cxl/region: Match auto-discovered region decoders by HPA range
authorAlison Schofield <alison.schofield@intel.com>
Tue, 5 Sep 2023 21:10:07 +0000 (14:10 -0700)
committerDan Williams <dan.j.williams@intel.com>
Fri, 15 Sep 2023 03:46:12 +0000 (20:46 -0700)
Currently, when the region driver attaches a region to a port, it
selects the ports next available decoder to program.

With the addition of auto-discovered regions, a port decoder has
already been programmed so grabbing the next available decoder can
be a mismatch when there is more than one region using the port.

The failure appears like this with CXL DEBUG enabled:

[] cxl_core:alloc_region_ref:754: cxl region0: endpoint9: HPA order violation region0:[mem 0x14780000000-0x1478fffffff flags 0x200] vs [mem 0x880000000-0x185fffffff flags 0x200]
[] cxl_core:cxl_port_attach_region:972: cxl region0: endpoint9: failed to allocate region reference

When CXL DEBUG is not enabled, there is no failure message. The region
just never materializes. Users can suspect this issue if they know their
firmware has programmed decoders so that more than one region is using
a port. Note that the problem may appear intermittently, ie not on
every reboot.

Add a matching method for auto-discovered regions that finds a decoder
based on an HPA range. The decoder range must exactly match the region
resource parameter.

Fixes: a32320b71f08 ("cxl/region: Add region autodiscovery")
Signed-off-by: Alison Schofield <alison.schofield@intel.com>
Reviewed-by: Dave Jiang <dave.jiang@intel.com>
Reviewed-by: Davidlohr Bueso <dave@stgolabs.net>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Link: https://lore.kernel.org/r/20230905211007.256385-1-alison.schofield@intel.com
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
drivers/cxl/core/region.c

index e115ba382e0445dc6c8ec1881e0e95ec307c4918..b4c6a749406f176605dba3a7ec93cdafca1ceaa7 100644 (file)
@@ -717,13 +717,35 @@ static int match_free_decoder(struct device *dev, void *data)
        return 0;
 }
 
+static int match_auto_decoder(struct device *dev, void *data)
+{
+       struct cxl_region_params *p = data;
+       struct cxl_decoder *cxld;
+       struct range *r;
+
+       if (!is_switch_decoder(dev))
+               return 0;
+
+       cxld = to_cxl_decoder(dev);
+       r = &cxld->hpa_range;
+
+       if (p->res && p->res->start == r->start && p->res->end == r->end)
+               return 1;
+
+       return 0;
+}
+
 static struct cxl_decoder *cxl_region_find_decoder(struct cxl_port *port,
                                                   struct cxl_region *cxlr)
 {
        struct device *dev;
        int id = 0;
 
-       dev = device_find_child(&port->dev, &id, match_free_decoder);
+       if (test_bit(CXL_REGION_F_AUTO, &cxlr->flags))
+               dev = device_find_child(&port->dev, &cxlr->params,
+                                       match_auto_decoder);
+       else
+               dev = device_find_child(&port->dev, &id, match_free_decoder);
        if (!dev)
                return NULL;
        /*