OFD locking broken on 32-bit Linux?

[bgamari]

I suspect that OFD locking is broken on 32-bit Linux in both lukko and base.

Specifically, we end up calling fcntl64 yet use the 32-bit struct flock. glibc suffered a similar bug some years ago. The following patch works around the issue although I'm still investigating the portability implications:

diff --git a/src-ofd/Lukko/OFD.hsc b/src-ofd/Lukko/OFD.hsc
index 8d04dae..024cb28 100644
--- a/src-ofd/Lukko/OFD.hsc
+++ b/src-ofd/Lukko/OFD.hsc
@@ -127,21 +127,21 @@ data FLock  = FLock { l_type   :: CShort
                     }
 
 instance Storable FLock where
-    sizeOf _ = #{size struct flock}
+    sizeOf _ = #{size struct flock64}
     alignment _ = #{alignmentcompat struct flock}
     poke ptr x = do
         fillBytes ptr 0 (sizeOf x)
-        #{poke struct flock, l_type}   ptr (l_type x)
-        #{poke struct flock, l_whence} ptr (l_whence x)
-        #{poke struct flock, l_start}  ptr (l_start x)
-        #{poke struct flock, l_len}    ptr (l_len x)
-        #{poke struct flock, l_pid}    ptr (l_pid x)
+        #{poke struct flock64, l_type}   ptr (l_type x)
+        #{poke struct flock64, l_whence} ptr (l_whence x)
+        #{poke struct flock64, l_start}  ptr (l_start x)
+        #{poke struct flock64, l_len}    ptr (l_len x)
+        #{poke struct flock64, l_pid}    ptr (l_pid x)
     peek ptr = do
-        x1 <- #{peek struct flock, l_type}   ptr
-        x2 <- #{peek struct flock, l_whence} ptr
-        x3 <- #{peek struct flock, l_start}  ptr
-        x4 <- #{peek struct flock, l_len}    ptr
-        x5 <- #{peek struct flock, l_pid}    ptr
+        x1 <- #{peek struct flock64, l_type}   ptr
+        x2 <- #{peek struct flock64, l_whence} ptr
+        x3 <- #{peek struct flock64, l_start}  ptr
+        x4 <- #{peek struct flock64, l_len}    ptr
+        x5 <- #{peek struct flock64, l_pid}    ptr
         return (FLock x1 x2 x3 x4 x5)
 
 lockImpl :: Maybe Handle -> FD -> String -> LockMode -> Bool -> IO Bool

[phadej]

Is there a way to test this, e.g. some Docker image? Or is it unreliable, if you don't really run 32bit kernel?

[phadej]

FYI, GHC also has the same in libraries/base/GHC/IO/Handle/Lock/LinuxOFD.hsc

instance Storable FLock where
    sizeOf _ = #{size struct flock}
    alignment _ = #{alignment struct flock}
    poke ptr x = do
        fillBytes ptr 0 (sizeOf x)
        #{poke struct flock, l_type}   ptr (l_type x)
        #{poke struct flock, l_whence} ptr (l_whence x)
        #{poke struct flock, l_start}  ptr (l_start x)
        #{poke struct flock, l_len}    ptr (l_len x)
        #{poke struct flock, l_pid}    ptr (l_pid x)

EDIT: I'm tempted to wait until GHC has the patch (or shows it is not required). I think it still has way to run by not using cabal-install which uses lukko. ATM there aren't any released cabal-install which uses it. (Except versions in HVR PPA).

[bgamari]

Actually, there are no portability implications by definition; OFD locking is Linux specific and fcntl64 has been long supported.

[bgamari]

FYI, GHC also has the same in libraries/base/GHC/IO/Handle/Lock/LinuxOFD.hsc

Yes, I know. Hence the mention of base in the ticket description.

Is there a way to test this, e.g. some Docker image? Or is it unreliable, if you don't really run 32bit kernel?

Yes, see the GHC ticket.

[phadej]

What that image is based on. It's manifest says file:594fa35cf803361e69d817fc867b6a4069c064ffd20ed50caf42ad9bb11ca999 in /

I assume it's some Ubuntu variant, but I'm not running it on my machine without knowing for sure.

[phadej]

The glibc bug seems to be fixed in 2.28 but on Ubuntu-18.04 it is

% dpkg -l |grep libc-bin
ii  libc-bin                                   2.27-3ubuntu1 ...

---

Can someone try on Ubuntu Eoan, it uses 2.30 https://packages.ubuntu.com/eoan/libc-bin, if it's so, then we can just use the glibc patch indeed.

---

To clarify: the fix for older glibc's shouldn't break newer ones.

Look at this diff: https://sourceware.org/git/?p=glibc.git;a=blobdiff;f=sysdeps/unix/sysv/linux/fcntl.c;h=cbab7b401a3875babb545681ba9128d850883f67;hp=e3992dc9d4d1e20ef8aa97e81ec1bd13ac170900;hb=06ab719d30b01da401150068054d3b8ea93dd12f;hpb=124e025864bb39732c71fc60c1443d5680881a0a

If we change Storage FLock as proposed, AFAICS we will break stuff too.

[phadej]

As the workaround for now: compile cabal with lukko -ofd-locking on 32-bit linuxes with old glibc.

[bgamari]

My thought here is that we should probably just use fcntl64 and struct flock64 explicitly.

[newhoggy]

If you're interested, I have i386 build support for CircleCI projects in our haskell-build CircleCI orb.

• Orb documentation: https://circleci.com/orbs/registry/orb/haskell-works/haskell-build
• Example project: https://github.com/haskell-works/cabal-cache
• Example build: https://circleci.com/gh/haskell-works/cabal-cache/1293

[leftaroundabout]

Problem persists on Debian 11, libc-bin version 2.13-13+deb11u2 i386.

Still needs -ofd-locking to avoid the fdLock invalid argument issue.

If no proper fix to this is on the horizon, how about default-setting -ofd-locking on i386? What are the drawbacks?