[PATCH v2 09/14] device core: Add ability to handle multiple dma offsets

Jim Quinlan james.quinlan at broadcom.com
Wed May 27 15:43:23 UTC 2020


Hi Nicolas,

On Wed, May 27, 2020 at 11:00 AM Nicolas Saenz Julienne
<nsaenzjulienne at suse.de> wrote:
>
> Hi Jim,
> one thing comes to mind, there is a small test suite in drivers/of/unittest.c
> (specifically of_unittest_pci_dma_ranges()) you could extend it to include your
> use cases.
Sure, will check out.
>
> On Tue, 2020-05-26 at 15:12 -0400, Jim Quinlan wrote:
> > The new field in struct device 'dma_pfn_offset_map' is used to facilitate
> > the use of multiple pfn offsets between cpu addrs and dma addrs.  It is
> > similar to 'dma_pfn_offset' except that the offset chosen depends on the
> > cpu or dma address involved.
> >
> > Signed-off-by: Jim Quinlan <james.quinlan at broadcom.com>
> > ---
> >  drivers/of/address.c        | 65 +++++++++++++++++++++++++++++++++++--
> >  drivers/usb/core/message.c  |  3 ++
> >  drivers/usb/core/usb.c      |  3 ++
> >  include/linux/device.h      | 10 +++++-
> >  include/linux/dma-direct.h  | 10 ++++--
> >  include/linux/dma-mapping.h | 46 ++++++++++++++++++++++++++
> >  kernel/dma/Kconfig          | 13 ++++++++
> >  7 files changed, 144 insertions(+), 6 deletions(-)
> >
>
> [...]
>
> > @@ -977,10 +1020,19 @@ int of_dma_get_range(struct device *dev, struct
> > device_node *np, u64 *dma_addr,
> >               pr_debug("dma_addr(%llx) cpu_addr(%llx) size(%llx)\n",
> >                        range.bus_addr, range.cpu_addr, range.size);
> >
> > +             num_ranges++;
> >               if (dma_offset && range.cpu_addr - range.bus_addr != dma_offset)
> > {
> > -                     pr_warn("Can't handle multiple dma-ranges with different
> > offsets on node(%pOF)\n", node);
> > -                     /* Don't error out as we'd break some existing DTs */
> > -                     continue;
> > +                     if (!IS_ENABLED(CONFIG_DMA_PFN_OFFSET_MAP)) {
> > +                             pr_warn("Can't handle multiple dma-ranges with
> > different offsets on node(%pOF)\n", node);
> > +                             pr_warn("Perhaps set DMA_PFN_OFFSET_MAP=y?\n");
> > +                             /*
> > +                              * Don't error out as we'd break some existing
> > +                              * DTs that are using configs w/o
> > +                              * CONFIG_DMA_PFN_OFFSET_MAP set.
> > +                              */
> > +                             continue;
>
> dev->bus_dma_limit is set in of_dma_configure(), this function's caller, based
> on dma_start's value (set after this continue). So you'd be effectively setting
> the dev->bus_dma_limit to whatever we get from the first dma-range.
I'm not seeing that at all.  On the  evaluation of each dma-range,
dma_start and dma_end are re-evaluated to be the lowest and highest
bus values of the  dma-ranges seen so far.  After all dma-ranges are
examined,  dev->bus_dma_limit being set to the highest.  In fact, the
current code -- ie before my commits -- already does this for multiple
dma-ranges as long as the cpu-bus offset is the same in the
dma-ranges.
>
> This can be troublesome depending on how the dma-ranges are setup, for example
> if the first dma-range doesn't include the CMA area, in arm64 generally set as
> high as possible in ZONE_DMA32, that would render it useless for
> dma/{direct/swiotlb}. Again depending on the bus_dma_limit value, if smaller
> than ZONE_DMA you'd be unable to allocate any DMA memory.
>
> IMO, a solution to this calls for a revamp of dma-direct's dma_capable(): match
> the target DMA memory area with each dma-range we have to see if it fits.
>
> > +                     }
> > +                     dma_multi_pfn_offset = true;
> >               }
> >               dma_offset = range.cpu_addr - range.bus_addr;
> >
> > @@ -991,6 +1043,13 @@ int of_dma_get_range(struct device *dev, struct
> > device_node *np, u64 *dma_addr,
> >                       dma_end = range.bus_addr + range.size;
> >       }
> >
> > +     if (dma_multi_pfn_offset) {
> > +             dma_offset = 0;
> > +             ret = attach_dma_pfn_offset_map(dev, node, num_ranges);
> > +             if (ret)
> > +                     return ret;
> > +     }
> > +
> >       if (dma_start >= dma_end) {
> >               ret = -EINVAL;
> >               pr_debug("Invalid DMA ranges configuration on node(%pOF)\n",
> > diff --git a/drivers/usb/core/message.c b/drivers/usb/core/message.c
> > index 6197938dcc2d..aaa3e58f5eb4 100644
> > --- a/drivers/usb/core/message.c
> > +++ b/drivers/usb/core/message.c
> > @@ -1960,6 +1960,9 @@ int usb_set_configuration(struct usb_device *dev, int
> > configuration)
> >                */
> >               intf->dev.dma_mask = dev->dev.dma_mask;
> >               intf->dev.dma_pfn_offset = dev->dev.dma_pfn_offset;
> > +#ifdef CONFIG_DMA_PFN_OFFSET_MAP
> > +             intf->dev.dma_pfn_offset_map = dev->dev.dma_pfn_offset_map;
> > +#endif
>
> Thanks for looking at this, that said, I see more instances of drivers changing
> dma_pfn_offset outside of the core code. Why not doing this there too?
>
> Also, are we 100% sure that dev->dev.dma_pfn_offset isn't going to be freed
> before we're done using intf->dev? Maybe it's safer to copy the ranges?
>
> >               INIT_WORK(&intf->reset_ws, __usb_queue_reset_device);
> >               intf->minor = -1;
> >               device_initialize(&intf->dev);
> > diff --git a/drivers/usb/core/usb.c b/drivers/usb/core/usb.c
> > index f16c26dc079d..d2ed4d90e56e 100644
> > --- a/drivers/usb/core/usb.c
> > +++ b/drivers/usb/core/usb.c
> > @@ -612,6 +612,9 @@ struct usb_device *usb_alloc_dev(struct usb_device
> > *parent,
> >        */
> >       dev->dev.dma_mask = bus->sysdev->dma_mask;
> >       dev->dev.dma_pfn_offset = bus->sysdev->dma_pfn_offset;
> > +#ifdef CONFIG_DMA_PFN_OFFSET_MAP
> > +     dev->dev.dma_pfn_offset_map = bus->sysdev->dma_pfn_offset_map;
> > +#endif
> >       set_dev_node(&dev->dev, dev_to_node(bus->sysdev));
> >       dev->state = USB_STATE_ATTACHED;
> >       dev->lpm_disable_count = 1;
> > diff --git a/include/linux/device.h b/include/linux/device.h
> > index ac8e37cd716a..67a240ad4fc5 100644
> > --- a/include/linux/device.h
> > +++ b/include/linux/device.h
> > @@ -493,6 +493,8 @@ struct dev_links_info {
> >   * @bus_dma_limit: Limit of an upstream bridge or bus which imposes a smaller
> >   *           DMA limit than the device itself supports.
> >   * @dma_pfn_offset: offset of DMA memory range relatively of RAM
> > + * @dma_pfn_offset_map:      Like dma_pfn_offset but used when there are
> > multiple
> > + *           pfn offsets for multiple dma-ranges.
> >   * @dma_parms:       A low level driver may set these to teach IOMMU code
> > about
> >   *           segment limitations.
> >   * @dma_pools:       Dma pools (if dma'ble device).
> > @@ -578,7 +580,13 @@ struct device {
> >                                            allocations such descriptors. */
> >       u64             bus_dma_limit;  /* upstream dma constraint */
> >       unsigned long   dma_pfn_offset;
> > -
> > +#ifdef CONFIG_DMA_PFN_OFFSET_MAP
> > +     const struct dma_pfn_offset_region *dma_pfn_offset_map;
> > +                                     /* Like dma_pfn_offset, but for
> > +                                      * the unlikely case of multiple
> > +                                      * offsets. If non-null, dma_pfn_offset
> > +                                      * will be set to 0. */
> > +#endif
>
> I'm still sad this doesn't fully replace dma_pfn_offset & bus_dma_limit. I feel
> the extra logic involved in incorporating this as default isn't going to be
> noticeable as far as performance is concerned to single dma-range users, and
> it'd make for a nicer DMA code. Also you'd force everyone to test their changes
> on the multi dma-ranges code path, as opposed to having this disabled 99.9% of
> the time (hence broken every so often).
Good point.
>
> Note that I sympathize with the amount of work involved on improving that, so
> better wait to hear what more knowledgeable people have to say about this :)
Yes, I agree.  I want to avoid coding and testing one solution only to
have a different reviewer NAK it.

Many thanks,
Jim
>
> Regards,
> Nicolas
>


More information about the iommu mailing list