Re: CVS commit: src/sys
> On Mar 29, 2020, at 8:03 AM, Joerg Sonnenberger wrote: > > Yes and I see no fundamental reason for not allowing buffering with 1ms > frames in that case. I expect Alpha to be fast enough to do that with > little overhead. That's fine, I just wanted to point it out. -- thorpej
Re: CVS commit: src/sys
On Sun, Mar 29, 2020 at 07:26:26AM -0700, Jason Thorpe wrote: > > > On Mar 29, 2020, at 6:11 AM, Joerg Sonnenberger wrote: > > > > I would allow at least 1/HZ as baseline. > > Be careful, some platforms have a relatively high HZ (e.g. Alpha HZ=1024). Yes and I see no fundamental reason for not allowing buffering with 1ms frames in that case. I expect Alpha to be fast enough to do that with little overhead. Joerg
Re: CVS commit: src/sys
> On Mar 29, 2020, at 6:11 AM, Joerg Sonnenberger wrote: > > I would allow at least 1/HZ as baseline. Be careful, some platforms have a relatively high HZ (e.g. Alpha HZ=1024). -- thorpej
Re: CVS commit: src/sys
On Sun, Mar 29, 2020 at 01:11:47PM +0900, Tetsuya Isaki wrote: > > I would prefer if blk_ms were kept below 5ms on amd64 and aarch64. > > We don't have to worry about the CPU load of playing audio on these > > platforms. > > CPU load or performance is not subject. I know that my > implementation will work on the most modern real hardware. > But I feel that at least 4msec is too rush to be default. > A default should not be for one game application. I would allow at least 1/HZ as baseline. Joerg
[vfs_cache] Re: CVS commit: src/sys/kern
Le 23/03/2020 à 21:02, Andrew Doran a écrit : > Module Name: src > Committed By: ad > Date: Mon Mar 23 20:02:14 UTC 2020 > > Modified Files: > src/sys/kern: vfs_cache.c > > Log Message: > cache_remove(): remove from the vnode list first, so cache_revlookup() > doesn't try to re-activate an entry no longer on the LRU list. > > > To generate a diff of this commit: > cvs rdiff -u -r1.133 -r1.134 src/sys/kern/vfs_cache.c > > Please note that diffs are not public domain; they are subject to the > copyright notices on the relevant files. It appears that your recent changes in vfs_cache.c have introduced a use-after-free. Booting KASAN gives: ASan: Unauthorized Access In 0x...: Addr 0x... [1 byte, read, PoolUseAfterFree] It seems that the problem is here: 557 if (nameiop == CREATE && (cnflags & ISLASTCN) != 0) { 558 /* 559 * Last component and we are preparing to create 560 * the named object, so flush the negative cache 561 * entry. 562 */ 563 COUNT(ncs_badhits); 564 cache_remove(ncp, true);< HERE 565 hit = false; 566 } else { 567 COUNT(ncs_neghits); 568 SDT_PROBE(vfs, namecache, lookup, hit, dvp, name, 569 namelen, 0, 0); 570 /* found neg entry; vn is already null from above */ 571 hit = true; 572 } 573 if (iswht_ret != NULL) { 574 /* 575 * Restore the ISWHITEOUT flag saved earlier. 576 */ 577 *iswht_ret = ncp->nc_whiteout; <-- ouch 578 } else { 579 KASSERT(!ncp->nc_whiteout); <-- ouch 580 } cache_remove() frees 'ncp', and then 'ncp->nc_whiteout' is read. Maxime
Re: CVS commit: src/sys
At Sat, 28 Mar 2020 10:40:26 +0100, Martin Husemann wrote: > It would be good to have a file somewhere in the audio code where the > default is selected based on some ifdefs - if we don't have anything > better something like: > > #if defined(__m68k__) || defined(__vax__) || \ > (defined(__sparc__) && !defined(_LP64) > > #else > ... 4ms > #endif > > Or create (if we don't have yet) a and define it there. I try it. Thanks, --- Tetsuya Isaki