Jeff King [off-list ref] writes:
On Tue, Feb 23, 2021 at 11:43:16PM -0500, Jeff King wrote:
quoted
The workaround here is enabled all the time, without a Makefile knob,
since it's a complete noop if open() never returns EINTR. I did push it
into its own compat/ source file, though, since it has to #undef our
macro redirection. Putting it in a file with other code risks confusion
if more code is added after that #undef.
Whoops, sorry, I had a last-minute change of heart here and just stuck
it in wrapper.c with a bit of preprocessor magic to guard it. It felt
weird having compat/open.c, when the rest of compat/ is always
conditional on Makefile knobs. But I'm happy to go the other way if
anybody feels strongly.
Hmph, just like other workarounds, shouldn't this be "if your
platform screws this up, define this knob to work it around"? That
would make it fit better with the rest of compat/.
I dunno. A no-op wrapper makes the code less transparent, which I
am somewhat unhappy. But a wrapper that is always used even on
platforms that do not need might give us a better chance of noticing
a bug in the wrapper itself, making it less likely for us to leave
only the users of minority broken platforms behind. So...
On Tue, Feb 23, 2021 at 09:36:25PM -0800, Junio C Hamano wrote:
Jeff King [off-list ref] writes:
quoted
On Tue, Feb 23, 2021 at 11:43:16PM -0500, Jeff King wrote:
quoted
The workaround here is enabled all the time, without a Makefile knob,
since it's a complete noop if open() never returns EINTR. I did push it
into its own compat/ source file, though, since it has to #undef our
macro redirection. Putting it in a file with other code risks confusion
if more code is added after that #undef.
Whoops, sorry, I had a last-minute change of heart here and just stuck
it in wrapper.c with a bit of preprocessor magic to guard it. It felt
weird having compat/open.c, when the rest of compat/ is always
conditional on Makefile knobs. But I'm happy to go the other way if
anybody feels strongly.
Hmph, just like other workarounds, shouldn't this be "if your
platform screws this up, define this knob to work it around"? That
would make it fit better with the rest of compat/.
I actually started that way, but then I got stuck deciding which
platforms in config.mak.uname to enable it for. We've seen it on Linux
and on macOS. I don't know how prevalent it is on those platforms (or
whether it is indeed a bug that may even get fixed). Since the check is
basically zero-cost at run-time, it seemed like it was worth just being
on guard for it so that nobody ever had to worry about it again.
But I don't feel that strongly. I'd be happy to do it with a Makefile
knob if you prefer.
I dunno. A no-op wrapper makes the code less transparent, which I
am somewhat unhappy.
Yeah. I think the run-time cost is essentially zero, but it does add a
bit of complexity if you are debugging. And handling the magic mode
vararg definitely makes it uglier.
The other thing that gives me pause is that there may be other syscalls
which need the same treatment. Here's a recent discussion that Big Sur
may also return EINTR from the opendir() family:
https://github.com/dotnet/runtime/issues/47584
Git is presumably susceptible to that, too; it's just that we call
open() a lot more frequently, so that was the obvious one that came up.
I don't think there is a more general way to address this, though. We'd
have to wrap each suspected syscall (but we could presumably at least
tie them all to a single Makefile knob). I'm not excited at the prospect
of preemptively adding such wrappers when we aren't even 100% sure that
there is a problem.
But a wrapper that is always used even on
platforms that do not need might give us a better chance of noticing
a bug in the wrapper itself, making it less likely for us to leave
only the users of minority broken platforms behind. So...
That part I'm less worried about. It sounds like we'd want it on for
recent versions of macOS, which would give us pretty wide coverage to
notice a bug.
-Peff
On Wed, Feb 24, 2021 at 01:20:44PM -0500, Jeff King wrote:
quoted
Hmph, just like other workarounds, shouldn't this be "if your
platform screws this up, define this knob to work it around"? That
would make it fit better with the rest of compat/.
I actually started that way, but then I got stuck deciding which
platforms in config.mak.uname to enable it for. We've seen it on Linux
and on macOS. I don't know how prevalent it is on those platforms (or
whether it is indeed a bug that may even get fixed). Since the check is
basically zero-cost at run-time, it seemed like it was worth just being
on guard for it so that nobody ever had to worry about it again.
But I don't feel that strongly. I'd be happy to do it with a Makefile
knob if you prefer.
So here's a v2 that turns it into a Makefile knob.
This was motivated by two things:
- it must not be enabled on Windows, where it overrides the existing
mingw_open()
- I couldn't replicate it on a Big Sur machine, using the same
sequence that Aleksey did (basically just cloning and checking out
dotnet/runtime.git from scratch).
So it's not entirely clear to me how widespread the problem is. A
conservative approach (at least in terms of change from the status
quo) is to introduce this knob but not enable it automatically
(which this patch does). And then if somebody runs into it, they can
try turning the knob.
It's less conservative in the sense that we won't be preemptively
protecting people. I dunno. Once we have the knob, it would be
pretty easy to just flip the default.
-- >8 --
Subject: [PATCH] Makefile: add OPEN_RETURNS_EINTR knob
On some platforms, open() reportedly returns EINTR when opening regular
files and we receive a signal (usually SIGALRM from our progress meter).
This shouldn't happen, as open() should be a restartable syscall, and we
specify SA_RESTART when setting up the alarm handler. So it may actually
be a kernel or libc bug for this to happen. But it has been reported on
at least one version of Linux (on a network filesystem):
https://lore.kernel.org/git/c8061cce-71e4-17bd-a56a-a5fed93804da@neanderfunk.de/
as well as on macOS starting with Big Sur even on a regular filesystem.
We can work around it by retrying open() calls that get EINTR, just as
we do for read(), etc. Since we don't ever _want_ to interrupt an open()
call, we can get away with just redefining open, rather than insisting
all callsites use xopen().
We actually do have an xopen() wrapper already (and it even does this
retry, though there's no indication of it being an observed problem back
then; it seems simply to have been lifted from xread(), etc). But it is
used hardly anywhere, and isn't suitable for general use because it will
die() on error. In theory we could combine the two, but it's awkward to
do so because of the variable-args interface of open().
This patch adds a Makefile knob for enabling the workaround. It's not
enabled by default for any platforms in config.mak.uname yet, as we
don't have enough data to decide how common this is (I have not been
able to reproduce on either Linux or Big Sur myself). It may be worth
enabling preemptively anyway, since the cost is pretty low (if we don't
see an EINTR, it's just an extra conditional).
However, note that we must not enable this on Windows. It doesn't do
anything there, and the macro overrides the existing mingw_open()
redirection. I've added a preemptive #undef here in the mingw header
(which is processed first) to just quietly disable it (we could also
make it an #error, but there is little point in being so aggressive).
Reported-by: Aleksey Kliger <redacted>
Signed-off-by: Jeff King <redacted>
---
Makefile | 7 +++++++
compat/mingw.h | 1 +
compat/open.c | 25 +++++++++++++++++++++++++
git-compat-util.h | 6 ++++++
4 files changed, 39 insertions(+)
create mode 100644 compat/open.c
diff --git a/Makefile b/Makefile
index 9b1bde2e0e..c4872ca16e 100644
--- a/Makefile
+++ b/Makefile
@@ -22,6 +22,9 @@ all::
# when attempting to read from an fopen'ed directory (or even to fopen
# it at all).
#
+# Define OPEN_RETURNS_EINTR if your open() system call may return EINTR
+# when a signal is received (as opposed to restarting).
+#
# Define NO_OPENSSL environment variable if you do not have OpenSSL.
#
# Define USE_LIBPCRE if you have and want to use libpcre. Various
@@ -1538,6 +1541,10 @@ ifdef FREAD_READS_DIRECTORIES
COMPAT_CFLAGS += -DFREAD_READS_DIRECTORIES
COMPAT_OBJS += compat/fopen.o
endif
+ifdef OPEN_RETURNS_EINTR
+ COMPAT_CFLAGS += -DOPEN_RETURNS_EINTR
+ COMPAT_OBJS += compat/open.o
+endif
ifdef NO_SYMLINK_HEAD
BASIC_CFLAGS += -DNO_SYMLINK_HEAD
endif
diff --git a/compat/mingw.h b/compat/mingw.h
index af8eddd73e..c9a52ad64a 100644
--- a/compat/mingw.h
+++ b/compat/mingw.h
@@ -227,6 +227,7 @@ int mingw_rmdir(const char *path);
int mingw_open (const char *filename, int oflags, ...);
#define open mingw_open
+#undef OPEN_RETURNS_EINTR
int mingw_fgetc(FILE *stream);
#define fgetc mingw_fgetc
diff --git a/compat/open.c b/compat/open.c
new file mode 100644
index 0000000000..eb3754a23b
--- /dev/null
+++ b/compat/open.c
@@ -0,0 +1,25 @@
+#include "git-compat-util.h"
+
+#undef open
+int git_open_with_retry(const char *path, int flags, ...)
+{
+ mode_t mode = 0;
+ int ret;
+
+ /*
+ * Also O_TMPFILE would take a mode, but it isn't defined everywhere.
+ * And anyway, we don't use it in our code base.
+ */
+ if (flags & O_CREAT) {
+ va_list ap;
+ va_start(ap, flags);
+ mode = va_arg(ap, int);
+ va_end(ap);
+ }
+
+ do {
+ ret = open(path, flags, mode);
+ } while (ret < 0 && errno == EINTR);
+
+ return ret;
+}diff --git a/git-compat-util.h b/git-compat-util.h
index 838246289c..551cc9f22f 100644
--- a/git-compat-util.h
+++ b/git-compat-util.h
@@ -788,6 +788,12 @@ int git_vsnprintf(char *str, size_t maxsize,
const char *format, va_list ap);
#endif
+#ifdef OPEN_RETURNS_EINTR
+#undef open
+#define open git_open_with_retry
+int git_open_with_retry(const char *path, int flag, ...);
+#endif
+
#ifdef __GLIBC_PREREQ
#if __GLIBC_PREREQ(2, 1)
#define HAVE_STRCHRNUL
--
2.31.0.rc0.520.g3ffb8d01f4