On 2018-08-28 14:30, Johannes Berg wrote:
Hi,
Sorry for the late reply, my work hours are limited right now.
I wanted to give a try with rhashtable and get your thoughts, since it
gave below flexibility to original patch,
1. avoids creating static memory when userspace starts accumulating
stats. 36 rate entries were used in original patch based on 10 (MCS
count) * 3 (entries per mcs)+ 6 escape entries, which would also
increase with HE supported now.
True, but rhashtable also has very significant overhead. I suspect it's
around the same order of magnitude as the allocation you need to keep
all the 36 entries?
struct rhashtable is probably already ~120 bytes on 64-bit systems
(very
rough counting), and a single bucket table is >64, so you already have
close to 200 bytes for just the basic structures, without *any*
entries.
And a significant amount of code too.
36 rate entries could probably be kept - I don't think you really need
to go more than that since you're not going to switch HT/VHT/HE all the
time, so that's only about 360 bytes total. You haven't gained much,
but
made it much more complex, doing much more allocations (create
new/destroy old entries) etc.
I don't really see much value in that.
Okay Johannes. I'll revive this patch based on your approach
and submit for review.
Thanks,
Sriram.R
2. avoids restricting to only 3 entries per MCS in the table. Using
hashtable gave flexibility to add/read the table dynamically based on
encoded rate key.
Yes but if you don't limit, you end up with run-away memory consumption
again.
Yes you're right ,it might grow but, as per this patch when Packet
count
is greater
than 65000 in an exntry or when the number of rate table/hashtable
entries exceed a count of MAX_RATE_TABLE_ELEMS (10 was used in this
patch), the complete table is dumped to userspace and new stats starts
getting collected in a new table after we flush the old one.
Hence with this approach also the memory consumption is limited
similar
to the original patch.
Right, so you limit to 10 entries, which is a total of
~120 + ~72 + 10 x (10 /* data */ + 3*8) = 524
in 12 different allocations. I don't think that's going to be much
better, and you seemed to think that the 10 would be commonly hit
(otherwise having a relatively small limit of 36 shouldn't be an issue)
So - I don't really see any advantage here, do you?
johannes
