On operating systems we support, the behavior is always such that the
kernel will do the right thing as far as invalidating the file
descriptor, regardless of the eventual return value. Therefore,
potentially looping and calling close multiple times is wrong.
At best, we call close again on an invalid FD and throw a spurious EBADF
error. At worst, we might close an FD which doesn't belong to us when a
multi-threaded application opens its own file descriptor between
iterations of the loop.
Signed-off-by: Dave Reisner <dreisner@archlinux.org>
Signed-off-by: Allan McRae <allan@archlinux.org>
I suspect that eventually we're going to end up returning a pointer to
an allocated struct to describe the download result, but that's for
another patch when the need arises...
Fixes FS#33508.
Signed-off-by: Dave Reisner <dreisner@archlinux.org>
Signed-off-by: Allan McRae <allan@archlinux.org>
Users have hit issues behind corporate firewalls that initially throttle
downloads to ~1B/sec.
Signed-off-by: Olivier Langlois < olivier.pis.langlois@transport.alstom.com>
Signed-off-by: Allan McRae <allan@archlinux.org>
RFC 2616 doesn't forbid a 301 or 302 repsonse from having a body, and
servers exist in the wild that show this behavior. In order to prevent
pacman from showing a progress bar when we aren't actually downloading a
package (and merely following one of these pain in the butt redirects),
capture the server response code in the response header, rather than
waiting to peel it off the handle after the download has finished.
Signed-off-by: Dave Reisner <dreisner@archlinux.org>
Reported-by: Alexandre Filgueira <alexfilgueira@cinnarch.com>
Signed-off-by: Allan McRae <allan@archlinux.org>
Apparently gcc 4.7 has decided that -Wshadow warnings aren't worth
reporting anymore even with the flag enabled. These were found on
an Ubuntu 10.04 install.
Signed-off-by: Dave Reisner <dreisner@archlinux.org>
Signed-off-by: Dan McGee <dan@archlinux.org>
Detected by clang scan-build static code analyzer.
* Don't attempt to free an uninitialized gpgme key variable
* Initialize answer variable before asking frontend a question
* Pass by reference instead of value if uninitialized fields are
possible in download signal handler code
* Ensure we never call strlen() on NULL payload->remote_name value
Signed-off-by: Dan McGee <dan@archlinux.org>
To avoid conflicts on reusing a payload after a failed download, ensure
that we reset the filename hints in the payload struct prior to the
download operation.
Signed-off-by: Dave Reisner <dreisner@archlinux.org>
Signed-off-by: Dan McGee <dan@archlinux.org>
It can happen that the to-be-downloaded file cannot be created in cachedir.
For example, I am an -Sup user, and it is comfortable to set --cachedir to
/mnt/pendrive, which is a FAT filesystem, so files like
capseo-1:0.3-2-i686.pkg.tar.xz cannot be downloaded to there.
Before this patch, pacman didn't give clear output about what happens when
the download code could not create the necessary file. This can be confusing
with -Su. An example output:
***
$ sudo pacman -S capseo bochs --cachedir /c/TEMP
resolving dependencies...
looking for inter-conflicts...
Targets (2): bochs-2.4.6-1 capseo-1:0.3-2
Total Download Size: 0.61 MiB
Total Installed Size: 2.61 MiB
Proceed with installation? [Y/n]
:: Retrieving packages from extra...
warning: failed to retrieve some files from extra
bochs-2.4.6-1-i686 611.5 KiB 118K/s 00:05 [------------------] 97%
error: failed to commit transaction (unexpected error)
Errors occurred, no packages were upgraded.
***
After the patch, pacman will give more informative error message (and
pm_errno is set properly):
***
error: could not open file '/c/TEMP/capseo-1:0.3-2-i686.pkg.tar.xz.part': Invalid argument
error: failed to commit transaction (failed to retrieve some files)
***
Unfortunately, the "could not open file" error message is printed for
every mirror (that can be dozens of lines), which is ugly, but at least
informative... Without modifying the download logic (for example, by
introducing -2 return value for _alpm_download() to indicate giving up),
this ugliness cannot be eliminated.
Signed-off-by: Nagy Gabor <ngaba@bibl.u-szeged.hu>
Signed-off-by: Dan McGee <dan@archlinux.org>
This is particularly important in the case of FTP control connections,
which may be closed by rogue NAT/firewall devices detecting idle
connections on larger transfers which may take 5-10+ minutes.
Signed-off-by: Dave Reisner <dreisner@archlinux.org>
Signed-off-by: Dan McGee <dan@archlinux.org>
We lost this logic somewhere between the libfetch and libcurl
transition, as it existed in the internal downloader, but was pulled
back only into the sync workflow. Add a helper function that will let us
check for existance in the filecache prior to calling the downloader.
Signed-off-by: Dave Reisner <dreisner@archlinux.org>
Signed-off-by: Dan McGee <dan@archlinux.org>
Ensures that config.h is always ordered correctly (first) in the
includes. Also means that new source files get this for free without
having to remember to add it.
We opt for -imacros over -include as its more portable, and the
added constraint by -imacros doesn't bother us for config.h.
This also touches the HACKING file to remove the explicit mention of
config.h as part of the includes.
Signed-off-by: Dave Reisner <dreisner@archlinux.org>
Signed-off-by: Dan McGee <dan@archlinux.org>
We had a 16 KiB limit on database signatures, we should do the same here
too to have a slight sanity check, even if we can't do so for the
package itself yet.
Signed-off-by: Dan McGee <dan@archlinux.org>
Replacing the strdup when after the first NULL check assures that we get
continue with payload->remote_name defined.
Signed-off-by: Dave Reisner <dreisner@archlinux.org>
Dan: fix mask calculation, add it to the success/fail block instead.
Signed-off-by: Dave Reisner <dreisner@archlinux.org>
Signed-off-by: Dan McGee <dan@archlinux.org>
This adds a logger to the CURLE_OK case so we can always know the return
code if it was >= 400, and debug log it regardless. Also adjust another
logger to use the cURL error message directly, as well as use fstat()
when we have an open file handle rather than stat().
Signed-off-by: Dan McGee <dan@archlinux.org>
This is in the realm of "probably not going to happen", but if someone
were to translate "disk" to a string longer than 256 characters, we
would have a smashed/corrupted stack due to our unchecked strcpy() call.
Rework the function to always length-check the value we copy into the
hostname buffer, and do it with memcpy rather than the more cumbersome
and unnecessary snprintf.
Finally, move the magic 256 value into a constant and pass it into the
function which is going to get inlined anyway.
Signed-off-by: Dan McGee <dan@archlinux.org>
Similar to what we did in edd9ed6a, disconnect the relationship with our
stack allocated error buffer from the curl handle. Just as an FTP
connection might have some network chatter on teardown causing the
progress callback to be triggered, we might also hit an error condition
that causes curl to write to our (now out of scope) error buffer.
I'm unable to reproduce FS#26327, but I have a suspicion that this
should fix it.
Signed-off-by: Dave Reisner <dreisner@archlinux.org>
Signed-off-by: Dan McGee <dan@archlinux.org>
This is a poor place for it, and it will likely move again in the
future, but it's better to have it here than as a static variable.
Initialization of this variable is now no longer necessary as its
zeroed on creation of the payload struct.
Signed-off-by: Dave Reisner <dreisner@archlinux.org>
Signed-off-by: Dan McGee <dan@archlinux.org>
This was done to squash a memory leak in the sync database download
code. When we downloaded a database and then reused the payload struct,
we could find ourselves calling get_fullpath() for the signatures and
overwriting non-freed values we had left over from the database
download.
Refactor the payload_free function into a payload_reset function that we
can call that does NOT free the payload itself, so we can reuse payload
structs. This also allows us to move the payload to the stack in some
call paths, relieving us of the need to alloc space.
Signed-off-by: Dan McGee <dan@archlinux.org>
Rather than always initializing it on any handle creation. There are
several frontend operations (search, info, etc.) that never need the
download code, so spending time initializing this every single time is a
bit silly. This makes it a bit more like the GPGME code init path.
Signed-off-by: Dan McGee <dan@archlinux.org>
In the sync code, we explicitly allocated a string for this field, while
in the dload code itself it was filled in with a pointer to another
string. This led to a memory leak in the sync download case.
Make remote_name non-const and always explicitly allocate it. This patch
ensures this as well as uses malloc + snprintf (rather than calloc) in
several codepaths, and eliminates the only use of PATH_MAX in the
download code.
Signed-off-by: Dan McGee <dan@archlinux.org>
This function doesn't exist on OSX. Since there aren't any other
candidates in alpm for which this function would make sense to use,
simply replace the function call with a loop that does the equivalent.
Signed-off-by: Dave Reisner <dreisner@archlinux.org>
if ~/.netrc exists and has credentials for the hostname requested in a
download, they will be provided in an http auth request. This can still
be overridden by explcitly declaring user:pass in the URL.
Signed-off-by: Dave Reisner <dreisner@archlinux.org>
Signed-off-by: Dan McGee <dan@archlinux.org>
This gives us some amount of room to grow in case we ever find another
reason that we might return with an error from the progress callback.
Signed-off-by: Dave Reisner <dreisner@archlinux.org>
We lost some of this output in the fetch->curl conversion, but I also
noticed in FS#25852 that we just lack some of this useful information
along the way.
Signed-off-by: Dave Reisner <dreisner@archlinux.org>
After commit 2e7d002315, we use off_t rather than long variables.
Use the _LARGE variants of the methods to indicate we are passing off_t
sized variables, and cast using (curl_off_t) accordingly.
Signed-off-by: Dan McGee <dan@archlinux.org>
This handles the no Content-Length header problem as stated in the
comments of FS#23413. We add a quick check to the callback that will
force an abort if the downloaded data exceeds the payload size, and then
check for this error in the post-download cleanup code.
Signed-off-by: Dan McGee <dan@archlinux.org>
Beautiful of libcurl to use floating point types for what are never
fractional values. We can do better, and we usually want these values in
their integer form anyway.
Signed-off-by: Dan McGee <dan@archlinux.org>
Since we store this directly in the download function, just rework
mask_signal() to take a pointer to a location to store the original.
Signed-off-by: Dan McGee <dan@archlinux.org>
This is a precursor to a following patch which will move the setting of
options to a separate function. With the open mode as part of the
struct, we can avoid modifying stack allocated variables.
Signed-off-by: Dave Reisner <dreisner@archlinux.org>
Signed-off-by: Dan McGee <dan@archlinux.org>
This is more in line with the menagerie of file name members that we now
have on the payload struct.
Signed-off-by: Dave Reisner <dreisner@archlinux.org>
Signed-off-by: Dan McGee <dan@archlinux.org>
These are private to the download operation already, so glob them onto
the struct. This is an ugly rename patch, with the only logical change
being that destfile and tempfile are now freed by the payload_free
function.
Signed-off-by: Dave Reisner <dreisner@archlinux.org>
Signed-off-by: Dan McGee <dan@archlinux.org>
