On Wed, May 27, 2020 at 9:43 AM Jim Quinlan [off-list ref] wrote:

Hi Nicolas,

On Wed, May 27, 2020 at 11:00 AM Nicolas Saenz Julienne
[off-list ref] wrote:

quoted

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.

quoted

On Tue, 2020-05-26 at 15:12 -0400, Jim Quinlan wrote:

quoted

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 <redacted>
---
 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(-)

[...]

quoted

@@ -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.

quoted

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.

quoted

+                     }
+                     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?

quoted

              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.

+1

quoted

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.

It's a pretty safe bet that everyone will prefer one code path over 2.

Rob