perf: answer chown and bound thermal paths through the proot fast path (#2203)

This commit is contained in:
Kurt Himebauch authored and GitHub committed 2026-11-07 01:08:06 -05:00
1 parent 18c7d56409
commit aba55031c3
7 files changed
+267 -3

No files matched your search

@@ -377,3 +377,77 @@ diagnostics remained enabled. Temporary probe files, executables and obsolete
partial staging were removed after validation. Original effective mount options
were restored; no app-specific Magisk root policy remained. The primary checkout's
unrelated `SessionService.kt` modification was preserved.
## Install phases on current main (2026-11-07)
A fresh install of 1000xRESIST (app 1675830, 2,216 files, 7.45 GB) to the same Thor SD
card, on the build from main at the time (`0652b0d`), with the fast path and storage
shim from the earlier work. Phase times come from Steam's `content_log.txt`.
| Run | Reserving space | Download and stage | Commit | Total |
|---|---|---|---|---|
| Clean | 78 s | 25 min 01 s | 205 s | 29 min 46 s |
| Profiled | 79 s | 18 min 43 s | 146 s | 22 min 28 s |
Reserving space was 404 s on 2026-10-04 and is now about 36 ms per file, which is the
file-creation floor measured above (25-36 ms). The two runs differ by 7 minutes because
the download phase follows the network: Steam logged 12 to 83 Mbps for the whole phase.
Raw sequential write to the card with `dd conv=fsync` was 15 MB/s (120 Mbps), so the
card is not the limit during most of the download. The run-to-run spread is larger than
any change a path-translation fix could make, so install totals are not used to judge
the fast path changes below.
Kernel stacks of Steam's blocked threads, sampled every 0.2 s from `/proc/<pid>/task/*`:
- **Commit:** every sampled thread was inside `renameat`, waiting on synchronous
directory entry writes (`exfat_rename`, `exfat_init_ext_entry`, `exfat_remove_entries`,
`exfat_free_dentry_set`, all through `__sync_dirty_buffer`). That is the `dirsync`
mount option, at about 65 to 92 ms per file. PRoot is not involved in that time.
- **Download:** blocked threads were in `pwrite64`, in `exfat_get_block`,
`exfat_alloc_cluster` and `cont_write_begin`, which allocate clusters as the unreserved
files grow.
PRoot's own handler time during the reserving phase (profiler, 1 ms slow threshold) was
7.3 s of the 79 s, of which 6.3 s was `openat` handler visits. Almost all of those came
from the HUD reading the three bound thermal files (22,326 fast path misses logged
against `/sys/class/thermal/thermal_zone{22,26,28}/temp`), not from Steam. `fchownat`
accounted for 2,795 visits at about 70 us. Steam's own calls were answered by the fast
path.
### Fast path changes
- `fchownat`, `chown` and `lchown` are answered in the fast path, through the libc
wrappers and the patched `svc` sites. The site cache name moves to `svc-v3` so cached
site lists from older builds are rebuilt.
- A path that names a bound file exactly, spelled through a symlinked directory
(`/sys/class/thermal/thermal_zone28/temp`, where `thermal_zone28` is a symlink), now
resolves to the bind's host path. Before, any path through a symlink with a binding
beneath it fell back to PRoot. Paths with `..`, a trailing slash or any other ending
still fall back.
Same workloads on the Thor, stock library against the new one (session started with
`droiddeckctl override libblfastpath.so`, PRoot profiler on):
| Workload | Stock | New |
|---|---|---|
| `chown` of 2,000 files, internal storage | 40-50 us each | 6 us each |
| `chown` of 2,000 files, SD card | 61-62 us each | 9 us each |
| Open of the three bound thermal files, 6,000 opens | 128-133 us each | 24-25 us each |
| PRoot `openat` handler visits for both workloads | 15,496 (1,205 ms) | 7,458 (366 ms) |
| PRoot `fchownat` handler visits | 6,873 (288 ms) | 0 |
`bench/chown.py` compares 87 `chown` results (paths, links, dangling and looping links,
descriptors, directory descriptors, foreign owners, trailing slashes) with the fast path
on and with `PROOT_FP_OFF=1`; they were identical. The thermal paths were compared the
same way for every spelling the HUD can use (absolute, doubled slashes, `.` and `..`
components, relative to a working directory or descriptor on the parent directory) and
matched. They differ only when the working directory or descriptor is inside the
symlinked directory, or the path uses its canonical `/sys/devices/virtual/...` name: PRoot
returns the bound file there and the fast path returns the real node. That difference
existed before these changes, and the HUD does not read those spellings.
What this does not change: the commit and the file creation costs above are the card's
synchronous directory updates, so SD installs stay bound by them (about 4 to 5 minutes per
2,000 files of reserving plus commit on this card). Removing those costs needs fewer
directory operations per file, not faster path translation.