Skip to content

tools/runqlat: Dynamically size pid/pidns histogram map#5342

Merged
yonghong-song merged 1 commit intoiovisor:masterfrom
ismhong:master
Jun 24, 2025
Merged

tools/runqlat: Dynamically size pid/pidns histogram map#5342
yonghong-song merged 1 commit intoiovisor:masterfrom
ismhong:master

Conversation

@ismhong
Copy link
Contributor

@ismhong ismhong commented Jun 23, 2025

When using runqlat with the --pids or --pidnss flags, the BPF histogram for tracking per-process or per-namespace latency is created with a fixed default size of 64 entries.

On systems with a large number of processes or active namespaces, this small, fixed size can be easily exhausted. This leads to silent data loss for any PIDs or PID namespaces beyond the 64th entry, making the tool's output incomplete and misleading.

To fix this, read the maximum possible PID and PID Namespace values from /proc/sys/kernel/pid_max and /proc/sys/user/max_user_namespaces respectively. Pass these values to the BPF program to dynamically size the histogram map, ensuring it is large enough to avoid dropping data.

This approach is similar to the one implemented for cpudist as discussed in PR #2568.

Link: #2568

tools/runqlat.py Outdated
exit()

max_pid = int(open("/proc/sys/kernel/pid_max").read())
max_pidns = int(open("/proc/sys/user/max_user_namespaces").read())
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Should the above be /proc/sys/user/max_pid_namespaces? Did I miss anything?

Copy link
Contributor Author

@ismhong ismhong Jun 24, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Hi @yonghong-song
Thanks for review. You are right, sorry for my carelessness.
https://docs.kernel.org/admin-guide/sysctl/user.html
I have fixed it, please help to review it, thanks.

When using `runqlat` with the `--pids` or `--pidnss` flags, the BPF
histogram for tracking per-process or per-namespace latency is created
with a fixed default size of 64 entries.

On systems with a large number of processes or active namespaces, this
small, fixed size can be easily exhausted. This leads to silent data
loss for any PIDs or PID namespaces beyond the 64th entry, making the
tool's output incomplete and misleading.

To fix this, read the maximum possible PID and PID Namespace values from
`/proc/sys/kernel/pid_max` and `/proc/sys/user/max_user_namespaces`
respectively. Pass these values to the BPF program to dynamically size
the histogram map, ensuring it is large enough to avoid dropping data.

This approach is similar to the one implemented for cpudist as discussed
in PR iovisor#2568.

Link: iovisor#2568
Signed-off-by: Ism Hong <ism.hong@gmail.com>
@yonghong-song yonghong-song merged commit 939828c into iovisor:master Jun 24, 2025
1 of 12 checks passed
ekyooo added a commit to ekyooo/bcc that referenced this pull request Jan 23, 2026
  * Support for kernel up to 6.18

  * New Tools
    tools/softirqslower: New tool to trace slow software interrupt handlers (iovisor#5356)

  * Enhanced Functionality
    libbpf-tools/opensnoop: Added full-path support with `-F` option (iovisor#5323, iovisor#5333)
    libbpf-tools/filelife: Added full-path support (iovisor#5347, ab8e061)
    libbpf-tools: Introduced path helpers (ab8e061)
    libbpf-tools/trace_helpers: Added str_loadavg() and str_timestamp() common functions (694de9f)
    libbpf-tools/filetop: Added directory filter capability (iovisor#5300)
    libbpf-tools/runqslower: Added `-c` option to filter by process name prefix (673911c)
    libbpf-tools/runqlat: Dynamically size pid/pidns histogram map (iovisor#5342)
    libbpf-tools/fsdist, fsslower: Added support for fuse filesystem (9691c56)
    libbpf-tools/tcptop: Major refactoring using fentry/fexit for better performance (75bb73a, e2c7917, d786eaa, da3a474)
    tools/opensnoop: Added full-path support with `-F` option (iovisor#5334, iovisor#5339)
    tools/kvmexit: Added AMD processor support and parallel post-processing (13a4e5a, c2af2ee)
    tools/offwaketime: Added raw tracepoint support to reduce overhead (380ee01)
    Python uprobe API: Added functionality to detach all uprobes for a binary (iovisor#5325)
    Python API: Added support for executing a program and tracing it (iovisor#5362)

  * Bug Fixes
    libbpf-tools/filelife: Fixed wrong full-path handling (iovisor#5347)
    libbpf-tools/filelife: Fixed problem when using perf-buffer (ec8415b)
    libbpf-tools/funclatency: Delete the element from the `starts` map after it has been used (06ce134)
    libbpf-tools/offcputime: Fixed min/max_block_ns unit conversion error (iovisor#5327, d507a53)
    libbpf-tools/syncsnoop: Added support for sync_file_range2 and arm_sync_file_range() (4287921)
    libbpf-tools/ksnoop: Fixed two invalid access to map value (iovisor#5361)
    libbpf-tools/klockstat: Allows kprobe fallback to work with lock debugging (iovisor#5359)
    libbpf-tools/biotop: Fixed segmentation fault with musl libc build (52d2d09)
    libbpf-tools/syscall_helpers, Python BCC: Updated syscall list (add file_getattr/file_setattr) (b63d7e3, a9c6650)
    tools/tcpaccept: Fixed on recent kernels (c208d0e)
    tools/tcpconnect: Fixed iov field for DNS with Linux>=6.4 (iovisor#5382)
    tools/javaobjnew: Use MIN macro instead of min function (fb8910a)
    tools/biolatency, biosnoop, biotop: Use TRACEPOINT_PROBE() for tracepoints (iovisor#5366)
    Various tools: Don't use the old bpf_probe_read() helper (1cc15c3)
    CC: Support versioned SONAME in shared library resolution (beb1fe4, c351210)
    Python TCP: Added state2str() and applied to tools (bfa05d2)
    s390 architecture: Prevent invalid mem access when reading PAGE_OFFSET (d8595ee)

  * Build & Test Fixes
    Fixed build failure with clang21 (iovisor#5369)
    Fixed build for LLVM 23 by avoiding deprecated TargetRegistry overloads (iovisor#5401)
    ci: Make version.cmake handle shallow clone (2232b7e)
    ci: Various test fixes for proper CI operation (blk probes, rss_stat, kmalloc, btrfs/f2fs) (a499181, c338547, 6b7dd5d, ea5cf83)
    tests: Added coverage for versioned SONAME resolution (c351210)
    Removed luajit options to ensure no errors (26eaf13)

  * Doc update, other bug fixes and tools improvement
ekyooo added a commit that referenced this pull request Jan 26, 2026
  * Support for kernel up to 6.18

  * New Tools
    tools/softirqslower: New tool to trace slow software interrupt handlers (#5356)

  * Enhanced Functionality
    libbpf-tools/opensnoop: Added full-path support with `-F` option (#5323, #5333)
    libbpf-tools/filelife: Added full-path support (#5347, ab8e061)
    libbpf-tools: Introduced path helpers (ab8e061)
    libbpf-tools/trace_helpers: Added str_loadavg() and str_timestamp() common functions (694de9f)
    libbpf-tools/filetop: Added directory filter capability (#5300)
    libbpf-tools/runqslower: Added `-c` option to filter by process name prefix (673911c)
    libbpf-tools/runqlat: Dynamically size pid/pidns histogram map (#5342)
    libbpf-tools/fsdist, fsslower: Added support for fuse filesystem (9691c56)
    libbpf-tools/tcptop: Major refactoring using fentry/fexit for better performance (75bb73a, e2c7917, d786eaa, da3a474)
    tools/opensnoop: Added full-path support with `-F` option (#5334, #5339)
    tools/kvmexit: Added AMD processor support and parallel post-processing (13a4e5a, c2af2ee)
    tools/offwaketime: Added raw tracepoint support to reduce overhead (380ee01)
    Python uprobe API: Added functionality to detach all uprobes for a binary (#5325)
    Python API: Added support for executing a program and tracing it (#5362)

  * Bug Fixes
    libbpf-tools/filelife: Fixed wrong full-path handling (#5347)
    libbpf-tools/filelife: Fixed problem when using perf-buffer (ec8415b)
    libbpf-tools/funclatency: Delete the element from the `starts` map after it has been used (06ce134)
    libbpf-tools/offcputime: Fixed min/max_block_ns unit conversion error (#5327, d507a53)
    libbpf-tools/syncsnoop: Added support for sync_file_range2 and arm_sync_file_range() (4287921)
    libbpf-tools/ksnoop: Fixed two invalid access to map value (#5361)
    libbpf-tools/klockstat: Allows kprobe fallback to work with lock debugging (#5359)
    libbpf-tools/biotop: Fixed segmentation fault with musl libc build (52d2d09)
    libbpf-tools/syscall_helpers, Python BCC: Updated syscall list (add file_getattr/file_setattr) (b63d7e3, a9c6650)
    tools/tcpaccept: Fixed on recent kernels (c208d0e)
    tools/tcpconnect: Fixed iov field for DNS with Linux>=6.4 (#5382)
    tools/javaobjnew: Use MIN macro instead of min function (fb8910a)
    tools/biolatency, biosnoop, biotop: Use TRACEPOINT_PROBE() for tracepoints (#5366)
    Various tools: Don't use the old bpf_probe_read() helper (1cc15c3)
    CC: Support versioned SONAME in shared library resolution (beb1fe4, c351210)
    Python TCP: Added state2str() and applied to tools (bfa05d2)
    s390 architecture: Prevent invalid mem access when reading PAGE_OFFSET (d8595ee)

  * Build & Test Fixes
    Fixed build failure with clang21 (#5369)
    Fixed build for LLVM 23 by avoiding deprecated TargetRegistry overloads (#5401)
    ci: Make version.cmake handle shallow clone (2232b7e)
    ci: Various test fixes for proper CI operation (blk probes, rss_stat, kmalloc, btrfs/f2fs) (a499181, c338547, 6b7dd5d, ea5cf83)
    tests: Added coverage for versioned SONAME resolution (c351210)
    Removed luajit options to ensure no errors (26eaf13)

  * Doc update, other bug fixes and tools improvement
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants