On 1/17/24 14:44, Andrew Lunn wrote:
On Wed, Jan 17, 2024 at 11:30:53AM +0100, Mirsad Todorovac wrote:
On 1/17/24 02:43, Jakub Kicinski wrote:
On Tue, 16 Jan 2024 14:38:47 -0500 Sasha Levin wrote:
Mirsad proposed a patch to reduce the number of spinlock lock/unlock
operations and the function code size. This can be further improved
because the function sets a consecutive register block.
Clearly a noop and a lot of LoC changed. I vote to drop this from
the backport.
Dear Jakub,
I will not argue with a senior developer, but please let me plead for the
cause.
There are a couple of issues here:
1. Heiner's patch generates smaller and faster code, with 100+
spin_lock_irqsave()/spin_unlock_restore() pairs less.
According to this table:
[1] https://mirrors.edge.kernel.org/pub/linux/kernel/people/paulmck/perfbook/perfbook-1c.2023.06.11a.pdf#table.3.1
The cost of single lock can be 15.4 - 101.9 ns (for the example CPU),
so total savings would be 1709 - 11310 ns. But as the event of PHY power
down is not frequent, this might be a insignificant saving indeed.
2. Why I had advertised atomic programming of RTL registers in the first
place?
The mac_ocp_lock was introduced recently:
commit 91c8643578a21e435c412ffbe902bb4b4773e262
Author: Heiner Kallweit <hkallweit1@xxxxxxxxx>
Date: Mon Mar 6 22:23:15 2023 +0100
r8169: use spinlock to protect mac ocp register access
For disabling ASPM during NAPI poll we'll have to access mac ocp
registers in atomic context. This could result in races because
a mac ocp read consists of a write to register OCPDR, followed
by a read from the same register. Therefore add a spinlock to
protect access to mac ocp registers.
Reviewed-by: Simon Horman <simon.horman@xxxxxxxxxxxx>
Tested-by: Kai-Heng Feng <kai.heng.feng@xxxxxxxxxxxxx>
Tested-by: Holger Hoffstätte <holger@xxxxxxxxxxxxxxxxxxxxxx>
Signed-off-by: Heiner Kallweit <hkallweit1@xxxxxxxxx>
Signed-off-by: David S. Miller <davem@xxxxxxxxxxxxx>
Well, the answer is in the question - the very need for protecting the access
to RTL_W(8|16|32) with locks comes from the fact that something was accessing
the RTL card asynchronously.
Forgive me if this is a stupid question ...
Now - do we have a guarantee that the card will not be used asynchronously
half-programmed from something else in that case, leading to another spurious
lockup?
IMHO, shouldn't the entire reprogramming of PHY down recovery of the RTL 8411b
be done atomically, under a single spin_lock_irqsave()/spin_unlock_irqrestore()
pair?
Hi Mirsad
Please take a read of:
https://www.kernel.org/doc/html/latest/process/stable-kernel-rules.html
Do you think this patch fulfils these criteria? In particularly, "It
must either fix a real bug that bothers people...".
I agree with Heiner, this appears to be just an optimisation,
Hi Andrew,
Yes, I wasn't aware of the 100 lines limit, and yes, this is not a bug fix,
but an improvement (optimisation).
I think by this I can join to consensus, this patch is not a candidate for
backporting. :-/
However, I am concerned about the possibility of two kernel threads accessing
the RTL NIC intermittently attempting to program the NIC over the RTL_(R|W)(8|16|32)
instructions (which are expanded to readl/writel and assembly).
AFAICS, nothing prevents two (or more) threads to decide in unlikely but
possible case to program the card at the same time. (Do we have a guard lock
against this?)
mac_ocp_lock appears to be acquired and released for each RTL_(R|W)(8|16|32),
with the exception of r8168_mac_ocp_modify().
To be true to the facts, each byte will go to the right port due to address/data
pairs used in each call - however, I am worried whether there could be a scenario
like this:
CPU 1 CPU 2
start programming NIC
programming NIC
(preempted in spin_lock_irqsave()
start programming NIC
programming NIC
programming NIC
programming NIC
preempted in spin_lock_irqsave()
(resume control in spin_unlock_irqrestore()
programming NIC
programming NIC
(preempted in spin_lock_irqsave()
continue programming NIC
programming NIC
programming NIC
end programming NIC
(resume control in spin_unlock_irqrestore()
programming NIC
end programming NIC
Now, every byte, word or longword will come to the right place, thanks to
RTL_(R|W)(8|16|32) having the address/data pairs - but I worry that this
jumping from sequence to sequence might confuse the NIC.
I mean, if those latches behind the addresses cause some physical effect, maybe
the ORDER is also important, not just that every byte, word or longword comes
to the right address?
r8168_mac_ocp_read()/r8168_mac_ocp_write() guarantee that every piece of
data will end being read or written at the right address, OK. But this
does not seem to guarantee the SEQUENTIAL ORDER of the programming.
I mean, if we are dealing with physical hardware like a NIC, the order
of (especially writing) data might be crucial. 8-/
Am I making any sense?
Are we algorithmically secured that two threads will never attempt to
write data to NIC hardware registers?
Thanks.
Best regards,
Mirsad Todorovac
Andrew
--
CARNet system engineer
Faculty of Graphic Arts | Academy of Fine Arts
University of Zagreb
CARNet sistem inženjer
Grafički fakultet | Akademija likovnih umjetnosti
Sveučilište u Zagrebu
