lkml.org 
[lkml]   [2010]   [May]   [20]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
Date
From
Subject[PATCH] arch/tile: new multi-core architecture for Linux
At Tilera we have been running Linux 2.6.26 on our architecture for a
while and distributing the sources to our customers. We just sync'ed up
our sources to 2.6.34 and would like to return it to the community more
widely, so I'm hoping to take advantage of the merge window for 2.6.35
to integrate support for our architecture.

The "tile" architecture supports the Tilera chips, both our current
32-bit chips and our upcoming 64-bit architecture. The chips are
multicore, with 64 (or 36) cores per chip on our current product line,
and up to 100 cores on the upcoming 64-bit architecture. They also
include multiple built-in memory controllers, 10 Gb Ethernet, PCIe,
and a number of other I/Os. There's more info at http://www.tilera.com.

The architecture is somewhat MIPS-like, but VLIW, with up to three
instructions per bundle. The system architecture is nicely orthogonal,
with four privilege levels that can be assigned to each of forty-odd
separate protection domains, many with an associated interrupt, e.g.
ITLB/DTLB misses, timer, performance counters, various interrupts
associated with the generic networks that connect the cores, etc.
A hypervisor (kind of like the Alpha PAL) runs at a higher privilege
level to support Linux via software-interrupt calls.

The Linux we ship has some additional performance and functionality
customization in the generic code, but appended is the patch that just
adds the minimum amount of functionality into the platform-independent
code to hook in the tile architecture code in arch/tile. We will
attempt to push the other changes to the platform-independent code
piece by piece, after the initial architecture support is in.
We will also push up the 64-bit TILE-Gx support once that architecture
is fully frozen (e.g. instruction encodings finalized).

We are using the http://www.tilera.com/scm/ web site to push
Tilera-modified sources back up to the community. At the moment, the
arch/tile hierarchy is there (as a bzipped tarball) as well as a copy
of the patch appended to this email. In addition, our gcc, binutils,
and gdb sources are available on the web site. We have not yet started
the community return process for gcc, binutils, and gdb, so they are in
a preliminary form at this point.

The git://www.tilera.com server is up, but without content yet, since
we realized this week that we need to upgrade the web server to
a 64-bit kernel to support a decent git server, so though we plan to
make the code available via git in the future, it isn't yet.

As far as the platform-independent changes go, two of the changes in the
appended patch are uncontroversial, one adding a stanza to MAINTAINERS,
and one adding a line to drivers/pci/Makefile to request "setup-bus.o
setup-irq.o" for tile PCI.

A slightly more interesting one-line change is to <linux/mm.h>,
to support lowmem PAs above the 4GB limit. We use NUMA to manage
the multiple memory controllers attached to the chip, and map some of
each controller into kernel LOWMEM to load-balance memory bandwidth for
kernel-intensive apps. The controllers can each manage up to 16GB, so we
use bits above the 4GB limit in the PA to indicate the controller number.
It turns out that generic Linux almost tolerates this, but requires one
cast in lowmem_page_address() to avoid shifting the high PA bits out of
a 32-bit PFN type.

The final change is just a PCI quirk for our TILEmpower platform, which
explains itself in the comment. This is not a critical change from our
point of view, but without it you can't use the SATA disks attached to
the PCI controller on that platform, so we're hoping it can be accepted
as part of the initial tile architecture submission as well.

I'd appreciate being cc'ed on any comments on the patch or the tile
architecture support, since although I try to follow LKML, the volume
can be somewhat overwhelming.


--- linux-2.6.34/MAINTAINERS 2010-05-16 17:17:36.000000000 -0400
+++ tilera-source/MAINTAINERS 2010-05-17 18:00:12.651112000 -0400
@@ -5436,6 +5436,12 @@
S: Maintained
F: sound/soc/codecs/twl4030*

+TILE ARCHITECTURE
+M: Chris Metcalf <cmetcalf@tilera.com>
+W: http://www.tilera.com/scm/
+S: Supported
+F: arch/tile/
+
TIPC NETWORK LAYER
M: Jon Maloy <jon.maloy@ericsson.com>
M: Allan Stephens <allan.stephens@windriver.com>
--- linux-2.6.34/include/linux/mm.h 2010-05-16 17:17:36.000000000 -0400
+++ tilera-source/include/linux/mm.h 2010-05-17 12:54:33.540145000 -0400
@@ -592,7 +592,7 @@

static __always_inline void *lowmem_page_address(struct page *page)
{
- return __va(page_to_pfn(page) << PAGE_SHIFT);
+ return __va((phys_addr_t)page_to_pfn(page) << PAGE_SHIFT);
}

#if defined(CONFIG_HIGHMEM) && !defined(WANT_PAGE_VIRTUAL)
--- linux-2.6.34/drivers/pci/Makefile 2010-05-09 21:36:28.000000000 -0400
+++ tilera-source/drivers/pci/Makefile 2010-05-13 15:03:05.615238000 -0400
@@ -49,6 +49,7 @@
obj-$(CONFIG_X86_VISWS) += setup-irq.o
obj-$(CONFIG_MN10300) += setup-bus.o
obj-$(CONFIG_MICROBLAZE) += setup-bus.o
+obj-$(CONFIG_TILE) += setup-bus.o setup-irq.o

#
# ACPI Related PCI FW Functions
--- linux-2.6.34/drivers/pci/quirks.c 2010-05-16 17:17:36.000000000 -0400
+++ tilera-source/drivers/pci/quirks.c 2010-05-17 13:26:22.347178000 -0400
@@ -2094,6 +2094,23 @@
quirk_unhide_mch_dev6);


+/*
+ * The Tilera Blade V1.0 platform needs to set the link speed
+ * to 2.5GT(Giga-Transfers)/s (Gen 1). The default link speed
+ * setting is 5GT/s (Gen 2). 0x98 is the Link Control2 PCIe
+ * capability register of the PEX8624 PCIe switch. The switch
+ * supports link speed auto negotiation. This is expected to
+ * be fixed in the next release of the Blade platform.
+ */
+static void __devinit quirk_tile_blade(struct pci_dev *dev)
+{
+ if (blade_pci) {
+ pci_write_config_dword(dev, 0x98, 0x1);
+ mdelay(50);
+ }
+}
+DECLARE_PCI_FIXUP_EARLY(PCI_VENDOR_ID_PLX, 0x8624, quirk_tile_blade);
+
#ifdef CONFIG_PCI_MSI
/* Some chipsets do not support MSI. We cannot easily rely on setting
* PCI_BUS_FLAGS_NO_MSI in its bus flags because there are actually

\
 
 \ /
  Last update: 2010-05-20 08:05    [from the cache]
©2003-2011 Jasper Spaans