RE: [PATCH V2 02/10] Drivers: hv: utils: run polling callback always in interrupt context
From: KY Srinivasan
Date: Thu Dec 10 2015 - 18:23:15 EST
> -----Original Message-----
> From: Dexuan Cui
> Sent: Thursday, December 10, 2015 2:19 AM
> To: KY Srinivasan <kys@xxxxxxxxxxxxx>; gregkh@xxxxxxxxxxxxxxxxxxx; linux-
> kernel@xxxxxxxxxxxxxxx; devel@xxxxxxxxxxxxxxxxxxxxxx; olaf@xxxxxxxxx;
> apw@xxxxxxxxxxxxx; vkuznets@xxxxxxxxxx; jasowang@xxxxxxxxxx
> Subject: RE: [PATCH V2 02/10] Drivers: hv: utils: run polling callback always in
> interrupt context
> > -----Original Message-----
> > From: devel [mailto:driverdev-devel-bounces@xxxxxxxxxxxxxxxxxxxxxx] On
> > Of K. Y. Srinivasan
> > Sent: Friday, October 30, 2015 9:13
> > To: gregkh@xxxxxxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx;
> > devel@xxxxxxxxxxxxxxxxxxxxxx; olaf@xxxxxxxxx; apw@xxxxxxxxxxxxx;
> > vkuznets@xxxxxxxxxx; jasowang@xxxxxxxxxx
> > Subject: [PATCH V2 02/10] Drivers: hv: utils: run polling callback always in
> > interrupt context
> > From: Olaf Hering <olaf@xxxxxxxxx>
> > All channel interrupts are bound to specific VCPUs in the guest
> > at the point channel is created. While currently, we invoke the
> > polling function on the correct CPU (the CPU to which the channel
> > is bound to) in some cases we may run the polling function in
> > a non-interrupt context. This potentially can cause an issue as the
> > polling function can be interrupted by the channel callback function.
> > Fix the issue by running the polling function on the appropriate CPU
> > at interrupt level. Additional details of the issue being addressed by
> > this patch are given below:
> > Currently hv_fcopy_onchannelcallback is called from interrupts and also
> > via the ->write function of hv_utils. Since the used global variables to
> > maintain state are not thread safe the state can get out of sync.
> > This affects the variable state as well as the channel inbound buffer.
> > As suggested by KY adjust hv_poll_channel to always run the given
> > callback on the cpu which the channel is bound to. This avoids the need
> > for locking because all the util services are single threaded and only
> > one transaction is active at any given point in time.
> > Additionally, remove the context variable, they will always be the same as
> > recv_channel.
> > Signed-off-by: Olaf Hering <olaf@xxxxxxxxx>
> > Signed-off-by: K. Y. Srinivasan <kys@xxxxxxxxxxxxx>
> > ---
> > V2: Added the check to catch unsolicited daemon writes - Vitaly
> > drivers/hv/hv_fcopy.c | 34 +++++++++++++---------------------
> > drivers/hv/hv_kvp.c | 28 ++++++++++------------------
> > drivers/hv/hv_snapshot.c | 29 +++++++++++------------------
> > drivers/hv/hyperv_vmbus.h | 6 +-----
> > 4 files changed, 35 insertions(+), 62 deletions(-)
> (Sorry for not joining the discussion when the patch was firstly made)
> It looks the patch has not been Greg's tree yet.
> I have 2 questions about the patch:
> 1. hv_poll_channel() is invoked in fcopy_handle_handshake(), but not in
> vss_handle_handshake() and kvp_handle_handshake().
> Why -- I guess we missed the vss/kvp cases somehow?
I will fix this.
> 2. With the patch, hv_fcopy_onchannelcallback() can be invoked in the
> tasklet (i.e., vmbus_on_event(). NB: local irq is enabled), and in the
> hard irq handler(the IPI handler, e.g.,
> fcopy_poll_wrapper() -> fcopy_poll_wrapper()).
> Can the former be interrupted by the latter?
> e.g., when the callback is running in the tasklet on vCPU0,
> fcopy_timeout_func() or fcopy_on_msg() could send the IPI to
> vCPU0 from another vCPU.
Keep in mind that when the poll function is run, the state will not be
HVUTIL_READY. The state will be set to HVUTIL_READY in the IPI
handler. So, it is ok if the tasklet is interrupted by the IPI handler.
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/