Re: svn commit: r202894 - in head/sys/i386: i386 include
Quoting John Baldwin j...@freebsd.org (from Tue, 26 Jan 2010 15:07:21 -0500): On Tuesday 26 January 2010 12:05:29 pm Ivan Voras wrote: 2010/1/26 Alexander Leidinger alexan...@leidinger.net: Quoting John Baldwin j...@freebsd.org (from Mon, 25 Jan 2010 07:57:49 -0500): On Saturday 23 January 2010 1:42:28 pm Alan Cox wrote: Author: alc Date: Sat Jan 23 18:42:28 2010 New Revision: 202894 URL: http://svn.freebsd.org/changeset/base/202894 Log: Handle a race between pmap_kextract() and pmap_promote_pde(). This race is known to cause a kernel crash in ZFS on i386 when superpage promotion is enabled. Tested by:netchild MFC after:1 week Yay! Should we enable pg_ps by default on i386 now? Data point, the i386 system where I have superpages enabled is a p4 with currently 11 jails active. Some webservers, squid, samba, bind, postfix and a lot of portupgrades (during the last days, I'm updating a jail with a desktop-setup inside). All jails are on ZFS. Do you have any method of comparing performance before-and-after? There was a thread on some of the mailing lists (stable@/curr...@?) about squid being the pessimal case for superpages. The thread did not say that superpages made squid slower, rather that due to the way it forked to execute the log rotation binaries, some superpages were demoted back to regular pages losing the gain of superpages for those pages. Additionally: - this is my server at home (limited amount of users accessing it) - I changed too much at once which may affect the performance (superpages, ATA_CAM, zfs-kernel-settings) Apart from that: I have no performance numbers for this machine. Bye, Alexander. -- Fresco's Discovery: If you knew what you were doing you'd probably be bored. http://www.Leidinger.netAlexander @ Leidinger.net: PGP ID = B0063FE7 http://www.FreeBSD.org netchild @ FreeBSD.org : PGP ID = 72077137 ___ svn-src-all@freebsd.org mailing list http://lists.freebsd.org/mailman/listinfo/svn-src-all To unsubscribe, send any mail to svn-src-all-unsubscr...@freebsd.org
Re: svn commit: r202894 - in head/sys/i386: i386 include
Quoting John Baldwin j...@freebsd.org (from Mon, 25 Jan 2010 07:57:49 -0500): On Saturday 23 January 2010 1:42:28 pm Alan Cox wrote: Author: alc Date: Sat Jan 23 18:42:28 2010 New Revision: 202894 URL: http://svn.freebsd.org/changeset/base/202894 Log: Handle a race between pmap_kextract() and pmap_promote_pde(). This race is known to cause a kernel crash in ZFS on i386 when superpage promotion is enabled. Tested by:netchild MFC after:1 week Yay! Should we enable pg_ps by default on i386 now? Data point, the i386 system where I have superpages enabled is a p4 with currently 11 jails active. Some webservers, squid, samba, bind, postfix and a lot of portupgrades (during the last days, I'm updating a jail with a desktop-setup inside). All jails are on ZFS. # uptime 2:01pm up 3 days, 21:52, 2 users, load averages: 1.34, 1.14, 1.09 It is not more because I switched to ATA_CAM after more than a day of uptime with the code of this commit. # sysctl vm.pmap vm.pmap.pmap_collect_active: 0 vm.pmap.pmap_collect_inactive: 0 vm.pmap.pv_entry_spare: 54584 vm.pmap.pv_entry_allocs: 504637278 vm.pmap.pv_entry_frees: 504575270 vm.pmap.pc_chunk_tryfail: 0 vm.pmap.pc_chunk_frees: 22566325 vm.pmap.pc_chunk_allocs: 22566672 vm.pmap.pc_chunk_count: 347 vm.pmap.pv_entry_count: 62008 vm.pmap.pde.promotions: 2080074 vm.pmap.pde.p_failures: 22736249 vm.pmap.pde.mappings: 0 vm.pmap.pde.demotions: 2080038 vm.pmap.shpgperproc: 200 vm.pmap.pv_entry_max: 1487136 vm.pmap.pg_ps_enabled: 1 Bye, Alexander. -- When eating an elephant take one bite at a time. http://www.Leidinger.netAlexander @ Leidinger.net: PGP ID = B0063FE7 http://www.FreeBSD.org netchild @ FreeBSD.org : PGP ID = 72077137 ___ svn-src-all@freebsd.org mailing list http://lists.freebsd.org/mailman/listinfo/svn-src-all To unsubscribe, send any mail to svn-src-all-unsubscr...@freebsd.org
Re: svn commit: r202894 - in head/sys/i386: i386 include
2010/1/26 Alexander Leidinger alexan...@leidinger.net: Quoting John Baldwin j...@freebsd.org (from Mon, 25 Jan 2010 07:57:49 -0500): On Saturday 23 January 2010 1:42:28 pm Alan Cox wrote: Author: alc Date: Sat Jan 23 18:42:28 2010 New Revision: 202894 URL: http://svn.freebsd.org/changeset/base/202894 Log: Handle a race between pmap_kextract() and pmap_promote_pde(). This race is known to cause a kernel crash in ZFS on i386 when superpage promotion is enabled. Tested by: netchild MFC after: 1 week Yay! Should we enable pg_ps by default on i386 now? Data point, the i386 system where I have superpages enabled is a p4 with currently 11 jails active. Some webservers, squid, samba, bind, postfix and a lot of portupgrades (during the last days, I'm updating a jail with a desktop-setup inside). All jails are on ZFS. Do you have any method of comparing performance before-and-after? There was a thread on some of the mailing lists (stable@/curr...@?) about squid being the pessimal case for superpages. ___ svn-src-all@freebsd.org mailing list http://lists.freebsd.org/mailman/listinfo/svn-src-all To unsubscribe, send any mail to svn-src-all-unsubscr...@freebsd.org
Re: svn commit: r202894 - in head/sys/i386: i386 include
On Tuesday 26 January 2010 12:05:29 pm Ivan Voras wrote: 2010/1/26 Alexander Leidinger alexan...@leidinger.net: Quoting John Baldwin j...@freebsd.org (from Mon, 25 Jan 2010 07:57:49 -0500): On Saturday 23 January 2010 1:42:28 pm Alan Cox wrote: Author: alc Date: Sat Jan 23 18:42:28 2010 New Revision: 202894 URL: http://svn.freebsd.org/changeset/base/202894 Log: Handle a race between pmap_kextract() and pmap_promote_pde(). This race is known to cause a kernel crash in ZFS on i386 when superpage promotion is enabled. Tested by:netchild MFC after:1 week Yay! Should we enable pg_ps by default on i386 now? Data point, the i386 system where I have superpages enabled is a p4 with currently 11 jails active. Some webservers, squid, samba, bind, postfix and a lot of portupgrades (during the last days, I'm updating a jail with a desktop-setup inside). All jails are on ZFS. Do you have any method of comparing performance before-and-after? There was a thread on some of the mailing lists (stable@/curr...@?) about squid being the pessimal case for superpages. The thread did not say that superpages made squid slower, rather that due to the way it forked to execute the log rotation binaries, some superpages were demoted back to regular pages losing the gain of superpages for those pages. -- John Baldwin ___ svn-src-all@freebsd.org mailing list http://lists.freebsd.org/mailman/listinfo/svn-src-all To unsubscribe, send any mail to svn-src-all-unsubscr...@freebsd.org
Re: svn commit: r202894 - in head/sys/i386: i386 include
On Saturday 23 January 2010 1:42:28 pm Alan Cox wrote: Author: alc Date: Sat Jan 23 18:42:28 2010 New Revision: 202894 URL: http://svn.freebsd.org/changeset/base/202894 Log: Handle a race between pmap_kextract() and pmap_promote_pde(). This race is known to cause a kernel crash in ZFS on i386 when superpage promotion is enabled. Tested by: netchild MFC after: 1 week Yay! Should we enable pg_ps by default on i386 now? -- John Baldwin ___ svn-src-all@freebsd.org mailing list http://lists.freebsd.org/mailman/listinfo/svn-src-all To unsubscribe, send any mail to svn-src-all-unsubscr...@freebsd.org
Re: svn commit: r202894 - in head/sys/i386: i386 include
John Baldwin wrote: On Saturday 23 January 2010 1:42:28 pm Alan Cox wrote: Author: alc Date: Sat Jan 23 18:42:28 2010 New Revision: 202894 URL: http://svn.freebsd.org/changeset/base/202894 Log: Handle a race between pmap_kextract() and pmap_promote_pde(). This race is known to cause a kernel crash in ZFS on i386 when superpage promotion is enabled. Tested by: netchild MFC after:1 week Yay! Should we enable pg_ps by default on i386 now? Yes, I think so. Alan ___ svn-src-all@freebsd.org mailing list http://lists.freebsd.org/mailman/listinfo/svn-src-all To unsubscribe, send any mail to svn-src-all-unsubscr...@freebsd.org
svn commit: r202894 - in head/sys/i386: i386 include
Author: alc
Date: Sat Jan 23 18:42:28 2010
New Revision: 202894
URL: http://svn.freebsd.org/changeset/base/202894
Log:
Handle a race between pmap_kextract() and pmap_promote_pde(). This race is
known to cause a kernel crash in ZFS on i386 when superpage promotion is
enabled.
Tested by:netchild
MFC after:1 week
Modified:
head/sys/i386/i386/pmap.c
head/sys/i386/include/pmap.h
Modified: head/sys/i386/i386/pmap.c
==
--- head/sys/i386/i386/pmap.c Sat Jan 23 17:58:40 2010(r202893)
+++ head/sys/i386/i386/pmap.c Sat Jan 23 18:42:28 2010(r202894)
@@ -243,8 +243,9 @@ struct sysmaps {
caddr_t CADDR2;
};
static struct sysmaps sysmaps_pcpu[MAXCPU];
-pt_entry_t *CMAP1 = 0;
+pt_entry_t *CMAP1 = 0, *KPTmap;
static pt_entry_t *CMAP3;
+static pd_entry_t *KPTD;
caddr_t CADDR1 = 0, ptvmmap = 0;
static caddr_t CADDR3;
struct msgbuf *msgbufp = 0;
@@ -421,6 +422,21 @@ pmap_bootstrap(vm_paddr_t firstaddr)
SYSMAP(struct msgbuf *, unused, msgbufp, atop(round_page(MSGBUF_SIZE)))
/*
+* KPTmap is used by pmap_kextract().
+*/
+ SYSMAP(pt_entry_t *, KPTD, KPTmap, KVA_PAGES)
+
+ for (i = 0; i NKPT; i++)
+ KPTD[i] = (KPTphys + (i PAGE_SHIFT)) | PG_RW | PG_V;
+
+ /*
+* Adjust the start of the KPTD and KPTmap so that the implementation
+* of pmap_kextract() and pmap_growkernel() can be made simpler.
+*/
+ KPTD -= KPTDI;
+ KPTmap -= i386_btop(KPTDI PDRSHIFT);
+
+ /*
* ptemap is used for pmap_pte_quick
*/
SYSMAP(pt_entry_t *, PMAP1, PADDR1, 1);
@@ -1839,6 +1855,7 @@ pmap_growkernel(vm_offset_t addr)
vm_page_t nkpg;
pd_entry_t newpdir;
pt_entry_t *pde;
+ boolean_t updated_PTD;
mtx_assert(kernel_map-system_mtx, MA_OWNED);
if (kernel_vm_end == 0) {
@@ -1878,14 +1895,20 @@ pmap_growkernel(vm_offset_t addr)
pmap_zero_page(nkpg);
ptppaddr = VM_PAGE_TO_PHYS(nkpg);
newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M);
- pdir_pde(PTD, kernel_vm_end) = newpdir;
+ pdir_pde(KPTD, kernel_vm_end) = newpdir;
+ updated_PTD = FALSE;
mtx_lock_spin(allpmaps_lock);
LIST_FOREACH(pmap, allpmaps, pm_list) {
+ if ((pmap-pm_pdir[PTDPTDI] PG_FRAME) == (PTDpde[0]
+ PG_FRAME))
+ updated_PTD = TRUE;
pde = pmap_pde(pmap, kernel_vm_end);
pde_store(pde, newpdir);
}
mtx_unlock_spin(allpmaps_lock);
+ KASSERT(updated_PTD,
+ (pmap_growkernel: current page table is not in allpmaps));
kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG)
~(PAGE_SIZE * NPTEPG - 1);
if (kernel_vm_end - 1 = kernel_map-max_offset) {
kernel_vm_end = kernel_map-max_offset;
Modified: head/sys/i386/include/pmap.h
==
--- head/sys/i386/include/pmap.hSat Jan 23 17:58:40 2010
(r202893)
+++ head/sys/i386/include/pmap.hSat Jan 23 18:42:28 2010
(r202894)
@@ -265,6 +265,16 @@ pte_load_store_ma(pt_entry_t *ptep, pt_e
#definepde_store_ma(ptep, pte) pte_load_store_ma((ptep),
(pt_entry_t)pte)
#elif !defined(XEN)
+
+/*
+ * KPTmap is a linear mapping of the kernel page table. It differs from the
+ * recursive mapping in two ways: (1) it only provides access to kernel page
+ * table pages, and not user page table pages, and (2) it provides access to
+ * a kernel page table page after the corresponding virtual addresses have
+ * been promoted to a 2/4MB page mapping.
+ */
+extern pt_entry_t *KPTmap;
+
/*
* Routine:pmap_kextract
* Function:
@@ -279,10 +289,17 @@ pmap_kextract(vm_offset_t va)
if ((pa = PTD[va PDRSHIFT]) PG_PS) {
pa = (pa PG_PS_FRAME) | (va PDRMASK);
} else {
- pa = *vtopte(va);
+ /*
+* Beware of a concurrent promotion that changes the PDE at
+* this point! For example, vtopte() must not be used to
+* access the PTE because it would use the new PDE. It is,
+* however, safe to use the old PDE because the page table
+* page is preserved by the promotion.
+*/
+ pa = KPTmap[i386_btop(va)];
pa = (pa PG_FRAME) | (va PAGE_MASK);
}
- return pa;
+ return (pa);
}
#define PT_UPDATES_FLUSH()
___
svn-src-all@freebsd.org mailing list
http://lists.freebsd.org/mailman/listinfo/svn-src-all
To unsubscribe, send any mail to
