On Fri, Oct 6, 2017 at 6:56 AM, Peter Maydell <peter.mayd...@linaro.org> wrote:
> On 18 September 2017 at 20:50, Andrey Smirnov <andrew.smir...@gmail.com>
> wrote:
>> In current implementation, packet queue flushing logic seem to suffer
>> from a deadlock like scenario if a packet is received by the interface
>> before before Rx ring is initialized by Guest's driver. Consider the
>> following sequence of events:
>>
>> 1. A QEMU instance is started against a TAP device on Linux
>> host, running Linux guest, e. g., something to the effect
>> of:
>>
>> qemu-system-arm \
>> -net nic,model=imx.fec,netdev=lan0 \
>> netdev tap,id=lan0,ifname=tap0,script=no,downscript=no \
>> ... rest of the arguments ...
>>
>> 2. Once QEMU starts, but before guest reaches the point where
>> FEC deriver is done initializing the HW, Guest, via TAP
>> interface, receives a number of multicast MDNS packets from
>> Host (not necessarily true for every OS, but it happens at
>> least on Fedora 25)
>>
>> 3. Recieving a packet in such a state results in
>> imx_eth_can_receive() returning '0', which in turn causes
>> tap_send() to disable corresponding event (tap.c:203)
>>
>> 4. Once Guest's driver reaches the point where it is ready to
>> recieve packets it prepares Rx ring descriptors and writes
>> ENET_RDAR_RDAR to ENET_RDAR register to indicate to HW that
>> more descriptors are ready. And at this points emulation
>> layer does this:
>>
>> s->regs[index] = ENET_RDAR_RDAR;
>> imx_eth_enable_rx(s);
>>
>> which, combined with:
>>
>> if (!s->regs[ENET_RDAR]) {
>> qemu_flush_queued_packets(qemu_get_queue(s->nic));
>> }
>>
>> results in Rx queue never being flushed and corresponding
>> I/O event beign disabled.
>>
>> Change the code to remember the fact that can_receive callback was
>> called before Rx ring was ready and use it to make a decision if
>> receive queue needs to be flushed.
>>
>> Cc: Peter Maydell <peter.mayd...@linaro.org>
>> Cc: Jason Wang <jasow...@redhat.com>
>> Cc: qemu-devel@nongnu.org
>> Cc: qemu-...@nongnu.org
>> Cc: yurov...@gmail.com
>> Signed-off-by: Andrey Smirnov <andrew.smir...@gmail.com>
>> ---
>> hw/net/imx_fec.c | 6 ++++--
>> include/hw/net/imx_fec.h | 1 +
>> 2 files changed, 5 insertions(+), 2 deletions(-)
>>
>> diff --git a/hw/net/imx_fec.c b/hw/net/imx_fec.c
>> index 84085afe09..767402909d 100644
>> --- a/hw/net/imx_fec.c
>> +++ b/hw/net/imx_fec.c
>> @@ -544,8 +544,9 @@ static void imx_eth_enable_rx(IMXFECState *s)
>>
>> if (rx_ring_full) {
>> FEC_PRINTF("RX buffer full\n");
>> - } else if (!s->regs[ENET_RDAR]) {
>> + } else if (s->needs_flush) {
>> qemu_flush_queued_packets(qemu_get_queue(s->nic));
>> + s->needs_flush = false;
>> }
>>
>> s->regs[ENET_RDAR] = rx_ring_full ? 0 : ENET_RDAR_RDAR;
>> @@ -930,7 +931,8 @@ static int imx_eth_can_receive(NetClientState *nc)
>>
>> FEC_PRINTF("\n");
>>
>> - return s->regs[ENET_RDAR] ? 1 : 0;
>> + s->needs_flush = !s->regs[ENET_RDAR];
>> + return !!s->regs[ENET_RDAR];
>> }
>>
>> static ssize_t imx_fec_receive(NetClientState *nc, const uint8_t *buf,
>> diff --git a/include/hw/net/imx_fec.h b/include/hw/net/imx_fec.h
>> index 62ad473b05..4bc8f03ec2 100644
>> --- a/include/hw/net/imx_fec.h
>> +++ b/include/hw/net/imx_fec.h
>> @@ -252,6 +252,7 @@ typedef struct IMXFECState {
>> uint32_t phy_int_mask;
>>
>> bool is_fec;
>> + bool needs_flush;
>> } IMXFECState;
>
> This looks odd -- I don't think you should need extra
> state here. Conceptually what you want is:
>
> * in the can_receive callback, test some function of
> various bits of device state to decide whether you can
> take data
> * in the rest of the device, whenever the device state
> changes such that you were previously not able to take
> data but now you can, call qemu_flush_queued_packets().
>
> You shouldn't need any extra state to do this, you just
> need to fix the bug where you have a code path that
> flips ENET_RDAR from 0 to 1 without calling flush
> (you might for instance have a helper function for
> "set ENET_RDAR" that encapsulates setting the state
> and arranging that flush is called).
>
I don't know if you've seen my response to Jason Wang, but I think he
was proposing something similar, and, as I said, that should work fine
and the only reason I didn't do it that way was to avoid doing a flush
every time that host driver drains full RX-ring and gives it back to
the IP block.
I'll give this a try in v2.
Thanks,
Andrey Smirnov
