and add a comment about related magic (many lines)).
__kernel_cos[f]() needs a trick to reduce the error to below 1 ulp
when |x| >= 0.3 for the range-reduced x. Modulo other bugs, naive
code that doesn't use the trick would have an error of >= 1 ulp
in about 0.00006% of cases when |x| >= 0.3 for the unreduced x,
with a maximum relative error of about 1.03 ulps. Mistransation
of the trick from the double precision case resulted in errors in
about 0.2% of cases, with a maximum relative error of about 1.3 ulps.
The mistranslation involved not doing implicit masking of the 32-bit
float word corresponding to to implicit masking of the lower 32-bit
double word by clearing it.
sinf() uses __kernel_cosf() for half of all cases so its errors from
this bug are similar. tanf() is not affected.
The error bounds in the above and in my other recent commit messages
are for amd64. Extra precision for floats on i386's accidentally masks
this bug, but only if k_cosf.c is compiled with -O. Although the extra
precision helps here, this is accidental and depends on longstanding
gcc precision bugs (not clipping extra precision on assignment...),
and the gcc bugs are mostly avoided by compiling without -O. I now
develop libm mainly on amd64 systems to simplify error detection and
debugging.
updated by a process holding the snapshot lock. Another process updating a
different inode in the same inodeblock will do copy on write checks and lock in
the opposite direction.
The snapshot code force a copy on write of these blocks manually (cf. start of
expunge_ufs[12]) and these inode blocks are later put on snapblklist.
This partial fix is to 'drain' the relevant ffs_copyonwrite() operation after
installing new snapblklist. This is not a 100% solution since a failed block
allocation can cause implicit fsync() which might deadlock before the new
snapblklist has been installed.
file is flushed by a process not holding snaplk (e.g. bufdaemon). Another
process might hold snaplk and try to access the block due to ffs_copyonwrite
processing.
the cg map buffer being held when writing indirect blocks. The process ends up
in ffs_copyonwrite(), attempting to get snaplk while holding the cg map buffer
lock.
Another process might be in ffs_copyonwrite(), trying to allocate a new block
for a copy. It would hold snaplk while trying to get the cg map buffer lock.
Release the cg map buffer early and use the copy for most of the cgaccount
processing to avoid this deadlock.
skipping the call from ffs_snapremove() if the block number is zero.
Simplify snapshot locking in ffs_copyonwrite() and ffs_snapblkfree() by using
the same locking protocol for low block numbers as for larger block numbers.
This removes a lock leak that could happen if vn_lock() succeeded after
lockmgr() failed in ffs_snapblkfree().
Check if snapshot is gone before retrying a lock in ffs_copyonwrite().
reclamation. If the vnode previously was a fifo then v_op would point to
ffs_fifoops[12] instead of the expected ffs_vnodeops[12], causing a panic at
the end of ffsext_strategy.
the UDF specification specifies a logical sectorsize of 2048.
Instead, get it from GEOM.
- When reading the UDF Anchor Volume Descriptor, use the logical
sectorsize of 2048 when calculating the offset to read from, but
use the actual sectorsize to determine how much to read.
- works with reading a DVD disk and a DVD disk image file via mdconfig
- correctly returns EINVAL if we try to mount_udf an audio CD, instead
of panicking inside GEOM when INVARIANTS is set
a functional change. I know this is a big no no, so this is a forced commit
to note the functional changes from my previous revision:
@@ -196,7 +176,7 @@ wait_for_lock(const char *name, int flag
int fd;
if ((fd = open(name, O_CREAT|O_RDONLY|O_EXLOCK|flags, 0666)) == -1) {
- if (errno == ENOENT || errno == EINTR || errno == EAGAIN)
+ if (errno == EINTR || errno == EAGAIN)
return (-1);
err(EX_CANTCREAT, "cannot open %s", name);
}
17+17+24 bit pi/2 must only be used when subtraction of the first 2
terms in it from the arg is exact. This happens iff the the arg in
bits is one of the 2**17[-1] values on each side of (float)(pi/2).
Revert to the algorithm in rev.1.7 and only fix its threshold for using
the 3-term pi/2. Use the threshold that maximizes the number of values
for which the 3-term pi/2 is used, subject to not changing the algorithm
for comparing with the threshold. The 3-term pi/2 ends up being used
for about half of its usable range (about 64K values on each side).
a maximum error of 2.905 ulps for cosf(), but the algorithm for cosf()
is good for < 1 ulps and happens to give perfect rounding (< 0.5 ulps)
near +-pi/2 except for the bug. The extra relative errors for tanf()
were similar (slightly larger). The bug didn't affect sinf() since
sinf'(+-pi/2) is 0.
For range reduction in ~[-3pi/4, -pi/4] and ~[pi/4, 3pi/4] we must
subtract +-pi/2 and the only complication is that this must be done
in extra precision. We have handy 17+24-bit and 17+17+24-bit
approximations to pi/2. If we always used the former then we would
lose up to 24 bits of accuracy due to cancelation of leading bits, but
we need to keep at least 24 bits plus a guard digit or 2, and should
keep as many guard bits as efficiency permits. So we used the
less-precise pi/2 not very near +-pi/2 and switched to using the
more-precise pi/2 very near +-pi/2. However, we got the threshold for
the switch wrong by allowing 19 bits to cancel, so we ended up with
only 21 or 22 bits of accuracy in some cases, which is even worse than
naively subtracting pi/2 would have done.
Exhaustive checking shows that allowing only 17 bits to cancel (min.
accuracy ~24 bits) is sufficient to reduce the maximum error for cosf()
near +-pi/2 to 0.726 ulps, but allowing only 6 bits to cancel (min.
accuracy ~35-bits) happens to give perfect rounding for cosf() at
little extra cost so we prefer that.
We actually (in effect) allow 0 bits to cancel and always use the
17+17+24-bit pi/2 (min. accuracy ~41 bits). This is simpler and
probably always more efficient too. Classifying args to avoid using
this pi/2 when it is not needed takes several extra integer operations
and a branch, but just using it takes only 1 FP operation.
The patch also fixes misspelling of 17 as 24 in many comments.
For the double-precision version, the magic numbers include 33+53 bits
for the less-precise pi/2 and (53-32-1 = 20) bits being allowed to
cancel, so there are ~33-20 = 13 guard bits. This is sufficient except
probably for perfect rounding. The more-precise pi/2 has 33+33+53
bits and we still waste time classifying args to avoid using it.
The bug is apparently from mistranslation of the magic 32 in 53-32-1.
The number of bits allowed to cancel is not critical and we use 32 for
double precision because it allows efficient classification using a
32-bit comparison. For float precision, we must use an explicit mask,
and there are fewer bits so there is less margin for error in their
allocation. The 32 got reduced to 4 but should have been reduced
almost in proportion to the reduction of mantissa bits.
the modified interface that they use. Changes include:
- Register a different interrupt handler for the new interface. This one is
INTR_MPSAFE, not INTR_FAST, and directly processes completions and AIFs.
- Add an event registration and callback mechanism for the ioctl and CAM
modules can know when a resource shortage clears. This condition was
previously fatal in CAM due to programming oversights.
- Fix locking to play better with newbus.
- Provide access methods for talking to cards with the NEWCOMM interface.
- Fix up the CAM module to be better suited for dealing with newer firmware
on the PERC Si/Di series that requires talking to plain SCSI via aac.
- Add a whole slew of new PCI Id's.
Thanks to Adaptec for providing an initial version of this work and for
answering countless questions about it. There are still some rough edges in
this, but it works well enough to commit and test for now.
Obtained from: Adaptec, Inc.
o Rather than just try to turn off EXCA_INTR_RESET, set the entire register
to 0. This is slightly faster, and a better hammer.
o Move attempted clearing of the output enable (EXCA_PWRCTL_OE) back to
after we turn off the power. Modify it to write 0 so that we don't get
Bad Vcc messages on TI bridges (untested, but ru@ sent me a similar patch)
while at the same time avoiding interrupt storms on Ricoh bridges (tested
by me on my Sony).
# Many of my observations of 'breakage' for this patch are due to some bug
# in the load/unload of cbb.ko unlreated to this change. I'll be investigating
# and fixing that bug in the fullness of time.
an allocation. This fixes the malloc 'use after free' panic on boot that
many were seeing. It doesn't solve the problem of the allocations being
cached and then written past their bounds later. That will take more work.
Submitted by: kan
in 1993 in rev.1.5 of the i386 a.out version (csu/i386/crt0.c).
Profiling uses a magic label "eprol" to delimit the start of the part
of the text section covered by profiling. This label must be placed
before the call to main() to get main() properly profiled. It was
placed there in rev.1.1 of crt0.c. Rev.1.5 imported the initial
implementation of shared libraries in FreeBSD and misplaced the label.
Fortunately, the misplaced label was misspelled and the old label
wasn't removed, so the new label had no effect. Unfortunately, when
profiling was implemented for the ELF in 1998 in rev.1.2 of
csu/i386-elf/crt1.c, only the incorrectly placed label was copied
(after fixing its name). The bug was then copied to all other arches.
The label seems to be still misplaced in NetBSD for most arches. It
is in common.c for most arches so it is even further from being inside
the function that calls main().
I think "eprol" is short for "end of prologue", but it must be placed
before the end of the prologue so that it covers main(). crt0.c has
it before the calls atexit(_mcleanup) and monstartup(...), but it
cannot affect these calls so I moved it after the call to monstartup().
It now also covers the call to _init() but not the newer call to
_init_tls(). Profiling of _init() seems to be harmless, and the call
to _init_tls() seems to be misplaced.
Reviewed by: jdp (long ago, for a slightly different i386 version)
http://lists.freebsd.org/pipermail/cvs-src/2004-October/033496.html
The same problem applies to if_bridge(4), too.
- Copy-and-paste the if_bridge(4) related block from
if_ethersubr.c to ng_ether.c
- Add XXXs, so that copy-and-paste would be noticed by
any future editors of this code.
- Also add XXXs near if_bridge(4) declarations.
Silence from: thompsa
appear to be never called:
(1) If a function is never called according to its call count but it
must have been called because its child time is nonzero, then print
it in the flat profile. Previously, if its call count was zero
then we only printed it in the flat profile if its self time was
nonzero.
(2) If a function has a zero call count but has a nonzero self or child
time, then print its total self time in the self time per call
column as a percentage of the total (self + child) time. It is
not possible to print the times per call in this case because the
call count is zero. Previously, this was handled by leaving both
per-call columns blank. The self time is printed in another column
but there was no way to recover the total time.
(1) partially fixes the case of the "never called" function main() and
prepares for (2) to apply to main() and other functions. Profiling
of main() was lost in the conversion from a.out to ELF, so main()'s
call count has always been zero for many years; then in the common
case where main() is a tiny function, it gets no profiling ticks, so
main() was completely lost in the flat profile.
(2) improves mainly cases like kernel threads. Most kernel threads
appear to be never called because they are always started before
userland can run to turn on profiling. As for main(), the fact that
they are called is not very interesting and their callers are
uninteresting, but their relative self time is interesting since they
are long-running.
Almost always printing percentages in the per-call columns would be
more useful than almost always printing 0.0ms. 0.1ms is now a long
time, so only very large functions take that long per call. The accuracy
per call can approach 1-10 nsec provided programs are run for about
100000 times as long as is necessary to get this accuracy with high
resolution kernel profiling.
you want to see, e.g., sendmail arguments mail(1) will use.
-H is not an independent flag, it's a modifier. Also explicitly
say that -H will cause mail(1) to exit as soon as it prints the headers.
MFC after: 5 days
