(+ Eric)
On Mon, 25 May 2020 at 11:01, Nicolas TOROMANOFF
<nicolas.toromanoff@xxxxxx> wrote:
>
> > -----Original Message-----
> > From: Ard Biesheuvel <ardb@xxxxxxxxxx>
> > Sent: Monday, May 25, 2020 9:46 AM
> > To: Nicolas TOROMANOFF <nicolas.toromanoff@xxxxxx>
> > Subject: Re: [PATCH 5/5] crypto: stm32/crc: protect from concurrent accesses
> >
> > On Mon, 25 May 2020 at 09:24, Nicolas TOROMANOFF
> > <nicolas.toromanoff@xxxxxx> wrote:
> > >
> > > Hello,
> > >
> > > > -----Original Message-----
> > > > From: Ard Biesheuvel <ardb@xxxxxxxxxx>
> > > > Sent: Friday, May 22, 2020 6:12 PM>
> > > > On Tue, 12 May 2020 at 16:13, Nicolas Toromanoff
> > > > <nicolas.toromanoff@xxxxxx> wrote:
> > > > >
> > > > > Protect STM32 CRC device from concurrent accesses.
> > > > >
> > > > > As we create a spinlocked section that increase with buffer size,
> > > > > we provide a module parameter to release the pressure by splitting
> > > > > critical section in chunks.
> > > > >
> > > > > Size of each chunk is defined in burst_size module parameter.
> > > > > By default burst_size=0, i.e. don't split incoming buffer.
> > > > >
> > > > > Signed-off-by: Nicolas Toromanoff <nicolas.toromanoff@xxxxxx>
> > > >
> > > > Would you mind explaining the usage model here? It looks like you
> > > > are sharing a CRC hardware accelerator with a synchronous interface
> > > > between different users by using spinlocks? You are aware that this
> > > > will tie up the waiting CPUs completely during this time, right? So
> > > > it would be much better to use a mutex here. Or perhaps it would
> > > > make more sense to fall back to a s/w based CRC routine if the h/w is tied up
> > working for another task?
> > >
> > > I know mutex are more acceptable here, but shash _update() and _init()
> > > may be call from any context, and so I cannot take a mutex.
> > > And to protect my concurrent HW access I only though about spinlock.
> > > Due to possible constraint on CPUs, I add a burst_size option to force
> > > slitting long buffer into smaller one, and so decrease time we take the lock.
> > > But I didn't though to fallback to software CRC.
> > >
> > > I'll do a patch on top.
> > > In in the burst_update() function I'll use a spin_trylock_irqsave() and use
> > software CRC32 if HW is already in use.
> > >
> >
> > Right. I didn't even notice that you were keeping interrupts disabled the whole
> > time when using the h/w block. That means that any serious use of this h/w
> > block will make IRQ latency go through the roof.
> >
> > I recommend that you go back to the drawing board on this driver, rather than
> > papering over the issues with a spin_trylock(). Perhaps it would be better to
> > model it as a ahash (even though the h/w block itself is synchronous) and use a
> > kthread to feed in the data.
>
> I thought when I updated the driver to move to a ahash interface, but the main usage
> of crc32 is the ext4 fs, that calls the shash API.
> Commit 877b5691f27a ("crypto: shash - remove shash_desc::flags") removed possibility
> to sleep in shash callback. (before this commit and with MAY_SLEEP option set, using
> a mutex may have been fine).
>
According to that commit's log, sleeping is never fine for shash(),
since it uses kmap_atomic() when iterating over the scatterlist.
> By now the solution I see is to use the spin_trylock_irqsave(), fallback to software crc *AND* capping burst_size
> to ensure the locked section stay reasonable.
>
> Does this seems acceptable ?
>
If the reason for disabling interrupts is to avoid deadlocks, wouldn't
the switch to trylock() with a software fallback allow us to keep
interrupts enabled?
