Skip to contentSkip to conversation

x86/resctrl: Enable user to view and select thread throttling mode

PATCH Superseded Reinette Chatre 1 participant 5 messages Apr 13, 2020

Subsystems (3)

v2 of 7 Superseded Latest is v7 →

lore
Share & tools
lore

Reinette Chatre

author maintainer
#0 lore raw

v2 cover x86/resctrl: Enable user to view and select thread throttling mode

V1 upstream submission available from:
https://lore.kernel.org/lkml/[email protected]

A notable change since V1 is the inclusion of two additional patches from Fenghua Yu that introduce the new per-thread MBA feature. These changes are added to this series because they are small and closely related to the original submission. The per-thread MBA feature is a hardware advancement that requires no software interface changes. The patches added just enumerate the feature and expose it to userspace by showing "per-thread" in the new resctrl file "thread_throttle_mode" to help user applications fine tune performance.

There are currently a few resctrl changes outstanding for upstream inclusion. To support their consideration all outstanding resctrl patches can be viewed at https://github.com/rchatre/linux.git (branch resctrl/next)

Changes since V1 (also documented within patches to which they apply):
- Rebased on top of James Morse's CDP fix
(https://lore.kernel.org/lkml/[email protected])
- Remove RF_UNINITIALIZED (having uninitialized be represented with ones creates too much confusion), replace with an explicit check of rft->fflags in rdtgroup_add_files() (Fenghua Yu)
- Rename MBA_THREAD_THROTTLE_MODE to MBA_THROTTLE_MODE_MASK to clarify its use as a mask (Tony Luck)
- Introduce explicit MBA_THROTTLE_MODE_MAX instead of implying it is the opposite of min and use these values (min and max) explicitly whenever testing/setting the throttle mode value (Tony Luck)
- Add __init attribute to thread_throttle_mode_init_intel_rw() and thread_throttle_mode_init_intel_ro() since they are only needed during initialization (Fenghua Yu)
- Remove MBA_CFG MSR reads and error checking so that the patch is simpler and easier to review (Fenghua Yu)
- Ensure CPU hotplug lock is taken when writing register on multiple CPUs (Fenghua Yu)
- Use CPU mask already maintained as part of domains to determine which CPUs to update MBA register on (Fenghua Yu)
- Maintain MBA configuration register contents to support use case when not all CPUs of a package are online when configuration is set from user space
- Use seq_puts() instead of seq_printf() when simple strings are printed
- Set MBA configuration to default when resctrl is unmounted
- Complete rewrite of "thread_throttle_mode" documentation (Tony Luck)
- Remove unnecessary checks on user input (Andy Shevchenko)
- Change code style surrounding usage of sysfs_match_string() (Andy Shevchenko)

From V1 submission:

The first patch in this series introduces a new resctrl file, "thread_throttle_mode", on Intel systems that exposes to the user how per-thread values are allocated to a core. This is added in support of newer Intel systems that can be configured to allocate either maximum or minimum throttling of the per-thread CLOS values to the core.

Details about the feature can be found in the commit description and in Chapter 9 of the most recent Intel ISE available from
https://software.intel.com/sites/default/files/managed/c5/15/architecture-instruction-set-extensions-programming-reference.pdf

The first patch parses user input with the appropriate sysfs API that has not previously been used in resctrl. The second (in V2 the fourth) patch is added as a subsequent cleanup that switches existing resctrl string parsing code to also use this appropriate API.

Fenghua Yu (2):
x86/resctrl: Enumerate per-thread MBA
x86/resctrl: Enable per-thread MBA

Reinette Chatre (2):
x86/resctrl: Enable user to view and select thread throttling mode x86/resctrl: Use appropriate API for strings terminated by newline

 Documentation/x86/resctrl_ui.rst       |  16 +-
 arch/x86/include/asm/cpufeatures.h     |   1 +
 arch/x86/kernel/cpu/cpuid-deps.c       |   1 +
 arch/x86/kernel/cpu/resctrl/core.c     |  32 ++++
 arch/x86/kernel/cpu/resctrl/internal.h |  13 ++
 arch/x86/kernel/cpu/resctrl/rdtgroup.c | 231 +++++++++++++++++++++++--
 arch/x86/kernel/cpu/scattered.c        |   1 +
 7 files changed, 277 insertions(+), 18 deletions(-)
Signature · 1 line
2.21.0
author maintainer
#1 lore raw

PATCH v2 1/4 x86/resctrl: Enable user to view and select thread throttling mode

Intel Memory Bandwidth Allocation (MBA) control is provided per processor core. At the same time different CLOS, configured with different bandwidth percentages, can be assigned to the hardware threads sharing a core. In the original implementation of MBA the maximum throttling of the per-thread CLOS is allocated to the core. Specifically, the lower bandwidth percentage is allocated to the core.

Newer systems can be configured to allocate either maximum or minimum throttling of the per-thread CLOS values to the core.

Introduce a new resctrl file, "thread_throttle_mode", on Intel systems that exposes to the user how per-thread values are allocated to a core. On systems that support the original MBA implementation the file will always display "max". On systems that can be configured the possible values are "min" or "max" that the user can modify by writing these same words to the file.

---
Changes since V1:
- Rebased on top of James Morse's CDP fix
(https://lore.kernel.org/lkml/[email protected])
- Remove RF_UNINITIALIZED (having uninitialized be represented with ones creates too much confusion), replace with an explicit check of rft->fflags in rdtgroup_add_files() (Fenghua Yu)
- Rename MBA_THREAD_THROTTLE_MODE to MBA_THROTTLE_MODE_MASK to clarify its use as a mask (Tony Luck)
- Introduce explicit MBA_THROTTLE_MODE_MAX instead of implying it is the opposite of min and use these values (min and max) explicitly whenever testing/setting the throttle mode value (Tony Luck)
- Add __init attribute to thread_throttle_mode_init_intel_rw() and thread_throttle_mode_init_intel_ro() since they are only needed during initialization (Fenghua Yu)
- Remove MBA_CFG MSR reads and error checking so that the patch is simpler and easier to review (Fenghua Yu)
- Ensure CPU hotplug lock is taken when writing register on multiple CPUs (Fenghua Yu)
- Use CPU mask already maintained as part of domains to determine which CPUs to update MBA register on (Fenghua Yu)
- Maintain MBA configuration register contents to support use case when not all CPUs of a package are online when configuration is set from user space.
- Use seq_puts() instead of seq_printf() when simple strings are printed
- Set MBA configuration to default when resctrl is unmounted
- Complete rewrite of "thread_throttle_mode" documentation (Tony Luck)

 Documentation/x86/resctrl_ui.rst       |  14 +-
 arch/x86/kernel/cpu/resctrl/core.c     |  32 +++++
 arch/x86/kernel/cpu/resctrl/internal.h |  13 ++
 arch/x86/kernel/cpu/resctrl/rdtgroup.c | 188 ++++++++++++++++++++++++-
 4 files changed, 244 insertions(+), 3 deletions(-)
Documentation/x86/resctrl_ui.rst +12 -2
index 5368cedfb530..1dc0611ec266 100644
@@ -138,6 +138,14 @@ with respect to allocation:
 		non-linear. This field is purely informational
 		only.
 
+"thread_throttle_mode":
+		Indicator (on some CPU models control) on Intel systems
+		of how tasks running on threads of a physical core are
+		throttled in cases where they request different memory
+		bandwidth percentages:
+		"min" - the largest percentage is applied to all threads
+		"max" - the smallest percentage is applied to all threads
+
 If RDT monitoring is available there will be an "L3_MON" directory
 with the following files:
 
@@ -364,8 +372,10 @@ to the next control step available on the hardware.
 
 The bandwidth throttling is a core specific mechanism on some of Intel
 SKUs. Using a high bandwidth and a low bandwidth setting on two threads
-sharing a core will result in both threads being throttled to use the
-low bandwidth. The fact that Memory bandwidth allocation(MBA) is a core
+sharing a core may result in both threads being throttled to use the
+low bandwidth (see "thread_throttle_mode").
+
+The fact that Memory bandwidth allocation(MBA) may be a core
 specific mechanism where as memory bandwidth monitoring(MBM) is done at
 the package level may lead to confusion when users try to apply control
 via the MBA and then monitor the bandwidth to see if the controls are
arch/x86/kernel/cpu/resctrl/core.c +32 -0
index d8cc5223b7ce..aca85117a653 100644
@@ -250,6 +250,30 @@ static inline bool rdt_get_mb_table(struct rdt_resource *r)
 	return false;
 }
 
+/*
+ * Model-specific test to determine if platform where memory bandwidth
+ * control is applied to a core can be configured to apply either the
+ * maximum or minimum of the per-thread delay values.
+ * By default, platforms where memory bandwidth control is applied to a
+ * core will select the maximum delay value of the per-thread CLOS.
+ *
+ * NOTE: delay value programmed to hardware is inverse of bandwidth
+ * percentage configured via user interface.
+ */
+bool mba_cfg_supports_min_max_intel(void)
+{
+	switch (boot_cpu_data.x86_model) {
+	case INTEL_FAM6_ATOM_TREMONT_D:
+	case INTEL_FAM6_ICELAKE_X:
+	case INTEL_FAM6_ICELAKE_D:
+		return true;
+	default:
+		return false;
+	}
+
+	return false;
+}
+
 static bool __get_mem_config_intel(struct rdt_resource *r)
 {
 	union cpuid_0x10_3_eax eax;
@@ -270,6 +294,11 @@ static bool __get_mem_config_intel(struct rdt_resource *r)
 	}
 	r->data_width = 3;
 
+	if (mba_cfg_supports_min_max_intel())
+		thread_throttle_mode_init_intel_rw();
+	else
+		thread_throttle_mode_init_intel_ro();
+
 	r->alloc_capable = true;
 	r->alloc_enabled = true;
 
@@ -580,6 +609,9 @@ static void domain_add_cpu(int cpu, struct rdt_resource *r)
 
 	rdt_domain_reconfigure_cdp(r);
 
+	if (mba_cfg_supports_min_max_intel())
+		wrmsrl(MSR_MBA_CFG, mba_cfg_msr);
+
 	if (r->alloc_capable && domain_setup_ctrlval(r, d)) {
 		kfree(d);
 		return;
arch/x86/kernel/cpu/resctrl/internal.h +13 -0
index f20a47d120b1..c5f4cb91009b 100644
@@ -9,6 +9,7 @@
 
 #define MSR_IA32_L3_QOS_CFG		0xc81
 #define MSR_IA32_L2_QOS_CFG		0xc82
+#define MSR_MBA_CFG			0xc84
 #define MSR_IA32_L3_CBM_BASE		0xc90
 #define MSR_IA32_L2_CBM_BASE		0xd10
 #define MSR_IA32_MBA_THRTL_BASE		0xd50
@@ -21,6 +22,9 @@
 
 #define L2_QOS_CDP_ENABLE		0x01ULL
 
+#define MBA_THROTTLE_MODE_MIN		0x01ULL
+#define MBA_THROTTLE_MODE_MAX		0x00ULL
+
 /*
  * Event IDs are used to program IA32_QM_EVTSEL before reading event
  * counter from IA32_QM_CTR
@@ -38,6 +42,8 @@
 #define MBA_MAX_MBPS			U32_MAX
 #define MAX_MBA_BW_AMD			0x800
 
+#define MBA_THROTTLE_MODE_MASK		BIT_ULL(0)
+
 #define RMID_VAL_ERROR			BIT_ULL(63)
 #define RMID_VAL_UNAVAIL		BIT_ULL(62)
 /*
@@ -47,6 +53,10 @@
  */
 #define MBM_CNTR_WIDTH_OFFSET_MAX (62 - MBM_CNTR_WIDTH_BASE)
 
+/*
+ * MSR_MBA_CFG cache
+ */
+extern u64 mba_cfg_msr;
 
 struct rdt_fs_context {
 	struct kernfs_fs_context	kfc;
@@ -611,5 +621,8 @@ void __check_limbo(struct rdt_domain *d, bool force_free);
 bool cbm_validate_intel(char *buf, u32 *data, struct rdt_resource *r);
 bool cbm_validate_amd(char *buf, u32 *data, struct rdt_resource *r);
 void rdt_domain_reconfigure_cdp(struct rdt_resource *r);
+bool mba_cfg_supports_min_max_intel(void);
+void thread_throttle_mode_init_intel_rw(void);
+void thread_throttle_mode_init_intel_ro(void);
 
 #endif /* _ASM_X86_RESCTRL_INTERNAL_H */
arch/x86/kernel/cpu/resctrl/rdtgroup.c +187 -1

230 lines not on this page · raw message

Signature · 1 line
2.21.0
author maintainer
#2 lore raw

PATCH v2 2/4 x86/resctrl: Enumerate per-thread MBA

In reply to Reinette Chatre

From: Fenghua Yu <[email protected]>

Some systems support per-thread Memory Bandwidth Allocation (MBA) which applies a throttling delay value to each hardware thread instead of to a core. Per-thread MBA is enumerated by CPUID.

No feature flag is shown in /proc/cpuinfo. User applications need to check a resctrl throttling mode info file to know if the feature is supported.

---
 arch/x86/include/asm/cpufeatures.h | 1 +
 arch/x86/kernel/cpu/cpuid-deps.c   | 1 +
 arch/x86/kernel/cpu/scattered.c    | 1 +
 3 files changed, 3 insertions(+)
arch/x86/include/asm/cpufeatures.h +1 -0
index db189945e9b0..d0ec26ce7f66 100644
@@ -286,6 +286,7 @@
 #define X86_FEATURE_FENCE_SWAPGS_USER	(11*32+ 4) /* "" LFENCE in user entry SWAPGS path */
 #define X86_FEATURE_FENCE_SWAPGS_KERNEL	(11*32+ 5) /* "" LFENCE in kernel entry SWAPGS path */
 #define X86_FEATURE_SPLIT_LOCK_DETECT	(11*32+ 6) /* #AC for split lock */
+#define X86_FEATURE_PER_THREAD_MBA	(11*32+ 7) /* "" Per-thread Memory Bandwidth Allocation */
 
 /* Intel-defined CPU features, CPUID level 0x00000007:1 (EAX), word 12 */
 #define X86_FEATURE_AVX512_BF16		(12*32+ 5) /* AVX512 BFLOAT16 instructions */
arch/x86/kernel/cpu/cpuid-deps.c +1 -0
index 3cbe24ca80ab..3e30b26c50ef 100644
@@ -69,6 +69,7 @@ static const struct cpuid_dep cpuid_deps[] = {
 	{ X86_FEATURE_CQM_MBM_TOTAL,		X86_FEATURE_CQM_LLC   },
 	{ X86_FEATURE_CQM_MBM_LOCAL,		X86_FEATURE_CQM_LLC   },
 	{ X86_FEATURE_AVX512_BF16,		X86_FEATURE_AVX512VL  },
+	{ X86_FEATURE_PER_THREAD_MBA,		X86_FEATURE_MBA       },
 	{}
 };
 
arch/x86/kernel/cpu/scattered.c +1 -0
index 62b137c3c97a..bccfc9ff3cc1 100644
@@ -35,6 +35,7 @@ static const struct cpuid_bit cpuid_bits[] = {
 	{ X86_FEATURE_CDP_L3,		CPUID_ECX,  2, 0x00000010, 1 },
 	{ X86_FEATURE_CDP_L2,		CPUID_ECX,  2, 0x00000010, 2 },
 	{ X86_FEATURE_MBA,		CPUID_EBX,  3, 0x00000010, 0 },
+	{ X86_FEATURE_PER_THREAD_MBA,	CPUID_ECX,  0, 0x00000010, 3 },
 	{ X86_FEATURE_HW_PSTATE,	CPUID_EDX,  7, 0x80000007, 0 },
 	{ X86_FEATURE_CPB,		CPUID_EDX,  9, 0x80000007, 0 },
 	{ X86_FEATURE_PROC_FEEDBACK,    CPUID_EDX, 11, 0x80000007, 0 },
Signature · 1 line
2.21.0
author maintainer
#3 lore raw

PATCH v2 3/4 x86/resctrl: Enable per-thread MBA

In reply to Reinette Chatre

From: Fenghua Yu <[email protected]>

Current Memory Bandwidth Allocation (MBA) hardware has a limitation:
all threads on the same core must have the same delay value. If there are different delay values across threads on one core, the original MBA implementation allocates the max delay value to the core and an updated implementation allocates either min or max delay value specified by a configuration MSR across threads on the core.

Newer systems support per-thread MBA such that each thread is allocated with its own delay value.

If per-thread MBA is supported, report "per-thread" in resctrl file "info/MB/thread_throttle_mode" to let user applications know memory bandwidth is allocated per thread and help them fine tune MBA on thread level.

---
 Documentation/x86/resctrl_ui.rst       |  2 ++
 arch/x86/kernel/cpu/resctrl/rdtgroup.c | 11 +++++++++++
 2 files changed, 13 insertions(+)
Documentation/x86/resctrl_ui.rst +2 -0
index 1dc0611ec266..4c0fba65da8e 100644
@@ -145,6 +145,8 @@ with respect to allocation:
 		bandwidth percentages:
 		"min" - the largest percentage is applied to all threads
 		"max" - the smallest percentage is applied to all threads
+		"per-thread" - bandwidth percentages are directly applied to
+			       the threads running on the core
 
 If RDT monitoring is available there will be an "L3_MON" directory
 with the following files:
arch/x86/kernel/cpu/resctrl/rdtgroup.c +11 -0
index 2d6707b52a5d..d4e9b8a204d6 100644
@@ -1038,12 +1038,23 @@ static int max_threshold_occ_show(struct kernfs_open_file *of,
  * with the maximum delay value that from the software interface will be
  * the minimum of the bandwidth percentages assigned to the hardware threads
  * sharing the core.
+ *
+ * Some systems (identified by X86_FEATURE_PER_THREAD_MBA enumerated via CPUID)
+ * support per-thread MBA. On these systems hardware doesn't apply the minimum
+ * or maximum delay value to all threads in a core. Instead, a thread is
+ * allocated with the delay value that is assigned to the thread.
  */
 static int rdt_thread_throttle_mode_show(struct kernfs_open_file *of,
 					 struct seq_file *seq, void *v)
 {
 	unsigned int throttle_mode = 0;
 
+	if (static_cpu_has(X86_FEATURE_PER_THREAD_MBA)) {
+		seq_puts(seq, "per-thread\n");
+
+		return 0;
+	}
+
 	if (mba_cfg_supports_min_max_intel())
 		throttle_mode = mba_cfg_msr & MBA_THROTTLE_MODE_MASK;
 
Signature · 1 line
2.21.0
author maintainer
#4 lore raw

PATCH v2 4/4 x86/resctrl: Use appropriate API for strings terminated by newline

In reply to Reinette Chatre

The user input to files in the resctrl filesystem are expected to be terminated with a newline. Testing the user input includes a test for the presence of a newline and then replacing the newline with NUL byte followed by comparison using strcmp().

sysfs_streq() exists to test if strings are equal, treating both NUL and newline-then-NUL as equivalent string terminations. Even more, sysfs_match_string() exists to match a given string in an array using sysfs_streq().

Replace existing strcmp() comparisons of strings that are terminated with a newline with more appropriate sysfs_streq() via the sysfs_match_string() API that can perform the match across the different mode strings that are already maintained in an array.

---
Changes since V1:
- Remove unnecessary checks on user input (Andy Shevchenko)
- Change code style surrounding usage of sysfs_match_string() (Andy Shevchenko)

 arch/x86/kernel/cpu/resctrl/rdtgroup.c | 32 ++++++++++++++------------
 1 file changed, 17 insertions(+), 15 deletions(-)
arch/x86/kernel/cpu/resctrl/rdtgroup.c +17 -15
index d4e9b8a204d6..a8cbec226633 100644
@@ -1408,12 +1408,11 @@ static ssize_t rdtgroup_mode_write(struct kernfs_open_file *of,
 {
 	struct rdtgroup *rdtgrp;
 	enum rdtgrp_mode mode;
-	int ret = 0;
+	int user_m;
+	int ret;
 
-	/* Valid input requires a trailing newline */
-	if (nbytes == 0 || buf[nbytes - 1] != '\n')
+	if (nbytes == 0)
 		return -EINVAL;
-	buf[nbytes - 1] = '\0';
 
 	rdtgrp = rdtgroup_kn_lock_live(of->kn);
 	if (!rdtgrp) {
@@ -1425,11 +1424,17 @@ static ssize_t rdtgroup_mode_write(struct kernfs_open_file *of,
 
 	mode = rdtgrp->mode;
 
-	if ((!strcmp(buf, "shareable") && mode == RDT_MODE_SHAREABLE) ||
-	    (!strcmp(buf, "exclusive") && mode == RDT_MODE_EXCLUSIVE) ||
-	    (!strcmp(buf, "pseudo-locksetup") &&
-	     mode == RDT_MODE_PSEUDO_LOCKSETUP) ||
-	    (!strcmp(buf, "pseudo-locked") && mode == RDT_MODE_PSEUDO_LOCKED))
+	ret = sysfs_match_string(rdt_mode_str, buf);
+	if (ret < 0) {
+		rdt_last_cmd_puts("Unknown or unsupported mode\n");
+		goto out;
+	}
+
+	user_m = ret;
+	ret = 0;
+
+	/* Do nothing and return success if user asks for current mode */
+	if (user_m == mode)
 		goto out;
 
 	if (mode == RDT_MODE_PSEUDO_LOCKED) {
@@ -1438,14 +1443,14 @@ static ssize_t rdtgroup_mode_write(struct kernfs_open_file *of,
 		goto out;
 	}
 
-	if (!strcmp(buf, "shareable")) {
+	if (user_m == RDT_MODE_SHAREABLE) {
 		if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP) {
 			ret = rdtgroup_locksetup_exit(rdtgrp);
 			if (ret)
 				goto out;
 		}
 		rdtgrp->mode = RDT_MODE_SHAREABLE;
-	} else if (!strcmp(buf, "exclusive")) {
+	} else if (user_m == RDT_MODE_EXCLUSIVE) {
 		if (!rdtgroup_mode_test_exclusive(rdtgrp)) {
 			ret = -EINVAL;
 			goto out;
@@ -1456,14 +1461,11 @@ static ssize_t rdtgroup_mode_write(struct kernfs_open_file *of,
 				goto out;
 		}
 		rdtgrp->mode = RDT_MODE_EXCLUSIVE;
-	} else if (!strcmp(buf, "pseudo-locksetup")) {
+	} else if (user_m == RDT_MODE_PSEUDO_LOCKSETUP) {
 		ret = rdtgroup_locksetup_enter(rdtgrp);
 		if (ret)
 			goto out;
 		rdtgrp->mode = RDT_MODE_PSEUDO_LOCKSETUP;
-	} else {
-		rdt_last_cmd_puts("Unknown or unsupported mode\n");
-		ret = -EINVAL;
 	}
 
 out:
Signature · 1 line
2.21.0