Re: [PATCH] Bluetooth: Deinline large functions

From: Joe Perches
Date: Sun Apr 10 2016 - 17:18:47 EST


On Sun, 2016-04-10 at 19:18 +0200, Denys Vlasenko wrote:
> On Sun, Apr 10, 2016 at 6:59 PM, Joe Perches <joe@xxxxxxxxxxx> wrote:
> >
> > On Sun, 2016-04-10 at 18:24 +0200, Denys Vlasenko wrote:
> > >
> > > Fastest existing Bluetooth standard's top speed is 2.4 MB/s.
> > > It is way off from being CPU limited, no need to squeeze
> > > last few cycles by excessive inlining.
> > >
> > > This patch delinlines the following functions:
> > >
> > > hci_conn_hash_lookup_handle: 345 bytes, 39 calls
> > > hci_conn_hash_lookup_ba: 372 bytes, 36 calls
> > > hci_conn_hash_lookup_le: 382 bytes, 8 calls
> > > hci_conn_hash_lookup_state: 356 bytes, 3 calls
> > > hci_lookup_le_connect: 378 bytes, 7 calls
> > > hci_conn_drop: 186 bytes, 30 calls
> > > hci_connect_cfm: 121 bytes, 15 calls
> > > hci_disconn_cfm: 121 bytes, 2 calls
> > > hci_auth_cfm: 156 bytes, 2 calls
> > > hci_encrypt_cfm: 156 bytes, 3 calls
> > >
> > > Size reduction is about 40k:
> > >
> > >     text     data      bss       dec     hex filename
> > > 95943139 20860256 35991552 152794947 91b7743 vmlinux_before
> > > 95903714 20860256 35991552 152755522 91add42 vmlinux
> > Hello Denys
> >
> > While removing unnecessary inlines is generally a
> > good thing, for extremely low power embedded systems
> > like a coin-battery operated bicycle computer or a
> > heart rate monitor, this might cause a throughput
> > reduction.
> Todays CPUs can push several GB/s over, say, Ethernet.
> BT is 2.4 MB/s, tops.
> Therefore total CPU time spent preparing BT I/O must be
> in the 0.1% ballpark. (Meaning, if we would have a magical
> infinitely fast CPU, throughput would possibly increase
> by about 0.1%). How much do you think this deinlining patch
> can possibly hurt here?

Today's extremely low power ARM chips that run
at a couple dozen MHz
cannot push gigabit speeds.

> > Can you please also show the size decrease when done
> > with a defconfig with bluetooth support?
> >
> > And for these types of patches in general, please
> > add a defconfig size reduction to the commit message.
> Okay, will do in the future.
>
> >
> > This is an x86-64 defconfig with bluetooth with and
> > without this patch:
> >
> > $ size vmlinux.defconfig.*
> >    text    data     bss     dec     hex filename
> > 10214414        4313816 1097728 15625958         ee6ee6 vmlinux.defconfig.new
> > 10224014        4313816 1097728 15635558         ee9466 vmlinux.defconfig.old
> >
> > ~10k total
> Looks good, right?

Size reductions are generally good.

The one function that might be appropriate to leave inline
is perhaps hci_conn_hash_lookup_handle.  Dunno.

It's be good to test though.