[NFS-Ganesha-Support] Re: Ref Leak in Chunk Cache

So, the chunking isn't perfect. There's an interaction between how many dirents the FSAL returns and how many dirents we can return to the client. It looks like, if we're at the end of a chunk when the client response is full, we'll start a new chunk on the next cycle, even if more dirents would fit in the current chunk. I think in your case, the only way to get perfect chunking would be for the readahead from the FSAL to be large enough to always return 500 dirents, so that each directory would contain exactly 1 chunk. Most FSALs, however, do readahead based on a buffer size, and fit as many dirents as they can into the buffer, so readahead count will change based on the names of the files in the directory. As for subsequent runs, the issue here is that the entire working set (total number of files and directories) is 12,500,000, which is much larger than the 500,000 of the hwmark. This means that handles for some of the directories early in the listing will be re-used later in the listing, clearing out the dirent cache for those directories (dirent cache it not a separate thing; the dirent cache for each directory is attached to that directory's handle, and so is freed when that handle is freed or re-used). If you always list the directories in the same order, then the next listing, the handles for later directories will be re-used when scanning the early entries, so you will get a new set of dirents and chunks every listing. Basically, this workload is a pathological case for the combination of handle cache and dirent cache, as implemented in Ganesha. Those caches were designed on the idea that the working set of files/directories would be a fairly constant subset of the total set of files and directories, that might change slowly over time, but that it wouldn't be the entire set. This workload of listing the entire tree over and over will be effectively reloading the entire cache every time. That said, it's not impossible that there's a leak somewhere. Could you run the listing maybe 20 times and see if the number only goes up? If so, that would be a strong hint that there's a leak. If it goes up and down, then it's likely just variability in readahead/client buffer in the listing. Daniel On 8/17/21 3:54 PM, None via Support wrote: