Re: somewhat important inteldrm fix
On Tue, Feb 4, 2014 at 11:55 PM, Mark Kettenis mark.kette...@xs4all.nl wrote: Running the updated xf86-video-intel driver uncovered a bug in the kernel drm code. The page fault handler wasn't handling some of the possible errors correctly. This made the X server die with a SIGSEGV. The diff below brings things closer to what Linux does, and seems to fix the crashes I was seeing. A bit more testing would be welcome though. Note that this needs the commit to drmP.h I just made, which might not have made it to all the anoncvs servers yet. I saw the X server crash yesterday, while watching a video in fullscreen using minitube. I will apply your diff and report back asap... ciao, David
exp() / expl() on Linux and OpenBSD (expl() bug?)
Hi! I hit this problem while working on updating math/R from version 2.15.3 to the latest version (3.0.2). It started happening since upstream switched from double functions to C99 long double functions (expl, fabsl, ...), during the R-3 development cycle. Take the following reduced test-case, adapted from what R's code does: ---8--- #include stdio.h #include stdlib.h #include math.h int main(void) { double theta = 1; long double lambda, pr, pr2; lambda = (0.5*theta); pr = exp(-lambda); pr2 = expl(-lambda); printf(theta == %g, pr == %Lg, pr2 == %Lg\n, theta, pr, pr2); exit(0); } ---8--- This produces the following output on Linux (x86_64): theta == 1, pr == 0.606531, pr2 == 0.606531 While on OpenBSD -current amd64: theta == 1, pr == 0.606531, pr2 == nan And indeed R-3's testsuite fails with the error message NaNs produced: Warning in pchisq(1e-300, df = 0, ncp = lam) : NaNs produced stopifnot(all.equal(p00, exp(-lam/2)), + all.equal(p.0, exp(-lam/2))) Error: all.equal(p.0, exp(-lam/2)) is not TRUE Execution halted Is this a bug in our expl() ? Ciao, David
Re: exp() / expl() on Linux and OpenBSD (expl() bug?)
Date: Wed, 5 Feb 2014 01:57:33 -0700 From: David Coppa dco...@openbsd.org Hi! I hit this problem while working on updating math/R from version 2.15.3 to the latest version (3.0.2). It started happening since upstream switched from double functions to C99 long double functions (expl, fabsl, ...), during the R-3 development cycle. Take the following reduced test-case, adapted from what R's code does: ---8--- #include stdio.h #include stdlib.h #include math.h int main(void) { double theta = 1; long double lambda, pr, pr2; lambda = (0.5*theta); pr = exp(-lambda); pr2 = expl(-lambda); printf(theta == %g, pr == %Lg, pr2 == %Lg\n, theta, pr, pr2); exit(0); } ---8--- This produces the following output on Linux (x86_64): theta == 1, pr == 0.606531, pr2 == 0.606531 While on OpenBSD -current amd64: theta == 1, pr == 0.606531, pr2 == nan And indeed R-3's testsuite fails with the error message NaNs produced: Warning in pchisq(1e-300, df = 0, ncp = lam) : NaNs produced stopifnot(all.equal(p00, exp(-lam/2)), + all.equal(p.0, exp(-lam/2))) Error: all.equal(p.0, exp(-lam/2)) is not TRUE Execution halted Is this a bug in our expl() ? Yes.
Re: exp() / expl() on Linux and OpenBSD (expl() bug?)
Date: Wed, 5 Feb 2014 01:57:33 -0700 From: David Coppa dco...@openbsd.org Hi! I hit this problem while working on updating math/R from version 2.15.3 to the latest version (3.0.2). It started happening since upstream switched from double functions to C99 long double functions (expl, fabsl, ...), during the R-3 development cycle. Take the following reduced test-case, adapted from what R's code does: ---8--- #include stdio.h #include stdlib.h #include math.h int main(void) { double theta = 1; long double lambda, pr, pr2; lambda = (0.5*theta); pr = exp(-lambda); pr2 = expl(-lambda); printf(theta == %g, pr == %Lg, pr2 == %Lg\n, theta, pr, pr2); exit(0); } ---8--- This produces the following output on Linux (x86_64): theta == 1, pr == 0.606531, pr2 == 0.606531 While on OpenBSD -current amd64: theta == 1, pr == 0.606531, pr2 == nan And indeed R-3's testsuite fails with the error message NaNs produced: Warning in pchisq(1e-300, df = 0, ncp = lam) : NaNs produced stopifnot(all.equal(p00, exp(-lam/2)), + all.equal(p.0, exp(-lam/2))) Error: all.equal(p.0, exp(-lam/2)) is not TRUE Execution halted Is this a bug in our expl() ? Oh, btw, the quad-precision code used on sparc64 gets this right. So the bug is probably somewhere in src/lib/limb/src/ld80.
Re: somewhat important inteldrm fix
From: David Coppa dco...@gmail.com Date: Wed, 5 Feb 2014 09:01:45 +0100 On Tue, Feb 4, 2014 at 11:55 PM, Mark Kettenis mark.kette...@xs4all.nl wrote: Running the updated xf86-video-intel driver uncovered a bug in the kernel drm code. The page fault handler wasn't handling some of the possible errors correctly. This made the X server die with a SIGSEGV. The diff below brings things closer to what Linux does, and seems to fix the crashes I was seeing. A bit more testing would be welcome though. Note that this needs the commit to drmP.h I just made, which might not have made it to all the anoncvs servers yet. I saw the X server crash yesterday, while watching a video in fullscreen using minitube. I will apply your diff and report back asap... Since matthieu@ confirmed it fixes his problems as well, I've committed the diff. So it should show up on your favourite anoncvs mirror shortly.
Re: somewhat important inteldrm fix
On Wed, Feb 5, 2014 at 11:43 AM, Mark Kettenis mark.kette...@xs4all.nl wrote: From: David Coppa dco...@gmail.com Date: Wed, 5 Feb 2014 09:01:45 +0100 On Tue, Feb 4, 2014 at 11:55 PM, Mark Kettenis mark.kette...@xs4all.nl wrote: Running the updated xf86-video-intel driver uncovered a bug in the kernel drm code. The page fault handler wasn't handling some of the possible errors correctly. This made the X server die with a SIGSEGV. The diff below brings things closer to what Linux does, and seems to fix the crashes I was seeing. A bit more testing would be welcome though. Note that this needs the commit to drmP.h I just made, which might not have made it to all the anoncvs servers yet. I saw the X server crash yesterday, while watching a video in fullscreen using minitube. I will apply your diff and report back asap... Since matthieu@ confirmed it fixes his problems as well, I've committed the diff. So it should show up on your favourite anoncvs mirror shortly. OK, thanks. cheers, david
Re: help needed from someone with an sk(4)
* David Higgs hig...@gmail.com [2014-01-25 18:25]: On Jan 25, 2014, at 12:48 AM, David Higgs hig...@gmail.com wrote: On Fri, Jan 24, 2014 at 4:24 AM, Henning Brauer lists-openbsdt...@bsws.de wrote: * Henning Brauer lists-openbsdt...@bsws.de [2014-01-24 05:50]: i need this tested on an sk(4). I don't have that hardware at all. this gets rif od a slight little bit more. Resurrected an old box, kernel compile w/ patch is underway. Should be able to provide feedback tomorrow. I'm kinda new to this - do I need to exercise forwarding, do some speed tests, or is it sufficient to just make sure that host-based usage doesn't break? Snapshot dmesg below. No problems seen with host usage, after downloading a couple files w/ FTP and updating source via CVS. awesome, thanks for verifying! a mistake would have shown up VERY early and clearly. -- Henning Brauer, h...@bsws.de, henn...@openbsd.org BS Web Services GmbH, http://bsws.de, Full-Service ISP Secure Hosting, Mail and DNS Services. Dedicated Servers, Root to Fully Managed Henning Brauer Consulting, http://henningbrauer.com/
Re: somewhat important inteldrm fix
Running the updated xf86-video-intel driver uncovered a bug in the kernel drm code. The page fault handler wasn't handling some of the possible errors correctly. This made the X server die with a SIGSEGV. The diff below brings things closer to what Linux does, and seems to fix the crashes I was seeing. A bit more testing would be welcome though. Note that this needs the commit to drmP.h I just made, which might not have made it to all the anoncvs servers yet. Index: i915_gem.c === RCS file: /cvs/src/sys/dev/pci/drm/i915/i915_gem.c,v retrieving revision 1.68 diff -u -p -r1.68 i915_gem.c --- i915_gem.c2 Feb 2014 10:54:10 - 1.68 +++ i915_gem.c4 Feb 2014 22:47:53 - @@ -1522,6 +1522,7 @@ i915_gem_fault(struct drm_gem_object *ge NULL, NULL); DRM_UNLOCK(); dev_priv-entries--; + pmap_update(ufi-orig_map-pmap); uvm_wait(intelflt); return (VM_PAGER_REFAULT); } @@ -1533,18 +1534,42 @@ unlock: DRM_UNLOCK(); dev_priv-entries--; pmap_update(ufi-orig_map-pmap); - if (ret == -EIO) { + + switch (ret) { + case -EIO: + /* If this -EIO is due to a gpu hang, give the reset code a + * chance to clean up the mess. Otherwise return the proper + * SIGBUS. */ + if (!atomic_read(dev_priv-mm.wedged)) + return VM_PAGER_ERROR; + case -EAGAIN: + /* Give the error handler a chance to run and move the + * objects off the GPU active list. Next time we service the + * fault, we should be able to transition the page into the + * GTT without touching the GPU (and so avoid further + * EIO/EGAIN). If the GPU is wedged, then there is no issue + * with coherency, just lost writes. + */ +#if 0 + set_need_resched(); +#endif + case 0: + case -ERESTART: + case -EINTR: + case -EBUSY: /* - * EIO means we're wedged, so upon resetting the gpu we'll - * be alright and can refault. XXX only on resettable chips. + * EBUSY is ok: this just means that another thread + * already did the job. */ - ret = VM_PAGER_REFAULT; - } else if (ret) { - ret = VM_PAGER_ERROR; - } else { - ret = VM_PAGER_OK; + return VM_PAGER_OK; + case -ENOMEM: + return VM_PAGER_ERROR; + case -ENOSPC: + return VM_PAGER_ERROR; + default: + WARN_ONCE(ret, unhandled error in i915_gem_fault: %i\n, ret); + return VM_PAGER_ERROR; } - return ret; } /** Thanks, this works good for me too -- no X segfaults anymore.
fail to boot snapshot 5.5 on MBPro8,2
Hi, I tried to install 5.5 snapshots from 2. Feb and 3. Feb onto my laptop MacBook Pro 8,2 - both failed. Then used an older snapshot from spacehopper.org/mirrmon, which claimed to be 12 days old. Failed as well. MacBook Pro runs perfectly well with 5.4, and 5.5 bsd.rd lets me go into the installation. After a reboot (bsd.mp), the system hangs after the line where the disk should be mounted, and the nvram/clock message normally appears (actually before going into userland). The screen then turns shortly but steady into white/grey, so the blue kernel lines can't be read anymore. I then tried to boot with bsd -c, with an external USB keyboard - but it hangs as well (as such no difference to the internal keyboard). Looks like I have two prompts (underlines) on the screen, that are flickering. With this status I cannot provide any dmesg:-( The first line after boot -c would show the booting hd0a:bsd: 7635180+1660460+1097336... line, and then entry point at The kernels blue lines would say: kbc: cmd word write error [ using 926384 bytes of bsd ELF symbol table ] Copyright (c) 1982, 1986, 1989, 1991, 1993 The Regents of the University ... OpenBSD 5.5-beta (GENERIC.MP) #284: Mon Feb 3 07:57:32 MST 2014 t...@amd64.openbsd.org:/usr/src/sys/arch/amd64/compile/GENERIC.MP RTC BIOS diagnostic error ffclock_battery,ROM_cksum,config_unit,memory_size,fixed_disk,invalid_time real mem = 4185079808 (3991MB) avail mem = 4065452032 (3877MB) User Kernel Config UKC The errors on RTC BIOS line also appears in the fully functional 5.4 version. But I couldn't see the line kbc: cmd word write error in other dmesgs on OpenBSD-tech. As mentioned, no dmesg possible, anyone else with MBPro issue? Any idea were to get old snapshots? I could try to build the kernel from scratch, I successfully built 5.4, and building with -current shows same behaviour as snapshots. Any ideas where to go from here? (No serial port obviously). Volker
Use explicit_bzero in login_*
Index: login_chpass/login_chpass.c === RCS file: /cvs/src/libexec/login_chpass/login_chpass.c,v retrieving revision 1.16 diff -u -p -r1.16 login_chpass.c --- login_chpass/login_chpass.c 4 Dec 2012 02:24:47 - 1.16 +++ login_chpass/login_chpass.c 5 Feb 2014 15:44:26 - @@ -208,7 +208,7 @@ yp_chpass(char *username) pwd_gensalt(salt, sizeof(salt), lc, 'y') == 0) strlcpy(salt, xx, sizeof(salt)); crypt(p, salt); - memset(p, 0, strlen(p)); + explicit_bzero(p, strlen(p)); } warnx(YP passwd database unchanged.); exit(1); Index: login_lchpass/login_lchpass.c === RCS file: /cvs/src/libexec/login_lchpass/login_lchpass.c,v retrieving revision 1.14 diff -u -p -r1.14 login_lchpass.c --- login_lchpass/login_lchpass.c 4 Dec 2012 02:24:47 - 1.14 +++ login_lchpass/login_lchpass.c 5 Feb 2014 15:44:27 - @@ -136,7 +136,7 @@ main(int argc, char *argv[]) exit(1); salt = crypt(p, salt); - memset(p, 0, strlen(p)); + explicit_bzero(p, strlen(p)); if (!pwd || strcmp(salt, pwd-pw_passwd) != 0) exit(1); Index: login_passwd/login.c === RCS file: /cvs/src/libexec/login_passwd/login.c,v retrieving revision 1.10 diff -u -p -r1.10 login.c --- login_passwd/login.c1 Jun 2012 01:43:19 - 1.10 +++ login_passwd/login.c5 Feb 2014 15:44:27 - @@ -158,7 +158,7 @@ main(int argc, char **argv) #endif if (password != NULL) - memset(password, 0, strlen(password)); + explicit_bzero(password, strlen(password)); if (ret != AUTH_OK) fprintf(back, BI_REJECT \n); Index: login_passwd/login_passwd.c === RCS file: /cvs/src/libexec/login_passwd/login_passwd.c,v retrieving revision 1.9 diff -u -p -r1.9 login_passwd.c --- login_passwd/login_passwd.c 9 Mar 2006 19:14:10 - 1.9 +++ login_passwd/login_passwd.c 5 Feb 2014 15:44:27 - @@ -62,7 +62,7 @@ pwd_login(char *username, char *password salt = crypt(password, salt); plen = strlen(password); - memset(password, 0, plen); + explicit_bzero(password, plen); /* * Authentication fails if the user does not exist in the password Index: login_tis/login_tis.c === RCS file: /cvs/src/libexec/login_tis/login_tis.c,v retrieving revision 1.11 diff -u -p -r1.11 login_tis.c --- login_tis/login_tis.c 4 Dec 2012 02:24:47 - 1.11 +++ login_tis/login_tis.c 5 Feb 2014 15:44:27 - @@ -394,8 +394,8 @@ tis_getkey(struct tis_connection *tc) } DES_string_to_key(key, cblock); error = DES_set_key(cblock, tc-keysched); - memset(key, 0, len); - memset(cblock, 0, sizeof(cblock)); + explicit_bzero(key, len); + explicit_bzero(cblock, sizeof(cblock)); free(tbuf); return (error); } @@ -507,10 +507,10 @@ tis_recv(struct tis_connection *tc, u_ch len, ks, iv, DES_DECRYPT); if (strlcpy(buf, tbuf, bufsiz) = bufsiz) { syslog(LOG_ERR, unencrypted data too large to store); - memset(tbuf, 0, sizeof(tbuf)); + explicit_bzero(tbuf, sizeof(tbuf)); return (-1); } - memset(tbuf, 0, sizeof(tbuf)); + explicit_bzero(tbuf, sizeof(tbuf)); } return (len); } @@ -656,7 +656,7 @@ tis_authorize(struct tis_connection *tc, syslog(LOG_ERR, unexpected response from authsrv: %s, obuf); resp = error; } - memset(buf, 0, sizeof(buf)); + explicit_bzero(buf, sizeof(buf)); return (resp); } @@ -684,10 +684,10 @@ tis_verify(struct tis_connection *tc, co if (strncmp(buf, ok, 2) == 0) { if (buf[2] != '\0') strlcpy(ebuf, buf + 3, TIS_BUFSIZ); - memset(buf, 0, sizeof(buf)); + explicit_bzero(buf, sizeof(buf)); return (0); } strlcpy(ebuf, buf, TIS_BUFSIZ); - memset(buf, 0, sizeof(buf)); + explicit_bzero(buf, sizeof(buf)); return (-1); } Index: login_yubikey/login_yubikey.c === RCS file: /cvs/src/libexec/login_yubikey/login_yubikey.c,v retrieving revision 1.8 diff -u -p -r1.8 login_yubikey.c --- login_yubikey/login_yubikey.c 27 Nov 2013 21:25:25 - 1.8 +++ login_yubikey/login_yubikey.c 5 Feb 2014 15:44:27 -
Re: ok to kill stdio.h in strsep.c?
On Sat, Jan 25, 2014 at 01:49:24AM -0500, Jean-Philippe Ouellet wrote: It appeared in revision 1.3 (Update from lite2.) It's the only one in the string family that has it, and nothing from it is used. I think this change is fine. I'll commit this soon if I don't hear objections. Index: strsep.c === RCS file: /cvs/src/lib/libc/string/strsep.c,v retrieving revision 1.6 diff -u -p -r1.6 strsep.c --- strsep.c8 Aug 2005 08:05:37 - 1.6 +++ strsep.c25 Jan 2014 06:41:18 - @@ -30,7 +30,6 @@ */ #include string.h -#include stdio.h /* * Get next token from string *stringp, where tokens are possibly-empty
Re: ok to kill stdio.h in strsep.c?
ok krw@ On 5 February 2014 13:15, Stefan Sperling s...@openbsd.org wrote: On Sat, Jan 25, 2014 at 01:49:24AM -0500, Jean-Philippe Ouellet wrote: It appeared in revision 1.3 (Update from lite2.) It's the only one in the string family that has it, and nothing from it is used. I think this change is fine. I'll commit this soon if I don't hear objections. Index: strsep.c === RCS file: /cvs/src/lib/libc/string/strsep.c,v retrieving revision 1.6 diff -u -p -r1.6 strsep.c --- strsep.c8 Aug 2005 08:05:37 - 1.6 +++ strsep.c25 Jan 2014 06:41:18 - @@ -30,7 +30,6 @@ */ #include string.h -#include stdio.h /* * Get next token from string *stringp, where tokens are possibly-empty
quick fix for uvm deadlocks
We are missing back pressure channels from uvm to the buf cache. The buf cache will happily sit on 9000 free pages while uvm churns around trying to scavenge up one more page. Fixing this is beyond the scope of a simple diff, but here's something that seems to help in a lot of the common cases, particularly the pla deadlock detected spin cycle that people see. If we're out of memory, kick the buf cache into releasing some clean pages. The buf cache may eventually find itself sorely in need of memory and unable to get it, but this is better than nothing. I've deliberately saved the back pressure until we're already on the about to die path to minimize regressions. uvm won't steal back memory unless it absolutely has to. Index: kern/vfs_bio.c === RCS file: /cvs/src/sys/kern/vfs_bio.c,v retrieving revision 1.154 diff -u -p -r1.154 vfs_bio.c --- kern/vfs_bio.c 25 Jan 2014 04:23:31 - 1.154 +++ kern/vfs_bio.c 5 Feb 2014 22:08:07 - @@ -305,6 +305,26 @@ bufadjust(int newbufpages) splx(s); } +int +buf_nukeclean(void) +{ + struct buf *bp; + int n; + + printf(nuking clean bufs\n); + n = 0; + while ((bp = TAILQ_FIRST(bufqueues[BQ_CLEAN])) n++ 10) { + bremfree(bp); + if (bp-b_vp) { + RB_REMOVE(buf_rb_bufs, + bp-b_vp-v_bufs_tree, bp); + brelvp(bp); + } + buf_put(bp); + } + return (n); +} + /* * Make the buffer cache back off from cachepct. */ Index: sys/buf.h === RCS file: /cvs/src/sys/sys/buf.h,v retrieving revision 1.93 diff -u -p -r1.93 buf.h --- sys/buf.h 21 Nov 2013 01:16:52 - 1.93 +++ sys/buf.h 5 Feb 2014 22:04:09 - @@ -312,6 +312,8 @@ voidbuf_fix_mapping(struct buf *, vsize void buf_alloc_pages(struct buf *, vsize_t); void buf_free_pages(struct buf *); +intbuf_nukeclean(void); + void minphys(struct buf *bp); intphysio(void (*strategy)(struct buf *), dev_t dev, int flags, Index: uvm/uvm_pdaemon.c === RCS file: /cvs/src/sys/uvm/uvm_pdaemon.c,v retrieving revision 1.64 diff -u -p -r1.64 uvm_pdaemon.c --- uvm/uvm_pdaemon.c 30 May 2013 16:29:46 - 1.64 +++ uvm/uvm_pdaemon.c 5 Feb 2014 22:04:15 - @@ -116,7 +116,7 @@ uvm_wait(const char *wmsg) * check for page daemon going to sleep (waiting for itself) */ - if (curproc == uvm.pagedaemon_proc) { + if (curproc == uvm.pagedaemon_proc buf_nukeclean() == 0) { /* * now we have a problem: the pagedaemon wants to go to * sleep until it frees more memory. but how can it Index: uvm/uvm_pmemrange.c === RCS file: /cvs/src/sys/uvm/uvm_pmemrange.c,v retrieving revision 1.36 diff -u -p -r1.36 uvm_pmemrange.c --- uvm/uvm_pmemrange.c 29 Jan 2013 19:55:48 - 1.36 +++ uvm/uvm_pmemrange.c 5 Feb 2014 22:07:37 - @@ -22,6 +22,7 @@ #include sys/malloc.h #include sys/proc.h /* XXX for atomic */ #include sys/kernel.h +#include sys/buf.h /* * 2 trees: addr tree and size tree. @@ -1883,6 +1884,13 @@ uvm_wait_pla(paddr_t low, paddr_t high, const char *wmsg = pmrwait; if (curproc == uvm.pagedaemon_proc) { + uvm_unlock_fpageq(); + if (buf_nukeclean() != 0) { + uvm_lock_fpageq(); + return 0; + } + uvm_lock_fpageq(); + /* * XXX detect pagedaemon deadlock - see comment in * uvm_wait(), as this is exactly the same issue.
Re: quick fix for uvm deadlocks
On Wed, Feb 5, 2014 at 3:17 PM, Ted Unangst t...@tedunangst.com wrote: We are missing back pressure channels from uvm to the buf cache. The buf cache will happily sit on 9000 free pages while uvm churns around trying to scavenge up one more page. Indeed, those are it's minimums (I presume in your case) and are exactly the amount of memory that uvm would never have even seen under the model of the static cache. So I don't agree with your statement we are missing back pressure channels from the uvm to the buf cache. It looks to me like the situation you are talking about is that the buffer cache has already backed off to it's minium (which is used to ensure things like avoiding deadlocks in the bufer cache on delayed writes and fun stuff like that). Or are you in a situation here where the cache has *not* backed off? Fixing this is beyond the scope of a simple diff, but here's something that seems to help in a lot of the common cases, particularly the pla deadlock detected spin cycle that people see. If we're out of memory, kick the buf cache into releasing some clean pages. The buf cache may eventually find itself sorely in need of memory and unable to get it, but this is better than nothing. I've deliberately saved the back pressure until we're already on the about to die path to minimize regressions. uvm won't steal back memory unless it absolutely has to. And for the reasons you say, I think this has great potential to move the deadlock into the buffer cache on small memory machines when we get the entire cache filled with delwri... Sure, we can make the miniums smaller, but in the end we still have to fix the page daemon or we are just delaying the inevitable or moving the deadlock to other subsystems. Index: kern/vfs_bio.c === RCS file: /cvs/src/sys/kern/vfs_bio.c,v retrieving revision 1.154 diff -u -p -r1.154 vfs_bio.c --- kern/vfs_bio.c 25 Jan 2014 04:23:31 - 1.154 +++ kern/vfs_bio.c 5 Feb 2014 22:08:07 - @@ -305,6 +305,26 @@ bufadjust(int newbufpages) splx(s); } +int +buf_nukeclean(void) +{ + struct buf *bp; + int n; + + printf(nuking clean bufs\n); + n = 0; + while ((bp = TAILQ_FIRST(bufqueues[BQ_CLEAN])) n++ 10) { + bremfree(bp); + if (bp-b_vp) { + RB_REMOVE(buf_rb_bufs, + bp-b_vp-v_bufs_tree, bp); + brelvp(bp); + } + buf_put(bp); + } + return (n); +} + /* * Make the buffer cache back off from cachepct. */ Index: sys/buf.h === RCS file: /cvs/src/sys/sys/buf.h,v retrieving revision 1.93 diff -u -p -r1.93 buf.h --- sys/buf.h 21 Nov 2013 01:16:52 - 1.93 +++ sys/buf.h 5 Feb 2014 22:04:09 - @@ -312,6 +312,8 @@ voidbuf_fix_mapping(struct buf *, vsize void buf_alloc_pages(struct buf *, vsize_t); void buf_free_pages(struct buf *); +intbuf_nukeclean(void); + void minphys(struct buf *bp); intphysio(void (*strategy)(struct buf *), dev_t dev, int flags, Index: uvm/uvm_pdaemon.c === RCS file: /cvs/src/sys/uvm/uvm_pdaemon.c,v retrieving revision 1.64 diff -u -p -r1.64 uvm_pdaemon.c --- uvm/uvm_pdaemon.c 30 May 2013 16:29:46 - 1.64 +++ uvm/uvm_pdaemon.c 5 Feb 2014 22:04:15 - @@ -116,7 +116,7 @@ uvm_wait(const char *wmsg) * check for page daemon going to sleep (waiting for itself) */ - if (curproc == uvm.pagedaemon_proc) { + if (curproc == uvm.pagedaemon_proc buf_nukeclean() == 0) { /* * now we have a problem: the pagedaemon wants to go to * sleep until it frees more memory. but how can it Index: uvm/uvm_pmemrange.c === RCS file: /cvs/src/sys/uvm/uvm_pmemrange.c,v retrieving revision 1.36 diff -u -p -r1.36 uvm_pmemrange.c --- uvm/uvm_pmemrange.c 29 Jan 2013 19:55:48 - 1.36 +++ uvm/uvm_pmemrange.c 5 Feb 2014 22:07:37 - @@ -22,6 +22,7 @@ #include sys/malloc.h #include sys/proc.h /* XXX for atomic */ #include sys/kernel.h +#include sys/buf.h /* * 2 trees: addr tree and size tree. @@ -1883,6 +1884,13 @@ uvm_wait_pla(paddr_t low, paddr_t high, const char *wmsg = pmrwait; if (curproc == uvm.pagedaemon_proc) { + uvm_unlock_fpageq(); + if (buf_nukeclean() != 0) { + uvm_lock_fpageq(); + return 0; + } + uvm_lock_fpageq(); + /* * XXX detect pagedaemon deadlock - see comment in * uvm_wait(), as this is exactly the same
Re: quick fix for uvm deadlocks
On Wed, Feb 05, 2014 at 17:53, Bob Beck wrote: On Wed, Feb 5, 2014 at 3:17 PM, Ted Unangst t...@tedunangst.com wrote: We are missing back pressure channels from uvm to the buf cache. The buf cache will happily sit on 9000 free pages while uvm churns around trying to scavenge up one more page. Or are you in a situation here where the cache has *not* backed off? Talked to Bob and hashed out better ideas of the problem. The page daemon does tell the buffer cache to make some room, but... If you have a huge mmap file, the pdaemon will try to flush it out via VOP_WRITE, which circles back via ffs into buf_get, which eats those previously freed pages, and then some, as the pagedaemon continues pushing more and more of the mmap file out. We discussed some other changes and fixes that this situation has clearly highlighted, but here's a slightly revised diff. It now uses the correct bufbackoff() function to communicate uvm's needs. Any other fix is rather precarious for this release, but as stated before, this keeps the change to the deadlock paths. You were already dead, but now you have a second chance. (We don't currently use the pmemrange argument; we'll have to adjust accordingly when the bufcache becomes range aware.) Index: uvm_pdaemon.c === RCS file: /cvs/src/sys/uvm/uvm_pdaemon.c,v retrieving revision 1.64 diff -u -p -r1.64 uvm_pdaemon.c --- uvm_pdaemon.c 30 May 2013 16:29:46 - 1.64 +++ uvm_pdaemon.c 6 Feb 2014 03:09:53 - @@ -117,6 +117,8 @@ uvm_wait(const char *wmsg) */ if (curproc == uvm.pagedaemon_proc) { + if (bufbackoff(NULL, 4) == 0) + return; /* * now we have a problem: the pagedaemon wants to go to * sleep until it frees more memory. but how can it Index: uvm_pmemrange.c === RCS file: /cvs/src/sys/uvm/uvm_pmemrange.c,v retrieving revision 1.36 diff -u -p -r1.36 uvm_pmemrange.c --- uvm_pmemrange.c 29 Jan 2013 19:55:48 - 1.36 +++ uvm_pmemrange.c 6 Feb 2014 03:10:32 - @@ -22,6 +22,7 @@ #include sys/malloc.h #include sys/proc.h /* XXX for atomic */ #include sys/kernel.h +#include sys/mount.h /* * 2 trees: addr tree and size tree. @@ -1883,6 +1884,13 @@ uvm_wait_pla(paddr_t low, paddr_t high, const char *wmsg = pmrwait; if (curproc == uvm.pagedaemon_proc) { + uvm_unlock_fpageq(); + if (bufbackoff(NULL, atop(size)) == 0) { + uvm_lock_fpageq(); + return 0; + } + uvm_lock_fpageq(); + /* * XXX detect pagedaemon deadlock - see comment in * uvm_wait(), as this is exactly the same issue.
Re: quick fix for uvm deadlocks
Yes, this is much better. although I think this problem related to big mmaps has been with us for a while. and appears to avoid the problem with the offending test programs on my machines. I'm ok with that going in for the moment, although I want some of your time to look at that nasty shit we talked about :) I think we can make that a lot better with some NOCACHE.. On Wed, Feb 5, 2014 at 9:03 PM, Ted Unangst t...@tedunangst.com wrote: On Wed, Feb 05, 2014 at 17:53, Bob Beck wrote: On Wed, Feb 5, 2014 at 3:17 PM, Ted Unangst t...@tedunangst.com wrote: We are missing back pressure channels from uvm to the buf cache. The buf cache will happily sit on 9000 free pages while uvm churns around trying to scavenge up one more page. Or are you in a situation here where the cache has *not* backed off? Talked to Bob and hashed out better ideas of the problem. The page daemon does tell the buffer cache to make some room, but... If you have a huge mmap file, the pdaemon will try to flush it out via VOP_WRITE, which circles back via ffs into buf_get, which eats those previously freed pages, and then some, as the pagedaemon continues pushing more and more of the mmap file out. We discussed some other changes and fixes that this situation has clearly highlighted, but here's a slightly revised diff. It now uses the correct bufbackoff() function to communicate uvm's needs. Any other fix is rather precarious for this release, but as stated before, this keeps the change to the deadlock paths. You were already dead, but now you have a second chance. (We don't currently use the pmemrange argument; we'll have to adjust accordingly when the bufcache becomes range aware.) Index: uvm_pdaemon.c === RCS file: /cvs/src/sys/uvm/uvm_pdaemon.c,v retrieving revision 1.64 diff -u -p -r1.64 uvm_pdaemon.c --- uvm_pdaemon.c 30 May 2013 16:29:46 - 1.64 +++ uvm_pdaemon.c 6 Feb 2014 03:09:53 - @@ -117,6 +117,8 @@ uvm_wait(const char *wmsg) */ if (curproc == uvm.pagedaemon_proc) { + if (bufbackoff(NULL, 4) == 0) + return; /* * now we have a problem: the pagedaemon wants to go to * sleep until it frees more memory. but how can it Index: uvm_pmemrange.c === RCS file: /cvs/src/sys/uvm/uvm_pmemrange.c,v retrieving revision 1.36 diff -u -p -r1.36 uvm_pmemrange.c --- uvm_pmemrange.c 29 Jan 2013 19:55:48 - 1.36 +++ uvm_pmemrange.c 6 Feb 2014 03:10:32 - @@ -22,6 +22,7 @@ #include sys/malloc.h #include sys/proc.h /* XXX for atomic */ #include sys/kernel.h +#include sys/mount.h /* * 2 trees: addr tree and size tree. @@ -1883,6 +1884,13 @@ uvm_wait_pla(paddr_t low, paddr_t high, const char *wmsg = pmrwait; if (curproc == uvm.pagedaemon_proc) { + uvm_unlock_fpageq(); + if (bufbackoff(NULL, atop(size)) == 0) { + uvm_lock_fpageq(); + return 0; + } + uvm_lock_fpageq(); + /* * XXX detect pagedaemon deadlock - see comment in * uvm_wait(), as this is exactly the same issue.