Fix segfault for ghc 9.2

ondrap commented 1 year ago

I tried to upgrade my compiler to ghc 9.2.5 and the who thing just stopped working. Ultimately, I found that the freeDbcIfNotAlready is being started practically at the first GC cycle. It seems the compiler was able to optimize the structure out by inlining the functions so that in the end there was nothing pointing to the DbcWrapper.

Marking one of the function that use the wrapperNOINLINE seems to help; though, it seems to me some other solution might be well preferable (in the end, using some kind of with function usually works better).

ak-coram commented 1 year ago

@ondrap This issue seems to be even more prevalent on newer GHC versions (e.g. 9.4.5) and statements seem to be affected as well. Additionally the NOINLINE workaround doesn't seem to work anymore.

flhorizon commented 9 months ago

Hello, I had a similar issue when upgrading from GHC 8.6 to 9.4.8.

I believed I've finally fixed the issue (had "Tried to use a disposed ODBC Statement" and sometimes the same about the connection handle) and finally had a break-through.

I've taken inspiration from this post recommending to attach the finalize to an embedded IORef:

https://discourse.haskell.org/t/using-the-garbage-collector-to-free-user-resources/5532/5

And reproduced the reaperRef pattern from:

https://hackage.haskell.org/package/resource-pool-0.4.0.0/docs/src/Data.Pool.Internal.html#newPool

It's only using the reaperRef pattern that I had a definite improvement. My previous tries were:

compiling with -fno-cse
adding StrictData extension
no-inline 'ing all IO functions in .hsc files
and settings __attribute__((noline)) in the cbits functions
masking async exceptions on alloc/free foreign calls and modifyMVar expressions

Though the stability improvement from my changes may be due to a mix of all this, I have yet to test whether any change besides the reaper ref ones are actually crucial

ak-coram commented 9 months ago

@flhorizon: I've replaced all calls to addFinalizer with mkWeakMVar on the relevant MVars. It seems to work well with several GHC versions.

flhorizon commented 9 months ago

@ak-coram Nice to know !

hdbc / hdbc-odbc

Fix segfault for ghc 9.2 #48