Merge git://git.kernel.org/pub/scm/linux/kernel/git/cmetcalf/linux-tile

* git://git.kernel.org/pub/scm/linux/kernel/git/cmetcalf/linux-tile: (27 commits)
  arch/tile: support newer binutils assembler shift semantics
  arch/tile: fix deadlock bugs in rwlock implementation
  drivers/edac: provide support for tile architecture
  tile on-chip network driver: sync up with latest fixes
  arch/tile: support 4KB page size as well as 64KB
  arch/tile: add some more VMSPLIT options and use consistent naming
  arch/tile: fix some comments and whitespace
  arch/tile: export some additional module symbols
  arch/tile: enhance existing finv_buffer_remote() routine
  arch/tile: fix two bugs in the backtracer code
  arch/tile: use extended assembly to inline __mb_incoherent()
  arch/tile: use a cleaner technique to enable interrupt for cpu_idle()
  arch/tile: sync up with <arch/sim.h> and <arch/sim_def.h> changes
  arch/tile: fix reversed test of strict_strtol() return value
  arch/tile: avoid a simulator warning during bootup
  arch/tile: export <asm/hardwall.h> to userspace
  arch/tile: warn and retry if an IPI is not accepted by the target cpu
  arch/tile: stop disabling INTCTRL_1 interrupts during hypervisor downcalls
  arch/tile: fix __ndelay etc to work better
  arch/tile: bug fix: exec'ed task thought it was still single-stepping
  ...

Fix up trivial conflict in arch/tile/kernel/vmlinux.lds.S (percpu
alignment vs section naming convention fix)

MAINTAINERS

@@ -6127,6 +6127,7 @@ S:	Supported
 F:	arch/tile/
 F:	drivers/tty/hvc/hvc_tile.c
 F:	drivers/net/tile/
+F:	drivers/edac/tile_edac.c
 
 TLAN NETWORK DRIVER
 M:	Samuel Chessman <chessman@tux.org>

README

@@ -24,7 +24,7 @@ ON WHAT HARDWARE DOES IT RUN?
   today Linux also runs on (at least) the Compaq Alpha AXP, Sun SPARC and
   UltraSPARC, Motorola 68000, PowerPC, PowerPC64, ARM, Hitachi SuperH, Cell,
   IBM S/390, MIPS, HP PA-RISC, Intel IA-64, DEC VAX, AMD x86-64, AXIS CRIS,
-  Xtensa, AVR32 and Renesas M32R architectures.
+  Xtensa, Tilera TILE, AVR32 and Renesas M32R architectures.
 
   Linux is easily portable to most general-purpose 32- or 64-bit architectures
   as long as they have a paged memory management unit (PMMU) and a port of the

arch/tile/Kconfig

@@ -1,5 +1,5 @@
 # For a description of the syntax of this configuration file,
-# see Documentation/kbuild/config-language.txt.
+# see Documentation/kbuild/kconfig-language.txt.
 
 config TILE
 	def_bool y
@@ -11,17 +11,18 @@ config TILE
 	select HAVE_GENERIC_HARDIRQS
 	select GENERIC_IRQ_PROBE
 	select GENERIC_PENDING_IRQ if SMP
+	select GENERIC_HARDIRQS_NO_DEPRECATED
 
 # FIXME: investigate whether we need/want these options.
 #	select HAVE_IOREMAP_PROT
-#       select HAVE_OPTPROBES
-#       select HAVE_REGS_AND_STACK_ACCESS_API
-#       select HAVE_HW_BREAKPOINT
-#       select PERF_EVENTS
-#       select HAVE_USER_RETURN_NOTIFIER
-#       config NO_BOOTMEM
-#       config ARCH_SUPPORTS_DEBUG_PAGEALLOC
-#       config HUGETLB_PAGE_SIZE_VARIABLE
+#	select HAVE_OPTPROBES
+#	select HAVE_REGS_AND_STACK_ACCESS_API
+#	select HAVE_HW_BREAKPOINT
+#	select PERF_EVENTS
+#	select HAVE_USER_RETURN_NOTIFIER
+#	config NO_BOOTMEM
+#	config ARCH_SUPPORTS_DEBUG_PAGEALLOC
+#	config HUGETLB_PAGE_SIZE_VARIABLE
 
 config MMU
 	def_bool y
@@ -39,7 +40,7 @@ config HAVE_SETUP_PER_CPU_AREA
 	def_bool y
 
 config NEED_PER_CPU_PAGE_FIRST_CHUNK
-        def_bool y
+	def_bool y
 
 config SYS_SUPPORTS_HUGETLBFS
 	def_bool y
@@ -201,12 +202,6 @@ config NODES_SHIFT
 	  By default, 2, i.e. 2^2 == 4 DDR2 controllers.
 	  In a system with more controllers, this value should be raised.
 
-# Need 16MB areas to enable hugetlb
-# See build-time check in arch/tile/mm/init.c.
-config FORCE_MAX_ZONEORDER
-	int
-	default 9
-
 choice
 	depends on !TILEGX
 	prompt "Memory split" if EXPERT
@@ -233,8 +228,12 @@ choice
 		bool "3.5G/0.5G user/kernel split"
 	config VMSPLIT_3G
 		bool "3G/1G user/kernel split"
-	config VMSPLIT_3G_OPT
-		bool "3G/1G user/kernel split (for full 1G low memory)"
+	config VMSPLIT_2_75G
+		bool "2.75G/1.25G user/kernel split (for full 1G low memory)"
+	config VMSPLIT_2_5G
+		bool "2.5G/1.5G user/kernel split"
+	config VMSPLIT_2_25G
+		bool "2.25G/1.75G user/kernel split"
 	config VMSPLIT_2G
 		bool "2G/2G user/kernel split"
 	config VMSPLIT_1G
@@ -245,7 +244,9 @@ config PAGE_OFFSET
 	hex
 	default 0xF0000000 if VMSPLIT_3_75G
 	default 0xE0000000 if VMSPLIT_3_5G
-	default 0xB0000000 if VMSPLIT_3G_OPT
+	default 0xB0000000 if VMSPLIT_2_75G
+	default 0xA0000000 if VMSPLIT_2_5G
+	default 0x90000000 if VMSPLIT_2_25G
 	default 0x80000000 if VMSPLIT_2G
 	default 0x40000000 if VMSPLIT_1G
 	default 0xC0000000

arch/tile/include/arch/interrupts_32.h

@@ -16,10 +16,11 @@
 #define __ARCH_INTERRUPTS_H__
 
 /** Mask for an interrupt. */
-#ifdef __ASSEMBLER__
 /* Note: must handle breaking interrupts into high and low words manually. */
-#define INT_MASK(intno) (1 << (intno))
-#else
+#define INT_MASK_LO(intno) (1 << (intno))
+#define INT_MASK_HI(intno) (1 << ((intno) - 32))
+
+#ifndef __ASSEMBLER__
 #define INT_MASK(intno) (1ULL << (intno))
 #endif
 
@@ -89,6 +90,7 @@
 
 #define NUM_INTERRUPTS 49
 
+#ifndef __ASSEMBLER__
 #define QUEUED_INTERRUPTS ( \
     INT_MASK(INT_MEM_ERROR) | \
     INT_MASK(INT_DMATLB_MISS) | \
@@ -301,4 +303,5 @@
     INT_MASK(INT_DOUBLE_FAULT) | \
     INT_MASK(INT_AUX_PERF_COUNT) | \
     0)
+#endif /* !__ASSEMBLER__ */
 #endif /* !__ARCH_INTERRUPTS_H__ */

arch/tile/include/arch/sim.h

@@ -152,16 +152,33 @@ sim_dump(unsigned int mask)
 /**
  * Print a string to the simulator stdout.
  *
- * @param str The string to be written; a newline is automatically added.
+ * @param str The string to be written.
+ */
+static __inline void
+sim_print(const char* str)
+{
+  for ( ; *str != '\0'; str++)
+  {
+    __insn_mtspr(SPR_SIM_CONTROL, SIM_CONTROL_PUTC |
+                 (*str << _SIM_CONTROL_OPERATOR_BITS));
+  }
+  __insn_mtspr(SPR_SIM_CONTROL, SIM_CONTROL_PUTC |
+               (SIM_PUTC_FLUSH_BINARY << _SIM_CONTROL_OPERATOR_BITS));
+}
+
+
+/**
+ * Print a string to the simulator stdout.
+ *
+ * @param str The string to be written (a newline is automatically added).
  */
 static __inline void
 sim_print_string(const char* str)
 {
-  int i;
-  for (i = 0; str[i] != 0; i++)
+  for ( ; *str != '\0'; str++)
   {
     __insn_mtspr(SPR_SIM_CONTROL, SIM_CONTROL_PUTC |
-                 (str[i] << _SIM_CONTROL_OPERATOR_BITS));
+                 (*str << _SIM_CONTROL_OPERATOR_BITS));
   }
   __insn_mtspr(SPR_SIM_CONTROL, SIM_CONTROL_PUTC |
                (SIM_PUTC_FLUSH_STRING << _SIM_CONTROL_OPERATOR_BITS));
@@ -203,7 +220,7 @@ sim_command(const char* str)
  * we are passing to the simulator are actually valid in the registers
  * (i.e. returned from memory) prior to the SIM_CONTROL spr.
  */
-static __inline int _sim_syscall0(int val)
+static __inline long _sim_syscall0(int val)
 {
   long result;
   __asm__ __volatile__ ("mtspr SIM_CONTROL, r0"
@@ -211,7 +228,7 @@ static __inline int _sim_syscall0(int val)
   return result;
 }
 
-static __inline int _sim_syscall1(int val, long arg1)
+static __inline long _sim_syscall1(int val, long arg1)
 {
   long result;
   __asm__ __volatile__ ("{ and zero, r1, r1; mtspr SIM_CONTROL, r0 }"
@@ -219,7 +236,7 @@ static __inline int _sim_syscall1(int val, long arg1)
   return result;
 }
 
-static __inline int _sim_syscall2(int val, long arg1, long arg2)
+static __inline long _sim_syscall2(int val, long arg1, long arg2)
 {
   long result;
   __asm__ __volatile__ ("{ and zero, r1, r2; mtspr SIM_CONTROL, r0 }"
@@ -233,7 +250,7 @@ static __inline int _sim_syscall2(int val, long arg1, long arg2)
    the register values for arguments 3 and up may still be in flight
    to the core from a stack frame reload. */
 
-static __inline int _sim_syscall3(int val, long arg1, long arg2, long arg3)
+static __inline long _sim_syscall3(int val, long arg1, long arg2, long arg3)
 {
   long result;
   __asm__ __volatile__ ("{ and zero, r3, r3 };"
@@ -244,7 +261,7 @@ static __inline int _sim_syscall3(int val, long arg1, long arg2, long arg3)
   return result;
 }
 
-static __inline int _sim_syscall4(int val, long arg1, long arg2, long arg3,
+static __inline long _sim_syscall4(int val, long arg1, long arg2, long arg3,
                                   long arg4)
 {
   long result;
@@ -256,7 +273,7 @@ static __inline int _sim_syscall4(int val, long arg1, long arg2, long arg3,
   return result;
 }
 
-static __inline int _sim_syscall5(int val, long arg1, long arg2, long arg3,
+static __inline long _sim_syscall5(int val, long arg1, long arg2, long arg3,
                                   long arg4, long arg5)
 {
   long result;
@@ -268,7 +285,6 @@ static __inline int _sim_syscall5(int val, long arg1, long arg2, long arg3,
   return result;
 }
 
-
 /**
  * Make a special syscall to the simulator itself, if running under
  * simulation. This is used as the implementation of other functions
@@ -281,7 +297,8 @@ static __inline int _sim_syscall5(int val, long arg1, long arg2, long arg3,
  */
 #define _sim_syscall(syscall_num, nr, args...) \
   _sim_syscall##nr( \
-    ((syscall_num) << _SIM_CONTROL_OPERATOR_BITS) | SIM_CONTROL_SYSCALL, args)
+    ((syscall_num) << _SIM_CONTROL_OPERATOR_BITS) | SIM_CONTROL_SYSCALL, \
+    ##args)
 
 
 /* Values for the "access_mask" parameters below. */
@@ -365,6 +382,13 @@ sim_validate_lines_evicted(unsigned long long pa, unsigned long length)
 }
 
 
+/* Return the current CPU speed in cycles per second. */
+static __inline long
+sim_query_cpu_speed(void)
+{
+  return _sim_syscall(SIM_SYSCALL_QUERY_CPU_SPEED, 0);
+}
+
 #endif /* !__DOXYGEN__ */
 
 

arch/tile/include/arch/sim_def.h

@@ -243,6 +243,9 @@
  */
 #define SIM_SYSCALL_VALIDATE_LINES_EVICTED 5
 
+/** Syscall number for sim_query_cpu_speed(). */
+#define SIM_SYSCALL_QUERY_CPU_SPEED 6
+
 
 /*
  * Bit masks which can be shifted by 8, combined with

arch/tile/include/asm/Kbuild

@@ -1,3 +1,4 @@
 include include/asm-generic/Kbuild.asm
 
 header-y += ucontext.h
+header-y += hardwall.h

arch/tile/include/asm/atomic.h

@@ -32,7 +32,7 @@
  */
 static inline int atomic_read(const atomic_t *v)
 {
-       return v->counter;
+	return ACCESS_ONCE(v->counter);
 }
 
 /**

arch/tile/include/asm/bitops_32.h

@@ -122,7 +122,7 @@ static inline int test_and_change_bit(unsigned nr,
 	return (_atomic_xor(addr, mask) & mask) != 0;
 }
 
-/* See discussion at smp_mb__before_atomic_dec() in <asm/atomic.h>. */
+/* See discussion at smp_mb__before_atomic_dec() in <asm/atomic_32.h>. */
 #define smp_mb__before_clear_bit()	smp_mb()
 #define smp_mb__after_clear_bit()	do {} while (0)
 

arch/tile/include/asm/cache.h

@@ -40,7 +40,7 @@
 #define INTERNODE_CACHE_BYTES   L2_CACHE_BYTES
 
 /* Group together read-mostly things to avoid cache false sharing */
-#define __read_mostly __attribute__((__section__(".data.read_mostly")))
+#define __read_mostly __attribute__((__section__(".data..read_mostly")))
 
 /*
  * Attribute for data that is kept read/write coherent until the end of

arch/tile/include/asm/cacheflush.h

@@ -138,55 +138,12 @@ static inline void finv_buffer(void *buffer, size_t size)
 }
 
 /*
- * Flush & invalidate a VA range that is homed remotely on a single core,
- * waiting until the memory controller holds the flushed values.
+ * Flush and invalidate a VA range that is homed remotely, waiting
+ * until the memory controller holds the flushed values.  If "hfh" is
+ * true, we will do a more expensive flush involving additional loads
+ * to make sure we have touched all the possible home cpus of a buffer
+ * that is homed with "hash for home".
  */
-static inline void finv_buffer_remote(void *buffer, size_t size)
-{
-	char *p;
-	int i;
-
-	/*
-	 * Flush and invalidate the buffer out of the local L1/L2
-	 * and request the home cache to flush and invalidate as well.
-	 */
-	__finv_buffer(buffer, size);
-
-	/*
-	 * Wait for the home cache to acknowledge that it has processed
-	 * all the flush-and-invalidate requests.  This does not mean
-	 * that the flushed data has reached the memory controller yet,
-	 * but it does mean the home cache is processing the flushes.
-	 */
-	__insn_mf();
-
-	/*
-	 * Issue a load to the last cache line, which can't complete
-	 * until all the previously-issued flushes to the same memory
-	 * controller have also completed.  If we weren't striping
-	 * memory, that one load would be sufficient, but since we may
-	 * be, we also need to back up to the last load issued to
-	 * another memory controller, which would be the point where
-	 * we crossed an 8KB boundary (the granularity of striping
-	 * across memory controllers).  Keep backing up and doing this
-	 * until we are before the beginning of the buffer, or have
-	 * hit all the controllers.
-	 */
-	for (i = 0, p = (char *)buffer + size - 1;
-	     i < (1 << CHIP_LOG_NUM_MSHIMS()) && p >= (char *)buffer;
-	     ++i) {
-		const unsigned long STRIPE_WIDTH = 8192;
-
-		/* Force a load instruction to issue. */
-		*(volatile char *)p;
-
-		/* Jump to end of previous stripe. */
-		p -= STRIPE_WIDTH;
-		p = (char *)((unsigned long)p | (STRIPE_WIDTH - 1));
-	}
-
-	/* Wait for the loads (and thus flushes) to have completed. */
-	__insn_mf();
-}
+void finv_buffer_remote(void *buffer, size_t size, int hfh);
 
 #endif /* _ASM_TILE_CACHEFLUSH_H */

arch/tile/include/asm/edac.h

@@ -0,0 +1,29 @@
+/*
+ * Copyright 2011 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.
+ */
+
+#ifndef _ASM_TILE_EDAC_H
+#define _ASM_TILE_EDAC_H
+
+/* ECC atomic, DMA, SMP and interrupt safe scrub function */
+
+static inline void atomic_scrub(void *va, u32 size)
+{
+	/*
+	 * These is nothing to be done here because CE is
+	 * corrected by the mshim.
+	 */
+	return;
+}
+
+#endif /* _ASM_TILE_EDAC_H */

arch/tile/include/asm/hugetlb.h

@@ -54,7 +54,7 @@ static inline void hugetlb_free_pgd_range(struct mmu_gather *tlb,
 static inline void set_huge_pte_at(struct mm_struct *mm, unsigned long addr,
 				   pte_t *ptep, pte_t pte)
 {
-	set_pte_order(ptep, pte, HUGETLB_PAGE_ORDER);
+	set_pte(ptep, pte);
 }
 
 static inline pte_t huge_ptep_get_and_clear(struct mm_struct *mm,

arch/tile/include/asm/irqflags.h

@@ -18,12 +18,24 @@
 #include <arch/interrupts.h>
 #include <arch/chip.h>
 
+#if !defined(__tilegx__) && defined(__ASSEMBLY__)
+
 /*
  * The set of interrupts we want to allow when interrupts are nominally
  * disabled.  The remainder are effectively "NMI" interrupts from
  * the point of view of the generic Linux code.  Note that synchronous
  * interrupts (aka "non-queued") are not blocked by the mask in any case.
  */
+#if CHIP_HAS_AUX_PERF_COUNTERS()
+#define LINUX_MASKABLE_INTERRUPTS_HI \
+       (~(INT_MASK_HI(INT_PERF_COUNT) | INT_MASK_HI(INT_AUX_PERF_COUNT)))
+#else
+#define LINUX_MASKABLE_INTERRUPTS_HI \
+       (~(INT_MASK_HI(INT_PERF_COUNT)))
+#endif
+
+#else
+
 #if CHIP_HAS_AUX_PERF_COUNTERS()
 #define LINUX_MASKABLE_INTERRUPTS \
 	(~(INT_MASK(INT_PERF_COUNT) | INT_MASK(INT_AUX_PERF_COUNT)))
@@ -32,6 +44,8 @@
 	(~(INT_MASK(INT_PERF_COUNT)))
 #endif
 
+#endif
+
 #ifndef __ASSEMBLY__
 
 /* NOTE: we can't include <linux/percpu.h> due to #include dependencies. */
@@ -224,11 +238,11 @@ DECLARE_PER_CPU(unsigned long long, interrupts_enabled_mask);
 #define IRQ_DISABLE(tmp0, tmp1)					\
 	{							\
 	 movei  tmp0, -1;					\
-	 moveli tmp1, lo16(LINUX_MASKABLE_INTERRUPTS)		\
+	 moveli tmp1, lo16(LINUX_MASKABLE_INTERRUPTS_HI)	\
 	};							\
 	{							\
 	 mtspr  SPR_INTERRUPT_MASK_SET_K_0, tmp0;		\
-	 auli   tmp1, tmp1, ha16(LINUX_MASKABLE_INTERRUPTS)	\
+	 auli   tmp1, tmp1, ha16(LINUX_MASKABLE_INTERRUPTS_HI)	\
 	};							\
 	mtspr   SPR_INTERRUPT_MASK_SET_K_1, tmp1
 

arch/tile/include/asm/page.h

@@ -16,10 +16,11 @@
 #define _ASM_TILE_PAGE_H
 
 #include <linux/const.h>
+#include <hv/pagesize.h>
 
 /* PAGE_SHIFT and HPAGE_SHIFT determine the page sizes. */
-#define PAGE_SHIFT	16
-#define HPAGE_SHIFT	24
+#define PAGE_SHIFT	HV_LOG2_PAGE_SIZE_SMALL
+#define HPAGE_SHIFT	HV_LOG2_PAGE_SIZE_LARGE
 
 #define PAGE_SIZE	(_AC(1, UL) << PAGE_SHIFT)
 #define HPAGE_SIZE	(_AC(1, UL) << HPAGE_SHIFT)
@@ -29,25 +30,18 @@
 
 #ifdef __KERNEL__
 
-#include <hv/hypervisor.h>
-#include <arch/chip.h>
-
 /*
- * The {,H}PAGE_SHIFT values must match the HV_LOG2_PAGE_SIZE_xxx
- * definitions in <hv/hypervisor.h>.  We validate this at build time
- * here, and again at runtime during early boot.  We provide a
- * separate definition since userspace doesn't have <hv/hypervisor.h>.
- *
- * Be careful to distinguish PAGE_SHIFT from HV_PTE_INDEX_PFN, since
- * they are the same on i386 but not TILE.
+ * If the Kconfig doesn't specify, set a maximum zone order that
+ * is enough so that we can create huge pages from small pages given
+ * the respective sizes of the two page types.  See <linux/mmzone.h>.
  */
-#if HV_LOG2_PAGE_SIZE_SMALL != PAGE_SHIFT
-# error Small page size mismatch in Linux
-#endif
-#if HV_LOG2_PAGE_SIZE_LARGE != HPAGE_SHIFT
-# error Huge page size mismatch in Linux
+#ifndef CONFIG_FORCE_MAX_ZONEORDER
+#define CONFIG_FORCE_MAX_ZONEORDER (HPAGE_SHIFT - PAGE_SHIFT + 1)
 #endif
 
+#include <hv/hypervisor.h>
+#include <arch/chip.h>
+
 #ifndef __ASSEMBLY__
 
 #include <linux/types.h>
@@ -81,12 +75,6 @@ static inline void copy_user_page(void *to, void *from, unsigned long vaddr,
  * Hypervisor page tables are made of the same basic structure.
  */
 
-typedef __u64 pteval_t;
-typedef __u64 pmdval_t;
-typedef __u64 pudval_t;
-typedef __u64 pgdval_t;
-typedef __u64 pgprotval_t;
-
 typedef HV_PTE pte_t;
 typedef HV_PTE pgd_t;
 typedef HV_PTE pgprot_t;

arch/tile/include/asm/pgalloc.h

@@ -41,9 +41,9 @@
 static inline void set_pmd(pmd_t *pmdp, pmd_t pmd)
 {
 #ifdef CONFIG_64BIT
-	set_pte_order(pmdp, pmd, L2_USER_PGTABLE_ORDER);
+	set_pte(pmdp, pmd);
 #else
-	set_pte_order(&pmdp->pud.pgd, pmd.pud.pgd, L2_USER_PGTABLE_ORDER);
+	set_pte(&pmdp->pud.pgd, pmd.pud.pgd);
 #endif
 }
 
@@ -100,6 +100,9 @@ pte_t *get_prealloc_pte(unsigned long pfn);
 /* During init, we can shatter kernel huge pages if needed. */
 void shatter_pmd(pmd_t *pmd);
 
+/* After init, a more complex technique is required. */
+void shatter_huge_page(unsigned long addr);
+
 #ifdef __tilegx__
 /* We share a single page allocator for both L1 and L2 page tables. */
 #if HV_L1_SIZE != HV_L2_SIZE

arch/tile/include/asm/pgtable.h

@@ -233,15 +233,23 @@ static inline void __pte_clear(pte_t *ptep)
 #define pgd_ERROR(e) \
 	pr_err("%s:%d: bad pgd 0x%016llx.\n", __FILE__, __LINE__, pgd_val(e))
 
+/* Return PA and protection info for a given kernel VA. */
+int va_to_cpa_and_pte(void *va, phys_addr_t *cpa, pte_t *pte);
+
+/*
+ * __set_pte() ensures we write the 64-bit PTE with 32-bit words in
+ * the right order on 32-bit platforms and also allows us to write
+ * hooks to check valid PTEs, etc., if we want.
+ */
+void __set_pte(pte_t *ptep, pte_t pte);
+
 /*
- * set_pte_order() sets the given PTE and also sanity-checks the
+ * set_pte() sets the given PTE and also sanity-checks the
  * requested PTE against the page homecaching.  Unspecified parts
  * of the PTE are filled in when it is written to memory, i.e. all
  * caching attributes if "!forcecache", or the home cpu if "anyhome".
  */
-extern void set_pte_order(pte_t *ptep, pte_t pte, int order);
-
-#define set_pte(ptep, pteval) set_pte_order(ptep, pteval, 0)
+extern void set_pte(pte_t *ptep, pte_t pte);
 #define set_pte_at(mm, addr, ptep, pteval) set_pte(ptep, pteval)
 #define set_pte_atomic(pteptr, pteval) set_pte(pteptr, pteval)
 
@@ -292,21 +300,6 @@ extern void check_mm_caching(struct mm_struct *prev, struct mm_struct *next);
 #define __pte_to_swp_entry(pte)	((swp_entry_t) { (pte).val >> 32 })
 #define __swp_entry_to_pte(swp)	((pte_t) { (((long long) ((swp).val)) << 32) })
 
-/*
- * clone_pgd_range(pgd_t *dst, pgd_t *src, int count);
- *
- *  dst - pointer to pgd range anwhere on a pgd page
- *  src - ""
- *  count - the number of pgds to copy.
- *
- * dst and src can be on the same page, but the range must not overlap,
- * and must not cross a page boundary.
- */
-static inline void clone_pgd_range(pgd_t *dst, pgd_t *src, int count)
-{
-       memcpy(dst, src, count * sizeof(pgd_t));
-}
-
 /*
  * Conversion functions: convert a page and protection to a page entry,
  * and a page entry and page directory to the page they refer to.

arch/tile/include/asm/pgtable_32.h

@@ -24,6 +24,7 @@
 #define PGDIR_SIZE	HV_PAGE_SIZE_LARGE
 #define PGDIR_MASK	(~(PGDIR_SIZE-1))
 #define PTRS_PER_PGD	(1 << (32 - PGDIR_SHIFT))
+#define SIZEOF_PGD	(PTRS_PER_PGD * sizeof(pgd_t))
 
 /*
  * The level-2 index is defined by the difference between the huge
@@ -33,6 +34,7 @@
  * this nomenclature is somewhat confusing.
  */
 #define PTRS_PER_PTE (1 << (HV_LOG2_PAGE_SIZE_LARGE - HV_LOG2_PAGE_SIZE_SMALL))
+#define SIZEOF_PTE	(PTRS_PER_PTE * sizeof(pte_t))
 
 #ifndef __ASSEMBLY__
 
@@ -94,7 +96,6 @@ static inline int pgd_addr_invalid(unsigned long addr)
  */
 #define __HAVE_ARCH_PTEP_TEST_AND_CLEAR_YOUNG
 #define __HAVE_ARCH_PTEP_SET_WRPROTECT
-#define __HAVE_ARCH_PTEP_GET_AND_CLEAR
 
 extern int ptep_test_and_clear_young(struct vm_area_struct *,
 				     unsigned long addr, pte_t *);
@@ -110,6 +111,11 @@ static inline pte_t ptep_get_and_clear(struct mm_struct *mm,
 	return pte;
 }
 
+static inline void __set_pmd(pmd_t *pmdp, pmd_t pmdval)
+{
+	set_pte(&pmdp->pud.pgd, pmdval.pud.pgd);
+}
+
 /* Create a pmd from a PTFN. */
 static inline pmd_t ptfn_pmd(unsigned long ptfn, pgprot_t prot)
 {

arch/tile/include/asm/processor.h

@@ -269,7 +269,6 @@ extern char chip_model[64];
 /* Data on which physical memory controller corresponds to which NUMA node. */
 extern int node_controller[];
 
-
 /* Do we dump information to the console when a user application crashes? */
 extern int show_crashinfo;
 

arch/tile/include/asm/ptrace.h

@@ -141,6 +141,9 @@ struct single_step_state {
 /* Single-step the instruction at regs->pc */
 extern void single_step_once(struct pt_regs *regs);
 
+/* Clean up after execve(). */
+extern void single_step_execve(void);
+
 struct task_struct;
 
 extern void send_sigtrap(struct task_struct *tsk, struct pt_regs *regs,

arch/tile/include/asm/spinlock_32.h

@@ -78,13 +78,6 @@ void arch_spin_unlock_wait(arch_spinlock_t *lock);
 #define _RD_COUNT_SHIFT 24
 #define _RD_COUNT_WIDTH 8
 
-/* Internal functions; do not use. */
-void arch_read_lock_slow(arch_rwlock_t *, u32);
-int arch_read_trylock_slow(arch_rwlock_t *);
-void arch_read_unlock_slow(arch_rwlock_t *);
-void arch_write_lock_slow(arch_rwlock_t *, u32);
-void arch_write_unlock_slow(arch_rwlock_t *, u32);
-
 /**
  * arch_read_can_lock() - would read_trylock() succeed?
  */
@@ -104,94 +97,32 @@ static inline int arch_write_can_lock(arch_rwlock_t *rwlock)
 /**
  * arch_read_lock() - acquire a read lock.
  */
-static inline void arch_read_lock(arch_rwlock_t *rwlock)
-{
-	u32 val = __insn_tns((int *)&rwlock->lock);
-	if (unlikely(val << _RD_COUNT_WIDTH)) {
-		arch_read_lock_slow(rwlock, val);
-		return;
-	}
-	rwlock->lock = val + (1 << _RD_COUNT_SHIFT);
-}
+void arch_read_lock(arch_rwlock_t *rwlock);
 
 /**
- * arch_read_lock() - acquire a write lock.
+ * arch_write_lock() - acquire a write lock.
  */
-static inline void arch_write_lock(arch_rwlock_t *rwlock)
-{
-	u32 val = __insn_tns((int *)&rwlock->lock);
-	if (unlikely(val != 0)) {
-		arch_write_lock_slow(rwlock, val);
-		return;
-	}
-	rwlock->lock = 1 << _WR_NEXT_SHIFT;
-}
+void arch_write_lock(arch_rwlock_t *rwlock);
 
 /**
  * arch_read_trylock() - try to acquire a read lock.
  */
-static inline int arch_read_trylock(arch_rwlock_t *rwlock)
-{
-	int locked;
-	u32 val = __insn_tns((int *)&rwlock->lock);
-	if (unlikely(val & 1))
-		return arch_read_trylock_slow(rwlock);
-	locked = (val << _RD_COUNT_WIDTH) == 0;
-	rwlock->lock = val + (locked << _RD_COUNT_SHIFT);
-	return locked;
-}
+int arch_read_trylock(arch_rwlock_t *rwlock);
 
 /**
  * arch_write_trylock() - try to acquire a write lock.
  */
-static inline int arch_write_trylock(arch_rwlock_t *rwlock)
-{
-	u32 val = __insn_tns((int *)&rwlock->lock);
-
-	/*
-	 * If a tns is in progress, or there's a waiting or active locker,
-	 * or active readers, we can't take the lock, so give up.
-	 */
-	if (unlikely(val != 0)) {
-		if (!(val & 1))
-			rwlock->lock = val;
-		return 0;
-	}
-
-	/* Set the "next" field to mark it locked. */
-	rwlock->lock = 1 << _WR_NEXT_SHIFT;
-	return 1;
-}
+int arch_write_trylock(arch_rwlock_t *rwlock);
 
 /**
  * arch_read_unlock() - release a read lock.
  */
-static inline void arch_read_unlock(arch_rwlock_t *rwlock)
-{
-	u32 val;
-	mb();  /* guarantee anything modified under the lock is visible */
-	val = __insn_tns((int *)&rwlock->lock);
-	if (unlikely(val & 1)) {
-		arch_read_unlock_slow(rwlock);
-		return;
-	}
-	rwlock->lock = val - (1 << _RD_COUNT_SHIFT);
-}
+void arch_read_unlock(arch_rwlock_t *rwlock);
 
 /**
  * arch_write_unlock() - release a write lock.
  */
-static inline void arch_write_unlock(arch_rwlock_t *rwlock)
-{
-	u32 val;
-	mb();  /* guarantee anything modified under the lock is visible */
-	val = __insn_tns((int *)&rwlock->lock);
-	if (unlikely(val != (1 << _WR_NEXT_SHIFT))) {
-		arch_write_unlock_slow(rwlock, val);
-		return;
-	}
-	rwlock->lock = 0;
-}
+void arch_write_unlock(arch_rwlock_t *rwlock);
 
 #define arch_read_lock_flags(lock, flags) arch_read_lock(lock)
 #define arch_write_lock_flags(lock, flags) arch_write_lock(lock)

arch/tile/include/asm/stack.h

@@ -18,13 +18,14 @@
 #include <linux/types.h>
 #include <linux/sched.h>
 #include <asm/backtrace.h>
+#include <asm/page.h>
 #include <hv/hypervisor.h>
 
 /* Everything we need to keep track of a backtrace iteration */
 struct KBacktraceIterator {
 	BacktraceIterator it;
 	struct task_struct *task;     /* task we are backtracing */
-	HV_PTE *pgtable;	      /* page table for user space access */
+	pte_t *pgtable;		      /* page table for user space access */
 	int end;		      /* iteration complete. */
 	int new_context;              /* new context is starting */
 	int profile;                  /* profiling, so stop on async intrpt */

arch/tile/include/asm/system.h

@@ -90,7 +90,24 @@
 #endif
 
 #if !CHIP_HAS_MF_WAITS_FOR_VICTIMS()
-int __mb_incoherent(void);  /* Helper routine for mb_incoherent(). */
+#include <hv/syscall_public.h>
+/*
+ * Issue an uncacheable load to each memory controller, then
+ * wait until those loads have completed.
+ */
+static inline void __mb_incoherent(void)
+{
+	long clobber_r10;
+	asm volatile("swint2"
+		     : "=R10" (clobber_r10)
+		     : "R10" (HV_SYS_fence_incoherent)
+		     : "r0", "r1", "r2", "r3", "r4",
+		       "r5", "r6", "r7", "r8", "r9",
+		       "r11", "r12", "r13", "r14",
+		       "r15", "r16", "r17", "r18", "r19",
+		       "r20", "r21", "r22", "r23", "r24",
+		       "r25", "r26", "r27", "r28", "r29");
+}
 #endif
 
 /* Fence to guarantee visibility of stores to incoherent memory. */

arch/tile/include/asm/thread_info.h

@@ -68,6 +68,7 @@ struct thread_info {
 #else
 #define THREAD_SIZE_ORDER (0)
 #endif
+#define THREAD_SIZE_PAGES (1 << THREAD_SIZE_ORDER)
 
 #define THREAD_SIZE (PAGE_SIZE << THREAD_SIZE_ORDER)
 #define LOG2_THREAD_SIZE (PAGE_SHIFT + THREAD_SIZE_ORDER)

arch/tile/include/asm/timex.h

@@ -38,6 +38,9 @@ static inline cycles_t get_cycles(void)
 
 cycles_t get_clock_rate(void);
 
+/* Convert nanoseconds to core clock cycles. */
+cycles_t ns2cycles(unsigned long nsecs);
+
 /* Called at cpu initialization to set some low-level constants. */
 void setup_clock(void);
 

arch/tile/include/hv/drv_mshim_intf.h

@@ -0,0 +1,50 @@
+/*
+ * Copyright 2011 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 drv_mshim_intf.h
+ * Interface definitions for the Linux EDAC memory controller driver.
+ */
+
+#ifndef _SYS_HV_INCLUDE_DRV_MSHIM_INTF_H
+#define _SYS_HV_INCLUDE_DRV_MSHIM_INTF_H
+
+/** Number of memory controllers in the public API. */
+#define TILE_MAX_MSHIMS 4
+
+/** Memory info under each memory controller. */
+struct mshim_mem_info
+{
+  uint64_t mem_size;     /**< Total memory size in bytes. */
+  uint8_t mem_type;      /**< Memory type, DDR2 or DDR3. */
+  uint8_t mem_ecc;       /**< Memory supports ECC. */
+};
+
+/**
+ * DIMM error structure.
+ * For now, only correctable errors are counted and the mshim doesn't record
+ * the error PA. HV takes panic upon uncorrectable errors.
+ */
+struct mshim_mem_error
+{
+  uint32_t sbe_count;     /**< Number of single-bit errors. */
+};
+
+/** Read this offset to get the memory info per mshim. */
+#define MSHIM_MEM_INFO_OFF 0x100
+
+/** Read this offset to check DIMM error. */
+#define MSHIM_MEM_ERROR_OFF 0x200
+
+#endif /* _SYS_HV_INCLUDE_DRV_MSHIM_INTF_H */

arch/tile/include/hv/hypervisor.h

@@ -338,9 +338,10 @@ typedef int HV_Errno;
 #define HV_ENOTREADY   -812  /**< Device not ready */
 #define HV_EIO         -813  /**< I/O error */
 #define HV_ENOMEM      -814  /**< Out of memory */
+#define HV_EAGAIN      -815  /**< Try again */
 
 #define HV_ERR_MAX     -801  /**< Largest HV error code */
-#define HV_ERR_MIN     -814  /**< Smallest HV error code */
+#define HV_ERR_MIN     -815  /**< Smallest HV error code */
 
 #ifndef __ASSEMBLER__
 
@@ -867,6 +868,43 @@ typedef struct
  */
 HV_PhysAddrRange hv_inquire_physical(int idx);
 
+/** Possible DIMM types. */
+typedef enum
+{
+  NO_DIMM                    = 0,  /**< No DIMM */
+  DDR2                       = 1,  /**< DDR2 */
+  DDR3                       = 2   /**< DDR3 */
+} HV_DIMM_Type;
+
+#ifdef __tilegx__
+
+/** Log2 of minimum DIMM bytes supported by the memory controller. */
+#define HV_MSH_MIN_DIMM_SIZE_SHIFT 29
+
+/** Max number of DIMMs contained by one memory controller. */
+#define HV_MSH_MAX_DIMMS 8
+
+#else
+
+/** Log2 of minimum DIMM bytes supported by the memory controller. */
+#define HV_MSH_MIN_DIMM_SIZE_SHIFT 26
+
+/** Max number of DIMMs contained by one memory controller. */
+#define HV_MSH_MAX_DIMMS 2
+
+#endif
+
+/** Number of bits to right-shift to get the DIMM type. */
+#define HV_DIMM_TYPE_SHIFT 0
+
+/** Bits to mask to get the DIMM type. */
+#define HV_DIMM_TYPE_MASK 0xf
+
+/** Number of bits to right-shift to get the DIMM size. */
+#define HV_DIMM_SIZE_SHIFT 4
+
+/** Bits to mask to get the DIMM size. */
+#define HV_DIMM_SIZE_MASK 0xf
 
 /** Memory controller information. */
 typedef struct
@@ -963,6 +1001,11 @@ HV_ASIDRange hv_inquire_asid(int idx);
 
 
 /** Waits for at least the specified number of nanoseconds then returns.
+ *
+ * NOTE: this deprecated function currently assumes a 750 MHz clock,
+ * and is thus not generally suitable for use.  New code should call
+ * hv_sysconf(HV_SYSCONF_CPU_SPEED), compute a cycle count to wait for,
+ * and delay by looping while checking the cycle counter SPR.
  *
  * @param nanosecs The number of nanoseconds to sleep.
  */
@@ -1038,6 +1081,7 @@ int hv_console_write(HV_VirtAddr bytes, int len);
  *  downcall:
  *
  *  INT_MESSAGE_RCV_DWNCL   (hypervisor message available)
+ *  INT_DEV_INTR_DWNCL      (device interrupt)
  *  INT_DMATLB_MISS_DWNCL   (DMA TLB miss)
  *  INT_SNITLB_MISS_DWNCL   (SNI TLB miss)
  *  INT_DMATLB_ACCESS_DWNCL (DMA TLB access violation)

arch/tile/kernel/entry.S

@@ -38,12 +38,6 @@ STD_ENTRY(kernel_execve)
 	jrp lr
 	STD_ENDPROC(kernel_execve)
 
-/* Delay a fixed number of cycles. */
-STD_ENTRY(__delay)
-	{ addi r0, r0, -1; bnzt r0, . }
-	jrp lr
-	STD_ENDPROC(__delay)
-
 /*
  * We don't run this function directly, but instead copy it to a page
  * we map into every user process.  See vdso_setup().
@@ -97,23 +91,17 @@ STD_ENTRY(smp_nap)
 
 /*
  * Enable interrupts racelessly and then nap until interrupted.
+ * Architecturally, we are guaranteed that enabling interrupts via
+ * mtspr to INTERRUPT_CRITICAL_SECTION only interrupts at the next PC.
  * This function's _cpu_idle_nap address is special; see intvec.S.
  * When interrupted at _cpu_idle_nap, we bump the PC forward 8, and
  * as a result return to the function that called _cpu_idle().
  */
 STD_ENTRY(_cpu_idle)
-	{
-	 lnk r0
-	 movei r1, KERNEL_PL
-	}
-	{
-	 addli r0, r0, _cpu_idle_nap - .
-	 mtspr INTERRUPT_CRITICAL_SECTION, r1
-	}
+	movei r1, 1
+	mtspr INTERRUPT_CRITICAL_SECTION, r1
 	IRQ_ENABLE(r2, r3)             /* unmask, but still with ICS set */
-	mtspr SPR_EX_CONTEXT_K_1, r1   /* Kernel PL, ICS clear */
-	mtspr SPR_EX_CONTEXT_K_0, r0
-	iret
+	mtspr INTERRUPT_CRITICAL_SECTION, zero
 	.global _cpu_idle_nap
 _cpu_idle_nap:
 	nap

arch/tile/kernel/head_32.S

@@ -133,7 +133,7 @@ ENTRY(_start)
 	}
 	ENDPROC(_start)
 
-.section ".bss.page_aligned","w"
+__PAGE_ALIGNED_BSS
 	.align PAGE_SIZE
 ENTRY(empty_zero_page)
 	.fill PAGE_SIZE,1,0
@@ -145,10 +145,10 @@ ENTRY(empty_zero_page)
 	.endif
 	.word HV_PTE_PAGE | HV_PTE_DIRTY | HV_PTE_PRESENT | HV_PTE_ACCESSED | \
 	      (HV_PTE_MODE_CACHE_NO_L3 << HV_PTE_INDEX_MODE)
-	.word (\bits1) | (HV_CPA_TO_PFN(\cpa) << HV_PTE_INDEX_PFN)
+	.word (\bits1) | (HV_CPA_TO_PFN(\cpa) << (HV_PTE_INDEX_PFN - 32))
 	.endm
 
-.section ".data.page_aligned","wa"
+__PAGE_ALIGNED_DATA
 	.align PAGE_SIZE
 ENTRY(swapper_pg_dir)
 	/*
@@ -158,12 +158,14 @@ ENTRY(swapper_pg_dir)
 	 */
 	.set addr, 0
 	.rept (MEM_USER_INTRPT - PAGE_OFFSET) >> PGDIR_SHIFT
-	PTE addr + PAGE_OFFSET, addr, HV_PTE_READABLE | HV_PTE_WRITABLE
+	PTE addr + PAGE_OFFSET, addr, (1 << (HV_PTE_INDEX_READABLE - 32)) | \
+				      (1 << (HV_PTE_INDEX_WRITABLE - 32))
 	.set addr, addr + PGDIR_SIZE
 	.endr
 
 	/* The true text VAs are mapped as VA = PA + MEM_SV_INTRPT */
-	PTE MEM_SV_INTRPT, 0, HV_PTE_READABLE | HV_PTE_EXECUTABLE
+	PTE MEM_SV_INTRPT, 0, (1 << (HV_PTE_INDEX_READABLE - 32)) | \
+			      (1 << (HV_PTE_INDEX_EXECUTABLE - 32))
 	.org swapper_pg_dir + HV_L1_SIZE
 	END(swapper_pg_dir)
 
@@ -176,6 +178,7 @@ ENTRY(swapper_pg_dir)
 	__INITDATA
 	.align CHIP_L2_LINE_SIZE()
 ENTRY(swapper_pgprot)
-	PTE	0, 0, HV_PTE_READABLE | HV_PTE_WRITABLE, 1
+	PTE	0, 0, (1 << (HV_PTE_INDEX_READABLE - 32)) | \
+		      (1 << (HV_PTE_INDEX_WRITABLE - 32)), 1
 	.align CHIP_L2_LINE_SIZE()
 	END(swapper_pgprot)

arch/tile/kernel/intvec_32.S

@@ -32,10 +32,6 @@
 # error "No support for kernel preemption currently"
 #endif
 
-#if INT_INTCTRL_K < 32 || INT_INTCTRL_K >= 48
-# error INT_INTCTRL_K coded to set high interrupt mask
-#endif
-
 #define PTREGS_PTR(reg, ptreg) addli reg, sp, C_ABI_SAVE_AREA_SIZE + (ptreg)
 
 #define PTREGS_OFFSET_SYSCALL PTREGS_OFFSET_REG(TREG_SYSCALL_NR)
@@ -1198,46 +1194,6 @@ STD_ENTRY(interrupt_return)
 
 	STD_ENDPROC(interrupt_return)
 
-	/*
-	 * This interrupt variant clears the INT_INTCTRL_K interrupt mask bit
-	 * before returning, so we can properly get more downcalls.
-	 */
-	.pushsection .text.handle_interrupt_downcall,"ax"
-handle_interrupt_downcall:
-	finish_interrupt_save handle_interrupt_downcall
-	check_single_stepping normal, .Ldispatch_downcall
-.Ldispatch_downcall:
-
-	/* Clear INTCTRL_K from the set of interrupts we ever enable. */
-	GET_INTERRUPTS_ENABLED_MASK_PTR(r30)
-	{
-	 addi   r30, r30, 4
-	 movei  r31, INT_MASK(INT_INTCTRL_K)
-	}
-	{
-	 lw     r20, r30
-	 nor    r21, r31, zero
-	}
-	and     r20, r20, r21
-	sw      r30, r20
-
-	{
-	 jalr   r0
-	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
-	}
-	FEEDBACK_REENTER(handle_interrupt_downcall)
-
-	/* Allow INTCTRL_K to be enabled next time we enable interrupts. */
-	lw      r20, r30
-	or      r20, r20, r31
-	sw      r30, r20
-
-	{
-	 movei  r30, 0   /* not an NMI */
-	 j      interrupt_return
-	}
-	STD_ENDPROC(handle_interrupt_downcall)
-
 	/*
 	 * Some interrupts don't check for single stepping
 	 */
@@ -1600,7 +1556,10 @@ STD_ENTRY(_sys_clone)
 	.align 64
 	/* Align much later jump on the start of a cache line. */
 #if !ATOMIC_LOCKS_FOUND_VIA_TABLE()
-	nop; nop
+	nop
+#if PAGE_SIZE >= 0x10000
+	nop
+#endif
 #endif
 ENTRY(sys_cmpxchg)
 
@@ -1628,9 +1587,13 @@ ENTRY(sys_cmpxchg)
 	 * about aliasing among multiple mappings of the same physical page,
 	 * and we ignore the low 3 bits so we have one lock that covers
 	 * both a cmpxchg64() and a cmpxchg() on either its low or high word.
-	 * NOTE: this code must match __atomic_hashed_lock() in lib/atomic.c.
+	 * NOTE: this must match __atomic_hashed_lock() in lib/atomic_32.c.
 	 */
 
+#if (PAGE_OFFSET & 0xffff) != 0
+# error Code here assumes PAGE_OFFSET can be loaded with just hi16()
+#endif
+
 #if ATOMIC_LOCKS_FOUND_VIA_TABLE()
 	{
 	 /* Check for unaligned input. */
@@ -1723,11 +1686,14 @@ ENTRY(sys_cmpxchg)
 	 lw	r26, r0
 	}
 	{
-	 /* atomic_locks is page aligned so this suffices to get its addr. */
-	 auli	r21, zero, hi16(atomic_locks)
+	 auli	r21, zero, ha16(atomic_locks)
 
 	 bbns   r23, .Lcmpxchg_badaddr
 	}
+#if PAGE_SIZE < 0x10000
+	/* atomic_locks is page-aligned so for big pages we don't need this. */
+	addli   r21, r21, lo16(atomic_locks)
+#endif
 	{
 	 /*
 	  * Insert the hash bits into the page-aligned pointer.
@@ -1762,7 +1728,7 @@ ENTRY(sys_cmpxchg)
 
 	/*
 	 * Perform the actual cmpxchg or atomic_update.
-	 * Note that __futex_mark_unlocked() in uClibc relies on
+	 * Note that the system <arch/atomic.h> header relies on
 	 * atomic_update() to always perform an "mf", so don't make
 	 * it optional or conditional without modifying that code.
 	 */
@@ -2014,17 +1980,17 @@ int_unalign:
 #endif
 	int_hand     INT_INTCTRL_0, INTCTRL_0, bad_intr
 	int_hand     INT_MESSAGE_RCV_DWNCL, MESSAGE_RCV_DWNCL, \
-		     hv_message_intr, handle_interrupt_downcall
+		     hv_message_intr
 	int_hand     INT_DEV_INTR_DWNCL, DEV_INTR_DWNCL, \
-		     tile_dev_intr, handle_interrupt_downcall
+		     tile_dev_intr
 	int_hand     INT_I_ASID, I_ASID, bad_intr
 	int_hand     INT_D_ASID, D_ASID, bad_intr
 	int_hand     INT_DMATLB_MISS_DWNCL, DMATLB_MISS_DWNCL, \
-		     do_page_fault, handle_interrupt_downcall
+		     do_page_fault
 	int_hand     INT_SNITLB_MISS_DWNCL, SNITLB_MISS_DWNCL, \
-		     do_page_fault, handle_interrupt_downcall
+		     do_page_fault
 	int_hand     INT_DMATLB_ACCESS_DWNCL, DMATLB_ACCESS_DWNCL, \
-		     do_page_fault, handle_interrupt_downcall
+		     do_page_fault
 	int_hand     INT_SN_CPL, SN_CPL, bad_intr
 	int_hand     INT_DOUBLE_FAULT, DOUBLE_FAULT, do_trap
 #if CHIP_HAS_AUX_PERF_COUNTERS()

arch/tile/kernel/irq.c

@@ -176,43 +176,43 @@ void disable_percpu_irq(unsigned int irq)
 EXPORT_SYMBOL(disable_percpu_irq);
 
 /* Mask an interrupt. */
-static void tile_irq_chip_mask(unsigned int irq)
+static void tile_irq_chip_mask(struct irq_data *d)
 {
-	mask_irqs(1UL << irq);
+	mask_irqs(1UL << d->irq);
 }
 
 /* Unmask an interrupt. */
-static void tile_irq_chip_unmask(unsigned int irq)
+static void tile_irq_chip_unmask(struct irq_data *d)
 {
-	unmask_irqs(1UL << irq);
+	unmask_irqs(1UL << d->irq);
 }
 
 /*
  * Clear an interrupt before processing it so that any new assertions
  * will trigger another irq.
  */
-static void tile_irq_chip_ack(unsigned int irq)
+static void tile_irq_chip_ack(struct irq_data *d)
 {
-	if ((unsigned long)get_irq_chip_data(irq) != IS_HW_CLEARED)
-		clear_irqs(1UL << irq);
+	if ((unsigned long)irq_data_get_irq_chip_data(d) != IS_HW_CLEARED)
+		clear_irqs(1UL << d->irq);
 }
 
 /*
  * For per-cpu interrupts, we need to avoid unmasking any interrupts
  * that we disabled via disable_percpu_irq().
  */
-static void tile_irq_chip_eoi(unsigned int irq)
+static void tile_irq_chip_eoi(struct irq_data *d)
 {
-	if (!(__get_cpu_var(irq_disable_mask) & (1UL << irq)))
-		unmask_irqs(1UL << irq);
+	if (!(__get_cpu_var(irq_disable_mask) & (1UL << d->irq)))
+		unmask_irqs(1UL << d->irq);
 }
 
 static struct irq_chip tile_irq_chip = {
 	.name = "tile_irq_chip",
-	.ack = tile_irq_chip_ack,
-	.eoi = tile_irq_chip_eoi,
-	.mask = tile_irq_chip_mask,
-	.unmask = tile_irq_chip_unmask,
+	.irq_ack = tile_irq_chip_ack,
+	.irq_eoi = tile_irq_chip_eoi,
+	.irq_mask = tile_irq_chip_mask,
+	.irq_unmask = tile_irq_chip_unmask,
 };
 
 void __init init_IRQ(void)
@@ -277,8 +277,10 @@ int show_interrupts(struct seq_file *p, void *v)
 	}
 
 	if (i < NR_IRQS) {
-		raw_spin_lock_irqsave(&irq_desc[i].lock, flags);
-		action = irq_desc[i].action;
+		struct irq_desc *desc = irq_to_desc(i);
+
+		raw_spin_lock_irqsave(&desc->lock, flags);
+		action = desc->action;
 		if (!action)
 			goto skip;
 		seq_printf(p, "%3d: ", i);
@@ -288,7 +290,7 @@ int show_interrupts(struct seq_file *p, void *v)
 		for_each_online_cpu(j)
 			seq_printf(p, "%10u ", kstat_irqs_cpu(i, j));
 #endif
-		seq_printf(p, " %14s", irq_desc[i].chip->name);
+		seq_printf(p, " %14s", get_irq_desc_chip(desc)->name);
 		seq_printf(p, "  %s", action->name);
 
 		for (action = action->next; action; action = action->next)
@@ -296,7 +298,7 @@ int show_interrupts(struct seq_file *p, void *v)
 
 		seq_putc(p, '\n');
 skip:
-		raw_spin_unlock_irqrestore(&irq_desc[i].lock, flags);
+		raw_spin_unlock_irqrestore(&desc->lock, flags);
 	}
 	return 0;
 }

arch/tile/kernel/machine_kexec.c

@@ -240,8 +240,11 @@ static void setup_quasi_va_is_pa(void)
 	pte = hv_pte(_PAGE_KERNEL | _PAGE_HUGE_PAGE);
 	pte = hv_pte_set_mode(pte, HV_PTE_MODE_CACHE_NO_L3);
 
-	for (i = 0; i < pgd_index(PAGE_OFFSET); i++)
-		pgtable[i] = pfn_pte(i << (HPAGE_SHIFT - PAGE_SHIFT), pte);
+	for (i = 0; i < pgd_index(PAGE_OFFSET); i++) {
+		unsigned long pfn = i << (HPAGE_SHIFT - PAGE_SHIFT);
+		if (pfn_valid(pfn))
+			__set_pte(&pgtable[i], pfn_pte(pfn, pte));
+	}
 }
 
 

arch/tile/kernel/pci-dma.c

@@ -86,6 +86,21 @@ EXPORT_SYMBOL(dma_free_coherent);
  * can count on nothing having been touched.
  */
 
+/* Flush a PA range from cache page by page. */
+static void __dma_map_pa_range(dma_addr_t dma_addr, size_t size)
+{
+	struct page *page = pfn_to_page(PFN_DOWN(dma_addr));
+	size_t bytesleft = PAGE_SIZE - (dma_addr & (PAGE_SIZE - 1));
+
+	while ((ssize_t)size > 0) {
+		/* Flush the page. */
+		homecache_flush_cache(page++, 0);
+
+		/* Figure out if we need to continue on the next page. */
+		size -= bytesleft;
+		bytesleft = PAGE_SIZE;
+	}
+}
 
 /*
  * dma_map_single can be passed any memory address, and there appear
@@ -97,26 +112,12 @@ EXPORT_SYMBOL(dma_free_coherent);
 dma_addr_t dma_map_single(struct device *dev, void *ptr, size_t size,
 	       enum dma_data_direction direction)
 {
-	struct page *page;
-	dma_addr_t dma_addr;
-	int thispage;
+	dma_addr_t dma_addr = __pa(ptr);
 
 	BUG_ON(!valid_dma_direction(direction));
 	WARN_ON(size == 0);
 
-	dma_addr = __pa(ptr);
-
-	/* We might have been handed a buffer that wraps a page boundary */
-	while ((int)size > 0) {
-		/* The amount to flush that's on this page */
-		thispage = PAGE_SIZE - ((unsigned long)ptr & (PAGE_SIZE - 1));
-		thispage = min((int)thispage, (int)size);
-		/* Is this valid for any page we could be handed? */
-		page = pfn_to_page(kaddr_to_pfn(ptr));
-		homecache_flush_cache(page, 0);
-		ptr += thispage;
-		size -= thispage;
-	}
+	__dma_map_pa_range(dma_addr, size);
 
 	return dma_addr;
 }
@@ -140,10 +141,8 @@ int dma_map_sg(struct device *dev, struct scatterlist *sglist, int nents,
 	WARN_ON(nents == 0 || sglist->length == 0);
 
 	for_each_sg(sglist, sg, nents, i) {
-		struct page *page;
 		sg->dma_address = sg_phys(sg);
-		page = pfn_to_page(sg->dma_address >> PAGE_SHIFT);
-		homecache_flush_cache(page, 0);
+		__dma_map_pa_range(sg->dma_address, sg->length);
 	}
 
 	return nents;
@@ -163,6 +162,7 @@ dma_addr_t dma_map_page(struct device *dev, struct page *page,
 {
 	BUG_ON(!valid_dma_direction(direction));
 
+	BUG_ON(offset + size > PAGE_SIZE);
 	homecache_flush_cache(page, 0);
 
 	return page_to_pa(page) + offset;

arch/tile/kernel/process.c

@@ -165,7 +165,7 @@ void free_thread_info(struct thread_info *info)
 		kfree(step_state);
 	}
 
-	free_page((unsigned long)info);
+	free_pages((unsigned long)info, THREAD_SIZE_ORDER);
 }
 
 static void save_arch_state(struct thread_struct *t);
@@ -574,6 +574,8 @@ SYSCALL_DEFINE4(execve, const char __user *, path,
 		goto out;
 	error = do_execve(filename, argv, envp, regs);
 	putname(filename);
+	if (error == 0)
+		single_step_execve();
 out:
 	return error;
 }
@@ -593,6 +595,8 @@ long compat_sys_execve(const char __user *path,
 		goto out;
 	error = compat_do_execve(filename, argv, envp, regs);
 	putname(filename);
+	if (error == 0)
+		single_step_execve();
 out:
 	return error;
 }

arch/tile/kernel/setup.c

@@ -59,6 +59,8 @@ unsigned long __initdata node_memmap_pfn[MAX_NUMNODES];
 unsigned long __initdata node_percpu_pfn[MAX_NUMNODES];
 unsigned long __initdata node_free_pfn[MAX_NUMNODES];
 
+static unsigned long __initdata node_percpu[MAX_NUMNODES];
+
 #ifdef CONFIG_HIGHMEM
 /* Page frame index of end of lowmem on each controller. */
 unsigned long __cpuinitdata node_lowmem_end_pfn[MAX_NUMNODES];
@@ -554,7 +556,6 @@ static void __init setup_bootmem_allocator(void)
 		reserve_bootmem(crashk_res.start,
 			crashk_res.end - crashk_res.start + 1, 0);
 #endif
-
 }
 
 void *__init alloc_remap(int nid, unsigned long size)
@@ -568,11 +569,13 @@ void *__init alloc_remap(int nid, unsigned long size)
 
 static int __init percpu_size(void)
 {
-	int size = ALIGN(__per_cpu_end - __per_cpu_start, PAGE_SIZE);
-#ifdef CONFIG_MODULES
-	if (size < PERCPU_ENOUGH_ROOM)
-		size = PERCPU_ENOUGH_ROOM;
-#endif
+	int size = __per_cpu_end - __per_cpu_start;
+	size += PERCPU_MODULE_RESERVE;
+	size += PERCPU_DYNAMIC_EARLY_SIZE;
+	if (size < PCPU_MIN_UNIT_SIZE)
+		size = PCPU_MIN_UNIT_SIZE;
+	size = roundup(size, PAGE_SIZE);
+
 	/* In several places we assume the per-cpu data fits on a huge page. */
 	BUG_ON(kdata_huge && size > HPAGE_SIZE);
 	return size;
@@ -589,7 +592,6 @@ static inline unsigned long alloc_bootmem_pfn(int size, unsigned long goal)
 static void __init zone_sizes_init(void)
 {
 	unsigned long zones_size[MAX_NR_ZONES] = { 0 };
-	unsigned long node_percpu[MAX_NUMNODES] = { 0 };
 	int size = percpu_size();
 	int num_cpus = smp_height * smp_width;
 	int i;
@@ -674,7 +676,7 @@ static void __init zone_sizes_init(void)
 		NODE_DATA(i)->bdata = NODE_DATA(0)->bdata;
 
 		free_area_init_node(i, zones_size, start, NULL);
-		printk(KERN_DEBUG "  DMA zone: %ld per-cpu pages\n",
+		printk(KERN_DEBUG "  Normal zone: %ld per-cpu pages\n",
 		       PFN_UP(node_percpu[i]));
 
 		/* Track the type of memory on each node */
@@ -1312,6 +1314,8 @@ static void *__init pcpu_fc_alloc(unsigned int cpu, size_t size, size_t align)
 
 	BUG_ON(size % PAGE_SIZE != 0);
 	pfn_offset[nid] += size / PAGE_SIZE;
+	BUG_ON(node_percpu[nid] < size);
+	node_percpu[nid] -= size;
 	if (percpu_pfn[cpu] == 0)
 		percpu_pfn[cpu] = pfn;
 	return pfn_to_kaddr(pfn);

arch/tile/kernel/single_step.c

@@ -56,7 +56,7 @@ enum mem_op {
 	MEMOP_STORE_POSTINCR
 };
 
-static inline tile_bundle_bits set_BrOff_X1(tile_bundle_bits n, int32_t offset)
+static inline tile_bundle_bits set_BrOff_X1(tile_bundle_bits n, s32 offset)
 {
 	tile_bundle_bits result;
 
@@ -254,6 +254,18 @@ P("\n");
 	return bundle;
 }
 
+/*
+ * Called after execve() has started the new image.  This allows us
+ * to reset the info state.  Note that the the mmap'ed memory, if there
+ * was any, has already been unmapped by the exec.
+ */
+void single_step_execve(void)
+{
+	struct thread_info *ti = current_thread_info();
+	kfree(ti->step_state);
+	ti->step_state = NULL;
+}
+
 /**
  * single_step_once() - entry point when single stepping has been triggered.
  * @regs: The machine register state
@@ -373,7 +385,7 @@ void single_step_once(struct pt_regs *regs)
 		/* branches */
 		case BRANCH_OPCODE_X1:
 		{
-			int32_t offset = signExtend17(get_BrOff_X1(bundle));
+			s32 offset = signExtend17(get_BrOff_X1(bundle));
 
 			/*
 			 * For branches, we use a rewriting trick to let the
@@ -731,4 +743,9 @@ void single_step_once(struct pt_regs *regs)
 	__insn_mtspr(SPR_SINGLE_STEP_EN_K_K, 1 << USER_PL);
 }
 
+void single_step_execve(void)
+{
+	/* Nothing */
+}
+
 #endif /* !__tilegx__ */

arch/tile/kernel/smp.c

@@ -36,6 +36,22 @@ static unsigned long __iomem *ipi_mappings[NR_CPUS];
 /* Set by smp_send_stop() to avoid recursive panics. */
 static int stopping_cpus;
 
+static void __send_IPI_many(HV_Recipient *recip, int nrecip, int tag)
+{
+	int sent = 0;
+	while (sent < nrecip) {
+		int rc = hv_send_message(recip, nrecip,
+					 (HV_VirtAddr)&tag, sizeof(tag));
+		if (rc < 0) {
+			if (!stopping_cpus)  /* avoid recursive panic */
+				panic("hv_send_message returned %d", rc);
+			break;
+		}
+		WARN_ONCE(rc == 0, "hv_send_message() returned zero\n");
+		sent += rc;
+	}
+}
+
 void send_IPI_single(int cpu, int tag)
 {
 	HV_Recipient recip = {
@@ -43,14 +59,13 @@ void send_IPI_single(int cpu, int tag)
 		.x = cpu % smp_width,
 		.state = HV_TO_BE_SENT
 	};
-	int rc = hv_send_message(&recip, 1, (HV_VirtAddr)&tag, sizeof(tag));
-	BUG_ON(rc <= 0);
+	__send_IPI_many(&recip, 1, tag);
 }
 
 void send_IPI_many(const struct cpumask *mask, int tag)
 {
 	HV_Recipient recip[NR_CPUS];
-	int cpu, sent;
+	int cpu;
 	int nrecip = 0;
 	int my_cpu = smp_processor_id();
 	for_each_cpu(cpu, mask) {
@@ -61,17 +76,7 @@ void send_IPI_many(const struct cpumask *mask, int tag)
 		r->x = cpu % smp_width;
 		r->state = HV_TO_BE_SENT;
 	}
-	sent = 0;
-	while (sent < nrecip) {
-		int rc = hv_send_message(recip, nrecip,
-					 (HV_VirtAddr)&tag, sizeof(tag));
-		if (rc <= 0) {
-			if (!stopping_cpus)  /* avoid recursive panic */
-				panic("hv_send_message returned %d", rc);
-			break;
-		}
-		sent += rc;
-	}
+	__send_IPI_many(recip, nrecip, tag);
 }
 
 void send_IPI_allbutself(int tag)

arch/tile/kernel/stack.c

@@ -44,13 +44,6 @@ static int in_kernel_stack(struct KBacktraceIterator *kbt, VirtualAddress sp)
 	return sp >= kstack_base && sp < kstack_base + THREAD_SIZE;
 }
 
-/* Is address in the specified kernel code? */
-static int in_kernel_text(VirtualAddress address)
-{
-	return (address >= MEM_SV_INTRPT &&
-		address < MEM_SV_INTRPT + HPAGE_SIZE);
-}
-
 /* Is address valid for reading? */
 static int valid_address(struct KBacktraceIterator *kbt, VirtualAddress address)
 {
@@ -63,6 +56,23 @@ static int valid_address(struct KBacktraceIterator *kbt, VirtualAddress address)
 	if (l1_pgtable == NULL)
 		return 0;	/* can't read user space in other tasks */
 
+#ifdef CONFIG_64BIT
+	/* Find the real l1_pgtable by looking in the l0_pgtable. */
+	pte = l1_pgtable[HV_L0_INDEX(address)];
+	if (!hv_pte_get_present(pte))
+		return 0;
+	pfn = hv_pte_get_pfn(pte);
+	if (pte_huge(pte)) {
+		if (!pfn_valid(pfn)) {
+			pr_err("L0 huge page has bad pfn %#lx\n", pfn);
+			return 0;
+		}
+		return hv_pte_get_present(pte) && hv_pte_get_readable(pte);
+	}
+	page = pfn_to_page(pfn);
+	BUG_ON(PageHighMem(page));  /* No HIGHMEM on 64-bit. */
+	l1_pgtable = (HV_PTE *)pfn_to_kaddr(pfn);
+#endif
 	pte = l1_pgtable[HV_L1_INDEX(address)];
 	if (!hv_pte_get_present(pte))
 		return 0;
@@ -92,7 +102,7 @@ static bool read_memory_func(void *result, VirtualAddress address,
 {
 	int retval;
 	struct KBacktraceIterator *kbt = (struct KBacktraceIterator *)vkbt;
-	if (in_kernel_text(address)) {
+	if (__kernel_text_address(address)) {
 		/* OK to read kernel code. */
 	} else if (address >= PAGE_OFFSET) {
 		/* We only tolerate kernel-space reads of this task's stack */
@@ -132,7 +142,7 @@ static struct pt_regs *valid_fault_handler(struct KBacktraceIterator* kbt)
 		}
 	}
 	if (EX1_PL(p->ex1) == KERNEL_PL &&
-	    in_kernel_text(p->pc) &&
+	    __kernel_text_address(p->pc) &&
 	    in_kernel_stack(kbt, p->sp) &&
 	    p->sp >= sp) {
 		if (kbt->verbose)

arch/tile/kernel/time.c

@@ -224,3 +224,13 @@ int setup_profiling_timer(unsigned int multiplier)
 {
 	return -EINVAL;
 }
+
+/*
+ * Use the tile timer to convert nsecs to core clock cycles, relying
+ * on it having the same frequency as SPR_CYCLE.
+ */
+cycles_t ns2cycles(unsigned long nsecs)
+{
+	struct clock_event_device *dev = &__get_cpu_var(tile_timer);
+	return ((u64)nsecs * dev->mult) >> dev->shift;
+}

arch/tile/kernel/vmlinux.lds.S

@@ -59,10 +59,7 @@ SECTIONS
 
   . = ALIGN(PAGE_SIZE);
   VMLINUX_SYMBOL(_sinitdata) = .;
-  .init.page : AT (ADDR(.init.page) - LOAD_OFFSET) {
-    *(.init.page)
-  } :data =0
-  INIT_DATA_SECTION(16)
+  INIT_DATA_SECTION(16) :data =0
   PERCPU(L2_CACHE_BYTES, PAGE_SIZE)
   . = ALIGN(PAGE_SIZE);
   VMLINUX_SYMBOL(_einitdata) = .;

arch/tile/lib/Makefile

@@ -2,9 +2,8 @@
 # Makefile for TILE-specific library files..
 #
 
-lib-y = cacheflush.o checksum.o cpumask.o delay.o \
-	mb_incoherent.o uaccess.o memmove.o \
-	memcpy_$(BITS).o memchr_$(BITS).o memset_$(BITS).o \
+lib-y = cacheflush.o checksum.o cpumask.o delay.o uaccess.o \
+	memmove.o memcpy_$(BITS).o memchr_$(BITS).o memset_$(BITS).o \
 	strchr_$(BITS).o strlen_$(BITS).o
 
 ifeq ($(CONFIG_TILEGX),y)

arch/tile/lib/atomic_32.c

@@ -46,14 +46,13 @@ struct atomic_locks_on_cpu *atomic_lock_ptr[ATOMIC_HASH_L1_SIZE]
 #else /* ATOMIC_LOCKS_FOUND_VIA_TABLE() */
 
 /* This page is remapped on startup to be hash-for-home. */
-int atomic_locks[PAGE_SIZE / sizeof(int) /* Only ATOMIC_HASH_SIZE is used */]
-  __attribute__((aligned(PAGE_SIZE), section(".bss.page_aligned")));
+int atomic_locks[PAGE_SIZE / sizeof(int)] __page_aligned_bss;
 
 #endif /* ATOMIC_LOCKS_FOUND_VIA_TABLE() */
 
 static inline int *__atomic_hashed_lock(volatile void *v)
 {
-	/* NOTE: this code must match "sys_cmpxchg" in kernel/intvec.S */
+	/* NOTE: this code must match "sys_cmpxchg" in kernel/intvec_32.S */
 #if ATOMIC_LOCKS_FOUND_VIA_TABLE()
 	unsigned long i =
 		(unsigned long) v & ((PAGE_SIZE-1) & -sizeof(long long));

arch/tile/lib/atomic_asm_32.S

@@ -14,7 +14,7 @@
  * Support routines for atomic operations.  Each function takes:
  *
  * r0: address to manipulate
- * r1: pointer to atomic lock guarding this operation (for FUTEX_LOCK_REG)
+ * r1: pointer to atomic lock guarding this operation (for ATOMIC_LOCK_REG)
  * r2: new value to write, or for cmpxchg/add_unless, value to compare against
  * r3: (cmpxchg/xchg_add_unless) new value to write or add;
  *     (atomic64 ops) high word of value to write

arch/tile/lib/cacheflush.c

@@ -21,3 +21,105 @@ void __flush_icache_range(unsigned long start, unsigned long end)
 {
 	invalidate_icache((const void *)start, end - start, PAGE_SIZE);
 }
+
+
+/* Force a load instruction to issue. */
+static inline void force_load(char *p)
+{
+	*(volatile char *)p;
+}
+
+/*
+ * Flush and invalidate a VA range that is homed remotely on a single
+ * core (if "!hfh") or homed via hash-for-home (if "hfh"), waiting
+ * until the memory controller holds the flushed values.
+ */
+void finv_buffer_remote(void *buffer, size_t size, int hfh)
+{
+	char *p, *base;
+	size_t step_size, load_count;
+	const unsigned long STRIPE_WIDTH = 8192;
+
+	/*
+	 * Flush and invalidate the buffer out of the local L1/L2
+	 * and request the home cache to flush and invalidate as well.
+	 */
+	__finv_buffer(buffer, size);
+
+	/*
+	 * Wait for the home cache to acknowledge that it has processed
+	 * all the flush-and-invalidate requests.  This does not mean
+	 * that the flushed data has reached the memory controller yet,
+	 * but it does mean the home cache is processing the flushes.
+	 */
+	__insn_mf();
+
+	/*
+	 * Issue a load to the last cache line, which can't complete
+	 * until all the previously-issued flushes to the same memory
+	 * controller have also completed.  If we weren't striping
+	 * memory, that one load would be sufficient, but since we may
+	 * be, we also need to back up to the last load issued to
+	 * another memory controller, which would be the point where
+	 * we crossed an 8KB boundary (the granularity of striping
+	 * across memory controllers).  Keep backing up and doing this
+	 * until we are before the beginning of the buffer, or have
+	 * hit all the controllers.
+	 *
+	 * If we are flushing a hash-for-home buffer, it's even worse.
+	 * Each line may be homed on a different tile, and each tile
+	 * may have up to four lines that are on different
+	 * controllers.  So as we walk backwards, we have to touch
+	 * enough cache lines to satisfy these constraints.  In
+	 * practice this ends up being close enough to "load from
+	 * every cache line on a full memory stripe on each
+	 * controller" that we simply do that, to simplify the logic.
+	 *
+	 * FIXME: See bug 9535 for some issues with this code.
+	 */
+	if (hfh) {
+		step_size = L2_CACHE_BYTES;
+		load_count = (STRIPE_WIDTH / L2_CACHE_BYTES) *
+			      (1 << CHIP_LOG_NUM_MSHIMS());
+	} else {
+		step_size = STRIPE_WIDTH;
+		load_count = (1 << CHIP_LOG_NUM_MSHIMS());
+	}
+
+	/* Load the last byte of the buffer. */
+	p = (char *)buffer + size - 1;
+	force_load(p);
+
+	/* Bump down to the end of the previous stripe or cache line. */
+	p -= step_size;
+	p = (char *)((unsigned long)p | (step_size - 1));
+
+	/* Figure out how far back we need to go. */
+	base = p - (step_size * (load_count - 2));
+	if ((long)base < (long)buffer)
+		base = buffer;
+
+	/*
+	 * Fire all the loads we need.  The MAF only has eight entries
+	 * so we can have at most eight outstanding loads, so we
+	 * unroll by that amount.
+	 */
+#pragma unroll 8
+	for (; p >= base; p -= step_size)
+		force_load(p);
+
+	/*
+	 * Repeat, but with inv's instead of loads, to get rid of the
+	 * data we just loaded into our own cache and the old home L3.
+	 * No need to unroll since inv's don't target a register.
+	 */
+	p = (char *)buffer + size - 1;
+	__insn_inv(p);
+	p -= step_size;
+	p = (char *)((unsigned long)p | (step_size - 1));
+	for (; p >= base; p -= step_size)
+		__insn_inv(p);
+
+	/* Wait for the load+inv's (and thus finvs) to have completed. */
+	__insn_mf();
+}

arch/tile/lib/delay.c

@@ -15,20 +15,31 @@
 #include <linux/module.h>
 #include <linux/delay.h>
 #include <linux/thread_info.h>
-#include <asm/fixmap.h>
-#include <hv/hypervisor.h>
+#include <asm/timex.h>
 
 void __udelay(unsigned long usecs)
 {
-	hv_nanosleep(usecs * 1000);
+	if (usecs > ULONG_MAX / 1000) {
+		WARN_ON_ONCE(usecs > ULONG_MAX / 1000);
+		usecs = ULONG_MAX / 1000;
+	}
+	__ndelay(usecs * 1000);
 }
 EXPORT_SYMBOL(__udelay);
 
 void __ndelay(unsigned long nsecs)
 {
-	hv_nanosleep(nsecs);
+	cycles_t target = get_cycles();
+	target += ns2cycles(nsecs);
+	while (get_cycles() < target)
+		cpu_relax();
 }
 EXPORT_SYMBOL(__ndelay);
 
-/* FIXME: should be declared in a header somewhere. */
+void __delay(unsigned long cycles)
+{
+	cycles_t target = get_cycles() + cycles;
+	while (get_cycles() < target)
+		cpu_relax();
+}
 EXPORT_SYMBOL(__delay);

arch/tile/lib/exports.c

@@ -29,6 +29,9 @@ EXPORT_SYMBOL(__put_user_8);
 EXPORT_SYMBOL(strnlen_user_asm);
 EXPORT_SYMBOL(strncpy_from_user_asm);
 EXPORT_SYMBOL(clear_user_asm);
+EXPORT_SYMBOL(flush_user_asm);
+EXPORT_SYMBOL(inv_user_asm);
+EXPORT_SYMBOL(finv_user_asm);
 
 /* arch/tile/kernel/entry.S */
 #include <linux/kernel.h>
@@ -45,9 +48,6 @@ EXPORT_SYMBOL(__copy_from_user_zeroing);
 EXPORT_SYMBOL(__copy_in_user_inatomic);
 #endif
 
-/* arch/tile/lib/mb_incoherent.S */
-EXPORT_SYMBOL(__mb_incoherent);
-
 /* hypervisor glue */
 #include <hv/hypervisor.h>
 EXPORT_SYMBOL(hv_dev_open);
@@ -85,4 +85,8 @@ int64_t __muldi3(int64_t, int64_t);
 EXPORT_SYMBOL(__muldi3);
 uint64_t __lshrdi3(uint64_t, unsigned int);
 EXPORT_SYMBOL(__lshrdi3);
+uint64_t __ashrdi3(uint64_t, unsigned int);
+EXPORT_SYMBOL(__ashrdi3);
+uint64_t __ashldi3(uint64_t, unsigned int);
+EXPORT_SYMBOL(__ashldi3);
 #endif

arch/tile/lib/mb_incoherent.S

@@ -1,34 +0,0 @@
-/*
- * 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.
- *
- * Assembly code for invoking the HV's fence_incoherent syscall.
- */
-
-#include <linux/linkage.h>
-#include <hv/syscall_public.h>
-#include <arch/abi.h>
-#include <arch/chip.h>
-
-#if !CHIP_HAS_MF_WAITS_FOR_VICTIMS()
-
-/*
- * Invoke the hypervisor's fence_incoherent syscall, which guarantees
- * that all victims for cachelines homed on this tile have reached memory.
- */
-STD_ENTRY(__mb_incoherent)
-	moveli TREG_SYSCALL_NR_NAME, HV_SYS_fence_incoherent
-	swint2
-	jrp lr
-	STD_ENDPROC(__mb_incoherent)
-
-#endif

arch/tile/lib/memcpy_tile64.c

@@ -96,7 +96,7 @@ static void memcpy_multicache(void *dest, const void *source,
 	newsrc = __fix_to_virt(idx) + ((unsigned long)source & (PAGE_SIZE-1));
 	pmdp = pmd_offset(pud_offset(pgd_offset_k(newsrc), newsrc), newsrc);
 	ptep = pte_offset_kernel(pmdp, newsrc);
-	*ptep = src_pte;   /* set_pte() would be confused by this */
+	__set_pte(ptep, src_pte);   /* set_pte() would be confused by this */
 	local_flush_tlb_page(NULL, newsrc, PAGE_SIZE);
 
 	/* Actually move the data. */
@@ -109,7 +109,7 @@ static void memcpy_multicache(void *dest, const void *source,
 	 */
 	src_pte = hv_pte_set_mode(src_pte, HV_PTE_MODE_CACHE_NO_L3);
 	src_pte = hv_pte_set_writable(src_pte); /* need write access for inv */
-	*ptep = src_pte;   /* set_pte() would be confused by this */
+	__set_pte(ptep, src_pte);   /* set_pte() would be confused by this */
 	local_flush_tlb_page(NULL, newsrc, PAGE_SIZE);
 
 	/*

arch/tile/lib/spinlock_32.c

@@ -15,6 +15,7 @@
 #include <linux/spinlock.h>
 #include <linux/module.h>
 #include <asm/processor.h>
+#include <arch/spr_def.h>
 
 #include "spinlock_common.h"
 
@@ -91,75 +92,75 @@ EXPORT_SYMBOL(arch_spin_unlock_wait);
 #define RD_COUNT_MASK   ((1 << RD_COUNT_WIDTH) - 1)
 
 
-/* Lock the word, spinning until there are no tns-ers. */
-static inline u32 get_rwlock(arch_rwlock_t *rwlock)
-{
-	u32 iterations = 0;
-	for (;;) {
-		u32 val = __insn_tns((int *)&rwlock->lock);
-		if (unlikely(val & 1)) {
-			delay_backoff(iterations++);
-			continue;
-		}
-		return val;
-	}
-}
-
-int arch_read_trylock_slow(arch_rwlock_t *rwlock)
-{
-	u32 val = get_rwlock(rwlock);
-	int locked = (val << RD_COUNT_WIDTH) == 0;
-	rwlock->lock = val + (locked << RD_COUNT_SHIFT);
-	return locked;
-}
-EXPORT_SYMBOL(arch_read_trylock_slow);
-
-void arch_read_unlock_slow(arch_rwlock_t *rwlock)
-{
-	u32 val = get_rwlock(rwlock);
-	rwlock->lock = val - (1 << RD_COUNT_SHIFT);
-}
-EXPORT_SYMBOL(arch_read_unlock_slow);
-
-void arch_write_unlock_slow(arch_rwlock_t *rwlock, u32 val)
+/*
+ * We can get the read lock if everything but the reader bits (which
+ * are in the high part of the word) is zero, i.e. no active or
+ * waiting writers, no tns.
+ *
+ * We guard the tns/store-back with an interrupt critical section to
+ * preserve the semantic that the same read lock can be acquired in an
+ * interrupt context.
+ */
+inline int arch_read_trylock(arch_rwlock_t *rwlock)
 {
-	u32 eq, mask = 1 << WR_CURR_SHIFT;
-	while (unlikely(val & 1)) {
-		/* Limited backoff since we are the highest-priority task. */
-		relax(4);
-		val = __insn_tns((int *)&rwlock->lock);
+	u32 val;
+	__insn_mtspr(SPR_INTERRUPT_CRITICAL_SECTION, 1);
+	val = __insn_tns((int *)&rwlock->lock);
+	if (likely((val << _RD_COUNT_WIDTH) == 0)) {
+		val += 1 << RD_COUNT_SHIFT;
+		rwlock->lock = val;
+		__insn_mtspr(SPR_INTERRUPT_CRITICAL_SECTION, 0);
+		BUG_ON(val == 0);  /* we don't expect wraparound */
+		return 1;
 	}
-	val = __insn_addb(val, mask);
-	eq = __insn_seqb(val, val << (WR_CURR_SHIFT - WR_NEXT_SHIFT));
-	val = __insn_mz(eq & mask, val);
-	rwlock->lock = val;
+	if ((val & 1) == 0)
+		rwlock->lock = val;
+	__insn_mtspr(SPR_INTERRUPT_CRITICAL_SECTION, 0);
+	return 0;
 }
-EXPORT_SYMBOL(arch_write_unlock_slow);
+EXPORT_SYMBOL(arch_read_trylock);
 
 /*
- * We spin until everything but the reader bits (which are in the high
- * part of the word) are zero, i.e. no active or waiting writers, no tns.
- *
+ * Spin doing arch_read_trylock() until we acquire the lock.
  * ISSUE: This approach can permanently starve readers.  A reader who sees
  * a writer could instead take a ticket lock (just like a writer would),
  * and atomically enter read mode (with 1 reader) when it gets the ticket.
- * This way both readers and writers will always make forward progress
+ * This way both readers and writers would always make forward progress
  * in a finite time.
  */
-void arch_read_lock_slow(arch_rwlock_t *rwlock, u32 val)
+void arch_read_lock(arch_rwlock_t *rwlock)
 {
 	u32 iterations = 0;
-	do {
-		if (!(val & 1))
-			rwlock->lock = val;
+	while (unlikely(!arch_read_trylock(rwlock)))
 		delay_backoff(iterations++);
+}
+EXPORT_SYMBOL(arch_read_lock);
+
+void arch_read_unlock(arch_rwlock_t *rwlock)
+{
+	u32 val, iterations = 0;
+
+	mb();  /* guarantee anything modified under the lock is visible */
+	for (;;) {
+		__insn_mtspr(SPR_INTERRUPT_CRITICAL_SECTION, 1);
 		val = __insn_tns((int *)&rwlock->lock);
-	} while ((val << RD_COUNT_WIDTH) != 0);
-	rwlock->lock = val + (1 << RD_COUNT_SHIFT);
+		if (likely(val & 1) == 0) {
+			rwlock->lock = val - (1 << _RD_COUNT_SHIFT);
+			__insn_mtspr(SPR_INTERRUPT_CRITICAL_SECTION, 0);
+			break;
+		}
+		__insn_mtspr(SPR_INTERRUPT_CRITICAL_SECTION, 0);
+		delay_backoff(iterations++);
+	}
 }
-EXPORT_SYMBOL(arch_read_lock_slow);
+EXPORT_SYMBOL(arch_read_unlock);
 
-void arch_write_lock_slow(arch_rwlock_t *rwlock, u32 val)
+/*
+ * We don't need an interrupt critical section here (unlike for
+ * arch_read_lock) since we should never use a bare write lock where
+ * it could be interrupted by code that could try to re-acquire it.
+ */
+void arch_write_lock(arch_rwlock_t *rwlock)
 {
 	/*
 	 * The trailing underscore on this variable (and curr_ below)
@@ -168,6 +169,12 @@ void arch_write_lock_slow(arch_rwlock_t *rwlock, u32 val)
 	 */
 	u32 my_ticket_;
 	u32 iterations = 0;
+	u32 val = __insn_tns((int *)&rwlock->lock);
+
+	if (likely(val == 0)) {
+		rwlock->lock = 1 << _WR_NEXT_SHIFT;
+		return;
+	}
 
 	/*
 	 * Wait until there are no readers, then bump up the next
@@ -206,23 +213,47 @@ void arch_write_lock_slow(arch_rwlock_t *rwlock, u32 val)
 			relax(4);
 	}
 }
-EXPORT_SYMBOL(arch_write_lock_slow);
+EXPORT_SYMBOL(arch_write_lock);
 
-int __tns_atomic_acquire(atomic_t *lock)
+int arch_write_trylock(arch_rwlock_t *rwlock)
 {
-	int ret;
-	u32 iterations = 0;
+	u32 val = __insn_tns((int *)&rwlock->lock);
 
-	BUG_ON(__insn_mfspr(SPR_INTERRUPT_CRITICAL_SECTION));
-	__insn_mtspr(SPR_INTERRUPT_CRITICAL_SECTION, 1);
+	/*
+	 * If a tns is in progress, or there's a waiting or active locker,
+	 * or active readers, we can't take the lock, so give up.
+	 */
+	if (unlikely(val != 0)) {
+		if (!(val & 1))
+			rwlock->lock = val;
+		return 0;
+	}
 
-	while ((ret = __insn_tns((void *)&lock->counter)) == 1)
-		delay_backoff(iterations++);
-	return ret;
+	/* Set the "next" field to mark it locked. */
+	rwlock->lock = 1 << _WR_NEXT_SHIFT;
+	return 1;
 }
+EXPORT_SYMBOL(arch_write_trylock);
 
-void __tns_atomic_release(atomic_t *p, int v)
+void arch_write_unlock(arch_rwlock_t *rwlock)
 {
-	p->counter = v;
-	__insn_mtspr(SPR_INTERRUPT_CRITICAL_SECTION, 0);
+	u32 val, eq, mask;
+
+	mb();  /* guarantee anything modified under the lock is visible */
+	val = __insn_tns((int *)&rwlock->lock);
+	if (likely(val == (1 << _WR_NEXT_SHIFT))) {
+		rwlock->lock = 0;
+		return;
+	}
+	while (unlikely(val & 1)) {
+		/* Limited backoff since we are the highest-priority task. */
+		relax(4);
+		val = __insn_tns((int *)&rwlock->lock);
+	}
+	mask = 1 << WR_CURR_SHIFT;
+	val = __insn_addb(val, mask);
+	eq = __insn_seqb(val, val << (WR_CURR_SHIFT - WR_NEXT_SHIFT));
+	val = __insn_mz(eq & mask, val);
+	rwlock->lock = val;
 }
+EXPORT_SYMBOL(arch_write_unlock);

arch/tile/mm/fault.c

@@ -654,14 +654,6 @@ struct intvec_state do_page_fault_ics(struct pt_regs *regs, int fault_num,
 		regs->ex1 = PL_ICS_EX1(KERNEL_PL, 0);
 	}
 
-	/*
-	 * NOTE: the one other type of access that might bring us here
-	 * are the memory ops in __tns_atomic_acquire/__tns_atomic_release,
-	 * but we don't have to check specially for them since we can
-	 * always safely return to the address of the fault and retry,
-	 * since no separate atomic locks are involved.
-	 */
-
 	/*
 	 * Now that we have released the atomic lock (if necessary),
 	 * it's safe to spin if the PTE that caused the fault was migrating.

arch/tile/mm/homecache.c

@@ -179,23 +179,46 @@ void flush_remote(unsigned long cache_pfn, unsigned long cache_control,
 	panic("Unsafe to continue.");
 }
 
+void flush_remote_page(struct page *page, int order)
+{
+	int i, pages = (1 << order);
+	for (i = 0; i < pages; ++i, ++page) {
+		void *p = kmap_atomic(page);
+		int hfh = 0;
+		int home = page_home(page);
+#if CHIP_HAS_CBOX_HOME_MAP()
+		if (home == PAGE_HOME_HASH)
+			hfh = 1;
+		else
+#endif
+			BUG_ON(home < 0 || home >= NR_CPUS);
+		finv_buffer_remote(p, PAGE_SIZE, hfh);
+		kunmap_atomic(p);
+	}
+}
+
 void homecache_evict(const struct cpumask *mask)
 {
 	flush_remote(0, HV_FLUSH_EVICT_L2, mask, 0, 0, 0, NULL, NULL, 0);
 }
 
-/* Return a mask of the cpus whose caches currently own these pages. */
-static void homecache_mask(struct page *page, int pages,
-			   struct cpumask *home_mask)
+/*
+ * Return a mask of the cpus whose caches currently own these pages.
+ * The return value is whether the pages are all coherently cached
+ * (i.e. none are immutable, incoherent, or uncached).
+ */
+static int homecache_mask(struct page *page, int pages,
+			  struct cpumask *home_mask)
 {
 	int i;
+	int cached_coherently = 1;
 	cpumask_clear(home_mask);
 	for (i = 0; i < pages; ++i) {
 		int home = page_home(&page[i]);
 		if (home == PAGE_HOME_IMMUTABLE ||
 		    home == PAGE_HOME_INCOHERENT) {
 			cpumask_copy(home_mask, cpu_possible_mask);
-			return;
+			return 0;
 		}
 #if CHIP_HAS_CBOX_HOME_MAP()
 		if (home == PAGE_HOME_HASH) {
@@ -203,11 +226,14 @@ static void homecache_mask(struct page *page, int pages,
 			continue;
 		}
 #endif
-		if (home == PAGE_HOME_UNCACHED)
+		if (home == PAGE_HOME_UNCACHED) {
+			cached_coherently = 0;
 			continue;
+		}
 		BUG_ON(home < 0 || home >= NR_CPUS);
 		cpumask_set_cpu(home, home_mask);
 	}
+	return cached_coherently;
 }
 
 /*
@@ -386,7 +412,7 @@ void homecache_change_page_home(struct page *page, int order, int home)
 		pte_t *ptep = virt_to_pte(NULL, kva);
 		pte_t pteval = *ptep;
 		BUG_ON(!pte_present(pteval) || pte_huge(pteval));
-		*ptep = pte_set_home(pteval, home);
+		__set_pte(ptep, pte_set_home(pteval, home));
 	}
 }
 

arch/tile/mm/init.c

@@ -53,22 +53,11 @@
 
 #include "migrate.h"
 
-/*
- * We could set FORCE_MAX_ZONEORDER to "(HPAGE_SHIFT - PAGE_SHIFT + 1)"
- * in the Tile Kconfig, but this generates configure warnings.
- * Do it here and force people to get it right to compile this file.
- * The problem is that with 4KB small pages and 16MB huge pages,
- * the default value doesn't allow us to group enough small pages
- * together to make up a huge page.
- */
-#if CONFIG_FORCE_MAX_ZONEORDER < HPAGE_SHIFT - PAGE_SHIFT + 1
-# error "Change FORCE_MAX_ZONEORDER in arch/tile/Kconfig to match page size"
-#endif
-
 #define clear_pgd(pmdptr) (*(pmdptr) = hv_pte(0))
 
 #ifndef __tilegx__
 unsigned long VMALLOC_RESERVE = CONFIG_VMALLOC_RESERVE;
+EXPORT_SYMBOL(VMALLOC_RESERVE);
 #endif
 
 DEFINE_PER_CPU(struct mmu_gather, mmu_gathers);
@@ -445,7 +434,7 @@ static pmd_t *__init get_pmd(pgd_t pgtables[], unsigned long va)
 
 /* Temporary page table we use for staging. */
 static pgd_t pgtables[PTRS_PER_PGD]
- __attribute__((section(".init.page")));
+ __attribute__((aligned(HV_PAGE_TABLE_ALIGN)));
 
 /*
  * This maps the physical memory to kernel virtual address space, a total
@@ -653,6 +642,17 @@ static void __init kernel_physical_mapping_init(pgd_t *pgd_base)
 	memcpy(pgd_base, pgtables, sizeof(pgtables));
 	__install_page_table(pgd_base, __get_cpu_var(current_asid),
 			     swapper_pgprot);
+
+	/*
+	 * We just read swapper_pgprot and thus brought it into the cache,
+	 * with its new home & caching mode.  When we start the other CPUs,
+	 * they're going to reference swapper_pgprot via their initial fake
+	 * VA-is-PA mappings, which cache everything locally.  At that
+	 * time, if it's in our cache with a conflicting home, the
+	 * simulator's coherence checker will complain.  So, flush it out
+	 * of our cache; we're not going to ever use it again anyway.
+	 */
+	__insn_finv(&swapper_pgprot);
 }
 
 /*
@@ -950,11 +950,7 @@ struct kmem_cache *pgd_cache;
 
 void __init pgtable_cache_init(void)
 {
-	pgd_cache = kmem_cache_create("pgd",
-				PTRS_PER_PGD*sizeof(pgd_t),
-				PTRS_PER_PGD*sizeof(pgd_t),
-				0,
-				NULL);
+	pgd_cache = kmem_cache_create("pgd", SIZEOF_PGD, SIZEOF_PGD, 0, NULL);
 	if (!pgd_cache)
 		panic("pgtable_cache_init(): Cannot create pgd cache");
 }
@@ -989,7 +985,7 @@ static long __write_once initfree = 1;
 static int __init set_initfree(char *str)
 {
 	long val;
-	if (strict_strtol(str, 0, &val)) {
+	if (strict_strtol(str, 0, &val) == 0) {
 		initfree = val;
 		pr_info("initfree: %s free init pages\n",
 			initfree ? "will" : "won't");

arch/tile/mm/migrate_32.S

@@ -18,6 +18,7 @@
 #include <linux/linkage.h>
 #include <linux/threads.h>
 #include <asm/page.h>
+#include <asm/thread_info.h>
 #include <asm/types.h>
 #include <asm/asm-offsets.h>
 #include <hv/hypervisor.h>

arch/tile/mm/pgtable.c

@@ -142,6 +142,76 @@ pte_t *_pte_offset_map(pmd_t *dir, unsigned long address)
 }
 #endif
 
+/**
+ * shatter_huge_page() - ensure a given address is mapped by a small page.
+ *
+ * This function converts a huge PTE mapping kernel LOWMEM into a bunch
+ * of small PTEs with the same caching.  No cache flush required, but we
+ * must do a global TLB flush.
+ *
+ * Any caller that wishes to modify a kernel mapping that might
+ * have been made with a huge page should call this function,
+ * since doing so properly avoids race conditions with installing the
+ * newly-shattered page and then flushing all the TLB entries.
+ *
+ * @addr: Address at which to shatter any existing huge page.
+ */
+void shatter_huge_page(unsigned long addr)
+{
+	pgd_t *pgd;
+	pud_t *pud;
+	pmd_t *pmd;
+	unsigned long flags = 0;  /* happy compiler */
+#ifdef __PAGETABLE_PMD_FOLDED
+	struct list_head *pos;
+#endif
+
+	/* Get a pointer to the pmd entry that we need to change. */
+	addr &= HPAGE_MASK;
+	BUG_ON(pgd_addr_invalid(addr));
+	BUG_ON(addr < PAGE_OFFSET);  /* only for kernel LOWMEM */
+	pgd = swapper_pg_dir + pgd_index(addr);
+	pud = pud_offset(pgd, addr);
+	BUG_ON(!pud_present(*pud));
+	pmd = pmd_offset(pud, addr);
+	BUG_ON(!pmd_present(*pmd));
+	if (!pmd_huge_page(*pmd))
+		return;
+
+	/*
+	 * Grab the pgd_lock, since we may need it to walk the pgd_list,
+	 * and since we need some kind of lock here to avoid races.
+	 */
+	spin_lock_irqsave(&pgd_lock, flags);
+	if (!pmd_huge_page(*pmd)) {
+		/* Lost the race to convert the huge page. */
+		spin_unlock_irqrestore(&pgd_lock, flags);
+		return;
+	}
+
+	/* Shatter the huge page into the preallocated L2 page table. */
+	pmd_populate_kernel(&init_mm, pmd,
+			    get_prealloc_pte(pte_pfn(*(pte_t *)pmd)));
+
+#ifdef __PAGETABLE_PMD_FOLDED
+	/* Walk every pgd on the system and update the pmd there. */
+	list_for_each(pos, &pgd_list) {
+		pmd_t *copy_pmd;
+		pgd = list_to_pgd(pos) + pgd_index(addr);
+		pud = pud_offset(pgd, addr);
+		copy_pmd = pmd_offset(pud, addr);
+		__set_pmd(copy_pmd, *pmd);
+	}
+#endif
+
+	/* Tell every cpu to notice the change. */
+	flush_remote(0, 0, NULL, addr, HPAGE_SIZE, HPAGE_SIZE,
+		     cpu_possible_mask, NULL, 0);
+
+	/* Hold the lock until the TLB flush is finished to avoid races. */
+	spin_unlock_irqrestore(&pgd_lock, flags);
+}
+
 /*
  * List of all pgd's needed so it can invalidate entries in both cached
  * and uncached pgd's. This is essentially codepath-based locking
@@ -184,9 +254,9 @@ static void pgd_ctor(pgd_t *pgd)
 	BUG_ON(((u64 *)swapper_pg_dir)[pgd_index(MEM_USER_INTRPT)] != 0);
 #endif
 
-	clone_pgd_range(pgd + KERNEL_PGD_INDEX_START,
-			swapper_pg_dir + KERNEL_PGD_INDEX_START,
-			KERNEL_PGD_PTRS);
+	memcpy(pgd + KERNEL_PGD_INDEX_START,
+	       swapper_pg_dir + KERNEL_PGD_INDEX_START,
+	       KERNEL_PGD_PTRS * sizeof(pgd_t));
 
 	pgd_list_add(pgd);
 	spin_unlock_irqrestore(&pgd_lock, flags);
@@ -220,8 +290,11 @@ void pgd_free(struct mm_struct *mm, pgd_t *pgd)
 
 struct page *pte_alloc_one(struct mm_struct *mm, unsigned long address)
 {
-	gfp_t flags = GFP_KERNEL|__GFP_REPEAT|__GFP_ZERO|__GFP_COMP;
+	gfp_t flags = GFP_KERNEL|__GFP_REPEAT|__GFP_ZERO;
 	struct page *p;
+#if L2_USER_PGTABLE_ORDER > 0
+	int i;
+#endif
 
 #ifdef CONFIG_HIGHPTE
 	flags |= __GFP_HIGHMEM;
@@ -231,6 +304,18 @@ struct page *pte_alloc_one(struct mm_struct *mm, unsigned long address)
 	if (p == NULL)
 		return NULL;
 
+#if L2_USER_PGTABLE_ORDER > 0
+	/*
+	 * Make every page have a page_count() of one, not just the first.
+	 * We don't use __GFP_COMP since it doesn't look like it works
+	 * correctly with tlb_remove_page().
+	 */
+	for (i = 1; i < L2_USER_PGTABLE_PAGES; ++i) {
+		init_page_count(p+i);
+		inc_zone_page_state(p+i, NR_PAGETABLE);
+	}
+#endif
+
 	pgtable_page_ctor(p);
 	return p;
 }
@@ -242,8 +327,15 @@ struct page *pte_alloc_one(struct mm_struct *mm, unsigned long address)
  */
 void pte_free(struct mm_struct *mm, struct page *p)
 {
+	int i;
+
 	pgtable_page_dtor(p);
-	__free_pages(p, L2_USER_PGTABLE_ORDER);
+	__free_page(p);
+
+	for (i = 1; i < L2_USER_PGTABLE_PAGES; ++i) {
+		__free_page(p+i);
+		dec_zone_page_state(p+i, NR_PAGETABLE);
+	}
 }
 
 void __pte_free_tlb(struct mmu_gather *tlb, struct page *pte,
@@ -252,18 +344,11 @@ void __pte_free_tlb(struct mmu_gather *tlb, struct page *pte,
 	int i;
 
 	pgtable_page_dtor(pte);
-	tlb->need_flush = 1;
-	if (tlb_fast_mode(tlb)) {
-		struct page *pte_pages[L2_USER_PGTABLE_PAGES];
-		for (i = 0; i < L2_USER_PGTABLE_PAGES; ++i)
-			pte_pages[i] = pte + i;
-		free_pages_and_swap_cache(pte_pages, L2_USER_PGTABLE_PAGES);
-		return;
-	}
-	for (i = 0; i < L2_USER_PGTABLE_PAGES; ++i) {
-		tlb->pages[tlb->nr++] = pte + i;
-		if (tlb->nr >= FREE_PTE_NR)
-			tlb_flush_mmu(tlb, 0, 0);
+	tlb_remove_page(tlb, pte);
+
+	for (i = 1; i < L2_USER_PGTABLE_PAGES; ++i) {
+		tlb_remove_page(tlb, pte + i);
+		dec_zone_page_state(pte + i, NR_PAGETABLE);
 	}
 }
 
@@ -346,35 +431,51 @@ int get_remote_cache_cpu(pgprot_t prot)
 	return x + y * smp_width;
 }
 
-void set_pte_order(pte_t *ptep, pte_t pte, int order)
+/*
+ * Convert a kernel VA to a PA and homing information.
+ */
+int va_to_cpa_and_pte(void *va, unsigned long long *cpa, pte_t *pte)
 {
-	unsigned long pfn = pte_pfn(pte);
-	struct page *page = pfn_to_page(pfn);
+	struct page *page = virt_to_page(va);
+	pte_t null_pte = { 0 };
 
-	/* Update the home of a PTE if necessary */
-	pte = pte_set_home(pte, page_home(page));
+	*cpa = __pa(va);
 
+	/* Note that this is not writing a page table, just returning a pte. */
+	*pte = pte_set_home(null_pte, page_home(page));
+
+	return 0; /* return non-zero if not hfh? */
+}
+EXPORT_SYMBOL(va_to_cpa_and_pte);
+
+void __set_pte(pte_t *ptep, pte_t pte)
+{
 #ifdef __tilegx__
 	*ptep = pte;
 #else
-	/*
-	 * When setting a PTE, write the high bits first, then write
-	 * the low bits.  This sets the "present" bit only after the
-	 * other bits are in place.  If a particular PTE update
-	 * involves transitioning from one valid PTE to another, it
-	 * may be necessary to call set_pte_order() more than once,
-	 * transitioning via a suitable intermediate state.
-	 * Note that this sequence also means that if we are transitioning
-	 * from any migrating PTE to a non-migrating one, we will not
-	 * see a half-updated PTE with the migrating bit off.
-	 */
-#if HV_PTE_INDEX_PRESENT >= 32 || HV_PTE_INDEX_MIGRATING >= 32
-# error Must write the present and migrating bits last
-#endif
-	((u32 *)ptep)[1] = (u32)(pte_val(pte) >> 32);
-	barrier();
-	((u32 *)ptep)[0] = (u32)(pte_val(pte));
-#endif
+# if HV_PTE_INDEX_PRESENT >= 32 || HV_PTE_INDEX_MIGRATING >= 32
+#  error Must write the present and migrating bits last
+# endif
+	if (pte_present(pte)) {
+		((u32 *)ptep)[1] = (u32)(pte_val(pte) >> 32);
+		barrier();
+		((u32 *)ptep)[0] = (u32)(pte_val(pte));
+	} else {
+		((u32 *)ptep)[0] = (u32)(pte_val(pte));
+		barrier();
+		((u32 *)ptep)[1] = (u32)(pte_val(pte) >> 32);
+	}
+#endif /* __tilegx__ */
+}
+
+void set_pte(pte_t *ptep, pte_t pte)
+{
+	struct page *page = pfn_to_page(pte_pfn(pte));
+
+	/* Update the home of a PTE if necessary */
+	pte = pte_set_home(pte, page_home(page));
+
+	__set_pte(ptep, pte);
 }
 
 /* Can this mm load a PTE with cached_priority set? */

drivers/edac/Kconfig

@@ -7,7 +7,7 @@
 menuconfig EDAC
 	bool "EDAC (Error Detection And Correction) reporting"
 	depends on HAS_IOMEM
-	depends on X86 || PPC
+	depends on X86 || PPC || TILE
 	help
 	  EDAC is designed to report errors in the core system.
 	  These are low-level errors that are reported in the CPU or
@@ -282,4 +282,12 @@ config EDAC_CPC925
 	  a companion chip to the PowerPC 970 family of
 	  processors.
 
+config EDAC_TILE
+	tristate "Tilera Memory Controller"
+	depends on EDAC_MM_EDAC && TILE
+	default y
+	help
+	  Support for error detection and correction on the
+	  Tilera memory controller.
+
 endif # EDAC

drivers/edac/Makefile

@@ -54,3 +54,4 @@ obj-$(CONFIG_EDAC_PPC4XX)		+= ppc4xx_edac.o
 obj-$(CONFIG_EDAC_AMD8111)		+= amd8111_edac.o
 obj-$(CONFIG_EDAC_AMD8131)		+= amd8131_edac.o
 
+obj-$(CONFIG_EDAC_TILE)			+= tile_edac.o

drivers/edac/tile_edac.c

@@ -0,0 +1,254 @@
+/*
+ * Copyright 2011 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.
+ * Tilera-specific EDAC driver.
+ *
+ * This source code is derived from the following driver:
+ *
+ * Cell MIC driver for ECC counting
+ *
+ * Copyright 2007 Benjamin Herrenschmidt, IBM Corp.
+ *                <benh@kernel.crashing.org>
+ *
+ */
+
+#include <linux/module.h>
+#include <linux/init.h>
+#include <linux/platform_device.h>
+#include <linux/io.h>
+#include <linux/uaccess.h>
+#include <linux/edac.h>
+#include <hv/hypervisor.h>
+#include <hv/drv_mshim_intf.h>
+
+#include "edac_core.h"
+
+#define DRV_NAME	"tile-edac"
+
+/* Number of cs_rows needed per memory controller on TILEPro. */
+#define TILE_EDAC_NR_CSROWS	1
+
+/* Number of channels per memory controller on TILEPro. */
+#define TILE_EDAC_NR_CHANS	1
+
+/* Granularity of reported error in bytes on TILEPro. */
+#define TILE_EDAC_ERROR_GRAIN	8
+
+/* TILE processor has multiple independent memory controllers. */
+struct platform_device *mshim_pdev[TILE_MAX_MSHIMS];
+
+struct tile_edac_priv {
+	int		hv_devhdl;	/* Hypervisor device handle. */
+	int		node;		/* Memory controller instance #. */
+	unsigned int	ce_count;	/*
+					 * Correctable-error counter
+					 * kept by the driver.
+					 */
+};
+
+static void tile_edac_check(struct mem_ctl_info *mci)
+{
+	struct tile_edac_priv	*priv = mci->pvt_info;
+	struct mshim_mem_error	mem_error;
+
+	if (hv_dev_pread(priv->hv_devhdl, 0, (HV_VirtAddr)&mem_error,
+		sizeof(struct mshim_mem_error), MSHIM_MEM_ERROR_OFF) !=
+		sizeof(struct mshim_mem_error)) {
+		pr_err(DRV_NAME ": MSHIM_MEM_ERROR_OFF pread failure.\n");
+		return;
+	}
+
+	/* Check if the current error count is different from the saved one. */
+	if (mem_error.sbe_count != priv->ce_count) {
+		dev_dbg(mci->dev, "ECC CE err on node %d\n", priv->node);
+		priv->ce_count = mem_error.sbe_count;
+		edac_mc_handle_ce(mci, 0, 0, 0, 0, 0, mci->ctl_name);
+	}
+}
+
+/*
+ * Initialize the 'csrows' table within the mci control structure with the
+ * addressing of memory.
+ */
+static int __devinit tile_edac_init_csrows(struct mem_ctl_info *mci)
+{
+	struct csrow_info	*csrow = &mci->csrows[0];
+	struct tile_edac_priv	*priv = mci->pvt_info;
+	struct mshim_mem_info	mem_info;
+
+	if (hv_dev_pread(priv->hv_devhdl, 0, (HV_VirtAddr)&mem_info,
+		sizeof(struct mshim_mem_info), MSHIM_MEM_INFO_OFF) !=
+		sizeof(struct mshim_mem_info)) {
+		pr_err(DRV_NAME ": MSHIM_MEM_INFO_OFF pread failure.\n");
+		return -1;
+	}
+
+	if (mem_info.mem_ecc)
+		csrow->edac_mode = EDAC_SECDED;
+	else
+		csrow->edac_mode = EDAC_NONE;
+	switch (mem_info.mem_type) {
+	case DDR2:
+		csrow->mtype = MEM_DDR2;
+		break;
+
+	case DDR3:
+		csrow->mtype = MEM_DDR3;
+		break;
+
+	default:
+		return -1;
+	}
+
+	csrow->first_page = 0;
+	csrow->nr_pages = mem_info.mem_size >> PAGE_SHIFT;
+	csrow->last_page = csrow->first_page + csrow->nr_pages - 1;
+	csrow->grain = TILE_EDAC_ERROR_GRAIN;
+	csrow->dtype = DEV_UNKNOWN;
+
+	return 0;
+}
+
+static int __devinit tile_edac_mc_probe(struct platform_device *pdev)
+{
+	char			hv_file[32];
+	int			hv_devhdl;
+	struct mem_ctl_info	*mci;
+	struct tile_edac_priv	*priv;
+	int			rc;
+
+	sprintf(hv_file, "mshim/%d", pdev->id);
+	hv_devhdl = hv_dev_open((HV_VirtAddr)hv_file, 0);
+	if (hv_devhdl < 0)
+		return -EINVAL;
+
+	/* A TILE MC has a single channel and one chip-select row. */
+	mci = edac_mc_alloc(sizeof(struct tile_edac_priv),
+		TILE_EDAC_NR_CSROWS, TILE_EDAC_NR_CHANS, pdev->id);
+	if (mci == NULL)
+		return -ENOMEM;
+	priv = mci->pvt_info;
+	priv->node = pdev->id;
+	priv->hv_devhdl = hv_devhdl;
+
+	mci->dev = &pdev->dev;
+	mci->mtype_cap = MEM_FLAG_DDR2;
+	mci->edac_ctl_cap = EDAC_FLAG_SECDED;
+
+	mci->mod_name = DRV_NAME;
+	mci->ctl_name = "TILEPro_Memory_Controller";
+	mci->dev_name = dev_name(&pdev->dev);
+	mci->edac_check = tile_edac_check;
+
+	/*
+	 * Initialize the MC control structure 'csrows' table
+	 * with the mapping and control information.
+	 */
+	if (tile_edac_init_csrows(mci)) {
+		/* No csrows found. */
+		mci->edac_cap = EDAC_FLAG_NONE;
+	} else {
+		mci->edac_cap = EDAC_FLAG_SECDED;
+	}
+
+	platform_set_drvdata(pdev, mci);
+
+	/* Register with EDAC core */
+	rc = edac_mc_add_mc(mci);
+	if (rc) {
+		dev_err(&pdev->dev, "failed to register with EDAC core\n");
+		edac_mc_free(mci);
+		return rc;
+	}
+
+	return 0;
+}
+
+static int __devexit tile_edac_mc_remove(struct platform_device *pdev)
+{
+	struct mem_ctl_info *mci = platform_get_drvdata(pdev);
+
+	edac_mc_del_mc(&pdev->dev);
+	if (mci)
+		edac_mc_free(mci);
+	return 0;
+}
+
+static struct platform_driver tile_edac_mc_driver = {
+	.driver		= {
+		.name	= DRV_NAME,
+		.owner	= THIS_MODULE,
+	},
+	.probe		= tile_edac_mc_probe,
+	.remove		= __devexit_p(tile_edac_mc_remove),
+};
+
+/*
+ * Driver init routine.
+ */
+static int __init tile_edac_init(void)
+{
+	char	hv_file[32];
+	struct platform_device *pdev;
+	int i, err, num = 0;
+
+	/* Only support POLL mode. */
+	edac_op_state = EDAC_OPSTATE_POLL;
+
+	err = platform_driver_register(&tile_edac_mc_driver);
+	if (err)
+		return err;
+
+	for (i = 0; i < TILE_MAX_MSHIMS; i++) {
+		/*
+		 * Not all memory controllers are configured such as in the
+		 * case of a simulator. So we register only those mshims
+		 * that are configured by the hypervisor.
+		 */
+		sprintf(hv_file, "mshim/%d", i);
+		if (hv_dev_open((HV_VirtAddr)hv_file, 0) < 0)
+			continue;
+
+		pdev = platform_device_register_simple(DRV_NAME, i, NULL, 0);
+		if (IS_ERR(pdev))
+			continue;
+		mshim_pdev[i] = pdev;
+		num++;
+	}
+
+	if (num == 0) {
+		platform_driver_unregister(&tile_edac_mc_driver);
+		return -ENODEV;
+	}
+	return 0;
+}
+
+/*
+ * Driver cleanup routine.
+ */
+static void __exit tile_edac_exit(void)
+{
+	int i;
+
+	for (i = 0; i < TILE_MAX_MSHIMS; i++) {
+		struct platform_device *pdev = mshim_pdev[i];
+		if (!pdev)
+			continue;
+
+		platform_set_drvdata(pdev, NULL);
+		platform_device_unregister(pdev);
+	}
+	platform_driver_unregister(&tile_edac_mc_driver);
+}
+
+module_init(tile_edac_init);
+module_exit(tile_edac_exit);

drivers/net/tile/tilepro.c

@@ -1,5 +1,5 @@
 /*
- * Copyright 2010 Tilera Corporation. All Rights Reserved.
+ * Copyright 2011 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
@@ -44,10 +44,6 @@
 #include <linux/tcp.h>
 
 
-/* There is no singlethread_cpu, so schedule work on the current cpu. */
-#define singlethread_cpu -1
-
-
 /*
  * First, "tile_net_init_module()" initializes all four "devices" which
  * can be used by linux.
@@ -73,15 +69,16 @@
  * return, knowing we will be called again later.  Otherwise, we
  * reenable the ingress interrupt, and call "napi_complete()".
  *
+ * HACK: Since disabling the ingress interrupt is not reliable, we
+ * ignore the interrupt if the global "active" flag is false.
+ *
  *
  * NOTE: The use of "native_driver" ensures that EPP exists, and that
- * "epp_sendv" is legal, and that "LIPP" is being used.
+ * we are using "LIPP" and "LEPP".
  *
  * NOTE: Failing to free completions for an arbitrarily long time
  * (which is defined to be illegal) does in fact cause bizarre
  * problems.  The "egress_timer" helps prevent this from happening.
- *
- * NOTE: The egress code can be interrupted by the interrupt handler.
  */
 
 
@@ -142,6 +139,7 @@
 MODULE_AUTHOR("Tilera");
 MODULE_LICENSE("GPL");
 
+
 /*
  * Queue of incoming packets for a specific cpu and device.
  *
@@ -177,7 +175,7 @@ struct tile_net_cpu {
 	struct tile_netio_queue queue;
 	/* Statistics. */
 	struct tile_net_stats_t stats;
-	/* ISSUE: Is this needed? */
+	/* True iff NAPI is enabled. */
 	bool napi_enabled;
 	/* True if this tile has succcessfully registered with the IPP. */
 	bool registered;
@@ -200,20 +198,20 @@ struct tile_net_cpu {
 struct tile_net_priv {
 	/* Our network device. */
 	struct net_device *dev;
-	/* The actual egress queue. */
-	lepp_queue_t *epp_queue;
-	/* Protects "epp_queue->cmd_tail" and "epp_queue->comp_tail" */
-	spinlock_t cmd_lock;
-	/* Protects "epp_queue->comp_head". */
-	spinlock_t comp_lock;
+	/* Pages making up the egress queue. */
+	struct page *eq_pages;
+	/* Address of the actual egress queue. */
+	lepp_queue_t *eq;
+	/* Protects "eq". */
+	spinlock_t eq_lock;
 	/* The hypervisor handle for this interface. */
 	int hv_devhdl;
 	/* The intr bit mask that IDs this device. */
 	u32 intr_id;
 	/* True iff "tile_net_open_aux()" has succeeded. */
-	int partly_opened;
-	/* True iff "tile_net_open_inner()" has succeeded. */
-	int fully_opened;
+	bool partly_opened;
+	/* True iff the device is "active". */
+	bool active;
 	/* Effective network cpus. */
 	struct cpumask network_cpus_map;
 	/* Number of network cpus. */
@@ -228,6 +226,10 @@ struct tile_net_priv {
 	struct tile_net_cpu *cpu[NR_CPUS];
 };
 
+/* Log2 of the number of small pages needed for the egress queue. */
+#define EQ_ORDER  get_order(sizeof(lepp_queue_t))
+/* Size of the egress queue's pages. */
+#define EQ_SIZE   (1 << (PAGE_SHIFT + EQ_ORDER))
 
 /*
  * The actual devices (xgbe0, xgbe1, gbe0, gbe1).
@@ -284,7 +286,11 @@ static void net_printk(char *fmt, ...)
  */
 static void dump_packet(unsigned char *data, unsigned long length, char *s)
 {
+	int my_cpu = smp_processor_id();
+
 	unsigned long i;
+	char buf[128];
+
 	static unsigned int count;
 
 	pr_info("dump_packet(data %p, length 0x%lx s %s count 0x%x)\n",
@@ -294,10 +300,12 @@ static void dump_packet(unsigned char *data, unsigned long length, char *s)
 
 	for (i = 0; i < length; i++) {
 		if ((i & 0xf) == 0)
-			sprintf(buf, "%8.8lx:", i);
+			sprintf(buf, "[%02d] %8.8lx:", my_cpu, i);
 		sprintf(buf + strlen(buf), " %2.2x", data[i]);
-		if ((i & 0xf) == 0xf || i == length - 1)
-			pr_info("%s\n", buf);
+		if ((i & 0xf) == 0xf || i == length - 1) {
+			strcat(buf, "\n");
+			pr_info("%s", buf);
+		}
 	}
 }
 #endif
@@ -351,60 +359,109 @@ static void tile_net_provide_linux_buffer(struct tile_net_cpu *info,
 
 /*
  * Provide a linux buffer for LIPP.
+ *
+ * Note that the ACTUAL allocation for each buffer is a "struct sk_buff",
+ * plus a chunk of memory that includes not only the requested bytes, but
+ * also NET_SKB_PAD bytes of initial padding, and a "struct skb_shared_info".
+ *
+ * Note that "struct skb_shared_info" is 88 bytes with 64K pages and
+ * 268 bytes with 4K pages (since the frags[] array needs 18 entries).
+ *
+ * Without jumbo packets, the maximum packet size will be 1536 bytes,
+ * and we use 2 bytes (NET_IP_ALIGN) of padding.  ISSUE: If we told
+ * the hardware to clip at 1518 bytes instead of 1536 bytes, then we
+ * could save an entire cache line, but in practice, we don't need it.
+ *
+ * Since CPAs are 38 bits, and we can only encode the high 31 bits in
+ * a "linux_buffer_t", the low 7 bits must be zero, and thus, we must
+ * align the actual "va" mod 128.
+ *
+ * We assume that the underlying "head" will be aligned mod 64.  Note
+ * that in practice, we have seen "head" NOT aligned mod 128 even when
+ * using 2048 byte allocations, which is surprising.
+ *
+ * If "head" WAS always aligned mod 128, we could change LIPP to
+ * assume that the low SIX bits are zero, and the 7th bit is one, that
+ * is, align the actual "va" mod 128 plus 64, which would be "free".
+ *
+ * For now, the actual "head" pointer points at NET_SKB_PAD bytes of
+ * padding, plus 28 or 92 bytes of extra padding, plus the sk_buff
+ * pointer, plus the NET_IP_ALIGN padding, plus 126 or 1536 bytes for
+ * the actual packet, plus 62 bytes of empty padding, plus some
+ * padding and the "struct skb_shared_info".
+ *
+ * With 64K pages, a large buffer thus needs 32+92+4+2+1536+62+88
+ * bytes, or 1816 bytes, which fits comfortably into 2048 bytes.
+ *
+ * With 64K pages, a small buffer thus needs 32+92+4+2+126+88
+ * bytes, or 344 bytes, which means we are wasting 64+ bytes, and
+ * could presumably increase the size of small buffers.
+ *
+ * With 4K pages, a large buffer thus needs 32+92+4+2+1536+62+268
+ * bytes, or 1996 bytes, which fits comfortably into 2048 bytes.
+ *
+ * With 4K pages, a small buffer thus needs 32+92+4+2+126+268
+ * bytes, or 524 bytes, which is annoyingly wasteful.
+ *
+ * Maybe we should increase LIPP_SMALL_PACKET_SIZE to 192?
+ *
+ * ISSUE: Maybe we should increase "NET_SKB_PAD" to 64?
  */
 static bool tile_net_provide_needed_buffer(struct tile_net_cpu *info,
 					   bool small)
 {
-	/* ISSUE: What should we use here? */
+#if TILE_NET_MTU <= 1536
+	/* Without "jumbo", 2 + 1536 should be sufficient. */
+	unsigned int large_size = NET_IP_ALIGN + 1536;
+#else
+	/* ISSUE: This has not been tested. */
 	unsigned int large_size = NET_IP_ALIGN + TILE_NET_MTU + 100;
+#endif
 
-	/* Round up to ensure to avoid "false sharing" with last cache line. */
-	unsigned int buffer_size =
+	/* Avoid "false sharing" with last cache line. */
+	/* ISSUE: This is already done by "dev_alloc_skb()". */
+	unsigned int len =
 		 (((small ? LIPP_SMALL_PACKET_SIZE : large_size) +
 		   CHIP_L2_LINE_SIZE() - 1) & -CHIP_L2_LINE_SIZE());
 
-	/*
-	 * ISSUE: Since CPAs are 38 bits, and we can only encode the
-	 * high 31 bits in a "linux_buffer_t", the low 7 bits must be
-	 * zero, and thus, we must align the actual "va" mod 128.
-	 */
-	const unsigned long align = 128;
+	unsigned int padding = 128 - NET_SKB_PAD;
+	unsigned int align;
 
 	struct sk_buff *skb;
 	void *va;
 
 	struct sk_buff **skb_ptr;
 
-	/* Note that "dev_alloc_skb()" adds NET_SKB_PAD more bytes, */
-	/* and also "reserves" that many bytes. */
-	/* ISSUE: Can we "share" the NET_SKB_PAD bytes with "skb_ptr"? */
-	int len = sizeof(*skb_ptr) + align + buffer_size;
-
-	while (1) {
-
-		/* Allocate (or fail). */
-		skb = dev_alloc_skb(len);
-		if (skb == NULL)
-			return false;
-
-		/* Make room for a back-pointer to 'skb'. */
-		skb_reserve(skb, sizeof(*skb_ptr));
+	/* Request 96 extra bytes for alignment purposes. */
+	skb = dev_alloc_skb(len + padding);
+	if (skb == NULL)
+		return false;
 
-		/* Make sure we are aligned. */
-		skb_reserve(skb, -(long)skb->data & (align - 1));
+	/* Skip 32 or 96 bytes to align "data" mod 128. */
+	align = -(long)skb->data & (128 - 1);
+	BUG_ON(align > padding);
+	skb_reserve(skb, align);
 
-		/* This address is given to IPP. */
-		va = skb->data;
+	/* This address is given to IPP. */
+	va = skb->data;
 
-		if (small)
-			break;
+	/* Buffers must not span a huge page. */
+	BUG_ON(((((long)va & ~HPAGE_MASK) + len) & HPAGE_MASK) != 0);
 
-		/* ISSUE: This has never been observed! */
-		/* Large buffers must not span a huge page. */
-		if (((((long)va & ~HPAGE_MASK) + 1535) & HPAGE_MASK) == 0)
-			break;
-		pr_err("Leaking unaligned linux buffer at %p.\n", va);
+#ifdef TILE_NET_PARANOIA
+#if CHIP_HAS_CBOX_HOME_MAP()
+	if (hash_default) {
+		HV_PTE pte = *virt_to_pte(current->mm, (unsigned long)va);
+		if (hv_pte_get_mode(pte) != HV_PTE_MODE_CACHE_HASH_L3)
+			panic("Non-HFH ingress buffer! VA=%p Mode=%d PTE=%llx",
+			      va, hv_pte_get_mode(pte), hv_pte_val(pte));
 	}
+#endif
+#endif
+
+	/* Invalidate the packet buffer. */
+	if (!hash_default)
+		__inv_buffer(va, len);
 
 	/* Skip two bytes to satisfy LIPP assumptions. */
 	/* Note that this aligns IP on a 16 byte boundary. */
@@ -415,23 +472,9 @@ static bool tile_net_provide_needed_buffer(struct tile_net_cpu *info,
 	skb_ptr = va - sizeof(*skb_ptr);
 	*skb_ptr = skb;
 
-	/* Invalidate the packet buffer. */
-	if (!hash_default)
-		__inv_buffer(skb->data, buffer_size);
-
 	/* Make sure "skb_ptr" has been flushed. */
 	__insn_mf();
 
-#ifdef TILE_NET_PARANOIA
-#if CHIP_HAS_CBOX_HOME_MAP()
-	if (hash_default) {
-		HV_PTE pte = *virt_to_pte(current->mm, (unsigned long)va);
-		if (hv_pte_get_mode(pte) != HV_PTE_MODE_CACHE_HASH_L3)
-			panic("Non-coherent ingress buffer!");
-	}
-#endif
-#endif
-
 	/* Provide the new buffer. */
 	tile_net_provide_linux_buffer(info, va, small);
 
@@ -469,47 +512,63 @@ oops:
  * Grab some LEPP completions, and store them in "comps", of size
  * "comps_size", and return the number of completions which were
  * stored, so the caller can free them.
- *
- * If "pending" is not NULL, it will be set to true if there might
- * still be some pending completions caused by this tile, else false.
  */
-static unsigned int tile_net_lepp_grab_comps(struct net_device *dev,
+static unsigned int tile_net_lepp_grab_comps(lepp_queue_t *eq,
 					     struct sk_buff *comps[],
 					     unsigned int comps_size,
-					     bool *pending)
+					     unsigned int min_size)
 {
-	struct tile_net_priv *priv = netdev_priv(dev);
-
-	lepp_queue_t *eq = priv->epp_queue;
-
 	unsigned int n = 0;
 
-	unsigned int comp_head;
-	unsigned int comp_busy;
-	unsigned int comp_tail;
-
-	spin_lock(&priv->comp_lock);
-
-	comp_head = eq->comp_head;
-	comp_busy = eq->comp_busy;
-	comp_tail = eq->comp_tail;
+	unsigned int comp_head = eq->comp_head;
+	unsigned int comp_busy = eq->comp_busy;
 
 	while (comp_head != comp_busy && n < comps_size) {
 		comps[n++] = eq->comps[comp_head];
 		LEPP_QINC(comp_head);
 	}
 
-	if (pending != NULL)
-		*pending = (comp_head != comp_tail);
+	if (n < min_size)
+		return 0;
 
 	eq->comp_head = comp_head;
 
-	spin_unlock(&priv->comp_lock);
-
 	return n;
 }
 
 
+/*
+ * Free some comps, and return true iff there are still some pending.
+ */
+static bool tile_net_lepp_free_comps(struct net_device *dev, bool all)
+{
+	struct tile_net_priv *priv = netdev_priv(dev);
+
+	lepp_queue_t *eq = priv->eq;
+
+	struct sk_buff *olds[64];
+	unsigned int wanted = 64;
+	unsigned int i, n;
+	bool pending;
+
+	spin_lock(&priv->eq_lock);
+
+	if (all)
+		eq->comp_busy = eq->comp_tail;
+
+	n = tile_net_lepp_grab_comps(eq, olds, wanted, 0);
+
+	pending = (eq->comp_head != eq->comp_tail);
+
+	spin_unlock(&priv->eq_lock);
+
+	for (i = 0; i < n; i++)
+		kfree_skb(olds[i]);
+
+	return pending;
+}
+
+
 /*
  * Make sure the egress timer is scheduled.
  *
@@ -544,21 +603,11 @@ static void tile_net_handle_egress_timer(unsigned long arg)
 	struct tile_net_cpu *info = (struct tile_net_cpu *)arg;
 	struct net_device *dev = info->napi.dev;
 
-	struct sk_buff *olds[32];
-	unsigned int wanted = 32;
-	unsigned int i, nolds = 0;
-	bool pending;
-
 	/* The timer is no longer scheduled. */
 	info->egress_timer_scheduled = false;
 
-	nolds = tile_net_lepp_grab_comps(dev, olds, wanted, &pending);
-
-	for (i = 0; i < nolds; i++)
-		kfree_skb(olds[i]);
-
-	/* Reschedule timer if needed. */
-	if (pending)
+	/* Free comps, and reschedule timer if more are pending. */
+	if (tile_net_lepp_free_comps(dev, false))
 		tile_net_schedule_egress_timer(info);
 }
 
@@ -636,8 +685,39 @@ static bool is_dup_ack(char *s1, char *s2, unsigned int len)
 
 
 
+static void tile_net_discard_aux(struct tile_net_cpu *info, int index)
+{
+	struct tile_netio_queue *queue = &info->queue;
+	netio_queue_impl_t *qsp = queue->__system_part;
+	netio_queue_user_impl_t *qup = &queue->__user_part;
+
+	int index2_aux = index + sizeof(netio_pkt_t);
+	int index2 =
+		((index2_aux ==
+		  qsp->__packet_receive_queue.__last_packet_plus_one) ?
+		 0 : index2_aux);
+
+	netio_pkt_t *pkt = (netio_pkt_t *)((unsigned long) &qsp[1] + index);
+
+	/* Extract the "linux_buffer_t". */
+	unsigned int buffer = pkt->__packet.word;
+
+	/* Convert "linux_buffer_t" to "va". */
+	void *va = __va((phys_addr_t)(buffer >> 1) << 7);
+
+	/* Acquire the associated "skb". */
+	struct sk_buff **skb_ptr = va - sizeof(*skb_ptr);
+	struct sk_buff *skb = *skb_ptr;
+
+	kfree_skb(skb);
+
+	/* Consume this packet. */
+	qup->__packet_receive_read = index2;
+}
+
+
 /*
- * Like "tile_net_handle_packets()", but just discard packets.
+ * Like "tile_net_poll()", but just discard packets.
  */
 static void tile_net_discard_packets(struct net_device *dev)
 {
@@ -650,32 +730,8 @@ static void tile_net_discard_packets(struct net_device *dev)
 
 	while (qup->__packet_receive_read !=
 	       qsp->__packet_receive_queue.__packet_write) {
-
 		int index = qup->__packet_receive_read;
-
-		int index2_aux = index + sizeof(netio_pkt_t);
-		int index2 =
-			((index2_aux ==
-			  qsp->__packet_receive_queue.__last_packet_plus_one) ?
-			 0 : index2_aux);
-
-		netio_pkt_t *pkt = (netio_pkt_t *)
-			((unsigned long) &qsp[1] + index);
-
-		/* Extract the "linux_buffer_t". */
-		unsigned int buffer = pkt->__packet.word;
-
-		/* Convert "linux_buffer_t" to "va". */
-		void *va = __va((phys_addr_t)(buffer >> 1) << 7);
-
-		/* Acquire the associated "skb". */
-		struct sk_buff **skb_ptr = va - sizeof(*skb_ptr);
-		struct sk_buff *skb = *skb_ptr;
-
-		kfree_skb(skb);
-
-		/* Consume this packet. */
-		qup->__packet_receive_read = index2;
+		tile_net_discard_aux(info, index);
 	}
 }
 
@@ -704,7 +760,8 @@ static bool tile_net_poll_aux(struct tile_net_cpu *info, int index)
 
 	netio_pkt_metadata_t *metadata = NETIO_PKT_METADATA(pkt);
 
-	/* Extract the packet size. */
+	/* Extract the packet size.  FIXME: Shouldn't the second line */
+	/* get subtracted?  Mostly moot, since it should be "zero". */
 	unsigned long len =
 		(NETIO_PKT_CUSTOM_LENGTH(pkt) +
 		 NET_IP_ALIGN - NETIO_PACKET_PADDING);
@@ -722,15 +779,6 @@ static bool tile_net_poll_aux(struct tile_net_cpu *info, int index)
 	/* Compare to "NETIO_PKT_CUSTOM_DATA(pkt)". */
 	unsigned char *buf = va + NET_IP_ALIGN;
 
-#ifdef IGNORE_DUP_ACKS
-
-	static int other;
-	static int final;
-	static int keep;
-	static int skip;
-
-#endif
-
 	/* Invalidate the packet buffer. */
 	if (!hash_default)
 		__inv_buffer(buf, len);
@@ -745,16 +793,8 @@ static bool tile_net_poll_aux(struct tile_net_cpu *info, int index)
 #ifdef TILE_NET_VERIFY_INGRESS
 	if (!NETIO_PKT_L4_CSUM_CORRECT_M(metadata, pkt) &&
 	    NETIO_PKT_L4_CSUM_CALCULATED_M(metadata, pkt)) {
-		/*
-		 * FIXME: This complains about UDP packets
-		 * with a "zero" checksum (bug 6624).
-		 */
-#ifdef TILE_NET_PANIC_ON_BAD
-		dump_packet(buf, len, "rx");
-		panic("Bad L4 checksum.");
-#else
+		/* Bug 6624: Includes UDP packets with a "zero" checksum. */
 		pr_warning("Bad L4 checksum on %d byte packet.\n", len);
-#endif
 	}
 	if (!NETIO_PKT_L3_CSUM_CORRECT_M(metadata, pkt) &&
 	    NETIO_PKT_L3_CSUM_CALCULATED_M(metadata, pkt)) {
@@ -769,90 +809,29 @@ static bool tile_net_poll_aux(struct tile_net_cpu *info, int index)
 		}
 		break;
 	case NETIO_PKT_STATUS_BAD:
-#ifdef TILE_NET_PANIC_ON_BAD
-		dump_packet(buf, len, "rx");
-		panic("Unexpected BAD packet.");
-#else
-		pr_warning("Unexpected BAD %d byte packet.\n", len);
-#endif
+		pr_warning("Unexpected BAD %ld byte packet.\n", len);
 	}
 #endif
 
 	filter = 0;
 
+	/* ISSUE: Filter TCP packets with "bad" checksums? */
+
 	if (!(dev->flags & IFF_UP)) {
 		/* Filter packets received before we're up. */
 		filter = 1;
+	} else if (NETIO_PKT_STATUS_M(metadata, pkt) == NETIO_PKT_STATUS_BAD) {
+		/* Filter "truncated" packets. */
+		filter = 1;
 	} else if (!(dev->flags & IFF_PROMISC)) {
-		/*
-		 * FIXME: Implement HW multicast filter.
-		 */
-		if (is_unicast_ether_addr(buf)) {
+		/* FIXME: Implement HW multicast filter. */
+		if (!is_multicast_ether_addr(buf)) {
 			/* Filter packets not for our address. */
 			const u8 *mine = dev->dev_addr;
 			filter = compare_ether_addr(mine, buf);
 		}
 	}
 
-#ifdef IGNORE_DUP_ACKS
-
-	if (len != 66) {
-		/* FIXME: Must check "is_tcp_ack(buf, len)" somehow. */
-
-		other++;
-
-	} else if (index2 ==
-		   qsp->__packet_receive_queue.__packet_write) {
-
-		final++;
-
-	} else {
-
-		netio_pkt_t *pkt2 = (netio_pkt_t *)
-			((unsigned long) &qsp[1] + index2);
-
-		netio_pkt_metadata_t *metadata2 =
-			NETIO_PKT_METADATA(pkt2);
-
-		/* Extract the packet size. */
-		unsigned long len2 =
-			(NETIO_PKT_CUSTOM_LENGTH(pkt2) +
-			 NET_IP_ALIGN - NETIO_PACKET_PADDING);
-
-		if (len2 == 66 &&
-		    NETIO_PKT_FLOW_HASH_M(metadata, pkt) ==
-		    NETIO_PKT_FLOW_HASH_M(metadata2, pkt2)) {
-
-			/* Extract the "linux_buffer_t". */
-			unsigned int buffer2 = pkt2->__packet.word;
-
-			/* Convert "linux_buffer_t" to "va". */
-			void *va2 =
-				__va((phys_addr_t)(buffer2 >> 1) << 7);
-
-			/* Extract the packet data pointer. */
-			/* Compare to "NETIO_PKT_CUSTOM_DATA(pkt)". */
-			unsigned char *buf2 = va2 + NET_IP_ALIGN;
-
-			/* Invalidate the packet buffer. */
-			if (!hash_default)
-				__inv_buffer(buf2, len2);
-
-			if (is_dup_ack(buf, buf2, len)) {
-				skip++;
-				filter = 1;
-			} else {
-				keep++;
-			}
-		}
-	}
-
-	if (net_ratelimit())
-		pr_info("Other %d Final %d Keep %d Skip %d.\n",
-			other, final, keep, skip);
-
-#endif
-
 	if (filter) {
 
 		/* ISSUE: Update "drop" statistics? */
@@ -877,10 +856,7 @@ static bool tile_net_poll_aux(struct tile_net_cpu *info, int index)
 		/* NOTE: This call also sets "skb->dev = dev". */
 		skb->protocol = eth_type_trans(skb, dev);
 
-		/* ISSUE: Discard corrupt packets? */
-		/* ISSUE: Discard packets with bad checksums? */
-
-		/* Avoid recomputing TCP/UDP checksums. */
+		/* Avoid recomputing "good" TCP/UDP checksums. */
 		if (NETIO_PKT_L4_CSUM_CORRECT_M(metadata, pkt))
 			skb->ip_summed = CHECKSUM_UNNECESSARY;
 
@@ -912,9 +888,14 @@ static bool tile_net_poll_aux(struct tile_net_cpu *info, int index)
 /*
  * Handle some packets for the given device on the current CPU.
  *
- * ISSUE: The "rotting packet" race condition occurs if a packet
- * arrives after the queue appears to be empty, and before the
- * hypervisor interrupt is re-enabled.
+ * If "tile_net_stop()" is called on some other tile while this
+ * function is running, we will return, hopefully before that
+ * other tile asks us to call "napi_disable()".
+ *
+ * The "rotting packet" race condition occurs if a packet arrives
+ * during the extremely narrow window between the queue appearing to
+ * be empty, and the ingress interrupt being re-enabled.  This happens
+ * a LOT under heavy network load.
  */
 static int tile_net_poll(struct napi_struct *napi, int budget)
 {
@@ -928,7 +909,7 @@ static int tile_net_poll(struct napi_struct *napi, int budget)
 
 	unsigned int work = 0;
 
-	while (1) {
+	while (priv->active) {
 		int index = qup->__packet_receive_read;
 		if (index == qsp->__packet_receive_queue.__packet_write)
 			break;
@@ -941,19 +922,24 @@ static int tile_net_poll(struct napi_struct *napi, int budget)
 
 	napi_complete(&info->napi);
 
-	/* Re-enable hypervisor interrupts. */
+	if (!priv->active)
+		goto done;
+
+	/* Re-enable the ingress interrupt. */
 	enable_percpu_irq(priv->intr_id);
 
-	/* HACK: Avoid the "rotting packet" problem. */
+	/* HACK: Avoid the "rotting packet" problem (see above). */
 	if (qup->__packet_receive_read !=
-	    qsp->__packet_receive_queue.__packet_write)
-		napi_schedule(&info->napi);
-
-	/* ISSUE: Handle completions? */
+	    qsp->__packet_receive_queue.__packet_write) {
+		/* ISSUE: Sometimes this returns zero, presumably */
+		/* because an interrupt was handled for this tile. */
+		(void)napi_reschedule(&info->napi);
+	}
 
 done:
 
-	tile_net_provide_needed_buffers(info);
+	if (priv->active)
+		tile_net_provide_needed_buffers(info);
 
 	return work;
 }
@@ -961,6 +947,12 @@ done:
 
 /*
  * Handle an ingress interrupt for the given device on the current cpu.
+ *
+ * ISSUE: Sometimes this gets called after "disable_percpu_irq()" has
+ * been called!  This is probably due to "pending hypervisor downcalls".
+ *
+ * ISSUE: Is there any race condition between the "napi_schedule()" here
+ * and the "napi_complete()" call above?
  */
 static irqreturn_t tile_net_handle_ingress_interrupt(int irq, void *dev_ptr)
 {
@@ -969,9 +961,15 @@ static irqreturn_t tile_net_handle_ingress_interrupt(int irq, void *dev_ptr)
 	int my_cpu = smp_processor_id();
 	struct tile_net_cpu *info = priv->cpu[my_cpu];
 
-	/* Disable hypervisor interrupt. */
+	/* Disable the ingress interrupt. */
 	disable_percpu_irq(priv->intr_id);
 
+	/* Ignore unwanted interrupts. */
+	if (!priv->active)
+		return IRQ_HANDLED;
+
+	/* ISSUE: Sometimes "info->napi_enabled" is false here. */
+
 	napi_schedule(&info->napi);
 
 	return IRQ_HANDLED;
@@ -1005,8 +1003,7 @@ static int tile_net_open_aux(struct net_device *dev)
 	 */
 	{
 		int epp_home = hv_lotar_to_cpu(epp_lotar);
-		struct page *page = virt_to_page(priv->epp_queue);
-		homecache_change_page_home(page, 0, epp_home);
+		homecache_change_page_home(priv->eq_pages, EQ_ORDER, epp_home);
 	}
 
 	/*
@@ -1015,9 +1012,9 @@ static int tile_net_open_aux(struct net_device *dev)
 	{
 		netio_ipp_address_t ea = {
 			.va = 0,
-			.pa = __pa(priv->epp_queue),
+			.pa = __pa(priv->eq),
 			.pte = hv_pte(0),
-			.size = PAGE_SIZE,
+			.size = EQ_SIZE,
 		};
 		ea.pte = hv_pte_set_lotar(ea.pte, epp_lotar);
 		ea.pte = hv_pte_set_mode(ea.pte, HV_PTE_MODE_CACHE_TILE_L3);
@@ -1043,7 +1040,7 @@ static int tile_net_open_aux(struct net_device *dev)
 
 
 /*
- * Register with hypervisor on each CPU.
+ * Register with hypervisor on the current CPU.
  *
  * Strangely, this function does important things even if it "fails",
  * which is especially common if the link is not up yet.  Hopefully
@@ -1092,7 +1089,8 @@ static void tile_net_register(void *dev_ptr)
 	priv->cpu[my_cpu] = info;
 
 	/*
-	 * Register ourselves with the IPP.
+	 * Register ourselves with LIPP.  This does a lot of stuff,
+	 * including invoking the LIPP registration code.
 	 */
 	ret = hv_dev_pwrite(priv->hv_devhdl, 0,
 			    (HV_VirtAddr)&config,
@@ -1101,8 +1099,11 @@ static void tile_net_register(void *dev_ptr)
 	PDEBUG("hv_dev_pwrite(NETIO_IPP_INPUT_REGISTER_OFF) returned %d\n",
 	       ret);
 	if (ret < 0) {
-		printk(KERN_DEBUG "hv_dev_pwrite NETIO_IPP_INPUT_REGISTER_OFF"
-		       " failure %d\n", ret);
+		if (ret != NETIO_LINK_DOWN) {
+			printk(KERN_DEBUG "hv_dev_pwrite "
+			       "NETIO_IPP_INPUT_REGISTER_OFF failure %d\n",
+			       ret);
+		}
 		info->link_down = (ret == NETIO_LINK_DOWN);
 		return;
 	}
@@ -1145,15 +1146,47 @@ static void tile_net_register(void *dev_ptr)
 			   NETIO_IPP_GET_FASTIO_OFF);
 	PDEBUG("hv_dev_pread(NETIO_IPP_GET_FASTIO_OFF) returned %d\n", ret);
 
-	netif_napi_add(dev, &info->napi, tile_net_poll, 64);
-
 	/* Now we are registered. */
 	info->registered = true;
 }
 
 
 /*
- * Unregister with hypervisor on each CPU.
+ * Deregister with hypervisor on the current CPU.
+ *
+ * This simply discards all our credits, so no more packets will be
+ * delivered to this tile.  There may still be packets in our queue.
+ *
+ * Also, disable the ingress interrupt.
+ */
+static void tile_net_deregister(void *dev_ptr)
+{
+	struct net_device *dev = (struct net_device *)dev_ptr;
+	struct tile_net_priv *priv = netdev_priv(dev);
+	int my_cpu = smp_processor_id();
+	struct tile_net_cpu *info = priv->cpu[my_cpu];
+
+	/* Disable the ingress interrupt. */
+	disable_percpu_irq(priv->intr_id);
+
+	/* Do nothing else if not registered. */
+	if (info == NULL || !info->registered)
+		return;
+
+	{
+		struct tile_netio_queue *queue = &info->queue;
+		netio_queue_user_impl_t *qup = &queue->__user_part;
+
+		/* Discard all our credits. */
+		__netio_fastio_return_credits(qup->__fastio_index, -1);
+	}
+}
+
+
+/*
+ * Unregister with hypervisor on the current CPU.
+ *
+ * Also, disable the ingress interrupt.
  */
 static void tile_net_unregister(void *dev_ptr)
 {
@@ -1162,35 +1195,23 @@ static void tile_net_unregister(void *dev_ptr)
 	int my_cpu = smp_processor_id();
 	struct tile_net_cpu *info = priv->cpu[my_cpu];
 
-	int ret = 0;
+	int ret;
 	int dummy = 0;
 
-	/* Do nothing if never registered. */
-	if (info == NULL)
-		return;
+	/* Disable the ingress interrupt. */
+	disable_percpu_irq(priv->intr_id);
 
-	/* Do nothing if already unregistered. */
-	if (!info->registered)
+	/* Do nothing else if not registered. */
+	if (info == NULL || !info->registered)
 		return;
 
-	/*
-	 * Unregister ourselves with LIPP.
-	 */
+	/* Unregister ourselves with LIPP/LEPP. */
 	ret = hv_dev_pwrite(priv->hv_devhdl, 0, (HV_VirtAddr)&dummy,
 			    sizeof(dummy), NETIO_IPP_INPUT_UNREGISTER_OFF);
-	PDEBUG("hv_dev_pwrite(NETIO_IPP_INPUT_UNREGISTER_OFF) returned %d\n",
-	       ret);
-	if (ret < 0) {
-		/* FIXME: Just panic? */
-		pr_err("hv_dev_pwrite NETIO_IPP_INPUT_UNREGISTER_OFF"
-		       " failure %d\n", ret);
-	}
+	if (ret < 0)
+		panic("Failed to unregister with LIPP/LEPP!\n");
 
-	/*
-	 * Discard all packets still in our NetIO queue.  Hopefully,
-	 * once the unregister call is complete, there will be no
-	 * packets still in flight on the IDN.
-	 */
+	/* Discard all packets still in our NetIO queue. */
 	tile_net_discard_packets(dev);
 
 	/* Reset state. */
@@ -1200,11 +1221,6 @@ static void tile_net_unregister(void *dev_ptr)
 	/* Cancel egress timer. */
 	del_timer(&info->egress_timer);
 	info->egress_timer_scheduled = false;
-
-	netif_napi_del(&info->napi);
-
-	/* Now we are unregistered. */
-	info->registered = false;
 }
 
 
@@ -1212,18 +1228,28 @@ static void tile_net_unregister(void *dev_ptr)
  * Helper function for "tile_net_stop()".
  *
  * Also used to handle registration failure in "tile_net_open_inner()",
- * when "fully_opened" is known to be false, and the various extra
- * steps in "tile_net_stop()" are not necessary.  ISSUE: It might be
- * simpler if we could just call "tile_net_stop()" anyway.
+ * when the various extra steps in "tile_net_stop()" are not necessary.
  */
 static void tile_net_stop_aux(struct net_device *dev)
 {
 	struct tile_net_priv *priv = netdev_priv(dev);
+	int i;
 
 	int dummy = 0;
 
-	/* Unregister all tiles, so LIPP will stop delivering packets. */
+	/*
+	 * Unregister all tiles, so LIPP will stop delivering packets.
+	 * Also, delete all the "napi" objects (sequentially, to protect
+	 * "dev->napi_list").
+	 */
 	on_each_cpu(tile_net_unregister, (void *)dev, 1);
+	for_each_online_cpu(i) {
+		struct tile_net_cpu *info = priv->cpu[i];
+		if (info != NULL && info->registered) {
+			netif_napi_del(&info->napi);
+			info->registered = false;
+		}
+	}
 
 	/* Stop LIPP/LEPP. */
 	if (hv_dev_pwrite(priv->hv_devhdl, 0, (HV_VirtAddr)&dummy,
@@ -1235,18 +1261,15 @@ static void tile_net_stop_aux(struct net_device *dev)
 
 
 /*
- * Disable ingress interrupts for the given device on the current cpu.
+ * Disable NAPI for the given device on the current cpu.
  */
-static void tile_net_disable_intr(void *dev_ptr)
+static void tile_net_stop_disable(void *dev_ptr)
 {
 	struct net_device *dev = (struct net_device *)dev_ptr;
 	struct tile_net_priv *priv = netdev_priv(dev);
 	int my_cpu = smp_processor_id();
 	struct tile_net_cpu *info = priv->cpu[my_cpu];
 
-	/* Disable hypervisor interrupt. */
-	disable_percpu_irq(priv->intr_id);
-
 	/* Disable NAPI if needed. */
 	if (info != NULL && info->napi_enabled) {
 		napi_disable(&info->napi);
@@ -1256,21 +1279,24 @@ static void tile_net_disable_intr(void *dev_ptr)
 
 
 /*
- * Enable ingress interrupts for the given device on the current cpu.
+ * Enable NAPI and the ingress interrupt for the given device
+ * on the current cpu.
+ *
+ * ISSUE: Only do this for "network cpus"?
  */
-static void tile_net_enable_intr(void *dev_ptr)
+static void tile_net_open_enable(void *dev_ptr)
 {
 	struct net_device *dev = (struct net_device *)dev_ptr;
 	struct tile_net_priv *priv = netdev_priv(dev);
 	int my_cpu = smp_processor_id();
 	struct tile_net_cpu *info = priv->cpu[my_cpu];
 
-	/* Enable hypervisor interrupt. */
-	enable_percpu_irq(priv->intr_id);
-
 	/* Enable NAPI. */
 	napi_enable(&info->napi);
 	info->napi_enabled = true;
+
+	/* Enable the ingress interrupt. */
+	enable_percpu_irq(priv->intr_id);
 }
 
 
@@ -1288,8 +1314,9 @@ static int tile_net_open_inner(struct net_device *dev)
 	int my_cpu = smp_processor_id();
 	struct tile_net_cpu *info;
 	struct tile_netio_queue *queue;
-	unsigned int irq;
+	int result = 0;
 	int i;
+	int dummy = 0;
 
 	/*
 	 * First try to register just on the local CPU, and handle any
@@ -1307,42 +1334,52 @@ static int tile_net_open_inner(struct net_device *dev)
 	/*
 	 * Now register everywhere else.  If any registration fails,
 	 * even for "link down" (which might not be possible), we
-	 * clean up using "tile_net_stop_aux()".
+	 * clean up using "tile_net_stop_aux()".  Also, add all the
+	 * "napi" objects (sequentially, to protect "dev->napi_list").
+	 * ISSUE: Only use "netif_napi_add()" for "network cpus"?
 	 */
 	smp_call_function(tile_net_register, (void *)dev, 1);
 	for_each_online_cpu(i) {
-		if (!priv->cpu[i]->registered) {
-			tile_net_stop_aux(dev);
-			return -EAGAIN;
-		}
+		struct tile_net_cpu *info = priv->cpu[i];
+		if (info->registered)
+			netif_napi_add(dev, &info->napi, tile_net_poll, 64);
+		else
+			result = -EAGAIN;
+	}
+	if (result != 0) {
+		tile_net_stop_aux(dev);
+		return result;
 	}
 
 	queue = &info->queue;
 
-	/*
-	 * Set the device intr bit mask.
-	 * The tile_net_register above sets per tile __intr_id.
-	 */
-	priv->intr_id = queue->__system_part->__intr_id;
-	BUG_ON(!priv->intr_id);
-
-	/*
-	 * Register the device interrupt handler.
-	 * The __ffs() function returns the index into the interrupt handler
-	 * table from the interrupt bit mask which should have one bit
-	 * and one bit only set.
-	 */
-	irq = __ffs(priv->intr_id);
-	tile_irq_activate(irq, TILE_IRQ_PERCPU);
-	BUG_ON(request_irq(irq, tile_net_handle_ingress_interrupt,
-			   0, dev->name, (void *)dev) != 0);
+	if (priv->intr_id == 0) {
+		unsigned int irq;
 
-	/* ISSUE: How could "priv->fully_opened" ever be "true" here? */
-
-	if (!priv->fully_opened) {
+		/*
+		 * Acquire the irq allocated by the hypervisor.  Every
+		 * queue gets the same irq.  The "__intr_id" field is
+		 * "1 << irq", so we use "__ffs()" to extract "irq".
+		 */
+		priv->intr_id = queue->__system_part->__intr_id;
+		BUG_ON(priv->intr_id == 0);
+		irq = __ffs(priv->intr_id);
 
-		int dummy = 0;
+		/*
+		 * Register the ingress interrupt handler for this
+		 * device, permanently.
+		 *
+		 * We used to call "free_irq()" in "tile_net_stop()",
+		 * and then re-register the handler here every time,
+		 * but that caused DNP errors in "handle_IRQ_event()"
+		 * because "desc->action" was NULL.  See bug 9143.
+		 */
+		tile_irq_activate(irq, TILE_IRQ_PERCPU);
+		BUG_ON(request_irq(irq, tile_net_handle_ingress_interrupt,
+				   0, dev->name, (void *)dev) != 0);
+	}
 
+	{
 		/* Allocate initial buffers. */
 
 		int max_buffers =
@@ -1359,18 +1396,21 @@ static int tile_net_open_inner(struct net_device *dev)
 		if (info->num_needed_small_buffers != 0 ||
 		    info->num_needed_large_buffers != 0)
 			panic("Insufficient memory for buffer stack!");
+	}
 
-		/* Start LIPP/LEPP and activate "ingress" at the shim. */
-		if (hv_dev_pwrite(priv->hv_devhdl, 0, (HV_VirtAddr)&dummy,
-				  sizeof(dummy), NETIO_IPP_INPUT_INIT_OFF) < 0)
-			panic("Failed to activate the LIPP Shim!\n");
+	/* We are about to be active. */
+	priv->active = true;
 
-		priv->fully_opened = 1;
-	}
+	/* Make sure "active" is visible to all tiles. */
+	mb();
 
-	/* On each tile, enable the hypervisor to trigger interrupts. */
-	/* ISSUE: Do this before starting LIPP/LEPP? */
-	on_each_cpu(tile_net_enable_intr, (void *)dev, 1);
+	/* On each tile, enable NAPI and the ingress interrupt. */
+	on_each_cpu(tile_net_open_enable, (void *)dev, 1);
+
+	/* Start LIPP/LEPP and activate "ingress" at the shim. */
+	if (hv_dev_pwrite(priv->hv_devhdl, 0, (HV_VirtAddr)&dummy,
+			  sizeof(dummy), NETIO_IPP_INPUT_INIT_OFF) < 0)
+		panic("Failed to activate the LIPP Shim!\n");
 
 	/* Start our transmit queue. */
 	netif_start_queue(dev);
@@ -1396,9 +1436,9 @@ static void tile_net_open_retry(struct work_struct *w)
 	 * ourselves to try again later; otherwise, tell Linux we now have
 	 * a working link.  ISSUE: What if the return value is negative?
 	 */
-	if (tile_net_open_inner(priv->dev))
-		schedule_delayed_work_on(singlethread_cpu, &priv->retry_work,
-					 TILE_NET_RETRY_INTERVAL);
+	if (tile_net_open_inner(priv->dev) != 0)
+		schedule_delayed_work(&priv->retry_work,
+				      TILE_NET_RETRY_INTERVAL);
 	else
 		netif_carrier_on(priv->dev);
 }
@@ -1412,8 +1452,8 @@ static void tile_net_open_retry(struct work_struct *w)
  * The open entry point is called when a network interface is made
  * active by the system (IFF_UP).  At this point all resources needed
  * for transmit and receive operations are allocated, the interrupt
- * handler is registered with the OS, the watchdog timer is started,
- * and the stack is notified that the interface is ready.
+ * handler is registered with the OS (if needed), the watchdog timer
+ * is started, and the stack is notified that the interface is ready.
  *
  * If the actual link is not available yet, then we tell Linux that
  * we have no carrier, and we keep checking until the link comes up.
@@ -1468,6 +1508,10 @@ static int tile_net_open(struct net_device *dev)
 #endif
 
 		priv->partly_opened = 1;
+
+	} else {
+		/* FIXME: Is this possible? */
+		/* printk("Already partly opened.\n"); */
 	}
 
 	/*
@@ -1487,57 +1531,17 @@ static int tile_net_open(struct net_device *dev)
 	 * and then remember to try again later.
 	 */
 	netif_carrier_off(dev);
-	schedule_delayed_work_on(singlethread_cpu, &priv->retry_work,
-				 TILE_NET_RETRY_INTERVAL);
+	schedule_delayed_work(&priv->retry_work, TILE_NET_RETRY_INTERVAL);
 
 	return 0;
 }
 
 
-/*
- * Disables a network interface.
- *
- * Returns 0, this is not allowed to fail.
- *
- * The close entry point is called when an interface is de-activated
- * by the OS.  The hardware is still under the drivers control, but
- * needs to be disabled.  A global MAC reset is issued to stop the
- * hardware, and all transmit and receive resources are freed.
- *
- * ISSUE: Can this can be called while "tile_net_poll()" is running?
- */
-static int tile_net_stop(struct net_device *dev)
+static int tile_net_drain_lipp_buffers(struct tile_net_priv *priv)
 {
-	struct tile_net_priv *priv = netdev_priv(dev);
-
-	bool pending = true;
-
-	PDEBUG("tile_net_stop()\n");
-
-	/* ISSUE: Only needed if not yet fully open. */
-	cancel_delayed_work_sync(&priv->retry_work);
-
-	/* Can't transmit any more. */
-	netif_stop_queue(dev);
-
-	/*
-	 * Disable hypervisor interrupts on each tile.
-	 */
-	on_each_cpu(tile_net_disable_intr, (void *)dev, 1);
-
-	/*
-	 * Unregister the interrupt handler.
-	 * The __ffs() function returns the index into the interrupt handler
-	 * table from the interrupt bit mask which should have one bit
-	 * and one bit only set.
-	 */
-	if (priv->intr_id)
-		free_irq(__ffs(priv->intr_id), dev);
-
-	/*
-	 * Drain all the LIPP buffers.
-	 */
+	int n = 0;
 
+	/* Drain all the LIPP buffers. */
 	while (true) {
 		int buffer;
 
@@ -1560,43 +1564,105 @@ static int tile_net_stop(struct net_device *dev)
 
 			kfree_skb(skb);
 		}
+
+		n++;
 	}
 
-	/* Stop LIPP/LEPP. */
-	tile_net_stop_aux(dev);
+	return n;
+}
 
 
-	priv->fully_opened = 0;
+/*
+ * Disables a network interface.
+ *
+ * Returns 0, this is not allowed to fail.
+ *
+ * The close entry point is called when an interface is de-activated
+ * by the OS.  The hardware is still under the drivers control, but
+ * needs to be disabled.  A global MAC reset is issued to stop the
+ * hardware, and all transmit and receive resources are freed.
+ *
+ * ISSUE: How closely does "netif_running(dev)" mirror "priv->active"?
+ *
+ * Before we are called by "__dev_close()", "netif_running()" will
+ * have been cleared, so no NEW calls to "tile_net_poll()" will be
+ * made by "netpoll_poll_dev()".
+ *
+ * Often, this can cause some tiles to still have packets in their
+ * queues, so we must call "tile_net_discard_packets()" later.
+ *
+ * Note that some other tile may still be INSIDE "tile_net_poll()",
+ * and in fact, many will be, if there is heavy network load.
+ *
+ * Calling "on_each_cpu(tile_net_stop_disable, (void *)dev, 1)" when
+ * any tile is still "napi_schedule()"'d will induce a horrible crash
+ * when "msleep()" is called.  This includes tiles which are inside
+ * "tile_net_poll()" which have not yet called "napi_complete()".
+ *
+ * So, we must first try to wait long enough for other tiles to finish
+ * with any current "tile_net_poll()" call, and, hopefully, to clear
+ * the "scheduled" flag.  ISSUE: It is unclear what happens to tiles
+ * which have called "napi_schedule()" but which had not yet tried to
+ * call "tile_net_poll()", or which exhausted their budget inside
+ * "tile_net_poll()" just before this function was called.
+ */
+static int tile_net_stop(struct net_device *dev)
+{
+	struct tile_net_priv *priv = netdev_priv(dev);
+
+	PDEBUG("tile_net_stop()\n");
 
+	/* Start discarding packets. */
+	priv->active = false;
+
+	/* Make sure "active" is visible to all tiles. */
+	mb();
 
 	/*
-	 * XXX: ISSUE: It appears that, in practice anyway, by the
-	 * time we get here, there are no pending completions.
+	 * On each tile, make sure no NEW packets get delivered, and
+	 * disable the ingress interrupt.
+	 *
+	 * Note that the ingress interrupt can fire AFTER this,
+	 * presumably due to packets which were recently delivered,
+	 * but it will have no effect.
 	 */
-	while (pending) {
+	on_each_cpu(tile_net_deregister, (void *)dev, 1);
 
-		struct sk_buff *olds[32];
-		unsigned int wanted = 32;
-		unsigned int i, nolds = 0;
+	/* Optimistically drain LIPP buffers. */
+	(void)tile_net_drain_lipp_buffers(priv);
 
-		nolds = tile_net_lepp_grab_comps(dev, olds,
-						 wanted, &pending);
+	/* ISSUE: Only needed if not yet fully open. */
+	cancel_delayed_work_sync(&priv->retry_work);
 
-		/* ISSUE: We have never actually seen this debug spew. */
-		if (nolds != 0)
-			pr_info("During tile_net_stop(), grabbed %d comps.\n",
-			       nolds);
+	/* Can't transmit any more. */
+	netif_stop_queue(dev);
 
-		for (i = 0; i < nolds; i++)
-			kfree_skb(olds[i]);
-	}
+	/* Disable NAPI on each tile. */
+	on_each_cpu(tile_net_stop_disable, (void *)dev, 1);
+
+	/*
+	 * Drain any remaining LIPP buffers.  NOTE: This "printk()"
+	 * has never been observed, but in theory it could happen.
+	 */
+	if (tile_net_drain_lipp_buffers(priv) != 0)
+		printk("Had to drain some extra LIPP buffers!\n");
 
+	/* Stop LIPP/LEPP. */
+	tile_net_stop_aux(dev);
+
+	/*
+	 * ISSUE: It appears that, in practice anyway, by the time we
+	 * get here, there are no pending completions, but just in case,
+	 * we free (all of) them anyway.
+	 */
+	while (tile_net_lepp_free_comps(dev, true))
+		/* loop */;
 
 	/* Wipe the EPP queue. */
-	memset(priv->epp_queue, 0, sizeof(lepp_queue_t));
+	memset(priv->eq, 0, sizeof(lepp_queue_t));
 
 	/* Evict the EPP queue. */
-	finv_buffer(priv->epp_queue, PAGE_SIZE);
+	finv_buffer(priv->eq, EQ_SIZE);
 
 	return 0;
 }
@@ -1620,7 +1686,7 @@ static unsigned int tile_net_tx_frags(lepp_frag_t *frags,
 	if (b_len != 0) {
 
 		if (!hash_default)
-			finv_buffer_remote(b_data, b_len);
+			finv_buffer_remote(b_data, b_len, 0);
 
 		cpa = __pa(b_data);
 		frags[n].cpa_lo = cpa;
@@ -1643,7 +1709,7 @@ static unsigned int tile_net_tx_frags(lepp_frag_t *frags,
 		if (!hash_default) {
 			void *va = pfn_to_kaddr(pfn) + f->page_offset;
 			BUG_ON(PageHighMem(f->page));
-			finv_buffer_remote(va, f->size);
+			finv_buffer_remote(va, f->size, 0);
 		}
 
 		cpa = ((phys_addr_t)pfn << PAGE_SHIFT) + f->page_offset;
@@ -1742,17 +1808,15 @@ static int tile_net_tx_tso(struct sk_buff *skb, struct net_device *dev)
 
 	unsigned long irqflags;
 
-	lepp_queue_t *eq = priv->epp_queue;
+	lepp_queue_t *eq = priv->eq;
 
-	struct sk_buff *olds[4];
-	unsigned int wanted = 4;
+	struct sk_buff *olds[8];
+	unsigned int wanted = 8;
 	unsigned int i, nolds = 0;
 
 	unsigned int cmd_head, cmd_tail, cmd_next;
 	unsigned int comp_tail;
 
-	unsigned int free_slots;
-
 
 	/* Paranoia. */
 	BUG_ON(skb->protocol != htons(ETH_P_IP));
@@ -1780,34 +1844,32 @@ static int tile_net_tx_tso(struct sk_buff *skb, struct net_device *dev)
 
 	/* Enqueue the command. */
 
-	spin_lock_irqsave(&priv->cmd_lock, irqflags);
+	spin_lock_irqsave(&priv->eq_lock, irqflags);
 
 	/*
 	 * Handle completions if needed to make room.
 	 * HACK: Spin until there is sufficient room.
 	 */
-	free_slots = lepp_num_free_comp_slots(eq);
-	if (free_slots < 1) {
-spin:
-		nolds += tile_net_lepp_grab_comps(dev, olds + nolds,
-						  wanted - nolds, NULL);
-		if (lepp_num_free_comp_slots(eq) < 1)
-			goto spin;
+	if (lepp_num_free_comp_slots(eq) == 0) {
+		nolds = tile_net_lepp_grab_comps(eq, olds, wanted, 0);
+		if (nolds == 0) {
+busy:
+			spin_unlock_irqrestore(&priv->eq_lock, irqflags);
+			return NETDEV_TX_BUSY;
+		}
 	}
 
 	cmd_head = eq->cmd_head;
 	cmd_tail = eq->cmd_tail;
 
-	/* NOTE: The "gotos" below are untested. */
-
 	/* Prepare to advance, detecting full queue. */
 	cmd_next = cmd_tail + cmd_size;
 	if (cmd_tail < cmd_head && cmd_next >= cmd_head)
-		goto spin;
+		goto busy;
 	if (cmd_next > LEPP_CMD_LIMIT) {
 		cmd_next = 0;
 		if (cmd_next == cmd_head)
-			goto spin;
+			goto busy;
 	}
 
 	/* Copy the command. */
@@ -1823,14 +1885,18 @@ spin:
 	eq->comp_tail = comp_tail;
 
 	/* Flush before allowing LEPP to handle the command. */
+	/* ISSUE: Is this the optimal location for the flush? */
 	__insn_mf();
 
 	eq->cmd_tail = cmd_tail;
 
-	spin_unlock_irqrestore(&priv->cmd_lock, irqflags);
-
+	/* NOTE: Using "4" here is more efficient than "0" or "2", */
+	/* and, strangely, more efficient than pre-checking the number */
+	/* of available completions, and comparing it to 4. */
 	if (nolds == 0)
-		nolds = tile_net_lepp_grab_comps(dev, olds, wanted, NULL);
+		nolds = tile_net_lepp_grab_comps(eq, olds, wanted, 4);
+
+	spin_unlock_irqrestore(&priv->eq_lock, irqflags);
 
 	/* Handle completions. */
 	for (i = 0; i < nolds; i++)
@@ -1870,10 +1936,10 @@ static int tile_net_tx(struct sk_buff *skb, struct net_device *dev)
 
 	unsigned int num_frags;
 
-	lepp_queue_t *eq = priv->epp_queue;
+	lepp_queue_t *eq = priv->eq;
 
-	struct sk_buff *olds[4];
-	unsigned int wanted = 4;
+	struct sk_buff *olds[8];
+	unsigned int wanted = 8;
 	unsigned int i, nolds = 0;
 
 	unsigned int cmd_size = sizeof(lepp_cmd_t);
@@ -1883,8 +1949,6 @@ static int tile_net_tx(struct sk_buff *skb, struct net_device *dev)
 
 	lepp_cmd_t cmds[LEPP_MAX_FRAGS];
 
-	unsigned int free_slots;
-
 
 	/*
 	 * This is paranoia, since we think that if the link doesn't come
@@ -1905,7 +1969,8 @@ static int tile_net_tx(struct sk_buff *skb, struct net_device *dev)
 	if (hash_default) {
 		HV_PTE pte = *virt_to_pte(current->mm, (unsigned long)data);
 		if (hv_pte_get_mode(pte) != HV_PTE_MODE_CACHE_HASH_L3)
-			panic("Non-coherent egress buffer!");
+			panic("Non-HFH egress buffer! VA=%p Mode=%d PTE=%llx",
+			      data, hv_pte_get_mode(pte), hv_pte_val(pte));
 	}
 #endif
 #endif
@@ -1958,37 +2023,35 @@ static int tile_net_tx(struct sk_buff *skb, struct net_device *dev)
 
 	/* Enqueue the commands. */
 
-	spin_lock_irqsave(&priv->cmd_lock, irqflags);
+	spin_lock_irqsave(&priv->eq_lock, irqflags);
 
 	/*
 	 * Handle completions if needed to make room.
 	 * HACK: Spin until there is sufficient room.
 	 */
-	free_slots = lepp_num_free_comp_slots(eq);
-	if (free_slots < 1) {
-spin:
-		nolds += tile_net_lepp_grab_comps(dev, olds + nolds,
-						  wanted - nolds, NULL);
-		if (lepp_num_free_comp_slots(eq) < 1)
-			goto spin;
+	if (lepp_num_free_comp_slots(eq) == 0) {
+		nolds = tile_net_lepp_grab_comps(eq, olds, wanted, 0);
+		if (nolds == 0) {
+busy:
+			spin_unlock_irqrestore(&priv->eq_lock, irqflags);
+			return NETDEV_TX_BUSY;
+		}
 	}
 
 	cmd_head = eq->cmd_head;
 	cmd_tail = eq->cmd_tail;
 
-	/* NOTE: The "gotos" below are untested. */
-
 	/* Copy the commands, or fail. */
 	for (i = 0; i < num_frags; i++) {
 
 		/* Prepare to advance, detecting full queue. */
 		cmd_next = cmd_tail + cmd_size;
 		if (cmd_tail < cmd_head && cmd_next >= cmd_head)
-			goto spin;
+			goto busy;
 		if (cmd_next > LEPP_CMD_LIMIT) {
 			cmd_next = 0;
 			if (cmd_next == cmd_head)
-				goto spin;
+				goto busy;
 		}
 
 		/* Copy the command. */
@@ -2005,14 +2068,18 @@ spin:
 	eq->comp_tail = comp_tail;
 
 	/* Flush before allowing LEPP to handle the command. */
+	/* ISSUE: Is this the optimal location for the flush? */
 	__insn_mf();
 
 	eq->cmd_tail = cmd_tail;
 
-	spin_unlock_irqrestore(&priv->cmd_lock, irqflags);
-
+	/* NOTE: Using "4" here is more efficient than "0" or "2", */
+	/* and, strangely, more efficient than pre-checking the number */
+	/* of available completions, and comparing it to 4. */
 	if (nolds == 0)
-		nolds = tile_net_lepp_grab_comps(dev, olds, wanted, NULL);
+		nolds = tile_net_lepp_grab_comps(eq, olds, wanted, 4);
+
+	spin_unlock_irqrestore(&priv->eq_lock, irqflags);
 
 	/* Handle completions. */
 	for (i = 0; i < nolds; i++)
@@ -2261,7 +2328,6 @@ static struct net_device *tile_net_dev_init(const char *name)
 	int ret;
 	struct net_device *dev;
 	struct tile_net_priv *priv;
-	struct page *page;
 
 	/*
 	 * Allocate the device structure.  This allocates "priv", calls
@@ -2285,23 +2351,21 @@ static struct net_device *tile_net_dev_init(const char *name)
 
 	INIT_DELAYED_WORK(&priv->retry_work, tile_net_open_retry);
 
-	spin_lock_init(&priv->cmd_lock);
-	spin_lock_init(&priv->comp_lock);
+	spin_lock_init(&priv->eq_lock);
 
-	/* Allocate "epp_queue". */
-	BUG_ON(get_order(sizeof(lepp_queue_t)) != 0);
-	page = alloc_pages(GFP_KERNEL | __GFP_ZERO, 0);
-	if (!page) {
+	/* Allocate "eq". */
+	priv->eq_pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, EQ_ORDER);
+	if (!priv->eq_pages) {
 		free_netdev(dev);
 		return NULL;
 	}
-	priv->epp_queue = page_address(page);
+	priv->eq = page_address(priv->eq_pages);
 
 	/* Register the network device. */
 	ret = register_netdev(dev);
 	if (ret) {
 		pr_err("register_netdev %s failed %d\n", dev->name, ret);
-		free_page((unsigned long)priv->epp_queue);
+		__free_pages(priv->eq_pages, EQ_ORDER);
 		free_netdev(dev);
 		return NULL;
 	}
@@ -2310,7 +2374,7 @@ static struct net_device *tile_net_dev_init(const char *name)
 	ret = tile_net_get_mac(dev);
 	if (ret < 0) {
 		unregister_netdev(dev);
-		free_page((unsigned long)priv->epp_queue);
+		__free_pages(priv->eq_pages, EQ_ORDER);
 		free_netdev(dev);
 		return NULL;
 	}
@@ -2321,6 +2385,9 @@ static struct net_device *tile_net_dev_init(const char *name)
 
 /*
  * Module cleanup.
+ *
+ * FIXME: If compiled as a module, this module cannot be "unloaded",
+ * because the "ingress interrupt handler" is registered permanently.
  */
 static void tile_net_cleanup(void)
 {
@@ -2331,8 +2398,8 @@ static void tile_net_cleanup(void)
 			struct net_device *dev = tile_net_devs[i];
 			struct tile_net_priv *priv = netdev_priv(dev);
 			unregister_netdev(dev);
-			finv_buffer(priv->epp_queue, PAGE_SIZE);
-			free_page((unsigned long)priv->epp_queue);
+			finv_buffer(priv->eq, EQ_SIZE);
+			__free_pages(priv->eq_pages, EQ_ORDER);
 			free_netdev(dev);
 		}
 	}
@@ -2355,7 +2422,12 @@ static int tile_net_init_module(void)
 }
 
 
+module_init(tile_net_init_module);
+module_exit(tile_net_cleanup);
+
+
 #ifndef MODULE
+
 /*
  * The "network_cpus" boot argument specifies the cpus that are dedicated
  * to handle ingress packets.
@@ -2391,8 +2463,5 @@ static int __init network_cpus_setup(char *str)
 	return 0;
 }
 __setup("network_cpus=", network_cpus_setup);
-#endif
-
 
-module_init(tile_net_init_module);
-module_exit(tile_net_cleanup);
+#endif