RFC: Keep a cache of frequently accessed files #517

nephros · 2025-10-29T11:01:53Z

RFC enhancement: Have the daemon keep track of files which are asked about by the preload lib, but are not patched.

This is the overwhelming majority of requests in normal runtime scenarios.

The idea is that calls to QFileInfo::exists() are minimized, and file names or unpatched files returned as soon as possible.

(Open question: Can we be smarter than the QFileInfo implementation and the kernel fs cacche? ;) )

The original idea was presented in
#431 (comment)
which suggests a Bloom filter to handle these file paths.

I experimented with bloom filters, but found the resulting logic too hard for
my small brain to handle.

So this implementation abuses Qt's QCache,
but uses a class which means the actual cache could be replaced with something better later.

@b100dian if you have the time, maybe look over the idea and review it.

Hotcache: Add a cache object to track nonexisting files
Hotcache: remove entry if it existed
Hotcache: simplify logic
Hotcache: store Object, retrieve one on lookup
Hotcache: remove log line
Hotcache: use object() not take()
Hotcache: move debugging messages after writing to the connection
Hotcache: set up the filter with some initial contents
Hotcache: Add a Settings switch
Hotcache: wire up settings switch
Hotcache: move into a class, rename to m_filter
Hotcache: Don't be a template class
Hotcache: Prepare for stats collecting
Hotcache: Static initializer lists
Hotcache: Implement a simpler stats method
Hotcache: Make the lists public
Hotcache: Use foreach again.
Hotcache: Missing braces
Hotcache: check for active filter in startReadingLocalServer
Hotcache: Use plain qstring for stats
Hotcache: Add some debug prints
Hotcache: Don't copy around strings, us a bool.
Print hotcache stats on PM exit
Resolve symlinks before adding to primed cache

take() gives ownership to the caller. object() does not.

- Try a Q_OBJECT macro for a template class... - Don't be a template class - Become Object-based, add contructor - Implement a constructor

nephros · 2025-10-29T12:50:25Z

One problem with evaluating this is getting a proper benchmark.

Suggestions are highly welcome of a test scenario which can somewhat reliably
compare "performance".

So the whole thing is currently a "should theoretically improve things" thing.

nephros · 2025-10-29T13:05:01Z

A comment about the logic.

Without this change, the communication goes something like this:

A process launches, preload library gets injected
Process tries to open a file /path/to/foo
Preload lib takes the file path string and writes it to the Socket
PM Daemon reads the file name from the Socket
PM Daemon checks whether the path exists in its tmp overlay, which means it is a patched file. This check uses QFileInfo::exists() which is potentially expensive.
If it exists, it writes the path of the patched file version back to the Socket
If not, it writes the original path name to the Socket
Preload lib reads the path string, and passes the result of open/open64 on the path received to the original Process,

With this change, the part of the Daemon becomes:

PM Daemon has a cache, which holds a list of unpatched files.
PM Daemon reads the file name from the Socket
PM Daemon checks whether the file name is contained in the cache (the cache notices this check/query).
If yes, write the original file name back.
If no, check that the file exists in the tmp overlay (same as above, but should happen more rarely).
If it exists, it writes the path of the patched file version back to the Socket
If it does not exist, register the path in the cache,

As the cache holds a list of unpatched files, it will continue to grow. we
rely on the QCache eviction logic to make sure it doesn't grow too large
(that's built-in).

QCache also has logic to evaluate which entries are useful to keep in the
cache, which is the most accessed.

This way, it should over time become a list of "hot", i.e. frequently accessed
file paths, improving efficiency.

See the QCache docs about those internals.

b100dian · 2025-10-31T22:29:27Z

Thanks for the detailed write-up and for starting working on this! Indeed, I remember people having races at initialization that only happen when patchmanager is installed so this is indeed of interest for the whole system.

In this phase, since I did not look at the code yet, my only questions are:

why are you caching the NON-patched files, that are larger number that the patched ones
are you also evicting the cache in some patchmanager related cases (such as: patch installed).

Thanks!

nephros · 2025-10-31T22:48:59Z

Thanks for the detailed write-up and for starting working on this! Indeed, I remember people having races at initialization that only happen when patchmanager is installed so this is indeed of interest for the whole system.

In this phase, since I did not look at the code yet, my only questions are:
1. why are you caching the NON-patched files, that are larger number that the patched ones

The idea is to have a list of "hot false positives", i.e. files that are accessed by almost all processes (like libc, /etc/passwd, and so on) but are usually not patched, and return early (quickly, in minimal time) for those.

And QCache with its auto-eviction logic (as well as a Bloom filter with its "maybe in set/definitely not in set" characteristic) isn't suitable for a list of patched files I think. Such a list would have to be persistent and managed by us.
But doing that reliably is hard, it would be a switch from "just check whether the file exists on the system" to monitoring things like app install/uninstall to maintain the list correctly.

2. are you also evicting the cache in some patchmanager related cases (such as: patch installed).

Not at the moment, but this could (really should) be done, every time the "list of active patches" changes.
Actually the cache could be cleared on every activate/deactivate event, and would be re-populated by the next processes accessing files. If this caching is at all useful, it shouldn't be impacted much by eviction, as these events are relatively rare.

nephros · 2025-10-31T23:02:30Z

Oh, and don't get scared by the long list of commits, the actual changes are not that many.

b100dian

Again, thanks for trying out this in-memory-do-not-touch-the-storage cache.

I have a few comments on the code, and one suggestion for testing:

That is, a simple find /usr -type f -exec head -n1 {} \; could probably be run a couple of times with FSCache on and off to see if there are differences.

If /usr is too big for the cache then maybe some narrower path.

Also, I would be interested in the memory usage - since its a number of hashed keys times one allocated object - what would that mean to fragmentation in the end (that last part) ?

b100dian · 2025-11-01T20:13:05Z

src/bin/patchmanager-daemon/patchmanagerobject.cpp

-            payload = fakePath.toLatin1();
+        bool passAsIs = true;
+        if (
+             (!m_failed) // return unaltered for failed


Oh boy, looking up where m_failed is set reminds me we should rewrite patchmanager in a simpler way :))

Lipstick crash? Journal match ? 😮

src/bin/patchmanager-daemon/patchmanagerobject.cpp

b100dian · 2025-11-01T20:32:29Z

src/bin/patchmanager-daemon/patchmanagerobject.cpp

+            insert(entry, new QObject(), HOTCACHE_COST_STRONG);
+        }
+    }
+    // they may be wrong, so use a higher cost than default


Not sure I follow why do we pre-load a set of sonames if we're going to make them be the first out anyways.

b100dian · 2025-11-01T20:33:24Z

src/bin/patchmanager-daemon/patchmanagerobject.cpp

+    QString list;
+    list + "Filter Stats:"
+         + "\n==========================="
+         + "\n  Hotcache entries:: .............." + size()


really curious some stats after a booted system and after some days of usage, how do they look like ?

Statistics have been approved a bit in the latest commit, and you can now trigger logging (and a string reply) from DBus:

busctl call org.SfietKonstantin.patchmanager /org/SfietKonstantin/patchmanager org.SfietKonstantin.patchmanager statistics| xargs echo -e

Example output:

Patchmanager version: 3.2.12+obs.20251103104027.132.g5670c8e Applied Patches: 20 Patched files: 97 Watched files: 50 Filter Stats: =========================== Hotcache entries:: ..............126 Hotcache cost: ..................259/5000 Hotcache hit/miss: ..............1284/106 (0.00%) =========================== Hotcache top entries: ........... /usr/share/patchmanager/patches/patch-keyboard-no-vibration/patch.json /usr/share/patchmanager/patches/patch-named-sims/patch.json /usr/share/patchmanager/patches/patch-email-headers/unified_diff.patch /usr/share/patchmanager/patches/patch-icon-filter/patch.json /etc/passwd /usr/lib64/libblkid.so.1.1.0

Cool! what happens if you run find / -type f -exec grep something {} \; with the cache?

(honest question, I do this all the times ;D)

(I should make my own build but I am away from any personal computer for the week)

Cool! what happens if you run find / -type f -exec grep something {} \; with the cache?

(honest question, I do this all the times ;D)

You need plocate ;)

So it goes to something like this.

busctl call "org.SfietKonstantin.patchmanager" /org/SfietKonstantin/patc hmanager org.SfietKonstantin.patchmanager statistics b true | xargs echo -e|head s Patchmanager version: 3.2.12+obs.20251103192017.134.g49ef3e1 Applied Patches: 20 Patched files: 97 Watched files: 50 Filter Stats: =========================== Hotcache entries:: ..............1512 Hotcache cost: ..................3033/5000 Hotcache hit/miss: ..............59162/1506 (0.98%) ===========================

Btw, don't actually do find /, that can crash or hang the device (because of stuff in /proc or /sys or so).

By the way, you MUST run find as non-root, otherwise the preload lib will do nothing :)

src/bin/patchmanager-daemon/patchmanagerobject.h

src/bin/patchmanager-daemon/patchmanagerobject.cpp

Co-authored-by: Vlad G. <[email protected]>

nephros added 24 commits October 29, 2025 11:25

Hotcache: Add a cache object to track nonexisting files

ad6c7b1

Hotcache: remove entry if it existed

37d44e9

Hotcache: simplify logic

e2fddc3

Hotcache: store Object, retrieve one on lookup

bb975f4

Hotcache: remove log line

5fb9b3f

Hotcache: use object() not take()

5951671

take() gives ownership to the caller. object() does not.

Hotcache: move debugging messages after writing to the connection

de94ed1

Hotcache: set up the filter with some initial contents

20925c1

Hotcache: Add a Settings switch

4ba2081

Hotcache: wire up settings switch

bb422ca

Hotcache: move into a class, rename to m_filter

a1826a4

Hotcache: Don't be a template class

3fd441c

- Try a Q_OBJECT macro for a template class... - Don't be a template class - Become Object-based, add contructor - Implement a constructor

Hotcache: Prepare for stats collecting

6884b81

Hotcache: Static initializer lists

b0c5479

Hotcache: Implement a simpler stats method

bdedbeb

Hotcache: Make the lists public

969a763

Hotcache: Use foreach again.

9cb1321

Hotcache: Missing braces

82c3901

Hotcache: check for active filter in startReadingLocalServer

b33ecff

Hotcache: Use plain qstring for stats

b65f653

Hotcache: Add some debug prints

8a3787d

Hotcache: Don't copy around strings, us a bool.

7df9ddb

Print hotcache stats on PM exit

475ac37

Resolve symlinks before adding to primed cache

5f9e2c9

nephros changed the title ~~Keep a cache of frequently accessed files~~ RFC: Keep a cache of frequently accessed files Oct 31, 2025

b100dian reviewed Nov 1, 2025

View reviewed changes

nephros and others added 4 commits November 2, 2025 00:34

Fix cache returning false the false way.

1a6e64f

Update src/bin/patchmanager-daemon/patchmanagerobject.cpp

18a3c44

Co-authored-by: Vlad G. <[email protected]>

Promote stat output to info level

64eb096

Split out Filter into .h

ec2aae2

nephros force-pushed the hotcache-as-class branch from e94e2e6 to ec2aae2 Compare November 2, 2025 00:15

nephros added 11 commits November 2, 2025 01:28

Split out Filter into .cpp

8b70747

fixup! Split out Filter into .h

29cd51e

Rework PM Filter class sources

b995f30

Print statistics on demand

59348ec

Rework Stats yet again

0d2fd91

Remove stats call from destructor

213b508

Stats: limit max output size

59ea818

Stats: lower max cost

7421eff

Default values go into header

48ccda6

Cache: Use small data type for cache entries

ebb671f

Add a comment about sizing the cache

e31d6bf

RFC: Keep a cache of frequently accessed files #517

Are you sure you want to change the base?

RFC: Keep a cache of frequently accessed files #517

Uh oh!

Conversation

nephros commented Oct 29, 2025

Uh oh!

nephros commented Oct 29, 2025

Uh oh!

nephros commented Oct 29, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

b100dian commented Oct 31, 2025

Uh oh!

nephros commented Oct 31, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

nephros commented Oct 31, 2025

Uh oh!

b100dian left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

nephros Nov 3, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

nephros commented Oct 29, 2025 •

edited

Loading

nephros commented Oct 31, 2025 •

edited

Loading

nephros Nov 3, 2025 •

edited

Loading