Date: Tue, 7 Sep 2004 16:45:18 +0300
From: "Michael S. Tsirkin" <mst@mellanox.co.il>
To: Andi Kleen <ak@suse.de>
Cc: discuss@x86-64.org, linux-kernel@vger.kernel.org
Subject: Re: [discuss] f_ops flag to speed up compatible ioctls in linux kernel
Message-ID: <20040907134517.GA1016@mellanox.co.il>
Reply-To: "Michael S. Tsirkin" <mst@mellanox.co.il>
References: <20040901072245.GF13749@mellanox.co.il> <20040903080058.GB2402@wotan.suse.de> <20040907104017.GB10096@mellanox.co.il> <20040907121418.GC25051@wotan.suse.de>
Mime-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20040907121418.GC25051@wotan.suse.de>
User-Agent: Mutt/1.4.1i
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 6586
Lines: 199

Hello!
Quoting r. Andi Kleen (ak@suse.de) "Re: [discuss] f_ops flag to speed up compatible ioctls in linux kernel":
> On Tue, Sep 07, 2004 at 01:40:17PM +0300, Michael S. Tsirkin wrote:
> > Hello!
> > Quoting Andi Kleen (ak@suse.de) "Re: [discuss] f_ops flag to speed up compatible ioctls in linux kernel":
> > > On Wed, Sep 01, 2004 at 10:22:45AM +0300, Michael S. Tsirkin wrote:
> > > > Hello!
> > > > Currently, on the x86_64 architecture, its quite tricky to make
> > > > a char device ioctl work for an x86 executables.
> > > > In particular,
> > > >    1. there is a requirement that ioctl number is unique -
> > > >       which is hard to guarantee especially for out of kernel modules
> > > 
> > > Yes, that is a problem for some people. But you should
> > > have used an unique number in the first place.
> > 
> > Do you mean the _IOC macro and friends?
> > But their uniqueness depends on allocating a unique magic number
> > in the first place.
> 
> Yep. It's not bullet proof, but works pretty well in practice with
> a little care.

Hrmp. I for one *would* like something moer bulletproof.

> > 
> > > There are some hackish ways to work around it for non modules[1], but at some
> > > point we should probably support it better.
> > > 
> > > [1] it can be handled, except for module unloading, so you have
> > > to disable that.
> > 
> > Why use the global hash at all?
> > Why not, for example, pass a parameter to the ioctl function
> > to make it possible to figure out this is a compat call?
> 
> The main reason is that traditionally there was some resistance
> to put compat code into the drivers itself because it "looked too
> ugly". So it was just put into a few centralized files. Patching 
> all the f_ops wouldn't have been practical for this. 
> 
> Maybe it could be added as an additional mechanism now though.

I'll try to add it and see what this does not performance,
if this helps I'll send a patch.


> > > >    2. there's a performance huge overhead for each compat call - there's
> > > >       a hash lookup in a global hash inside a lock_kernel -
> > > >       and I think compat performance *is* important.
> > > 
> > > Did you actually measure it? I doubt it is a big issue.
> > > 
> > 
> > But that would depend on what the driver actually does inside
> > the ioctl and on how many ioctls are already registered, would it not?
> 
> Most ioctls should be registered at boot, the additional ones
> are probably negligible.

But this does not matter - the hash collision will add overhead
on each lookup - and whether you have collisions is a matter of luck -
theoretically, some users may use such drivers that you may always have
collisions.

> > 
> > I built a silly driver example which just used a semaphore and a switch
> > statement inside the ioctl.
> > 
> > ~/<1>tavor/tools/driver_new>time /tmp/ioctltest64 /dev/mst/mt23108_pci_cr0
> > 0.357u 4.760s 0:05.11 100.0%    0+0k 0+0io 0pf+0w
> > ~/<1>tavor/tools/driver_new>time /tmp/ioctltest32 /dev/mst/mt23108_pci_cr0
> > 0.641u 6.486s 0:07.12 100.0%    0+0k 0+0io 0pf+0w
> > 
> > So just looking at system time there seems to be an overhead of
> > about 20%.
> 
> That's with an empty ioctl?

Not exactly empty - below's the code snippet.


***

static int ioctl (struct inode *inode, struct file *file, unsigned int opcode, unsigned long udata_l)
{
  void* udata=(void*)udata_l;
  int minor=MINOR(inode->i_rdev);
  struct dev_data* dev=&devices[minor];
  int ret=0;

  /* By convention, any user gets read access
   * and is allowed to use the device.
   * Commands with no direction are administration
   * commands, and you need write permission
   * for this */

  if ( _IOC_DIR(opcode) == _IOC_NONE ) {
    if (! ( file->f_mode & FMODE_WRITE) ) return -EPERM;
  } else {
    if (! ( file->f_mode & FMODE_READ) ) return -EPERM;
  }

  if (down_interruptible(&devices[minor].sem)) {
    return -ERESTARTSYS;
  }


  switch (opcode) {

    /* .. snip .. */

    case PARAMS:
      {
        struct mst_pci_params_st paramsd;
        paramsd.bar=dev->bar;
        paramsd.size=dev->size;

        if (copy_to_user(udata, &paramsd, sizeof(paramsd))) {
          ret=-EFAULT;
        }
        goto fin;
      }

    default:
      ret= -ENOTTY;
      goto fin;
  }

  fin:
  up(&devices[minor].sem);
  return ret;
}

***


> I would expect most ioctls to do
> more work, so the overhead would be less.
> Still it could be probably made better. 

Then I expect you'll get bitten by the BKL taken while ioctl runs.
That's another issue that needs addressing, in my opinion.

> > The overhead is bigger if there are collisions in the hash.
> > 
> > For muti-processor scenarious, the difference is much more pronounced
> > (note I have dual-cpu Opteron system):
> > 
> > ~>time /tmp/ioctltest32 /dev/mst/mt23108_pci_cr0 & ;time /tmp/ioctltest32
> > /dev/mst/mt23108_pci_cr0 &
> > [2] 10829
> > [3] 10830
> > [2]    Done                          /tmp/ioctltest32 /dev/mst/mt23108_pci_cr0
> > 0.435u 21.322s 0:21.76 99.9%    0+0k 0+0io 0pf+0w
> > [3]    Done                          /tmp/ioctltest32 /dev/mst/mt23108_pci_cr0
> > 0.683u 21.231s 0:21.92 99.9%    0+0k 0+0io 0pf+0w
> > ~>
> > 
> > 
> > ~>time /tmp/ioctltest64 /dev/mst/mt23108_pci_cr0 & ;time /tmp/ioctltest64
> > /dev/mst/mt23108_pci_cr0 &
> > [2] 10831
> > [3] 10832
> > [3]    Done                          /tmp/ioctltest64 /dev/mst/mt23108_pci_cr0
> > 0.474u 11.194s 0:11.70 99.6%    0+0k 0+0io 0pf+0w
> > [2]    Done                          /tmp/ioctltest64 /dev/mst/mt23108_pci_cr0
> > 0.476u 11.277s 0:11.75 99.9%    0+0k 0+0io 0pf+0w
> > ~>
> > 
> > So we get 50% slowdown.
> > I imagine this is the result of BKL contention during the hash lookup.
> 
> 
> Ok, this could be improved agreed (although I still think your microbenchmark
> is a bit too artificial) 
>
> In theory the BKL could be dropped from the lookup anyways
> if RCU is needed for the cleanup. For locking the handler 
> itself into memory it doesn't make any difference.
> 
> What happens when you just remove the lock_kernel() there? 
> (as long as you don't unload any modules this should be safe) 
> 
> -Andi

Well, I personally do want to enable module unloading.
I think I'll add a new entry point to f_ops  and see what *this* does
to speed. That would be roughly equivalent, and cleaner, right?

MST
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/