x86 platform driver: intelligent power sharing driver
authorJesse Barnes <jbarnes@virtuousgeek.org>
Fri, 14 May 2010 22:41:14 +0000 (15:41 -0700)
committerMatthew Garrett <mjg@redhat.com>
Tue, 3 Aug 2010 13:48:45 +0000 (09:48 -0400)
Intel Core i3/5 platforms with integrated graphics support both CPU and
GPU turbo mode.  CPU turbo mode is opportunistic: the CPU will use any
available power to increase core frequencies if thermal headroom is
available.  The GPU side is more manual however; the graphics driver
must monitor GPU power and temperature and coordinate with a core
thermal driver to take advantage of available thermal and power headroom
in the package.

The intelligent power sharing (IPS) driver is intended to coordinate
this activity by monitoring MCP (multi-chip package) temperature and
power, allowing the CPU and/or GPU to increase their power consumption,
and thus performance, when possible.  The goal is to maximize
performance within a given platform's TDP (thermal design point).

Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Matthew Garrett <mjg@redhat.com>
drivers/platform/x86/Kconfig
drivers/platform/x86/Makefile
drivers/platform/x86/intel_ips.c [new file with mode: 0644]
include/drm/i915_drm.h

index fd060016b7e9644569127a43f85cc46ca9aaf8bb..724b2ed1a3cb03bf108796186eff66ce796d7557 100644 (file)
@@ -539,4 +539,14 @@ config INTEL_SCU_IPC
          some embedded Intel x86 platforms. This is not needed for PC-type
          machines.
 
+config INTEL_IPS
+       tristate "Intel Intelligent Power Sharing"
+       depends on ACPI
+       ---help---
+         Intel Calpella platforms support dynamic power sharing between the
+         CPU and GPU, maximizing performance in a given TDP.  This driver,
+         along with the CPU frequency and i915 drivers, provides that
+         functionality.  If in doubt, say Y here; it will only load on
+         supported platforms.
+
 endif # X86_PLATFORM_DEVICES
index 8770bfe71431ec6435feacd02b4d2211741b7a39..7318fc2c1629f1a1f69cd894686b4a5b47a886ff 100644 (file)
@@ -26,3 +26,4 @@ obj-$(CONFIG_TOPSTAR_LAPTOP)  += topstar-laptop.o
 obj-$(CONFIG_ACPI_TOSHIBA)     += toshiba_acpi.o
 obj-$(CONFIG_TOSHIBA_BT_RFKILL)        += toshiba_bluetooth.o
 obj-$(CONFIG_INTEL_SCU_IPC)    += intel_scu_ipc.o
+obj-$(CONFIG_INTEL_IPS)                += intel_ips.o
diff --git a/drivers/platform/x86/intel_ips.c b/drivers/platform/x86/intel_ips.c
new file mode 100644 (file)
index 0000000..f1dce3b
--- /dev/null
@@ -0,0 +1,1655 @@
+/*
+ * Copyright (c) 2009-2010 Intel Corporation
+ *
+ * This program is free software; you can redistribute it and/or modify it
+ * under the terms and conditions of the GNU General Public License,
+ * version 2, as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
+ * more details.
+ *
+ * You should have received a copy of the GNU General Public License along with
+ * this program; if not, write to the Free Software Foundation, Inc.,
+ * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
+ *
+ * The full GNU General Public License is included in this distribution in
+ * the file called "COPYING".
+ *
+ * Authors:
+ *     Jesse Barnes <jbarnes@virtuousgeek.org>
+ */
+
+/*
+ * Some Intel Ibex Peak based platforms support so-called "intelligent
+ * power sharing", which allows the CPU and GPU to cooperate to maximize
+ * performance within a given TDP (thermal design point).  This driver
+ * performs the coordination between the CPU and GPU, monitors thermal and
+ * power statistics in the platform, and initializes power monitoring
+ * hardware.  It also provides a few tunables to control behavior.  Its
+ * primary purpose is to safely allow CPU and GPU turbo modes to be enabled
+ * by tracking power and thermal budget; secondarily it can boost turbo
+ * performance by allocating more power or thermal budget to the CPU or GPU
+ * based on available headroom and activity.
+ *
+ * The basic algorithm is driven by a 5s moving average of tempurature.  If
+ * thermal headroom is available, the CPU and/or GPU power clamps may be
+ * adjusted upwards.  If we hit the thermal ceiling or a thermal trigger,
+ * we scale back the clamp.  Aside from trigger events (when we're critically
+ * close or over our TDP) we don't adjust the clamps more than once every
+ * five seconds.
+ *
+ * The thermal device (device 31, function 6) has a set of registers that
+ * are updated by the ME firmware.  The ME should also take the clamp values
+ * written to those registers and write them to the CPU, but we currently
+ * bypass that functionality and write the CPU MSR directly.
+ *
+ * UNSUPPORTED:
+ *   - dual MCP configs
+ *
+ * TODO:
+ *   - handle CPU hotplug
+ *   - provide turbo enable/disable api
+ *   - make sure we can write turbo enable/disable reg based on MISC_EN
+ *
+ * Related documents:
+ *   - CDI 403777, 403778 - Auburndale EDS vol 1 & 2
+ *   - CDI 401376 - Ibex Peak EDS
+ *   - ref 26037, 26641 - IPS BIOS spec
+ *   - ref 26489 - Nehalem BIOS writer's guide
+ *   - ref 26921 - Ibex Peak BIOS Specification
+ */
+
+#include <linux/debugfs.h>
+#include <linux/delay.h>
+#include <linux/interrupt.h>
+#include <linux/kernel.h>
+#include <linux/kthread.h>
+#include <linux/module.h>
+#include <linux/pci.h>
+#include <linux/sched.h>
+#include <linux/seq_file.h>
+#include <linux/string.h>
+#include <linux/tick.h>
+#include <linux/timer.h>
+#include <drm/i915_drm.h>
+#include <asm/msr.h>
+#include <asm/processor.h>
+
+#define PCI_DEVICE_ID_INTEL_THERMAL_SENSOR 0x3b32
+
+/*
+ * Package level MSRs for monitor/control
+ */
+#define PLATFORM_INFO  0xce
+#define   PLATFORM_TDP         (1<<29)
+#define   PLATFORM_RATIO       (1<<28)
+
+#define IA32_MISC_ENABLE       0x1a0
+#define   IA32_MISC_TURBO_EN   (1ULL<<38)
+
+#define TURBO_POWER_CURRENT_LIMIT      0x1ac
+#define   TURBO_TDC_OVR_EN     (1UL<<31)
+#define   TURBO_TDC_MASK       (0x000000007fff0000UL)
+#define   TURBO_TDC_SHIFT      (16)
+#define   TURBO_TDP_OVR_EN     (1UL<<15)
+#define   TURBO_TDP_MASK       (0x0000000000003fffUL)
+
+/*
+ * Core/thread MSRs for monitoring
+ */
+#define IA32_PERF_CTL          0x199
+#define   IA32_PERF_TURBO_DIS  (1ULL<<32)
+
+/*
+ * Thermal PCI device regs
+ */
+#define THM_CFG_TBAR   0x10
+#define THM_CFG_TBAR_HI        0x14
+
+#define THM_TSIU       0x00
+#define THM_TSE                0x01
+#define   TSE_EN       0xb8
+#define THM_TSS                0x02
+#define THM_TSTR       0x03
+#define THM_TSTTP      0x04
+#define THM_TSCO       0x08
+#define THM_TSES       0x0c
+#define THM_TSGPEN     0x0d
+#define   TSGPEN_HOT_LOHI      (1<<1)
+#define   TSGPEN_CRIT_LOHI     (1<<2)
+#define THM_TSPC       0x0e
+#define THM_PPEC       0x10
+#define THM_CTA                0x12
+#define THM_PTA                0x14
+#define   PTA_SLOPE_MASK       (0xff00)
+#define   PTA_SLOPE_SHIFT      8
+#define   PTA_OFFSET_MASK      (0x00ff)
+#define THM_MGTA       0x16
+#define   MGTA_SLOPE_MASK      (0xff00)
+#define   MGTA_SLOPE_SHIFT     8
+#define   MGTA_OFFSET_MASK     (0x00ff)
+#define THM_TRC                0x1a
+#define   TRC_CORE2_EN (1<<15)
+#define   TRC_THM_EN   (1<<12)
+#define   TRC_C6_WAR   (1<<8)
+#define   TRC_CORE1_EN (1<<7)
+#define   TRC_CORE_PWR (1<<6)
+#define   TRC_PCH_EN   (1<<5)
+#define   TRC_MCH_EN   (1<<4)
+#define   TRC_DIMM4    (1<<3)
+#define   TRC_DIMM3    (1<<2)
+#define   TRC_DIMM2    (1<<1)
+#define   TRC_DIMM1    (1<<0)
+#define THM_TES                0x20
+#define THM_TEN                0x21
+#define   TEN_UPDATE_EN        1
+#define THM_PSC                0x24
+#define   PSC_NTG      (1<<0) /* No GFX turbo support */
+#define   PSC_NTPC     (1<<1) /* No CPU turbo support */
+#define   PSC_PP_DEF   (0<<2) /* Perf policy up to driver */
+#define   PSP_PP_PC    (1<<2) /* BIOS prefers CPU perf */
+#define   PSP_PP_BAL   (2<<2) /* BIOS wants balanced perf */
+#define   PSP_PP_GFX   (3<<2) /* BIOS prefers GFX perf */
+#define   PSP_PBRT     (1<<4) /* BIOS run time support */
+#define THM_CTV1       0x30
+#define   CTV_TEMP_ERROR (1<<15)
+#define   CTV_TEMP_MASK        0x3f
+#define   CTV_
+#define THM_CTV2       0x32
+#define THM_CEC                0x34 /* undocumented power accumulator in joules */
+#define THM_AE         0x3f
+#define THM_HTS                0x50 /* 32 bits */
+#define   HTS_PCPL_MASK        (0x7fe00000)
+#define   HTS_PCPL_SHIFT 21
+#define   HTS_GPL_MASK  (0x001ff000)
+#define   HTS_GPL_SHIFT 12
+#define   HTS_PP_MASK  (0x00000c00)
+#define   HTS_PP_SHIFT  10
+#define   HTS_PP_DEF   0
+#define   HTS_PP_PROC  1
+#define   HTS_PP_BAL   2
+#define   HTS_PP_GFX   3
+#define   HTS_PCTD_DIS (1<<9)
+#define   HTS_GTD_DIS  (1<<8)
+#define   HTS_PTL_MASK  (0x000000fe)
+#define   HTS_PTL_SHIFT 1
+#define   HTS_NVV      (1<<0)
+#define THM_HTSHI      0x54 /* 16 bits */
+#define   HTS2_PPL_MASK                (0x03ff)
+#define   HTS2_PRST_MASK       (0x3c00)
+#define   HTS2_PRST_SHIFT      10
+#define   HTS2_PRST_UNLOADED   0
+#define   HTS2_PRST_RUNNING    1
+#define   HTS2_PRST_TDISOP     2 /* turbo disabled due to power */
+#define   HTS2_PRST_TDISHT     3 /* turbo disabled due to high temp */
+#define   HTS2_PRST_TDISUSR    4 /* user disabled turbo */
+#define   HTS2_PRST_TDISPLAT   5 /* platform disabled turbo */
+#define   HTS2_PRST_TDISPM     6 /* power management disabled turbo */
+#define   HTS2_PRST_TDISERR    7 /* some kind of error disabled turbo */
+#define THM_PTL                0x56
+#define THM_MGTV       0x58
+#define   TV_MASK      0x000000000000ff00
+#define   TV_SHIFT     8
+#define THM_PTV                0x60
+#define   PTV_MASK     0x00ff
+#define THM_MMGPC      0x64
+#define THM_MPPC       0x66
+#define THM_MPCPC      0x68
+#define THM_TSPIEN     0x82
+#define   TSPIEN_AUX_LOHI      (1<<0)
+#define   TSPIEN_HOT_LOHI      (1<<1)
+#define   TSPIEN_CRIT_LOHI     (1<<2)
+#define   TSPIEN_AUX2_LOHI     (1<<3)
+#define THM_TSLOCK     0x83
+#define THM_ATR                0x84
+#define THM_TOF                0x87
+#define THM_STS                0x98
+#define   STS_PCPL_MASK                (0x7fe00000)
+#define   STS_PCPL_SHIFT       21
+#define   STS_GPL_MASK         (0x001ff000)
+#define   STS_GPL_SHIFT                12
+#define   STS_PP_MASK          (0x00000c00)
+#define   STS_PP_SHIFT         10
+#define   STS_PP_DEF           0
+#define   STS_PP_PROC          1
+#define   STS_PP_BAL           2
+#define   STS_PP_GFX           3
+#define   STS_PCTD_DIS         (1<<9)
+#define   STS_GTD_DIS          (1<<8)
+#define   STS_PTL_MASK         (0x000000fe)
+#define   STS_PTL_SHIFT                1
+#define   STS_NVV              (1<<0)
+#define THM_SEC                0x9c
+#define   SEC_ACK      (1<<0)
+#define THM_TC3                0xa4
+#define THM_TC1                0xa8
+#define   STS_PPL_MASK         (0x0003ff00)
+#define   STS_PPL_SHIFT                16
+#define THM_TC2                0xac
+#define THM_DTV                0xb0
+#define THM_ITV                0xd8
+#define   ITV_ME_SEQNO_MASK 0x000f0000 /* ME should update every ~200ms */
+#define   ITV_ME_SEQNO_SHIFT (16)
+#define   ITV_MCH_TEMP_MASK 0x0000ff00
+#define   ITV_MCH_TEMP_SHIFT (8)
+#define   ITV_PCH_TEMP_MASK 0x000000ff
+
+#define thm_readb(off) readb(ips->regmap + (off))
+#define thm_readw(off) readw(ips->regmap + (off))
+#define thm_readl(off) readl(ips->regmap + (off))
+#define thm_readq(off) readq(ips->regmap + (off))
+
+#define thm_writeb(off, val) writeb((val), ips->regmap + (off))
+#define thm_writew(off, val) writew((val), ips->regmap + (off))
+#define thm_writel(off, val) writel((val), ips->regmap + (off))
+
+static const int IPS_ADJUST_PERIOD = 5000; /* ms */
+
+/* For initial average collection */
+static const int IPS_SAMPLE_PERIOD = 200; /* ms */
+static const int IPS_SAMPLE_WINDOW = 5000; /* 5s moving window of samples */
+#define IPS_SAMPLE_COUNT (IPS_SAMPLE_WINDOW / IPS_SAMPLE_PERIOD)
+
+/* Per-SKU limits */
+struct ips_mcp_limits {
+       int cpu_family;
+       int cpu_model; /* includes extended model... */
+       int mcp_power_limit; /* mW units */
+       int core_power_limit;
+       int mch_power_limit;
+       int core_temp_limit; /* degrees C */
+       int mch_temp_limit;
+};
+
+/* Max temps are -10 degrees C to avoid PROCHOT# */
+
+struct ips_mcp_limits ips_sv_limits = {
+       .mcp_power_limit = 35000,
+       .core_power_limit = 29000,
+       .mch_power_limit = 20000,
+       .core_temp_limit = 95,
+       .mch_temp_limit = 90
+};
+
+struct ips_mcp_limits ips_lv_limits = {
+       .mcp_power_limit = 25000,
+       .core_power_limit = 21000,
+       .mch_power_limit = 13000,
+       .core_temp_limit = 95,
+       .mch_temp_limit = 90
+};
+
+struct ips_mcp_limits ips_ulv_limits = {
+       .mcp_power_limit = 18000,
+       .core_power_limit = 14000,
+       .mch_power_limit = 11000,
+       .core_temp_limit = 95,
+       .mch_temp_limit = 90
+};
+
+struct ips_driver {
+       struct pci_dev *dev;
+       void *regmap;
+       struct task_struct *monitor;
+       struct task_struct *adjust;
+       struct dentry *debug_root;
+
+       /* Average CPU core temps (all averages in .01 degrees C for precision) */
+       u16 ctv1_avg_temp;
+       u16 ctv2_avg_temp;
+       /* GMCH average */
+       u16 mch_avg_temp;
+       /* Average for the CPU (both cores?) */
+       u16 mcp_avg_temp;
+       /* Average power consumption (in mW) */
+       u32 cpu_avg_power;
+       u32 mch_avg_power;
+
+       /* Offset values */
+       u16 cta_val;
+       u16 pta_val;
+       u16 mgta_val;
+
+       /* Maximums & prefs, protected by turbo status lock */
+       spinlock_t turbo_status_lock;
+       u16 mcp_temp_limit;
+       u16 mcp_power_limit;
+       u16 core_power_limit;
+       u16 mch_power_limit;
+       bool cpu_turbo_enabled;
+       bool __cpu_turbo_on;
+       bool gpu_turbo_enabled;
+       bool __gpu_turbo_on;
+       bool gpu_preferred;
+       bool poll_turbo_status;
+       bool second_cpu;
+       struct ips_mcp_limits *limits;
+
+       /* Optional MCH interfaces for if i915 is in use */
+       unsigned long (*read_mch_val)(void);
+       bool (*gpu_raise)(void);
+       bool (*gpu_lower)(void);
+       bool (*gpu_busy)(void);
+       bool (*gpu_turbo_disable)(void);
+
+       /* For restoration at unload */
+       u64 orig_turbo_limit;
+       u64 orig_turbo_ratios;
+};
+
+/**
+ * ips_cpu_busy - is CPU busy?
+ * @ips: IPS driver struct
+ *
+ * Check CPU for load to see whether we should increase its thermal budget.
+ *
+ * RETURNS:
+ * True if the CPU could use more power, false otherwise.
+ */
+static bool ips_cpu_busy(struct ips_driver *ips)
+{
+       if ((avenrun[0] >> FSHIFT) > 1)
+               return true;
+
+       return false;
+}
+
+/**
+ * ips_cpu_raise - raise CPU power clamp
+ * @ips: IPS driver struct
+ *
+ * Raise the CPU power clamp by %IPS_CPU_STEP, in accordance with TDP for
+ * this platform.
+ *
+ * We do this by adjusting the TURBO_POWER_CURRENT_LIMIT MSR upwards (as
+ * long as we haven't hit the TDP limit for the SKU).
+ */
+static void ips_cpu_raise(struct ips_driver *ips)
+{
+       u64 turbo_override;
+       u16 cur_tdp_limit, new_tdp_limit;
+
+       if (!ips->cpu_turbo_enabled)
+               return;
+
+       rdmsrl(TURBO_POWER_CURRENT_LIMIT, turbo_override);
+
+       cur_tdp_limit = turbo_override & TURBO_TDP_MASK;
+       new_tdp_limit = cur_tdp_limit + 8; /* 1W increase */
+
+       /* Clamp to SKU TDP limit */
+       if (((new_tdp_limit * 10) / 8) > ips->core_power_limit)
+               new_tdp_limit = cur_tdp_limit;
+
+       thm_writew(THM_MPCPC, (new_tdp_limit * 10) / 8);
+
+       turbo_override |= TURBO_TDC_OVR_EN | TURBO_TDC_OVR_EN;
+       wrmsrl(TURBO_POWER_CURRENT_LIMIT, turbo_override);
+
+       turbo_override &= ~TURBO_TDP_MASK;
+       turbo_override |= new_tdp_limit;
+
+       wrmsrl(TURBO_POWER_CURRENT_LIMIT, turbo_override);
+}
+
+/**
+ * ips_cpu_lower - lower CPU power clamp
+ * @ips: IPS driver struct
+ *
+ * Lower CPU power clamp b %IPS_CPU_STEP if possible.
+ *
+ * We do this by adjusting the TURBO_POWER_CURRENT_LIMIT MSR down, going
+ * as low as the platform limits will allow (though we could go lower there
+ * wouldn't be much point).
+ */
+static void ips_cpu_lower(struct ips_driver *ips)
+{
+       u64 turbo_override;
+       u16 cur_limit, new_limit;
+
+       rdmsrl(TURBO_POWER_CURRENT_LIMIT, turbo_override);
+
+       cur_limit = turbo_override & TURBO_TDP_MASK;
+       new_limit = cur_limit - 8; /* 1W decrease */
+
+       /* Clamp to SKU TDP limit */
+       if (((new_limit * 10) / 8) < (ips->orig_turbo_limit & TURBO_TDP_MASK))
+               new_limit = ips->orig_turbo_limit & TURBO_TDP_MASK;
+
+       thm_writew(THM_MPCPC, (new_limit * 10) / 8);
+
+       turbo_override |= TURBO_TDC_OVR_EN | TURBO_TDC_OVR_EN;
+       wrmsrl(TURBO_POWER_CURRENT_LIMIT, turbo_override);
+
+       turbo_override &= ~TURBO_TDP_MASK;
+       turbo_override |= new_limit;
+
+       wrmsrl(TURBO_POWER_CURRENT_LIMIT, turbo_override);
+}
+
+/**
+ * do_enable_cpu_turbo - internal turbo enable function
+ * @data: unused
+ *
+ * Internal function for actually updating MSRs.  When we enable/disable
+ * turbo, we need to do it on each CPU; this function is the one called
+ * by on_each_cpu() when needed.
+ */
+static void do_enable_cpu_turbo(void *data)
+{
+       u64 perf_ctl;
+
+       rdmsrl(IA32_PERF_CTL, perf_ctl);
+       if (perf_ctl & IA32_PERF_TURBO_DIS) {
+               perf_ctl &= ~IA32_PERF_TURBO_DIS;
+               wrmsrl(IA32_PERF_CTL, perf_ctl);
+       }
+}
+
+/**
+ * ips_enable_cpu_turbo - enable turbo mode on all CPUs
+ * @ips: IPS driver struct
+ *
+ * Enable turbo mode by clearing the disable bit in IA32_PERF_CTL on
+ * all logical threads.
+ */
+static void ips_enable_cpu_turbo(struct ips_driver *ips)
+{
+       /* Already on, no need to mess with MSRs */
+       if (ips->__cpu_turbo_on)
+               return;
+
+       on_each_cpu(do_enable_cpu_turbo, ips, 1);
+
+       ips->__cpu_turbo_on = true;
+}
+
+/**
+ * do_disable_cpu_turbo - internal turbo disable function
+ * @data: unused
+ *
+ * Internal function for actually updating MSRs.  When we enable/disable
+ * turbo, we need to do it on each CPU; this function is the one called
+ * by on_each_cpu() when needed.
+ */
+static void do_disable_cpu_turbo(void *data)
+{
+       u64 perf_ctl;
+
+       rdmsrl(IA32_PERF_CTL, perf_ctl);
+       if (!(perf_ctl & IA32_PERF_TURBO_DIS)) {
+               perf_ctl |= IA32_PERF_TURBO_DIS;
+               wrmsrl(IA32_PERF_CTL, perf_ctl);
+       }
+}
+
+/**
+ * ips_disable_cpu_turbo - disable turbo mode on all CPUs
+ * @ips: IPS driver struct
+ *
+ * Disable turbo mode by setting the disable bit in IA32_PERF_CTL on
+ * all logical threads.
+ */
+static void ips_disable_cpu_turbo(struct ips_driver *ips)
+{
+       /* Already off, leave it */
+       if (!ips->__cpu_turbo_on)
+               return;
+
+       on_each_cpu(do_disable_cpu_turbo, ips, 1);
+
+       ips->__cpu_turbo_on = false;
+}
+
+/**
+ * ips_gpu_busy - is GPU busy?
+ * @ips: IPS driver struct
+ *
+ * Check GPU for load to see whether we should increase its thermal budget.
+ * We need to call into the i915 driver in this case.
+ *
+ * RETURNS:
+ * True if the GPU could use more power, false otherwise.
+ */
+static bool ips_gpu_busy(struct ips_driver *ips)
+{
+       return false;
+}
+
+/**
+ * ips_gpu_raise - raise GPU power clamp
+ * @ips: IPS driver struct
+ *
+ * Raise the GPU frequency/power if possible.  We need to call into the
+ * i915 driver in this case.
+ */
+static void ips_gpu_raise(struct ips_driver *ips)
+{
+       if (!ips->gpu_turbo_enabled)
+               return;
+
+       if (!ips->gpu_raise())
+               ips->gpu_turbo_enabled = false;
+
+       return;
+}
+
+/**
+ * ips_gpu_lower - lower GPU power clamp
+ * @ips: IPS driver struct
+ *
+ * Lower GPU frequency/power if possible.  Need to call i915.
+ */
+static void ips_gpu_lower(struct ips_driver *ips)
+{
+       if (!ips->gpu_turbo_enabled)
+               return;
+
+       if (!ips->gpu_lower())
+               ips->gpu_turbo_enabled = false;
+
+       return;
+}
+
+/**
+ * ips_enable_gpu_turbo - notify the gfx driver turbo is available
+ * @ips: IPS driver struct
+ *
+ * Call into the graphics driver indicating that it can safely use
+ * turbo mode.
+ */
+static void ips_enable_gpu_turbo(struct ips_driver *ips)
+{
+       if (ips->__gpu_turbo_on)
+               return;
+       ips->__gpu_turbo_on = true;
+}
+
+/**
+ * ips_disable_gpu_turbo - notify the gfx driver to disable turbo mode
+ * @ips: IPS driver struct
+ *
+ * Request that the graphics driver disable turbo mode.
+ */
+static void ips_disable_gpu_turbo(struct ips_driver *ips)
+{
+       /* Avoid calling i915 if turbo is already disabled */
+       if (!ips->__gpu_turbo_on)
+               return;
+
+       if (!ips->gpu_turbo_disable())
+               dev_err(&ips->dev->dev, "failed to disable graphis turbo\n");
+       else
+               ips->__gpu_turbo_on = false;
+}
+
+/**
+ * mcp_exceeded - check whether we're outside our thermal & power limits
+ * @ips: IPS driver struct
+ *
+ * Check whether the MCP is over its thermal or power budget.
+ */
+static bool mcp_exceeded(struct ips_driver *ips)
+{
+       unsigned long flags;
+       bool ret = false;
+
+       spin_lock_irqsave(&ips->turbo_status_lock, flags);
+       if (ips->mcp_avg_temp > (ips->mcp_temp_limit * 100))
+               ret = true;
+       if (ips->cpu_avg_power + ips->mch_avg_power > ips->mcp_power_limit)
+               ret = true;
+       spin_unlock_irqrestore(&ips->turbo_status_lock, flags);
+
+       if (ret)
+               dev_warn(&ips->dev->dev,
+                        "MCP power or thermal limit exceeded\n");
+
+       return ret;
+}
+
+/**
+ * cpu_exceeded - check whether a CPU core is outside its limits
+ * @ips: IPS driver struct
+ * @cpu: CPU number to check
+ *
+ * Check a given CPU's average temp or power is over its limit.
+ */
+static bool cpu_exceeded(struct ips_driver *ips, int cpu)
+{
+       unsigned long flags;
+       int avg;
+       bool ret = false;
+
+       spin_lock_irqsave(&ips->turbo_status_lock, flags);
+       avg = cpu ? ips->ctv2_avg_temp : ips->ctv1_avg_temp;
+       if (avg > (ips->limits->core_temp_limit * 100))
+               ret = true;
+       if (ips->cpu_avg_power > ips->core_power_limit)
+               ret = true;
+       spin_unlock_irqrestore(&ips->turbo_status_lock, flags);
+
+       if (ret)
+               dev_warn(&ips->dev->dev,
+                        "CPU power or thermal limit exceeded\n");
+
+       return ret;
+}
+
+/**
+ * mch_exceeded - check whether the GPU is over budget
+ * @ips: IPS driver struct
+ *
+ * Check the MCH temp & power against their maximums.
+ */
+static bool mch_exceeded(struct ips_driver *ips)
+{
+       unsigned long flags;
+       bool ret = false;
+
+       spin_lock_irqsave(&ips->turbo_status_lock, flags);
+       if (ips->mch_avg_temp > (ips->limits->mch_temp_limit * 100))
+               ret = true;
+       spin_unlock_irqrestore(&ips->turbo_status_lock, flags);
+
+       return ret;
+}
+
+/**
+ * update_turbo_limits - get various limits & settings from regs
+ * @ips: IPS driver struct
+ *
+ * Update the IPS power & temp limits, along with turbo enable flags,
+ * based on latest register contents.
+ *
+ * Used at init time and for runtime BIOS support, which requires polling
+ * the regs for updates (as a result of AC->DC transition for example).
+ *
+ * LOCKING:
+ * Caller must hold turbo_status_lock (outside of init)
+ */
+static void update_turbo_limits(struct ips_driver *ips)
+{
+       u32 hts = thm_readl(THM_HTS);
+
+       ips->cpu_turbo_enabled = !(hts & HTS_PCTD_DIS);
+       ips->gpu_turbo_enabled = !(hts & HTS_GTD_DIS);
+       ips->core_power_limit = thm_readw(THM_MPCPC);
+       ips->mch_power_limit = thm_readw(THM_MMGPC);
+       ips->mcp_temp_limit = thm_readw(THM_PTL);
+       ips->mcp_power_limit = thm_readw(THM_MPPC);
+
+       /* Ignore BIOS CPU vs GPU pref */
+}
+
+/**
+ * ips_adjust - adjust power clamp based on thermal state
+ * @data: ips driver structure
+ *
+ * Wake up every 5s or so and check whether we should adjust the power clamp.
+ * Check CPU and GPU load to determine which needs adjustment.  There are
+ * several things to consider here:
+ *   - do we need to adjust up or down?
+ *   - is CPU busy?
+ *   - is GPU busy?
+ *   - is CPU in turbo?
+ *   - is GPU in turbo?
+ *   - is CPU or GPU preferred? (CPU is default)
+ *
+ * So, given the above, we do the following:
+ *   - up (TDP available)
+ *     - CPU not busy, GPU not busy - nothing
+ *     - CPU busy, GPU not busy - adjust CPU up
+ *     - CPU not busy, GPU busy - adjust GPU up
+ *     - CPU busy, GPU busy - adjust preferred unit up, taking headroom from
+ *       non-preferred unit if necessary
+ *   - down (at TDP limit)
+ *     - adjust both CPU and GPU down if possible
+ *
+               cpu+ gpu+       cpu+gpu-        cpu-gpu+        cpu-gpu-
+cpu < gpu <    cpu+gpu+        cpu+            gpu+            nothing
+cpu < gpu >=   cpu+gpu-(mcp<)  cpu+gpu-(mcp<)  gpu-            gpu-
+cpu >= gpu <   cpu-gpu+(mcp<)  cpu-            cpu-gpu+(mcp<)  cpu-
+cpu >= gpu >=  cpu-gpu-        cpu-gpu-        cpu-gpu-        cpu-gpu-
+ *
+ */
+static int ips_adjust(void *data)
+{
+       struct ips_driver *ips = data;
+       unsigned long flags;
+
+       dev_dbg(&ips->dev->dev, "starting ips-adjust thread\n");
+
+       /*
+        * Adjust CPU and GPU clamps every 5s if needed.  Doing it more
+        * often isn't recommended due to ME interaction.
+        */
+       do {
+               bool cpu_busy = ips_cpu_busy(ips);
+               bool gpu_busy = ips_gpu_busy(ips);
+
+               spin_lock_irqsave(&ips->turbo_status_lock, flags);
+               if (ips->poll_turbo_status)
+                       update_turbo_limits(ips);
+               spin_unlock_irqrestore(&ips->turbo_status_lock, flags);
+
+               /* Update turbo status if necessary */
+               if (ips->cpu_turbo_enabled)
+                       ips_enable_cpu_turbo(ips);
+               else
+                       ips_disable_cpu_turbo(ips);
+
+               if (ips->gpu_turbo_enabled)
+                       ips_enable_gpu_turbo(ips);
+               else
+                       ips_disable_gpu_turbo(ips);
+
+               /* We're outside our comfort zone, crank them down */
+               if (!mcp_exceeded(ips)) {
+                       ips_cpu_lower(ips);
+                       ips_gpu_lower(ips);
+                       goto sleep;
+               }
+
+               if (!cpu_exceeded(ips, 0) && cpu_busy)
+                       ips_cpu_raise(ips);
+               else
+                       ips_cpu_lower(ips);
+
+               if (!mch_exceeded(ips) && gpu_busy)
+                       ips_gpu_raise(ips);
+               else
+                       ips_gpu_lower(ips);
+
+sleep:
+               schedule_timeout_interruptible(msecs_to_jiffies(IPS_ADJUST_PERIOD));
+       } while (!kthread_should_stop());
+
+       dev_dbg(&ips->dev->dev, "ips-adjust thread stopped\n");
+
+       return 0;
+}
+
+/*
+ * Helpers for reading out temp/power values and calculating their
+ * averages for the decision making and monitoring functions.
+ */
+
+static u16 calc_avg_temp(struct ips_driver *ips, u16 *array)
+{
+       u64 total = 0;
+       int i;
+       u16 avg;
+
+       for (i = 0; i < IPS_SAMPLE_COUNT; i++)
+               total += (u64)(array[i] * 100);
+
+       do_div(total, IPS_SAMPLE_COUNT);
+
+       avg = (u16)total;
+
+       return avg;
+}
+
+static u16 read_mgtv(struct ips_driver *ips)
+{
+       u16 ret;
+       u64 slope, offset;
+       u64 val;
+
+       val = thm_readq(THM_MGTV);
+       val = (val & TV_MASK) >> TV_SHIFT;
+
+       slope = offset = thm_readw(THM_MGTA);
+       slope = (slope & MGTA_SLOPE_MASK) >> MGTA_SLOPE_SHIFT;
+       offset = offset & MGTA_OFFSET_MASK;
+
+       ret = ((val * slope + 0x40) >> 7) + offset;
+
+
+       return ret;
+}
+
+static u16 read_ptv(struct ips_driver *ips)
+{
+       u16 val, slope, offset;
+
+       slope = (ips->pta_val & PTA_SLOPE_MASK) >> PTA_SLOPE_SHIFT;
+       offset = ips->pta_val & PTA_OFFSET_MASK;
+
+       val = thm_readw(THM_PTV) & PTV_MASK;
+
+       return val;
+}
+
+static u16 read_ctv(struct ips_driver *ips, int cpu)
+{
+       int reg = cpu ? THM_CTV2 : THM_CTV1;
+       u16 val;
+
+       val = thm_readw(reg);
+       if (!(val & CTV_TEMP_ERROR))
+               val = (val) >> 6; /* discard fractional component */
+       else
+               val = 0;
+
+       return val;
+}
+
+static u32 get_cpu_power(struct ips_driver *ips, u32 *last, int period)
+{
+       u32 val;
+       u32 ret;
+
+       /*
+        * CEC is in joules/65535.  Take difference over time to
+        * get watts.
+        */
+       val = thm_readl(THM_CEC);
+
+       /* period is in ms and we want mW */
+       ret = (((val - *last) * 1000) / period);
+       ret = (ret * 1000) / 65535;
+       *last = val;
+
+       return ret;
+}
+
+static const u16 temp_decay_factor = 2;
+static u16 update_average_temp(u16 avg, u16 val)
+{
+       u16 ret;
+
+       /* Multiply by 100 for extra precision */
+       ret = (val * 100 / temp_decay_factor) +
+               (((temp_decay_factor - 1) * avg) / temp_decay_factor);
+       return ret;
+}
+
+static const u16 power_decay_factor = 2;
+static u16 update_average_power(u32 avg, u32 val)
+{
+       u32 ret;
+
+       ret = (val / power_decay_factor) +
+               (((power_decay_factor - 1) * avg) / power_decay_factor);
+
+       return ret;
+}
+
+static u32 calc_avg_power(struct ips_driver *ips, u32 *array)
+{
+       u64 total = 0;
+       u32 avg;
+       int i;
+
+       for (i = 0; i < IPS_SAMPLE_COUNT; i++)
+               total += array[i];
+
+       do_div(total, IPS_SAMPLE_COUNT);
+       avg = (u32)total;
+
+       return avg;
+}
+
+static void monitor_timeout(unsigned long arg)
+{
+       wake_up_process((struct task_struct *)arg);
+}
+
+/**
+ * ips_monitor - temp/power monitoring thread
+ * @data: ips driver structure
+ *
+ * This is the main function for the IPS driver.  It monitors power and
+ * tempurature in the MCP and adjusts CPU and GPU power clams accordingly.
+ *
+ * We keep a 5s moving average of power consumption and tempurature.  Using
+ * that data, along with CPU vs GPU preference, we adjust the power clamps
+ * up or down.
+ */
+static int ips_monitor(void *data)
+{
+       struct ips_driver *ips = data;
+       struct timer_list timer;
+       unsigned long seqno_timestamp, expire, last_msecs, last_sample_period;
+       int i;
+       u32 *cpu_samples = NULL, *mchp_samples = NULL, old_cpu_power;
+       u16 *mcp_samples = NULL, *ctv1_samples = NULL, *ctv2_samples = NULL,
+               *mch_samples = NULL;
+       u8 cur_seqno, last_seqno;
+
+       mcp_samples = kzalloc(sizeof(u16) * IPS_SAMPLE_COUNT, GFP_KERNEL);
+       ctv1_samples = kzalloc(sizeof(u16) * IPS_SAMPLE_COUNT, GFP_KERNEL);
+       ctv2_samples = kzalloc(sizeof(u16) * IPS_SAMPLE_COUNT, GFP_KERNEL);
+       mch_samples = kzalloc(sizeof(u16) * IPS_SAMPLE_COUNT, GFP_KERNEL);
+       cpu_samples = kzalloc(sizeof(u32) * IPS_SAMPLE_COUNT, GFP_KERNEL);
+       mchp_samples = kzalloc(sizeof(u32) * IPS_SAMPLE_COUNT, GFP_KERNEL);
+       if (!mcp_samples || !ctv1_samples || !ctv2_samples || !mch_samples) {
+               dev_err(&ips->dev->dev,
+                       "failed to allocate sample array, ips disabled\n");
+               kfree(mcp_samples);
+               kfree(ctv1_samples);
+               kfree(ctv2_samples);
+               kfree(mch_samples);
+               kfree(cpu_samples);
+               kthread_stop(ips->adjust);
+               return -ENOMEM;
+       }
+
+       last_seqno = (thm_readl(THM_ITV) & ITV_ME_SEQNO_MASK) >>
+               ITV_ME_SEQNO_SHIFT;
+       seqno_timestamp = get_jiffies_64();
+
+       old_cpu_power = thm_readl(THM_CEC) / 65535;
+       schedule_timeout_interruptible(msecs_to_jiffies(IPS_SAMPLE_PERIOD));
+
+       /* Collect an initial average */
+       for (i = 0; i < IPS_SAMPLE_COUNT; i++) {
+               u32 mchp, cpu_power;
+               u16 val;
+
+               mcp_samples[i] = read_ptv(ips);
+
+               val = read_ctv(ips, 0);
+               ctv1_samples[i] = val;
+
+               val = read_ctv(ips, 1);
+               ctv2_samples[i] = val;
+
+               val = read_mgtv(ips);
+               mch_samples[i] = val;
+
+               cpu_power = get_cpu_power(ips, &old_cpu_power,
+                                         IPS_SAMPLE_PERIOD);
+               cpu_samples[i] = cpu_power;
+
+               if (ips->read_mch_val) {
+                       mchp = ips->read_mch_val();
+                       mchp_samples[i] = mchp;
+               }
+
+               schedule_timeout_interruptible(msecs_to_jiffies(IPS_SAMPLE_PERIOD));
+               if (kthread_should_stop())
+                       break;
+       }
+
+       ips->mcp_avg_temp = calc_avg_temp(ips, mcp_samples);
+       ips->ctv1_avg_temp = calc_avg_temp(ips, ctv1_samples);
+       ips->ctv2_avg_temp = calc_avg_temp(ips, ctv2_samples);
+       ips->mch_avg_temp = calc_avg_temp(ips, mch_samples);
+       ips->cpu_avg_power = calc_avg_power(ips, cpu_samples);
+       ips->mch_avg_power = calc_avg_power(ips, mchp_samples);
+       kfree(mcp_samples);
+       kfree(ctv1_samples);
+       kfree(ctv2_samples);
+       kfree(mch_samples);
+       kfree(cpu_samples);
+       kfree(mchp_samples);
+
+       /* Start the adjustment thread now that we have data */
+       wake_up_process(ips->adjust);
+
+       /*
+        * Ok, now we have an initial avg.  From here on out, we track the
+        * running avg using a decaying average calculation.  This allows
+        * us to reduce the sample frequency if the CPU and GPU are idle.
+        */
+       old_cpu_power = thm_readl(THM_CEC);
+       schedule_timeout_interruptible(msecs_to_jiffies(IPS_SAMPLE_PERIOD));
+       last_sample_period = IPS_SAMPLE_PERIOD;
+
+       setup_deferrable_timer_on_stack(&timer, monitor_timeout,
+                                       (unsigned long)current);
+       do {
+               u32 cpu_val, mch_val;
+               u16 val;
+
+               /* MCP itself */
+               val = read_ptv(ips);
+               ips->mcp_avg_temp = update_average_temp(ips->mcp_avg_temp, val);
+
+               /* Processor 0 */
+               val = read_ctv(ips, 0);
+               ips->ctv1_avg_temp =
+                       update_average_temp(ips->ctv1_avg_temp, val);
+               /* Power */
+               cpu_val = get_cpu_power(ips, &old_cpu_power,
+                                       last_sample_period);
+               ips->cpu_avg_power =
+                       update_average_power(ips->cpu_avg_power, cpu_val);
+
+               if (ips->second_cpu) {
+                       /* Processor 1 */
+                       val = read_ctv(ips, 1);
+                       ips->ctv2_avg_temp =
+                               update_average_temp(ips->ctv2_avg_temp, val);
+               }
+
+               /* MCH */
+               val = read_mgtv(ips);
+               ips->mch_avg_temp = update_average_temp(ips->mch_avg_temp, val);
+               /* Power */
+               if (ips->read_mch_val) {
+                       mch_val = ips->read_mch_val();
+                       ips->mch_avg_power =
+                               update_average_power(ips->mch_avg_power,
+                                                    mch_val);
+               }
+
+               /*
+                * Make sure ME is updating thermal regs.
+                * Note:
+                * If it's been more than a second since the last update,
+                * the ME is probably hung.
+                */
+               cur_seqno = (thm_readl(THM_ITV) & ITV_ME_SEQNO_MASK) >>
+                       ITV_ME_SEQNO_SHIFT;
+               if (cur_seqno == last_seqno &&
+                   time_after(jiffies, seqno_timestamp + HZ)) {
+                       dev_warn(&ips->dev->dev, "ME failed to update for more than 1s, likely hung\n");
+               } else {
+                       seqno_timestamp = get_jiffies_64();
+                       last_seqno = cur_seqno;
+               }
+
+               last_msecs = jiffies_to_msecs(jiffies);
+               expire = jiffies + msecs_to_jiffies(IPS_SAMPLE_PERIOD);
+
+               __set_current_state(TASK_UNINTERRUPTIBLE);
+               mod_timer(&timer, expire);
+               schedule();
+
+               /* Calculate actual sample period for power averaging */
+               last_sample_period = jiffies_to_msecs(jiffies) - last_msecs;
+               if (!last_sample_period)
+                       last_sample_period = 1;
+       } while (!kthread_should_stop());
+
+       del_timer_sync(&timer);
+       destroy_timer_on_stack(&timer);
+
+       dev_dbg(&ips->dev->dev, "ips-monitor thread stopped\n");
+
+       return 0;
+}
+
+#if 0
+#define THM_DUMPW(reg) \
+       { \
+       u16 val = thm_readw(reg); \
+       dev_dbg(&ips->dev->dev, #reg ": 0x%04x\n", val); \
+       }
+#define THM_DUMPL(reg) \
+       { \
+       u32 val = thm_readl(reg); \
+       dev_dbg(&ips->dev->dev, #reg ": 0x%08x\n", val); \
+       }
+#define THM_DUMPQ(reg) \
+       { \
+       u64 val = thm_readq(reg); \
+       dev_dbg(&ips->dev->dev, #reg ": 0x%016x\n", val); \
+       }
+
+static void dump_thermal_info(struct ips_driver *ips)
+{
+       u16 ptl;
+
+       ptl = thm_readw(THM_PTL);
+       dev_dbg(&ips->dev->dev, "Processor temp limit: %d\n", ptl);
+
+       THM_DUMPW(THM_CTA);
+       THM_DUMPW(THM_TRC);
+       THM_DUMPW(THM_CTV1);
+       THM_DUMPL(THM_STS);
+       THM_DUMPW(THM_PTV);
+       THM_DUMPQ(THM_MGTV);
+}
+#endif
+
+/**
+ * ips_irq_handler - handle temperature triggers and other IPS events
+ * @irq: irq number
+ * @arg: unused
+ *
+ * Handle temperature limit trigger events, generally by lowering the clamps.
+ * If we're at a critical limit, we clamp back to the lowest possible value
+ * to prevent emergency shutdown.
+ */
+static irqreturn_t ips_irq_handler(int irq, void *arg)
+{
+       struct ips_driver *ips = arg;
+       u8 tses = thm_readb(THM_TSES);
+       u8 tes = thm_readb(THM_TES);
+
+       if (!tses && !tes)
+               return IRQ_NONE;
+
+       dev_info(&ips->dev->dev, "TSES: 0x%02x\n", tses);
+       dev_info(&ips->dev->dev, "TES: 0x%02x\n", tes);
+
+       /* STS update from EC? */
+       if (tes & 1) {
+               u32 sts, tc1;
+
+               sts = thm_readl(THM_STS);
+               tc1 = thm_readl(THM_TC1);
+
+               if (sts & STS_NVV) {
+                       spin_lock(&ips->turbo_status_lock);
+                       ips->core_power_limit = (sts & STS_PCPL_MASK) >>
+                               STS_PCPL_SHIFT;
+                       ips->mch_power_limit = (sts & STS_GPL_MASK) >>
+                               STS_GPL_SHIFT;
+                       /* ignore EC CPU vs GPU pref */
+                       ips->cpu_turbo_enabled = !(sts & STS_PCTD_DIS);
+                       ips->gpu_turbo_enabled = !(sts & STS_GTD_DIS);
+                       ips->mcp_temp_limit = (sts & STS_PTL_MASK) >>
+                               STS_PTL_SHIFT;
+                       ips->mcp_power_limit = (tc1 & STS_PPL_MASK) >>
+                               STS_PPL_SHIFT;
+                       spin_unlock(&ips->turbo_status_lock);
+
+                       thm_writeb(THM_SEC, SEC_ACK);
+               }
+               thm_writeb(THM_TES, tes);
+       }
+
+       /* Thermal trip */
+       if (tses) {
+               dev_warn(&ips->dev->dev,
+                        "thermal trip occurred, tses: 0x%04x\n", tses);
+               thm_writeb(THM_TSES, tses);
+       }
+
+       return IRQ_HANDLED;
+}
+
+#ifndef CONFIG_DEBUG_FS
+static void ips_debugfs_init(struct ips_driver *ips) { return; }
+static void ips_debugfs_cleanup(struct ips_driver *ips) { return; }
+#else
+
+/* Expose current state and limits in debugfs if possible */
+
+struct ips_debugfs_node {
+       struct ips_driver *ips;
+       char *name;
+       int (*show)(struct seq_file *m, void *data);
+};
+
+static int show_cpu_temp(struct seq_file *m, void *data)
+{
+       struct ips_driver *ips = m->private;
+
+       seq_printf(m, "%d.%02d\n", ips->ctv1_avg_temp / 100,
+                  ips->ctv1_avg_temp % 100);
+
+       return 0;
+}
+
+static int show_cpu_power(struct seq_file *m, void *data)
+{
+       struct ips_driver *ips = m->private;
+
+       seq_printf(m, "%dmW\n", ips->cpu_avg_power);
+
+       return 0;
+}
+
+static int show_cpu_clamp(struct seq_file *m, void *data)
+{
+       u64 turbo_override;
+       int tdp, tdc;
+
+       rdmsrl(TURBO_POWER_CURRENT_LIMIT, turbo_override);
+
+       tdp = (int)(turbo_override & TURBO_TDP_MASK);
+       tdc = (int)((turbo_override & TURBO_TDC_MASK) >> TURBO_TDC_SHIFT);
+
+       /* Convert to .1W/A units */
+       tdp = tdp * 10 / 8;
+       tdc = tdc * 10 / 8;
+
+       /* Watts Amperes */
+       seq_printf(m, "%d.%dW %d.%dA\n", tdp / 10, tdp % 10,
+                  tdc / 10, tdc % 10);
+
+       return 0;
+}
+
+static int show_mch_temp(struct seq_file *m, void *data)
+{
+       struct ips_driver *ips = m->private;
+
+       seq_printf(m, "%d.%02d\n", ips->mch_avg_temp / 100,
+                  ips->mch_avg_temp % 100);
+
+       return 0;
+}
+
+static int show_mch_power(struct seq_file *m, void *data)
+{
+       struct ips_driver *ips = m->private;
+
+       seq_printf(m, "%dmW\n", ips->mch_avg_power);
+
+       return 0;
+}
+
+static struct ips_debugfs_node ips_debug_files[] = {
+       { NULL, "cpu_temp", show_cpu_temp },
+       { NULL, "cpu_power", show_cpu_power },
+       { NULL, "cpu_clamp", show_cpu_clamp },
+       { NULL, "mch_temp", show_mch_temp },
+       { NULL, "mch_power", show_mch_power },
+};
+
+static int ips_debugfs_open(struct inode *inode, struct file *file)
+{
+       struct ips_debugfs_node *node = inode->i_private;
+
+       return single_open(file, node->show, node->ips);
+}
+
+static const struct file_operations ips_debugfs_ops = {
+       .owner = THIS_MODULE,
+       .open = ips_debugfs_open,
+       .read = seq_read,
+       .llseek = seq_lseek,
+       .release = single_release,
+};
+
+static void ips_debugfs_cleanup(struct ips_driver *ips)
+{
+       if (ips->debug_root)
+               debugfs_remove_recursive(ips->debug_root);
+       return;
+}
+
+static void ips_debugfs_init(struct ips_driver *ips)
+{
+       int i;
+
+       ips->debug_root = debugfs_create_dir("ips", NULL);
+       if (!ips->debug_root) {
+               dev_err(&ips->dev->dev,
+                       "failed to create debugfs entries: %ld\n",
+                       PTR_ERR(ips->debug_root));
+               return;
+       }
+
+       for (i = 0; i < ARRAY_SIZE(ips_debug_files); i++) {
+               struct dentry *ent;
+               struct ips_debugfs_node *node = &ips_debug_files[i];
+
+               node->ips = ips;
+               ent = debugfs_create_file(node->name, S_IFREG | S_IRUGO,
+                                         ips->debug_root, node,
+                                         &ips_debugfs_ops);
+               if (!ent) {
+                       dev_err(&ips->dev->dev,
+                               "failed to create debug file: %ld\n",
+                               PTR_ERR(ent));
+                       goto err_cleanup;
+               }
+       }
+
+       return;
+
+err_cleanup:
+       ips_debugfs_cleanup(ips);
+       return;
+}
+#endif /* CONFIG_DEBUG_FS */
+
+/**
+ * ips_detect_cpu - detect whether CPU supports IPS
+ *
+ * Walk our list and see if we're on a supported CPU.  If we find one,
+ * return the limits for it.
+ */
+static struct ips_mcp_limits *ips_detect_cpu(struct ips_driver *ips)
+{
+       u64 turbo_power, misc_en;
+       struct ips_mcp_limits *limits = NULL;
+       u16 tdp;
+
+       if (!(boot_cpu_data.x86 == 6 && boot_cpu_data.x86_model == 37)) {
+               dev_info(&ips->dev->dev, "Non-IPS CPU detected.\n");
+               goto out;
+       }
+
+       rdmsrl(IA32_MISC_ENABLE, misc_en);
+       /*
+        * If the turbo enable bit isn't set, we shouldn't try to enable/disable
+        * turbo manually or we'll get an illegal MSR access, even though
+        * turbo will still be available.
+        */
+       if (!(misc_en & IA32_MISC_TURBO_EN))
+               ; /* add turbo MSR write allowed flag if necessary */
+
+       if (strstr(boot_cpu_data.x86_model_id, "CPU       M"))
+               limits = &ips_sv_limits;
+       else if (strstr(boot_cpu_data.x86_model_id, "CPU       L"))
+               limits = &ips_lv_limits;
+       else if (strstr(boot_cpu_data.x86_model_id, "CPU       U"))
+               limits = &ips_ulv_limits;
+       else
+               dev_info(&ips->dev->dev, "No CPUID match found.\n");
+
+       rdmsrl(TURBO_POWER_CURRENT_LIMIT, turbo_power);
+       tdp = turbo_power & TURBO_TDP_MASK;
+
+       /* Sanity check TDP against CPU */
+       if (limits->mcp_power_limit != (tdp / 8) * 1000) {
+               dev_warn(&ips->dev->dev, "Warning: CPU TDP doesn't match expected value (found %d, expected %d)\n",
+                        tdp / 8, limits->mcp_power_limit / 1000);
+       }
+
+out:
+       return limits;
+}
+
+/**
+ * ips_get_i915_syms - try to get GPU control methods from i915 driver
+ * @ips: IPS driver
+ *
+ * The i915 driver exports several interfaces to allow the IPS driver to
+ * monitor and control graphics turbo mode.  If we can find them, we can
+ * enable graphics turbo, otherwise we must disable it to avoid exceeding
+ * thermal and power limits in the MCP.
+ */
+static bool ips_get_i915_syms(struct ips_driver *ips)
+{
+       ips->read_mch_val = symbol_get(i915_read_mch_val);
+       if (!ips->read_mch_val)
+               goto out_err;
+       ips->gpu_raise = symbol_get(i915_gpu_raise);
+       if (!ips->gpu_raise)
+               goto out_put_mch;
+       ips->gpu_lower = symbol_get(i915_gpu_lower);
+       if (!ips->gpu_lower)
+               goto out_put_raise;
+       ips->gpu_busy = symbol_get(i915_gpu_busy);
+       if (!ips->gpu_busy)
+               goto out_put_lower;
+       ips->gpu_turbo_disable = symbol_get(i915_gpu_turbo_disable);
+       if (!ips->gpu_turbo_disable)
+               goto out_put_busy;
+
+       return true;
+
+out_put_busy:
+       symbol_put(i915_gpu_turbo_disable);
+out_put_lower:
+       symbol_put(i915_gpu_lower);
+out_put_raise:
+       symbol_put(i915_gpu_raise);
+out_put_mch:
+       symbol_put(i915_read_mch_val);
+out_err:
+       return false;
+}
+
+static DEFINE_PCI_DEVICE_TABLE(ips_id_table) = {
+       { PCI_DEVICE(PCI_VENDOR_ID_INTEL,
+                    PCI_DEVICE_ID_INTEL_THERMAL_SENSOR), },
+       { 0, }
+};
+
+MODULE_DEVICE_TABLE(pci, ips_id_table);
+
+static int ips_probe(struct pci_dev *dev, const struct pci_device_id *id)
+{
+       u64 platform_info;
+       struct ips_driver *ips;
+       u32 hts;
+       int ret = 0;
+       u16 htshi, trc, trc_required_mask;
+       u8 tse;
+
+       ips = kzalloc(sizeof(struct ips_driver), GFP_KERNEL);
+       if (!ips)
+               return -ENOMEM;
+
+       pci_set_drvdata(dev, ips);
+       ips->dev = dev;
+
+       ips->limits = ips_detect_cpu(ips);
+       if (!ips->limits) {
+               dev_info(&dev->dev, "IPS not supported on this CPU\n");
+               ret = -ENXIO;
+               goto error_free;
+       }
+
+       spin_lock_init(&ips->turbo_status_lock);
+
+       if (!pci_resource_start(dev, 0)) {
+               dev_err(&dev->dev, "TBAR not assigned, aborting\n");
+               ret = -ENXIO;
+               goto error_free;
+       }
+
+       ret = pci_request_regions(dev, "ips thermal sensor");
+       if (ret) {
+               dev_err(&dev->dev, "thermal resource busy, aborting\n");
+               goto error_free;
+       }
+
+       ret = pci_enable_device(dev);
+       if (ret) {
+               dev_err(&dev->dev, "can't enable PCI device, aborting\n");
+               goto error_free;
+       }
+
+       ips->regmap = ioremap(pci_resource_start(dev, 0),
+                             pci_resource_len(dev, 0));
+       if (!ips->regmap) {
+               dev_err(&dev->dev, "failed to map thermal regs, aborting\n");
+               ret = -EBUSY;
+               goto error_release;
+       }
+
+       tse = thm_readb(THM_TSE);
+       if (tse != TSE_EN) {
+               dev_err(&dev->dev, "thermal device not enabled (0x%02x), aborting\n", tse);
+               ret = -ENXIO;
+               goto error_unmap;
+       }
+
+       trc = thm_readw(THM_TRC);
+       trc_required_mask = TRC_CORE1_EN | TRC_CORE_PWR | TRC_MCH_EN;
+       if ((trc & trc_required_mask) != trc_required_mask) {
+               dev_err(&dev->dev, "thermal reporting for required devices not enabled, aborting\n");
+               ret = -ENXIO;
+               goto error_unmap;
+       }
+
+       if (trc & TRC_CORE2_EN)
+               ips->second_cpu = true;
+
+       if (!ips_get_i915_syms(ips)) {
+               dev_err(&dev->dev, "failed to get i915 symbols, graphics turbo disabled\n");
+               ips->gpu_turbo_enabled = false;
+       } else {
+               dev_dbg(&dev->dev, "graphics turbo enabled\n");
+               ips->gpu_turbo_enabled = true;
+       }
+
+       update_turbo_limits(ips);
+       dev_dbg(&dev->dev, "max cpu power clamp: %dW\n",
+               ips->mcp_power_limit / 10);
+       dev_dbg(&dev->dev, "max core power clamp: %dW\n",
+               ips->core_power_limit / 10);
+       /* BIOS may update limits at runtime */
+       if (thm_readl(THM_PSC) & PSP_PBRT)
+               ips->poll_turbo_status = true;
+
+       /*
+        * Check PLATFORM_INFO MSR to make sure this chip is
+        * turbo capable.
+        */
+       rdmsrl(PLATFORM_INFO, platform_info);
+       if (!(platform_info & PLATFORM_TDP)) {
+               dev_err(&dev->dev, "platform indicates TDP override unavailable, aborting\n");
+               ret = -ENODEV;
+               goto error_unmap;
+       }
+
+       /*
+        * IRQ handler for ME interaction
+        * Note: don't use MSI here as the PCH has bugs.
+        */
+       pci_disable_msi(dev);
+       ret = request_irq(dev->irq, ips_irq_handler, IRQF_SHARED, "ips",
+                         ips);
+       if (ret) {
+               dev_err(&dev->dev, "request irq failed, aborting\n");
+               goto error_unmap;
+       }
+
+       /* Enable aux, hot & critical interrupts */
+       thm_writeb(THM_TSPIEN, TSPIEN_AUX2_LOHI | TSPIEN_CRIT_LOHI |
+                  TSPIEN_HOT_LOHI | TSPIEN_AUX_LOHI);
+       thm_writeb(THM_TEN, TEN_UPDATE_EN);
+
+       /* Collect adjustment values */
+       ips->cta_val = thm_readw(THM_CTA);
+       ips->pta_val = thm_readw(THM_PTA);
+       ips->mgta_val = thm_readw(THM_MGTA);
+
+       /* Save turbo limits & ratios */
+       rdmsrl(TURBO_POWER_CURRENT_LIMIT, ips->orig_turbo_limit);
+
+       ips_enable_cpu_turbo(ips);
+       ips->cpu_turbo_enabled = true;
+
+       /* Set up the work queue and monitor/adjust threads */
+       ips->monitor = kthread_run(ips_monitor, ips, "ips-monitor");
+       if (IS_ERR(ips->monitor)) {
+               dev_err(&dev->dev,
+                       "failed to create thermal monitor thread, aborting\n");
+               ret = -ENOMEM;
+               goto error_free_irq;
+       }
+
+       ips->adjust = kthread_create(ips_adjust, ips, "ips-adjust");
+       if (IS_ERR(ips->adjust)) {
+               dev_err(&dev->dev,
+                       "failed to create thermal adjust thread, aborting\n");
+               ret = -ENOMEM;
+               goto error_thread_cleanup;
+       }
+
+       hts = (ips->core_power_limit << HTS_PCPL_SHIFT) |
+               (ips->mcp_temp_limit << HTS_PTL_SHIFT) | HTS_NVV;
+       htshi = HTS2_PRST_RUNNING << HTS2_PRST_SHIFT;
+
+       thm_writew(THM_HTSHI, htshi);
+       thm_writel(THM_HTS, hts);
+
+       ips_debugfs_init(ips);
+
+       dev_info(&dev->dev, "IPS driver initialized, MCP temp limit %d\n",
+                ips->mcp_temp_limit);
+       return ret;
+
+error_thread_cleanup:
+       kthread_stop(ips->monitor);
+error_free_irq:
+       free_irq(ips->dev->irq, ips);
+error_unmap:
+       iounmap(ips->regmap);
+error_release:
+       pci_release_regions(dev);
+error_free:
+       kfree(ips);
+       return ret;
+}
+
+static void ips_remove(struct pci_dev *dev)
+{
+       struct ips_driver *ips = pci_get_drvdata(dev);
+       u64 turbo_override;
+
+       if (!ips)
+               return;
+
+       ips_debugfs_cleanup(ips);
+
+       /* Release i915 driver */
+       if (ips->read_mch_val)
+               symbol_put(i915_read_mch_val);
+       if (ips->gpu_raise)
+               symbol_put(i915_gpu_raise);
+       if (ips->gpu_lower)
+               symbol_put(i915_gpu_lower);
+       if (ips->gpu_busy)
+               symbol_put(i915_gpu_busy);
+       if (ips->gpu_turbo_disable)
+               symbol_put(i915_gpu_turbo_disable);
+
+       rdmsrl(TURBO_POWER_CURRENT_LIMIT, turbo_override);
+       turbo_override &= ~(TURBO_TDC_OVR_EN | TURBO_TDP_OVR_EN);
+       wrmsrl(TURBO_POWER_CURRENT_LIMIT, turbo_override);
+       wrmsrl(TURBO_POWER_CURRENT_LIMIT, ips->orig_turbo_limit);
+
+       free_irq(ips->dev->irq, ips);
+       if (ips->adjust)
+               kthread_stop(ips->adjust);
+       if (ips->monitor)
+               kthread_stop(ips->monitor);
+       iounmap(ips->regmap);
+       pci_release_regions(dev);
+       kfree(ips);
+       dev_dbg(&dev->dev, "IPS driver removed\n");
+}
+
+#ifdef CONFIG_PM
+static int ips_suspend(struct pci_dev *dev, pm_message_t state)
+{
+       return 0;
+}
+
+static int ips_resume(struct pci_dev *dev)
+{
+       return 0;
+}
+#else
+#define ips_suspend NULL
+#define ips_resume NULL
+#endif /* CONFIG_PM */
+
+static void ips_shutdown(struct pci_dev *dev)
+{
+}
+
+static struct pci_driver ips_pci_driver = {
+       .name = "intel ips",
+       .id_table = ips_id_table,
+       .probe = ips_probe,
+       .remove = ips_remove,
+       .suspend = ips_suspend,
+       .resume = ips_resume,
+       .shutdown = ips_shutdown,
+};
+
+static int __init ips_init(void)
+{
+       return pci_register_driver(&ips_pci_driver);
+}
+module_init(ips_init);
+
+static void ips_exit(void)
+{
+       pci_unregister_driver(&ips_pci_driver);
+       return;
+}
+module_exit(ips_exit);
+
+MODULE_LICENSE("GPL");
+MODULE_AUTHOR("Jesse Barnes <jbarnes@virtuousgeek.org>");
+MODULE_DESCRIPTION("Intelligent Power Sharing Driver");
index 7f0028e1010b97e253b29dffdaaa8edb152bab54..8f8b072c4c7b6013e12c10e35bbf1da24f6a9d79 100644 (file)
  * subject to backwards-compatibility constraints.
  */
 
+#ifdef __KERNEL__
+/* For use by IPS driver */
+extern unsigned long i915_read_mch_val(void);
+extern bool i915_gpu_raise(void);
+extern bool i915_gpu_lower(void);
+extern bool i915_gpu_busy(void);
+extern bool i915_gpu_turbo_disable(void);
+#endif
+
 /* Each region is a minimum of 16k, and there are at most 255 of them.
  */
 #define I915_NR_TEX_REGIONS 255        /* table size 2k - maximum due to use