| Age | Commit message (Collapse) | Author |
|---|
|
The maximum parallelization soft limit is the number of CPU
cores * 2 on systems defining _SC_NPROCESSORS_ONLN. The hard
limit in all cases is Acquire::QueueHost::Limit.
|
|
This is a multiple of the page size and thus results in less
page faults, speeding up copying.
Also, while we're at at, unify all uses of that size in a
constant variable APT_BUFFER_SIZE.
|
|
Implement native support for LZ4 compression, using the official
lz4 library.
|
|
This drop the hash table utilization from a high 98%
to acceptable 74% on unstable, and the average bucket length
from 4.6 to 1.8.
This improves performance by about 5%, while increasing
the size of the cache by 0.2 out of 38MB, that is 0.5%.
48481 is a nice number
|
|
Gbp-Dch: ignore
|
|
This will give us the freedom to insert more compressors at
positions in between.
Also change the cost of uncompressed to 0, as that really has
no overhead, and the values do not really mean much.
|
|
This makes code easier to read, and somewhat more correct.
Gbp-Dch: ignore
|
|
Gbp-Dch: ignore
|
|
apt_preferences and deb822-style sources used the specialized class
pkgUserTagSection to deal with comments before/after a given stanza, but
it couldn't deal with comments in the stanza at all.
codesearch suggests that nobody else does and a vastely superior way of
working with potentially commented files is implemented now, so we can
officially discourage the use of the old incomplete hack class.
|
|
Now (55153bf94ff28a23318e79aa48242244c4d82b3c) that pkgTagFile can be
told to deal with all sorts of comments we can use this mode to parse
dsc (as by catch) and debian/control files properly even in the wake of
multiline fields spliced with comments like Build-Depends.
Closes: 806775
|
|
APT usually deals with perfectly formatted files generated automatically
be other programs – and as it has to parse multiple MBs of such files it
tries to be fast rather than forgiving.
This was always a problem if we reused this parser for files with a
deb822 syntax which are mostly written by hand however, like
apt_preferences or the deb822-style sources as these can include stray
newlines and more importantly comments all over the place.
As a stopgap we had pkgUserTagSection which deals at least with comments
before and after a given stanza, but comments in between weren't really
supported and now that we support parsing debian/control for e.g.
build-dep we face the full comment problem e.g. with comments inbetween
multi-line fields (like Build-Depends).
We can't easily deal with this on the pkgTagSection level as the interface
gives access to 'raw' char-pointers for performance reasons so we would
need to optionally add a buffer here on which we could remove comments
to hand out pointers into this buffer instead. The interface is quite
large already and supports writing stanzas as well, which does not
support comments at all either. So while in future it might make sense
to have a parser setup which deals with and keeps comments in this
commit we opt for the simpler solution for now: We officially declare
that pkgTagSection does not support comments and instead expect the
caller to deal with them, which in our case is pkgTagFile:
pkgTagFile is extended with an additional mode which can deal with
comments by dropping them from the buffer which will later form the
input of pkgTagSection. The actual implementation is slightly more
complex than this sentence suggests at first on one hand to have good
performance and on the other to allow jumping directly to stanzas with
offsets collected in a previous run (like our cache generation does it
for example).
|
|
Integrity is taken care of by the checksum now.
|
|
|
|
If we already have opened a cache, there is no point in having
to open it again.
|
|
We ignored the boundary of the buffer we were reading in
while scanning for spaces.
|
|
This shuts up gcc
Gbp-Dch: ignore
|
|
To preserve compatibility, the new inline functions have _inline
as a suffix, and a macro defines the old names to refer to the
inline variants.
The old functions are still preserved for binary compatibility.
Also simplify the implementation of both functions.
|
|
On my testing system, consisting of unstable and experimental,
this reduces the average chain from 6.5 to 4.5, and the longest
chain from 17 to 15.
|
|
Previously, if flush errored inside the loop, data could have
already been written to the wrapped descriptor without having
been removed from the buffer.
Also try to work around EINTR here. A better solution might be
to have the individual privates detect an interrupt and return
0 in such a case, instead of relying on errno being untouched
in between the syscall and the return from InternalWrite.
|
|
Since commit 7a68effcb904b4424b54a30e448b6f2560cd1078, the xz
and lzma compressors read the level of compression they shall
use.
A default of -9 is too much for them, this will use 674 MB,
according to the xz manual page. Level -6 on the other hand
only needs 94 MB memory for compression.
This causes autopkgtest failures in the test-compressed-indexes
test, as not enough memory exists to proceed.
Change the other compression levels to 6 as well: The gzip
and bzip2 FileFd backends do not read them, and use their
code's default level which is 6, so do the same for external
methods.
|
|
This avoids some issues with InternalWrite returning 0 because
it just cannot write stuff at the moment.
|
|
Regression introduced in 8710a36a01c0cb1648926792c2ad05185535558e,
but such fields are unlikely in practice as it is just as simple to not
have a field at all with the same result of not having a value.
Closes: 808102
|
|
Debian has a Packages file for arch:all already, but the arch:any files
contain arch:all packages as well, so downloading it would be a total
waste of resources. Getting this solved is on the list of things to do,
but it is also the hardest part – for index targets like Contents the
situation is much easier and less server/client implementations are
involved so we might not want to stall them.
A repository can now declare via:
No-Support-for-Architecture-all: Packages
that even if an arch:all Packages exists, it shouldn't be downloaded, so
that support for Contents files can be added now.
See also 1dd20368486820efb6ef4476ad739e967174bec4 for the implementation
of downloading arch:all index targets, which this is limiting.
The field uses the name of the target from the apt configuration for
simplicity and is negative by design as this field is intended to be
supported/needed only for a "short" time (one or two Debian releases).
While this commit theoretically supports any target, its expected to
only see "Packages" as a value in reality.
|
|
This is for public users only, which cannot use the class at all,
except for the static methods.
|
|
This is somewhat experimental right now, and might not work
for everyone, so it is on an opt-in basis.
|
|
The flush function can be used for buffered writers.
|
|
We will soon implement a buffered writing decorator and we will
need to forward attribute changes to those.
|
|
These can be used to implement write buffering
Gbp-Dch: ignore
|
|
Suggested by David.
Gbp-Dch: ignore
|
|
Gbp-Dch: ignore
|
|
We do not see those branches at all during normal mode of
operation (that is, during cache generation), so tell the
compiler about it.
|
|
The Set() method returns false if the input is no hex number,
so simply use that.
|
|
We directly check if we are a hex digit in HexDigit, so use that
information.
[jak@debian.org: Commit message wording]
|
|
This makes the code parsing architecture lists slower, but on
the other hand, improves the more generic case of reading
dependencies from Packages files.
|
|
std::unordered_map is faster than std::map in our use case,
reducing cache generation time by about 10% in my benchmark.
|
|
This converts all callers that read machine-generated data,
callers that might work with user input are not converted.
|
|
This is like isspace(), but ignores the current locale.
|
|
There is not much point and this is more readable.
Gbp-Dch: ignore
|
|
This is mostly a documentation issue, as the size we want to
read is always less than or equal to the size of the buffer,
so the return value will be the same as the size argument.
Nonetheless, people wondered about it, and it seems clearer
to just always use the return value.
|
|
This further improves our performance, and rred on uncompressed
files now spents 78% of its time in writing. Which means that
we should really look at buffering those.
|
|
The code uses memmove() to move parts of the buffer to the
front when the buffer is only partially read. By simply
reading one page at a time, the maximum size of bytes that
must be moved has a hard limit, and performance improves:
In one test case, consisting of a 430 MB Contents file,
and a 75K PDiff, applying the PDiff previously took about
48 seconds and now completes in 2 seconds.
Further speed up can be achieved by buffering writes, they
account for about 60% of the run-time now.
|
|
Gbp-Dch: ignore
|
|
And as we are at it lets fix the 'style' issue I introduced with the
filefd changes as well.
Reported-By: gcc -fsanitize's & cppcheck
Git-Dch: Ignore
|
|
We don't need the buffer that often - only for ReadLine - as it is only
occasionally used, so it is actually more efficient to allocate it if
needed instead of statically by default. It also allows the caller to
influence the buffer size instead of hardcoding it.
Git-Dch: Ignore
|
|
The default implementation of ReadLine was very naive by just reading
each character one-by-one. That is kinda okay for libraries implementing
compression as they have internal buffers (but still not great), but
while working with files directly or via a pipe as there is no buffer
there so all those reads are in fact system calls.
This commit introduces an internal buffer in the FileFd implementation
which is only used by ReadLine. The more low-level Read and all other
actions remain unbuffered – they just changed to deal with potential
"left-overs" in the buffer correctly.
Closes: 808579
|
|
If we use the library to compress xz, still try to understand and pick
up the arguments we would have used to call xz to figure out which level
the user wants us to use instead of defaulting to level 6 (which is the
default level of xz).
|
|
dpkg switched from CRC32 to CRC64 in
777915108d9d36d022dc4fc4151a615fc95e5032 with the message:
| This is the default CRC used by the xz command-line tool, align with
| it and switch from CRC32 to CRC64. It should provide slightly better
| detection against damaged data, at a negligible speed difference.
|
|
This isn't implementing any new features, it is "just" moving code
around from FileFd methods which decided on each call how to handle the
request by including all logic for all possible compressor backends in
the method body to a model in which backend-specifics are implemented in
a FileFdPrivate subclass. This avoids a big chunk of #ifdef's and should
make it a tiny bit more obvious which backend uses which code.
The execution of the idea is slightly uglified by the need to preserve
ABI and API which causes liberal befriending.
Git-Dch: Ignore
|
|
There's no point trying to read 0 bytes, so let's just not
do this and switch to a while loop like in Write().
Gbp-Dch: ignore
|
|
Turn the do-while loop into while loops, so it simply does nothing
if the Size is already 0.
This reverts commit c0b271edc2f6d9e5dea5ac82fbc911f0e3adfa7a which
introduced a fix for a specific instance of the issue in the
CopyFile() function.
Closes: #808381
|
