arch/tile: Enable more sophisticated IRQ model for 32-bit chips.

This model is based on the on-chip interrupt model used by the
TILE-Gx next-generation hardware, and interacts much more cleanly
with the Linux generic IRQ layer.

The change includes modifications to the Tilera hypervisor, which
are reflected in the hypervisor headers in arch/tile/include/arch/.

Signed-off-by: Chris Metcalf <cmetcalf@tilera.com>
Acked-by: Thomas Gleixner <tglx@linutronix.de>

6 files changed, 181 insertions, 63 deletions

diff --git a/arch/tile/include/arch/chip_tile64.h b/arch/tile/include/arch/chip_tile64.h
index 18b5bc8e563f..1246573be59e 100644
--- a/arch/tile/include/arch/chip_tile64.h
+++ b/arch/tile/include/arch/chip_tile64.h
@@ -248,5 +248,8 @@
 /** Does the chip support rev1 DMA packets? */
 #define CHIP_HAS_REV1_DMA_PACKETS() 0
 
+/** Does the chip have an IPI shim? */
+#define CHIP_HAS_IPI() 0
+
 #endif /* !__OPEN_SOURCE__ */
 #endif /* __ARCH_CHIP_H__ */
diff --git a/arch/tile/include/arch/chip_tilepro.h b/arch/tile/include/arch/chip_tilepro.h
index 9852af163862..e864c47fc89c 100644
--- a/arch/tile/include/arch/chip_tilepro.h
+++ b/arch/tile/include/arch/chip_tilepro.h
@@ -248,5 +248,8 @@
 /** Does the chip support rev1 DMA packets? */
 #define CHIP_HAS_REV1_DMA_PACKETS() 1
 
+/** Does the chip have an IPI shim? */
+#define CHIP_HAS_IPI() 0
+
 #endif /* !__OPEN_SOURCE__ */
 #endif /* __ARCH_CHIP_H__ */
diff --git a/arch/tile/include/asm/irq.h b/arch/tile/include/asm/irq.h
index 9be1f849fac9..572fd3ef1d73 100644
--- a/arch/tile/include/asm/irq.h
+++ b/arch/tile/include/asm/irq.h
@@ -23,15 +23,65 @@
 /* IRQ numbers used for linux IPIs. */
 #define IRQ_RESCHEDULE 1
 
-/* The HV interrupt state object. */
-DECLARE_PER_CPU(HV_IntrState, dev_intr_state);
-
 void ack_bad_irq(unsigned int irq);
 
 /*
- * Paravirtualized drivers should call this when their init calls
- * discover a valid HV IRQ.
+ * Different ways of handling interrupts.  Tile interrupts are always
+ * per-cpu; there is no global interrupt controller to implement
+ * enable/disable.  Most onboard devices can send their interrupts to
+ * many tiles at the same time, and Tile-specific drivers know how to
+ * deal with this.
+ *
+ * However, generic devices (usually PCIE based, sometimes GPIO)
+ * expect that interrupts will fire on a single core at a time and
+ * that the irq can be enabled or disabled from any core at any time.
+ * We implement this by directing such interrupts to a single core.
+ *
+ * One added wrinkle is that PCI interrupts can be either
+ * hardware-cleared (legacy interrupts) or software cleared (MSI).
+ * Other generic device systems (GPIO) are always software-cleared.
+ *
+ * The enums below are used by drivers for onboard devices, including
+ * the internals of PCI root complex and GPIO.  They allow the driver
+ * to tell the generic irq code what kind of interrupt is mapped to a
+ * particular IRQ number.
+ */
+enum {
+	/* per-cpu interrupt; use enable/disable_percpu_irq() to mask */
+	TILE_IRQ_PERCPU,
+	/* global interrupt, hardware responsible for clearing. */
+	TILE_IRQ_HW_CLEAR,
+	/* global interrupt, software responsible for clearing. */
+	TILE_IRQ_SW_CLEAR,
+};
+
+
+/*
+ * Paravirtualized drivers should call this when they dynamically
+ * allocate a new IRQ or discover an IRQ that was pre-allocated by the
+ * hypervisor for use with their particular device.  This gives the
+ * IRQ subsystem an opportunity to do interrupt-type-specific
+ * initialization.
+ *
+ * ISSUE: We should modify this API so that registering anything
+ * except percpu interrupts also requires providing callback methods
+ * for enabling and disabling the interrupt.  This would allow the
+ * generic IRQ code to proxy enable/disable_irq() calls back into the
+ * PCI subsystem, which in turn could enable or disable the interrupt
+ * at the PCI shim.
  */
-void tile_irq_activate(unsigned int irq);
+void tile_irq_activate(unsigned int irq, int tile_irq_type);
+
+/*
+ * For onboard, non-PCI (e.g. TILE_IRQ_PERCPU) devices, drivers know
+ * how to use enable/disable_percpu_irq() to manage interrupts on each
+ * core.  We can't use the generic enable/disable_irq() because they
+ * use a single reference count per irq, rather than per cpu per irq.
+ */
+void enable_percpu_irq(unsigned int irq);
+void disable_percpu_irq(unsigned int irq);
+
+
+void setup_irq_regs(void);
 
 #endif /* _ASM_TILE_IRQ_H */
diff --git a/arch/tile/include/asm/smp.h b/arch/tile/include/asm/smp.h
index da24858a7392..532124ae4b12 100644
--- a/arch/tile/include/asm/smp.h
+++ b/arch/tile/include/asm/smp.h
@@ -20,6 +20,7 @@
 #include <asm/processor.h>
 #include <linux/cpumask.h>
 #include <linux/irqreturn.h>
+#include <hv/hypervisor.h>
 
 /* Set up this tile to support receiving hypervisor messages */
 void init_messaging(void);
@@ -39,9 +40,6 @@ void send_IPI_single(int dest, int tag);
 /* Process an IPI message */
 void evaluate_message(int tag);
 
-/* Process an IRQ_RESCHEDULE IPI. */
-irqreturn_t handle_reschedule_ipi(int irq, void *token);
-
 /* Boot a secondary cpu */
 void online_secondary(void);
 
@@ -56,6 +54,20 @@ extern HV_Topology smp_topology;
 #define smp_height		(smp_topology.height)
 #define smp_width		(smp_topology.width)
 
+/* Convenience functions for converting cpu <-> coords. */
+static inline int cpu_x(int cpu)
+{
+	return cpu % smp_width;
+}
+static inline int cpu_y(int cpu)
+{
+	return cpu / smp_width;
+}
+static inline int xy_to_cpu(int x, int y)
+{
+	return y * smp_width + x;
+}
+
 /* Hypervisor message tags sent via the tile send_IPI*() routines. */
 #define MSG_TAG_START_CPU		1
 #define MSG_TAG_STOP_CPU		2
@@ -85,6 +97,9 @@ void print_disabled_cpus(void);
 #define smp_master_cpu		0
 #define smp_height		1
 #define smp_width		1
+#define cpu_x(cpu)		0
+#define cpu_y(cpu)		0
+#define xy_to_cpu(x, y)		0
 
 #endif /* !CONFIG_SMP */
 
@@ -123,4 +138,10 @@ static inline int __cpulist_parse_crop(const char *buf, struct cpumask *dstp,
 	return bitmap_parselist_crop(buf, cpumask_bits(dstp), nbits);
 }
 
+/* Initialize the IPI subsystem. */
+void ipi_init(void);
+
+/* Function for start-cpu message to cause us to jump to. */
+extern unsigned long start_cpu_function_addr;
+
 #endif /* _ASM_TILE_SMP_H */
diff --git a/arch/tile/include/hv/hypervisor.h b/arch/tile/include/hv/hypervisor.h
index 84b31551080a..a90d2989587a 100644
--- a/arch/tile/include/hv/hypervisor.h
+++ b/arch/tile/include/hv/hypervisor.h
@@ -20,12 +20,9 @@
 #ifndef _TILE_HV_H
 #define _TILE_HV_H
 
-#ifdef __tile__
 #include <arch/chip.h>
-#else
-/* HACK: Allow use by "tools/cpack/". */
-#include "install/include/arch/chip.h"
-#endif
+
+#include <hv/pagesize.h>
 
 /* Linux builds want unsigned long constants, but assembler wants numbers */
 #ifdef __ASSEMBLER__
@@ -39,7 +36,6 @@
 #define __HV_SIZE_ONE 1UL
 #endif
 
-
 /** The log2 of the span of a level-1 page table, in bytes.
  */
 #define HV_LOG2_L1_SPAN 32
@@ -48,21 +44,11 @@
  */
 #define HV_L1_SPAN (__HV_SIZE_ONE << HV_LOG2_L1_SPAN)
 
-/** The log2 of the size of small pages, in bytes. This value should
- * be verified at runtime by calling hv_sysconf(HV_SYSCONF_PAGE_SIZE_SMALL).
- */
-#define HV_LOG2_PAGE_SIZE_SMALL 16
-
 /** The size of small pages, in bytes. This value should be verified
  * at runtime by calling hv_sysconf(HV_SYSCONF_PAGE_SIZE_SMALL).
  */
 #define HV_PAGE_SIZE_SMALL (__HV_SIZE_ONE << HV_LOG2_PAGE_SIZE_SMALL)
 
-/** The log2 of the size of large pages, in bytes. This value should be
- * verified at runtime by calling hv_sysconf(HV_SYSCONF_PAGE_SIZE_LARGE).
- */
-#define HV_LOG2_PAGE_SIZE_LARGE 24
-
 /** The size of large pages, in bytes. This value should be verified
  * at runtime by calling hv_sysconf(HV_SYSCONF_PAGE_SIZE_LARGE).
  */
@@ -93,7 +79,7 @@
 #define HV_DISPATCH_ENTRY_SIZE 32
 
 /** Version of the hypervisor interface defined by this file */
-#define _HV_VERSION 10
+#define _HV_VERSION 11
 
 /* Index into hypervisor interface dispatch code blocks.
  *
@@ -253,8 +239,10 @@
 /** hv_set_command_line */
 #define HV_DISPATCH_SET_COMMAND_LINE              47
 
-/** hv_dev_register_intr_state */
-#define HV_DISPATCH_DEV_REGISTER_INTR_STATE       48
+#if !CHIP_HAS_IPI()
+
+/** hv_clear_intr */
+#define HV_DISPATCH_CLEAR_INTR                    48
 
 /** hv_enable_intr */
 #define HV_DISPATCH_ENABLE_INTR                   49
@@ -262,20 +250,30 @@
 /** hv_disable_intr */
 #define HV_DISPATCH_DISABLE_INTR                  50
 
+/** hv_raise_intr */
+#define HV_DISPATCH_RAISE_INTR                    51
+
 /** hv_trigger_ipi */
-#define HV_DISPATCH_TRIGGER_IPI                   51
+#define HV_DISPATCH_TRIGGER_IPI                   52
+
+#endif /* !CHIP_HAS_IPI() */
 
 /** hv_store_mapping */
-#define HV_DISPATCH_STORE_MAPPING                 52
+#define HV_DISPATCH_STORE_MAPPING                 53
 
 /** hv_inquire_realpa */
-#define HV_DISPATCH_INQUIRE_REALPA                53
+#define HV_DISPATCH_INQUIRE_REALPA                54
 
 /** hv_flush_all */
-#define HV_DISPATCH_FLUSH_ALL                     54
+#define HV_DISPATCH_FLUSH_ALL                     55
+
+#if CHIP_HAS_IPI()
+/** hv_get_ipi_pte */
+#define HV_DISPATCH_GET_IPI_PTE                   56
+#endif
 
 /** One more than the largest dispatch value */
-#define _HV_DISPATCH_END                          55
+#define _HV_DISPATCH_END                          57
 
 
 #ifndef __ASSEMBLER__
@@ -484,21 +482,6 @@ typedef enum {
  */
 int hv_confstr(HV_ConfstrQuery query, HV_VirtAddr buf, int len);
 
-/** State object used to enable and disable one-shot and level-sensitive
- *  interrupts. */
-typedef struct
-{
-#if CHIP_VA_WIDTH() > 32
-  __hv64 opaque[2]; /**< No user-serviceable parts inside */
-#else
-  __hv32 opaque[2]; /**< No user-serviceable parts inside */
-#endif
-}
-HV_IntrState;
-
-/** A set of interrupts. */
-typedef __hv32 HV_IntrMask;
-
 /** Tile coordinate */
 typedef struct
 {
@@ -509,34 +492,51 @@ typedef struct
   int y;
 } HV_Coord;
 
+
+#if CHIP_HAS_IPI()
+
+/** Get the PTE for sending an IPI to a particular tile.
+ *
+ * @param tile Tile which will receive the IPI.
+ * @param pl Indicates which IPI registers: 0 = IPI_0, 1 = IPI_1.
+ * @param pte Filled with resulting PTE.
+ * @result Zero if no error, non-zero for invalid parameters.
+ */
+int hv_get_ipi_pte(HV_Coord tile, int pl, HV_PTE* pte);
+
+#else /* !CHIP_HAS_IPI() */
+
+/** A set of interrupts. */
+typedef __hv32 HV_IntrMask;
+
 /** The low interrupt numbers are reserved for use by the client in
  *  delivering IPIs.  Any interrupt numbers higher than this value are
  *  reserved for use by HV device drivers. */
 #define HV_MAX_IPI_INTERRUPT 7
 
-/** Register an interrupt state object.  This object is used to enable and
- *  disable one-shot and level-sensitive interrupts.  Once the state is
- *  registered, the client must not read or write the state object; doing
- *  so will cause undefined results.
+/** Enable a set of device interrupts.
  *
- * @param intr_state Pointer to interrupt state object.
- * @return HV_OK on success, or a hypervisor error code.
+ * @param enab_mask Bitmap of interrupts to enable.
  */
-HV_Errno hv_dev_register_intr_state(HV_IntrState* intr_state);
+void hv_enable_intr(HV_IntrMask enab_mask);
 
-/** Enable a set of one-shot and level-sensitive interrupts.
+/** Disable a set of device interrupts.
  *
- * @param intr_state Pointer to interrupt state object.
- * @param enab_mask Bitmap of interrupts to enable.
+ * @param disab_mask Bitmap of interrupts to disable.
  */
-void hv_enable_intr(HV_IntrState* intr_state, HV_IntrMask enab_mask);
+void hv_disable_intr(HV_IntrMask disab_mask);
 
-/** Disable a set of one-shot and level-sensitive interrupts.
+/** Clear a set of device interrupts.
  *
- * @param intr_state Pointer to interrupt state object.
- * @param disab_mask Bitmap of interrupts to disable.
+ * @param clear_mask Bitmap of interrupts to clear.
  */
-void hv_disable_intr(HV_IntrState* intr_state, HV_IntrMask disab_mask);
+void hv_clear_intr(HV_IntrMask clear_mask);
+
+/** Assert a set of device interrupts.
+ *
+ * @param assert_mask Bitmap of interrupts to clear.
+ */
+void hv_assert_intr(HV_IntrMask assert_mask);
 
 /** Trigger a one-shot interrupt on some tile
  *
@@ -547,6 +547,8 @@ void hv_disable_intr(HV_IntrState* intr_state, HV_IntrMask disab_mask);
  */
 HV_Errno hv_trigger_ipi(HV_Coord tile, int interrupt);
 
+#endif // !CHIP_HAS_IPI()
+
 /** Store memory mapping in debug memory so that external debugger can read it.
  * A maximum of 16 entries can be stored.
  *
@@ -1010,6 +1012,13 @@ int hv_console_write(HV_VirtAddr bytes, int len);
  *  it will return to the code which was interrupted by the INTCTRL_1
  *  interrupt.
  *
+ *  Under some circumstances, the firing of INTCTRL_1 can race with
+ *  the lowering of a device interrupt.  In such a case, the
+ *  hv_downcall_dispatch service may issue an iret instruction instead
+ *  of entering one of the client's actual downcall-handling interrupt
+ *  vectors.  This will return execution to the location that was
+ *  interrupted by INTCTRL_1.
+ *
  *  Any saving of registers should be done by the actual handling
  *  vectors; no registers should be changed by the INTCTRL_1 handler.
  *  In particular, the client should not use a jal instruction to invoke
diff --git a/arch/tile/include/hv/pagesize.h b/arch/tile/include/hv/pagesize.h
new file mode 100644
index 000000000000..58bed114fedd
--- /dev/null
+++ b/arch/tile/include/hv/pagesize.h
@@ -0,0 +1,32 @@
+/*
+ * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   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, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+/**
+ * @file pagesize.h
+ */
+
+#ifndef _HV_PAGESIZE_H
+#define _HV_PAGESIZE_H
+
+/** The log2 of the size of small pages, in bytes. This value should
+ * be verified at runtime by calling hv_sysconf(HV_SYSCONF_PAGE_SIZE_SMALL).
+ */
+#define HV_LOG2_PAGE_SIZE_SMALL 16
+
+/** The log2 of the size of large pages, in bytes. This value should be
+ * verified at runtime by calling hv_sysconf(HV_SYSCONF_PAGE_SIZE_LARGE).
+ */
+#define HV_LOG2_PAGE_SIZE_LARGE 24
+
+#endif /* _HV_PAGESIZE_H */