| Message ID | 20220803140441.1196056-1-JPEWhacker@gmail.com |
|---|---|
| State | New |
| Headers | show |
| Series | [bitbake-devel] siggen: Fix insufficent entropy in sigtask file names | expand |
Hi Joshua, On Wed, 3 Aug 2022 09:04:41 -0500 "Joshua Watt" <JPEWhacker@gmail.com> wrote: > Signature generation uses mkstemp() to get a file descriptor to a unique > file and then write the signature into it. However, the unique file name > generation in glibc is based on the system timestamp, which means that > with highly parallel builds it is more likely than one might expect > expected that a conflict will occur between two different builder nodes. s/expect expected/expect/ Taken in my testing branch with the above fix.
On Thu, Aug 4, 2022 at 4:26 AM Luca Ceresoli <luca.ceresoli@bootlin.com> wrote: > > Hi Joshua, > > On Wed, 3 Aug 2022 09:04:41 -0500 > "Joshua Watt" <JPEWhacker@gmail.com> wrote: > > > Signature generation uses mkstemp() to get a file descriptor to a unique > > file and then write the signature into it. However, the unique file name > > generation in glibc is based on the system timestamp, which means that > > with highly parallel builds it is more likely than one might expect > > expected that a conflict will occur between two different builder nodes. > > s/expect expected/expect/ > > Taken in my testing branch with the above fix. Steve, can you please backport this to kirkstone and dunfell? There may be a context failure backporting it to dunfell, let me know if you want me to send a new patch with the correct context. Thanks > > -- > Luca Ceresoli, Bootlin > Embedded Linux and Kernel engineering > https://bootlin.com
On Tue, Aug 9, 2022 at 11:44 AM Joshua Watt <jpewhacker@gmail.com> wrote: > > On Thu, Aug 4, 2022 at 4:26 AM Luca Ceresoli <luca.ceresoli@bootlin.com> wrote: > > > > Hi Joshua, > > > > On Wed, 3 Aug 2022 09:04:41 -0500 > > "Joshua Watt" <JPEWhacker@gmail.com> wrote: > > > > > Signature generation uses mkstemp() to get a file descriptor to a unique > > > file and then write the signature into it. However, the unique file name > > > generation in glibc is based on the system timestamp, which means that > > > with highly parallel builds it is more likely than one might expect > > > expected that a conflict will occur between two different builder nodes. > > > > s/expect expected/expect/ > > > > Taken in my testing branch with the above fix. > > Steve, can you please backport this to kirkstone and dunfell? There > may be a context failure backporting it to dunfell, let me know if you > want me to send a new patch with the correct context. Yes! I'll probably get to that tomorrow after I finish up testing for the 4.0.3 release. I'll let you know if I run into any issues I'm not confident about resolving. Steve
On Tue, Aug 9, 2022 at 4:48 PM Steve Sakoman <steve@sakoman.com> wrote: > > On Tue, Aug 9, 2022 at 11:44 AM Joshua Watt <jpewhacker@gmail.com> wrote: > > > > On Thu, Aug 4, 2022 at 4:26 AM Luca Ceresoli <luca.ceresoli@bootlin.com> wrote: > > > > > > Hi Joshua, > > > > > > On Wed, 3 Aug 2022 09:04:41 -0500 > > > "Joshua Watt" <JPEWhacker@gmail.com> wrote: > > > > > > > Signature generation uses mkstemp() to get a file descriptor to a unique > > > > file and then write the signature into it. However, the unique file name > > > > generation in glibc is based on the system timestamp, which means that > > > > with highly parallel builds it is more likely than one might expect > > > > expected that a conflict will occur between two different builder nodes. > > > > > > s/expect expected/expect/ > > > > > > Taken in my testing branch with the above fix. > > > > Steve, can you please backport this to kirkstone and dunfell? There > > may be a context failure backporting it to dunfell, let me know if you > > want me to send a new patch with the correct context. > > Yes! I'll probably get to that tomorrow after I finish up testing for > the 4.0.3 release. I'll let you know if I run into any issues I'm not > confident about resolving. That's fine. I also just realized this relies on Python 3.6 functionality, so hold off on backporting it to dunfell for the time being; Hardknott should be OK still > > Steve
diff --git a/bitbake/lib/bb/siggen.py b/bitbake/lib/bb/siggen.py index 3f3d6df54d..bb80343a97 100644 --- a/bitbake/lib/bb/siggen.py +++ b/bitbake/lib/bb/siggen.py @@ -425,7 +425,7 @@ class SignatureGeneratorBasic(SignatureGenerator): bb.error("Taskhash mismatch %s versus %s for %s" % (computed_taskhash, self.taskhash[tid], tid)) sigfile = sigfile.replace(self.taskhash[tid], computed_taskhash) - fd, tmpfile = tempfile.mkstemp(dir=os.path.dirname(sigfile), prefix="sigtask.") + fd, tmpfile = bb.utils.mkstemp(dir=os.path.dirname(sigfile), prefix="sigtask.") try: with bb.compress.zstd.open(fd, "wt", encoding="utf-8", num_threads=1) as f: json.dump(data, f, sort_keys=True, separators=(",", ":"), cls=SetEncoder) diff --git a/bitbake/lib/bb/utils.py b/bitbake/lib/bb/utils.py index 19ed68ea62..b8b90df8d3 100644 --- a/bitbake/lib/bb/utils.py +++ b/bitbake/lib/bb/utils.py @@ -28,6 +28,8 @@ import signal import collections import copy import ctypes +import random +import tempfile from subprocess import getstatusoutput from contextlib import contextmanager from ctypes import cdll @@ -1754,3 +1756,22 @@ def is_local_uid(uid=''): if str(uid) == line_split[2]: return True return False + +def mkstemp(suffix=None, prefix=None, dir=None, text=False): + """ + Generates a unique filename, independent of time. + + mkstemp() in glibc (at least) generates unique file names based on the + current system time. When combined with highly parallel builds, and + operating over NFS (e.g. shared sstate/downloads) this can result in + conflicts and race conditions. + + This function adds additional entropy to the file name so that a collision + is independent of time and thus extremely unlikely. + """ + entropy = "".join(random.choices("abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ1234567890", k=20)) + if prefix: + prefix = prefix + entropy + else: + prefix = tempfile.gettempprefix() + entropy + return tempfile.mkstemp(suffix=suffix, prefix=prefix, dir=dir, text=text)
Signature generation uses mkstemp() to get a file descriptor to a unique file and then write the signature into it. However, the unique file name generation in glibc is based on the system timestamp, which means that with highly parallel builds it is more likely than one might expect expected that a conflict will occur between two different builder nodes. When operating over NFS (such as a shared sstate cache), this can cause race conditions and rare failures (particularly with NFS servers that may not correctly implement O_EXCL). The signature generation code is particularly susceptible to races since a single "sigtask." prefix used for all signatures from all tasks, which makes collision even more likely. To work around this, add an internal implementation of mkstemp() that adds additional truly random entropy to the file name to eliminate conflicts. Signed-off-by: Joshua Watt <JPEWhacker@gmail.com> --- bitbake/lib/bb/siggen.py | 2 +- bitbake/lib/bb/utils.py | 21 +++++++++++++++++++++ 2 files changed, 22 insertions(+), 1 deletion(-)