On 2/22/17 8:52 AM, Vaibhaw Pandey wrote: > Thanks for replying Eric. > >> So I'm curious - what happened, and how did you resolve it? > Nothing really. The application in question is the Redis key value > store. Reading through it's code is how I began to suspect rename as > the potential problem. It does a fopen/write/fsync/rename sequence. I > mailed Antirez (Redis's primary author) about it and meanwhile added a > sync call in my code. :) ok, cool. Good answer! >> I agree with Carlos that direct comparisons to other filesystems are >> less useful, if nothing else because other filesystems may change. > As I said, I probably shouldn't have framed the questions like that. > But, IMHO, those are valid inquiries that people would make about XFS, > even when not comparing it with filesystems they are familiar with. I > was only trying to suggest what could be useful for other developers > looking to understand enough XFS work with it confidently. The current > XFS FAQ does seem to have gaps IMO. But you guys would know best on > what should be documented. :) Well, don't get me wrong. If there are gaps in FAQ, let's fix them. And it's a wiki, so as long as it's functioning correctly (?), then /you/ can fix them! :) But we do want to keep the style clean and consistent, and keep it accurate and factual, and limit it to actual FAQs. If you'd like to put together proposed new entries and run them by the list for critique, that'd be great. >> which controls data flushing. Some of this is filesystems 101; some of >> it is specific to xfs... > > That is a rather important difference, thanks for pointing it out! > Again, something that should be added to the FAQ. :) I just don't want to go too far into "document the world" in the faq. Some things like kernel/vfs flushing behavior, which is not xfs-specific, probably doesn't belong there. And tunables which should almost never be tuned without a deep understanding of a specific workload's performance probably don't belong in the faq either. There is a strong tendency to turn knobs without really understanding what they do; highlighting, say, the xfssyncd_centisecs sysctl in the faq will probably lead to people fiddling with that knob without real understanding and without good reason. Or, to put it another way, "what is xfssyncd_centisecs?" is not actually a Frequently Asked Question; it is a low-level behavior control which few people will need to touch. >> ;) IOWs, why were you messing with the log size in the first place? :) > I wasn't at all :) We are a Database-as-a-Service startup and are > looking to start deploying MongoDB on XFS for better performance. & > understanding the sizing requirements and some level of internals is > important for us. I guess my point was: before documenting max log size, we'd document "why would I want a larger than default log?" And given that the answer is almost always "you don't; use the defaults" (tweaking something like the log size would almost always be the result of fairly low level perf investigation, not a general suggestion in the faq), we probably don't need to address log size in the FAQ at all. Anyway - if there are FAQs you feel are not present in the FAQ list, feel free to send proposals through the list, and edit the wiki as appropriate. Thanks! -Eric > On Wed, Feb 22, 2017 at 7:11 PM, Eric Sandeen <sandeen@xxxxxxxxxxx> wrote: >> On 2/22/17 5:55 AM, Vaibhaw Pandey wrote: >>> Hey, >>> >>> I had recently run into the ext4 auto_da_alloc delayed allocation type >>> behavior with XFS i.e. replace by rename leaving an empty file behind. >>> It took me forever to debug it cause I couldn't find answers to some >>> simple questions right away. >> >> So I'm curious - what happened, and how did you resolve it? >> >>> You guys are the experts but I would like to suggest adding some >>> questions (& answers) to the XFS FAQ doc for the clueless folks like >>> me. >>> >>> I would suggest the following questions: >>> >>> 0. Does XFS support a mount option equivalent to ext4's auto_da_alloc? >>> i.e. Does XFS have the workarounds to support the replace by truncate >>> and replace by rename? >>> Ans: >>> Answered exactly in http://oss.sgi.com/archives/xfs/2015-12/msg00553.html >> >> http://xfs.org/index.php/XFS_FAQ#Q:_Why_do_I_see_binary_NULLS_in_some_files_after_recovery_when_I_unplugged_the_power.3F' >> alludes to this, but I suppose a better explanation of the existing >> heuristic might be nice for those who want the details. The "binary NULLs" >> thing is ancient history. >> >> I agree with Carlos that direct comparisons to other filesystems are >> less useful, if nothing else because other filesystems may change. >> >> Documenting what XFS does should be the goal of the FAQ. >> >>> 1. Does XFS support a mount option equivalent to ext4's commit? i.e. >>> How do I control how often does XFS sync to disk? Or Does XFS never >>> sync to disk until a sync/fsync is called? >> >> ext4's commit= doesn't control how often it "syncs to disk", exactly. >> (that's a bit vague). >> >> It controls the journal commit time, which may or may not (depending on >> other options) control data vs. metadata, etc. Again, we'd need to >> document ext4 in the faq before we started making comparisons to it. :) >> >>> Ans: >>> Answered here: http://article.gmane.org/gmane.comp.file-systems.xfs.general/53376 >>> Reproducing from source: >>> <snip> >>> By and large, buffered IO in a filesystem is flushed out by the vm, >>> due to either age or memory pressure. The filesystem then responds >>> to these requests by the VM, writing data as requested. >> >> So that's about dirty data flushing, whereas ext4's commit= is more related >> to metadata flushing, which may or may not lead to data flushing for some files. >> We do document a sysctl: >> >> fs.xfs.xfssyncd_centisecs (Min: 100 Default: 3000 Max: 720000) >> The interval at which the filesystem flushes metadata >> out to disk and runs internal cache cleanup routines. >> >> which is different from: >> >>> You can read all about it in >>> https://www.kernel.org/doc/Documentation/sysctl/vm.txt See >>> dirty_expire_centisecs and dirty_writeback_centisecs - flushers wake >>> up every 30s and push on data more than 5s old, by default. >>> </snip> >> >> which controls data flushing. Some of this is filesystems 101; some of >> it is specific to xfs... >> >>> 2. What is the maximum size of the XFS journal? >>> Ans: Not sure. But this is the closest answer I could find: >>> https://serverfault.com/questions/367973/xfs-maximum-log-size-sw-raid-10-mdadm-sles-11-sp1 >>> I could read through the code and find a better answer in case you >>> folks wouldn't have the time. >> >> As you can see from the URL there, the answer is "it's complicated" >> and depends on filesystem geometry. I think it's probably best answered by: >> >> http://xfs.org/index.php/XFS_FAQ#Q:_I_want_to_tune_my_XFS_filesystems_for_.3Csomething.3E >> >> ;) IOWs, why were you messing with the log size in the first place? :) >> >> -Eric >> >> >> >>> >>> Lemme know what you think. >>> >>> Thanks, >>> Vaibhaw >>> -- >>> To unsubscribe from this list: send the line "unsubscribe linux-xfs" in >>> the body of a message to majordomo@xxxxxxxxxxxxxxx >>> More majordomo info at http://vger.kernel.org/majordomo-info.html >>> > -- To unsubscribe from this list: send the line "unsubscribe linux-xfs" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
