #738471 - ReFS file system is not recognised
get_filesystem_object() takes a constant reference to a FILESYSTEM, but
FILESYSTEM is just an enumeration. So that's a pointer to a constant
int. Just pass by value instead.
RHEL / CentOS 5.6 and later officially support ext4 file system [1].
From RHEL / CentOS 5.3 ext4 file system was included as a technology
preview. Ext4 file system tools are in a separate package e4fsprogs,
using uniquely named commands. The standard e2fsprogs commands only
support ext2 and ext3 file systems.
# mkfs.ext4 /dev/sdb3
# tune2fs -l /dev/sdb3
tune2fs 1.39 (29-May-2006)
tune2fs: Filesystem has unsupported feature(s) while trying to open /dev/sdb3
Couldn't find valid filesystem superblock.
# echo $?
1
# tune4fs -l /dev/sdb3
tune4fs 1.41.12 (17-May-2010)
Filesystem volume name: <none>
Last mounted on: <not available>
Filesystem UUID: ba4a9d58-7728-4b47-8a90-80e772615637
Filesystem magic number: 0xEF53
Filesystem revision #: 1 (dynamic)
Filesystem features: has_journal ext_attr resize_inode dir_index filetype extent flex_bg sparse_super large_file huge_file uninit_bg dir_nlink extra_isize
...
For ext4 only, search for the e4fsprogs specific commands first and the
standard e2fsprogs commands second.
[1] RHEL 5.6 Release Notes, 5. Filesystems and Storage
https://access.redhat.com/documentation/en-US/Red_Hat_Enterprise_Linux/5/html/5.6_Release_Notes/ar01s05.html
Bug #738706 - GParted doesn't support ext4 on RHEL/CentOS 5.x
The device combobox was getting drawn blank, then getting drawn again
with the selected device. This was happening because at the start of
Win_GParted::refresh_combo_devices() the GTK model behind the combobox,
liststore_devices, was cleared, changing the active item, causing the
combobox to get redrawn empty. After the GTK model had been repopulated
the active item was reset causing the comboxbox to get redrawn again,
now showing the selected device. Call flow:
Win_GParted::refresh_combo_devices()
liststore_devices->clear()
//Gtk::Combobox emits signal_change. Registered callbacks
//called.
Win_GParted::combo_devices_changed()
Win_GParted::Refresh_Visual()
...
...
combo_devices.set_active(current_device);
//Gtk::Combobox emits signal_change. Registered callbacks
//called.
Win_GParted::combo_devices_changed()
Win_GParted::Refresh_Visual()
...
This has always been the case, since the device combobox was first added
to GParted before version 0.1 by commit:
3a4b43e0ad
replaced deprecated OptionMenu with ComboBox ...
Fix by temporarily blocking the devices comboxbox from emitting
signal_changed while the GTK model behind the combobox is recreated.
However, since automatic selection of the largest free space was added
[1] in GParted 0.15.0, a more noticeable flashing redraw issue was
caused in which the partition graphic and partition list were both drawn
blank then redrawn fully populated. Some distributions were not
affected by this at all, some only experienced a single flash and others
suffered from two or more flashing redraws. Some affected
distributions: CentOS 5.10, 6.5, 7.0, Debian 6, Fedora 14, 19, 20,
Ubuntu 13.10, Xubuntu 14.04 LTS. Did not occur on Kubuntu 12.04 LTS.
[1] 5b53c12f6e
Select largest unallocated partition by default (#667365)
Bug #696149 - Double refresh of display introduced with default
unallocated space
Remove HAVE_LIBPARTED_3_1_0_PLUS definition and replace Autoconf
check for libparted >= 3.1
Currently uses custom check which compiles an executable to check for
libparted version >= 3.1 to determine the availability of the library
parted-fs-resize and the need to include the header <parted/filesys.h>.
Change to use a direct Autoconf check to determine the availability of
the ped_file_system_resize() function in the parted-fs-resize library.
Remove inclusion of the header <parted/filesys.h> as it has always been
included via <parted/parted.h>, at least as far back as parted 1.8.0,
and even in parted 3.0 when ped_file_system_resize() and libparted file
system resize capability didn't exist.
Bug #734718 - Update Autoconf version specific libparted checks and
defines to feature specific ones
While one partition is busy, reformat another partition from the command
line. Afterwards parted/libparted still detects the original file
system and GParted shows errors from the file system specific tools
reporting the new file system doesn't exist. Only limitation is that
the new new file system must be recognised by libparted (or by GParted's
fallback file system signature detection).
Case #1, File system reformatting:
# parted /dev/sdb print
Model: ATA SAMSUNG SSD UM41 (scsi)
Disk /dev/sdb: 8012MB
Sector size (logical/physical): 512B/512B
Partition Table: msdos
Number Start End Size Type File system Flags
1 1049kB 2149MB 2147MB primary ext2
2 2149MB 4296MB 2147MB primary ext2
# mount | fgrep sdb
/dev/sdb1 on /mnt/1 type ext2 (rw)
# mkfs.xfs -f /dev/sdb2
# blkid /dev/sdb2
/dev/sdb2: UUID="c31823a2-b81b-46fa-8246-0a59695e4834" TYPE="xfs"
# parted /dev/sdb print
Model: ATA SAMSUNG SSD UM41 (scsi)
Disk /dev/sdb: 8012MB
Sector size (logical/physical): 512B/512B
Partition Table: msdos
Number Start End Size Type File system Flags
1 1049kB 2149MB 2147MB primary ext2
2 2149MB 4296MB 2147MB primary ext2
# e2label /dev/sdb2
e2label: Bad magic number in super-block while trying to open /dev/sdb2
Couldn't find valid filesystem superblock.
# dumpe2fs /dev/sdb2
dumpe2fs 1.41.12 (17-May-2010)
dumpe2fs: Bad magic number in super-block while trying to open /dev/sdb2
Couldn't find valid filesystem superblock.
Case #2, Removing device from multi-device btrfs:
# btrfs filesystem show /dev/sdb1
Label: none uuid: a05db434-efd5-4e8c-902f-05f89a88b610
Total devices 2 FS bytes used 156.00KB
devid 2 size 2.00GB used 512.00MB path /dev/sdb2
devid 1 size 2.00GB used 240.75MB path /dev/sdb1
# mount /dev/sdb1 /mnt/1
# btrfs device delete /dev/sdb2
# btrfs filesystem show /dev/sdb1
Label: none uuid: a05db434-efd5-4e8c-902f-05f89a88b610
Total devices 1 FS bytes used 92.00KB
devid 1 size 2.00GB used 714.25MB path /dev/sdb1
# btrfs filesystem show /dev/sdb2
and GParted reports this error for partition /dev/sdb2:
Unable to read the contents of this file system!
Because of this some operations may be unavailable.
The cause might be a missing software package.
The following list of software packages is required for btrfs
file system support: btrfs-tools.
This is another case of libparted reading from the whole disk device
(/dev/sdb) yet the file system tools use the partition specific block
device (/dev/sdb2), and the Linux buffer cache not providing cache
coherency. Previous scenario was fixed with:
797f0b8eeb
Flush device after wiping a file system (#688882)
This affects libparted 2.0 to 3.1 inclusive and is fixed by:
http://git.savannah.gnu.org/cgit/parted.git/commit/?id=fb99ba5ebd0dc34204fc9f1014131d5d494805bc
Revert "linux-commit: do not unnecessarily open partition device nodes"
Fix by calling ped_device_sync() to guarantee cache coherency for each
device during scanning.
Bug #723842 - GParted resizes the wrong filesystem (does not pass the
devid to btrfs filesystem resize)
With linux 3.5 and later, the device used to mount a btrfs file system
is updated in /proc/mounts when the previous mounting device is removed
from the file system. Most recent distributions make /etc/mtab a
symbolic link to /proc/mounts. However some still have /etc/mtab as a
plain file only updated by mount and umount, thus showing the old device
name which is no longer part of the file system.
On Ubuntu 13.10, which has /etc/mtab as a plain file managed by mount
and umount:
# mkfs.btrfs /dev/sdb1
# mount /dev/sdb1 /mnt/1
# btrfs device add /dev/sdb2 /mnt/1
# btrfs device delete /dev/sdb1 /mnt/1
# sync
# btrfs filesystem show /dev/sdb1
# btrfs filesystem show /dev/sdb2
Label: none uuid: e47775a6-e5ad-4fb4-9ea4-1570aa5b4009
Total devices 2 FS bytes used 28.00KB
devid 2 size 2.00GB used 272.00MB path /dev/sdb2
# fgrep btrfs /proc/mounts
/dev/sdb2 /mnt/1 btrfs rw,relatime,space_cache 0 0
# ls -l /etc/mtab
-rw-r--r-- 1 root root 842 Apr 15 19:41 /etc/mtab
# fgrep btrfs /etc/mtab
/dev/sdb1 /mnt/1 btrfs rw 0 0
This causes GParted to report /dev/sdb1 as busy and mounted at /mnt/1
when it is no longer mounted. This effects recent releases of Ubuntu,
13.04, 13.10 and 14.04.
Either /etc/mtab is a symlink and is identical to /proc/mounts or
/etc/mtab is a plain file with wrong information. Fix by not reading
mounted file systems from /etc/mtab.
However old distributions only contain 'rootfs' and '/dev/root' device
names for the / (root) file system with '/dev/root' being a block device
rather than a symlink to the true device. For example from CentOS 5.x:
# fgrep ' / ' /proc/mounts
rootfs / rootfs rw 0 0
/dev/root / ext3 rw,data=ordered 0 0
# ls -l /dev/root
brw------- 1 root root 8, 3 Jun 4 2013 /dev/root
This prevents identification, and therefore busy detection, of the
device containing the / (root) file system. Used to read /etc/mtab to
get the root file system device name.
# fgrep ' / ' /etc/mtab
/dev/sda3 / ext3 rw 0 0
# ls -l /dev/sda3
brw-r----- 1 root disk 8, 3 Jun 4 2013 /dev/sda3
As per commit:
409096f739
improved scanning for root mountpoint (/) ...
but, as discussed above, this contains an out of date device name after
the mounting device has been dynamically removed from a multi-device
btrfs, thus identifying the wrong device as busy. Instead fall back to
reading mounted file systems from the output of the mount command, but
only when required.
# mount | fgrep ' / '
/dev/sda3 on / type ext3 (rw)
Bug #723842 - GParted resizes the wrong filesystem (does not pass the
devid to btrfs filesystem resize)
Linux can only show a single device name in /proc/mounts and /etc/mtab
for each mounted btrfs, even if it is a multi-device file system. So
GParted only shows a mount point for one of the devices in the btrfs, no
matter how many devices are part of the file system.
# mkfs.btrfs /dev/sdb1 /dev/sdb2
# btrfs filesystem show /dev/sdb1
Label: none uuid: 36eb51a2-2927-4c92-820f-b2f0b5cdae50
Total devices 2 FS bytes used 156.00KB
devid 2 size 2.00GB used 512.00MB path /dev/sdb2
devid 1 size 2.00GB used 240.75MB path /dev/sdb1
# mount /dev/sdb1 /mnt/1
# grep btrfs /proc/mounts
/dev/sdb1 /mnt/1 btrfs rw,seclabel,relatime,ssd,space_cache 0 0
GParted only shows the mount point for /dev/sdb1 as /mnt/1, but nothing
for /dev/sdb2.
Make GParted report the same mount point for all devices included in a
multi-device btrfs file system.
Add btrfs specific get_mount_device() method to report the mounting
device, if any, for the btrfs file system in the occupying the device in
question. Uses the existing cache of btrfs file system device
membership. Also extract common code from GParted_Core::
set_mountpoints() into set_mountpoints_helper().
Bug #723842 - GParted resizes the wrong filesystem (does not pass the
devid to btrfs filesystem resize)
Busy detection of file systems works by checking if the device is
mounted (appears in the mount_info map). For a multi-device btrfs file
system this will only report one of the devices as busy, not all of
them.
# btrfs filesystem show /dev/sdb1
Label: none uuid: 36eb51a2-2927-4c92-820f-b2f0b5cdae50
Total devices 2 FS bytes used 156.00KB
devid 2 size 2.00GB used 512.00MB path /dev/sdb2
devid 1 size 2.00GB used 240.75MB path /dev/sdb1
# mount /dev/sdb1 /mnt/1
# grep btrfs /proc/mounts
/dev/sdb1 /mnt/1 btrfs rw,seclabel,relatime,ssd,space_cache 0 0
GParted will only report /dev/sdb1 as busy, but not /dev/sdb2.
Add btrfs specific is_busy() method which reports the device as busy if
any of the devices in the btrfs file system are mounted. This uses a
cache which maps device membership in all btrfs file systems. The cache
is cleared on GParted refresh and incrementally populated as each btrfs
partition is checked for busy status.
WARNING:
Removal of the mounting device from a btrfs file system makes it
impossible to determine whether the file system is mounted or not for
linux <= 3.4. This is because /proc/mounts continues to show the old
device which is no longer a member of the file system.
# btrfs device delete /dev/sdb1 /mnt/1
# sync
# grep btrfs /proc/mounts
/dev/sdb1 /mnt/1 btrfs rw,seclabel,relatime,ssd,space_cache 0 0
# btrfs filesystem show /dev/sdb1
# btrfs filesystem show /dev/sdb2
Label: none uuid: 36eb51a2-2927-4c92-820f-b2f0b5cdae50
Total devices 1 FS bytes used 28.00KB
devid 2 size 2.00GB used 1.02GB path /dev/sdb2
Fixed in linux 3.5 by commit:
Btrfs: implement ->show_devname
https://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/commit/?id=9c5085c147989d48dfe74194b48affc23f376650
Bug #723842 - GParted resizes the wrong filesystem (does not pass the
devid to btrfs filesystem resize)
Add static member function GParted_Core::is_dev_mounted() so that other
modules can determine if a particular partition contains a mounted file
system or not.
Make it a static member function so that it can be called without
needing the gparted_core object. Extend to make the group of
manipulated variables (mount_info, fstab_info) and manipulating
functions (init_maps(), read_mountpoints_from_file(),
read_mountpoints_from_file_swaps(), get_all_mountpoints()) static too.
Bug #723842 - GParted resizes the wrong filesystem (does not pass the
devid to btrfs filesystem resize)
GParted's primary inbuilt busy detection method is "is the partition
mounted?". A custom method is used for LVM2 PV because its not a
mounted file system.
Make busy detection selectable per file system type.
.fs.busy = FS::NONE (default)
No busy detection.
.fs.busy = FS::GPARTED
Use internal GParted method which checks if the partition is
mounted.
.fs.busy = FS:EXTERNAL
Call the file system type's member function is_busy().
LVM2 PV busy detection changes from a special case to just electing to
call the lvm2_pv::is_busy() method. Linux Software RAID remains a
special case because it's only recognised, but not otherwise supported.
Bug #723842 - GParted resizes the wrong filesystem (does not pass the
devid to btrfs filesystem resize)
Fragment of debugging and valgrind output:
D: tid=2193 main()
...
D: tid=2202 GParted_Core::set_devices_thread()
...
D: tid=2202 Utils::execute_command(command="dumpe2fs -h /dev/sda1", output, error, use_C_locale=1)
D: tid=2202 this=0x13fef4a0 PipeCapture::PipeCapture()
D: tid=2202 this=0x13fef4f0 PipeCapture::PipeCapture()
D: tid=2202 this=0x13fef4a0 PipeCapture::connect_signal()
D: sourceid=77
D: tid=2202 this=0x13fef4f0 PipeCapture::connect_signal()
D: sourceid=78
D: tid=2193 data=0x13fef4a0 PipeCapture::_OnReadable()
D: tid=2193 this=0x13fef4a0 PipeCapture::OnReadable()
D: signal_update.emit()
D: return true
D: tid=2193 data=0x13fef4f0 PipeCapture::_OnReadable()
D: tid=2193 this=0x13fef4f0 PipeCapture::OnReadable()
D: signal_update.emit()
D: return true
D: tid=2193 data=0x13fef4a0 PipeCapture::_OnReadable()
D: tid=2193 this=0x13fef4a0 PipeCapture::OnReadable()
D: signal_update.emit()
D: return true
D: tid=2193 data=0x13fef4f0 PipeCapture::_OnReadable()
D: tid=2193 this=0x13fef4f0 PipeCapture::OnReadable()
D: signal_eof.emit()
D: return false
D: (!rc) &(pc->sourceid)=0x13fef518
D: tid=2193 data=0x13fef4a0 PipeCapture::_OnReadable()
D: tid=2193 this=0x13fef4a0 PipeCapture::OnReadable()
D: signal_update.emit()
D: return true
D: tid=2193 data=0x13fef4a0 PipeCapture::_OnReadable()
D: tid=2193 this=0x13fef4a0 PipeCapture::OnReadable()
D: signal_update.emit()
D: return true
D: tid=2193 data=0x13fef4a0 PipeCapture::_OnReadable()
D: tid=2193 this=0x13fef4a0 PipeCapture::OnReadable()
D: signal_eof.emit()
D: tid=2202 this=0x13fef4f0 PipeCapture::~PipeCapture()
D: sourceid=0
D: tid=2202 this=0x13fef4a0 PipeCapture::~PipeCapture()
D: sourceid=77
D: return false
D: (!rc) &(pc->sourceid)=0x13fef4c8
==2193== Thread 1:
==2193== Invalid write of size 4
==2193== at 0x490580: GParted::PipeCapture::_OnReadable(_GIOChannel*, GIOCondition, void*) (PipeCapture.cc:56)
==2193== by 0x38662492A5: g_main_context_dispatch (gmain.c:3066)
==2193== by 0x3866249627: g_main_context_iterate.isra.24 (gmain.c:3713)
==2193== by 0x3866249A39: g_main_loop_run (gmain.c:3907)
==2193== by 0x3D7FD45C26: gtk_main (gtkmain.c:1257)
==2193== by 0x469743: GParted::GParted_Core::set_devices(std::vector<GParted::Device, std::allocator<GParted::Device> >&) (GParted_Core.cc:155)
==2193== by 0x4A78F1: GParted::Win_GParted::menu_gparted_refresh_devices() (Win_GParted.cc:1259)
==2193== by 0x4A7886: GParted::Win_GParted::on_show() (Win_GParted.cc:1253)
==2193== by 0x3D82B2009C: Gtk::Widget_Class::show_callback(_GtkWidget*) (widget.cc:3855)
==2193== by 0x3867210297: g_closure_invoke (gclosure.c:777)
==2193== by 0x3867221B86: signal_emit_unlocked_R (gsignal.c:3516)
==2193== by 0x386722A0F1: g_signal_emit_valist (gsignal.c:3330)
==2193== Address 0x13fef4c8 is not stack'd, malloc'd or (recently) free'd
==2193==
PipeCapture.cc (with debugging):
46 gboolean PipeCapture::_OnReadable( GIOChannel *source,
47 GIOCondition condition,
48 gpointer data )
49 {
50 std::cout << "D: tid=" << (long int)syscall(SYS_gettid) << " data=" << data << " PipeCapture::_OnReadable()" << std::endl;
51 PipeCapture *pc = static_cast<PipeCapture *>(data);
52 gboolean rc = pc->OnReadable( Glib::IOCondition(condition) );
53 if (!rc)
54 {
55 std::cout << "D: (!rc) &(pc->sourceid)=" << &(pc->sourceid) << std::endl;
56 pc->sourceid = 0;
57 }
58 return rc;
59 }
The use after free across threads only happens when an external program
is being executed from a thread other than the main() thread. This is
because by default glib registered callbacks are run by the glib main
loop, which is only called from the main() thread with Gtk::Main::run().
Event sequence:
tid=2193 tid=2202
main()
...
GParted_Core::set_devices()
Glib::Thread::create(... set_devices_thread ...)
Gtk::Main::run() GParted_Core::set_devices_thread()
...
Utils::execute_command("dumpe2fs ... /dev/sda1" ...)
Glib::spawn_async_with_pipes()
PipeCapture outputcapture(out, output)
outputcapture.connect_signal()
//Glib main loop runs callback
PipeCapture::_OnReadable()
pc->OnReadable()
//output read
signal_update.emit()
return true
...
//Glib main loop runs callback
PipeCapture::_OnReadable()
pc->OnReadable()
//eof reached
[1] signal_eof.emit()
return status.exit_status
[2] PipeCapture::~PipeCapture()
[3] return false
[4] pc->sourceid = 0
What is happening is that the PipeCapture destructor [2] is running in
the set_devices_thread() thread and freeing the object's memory as soon
as signal_eof.emit() [1] has been called. Then signal_eof.emit()
returns back to OnReadable() which then returns false [3] back to the
_OnReadable() callback function which then assigns 0 to sourceid member
variable [4] in the already freed object, detected by valgrind as:
Invalid write of size 4
at ... GParted::PipeCapture::_OnReadable(...) (PipeCapture.cc:56)
This is happening because PipeCapture member variable sourceid is being
saved, in a different thread, just so the _OnReadable() callback can be
removed. However a glib IOChannel callback, type GIOFunc(), returning
false will be automatically removed.
GLib Reference Manual 2.26 / IO Channels
https://developer.gnome.org/glib/2.26/glib-IO-Channels.html#GIOFunc
GIOFunc()
Returns : the function should return FALSE if the event source
should be removed
Therefore fix by just not saving the event sourceid at all, and not
calling g_source_remove() to manually remove the callback, but instead
letting glib automatically remove the callback when it returns false.
Bug #731752 - Write after free cross thread race in
PipeCapture::_OnReadable()
Mike Fleetwood