diff options
Diffstat (limited to 'open_issues/memory_object_model_vs_block-level_cache.mdwn')
-rw-r--r-- | open_issues/memory_object_model_vs_block-level_cache.mdwn | 273 |
1 files changed, 273 insertions, 0 deletions
diff --git a/open_issues/memory_object_model_vs_block-level_cache.mdwn b/open_issues/memory_object_model_vs_block-level_cache.mdwn new file mode 100644 index 00000000..7da5dea4 --- /dev/null +++ b/open_issues/memory_object_model_vs_block-level_cache.mdwn @@ -0,0 +1,273 @@ +[[!meta copyright="Copyright © 2012 Free Software Foundation, Inc."]] + +[[!meta license="""[[!toggle id="license" text="GFDL 1.2+"]][[!toggleable +id="license" text="Permission is granted to copy, distribute and/or modify this +document under the terms of the GNU Free Documentation License, Version 1.2 or +any later version published by the Free Software Foundation; with no Invariant +Sections, no Front-Cover Texts, and no Back-Cover Texts. A copy of the license +is included in the section entitled [[GNU Free Documentation +License|/fdl]]."]]"""]] + +[[!tag open_issue_documentation open_issue_hurd open_issue_gnumach]] + + +# IRC, freenode, #hurd, 2012-02-14 + + <slpz> Open question: what do you think about dropping the memory object + model and implementing a simple block-level cache? + +[[microkernel/mach/memory_object]]. + + <kilobug> slpz: AFAIK the memory object has more purpose than just cache, + it's allow used for passing chunk of data between processes, handling + swap (which similar to cache, but still slightly different), ... + <slpz> kilobug: user processes usually make their way to data with POSIX + operations, so memory objects are only needed for mmap'ed files + <slpz> kilobug: and swap can be replaced for an in-kernel system or even + could still use the memory object + <braunr> slpz: memory objects are used for the page cache + <kilobug> slpz: translators (especially diskfs based) make heavy use of + memory objects, and if "user processes" use POSIX semantics, Hurd process + (translators, pagers, ...) shouldn't be bound to POSIX + <slpz> braunr: and page cache could be moved to a lower level, near to the + devices + <braunr> not likely + <braunr> well, it could, but then you'd still have the file system overhead + <slpz> kilobug: but the use of memory objects it's not compulsory, you can + easily write a fs translator without implementing memory objects at all + (except to mmap) + <braunr> a unified buffer/VM cache as all modern systems have is probably + the most efficient approach + <slpz> braunr: I agree. I want to look at *BSD/Linux vfs systems to seem + how much cache policy depends on the filesystem + <slpz> braunr: Are you aware of any good papers on this matter? + <braunr> netbsd UVM, the linux virtual memory system + <braunr> both a bit old bit still relevant + <slpz> braunr: Thanks. + <slpz> the problem in our case is that having FS and cache information at + different contexts (kernel vs. translator), I find hard to coordinate + them. + <slpz> that's why I though about a block-level cache that GNU Mach could + manage by itself + <slpz> I wonder how QNX deals with this + <braunr> the point of having a simple page cache is explicitely about not + caring if those pages are blocks or files or whatever + <braunr> the kernel (at least, mach) normally has all the accounting + information it needs to implement its cache policy + <braunr> file system translators shouldn't cache much + <braunr> the pager interface could be refined, but it looks ok to me as it + is + <slpz> Mach has the accounting info, but it's not able to purge the cache + without coordination with translators + <braunr> which is normal + <slpz> And this is a big problem when memory pressure increases, as it + doesn't know for sure when memory is going to be freed + <braunr> Mach flushes its cache when it decides to, and sends back dirty + pages if needed by the pager + <braunr> that's the case with every paging implementation + <braunr> the main difference is security with untrusted pagers + <braunr> but that's another issue + <slpz> but in a monolithic implementation, the kernel is able for force a + chunk of cache memory to be freed without hoping for other process to do + the job + <braunr> that's not true + <braunr> they're not process, they're threads, but the timing issue is the + same + <braunr> see pdflush on linux + <slpz> no, it isn't. + <braunr> when memory is scarce, threads that request memory can either wait + or immediately fail, and if they wait, they're usually woken by one of + the vm threads once flushing is done + <slpz> a kernel thread can access all the information in the kernel, and + synchronization is pretty easy. + <braunr> on mach, synchronization is done with messages, that's even easier + than shared kernel locks + <slpz> with processes in different spaces, resource coordination becomes + really difficult + <braunr> and what kind of info would an external pager need when simply + asked to take back its dirty pages + <braunr> what resources ? + <slpz> just take a look at the thread storm problem when GNU Mach needs to + clean a bunch of pages + <braunr> Mach is big enough to correctly account memory + <braunr> there can be thread storms on monolithic systems + <braunr> that's a Mach issue, not a microkernel issue + <braunr> that's why linux limits the number of pdflush thread instances + <slpz> Mach can account memory, but can't assure when be freed by any + means, in a lesser degree than a monolithic system + <braunr> again i disagree + <braunr> no system can guarantee when memory will be freed with paging + <slpz> a block level cache can, for most situations + <braunr> slpz: why ? + <braunr> slpz: or how i mean ? + <slpz> braunr: with a block-level page cache, GNU Mach should be able to + flush dirty pages directly to the underlaying device without all the + complexity and resource cost involved in a m_o_data_return message. It + can also throttle the rate at which pages are being cleaned, and do all + this while blocking new page allocations to deal with memory exhaustion + cases. + <slpz> braunr: in the current state, when cleaning dirty pages, GNU Mach + sends a bunch on m_o_data_return to the corresponding pagers, hoping they + will do their job as soon and as fast as possible. + <slpz> memory is not really freed, but transformed from page cache to + anonymous memory pertaining to the corresponding translator + <slpz> and GNU Mach never knows for sure when this memory is released, if + it ever is. + <slpz> not being able to flush dirty pages synchronously is a big problem + when you need to throttle memory usage + <slpz> and needing allocating more memory when you're trying to free (which + is the case for the m_o_data_return mechanism) makes the problem even + worse + <braunr> your idea of a block level cache means in kernel block drivers + <braunr> that's not the direction we're taking + <braunr> i agree flushing should be a synchronous process, which was one of + the proposed improvements in the thread migration papers + <braunr> (they didn't achieve it but thought about it for future works, so + that the thread at the origin of the fault would handle it itself) + <braunr> but it should be possible to have kernel threads similar to + pdflush and throttle flush requests too + <braunr> again, i really think it's a mach bug, and having a buffer cache + would be stepping backward + <braunr> the real design issue is allocating memory while trying to free + it, yes + <slpz> braunr: thread migration doesn't apply to asynchronous IPC, and the + entire paging mechanism is implemented this way + <slpz> in fact, trying to do a synchronous m_o_data_return will trigger a + deadlock for sure + <slpz> to achieve synchronous flushing with translators, the entire paging + model must be redesigned + <slpz> It's true that I'm not very confident of the viability of user space + drivers + <slpz> at least, not for every device + <slpz> I know this is against the current ideas for most ukernel designs, + but if we want to achieve real work functionality, I think some + sacrifices must be done. Or at least a reasonable compromise. + <braunr> slpz: thread migration for paging requests implies synchronous + RPC, we don't care much about the IPC layer there + <braunr> and it requires large changes of the VM code in addition, yes + <braunr> let's not talk about this, we don't have thread migration anyway + :p + <braunr> except the allocation-on-free-path issue, i really don't see how + the current pager interface or the page cache creates problems wrt + flushing .. + <braunr> monolithic systems also have that problem, with lower impacts + though, but still + <slpz> braunr: because as it doesn't know when memory is really freed, 1) + it just blindly sends a bunch of m_o_data_return to the pagers, usually + overloading them (causing thread storms), and 2) it can't properly + throttle new page requests to deal with resource exhaustion + <braunr> it does know when memory is really freed + <braunr> and yes, it blindly sends a bunch of requests, they can and should + be trottled + <slpz> but dirty pages freed become indistinguishable from common anonymous + chunks released, so it doesn't really know if page flushes are really + working or not (i.e. doesn't know how fast a device is processing write + requests) + <braunr> memory is freed when the pager deallocates it + <braunr> the speed of the operation is irrelevant + <braunr> no system can rely on disk speed to guarantee correct page flushes + <braunr> disk or anything else + <slpz> requests can't be throttled if Mach doesn't know when they are being + processed + <braunr> it can easily know it + <braunr> they are processed as soon as the request is sent from the kernel + <braunr> and processing is done when the pager acknowledges the end of the + flush + <braunr> memory backing the flushed pages should be released before + acknowleding that to avoid starting new requests too soon + <slpz> AFAIK pagers doesn't acknowledge the end of the flush + <braunr> well that's where the interface should be refined + <slpz> Mach just sends the m_o_data_return and continues on its own + <braunr> that's why flushing should be synrhconous + <braunr> are you sure about that however ? + <slpz> so the entire paging system needs a new design... :) + <slpz> pretty sure + <braunr> not a new design .. + <braunr> there is m_o_supply_completed, i don't see how difficult it would + be to add m_o_data_return_completed + <braunr> it's not a small change, but not a difficult one either + <braunr> i'm more worried about the allocation problem + <braunr> the default pager should probably be wired in memory + <braunr> maybe others + <slpz> let's suppose a case in which Mach needs to free memory due to an + increase in its pressure. vm_pageout_daemon starts running, clean pages + are freed easily, but for each dirty one a m_o_data_return in sent. 1) + when should this daemon stop sending m_o_data_return and start waiting + for m_o_data_return_completed? 2) what happens if the translator needs to + read new blocks to fulfill a write request (pretty common in ext2fs)? + <braunr> it should stop after an arbitrary limit is reached + <braunr> a reasonable one + <braunr> linux limits the number of pdflush threads for that reason as i + mentioned (to 8 iirc) + <braunr> the problem of reading blocks while flushing is what i'm worried + about too, hence the need to wire that code + <braunr> well, i'm nto sure it's needed + <braunr> again, a reasonable about of free memory should be reserved for + that at all times + <slpz> but the work for pdflush seems to be a lot easier, as it only deals + directly with block devices (if I understood it correctly, I just started + looking at it). + <braunr> i don't know how other systems compute that, but this is how they + seem to do as well + <braunr> no, i don't think so + <slpz> well, I'll try to invest a few days understanding how pdflush work, + to see if some ideas can be borrowed for Hurd + <braunr> iirc, freebsd has thresholds in percent for each part of its cache + (active, inactive, free, dirty) + <slpz> but I still think simple solutions work better, and using the memory + object for page cache is tremendously complex. + <braunr> the amount of free cache pages is generally sufficient to + guarantee much memory can be released at once if needed, without flushing + anything + <braunr> yes but that's the whole point of the Mach VM + <braunr> and its greatest advance .. + <slpz> what, memory objects? + <braunr> yes + <braunr> using physical memory as a cache for anything, not just block + buffers + <slpz> memory objects work great as a way to provide a shared image of + objects between processes, but as page cache they are an overkill (IMHO). + <slpz> or, at least, in the way we're using them + <braunr> probably + <braunr> http://lwn.net/Articles/326552/ + <braunr> this can help udnerstand the problems we may have without better + knowledge of the underlying devices, yes + <braunr> (e.g. not being able to send multiple requests to pagers that + don't share the same disk) + <braunr> slpz: actually i'm not sure it's that overkill + <braunr> the linux vm uses struct vm_file to represent memory objects iirc + <braunr> there are many links between that structure and some vfs related + subsystems + <braunr> when a system very actively uses the page cache, the kernel has to + maintain a lot of objects to accurately describe the cache content + <braunr> you could consider this overkill at first too + <braunr> the mach way of doing it just implies some ipc messages instead of + function calls, it's not that overkill for me + <braunr> the main problems are recursion (allocation while freeing, + handling page faults in order to handle flushes, that sort of things) + <braunr> struct file and struct address_space actually + <braunr> slpz: see struct address_space, it contains a set of function + pointers that can help understanding the linux pager interface + <braunr> they probably sufferred from similar caveats and worked around + them, adjusting that interface on the way + <slpz> but their strategy makes them able to treat the relationship between + the page cache and the block devices in a really simple way, almost as a + traditional storage cache. + <slpz> meanwhile on Mach+pager scenario, the relationship between a block + in a file and its underlying storage becomes really blurry + <slpz> this is a huge advantage when flusing out data, specially when + resources are scarce + <slpz> I think the idea of using abstract objects for page cache, loses a + bit the point that we just want to avoid accessing constantly to a slow + device + <slpz> and breaking the tight relationship between the device and its + cache, makes things a lot harder + <slpz> this also manifest itself when flushing clean pages, as things like + having an static maximum for cached memory objects + <slpz> we shouldn't care about the number of objects, we just need to + control the number of pages + <slpz> but as we need the pager to flush pages, we need to keep alive a lot + of control ports to them + <mcsim> slpz: When mo_data_return is called, once the memory manager no + longer needs supplied data, it should be deallocated using + vm_deallocate. So this way pagers acknowledges the end of flush. |