Re: [PATCH v3 4/7] swiotlb: if swiotlb is full, fall back to a transient memory pool

[Petr Tesařík]

On Sat, 8 Jul 2023 15:18:32 +0000
"Michael Kelley (LINUX)" <mikelley@xxxxxxxxxxxxx> wrote:

> From: Petr Tesařík <petr@xxxxxxxxxxx> Sent: Friday, July 7, 2023 3:22 AM
> > 
> > On Fri, 7 Jul 2023 10:29:00 +0100
> > Greg Kroah-Hartman <gregkh@xxxxxxxxxxxxxxxxxxx> wrote:
> >   
> > > On Thu, Jul 06, 2023 at 02:22:50PM +0000, Michael Kelley (LINUX) wrote:  
> > > > From: Greg Kroah-Hartman <gregkh@xxxxxxxxxxxxxxxxxxx> Sent: Thursday, July 6,  
> > 2023 1:07 AM  
> > > > >
> > > > > On Thu, Jul 06, 2023 at 03:50:55AM +0000, Michael Kelley (LINUX) wrote:  
> > > > > > From: Petr Tesarik <petrtesarik@xxxxxxxxxxxxxxx> Sent: Tuesday, June 27, 2023  
> > > > > 2:54 AM  
> > > > > > >
> > > > > > > Try to allocate a transient memory pool if no suitable slots can be found,
> > > > > > > except when allocating from a restricted pool. The transient pool is just
> > > > > > > enough big for this one bounce buffer. It is inserted into a per-device
> > > > > > > list of transient memory pools, and it is freed again when the bounce
> > > > > > > buffer is unmapped.
> > > > > > >
> > > > > > > Transient memory pools are kept in an RCU list. A memory barrier is
> > > > > > > required after adding a new entry, because any address within a transient
> > > > > > > buffer must be immediately recognized as belonging to the SWIOTLB, even if
> > > > > > > it is passed to another CPU.
> > > > > > >
> > > > > > > Deletion does not require any synchronization beyond RCU ordering
> > > > > > > guarantees. After a buffer is unmapped, its physical addresses may no
> > > > > > > longer be passed to the DMA API, so the memory range of the corresponding
> > > > > > > stale entry in the RCU list never matches. If the memory range gets
> > > > > > > allocated again, then it happens only after a RCU quiescent state.
> > > > > > >
> > > > > > > Since bounce buffers can now be allocated from different pools, add a
> > > > > > > parameter to swiotlb_alloc_pool() to let the caller know which memory pool
> > > > > > > is used. Add swiotlb_find_pool() to find the memory pool corresponding to
> > > > > > > an address. This function is now also used by is_swiotlb_buffer(), because
> > > > > > > a simple boundary check is no longer sufficient.
> > > > > > >
> > > > > > > The logic in swiotlb_alloc_tlb() is taken from __dma_direct_alloc_pages(),
> > > > > > > simplified and enhanced to use coherent memory pools if needed.
> > > > > > >
> > > > > > > Note that this is not the most efficient way to provide a bounce buffer,
> > > > > > > but when a DMA buffer can't be mapped, something may (and will) actually
> > > > > > > break. At that point it is better to make an allocation, even if it may be
> > > > > > > an expensive operation.  
> > > > > >
> > > > > > I continue to think about swiotlb memory management from the standpoint
> > > > > > of CoCo VMs that may be quite large with high network and storage loads.
> > > > > > These VMs are often running mission-critical workloads that can't tolerate
> > > > > > a bounce buffer allocation failure.  To prevent such failures, the swiotlb
> > > > > > memory size must be overly large, which wastes memory.  
> > > > >
> > > > > If "mission critical workloads" are in a vm that allowes overcommit and
> > > > > no control over other vms in that same system, then you have worse
> > > > > problems, sorry.
> > > > >
> > > > > Just don't do that.
> > > > >  
> > > >
> > > > No, the cases I'm concerned about don't involve memory overcommit.
> > > >
> > > > CoCo VMs must use swiotlb bounce buffers to do DMA I/O.  Current swiotlb
> > > > code in the Linux guest allocates a configurable, but fixed, amount of guest
> > > > memory at boot time for this purpose.  But it's hard to know how much
> > > > swiotlb bounce buffer memory will be needed to handle peak I/O loads.
> > > > This patch set does dynamic allocation of swiotlb bounce buffer memory,
> > > > which can help avoid needing to configure an overly large fixed size at boot.  
> > >
> > > But, as you point out, memory allocation can fail at runtime, so how can
> > > you "guarantee" that this will work properly anymore if you are going to
> > > make it dynamic?  
> > 
> > In general, there is no guarantee, of course, because bounce buffers
> > may be requested from interrupt context. I believe Michael is looking
> > for the SWIOTLB_MAY_SLEEP flag that was introduced in my v2 series, so
> > new pools can be allocated with GFP_KERNEL instead of GFP_NOWAIT if
> > possible, and then there is no need to dip into the coherent pool.
> > 
> > Well, I have deliberately removed all complexities from my v3 series,
> > but I have more WIP local topic branches in my local repo:
> > 
> > - allow blocking allocations if possible
> > - allocate a new pool before existing pools are full
> > - free unused memory pools
> > 
> > I can make a bigger series, or I can send another series as RFC if this
> > is desired. ATM I don't feel confident enough that my v3 series will be
> > accepted without major changes, so I haven't invested time into
> > finalizing the other topic branches.
> > 
> > @Michael: If you know that my plan is to introduce blocking allocations
> > with a follow-up patch series, is the present approach acceptable?
> >   
> 
> Yes, I think the present approach is acceptable as a first step.  But
> let me elaborate a bit on my thinking.
> 
> I was originally wondering if it is possible for swiotlb_map() to detect
> whether it is called from a context that allows sleeping, without the use
> of SWIOTLB_MAY_SLEEP.   This would get the benefits without having to
> explicitly update drivers to add the flag.  But maybe that's too risky.

This is a recurring topic and it has been discussed several times in
the mailing lists. If you ask me, the best answer is this one by Andrew
Morton, albeit a bit dated:

https://lore.kernel.org/lkml/20080320201723.b87b3732.akpm@xxxxxxxxxxxxxxxxxxxx/

> For
> the CoCo VM scenario that I'm most interested in, being a VM implicitly
> reduces the set of drivers that are being used, and so it's not that hard
> to add the flag in the key drivers that generate most of the bounce
> buffer traffic.

Yes, that's my thinking as well.

> Then I was thinking about a slightly different usage for the flag than what
> you implemented in v2 of the series.   In the case where swiotlb_map()
> can't allocate slots because of the swiotlb pool being full (or mostly full),
> kick the background thread (if it is not already awake) to allocate a
> dynamic pool and grow the total size of the swiotlb.  Then if
> SWIOTLB_MAY_SLEEP is *not* set, allocate a transient pool just as you
> have implemented in this v3 of the series.  But if SWIOTLB_MAY_SLEEP
> *is* set, swiotlb_map() should sleep until the background thread has
> completed the memory allocation and grown the size of the swiotlb.
> After the sleep, retry the slot allocation.  Maybe what I'm describing
> is what you mean by "allow blocking allocations".  :-)

Not really, but I like the idea. After all, the only reason to have
transient pools is when something is needed immediately while the
background allocation is running.

> This approach effectively throttles incoming swiotlb requests when space
> is exhausted, and gives the dynamic sizing mechanism a chance to catch
> up in an efficient fashion.  Limiting transient pools to requests that can't
> sleep will reduce the likelihood of exhausting the coherent memory
> pools.  And as you mentioned above, kicking the background thread at the
> 90% full mark (or some such heuristic) also helps the dynamic sizing
> mechanism keep up with demand.

FWIW I did some testing, and my systems were not able to survive a
sudden I/O peak without transient pools, no matter how low I set the
threshold for kicking a background. OTOH I always tested with the
smallest possible SWIOTLB (256 KiB * rounded up number of CPUs, e.g. 16
MiB on my VM with 48 CPUs). Other sizes may lead to different results.

As a matter of fact, the size of the initial SWIOTLB memory pool and the
size(s) of additional pool(s) sound like interesting tunable parameters
that I haven't explored in depth yet.

Petr T