They aren't.

No sane person should use frame buffers if they have the choice.

Like your mama told you: "Just say no". Use text-mode and X11, and be happy.

Some people don't have the choice, of course.

This is regarding a change I had to make to the kernel 2.2.17-14 to support really large drives. In our project we deal with huge drives, sometimes drives larger than 128GB. The file system is FAT32 and only one partition is create on any drive. During our testing we realized that linux fat implementation doesn't support partitions larger than 128GB(actually 64GB because of a bug in fat implementation).

This limitation is imposed by the data structures used by the linux fat implementation to read/write directory entries. A 'long' data type is used to access directory entries on the disk, of which only 28 bits are used to identify the sector that contains the directory entry and the rest 4 bits are used to specify offset of the directory entry inside the sector. Using 28 bits to identify a sector means we cannot access sectors beyond 128GB (2^28*512), thus limiting us from creating partitions larger than 128GB on large disk drives.

I have made changes to fat, vfat and msdos file system implementations in the kernel to use larger data types, thus allowing us to create larger partitions. As per the GPL I would like to make the patch available to everyone and also in case somebody has run into the same problem(who cares about fat in the linux world). The patch has been fairly well tested only on our systems(p3, 700MHz with FC). I truly appreciate if you & anybody in the kernel mailing list have any feedback about the changes.

The patch is applicable to only 2.2.17-14 kernel and may require changes to use with other kernel versions. As far as I know none of the kernel versions provide any fix for this problem.

That's Microsoft's problem -- that's a fundamental limit of the format they defined. The fact that they defined it in the first place is part of the problem (the only way you can make a FAT filesystem work *well* is by loading the entire FAT into memory ahead of time, and "FAT32" breaks that), instead of creating a more sensible replacement.

(FWIW, the reason they used to justify FAT32 was "it would be too much work to make DOS handle NTFS", as if those were the only options...)

now that new-year's parties are over things are getting boring again. For those who want to see and perhaps even try something more complex, i'm announcing this patch that is a pretty radical rewrite of the Linux scheduler for 2.5.2-pre6:

http://redhat.com/~mingo/O(1)-scheduler/sched-O1-2.5.2-A0.patch

for 2.4.17:

http://redhat.com/~mingo/O(1)-scheduler/sched-O1-2.4.17-A0.patch

Goal

The main goal of the new scheduler is to keep all the good things we know and love about the current Linux scheduler:

• good interactive performance even during high load: if the user types or clicks then the system must react instantly and must execute the user tasks smoothly, even during considerable background load.
• good scheduling/wakeup performance with 1-2 runnable processes.
• fairness: no process should stay without any timeslice for any unreasonable amount of time. No process should get an unjustly high amount of CPU time.
• priorities: less important tasks can be started with lower priority, more important tasks with higher priority.
• SMP efficiency: no CPU should stay idle if there is work to do.
• SMP affinity: processes which run on one CPU should stay affine to that CPU. Processes should not bounce between CPUs too frequently.
• plus additional scheduler features: RT scheduling, CPU binding.

and the goal is also to add a few new things:

• fully O(1) scheduling. Are you tired of the recalculation loop blowing the L1 cache away every now and then? Do you think the goodness loop is taking a bit too long to finish if there are lots of runnable processes? This new scheduler takes no prisoners: wakeup(), schedule(), the timer interrupt are all O(1) algorithms. There is no recalculation loop. There is no goodness loop either.
• 'perfect' SMP scalability. With the new scheduler there is no 'big' runqueue_lock anymore - it's all per-CPU runqueues and locks - two tasks on two separate CPUs can wake up, schedule and context-switch completely in parallel, without any interlocking. All scheduling-relevant data is structured for maximum scalability. (see the benchmark section later on.)

• better SMP affinity. The old scheduler has a particular weakness that causes the random bouncing of tasks between CPUs if/when higher priority/interactive tasks, this was observed and reported by many people. The reason is that the timeslice recalculation loop first needs every currently running task to consume its timeslice. But when this happens on eg. an 8-way system, then this property starves an increasing number of CPUs from executing any process. Once the last task that has a timeslice left has finished using up that timeslice, the recalculation loop is triggered and other CPUs can start executing tasks again - after having idled around for a number of timer ticks. The more CPUs, the worse this effect.

  Furthermore, this same effect causes the bouncing effect as well: whenever there is such a 'timeslice squeeze' of the global runqueue, idle processors start executing tasks which are not affine to that CPU. (because the affine tasks have finished off their timeslices already.)

  The new scheduler solves this problem by distributing timeslices on a per-CPU basis, without having any global synchronization or recalculation.

• batch scheduling. A significant proportion of computing-intensive tasks benefit from batch-scheduling, where timeslices are long and processes are roundrobin scheduled. The new scheduler does such batch-scheduling of the lowest priority tasks - so nice +19 jobs will get 'batch-scheduled' automatically. With this scheduler, nice +19 jobs are in essence SCHED_IDLE, from an interactiveness point of view.
• handle extreme loads more smoothly, without breakdown and scheduling storms.
• O(1) RT scheduling. For those RT folks who are paranoid about the O(nr_running) property of the goodness loop and the recalculation loop.
• run fork()ed children before the parent. Andrea has pointed out the advantages of this a few months ago, but patches for this feature do not work with the old scheduler as well as they should, because idle processes often steal the new child before the fork()ing CPU gets to execute it.

Design

(those who find the following design issues boring can skip to the next, 'Benchmarks' section.)

the core of the new scheduler are the following mechanizms:

• *two*, priority-ordered 'priority arrays' per CPU. There is an 'active' array and an 'expired' array. The active array contains all tasks that are affine to this CPU and have timeslices left. The expired array contains all tasks which have used up their timeslices - but this array is kept sorted as well. The active and expired array is not accessed directly, it's accessed through two pointers in the per-CPU runqueue structure. If all active tasks are used up then we 'switch' the two pointers and from now on the ready-to-go (former-) expired array is the active array - and the empty active array serves as the new collector for expired tasks.
• there is a 64-bit bitmap cache for array indices. Finding the highest priority task is thus a matter of two x86 BSFL bit-search instructions.

the split-array solution enables us to have an arbitrary number of active and expired tasks, and the recalculation of timeslices can be done immediately when the timeslice expires. Because the arrays are always access through the pointers in the runqueue, switching the two arrays can be done very quickly.

this is a hybride priority-list approach coupled with roundrobin scheduling and the array-switch method of distributing timeslices.

• there is a per-task 'load estimator'.

one of the toughest things to get right is good interactive feel during heavy system load. While playing with various scheduler variants i found that the best interactive feel is achieved not by 'boosting' interactive tasks, but by 'punishing' tasks that want to use more CPU time than there is available. This method is also much easier to do in an O(1) fashion.

to establish the actual 'load' the task contributes to the system, a complex-looking but pretty accurate method is used: there is a 4-entry 'history' ringbuffer of the task's activities during the last 4 seconds. This ringbuffer is operated without much overhead. The entries tell the scheduler a pretty accurate load-history of the task: has it used up more CPU time or less during the past N seconds. [the size '4' and the interval of 4x 1 seconds was found by lots of experimentation - this part is flexible and can be changed in both directions.]

the penalty a task gets for generating more load than the CPU can handle is a priority decrease - there is a maximum amount to this penalty relative to their static priority, so even fully CPU-bound tasks will observe each other's priorities, and will share the CPU accordingly.

I've separated the RT scheduler into a different codebase, while still keeping some of the scheduling codebase common. This does not look pretty in certain places such as __sched_tail() or activate_task(), but i dont think it can be avoided. RT scheduling is different, it uses a global runqueue (and global spinlock) and it needs global decisions. To make RT scheduling more instant, i've added a broadcast-reschedule message as well, to make it absolutely sure that RT tasks of the right priority are scheduled apropriately, even on SMP systems. The RT-scheduling part is O(1) as well.

the SMP load-balancer can be extended/switched with additional parallel computing and cache hierarchy concepts: NUMA scheduling, multi-core CPUs can be supported easily by changing the load-balancer. Right now it's tuned for my SMP systems.

i skipped the prev->mm == next->mm advantage - no workload i know of shows any sensitivity to this. It can be added back by sacrificing O(1) schedule() [the current and one-lower priority list can be searched for a that->mm == current->mm condition], but costs a fair number of cycles during a number of important workloads, so i wanted to avoid this as much as possible.

• the SMP idle-task startup code was still racy and the new scheduler triggered this. So i streamlined the idle-setup code a bit. We do not call into schedule() before all processors have started up fully and all idle threads are in place.
• the patch also cleans up a number of aspects of sched.c - moves code into other areas of the kernel where it's appropriate, and simplifies certain code paths and data constructs. As a result, the new scheduler's code is smaller than the old one.

(i'm sure there are other details i forgot to explain. I've commented some of the more important code paths and data constructs. If you think some aspect of this design is faulty or misses some important issue then please let me know.)

(the current code is by no means perfect, my main goal right now, besides fixing bugs is to make the code cleaner. Any suggestions for simplifications are welcome.)

Benchmarks

i've performed two major groups of benchmarks: first i've verified the interactive performance (interactive 'feel') of the new scheduler on UP and SMP systems as well. While this is a pretty subjective thing, i found that the new scheduler is at least as good as the old one in all areas, and in a number of high load workloads it feels visibly smoother. I've tried a number of workloads, such as make -j background compilation or 1000 background processes. Interactive performance can also be verified via tracing both schedulers, and i've done that and found no areas of missed wakeups or imperfect SMP scheduling latencies in either of the two schedulers.

the other group of benchmarks was the actual performance of the scheduler. I picked the following ones (some were intentionally picked to load the scheduler, others were picked to make the benchmark spectrum more complete):

• compilation benchmarks
• thr chat-server workload simulator written by Bill Hartner
• the usual components from the lmbench suite
• a heavily sched_yield()-ing testcode to measure yield() performance.

[ i can test any other workload too that anyone would find interesting. ]

i ran these benchmarks on a 1-CPU box using a UP kernel, a 2-CPU and a 8-CPU box as well, using the SMP kernel.

The chat-server simulator creates a number of processes that are connected to each other via TCP sockets, the processes send messages to each other randomly, in a way that simulates actual chat server designs and workloads.

3 successive runs of './chat_c 127.0.0.1 10 1000' produce the following message throughput:

vanilla-2.5.2-pre6:

Average throughput : 110619 messages per second
Average throughput : 107813 messages per second
Average throughput : 120558 messages per second

O(1)-schedule-2.5.2-pre6:

Average throughput : 131250 messages per second
Average throughput : 116333 messages per second
Average throughput : 179686 messages per second

this is a rougly 20% improvement.

To get all benefits of the new scheduler, i ran it reniced, which in essence triggers round-robin batch scheduling for the chat server tasks:

3 successive runs of 'nice -n 19 ./chat_c 127.0.0.1 10 1000' produce the following throughput:

vanilla-2.5.2-pre6:

Average throughput : 77719 messages per second
Average throughput : 83460 messages per second
Average throughput : 90029 messages per second

O(1)-schedule-2.5.2-pre6:

Average throughput : 609942 messages per second
Average throughput : 610221 messages per second
Average throughput : 609570 messages per second

throughput improved by more than 600%. The UP and 2-way SMP tests show a similar edge for the new scheduler. Furthermore, during these chatserver tests, the old scheduler doesnt handle interactive tasks very well, and the system is very jerky. (which is a side-effect of the overscheduling situation the scheduler gets into.)

the 1-CPU UP numbers are interesting as well:

vanilla-2.5.2-pre6:

./chat_c 127.0.0.1 10 100
Average throughput : 102885 messages per second
Average throughput : 95319 messages per second
Average throughput : 99076 messages per second

nice -n 19 ./chat_c 127.0.0.1 10 1000
Average throughput : 161205 messages per second
Average throughput : 151785 messages per second
Average throughput : 152951 messages per second

O(1)-schedule-2.5.2-pre6:

./chat_c 127.0.0.1 10 100 # NEW
Average throughput : 128865 messages per second
Average throughput : 115240 messages per second
Average throughput : 99034 messages per second

nice -n 19 ./chat_c 127.0.0.1 10 1000 # NEW
Average throughput : 163112 messages per second
Average throughput : 163012 messages per second
Average throughput : 163652 messages per second

this shows that while on UP we dont have the scalability improvements, the O(1) scheduler is still slightly ahead.

another benchmark measures sched_yield() performance. (which the pthreads code relies on pretty heavily.)

on a 2-way system, starting 4 instances of ./loop_yield gives the following context-switch throughput:

vanilla-2.5.2-pre6

  # vmstat 5 | cut -c57-
     system         cpu
   in    cs  us  sy  id
  102 241247   6  94   0
  101 240977   5  95   0
  101 241051   6  94   0
  101 241176   7  93   0

O(1)-schedule-2.5.2-pre6

  # vmstat 5 | cut -c57-
     system         cpu
   in     cs  us  sy  id
  101 977530  31  69   0
  101 977478  28  72   0
  101 977538  27  73   0

the O(1) scheduler is 300% faster, and we do nearly 1 million context switches per second!

this test is even more interesting on the 8-way system, running 16 instances of loop_yield:

vanilla-2.5.2-pre6:

   vmstat 5 | cut -c57-
      system         cpu
    in     cs  us  sy  id
   106 108498   2  98   0
   101 108333   1  99   0
   102 108437   1  99   0

100K/sec context switches - the overhead of the global runqueue makes the scheduler slower than the 2-way box!

O(1)-schedule-2.5.2-pre6:

    vmstat 5 | cut -c57-
     system         cpu
    in      cs  us  sy  id
   102 6120358  34  66   0
   101 6117063  33  67   0
   101 6117124  34  66   0

this is more than 6 million context switches per second! (i think this is a first, no Linux box in existence did so many context switches per second before.) This is one workload where the per-CPU runqueues and scalability advantages show up big time.

here are the lat_proc and lat_ctx comparisons (the results quoted here are the best numbers from a series of tests):

vanilla-2.5.2-pre6:

./lat_proc fork
Process fork+exit: 440.0000 microseconds
./lat_proc exec
Process fork+execve: 491.6364 microseconds
./lat_proc shell
Process fork+/bin/sh -c: 3545.0000 microseconds

O(1)-schedule-2.5.2-pre6:

./lat_proc fork
Process fork+exit: 168.6667 microseconds
./lat_proc exec
Process fork+execve: 279.6500 microseconds
./lat_proc shell
Process fork+/bin/sh -c: 2874.0000 microseconds

the difference is pretty dramatic - it's mostly due to avoiding much of the COW overhead that comes from fork()+execve(). The fork()+exit() improvement is mostly due to better CPU affinity - parent and child are running on the same CPU, while the old scheduler pushes the child to another, idle CPU, which creates heavy interlocking traffic between the MM structures.

the compilation benchmarks i ran gave very similar results for both schedulers. The O(1) scheduler has a small 2% advantage in make -j benchmarks (not accounting statistical noise - it's hard to produce reliable compilation benchmarks) - probably due to better SMP affinity again.

Status

i've tested the new scheduler under the aforementioned range of systems and workloads, but it's still experimental code nevertheless. I've developed it on SMP systems using the 2.5.2-pre kernels, so it has the most testing there, but i did a fair number of UP and 2.4.17 tests as well. NOTE! For the 2.5.2-pre6 kernel to be usable you should apply Andries' latest 2.5.2pre6-kdev_t patch available at:

http://www.kernel.org/pub/linux/kernel/people/aeb/

i also tested the RT scheduler for various situations such as sched_yield()-ing of RT tasks, strace-ing RT tasks and other details, and it's all working as expected. There might be some rough edges though.

Please don't take it bad, I don't want to flame you nor anybody, but I think that if everyone publicly announces his own tree with his own set of changes against the main kernel, many users will be lost quickly.

Once, we had "only" 3 main trees for the stable release :

• Linus' official kernels
• Alan's who did an excellent job at combining a stable core with experimental drivers
• Andrea's kernel which is more oriented towards big servers and very high loads.

Now that Alan is working on something else, I can easily understand that people need a branch like the one he maintained, even if the majority of his work has been merged into the main tree.

But there is now an -mjc tree, an -nj tree, perhaps others I missed, and many more not announced here (-wt, I remember of Mathias Andree's -ma for example who may have been the first one to merge ext3 and reiserfs into a same kernel). All of them include nearly the same set of fixes that have not yet got into the official kernel, plus a set of more-or-less stable, more-or-less interesting features (depending who the targets are). At least, each one should announce the goal of his kernel, and who it is intended to : developpers, production users, desktop users, testers, all of them ? With your announce, nobody even knows if he takes some particular risks using features from your kernel. Example: I believe that at least Andre Pavenis still has problems with Doug's i810_audio driver, so this cannot be annouced simply as a "fix" without a little note.

I sure know it takes a lot of time maintaining a set of patches against a mainstream kernel, and it even takes more time reading bug reports and determining what is stable and what isn't. Not to tell about the dozens of compilations before an announcement (because you compile them, don't you?), and occasional porting of some fixes to other architectures.

So I really think it would be more productive if people worked around a smaller set of trees, stopped editing patches by hand again and again to resolve the same conflicts and tried to be a bit more understandable for newbies who are a bit lost when they don't know what kernel to try.

Perhaps you could have sent your patches to Mickael Cohen to help him release an -mjc2 more quickly, like we all did with Alan ? Or perhaps you have a clear idea in mind about what your kernel tree will be, but in this case, please elaborate on this a bit more than this simple post, and prepare to cope with bug reports from people who will trust your tree. This last point may be the major reason why I chose to keep my kernels for my friends and I...

I hope you didn't take it as an attack, it wasn't really against you, nor to start a flame war, but just to make people think a bit about something more constructive.

I am the author of the 'Linux kernel source finder' web page that lists for each architecture the place to get appropriate Linux kernel patches - see:

http://www.treblig.org/Linux_kernel_source_finder.html

I wish to extend this to include pointers to the best/latest/most appropriate binutils for each architecture. I've put links in for x86, Alpha and MIPS to H.J.Lu's ftp site, since he tests for those 3 platforms prior to release.

I'd appreciate recommendations and comments from those using binutils on Linux for other platforms, with links to ftp, cvs or web pages describing the solutions for those architectures.