[Top] [All Lists]

Re: serious netpoll bug w/NAPI

To: Matt Mackall <mpm@xxxxxxxxxxx>
Subject: Re: serious netpoll bug w/NAPI
From: Jeff Moyer <jmoyer@xxxxxxxxxx>
Date: Tue, 15 Feb 2005 17:49:50 -0500
Cc: "David S. Miller" <davem@xxxxxxxxxxxxx>, netdev@xxxxxxxxxxx
In-reply-to: <20050210011104.GF2366@xxxxxxxxx>
References: <20050208201634.03074349.davem@xxxxxxxxxxxxx> <20050209183219.GA2366@xxxxxxxxx> <20050209164658.409f8950.davem@xxxxxxxxxxxxx> <20050210011104.GF2366@xxxxxxxxx>
Reply-to: jmoyer@xxxxxxxxxx
Sender: netdev-bounce@xxxxxxxxxxx
==> Regarding Re: serious netpoll bug w/NAPI; Matt Mackall <mpm@xxxxxxxxxxx> 

Sorry, Matt, I'm just now getting to this.

mpm> On Wed, Feb 09, 2005 at 04:46:58PM -0800, David S. Miller wrote:
>> On Wed, 9 Feb 2005 10:32:19 -0800 Matt Mackall <mpm@xxxxxxxxxxx> wrote:
>> > On closer inspection, there's a couple other related failure cases >
>> with the new ->poll logic in netpoll. I'm afraid it looks like >
>> CONFIG_NETPOLL will need to guard ->poll() with a per-device spinlock >
>> on netpoll-enabled devices.
>> > 
>> > This will mean putting a pointer to struct netpoll in struct >
>> net_device (which I should have done in the first place) and will take >
>> a few patches to sort out.
>> Will this ->poll() guarding lock be acquired only in the netpoll code or
>> system-wide?  If the latter, this introduced an incredible performance
>> regression for devices using the LLTX locking scheme (ie. the most
>> important high-performance gigabit drivers in the tree use this).

mpm> The lock will only be taken in net_rx_action iff netpoll is enabled
mpm> for the given device. Lock contention should be basically nil.

mpm> Here's my current patch (on top of -mm), which I'm struggling to find
mpm> an appropriate test box for (my dual box with NAPI got pressed into
mpm> service as a web server a couple weeks ago). I've attached the other
mpm> two patches it relies on as well.

mpm> --------------

mpm> Introduce a per-client poll lock and flag. The lock assures we never
mpm> have more than one caller in dev->poll(). The flag provides recursion
mpm> avoidance on UP where the lock disappears.

|  /*
| - * Check whether delayed processing was scheduled for our current CPU,
| - * and then manually invoke NAPI polling to pump data off the card.
| + * Check whether delayed processing was scheduled for our NIC. If so,
| + * we attempt to grab the poll lock and use ->poll() to pump the card.
| + * If this fails, either we've recursed in ->poll() or it's already
| + * running on another CPU.
| + *
| + * Note: we don't mask interrupts with this lock because we're using
| + * trylock here and interrupts are already disabled in the softirq
| + * case. Further, we test the poll_flag to avoid recursion on UP
| + * systems where the lock doesn't exist.
|   *
|   * In cases where there is bi-directional communications, reading only
|   * one message at a time can lead to packets being dropped by the
| @@ -74,13 +80,9 @@
|  static void poll_napi(struct netpoll *np)
|  {
|       int budget = 16;
| -     unsigned long flags;
| -     struct softnet_data *queue;
| -     spin_lock_irqsave(&netpoll_poll_lock, flags);
| -     queue = &__get_cpu_var(softnet_data);
|       if (test_bit(__LINK_STATE_RX_SCHED, &np->dev->state) &&
| -         !list_empty(&queue->poll_list)) {
| +         !np->poll_flag && spin_trylock(&np->poll_lock)) {
|               np->rx_flags |= NETPOLL_RX_DROP;
|               atomic_inc(&trapped);
| @@ -88,8 +90,8 @@
|               atomic_dec(&trapped);
|               np->rx_flags &= ~NETPOLL_RX_DROP;
| +             spin_unlock(&np->poll_lock);
|       }
| -     spin_unlock_irqrestore(&netpoll_poll_lock, flags);
|  }

Okay, I've only taken a quick glance at this, but I don't quite understand
why it's safe to take out the check for the per-cpu queue.  Realize that an
interrupt may have been serviced on another processor, and a NAPI poll may
have been scheduled there.

Also, could you use the -p flag to diff when you generate your next patch?
It makes it *much* easier to review.

I'll look the patch over more closely tomorrow.



<Prev in Thread] Current Thread [Next in Thread>