os/doc
3
0
Fork 0
Code Issues Pull requests Projects Releases Wiki Activity

Add process-shared locks report from Konstantin Belousov

Browse source 
<kostikbel@gmail.com>.
This commit is contained in:
Warren Block 2016-04-12 21:50:01 +00:00
parent 97695eaa2a
commit a4a29bb20a
Notes: svn2git 2020-12-08 03:00:23 +00:00 
svn path=/head/; revision=48585
1 changed files with 85 additions and 0 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

85
en_US.ISO8859-1/htdocs/news/status/report-2016-01-2016-03.xml
View file

@ -748,4 +748,89 @@
</task> </task>
</help> </help>
</project> </project>
<project cat='proj'>
<title>Process-Shared locks for libthr</title>
<contact>
<person>
<name>
<given>Konstantin</given>
<common>Belousov</common>
</name>
<email>kib@FreeBSD.org</email>
</person>
</contact>
<body>
<p>POSIX specifies several kinds of pthread locks, for this
report the private and process-shared variants are considered.
Private locks can be used only by the threads of the same
process, which share the address space. Process-shared locks
can be used by threads from any process, assuming the process
can map the lock memory into its address space.</p>
<p>Our libthr, the library implementing the POSIX threads and
locking operations, uses a pointer as the internal
representation behind a lock. The pointer contains the
address of the actual structure carrying the lock. This has
unfortunate consequences for implementing the
<tt>PTHREAD_PROCESS_SHARED</tt> attribute for locks, since
really only the pointer is shared when the lock is mapped into
distinct address spaces.</p>
<p>A common opinion was that we have no choice but to break the
libthr Application Binary Interface (ABI) by changing the lock
types to be the actual lock structures (and padding for future
ABI extension). This is very painful for users, as our
previous experience with non-versioned libc and libc_r
shown.</p>
<p>Instead, I proposed and implemented a scheme where
process-shared locks can be implemented without breaking the
ABI. The lock memory is used as a key into the system-global
hash of the shared memory objects (off-pages), which carry the
real lock structures.</p>
<p>New umtx operations to create or look up the shared object,
by the memory key, were added. Libthr is modified to lookup
the object and use it for shared locks, instead of using
malloc() as for private locks.</p>
<p>The pointer value in the user-visible lock type contains a
canary for shared locks. Libthr detects the canary and
switches into the shared-lock mode.</p>
<p>The proposal of inlining the lock structures, besides the
drawbacks of breaking ABI, has its merits. Most important,
the inlining avoids the need of indirection. Another
important advantage over the off-page page approach is that no
off-page object needs to be maintained, and the lifecycle of
the shared lock naturally finishes with the destruction of the
shared memory, without explicit cleanup. Right now, off-pages
hook into vm object termination to avoid leakage, but long
liviness of the vnode vm object prolonges the off-page
existence for shared locks backed by files, however unlikely
they may be.</p>
<p>Libthr with inlined locks become informally known as libthr2
project, since the library name better be changed instead of
only bumping the library version. The rtld should ensure that
libthr and libthr2 do not become simultaneously loaded into a
single address space.</p>
</body>
<sponsor>The FreeBSD Foundation</sponsor>
<help>
<task>
<p>Implement robust mutexes.</p>
</task>
<task>
<p>Evaluate and implement libthr2.</p>
</task>
</help>
</project>
</report> </report>