Module Name: src
Committed By: riastradh
Date: Thu Feb 23 02:58:41 UTC 2023
Modified Files:
src/sys/kern: kern_descrip.c
Log Message:
kern_descrip.c: Use atomic_store_relaxed/release for ff->ff_file.
1. atomic_store_relaxed in fd_close avoids the appearance of race in
sanitizers (minor bug).
2. atomic_store_release in fd_affix is necessary because the lock
activity was not, in fact, enough to guarantee ordering (real bug
some architectures like aarch64).
The premise appears to have been that the mutex_enter/exit earlier
in fd_affix is enough to guarantee that initialization of fp (A)
happens before use of fp by a user once fp is published (B):
fp->f_... = ...; // A
/* fd_affix */
mutex_enter(&fp->f_lock);
fp->f_count++;
mutex_exit(&fp->f_lock);
...
ff->ff_file = fp; // B
But actually mutex_enter/exit allow the following reordering by
the CPU:
mutex_enter(&fp->f_lock);
ff->ff_file = fp; // B
fp->f_count++;
fp->f_... = ...; // A
mutex_exit(&fp->f_lock);
The only constraints they imply are:
1. fp->f_count++ and B cannot precede mutex_enter
2. mutex_exit cannot precede A and fp->f_count++
They imply no constraint on the relative ordering of A, B, and
fp->f_count++ amongst each other, however.
This affects any architecture that has a native load-acquire or
store-release operation in mutex_enter/exit, like aarch64, instead
of explicit load-before-load/store and load/store-before-store
barrier.
No need for atomic_store_* in fd_copy or fd_free because we have
exclusive access to ff as is.
XXX pullup-9
XXX pullup-10
To generate a diff of this commit:
cvs rdiff -u -r1.252 -r1.253 src/sys/kern/kern_descrip.c
Please note that diffs are not public domain; they are subject to the
copyright notices on the relevant files.
