[GHC] #15153: GHC uses O_NONBLOCK on regular files, which has no effect, and blocks the runtime

[GHC]

#15153: GHC uses O_NONBLOCK on regular files, which has no effect, and blocks the
runtime
-------------------------------------+-------------------------------------
        Reporter:  nh2               |                Owner:  (none)
            Type:  bug               |               Status:  new
        Priority:  normal            |            Milestone:  8.6.1
       Component:  Runtime System    |              Version:  8.2.2
      Resolution:                    |             Keywords:
Operating System:  Linux             |         Architecture:
 Type of failure:  Runtime           |  Unknown/Multiple
  performance bug                    |            Test Case:
      Blocked By:                    |             Blocking:
 Related Tickets:                    |  Differential Rev(s):
       Wiki Page:                    |
-------------------------------------+-------------------------------------

Comment (by nh2):

 Replying to [comment:3 simonmar]:
 > It's not usually a big problem except when the filesystem actually has
 significant latency, like with NFS or an arbitrary fuse filesystem.

 What about my example though?

 I'm just reading a 3GB file using `BS.readFile` and it blocks other
 threads for 30 seconds.

 (My file system doens't have unexpected latency in terms of access time,
 it's a normal spinning disk, and the problem is the time it takes to read
 the data. Though I find it reasonablye to argue that the ~8ms it may take
 to seek this disk is ''also'' a lot of wasted CPU time, that's not the key
 problem in my example.)

 > I think I was worried about the overhead of using safe calls here, the
 problem is that if you have lots of threads all reading from Handles (e.g.
 sockets) then if those reads are safe calls we'll get a large number of
 real OS threads created.

 The concern about overhead makes sense to me, but I find it easy to argue
 that there's more overheard in having all other resources of a machine
 (CPU, network, other disks) idle while a single disk arm is moving. And
 the time for a `pthread_create` is rather small compared to a read of any
 physical storage device.

 Another argument is that if examples like the one I gave don't work, it is
 very hard to build reasonable heartbeating of long-running IO tasks in
 Haskell.

 > O_NONBLOCK doesn't do anything for local disk files, but it also doesn't
 do any harm

 What I meant is that from a quick look at the code, it looked like if
 O_NONBLOCK isn't set, a `safe` call would be made instead.

 In the spirit of https://ghc.haskell.org/trac/ghc/ticket/13296#comment:3
 and the followup comment, would it be a good default to have reads to be
 `safe`, and offer `unsafe` equivalents when you know that the read will be
 very fast?

-- 
Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/15153#comment:4>
GHC <http://www.haskell.org/ghc/>
The Glasgow Haskell Compiler