/linux/arch/arm/mach-rpc/ |
H A D | io-acorn.S | diff 337015573718b161891a3473d25f59273f2e626b Tue Jun 15 18:52:53 CEST 2021 Chris Down <chris@chrisdown.name> printk: Userspace format indexing support
We have a number of systems industry-wide that have a subset of their functionality that works as follows:
1. Receive a message from local kmsg, serial console, or netconsole; 2. Apply a set of rules to classify the message; 3. Do something based on this classification (like scheduling a remediation for the machine), rinse, and repeat.
As a couple of examples of places we have this implemented just inside Facebook, although this isn't a Facebook-specific problem, we have this inside our netconsole processing (for alarm classification), and as part of our machine health checking. We use these messages to determine fairly important metrics around production health, and it's important that we get them right.
While for some kinds of issues we have counters, tracepoints, or metrics with a stable interface which can reliably indicate the issue, in order to react to production issues quickly we need to work with the interface which most kernel developers naturally use when developing: printk.
Most production issues come from unexpected phenomena, and as such usually the code in question doesn't have easily usable tracepoints or other counters available for the specific problem being mitigated. We have a number of lines of monitoring defence against problems in production (host metrics, process metrics, service metrics, etc), and where it's not feasible to reliably monitor at another level, this kind of pragmatic netconsole monitoring is essential.
As one would expect, monitoring using printk is rather brittle for a number of reasons -- most notably that the message might disappear entirely in a new version of the kernel, or that the message may change in some way that the regex or other classification methods start to silently fail.
One factor that makes this even harder is that, under normal operation, many of these messages are never expected to be hit. For example, there may be a rare hardware bug which one wants to detect if it was to ever happen again, but its recurrence is not likely or anticipated. This precludes using something like checking whether the printk in question was printed somewhere fleetwide recently to determine whether the message in question is still present or not, since we don't anticipate that it should be printed anywhere, but still need to monitor for its future presence in the long-term.
This class of issue has happened on a number of occasions, causing unhealthy machines with hardware issues to remain in production for longer than ideal. As a recent example, some monitoring around blk_update_request fell out of date and caused semi-broken machines to remain in production for longer than would be desirable.
Searching through the codebase to find the message is also extremely fragile, because many of the messages are further constructed beyond their callsite (eg. btrfs_printk and other module-specific wrappers, each with their own functionality). Even if they aren't, guessing the format and formulation of the underlying message based on the aesthetics of the message emitted is not a recipe for success at scale, and our previous issues with fleetwide machine health checking demonstrate as much.
This provides a solution to the issue of silently changed or deleted printks: we record pointers to all printk format strings known at compile time into a new .printk_index section, both in vmlinux and modules. At runtime, this can then be iterated by looking at <debugfs>/printk/index/<module>, which emits the following format, both readable by humans and able to be parsed by machines:
$ head -1 vmlinux; shuf -n 5 vmlinux # <level[,flags]> filename:line function "format" <5> block/blk-settings.c:661 disk_stack_limits "%s: Warning: Device %s is misaligned\n" <4> kernel/trace/trace.c:8296 trace_create_file "Could not create tracefs '%s' entry\n" <6> arch/x86/kernel/hpet.c:144 _hpet_print_config "hpet: %s(%d):\n" <6> init/do_mounts.c:605 prepare_namespace "Waiting for root device %s...\n" <6> drivers/acpi/osl.c:1410 acpi_no_auto_serialize_setup "ACPI: auto-serialization disabled\n"
This mitigates the majority of cases where we have a highly-specific printk which we want to match on, as we can now enumerate and check whether the format changed or the printk callsite disappeared entirely in userspace. This allows us to catch changes to printks we monitor earlier and decide what to do about it before it becomes problematic.
There is no additional runtime cost for printk callers or printk itself, and the assembly generated is exactly the same.
Signed-off-by: Chris Down <chris@chrisdown.name> Cc: Petr Mladek <pmladek@suse.com> Cc: Jessica Yu <jeyu@kernel.org> Cc: Sergey Senozhatsky <sergey.senozhatsky@gmail.com> Cc: John Ogness <john.ogness@linutronix.de> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Kees Cook <keescook@chromium.org> Reviewed-by: Petr Mladek <pmladek@suse.com> Tested-by: Petr Mladek <pmladek@suse.com> Reported-by: kernel test robot <lkp@intel.com> Acked-by: Andy Shevchenko <andy.shevchenko@gmail.com> Acked-by: Jessica Yu <jeyu@kernel.org> # for module.{c,h} Signed-off-by: Petr Mladek <pmladek@suse.com> Link: https://lore.kernel.org/r/e42070983637ac5e384f17fbdbe86d19c7b212a5.1623775748.git.chris@chrisdown.name
|
/linux/kernel/printk/ |
H A D | index.c | 337015573718b161891a3473d25f59273f2e626b Tue Jun 15 18:52:53 CEST 2021 Chris Down <chris@chrisdown.name> printk: Userspace format indexing support
We have a number of systems industry-wide that have a subset of their functionality that works as follows:
1. Receive a message from local kmsg, serial console, or netconsole; 2. Apply a set of rules to classify the message; 3. Do something based on this classification (like scheduling a remediation for the machine), rinse, and repeat.
As a couple of examples of places we have this implemented just inside Facebook, although this isn't a Facebook-specific problem, we have this inside our netconsole processing (for alarm classification), and as part of our machine health checking. We use these messages to determine fairly important metrics around production health, and it's important that we get them right.
While for some kinds of issues we have counters, tracepoints, or metrics with a stable interface which can reliably indicate the issue, in order to react to production issues quickly we need to work with the interface which most kernel developers naturally use when developing: printk.
Most production issues come from unexpected phenomena, and as such usually the code in question doesn't have easily usable tracepoints or other counters available for the specific problem being mitigated. We have a number of lines of monitoring defence against problems in production (host metrics, process metrics, service metrics, etc), and where it's not feasible to reliably monitor at another level, this kind of pragmatic netconsole monitoring is essential.
As one would expect, monitoring using printk is rather brittle for a number of reasons -- most notably that the message might disappear entirely in a new version of the kernel, or that the message may change in some way that the regex or other classification methods start to silently fail.
One factor that makes this even harder is that, under normal operation, many of these messages are never expected to be hit. For example, there may be a rare hardware bug which one wants to detect if it was to ever happen again, but its recurrence is not likely or anticipated. This precludes using something like checking whether the printk in question was printed somewhere fleetwide recently to determine whether the message in question is still present or not, since we don't anticipate that it should be printed anywhere, but still need to monitor for its future presence in the long-term.
This class of issue has happened on a number of occasions, causing unhealthy machines with hardware issues to remain in production for longer than ideal. As a recent example, some monitoring around blk_update_request fell out of date and caused semi-broken machines to remain in production for longer than would be desirable.
Searching through the codebase to find the message is also extremely fragile, because many of the messages are further constructed beyond their callsite (eg. btrfs_printk and other module-specific wrappers, each with their own functionality). Even if they aren't, guessing the format and formulation of the underlying message based on the aesthetics of the message emitted is not a recipe for success at scale, and our previous issues with fleetwide machine health checking demonstrate as much.
This provides a solution to the issue of silently changed or deleted printks: we record pointers to all printk format strings known at compile time into a new .printk_index section, both in vmlinux and modules. At runtime, this can then be iterated by looking at <debugfs>/printk/index/<module>, which emits the following format, both readable by humans and able to be parsed by machines:
$ head -1 vmlinux; shuf -n 5 vmlinux # <level[,flags]> filename:line function "format" <5> block/blk-settings.c:661 disk_stack_limits "%s: Warning: Device %s is misaligned\n" <4> kernel/trace/trace.c:8296 trace_create_file "Could not create tracefs '%s' entry\n" <6> arch/x86/kernel/hpet.c:144 _hpet_print_config "hpet: %s(%d):\n" <6> init/do_mounts.c:605 prepare_namespace "Waiting for root device %s...\n" <6> drivers/acpi/osl.c:1410 acpi_no_auto_serialize_setup "ACPI: auto-serialization disabled\n"
This mitigates the majority of cases where we have a highly-specific printk which we want to match on, as we can now enumerate and check whether the format changed or the printk callsite disappeared entirely in userspace. This allows us to catch changes to printks we monitor earlier and decide what to do about it before it becomes problematic.
There is no additional runtime cost for printk callers or printk itself, and the assembly generated is exactly the same.
Signed-off-by: Chris Down <chris@chrisdown.name> Cc: Petr Mladek <pmladek@suse.com> Cc: Jessica Yu <jeyu@kernel.org> Cc: Sergey Senozhatsky <sergey.senozhatsky@gmail.com> Cc: John Ogness <john.ogness@linutronix.de> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Kees Cook <keescook@chromium.org> Reviewed-by: Petr Mladek <pmladek@suse.com> Tested-by: Petr Mladek <pmladek@suse.com> Reported-by: kernel test robot <lkp@intel.com> Acked-by: Andy Shevchenko <andy.shevchenko@gmail.com> Acked-by: Jessica Yu <jeyu@kernel.org> # for module.{c,h} Signed-off-by: Petr Mladek <pmladek@suse.com> Link: https://lore.kernel.org/r/e42070983637ac5e384f17fbdbe86d19c7b212a5.1623775748.git.chris@chrisdown.name
|
H A D | Makefile | diff 337015573718b161891a3473d25f59273f2e626b Tue Jun 15 18:52:53 CEST 2021 Chris Down <chris@chrisdown.name> printk: Userspace format indexing support
We have a number of systems industry-wide that have a subset of their functionality that works as follows:
1. Receive a message from local kmsg, serial console, or netconsole; 2. Apply a set of rules to classify the message; 3. Do something based on this classification (like scheduling a remediation for the machine), rinse, and repeat.
As a couple of examples of places we have this implemented just inside Facebook, although this isn't a Facebook-specific problem, we have this inside our netconsole processing (for alarm classification), and as part of our machine health checking. We use these messages to determine fairly important metrics around production health, and it's important that we get them right.
While for some kinds of issues we have counters, tracepoints, or metrics with a stable interface which can reliably indicate the issue, in order to react to production issues quickly we need to work with the interface which most kernel developers naturally use when developing: printk.
Most production issues come from unexpected phenomena, and as such usually the code in question doesn't have easily usable tracepoints or other counters available for the specific problem being mitigated. We have a number of lines of monitoring defence against problems in production (host metrics, process metrics, service metrics, etc), and where it's not feasible to reliably monitor at another level, this kind of pragmatic netconsole monitoring is essential.
As one would expect, monitoring using printk is rather brittle for a number of reasons -- most notably that the message might disappear entirely in a new version of the kernel, or that the message may change in some way that the regex or other classification methods start to silently fail.
One factor that makes this even harder is that, under normal operation, many of these messages are never expected to be hit. For example, there may be a rare hardware bug which one wants to detect if it was to ever happen again, but its recurrence is not likely or anticipated. This precludes using something like checking whether the printk in question was printed somewhere fleetwide recently to determine whether the message in question is still present or not, since we don't anticipate that it should be printed anywhere, but still need to monitor for its future presence in the long-term.
This class of issue has happened on a number of occasions, causing unhealthy machines with hardware issues to remain in production for longer than ideal. As a recent example, some monitoring around blk_update_request fell out of date and caused semi-broken machines to remain in production for longer than would be desirable.
Searching through the codebase to find the message is also extremely fragile, because many of the messages are further constructed beyond their callsite (eg. btrfs_printk and other module-specific wrappers, each with their own functionality). Even if they aren't, guessing the format and formulation of the underlying message based on the aesthetics of the message emitted is not a recipe for success at scale, and our previous issues with fleetwide machine health checking demonstrate as much.
This provides a solution to the issue of silently changed or deleted printks: we record pointers to all printk format strings known at compile time into a new .printk_index section, both in vmlinux and modules. At runtime, this can then be iterated by looking at <debugfs>/printk/index/<module>, which emits the following format, both readable by humans and able to be parsed by machines:
$ head -1 vmlinux; shuf -n 5 vmlinux # <level[,flags]> filename:line function "format" <5> block/blk-settings.c:661 disk_stack_limits "%s: Warning: Device %s is misaligned\n" <4> kernel/trace/trace.c:8296 trace_create_file "Could not create tracefs '%s' entry\n" <6> arch/x86/kernel/hpet.c:144 _hpet_print_config "hpet: %s(%d):\n" <6> init/do_mounts.c:605 prepare_namespace "Waiting for root device %s...\n" <6> drivers/acpi/osl.c:1410 acpi_no_auto_serialize_setup "ACPI: auto-serialization disabled\n"
This mitigates the majority of cases where we have a highly-specific printk which we want to match on, as we can now enumerate and check whether the format changed or the printk callsite disappeared entirely in userspace. This allows us to catch changes to printks we monitor earlier and decide what to do about it before it becomes problematic.
There is no additional runtime cost for printk callers or printk itself, and the assembly generated is exactly the same.
Signed-off-by: Chris Down <chris@chrisdown.name> Cc: Petr Mladek <pmladek@suse.com> Cc: Jessica Yu <jeyu@kernel.org> Cc: Sergey Senozhatsky <sergey.senozhatsky@gmail.com> Cc: John Ogness <john.ogness@linutronix.de> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Kees Cook <keescook@chromium.org> Reviewed-by: Petr Mladek <pmladek@suse.com> Tested-by: Petr Mladek <pmladek@suse.com> Reported-by: kernel test robot <lkp@intel.com> Acked-by: Andy Shevchenko <andy.shevchenko@gmail.com> Acked-by: Jessica Yu <jeyu@kernel.org> # for module.{c,h} Signed-off-by: Petr Mladek <pmladek@suse.com> Link: https://lore.kernel.org/r/e42070983637ac5e384f17fbdbe86d19c7b212a5.1623775748.git.chris@chrisdown.name
|
H A D | printk.c | diff 337015573718b161891a3473d25f59273f2e626b Tue Jun 15 18:52:53 CEST 2021 Chris Down <chris@chrisdown.name> printk: Userspace format indexing support
We have a number of systems industry-wide that have a subset of their functionality that works as follows:
1. Receive a message from local kmsg, serial console, or netconsole; 2. Apply a set of rules to classify the message; 3. Do something based on this classification (like scheduling a remediation for the machine), rinse, and repeat.
As a couple of examples of places we have this implemented just inside Facebook, although this isn't a Facebook-specific problem, we have this inside our netconsole processing (for alarm classification), and as part of our machine health checking. We use these messages to determine fairly important metrics around production health, and it's important that we get them right.
While for some kinds of issues we have counters, tracepoints, or metrics with a stable interface which can reliably indicate the issue, in order to react to production issues quickly we need to work with the interface which most kernel developers naturally use when developing: printk.
Most production issues come from unexpected phenomena, and as such usually the code in question doesn't have easily usable tracepoints or other counters available for the specific problem being mitigated. We have a number of lines of monitoring defence against problems in production (host metrics, process metrics, service metrics, etc), and where it's not feasible to reliably monitor at another level, this kind of pragmatic netconsole monitoring is essential.
As one would expect, monitoring using printk is rather brittle for a number of reasons -- most notably that the message might disappear entirely in a new version of the kernel, or that the message may change in some way that the regex or other classification methods start to silently fail.
One factor that makes this even harder is that, under normal operation, many of these messages are never expected to be hit. For example, there may be a rare hardware bug which one wants to detect if it was to ever happen again, but its recurrence is not likely or anticipated. This precludes using something like checking whether the printk in question was printed somewhere fleetwide recently to determine whether the message in question is still present or not, since we don't anticipate that it should be printed anywhere, but still need to monitor for its future presence in the long-term.
This class of issue has happened on a number of occasions, causing unhealthy machines with hardware issues to remain in production for longer than ideal. As a recent example, some monitoring around blk_update_request fell out of date and caused semi-broken machines to remain in production for longer than would be desirable.
Searching through the codebase to find the message is also extremely fragile, because many of the messages are further constructed beyond their callsite (eg. btrfs_printk and other module-specific wrappers, each with their own functionality). Even if they aren't, guessing the format and formulation of the underlying message based on the aesthetics of the message emitted is not a recipe for success at scale, and our previous issues with fleetwide machine health checking demonstrate as much.
This provides a solution to the issue of silently changed or deleted printks: we record pointers to all printk format strings known at compile time into a new .printk_index section, both in vmlinux and modules. At runtime, this can then be iterated by looking at <debugfs>/printk/index/<module>, which emits the following format, both readable by humans and able to be parsed by machines:
$ head -1 vmlinux; shuf -n 5 vmlinux # <level[,flags]> filename:line function "format" <5> block/blk-settings.c:661 disk_stack_limits "%s: Warning: Device %s is misaligned\n" <4> kernel/trace/trace.c:8296 trace_create_file "Could not create tracefs '%s' entry\n" <6> arch/x86/kernel/hpet.c:144 _hpet_print_config "hpet: %s(%d):\n" <6> init/do_mounts.c:605 prepare_namespace "Waiting for root device %s...\n" <6> drivers/acpi/osl.c:1410 acpi_no_auto_serialize_setup "ACPI: auto-serialization disabled\n"
This mitigates the majority of cases where we have a highly-specific printk which we want to match on, as we can now enumerate and check whether the format changed or the printk callsite disappeared entirely in userspace. This allows us to catch changes to printks we monitor earlier and decide what to do about it before it becomes problematic.
There is no additional runtime cost for printk callers or printk itself, and the assembly generated is exactly the same.
Signed-off-by: Chris Down <chris@chrisdown.name> Cc: Petr Mladek <pmladek@suse.com> Cc: Jessica Yu <jeyu@kernel.org> Cc: Sergey Senozhatsky <sergey.senozhatsky@gmail.com> Cc: John Ogness <john.ogness@linutronix.de> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Kees Cook <keescook@chromium.org> Reviewed-by: Petr Mladek <pmladek@suse.com> Tested-by: Petr Mladek <pmladek@suse.com> Reported-by: kernel test robot <lkp@intel.com> Acked-by: Andy Shevchenko <andy.shevchenko@gmail.com> Acked-by: Jessica Yu <jeyu@kernel.org> # for module.{c,h} Signed-off-by: Petr Mladek <pmladek@suse.com> Link: https://lore.kernel.org/r/e42070983637ac5e384f17fbdbe86d19c7b212a5.1623775748.git.chris@chrisdown.name
|
/linux/arch/arm/lib/ |
H A D | backtrace-clang.S | diff 337015573718b161891a3473d25f59273f2e626b Tue Jun 15 18:52:53 CEST 2021 Chris Down <chris@chrisdown.name> printk: Userspace format indexing support
We have a number of systems industry-wide that have a subset of their functionality that works as follows:
1. Receive a message from local kmsg, serial console, or netconsole; 2. Apply a set of rules to classify the message; 3. Do something based on this classification (like scheduling a remediation for the machine), rinse, and repeat.
As a couple of examples of places we have this implemented just inside Facebook, although this isn't a Facebook-specific problem, we have this inside our netconsole processing (for alarm classification), and as part of our machine health checking. We use these messages to determine fairly important metrics around production health, and it's important that we get them right.
While for some kinds of issues we have counters, tracepoints, or metrics with a stable interface which can reliably indicate the issue, in order to react to production issues quickly we need to work with the interface which most kernel developers naturally use when developing: printk.
Most production issues come from unexpected phenomena, and as such usually the code in question doesn't have easily usable tracepoints or other counters available for the specific problem being mitigated. We have a number of lines of monitoring defence against problems in production (host metrics, process metrics, service metrics, etc), and where it's not feasible to reliably monitor at another level, this kind of pragmatic netconsole monitoring is essential.
As one would expect, monitoring using printk is rather brittle for a number of reasons -- most notably that the message might disappear entirely in a new version of the kernel, or that the message may change in some way that the regex or other classification methods start to silently fail.
One factor that makes this even harder is that, under normal operation, many of these messages are never expected to be hit. For example, there may be a rare hardware bug which one wants to detect if it was to ever happen again, but its recurrence is not likely or anticipated. This precludes using something like checking whether the printk in question was printed somewhere fleetwide recently to determine whether the message in question is still present or not, since we don't anticipate that it should be printed anywhere, but still need to monitor for its future presence in the long-term.
This class of issue has happened on a number of occasions, causing unhealthy machines with hardware issues to remain in production for longer than ideal. As a recent example, some monitoring around blk_update_request fell out of date and caused semi-broken machines to remain in production for longer than would be desirable.
Searching through the codebase to find the message is also extremely fragile, because many of the messages are further constructed beyond their callsite (eg. btrfs_printk and other module-specific wrappers, each with their own functionality). Even if they aren't, guessing the format and formulation of the underlying message based on the aesthetics of the message emitted is not a recipe for success at scale, and our previous issues with fleetwide machine health checking demonstrate as much.
This provides a solution to the issue of silently changed or deleted printks: we record pointers to all printk format strings known at compile time into a new .printk_index section, both in vmlinux and modules. At runtime, this can then be iterated by looking at <debugfs>/printk/index/<module>, which emits the following format, both readable by humans and able to be parsed by machines:
$ head -1 vmlinux; shuf -n 5 vmlinux # <level[,flags]> filename:line function "format" <5> block/blk-settings.c:661 disk_stack_limits "%s: Warning: Device %s is misaligned\n" <4> kernel/trace/trace.c:8296 trace_create_file "Could not create tracefs '%s' entry\n" <6> arch/x86/kernel/hpet.c:144 _hpet_print_config "hpet: %s(%d):\n" <6> init/do_mounts.c:605 prepare_namespace "Waiting for root device %s...\n" <6> drivers/acpi/osl.c:1410 acpi_no_auto_serialize_setup "ACPI: auto-serialization disabled\n"
This mitigates the majority of cases where we have a highly-specific printk which we want to match on, as we can now enumerate and check whether the format changed or the printk callsite disappeared entirely in userspace. This allows us to catch changes to printks we monitor earlier and decide what to do about it before it becomes problematic.
There is no additional runtime cost for printk callers or printk itself, and the assembly generated is exactly the same.
Signed-off-by: Chris Down <chris@chrisdown.name> Cc: Petr Mladek <pmladek@suse.com> Cc: Jessica Yu <jeyu@kernel.org> Cc: Sergey Senozhatsky <sergey.senozhatsky@gmail.com> Cc: John Ogness <john.ogness@linutronix.de> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Kees Cook <keescook@chromium.org> Reviewed-by: Petr Mladek <pmladek@suse.com> Tested-by: Petr Mladek <pmladek@suse.com> Reported-by: kernel test robot <lkp@intel.com> Acked-by: Andy Shevchenko <andy.shevchenko@gmail.com> Acked-by: Jessica Yu <jeyu@kernel.org> # for module.{c,h} Signed-off-by: Petr Mladek <pmladek@suse.com> Link: https://lore.kernel.org/r/e42070983637ac5e384f17fbdbe86d19c7b212a5.1623775748.git.chris@chrisdown.name
|
H A D | backtrace.S | diff 337015573718b161891a3473d25f59273f2e626b Tue Jun 15 18:52:53 CEST 2021 Chris Down <chris@chrisdown.name> printk: Userspace format indexing support
We have a number of systems industry-wide that have a subset of their functionality that works as follows:
1. Receive a message from local kmsg, serial console, or netconsole; 2. Apply a set of rules to classify the message; 3. Do something based on this classification (like scheduling a remediation for the machine), rinse, and repeat.
As a couple of examples of places we have this implemented just inside Facebook, although this isn't a Facebook-specific problem, we have this inside our netconsole processing (for alarm classification), and as part of our machine health checking. We use these messages to determine fairly important metrics around production health, and it's important that we get them right.
While for some kinds of issues we have counters, tracepoints, or metrics with a stable interface which can reliably indicate the issue, in order to react to production issues quickly we need to work with the interface which most kernel developers naturally use when developing: printk.
Most production issues come from unexpected phenomena, and as such usually the code in question doesn't have easily usable tracepoints or other counters available for the specific problem being mitigated. We have a number of lines of monitoring defence against problems in production (host metrics, process metrics, service metrics, etc), and where it's not feasible to reliably monitor at another level, this kind of pragmatic netconsole monitoring is essential.
As one would expect, monitoring using printk is rather brittle for a number of reasons -- most notably that the message might disappear entirely in a new version of the kernel, or that the message may change in some way that the regex or other classification methods start to silently fail.
One factor that makes this even harder is that, under normal operation, many of these messages are never expected to be hit. For example, there may be a rare hardware bug which one wants to detect if it was to ever happen again, but its recurrence is not likely or anticipated. This precludes using something like checking whether the printk in question was printed somewhere fleetwide recently to determine whether the message in question is still present or not, since we don't anticipate that it should be printed anywhere, but still need to monitor for its future presence in the long-term.
This class of issue has happened on a number of occasions, causing unhealthy machines with hardware issues to remain in production for longer than ideal. As a recent example, some monitoring around blk_update_request fell out of date and caused semi-broken machines to remain in production for longer than would be desirable.
Searching through the codebase to find the message is also extremely fragile, because many of the messages are further constructed beyond their callsite (eg. btrfs_printk and other module-specific wrappers, each with their own functionality). Even if they aren't, guessing the format and formulation of the underlying message based on the aesthetics of the message emitted is not a recipe for success at scale, and our previous issues with fleetwide machine health checking demonstrate as much.
This provides a solution to the issue of silently changed or deleted printks: we record pointers to all printk format strings known at compile time into a new .printk_index section, both in vmlinux and modules. At runtime, this can then be iterated by looking at <debugfs>/printk/index/<module>, which emits the following format, both readable by humans and able to be parsed by machines:
$ head -1 vmlinux; shuf -n 5 vmlinux # <level[,flags]> filename:line function "format" <5> block/blk-settings.c:661 disk_stack_limits "%s: Warning: Device %s is misaligned\n" <4> kernel/trace/trace.c:8296 trace_create_file "Could not create tracefs '%s' entry\n" <6> arch/x86/kernel/hpet.c:144 _hpet_print_config "hpet: %s(%d):\n" <6> init/do_mounts.c:605 prepare_namespace "Waiting for root device %s...\n" <6> drivers/acpi/osl.c:1410 acpi_no_auto_serialize_setup "ACPI: auto-serialization disabled\n"
This mitigates the majority of cases where we have a highly-specific printk which we want to match on, as we can now enumerate and check whether the format changed or the printk callsite disappeared entirely in userspace. This allows us to catch changes to printks we monitor earlier and decide what to do about it before it becomes problematic.
There is no additional runtime cost for printk callers or printk itself, and the assembly generated is exactly the same.
Signed-off-by: Chris Down <chris@chrisdown.name> Cc: Petr Mladek <pmladek@suse.com> Cc: Jessica Yu <jeyu@kernel.org> Cc: Sergey Senozhatsky <sergey.senozhatsky@gmail.com> Cc: John Ogness <john.ogness@linutronix.de> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Kees Cook <keescook@chromium.org> Reviewed-by: Petr Mladek <pmladek@suse.com> Tested-by: Petr Mladek <pmladek@suse.com> Reported-by: kernel test robot <lkp@intel.com> Acked-by: Andy Shevchenko <andy.shevchenko@gmail.com> Acked-by: Jessica Yu <jeyu@kernel.org> # for module.{c,h} Signed-off-by: Petr Mladek <pmladek@suse.com> Link: https://lore.kernel.org/r/e42070983637ac5e384f17fbdbe86d19c7b212a5.1623775748.git.chris@chrisdown.name
|
/linux/arch/um/include/shared/ |
H A D | user.h | diff 337015573718b161891a3473d25f59273f2e626b Tue Jun 15 18:52:53 CEST 2021 Chris Down <chris@chrisdown.name> printk: Userspace format indexing support
We have a number of systems industry-wide that have a subset of their functionality that works as follows:
1. Receive a message from local kmsg, serial console, or netconsole; 2. Apply a set of rules to classify the message; 3. Do something based on this classification (like scheduling a remediation for the machine), rinse, and repeat.
As a couple of examples of places we have this implemented just inside Facebook, although this isn't a Facebook-specific problem, we have this inside our netconsole processing (for alarm classification), and as part of our machine health checking. We use these messages to determine fairly important metrics around production health, and it's important that we get them right.
While for some kinds of issues we have counters, tracepoints, or metrics with a stable interface which can reliably indicate the issue, in order to react to production issues quickly we need to work with the interface which most kernel developers naturally use when developing: printk.
Most production issues come from unexpected phenomena, and as such usually the code in question doesn't have easily usable tracepoints or other counters available for the specific problem being mitigated. We have a number of lines of monitoring defence against problems in production (host metrics, process metrics, service metrics, etc), and where it's not feasible to reliably monitor at another level, this kind of pragmatic netconsole monitoring is essential.
As one would expect, monitoring using printk is rather brittle for a number of reasons -- most notably that the message might disappear entirely in a new version of the kernel, or that the message may change in some way that the regex or other classification methods start to silently fail.
One factor that makes this even harder is that, under normal operation, many of these messages are never expected to be hit. For example, there may be a rare hardware bug which one wants to detect if it was to ever happen again, but its recurrence is not likely or anticipated. This precludes using something like checking whether the printk in question was printed somewhere fleetwide recently to determine whether the message in question is still present or not, since we don't anticipate that it should be printed anywhere, but still need to monitor for its future presence in the long-term.
This class of issue has happened on a number of occasions, causing unhealthy machines with hardware issues to remain in production for longer than ideal. As a recent example, some monitoring around blk_update_request fell out of date and caused semi-broken machines to remain in production for longer than would be desirable.
Searching through the codebase to find the message is also extremely fragile, because many of the messages are further constructed beyond their callsite (eg. btrfs_printk and other module-specific wrappers, each with their own functionality). Even if they aren't, guessing the format and formulation of the underlying message based on the aesthetics of the message emitted is not a recipe for success at scale, and our previous issues with fleetwide machine health checking demonstrate as much.
This provides a solution to the issue of silently changed or deleted printks: we record pointers to all printk format strings known at compile time into a new .printk_index section, both in vmlinux and modules. At runtime, this can then be iterated by looking at <debugfs>/printk/index/<module>, which emits the following format, both readable by humans and able to be parsed by machines:
$ head -1 vmlinux; shuf -n 5 vmlinux # <level[,flags]> filename:line function "format" <5> block/blk-settings.c:661 disk_stack_limits "%s: Warning: Device %s is misaligned\n" <4> kernel/trace/trace.c:8296 trace_create_file "Could not create tracefs '%s' entry\n" <6> arch/x86/kernel/hpet.c:144 _hpet_print_config "hpet: %s(%d):\n" <6> init/do_mounts.c:605 prepare_namespace "Waiting for root device %s...\n" <6> drivers/acpi/osl.c:1410 acpi_no_auto_serialize_setup "ACPI: auto-serialization disabled\n"
This mitigates the majority of cases where we have a highly-specific printk which we want to match on, as we can now enumerate and check whether the format changed or the printk callsite disappeared entirely in userspace. This allows us to catch changes to printks we monitor earlier and decide what to do about it before it becomes problematic.
There is no additional runtime cost for printk callers or printk itself, and the assembly generated is exactly the same.
Signed-off-by: Chris Down <chris@chrisdown.name> Cc: Petr Mladek <pmladek@suse.com> Cc: Jessica Yu <jeyu@kernel.org> Cc: Sergey Senozhatsky <sergey.senozhatsky@gmail.com> Cc: John Ogness <john.ogness@linutronix.de> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Kees Cook <keescook@chromium.org> Reviewed-by: Petr Mladek <pmladek@suse.com> Tested-by: Petr Mladek <pmladek@suse.com> Reported-by: kernel test robot <lkp@intel.com> Acked-by: Andy Shevchenko <andy.shevchenko@gmail.com> Acked-by: Jessica Yu <jeyu@kernel.org> # for module.{c,h} Signed-off-by: Petr Mladek <pmladek@suse.com> Link: https://lore.kernel.org/r/e42070983637ac5e384f17fbdbe86d19c7b212a5.1623775748.git.chris@chrisdown.name
|
/linux/arch/arm/kernel/ |
H A D | entry-v7m.S | diff 337015573718b161891a3473d25f59273f2e626b Tue Jun 15 18:52:53 CEST 2021 Chris Down <chris@chrisdown.name> printk: Userspace format indexing support
We have a number of systems industry-wide that have a subset of their functionality that works as follows:
1. Receive a message from local kmsg, serial console, or netconsole; 2. Apply a set of rules to classify the message; 3. Do something based on this classification (like scheduling a remediation for the machine), rinse, and repeat.
As a couple of examples of places we have this implemented just inside Facebook, although this isn't a Facebook-specific problem, we have this inside our netconsole processing (for alarm classification), and as part of our machine health checking. We use these messages to determine fairly important metrics around production health, and it's important that we get them right.
While for some kinds of issues we have counters, tracepoints, or metrics with a stable interface which can reliably indicate the issue, in order to react to production issues quickly we need to work with the interface which most kernel developers naturally use when developing: printk.
Most production issues come from unexpected phenomena, and as such usually the code in question doesn't have easily usable tracepoints or other counters available for the specific problem being mitigated. We have a number of lines of monitoring defence against problems in production (host metrics, process metrics, service metrics, etc), and where it's not feasible to reliably monitor at another level, this kind of pragmatic netconsole monitoring is essential.
As one would expect, monitoring using printk is rather brittle for a number of reasons -- most notably that the message might disappear entirely in a new version of the kernel, or that the message may change in some way that the regex or other classification methods start to silently fail.
One factor that makes this even harder is that, under normal operation, many of these messages are never expected to be hit. For example, there may be a rare hardware bug which one wants to detect if it was to ever happen again, but its recurrence is not likely or anticipated. This precludes using something like checking whether the printk in question was printed somewhere fleetwide recently to determine whether the message in question is still present or not, since we don't anticipate that it should be printed anywhere, but still need to monitor for its future presence in the long-term.
This class of issue has happened on a number of occasions, causing unhealthy machines with hardware issues to remain in production for longer than ideal. As a recent example, some monitoring around blk_update_request fell out of date and caused semi-broken machines to remain in production for longer than would be desirable.
Searching through the codebase to find the message is also extremely fragile, because many of the messages are further constructed beyond their callsite (eg. btrfs_printk and other module-specific wrappers, each with their own functionality). Even if they aren't, guessing the format and formulation of the underlying message based on the aesthetics of the message emitted is not a recipe for success at scale, and our previous issues with fleetwide machine health checking demonstrate as much.
This provides a solution to the issue of silently changed or deleted printks: we record pointers to all printk format strings known at compile time into a new .printk_index section, both in vmlinux and modules. At runtime, this can then be iterated by looking at <debugfs>/printk/index/<module>, which emits the following format, both readable by humans and able to be parsed by machines:
$ head -1 vmlinux; shuf -n 5 vmlinux # <level[,flags]> filename:line function "format" <5> block/blk-settings.c:661 disk_stack_limits "%s: Warning: Device %s is misaligned\n" <4> kernel/trace/trace.c:8296 trace_create_file "Could not create tracefs '%s' entry\n" <6> arch/x86/kernel/hpet.c:144 _hpet_print_config "hpet: %s(%d):\n" <6> init/do_mounts.c:605 prepare_namespace "Waiting for root device %s...\n" <6> drivers/acpi/osl.c:1410 acpi_no_auto_serialize_setup "ACPI: auto-serialization disabled\n"
This mitigates the majority of cases where we have a highly-specific printk which we want to match on, as we can now enumerate and check whether the format changed or the printk callsite disappeared entirely in userspace. This allows us to catch changes to printks we monitor earlier and decide what to do about it before it becomes problematic.
There is no additional runtime cost for printk callers or printk itself, and the assembly generated is exactly the same.
Signed-off-by: Chris Down <chris@chrisdown.name> Cc: Petr Mladek <pmladek@suse.com> Cc: Jessica Yu <jeyu@kernel.org> Cc: Sergey Senozhatsky <sergey.senozhatsky@gmail.com> Cc: John Ogness <john.ogness@linutronix.de> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Kees Cook <keescook@chromium.org> Reviewed-by: Petr Mladek <pmladek@suse.com> Tested-by: Petr Mladek <pmladek@suse.com> Reported-by: kernel test robot <lkp@intel.com> Acked-by: Andy Shevchenko <andy.shevchenko@gmail.com> Acked-by: Jessica Yu <jeyu@kernel.org> # for module.{c,h} Signed-off-by: Petr Mladek <pmladek@suse.com> Link: https://lore.kernel.org/r/e42070983637ac5e384f17fbdbe86d19c7b212a5.1623775748.git.chris@chrisdown.name
|
/linux/arch/openrisc/kernel/ |
H A D | entry.S | diff 337015573718b161891a3473d25f59273f2e626b Tue Jun 15 18:52:53 CEST 2021 Chris Down <chris@chrisdown.name> printk: Userspace format indexing support
We have a number of systems industry-wide that have a subset of their functionality that works as follows:
1. Receive a message from local kmsg, serial console, or netconsole; 2. Apply a set of rules to classify the message; 3. Do something based on this classification (like scheduling a remediation for the machine), rinse, and repeat.
As a couple of examples of places we have this implemented just inside Facebook, although this isn't a Facebook-specific problem, we have this inside our netconsole processing (for alarm classification), and as part of our machine health checking. We use these messages to determine fairly important metrics around production health, and it's important that we get them right.
While for some kinds of issues we have counters, tracepoints, or metrics with a stable interface which can reliably indicate the issue, in order to react to production issues quickly we need to work with the interface which most kernel developers naturally use when developing: printk.
Most production issues come from unexpected phenomena, and as such usually the code in question doesn't have easily usable tracepoints or other counters available for the specific problem being mitigated. We have a number of lines of monitoring defence against problems in production (host metrics, process metrics, service metrics, etc), and where it's not feasible to reliably monitor at another level, this kind of pragmatic netconsole monitoring is essential.
As one would expect, monitoring using printk is rather brittle for a number of reasons -- most notably that the message might disappear entirely in a new version of the kernel, or that the message may change in some way that the regex or other classification methods start to silently fail.
One factor that makes this even harder is that, under normal operation, many of these messages are never expected to be hit. For example, there may be a rare hardware bug which one wants to detect if it was to ever happen again, but its recurrence is not likely or anticipated. This precludes using something like checking whether the printk in question was printed somewhere fleetwide recently to determine whether the message in question is still present or not, since we don't anticipate that it should be printed anywhere, but still need to monitor for its future presence in the long-term.
This class of issue has happened on a number of occasions, causing unhealthy machines with hardware issues to remain in production for longer than ideal. As a recent example, some monitoring around blk_update_request fell out of date and caused semi-broken machines to remain in production for longer than would be desirable.
Searching through the codebase to find the message is also extremely fragile, because many of the messages are further constructed beyond their callsite (eg. btrfs_printk and other module-specific wrappers, each with their own functionality). Even if they aren't, guessing the format and formulation of the underlying message based on the aesthetics of the message emitted is not a recipe for success at scale, and our previous issues with fleetwide machine health checking demonstrate as much.
This provides a solution to the issue of silently changed or deleted printks: we record pointers to all printk format strings known at compile time into a new .printk_index section, both in vmlinux and modules. At runtime, this can then be iterated by looking at <debugfs>/printk/index/<module>, which emits the following format, both readable by humans and able to be parsed by machines:
$ head -1 vmlinux; shuf -n 5 vmlinux # <level[,flags]> filename:line function "format" <5> block/blk-settings.c:661 disk_stack_limits "%s: Warning: Device %s is misaligned\n" <4> kernel/trace/trace.c:8296 trace_create_file "Could not create tracefs '%s' entry\n" <6> arch/x86/kernel/hpet.c:144 _hpet_print_config "hpet: %s(%d):\n" <6> init/do_mounts.c:605 prepare_namespace "Waiting for root device %s...\n" <6> drivers/acpi/osl.c:1410 acpi_no_auto_serialize_setup "ACPI: auto-serialization disabled\n"
This mitigates the majority of cases where we have a highly-specific printk which we want to match on, as we can now enumerate and check whether the format changed or the printk callsite disappeared entirely in userspace. This allows us to catch changes to printks we monitor earlier and decide what to do about it before it becomes problematic.
There is no additional runtime cost for printk callers or printk itself, and the assembly generated is exactly the same.
Signed-off-by: Chris Down <chris@chrisdown.name> Cc: Petr Mladek <pmladek@suse.com> Cc: Jessica Yu <jeyu@kernel.org> Cc: Sergey Senozhatsky <sergey.senozhatsky@gmail.com> Cc: John Ogness <john.ogness@linutronix.de> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Kees Cook <keescook@chromium.org> Reviewed-by: Petr Mladek <pmladek@suse.com> Tested-by: Petr Mladek <pmladek@suse.com> Reported-by: kernel test robot <lkp@intel.com> Acked-by: Andy Shevchenko <andy.shevchenko@gmail.com> Acked-by: Jessica Yu <jeyu@kernel.org> # for module.{c,h} Signed-off-by: Petr Mladek <pmladek@suse.com> Link: https://lore.kernel.org/r/e42070983637ac5e384f17fbdbe86d19c7b212a5.1623775748.git.chris@chrisdown.name
|
/linux/arch/arm/vfp/ |
H A D | vfphw.S | diff 337015573718b161891a3473d25f59273f2e626b Tue Jun 15 18:52:53 CEST 2021 Chris Down <chris@chrisdown.name> printk: Userspace format indexing support
We have a number of systems industry-wide that have a subset of their functionality that works as follows:
1. Receive a message from local kmsg, serial console, or netconsole; 2. Apply a set of rules to classify the message; 3. Do something based on this classification (like scheduling a remediation for the machine), rinse, and repeat.
As a couple of examples of places we have this implemented just inside Facebook, although this isn't a Facebook-specific problem, we have this inside our netconsole processing (for alarm classification), and as part of our machine health checking. We use these messages to determine fairly important metrics around production health, and it's important that we get them right.
While for some kinds of issues we have counters, tracepoints, or metrics with a stable interface which can reliably indicate the issue, in order to react to production issues quickly we need to work with the interface which most kernel developers naturally use when developing: printk.
Most production issues come from unexpected phenomena, and as such usually the code in question doesn't have easily usable tracepoints or other counters available for the specific problem being mitigated. We have a number of lines of monitoring defence against problems in production (host metrics, process metrics, service metrics, etc), and where it's not feasible to reliably monitor at another level, this kind of pragmatic netconsole monitoring is essential.
As one would expect, monitoring using printk is rather brittle for a number of reasons -- most notably that the message might disappear entirely in a new version of the kernel, or that the message may change in some way that the regex or other classification methods start to silently fail.
One factor that makes this even harder is that, under normal operation, many of these messages are never expected to be hit. For example, there may be a rare hardware bug which one wants to detect if it was to ever happen again, but its recurrence is not likely or anticipated. This precludes using something like checking whether the printk in question was printed somewhere fleetwide recently to determine whether the message in question is still present or not, since we don't anticipate that it should be printed anywhere, but still need to monitor for its future presence in the long-term.
This class of issue has happened on a number of occasions, causing unhealthy machines with hardware issues to remain in production for longer than ideal. As a recent example, some monitoring around blk_update_request fell out of date and caused semi-broken machines to remain in production for longer than would be desirable.
Searching through the codebase to find the message is also extremely fragile, because many of the messages are further constructed beyond their callsite (eg. btrfs_printk and other module-specific wrappers, each with their own functionality). Even if they aren't, guessing the format and formulation of the underlying message based on the aesthetics of the message emitted is not a recipe for success at scale, and our previous issues with fleetwide machine health checking demonstrate as much.
This provides a solution to the issue of silently changed or deleted printks: we record pointers to all printk format strings known at compile time into a new .printk_index section, both in vmlinux and modules. At runtime, this can then be iterated by looking at <debugfs>/printk/index/<module>, which emits the following format, both readable by humans and able to be parsed by machines:
$ head -1 vmlinux; shuf -n 5 vmlinux # <level[,flags]> filename:line function "format" <5> block/blk-settings.c:661 disk_stack_limits "%s: Warning: Device %s is misaligned\n" <4> kernel/trace/trace.c:8296 trace_create_file "Could not create tracefs '%s' entry\n" <6> arch/x86/kernel/hpet.c:144 _hpet_print_config "hpet: %s(%d):\n" <6> init/do_mounts.c:605 prepare_namespace "Waiting for root device %s...\n" <6> drivers/acpi/osl.c:1410 acpi_no_auto_serialize_setup "ACPI: auto-serialization disabled\n"
This mitigates the majority of cases where we have a highly-specific printk which we want to match on, as we can now enumerate and check whether the format changed or the printk callsite disappeared entirely in userspace. This allows us to catch changes to printks we monitor earlier and decide what to do about it before it becomes problematic.
There is no additional runtime cost for printk callers or printk itself, and the assembly generated is exactly the same.
Signed-off-by: Chris Down <chris@chrisdown.name> Cc: Petr Mladek <pmladek@suse.com> Cc: Jessica Yu <jeyu@kernel.org> Cc: Sergey Senozhatsky <sergey.senozhatsky@gmail.com> Cc: John Ogness <john.ogness@linutronix.de> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Kees Cook <keescook@chromium.org> Reviewed-by: Petr Mladek <pmladek@suse.com> Tested-by: Petr Mladek <pmladek@suse.com> Reported-by: kernel test robot <lkp@intel.com> Acked-by: Andy Shevchenko <andy.shevchenko@gmail.com> Acked-by: Jessica Yu <jeyu@kernel.org> # for module.{c,h} Signed-off-by: Petr Mladek <pmladek@suse.com> Link: https://lore.kernel.org/r/e42070983637ac5e384f17fbdbe86d19c7b212a5.1623775748.git.chris@chrisdown.name
|
/linux/arch/x86/kernel/ |
H A D | head_32.S | diff 337015573718b161891a3473d25f59273f2e626b Tue Jun 15 18:52:53 CEST 2021 Chris Down <chris@chrisdown.name> printk: Userspace format indexing support
We have a number of systems industry-wide that have a subset of their functionality that works as follows:
1. Receive a message from local kmsg, serial console, or netconsole; 2. Apply a set of rules to classify the message; 3. Do something based on this classification (like scheduling a remediation for the machine), rinse, and repeat.
As a couple of examples of places we have this implemented just inside Facebook, although this isn't a Facebook-specific problem, we have this inside our netconsole processing (for alarm classification), and as part of our machine health checking. We use these messages to determine fairly important metrics around production health, and it's important that we get them right.
While for some kinds of issues we have counters, tracepoints, or metrics with a stable interface which can reliably indicate the issue, in order to react to production issues quickly we need to work with the interface which most kernel developers naturally use when developing: printk.
Most production issues come from unexpected phenomena, and as such usually the code in question doesn't have easily usable tracepoints or other counters available for the specific problem being mitigated. We have a number of lines of monitoring defence against problems in production (host metrics, process metrics, service metrics, etc), and where it's not feasible to reliably monitor at another level, this kind of pragmatic netconsole monitoring is essential.
As one would expect, monitoring using printk is rather brittle for a number of reasons -- most notably that the message might disappear entirely in a new version of the kernel, or that the message may change in some way that the regex or other classification methods start to silently fail.
One factor that makes this even harder is that, under normal operation, many of these messages are never expected to be hit. For example, there may be a rare hardware bug which one wants to detect if it was to ever happen again, but its recurrence is not likely or anticipated. This precludes using something like checking whether the printk in question was printed somewhere fleetwide recently to determine whether the message in question is still present or not, since we don't anticipate that it should be printed anywhere, but still need to monitor for its future presence in the long-term.
This class of issue has happened on a number of occasions, causing unhealthy machines with hardware issues to remain in production for longer than ideal. As a recent example, some monitoring around blk_update_request fell out of date and caused semi-broken machines to remain in production for longer than would be desirable.
Searching through the codebase to find the message is also extremely fragile, because many of the messages are further constructed beyond their callsite (eg. btrfs_printk and other module-specific wrappers, each with their own functionality). Even if they aren't, guessing the format and formulation of the underlying message based on the aesthetics of the message emitted is not a recipe for success at scale, and our previous issues with fleetwide machine health checking demonstrate as much.
This provides a solution to the issue of silently changed or deleted printks: we record pointers to all printk format strings known at compile time into a new .printk_index section, both in vmlinux and modules. At runtime, this can then be iterated by looking at <debugfs>/printk/index/<module>, which emits the following format, both readable by humans and able to be parsed by machines:
$ head -1 vmlinux; shuf -n 5 vmlinux # <level[,flags]> filename:line function "format" <5> block/blk-settings.c:661 disk_stack_limits "%s: Warning: Device %s is misaligned\n" <4> kernel/trace/trace.c:8296 trace_create_file "Could not create tracefs '%s' entry\n" <6> arch/x86/kernel/hpet.c:144 _hpet_print_config "hpet: %s(%d):\n" <6> init/do_mounts.c:605 prepare_namespace "Waiting for root device %s...\n" <6> drivers/acpi/osl.c:1410 acpi_no_auto_serialize_setup "ACPI: auto-serialization disabled\n"
This mitigates the majority of cases where we have a highly-specific printk which we want to match on, as we can now enumerate and check whether the format changed or the printk callsite disappeared entirely in userspace. This allows us to catch changes to printks we monitor earlier and decide what to do about it before it becomes problematic.
There is no additional runtime cost for printk callers or printk itself, and the assembly generated is exactly the same.
Signed-off-by: Chris Down <chris@chrisdown.name> Cc: Petr Mladek <pmladek@suse.com> Cc: Jessica Yu <jeyu@kernel.org> Cc: Sergey Senozhatsky <sergey.senozhatsky@gmail.com> Cc: John Ogness <john.ogness@linutronix.de> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Kees Cook <keescook@chromium.org> Reviewed-by: Petr Mladek <pmladek@suse.com> Tested-by: Petr Mladek <pmladek@suse.com> Reported-by: kernel test robot <lkp@intel.com> Acked-by: Andy Shevchenko <andy.shevchenko@gmail.com> Acked-by: Jessica Yu <jeyu@kernel.org> # for module.{c,h} Signed-off-by: Petr Mladek <pmladek@suse.com> Link: https://lore.kernel.org/r/e42070983637ac5e384f17fbdbe86d19c7b212a5.1623775748.git.chris@chrisdown.name
|
/linux/include/linux/ |
H A D | printk.h | diff 337015573718b161891a3473d25f59273f2e626b Tue Jun 15 18:52:53 CEST 2021 Chris Down <chris@chrisdown.name> printk: Userspace format indexing support
We have a number of systems industry-wide that have a subset of their functionality that works as follows:
1. Receive a message from local kmsg, serial console, or netconsole; 2. Apply a set of rules to classify the message; 3. Do something based on this classification (like scheduling a remediation for the machine), rinse, and repeat.
As a couple of examples of places we have this implemented just inside Facebook, although this isn't a Facebook-specific problem, we have this inside our netconsole processing (for alarm classification), and as part of our machine health checking. We use these messages to determine fairly important metrics around production health, and it's important that we get them right.
While for some kinds of issues we have counters, tracepoints, or metrics with a stable interface which can reliably indicate the issue, in order to react to production issues quickly we need to work with the interface which most kernel developers naturally use when developing: printk.
Most production issues come from unexpected phenomena, and as such usually the code in question doesn't have easily usable tracepoints or other counters available for the specific problem being mitigated. We have a number of lines of monitoring defence against problems in production (host metrics, process metrics, service metrics, etc), and where it's not feasible to reliably monitor at another level, this kind of pragmatic netconsole monitoring is essential.
As one would expect, monitoring using printk is rather brittle for a number of reasons -- most notably that the message might disappear entirely in a new version of the kernel, or that the message may change in some way that the regex or other classification methods start to silently fail.
One factor that makes this even harder is that, under normal operation, many of these messages are never expected to be hit. For example, there may be a rare hardware bug which one wants to detect if it was to ever happen again, but its recurrence is not likely or anticipated. This precludes using something like checking whether the printk in question was printed somewhere fleetwide recently to determine whether the message in question is still present or not, since we don't anticipate that it should be printed anywhere, but still need to monitor for its future presence in the long-term.
This class of issue has happened on a number of occasions, causing unhealthy machines with hardware issues to remain in production for longer than ideal. As a recent example, some monitoring around blk_update_request fell out of date and caused semi-broken machines to remain in production for longer than would be desirable.
Searching through the codebase to find the message is also extremely fragile, because many of the messages are further constructed beyond their callsite (eg. btrfs_printk and other module-specific wrappers, each with their own functionality). Even if they aren't, guessing the format and formulation of the underlying message based on the aesthetics of the message emitted is not a recipe for success at scale, and our previous issues with fleetwide machine health checking demonstrate as much.
This provides a solution to the issue of silently changed or deleted printks: we record pointers to all printk format strings known at compile time into a new .printk_index section, both in vmlinux and modules. At runtime, this can then be iterated by looking at <debugfs>/printk/index/<module>, which emits the following format, both readable by humans and able to be parsed by machines:
$ head -1 vmlinux; shuf -n 5 vmlinux # <level[,flags]> filename:line function "format" <5> block/blk-settings.c:661 disk_stack_limits "%s: Warning: Device %s is misaligned\n" <4> kernel/trace/trace.c:8296 trace_create_file "Could not create tracefs '%s' entry\n" <6> arch/x86/kernel/hpet.c:144 _hpet_print_config "hpet: %s(%d):\n" <6> init/do_mounts.c:605 prepare_namespace "Waiting for root device %s...\n" <6> drivers/acpi/osl.c:1410 acpi_no_auto_serialize_setup "ACPI: auto-serialization disabled\n"
This mitigates the majority of cases where we have a highly-specific printk which we want to match on, as we can now enumerate and check whether the format changed or the printk callsite disappeared entirely in userspace. This allows us to catch changes to printks we monitor earlier and decide what to do about it before it becomes problematic.
There is no additional runtime cost for printk callers or printk itself, and the assembly generated is exactly the same.
Signed-off-by: Chris Down <chris@chrisdown.name> Cc: Petr Mladek <pmladek@suse.com> Cc: Jessica Yu <jeyu@kernel.org> Cc: Sergey Senozhatsky <sergey.senozhatsky@gmail.com> Cc: John Ogness <john.ogness@linutronix.de> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Kees Cook <keescook@chromium.org> Reviewed-by: Petr Mladek <pmladek@suse.com> Tested-by: Petr Mladek <pmladek@suse.com> Reported-by: kernel test robot <lkp@intel.com> Acked-by: Andy Shevchenko <andy.shevchenko@gmail.com> Acked-by: Jessica Yu <jeyu@kernel.org> # for module.{c,h} Signed-off-by: Petr Mladek <pmladek@suse.com> Link: https://lore.kernel.org/r/e42070983637ac5e384f17fbdbe86d19c7b212a5.1623775748.git.chris@chrisdown.name
|
H A D | module.h | diff 337015573718b161891a3473d25f59273f2e626b Tue Jun 15 18:52:53 CEST 2021 Chris Down <chris@chrisdown.name> printk: Userspace format indexing support
We have a number of systems industry-wide that have a subset of their functionality that works as follows:
1. Receive a message from local kmsg, serial console, or netconsole; 2. Apply a set of rules to classify the message; 3. Do something based on this classification (like scheduling a remediation for the machine), rinse, and repeat.
As a couple of examples of places we have this implemented just inside Facebook, although this isn't a Facebook-specific problem, we have this inside our netconsole processing (for alarm classification), and as part of our machine health checking. We use these messages to determine fairly important metrics around production health, and it's important that we get them right.
While for some kinds of issues we have counters, tracepoints, or metrics with a stable interface which can reliably indicate the issue, in order to react to production issues quickly we need to work with the interface which most kernel developers naturally use when developing: printk.
Most production issues come from unexpected phenomena, and as such usually the code in question doesn't have easily usable tracepoints or other counters available for the specific problem being mitigated. We have a number of lines of monitoring defence against problems in production (host metrics, process metrics, service metrics, etc), and where it's not feasible to reliably monitor at another level, this kind of pragmatic netconsole monitoring is essential.
As one would expect, monitoring using printk is rather brittle for a number of reasons -- most notably that the message might disappear entirely in a new version of the kernel, or that the message may change in some way that the regex or other classification methods start to silently fail.
One factor that makes this even harder is that, under normal operation, many of these messages are never expected to be hit. For example, there may be a rare hardware bug which one wants to detect if it was to ever happen again, but its recurrence is not likely or anticipated. This precludes using something like checking whether the printk in question was printed somewhere fleetwide recently to determine whether the message in question is still present or not, since we don't anticipate that it should be printed anywhere, but still need to monitor for its future presence in the long-term.
This class of issue has happened on a number of occasions, causing unhealthy machines with hardware issues to remain in production for longer than ideal. As a recent example, some monitoring around blk_update_request fell out of date and caused semi-broken machines to remain in production for longer than would be desirable.
Searching through the codebase to find the message is also extremely fragile, because many of the messages are further constructed beyond their callsite (eg. btrfs_printk and other module-specific wrappers, each with their own functionality). Even if they aren't, guessing the format and formulation of the underlying message based on the aesthetics of the message emitted is not a recipe for success at scale, and our previous issues with fleetwide machine health checking demonstrate as much.
This provides a solution to the issue of silently changed or deleted printks: we record pointers to all printk format strings known at compile time into a new .printk_index section, both in vmlinux and modules. At runtime, this can then be iterated by looking at <debugfs>/printk/index/<module>, which emits the following format, both readable by humans and able to be parsed by machines:
$ head -1 vmlinux; shuf -n 5 vmlinux # <level[,flags]> filename:line function "format" <5> block/blk-settings.c:661 disk_stack_limits "%s: Warning: Device %s is misaligned\n" <4> kernel/trace/trace.c:8296 trace_create_file "Could not create tracefs '%s' entry\n" <6> arch/x86/kernel/hpet.c:144 _hpet_print_config "hpet: %s(%d):\n" <6> init/do_mounts.c:605 prepare_namespace "Waiting for root device %s...\n" <6> drivers/acpi/osl.c:1410 acpi_no_auto_serialize_setup "ACPI: auto-serialization disabled\n"
This mitigates the majority of cases where we have a highly-specific printk which we want to match on, as we can now enumerate and check whether the format changed or the printk callsite disappeared entirely in userspace. This allows us to catch changes to printks we monitor earlier and decide what to do about it before it becomes problematic.
There is no additional runtime cost for printk callers or printk itself, and the assembly generated is exactly the same.
Signed-off-by: Chris Down <chris@chrisdown.name> Cc: Petr Mladek <pmladek@suse.com> Cc: Jessica Yu <jeyu@kernel.org> Cc: Sergey Senozhatsky <sergey.senozhatsky@gmail.com> Cc: John Ogness <john.ogness@linutronix.de> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Kees Cook <keescook@chromium.org> Reviewed-by: Petr Mladek <pmladek@suse.com> Tested-by: Petr Mladek <pmladek@suse.com> Reported-by: kernel test robot <lkp@intel.com> Acked-by: Andy Shevchenko <andy.shevchenko@gmail.com> Acked-by: Jessica Yu <jeyu@kernel.org> # for module.{c,h} Signed-off-by: Petr Mladek <pmladek@suse.com> Link: https://lore.kernel.org/r/e42070983637ac5e384f17fbdbe86d19c7b212a5.1623775748.git.chris@chrisdown.name
|
/linux/include/asm-generic/ |
H A D | vmlinux.lds.h | diff 337015573718b161891a3473d25f59273f2e626b Tue Jun 15 18:52:53 CEST 2021 Chris Down <chris@chrisdown.name> printk: Userspace format indexing support
We have a number of systems industry-wide that have a subset of their functionality that works as follows:
1. Receive a message from local kmsg, serial console, or netconsole; 2. Apply a set of rules to classify the message; 3. Do something based on this classification (like scheduling a remediation for the machine), rinse, and repeat.
As a couple of examples of places we have this implemented just inside Facebook, although this isn't a Facebook-specific problem, we have this inside our netconsole processing (for alarm classification), and as part of our machine health checking. We use these messages to determine fairly important metrics around production health, and it's important that we get them right.
While for some kinds of issues we have counters, tracepoints, or metrics with a stable interface which can reliably indicate the issue, in order to react to production issues quickly we need to work with the interface which most kernel developers naturally use when developing: printk.
Most production issues come from unexpected phenomena, and as such usually the code in question doesn't have easily usable tracepoints or other counters available for the specific problem being mitigated. We have a number of lines of monitoring defence against problems in production (host metrics, process metrics, service metrics, etc), and where it's not feasible to reliably monitor at another level, this kind of pragmatic netconsole monitoring is essential.
As one would expect, monitoring using printk is rather brittle for a number of reasons -- most notably that the message might disappear entirely in a new version of the kernel, or that the message may change in some way that the regex or other classification methods start to silently fail.
One factor that makes this even harder is that, under normal operation, many of these messages are never expected to be hit. For example, there may be a rare hardware bug which one wants to detect if it was to ever happen again, but its recurrence is not likely or anticipated. This precludes using something like checking whether the printk in question was printed somewhere fleetwide recently to determine whether the message in question is still present or not, since we don't anticipate that it should be printed anywhere, but still need to monitor for its future presence in the long-term.
This class of issue has happened on a number of occasions, causing unhealthy machines with hardware issues to remain in production for longer than ideal. As a recent example, some monitoring around blk_update_request fell out of date and caused semi-broken machines to remain in production for longer than would be desirable.
Searching through the codebase to find the message is also extremely fragile, because many of the messages are further constructed beyond their callsite (eg. btrfs_printk and other module-specific wrappers, each with their own functionality). Even if they aren't, guessing the format and formulation of the underlying message based on the aesthetics of the message emitted is not a recipe for success at scale, and our previous issues with fleetwide machine health checking demonstrate as much.
This provides a solution to the issue of silently changed or deleted printks: we record pointers to all printk format strings known at compile time into a new .printk_index section, both in vmlinux and modules. At runtime, this can then be iterated by looking at <debugfs>/printk/index/<module>, which emits the following format, both readable by humans and able to be parsed by machines:
$ head -1 vmlinux; shuf -n 5 vmlinux # <level[,flags]> filename:line function "format" <5> block/blk-settings.c:661 disk_stack_limits "%s: Warning: Device %s is misaligned\n" <4> kernel/trace/trace.c:8296 trace_create_file "Could not create tracefs '%s' entry\n" <6> arch/x86/kernel/hpet.c:144 _hpet_print_config "hpet: %s(%d):\n" <6> init/do_mounts.c:605 prepare_namespace "Waiting for root device %s...\n" <6> drivers/acpi/osl.c:1410 acpi_no_auto_serialize_setup "ACPI: auto-serialization disabled\n"
This mitigates the majority of cases where we have a highly-specific printk which we want to match on, as we can now enumerate and check whether the format changed or the printk callsite disappeared entirely in userspace. This allows us to catch changes to printks we monitor earlier and decide what to do about it before it becomes problematic.
There is no additional runtime cost for printk callers or printk itself, and the assembly generated is exactly the same.
Signed-off-by: Chris Down <chris@chrisdown.name> Cc: Petr Mladek <pmladek@suse.com> Cc: Jessica Yu <jeyu@kernel.org> Cc: Sergey Senozhatsky <sergey.senozhatsky@gmail.com> Cc: John Ogness <john.ogness@linutronix.de> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Kees Cook <keescook@chromium.org> Reviewed-by: Petr Mladek <pmladek@suse.com> Tested-by: Petr Mladek <pmladek@suse.com> Reported-by: kernel test robot <lkp@intel.com> Acked-by: Andy Shevchenko <andy.shevchenko@gmail.com> Acked-by: Jessica Yu <jeyu@kernel.org> # for module.{c,h} Signed-off-by: Petr Mladek <pmladek@suse.com> Link: https://lore.kernel.org/r/e42070983637ac5e384f17fbdbe86d19c7b212a5.1623775748.git.chris@chrisdown.name
|
/linux/init/ |
H A D | Kconfig | diff 337015573718b161891a3473d25f59273f2e626b Tue Jun 15 18:52:53 CEST 2021 Chris Down <chris@chrisdown.name> printk: Userspace format indexing support
We have a number of systems industry-wide that have a subset of their functionality that works as follows:
1. Receive a message from local kmsg, serial console, or netconsole; 2. Apply a set of rules to classify the message; 3. Do something based on this classification (like scheduling a remediation for the machine), rinse, and repeat.
As a couple of examples of places we have this implemented just inside Facebook, although this isn't a Facebook-specific problem, we have this inside our netconsole processing (for alarm classification), and as part of our machine health checking. We use these messages to determine fairly important metrics around production health, and it's important that we get them right.
While for some kinds of issues we have counters, tracepoints, or metrics with a stable interface which can reliably indicate the issue, in order to react to production issues quickly we need to work with the interface which most kernel developers naturally use when developing: printk.
Most production issues come from unexpected phenomena, and as such usually the code in question doesn't have easily usable tracepoints or other counters available for the specific problem being mitigated. We have a number of lines of monitoring defence against problems in production (host metrics, process metrics, service metrics, etc), and where it's not feasible to reliably monitor at another level, this kind of pragmatic netconsole monitoring is essential.
As one would expect, monitoring using printk is rather brittle for a number of reasons -- most notably that the message might disappear entirely in a new version of the kernel, or that the message may change in some way that the regex or other classification methods start to silently fail.
One factor that makes this even harder is that, under normal operation, many of these messages are never expected to be hit. For example, there may be a rare hardware bug which one wants to detect if it was to ever happen again, but its recurrence is not likely or anticipated. This precludes using something like checking whether the printk in question was printed somewhere fleetwide recently to determine whether the message in question is still present or not, since we don't anticipate that it should be printed anywhere, but still need to monitor for its future presence in the long-term.
This class of issue has happened on a number of occasions, causing unhealthy machines with hardware issues to remain in production for longer than ideal. As a recent example, some monitoring around blk_update_request fell out of date and caused semi-broken machines to remain in production for longer than would be desirable.
Searching through the codebase to find the message is also extremely fragile, because many of the messages are further constructed beyond their callsite (eg. btrfs_printk and other module-specific wrappers, each with their own functionality). Even if they aren't, guessing the format and formulation of the underlying message based on the aesthetics of the message emitted is not a recipe for success at scale, and our previous issues with fleetwide machine health checking demonstrate as much.
This provides a solution to the issue of silently changed or deleted printks: we record pointers to all printk format strings known at compile time into a new .printk_index section, both in vmlinux and modules. At runtime, this can then be iterated by looking at <debugfs>/printk/index/<module>, which emits the following format, both readable by humans and able to be parsed by machines:
$ head -1 vmlinux; shuf -n 5 vmlinux # <level[,flags]> filename:line function "format" <5> block/blk-settings.c:661 disk_stack_limits "%s: Warning: Device %s is misaligned\n" <4> kernel/trace/trace.c:8296 trace_create_file "Could not create tracefs '%s' entry\n" <6> arch/x86/kernel/hpet.c:144 _hpet_print_config "hpet: %s(%d):\n" <6> init/do_mounts.c:605 prepare_namespace "Waiting for root device %s...\n" <6> drivers/acpi/osl.c:1410 acpi_no_auto_serialize_setup "ACPI: auto-serialization disabled\n"
This mitigates the majority of cases where we have a highly-specific printk which we want to match on, as we can now enumerate and check whether the format changed or the printk callsite disappeared entirely in userspace. This allows us to catch changes to printks we monitor earlier and decide what to do about it before it becomes problematic.
There is no additional runtime cost for printk callers or printk itself, and the assembly generated is exactly the same.
Signed-off-by: Chris Down <chris@chrisdown.name> Cc: Petr Mladek <pmladek@suse.com> Cc: Jessica Yu <jeyu@kernel.org> Cc: Sergey Senozhatsky <sergey.senozhatsky@gmail.com> Cc: John Ogness <john.ogness@linutronix.de> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Kees Cook <keescook@chromium.org> Reviewed-by: Petr Mladek <pmladek@suse.com> Tested-by: Petr Mladek <pmladek@suse.com> Reported-by: kernel test robot <lkp@intel.com> Acked-by: Andy Shevchenko <andy.shevchenko@gmail.com> Acked-by: Jessica Yu <jeyu@kernel.org> # for module.{c,h} Signed-off-by: Petr Mladek <pmladek@suse.com> Link: https://lore.kernel.org/r/e42070983637ac5e384f17fbdbe86d19c7b212a5.1623775748.git.chris@chrisdown.name
|
/linux/ |
H A D | MAINTAINERS | diff 337015573718b161891a3473d25f59273f2e626b Tue Jun 15 18:52:53 CEST 2021 Chris Down <chris@chrisdown.name> printk: Userspace format indexing support
We have a number of systems industry-wide that have a subset of their functionality that works as follows:
1. Receive a message from local kmsg, serial console, or netconsole; 2. Apply a set of rules to classify the message; 3. Do something based on this classification (like scheduling a remediation for the machine), rinse, and repeat.
As a couple of examples of places we have this implemented just inside Facebook, although this isn't a Facebook-specific problem, we have this inside our netconsole processing (for alarm classification), and as part of our machine health checking. We use these messages to determine fairly important metrics around production health, and it's important that we get them right.
While for some kinds of issues we have counters, tracepoints, or metrics with a stable interface which can reliably indicate the issue, in order to react to production issues quickly we need to work with the interface which most kernel developers naturally use when developing: printk.
Most production issues come from unexpected phenomena, and as such usually the code in question doesn't have easily usable tracepoints or other counters available for the specific problem being mitigated. We have a number of lines of monitoring defence against problems in production (host metrics, process metrics, service metrics, etc), and where it's not feasible to reliably monitor at another level, this kind of pragmatic netconsole monitoring is essential.
As one would expect, monitoring using printk is rather brittle for a number of reasons -- most notably that the message might disappear entirely in a new version of the kernel, or that the message may change in some way that the regex or other classification methods start to silently fail.
One factor that makes this even harder is that, under normal operation, many of these messages are never expected to be hit. For example, there may be a rare hardware bug which one wants to detect if it was to ever happen again, but its recurrence is not likely or anticipated. This precludes using something like checking whether the printk in question was printed somewhere fleetwide recently to determine whether the message in question is still present or not, since we don't anticipate that it should be printed anywhere, but still need to monitor for its future presence in the long-term.
This class of issue has happened on a number of occasions, causing unhealthy machines with hardware issues to remain in production for longer than ideal. As a recent example, some monitoring around blk_update_request fell out of date and caused semi-broken machines to remain in production for longer than would be desirable.
Searching through the codebase to find the message is also extremely fragile, because many of the messages are further constructed beyond their callsite (eg. btrfs_printk and other module-specific wrappers, each with their own functionality). Even if they aren't, guessing the format and formulation of the underlying message based on the aesthetics of the message emitted is not a recipe for success at scale, and our previous issues with fleetwide machine health checking demonstrate as much.
This provides a solution to the issue of silently changed or deleted printks: we record pointers to all printk format strings known at compile time into a new .printk_index section, both in vmlinux and modules. At runtime, this can then be iterated by looking at <debugfs>/printk/index/<module>, which emits the following format, both readable by humans and able to be parsed by machines:
$ head -1 vmlinux; shuf -n 5 vmlinux # <level[,flags]> filename:line function "format" <5> block/blk-settings.c:661 disk_stack_limits "%s: Warning: Device %s is misaligned\n" <4> kernel/trace/trace.c:8296 trace_create_file "Could not create tracefs '%s' entry\n" <6> arch/x86/kernel/hpet.c:144 _hpet_print_config "hpet: %s(%d):\n" <6> init/do_mounts.c:605 prepare_namespace "Waiting for root device %s...\n" <6> drivers/acpi/osl.c:1410 acpi_no_auto_serialize_setup "ACPI: auto-serialization disabled\n"
This mitigates the majority of cases where we have a highly-specific printk which we want to match on, as we can now enumerate and check whether the format changed or the printk callsite disappeared entirely in userspace. This allows us to catch changes to printks we monitor earlier and decide what to do about it before it becomes problematic.
There is no additional runtime cost for printk callers or printk itself, and the assembly generated is exactly the same.
Signed-off-by: Chris Down <chris@chrisdown.name> Cc: Petr Mladek <pmladek@suse.com> Cc: Jessica Yu <jeyu@kernel.org> Cc: Sergey Senozhatsky <sergey.senozhatsky@gmail.com> Cc: John Ogness <john.ogness@linutronix.de> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Kees Cook <keescook@chromium.org> Reviewed-by: Petr Mladek <pmladek@suse.com> Tested-by: Petr Mladek <pmladek@suse.com> Reported-by: kernel test robot <lkp@intel.com> Acked-by: Andy Shevchenko <andy.shevchenko@gmail.com> Acked-by: Jessica Yu <jeyu@kernel.org> # for module.{c,h} Signed-off-by: Petr Mladek <pmladek@suse.com> Link: https://lore.kernel.org/r/e42070983637ac5e384f17fbdbe86d19c7b212a5.1623775748.git.chris@chrisdown.name
|