LWN Weekly Edition Front pageSecurity Kernel development Distributions Development Linux in the news Announcements Letters to the editor ->One big page
This page Previous weekFollowing week |
Kernel developmentRelease status Kernel release status The current development kernel is 2.6.0-test6, which was finally released by Linus on September 27. The most significant change in this kernel, perhaps, is the long-awaited expansion of dev_t to 32 bits (as covered in last week's LWN Kernel Page). Other patches which have been merged include a device mapper update, some NFS updates, a big I2C update, Con Kolivas's and Ingo Molnar's scheduler interactivity patches, a Coda filesystem update, some initramfs tweaks, improvements in random driver locking, the removal of some ext3 debugging hooks, direct I/O support for reiserfs, some CPU frequency work, numerous power management patches, an Intel SpeedStep-SMI driver, a substantial amount of janitorial work, and lots of fixes. The long-format changelog has the details.Note that 2.6.0-test6 changed the semantics of /proc slightly as part of an effort to get thread information represented properly. That change may be reverted in a future release. Linus's BitKeeper tree contains, as of this writing, a janitorial effort to move EXPORT_SYMBOL calls out of ksyms.c to the places where the exported symbols are actually defined, some IDE driver updates, and lots of fixes. The current stable kernel is 2.4.22. Marcelo released 2.4.23-pre6 on October 1; it fixes a number of ACPI problems, adds SCTP support, and includes the usual pile of fixes and updates.
Kernel development news How many groups do you need? The Linux kernel currently allows any process to be a member of up to 32 groups at the same time. As with many such limits, the 32-group allowance is sufficient for most users, but not for all. In fact, some users have a need for significantly more group memberships than that.This problem came to Rusty Russell's attention; it seems there is a client down under with a need for almost 200 simultaneous groups. The client is running Samba-based servers and having to deal with NT's hierarchical group structure. So Rusty sent out a patch making the maximum number of groups adjustable via a sysctl file. If a process expands to more than 32 groups, the code will dynamicly allocate an array to store them all. Rusty's patch asked for comments, and got them in the form of a rather exasperated note from Tim Hockin. Tim has been posting a very similar patch for quite some time; LWN included a version of this patch back in October, 2002. Despite repeated posts, the patch had never been picked up and merged. In response to Tim's note, Rusty set about the task of merging the two patches. It turns out that Tim's needs were a little different, however; his customers need, for some reason, to have processes be members of up to 10,000 groups. That changes things a bit. So, among other things, the combined patch will use vmalloc() to allocate the array of group IDs should an attempt with kmalloc() fail. Since vmalloc() can create large virtual arrays in kernel space through the use of page tables, it works better when large, virtually contiguous chunks of memory are required. That approach didn't fly with Linus, however; he asked: "Have you been looking at glibc sources lately, or why do you believe that we should encourage insane usage?" These comments have led Tim to take a turn reworking the patch. In his version, the kernel maintains a list of individual pages full of group IDs when the need arises. The lists are shared and reference counted (to help avoid unnecessary copying); they are also sorted so that searches are quick. Given that Linus is "definitely happier about this one," there is a distinct possibility that it may yet find its way into the 2.6 kernel. Sometimes, getting your patch in is just a matter of waiting for somebody else to reimplement it.
Defining the kernel/user space interface Joerg Schilling has a complaint. If you try to include the kernel header file include/linux/scsi/scsi.h in a user-space program, a number of compilation errors are the result. Joerg had reported the problem a good year ago, but it remains unfixed. "Is there no interest in user applications for kernel features or is there just no kernel maintainer left over who makes the needed work?"The initial reaction from the kernel developers was rather unsympathetic. In the modern world, applications are not supposed to be including kernel header files directly. Instead, they should use whatever sanitized version is provided by glibc. Anybody who tries to include kernel headers into user-space applications should not complain when things fail to work. In the real world, however, such a simple answer is not sufficient. The kernel exports a vast number of interfaces; consider just the various ioctl() calls offered by innumerable device drivers and subsystems, for example. Even if the communications between the kernel and glibc developers were better than they are, it would be difficult to expect the glibc people to keep up with every obscure interface that gets added to the kernel. As David Miller put it:
Even if one is of the opinion that nobody should be including the
kernel headers, you must fully realize that as a matter of
practicality people absolutely must do this to use many kernel
interfaces to their full extent. Suggest changes to fix the
problems, but just saying "don't include kernel header in your user
apps, NYAH NYAH NYAH!" does not help anyone at all.
For the short term, David suggests that the relevant kernel header files should simply be fixed so that they work when included into user-space programs. In the longer term, there is clearly a need to get a better handle on what, exactly, the interface between kernel and user space is. One approach that is being taken in this direction is to push kernel interfaces into virtual filesystems such as sysfs. Interfaces defined in this way are explicit and visible, and they rarely require any C header file support at all. The virtual filesystem approach works well in many cases, but it is unlikely to replace other kernel interfaces entirely. So it is still necessary to, somehow, better define the kernel's binary application interface. To that end, Andries Brouwer has posted a patch bringing back the idea of a separate set of "Linux ABI" include files. In this rendition, there would be a new include/linuxabi directory (along with architecture-specific variations) which would contain header files defining constants and data structures used for communication with user space. They would be specifically intended to be included from user space. Andries starts by removing the various mount option flags from <linux/fs.h> and putting them in <linuxabi/mountflags.h>. This idea has come up before, but it has never been adopted in the mainline kernel. Getting to a point where a significant part of the kernel binary interface is documented in include/linuxabi will clearly take a long time. Rearranging the kernel include file hierarchy may not be an appropriate thing to do at this point in the development cycle. It might not be a bad idea for early in the 2.7 series, however.
The Alan's Diary Translation Project For those wanting to follow Alan Cox's progress as he disconnects from kernel development and heads off to pursue an MBA: Paul Sladen has been translating his diary from the original Welsh. "So, the next three days to finish hacking the VIA CLE266, unsubscribe from all the lists and buy some paper folders. Except in 3 days, awful food, woe and having to get up in the morning... I hope I can remember the hang of studying."
Software suspend work funded Nigel Cunningham, the developer responsible for making software suspend actually work in the 2.4 kernel, has recently announced that his work is now being funded by LinuxFund.org. This grant will enable him to complete his 2.4 work, and to work at bringing it forward to the 2.6 kernel as well. For those who are interested in following Nigel's progress, he has set up a weblog on the LinuxFund.org site.
Driver porting The zen of kobjects
Part of the difficulty in understanding the driver model - and the kobject abstraction upon which it is built - is that there is no obvious starting place. Dealing with kobjects requires understanding a few different types, all of which make reference to each other. In an attempt to make things easier, we'll take a multi-pass approach, starting with vague terms and adding detail as we go. To that end, here are some quick definitions of some terms we will be working with.
We'll look at how to create and manipulate all of these types. A bottom-up approach will be taken, so we'll go back to kobjects.
Embedding kobjectsIt is rare (even unknown) for kernel code to create a standalone kobject; instead, kobjects are used to control access to a larger, domain-specific object. To this end, kobjects will be found embedded in other structures. If you are used to thinking of things in object-oriented terms, kobjects can be seen as a top-level, abstract class from which other classes are derived. A kobject implements a set of capabilities which are not particularly useful by themselves, but which are nice to have in other objects. The C language does not allow for the direct expression of inheritance, so other techniques - such as structure embedding - must be used.So, for example, the 2.6.0-test6 version of struct cdev, the structure describing a char device, is:
struct cdev {
struct kobject kobj;
struct module *owner;
struct file_operations *ops;
struct list_head list;
};
If you have a struct cdev structure, finding its embedded kobject is just a matter of using the kobj pointer. Code that works with kobjects will often have the opposite problem, however: given a struct kobject pointer, what is the pointer to the containing structure? You should avoid tricks (such as assuming that the kobject is at the beginning of the structure) and, instead, use the container_of() macro, found in <linux/kernel.h>:
container_of(pointer, type, member) where pointer is the pointer to the embedded kobject, type is the type of the containing structure, and member is the name of the structure field to which pointer points. The return value from container_of() is a pointer to the given type. So, for example, a pointer to a struct kobject embedded within a struct cdev called "kp" could be converted to a pointer to the containing structure with:
struct cdev *device = container_of(kp, struct cdev, kobj);
Programmers will often define a simple macro for "back-casting" kobject pointers to the containing type.
Initialization of kobjectsCode which creates a kobject must, of course, initialize that object. Some of the internal fields are setup with a (mandatory) call to kobject_init():
void kobject_init(struct kobject *kobj);
Among other things, kobject_init() sets the kobject's reference count to one. Calling kobject_init() is not sufficient, however. Kobject users must, at a minimum, set the name of the kobject; this is the name that will be used in sysfs entries. If you dig through the kernel source, you will find code which copies a string directly into the kobject's name field, but that approach should be avoided. Instead, use:
int kobject_set_name(struct kobject *kobj, const char *format, ...);
This function takes a printk-style variable argument list. Believe it or not, it is actually possible for this operation to fail; conscientious code should check the return value and react accordingly. The other kobject fields which should be set, directly or indirectly, by the creator are its ktype, kset, and parent. We will get to those shortly.
Reference countsOne of the key functions of a kobject is to serve as a reference counter for the object in which it is embedded. As long as references to the object exist, the object (and the code which supports it) must continue to exist. The low-level functions for manipulating a kobject's reference counts are:
struct kobject *kobject_get(struct kobject *kobj);
void kobject_put(struct kobject *kobj);
A successful call to kobject_get() will increment the kobject's reference counter and return the pointer to the kobject. If, however, the kobject is already in the process of being destroyed, the operation will fail and kobject_get() will return NULL. This return value must always be tested, or no end of unpleasant race conditions could result. When a reference is released, the call to kobject_put() will decrement the reference count and, possibly, free the object. Note that kobject_init() sets the reference count to one, so the code which sets up the kobject will need to do a kobject_put() eventually to release that reference. Note that, in many cases, the reference count in the kobject itself may not be sufficient to prevent race conditions. The existence of a kobject (and its containing structure) may well, for example, require the continued existence of the module which created that kobject. It would not do to unload that module while the kobject is still being passed around. That is why the cdev structure we saw above contains a struct module pointer. The reference counting for struct cdev is implemented as follows:
struct kobject *cdev_get(struct cdev *p)
{
struct module *owner = p->owner;
struct kobject *kobj;
if (owner && !try_module_get(owner))
return NULL;
kobj = kobject_get(&p->kobj);
if (!kobj)
module_put(owner);
return kobj;
}
Creating a reference to a cdev structure requires creating a reference also to the module which owns it. So cdev_get() uses try_module_get() to attempt to increment that module's usage count. If that operation succeeds, kobject_get() is used to increment the kobject's reference count as well. That operation could fail, of course, so the code checks the return value from kobject_get() and releases its reference to the module if things don't work out.
Hooking into sysfsAn initialized kobject will perform reference counting without trouble, but it will not appear in sysfs. To create sysfs entries, kernel code must pass the object to kobject_add():
int kobject_add(struct kobject *kobj);
As always, this operation can fail. The function:
void kobject_del(struct kobject *kobj);
will remove the kobject from sysfs. There is a kobject_register() function, which is really just the combination of the calls to kobject_init() and kobject_add(). Similarly, kobject_unregister() will call kobject_del(), then call kobject_put() to release the initial reference created with kobject_register() (or really kobject_init()).
ktypes and release methodsOne important thing still missing from the discussion is what happens to a kobject when its reference count reaches zero. The code which created the kobject generally does not know when that will happen; if it did, there would be little point in using a kobject in the first place. Even predicatable object lifecycles become more complicated when sysfs is brought in; user-space programs can keep a reference to a kobject (by keeping one of its associated sysfs files open) for an arbitrary period of time.The end result is that a structure protected by a kobject cannot be freed before its reference count goes to zero. The reference count is not under the direct control of the code which created the kobject. So that code must be notified asynchronously whenever the last reference to one of its kobjects goes away. This notification is done through a kobject's release() method. Usually such a method has a form like:
void my_object_release(struct kobject *kobj)
{
struct my_object *mine = container_of(kobj, struct my_object, kobj);
/* Perform any additional cleanup on this object, then... */
kfree (mine);
}
One important point cannot be overstated: every kobject must have a release() method, and the kobject must persist (in a consistent state) until that method is called. If these constraints are not met, the code is flawed. Interestingly, the release() method is not stored in the kobject itself; instead, it is associated with the ktype. So let us introduce struct kobj_type:
struct kobj_type {
void (*release)(struct kobject *);
struct sysfs_ops *sysfs_ops;
struct attribute **default_attrs;
};
This structure is used to describe a particular type of kobject (or, more correctly, of containing object). Every kobject needs to have an associated kobj_type structure; a pointer to that structure can be placed in the kobject's ktype field at initialization time, or (more likely) it can be defined by the kobject's containing kset. The release field in struct kobj_type is, of course, a pointer to the release() method for this type of kobject. The other two fields (sysfs_ops and default_attrs) control how objects of this type are represented in sysfs; they are beyond the scope of this document.
ksetsIn many ways, a kset looks like an extension of the kobj_type structure; a kset is a collection of identical kobjects. But, while struct kobj_type concerns itself with the type of an object, struct kset is concerned with aggregation and collection. The two concepts have been separated so that objects of identical type can appear in distinct sets.A kset serves these functions:
In object-oriented terms, "kset" is the top-level container class; ksets inherit their own kobject, and can be treated as a kobject as well.
For initialization and setup, ksets have an interface very similar to that
of kobjects. The following functions exist:
void kset_init(struct kset *kset);
int kset_add(struct kset *kset);
int kset_register(struct kset *kset);
void kset_unregister(struct kset *kset);
For the most part, these functions just call the analogous kobject_ function on the kset's embedded kobject. For managing the reference counts of ksets, the situation is about the same:
struct kset *kset_get(struct kset *kset);
void kset_put(struct kset *kset);
A kset, too, has a name, which is stored in the embedded kobject. So, if you have a kset called my_set, you would set its name with:
kobject_set_name(my_set->kobj, "The name");
Ksets also have a pointer (in the ktype field) to the kobj_type structure describing the kobjects it contains. This type will be applied to any kobject which does not contain a pointer to its own kobj_type structure. Another attribute of a kset is a set of hotplug operations; these operations are invoked whenever a kobject enters or leaves the kset. They are able to determine whether a user-space hotplug event is generated for this change, and to affect how that event is presented. The hotplug operations are beyond the scope of this document; they will be discussed later with sysfs. One might ask how, exactly, a kobject is added to a kset, given that no functions which perform that function have been presented. The answer is that this task is handled by kobject_add(). When a kobject is passed to kobject_add(), its kset member should point to the kset to which the kobject will belong. kobject_add() will handle the rest. There is currently no other way to add a kobject to a kset without directly messing with the list pointers. Finally, a kset contains a subsystem pointer (called subsys). So it must be time to talk about subsystems.
SubsystemsA subsystem is a representation for a high-level portion of the kernel as a whole. It is actually a simple structure:
struct subsystem {
struct kset kset;
struct rw_semaphore rwsem;
};
A subsystem, thus, is really just a wrapper around a kset. In fact, life is not quite that simple; a single subsystem can contain multiple ksets. This containment is represented by the subsys pointer in struct kset; so, if there are multiple ksets in a subsystem, it will not be possible to find all of them directly from the subsystem structure. Every kset must belong to a subsystem; the subsystem's rwsem semaphore is used to serialize access to a kset's internal linked list. Subsystems are often declared with a special macro:
decl_subsys(char *name, struct kobj_type *type,
struct kset_hotplug_ops *hotplug_ops);
This macro just creates a struct subsystem (its name is the name given to the macro with _subsys appended) with the internal kset initialized with the given type and hotplug_ops. Subsystems have the usual set of setup and teardown functions:
void subsystem_init(struct subsystem *subsys);
int subsystem_register(struct subsystem *subsys);
void subsystem_unregister(struct subsystem *subsys);
struct subsystem *subsys_get(struct subsystem *subsys)
void subsys_put(struct subsystem *subsys);
Most of these operations just act upon the subsystem's kset.
Kobject initialization againNow that we have covered all of that stuff, we can talk in detail about how a kobject should be prepared for its existence in the kernel. Here are all of the struct kobject fields which must be initialized somehow:
Often, much of the initialization of a kobject is handled by the layer that manages the containing kset. Thus, to get back to our old example, a char driver might create a struct cdev, but it need not worry about setting any of the fields in the embedded kobject - except for the name. Everything else is handled by the char device layer.
Looking forwardSo far, we have covered the operations used to set up and manipulate kobjects. The core concept is relatively simple: kobjects can be used to (1) maintain a reference count for an object and clean up when the object is no longer used, and (2) create a hierarchical data structure through kset membership.What is missing so far is how kobjects represent themselves to user space. The sysfs interface to kobjects makes it easy to export information to (and to receive information from) user space. The symbolic linking features of sysfs allow the creation of pointers across distinct kobject hierarchies. Stay tuned for a description of how all that works.
Patches and updates Kernel trees
Core kernel code
Device drivers
Filesystems and block I/O
Kernel building
Networking
Architecture-specific
Security-related
Benchmarks and bugs
Miscellaneous
Page editor: Jonathan Corbet |
Copyright © 2003, Eklektix, Inc.
Comments and public postings are copyrighted by their creators.
Linux is a registered trademark of Linus Torvalds
Powered by Rackspace Managed Hosting.