We could make some operations lockless, like read and store, but I don't know how to make read-modify-write operations lockless generically. What do you have in mind?
On Sat, Jul 19, 2014 at 11:37:42AM +0300, Jarno Rajahalme wrote: > Would it be a bad idea to make the trivial types with sizes less than > void pointer lockless in ovs-atomic-pthread just like in gcc4+, at > least for x86/64? > > Jarno > > > On Jul 19, 2014, at 10:05 AM, Ben Pfaff <b...@nicira.com> wrote: > > > > I guess not, I see that xenddk-6.0.0-50762, at least, doesn't have a > > libatomic. > > > >> On Sat, Jul 19, 2014 at 12:04:25AM -0700, Ben Pfaff wrote: > >> Does it make a difference if you add -latomic to the linker flags? > >> > >>> On Fri, Jul 18, 2014 at 05:32:39PM -0700, Jarno Rajahalme wrote: > >>> From config.log: > >>> > >>> configure:13921: checking whether gcc -std=gnu99 supports GCC 4.0+ atomic > >>> built-ins > >>> configure:13988: gcc -std=gnu99 -o conftest -g -O2 conftest.c -lrt -lm > >>> >&5 > >>> /tmp/ccZ2M4lr.o: In function `main': > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:57: undefined > >>> reference to `__sync_fetch_and_add_1' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:57: undefined > >>> reference to `__sync_fetch_and_sub_1' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:57: undefined > >>> reference to `__sync_fetch_and_or_1' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:57: undefined > >>> reference to `__sync_fetch_and_and_1' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:57: undefined > >>> reference to `__sync_fetch_and_xor_1' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:58: undefined > >>> reference to `__sync_fetch_and_add_1' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:58: undefined > >>> reference to `__sync_fetch_and_sub_1' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:58: undefined > >>> reference to `__sync_fetch_and_or_1' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:58: undefined > >>> reference to `__sync_fetch_and_and_1' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:58: undefined > >>> reference to `__sync_fetch_and_xor_1' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:59: undefined > >>> reference to `__sync_fetch_and_add_1' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:59: undefined > >>> reference to `__sync_fetch_and_sub_1' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:59: undefined > >>> reference to `__sync_fetch_and_or_1' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:59: undefined > >>> reference to `__sync_fetch_and_and_1' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:59: undefined > >>> reference to `__sync_fetch_and_xor_1' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:60: undefined > >>> reference to `__sync_fetch_and_add_2' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:60: undefined > >>> reference to `__sync_fetch_and_sub_2' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:60: undefined > >>> reference to `__sync_fetch_and_or_2' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:60: undefined > >>> reference to `__sync_fetch_and_and_2' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:60: undefined > >>> reference to `__sync_fetch_and_xor_2' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:61: undefined > >>> reference to `__sync_fetch_and_add_2' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:61: undefined > >>> reference to `__sync_fetch_and_sub_2' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:61: undefined > >>> reference to `__sync_fetch_and_or_2' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:61: undefined > >>> reference to `__sync_fetch_and_and_2' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:61: undefined > >>> reference to `__sync_fetch_and_xor_2' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:62: undefined > >>> reference to `__sync_fetch_and_add_4' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:62: undefined > >>> reference to `__sync_fetch_and_sub_4' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:62: undefined > >>> reference to `__sync_fetch_and_or_4' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:62: undefined > >>> reference to `__sync_fetch_and_and_4' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:62: undefined > >>> reference to `__sync_fetch_and_xor_4' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:63: undefined > >>> reference to `__sync_fetch_and_add_4' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:63: undefined > >>> reference to `__sync_fetch_and_sub_4' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:63: undefined > >>> reference to `__sync_fetch_and_or_4' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:63: undefined > >>> reference to `__sync_fetch_and_and_4' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:63: undefined > >>> reference to `__sync_fetch_and_xor_4' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:64: undefined > >>> reference to `__sync_fetch_and_add_4' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:64: undefined > >>> reference to `__sync_fetch_and_sub_4' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:64: undefined > >>> reference to `__sync_fetch_and_or_4' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:64: undefined > >>> reference to `__sync_fetch_and_and_4' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:64: undefined > >>> reference to `__sync_fetch_and_xor_4' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:65: undefined > >>> reference to `__sync_fetch_and_add_4' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:65: undefined > >>> reference to `__sync_fetch_and_sub_4' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:65: undefined > >>> reference to `__sync_fetch_and_or_4' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:65: undefined > >>> reference to `__sync_fetch_and_and_4' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:65: undefined > >>> reference to `__sync_fetch_and_xor_4' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:66: undefined > >>> reference to `__sync_fetch_and_add_8' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:66: undefined > >>> reference to `__sync_fetch_and_sub_8' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:66: undefined > >>> reference to `__sync_fetch_and_or_8' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:66: undefined > >>> reference to `__sync_fetch_and_and_8' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:66: undefined > >>> reference to `__sync_fetch_and_xor_8' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:67: undefined > >>> reference to `__sync_fetch_and_add_8' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:67: undefined > >>> reference to `__sync_fetch_and_sub_8' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:67: undefined > >>> reference to `__sync_fetch_and_or_8' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:67: undefined > >>> reference to `__sync_fetch_and_and_8' > >>> /usr/src/redhat/BUILD/openvswitch-2.3.90.37691/conftest.c:67: undefined > >>> reference to `__sync_fetch_and_xor_8' > >>> collect2: ld returned 1 exit status > >>> > >>>> On Jul 18, 2014, at 5:00 PM, Jarno Rajahalme <jrajaha...@nicira.com> > >>>> wrote: > >>>> > >>>> I think I found the reason for the slowness of cmap on XenServer. > >>>> > >>>> I was debugging on a XenServer tonight and found out that the configure > >>>> test for GCC atomics had apparently failed, as I noticed actual lock and > >>>> unlock operations on gdb disassembly, and then by the fact the changes I > >>>> made to the ovs-atomic-gcc4+.h did not trigger any recompilations with > >>>> ?make? (so there was no dependency on it). > >>>> > >>>> If the configure check is correct, then there must be something wrong > >>>> with the XenServer setup? > >>>> > >>>> Jarno > >>>> > >>>>> On Jun 25, 2014, at 11:09 AM, Ben Pfaff <b...@nicira.com> wrote: > >>>>> > >>>>> I am really surprised to hear that it takes over a minute. On my > >>>>> workstation (which is a few years old), "ovstest test-cmap check 1" > >>>>> only takes 6.5 seconds. Do you have any idea what's different on > >>>>> XenServer? It might be the GCC version, which version is it using? > >>>>> > >>>>>> On Wed, Jun 25, 2014 at 09:40:12AM -0700, Gurucharan Shetty wrote: > >>>>>> I had a quick comment. I noticed while running unit tests on Xenserver > >>>>>> that the cuckoo hash unit test was very slow. It takes a little more > >>>>>> than one minute to complete. I wonder whether using more cmap would > >>>>>> hurt Xenserver's performance? > >>>>>> > >>>>>>> On Wed, Jun 25, 2014 at 4:19 AM, Jarno Rajahalme > >>>>>>> <jrajaha...@nicira.com> wrote: > >>>>>>> Ben, > >>>>>>> > >>>>>>> Thanks for the reviews! > >>>>>>> > >>>>>>> I have addressed the (minor) comments on the series, will push once > >>>>>>> the vector v2 is ack???ed. I???ll keep this last patch as the first > >>>>>>> in the classifier RCU series. > >>>>>>> > >>>>>>> Jarno > >>>>>>> > >>>>>>>> On Jun 13, 2014, at 10:37 AM, Ben Pfaff <b...@nicira.com> wrote: > >>>>>>>> > >>>>>>>>> On Mon, Jun 09, 2014 at 11:53:55AM -0700, Jarno Rajahalme wrote: > >>>>>>>>> Use cmap instead of hmap & hindex in classifier. Performance impact > >>>>>>>>> with current locking strategy is not yet tested. Later patches will > >>>>>>>>> introduce RCU into the classifer. > >>>>>>>> > >>>>>>>> into the "classifier". > >>>>>>>> > >>>>>>>>> Signed-off-by: Jarno Rajahalme <jrajaha...@nicira.com> > >>>>>>>> > >>>>>>>> As long as this passes the tests (I assume you ran them) it seems OK > >>>>>>>> to me. The tricky review will be when we start actually making the > >>>>>>>> classifier RCU. (I don't know if it makes sense to push this patch > >>>>>>>> before then. It might be better as the first patch of a series that > >>>>>>>> makes the classifier RCU. But I'll leave that decision to you.) > >>>>>>>> > >>>>>>>> Acked-by: Ben Pfaff <b...@nicira.com> > >>>>>>> > >>>>>>> _______________________________________________ > >>>>>>> dev mailing list > >>>>>>> dev@openvswitch.org > >>>>>>> http://openvswitch.org/mailman/listinfo/dev > >>> _______________________________________________ dev mailing list dev@openvswitch.org http://openvswitch.org/mailman/listinfo/dev
