Git, quilt and binary searching
Generally, kernel source trees are maintained using two basic tools, git and quilt. While it is true that you can also use some git ”frontends”, like Cogito, Stacked GIT, Patchy Git (pg), (h)gct, which simplify user interactions with this tool, we will focus on git itself and on quilt, since each of them is strictly related to a specific way of maintaining a source tree.
Git is a versioning control system written by Linus Torvalds specifically for the maintenance of the main Linux kernel source tree (ie. the mainline). The current maintainer of git itself is Junio C. Hamano and the latest version is git-2.30.0.
Of course, to use git you need to install it in your system, but all of the contemporary Linux distributions provide git packages that can be installed in usual ways. The most important of these packages is usually called git-core, as it contains the most basic git components. Alternatively, you can download the source code of git from http://www.kernel.org/pub/software/scm/git/ and build it yourself.
Like some other versioning control systems (e.g. CVS), git assumes that there is a central ”master” directory tree used as a reference for all changes made to the maintained source code. This directory tree contains the source code itself along with the history of all changes made to it since certain point in time. To be registered in the ”master” tree, the changes must be accepted by its maintainer (e.g. by Linus, in the case of the Linux kernel ”master” tree) and after they have been registered (”committed” in the git-related terminology) all of the users of this tree can ”see” the source code with these changes applied. Still, they can also see how the source code looked like before specific modifications and they can follow its history (i.e. view all modifications made to it after given point that may be specified in many different ways). For this reason each modification of the source code is registered in the ”master” tree along with additional information including, among other things, the name and email address of its author and a short description of the purpose and scope of the modification (known as the ”changelog”).
If you create a copy of the Linux kernel ”master” tree, all of the information contained in it will be available to you locally and can be used for many different purposes. In particular, you can use them to identify the modifications that have introduced bugs.
To make a copy of this tree, or to ”clone” it in the git language, you can run
$ git clone git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git linux-git
This will download the entire Linus’ git tree to your machine and the local copy of it will be located in the subdirectory linux-git of the current directory (i.e. the one that was current when the above command was being executed). You should remember, however, that the Linus’ tree is huge, so the download may take quite a lot of time, depending on your Internet connection’s bandwidth and the current load of the kernel.org’s git server. If you are not going to use the git capabilities described below, it might be less cumbersome to download stand-alone patches and apply them as described in Section 1.2.
Once you have downloaded the Linus’ git tree, it generally is a good idea to tell git that it is the the tree that you want to start with. To do this, change directory to linux-git and run
$ git checkout -f
Now, you are able to do some really nice things with the help of git. For example, you can synchronize your local copy of the Linus’ tree with the original one by downloading only the changes that were committed by Linus after you had run git clone. For this purpose, run
$ git pull git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux-2.6.git $ git checkout
Of course, this means that you don’t need to run git clone every time the Linus’ tree changes. Moreover, you should not do this. In fact, it is only necessary do download the Linus’ tree once and then use git pull to update your local copy. Naturally, if you make your own changes to the local copy of the tree, git pull may fail, but then you can use ’git-checkout -f’ to revert all of the conflicting changes. Still, if you want to save your modifications, it is not that easy any more, but we won’t discuss this case any further (for more information see, for example, the tutorial at http://www.kernel.org/pub/software/scm/git/docs/tutorial.html).
Apart from the git commands presented above, you will probably find the following ones useful:
git whatchanged <file> – shows all changes aﬀecting given source ﬁle (the patch relative to the git tree’s root directory should be provided)
git bisect * – used for binary searching for buggy patches (see below)
git revert * – reverts given change (”commit” in the git terminology)
gitk – graphically visualizes the tree
git show – prints given commit
git log – shows the list of changes made to the tree
In our opinion particularly useful is the git log command that allows you to track the history of the source code. By running
you can see all changes made since the Linus’ tree was created. Well, albeit impressive, this is not particularly useful, because there are very many of them, but you can easily narrow the scope of its output. For example, the command
$ git-log v2.6.19..+
will print the changes made since the version 2.6.19 of the kernel (this won’t work for v2.6.21.., but we don’t know why), and the following one:
$ git-log --since="7 days ago"
will give you all changes made since – you have guessed it – 7 days ago. Isn’t it nice?
While playing with git log you can see that each change, or commit, in the log starts from the word commit and a long hexadecimal number. These numbers are unique commit identifiers that can be used for referring to specific commits in many git commands. For example, if you want to see the source code modifications associated with commit b5bf28cde894b3bb3bd25c13a7647020562f9ea0 in the form of a patch, run
$ git show b5bf28cde894b3bb3bd25c13a7647020562f9ea0
It sometimes is unnecessary to use the entire commit identifier here, since several initial characters may be sufficient to identify the object. Thus to get the same result as from the above command, it may be sufficient to run
$ git show b5bf28
Of course, if you want to save the resulting patch in a ﬁle, it is only necessary to do
$ git show b5bf28 > b5bf28.patch
where b5bf28.patch can be replaced with an arbitrary ﬁle name.
When you run such git commands as git clone or git pull, some data are being transferred over the Internet. In such cases you should always use the git:// protocol designed specifically for transferring git objects. Still, sometimes you may not be able to do this (e.g. if you are behind a ﬁrewall that you cannot conﬁgure) and then you can use the ”ordinary” http:// protocol, but it is not generally recommended.
It is worthy of noting that the Linus’ tree is not the only git tree downloadable from kernel.org . The other ones correspond to various development branches of the kernel introduced in Section 1.3. At http://git.kernel.org there is the list of all git trees available from kernel.org . Another list of kernel git tree is available from http://git.infradead.org .
To learn more about git, you can read the documentation distributed along with it or visit, for instance, the web page at http://www.kernel.org/pub/software/scm/git/docs
Quilt is a tool to manage series of patches. First of all, it makes the creation of patches easy and allows one to preserve the right ordering of patches within the series. It also automates the application and reverting of patches as well as the updating of patches in accordance with manual modifications of the source code. Usually, it can be installed from a package provided by your distribution, but you can also install it from the source code available at http://savannah.nongnu.org/projects/quilt/ . In principle, it can be used for managing modifications of an arbitrary set of text ﬁles located in a single directory tree, but we will focus on using it with respect to the Linux kernel source code.
Some Linux kernel developers use quilt as their primary patch management tool and some of the patchsets discussed in Section 1.4 are distributed as quilt-friendly series of patches. The most important of them is the -mm tree maintained by Andrew Morton (see Section 1.5).
A quilt-friendly series of patches consists of the patches themselves and the ﬁle series, in which the patch names are saved in the right order (they must be the same as the names of the ﬁles that contain the patches). To apply the patches, you only need to create a subdirectory called patches in the root of the kernel source tree and place the entire patch series, including the series ﬁle, in it. Then, you can use the ’quilt push -a’ command to apply them all in one shot.
Suppose, for example, that you want to apply the series of patches available at http://www.sisk.pl/kernel/hibernation_and_suspend/2.6.22-rc1/patches/ and consisting of the following ﬁles:
01-freezer-close-theoretical-race-between-refrigerator-and-thaw_tasks.patch 02-freezer-fix-vfork-problem.patch 03-freezer-take-kernel_execve-into-consideration.patch 04-fix-kthread_create-vs-freezer-theoretical-race.patch 05-fix-pf_nofreeze-and-freezeable-race-2.patch 06-move-frozen_process-to-kernel-power-processc.patch 07-power-management-use-mutexes-instead-of-semaphores.patch 08-swsusp-fix-sysfs-interface.patch 09-acpi-fix-suspend-resume-ordering.patch 10-swsusp-remove-platform-callbacks-from-resume-code.patch 11-swsusp-reduce-code-duplication-between-user_c-and-disk_c.patch series
on top of the 2.6.22-rc1 kernel source. For this purpose, go to the directory that contains the kernel sources, create the directory patches and copy the ﬁles that the series consists of into it. Next, run
$ quilt push -a
and all patches in the series will be applied. Alternatively, you can run
$ quilt push 11-swsusp-reduce-code-duplication-between-user_c-and-disk_c.patch
and quilt will apply all patches in the series up to and including the one given as the argument to ’quilt push’. The ordering of patches is based on the contents of the series ﬁle. That is, the patches the names of which are at the beginning of the series ﬁle are applied ﬁrst.
Generally, you should not modify the series ﬁle manually, but sometimes it is convenient to do that. Suppose, for instance, that you do not want the patch 09-acpi-fix-suspend-resume-ordering.patch
from the above series to be applied. In that case, you can place ’#’ before the name of the patch in the series ﬁle:
01-freezer-close-theoretical-race-between-refrigerator-and-thaw_tasks.patch 02-freezer-fix-vfork-problem.patch 03-freezer-take-kernel_execve-into-consideration.patch 04-fix-kthread_create-vs-freezer-theoretical-race.patch 05-fix-pf_nofreeze-and-freezeable-race-2.patch 06-move-frozen_process-to-kernel-power-processc.patch 07-power-management-use-mutexes-instead-of-semaphores.patch 08-swsusp-fix-sysfs-interface.patch #09-acpi-fix-suspend-resume-ordering.patch 10-swsusp-remove-platform-callbacks-from-resume-code.patch 11-swsusp-reduce-code-duplication-between-user_c-and-disk_c.patch
and that will make quilt regard this line as a comment and skip the patch. You can also change the ordering of patches by changing the ordering of names in the series ﬁle manually. Remember, however, not to change the ordering of names of the patches that have already been applied, because that will confuse quilt.
The applied patches are treated by quilt as though they were on a stack. The ﬁrst applied patch is on the bottom of the stack, while the last applied patch is on the top of it. That’s why ’quilt push’ is used to apply patches. Specifically, ’quilt push’ applies the next patch in the series that has not been applied yet and places it on the top of the stack. Analogously, ’quilt pop’ reverts the most recently applied patch (i.e. the one on the top of the stack) and removes it from the stack. The command ’quilt push -a’ applies all patches in the series that have not been applied yet and places them on the stack in the right order, while ’quilt pop -a’ reverts all of the applied patches and removes them from the stack, one by one.
You can make quilt print the name of the most recently applied patch (i.e. the one on the top of the stack) by running ’quilt top’, while ’quilt next’ will show you the name of the next patch to be applied. Similarly, ’quilt previous’ prints the name of the patch that has been applied right before the last one, ’quilt applied’ prints the names of all the currently applied patches (in the order in which they have been applied, so the name of the ”top” patch is printed last), and ’quilt series’ prints the names of all patches in the series.
You can also provide quilt with the number of patches to be applied or reverted. Namely, if you want it to apply the next two patches in the series, run
$ quilt push 2
Similarly, to make it revert the last two most recently applied patches, use
$ quilt pop 2
(these commands are very useful in binary searching for ”bad” patches, discussed in the next section). For example, having applied the ﬁrst two patches from the series introduced above, you may want to apply the next three patches:
$ quilt push 3 Applying patch 03-freezer-take-kernel_execve-into-consideration.patch patching file kernel/power/process.c Applying patch 04-fix-kthread_create-vs-freezer-theoretical-race.patch patching file kernel/kthread.c Applying patch 05-fix-pf_nofreeze-and-freezeable-race-2.patch patching file include/linux/freezer.h Now at patch 05-fix-pf_nofreeze-and-freezeable-race-2.patch
Now, quilt tells you that 05-fix-pf_nofreeze-and-freezeable-race-2.patch is on the top of the stack:
$ quilt top 05-fix-pf_nofreeze-and-freezeable-race-2.patch $ quilt next 06-move-frozen_process-to-kernel-power-processc.patch $ quilt previous 04-fix-kthread_create-vs-freezer-theoretical-race.patch
Next, suppose that you want to apply two patches more:
$ quilt push 2 Applying patch 06-move-frozen_process-to-kernel-power-processc.patch patching file include/linux/freezer.h patching file kernel/power/process.c Applying patch 07-power-management-use-mutexes-instead-of-semaphores.patch patching file drivers/base/power/main.c patching file drivers/base/power/power.h patching file drivers/base/power/resume.c patching file drivers/base/power/runtime.c patching file drivers/base/power/suspend.c Now at patch 07-power-management-use-mutexes-instead-of-semaphores.patch
Reverting of the three most recently applied patches is also simple:
$ quilt pop 3 Removing patch 07-power-management-use-mutexes-instead-of-semaphores.patch Restoring drivers/base/power/resume.c Restoring drivers/base/power/main.c Restoring drivers/base/power/runtime.c Restoring drivers/base/power/power.h Restoring drivers/base/power/suspend.c Removing patch 06-move-frozen_process-to-kernel-power-processc.patch Restoring kernel/power/process.c Restoring include/linux/freezer.h Removing patch 05-fix-pf_nofreeze-and-freezeable-race-2.patch Restoring include/linux/freezer.h Now at patch 04-fix-kthread_create-vs-freezer-theoretical-race.patch
Sometimes you may want quilt to revert patches until a specific patch is on the top of the stack. To do this, use ’quilt pop’ with an argument being the name of the patch that you want to be on the top of the stack after the operation. Suppose, for instance, that you have applied the ﬁrst ten patches from our example series, but now you want 04-fix-kthread_create-vs-freezer-theoretical-race.patch to be on the top of the stack (i.e. to become the last recently applied one). You can make this happen in the following way:
$ quilt pop 04-fix-kthread_create-vs-freezer-theoretical-race.patch Removing patch 10-swsusp-remove-platform-callbacks-from-resume-code.patch Restoring kernel/power/user.c Removing patch 09-acpi-fix-suspend-resume-ordering.patch Restoring kernel/power/main.c Removing patch 08-swsusp-fix-sysfs-interface.patch Restoring kernel/power/disk.c Restoring kernel/power/main.c Removing patch 07-power-management-use-mutexes-instead-of-semaphores.patch Restoring drivers/base/power/resume.c Restoring drivers/base/power/main.c Restoring drivers/base/power/runtime.c Restoring drivers/base/power/power.h Restoring drivers/base/power/suspend.c Removing patch 06-move-frozen_process-to-kernel-power-processc.patch Restoring kernel/power/process.c Restoring include/linux/freezer.h Removing patch 05-fix-pf_nofreeze-and-freezeable-race-2.patch Restoring include/linux/freezer.h Now at patch 04-fix-kthread_create-vs-freezer-theoretical-race.patch
Analogously, you can make quilt apply patches until a specific one is on the top of the stack:
$ quilt push 09-acpi-fix-suspend-resume-ordering.patch Applying patch 05-fix-pf_nofreeze-and-freezeable-race-2.patch patching file include/linux/freezer.h Applying patch 06-move-frozen_process-to-kernel-power-processc.patch patching file include/linux/freezer.h patching file kernel/power/process.c Applying patch 07-power-management-use-mutexes-instead-of-semaphores.patch patching file drivers/base/power/main.c patching file drivers/base/power/power.h patching file drivers/base/power/resume.c patching file drivers/base/power/runtime.c patching file drivers/base/power/suspend.c Applying patch 08-swsusp-fix-sysfs-interface.patch patching file kernel/power/disk.c patching file kernel/power/main.c Applying patch 09-acpi-fix-suspend-resume-ordering.patch patching file kernel/power/main.c Now at patch 09-acpi-fix-suspend-resume-ordering.patch
As you can see in the above examples, for each applied or reverted patch quilt prints the names of the ﬁles modified in the process (in fact, these names are printed by the patch program introduced in Sec. 1.2, used by quilt). You can also make it print the names of the ﬁles modified by the most recently applied patch (i.e. the one on the top of the stack) by using ’quilt files’:
$ quilt top 09-acpi-fix-suspend-resume-ordering.patch $ quilt files kernel/power/main.c
Additionally, it is possible to add some ”external” patches to a quilt series. The recommended way of doing this is to use the ’quilt import’ command with the additional argument being the name of the ﬁle containing the patch that you want to add to the series.
If this command is used, the ﬁle with the patch is automatically copied to the patches directory and the series ﬁle is updated by adding the name of this ﬁle right after the most recently applied one. Thus the patch becomes the next one to be applied by quilt, although it is not applied automatically.
Our description of quilt is by no means a complete one. It may only allow you to get a general idea of how quilt works and how ﬂexible it is. Generally speaking, it allows one to do much more than we have shown above. In particular, it automates the creation and updating of patches in a very convenient way, but this part of its functionality is beyond the scope of our discussion. For more information about quilt refer to the documentation distributed with it, available at http://download.savannah.gnu.org/releases/quilt/ .
4.3 General idea of binary searching
Imagine that you have found a kernel bug, but you have no idea which of the thousand or more patches included in the tree that you are testing might have caused it to appear. In principle, you could revert the patches one by one and test the kernel after reverting each of them, but that would take way too much time. The solution is to use binary searching, also known as bisection.
To explain what it is we ﬁrst assume that the kernel does not work for you any more after you have applied a series of n patches (n may be of the order of 1000). We also assume that it takes ∆t minutes on the average to compile and run the kernel on your system after reverting one patch (∆t may be of the order of 10). Under these assumptions, if you tried to revert patches one by one and test the kernel each time, you could spend about ∆t * n minutes worst-case before ﬁnding the oﬀending patch. Of course, you might be lucky and the bug might have been introduced by the last patch in the series, but generally this is not very probable. More precisely, if you know nothing about the patches in question, you should assume that each of them is equally likely to have introduced the bug and therefore the probability of the bug being introduced by a specific patch is equal to 1/n (hence, the worst case is not very probable either, but that does not improve the situation very much).
Naturally, you can revert k patches out of n and then test the kernel to see whether or not one of these k patches has introduced the bug. Still, the question arises what number k should be equal to. To answer it, let’s estimate the maximum expected time needed to find the patch that has introduced the bug under the assumption that k patches are reverted in the first step. For this purpose, let A(k) denote the event that one of the k patches reverted in the first step has introduced the bug and let B(k) denote the event that the bug has been introduced by one of the remaining n − k patches. If A(k) occurs, the time needed to find the buggy patch will not be greater than TA (k) = ∆t · k (we need to search a series of k patches and it takes ∆t minutes to check one patch on the average) and the probability of occurrence of A(k) is pA (k) = k/n. In turn, if B(k) occurs, the time needed to ﬁnd the buggy patch will not be greater than TB (k) = ∆t · (n − k) and the probability of occurrence of B(k) is pB (k) = (n − k)/n. Hence, since A(k) and B(k) are mutually exclusive and one of them must occur, the maximum average time needed to ﬁnd the buggy patch is given by
T (k) = TA (k)pA (k) + TB (k)pB (k) = ∆t/n [ k + (n − k)^2 ]
Now, it takes a little algebra to show that T (k) attains minimum for k = n/2, so it is most reasonable to revert n/2 patches in the ﬁrst step.
Further, if the kernel still doesn’t work after reverting n/2 patches from the series, we get the problem that is formally equivalent to the initial one with the number of patches to search reduced by half. Then, we can repeat the above reasoning for n = n/2 and it will turn out that the most reasonable thing we can do is to revert n /2 patches and test the kernel.
On the other hand, if the kernel works after we have reverted n/2 patches from the series, we need to reapply some of the reverted patches and test it once again. In that case, however, we still have n = n/2 patches to check and a reasoning analogous to the above one leads to the conclusion that the most reasonable number of patches to reapply before testing the kernel once again is n /2.
Thus we can establish a procedure, referred to as bisection or binary searching, allowing us to reduce the number of patches to check by half in every step. Namely, if the initial number of patches to check is n, we revert n/2 patches and test the kernel. Next, depending on the result of the previous test, we either revert or reapply n/4 patches and test the kernel once again. Subsequently, depending on the result of the previous test, we either revert or reapply n/8 patches and test the kernel once again. By repeating this approximately log2 n times we can reduce the number of suspicious patches to one and thus ﬁnd the patch that has introduced the bug. It is not very difficult to observe that the average time needed to ﬁnd the buggy patch this way is proportional to log2 n.
4.4 Binary searching with the help of quilt
Suppose that we have a quilt series of patches with the following contents of the series file:
01-kmemleak-base.patch 02-kmemleak-doc.patch 03-kmemleak-hooks.patch 04-kmemleak-modules.patch 05-kmemleak-i386.patch 06-kmemleak-arm.patch 07-kmemleak-false-positives.patch 08-kmemleak-keep-init.patch 09-kmemleak-test.patch 10-kmemleak-maintainers.patch #11-new-locking.patch #12-new-locking-fix.patch 13-vt-memleak-fix.patch 14-new-locking-fix.patch 15-fix-for-possible-leak-in-delayacctc.patch
and we cannot build the kernel after applying all of the above patches. To find the patch that causes the problem to appear, we can carry out a binary search.
Of course, having only 13 patches to check, we can use the ’quilt patches <file_name>’ command to identify patches modifying the ﬁle which fails to compile, so that we can revert them and test the kernel without them, but if the tested patchset is huge, it usually is not a good idea to revert random patches from the middle of it.
Suppose that we have applied all of the patches from the series:
$ quilt push -a Applying patch patches/01-kmemleak-base.patch patching file include/linux/kernel.h [..] patching file kernel/delayacct.c Now at patch patches/15-fix-for-possible-leak-in-delayacctc.patch
and it turns out that something is wrong, because we cannot compile the kernel any more. First, we obtain the number of applied patches:
$ quilt applied | wc -l 13
and we revert one half of them:
$ quilt pop 6 Removing patch patches/15-fix-for-possible-leak-in-delayacctc.patch Restoring include/linux/delayacct.h [..] Restoring lib/Kconfig.debug Now at patch patches/07-kmemleak-false-positives.patch
Then, the number of patches to revert or apply in the next step is 3 (we must remember this number, so it is best to write it down somewhere). We test the kernel and ﬁnd that the problem is still appearing, so we revert 3 patches:
$ quilt pop 3 [...] Now at patch patches/04-kmemleak-modules.patch
and note that the number of patches to check in the next step is 1.
Next, we test the kernel again and ﬁnd that the problem has disappeared. We thus apply one patch:
$ quilt push [...] Now at patch patches/05-kmemleak-i386.patch
and note that the number of patches to check in the next step will depend on the result of the subsequent kernel test. Namely, if we test the kernel and the problem reappears, we will know that 05-kmemleak-i386.patch is the source of it, since none of the ”earlier” patches has caused it to be present. Otherwise, we will need to apply one patch more
$ quilt push [...] Now at patch patches/06-kmemleak-arm.patch
and if the problem reappears, we will know that this patch is ”guilty”. Otherwise, the next patch, which is 07-kmemleak-false-positives.patch, will have to be the offending one, since the problem is not present with all of the patches ”below” it applied.
Some more useful hints about using quilt for binary searching can be found in the Andrew Morton’s article How to perform bisection searches on -mm trees (http://www.zip.com.au/~akpm/linux/patches/stuff/bisecting-mm-trees.txt). In particular, you can learn from it that if the -mm tree contains a series of patches similar to the following one:
patch-blah.patch patch-blah-blah.patch patch-blah-blah-fix1.patch patch-blah-blah-fix2.patch patch-blah-blah-fix3.patch
you should treat them as a single patch. That is, you should always revert them all or apply them all at once.
A practical example of binary searching with the help of quilt is shown in the movie available at http://www.youtube.com/watch?v=LS_hTnBDIYk .
4.5 Binary searching with the help of git-bisect
Since binary searching for buggy patches is a very important debugging technique, as far as the Linux kernel is concerned, and git is the primary versioning control system used by the kernel developers, the git package contains a tool that automates binary searching, called git bisect. It is powerful and easy to use, so it can be recommended to inexperienced kernel testers.
To explain how to use git bisect, we suppose that there is a problem in the 2.6.18-rc5 kernel and we do not know which commit has introduced it. We know, however, that it is not present in the 2.6.18-rc4 kernel.
We start the bisection by running
$ git bisect start
and mark the current kernel version as the ﬁrst known bad one:
$ git bisect bad
Next, we use gitk to get the commit identifier associated with the 2.6.18-rc4 version of the kernel, which is 9f737633e6ee54fc174282d49b2559bd2208391d, and mark this version as the last known good one:
$ git bisect good 9f737633e6ee54fc174282d49b2559bd2208391d
Now, git bisect will select a commit, more or less equally distant from the two corresponding to the first known bad and the last known good kernel versions, that we should test:
Bisecting: 202 revisions left to test after this [c5ab964debe92d0ec7af330f350a3433c1b5b61e] spectrum_cs: Fix firmware uploading errors
(using ’git bisect visualize’ we can see the current status of the bisection in gitk). Then, we need to compile, install and test the kernel.
Suppose that we have done it and the problem is still appearing, so we mark the current kernel version (i.e. the one corresponding to the commit previously selected by git-bisect) as the ﬁrst known bad one and git-bisect selects the next commit to test:
$ git bisect bad Bisecting: 101 revisions left to test after this [1d7ea7324ae7a59f8e17e4ba76a2707c1e6f24d2] fuse: fix error case in fuse_readpages
We compile the kernel, install and test it. Suppose that this time the problem is not present, so we mark the current kernel version, corresponding to the last commit selected by git bisect, as the last known good one and let git bisect select another commit:
git bisect good Bisecting: 55 revisions left to test after this [a4657141091c2f975fa35ac1ad28fffdd756091e] Merge email@example.com:/pub/scm/linux/kernel/git/davem/net-2.6
Next, we compile, install and test the kernel, and so on, until we get a message similar to the following one:
$ git bisect good 1d7ea7324ae7a59f8e17e4ba76a2707c1e6f24d2 is first bad commit commit 1d7ea7324ae7a59f8e17e4ba76a2707c1e6f24d2 Author: Jan Kowalski <a@b> Date: Sun Aug 13 23:24:27 2006 -0700 [PATCH] fuse: fix error case in fuse_readpages Don’t let fuse_readpages leave the @pages list not empty when exiting on error. [...]
which contains the identifier of the commit that, most probably, has introduced the bug.
Still, we need to make sure that the bug has really been introduced by this particular commit. For this purpose we return to the initial kernel version:
$ git bisect reset
and try to revert the commit that we have identified as the source of the problem with the help of git-bisect:
git revert 1d7ea7324ae7a59f8e17e4ba76a2707c1e6f24d2
If we are lucky, the commit will be reverted cleanly and we will be able to test the kernel without it to make sure that it is buggy. Otherwise, there are some commits that depend on this one and in fact we should revert them all for the final testing.
Although the binary searching in the above example is pretty straightforward, generally it can be more complicated. For example, we may be unable to test the commit selected by git bisect, because some more commits must be applied so that we can build the kernel. In that case we need to tell git bisect where to continue and ’git reset --hard’ can be used for this purpose.
Suppose, for instance, that you have:
$ git bisect start $ git bisect bad 5ecd3100e695228ac5e0ce0e325e252c0f11806f $ git bisect good f285e3d329ce68cc355fadf4ab2c8f34d7f264cb Bisecting: 41 revisions left to test after this [c1a13ff57ab1ce52a0aae9984594dbfcfbaf68c0] Merge branch ’upstream-linus’ of master.kernel.org:/pub/scm/linux/kernel/git/jgarzik/netdev-2.6
but you cannot compile the kernel version corresponding to the selected commit. Then, you can use ’git reset --hard’ to manually select the next commit for git bisect:
$ git reset --hard c4d36a822e7c51cd6ffcf9133854d5e32489d269 HEAD is now at c4d36a8... Pull osi-now into release branch
Now, after running ’git bisect visualize’ you will see that the commits 5ecd3100e695... and f285e3d329ce... are still marked as ”bad” and ”good”, respectively, but the commit c4d36a822e7c51cd6ffcf9133854d5e32489d269 that you have selected is marked as ”bisect” instead of c1a13ff57ab1ce52a0aae9984594dbfcfbaf68c0.
It is also possible that a new version of the kernel appears on the kernel.org server while you are carrying out a bisection search. In that case you may suspect that the bug under investigation has been fixed in the new kernel version, but nevertheless you may want to continue the bisection if it turns out not to be the case. Then, you can run
$ cp .git/BISECT_LOG ../bisect.replay $ git bisect reset
and use git pull, as usual, to get the new kernel version (see Section 4.1). After testing it, if the bug is still present, you can do:
$ git bisect replay ../bisect.replay
to return to the point at which you have ”suspended” the binary search.
To summarize, the most often used commands related to binary searching with the help of git bisect are:
git bisect start – starts a new binary search
git bisect reset – goes back to the initial kernel version and ﬁnishes the bisection
git bisect good <commit> – marks the commit given by <commit>, or the current commit, as corresponding to the last known good kernel version
git bisect bad <commit> – marks the commit given by <commit>, or the current commit, as corresponding to the ﬁrst known bad kernel version
git bisect visualize – uses gitk to show changes between the last known good and the first known bad kernel versions
git bisect replay – may be used to return to the point at which the bisection has been ”suspended” in order to test a new version of the kernel
git reset --hard – can be used to select the next commit for git-bisect manually
For more information about git bisect see its manual page (’man git-bisect’) that contains more detailed description of it. You can also refer to the git documentation available on the Web (e.g. at http://www.kernel.org/pub/software/scm/git/docs).
A practical example of binary searching with the help of git bisect is shown in the movie available at http://www.youtube.com/watch?v=R7_LY-ceFbE
Although in general binary searching allows one to ﬁnd patches that introduce reproducible bugs relatively quickly, you should be aware that it may fail, as well as any other debugging technique.
First of all, it tends to single out patches that expose problems, but they need not be the same as the patches that actually introduce them. Usually, the patch that introduces bugs breaks the kernel, but sometimes the breakage results from some other changes that would work just ﬁne if the bugs were not present. You should always remember about it, especially when you report the problem to the kernel developers (ie. never assume blindly that the patch identified as ”bad” by a bisection must be buggy).
Second, in some quite rare situations, the patch that introduces the problem observed by you simply cannot be identified. For example, there may be two or more patches that introduce problems with similar symptoms and in that case it is difficult to say what you are really observing. Apart from this, the buggy patch may belong to a series of several patches depending on each other and you can only compile the kernel with either all of them, or none of them applied. Moreover, two or more such series of patches may be mixed within the tree in a fashion that makes them impossible to untangle. If that happens, you will only be able to identify a range of commits including the buggy patch and that need not be very useful.
It is also possible to make a mistake in binary searching, even if you are using git-bisect. Yes, it is. For instance, if you run ’git-bisect good’ instead of ’git-bisect bad’, or vice versa, by mistake at one point, the entire procedure will lead to nowhere. Another common mistake is to run ’make oldconfig’ (see Section 1.6) in every step of bisection in such a way that the kernel conﬁguration is slightly different each time. This may cause the options that actually trigger the bug to be turned on and oﬀ in the process, in which case the binary search will not lead to any consistent result. To avoid this, it is recommended to preserve the kernel conﬁguration ﬁle known to trigger the bug and copy it to the build directory (usually it is the same as the kernel source directory – see Section 1.6 for details) before each compilation of the kernel (you may be asked to set or unset some options anyway and in that case you will need to remember the appropriate settings and apply them every time in the same way).
Finally, if you know something about the series of patches containing the buggy one, you can make the binary search for it end (or converge, as it is usually said) faster. Namely, knowing what changes are made by different patches in the series, you can manually skip the patches that are surely not related to the observed problem. For example, if you have found a bug related to a ﬁle system, you should be able to omit the patches that are only related to networking, since most likely they have nothing to do with the bug. This way you can save several kernel compilations, which may be equivalent to a fair amount of time.