A clean approach to writeout throttling

Good afternoon,

According to me, each line of code removed from the kernel is worth ten lines added. If lines can be removed while at the same time improving performance, that is worth, ah, about 1,000 times more than lines added, correct? Or maybe 1,000,000 times as much, if removing lines removes a deadlock at the same time, something like that.

Today's patch is a step towards removing many lines of code from mainline and also aims to improve write cache performance, eliminate a troublesome class of deadlocks, save kernel memory and make the kernel easier to read.

Background

We seriously began to tackle the issue of block writeout vm deadlock more than three years ago, and by now we have acreted a creaky agglomeration of dirty page limits, dirty page balancing between block devices, and miscellaneous other hacks attempting to solve these deadlocks. These bandaids as of 2.6.23 do not in themselves address the underlying problem, but they do add a lot of code to core kernel and they do impair write Linux's write cache performance. This is a classic case of fixing symptoms instead of the cause of a problem. The worst part of it? The dirty page limit idea did not actually fix the deadlock it was supposed to, but instead generated a new class of deadlocks that are easier to trigger.

The basis of the writeout deadlock scenario is easy to see: a task requests a page of memory, but all pages are currently in use for caching among other things. To recover memory, some disk-backed pages must be evicted. Dirty pages must be written to disk before being evicted, so they are passed to the block layer for writeout. If any code in the block writeout path needs to allocate memory to do its work, then we can deadlock because the shortage of memory prevents the block layer from making progress to recover memory.

So far, I have just summarized what everybody knows. This deadlock scenario has been present in Linux since day one, and has been solved for local block devices since very early on, by the simple expedient of providing a reserve of "memalloc" pages that a task may access only if it is in the call chain of a memory manager task engaged in writing out dirty pages.

What was not commonly known three years ago is that the memalloc reserve solution fails in some cases, all of which share the common attribute that the task trying to allocate memory for block writeout is not the same as the task that initiated the writeout. In this case, the "PF_MEMALLOC" task flag strategy fails because the consumer of the memory is not in the call chain of the submitter, and thus does not inherit the flag. Examples of such deadlock-prone use cases include:

* Fancy virtual block devices with helper daemons
* Network block device accessed locally
* Swap over network
* Remote block devices in general
* Any block device with a user space component
* Filesystems implemented in user space

To put it bluntly: without solving these deadlocks, Linux is pretty much useless in many modern storage roles.

When I started working on this problem years go, I went at it by tackling the nastiest, most icky manifestation of it first, namely the possibility that the network layer might be unable to allocate memory to receive a reply packet from a remote block device. Unfortunately, the tricky details of the solution to this problem had the unintended effect of overshadowing the main issues, just because the details of the network receive deadlock are so deliciously obscure. Ironically, we found that the network read readlock scenario does not occur in practice. It is high time to draw attention back to the main issue.

The meta-mistake I made while tackling this problem was to focus mainly on the logistics of providing "writeout helper" tasks with access to memory reserves, and just assume that it would be easy to place limits on how deep the helpers would dip into the reserves, which restriction is obviously necessary to prevent deadlock. This was so obvious in fact that I (we) did not get around to implementing it until quite recently. Then it became abundantly clear that limiting resource requirements is actually the main ingredient of any correct solution, and that the various complexities of providing access to memory reserves do not amount to much more than an interesting side show. We (Zumastor team) proved this to ourselves by removing the entire "PeterZ" patch set we were carrying (a distant descendant of my original network deadlock prevention patch) and lo! No deadlocks. It seems that throttling writeout traffic in an organized way amounts to powerful magic indeed.

So that is all by way of saying, today's patch to limit the amount of data in flight to a block device is an Important Patch. We find that our own storage project needs very little more than this hundred lines of code or so to wave goodbye permanently to writeout deadlocks, and thereby make our storage software actually useful. Extension to the other use cases listed above is Obvious[tm].

Theory

Terje Mathisen said "almost all programming can be viewed as an exercise in caching." In fact, almost all of the Linux kernel can be viewed as an exercise in caching. The main job of the kernel (granted, there are other side jobs) is to move data back and forth between disk and memory. We can usefully define the from-memory-to-disk half of this task as "progress". So the main reason for the kernel to exist at all is to make progress by writing dirty memory to disk. OK?

With that definition in hand, we can see right away what must be done to guarantee progress: the part of the kernel that makes progress must be guaranteed to have enough resources to make progress. Like a shark, when the kernel stops swimming it dies. (Ok, that analogy is admittedly stupid but I still cannot get Linus's rabbits and bazookas out of my head, so retaliation is in order.)

To guarantee resources, we need to do two things:

1) Provide a dedicated pool of resources

2) Ensure that the consumer never requires more than its share of the dedicated resource pool in order to make progress

For the problem at hand a suitable resource pool already exists, namely the memalloc reserve. However, examination of the attached patch will show that it knows nothing about that particular form of dedicated
reserve: in fact any form of reserve, for example, Ingo's mempool, or any other will do. The key item is (2) above: we require a sane way of ensuring that the consumer (the block writeout path) never exceeds its allotment of resources. This we accomplish simply, by imposing a limit on the amount of data that can be in flight to any particular block device.

Practice

So here is the key idea of today's patch: it provides a simple mechanism for imposing a limit on the amount of data that can be in flight to any particular block device.

The limit on in-flight data is in fact expressed generically, as it is hard to set down any single rule to specify the amount of resources any particular bio transfer will require. Instead, the block layer provides a method that a block driver may optionally fill in, to calculate the resource bound in units of the block driver's choosing. The block driver thus takes upon itself the task of translating its own, self-imposed bound into generic resource units that can be treated generically by the kernel. In simple terms, the block driver looks at each bio and decides how many pages (at most) of memalloc reserve could be needed to fully service the transfer, and translates that requirement into generic units for use by the block layer. The block layer compares the generic units to a generic bound provided by the block driver at initialization time and decides whether to let the bio transfer in question proceed, or hold it back.

This idea is Simplicity itself. Some not so obvious details follow.

For one thing, a block device these days may not be just a single device, but may be a stack of devices connected together by a generic mechanism such as device mapper, or a hardcoded stack such as multi-disk or network block device. It is necessary to consider the resource requirements of the stack as a whole _before_ letting a transfer proceed into any layer of the stack, otherwise deadlock on many partially completed transfers becomes a possibility. For this reason, the bio throttling is only implemented at the initial, highest level submission of the bio to the block layer and not for any recursive submission of the same bio to a lower level block device in a stack.

This in turn has rather far reaching implications: the top level device in a stack must take care of inspecting the entire stack in order to determine how to calculate its resource requirements, thus becoming the boss device for the entire stack. Though this intriguing idea could easily become the cause of endless design work and many thousands of lines of fancy code, today I sidestep the question entirely using the "just provide lots of reserve" strategy. Horrifying as it may seem to some, this is precisely the strategy that Linux has used in the context of resource management in general, from the very beginning and likely continuing for quite some time into the future My strongly held opinion in this matter is that we need to solve the real, underlying problems definitively with nice code before declaring the opening of fancy patch season. So I am leaving further discussion of automatic resource discovery algorithms and the like out of this post.

Today's patch implements _only_ the inflight-data limiting, and does not include any code to provide access to reserves. In practice, simply oring PF_MEMALLOC into the flags of each writeout helper daemon does the trick. We also found that we had to disable or work around the code that enforces memory dirty limits (contestion_wait) which otherwise causes a whole new class of deadlocks, which are in fact easier to trigger than the deadlocks it attempts to fix.

Overhead.

This bio patch adds a small amount of overhead to the bio processing path, consisting of one new test in submit_bio (generic_make_request) and one new test in bio->bi_endio, just to see if thottling methods are present. I do not think that the extra cpu consumed is measurable. There is also a slight increase in the size of struct bio to hold two new fields, one to reference the struct queue of the block device in question and another to remember how many units of resources were assigned to the bio in order that exactly that many may be released on completion. In fact, total memory use due to struct bio is typically reduced by a not insignificant amount by limiting the number of bios concurrently in flight. Provided of course that local block devices also adopt this mechanism, which in fact would be a Very Good Thing[1] for reasons I will not get into here.

That is enough for today, and arguably too much, because in the past, communicating these simple ideas has always seemed to founder on the intrusion of many peripherally related ideas, hijacking the discussion. Either that, or appear to be such an involved subject that nobody will risk anything more than a cosmetic reply. Hopefully not this time.

Let me close with perhaps the most relevant remarks: the attached code has been in heavy testing and in production for months now. Thus there is nothing theoretical when I say it works, and the patch speaks for itself in terms of obvious correctness. What I hope to add to this in the not too distant future is the news that we have removed hundreds of lines of existing kernel code, maintaining stability and improving performance.

Regards,

Daniel

@@ -3210,7 +3210,7 @@ static inline int bio_check_eod(struct b
  */
 static inline void __generic_make_request(struct bio *bio)
 {
-	struct request_queue *q;
+	request_queue_t *q = bdev_get_queue(bio->bi_bdev);
 	sector_t old_sector;
 	int ret, nr_sectors = bio_sectors(bio);
 	dev_t old_dev;

@@ -3221,6 +3221,13 @@ static inline void __generic_make_reques
 	if (bio_check_eod(bio, nr_sectors))
 		goto end_io;
 
+	if (q && q->metric && !bio->bi_queue) {
+		int need = bio->bi_throttle = q->metric(bio);
+		bio->bi_queue = q;
+		/* FIXME: potential race if atomic_sub is called in the middle of condition check */
+		wait_event_interruptible(q->throttle_wait, atomic_read(&q->available) >= need);
+		atomic_sub(need, &q->available);
+	}
 	/*
 	 * Resolve the mapping until finished. (drivers are
 	 * still free to implement/resolve their own stacking

@@ -3234,7 +3241,6 @@ static inline void __generic_make_reques
 	do {
 		char b[BDEVNAME_SIZE];
 
-		q = bdev_get_queue(bio->bi_bdev);
 		if (!q) {
 			printk(KERN_ERR
 			       "generic_make_request: Trying to access "

@@ -889,6 +889,11 @@ static int dm_any_congested(void *conges
 	return r;
 }
 
+static unsigned dm_metric(struct bio *bio)
+{
+	return bio->bi_vcnt;
+}
+
 /*-----------------------------------------------------------------
  * An IDR is used to keep track of allocated minor numbers.
  *---------------------------------------------------------------*/

@@ -967,6 +972,7 @@ out:
 
 static struct block_device_operations dm_blk_dops;
 
+#define DEFAULT_THROTTLE_CAPACITY 1000
 /*
  * Allocate and initialise a blank device with a given minor.
  */

@@ -1009,6 +1015,11 @@ static struct mapped_device *alloc_dev(i
 		goto bad1_free_minor;
 
 	md->queue->queuedata = md;
+	md->queue->metric = dm_metric;
+	/* A dm device constructor may change the throttle capacity */
+	atomic_set(&md->queue->available, md->queue->capacity = DEFAULT_THROTTLE_CAPACITY);
+	init_waitqueue_head(&md->queue->throttle_wait);
+
 	md->queue->backing_dev_info.congested_fn = dm_any_congested;
 	md->queue->backing_dev_info.congested_data = md;
 	blk_queue_make_request(md->queue, dm_request);

@@ -1007,6 +1007,13 @@ void bio_endio(struct bio *bio, int erro
 	else if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
 		error = -EIO;
 
+	if (bio->bi_throttle) {
+		struct request_queue *q = bio->bi_queue;
+		bio->bi_throttle = 0; /* or detect multiple endio and err? */
+		atomic_add(bio->bi_throttle, &q->available);
+		wake_up(&q->throttle_wait);
+	}
+
 	if (bio->bi_end_io)
 		bio->bi_end_io(bio, error);
 }

@@ -111,6 +111,9 @@ struct bio {
 	bio_end_io_t		*bi_end_io;
 	atomic_t		bi_cnt;		/* pin count */
 
+	struct request_queue	*bi_queue;	/* for throttling */
+	unsigned		bi_throttle;	/* throttle metric */
+
 	void			*bi_private;
 
 	bio_destructor_t	*bi_destructor;	/* destructor */

@@ -383,6 +383,10 @@ struct request_queue
 	struct work_struct	unplug_work;
 
 	struct backing_dev_info	backing_dev_info;
+	unsigned (*metric)(struct bio *bio);	/* bio throttle metric */
+	wait_queue_head_t	throttle_wait;
+	atomic_t		available;
+	unsigned		capacity;
 
 	/*
 	 * The queue owner gets to use this for whatever they like.

Rather than asking the stack "how much memory will this request consume" you could instead ask "how much memory are you currently using".

ie: on entry to the stack, do

	current->account_block_allocations = 1;
	make_request(...);
	rq->used_memory += current->pages_used_for_block_allocations;

and in the page allocator do

	if (!in_interrupt() && current->account_block_allocations)
		current->pages_used_for_block_allocations++;

and then somehow handle deallocation too ;)

The basic idea being to know in real time how much memory a particular block stack is presently using. Then, on entry to that stack, if the stack's current usage is too high, wait for it to subside.

otoh we already have mechanisms for limiting the number of requests in flight. This is approximately proportional to the amount of memory which was allocated to service those requests. Why not just use that?

Ah, and how do you ensure that you do not deadlock while making this inquiry? Perhaps send a dummy transaction down the pipe? Even so, deadlock is possible, quite evidently so in the real life example I have at hand.

Yours is essentially one of the strategies I had in mind, the other major one being simply to examine the whole stack, which presupposes some as-yet-nonexistant kernel wide method of representing block device stacks in all there glorious possible topology variations.

We do not wait for high block device resource usage to subside before submitting more requests. The improvement you suggest is aimed at automatically determining resource requirements by sampling a running system, rather than requiring a programmer to determine them arduously by hand. Something like automatically determining a workable locking strategy by analyzing running code, wouldn't that be a treat? I will hope for one of those under my tree at Christmas.

More practically, I can see a debug mode implemented along the lines you describe where we automatically detect that a writeout path has violated its covenant as expressed by its throttle_metric.

Two reasons. The minor one is that device mapper bypasses that mechanism (no elevator) and the major one is that number of requests does not map well to the amount of resources consumed. In ddsnap for example, the amount of memory used by the userspace ddsnapd is roughly linear vs the number of pages transferred, not the number of requests.

Yes, as a first order repair. To be done properly I need to express this in terms of the guts of wait_event_*, and get rid of that race, maybe that changes the equation.

It would be nice if threads didn't get stuck in D state here, though _interruptible is probably the wrong idea, we should instead ensure that whatever is being waited on must respond to, e.g., SIGKILL. This at the limits of my scheduler knowledge, l would appreciate a better suggestion. I do detest hang in D state with SIGKILL immunity, which behavior unfortunately does not seem all that rare.

Daniel

It isn't an inquiry - it's a plain old submit_bio() and it runs to completion in the usual fashion.

Thing is, we wouldn't have called it at all if this queue was already over its allocation limit. IOW, we know that it's below its allocation limit, so we know it won't deadlock. Given, of course, reasonably pessimistc error margins.

Which margins can even be observed at runtime: keep a running "max" of this stack's most-ever memory consumption (for a single call), and only submit a bio into it when its current allocation is less than (limit - that-max).

We already have that, I think: blk_run_backing_dev(). One could envisage a similar thing which runs up and down the stack accumulating "how much memory do you need for this request" data, but I think that would be hard to implement and plain dumb.

I don't see any unviability.

oh.

Yeah, one would need to be pretty pessimal. Perhaps unacceptably inaccurate, dunno.

I don't think so. If you're going to sleep in state TASK_INTERRUPTIBLE then you *have* to bale out and return to userspace (or whatever) when signal_pending(). Because schedule() becomes a no-op.

Well.. See

add-lock_page_killable.patch
kernel-add-mutex_lock_killable.patch
vfs-use-mutex_lock_killable-in-vfs_readdir.patch

in 2.6.24-rc4-mm1.

But for now, TASK_UNINTERRUPTIBLE is the honest solution.

OK, I see what you are suggesting. Yes, one could set the inflight limit very low and the reserve very high, and run a bio through the stack (what I meant by "inquiry") to discover the actual usage, then shrink the reserve accordingly. By also running a real bio through the stack we can discover something about the latency. So we would then know roughly how high the inflight limit should be set and how much the memalloc reserve should be increased to handle that particular driver instance.

The big fly in this ointment is that we cannot possibly know that our bio followed the worst case resource consumption path, whereas it is fairly easy (hopefully) for a programmer to determine this statically.

Actually, your mechanism would always have to be operable at runtime, since inserting a new driver while the system is under heavy memory load is a perfectly valid operation and has to be reliable.

Anyway, even if you run a bio through the stack lots of times (insert definition of "lots" here) you still cannot be sure that it has explored the worst case path. To put this in perspective, some of the deadlocks we have hunted down recently have taken days to manifest under artificially high load. It just takes that long to randomly explore a sufficient number of corner cases.

I don't think I quite communicated there. We don't actually have any generic notion of "the block device stack". Device mapper has its own model, md has another model, and other stacking devices may have no model at all, just some through-coded hack. It would be worth fixing this problem as part of an effort to generalize the block IO model and make block devices in general look more like device mapper devices. But that would be a pretty big project, the need for which is not generally recognized.

A small matter of coding. It would be a legendary hack.

Orders of magnitude more reserve would need to be allocated in the case of ddsnap, since bio payload can vary through a big range, which is expected to get bigger as time goes by. So the few lines of extra code and the extra bio field needed to get a better fit is well worth the effort, or even indispensable.

Thanks for clearing that up. No I did not know that important fact about signal handling, and now it is obvious why it must be so. Got a pointer to a doc for that, and other similar scheduler factoids that must surely be known to those in the know?

Right, and easy too. I think we ought to be able to do a fairly good job of making the in-flight items go away in the presence of a signal, thus ending the uninterruptible sleep reliably. Probably.

Regards,

Daniel

nonono...

Consider an example.

- We a-priori decide to limit a particular stack's peak memory usage to 1MB

- We empirically discover that the maximum amount of memory which is allocated by that stack on behalf of a single BIO is 16kb. (ie: that's the most it has ever used for a single BIO).

- Now, we refuse to feed any more BIOs into the stack when its instantaneous memory usage exceeds (1MB - 16kb).

Of course, the _average_ memory-per-BIO is much less than 16kb. So there are a lot of BIOs in flight - probably hundreds, but a minimum of 63.

There is a teeny so-small-it-doesn't-matter chance that the stack will exceed the 1MB limit. If it happens to be at its (1MB-16kb) limit and all the memory in the machine is AWOL and then someone throws a never-seen-before twirly BIO at it. Not worth worrying about, surely.

There is only one problem I can see with this. With network block IO, some memory will be consumed upon IO completion. We need to make sure we reserve (number of in flight BIOs * maximum amount of memory consumed upon IO completion) memory, in addition to the memory you're accounting in your example above.

All Rights Reversed

hm, yeah, drat.

What we could do is

- in do_IRQ(): set up a per-cpu pointer to some counter which corresponds to this IRQ.

- in the page allocator, if in_irq(), retrieve that per-cpu pointer and increment the counter.

- in the network block-io stack we can now look at the number of interrupts, number of packets, size of packets and amount of memory allocated and work out the max amount of memory which needs to be allocated for each frame.

That's all rather handwavy and misses a lot of details and might be inaccurate too. Probably sufficient to just work out by hand the amount of memory which the network stack will need to allocate. I expect it'll be two pages..

Doesn't Peter Zijlstra's patch series take care of all those nasty details already?

All Rights Reversed

Ah, this is the piece I was missing.

And printk a warning, because the programmer's static analysis was wrong.

And progress is being made, everybody is happy.

OK, I see where you are going. The programmer statically determines a total reservation for the stack, which is big enough that progress is guaranteed. We then throttle based on actual memory consumption, and essentially use a heuristic to decide when we are near the upper limit. Workable I think, but...

The main idea, current->pages_used_for_block_allocations++ is valid only in direct call context. If a daemon needs to allocate memory on behalf of the IO transfer (not unusual) it won't get accounted, which is actually the central issue in this whole class of deadlocks. Any idea how to extend the accounting idea to all tasks involved in a particular block device stack?

Regards,

Daniel

SMOP, I'd have thought. As long as each piece of code which handles data for this stack knows that it's handling data for that stack it should be able to account its memory allocations.

The tricky part will be networking allocations because a NIC can of course handle data for all sorts of consumers. But I expect this can be greatly simplified with a few heuristics - work out how much memory your typical networking stack will allocate for a frame and tack that onto the total. Couple of pages worst case..

Agreed, which I realized as soon as the post was one minute old. Sure, each helper for the device registers as a helper which puts a pointer in the task struct, which points to the accounting info so only one new field in task struct. The more I ponder, the more doable it seems.

Don't forget that we do not actually have a usable notion of "block device stack" yet. Perhaps you are just assuming that is easy/imminent?

Actually, the same pattern that Peter and I developed for handling network deadlock extends to this accounting concept. As you say, it's a SMOP.

Daniel

The problem is that you (a) may or may not know just how bad a worst case can be, and (b) may block unnecessarily by being pessimistic.

The dummy transaction would be nice, but it would be perfect if you could send the real transaction down with a max memory limit and a flag, have each level check and decrement the max by what's actually needed, and then return some pass/fail status for that particular transaction. Clearly every level in the stack would have to know how to do that. It would seem that once excess memory use was detected the transaction could be failed without deadlock.

Bill Davidsen <[email protected]>
  "We have more to fear from the bungling of the incompetent than from
the machinations of the wicked."  - from Slashdot

True, but after a quick introspect I realized that that issue (it's really a single issue) is not any worse than the way I planned to wave my hands at the issue of programmers constructing their metrics wrongly and thereby breaking the throttling assumptions.

Which is to say that I am now entirely convince by Andrew's argument and am prepardc to reroll the patch along the lines he suggests. The result will be somewhat bigger. Only a minor change is required to the main mechanism: we will now account things entirely in units of pages instead of abstract units, eliminating a whole class of things to go wrong. I like that. Accounting variables get shifted to a new home, maybe. Must try a few ideas and see what works.

Anyway, the key idea is that task struct will gain a field pointing at a handle for the "block device stack", whatever that is (this is sure to evolve over time) and alloc_pages will know how to account pages to that object. The submit_bio and bio->endio bits change hardly at all.

The runner up key idea is that we will gain a notion of "block device stack" (or block stack for short, so that we may implement block stackers) which for the time being will simply be Device Mapper's notion of device stack, however many warts that may have. It's there now and we use it for ddsnap.

The other player in this is Peterz's swap over network use case, which does not involve a device mapper device. Maybe it should? Otherwise we will need a variant notion of block device stack, and the two threads of work should merge eventually. There is little harm in starting this effort in two different places, quite the contrary.

In the meantime we do have a strategy that works, posted at the head of this thread, for anybody who needs it now.

The function of the dummy transaction will be to establish roughly what kind of footprint for a single transaction we see on that block IO path. Then we will make the reservation _hugely_ greater than that, to accommodate 1000 or so of those. A transaction blocks if it actually tries to use more than that. We close the "many partial submissions all deadlock together" hole by ... <insert handwaving here>. Can't go wrong, right?

Agreed that drivers should pay special attention to our dummy transaction and try to use the maximum possible resources when they see one. More handwaving, but this is progress.

Regards,

Daniel

Perhaps all we need to track is the outermost point?

submit_bio(...)

{
	bool remove_the_rq = false;

	...
	if (current->the_rq == NULL) {
		current->the_rq = rq;
		remove_the_rq = true;
	}
	...
	if (remove_the_rq)
		current->the_rq = NULL;
}

?

The parent patch already has that crucial property in a simple say, see

       if (q && q->metric && !bio->bi_queue) {
               bio->bi_queue = q;

Regards,

Daniel

Hi Andrew,

Unfortunately, I agreed with your suggestion too hastily. Not only
would it be complex to implement, It does not work. It took me several days to put my finger on exactly why. Here it is in a nutshell:
resources may be consumed _after_ the gatekeeper runs the "go, no go" throttling decision. To illustrate, throw 10,000 bios simultaneously at a block stack that is supposed to allow only about 1,000 in flight at a time. If the block stack allocates memory somewhat late in its servicing scheme (for example, when it sends a network message) then it is possible that no actual resource consumption will have taken place before all 10,000 bios are allowed past the gate keeper, and deadlock is sure to occur sooner or later.

In general, we must throttle against the maximum requirement of inflight bios rather than against the measured consumption. This achieves the invariant I have touted, namely that memory consumption on the block writeout path must be bounded. We could therefore possibly use your suggestion or something resembling it to implement a debug check that the programmer did in fact do their bounds arithmetic correctly, but it is not useful for enforcing the bound itself.

In case that coffin needs more nails in it, consider that we would not only need to account page allocations, but frees as well. So what tells us that a page has returned to the reserve pool? Oops, tough one. The page may have been returned to a slab and thus not actually freed, though it remains available for satisfying new bio transactions. Because of such caching, your algorithm would quickly lose track of available resources and grind to a halt.

Never mind that keeping track of page frees is a nasty problem in
itself. They can occur in interrupt context, so forget the current->
idea. Even keeping track of page allocations for bio transactions in normal context will be a mess, and that is the easy part. I can just imagine the code attempting to implement this approach acreting into a monster that gets confusingly close to working without ever actually getting there.

We do have a simple, elegant solution posted at the head of this thread, which is known to work.

Regards,

Daniel

Such paranoia. Yes, the patch was compiled. Yes, the warning was
slipped through. No, it is not substantive, and in fact was removed from another branch of our tree already.

Ignoring the rhetoric, apparently you missed the line:

+ if (q && q->metric && !bio->bi_queue) {

The prevents any reference ti bi_bdev after the intial call to generic_make_request. Thanks to Evgeniy for pointing out the need for this measure on the last go round.

"So broken" is a gross exaggeration. Substantive comments welcome.

Daniel

Which saves the initial target, for ease of accounting at end io time - that's not the point. What happens when ->make_request_fn() changes bio->bi_bdev and returns 1, causing another iteration of the __generic_make_request() loop? 'q' is no longer the valid target, bdev_get_queue(bio->bi_bdev) is.

Or you could try and make an effort to understand the comment instead of just glancing over it.

Jens Axboe

What happens on the second iteration of a recursive submission loop is exactly nothing, as is right and proper. The throttling has already been done, and all the state necessary to perform the unthrottle was recorded in the bio. Everything seems to be in order there, and the algorithm does indeed perform its function as designed, though to be sure we have not tested it on -mm branch, only on mainline.

Regards,

Daniel

OK, let me get the neon out then. This has nothing to do with throttling, I thought I made it clear that I get why you store the origin queue in ->bi_queue. I'm concerned with the workings of redirecting a bio. Previously we looked up the queue associated with bio->bi_bdev inside the loop in __generic_make_request(), as is REQUIRED to correctly locate a DIFFERENT queue if bio->bi_bdev has been changed to point somewhere else.

Clear?

Jens Axboe

Rhetoric aside, again.

We are only interested in throttling against the bio->bi_bdev that was stored in the bio at the time of the call to generic_make_request, why should we care about the redirected value?

Regards,

Daniel

Let me repeat - this has nothing to do with throttling! You are breaking the bio redirection by killing that bdev_get_queue() in the __generic_make_request().

I honestly don't know how to make this any clearer than I already did above. Sleep on it.

Jens Axboe

Sure you do, you could cut out the rhetoric and save lots of bandwidth thereby.

Yes, the q = bdev_get_queue(bio->bi_bdev) needs to be repeated inside the submission loop, that was a flaw, thanks for the catch.

Regards,

Daniel

@@ -3210,9 +3210,9 @@ static inline int bio_check_eod(struct b
  */
 static inline void __generic_make_request(struct bio *bio)
 {
-	struct request_queue *q;
+	struct request_queue *q = bdev_get_queue(bio->bi_bdev);
 	sector_t old_sector;
-	int ret, nr_sectors = bio_sectors(bio);
+	int nr_sectors = bio_sectors(bio);
 	dev_t old_dev;
 	int err = -EIO;
 

@@ -3221,6 +3221,13 @@ static inline void __generic_make_reques
 	if (bio_check_eod(bio, nr_sectors))
 		goto end_io;
 
+	if (q && q->metric && !bio->bi_queue) {
+		int need = bio->bi_throttle = q->metric(bio);
+		bio->bi_queue = q;
+		/* FIXME: potential race if atomic_sub is called in the middle of condition check */
+		wait_event(q->throttle_wait, atomic_read(&q->available) >= need);
+		atomic_sub(need, &q->available);
+	}
 	/*
 	 * Resolve the mapping until finished. (drivers are
 	 * still free to implement/resolve their own stacking

@@ -3231,10 +3238,9 @@ static inline void __generic_make_reques
 	 */
 	old_sector = -1;
 	old_dev = 0;
-	do {
+	while (1) {
 		char b[BDEVNAME_SIZE];
 
-		q = bdev_get_queue(bio->bi_bdev);
 		if (!q) {
 			printk(KERN_ERR
 			       "generic_make_request: Trying to access "

@@ -3282,8 +3288,10 @@ end_io:
 			goto end_io;
 		}
 
-		ret = q->make_request_fn(q, bio);
-	} while (ret);
+		if (!q->make_request_fn(q, bio))
+			return;
+		q = bdev_get_queue(bio->bi_bdev);
+	}
 }
 
 /*

@@ -889,6 +889,11 @@ static int dm_any_congested(void *conges
 	return r;
 }
 
+static unsigned dm_metric(struct bio *bio)
+{
+	return bio->bi_vcnt;
+}
+
 /*-----------------------------------------------------------------
  * An IDR is used to keep track of allocated minor numbers.
  *---------------------------------------------------------------*/

@@ -967,6 +972,7 @@ out:
 
 static struct block_device_operations dm_blk_dops;
 
+#define DEFAULT_THROTTLE_CAPACITY 1000
 /*
  * Allocate and initialise a blank device with a given minor.
  */

@@ -1009,6 +1015,11 @@ static struct mapped_device *alloc_dev(i
 		goto bad1_free_minor;
 
 	md->queue->queuedata = md;
+	md->queue->metric = dm_metric;
+	/* A dm device constructor may change the throttle capacity */
+	atomic_set(&md->queue->available, md->queue->capacity = DEFAULT_THROTTLE_CAPACITY);
+	init_waitqueue_head(&md->queue->throttle_wait);
+
 	md->queue->backing_dev_info.congested_fn = dm_any_congested;
 	md->queue->backing_dev_info.congested_data = md;
 	blk_queue_make_request(md->queue, dm_request);

@@ -1007,6 +1007,13 @@ void bio_endio(struct bio *bio, int erro
 	else if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
 		error = -EIO;
 
+	if (bio->bi_throttle) {
+		struct request_queue *q = bio->bi_queue;
+		bio->bi_throttle = 0; /* or detect multiple endio and err? */
+		atomic_add(bio->bi_throttle, &q->available);
+		wake_up(&q->throttle_wait);
+	}
+
 	if (bio->bi_end_io)
 		bio->bi_end_io(bio, error);
 }

@@ -111,6 +111,9 @@ struct bio {
 	bio_end_io_t		*bi_end_io;
 	atomic_t		bi_cnt;		/* pin count */
 
+	struct request_queue	*bi_queue;	/* for throttling */
+	unsigned		bi_throttle;	/* throttle metric */
+
 	void			*bi_private;
 
 	bio_destructor_t	*bi_destructor;	/* destructor */

@@ -383,6 +383,10 @@ struct request_queue
 	struct work_struct	unplug_work;
 
 	struct backing_dev_info	backing_dev_info;
+	unsigned (*metric)(struct bio *bio);	/* bio throttle metric */
+	wait_queue_head_t	throttle_wait;
+	atomic_t		available;
+	unsigned		capacity;
 
 	/*
 	 * The queue owner gets to use this for whatever they like.

I spent 3 mail explaining it as clearly as I could. So you're welcome for the review and the reminder of why it's impossible to have a normal conversation with you.

Precisely. So forgive me for thinking this patch hasn't seen very varied testing, that's 2 errors (one simple, one bad - broken was NOT a gross exageration, thanks) in very few lines.

Jens Axboe

See the [RFC]? If I had meant Request For Flaming, I would have written that. Thankyou for the catch.

Regards,

Daniel

The same mail that contained this part, copied verbatim:

"Let me close with perhaps the most relevant remarks: the attached code
has been in heavy testing and in production for months now. Thus there is nothing theoretical when I say it works, and the patch speaks for itself in terms of obvious correctness."

We must have differing opinions on what obvious correctness is.

Future replies to /dev/null, please.

Jens Axboe

That is quite correct, even without the redirect the code passed all our tests. Remember, we were testing for deadlock, not every possible block IO configuration.

Yes we do. You appear to have missed the plot entirely. I suppose I should remind you: this is about deadlock in _your_ subsystem that has been creating bug reports for years. Block writeout deadlock. Caused by a deficiency in _your_ subsystem.

Got a plan? Or does endless, pointless flaming feel more like progress to you?

And round we go.

And I may remind you that I have participated in this discussion before and made my points clear there. I suppose I should remind you how the development process works? Just because I happen to maintain some piece of code does not mean I'm under some sort of contractual obligation to fix and write new code for users. I'll happily review patches and integrate stuff I agree with, as I have been doing for years. This bug may seem really important to you - guess what, that's the normal nature of a bug to a user that runs into them.

Please, I'm not flaming you. I reviewed your code and pointed out errors, which was followed by lots of hand waving on your side instead of just sitting down and reading/fixing the bug.

Jens Axboe

Well that is indeed more civil language, if somewhat revisionist, since I distinctly remember being flamed by you from the word go. Once again, thankyou for the catch. A fairly trivial oversight that would have been caught sooner of later. As for the typedef thing, that was just a spelling flame, admit it.

Truly, the way you were yelling I thought you had picked up a fundamental flaw instead of a simple misplaced line of code.

Now about that block writeout deadlock... it doesn't just affect my code, it basically breaks Linux as a storage platform, among other things.

Regards,

Daniel

As written in other similar threads in the past in which you also participated, I still of the opinion that this is a vm issue and should be solved as such.

As to the patch in question "fixing" it in the block layer, it's a fairly simple work around and I'm not totally against it. If you get rid of the ->bi_throttle stuff and just do sanity checks on the count, then we could look at getting some testing done.

Jens Axboe

The problem is solved. The main cornerstone of the solution is bio throttling, simply because the resources in question are consumed by bio transactions.

Testing is already progressing fine without you, thankyou. If you do want to participate, welcome, otherwise it is not a problem. Thanks for picking up that bug yesterday.

Daniel

... because too much is pushed out. This isn't a mathematica problem, there's more than one solution to this problem. Throttling the bio count is one.

Here we go again, thanks for picking up your jerky attitude again. I'm trying to suggest a way to get the patch in a state to be included, but apparently you are not interested. With 3 bugs so far exposed in your really short patch, seems you should take all the testing help you can get.

For what it's worth, your behind the doors testing is worth basically nothing. It's already been shown that your test coverage wasn't very wide, if this patch/idea is to have any hope of proceeding further you need user testing. Period.

Stop cc'ing or replying, what little interest I had is totally gone.

Jens Axboe

And nobody has been able to find another. Funny that. In fact, every solution proposed so far has implicitly required the writeout traffic to be throttled, even if that throttling was not part of the patch. Without throttling, deadlock. Simple as that.

Oh, sorry I missed that olive branch on first read. Getting rid of those 8 bytes that bother you requires an extensive rethink of bio handling in order to make some fields that are not now constant, constant or at least restored on change. Which would be a good thing in itself. There are lots of good improvements that can be made to this subsystem along those lines.

But that is properly a separate project. Quite some time will be needed to get it right, and should I mention it, everybody needs to be on the same page or the work will never start. It is therefore a theoretical solution. We have a practical, tested solution, here and now, and it is short enough to be understood, unlike any of the previous attempts.

Your argument seems to be that adding 8 bytes to struct bio turns this beautiful swan into an ugly duck. Actually, because the throttling reduces the number of bios in flight in a busy system, total memory use is reduced. When the system is not busy, there are few bios hanging around so that is not a problem either. Nice, hmm?

Daniel

There you go, Jens, service with a smile.

Regards,

Daniel

@@ -3210,7 +3210,7 @@ static inline int bio_check_eod(struct b
  */
 static inline void __generic_make_request(struct bio *bio)
 {
-	struct request_queue *q;
+	struct request_queue *q = bdev_get_queue(bio->bi_bdev);
 	sector_t old_sector;
 	int ret, nr_sectors = bio_sectors(bio);
 	dev_t old_dev;

@@ -3221,6 +3221,13 @@ static inline void __generic_make_reques
 	if (bio_check_eod(bio, nr_sectors))
 		goto end_io;
 
+	if (q && q->metric && !bio->bi_queue) {
+		int need = bio->bi_throttle = q->metric(bio);
+		bio->bi_queue = q;
+		/* FIXME: potential race if atomic_sub is called in the middle of condition check */
+		wait_event_interruptible(q->throttle_wait, atomic_read(&q->available) >= need);
+		atomic_sub(need, &q->available);
+	}
 	/*
 	 * Resolve the mapping until finished. (drivers are
 	 * still free to implement/resolve their own stacking

@@ -3234,7 +3241,6 @@ static inline void __generic_make_reques
 	do {
 		char b[BDEVNAME_SIZE];
 
-		q = bdev_get_queue(bio->bi_bdev);
 		if (!q) {
 			printk(KERN_ERR
 			       "generic_make_request: Trying to access "

@@ -889,6 +889,11 @@ static int dm_any_congested(void *conges
 	return r;
 }
 
+static unsigned dm_metric(struct bio *bio)
+{
+	return bio->bi_vcnt;
+}
+
 /*-----------------------------------------------------------------
  * An IDR is used to keep track of allocated minor numbers.
  *---------------------------------------------------------------*/

@@ -967,6 +972,7 @@ out:
 
 static struct block_device_operations dm_blk_dops;
 
+#define DEFAULT_THROTTLE_CAPACITY 1000
 /*
  * Allocate and initialise a blank device with a given minor.
  */

@@ -1009,6 +1015,11 @@ static struct mapped_device *alloc_dev(i
 		goto bad1_free_minor;
 
 	md->queue->queuedata = md;
+	md->queue->metric = dm_metric;
+	/* A dm device constructor may change the throttle capacity */
+	atomic_set(&md->queue->available, md->queue->capacity = DEFAULT_THROTTLE_CAPACITY);
+	init_waitqueue_head(&md->queue->throttle_wait);
+
 	md->queue->backing_dev_info.congested_fn = dm_any_congested;
 	md->queue->backing_dev_info.congested_data = md;
 	blk_queue_make_request(md->queue, dm_request);

@@ -1007,6 +1007,13 @@ void bio_endio(struct bio *bio, int erro
 	else if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
 		error = -EIO;
 
+	if (bio->bi_throttle) {
+		struct request_queue *q = bio->bi_queue;
+		bio->bi_throttle = 0; /* or detect multiple endio and err? */
+		atomic_add(bio->bi_throttle, &q->available);
+		wake_up(&q->throttle_wait);
+	}
+
 	if (bio->bi_end_io)
 		bio->bi_end_io(bio, error);
 }

@@ -111,6 +111,9 @@ struct bio {
 	bio_end_io_t		*bi_end_io;
 	atomic_t		bi_cnt;		/* pin count */
 
+	struct request_queue	*bi_queue;	/* for throttling */
+	unsigned		bi_throttle;	/* throttle metric */
+
 	void			*bi_private;
 
 	bio_destructor_t	*bi_destructor;	/* destructor */

@@ -383,6 +383,10 @@ struct request_queue
 	struct work_struct	unplug_work;
 
 	struct backing_dev_info	backing_dev_info;
+	unsigned (*metric)(struct bio *bio);	/* bio throttle metric */
+	wait_queue_head_t	throttle_wait;
+	atomic_t		available;
+	unsigned		capacity;
 
 	/*
 	 * The queue owner gets to use this for whatever they like.

Hey, Daniel,

I'm just getting around to looking at this. One thing jumped out at me:

Hi,

Heh, well, that's ok as long as bio->bi_vcnt is set to zero and I think we have some md raid drivers do just that... ;-)

Pekka

Hi Jon,

Don't you know? These days we optimize all our code for modern processors with tunnelling instructions and metaphysical cache. On such processors, setting a register to zero does not entirely destroy all the data that used to be in the register, so subsequent instructions can make further use of the overwritten data by reconstructing it from remnants of bits left attached to the edges of the register.

Um, yeah, that's it.

Actually, I fat-fingered it in the merge to -mm. Thanks for the catch, corrected patch attached.

The offending line isn't even a functional part of the algorithm, it is just supposed to defend against the possibility that, somehow, ->bi_endio gets called multiple times. Probably it should really be something like:

		BUG_ON(bio->bi_throttle == -1);
		if (bio->bi_throttle) {
			...
			bio->bi_throttle = -1;

Or perhaps we should just rely on nobody ever making that mistake and let somebody else catch it if it does.

Regards,

Daniel

@@ -3210,9 +3210,9 @@ static inline int bio_check_eod(struct b
  */
 static inline void __generic_make_request(struct bio *bio)
 {
-	struct request_queue *q;
+	struct request_queue *q = bdev_get_queue(bio->bi_bdev);
 	sector_t old_sector;
-	int ret, nr_sectors = bio_sectors(bio);
+	int nr_sectors = bio_sectors(bio);
 	dev_t old_dev;
 	int err = -EIO;
 

@@ -3221,6 +3221,13 @@ static inline void __generic_make_reques
 	if (bio_check_eod(bio, nr_sectors))
 		goto end_io;
 
+	if (q && q->metric && !bio->bi_queue) {
+		int need = bio->bi_throttle = q->metric(bio);
+		bio->bi_queue = q;
+		/* FIXME: potential race if atomic_sub is called in the middle of condition check */
+		wait_event(q->throttle_wait, atomic_read(&q->available) >= need);
+		atomic_sub(need, &q->available);
+	}
 	/*
 	 * Resolve the mapping until finished. (drivers are
 	 * still free to implement/resolve their own stacking

@@ -3231,10 +3238,9 @@ static inline void __generic_make_reques
 	 */
 	old_sector = -1;
 	old_dev = 0;
-	do {
+	while (1) {
 		char b[BDEVNAME_SIZE];
 
-		q = bdev_get_queue(bio->bi_bdev);
 		if (!q) {
 			printk(KERN_ERR
 			       "generic_make_request: Trying to access "

@@ -3282,8 +3288,10 @@ end_io:
 			goto end_io;
 		}
 
-		ret = q->make_request_fn(q, bio);
-	} while (ret);
+		if (!q->make_request_fn(q, bio))
+			return;
+		q = bdev_get_queue(bio->bi_bdev);
+	}
 }
 
 /*

@@ -889,6 +889,11 @@ static int dm_any_congested(void *conges
 	return r;
 }
 
+static unsigned dm_metric(struct bio *bio)
+{
+	return bio->bi_vcnt;
+}
+
 /*-----------------------------------------------------------------
  * An IDR is used to keep track of allocated minor numbers.
  *---------------------------------------------------------------*/

@@ -967,6 +972,7 @@ out:
 
 static struct block_device_operations dm_blk_dops;
 
+#define DEFAULT_THROTTLE_CAPACITY 1000
 /*
  * Allocate and initialise a blank device with a given minor.
  */

@@ -1009,6 +1015,11 @@ static struct mapped_device *alloc_dev(i
 		goto bad1_free_minor;
 
 	md->queue->queuedata = md;
+	md->queue->metric = dm_metric;
+	/* A dm device constructor may change the throttle capacity */
+	atomic_set(&md->queue->available, md->queue->capacity = DEFAULT_THROTTLE_CAPACITY);
+	init_waitqueue_head(&md->queue->throttle_wait);
+
 	md->queue->backing_dev_info.congested_fn = dm_any_congested;
 	md->queue->backing_dev_info.congested_data = md;
 	blk_queue_make_request(md->queue, dm_request);

@@ -1007,6 +1007,13 @@ void bio_endio(struct bio *bio, int erro
 	else if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
 		error = -EIO;
 
+	if (bio->bi_throttle) {
+		struct request_queue *q = bio->bi_queue;
+		atomic_add(bio->bi_throttle, &q->available);
+		bio->bi_throttle = 0; /* or detect multiple endio and err? */
+		wake_up(&q->throttle_wait);
+	}
+
 	if (bio->bi_end_io)
 		bio->bi_end_io(bio, error);
 }

@@ -111,6 +111,9 @@ struct bio {
 	bio_end_io_t		*bi_end_io;
 	atomic_t		bi_cnt;		/* pin count */
 
+	struct request_queue	*bi_queue;	/* for throttling */
+	unsigned		bi_throttle;	/* throttle metric */
+
 	void			*bi_private;
 
 	bio_destructor_t	*bi_destructor;	/* destructor */

@@ -383,6 +383,10 @@ struct request_queue
 	struct work_struct	unplug_work;
 
 	struct backing_dev_info	backing_dev_info;
+	unsigned (*metric)(struct bio *bio);	/* bio throttle metric */
+	wait_queue_head_t	throttle_wait;
+	atomic_t		available;
+	unsigned		capacity;
 
 	/*
 	 * The queue owner gets to use this for whatever they like.