Hi, Alex,


On 2017/2/1 3:37, Alexander Graf wrote:

On 31/01/2017 14:32, John Garry wrote:

quoted

On 30/01/2017 17:12, Alexander Graf wrote:

quoted

On 01/24/2017 08:05 AM, zhichang.yuan wrote:

quoted

Low-pin-count interface is integrated into some SoCs. The accesses to
those
peripherals under LPC make use of I/O ports rather than the memory
mapped I/O.

To drive these devices, this patch introduces a method named
indirect-IO.
In this method the in/out() accessor in include/asm-generic/io.h will be
redefined. When upper layer drivers call in/out() with those known
legacy port
addresses to access the peripherals, the I/O operations will be routed
to the
right hooks which are registered specific to the host device, such as
LPC.
Then the hardware relevant manupulations are finished by the
corresponding
host.

According to the comments on V5, this patch adds a common indirect-IO
driver
which support this I/O indirection to the generic directory.

In the later pathches, some host-relevant drivers are implemented to
support
the specific I/O hooks and register them.
Based on these, the upper layer drivers which depend on in/out() can
work well
without any extra work or any changes.

Signed-off-by: zhichang.yuan <redacted>
Signed-off-by: Gabriele Paoloni <redacted>
Signed-off-by: John Garry <redacted>

I like the extio idea. That allows us to handle all PIO requests on
platforms that don't have native PIO support via different routes
depending on the region they're in. Unfortunately we now we have 2
frameworks for handling sparse PIO regions: One in extio, one in PCI.

Why don't we just merge the two? Most of the code that has #ifdef
PCI_IOBASE throughout the code base sounds like an ideal candidate to
get migrated to extio instead. Then we only have a single framework to
worry about ...

To be clear, are you suggesting we merge the functionality from
pci_register_io_range(), pci_pio_to_address(), pci_address_to_pio() into
extio, so extio manages all PIO?

Yes, I guess so.

quoted

And having a single type of node to
register PIO ranges, by amalgamating struct extio_node and io_range (as
Bjorn mentioned)?

I'm not quite sure I follow you here. Basically I think you want a generic "non-x86 PIO" framework that PCI just plugs into.

I don't think that necessarily means you want to statically allocate regions of that PIO space to separate (pseudo-)devices. Instead, everyone shares that space and should be able to fail gracefully if some space is already occupied.

quoted

It would make sense. We would be somewhat decoupling PIO from PCI.

Yes :).

quoted

I think that other architectures, like PPC, and other code would need to
be fixed up to handle this.

I think only PPC, Microblaze and ARM are using this. Grep for PCI_IOBASE. It's not that many.

quoted

We need to consider all the other challenges/obstacles to this.

Well, getting our abstraction levels right to me sounds like it's worth the obstacles.

quoted

quoted

quoted

---
  include/asm-generic/io.h |  50 ++++++++++++++++
  include/linux/extio.h    |  85 +++++++++++++++++++++++++++
  include/linux/io.h       |   1 +
  lib/Kconfig              |   8 +++
  lib/Makefile             |   2 +
  lib/extio.c              | 147
+++++++++++++++++++++++++++++++++++++++++++++++ xc>>   create mode
100644 include/linux/extio.h
  create mode 100644 lib/extio.c

<snip>

quoted

quoted

+ * Copyright (C) 2016 Hisilicon Limited, All Rights Reserved.
+ * Author: Zhichang Yuan [off-list ref]
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License version 2 as
+ * published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program.  If not, see
<http://www.gnu.org/licenses/>.
+ */
+
+#include <linux/io.h>
+#include <linux/spinlock.h>
+
+static LIST_HEAD(extio_dev_list);
+static DEFINE_RWLOCK(extio_list_lock);

Why not just make the list an RCU list? Then you don't need read locks.
We also wouldn't create potential lock contention between devices that
could easily have parallel PIO operations (say a PCI device and an LPC
device).

OK

quoted

quoted

+
+void register_extio(struct extio_node *node)
+{
+    write_lock(&extio_list_lock);
+    list_add_tail(&node->list, &extio_dev_list);
+    write_unlock(&extio_list_lock);
+}
+
+static struct extio_node *find_extio_token(unsigned long addr)
+{
+    struct extio_node *extio_entry;
+
+    read_lock(&extio_list_lock);
+    list_for_each_entry(extio_entry, &extio_dev_list, list) {
+        if ((addr < extio_entry->io_start + extio_entry->range_size) &&
+            (addr >= extio_entry->io_start))
+            break;
+    }
+    read_unlock(&extio_list_lock);
+    return (&extio_entry->list == &extio_dev_list) ? NULL :
extio_entry;
+}
+
+struct extio_node *extio_find_node(struct fwnode_handle *node)
+{
+    struct extio_node *entry;
+
+    read_lock(&extio_list_lock);
+    list_for_each_entry(entry, &extio_dev_list, list) {
+        if (entry->fwnode == node)
+            break;
+    }
+    read_unlock(&extio_list_lock);
+
+    return (&entry->list == &extio_dev_list) ? NULL : entry;
+}
+
+unsigned long extio_translate(struct fwnode_handle *node,
+        unsigned long bus_addr)
+{
+    struct extio_node *entry;
+    unsigned long port_id = -1;
+
+    read_lock(&extio_list_lock);
+    list_for_each_entry(entry, &extio_dev_list, list) {
+        if (entry->fwnode == node &&
+            bus_addr >= entry->bus_start &&
+            bus_addr - entry->bus_start < entry->range_size)
+            port_id = entry->io_start + bus_addr -
+                    entry->bus_start;
+    }
+    read_unlock(&extio_list_lock);
+
+    return port_id;
+}
+
+#ifdef PCI_IOBASE
+
+#define BUILD_EXTIO(bw, type)                        \
+type extio_in##bw(unsigned long addr)                    \
+{                                    \
+    struct extio_node *extio_entry = find_extio_token(addr);    \
+                                    \
+    if (!extio_entry)                        \
+        return read##bw(PCI_IOBASE + addr);            \
+    return extio_entry->ops->pfin ?                    \
+            extio_entry->ops->pfin(extio_entry->devpara,    \
+            addr, sizeof(type)) : -1;            \
+}                                    \
+                                    \
+void extio_out##bw(type value, unsigned long addr)            \
+{                                    \
+    struct extio_node *extio_entry = find_extio_token(addr);    \
+                                    \
+    if (!extio_entry)                        \
+        write##bw(value, PCI_IOBASE + addr);            \

All of the fallback code would also disappear as a nice side effect of
making pci pio handling a user of extio :).

Is your idea that PCI IO space will also register accessors, which would
be the same read{b,w,l}/write{b,w,l}?

I am not so sure what is your ideas on this. Do you mean the snippet like these:

#define BUILD_IO(bw, type)					\
type extio_in##bw(unsigned long addr) 				\
{ 								\
	struct io_range *entry = find_io_range(addr); 		\
								\
	if (entry) 						\
		return entry->ops->pfin(entry->devpara, 	\
			addr, sizeof(type)); 			\
	return read##bw(PCI_IOBASE + addr);			\
}

we add the last 'return read##bw(PCI_IOBASE + addr);' to keep the original logic of inX() in asm-generic/io.h;
In above snippet, all the hosts applied extio should register their own ops->pfin().


Thanks,
Zhichang

Yes. If you need to later on accelerate that bit, you can always do something like

  if (extio_entry->ops->pfin == pci_extio_in)
    return pci_extio_in(...);

which should get you all the prefetcher and branch prediction benefits that the current version gives you. But for starters I'd leave that out, since I doubt it'll have measurable performance impact to go via an indirect function call.

quoted

quoted

It would be nice to have a quicker way to so the lookup from address to
node, as we loop all nodes in find_extio_token() every single time.

You can always replace the search with a tree. But to me that's an implementation detail that's easy enough to replace in a follow-up patch series.


Alex

.