[BUG] Segmentation fault evaluating hackage/hpack-0.34.3-r0

[tdiesler]

It seems that hackage.nix is generally not usable on ARM64 - it always crashes on a variety HW platforms and OSs

downloading 'https://github.com/input-output-hk/hackage.nix/archive/972cf011638e3bb4509bdb1edd731aab18a31ec8.tar.gz'...
unpacking 'https://github.com/input-output-hk/hackage.nix/archive/972cf011638e3bb4509bdb1edd731aab18a31ec8.tar.gz'...
warning: dumping very large path (> 256 MiB); this may run out of memory
...
building '/nix/store/r59apm6jqkvamkvxj254jgb7255mxz7f-nix-tools-plan-to-nix-pkgs.drv'...
evaluating file '/nix/store/pp0bq2z970pybwidxi5vxc72p3ig81iz-nix-tools-plan-to-nix-pkgs/default.nix'
evaluating file '/nix/store/gb1dxbbdcrjnrw8n480g4asqfwzgml1q-haskellNix-src/patches/ghc865/default.nix'
evaluating file '/nix/store/pp0bq2z970pybwidxi5vxc72p3ig81iz-nix-tools-plan-to-nix-pkgs/.plan.nix/nix-tools.nix'
evaluating file '/nix/store/zz70g6bklfp5lgx629ccihjpnrdq8jwv-hackage-exprs-source/hackage/hpack-0.34.3-r0-ec6d6734b567acc30af5568fea45fbff33885be73fafd59413137c7cf19c3ece.nix'
Segmentation fault

This happens on large AWS (ARM64) machines (32GB RAM) with ...

• CentOS8
• Debian9

It is also reproducible with RaspberryPi 4 (8GB)

The version of hackage.nix that I used is that referenced by cardano-node-1.25.1

Without hackage.nix, I can successfully install ghc + cabal into nix-env and nix-shell

[angerman]

We've opened NixOS/nix#4510 as well.

[tdiesler]

I suppose a proper solution to this and possibly other issues would be to integrate the arm64 build with Hydra, then we could pull down the binaries from caches and actually do a nix based cardano-node build on Raspi4. (i.e. instead of doing this on a large arm64 machine first)

[dhess]

Does NixOS/nix#4264 make this work?

[angerman]

@dhess that ticket claims that setting GC_DONT_GC=1 would also fix the issue, which it does not for me. I guess I could try a patched nix.

[angerman]

I suppose a proper solution to this and possibly other issues would be to integrate the arm64 build with Hydra, then we could pull down the binaries from caches and actually do a nix based cardano-node build on Raspi4. (i.e. instead of doing this on a large arm64 machine first)

I'm afraid this won't help. The it blows up on beefy aarch64 machines even with 64GB or ram. Again for cardano-node, we can perfectly fine cross compile aarch64 binaries to use on raspberry pis. And other low end aarch64 scbs. Though I'm going to caution against this here as memory requirements might go up (for running cardano-node, most scb's today have 4GB, and the RPi4 is limited to 8GB right now).

There is a telegram group, where I frequently post fully static cardano-node builds (e.g. 1.25.1). Those are built using haskell.nix, but on an x86_64 machine.

The only caveat is that a tiny ghc patch is still needed for 8.10.3 to make musl builds possible. Specifically this haskell.nix modification:

diff --git a/overlays/bootstrap.nix b/overlays/bootstrap.nix
index 6b7d030..db46ff5 100644
--- a/overlays/bootstrap.nix
+++ b/overlays/bootstrap.nix
@@ -135,6 +135,7 @@ in {
                 ++ final.lib.optional (versionAtLeast "8.6.4" && versionLessThan "8.8") ./patches/ghc/ghc-no-system-linker.patch
                 ++ fromUntil "8.10.2" "8.10.3" ./patches/ghc/MR3714-backported-to-8.10.2.patch
+                ++ final.lib.optional (version == "8.10.3" && final.targetPlatform.isAarch64) ./patches/ghc/3434.patch
                 ++ from      "8.10.1"          ./patches/ghc/ghc-acrt-iob-func.patch
diff --git a/overlays/patches/ghc/3434.patch b/overlays/patches/ghc/3434.patch
new file mode 100644
index 0000000..3feead9
--- /dev/null
+++ b/overlays/patches/ghc/3434.patch
@@ -0,0 +1,68 @@
+diff --git a/libraries/ghc-prim/ghc-prim.cabal b/libraries/ghc-prim/ghc-prim.cabal
+index c633124..2bd51c0 100644
+--- a/libraries/ghc-prim/ghc-prim.cabal
++++ b/libraries/ghc-prim/ghc-prim.cabal
+@@ -70,7 +70,7 @@ Library
+     if os(linux)
+         -- we need libm, but for musl and other's we might need libc, as libm
+         -- is just an empty shell.
+-        extra-libraries: c, m
++        extra-libraries: gcc, c, m
+
+     c-sources:
+         cbits/atomic.c
+diff --git a/rts/RtsSymbols.c b/rts/RtsSymbols.c
+index 9ca696c..b4a85e5 100644
+--- a/rts/RtsSymbols.c
++++ b/rts/RtsSymbols.c
+@@ -1000,6 +1000,18 @@
+ #define RTS_LIBGCC_SYMBOLS
+ #endif
+
++
++#if !defined(DYNAMIC) && defined(linux_HOST_OS)
++// we need these for static musl builds. However when
++// linking shared objects (DLLs) this will fail, hence
++// we do not include them when building with -DDYNAMIC
++#define RTS_LINKER_SYMBOLS                             \
++      SymI_NeedsProto(__fini_array_start)              \
++      SymI_NeedsProto(__fini_array_end)
++#else
++#define RTS_LINKER_SYMBOLS
++#endif
++
+ /* entirely bogus claims about types of these symbols */
+ #define SymI_NeedsProto(vvv)  extern void vvv(void);
+ #define SymI_NeedsDataProto(vvv)  extern StgWord vvv[];
+@@ -1028,6 +1040,7 @@ RTS_DARWIN_ONLY_SYMBOLS
+ RTS_OPENBSD_ONLY_SYMBOLS
+ RTS_LIBGCC_SYMBOLS
+ RTS_LIBFFI_SYMBOLS
++RTS_LINKER_SYMBOLS
+ #undef SymI_NeedsProto
+ #undef SymI_NeedsDataProto
+ #undef SymI_HasProto
+@@ -1068,6 +1081,8 @@ RTS_LIBFFI_SYMBOLS
+ #define SymI_HasProto_deprecated(vvv)   \
+    { #vvv, (void*)0xBAADF00D, true },
+
++void *RTS_DYNAMIC = NULL;
++
+ RtsSymbolVal rtsSyms[] = {
+       RTS_SYMBOLS
+       RTS_RET_SYMBOLS
+@@ -1078,6 +1093,7 @@ RtsSymbolVal rtsSyms[] = {
+       RTS_OPENBSD_ONLY_SYMBOLS
+       RTS_LIBGCC_SYMBOLS
+       RTS_LIBFFI_SYMBOLS
++      RTS_LINKER_SYMBOLS
+       SymI_HasDataProto(nonmoving_write_barrier_enabled)
+ #if defined(darwin_HOST_OS) && defined(i386_HOST_ARCH)
+       // dyld stub code contains references to this,
+@@ -1085,5 +1101,6 @@ RtsSymbolVal rtsSyms[] = {
+       // lazy pointers as nonlazy.
+       { "dyld_stub_binding_helper", (void*)0xDEADBEEF, false },
+ #endif
++      { "_DYNAMIC", (void*)(&RTS_DYNAMIC), false },
+       { 0, 0, false } /* sentinel */
+ };
\ No newline at end of file

[hamishmack]

@angerman does ghc 8.10.4 need this patch? or is it included?

[angerman]

@hamishmack it's not in 8.10.4, nor in 9.0 yet.