Merge branch 'upstream' of git://git.linux-mips.org/pub/scm/ralf/upstream-linus

Pull MIPS updates from Ralf Baechle:
 "This is the main pull request for MIPS for 4.7.  Here's the summary of
  the changes:

   - ATH79: Support for DTB passuing using the UHI boot protocol
   - ATH79: Remove support for builtin DTB.
   - ATH79: Add zboot debug serial support.
   - ATH79: Add initial support for Dragino MS14 (Dragine 2), Onion Omega
            and DPT-Module.
   - ATH79: Update devicetree clock support for AR9132 and AR9331.
   - ATH79: Cleanup the DT code.
   - ATH79: Support newer SOCs in ath79_ddr_ctrl_init.
   - ATH79: Fix regression in PCI window initialization.
   - BCM47xx: Move SPROM driver to drivers/firmware/
   - BCM63xx: Enable partition parser in defconfig.
   - BMIPS: BMIPS5000 has I cache filing from D cache
   - BMIPS: BMIPS: Add cpu-feature-overrides.h
   - BMIPS: Add Whirlwind support
   - BMIPS: Adjust mips-hpt-frequency for BCM7435
   - BMIPS: Remove maxcpus from BCM97435SVMB DTS
   - BMIPS: Add missing 7038 L1 register cells to BCM7435
   - BMIPS: Various tweaks to initialization code.
   - BMIPS: Enable partition parser in defconfig.
   - BMIPS: Cache tweaks.
   - BMIPS: Add UART, I2C and SATA devices to DT.
   - BMIPS: Add BCM6358 and BCM63268support
   - BMIPS: Add device tree example for BCM6358.
   - BMIPS: Improve Improve BCM6328 and BCM6368 device trees
   - Lantiq: Add support for device tree file from boot loader
   - Lantiq: Allow build with no built-in DT.
   - Loongson 3: Reserve 32MB for RS780E integrated GPU.
   - Loongson 3: Fix build error after ld-version.sh modification
   - Loongson 3: Move chipset ACPI code from drivers to arch.
   - Loongson 3: Speedup irq processing.
   - Loongson 3: Add basic Loongson 3A support.
   - Loongson 3: Set cache flush handlers to nop.
   - Loongson 3: Invalidate special TLBs when needed.
   - Loongson 3: Fast TLB refill handler.
   - MT7620: Fallback strategy for invalid syscfg0.
   - Netlogic: Fix CP0_EBASE redefinition warnings
   - Octeon: Initialization fixes
   - Octeon: Add DTS files for the D-Link DSR-1000N and EdgeRouter Lite
   - Octeon: Enable add Octeon-drivers in cavium_octeon_defconfig
   - Octeon: Correctly handle endian-swapped initramfs images.
   - Octeon: Support CN73xx, CN75xx and CN78xx.
   - Octeon: Remove dead code from cvmx-sysinfo.
   - Octeon: Extend number of supported CPUs past 32.
   - Octeon: Remove some code limiting NR_IRQS to 255.
   - Octeon: Simplify octeon_irq_ciu_gpio_set_type.
   - Octeon: Mark some functions __init in smp.c
   - Octeon: Octeon: Add Octeon III CN7xxx interface detection
   - PIC32: Add serial driver and bindings for it.
   - PIC32: Add PIC32 deadman timer driver and bindings.
   - PIC32: Add PIC32 clock timer driver and bindings.
   - Pistachio: Determine SoC revision during boot
   - Sibyte: Fix Kconfig dependencies of SIBYTE_BUS_WATCHER.
   - Sibyte: Strip redundant comments from bcm1480_regs.h.
   - Panic immediately if panic_on_oops is set.
   - module: fix incorrect IS_ERR_VALUE macro usage.
   - module: Make consistent use of pr_*
   - Remove no longer needed work_on_cpu() call.
   - Remove CONFIG_IPV6_PRIVACY from defconfigs.
   - Fix registers of non-crashing CPUs in dumps.
   - Handle MIPSisms in new vmcore_elf32_check_arch.
   - Select CONFIG_HANDLE_DOMAIN_IRQ and make it work.
   - Allow RIXI to be used on non-R2 or R6 cores.
   - Reserve nosave data for hibernation
   - Fix siginfo.h to use strict POSIX types.
   - Don't unwind user mode with EVA.
   - Fix watchpoint restoration
   - Ptrace watchpoints for R6.
   - Sync icache when it fills from dcache
   - I6400 I-cache fills from dcache.
   - Various MSA fixes.
   - Cleanup MIPS_CPU_* definitions.
   - Signal: Move generic copy_siginfo to signal.h
   - Signal: Fix uapi include in exported asm/siginfo.h
   - Timer fixes for sake of KVM.
   - XPA TLB refill fixes.
   - Treat perf counter feature
   - Update John Crispin's email address
   - Add PIC32 watchdog and bindings.
   - Handle R10000 LL/SC bug in set_pte()
   - cpufreq: Various fixes for Longson1.
   - R6: Fix R2 emulation.
   - mathemu: Cosmetic fix to ADDIUPC emulation, plenty of other small fixes
   - ELF: ABI and FP fixes.
   - Allow for relocatable kernel and use that to support KASLR.
   - Fix CPC_BASE_ADDR mask
   - Plenty fo smp-cps, CM, R6 and M6250 fixes.
   - Make reset_control_ops const.
   - Fix kernel command line handling of leading whitespace.
   - Cleanups to cache handling.
   - Add brcm, bcm6345-l1-intc device tree bindings.
   - Use generic clkdev.h header
   - Remove CLK_IS_ROOT usage.
   - Misc small cleanups.
   - CM: Fix compilation error when !MIPS_CM
   - oprofile: Fix a preemption issue
   - Detect DSP ASE v3 support:1"

* 'upstream' of git://git.linux-mips.org/pub/scm/ralf/upstream-linus: (275 commits)
  MIPS: pic32mzda: fix getting timer clock rate.
  MIPS: ath79: fix regression in PCI window initialization
  MIPS: ath79: make ath79_ddr_ctrl_init() compatible for newer SoCs
  MIPS: Fix VZ probe gas errors with binutils <2.24
  MIPS: perf: Fix I6400 event numbers
  MIPS: DEC: Export `ioasic_ssr_lock' to modules
  MIPS: MSA: Fix a link error on `_init_msa_upper' with older GCC
  MIPS: CM: Fix compilation error when !MIPS_CM
  MIPS: Fix genvdso error on rebuild
  USB: ohci-jz4740: Remove obsolete driver
  MIPS: JZ4740: Probe OHCI platform device via DT
  MIPS: JZ4740: Qi LB60: Remove support for AVT2 variant
  MIPS: pistachio: Determine SoC revision during boot
  MIPS: BMIPS: Adjust mips-hpt-frequency for BCM7435
  mips: mt7620: fallback to SDRAM when syscfg0 does not have a valid value for the memory type
  MIPS: Prevent "restoration" of MSA context in non-MSA kernels
  MIPS: cevt-r4k: Dynamically calculate min_delta_ns
  MIPS: malta-time: Take seconds into account
  MIPS: malta-time: Start GIC count before syncing to RTC
  MIPS: Force CPUs to lose FP context during mode switches
  ...

1 files changed, 80 insertions, 2 deletions

diff --git a/arch/mips/kernel/cevt-r4k.c b/arch/mips/kernel/cevt-r4k.c
index 8dfe6a6e1480..e4c21bbf9422 100644
--- a/arch/mips/kernel/cevt-r4k.c
+++ b/arch/mips/kernel/cevt-r4k.c
@@ -28,6 +28,83 @@ static int mips_next_event(unsigned long delta,
 	return res;
 }
 
+/**
+ * calculate_min_delta() - Calculate a good minimum delta for mips_next_event().
+ *
+ * Running under virtualisation can introduce overhead into mips_next_event() in
+ * the form of hypervisor emulation of CP0_Count/CP0_Compare registers,
+ * potentially with an unnatural frequency, which makes a fixed min_delta_ns
+ * value inappropriate as it may be too small.
+ *
+ * It can also introduce occasional latency from the guest being descheduled.
+ *
+ * This function calculates a good minimum delta based roughly on the 75th
+ * percentile of the time taken to do the mips_next_event() sequence, in order
+ * to handle potentially higher overhead while also eliminating outliers due to
+ * unpredictable hypervisor latency (which can be handled by retries).
+ *
+ * Return:	An appropriate minimum delta for the clock event device.
+ */
+static unsigned int calculate_min_delta(void)
+{
+	unsigned int cnt, i, j, k, l;
+	unsigned int buf1[4], buf2[3];
+	unsigned int min_delta;
+
+	/*
+	 * Calculate the median of 5 75th percentiles of 5 samples of how long
+	 * it takes to set CP0_Compare = CP0_Count + delta.
+	 */
+	for (i = 0; i < 5; ++i) {
+		for (j = 0; j < 5; ++j) {
+			/*
+			 * This is like the code in mips_next_event(), and
+			 * directly measures the borderline "safe" delta.
+			 */
+			cnt = read_c0_count();
+			write_c0_compare(cnt);
+			cnt = read_c0_count() - cnt;
+
+			/* Sorted insert into buf1 */
+			for (k = 0; k < j; ++k) {
+				if (cnt < buf1[k]) {
+					l = min_t(unsigned int,
+						  j, ARRAY_SIZE(buf1) - 1);
+					for (; l > k; --l)
+						buf1[l] = buf1[l - 1];
+					break;
+				}
+			}
+			if (k < ARRAY_SIZE(buf1))
+				buf1[k] = cnt;
+		}
+
+		/* Sorted insert of 75th percentile into buf2 */
+		for (k = 0; k < i; ++k) {
+			if (buf1[ARRAY_SIZE(buf1) - 1] < buf2[k]) {
+				l = min_t(unsigned int,
+					  i, ARRAY_SIZE(buf2) - 1);
+				for (; l > k; --l)
+					buf2[l] = buf2[l - 1];
+				break;
+			}
+		}
+		if (k < ARRAY_SIZE(buf2))
+			buf2[k] = buf1[ARRAY_SIZE(buf1) - 1];
+	}
+
+	/* Use 2 * median of 75th percentiles */
+	min_delta = buf2[ARRAY_SIZE(buf2) - 1] * 2;
+
+	/* Don't go too low */
+	if (min_delta < 0x300)
+		min_delta = 0x300;
+
+	pr_debug("%s: median 75th percentile=%#x, min_delta=%#x\n",
+		 __func__, buf2[ARRAY_SIZE(buf2) - 1], min_delta);
+	return min_delta;
+}
+
 DEFINE_PER_CPU(struct clock_event_device, mips_clockevent_device);
 int cp0_timer_irq_installed;
 
@@ -177,7 +254,7 @@ int r4k_clockevent_init(void)
 {
 	unsigned int cpu = smp_processor_id();
 	struct clock_event_device *cd;
-	unsigned int irq;
+	unsigned int irq, min_delta;
 
 	if (!cpu_has_counter || !mips_hpt_frequency)
 		return -ENXIO;
@@ -203,7 +280,8 @@ int r4k_clockevent_init(void)
 
 	/* Calculate the min / max delta */
 	cd->max_delta_ns	= clockevent_delta2ns(0x7fffffff, cd);
-	cd->min_delta_ns	= clockevent_delta2ns(0x300, cd);
+	min_delta		= calculate_min_delta();
+	cd->min_delta_ns	= clockevent_delta2ns(min_delta, cd);
 
 	cd->rating		= 300;
 	cd->irq			= irq;