Hard IRQ latency tracking

includes/bpf_helpers.h

@@ -191,4 +191,25 @@ static int (*bpf_skb_change_head)(void *, int len, int flags) =
 				(void *)(PT_REGS_FP(ctx) + sizeof(ip))); })
 #endif
 
+/*
+the TP_DATA_LOC_READ_* macros are used for reading from a field that's pointed
+to by a __data_loc variable.
+
+FYI, a __data_loc variable is really an int that contains within it the data
+needed to get the location of the actual value. these macros do the
+transformation needed to get that final location and then read from it.
+
+this code is from iovisor/bcc file src/cc/exports/helpers.h and modified by
+Netdata's Agent team for inclusion in Netdata.
+*/
+#define TP_DATA_LOC_READ_CONST(_dst, _arg, _data_loc, _length) do {           \
+    unsigned short __offset = _data_loc & 0xFFFF;                             \
+    bpf_probe_read((void *)_dst, _length, (char *)_arg + __offset);           \
+} while (0)
+#define TP_DATA_LOC_READ(_dst, _arg, _data_loc) do {                          \
+    unsigned short __offset = _data_loc & 0xFFFF;                             \
+    unsigned short __length = _data_loc >> 16;                                \
+    bpf_probe_read((void *)_dst, __length, (char *)_arg + __offset);          \
+} while (0)
+
 #endif

includes/netdata_ebpf.h

@@ -17,6 +17,7 @@ This header has the common definitions for all `.c` files.
 #include "netdata_disk.h"
 #include "netdata_fd.h"
 #include "netdata_fs.h"
+#include "netdata_hardirq.h"
 #include "netdata_mount.h"
 #include "netdata_network.h"
 #include "netdata_process.h"

includes/netdata_hardirq.h

@@ -0,0 +1,85 @@
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+#ifndef _NETDATA_HARDIRQ_H_
+#define _NETDATA_HARDIRQ_H_ 1
+
+#define NETDATA_HARDIRQ_MAX_IRQS 1024L
+#define NETDATA_HARDIRQ_NAME_LEN 32
+
+// /sys/kernel/debug/tracing/events/irq/irq_handler_entry/
+struct netdata_irq_handler_entry {
+    u64 pad;                    // This is not used with eBPF
+    int irq;                    // offset:8;       size:4; signed:1;
+    int data_loc_name;          // offset:12;      size:4; signed:1; (https://github.com/iovisor/bpftrace/issues/385)
+                                // (https://lists.linuxfoundation.org/pipermail/iovisor-dev/2017-February/000627.html)
+};
+
+// /sys/kernel/debug/tracing/events/irq/irq_handler_exit/
+struct netdata_irq_handler_exit {
+    u64 pad;                    // This is not used with eBPF
+    int irq;                    // offset:8;       size:4; signed:1;
+    int ret;                    // offset:12;      size:4; signed:1;
+};
+
+typedef struct hardirq_key {
+    int irq;
+} hardirq_key_t;
+
+typedef struct hardirq_val {
+    // incremental counter storing the total latency so far.
+    u64 latency;
+
+    // temporary timestamp stored at the IRQ entry handler, to be diff'd with a
+    // timestamp at the IRQ exit handler, to get the latency to add to the
+    // `latency` field.
+    u64 ts;
+
+    // identifies the IRQ with a human-readable string.
+    char name[NETDATA_HARDIRQ_NAME_LEN];
+} hardirq_val_t;
+
+/************************************************************************************
+ *                                HARDIRQ STATIC
+ ***********************************************************************************/
+
+// all of the `irq_vectors` events, except `vector_*`, have the same format.
+// cat /sys/kernel/debug/tracing/available_events | grep 'irq_vectors' | grep -v ':vector_'
+struct netdata_irq_vectors_entry {
+    u64 pad;                    // This is not used with eBPF
+    int vector;                 // offset:8;       size:4; signed:1;
+};
+struct netdata_irq_vectors_exit {
+    u64 pad;                    // This is not used with eBPF
+    int vector;                 // offset:8;       size:4; signed:1;
+};
+
+// these represent static IRQs that aren't given an IRQ ID like the ones above.
+// they each require separate entry/exit tracepoints to track.
+enum netdata_hardirq_static {
+    NETDATA_HARDIRQ_STATIC_APIC_THERMAL,
+    NETDATA_HARDIRQ_STATIC_APIC_THRESHOLD,
+    NETDATA_HARDIRQ_STATIC_APIC_ERROR,
+    NETDATA_HARDIRQ_STATIC_APIC_DEFERRED_ERROR,
+    NETDATA_HARDIRQ_STATIC_APIC_SPURIOUS,
+    NETDATA_HARDIRQ_STATIC_FUNC_CALL,
+    NETDATA_HARDIRQ_STATIC_FUNC_CALL_SINGLE,
+    NETDATA_HARDIRQ_STATIC_RESCHEDULE,
+    NETDATA_HARDIRQ_STATIC_LOCAL_TIMER,
+    NETDATA_HARDIRQ_STATIC_IRQ_WORK,
+    NETDATA_HARDIRQ_STATIC_X86_PLATFORM_IPI,
+
+    // must be last; used as counter.
+    NETDATA_HARDIRQ_STATIC_END
+};
+
+typedef struct hardirq_static_val {
+    // incremental counter storing the total latency so far.
+    u64 latency;
+
+    // temporary timestamp stored at the IRQ entry handler, to be diff'd with a
+    // timestamp at the IRQ exit handler, to get the latency to add to the
+    // `latency` field.
+    u64 ts;
+} hardirq_static_val_t;
+
+#endif /* _NETDATA_HARDIRQ_H_ */

kernel/Makefile

@@ -40,6 +40,7 @@ NETDATA_APPS= btrfs \
 	      fd \
 	      fdatasync \
 	      fsync \
+	      hardirq \
 	      mount \
 	      msync \
 	      nfs \

kernel/README.md

@@ -16,21 +16,22 @@ of your distribution.
 
 Right now we have the following `eBPF` program collectors:
 
--  `cachestat_kern.c`      : provides Linux page cache monitoring.
--  `dc_kern.c`             : provides Linux directory cache monitoring.
--  `disk_kern.c`           : provides disk latency monitoring.
--  `ext4_kern.c`           : provides ext4 monitoring.
--  `fdatasync_kern.c`      : monitor calls for syscall `fdatasync`.
--  `fsync_kern.c`          : monitor calls for syscall `fsync`.
--  `mount_kern.c`          : monitor calls for syscalls `mount` and `umount`.
--  `msync_kern.c`          : monitor calls for syscall `msync`.
--  `nfs_kern.c`            : provides nfs monitoring.
--  `process_kern.c`        : provides process, file and VFS stats.
--  `socket_kern.c`         : provides network stats;
--  `swap_kern.c`           : provides swap stats;
--  `sync_file_range_kern.c`: monitor calls for syscall `sync_file_range`.
--  `sync_kern.c`           : monitor calls for syscall `sync`.
--  `syncfs_kern.c`         : monitor calls for syscall `syncfs`.
--  `vfs_kern.c`            : monitor Virtual Filesystem functions.
--  `xfs_kern.c`            : provides XFS monitoring.
--  `zfs_kern.c`            : provides ZFS monitoring.
+- `cachestat_kern.c`      : provides Linux page cache monitoring.
+- `dc_kern.c`             : provides Linux directory cache monitoring.
+- `disk_kern.c`           : provides disk latency monitoring.
+- `ext4_kern.c`           : provides ext4 monitoring.
+- `fdatasync_kern.c`      : monitor calls for syscall `fdatasync`.
+- `fsync_kern.c`          : monitor calls for syscall `fsync`.
+- `hardirq_kern.c`        : provides hard interrupt (hard IRQ) latency monitoring.
+- `mount_kern.c`          : monitor calls for syscalls `mount` and `umount`.
+- `msync_kern.c`          : monitor calls for syscall `msync`.
+- `nfs_kern.c`            : provides nfs monitoring.
+- `process_kern.c`        : provides process, file and VFS stats.
+- `socket_kern.c`         : provides network stats;
+- `swap_kern.c`           : provides swap stats;
+- `sync_file_range_kern.c`: monitor calls for syscall `sync_file_range`.
+- `sync_kern.c`           : monitor calls for syscall `sync`.
+- `syncfs_kern.c`         : monitor calls for syscall `syncfs`.
+- `vfs_kern.c`            : monitor Virtual Filesystem functions.
+- `xfs_kern.c`            : provides XFS monitoring.
+- `zfs_kern.c`            : provides ZFS monitoring.