2.2.18pre2aa2 and patches for 2.2.18pre3

From: Andrea Arcangeli (andrea@suse.de)
Date: Wed Sep 06 2000 - 14:52:07 EST

The main features of 2.2.18pre2aa2 are:

o Support for 4Gigabyte of RAM on IA32 (me and Gerhard Wichert)
o Support for 2T of RAM on alpha (me)
o Improved VM for high end machines with enough ram and doing
        heavy I/O under high memory pressure plus fixes
        for the MAP_SHARED oom by killing kpiod. (me)
o RAW-IO (doable with bigmem enabled too). (Stephen C. Tweedie)
o SMP scheduler improvements. (me and partly from 2.3.x contributed
        by Ingo Molnar)
o LFS (>2G file on 32bit architectures)
o misc fixes

Detailed description of 2.2.18pre2aa2 follows.


        Increase the min percent of the buffer cache and page cache to 4%.
        (it wouldn't matter if the VM algorithms were better). (me)


        Avoid suprious unplug of the I/O queue. (me)


        Better SMP reschedule_idle. (partly backported from 2.3.x, 2.3.x
        version was contributed by Ingo Molnar)


        Account also the failed buffer tries during shrink_mmap. (me)


        Put in sync the RTC parsing done by the kernel at boot
        while setting the system time, with the parsing done
        by the RTC driver during initialization. (Jan-Benedict Glaw
        <jbglaw@lug-owl.de> fixed it in 2.4.x)


        SMP race fix in the read_unlock implementation of alpha. (me)


        Avoid suprious unplug of the I/O queue. (me)


        Additional paranoid stuff in TCP delayed acks code (if somebody
        can still reproduce lockups under heavy TCP load, the
        delack-timer-5 patch can be the solution). In 2.2.15 only what it
        was obviously wrong is been fixed, but at the time I did
        delack-timer-5 I was convinced the stuff in 2.2.15 _might_ not be
        enough, so if you have still lockups under 2.2.15 let us know.
        delack-timer-5 is reported to definitely fix the TCP lockup. (me)


        It simplify the elevator algorithm to remove the careful code
        that was accounting also the place in the queue where the
        request was inserted while setting its initial latency.
        This changed makes the systems much less interactive, but
        it fixes the tiotest numbers of course still avoiding indefinite
        starvation as it was happeneing before the elevator rewrite. (me,
        tested also by Jens Axboe)


        Minor inode stuff cleanedup. (me)


        Revertd the semantic change that make differences between
        /proc/00000$$ and /proc/$$, this allows backwards compatibilty of
        a misfeature and it _won't_ hurt security. There's no downside
        in reverting the 2.2.13 semantic change.


        Allows kupdate to be stopped via SIGSTOP (currently it must be
        stopped by setting interval to zero via sysctl). This returns
        useful in the airplane where you do killall -STOP
        kupdate... STOP kupdate at your own risk of course ;) (me)


        Avoids mremap to grow in virtual addresses (and so in turn
        it avoids to waste virtual address space).


        Provide nanosleep usec resolution so that a signal flood doesn't
        hang glibc folks that correctly trust the rem field to resume
        the nanosleep after a syscall interruption. (without the patch
        nanosleep resolution is instead 10msec on IA32 and around 1msec on
        alpha) (me)


        Make sure to not understimate the available memory (the cache and
        buffers may be under the min percent) (me)


        Fix set_rtc_mss SMP race between timer irq and rtc irq. (me)


        Avoid stack to silenty overflow on the heap (make life easier
        to track down userspace issues). Perfect approch is not possible
        even using special LDT for the stack segment. The approch in the
        patch will work 99% of the time and it enforces a GAP between heap
        and stack of 1 page as default (configurable via sysctl, if
        turned to zero the feature is disabled completly). (me)


        Fixes an SMP race in slow gettimeofday. (me)


        Check lost_ticks in settimeofday to be more precise.


        TSC calibration must be dynamic and not a compile time thing
        because gettimeofday is dynamic and it depends on the TSCs
        to be in sync. (me)


        Change to the Makefile EXTRAVERSION.


        BIGMEM production code (IA32 and alpha). This one fixes
        a memory leak during swap (almost unreproducibe, found it
        while reading the sources only) and some alpha-bigmem
        bug. It's now possible to generate core dumps larger
        than 2G (the previous limitation was due a few ext2
        bugs and a binfmt_elf.c bug). I just includes the large_shm
        patch so it should be possible to allocate more than 2g in
        one shm segment. (me and Gerhard Wichert)


        SCT's rawio bigmem capable. (Stephen C. Tweedie)


        LVM compatible with the 2.3.x one. (Heinz Mauelshagen)


        Needed by alpha SMP. (won't be needed anymore in 2.4.x)


        LFS: >2G file support for 32bit archs. This also includes
        the new getdents64 and fcntl64 syscalls in sync with the
        2.4.0-test8-pre5 user<-> kernel API. It also fixes some bug
        present in the previous 2.2.x LFS patches.


        Relocate the initrd image ASAP (btw, on alpha ASAP is sooner
        than on IA32), to avoid corrupting it while allocating the
        boot memory (like the mem_map_t array). This is good for
        robusteness and it's mainly interesting when bigmem
        is enabled.


        Kills kpiod removing the OOM while paging out MAP_SHARED
        segments during a low memory condition. Fixes
        mmap00[123] on 8mbyte of RAM, and I bet I fixes as well
        Alan's DBMS OOM troubles. The way I could remove
        kpiod is to add a per-process fs_locks information
        and to check that information instead of deadlocking
        into the semaphore or into the superblock lock.

        Fixes in a alternate (more efficient) way the GFP race that
        is currently handled by the free_before_allocate logic.
        The alternate way is to have a per-process freelist
        and to refill the global queue only once the tasks
        allocate something for itself.

        Accounts protected buffers as unfreeable (fixes
        OOM with big ramdisk in memory).

        Rewrite the sync_page_buffers logic. Now it's a per-bh
        bitflag information that tell us if a dirty buffer is hanging
        around for a too long time w/o becoming freeable (so we block
        on such buffer only once necessary and never when
        there's no high pressure). If there's no high pressure and
        __GFP_IO is set we try to write with WRITEA that doesn't block
        but that allows us to fill the I/O queue with the
        less interesting buffers in the VM. Never waits on the
        buffer to become unlocked (that's too aggressive and that was
        necessary only to workaround the pagingout of MAP_SHARED segments
        that wasn't able to do proper write throttling due the always
        async kpiod).

        Rewrite the async buffer flushing wakeup points ala 2.4.x way.

        Avoids swain/swapout while writing heavily to disk (no-swapout).

        Adds lowlatency control checks in shrink_mmap and swap_out.

        The same VM-global-patch backported to clean 2.2.18pre2 is
        here (the only difference is that it includes the above
        account-buffer-tries and it doesn't remove the suprious
        tq_disk run during readahead that isn't a VM related change):



        Implements conditional_schedule(). (Ingo Molnar)


        Add low latency schedule entry points into the copy_user(), this fixes
        read(2)/write(2) hangs. (Originally implemented by Ingo Molnar
        in the lowlatency patch, but this patch is not equivalent, it's
        a very subset and it's different the reason I put the
        conditional schedule _after_ the copy is because we want the
        higher timeslice possible once we return to userspace to avoid
        live locks during swap)

The 2.2.18pre2aa1 patch is here:


As usual all the separate patches are in:


and they can be applyed in `ls` order to generate the 2.2.18pre2aa2

All the patches that I propose for inclusion into 2.2.18pre3 are here
(the directory is called 2.2.18pre2 because the patches are against


They are all orthogonal and they can be applied in any order regardless of
each other.


PS. If you'll run with more than 4giga of RAM on alpha, you'll need also
to fix procps if you want to read the correct MM values with `free`, `top`
and `vmstat`. procps takes all the memory information with byte
granularity in `int` words (and not long words). I fixed procps this way:


PPS. 2.2.18pre2 doesn't included riskious stuff compared to 2.2.17, thus
     don't worry applying that pre-patch. In case you need to be
     based on 2.2.17 for whatever reason the 2.2.18pre2aa2.bz2 patch will
     just apply cleanly to 2.2.17 too.

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

This archive was generated by hypermail 2b29 : Thu Sep 07 2000 - 21:00:27 EST