> -----Original Message-----
> From: Burakov, Anatoly
> Sent: Monday, September 17, 2018 12:34 PM
> To: Stojaczyk, Dariusz <dariusz.stojac...@intel.com>; dev@dpdk.org;
> Santosh Shukla <santosh.shu...@caviumnetworks.com>; Hemant Agrawal
> <hemant.agra...@nxp.com>; Jerin Jacob
> <jerin.ja...@caviumnetworks.com>
> Cc: Maxime Coquelin <maxime.coque...@redhat.com>; Chas Williams
> <ch...@att.com>
> Subject: Re: [PATCH v2] eal/bus: use RTE_IOVA_PA only if phys addresses
> are available
>
> On 07-Sep-18 4:58 PM, Darek Stojaczyk wrote:
> > When neither RTE_IOVA_VA nor RTE_IOVA_PA was explicitly requested,
> > DPDK would currently fallback to the default RTE_IOVA_PA mode and
> > possibly encounter a failure later on if running as a non-priviledged
> > user. Attempting to use RTE_IOVA_VA if no phys addresses are available
> > may help in this case.
> >
> > Signed-off-by: Darek Stojaczyk <dariusz.stojac...@intel.com>
> > ---
> > Changes since v1:
> > * added a missing rte_memory.h include
> >
> > lib/librte_eal/common/eal_common_bus.c | 19 +++++++++++++++----
> > 1 file changed, 15 insertions(+), 4 deletions(-)
> >
> > diff --git a/lib/librte_eal/common/eal_common_bus.c
> > b/lib/librte_eal/common/eal_common_bus.c
> > index 0943851cc..68c581b8a 100644
> > --- a/lib/librte_eal/common/eal_common_bus.c
> > +++ b/lib/librte_eal/common/eal_common_bus.c
> > @@ -37,6 +37,7 @@
> > #include <rte_bus.h>
> > #include <rte_debug.h>
> > #include <rte_string_fns.h>
> > +#include <rte_memory.h>
> >
> > #include "eal_private.h"
> >
> > @@ -236,9 +237,19 @@ rte_bus_get_iommu_class(void)
> > mode |= bus->get_iommu_class();
> > }
> >
> > - if (mode != RTE_IOVA_VA) {
> > - /* Use default IOVA mode */
> > - mode = RTE_IOVA_PA;
> > + if (mode == RTE_IOVA_VA)
> > + return RTE_IOVA_VA;
> > +
> > + if (mode & RTE_IOVA_PA) {
> > + /* Not all buses support RTE_IOVA_VA, fallback to
> RTE_IOVA_PA */
> > + return RTE_IOVA_PA;
> > + }
> > +
> > + if (rte_eal_using_phys_addrs()) {
> > + /* Default to RTE_IOVA_PA only if it's supported */
> > + return RTE_IOVA_PA;
> > }
> > - return mode;
> > +
> > + /* Since RTE_IOVA_PA is unsupported, fallback to RTE_IOVA_VA */
> > + return RTE_IOVA_VA;
> > }
> >
>
> This is a good change, however I think that this is too pessimistic. If i
> don't
> have any devices that explictly require IOVA_PA, i should be running in
> IOVA_VA mode.
Another problem may occur when trying to hotplug devices that support only
39bit DMA. You may not be able to map any memory with vfio when in RTE_IOVA_VA
mode, as virtual addresses likely occupy more than 39 bits.
The rte_pci bus enforces RTE_IOVA_PA whenever it finds such devices on init.
I have no doubt the logic can be improved here, but for now RTE_IOVA_PA is the
only safe default.
D.
>
> This of course doesn't take hotplug into account, so a command-line switch
> to force one or the other should also be available.
>
> For example, at startup, i might have devices bound to VFIO, so IOVA_VA
> mode is picked. However, even though at a time of startup none of the
> devices require physical addresses, i also know that i might later hotplug a
> device that requires IOVA_PA (leaving the question of hotplug brokenness
> aside for now...) - currently, this scenario will not work, as i will be
> forced to
> use IOVA_VA mode unless i happen to have a IOVA_PA device available at
> startup.
>
> Similarly, if i'm running DPDK as root but am only using virtual devices like
> pcap, i should be able to force DPDK into using VA addresses [*], yet
> currently i will be forced to use IOVA_PA if i don't *also* have a few devices
> bound exclusively to VFIO.
>
> [*] Do we have vdev devices that require IOVA_PA? I can't think of any...
>
> --
> Thanks,
> Anatoly
