You've already forked linux-apfs
mirror of
https://github.com/linux-apfs/linux-apfs.git
synced 2026-05-01 15:00:59 -07:00
Paul Jackson
9276b1bc96
Optimize the critical zonelist scanning for free pages in the kernel memory
allocator by caching the zones that were found to be full recently, and
skipping them.
Remembers the zones in a zonelist that were short of free memory in the
last second. And it stashes a zone-to-node table in the zonelist struct,
to optimize that conversion (minimize its cache footprint.)
Recent changes:
This differs in a significant way from a similar patch that I
posted a week ago. Now, instead of having a nodemask_t of
recently full nodes, I have a bitmask of recently full zones.
This solves a problem that last weeks patch had, which on
systems with multiple zones per node (such as DMA zone) would
take seeing any of these zones full as meaning that all zones
on that node were full.
Also I changed names - from "zonelist faster" to "zonelist cache",
as that seemed to better convey what we're doing here - caching
some of the key zonelist state (for faster access.)
See below for some performance benchmark results. After all that
discussion with David on why I didn't need them, I went and got
some ;). I wanted to verify that I had not hurt the normal case
of memory allocation noticeably. At least for my one little
microbenchmark, I found (1) the normal case wasn't affected, and
(2) workloads that forced scanning across multiple nodes for
memory improved up to 10% fewer System CPU cycles and lower
elapsed clock time ('sys' and 'real'). Good. See details, below.
I didn't have the logic in get_page_from_freelist() for various
full nodes and zone reclaim failures correct. That should be
fixed up now - notice the new goto labels zonelist_scan,
this_zone_full, and try_next_zone, in get_page_from_freelist().
There are two reasons I persued this alternative, over some earlier
proposals that would have focused on optimizing the fake numa
emulation case by caching the last useful zone:
1) Contrary to what I said before, we (SGI, on large ia64 sn2 systems)
have seen real customer loads where the cost to scan the zonelist
was a problem, due to many nodes being full of memory before
we got to a node we could use. Or at least, I think we have.
This was related to me by another engineer, based on experiences
from some time past. So this is not guaranteed. Most likely, though.
The following approach should help such real numa systems just as
much as it helps fake numa systems, or any combination thereof.
2) The effort to distinguish fake from real numa, using node_distance,
so that we could cache a fake numa node and optimize choosing
it over equivalent distance fake nodes, while continuing to
properly scan all real nodes in distance order, was going to
require a nasty blob of zonelist and node distance munging.
The following approach has no new dependency on node distances or
zone sorting.
See comment in the patch below for a description of what it actually does.
Technical details of note (or controversy):
- See the use of "zlc_active" and "did_zlc_setup" below, to delay
adding any work for this new mechanism until we've looked at the
first zone in zonelist. I figured the odds of the first zone
having the memory we needed were high enough that we should just
look there, first, then get fancy only if we need to keep looking.
- Some odd hackery was needed to add items to struct zonelist, while
not tripping up the custom zonelists built by the mm/mempolicy.c
code for MPOL_BIND. My usual wordy comments below explain this.
Search for "MPOL_BIND".
- Some per-node data in the struct zonelist is now modified frequently,
with no locking. Multiple CPU cores on a node could hit and mangle
this data. The theory is that this is just performance hint data,
and the memory allocator will work just fine despite any such mangling.
The fields at risk are the struct 'zonelist_cache' fields 'fullzones'
(a bitmask) and 'last_full_zap' (unsigned long jiffies). It should
all be self correcting after at most a one second delay.
- This still does a linear scan of the same lengths as before. All
I've optimized is making the scan faster, not algorithmically
shorter. It is now able to scan a compact array of 'unsigned
short' in the case of many full nodes, so one cache line should
cover quite a few nodes, rather than each node hitting another
one or two new and distinct cache lines.
- If both Andi and Nick don't find this too complicated, I will be
(pleasantly) flabbergasted.
- I removed the comment claiming we only use one cachline's worth of
zonelist. We seem, at least in the fake numa case, to have put the
lie to that claim.
- I pay no attention to the various watermarks and such in this performance
hint. A node could be marked full for one watermark, and then skipped
over when searching for a page using a different watermark. I think
that's actually quite ok, as it will tend to slightly increase the
spreading of memory over other nodes, away from a memory stressed node.
===============
Performance - some benchmark results and analysis:
This benchmark runs a memory hog program that uses multiple
threads to touch alot of memory as quickly as it can.
Multiple runs were made, touching 12, 38, 64 or 90 GBytes out of
the total 96 GBytes on the system, and using 1, 19, 37, or 55
threads (on a 56 CPU system.) System, user and real (elapsed)
timings were recorded for each run, shown in units of seconds,
in the table below.
Two kernels were tested - 2.6.18-mm3 and the same kernel with
this zonelist caching patch added. The table also shows the
percentage improvement the zonelist caching sys time is over
(lower than) the stock *-mm kernel.
number 2.6.18-mm3 zonelist-cache delta (< 0 good) percent
GBs N ------------ -------------- ---------------- systime
mem threads sys user real sys user real sys user real better
12 1 153 24 177 151 24 176 -2 0 -1 1%
12 19 99 22 8 99 22 8 0 0 0 0%
12 37 111 25 6 112 25 6 1 0 0 -0%
12 55 115 25 5 110 23 5 -5 -2 0 4%
38 1 502 74 576 497 73 570 -5 -1 -6 0%
38 19 426 78 48 373 76 39 -53 -2 -9 12%
38 37 544 83 36 547 82 36 3 -1 0 -0%
38 55 501 77 23 511 80 24 10 3 1 -1%
64 1 917 125 1042 890 124 1014 -27 -1 -28 2%
64 19 1118 138 119 965 141 103 -153 3 -16 13%
64 37 1202 151 94 1136 150 81 -66 -1 -13 5%
64 55 1118 141 61 1072 140 58 -46 -1 -3 4%
90 1 1342 177 1519 1275 174 1450 -67 -3 -69 4%
90 19 2392 199 192 2116 189 176 -276 -10 -16 11%
90 37 3313 238 175 2972 225 145 -341 -13 -30 10%
90 55 1948 210 104 1843 213 100 -105 3 -4 5%
Notes:
1) This test ran a memory hog program that started a specified number N of
threads, and had each thread allocate and touch 1/N'th of
the total memory to be used in the test run in a single loop,
writing a constant word to memory, one store every 4096 bytes.
Watching this test during some earlier trial runs, I would see
each of these threads sit down on one CPU and stay there, for
the remainder of the pass, a different CPU for each thread.
2) The 'real' column is not comparable to the 'sys' or 'user' columns.
The 'real' column is seconds wall clock time elapsed, from beginning
to end of that test pass. The 'sys' and 'user' columns are total
CPU seconds spent on that test pass. For a 19 thread test run,
for example, the sum of 'sys' and 'user' could be up to 19 times the
number of 'real' elapsed wall clock seconds.
3) Tests were run on a fresh, single-user boot, to minimize the amount
of memory already in use at the start of the test, and to minimize
the amount of background activity that might interfere.
4) Tests were done on a 56 CPU, 28 Node system with 96 GBytes of RAM.
5) Notice that the 'real' time gets large for the single thread runs, even
though the measured 'sys' and 'user' times are modest. I'm not sure what
that means - probably something to do with it being slow for one thread to
be accessing memory along ways away. Perhaps the fake numa system, running
ostensibly the same workload, would not show this substantial degradation
of 'real' time for one thread on many nodes -- lets hope not.
6) The high thread count passes (one thread per CPU - on 55 of 56 CPUs)
ran quite efficiently, as one might expect. Each pair of threads needed
to allocate and touch the memory on the node the two threads shared, a
pleasantly parallizable workload.
7) The intermediate thread count passes, when asking for alot of memory forcing
them to go to a few neighboring nodes, improved the most with this zonelist
caching patch.
Conclusions:
* This zonelist cache patch probably makes little difference one way or the
other for most workloads on real numa hardware, if those workloads avoid
heavy off node allocations.
* For memory intensive workloads requiring substantial off-node allocations
on real numa hardware, this patch improves both kernel and elapsed timings
up to ten per-cent.
* For fake numa systems, I'm optimistic, but will have to leave that up to
Rohit Seth to actually test (once I get him a 2.6.18 backport.)
Signed-off-by: Paul Jackson <pj@sgi.com>
Cc: Rohit Seth <rohitseth@google.com>
Cc: Christoph Lameter <clameter@engr.sgi.com>
Cc: David Rientjes <rientjes@cs.washington.edu>
Cc: Paul Menage <menage@google.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Linux kernel release 2.6.xx <http://kernel.org/>
These are the release notes for Linux version 2.6. Read them carefully,
as they tell you what this is all about, explain how to install the
kernel, and what to do if something goes wrong.
WHAT IS LINUX?
Linux is a clone of the operating system Unix, written from scratch by
Linus Torvalds with assistance from a loosely-knit team of hackers across
the Net. It aims towards POSIX and Single UNIX Specification compliance.
It has all the features you would expect in a modern fully-fledged Unix,
including true multitasking, virtual memory, shared libraries, demand
loading, shared copy-on-write executables, proper memory management,
and multistack networking including IPv4 and IPv6.
It is distributed under the GNU General Public License - see the
accompanying COPYING file for more details.
ON WHAT HARDWARE DOES IT RUN?
Although originally developed first for 32-bit x86-based PCs (386 or higher),
today Linux also runs on (at least) the Compaq Alpha AXP, Sun SPARC and
UltraSPARC, Motorola 68000, PowerPC, PowerPC64, ARM, Hitachi SuperH, Cell,
IBM S/390, MIPS, HP PA-RISC, Intel IA-64, DEC VAX, AMD x86-64, AXIS CRIS,
Cris, Xtensa, AVR32 and Renesas M32R architectures.
Linux is easily portable to most general-purpose 32- or 64-bit architectures
as long as they have a paged memory management unit (PMMU) and a port of the
GNU C compiler (gcc) (part of The GNU Compiler Collection, GCC). Linux has
also been ported to a number of architectures without a PMMU, although
functionality is then obviously somewhat limited.
Linux has also been ported to itself. You can now run the kernel as a
userspace application - this is called UserMode Linux (UML).
DOCUMENTATION:
- There is a lot of documentation available both in electronic form on
the Internet and in books, both Linux-specific and pertaining to
general UNIX questions. I'd recommend looking into the documentation
subdirectories on any Linux FTP site for the LDP (Linux Documentation
Project) books. This README is not meant to be documentation on the
system: there are much better sources available.
- There are various README files in the Documentation/ subdirectory:
these typically contain kernel-specific installation notes for some
drivers for example. See Documentation/00-INDEX for a list of what
is contained in each file. Please read the Changes file, as it
contains information about the problems, which may result by upgrading
your kernel.
- The Documentation/DocBook/ subdirectory contains several guides for
kernel developers and users. These guides can be rendered in a
number of formats: PostScript (.ps), PDF, and HTML, among others.
After installation, "make psdocs", "make pdfdocs", or "make htmldocs"
will render the documentation in the requested format.
INSTALLING the kernel:
- If you install the full sources, put the kernel tarball in a
directory where you have permissions (eg. your home directory) and
unpack it:
gzip -cd linux-2.6.XX.tar.gz | tar xvf -
or
bzip2 -dc linux-2.6.XX.tar.bz2 | tar xvf -
Replace "XX" with the version number of the latest kernel.
Do NOT use the /usr/src/linux area! This area has a (usually
incomplete) set of kernel headers that are used by the library header
files. They should match the library, and not get messed up by
whatever the kernel-du-jour happens to be.
- You can also upgrade between 2.6.xx releases by patching. Patches are
distributed in the traditional gzip and the newer bzip2 format. To
install by patching, get all the newer patch files, enter the
top level directory of the kernel source (linux-2.6.xx) and execute:
gzip -cd ../patch-2.6.xx.gz | patch -p1
or
bzip2 -dc ../patch-2.6.xx.bz2 | patch -p1
(repeat xx for all versions bigger than the version of your current
source tree, _in_order_) and you should be ok. You may want to remove
the backup files (xxx~ or xxx.orig), and make sure that there are no
failed patches (xxx# or xxx.rej). If there are, either you or me has
made a mistake.
Unlike patches for the 2.6.x kernels, patches for the 2.6.x.y kernels
(also known as the -stable kernels) are not incremental but instead apply
directly to the base 2.6.x kernel. Please read
Documentation/applying-patches.txt for more information.
Alternatively, the script patch-kernel can be used to automate this
process. It determines the current kernel version and applies any
patches found.
linux/scripts/patch-kernel linux
The first argument in the command above is the location of the
kernel source. Patches are applied from the current directory, but
an alternative directory can be specified as the second argument.
- If you are upgrading between releases using the stable series patches
(for example, patch-2.6.xx.y), note that these "dot-releases" are
not incremental and must be applied to the 2.6.xx base tree. For
example, if your base kernel is 2.6.12 and you want to apply the
2.6.12.3 patch, you do not and indeed must not first apply the
2.6.12.1 and 2.6.12.2 patches. Similarly, if you are running kernel
version 2.6.12.2 and want to jump to 2.6.12.3, you must first
reverse the 2.6.12.2 patch (that is, patch -R) _before_ applying
the 2.6.12.3 patch.
You can read more on this in Documentation/applying-patches.txt
- Make sure you have no stale .o files and dependencies lying around:
cd linux
make mrproper
You should now have the sources correctly installed.
SOFTWARE REQUIREMENTS
Compiling and running the 2.6.xx kernels requires up-to-date
versions of various software packages. Consult
Documentation/Changes for the minimum version numbers required
and how to get updates for these packages. Beware that using
excessively old versions of these packages can cause indirect
errors that are very difficult to track down, so don't assume that
you can just update packages when obvious problems arise during
build or operation.
BUILD directory for the kernel:
When compiling the kernel all output files will per default be
stored together with the kernel source code.
Using the option "make O=output/dir" allow you to specify an alternate
place for the output files (including .config).
Example:
kernel source code: /usr/src/linux-2.6.N
build directory: /home/name/build/kernel
To configure and build the kernel use:
cd /usr/src/linux-2.6.N
make O=/home/name/build/kernel menuconfig
make O=/home/name/build/kernel
sudo make O=/home/name/build/kernel modules_install install
Please note: If the 'O=output/dir' option is used then it must be
used for all invocations of make.
CONFIGURING the kernel:
Do not skip this step even if you are only upgrading one minor
version. New configuration options are added in each release, and
odd problems will turn up if the configuration files are not set up
as expected. If you want to carry your existing configuration to a
new version with minimal work, use "make oldconfig", which will
only ask you for the answers to new questions.
- Alternate configuration commands are:
"make config" Plain text interface.
"make menuconfig" Text based color menus, radiolists & dialogs.
"make xconfig" X windows (Qt) based configuration tool.
"make gconfig" X windows (Gtk) based configuration tool.
"make oldconfig" Default all questions based on the contents of
your existing ./.config file and asking about
new config symbols.
"make silentoldconfig"
Like above, but avoids cluttering the screen
with questions already answered.
"make defconfig" Create a ./.config file by using the default
symbol values from arch/$ARCH/defconfig.
"make allyesconfig"
Create a ./.config file by setting symbol
values to 'y' as much as possible.
"make allmodconfig"
Create a ./.config file by setting symbol
values to 'm' as much as possible.
"make allnoconfig" Create a ./.config file by setting symbol
values to 'n' as much as possible.
"make randconfig" Create a ./.config file by setting symbol
values to random values.
The allyesconfig/allmodconfig/allnoconfig/randconfig variants can
also use the environment variable KCONFIG_ALLCONFIG to specify a
filename that contains config options that the user requires to be
set to a specific value. If KCONFIG_ALLCONFIG=filename is not used,
"make *config" checks for a file named "all{yes/mod/no/random}.config"
for symbol values that are to be forced. If this file is not found,
it checks for a file named "all.config" to contain forced values.
NOTES on "make config":
- having unnecessary drivers will make the kernel bigger, and can
under some circumstances lead to problems: probing for a
nonexistent controller card may confuse your other controllers
- compiling the kernel with "Processor type" set higher than 386
will result in a kernel that does NOT work on a 386. The
kernel will detect this on bootup, and give up.
- A kernel with math-emulation compiled in will still use the
coprocessor if one is present: the math emulation will just
never get used in that case. The kernel will be slightly larger,
but will work on different machines regardless of whether they
have a math coprocessor or not.
- the "kernel hacking" configuration details usually result in a
bigger or slower kernel (or both), and can even make the kernel
less stable by configuring some routines to actively try to
break bad code to find kernel problems (kmalloc()). Thus you
should probably answer 'n' to the questions for
"development", "experimental", or "debugging" features.
COMPILING the kernel:
- Make sure you have at least gcc 3.2 available.
For more information, refer to Documentation/Changes.
Please note that you can still run a.out user programs with this kernel.
- Do a "make" to create a compressed kernel image. It is also
possible to do "make install" if you have lilo installed to suit the
kernel makefiles, but you may want to check your particular lilo setup first.
To do the actual install you have to be root, but none of the normal
build should require that. Don't take the name of root in vain.
- If you configured any of the parts of the kernel as `modules', you
will also have to do "make modules_install".
- Keep a backup kernel handy in case something goes wrong. This is
especially true for the development releases, since each new release
contains new code which has not been debugged. Make sure you keep a
backup of the modules corresponding to that kernel, as well. If you
are installing a new kernel with the same version number as your
working kernel, make a backup of your modules directory before you
do a "make modules_install".
Alternatively, before compiling, use the kernel config option
"LOCALVERSION" to append a unique suffix to the regular kernel version.
LOCALVERSION can be set in the "General Setup" menu.
- In order to boot your new kernel, you'll need to copy the kernel
image (e.g. .../linux/arch/i386/boot/bzImage after compilation)
to the place where your regular bootable kernel is found.
- Booting a kernel directly from a floppy without the assistance of a
bootloader such as LILO, is no longer supported.
If you boot Linux from the hard drive, chances are you use LILO which
uses the kernel image as specified in the file /etc/lilo.conf. The
kernel image file is usually /vmlinuz, /boot/vmlinuz, /bzImage or
/boot/bzImage. To use the new kernel, save a copy of the old image
and copy the new image over the old one. Then, you MUST RERUN LILO
to update the loading map!! If you don't, you won't be able to boot
the new kernel image.
Reinstalling LILO is usually a matter of running /sbin/lilo.
You may wish to edit /etc/lilo.conf to specify an entry for your
old kernel image (say, /vmlinux.old) in case the new one does not
work. See the LILO docs for more information.
After reinstalling LILO, you should be all set. Shutdown the system,
reboot, and enjoy!
If you ever need to change the default root device, video mode,
ramdisk size, etc. in the kernel image, use the 'rdev' program (or
alternatively the LILO boot options when appropriate). No need to
recompile the kernel to change these parameters.
- Reboot with the new kernel and enjoy.
IF SOMETHING GOES WRONG:
- If you have problems that seem to be due to kernel bugs, please check
the file MAINTAINERS to see if there is a particular person associated
with the part of the kernel that you are having trouble with. If there
isn't anyone listed there, then the second best thing is to mail
them to me (torvalds@osdl.org), and possibly to any other relevant
mailing-list or to the newsgroup.
- In all bug-reports, *please* tell what kernel you are talking about,
how to duplicate the problem, and what your setup is (use your common
sense). If the problem is new, tell me so, and if the problem is
old, please try to tell me when you first noticed it.
- If the bug results in a message like
unable to handle kernel paging request at address C0000010
Oops: 0002
EIP: 0010:XXXXXXXX
eax: xxxxxxxx ebx: xxxxxxxx ecx: xxxxxxxx edx: xxxxxxxx
esi: xxxxxxxx edi: xxxxxxxx ebp: xxxxxxxx
ds: xxxx es: xxxx fs: xxxx gs: xxxx
Pid: xx, process nr: xx
xx xx xx xx xx xx xx xx xx xx
or similar kernel debugging information on your screen or in your
system log, please duplicate it *exactly*. The dump may look
incomprehensible to you, but it does contain information that may
help debugging the problem. The text above the dump is also
important: it tells something about why the kernel dumped code (in
the above example it's due to a bad kernel pointer). More information
on making sense of the dump is in Documentation/oops-tracing.txt
- If you compiled the kernel with CONFIG_KALLSYMS you can send the dump
as is, otherwise you will have to use the "ksymoops" program to make
sense of the dump (but compiling with CONFIG_KALLSYMS is usually preferred).
This utility can be downloaded from
ftp://ftp.<country>.kernel.org/pub/linux/utils/kernel/ksymoops/ .
Alternately you can do the dump lookup by hand:
- In debugging dumps like the above, it helps enormously if you can
look up what the EIP value means. The hex value as such doesn't help
me or anybody else very much: it will depend on your particular
kernel setup. What you should do is take the hex value from the EIP
line (ignore the "0010:"), and look it up in the kernel namelist to
see which kernel function contains the offending address.
To find out the kernel function name, you'll need to find the system
binary associated with the kernel that exhibited the symptom. This is
the file 'linux/vmlinux'. To extract the namelist and match it against
the EIP from the kernel crash, do:
nm vmlinux | sort | less
This will give you a list of kernel addresses sorted in ascending
order, from which it is simple to find the function that contains the
offending address. Note that the address given by the kernel
debugging messages will not necessarily match exactly with the
function addresses (in fact, that is very unlikely), so you can't
just 'grep' the list: the list will, however, give you the starting
point of each kernel function, so by looking for the function that
has a starting address lower than the one you are searching for but
is followed by a function with a higher address you will find the one
you want. In fact, it may be a good idea to include a bit of
"context" in your problem report, giving a few lines around the
interesting one.
If you for some reason cannot do the above (you have a pre-compiled
kernel image or similar), telling me as much about your setup as
possible will help. Please read the REPORTING-BUGS document for details.
- Alternately, you can use gdb on a running kernel. (read-only; i.e. you
cannot change values or set break points.) To do this, first compile the
kernel with -g; edit arch/i386/Makefile appropriately, then do a "make
clean". You'll also need to enable CONFIG_PROC_FS (via "make config").
After you've rebooted with the new kernel, do "gdb vmlinux /proc/kcore".
You can now use all the usual gdb commands. The command to look up the
point where your system crashed is "l *0xXXXXXXXX". (Replace the XXXes
with the EIP value.)
gdb'ing a non-running kernel currently fails because gdb (wrongly)
disregards the starting offset for which the kernel is compiled.