On 15/01/2021 19:21, Robin Murphy wrote:
It would be good to understand why the rcache doesn't stabilize. Could be
a bug, or just need some tuning
In strict mode, if a driver does Alloc-Free-Alloc and the first alloc
misses the rcache, the second allocation hits it. The same sequence in
non-strict mode misses the cache twice, because the IOVA is added to the
flush queue on Free.
So rather than AFAFAF.. we get AAA..FFF.., only once the fq_timer
triggers
or the FQ is full.
Sounds right
Interestingly the FQ size is 2x IOVA_MAG_SIZE, so we
could allocate 2 magazines worth of fresh IOVAs before alloc starts
hitting the cache. If a job allocates more than that, some magazines are
going to the depot, and with multi-CPU jobs those will get used on other
CPUs during the next alloc bursts, causing the progressive increase in
rcache consumption. I wonder if setting IOVA_MAG_SIZE > IOVA_FQ_SIZE
helps
reuse of IOVAs?
Looking back through the lore history, I don't know where the
IOVA_FQ_SIZE = 256 came from. I guess it's size of 2x IOVA_MAG_SIZE (1x
for loaded and 1x for prev) for the reason you mention.
Then again I haven't worked out the details, might be entirely wrong.
I'll
have another look next week.
cheers
I did start digging into the data (thanks for that!) before Christmas,
but between being generally frazzled and trying to remember how to write
Perl to massage the numbers out of the log dump I never got round to
responding, sorry.
As you may have seen:
https://raw.githubusercontent.com/hisilicon/kernel-dev/064c4dc8869b3f2ad07edffceafde0b129f276b0/lsi3008_dmesg
I had to change some block configs via sysfs to ever get IOVA locations
for size > 0. And even then, I still got none bigger than
IOVA_RANGE_CACHE_MAX_SIZE.
Note: For a log like:
[13175.361915] print_iova2 iova_allocs(=5000000 ... too_big=47036
47036 is number of IOVA size > IOVA_RANGE_CACHE_MAX_SIZE, in case it was
not clear.
And I never hit the critical point of a depot bin filling, but it may
just take even longer.
However with IOVA size = 0 always occurring, then I noticed that the
depot size = 0 bin fills up relatively quickly. As such, I am now
slightly skeptical of the approach I have taken here, i.e purge the
whole rcache.
The partial thoughts that I can recall right now are firstly that the
total numbers of IOVAs are actually pretty meaningless, it really needs
to be broken down by size (that's where my Perl-hacking stalled...);
secondly that the pattern is far more than just a steady increase - the
CPU rcache count looks to be heading asymptotically towards ~65K IOVAs
all the time, representing (IIRC) two sizes being in heavy rotation,
while the depot is happily ticking along in a steady state as expected,
until it suddenly explodes out of nowhere; thirdly, I'd really like to
see instrumentation of the flush queues at the same time, since I think
they're the real culprit.
My theory so far is that everyone is calling queue_iova() frequently
enough to keep the timer at bay and their own queues drained. Then at
the ~16H mark, *something* happens that pauses unmaps long enough for
the timer to fire, and at that point all hell breaks loose.
So do you think that the freeing the IOVA magazines when the depot fills
is the cause of this? That was our analysis.
The depot is
suddenly flooded with IOVAs of *all* sizes, indicative of all the queues
being flushed at once (note that the two most common sizes have been
hovering perilously close to "full" the whole time), but then,
crucially, *that keeps happening*. My guess is that the load of
fq_flush_timeout() slows things down enough that the the timer then
keeps getting the chance to expire and repeat the situation.
Not sure on that one.
The main conclusion I draw from this is the same one that was my initial
gut feeling; that MAX_GLOBAL_MAGS = 32 is utter bollocks.
Yeah, I tend to agree with that. Or, more specifically, how things work
today is broken, and MAX_GLOBAL_MAGS = 32 is very much involved with that.
The CPU rcache
capacity scales with the number of CPUs; the flush queue capacity scales
with the number of CPUs; it is nonsensical that the depot size does not
correspondingly scale with the number of CPUs (I note that the testing
on the original patchset cites a 16-CPU system, where that depot
capacity is conveniently equal to the total rcache capacity).
Now yes, purging the rcaches when the depot is full does indeed help
mitigate this scenario - I assume it provides enough of a buffer where
the regular free_iova_fast() calls don't hit queue_iova() for a while
(and gives fq_ring_free() some reprieve on the CPU handling the
timeout), giving enough leeway for the flood to finish before anyone
starts hitting queues/locks/etc. and stalling again, and thus break the
self-perpetuating timeout cycle. But that's still only a damage
limitation exercise! It's planning for failure to just lie down and
assume that the depot is going to be full if fq_flush_timeout() ever
fires because it's something like an order of magnitude smaller than the
flush queue capacity (even for a uniform distribution of IOVA sizes) on
super-large systems.
I'm honestly tempted to move my position further towards a hard NAK on
this approach, because all the evidence so far points to it being a
bodge around a clear and easily-fixed scalability oversight. At the very
least I'd now want to hear a reasoned justification for why you want to
keep the depot at an arbitrary fixed size while the whole rest of the
system scales
(I'm assuming that since my previous suggestion to try
changes in that area seems to have been ignored).
So I said that it should fix the problem of the throughput going through
the floor at this 16h mark.
But we see 2x tightly coupled problems:
a. leading up to the ~16H critical point, throughput is slowly degrading
and becomes quite unstable (not shown in the log)
For the LSI3008 card, we don't see that. But then no IOVA size >
IOVA_RANGE_CACHE_MAX_SIZE occur there.
b. at the critical point, throughput goes through the floor
So b. should be fixed with the suggestion to have unlimited/higher depot
max bin size, but I reckon that we would still see a. And I put that
down to the fact that we have IOVA sizes > IOVA_RANGE_CACHE_MAX_SIZE at
a certain rate always. As the rb tree grows over time, they become
slower and slower to alloc+free - that's our theory. Allowing the depot
to grow further isn’t going to help that.
Maybe Leizhen's idea to trim the rcache periodically is overall better,
but I am concerned on implementation.
If not, then if we allow depot bin size to scale/grow, I would like to
see more efficient handling for IOVA size > IOVA_RANGE_CACHE_MAX_SIZE.
Thanks,
John
