2.4.21pre3aa1

From: Andrea Arcangeli (andrea@suse.de)
Date: Mon Jan 20 2003 - 23:06:28 EST


URL:

        http://www.us.kernel.org/pub/linux/kernel/people/andrea/kernels/v2.4/2.4.21pre3aa1.gz
        http://www.us.kernel.org/pub/linux/kernel/people/andrea/kernels/v2.4/2.4.21pre3aa1/

diff between 2.4.21pre2aa2 and 2.4.21pre3aa1:

Only in 2.4.21pre2aa2: 00_apm-do_read-1
Only in 2.4.21pre2aa2: 20_apm-o1-sched-1
Only in 2.4.21pre2aa2: 9970_corename-2.gz

        Merged in mainline.

Only in 2.4.21pre2aa2: 00_extraversion-15
Only in 2.4.21pre3aa1: 00_extraversion-16
Only in 2.4.21pre2aa2: 00_sched-O1-aa-2.4.19rc3-7.gz
Only in 2.4.21pre3aa1: 00_sched-O1-aa-2.4.19rc3-8.gz
Only in 2.4.21pre2aa2: 10_o_direct-open-check-2
Only in 2.4.21pre3aa1: 10_o_direct-open-check-3
Only in 2.4.21pre2aa2: 30_09_o_direct-1
Only in 2.4.21pre3aa1: 30_09_o_direct-2
Only in 2.4.21pre2aa2: 70_xfs-sysctl-1
Only in 2.4.21pre3aa1: 70_xfs-sysctl-2
Only in 2.4.21pre3aa1: 9910_shm-largepage-10.gz
Only in 2.4.21pre2aa2: 9910_shm-largepage-9.gz

        Rediffed.

Only in 2.4.21pre3aa1: 00_o_direct-align-1

        Fix alignment check.

Only in 2.4.21pre3aa1: 00_one_highpage_init-cleanup-1

        Cleanup from J.A. Magallon backported from 2.5.

Only in 2.4.21pre2aa2: 00_semop-timeout-1
Only in 2.4.21pre3aa1: 00_semop-timeout-2

        Add ia32 emulation of the syscall to ia64 (forwarded by Andrew).

Only in 2.4.21pre3aa1: 00_vmalloc-ltp-crash-1

        Fix oops with ltp (still log, doesn't need to be rate limited
        since it's a privilegied call and printing it is useful
        for debugging but it doesn't crash).

Only in 2.4.21pre3aa1: 05_vm_22_vm-anon-lru-1

        SMP scalability feature that increase performance of 20% on some
        16-way. VM will be less aware when it's time to start swapping, this
        may be even better for some DB workload, but it could be a little worse
        for very small boxes. The first seconds of swap could be less smooth.
        If you want the VM to always have visibility of anon pages just set
        this sysctl to 1: /proc/sys/vm/vm_anon_lru. Default is set to zero.
        After the first swapouts the anon pages will be left into the lru and
        the behaviour will be equivalent to vm_anon_lru set to 1.

Only in 2.4.21pre3aa1: 05_vm_23_per-cpu-pages-1

        Add per-cpu-pages in front of the allocator, this brings several dozen
        points percent to some 16-way too, this along with the vm-anon-lru
        drops all the global locks in the anonymous memory allocation page fault
        fast path, and if applied together they nearly doubles performance of
        very parallel SMP workloads.

        This patch is based on the original code from Ingo Molnar but the
        refill/allocation decisions are quite different, and this as well is
        made to work fine with the per-point-of-view watermarks.

Only in 2.4.21pre2aa2: 10_rawio-vary-io-16
Only in 2.4.21pre3aa1: 10_rawio-vary-io-17

        Add some exports for modules needing these functionalities.

Only in 2.4.21pre2aa2: 93_NUMAQ-8
Only in 2.4.21pre3aa1: 93_NUMAQ-9

        Should fix compilation (reported by William Lee Irwin III)

Only in 2.4.21pre2aa2: 9985_blk-atomic-aa4
Only in 2.4.21pre3aa1: 9985_blk-atomic-aa5

        Remeber to start with blk_elv_seq set to 0 by default (or it can
        hang with synchronous-rawio followed by async-io-rawio).

Only in 2.4.21pre3aa1: 9995_frlock-gettimeofday-1

        Added patch from Stephen Hemminger that uses the two sequence
        numbers to bring total SMP scalability to gettimeofday and
        implements a frlock framework that can replace read/write locks.
        This way the fr_write_lock is also fair, and it cannot get starved
        by the readers (unlike the plain read_lock where we depend on the
        unfariness to avoid clearing irq in the readers if the writes will
        never run from irqs). Avoiding starvation of writes make sure not to
        lose timer ticks (and in turn system time), in particular on numa or
        big smp. The locking design is the same of the one implemented in the
        x86-64 vsyscalls, except here the reader is still in kernel so it's
        still some order of magnitude slower than x86-64, but at least the smp
        scalability is total and the starvation is gone. This also includes a
        critical race fix from me to make the code stable, and some porting
        work from Andreas Schwab.

Andrea
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/



This archive was generated by hypermail 2b29 : Thu Jan 23 2003 - 22:00:25 EST