| commit | e7cb11ec46aea8b0515575369f26c95747b67d53 | |
| author | Matthew Dillon <dillon@apollo.backplane.com> | |
| Fri, 20 Oct 2017 19:01:43 +0000 (20 12:01 -0700) | ||
| committer | Matthew Dillon <dillon@apollo.backplane.com> | |
| Tue, 31 Oct 2017 17:49:47 +0000 (31 10:49 -0700) | ||
| tree | 0a1e871f256f56b43f3b8153277d40717bb213e3 | treesnapshot (tar.gz zip) |
| parent | eae4df8827c6d4a0674ae320ce85170ccaf651af | commitdiff |
kernel - Refactor kern_sendfile()
* Refactor kern_sendfile() to greatly improve performance.
* Use vm_page_lookup_sbusy_try() exclusively to acquire VM pages
to assign to the mbufs.
* Instead of holding pages in a fancy manner, just issue the
UIO_NOCOPY / VMIO VOP_READ() in the blind and loop up.
* The VOP_READ() is still synchronous. It is really unclear
whether asynchronizing VOP_READ() via the pagerops would
really improve performance verses simply implementing a
limited number of connections per worker. At least in
localhost tests, we seem to be hitting a hardware memory
bottleneck long before we hit a cpu bottleneck.
* Refactor kern_sendfile() to greatly improve performance.
* Use vm_page_lookup_sbusy_try() exclusively to acquire VM pages
to assign to the mbufs.
* Instead of holding pages in a fancy manner, just issue the
UIO_NOCOPY / VMIO VOP_READ() in the blind and loop up.
* The VOP_READ() is still synchronous. It is really unclear
whether asynchronizing VOP_READ() via the pagerops would
really improve performance verses simply implementing a
limited number of connections per worker. At least in
localhost tests, we seem to be hitting a hardware memory
bottleneck long before we hit a cpu bottleneck.
| sys/kern/uipc_syscalls.c | diffblobblamehistory |