The current -mm prepatch is 2.6.11-rc4-mm1. Recent changes to -mm include device mapper multipath support (see below), the cpushare "secure computing" patch, a SCSI changer driver, a new set of BIO support functions, some performance counter updates, and various fixes.
The current 2.4 prepatch is 2.4.30-pre2, released by Marcelo on February 23. This prepatch adds a new set of fixes (mostly in the networking subsystem) and a few filesystem and driver updates.
Kernel development newsa patch which adds a new degree of user-space control over memory management policy. In particular, it creates a new /proc entry:
If a suitably privileged process writes one or more NUMA node numbers to that file, all pages belonging to that node which are found in the page cache will be flushed out. Essentially, this operation causes a node to forget about all locally-cached pages from files in the filesystem.
Clearing the page cache in this way would normally be bad for performance. The page cache exists to allow the filesystem to satisfy common filesystem requests without going to the disk; clearing the cache defeats that functionality and would normally be undesirable. There are exceptions to everything, however. This patch is aimed at large-scale high-performance computing tasks running in a cluster environment. Such jobs typically do best if they can start with a clean system; they have no real use for whatever may have been cached for the previous user. More to the point, a full page cache can cause memory allocations to be satisfied with non-local (slower) memory, resulting in significantly worse performance. By clearing the cache before starting a new job, a system administrator can ensure that local memory is available for that job.
Not everybody likes the patch. Ingo Molnar thinks that this capability will create confusion and make the debugging of memory problems even harder.
Andrew Morton, instead, sees the value of the patch for some users, but doesn't like the implementation. He would like to see this capability made useful for other classes of users, such as kernel developers who want to put the system into a known state before running tests. He also doesn't like the /proc interface, and argues for a new system call instead. His suggestion was:
sys_free_node_memory(long node_id, long pages_to_make_free, long what_to_free);
This form of the call would allow the clearing of something less than the entire page cache, making the tool a bit less crude. The what_to_free argument would be a bitmask specifying which types of memory to free; beyond the page cache, this call could cause the kernel to reclaim anonymous memory or slab caches.
The system call approach would seem to make sense; there is one remaining glitch, however: SUSE already shipped the /proc interface in SLES9. That revelation drew a complaint from Andrew:
An explicit purpose behind the 2.6 development model is to get patches into the mainline quickly so that their form can be stabilized before distributors ship them. As the developers become used to this mode of operation, this sort of issue should become relatively rare.
Support for multipath in Linux has traditionally been spotty, at best. Some low-level block drivers have included support for their specific devices, but support at that level leads to duplicated functionality and difficulties for administrators. Some thought has gone into how multipath is best supported: does that logic belong at the driver layer, the SCSI mid-layer, the block layer, or somewhere else? The conclusion that was reached at last year's Kernel Summit was that the device mapper was the best place for multipath support.
That support has now been coded up and posted for review; it was added to the 2.6.11-rc4-mm1 kernel. When used with the user-space multipath tools distribution, the device mapper can now provide proper multipath support - for some hardware, at least.
Internally, the multipath code creates a data structure, attached to a device mapper target, which looks like this:
When time comes to transfer blocks to or from a device mapper target representing a multipath device, the code goes to the first priority group in the list. Each group represents a set of paths to the device, each of which is considered equal to the others; the preferred paths (being the fastest and/or most reliable) should be contained in the first group in the list. Priority groups include a path selector - a function which determines which path should be used for each I/O request. The current patches include a round-robin selector which simply rotates through the paths to balance the load across them. Should situations arise which require more complicated policies, it should not be tremendously difficult to create an appropriate path selector.
If a given path starts to generate errors, it is marked as failed and the path selector will pass over it. Should all paths in a priority group fail, the next group in the list (if it exists) will be used. The multipath tools include a management daemon which is informed of failed paths; its job is to scream for help and retry the failed paths. If a path starts to work again, the daemon will inform the device mapper, which will resume using that path.
There may be times when no paths are available; this can happen, for example, when a new priority group has been selected and is in the process of initializing itself. In this situation, the multipath target will maintain a queue of pending BIO structures. Once a path becomes available, a special worker thread works through the pending I/O list and sees to it that all requests are executed.
At the lower level, the multipath code includes a set of hardware hooks for dealing with hardware-specific events. These hooks include a status function, an initialization function, and an error handler. The patch set includes a hardware handler for EMC CLARiiON devices.
Comments on the patches have been relatively few, and have dealt mostly with trivial issues. The multipath patches are unintrusive; they add new functionality, but do not make significant changes to existing code. So chances are good that they could find their way into the 2.6.12 kernel.stumbled across a case where the FUTEX code can deadlock the system (thus failing the "fast" test) which shows how hard it can be to get concurrency issues right.
One of the many locking primitives provided by the kernel is the reader-writer semaphore, or "rwsem". An rwsem can be obtained for either read or write access. Any number of readers will be allowed to hold the semaphore concurrently. Any thread which must change the protected data structures must, however, obtain the semaphore for write access. Only one writer is allowed at any given time, and no readers may be in the critical section while the writer is at work.
If a thread tries to obtain an rwsem for write access, and that semaphore is currently held (by somebody else) for read access, the writer will be put to sleep. Once the writer gets in line, however, no more readers will be allowed in. Once the existing readers have gotten out of the way, the writer will be allowed to proceed. The queued readers will only wake up after the writer is done. This implementation makes rwsems fair, in that readers cannot starve writers indefinitely. It also makes certain types of subtle faults possible, however.
If a process might have to wait on a FUTEX, the kernel must obtain that process's memory map semaphore (mmap_sem). This semaphore, which is an rwsem, controls access to the internal FUTEX data structures; it is taken for read access. The kernel must also query the value of the FUTEX itself, which is done through a call to get_user(). Should that access generate a page fault, the fault handler will obtain mmap_sem for read access a second time. This double access works just fine; the second down_read() call simply looks like another reader, which can run concurrently with the first.
Life gets complicated, however, when other processes share the same address space. Since the FUTEX mechanism is aimed at threads, this is a situation which comes about frequently. Consider the following series of events:
|Thread 1||Thread 2|
|(goes to sleep)|
|(everything comes to a halt)|
When the second process calls mmap(), it must obtain mmap_sem for write access. Since the first process is already a reader, the down_write() call is queued and the process is put to sleep. When the first process makes its get_user() call, it tries to obtain the rwsem for read access for the second time. Since there is now a writer waiting, however, the first process also is put to sleep. Since the first process is the one holding the initial read lock, this situation will never resolve itself; it is a deadlock. This particular type of deadlock is nasty in that it requires a race condition to become visible; things usually just work.
Several possible solutions have been proposed. The rwsem "lock depth" could be explicitly tracked so that a second attempt to obtain read access simply implements a counter and does not sleep. Processes holding mmap_sem could be marked with a special PF_MMAP_SEM flag; the page fault code would see that flag, realize that the semaphore is already held, and not take it again. Olof's initial report included a patch which tries to explicitly fault in the page before taking the semaphore so that the get_user() call would not generate a fault.
The solution which will eventually be adopted will likely take a different approach, however. Conventional wisdom has long said that functions like get_user() cannot be called in atomic context (in an interrupt handler or when a spinlock is held), since they might sleep. In fact, if the user-space access functions generate a page fault in atomic context, the fault handler simply refuses to bring in the page and the function returns an error code. So the solution, first suggested by Linus, is to put the process into an atomic mode (by calling inc_preempt_count()) just before the get_user() call. If get_user() fails, the page must be faulted in. So the mmap_sem is dropped, the page is explicitly faulted, and the whole process starts over again.
As often happens, the full solution turned out to be a bit more complicated than initially thought. So Olof put together a patch implementing a new user-space access function:
int get_user_inatomic(value, user_pointer);
This function is atomic; it may succeed or fail, but it will always return without sleeping. Like get_user(), it is implemented as a macro which tries to do the right thing regardless of the data type of the value to be fetched. That implementation drew a complaint from one developer, who would rather see new interfaces done in a more strongly-typed manner. So the details of the patch that eventually gets merged (presumably after 2.6.11) may change, but it will likely follow this approach.
Patches and updates
Core kernel code
Filesystems and block I/O
Page editor: Jonathan Corbet
Next page: Distributions>>
Copyright © 2005, Eklektix, Inc.
Comments and public postings are copyrighted by their creators.
Linux is a registered trademark of Linus Torvalds