linux/Documentation/vm/frontswap.rst
Christoph Hellwig 0b364446d7 frontswap: remove frontswap_shrink
frontswap_shrink is never called, so remove it.

Link: https://lkml.kernel.org/r/20211224062246.1258487-5-hch@lst.de
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Juergen Gross <jgross@suse.com>
Cc: Dan Streetman <ddstreet@ieee.org>
Cc: Geert Uytterhoeven <geert@linux-m68k.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Konrad Rzeszutek Wilk <Konrad.wilk@oracle.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Seth Jennings <sjenning@redhat.com>
Cc: Vitaly Wool <vitaly.wool@konsulko.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-01-22 08:33:38 +02:00

267 lines
14 KiB
ReStructuredText

.. _frontswap:
=========
Frontswap
=========
Frontswap provides a "transcendent memory" interface for swap pages.
In some environments, dramatic performance savings may be obtained because
swapped pages are saved in RAM (or a RAM-like device) instead of a swap disk.
.. _Transcendent memory in a nutshell: https://lwn.net/Articles/454795/
Frontswap is so named because it can be thought of as the opposite of
a "backing" store for a swap device. The storage is assumed to be
a synchronous concurrency-safe page-oriented "pseudo-RAM device" conforming
to the requirements of transcendent memory (such as Xen's "tmem", or
in-kernel compressed memory, aka "zcache", or future RAM-like devices);
this pseudo-RAM device is not directly accessible or addressable by the
kernel and is of unknown and possibly time-varying size. The driver
links itself to frontswap by calling frontswap_register_ops to set the
frontswap_ops funcs appropriately and the functions it provides must
conform to certain policies as follows:
An "init" prepares the device to receive frontswap pages associated
with the specified swap device number (aka "type"). A "store" will
copy the page to transcendent memory and associate it with the type and
offset associated with the page. A "load" will copy the page, if found,
from transcendent memory into kernel memory, but will NOT remove the page
from transcendent memory. An "invalidate_page" will remove the page
from transcendent memory and an "invalidate_area" will remove ALL pages
associated with the swap type (e.g., like swapoff) and notify the "device"
to refuse further stores with that swap type.
Once a page is successfully stored, a matching load on the page will normally
succeed. So when the kernel finds itself in a situation where it needs
to swap out a page, it first attempts to use frontswap. If the store returns
success, the data has been successfully saved to transcendent memory and
a disk write and, if the data is later read back, a disk read are avoided.
If a store returns failure, transcendent memory has rejected the data, and the
page can be written to swap as usual.
Note that if a page is stored and the page already exists in transcendent memory
(a "duplicate" store), either the store succeeds and the data is overwritten,
or the store fails AND the page is invalidated. This ensures stale data may
never be obtained from frontswap.
If properly configured, monitoring of frontswap is done via debugfs in
the `/sys/kernel/debug/frontswap` directory. The effectiveness of
frontswap can be measured (across all swap devices) with:
``failed_stores``
how many store attempts have failed
``loads``
how many loads were attempted (all should succeed)
``succ_stores``
how many store attempts have succeeded
``invalidates``
how many invalidates were attempted
A backend implementation may provide additional metrics.
FAQ
===
* Where's the value?
When a workload starts swapping, performance falls through the floor.
Frontswap significantly increases performance in many such workloads by
providing a clean, dynamic interface to read and write swap pages to
"transcendent memory" that is otherwise not directly addressable to the kernel.
This interface is ideal when data is transformed to a different form
and size (such as with compression) or secretly moved (as might be
useful for write-balancing for some RAM-like devices). Swap pages (and
evicted page-cache pages) are a great use for this kind of slower-than-RAM-
but-much-faster-than-disk "pseudo-RAM device".
Frontswap with a fairly small impact on the kernel,
provides a huge amount of flexibility for more dynamic, flexible RAM
utilization in various system configurations:
In the single kernel case, aka "zcache", pages are compressed and
stored in local memory, thus increasing the total anonymous pages
that can be safely kept in RAM. Zcache essentially trades off CPU
cycles used in compression/decompression for better memory utilization.
Benchmarks have shown little or no impact when memory pressure is
low while providing a significant performance improvement (25%+)
on some workloads under high memory pressure.
"RAMster" builds on zcache by adding "peer-to-peer" transcendent memory
support for clustered systems. Frontswap pages are locally compressed
as in zcache, but then "remotified" to another system's RAM. This
allows RAM to be dynamically load-balanced back-and-forth as needed,
i.e. when system A is overcommitted, it can swap to system B, and
vice versa. RAMster can also be configured as a memory server so
many servers in a cluster can swap, dynamically as needed, to a single
server configured with a large amount of RAM... without pre-configuring
how much of the RAM is available for each of the clients!
In the virtual case, the whole point of virtualization is to statistically
multiplex physical resources across the varying demands of multiple
virtual machines. This is really hard to do with RAM and efforts to do
it well with no kernel changes have essentially failed (except in some
well-publicized special-case workloads).
Specifically, the Xen Transcendent Memory backend allows otherwise
"fallow" hypervisor-owned RAM to not only be "time-shared" between multiple
virtual machines, but the pages can be compressed and deduplicated to
optimize RAM utilization. And when guest OS's are induced to surrender
underutilized RAM (e.g. with "selfballooning"), sudden unexpected
memory pressure may result in swapping; frontswap allows those pages
to be swapped to and from hypervisor RAM (if overall host system memory
conditions allow), thus mitigating the potentially awful performance impact
of unplanned swapping.
A KVM implementation is underway and has been RFC'ed to lkml. And,
using frontswap, investigation is also underway on the use of NVM as
a memory extension technology.
* Sure there may be performance advantages in some situations, but
what's the space/time overhead of frontswap?
If CONFIG_FRONTSWAP is disabled, every frontswap hook compiles into
nothingness and the only overhead is a few extra bytes per swapon'ed
swap device. If CONFIG_FRONTSWAP is enabled but no frontswap "backend"
registers, there is one extra global variable compared to zero for
every swap page read or written. If CONFIG_FRONTSWAP is enabled
AND a frontswap backend registers AND the backend fails every "store"
request (i.e. provides no memory despite claiming it might),
CPU overhead is still negligible -- and since every frontswap fail
precedes a swap page write-to-disk, the system is highly likely
to be I/O bound and using a small fraction of a percent of a CPU
will be irrelevant anyway.
As for space, if CONFIG_FRONTSWAP is enabled AND a frontswap backend
registers, one bit is allocated for every swap page for every swap
device that is swapon'd. This is added to the EIGHT bits (which
was sixteen until about 2.6.34) that the kernel already allocates
for every swap page for every swap device that is swapon'd. (Hugh
Dickins has observed that frontswap could probably steal one of
the existing eight bits, but let's worry about that minor optimization
later.) For very large swap disks (which are rare) on a standard
4K pagesize, this is 1MB per 32GB swap.
When swap pages are stored in transcendent memory instead of written
out to disk, there is a side effect that this may create more memory
pressure that can potentially outweigh the other advantages. A
backend, such as zcache, must implement policies to carefully (but
dynamically) manage memory limits to ensure this doesn't happen.
* OK, how about a quick overview of what this frontswap patch does
in terms that a kernel hacker can grok?
Let's assume that a frontswap "backend" has registered during
kernel initialization; this registration indicates that this
frontswap backend has access to some "memory" that is not directly
accessible by the kernel. Exactly how much memory it provides is
entirely dynamic and random.
Whenever a swap-device is swapon'd frontswap_init() is called,
passing the swap device number (aka "type") as a parameter.
This notifies frontswap to expect attempts to "store" swap pages
associated with that number.
Whenever the swap subsystem is readying a page to write to a swap
device (c.f swap_writepage()), frontswap_store is called. Frontswap
consults with the frontswap backend and if the backend says it does NOT
have room, frontswap_store returns -1 and the kernel swaps the page
to the swap device as normal. Note that the response from the frontswap
backend is unpredictable to the kernel; it may choose to never accept a
page, it could accept every ninth page, or it might accept every
page. But if the backend does accept a page, the data from the page
has already been copied and associated with the type and offset,
and the backend guarantees the persistence of the data. In this case,
frontswap sets a bit in the "frontswap_map" for the swap device
corresponding to the page offset on the swap device to which it would
otherwise have written the data.
When the swap subsystem needs to swap-in a page (swap_readpage()),
it first calls frontswap_load() which checks the frontswap_map to
see if the page was earlier accepted by the frontswap backend. If
it was, the page of data is filled from the frontswap backend and
the swap-in is complete. If not, the normal swap-in code is
executed to obtain the page of data from the real swap device.
So every time the frontswap backend accepts a page, a swap device read
and (potentially) a swap device write are replaced by a "frontswap backend
store" and (possibly) a "frontswap backend loads", which are presumably much
faster.
* Can't frontswap be configured as a "special" swap device that is
just higher priority than any real swap device (e.g. like zswap,
or maybe swap-over-nbd/NFS)?
No. First, the existing swap subsystem doesn't allow for any kind of
swap hierarchy. Perhaps it could be rewritten to accommodate a hierarchy,
but this would require fairly drastic changes. Even if it were
rewritten, the existing swap subsystem uses the block I/O layer which
assumes a swap device is fixed size and any page in it is linearly
addressable. Frontswap barely touches the existing swap subsystem,
and works around the constraints of the block I/O subsystem to provide
a great deal of flexibility and dynamicity.
For example, the acceptance of any swap page by the frontswap backend is
entirely unpredictable. This is critical to the definition of frontswap
backends because it grants completely dynamic discretion to the
backend. In zcache, one cannot know a priori how compressible a page is.
"Poorly" compressible pages can be rejected, and "poorly" can itself be
defined dynamically depending on current memory constraints.
Further, frontswap is entirely synchronous whereas a real swap
device is, by definition, asynchronous and uses block I/O. The
block I/O layer is not only unnecessary, but may perform "optimizations"
that are inappropriate for a RAM-oriented device including delaying
the write of some pages for a significant amount of time. Synchrony is
required to ensure the dynamicity of the backend and to avoid thorny race
conditions that would unnecessarily and greatly complicate frontswap
and/or the block I/O subsystem. That said, only the initial "store"
and "load" operations need be synchronous. A separate asynchronous thread
is free to manipulate the pages stored by frontswap. For example,
the "remotification" thread in RAMster uses standard asynchronous
kernel sockets to move compressed frontswap pages to a remote machine.
Similarly, a KVM guest-side implementation could do in-guest compression
and use "batched" hypercalls.
In a virtualized environment, the dynamicity allows the hypervisor
(or host OS) to do "intelligent overcommit". For example, it can
choose to accept pages only until host-swapping might be imminent,
then force guests to do their own swapping.
There is a downside to the transcendent memory specifications for
frontswap: Since any "store" might fail, there must always be a real
slot on a real swap device to swap the page. Thus frontswap must be
implemented as a "shadow" to every swapon'd device with the potential
capability of holding every page that the swap device might have held
and the possibility that it might hold no pages at all. This means
that frontswap cannot contain more pages than the total of swapon'd
swap devices. For example, if NO swap device is configured on some
installation, frontswap is useless. Swapless portable devices
can still use frontswap but a backend for such devices must configure
some kind of "ghost" swap device and ensure that it is never used.
* Why this weird definition about "duplicate stores"? If a page
has been previously successfully stored, can't it always be
successfully overwritten?
Nearly always it can, but no, sometimes it cannot. Consider an example
where data is compressed and the original 4K page has been compressed
to 1K. Now an attempt is made to overwrite the page with data that
is non-compressible and so would take the entire 4K. But the backend
has no more space. In this case, the store must be rejected. Whenever
frontswap rejects a store that would overwrite, it also must invalidate
the old data and ensure that it is no longer accessible. Since the
swap subsystem then writes the new data to the read swap device,
this is the correct course of action to ensure coherency.
* Why does the frontswap patch create the new include file swapfile.h?
The frontswap code depends on some swap-subsystem-internal data
structures that have, over the years, moved back and forth between
static and global. This seemed a reasonable compromise: Define
them as global but declare them in a new include file that isn't
included by the large number of source files that include swap.h.
Dan Magenheimer, last updated April 9, 2012