Change Details

== Feature Overview == Boottrace is a facility for capturing trace events during boot and shutdown. This includes kernel initialization, as well as rc. It has been used by NetApp internally for several years, for catching and diagnosing slow devices or subsystems. Event annotations will be added to the boot/shutdown paths in the kernel, to some key system utilities (init(8), shutdown(8), reboot(8)), and surrounding the execution of individual rc(8) scripts. This gives a holistic log of events covering the entire boot process. A sample output is available at: Events are stored into three tables: boot-time events, run-time events, and shutdown-events. Events are stored in the boot-time table, until the time that control is handed over to init(8). After this point, events are stored in the run-time table, which will wrap around if it fills. Upon initiating a shutdown, reboot, or panic, events will be stored in the shutdown-time table. For now, boottrace is unconditionally compiled into the kernel but disabled by default. There is some overlap in functionality between boottrace and the existing boot-time event tracing facility, TSLOG. TSLOG captures kernel events from early boot (hammer_time) to mountroot, at a slightly finer granularity.. Boottrace's scope is wider, and it is currently initialized slightly later, at SI_SUB_CPU. For this reason, I think it's fine for the two to co-exist. TSLog is nicely self-contained, and extending it beyond its initial purpose would likely be messier than adding this as a separate facility. === sysctl interface === Userland is expected to create/query events via sysctl. This module introduces the following: `kern.boottrace.enabled` (read-only tunable): enable/disable tracing. Default: false. `kern.boottrace.shutdown_trace`: log shutdown-events to console before the system halts `kern.boottrace.shutdown_trace_threshold`: Set to a millisecond value. Events that occur (as compared to the previous) within less time than this value will not be printed. The default of zero logs all events. `kern.boottrace.log` (read-only, CTLFLAG_SKIP): print the boot-time and run-time tables. `kern.boottrace.boottrace` (write-only): Accepts a string, which creates a new boot-time trace event. The event will be logged as "${procname}: name". `kern.boottrace.runtrace` (write-only): Same as `kern.boottrace.boottrace`, but writes to the runtime table. `kern.boottrace.shuttrace` (write-only): Same as above, but writes to the shutdown table. === This Patch === This revision adds the core boottrace facility implementing these interfaces. Adding the trace annotations themselves to kernel and userland will happen in follow-up reviews.

== Feature Overview == Boottrace is a facility for capturing trace events during boot and shutdown. This includes kernel initialization, as well as rc. It has been used by NetApp internally for several years, for catching and diagnosing slow devices or subsystems. Event annotations will be added to the boot/shutdown paths in the kernel, to some key system utilities (init(8), shutdown(8), reboot(8)), and surrounding the execution of individual rc(8) scripts. This gives a holistic log of events covering the entire boot process. A sample output is available at: Events are stored into three tables: boot-time events, run-time events, and shutdown-events. Events are stored in the boot-time table, until the time that control is handed over to init(8). After this point, events are stored in the run-time table, which will wrap around if it fills. Upon initiating a shutdown, reboot, or panic, events will be stored in the shutdown-time table. For now, boottrace is unconditionally compiled into the kernel but disabled by default. There is some overlap in functionality between boottrace and the existing boot-time event tracing facility, TSLOG. TSLOG captures kernel events from early boot (hammer_time) to mountroot, at a slightly finer granularity.. Boottrace's scope is wider, and it is currently initialized slightly later, at SI_SUB_CPU. For this reason, I think it's fine for the two to co-exist. TSLog is nicely self-contained, and extending it beyond its initial purpose would likely be messier than adding this as a separate facility. === sysctl interface === Userland is expected to create/query events via sysctl. This module introduces the following: `kern.boottrace.enabled` (read-only tunable): enable/disable tracing. Default: false. `kern.boottrace.shutdown_trace`: log shutdown-events to console before the system halts `kern.boottrace.shutdown_trace_threshold`: Set to a millisecond value. Events that occur (as compared to the previous) within less time than this value will not be printed. The default of zero logs all events. `kern.boottrace.log` (read-only, CTLFLAG_SKIP): print the boot-time and run-time tables. `kern.boottrace.boottrace` (write-only): Accepts a string, which creates a new boot-time trace event. The event will be logged as `${procname}: name`. `kern.boottrace.runtrace` (write-only): Same as `kern.boottrace.boottrace`, but writes to the runtime table. `kern.boottrace.shuttrace` (write-only): Same as above, but writes to the shutdown table. === This Patch === This revision adds the core boottrace facility implementing these interfaces. Adding the trace annotations themselves to kernel and userland will happen in follow-up reviews.

== Feature Overview == Boottrace is a facility for capturing trace events during boot and shutdown. This includes kernel initialization, as well as rc. It has been used by NetApp internally for several years, for catching and diagnosing slow devices or subsystems. Event annotations will be added to the boot/shutdown paths in the kernel, to some key system utilities (init(8), shutdown(8), reboot(8)), and surrounding the execution of individual rc(8) scripts. This gives a holistic log of events covering the entire boot process. A sample output is available at: Events are stored into three tables: boot-time events, run-time events, and shutdown-events. Events are stored in the boot-time table, until the time that control is handed over to init(8). After this point, events are stored in the run-time table, which will wrap around if it fills. Upon initiating a shutdown, reboot, or panic, events will be stored in the shutdown-time table. For now, boottrace is unconditionally compiled into the kernel but disabled by default. There is some overlap in functionality between boottrace and the existing boot-time event tracing facility, TSLOG. TSLOG captures kernel events from early boot (hammer_time) to mountroot, at a slightly finer granularity.. Boottrace's scope is wider, and it is currently initialized slightly later, at SI_SUB_CPU. For this reason, I think it's fine for the two to co-exist. TSLog is nicely self-contained, and extending it beyond its initial purpose would likely be messier than adding this as a separate facility. === sysctl interface === Userland is expected to create/query events via sysctl. This module introduces the following: `kern.boottrace.enabled` (read-only tunable): enable/disable tracing. Default: false. `kern.boottrace.shutdown_trace`: log shutdown-events to console before the system halts `kern.boottrace.shutdown_trace_threshold`: Set to a millisecond value. Events that occur (as compared to the previous) within less time than this value will not be printed. The default of zero logs all events. `kern.boottrace.log` (read-only, CTLFLAG_SKIP): print the boot-time and and run-time tables. `kern.boottrace.boottrace` (write-only): Accepts a string, which creates a new boot-time trace event. The event will be logged as "as `${procname}: name"`. `kern.boottrace.runtrace` (write-only): Same as `kern.boottrace.boottrace`, but writes to the runtime table. `kern.boottrace.shuttrace` (write-only): Same as above, but writes to the shutdown table. === This Patch === This revision adds the core boottrace facility implementing these interfaces. Adding the trace annotations themselves to kernel and userland will happen in follow-up reviews.