UBC: An Efficient Unified I/O and

Memory Caching Subsystem for NetBSD

Chuck Silvers
The NetBSD Project
chuq@chuq.com, http://www.netbsd.org/

Abstract ¯ “buffer cache”:

A pool of memory allocated during system startup
This paper introduces UBC (“Unified Buffer Cache”), a which is dedicated to caching filesystem data and is
design for unifying the filesystem and virtual memory managed by special-purpose routines.
caches of file data, thereby providing increased system
performance. In this paper we discuss both the tradi- The memory is organized into “buffers,” which are
tional BSD caching interfaces and new UBC interfaces, variable-sized chunks of file data that are mapped to
concentrating on the design decisions that were made as kernel virtual addresses as long as they retain their
the design progressed. We also discuss the designs used identity.
by other operating systems to solve the same problems
that UBC solves, with emphasis on the practical impli- ¯ “page cache”:
cations of the differences between these designs. This The portion of available system memory which is
project is still in progress, and once completed will be used for cached file data and is managed by the VM
part of a future release of NetBSD. system.
The amount of memory used by the page cache can
1 Introduction vary from nearly 0% to nearly 100% of the physical
memory that isn’t locked into some other use.
Modern operating systems allow filesystem data to be
¯ “vnode”:
accessed using two mechanisms: memory mapping, and
I/O system calls such as read() and write(). In The kernel abstraction which represents a file.
traditional UNIX-like operating systems, memory map-
ping requests are handled by the virtual memory sub- Most manipulation of vnodes and their associated
system while I/O calls are handled by the I/O subsys- data are performed via “VOPs” (short for “vnode
tem. Traditionally these two subsystems were developed operations”).
separately and were not tightly integrated. For example,
in the NetBSD operating system[1], the VM subsystem
(“UVM”[2]) and I/O subsystem each have their own data The major interfaces for accessing file data are:
caching mechanisms that operate semi-independently of
each other. This lack of integration leads to inefficient
overall system performance and a lack of flexibility. To ¯ read() and write():
achieve good performance it is important for the vir- The read() system call reads data from disk into
tual memory and I/O subsystems to be highly integrated. the kernel’s cache if necessary, then copies data
This integration is the function of UBC. from the kernel’s cached copy to the application’s
address space. The write() system call moves
data the opposite direction, copying from the appli-
2 Background cation’s address space into the kernel’s cache and
eventually writing the data from the cache to disk.
In order to understand the improvements made in UBC, These interfaces can be implemented using either
it is important to first understand how things work with- the buffer cache or the page cache to store the data
out UBC. First, some terms: in the kernel.
 ¯ mmap(): application
The mmap() system call gives the application di-
rect memory-mapped access to the kernel’s page
cache data. File data is read into the page cache
lazily as processes attempt to access the mappings
created with mmap() and generate page faults. mmap read/write

In NetBSD without UBC, read() and write() are

implemented using the buffer cache. The read() sys-
tem call reads file data into a buffer cache buffer and then
copies it to the application. The mmap() system call, page cache
however, has to use the page cache to store its data since
the buffer cache memory is not managed by the VM sys-
tem and thus not cannot be mapped into an application extra data
address space. Therefore the file data in the buffer cache copy
is copied into page cache pages, which are then used
to satisfy page faults on the application mappings. To
write modified data in page cache pages back to disk, buffer cache
the new version is copied back to the buffer cache and
from there is written to disk. Figure 1 shows the flow of
data between the disk and the application with a tradi-
tional buffer cache.
raw device
This double-caching of data is a major source of ineffi-
ciency. Having two copies of file data means that twice
Figure 1: NetBSD before UBC.
as much memory is used, which reduces the amount of
memory available for applications. Copying the data
back and forth between the buffer cache and the page
cache wastes CPU cycles, clobbers CPU caches and is 3 So what is UBC anyway?
generally bad for performance. Having two copies of
the data also allows the possibility that the two copies
UBC is a new subsystem which solves the problems with
will become inconsistent, which can lead to application
the two-cache model. In the UBC model, we store file
problems which are difficult to debug.
data in the page cache for both read()/write() and
mmap() accesses. File data is read directly into the
The use of the buffer cache for large amounts of data is
page cache without going through the buffer cache by
generally bad, since the static sizing of the buffer cache
creating two new VOPs which return page cache pages
means that the buffer cache is often either too small (re-
with the desired data, calling into the device driver to
sulting in excessive cache misses), or too large (resulting
read the data from disk if necessary. Since page cache
in too little memory left for other uses).
pages aren’t always mapped, we created a new mecha-
nism for providing temporary mappings of page cache
The buffer cache also has the limitation that cached data
pages, which is used by read() and write() while
must always be mapped into kernel virtual space, which
copying the file data to the application’s address space.
puts an additional artificial limit on the amount of data
Figure 2 shows the changed data flow with UBC.
which can be cached since modern hardware can easily
have more RAM than kernel virtual address space.
UBC introduces these new interfaces:
To solve these problems, many operating systems have
changed their usage of the page cache and the buffer
¯ VOP GETPAGES(), VOP PUTPAGES()
cache. Each system has its own variation, so we will
describe UBC first and then some other popular operat- These new VOPs are provided by the filesystems
ing systems. to allow the VM system to request ranges of pages
 application the mappings created by ubc alloc(). (A UVM
pager is an abstraction which embodies knowledge
of page-fault resolution and other VM data man-
agement. See the UVM paper[2] for more informa-
tion on pagers.) Since its only purpose is to han-
dle those page faults, the only action performed by
mmap read/write ubc pager is to call the new VOP GETPAGES()
operation to get pages as needed to resolve the
faults.

In addition to these new interfaces, several changes were

made to the existing UVM design to fix problems which
page cache were glossed over in the original design.

Previously in UVM, vnodes and uvm objects were

not interchangeable, and in fact several fields were du-
plicated and maintained separately in each. These du-
plicate fields were combined. At this time there’s still a
raw device bit of extra initialization the first time a struct vn-
ode is used as a struct uvm object, but that will
Figure 2: NetBSD with UBC. be removed eventually.

Previously UVM only supported 32-bit offsets into

uvm objects, which meant that data could only be
to be read into memory from disk or written from
stored in the page cache for the first 4 GB of a file.
memory back to disk. VOP GETPAGES() must al-
This wasn’t much of a problem before since the num-
locate pages from the VM system for data which is
ber of programs which wanted to access file offsets past
not already cached and then initiate device I/O op-
4 GB via mmap() was small, but now that read() and
erations to read all the disk blocks which contain
write() also use the page cache interfaces to access
the data for those pages. VOP PUTPAGES() must
data, we had to support 64-bit uvm object offsets in
initiate device I/Os to write dirty pages back to disk.
order to continue to allow any access to file offsets past
¯ ubc alloc(), ubc release() 4 GB.
These functions allocate and free temporary map-
pings of page cache file data. These are the page
cache equivalents of the buffer cache functions 4 What do other operating systems do?
getblk() and brelse()[3]. These tempo-
rary mappings are not wired, but they are cached The problems addressed by UBC have been around for
to speed repeated access to the same file. The a long time, ever since memory-mapped access to files
selection of which virtual addresses to use for was first introduced in the late 1980’s. Most UNIX-like
these temporary mappings is important on hard- operating systems have addressed this issue one way or
ware which has a virtually-addressed CPU data another, but there are considerable differences in how
cache, so the addresses are carefully chosen to be they go about it.
correctly aligned with the preferred addresses for
user file mappings, so that both kinds of mappings The first operating system to address these problems
can be present at the same time without creating co- was SunOS[4, 5], and UBC is largely modeled after
herency problems in the CPU cache. It is still pos- this design. The main differences in the design of the
sible for applications to create unaligned file map- SunOS cache and UBC result from the differences be-
pings, but if the application lets the operating sys- tween the SunOS VM system and UVM. Since UVM’s
tem choose the mapping address then all mappings pager abstraction and SunOS’s segment-driver abstrac-
will always be aligned. tion are similar, this didn’t change the design much at
all.
¯ ubc pager
This is a UVM pager which handles page faults on When work on UBC first began over two years ago, the
other design that we examined was that of FreeBSD[6], 5 Performance
which had also already dealt with this problem. The
model in FreeBSD was to keep the same buffer cache in-
terfaces to access file data, but to use page cache pages as Since UBC is unfortunately not yet finished, a detailed
the memory for a buffer’s data rather than memory from performance analysis would be premature. However, we
a separate pool. The result is that the same physical page have made some simple comparisons just to see where
is accessed to retrieve a given range of the file regardless we stand. The hardware used for this test was a 333MHz
of whether the access is made via the buffer cache inter- Pentium II with 64MB of RAM and a 12GB IDE disk.
face or the page cache interface. This had the advantage The operations performed were a series of “dd” com-
that filesystems did not need to be changed in order to mands designed to expose the behaviour of sequential
take benefit from the changes. However, the glue to do reads and writes. We create a 1GB file (which is much
the translation between the interfaces was just as com- larger than the physical memory available for caching),
plicated as the glue in the SunOS design and failed to then overwrite this file to see the speed at which the
address certain deadlock problems (such as an applica- data modifications caused by the write() are flushed
tion calling write() with a buffer which was a mem- to disk without the overhead of allocating blocks to the
ory mapping of the same file being written to), so we file. Then we read back the entire file to get an idea of
chose the SunOS approach over this one. how fast the filesystem can get data from the disk. Fi-
nally, we read the first 50MB of the file (which should
The approach taken by Linux[7] (as of kernel version fit entirely in physical memory) several times to deter-
2.3.44, the latest version at the time this paper was writ- mine the speed of access to cached data. See Table 1 for
ten) is actually fairly similar to the SunOS design also. the results of these tests.
File data is stored only in the page cache. Temporary
mappings of page cache pages to support read() and The great disparity in the results of the first four tests
write() usually aren’t needed since Linux usually on the three non-UBC operating systems is due to dif-
maps all of physical memory into the kernel’s virtual ad- ferences in performance of their IDE disk drivers. All of
dress space all the time. One interesting twist that Linux the operating systems tested except NetBSD with UBC
adds is that the device block numbers where a page is do sequential buffered reads from a large file at the same
stored on disk are cached with the page in the form of a speed as reads from the raw device, so all we can really
list of buffer head structures. When a modified page say from this is that the other caching designs don’t add
is to be written back to disk, the I/O requests can be sent any noticable overhead. For reads, the UBC system is
to the device driver right away, without needing to read not yet running at device speed, so there’s still room for
any indirect blocks to determine where the page’s data improvement. Further analysis is required to determine
should be written. the cause of the slowdown.

The last of the operating systems we examined, HP- UBC obviously needs much improvement in the area of
UX, takes a completely different stance on the issue of write performance. This is partly due to UVM not being
how to cache filesystem data. HP-UX continues to store very efficient about flushing modified pages when mem-
file data in both the buffer cache and the page cache, ory is low and partly because the filesystem code cur-
though it does avoid the extra of copying of data that is rently doesn’t trigger any asynchronous writes to disk
present in pre-UBC NetBSD by reading data from disk during a big sequence of writes, so the writes to disk
directly into the page cache. The reasoning behind this are all started by the inefficient UVM code. We’ve been
is apparently that most files are only accessed by either concentrating on read performance so far, so this poor
read()/write() or mmap(), but not both, so as long write performance is not surprising.
as both mechanisms perform well individually, there’s
no need to redesign HP-UX just to fix the coherency is- The interesting part of this test series is the set of tests
sue. There is some attempt made to avoid incoherency where we read the same 50MB file five times. This
between the two caches, but locking constraints prevent clearly shows the benefit of the increased memory avail-
this from being completely effective. able for caching in the UBC system over NetBSD with-
out UBC. In NetBSD 1.4.2, all five reads occured at the
There are other operating systems which have imple- speed of the device, whereas in all the other systems
mented a unified cache (eg. Compaq’s Tru64 UNIX and the reads were completed at memory speed after sev-
IBM’s AIX), but we were unable to find information on eral runs. We have no explanation for why FreeBSD and
the design of these operating systems for comparison. Linux didn’t complete the second 50MB read at memory
speed, or why Linux didn’t complete even the third read
 Experiment Run Time (seconds)
Input Output Size NetBSD NetBSD FreeBSD Linux
1.4.2 with UBC 3.4 2.2.12-20smp
raw device /dev/null 1GB 72.8 72.7 279.3 254.6
/dev/zero new file 1GB 83.8 193.0 194.3 163.9
/dev/zero overwrite file 1GB 79.4 186.6 192.2 167.3
non-resident file /dev/null 1GB 72.7 86.7 279.3 254.5
non-resident file /dev/null 50MB 3.6 4.3 13.7 12.8
resident file /dev/null 50MB 3.6 0.8 4.1 11.5
repeat above /dev/null 50MB 3.6 0.8 0.7 4.5
repeat above /dev/null 50MB 3.6 0.8 0.7 0.8
repeat above /dev/null 50MB 3.6 0.8 0.7 0.8

Table 1: UBC performance comparison.

at memory speed. much larger amount of page-cache data which will

be dirty.

6 Conclusion ¯ Elimination of the data copying in read() and

write() via UVM page loanout when possible.
This could be done without UBC too, but with
In this paper we introduced UBC, a improved design for UBC it will be zero-copy instead of one-copy (from
filesystem and virtual memory caching in NetBSD. This buffer cache to page cache).
design includes many improvements over the previous
design used in NetBSD by: ¯ Elimination of the need to map pages to do I/O to
them by adding a page list to struct buf and
adding glue in bus dma to map pages temporarily
¯ Eliminating double caching of file data in the kernel for hardware that actually needs that.
(and the possibility of cache incoherency that goes
with it) when the same file is accessed via different ¯ Adding support for “XIP” (eXecute In Place). This
interfaces. will allow zero-copy access to filesystem images
¯ Allowing more flexibility in how physical memory stored in flash roms or other memory-mapped stor-
is used, which can greatly improve performance for age devices.
applications whose data fits in physical memory.
¯ Adding support for cache coherency in layered
filesystems. (The current UBC design does not ad-
dress caching in layered filesystems.)
7 Availability

This work will be part of a future release of NetBSD Acknowledgments

once it is completed. Until then, the source code is
available in the “chs-ubc2” branch in the NetBSD
CVS tree, accessible via anonymous CVS. See We would like to thank everyone who helped review
http://www.netbsd.org/Sites/net.html drafts of this paper. Special thanks to Chuck Cranor!
for details.

This being a work-in-progress, there is naturally much References

more work to do! Planned work includes:
[1] The NetBSD Project. The NetBSD Operating System. See
http://www.netbsd.org/ for more information.
¯ Integration of UBC into the NetBSD development [2] C. Cranor and G. Parulkar. The UVM Virtual Memory
source tree and performance improvement. The System. In Proceedings of the 1999 USENIX Technical
pagedaemon needs to be enhanced to deal with the Conference, June 1999.
[3] Marice J. Bach. The Design of the UNIX Operating Sys-
tem. Prentice Hall, February 1987.
[4] J. Moran, R. Gingell and W. Shannon. Virtual Memory
Architecture in SunOS. In Proceedings of USENIX Sum-
mer Conference, pages 81-94. USENIX, June 1987.
[5] J. Moran. SunOS Virtual Memory Implementation. In
Proceedings of the Spring 1988 European UNIX Users
Group Conference, April 1988.
[6] The FreeBSD Project. The FreeBSD Operating System.
See http://www.freebsd.org/ for more informa-
tion.
[7] L. Torvalds, et al. The Linux Operating System. See
http://www.linux.org/ for more information.