On Tue, May 5, 2020 at 7:20 PM Amir Goldstein <amir73il@xxxxxxxxx> wrote: > > Jan, > > In the v3 posting of the name info patches [1] I dropped the > FAN_REPORT_NAME patches as agreed to defer them to next cycle. > > Following is remainder of the series to complement the FAN_DIR_MODIFY > patches that were merged to v5.7-rc1. > > The v3 patches are available on my github branch fanotify_name [2]. > Same branch names for LTP tests [3], man page draft [4] and a demo [5]. > > Patches 1-4 are cleanup and minor re-factoring in prep for the name > info patches. > > Patch 5 adds the FAN_REPORT_NAME flag and the new event reporting format > combined of FAN_EVENT_INFO_TYPE_DFID_NAME and FAN_EVENT_INFO_TYPE_FID > info records, but provides not much added value beyond inotify. > > Patches 6-7 add the new capability of filesystem/mount watch with events > including name info. > > I have made an API decision that stems from consolidating the > implementation with fsnotify_parent() that requires your approval - > A filesystem/mount mark with FAN_REPORT_NAME behaves as if all the > directories and inodes are marked. This results in user getting all > relevant events in two flavors - one with both info records and one with just > FAN_EVENT_INFO_TYPE_FID. I have tries several approaches to work around this > bizarrity, but in the end I decided that would be the lesser evil and that > bizarre behavior is at least easy to document. > Hi Jan, Were you able to give some thought to the API question above? I would really like to be able to finalize the design of the API, so that I will be able to continue working on the man page patches. Re-phrasing the API question that needs addressing: With FAN_REPORT_NAME, filesystem/mount watches get - 1. ONLY events with name (no events on root and no SELF events) 2. Each event in one flavor (with name info when available) 3. ALL events in both flavors (where name info is available) 4. Something else? The current v3 patches implement API choice #3, which is derived from the implementation choice to emit two events via fsnotify_parent(). My v2 patches implemented API choice #2, which was the reason for duplicating name snapshot code inside handle_event(). I considered implementing merge of event with and without name, but it seemed too ugly to live. We could also go for an API that allows any combination of FAN_REPORT_NAME and FAN_REPORT_FID: FAN_REPORT_FID - current upstream behavior FAN_REPORT_FID_NAME - choice #3 above as implemented in v3 FAN_REPORT_NAME - choice #1 above At one point, I also considered that user needs to opt-in for named events per filesystem/mount mark with FAN_EVENT_ON_CHILD. This flag is implied in v3 for all filesystem/mount marks by FAN_REPORT_NAME, while for directory marks it is opt-in as it has always been. At the moment I went with choice #3 in v3 because I felt it would be the simplest of all choices to document. I didn't want to invest time in documenting behavior if you find it unacceptable. If you are swaying between more than one option, then I am willing to try documenting more than one option, so we can see what the result looks like. Thanks, Amir.
