uprobes: turn trace_uprobe's nhit counter to be per-CPU one
Andrii Nakryiko
maintainerPATCH v3 1/1 uprobes: turn trace_uprobe's nhit counter to be per-CPU one
trace_uprobe->nhit counter is not incremented atomically, so its value is questionable in when uprobe is hit on multiple CPUs simultaneously.
Also, doing this shared counter increment across many CPUs causes heavy cache line bouncing, limiting uprobe/uretprobe performance scaling with number of CPUs.
Solve both problems by making this a per-CPU counter.
--- kernel/trace/trace_uprobe.c | 24 +++++++++++++++++++++--- 1 file changed, 21 insertions(+), 3 deletions(-)
kernel/trace/trace_uprobe.c +21 -3
@@ -17,6 +17,7 @@#include <linux/string.h> #include <linux/rculist.h> #include <linux/filter.h> +#include <linux/percpu.h> #include "trace_dynevent.h" #include "trace_probe.h"
@@ -62,7 +63,7 @@ struct trace_uprobe {char *filename;unsigned long offset;unsigned long ref_ctr_offset; -unsigned long nhit; +unsigned long __percpu *nhits;struct trace_probe tp; };
@@ -337,6 +338,12 @@ alloc_trace_uprobe(const char *group, const char *event, int nargs, bool is_ret)if (!tu)return ERR_PTR (-ENOMEM ); + tu->nhits = alloc_percpu(unsigned long); +if (!tu->nhits) { + ret = -ENOMEM ; + goto error; + } + ret =trace_probe_init (&tu->tp, event, group, true, nargs);if (ret <0 ) goto error;
@@ -349,6 +356,7 @@ alloc_trace_uprobe(const char *group, const char *event, int nargs, bool is_ret)return tu; error: +free_percpu (tu->nhits);kfree (tu);return ERR_PTR (ret);
@@ -362,6 +370,7 @@ static void free_trace_uprobe(struct trace_uprobe *tu)path_put (&tu->path);trace_probe_cleanup (&tu->tp);kfree (tu->filename); +free_percpu (tu->nhits);kfree (tu); }
@@ -815,13 +824,21 @@ static int probes_profile_seq_show(struct seq_file *m, void *v) {struct dyn_event *ev = v;struct trace_uprobe *tu; +unsigned long nhits; +int cpu;if (!is_trace_uprobe (ev))return 0 ; tu =to_trace_uprobe (ev); + + nhits =0 ; +for_each_possible_cpu (cpu) { + nhits +=per_cpu (*tu->nhits, cpu); + } +seq_printf (m," %s %-44s %15lu\n" , tu->filename, -trace_probe_name (&tu->tp), tu->nhit); +trace_probe_name (&tu->tp), nhits);return 0 ; }
@@ -1512,7 +1529,8 @@ static int uprobe_dispatcher(struct uprobe_consumer *con, struct pt_regs *regs)int ret =0 ; tu = container_of(con, struct trace_uprobe, consumer); - tu->nhit++; + +this_cpu_inc (*tu->nhits); udd.tu = tu; udd.bp_addr =instruction_pointer (regs);
Signature · 1 line
2.43.5
Quoted 8 lines · Andrii Nakryiko
trace_uprobe->nhitcounter is not incremented atomically, so its value is questionable in when uprobe is hit on multiple CPUs simultaneously.Also, doing this shared counter increment across many CPUs causes heavy cache line bouncing, limiting uprobe/uretprobe performance scaling with number of CPUs.
Quoted 8 lines
Solve both problems by making this a per-CPU counter.
Reviewed-by Oleg Nesterov <[email protected]>Signed-off-by Andrii Nakryiko <[email protected]>--- kernel/trace/trace_uprobe.c | 24 +++++++++++++++++++++--- 1 file changed, 21 insertions(+), 3 deletions(-)
Is there anything else I'm expected to do about this patch? If not, can this please be applied? Thanks!
Quoted 86 lines
kernel/trace/trace_uprobe.c +21
-4@@ -17,6 +17,7 @@#include <linux/string.h> #include <linux/rculist.h> #include <linux/filter.h> +#include <linux/percpu.h> #include "trace_dynevent.h" #include "trace_probe.h" @@ -62,7 +63,7 @@ struct trace_uprobe {char *filename;unsigned long offset;unsigned long ref_ctr_offset; -unsigned long nhit; +unsigned long __percpu *nhits;struct trace_probe tp; };@@ -337,6 +338,12 @@ alloc_trace_uprobe(const char *group, const char *event, int nargs, bool is_ret)if (!tu)return ERR_PTR (-ENOMEM ); + tu->nhits = alloc_percpu(unsigned long); +if (!tu->nhits) { + ret = -ENOMEM ; + goto error; + } + ret =trace_probe_init (&tu->tp, event, group, true, nargs);if (ret <0 ) goto error;@@ -349,6 +356,7 @@ alloc_trace_uprobe(const char *group, const char *event, int nargs, bool is_ret)return tu; error: +free_percpu (tu->nhits);kfree (tu);return ERR_PTR (ret);@@ -362,6 +370,7 @@ static void free_trace_uprobe(struct trace_uprobe *tu)path_put (&tu->path);trace_probe_cleanup (&tu->tp);kfree (tu->filename); +free_percpu (tu->nhits);kfree (tu); }@@ -815,13 +824,21 @@ static int probes_profile_seq_show(struct seq_file *m, void *v) {struct dyn_event *ev = v;struct trace_uprobe *tu; +unsigned long nhits; +int cpu;if (!is_trace_uprobe (ev))return 0 ; tu =to_trace_uprobe (ev); + + nhits =0 ; +for_each_possible_cpu (cpu) { + nhits +=per_cpu (*tu->nhits, cpu); + } +seq_printf (m," %s %-44s %15lu\n" , tu->filename, -trace_probe_name (&tu->tp), tu->nhit); +trace_probe_name (&tu->tp), nhits);return 0 ; }@@ -1512,7 +1529,8 @@ static int uprobe_dispatcher(struct uprobe_consumer *con, struct pt_regs *regs)int ret =0 ; tu = container_of(con, struct trace_uprobe, consumer); - tu->nhit++; + +this_cpu_inc (*tu->nhits); udd.tu = tu; udd.bp_addr =instruction_pointer (regs); --2.43.5
Quoted 11 lines · Andrii Nakryiko
trace_uprobe->nhitcounter is not incremented atomically, so its value is questionable in when uprobe is hit on multiple CPUs simultaneously.Also, doing this shared counter increment across many CPUs causes heavy cache line bouncing, limiting uprobe/uretprobe performance scaling with number of CPUs.
Solve both problems by making this a per-CPU counter.
Reviewed-by Oleg Nesterov <[email protected]>Signed-off-by Andrii Nakryiko <[email protected]>
lgtm, fwiw
jirka
Quoted 90 lines
--- kernel/trace/trace_uprobe.c | 24 +++++++++++++++++++++--- 1 file changed, 21 insertions(+), 3 deletions(-)kernel/trace/trace_uprobe.c +21
-3@@ -17,6 +17,7 @@#include <linux/string.h> #include <linux/rculist.h> #include <linux/filter.h> +#include <linux/percpu.h> #include "trace_dynevent.h" #include "trace_probe.h" @@ -62,7 +63,7 @@ struct trace_uprobe {char *filename;unsigned long offset;unsigned long ref_ctr_offset; -unsigned long nhit; +unsigned long __percpu *nhits;struct trace_probe tp; };@@ -337,6 +338,12 @@ alloc_trace_uprobe(const char *group, const char *event, int nargs, bool is_ret)if (!tu)return ERR_PTR (-ENOMEM ); + tu->nhits = alloc_percpu(unsigned long); +if (!tu->nhits) { + ret = -ENOMEM ; + goto error; + } + ret =trace_probe_init (&tu->tp, event, group, true, nargs);if (ret <0 ) goto error;@@ -349,6 +356,7 @@ alloc_trace_uprobe(const char *group, const char *event, int nargs, bool is_ret)return tu; error: +free_percpu (tu->nhits);kfree (tu);return ERR_PTR (ret);@@ -362,6 +370,7 @@ static void free_trace_uprobe(struct trace_uprobe *tu)path_put (&tu->path);trace_probe_cleanup (&tu->tp);kfree (tu->filename); +free_percpu (tu->nhits);kfree (tu); }@@ -815,13 +824,21 @@ static int probes_profile_seq_show(struct seq_file *m, void *v) {struct dyn_event *ev = v;struct trace_uprobe *tu; +unsigned long nhits; +int cpu;if (!is_trace_uprobe (ev))return 0 ; tu =to_trace_uprobe (ev); + + nhits =0 ; +for_each_possible_cpu (cpu) { + nhits +=per_cpu (*tu->nhits, cpu); + } +seq_printf (m," %s %-44s %15lu\n" , tu->filename, -trace_probe_name (&tu->tp), tu->nhit); +trace_probe_name (&tu->tp), nhits);return 0 ; }@@ -1512,7 +1529,8 @@ static int uprobe_dispatcher(struct uprobe_consumer *con, struct pt_regs *regs)int ret =0 ; tu = container_of(con, struct trace_uprobe, consumer); - tu->nhit++; + +this_cpu_inc (*tu->nhits); udd.tu = tu; udd.bp_addr =instruction_pointer (regs);Signature · 1 line
2.43.5
Masami Hiramatsu (Google)
maintainerQuoted 9 lines · On Tue, 13 Aug 2024 13:34:09 -0700 Andrii Nakryiko <[email protected]> wrote:
trace_uprobe->nhitcounter is not incremented atomically, so its value is questionable in when uprobe is hit on multiple CPUs simultaneously.Also, doing this shared counter increment across many CPUs causes heavy cache line bouncing, limiting uprobe/uretprobe performance scaling with number of CPUs.
Solve both problems by making this a per-CPU counter.
Looks good to me. Let me pick it.
Quoted 92 lines
Reviewed-by Oleg Nesterov <[email protected]>Signed-off-by Andrii Nakryiko <[email protected]>--- kernel/trace/trace_uprobe.c | 24 +++++++++++++++++++++--- 1 file changed, 21 insertions(+), 3 deletions(-)kernel/trace/trace_uprobe.c +21
-3@@ -17,6 +17,7 @@#include <linux/string.h> #include <linux/rculist.h> #include <linux/filter.h> +#include <linux/percpu.h> #include "trace_dynevent.h" #include "trace_probe.h" @@ -62,7 +63,7 @@ struct trace_uprobe {char *filename;unsigned long offset;unsigned long ref_ctr_offset; -unsigned long nhit; +unsigned long __percpu *nhits;struct trace_probe tp; };@@ -337,6 +338,12 @@ alloc_trace_uprobe(const char *group, const char *event, int nargs, bool is_ret)if (!tu)return ERR_PTR (-ENOMEM ); + tu->nhits = alloc_percpu(unsigned long); +if (!tu->nhits) { + ret = -ENOMEM ; + goto error; + } + ret =trace_probe_init (&tu->tp, event, group, true, nargs);if (ret <0 ) goto error;@@ -349,6 +356,7 @@ alloc_trace_uprobe(const char *group, const char *event, int nargs, bool is_ret)return tu; error: +free_percpu (tu->nhits);kfree (tu);return ERR_PTR (ret);@@ -362,6 +370,7 @@ static void free_trace_uprobe(struct trace_uprobe *tu)path_put (&tu->path);trace_probe_cleanup (&tu->tp);kfree (tu->filename); +free_percpu (tu->nhits);kfree (tu); }@@ -815,13 +824,21 @@ static int probes_profile_seq_show(struct seq_file *m, void *v) {struct dyn_event *ev = v;struct trace_uprobe *tu; +unsigned long nhits; +int cpu;if (!is_trace_uprobe (ev))return 0 ; tu =to_trace_uprobe (ev); + + nhits =0 ; +for_each_possible_cpu (cpu) { + nhits +=per_cpu (*tu->nhits, cpu); + } +seq_printf (m," %s %-44s %15lu\n" , tu->filename, -trace_probe_name (&tu->tp), tu->nhit); +trace_probe_name (&tu->tp), nhits);return 0 ; }@@ -1512,7 +1529,8 @@ static int uprobe_dispatcher(struct uprobe_consumer *con, struct pt_regs *regs)int ret =0 ; tu = container_of(con, struct trace_uprobe, consumer); - tu->nhit++; + +this_cpu_inc (*tu->nhits); udd.tu = tu; udd.bp_addr =instruction_pointer (regs);Signature · 1 line
2.43.5
Signature · 1 line
Masami Hiramatsu (Google) <[email protected]>
Masami Hiramatsu (Google)
maintainerQuoted 9 lines · On Tue, 13 Aug 2024 13:34:09 -0700 Andrii Nakryiko <[email protected]> wrote:
trace_uprobe->nhitcounter is not incremented atomically, so its value is questionable in when uprobe is hit on multiple CPUs simultaneously.Also, doing this shared counter increment across many CPUs causes heavy cache line bouncing, limiting uprobe/uretprobe performance scaling with number of CPUs.
Solve both problems by making this a per-CPU counter.
Looks good to me. Let me pick it to linux-trace probes/for-next.
Thank you,
Quoted 92 lines
Reviewed-by Oleg Nesterov <[email protected]>Signed-off-by Andrii Nakryiko <[email protected]>--- kernel/trace/trace_uprobe.c | 24 +++++++++++++++++++++--- 1 file changed, 21 insertions(+), 3 deletions(-)kernel/trace/trace_uprobe.c +21
-3@@ -17,6 +17,7 @@#include <linux/string.h> #include <linux/rculist.h> #include <linux/filter.h> +#include <linux/percpu.h> #include "trace_dynevent.h" #include "trace_probe.h" @@ -62,7 +63,7 @@ struct trace_uprobe {char *filename;unsigned long offset;unsigned long ref_ctr_offset; -unsigned long nhit; +unsigned long __percpu *nhits;struct trace_probe tp; };@@ -337,6 +338,12 @@ alloc_trace_uprobe(const char *group, const char *event, int nargs, bool is_ret)if (!tu)return ERR_PTR (-ENOMEM ); + tu->nhits = alloc_percpu(unsigned long); +if (!tu->nhits) { + ret = -ENOMEM ; + goto error; + } + ret =trace_probe_init (&tu->tp, event, group, true, nargs);if (ret <0 ) goto error;@@ -349,6 +356,7 @@ alloc_trace_uprobe(const char *group, const char *event, int nargs, bool is_ret)return tu; error: +free_percpu (tu->nhits);kfree (tu);return ERR_PTR (ret);@@ -362,6 +370,7 @@ static void free_trace_uprobe(struct trace_uprobe *tu)path_put (&tu->path);trace_probe_cleanup (&tu->tp);kfree (tu->filename); +free_percpu (tu->nhits);kfree (tu); }@@ -815,13 +824,21 @@ static int probes_profile_seq_show(struct seq_file *m, void *v) {struct dyn_event *ev = v;struct trace_uprobe *tu; +unsigned long nhits; +int cpu;if (!is_trace_uprobe (ev))return 0 ; tu =to_trace_uprobe (ev); + + nhits =0 ; +for_each_possible_cpu (cpu) { + nhits +=per_cpu (*tu->nhits, cpu); + } +seq_printf (m," %s %-44s %15lu\n" , tu->filename, -trace_probe_name (&tu->tp), tu->nhit); +trace_probe_name (&tu->tp), nhits);return 0 ; }@@ -1512,7 +1529,8 @@ static int uprobe_dispatcher(struct uprobe_consumer *con, struct pt_regs *regs)int ret =0 ; tu = container_of(con, struct trace_uprobe, consumer); - tu->nhit++; + +this_cpu_inc (*tu->nhits); udd.tu = tu; udd.bp_addr =instruction_pointer (regs);Signature · 1 line
2.43.5
Signature · 1 line
Masami Hiramatsu (Google) <[email protected]>
Quoted 7 lines · Masami Hiramatsu
Quoted 3 lines · On Tue, 13 Aug 2024 13:34:09 -0700 Andrii Nakryiko <[email protected]> wrote:
trace_uprobe->nhitcounter is not incremented atomically, so its value is questionable in when uprobe is hit on multiple CPUs simultaneously.
Quoted 9 lines
Quoted 6 lines
Also, doing this shared counter increment across many CPUs causes heavy cache line bouncing, limiting uprobe/uretprobe performance scaling with number of CPUs.
Solve both problems by making this a per-CPU counter.
Looks good to me. Let me pick it to linux-trace probes/for-next.
Thanks! I just checked linux-trace repo, doesn't seem like this was applied yet, is that right? Or am I checking in the wrong place?
Quoted 100 lines
Thank you,
Quoted 90 lines
Reviewed-by Oleg Nesterov <[email protected]>Signed-off-by Andrii Nakryiko <[email protected]>--- kernel/trace/trace_uprobe.c | 24 +++++++++++++++++++++--- 1 file changed, 21 insertions(+), 3 deletions(-)kernel/trace/trace_uprobe.c +21
-4@@ -17,6 +17,7 @@#include <linux/string.h> #include <linux/rculist.h> #include <linux/filter.h> +#include <linux/percpu.h> #include "trace_dynevent.h" #include "trace_probe.h" @@ -62,7 +63,7 @@ struct trace_uprobe {char *filename;unsigned long offset;unsigned long ref_ctr_offset; -unsigned long nhit; +unsigned long __percpu *nhits;struct trace_probe tp; };@@ -337,6 +338,12 @@ alloc_trace_uprobe(const char *group, const char *event, int nargs, bool is_ret)if (!tu)return ERR_PTR (-ENOMEM ); + tu->nhits = alloc_percpu(unsigned long); +if (!tu->nhits) { + ret = -ENOMEM ; + goto error; + } + ret =trace_probe_init (&tu->tp, event, group, true, nargs);if (ret <0 ) goto error;@@ -349,6 +356,7 @@ alloc_trace_uprobe(const char *group, const char *event, int nargs, bool is_ret)return tu; error: +free_percpu (tu->nhits);kfree (tu);return ERR_PTR (ret);@@ -362,6 +370,7 @@ static void free_trace_uprobe(struct trace_uprobe *tu)path_put (&tu->path);trace_probe_cleanup (&tu->tp);kfree (tu->filename); +free_percpu (tu->nhits);kfree (tu); }@@ -815,13 +824,21 @@ static int probes_profile_seq_show(struct seq_file *m, void *v) {struct dyn_event *ev = v;struct trace_uprobe *tu; +unsigned long nhits; +int cpu;if (!is_trace_uprobe (ev))return 0 ; tu =to_trace_uprobe (ev); + + nhits =0 ; +for_each_possible_cpu (cpu) { + nhits +=per_cpu (*tu->nhits, cpu); + } +seq_printf (m," %s %-44s %15lu\n" , tu->filename, -trace_probe_name (&tu->tp), tu->nhit); +trace_probe_name (&tu->tp), nhits);return 0 ; }@@ -1512,7 +1529,8 @@ static int uprobe_dispatcher(struct uprobe_consumer *con, struct pt_regs *regs)int ret =0 ; tu = container_of(con, struct trace_uprobe, consumer); - tu->nhit++; + +this_cpu_inc (*tu->nhits); udd.tu = tu; udd.bp_addr =instruction_pointer (regs); --Quoted 2 lines
2.43.5
--
Masami Hiramatsu (Google) <[email protected]>
Masami Hiramatsu (Google)
maintainerQuoted 12 lines · On Thu, 29 Aug 2024 10:28:24 -0700 Andrii Nakryiko <[email protected]>…
Quoted 11 lines · Masami Hiramatsu
Quoted 7 lines · On Tue, 13 Aug 2024 13:34:09 -0700 Andrii Nakryiko <[email protected]> wrote:
trace_uprobe->nhitcounter is not incremented atomically, so its value is questionable in when uprobe is hit on multiple CPUs simultaneously.Also, doing this shared counter increment across many CPUs causes heavy cache line bouncing, limiting uprobe/uretprobe performance scaling with number of CPUs.
Quoted 8 lines
Quoted 5 lines
Quoted 2 lines
Solve both problems by making this a per-CPU counter.
Looks good to me. Let me pick it to linux-trace probes/for-next.
Thanks! I just checked linux-trace repo, doesn't seem like this was applied yet, is that right? Or am I checking in the wrong place?
Sorry, I missed to push probes/for-next. Let me push it.
Thank you,
Quoted 101 lines
Quoted 101 lines
Thank you,
Quoted 90 lines
Reviewed-by Oleg Nesterov <[email protected]>Signed-off-by Andrii Nakryiko <[email protected]>--- kernel/trace/trace_uprobe.c | 24 +++++++++++++++++++++--- 1 file changed, 21 insertions(+), 3 deletions(-)kernel/trace/trace_uprobe.c +21
-4@@ -17,6 +17,7 @@#include <linux/string.h> #include <linux/rculist.h> #include <linux/filter.h> +#include <linux/percpu.h> #include "trace_dynevent.h" #include "trace_probe.h" @@ -62,7 +63,7 @@ struct trace_uprobe {char *filename;unsigned long offset;unsigned long ref_ctr_offset; -unsigned long nhit; +unsigned long __percpu *nhits;struct trace_probe tp; };@@ -337,6 +338,12 @@ alloc_trace_uprobe(const char *group, const char *event, int nargs, bool is_ret)if (!tu)return ERR_PTR (-ENOMEM ); + tu->nhits = alloc_percpu(unsigned long); +if (!tu->nhits) { + ret = -ENOMEM ; + goto error; + } + ret =trace_probe_init (&tu->tp, event, group, true, nargs);if (ret <0 ) goto error;@@ -349,6 +356,7 @@ alloc_trace_uprobe(const char *group, const char *event, int nargs, bool is_ret)return tu; error: +free_percpu (tu->nhits);kfree (tu);return ERR_PTR (ret);@@ -362,6 +370,7 @@ static void free_trace_uprobe(struct trace_uprobe *tu)path_put (&tu->path);trace_probe_cleanup (&tu->tp);kfree (tu->filename); +free_percpu (tu->nhits);kfree (tu); }@@ -815,13 +824,21 @@ static int probes_profile_seq_show(struct seq_file *m, void *v) {struct dyn_event *ev = v;struct trace_uprobe *tu; +unsigned long nhits; +int cpu;if (!is_trace_uprobe (ev))return 0 ; tu =to_trace_uprobe (ev); + + nhits =0 ; +for_each_possible_cpu (cpu) { + nhits +=per_cpu (*tu->nhits, cpu); + } +seq_printf (m," %s %-44s %15lu\n" , tu->filename, -trace_probe_name (&tu->tp), tu->nhit); +trace_probe_name (&tu->tp), nhits);return 0 ; }@@ -1512,7 +1529,8 @@ static int uprobe_dispatcher(struct uprobe_consumer *con, struct pt_regs *regs)int ret =0 ; tu = container_of(con, struct trace_uprobe, consumer); - tu->nhit++; + +this_cpu_inc (*tu->nhits); udd.tu = tu; udd.bp_addr =instruction_pointer (regs); --Quoted 2 lines
2.43.5
--
Masami Hiramatsu (Google) <[email protected]>
Signature · 1 line
Masami Hiramatsu (Google) <[email protected]>