The tempfile API keeps to-be-cleaned tempfiles in a
singly-linked list and never removes items from the list. A
future patch would like to start removing items, but removal
from a singly linked list is O(n), as we have to walk the
list to find the predecessor element. This means that a
process which takes "n" simultaneous lockfiles (for example,
an atomic transaction on "n" refs) may end up quadratic in
"n".
Before we start allowing items to be removed, it would be
nice to have a way to cover this case in linear time.
The simplest solution is to make an assumption about the
order in which tempfiles are added and removed from the
list. If both operations iterate over the tempfiles in the
same order, then by putting new items at the end of the list
our removal search will always find its items at the
beginning of the list. And indeed, that would work for the
case of refs. But it creates a hidden dependency between
unrelated parts of the code. If anybody changes the ref code
(or if we add a new caller that opens multiple simultaneous
tempfiles) they may unknowingly introduce a performance
regression.
Another solution is to use a better data structure. A
doubly-linked list works fine, and we already have an
implementation in list.h. But there's one snag: the elements
of "struct tempfile" are all marked as "volatile", since a
signal handler may interrupt us and iterate over the list at
any moment (even if we were in the middle of adding a new
entry).
We can declare a "volatile struct list_head", but we can't
actually use it with the normal list functions. The compiler
complains about passing a pointer-to-volatile via a regular
pointer argument. And rightfully so, as the sub-function
would potentially need different code to deal with the
volatile case.
That leaves us with a few options:
1. Drop the "volatile" modifier for the list items.
This is probably a bad idea. I checked the assembly
output from "gcc -O2", and the "volatile" really does
impact the order in which it updates memory.
2. Use macros instead of inline functions. The irony here
is that list.h is entirely implemented as trivial
inline functions. So we basically are already
generating custom code for each call. But sadly there's no
way in C to declare the inline function to take a more
generic type.
We could do so by switching the inline functions to
macros, but it does make the end result harder to read.
And it doesn't fully solve the problem (for instance,
the declaration of list_head needs to change so that
its "prev" and "next" pointers point to other volatile
structs).
3. Don't use list.h, and just make our own ad-hoc
doubly-linked list. It's not that much code to
implement the basics that we need here. But if we're
going to do so, why not add the few extra lines
required to model it after the actual list.h interface?
We can even reuse a few of the macro helpers.
So this patch takes option 3, but actually implements a
parallel "volatile list" interface in list.h, where it could
potentially be reused by other code. This implements just
enough for tempfile.c's use, though we could easily port
other functions later if need be.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When a tempfile is deactivated, we reset its strbuf to the
empty string, which means we hold onto the memory for later
reuse.
Since we'd like to move to a system where tempfile structs
can actually be freed, deactivating one should drop all
resources it is currently using. And thus "release" rather
than "reset" is the appropriate function to call.
In theory the reset may have saved a malloc() when a
tempfile (or a lockfile) is reused multiple times. But in
practice this happened rarely. Most of our tempfiles are
single-use, since in cases where we might actually use many
(like ref locking) we xcalloc() a fresh one for each ref. In
fact, we leak those locks (to appease the rule that tempfile
storage can never be freed). Which means that using reset is
actively hurting us: instead of leaking just the tempfile
struct, we're leaking the extra heap chunk for the filename,
too.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We may call remove_tempfiles() from an atexit handler, or
from a signal handler. In the latter case we must take care
to avoid functions which may deadlock if the process is in
an unknown state, including looking at any stdio handles
(which may be in the middle of doing I/O and locked) or
calling malloc() or free().
The current implementation calls delete_tempfile(). We unset
the tempfile's stdio handle (if any) to avoid deadlocking
there. But delete_tempfile() still calls unlink_or_warn(),
which can deadlock writing to stderr if the unlink fails.
Since delete_tempfile() isn't very long, let's just
open-code our own simple conservative version of the same
thing. Notably:
1. The "skip_fclose" flag is now called "in_signal_handler",
because it should inform more decisions than just the
fclose handling.
2. We can replace close_tempfile() with just close(fd).
That skips the fclose() question altogether. This is
fine for the atexit() case, too; there's no point
flushing data to a file which we're about to delete
anyway.
3. We can choose between unlink/unlink_or_warn based on
whether it's safe to use stderr.
4. We can replace the deactivate_tempfile() call with a
simple setting of the active flag. There's no need to
do any further cleanup since we know the program is
exiting. And even though the current deactivation code
is safe in a signal handler, this frees us up in future
patches to make non-signal deactivation more
complicated (e.g., by freeing resources).
5. There's no need to remove items from the tempfile_list.
The "active" flag is the ultimate answer to whether an
entry has been handled or not. Manipulating the list
just introduces more chance of recursive signals
stomping on each other, and the whole list will go away
when the program exits anyway. Less is more.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When we deactivate a tempfile, we also have to clean up the
"filename" strbuf. Let's pull this out into its own function
to keep the logic in one place (which will become more
important when a future patch makes it more complicated).
Note that we can use the same function when deactivating an
object that _isn't_ actually active yet (like when we hit an
error creating a tempfile). These callsites don't currently
reset the "active" flag to 0, but it's OK to do so (it's
just a noop for these cases).
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
There are a few steps required to "activate" a tempfile
struct. Let's pull these out into a function. That saves a
few repeated lines now, but more importantly will make it
easier to change the activation scheme later.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Compared to die(), using BUG() triggers abort(). That may
give us an actual coredump, which should make it easier to
get a stack trace. And since the programming error for these
assertions is not in the functions themselves but in their
callers, such a stack trace is needed to actually find the
source of the bug.
In addition, abort() raises SIGABRT, which is more likely to
be caught by our test suite.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The tempfile functions all take pointers to tempfile
objects, but do not check whether the argument is NULL.
This isn't a big deal in practice, since the lifetime of any
tempfile object is defined to last for the whole program. So
even if we try to call delete_tempfile() on an
already-deleted tempfile, our "active" check will tell us
that it's a noop.
In preparation for transitioning to a new system that
loosens the "tempfile objects can never be freed" rule,
let's tighten up our active checks:
1. A NULL pointer is now defined as "inactive" (so it will
BUG for most functions, but works as a silent noop for
things like delete_tempfile).
2. Functions should always do the "active" check before
looking at any of the struct fields.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The tempfile code keeps an "active" flag, and we have a
number of assertions to make sure that the objects are being
used in the right order. Most of these directly check
"active" rather than using the is_tempfile_active()
accessor.
Let's prefer using the accessor, in preparation for it
growing more complicated logic (like checking for NULL).
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When close_tempfile() fails, we delete the tempfile and
reset the fields of the tempfile struct. This makes it
easier for callers to return without cleaning up, but it
also makes this common pattern:
if (close_tempfile(tempfile))
return error_errno("error closing %s", tempfile->filename.buf);
wrong, because the "filename" field has been reset after the
failed close. And it's not easy to fix, as in many cases we
don't have another copy of the filename (e.g., if it was
created via one of the mks_tempfile functions, and we just
have the original template string).
Let's drop the feature that a failed close automatically
deletes the file. This puts the burden on the caller to do
the deletion themselves, but this isn't that big a deal.
Callers which do:
if (write(...) || close_tempfile(...)) {
delete_tempfile(...);
return -1;
}
already had to call delete when the write() failed, and so
aren't affected. Likewise, any caller which just calls die()
in the error path is OK; we'll delete the tempfile during
the atexit handler.
Because this patch changes the semantics of close_tempfile()
without changing its signature, all callers need to be
manually checked and converted to the new scheme. This patch
covers all in-tree callers, but there may be others for
not-yet-merged topics. To catch these, we rename the
function to close_tempfile_gently(), which will attract
compile-time attention to new callers. (Technically the
original could be considered "gentle" already in that it
didn't die() on errors, but this one is even more so).
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
In close_tempfile(), we return an error if ferror()
indicated a previous failure, or if fclose() failed. In the
latter case, errno is set and it is useful for callers to
report it.
However, if _only_ ferror() triggers, then the value of
errno is based on whatever syscall happened to last fail,
which may not be related to our filehandle at all. A caller
cannot tell the difference between the two cases, and may
use "die_errno()" or similar to report a nonsense errno value.
One solution would be to actually pass back separate return
values for the two cases, so a caller can write a more
appropriate message for each case. But that makes the
interface clunky.
Instead, let's just set errno to the generic EIO in this case.
That's not as descriptive as we'd like, but at least it's
predictable. So it's better than the status quo in all cases
but one: when the last syscall really did involve a failure
on our filehandle, we'll be wiping that out. But that's a
fragile thing for us to rely on.
In any case, we'll let the errno result from fclose() take
precedence over our value, as we know that's recent and
accurate (and many I/O errors will persist through the
fclose anyway).
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The current code wants to record an error condition from
either ferror() or fclose(), but makes sure that we always
call both functions. So it can't use logical-OR "||", which
would short-circuit when ferror() is true. Instead, it uses
bitwise-OR "|" to evaluate both functions and set one or
more bits in the "err" flag if they reported a failure.
Unlike logical-OR, though, bitwise-OR does not introduce a
sequence point, and the order of evaluation for its operands
is unspecified. So a compiler would be free to generate code
which calls fclose() first, and then ferror() on the
now-freed filehandle.
There's no indication that this has happened in practice,
but let's write it out in a way that follows the standard.
Noticed-by: Andreas Schwab <schwab@linux-m68k.org>
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When the index is locked and child processes inherit the handle to
said lock and the parent process wants to remove the lock before the
child process exits, on Windows there is a problem: it won't work
because files cannot be deleted if a process holds a handle on them.
The symptom:
Rename from 'xxx/.git/index.lock' to 'xxx/.git/index' failed.
Should I try again? (y/n)
Spawning child processes with bInheritHandles==FALSE would not work
because no file handles would be inherited, not even the hStdXxx
handles in STARTUPINFO (stdin/stdout/stderr).
Opening every file with O_NOINHERIT does not work, either, as e.g.
git-upload-pack expects inherited file handles.
This leaves us with the only way out: creating temp files with the
O_NOINHERIT flag. This flag is Windows-specific, however. For our
purposes, it is equivalent to O_CLOEXEC (which does not exist on
Windows), so let's just open temporary files with the O_CLOEXEC flag and
map that flag to O_NOINHERIT on Windows.
As Eric Wong pointed out, we need to be careful to handle the case where
the Linux headers used to compile Git support O_CLOEXEC but the Linux
kernel used to run Git does not: it returns an EINVAL.
This fixes the test that we just introduced to demonstrate the problem.
Signed-off-by: Ben Wijen <ben@wijen.net>
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Allow an existing file to be registered with the tempfile-handling
infrastructure; in particular, arrange for it to be deleted on program
exit. This can be used if the temporary file has to be created in a
more complicated way than just open(). For example:
* If the file itself needs to be created via the lockfile API
* If it is not a regular file (e.g., a socket)
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Add several functions for creating temporary files with
automatically-generated names, analogous to mkstemps(), but also
arranging for the files to be deleted on program exit.
The functions are named according to a pattern depending how they
operate. They will be used to replace many places in the code where
temporary files are created and cleaned up ad-hoc.
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This makes the next step easier.
The old code used to use "path" to set the initial length of
tempfile->filename. This was not helpful because path was usually
relative whereas the value stored to filename will be absolute. So
just initialize the length to 0.
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
A lot of work went into defining the state diagram for lockfiles and
ensuring correct, race-resistant cleanup in all circumstances.
Most of that infrastructure can be applied directly to *any* temporary
file. So extract a new "tempfile" module from the "lockfile" module.
Reimplement lockfile on top of tempfile.
Subsequent commits will add more users of the new module.
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Jeff King
Ben Wijen
Michael Haggerty