[LLVMdev] [RFC] Parsing runtime flags in sanitizers (ASan/LSan/UBSan)

[Alexey Samsonov]

Hi all,

TL;DR
1) We should change the way we parse common runtime flags in sanitizers.
2) We should make ASan aware of the tools it can be combined with (LSan and
UBSan).
3) We may have to restrict the tools UBSan can be combined with (currently
to ASan) (see [1])

Currently we have two kinds of sanitizer runtime flags: tool-specific flags
and "common flags", defined in sanitizer_common library and shared across
all the sanitizers. Many of these common flags are used early during tool
initialization (for example, we use "suppressions" flag to create
SuppressionContext singleton object). That's the reason why we parse both
tool-specific flags and common flags as early as possible at program
startup, before running the rest of initialization code. It all works fine
until we have a single sanitizer - e.g. for TSan or MSan.

The situation gets crazy when we combine multiple sanitizers in a single
process, for instance use ASan+LSan+UBSan (the default use case in some
setups). Each tool has its own defaults for common flag values, and each of
ASAN_OPTIONS, LSAN_OPTIONS and UBSAN_OPTIONS can define both tool-specific
and common flags. These environment variables are parsed at different time,
sometimes in undefined order. We can easily end up in situation where ASan
initializes some parts of sanitizer_common assuming certain values of
common runtime flags, but then these flags are overwritten by LSAN_OPTIONS.
All this is very complicated and fragile.

I propose to implement the following:

*ASan flag parsing*:
1) Detect all sanitizers ASan is combined with (We know that we have LSan
if "CAN_SANITIZE_LEAKS" is true. We can detect if we have UBSan in the
process somehow.)
2) Setup defaults for ASan-specific flags and override them from
ASAN_OPTIONS.
3) Setup defaults for common flags and override them from *all of
*ASAN_OPTIONS,
LSAN_OPTIONS and UBSAN_OPTIONS (if corresponding sanitizers are enabled for
the process).
4) Proceed with initialization, and call LSan/UBSan initializers when
appropriate.

*LSan/UBSan flag parsing:*
1) Learn if LSan/UBSan are combined with ASan (the "main" tool) or run as
standalone tool. This is already done for LSan and can be done for UBSan by
slightly modifying the way we structure runtimes.
2) Setup defaults for tool-specific flags and override them from
LSAN/UBSAN_OPTIONS.
3) If the tools run in standalone mode, setup the defaults for common flags
and override them from LSAN/UBSAN_OPTIONS. If the tools are combined with
"main" tool, do nothing.
4) Proceed with initialization.

*[1] Conjecture*:
1) We will have to add the hook to initialize UBSan that we would invoke
from ASan (much like we do for LSan). It means that we would have to
restrict the set of sanitizers that can be combined with UBSan (allowing it
for another sanitizers would be very easy, but manual process). The
alternative would be to make UBSan setup no-op except for parsing the
flags, so that UBSan initializer could be safely called before main tool
initializer.

Bonus:
1) We will be able to significantly improve runtime flag parsing
diagnostic. For instance, we would be able to report conflicting flag
definitions (e.g. when one provides ASAN_OPTIONS=symbolize=0
LSAN_OPTIONS=symbolize=1).

Comments/objections?

-- 
Alexey Samsonov
vonosmas at gmail.com
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20141205/1b4e98f1/attachment.html>