Andrea Arcangeli: 2.2.20pre2aa1
May 15, 2001, 22:29 (0 Talkback[s])
WEBINAR: On-demand Event
Replace Oracle with the NoSQL Engagement Database: Why and how leading companies are making the switch REGISTER >
Date: Tue, 15 May 2001 23:58:59 +0200
From: Andrea Arcangeli <email@example.com>
The main features of 2.2.20pre2aa1 are:
o Support for 4Gigabyte of RAM on IA32 (me and Gerhard Wichert)
o Support for 2T of RAM on alpha (me)
o RAW-IO (doable with bigmem enabled too). Improvements are also been
backported from 2.4.
o SMP scheduler improvements. (me and partly from 2.3.x contributed
by Ingo Molnar)
o LFS (>2G file on 32bit architectures) also NFSv3 works over 2G
(nfsv3-lfs work from me, Andi and fix from Jay Weber)
o fixed race in wake-one LIFO in accept(2). Apache must be compiled
with -DSINGLE_LISTEN_UNSERIALIZED_ACCEPT to take advantage of that.
o lowlatency and SMP scalability in all copy-user and tcp_sendmsg
o GFS support.
o various fixes
Detailed description of 2.2.20pre2aa1 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)
Allows the kernel to be compiled for K7 (AMD Athlon) or Pentium4.
This compilations options _only_ make the kernel to assume
respectively 64byte cachelines or 128byte cachelines.
Those assumptions are critical mostly on SMP systems, but even UP
will take advantage of it because it will make most performance
critical slab allocations to start on a cacheline boundary.
Since the only difference between Ppro/K7/P4 compilations
is the cacheline size assumed by the kernel you can safely boot
a K7/P4 compiled kernel on a Ppro, it won't obviously generate
cacheline pinpongs in SMP. There are only two downsides in
running on a Ppro (PII/PIII included) a K7/P4 compiled kernel:
1) some byte of memory wasted due the larger paddings (really not
a big deal)
2) potential waste of cacheline sets. On PII and PIII and PPro
the L1 dcache is 8kbyte or 16kbytes 2-way set associative (so
there are 128 or 256 sets of cachelines) and by stressing the
first 32bytes of the 128 byte aligned data structures (for
example if the kernel is compiled for P4) you would take
advantage of only 1/4 of the available L1 cache. (you would
stress set 0, 4, 8, ... only) This is probably quite serious
issue in terms of performance. So for Ppro/PII/PIII you're
still suggested to use Ppro compilation option (if care to
optimize the L1 cache usage).
(me and Andi)
Fix a local APIC initaliziation ordering bug that triggers on the
SSE/SSE2 support (unmasked exception via mxcsr included).
(mix of Doug Ledford, Ingo Molnar, Gabriel Paubert PIII patch for
2.2.x and 2.4.x PIII support from Gareth Hughes, audited, fixed and
changed by me to be dynamic. At the end it's very similar to the
Kernel now understands the `nofxsr' boot time parameter and it
doesn't enable fxsr in that case (if there's any CPU that
crashes at boot because it's buggy, nofxsr will workaround
the hardware bug; it's also useful for asymetric multiprocessing
where boot cpu can have fxsr capabilities and the other cpus hasn't)
Pass the reason for the sigio in the si_code (and a duplicate info
in si_band) with the same API of 2.4.x. This avoids people
having to poll a set of fd during the sigio handler. (current
2.4.x has two bugs in that area but fixes are in Linus's mailbox)
Better SMP reschedule_idle. (partly backported from 2.3.x, 2.3.x
version was contributed by Ingo Molnar)
Fixes the wmb() in schedule_tail that should really be a mb(), in
theory one of the last reads in reschedule_idle could return garbage
(in practice I think it can't trigger... at least on x86 :) (me)
wait I/O completion as well while doing the Wait_IO second pass on
the dirty cache (should fix last VM problem reported by VA while
creating very large ext2 fs on lowmem machines)
Allows device drivers to set the VMA as reserved to avoid swap_out
to try to unmap stuff from them (this avoid device drivers using
->nopage for lazily mapping scatter gather dma areas in userspace,
to implement a noop swapout and it also avoids useless page faults).
This is much more efficient than setting the physical pages as
Fixes the RTC parsing done by the kernel at boot. (backported from
Fix from Andi for the compatibility 2.4 interface (mark_bh was
Avoid a suprious useless 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
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)
Fix elevator merging when tasks blocks waiting for new requests.
Declare rtc_lock external spinlock to fix compilation. (me)
Make the inode-max and ihash dynamic. Mean distribution in the
hash is of 1. inode-max remains the same until 256M of ram.
It then reaches 32k inodes with 512Mbyte of ram and 64k inodes
with 1G of ram and then it stops growing. It's true dcache
avoids most of the iget when everything fits in the working set,
but as soon there's icache pollution iget will run all the time
so it's not true iget hash table doesn't matter for performance.
The ihash hasnfn is very bad in both 2.2.x and 2.4.x:
unsigned long tmp = i_ino | (unsigned long) sb;
Think what happens if `sb' has many bits set to 1, it will generate
an huge collision chain. It should have been at least an xor, not an
or. So I changed it this way:
unsigned long tmp = i_ino + (unsigned long) sb / (sizeof(struct
super_block) & ~(sizeof(struct super_block) - 1));
This will shift right sb with around `super_block size' granularity
and it will share hash cachelines if the user asks for consecutive
inode numbers (can be the case).
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 large kupdate intervals (don't change the interval unless
you know what you're doing!)
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)
Instead of making a single threaded async daemon to avoid unplugging
the queue all the time, the blockdevice code is been changed
to allow the loop device to handle merged requests. (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
ext2 64bit fix
Make sure to not understimate the available memory (the cache and
buffers may be under the min percent) (me)
If child exists _before_ it got the timeslice refresh parents
gets its timeslice back. 2.4.x does that even if child got
the refresh and it's wrong.
This feature avoids programs like `bash scripts' to be starved while
forking short-lived childs at high frequency.
Don't generate suprious unplug of the queue during readahead.
Some more info out of SYSRQ+P.
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)
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.
Backport 2.4 waitqueues and in turn fixes an hanging condition in
Use only ordered string operations while writing to IOMMU areas.
the non __ versions can be reordered at compile time. (Doug Ledford)
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 that need to be fixed in 2.4.x too).
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)
Includes >4G mtrr support developed by Boszormenyi Zoltan.
Recently fixed sysinfo structure to be defined as:
so that sysinfo(2) doesn't overflow the user buffer on 64bit
Also 2.4.x is buggy in that respect and needs to be fixed, btw.
SCT's rawio bigmem capable. (Stephen C. Tweedie)
Backported latest 2.4.4 improvements.
beta4 LVM compatible with the 2.4.5pre2aa1 beta7 one. (Heinz
Implements the consistent snapshot API so that for example we can
live snaphost reiserfs.
Fixes hardblocksize bug present in beta4 lvm. (me)
(the fix is been merged in lvm beta7 btw)
use the preallocated block array in the kiobuf to save stack. (me)
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.x user <-> kernel API. It also fixes some bug
present in the previous 2.2.x LFS patches. Support for
nfsv3 is complete as well on both server and client side.
Avoid duplicating disconnect dentries. (Chip)
Relocate the initrd image ASAP (btw, on alpha `ASAP' happens sooner
than on IA32 due kseg), 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.
Implements conditional_schedule(). (Ingo Molnar from lowlatency
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)
Drop the big kernel lock during copy-user that can release
it anyways while getting faults. Improves SMP scalability. (me)
Drop big kernel lock during TCP checksum (in tcp_sendmsg). Receiver
side can't drop the big lock as it doesn't have a process context.
Same as above for alpha architecture.
Same as above for ppc, sparc, sparc64 architectures.
Fix implementation bugs in spinlocks and other SMP atomic operations
that needs to declare "memory" as clobbered. Didn't removed the
__dummy structures in 2.2.x to make sure not to trigger any compiler
NMI wathdog is compiled in without configuration options.
It's disabled by default. To enable it dynamically at boot
pass the kernel the parameter "nmi_irq=0". nmi_irq=0 makes sense
on SMP only. (original patch from Ingo Molnar, fixed to work again
in late 2.2.x kernels and made it dynamic by me)
RAID 0.90 updates from Ingo (it needs raidtools, not mdutils)
During reads from raid1 keep the two disks working on different
areas of the raid1. (Mika Kuoppala)
Add hooks in the buffer head that GFS needs.
make scsi command length dynamic (max 16) instead of fixed to 12.
(needed by GFS)
export rscsi_disks structure from sd.c for GFS. Note the equivalent
patch in GFS latest release is buggy and I had to fix it up to allow
sd_mod to be compied correctly. (just reported to GFS folks)
Calls lowlevel f_ops->lock also from flock (not only from
fcntl(F_*K)). needed by GFS.
Export a few kernel functions for GFS module. (This was buggy
too in the GFS latest release and it's just reported and fixed here)
Added device_names to generic disk structure (for GFS).
Added few ioctl to sparc64 32bit ioctl wrapper for GFS.
Export change_termios to modules (for GFS).
Added fs_transaction field to the task structure (for GFS).
Added generic map_fn/request_fn to ll_rw_block. The interface with
MD and LVM got broken so it had to be fixed too. (for GFS).
Insert the GFS inode in the inode structure.
GFS uses same inode for named pipes/unix sockets and
GFS itself, so left some space before putting generic_ip pointer
in the gfs_inode_info.
Implements hard_panic() function that will attempt to show the
message nad reboot the machine after some second.
Pass the opaque cookie of iget4 back to the filesystem during
the read_inode callback.
The 2.2.20pre2aa1 patch is here:
As usual all the separate patches are in:
and they can be applyed in `ls` order to generate the
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: