Patch Name: PHKL_34926

Patch Description: s700_800 11.11 Buffer cache performance improvement patch

Creation Date: 06/05/30

Post Date: 06/09/15

Hardware Platforms - OS Releases:
	s700: 11.11
	s800: 11.11

Products: N/A

Filesets:
	OS-Core.CORE2-KRN,fr=B.11.11,fa=HP-UX_B.11.11_32,v=HP
	OS-Core.CORE2-KRN,fr=B.11.11,fa=HP-UX_B.11.11_64,v=HP

Automatic Reboot?: Yes

Status: General Release

Critical:
	No (superseded patches were critical)
	PHKL_32165: HANG
	PHKL_27266: OTHER
		Defect leads to unmountable NFS filesystems.

	PHKL_26518: HANG
		Hang - Filesystem or Process Hang
	PHKL_23154: PANIC

Category Tags:
	defect_repair enhancement general_release critical panic
	halts_system

Path Name: /hp-ux_patches/s700_800/11.X/PHKL_34926

Symptoms:
	PHKL_34926:
	( SR:8606432767 CR:JAGaf92206 )
	After experiencing heavy memory usage, the Dynamic Buffer
	Cache was reduced from dbc_max_pct down to the dbc_min_pct.
	However, after memory was freed up, the Dynamic Buffer Cache
	would not grow back to dbc_max_pct and remained at
	dbc_min_pct, causing filesystem performance problems.

	PHKL_33367:
	( SR:8606384017 CR:JAGaf44208 )
	Unmounting a CacheFS may fail with the device busy error.

	PHKL_32165:
	( SR:8606369345 CR:JAGaf29889 )
	Applications that read a file from filesystems larger
	than or equal to 2TB or read block devices larger than
	or equal to 2TB will hang if the device block address
	where the data resides is close to the 2TB area on the
	block device.

	PHKL_30920:
	( SR:8606312681 CR:JAGae75497 )
	syncer(1M) consumes a lot of CPU time.  A top report will
	indicate a 90 plus percentage usage of CPU.

	PHKL_27266:
	( SR:8606203127 CR:JAGad72301 )
	NFS File Systems fail to unmount with error "NFS umount
	failed: device busy". The File System can stay in this
	state forever or recover after a while.

	PHKL_27211:
	( SR:8606247419 CR:JAGae13857 )
	Enhancement:  This product update is a member of a set
	which improve the performance of mount(2) and umount(2).
	The full list of product updates that address this
	performance improvement are: PHKL_26239, PHKL_26694,
	PHKL_26695, PHKL_26698, and PHKL_27211.

	Each of these product updates may be installed independently
	of the others.  Each provides an incremental performance
	improvement.

	( SR:8606261834 CR:JAGae26152 )
	Enhancement: Improve performance of final close.
	In particular, this improves fsck(1M) execution time.

	PHKL_26518:
	( SR:8606251259 CR:JAGae17325 )
	( SR:8606241742 CR:JAGae09000 )
	Processes may hang related to Filesystem access.

	PHKL_24071:
	( SR:8606173812 CR:JAGad43065 )
	Enhancement: Provide real time (RTSCHED) priority scheduling
	of threads. This is being provided because certain time
	sensitive applications were aborting (timing out), due
	to some long running kernel paths.

	PHKL_23154:
	( SR:8606173404 CR:JAGad42662 )
	System may panic when using Instant Capacity On Demand
	(iCOD) to do instant processor deactivation.  The stack
	trace reads:

	    panic+0x6c
	    report_trap_or_int_and_panic+0x94
	    trap+0xed4
	    nokgdb+0x8
	    dflush+0x88
	    sync_buffer_cache+0x14
	    tsync+0xe8
	    syscall+0x750
	    $syscallrtn+0x0

Defect Description:
	PHKL_34926:
	( SR:8606432767 CR:JAGaf92206 )
	The buffer cache routines are unable to reuse the buffer
	headers on the empty queue when they try to obtain a new
	buffer. Hence, expansion of the buffer cache fails, and an
	existing buffer is reused.

	Resolution:
	The buffers on the empty queue are now properly reused so
	that the buffer cache can be expanded.

	PHKL_33367:
	( SR:8606384017 CR:JAGaf44208 )
	The unmount failure is caused by not releasing a hold on
	the CacheFS vnode when the buffer associated with the
	CacheFS was released by the garbage collector.

	Resolution:
	Kernel changes are made to release the hold on the CacheFS
	vnode appropriately.

	PHKL_32165:
	( SR:8606369345 CR:JAGaf29889 )
	The defect is caused due to an integer overflow during
	detection of buffer overlaps that causes the kernel routine
	to go into an infinite loop and become unkillable.

	Resolution:
	variable used in detecting buffer overlaps was changed
	from a signed 32-bit to unsigned 32-bit integer.

	PHKL_30920:
	( SR:8606312681 CR:JAGae75497 )
	syncer(1M) consumes a lot of CPU time.  This happens as the
	thread does not sleep at all.  The syncer(1M) is designed
	such that each thread will sleep for a calculated period of
	time after flushing it's respective SPU list.  However,
	under some circumstances, we could have many buffers which
	are marked as B_BUSY, idling in the sync queue.  Due to this
	, the time computation routine of the syncer(1M) computes
	the sleep time as zero.  So, the thread wakes up immediatly.

	Resolution:
	Modified the time calculation routine to exclude B_BUSY
	buffers.

	PHKL_27266:
	( SR:8606203127 CR:JAGad72301 )
	The umount failures occur because some vnodes in the
	NFS filesystem have active references which cannot be
	removed at the time of umount. These references are held
	by buffers which have been removed inappropriately from
	the list that NFS uses to manage buffer flush and
	buffer-to-vnode references. The vnode reference of the
	buffers is not being reduced before removing the buffers
	from the list. The umounts will continue to fail
	indefinitely until all such buffers are re-used.

	Resolution:
	The fix was to release the vnode references held by
	the buffers when they are being removed from the
	list.

	PHKL_27211:
	( SR:8606247419 CR:JAGae13857 )
	This product update contains minor changes necessary to
	improve the performance of mount(2) and umount(2) for all
	filesystems.  We incorrectly held a lock to serialize all
	access to all filesystems for mount/umount service calls.
	These filesystems specific operations are already serialized
	where necessary, and the additional higher level
	serialization results in unnecessary performance
	degradation.

	Resolution:
	Removed the unnecessary use of a filesystem lock that used
	to unconditionally serialize all access to all filesystems
	for mount/umount service calls.

	( SR:8606261834 CR:JAGae26152 )
	1. All of buffer cache is searched for buffers associated
	   with the vnode being closed, even when there are none.
	2. After each buffer is inspected, we check to see if
	   it is time for a kernel preemption, in a very
	   expensive manner.

	Resolution:
	1. Before starting to loop through buffer cache for
	   the vnode, verify that the vnode has at least one
	   buffer associated with it.
	2. Use a less expensive way to check if it is time
	   for a kernel preemption.

	PHKL_26518:
	( SR:8606251259 CR:JAGae17325 )
	( SR:8606241742 CR:JAGae09000 )
	The problem is caused by a thread/process waiting on a
	buffer that it incorrectly believes is still owned by
	another process.  The previous process failed to update the
	internal data structure used to show that the buffer is no
	longer in use.  The new process sees this stale information,
	goes to sleep believing that the buffer is still in use, and
	never wakes up, which results in a filesystem hang.  This
	was seen with a process accessing a VxFS 3.3 filessytem with
	Veritas NetBackup running.  However, this defect is not
	specific to any one filesystem; it applies to all block
	filesystem access.

	Resolution:
	Change the code to modify the internal data before releasing
	the buffer and doing the wakeup.

	PHKL_24071:
	( SR:8606173812 CR:JAGad43065 )
	This is an enhancement to provide realtime (RTSCHED)
	scheduling of threads. Threads that were executing long
	running kernel paths kept running on the CPU while there
	were other threads waiting. If the waiting threads were
	time sensitive, they would abort without receiving any
	CPU time.

	Resolution:
	Long running kernel code paths were identified and
	preemption points were introduced in these paths. At these
	preemption points if there are any realtime threads waiting
	then the current thread is switched out and the real time
	thread is given an opportunity to run on the CPU.

	PHKL_23154:
	( SR:8606173404 CR:JAGad42662 )
	When a syncer thread is processing a dirty buffer list, and
	it gets preempted for a long time (>30 seconds) because of
	the instant processor deactivation activities, a new thread
	could be started by the syncer to process the same dirty
	list.  If the new syncer thread cleans up some buffers from
	the dirty list, the old syncer thread may still have an old
	link address of the dirty list which has been set to NULL by
	the new syncer thread.  The operating system will panic when
	the old syncer thread resumes and tries to access the NULL
	address.

	Resolution:
	The syncer will now validate an address before referencing
	it, and terminate if the address is NULL.

Enhancement:
	No (superseded patches contained enhancements)
	PHKL_30920:
		Enhancements were delivered in a patch this one has
		superseded.  Please review the Defect Description
		text for more information.

SR:
	8606432767 8606173404 8606173812 8606203127 8606241742
	8606247419 8606251259 8606261834 8606312681 8606369345
	8606384017

Patch Files:
	
	OS-Core.CORE2-KRN,fr=B.11.11,fa=HP-UX_B.11.11_32,v=HP:
	/usr/conf/lib/libfs.a(vfs_bio.o)

	OS-Core.CORE2-KRN,fr=B.11.11,fa=HP-UX_B.11.11_64,v=HP:
	/usr/conf/lib/libfs.a(vfs_bio.o)

what(1) Output:
	
	OS-Core.CORE2-KRN,fr=B.11.11,fa=HP-UX_B.11.11_32,v=HP:
	/usr/conf/lib/libfs.a(vfs_bio.o):
		vfs_bio.c $Date: 2006/05/21 23:02:08 $Revision: r11.
			11/10 PATCH_11.11 (PHKL_34926)

	OS-Core.CORE2-KRN,fr=B.11.11,fa=HP-UX_B.11.11_64,v=HP:
	/usr/conf/lib/libfs.a(vfs_bio.o):
		vfs_bio.c $Date: 2006/05/21 23:02:08 $Revision: r11.
			11/10 PATCH_11.11 (PHKL_34926)

cksum(1) Output:
	
	OS-Core.CORE2-KRN,fr=B.11.11,fa=HP-UX_B.11.11_32,v=HP:
	2280099771 44392 /usr/conf/lib/libfs.a(vfs_bio.o)

	OS-Core.CORE2-KRN,fr=B.11.11,fa=HP-UX_B.11.11_64,v=HP:
	1569716107 102280 /usr/conf/lib/libfs.a(vfs_bio.o)

Patch Conflicts: None

Patch Dependencies: None

Hardware Dependencies: None

Other Dependencies: None

Supersedes:
	PHKL_33367 PHKL_32165 PHKL_30920 PHKL_27266 PHKL_27211 PHKL_26518
	PHKL_24071 PHKL_23154

Equivalent Patches:
	PHKL_30709:
	s700: 11.00
	s800: 11.00

Patch Package Size: 90 KBytes

Installation Instructions:
	Please review all instructions and the Hewlett-Packard
	SupportLine User Guide or your Hewlett-Packard support terms
	and conditions for precautions, scope of license,
	restrictions, and, limitation of liability and warranties,
	before installing this patch.
	------------------------------------------------------------
	1. Back up your system before installing a patch.

	2. Login as root.

	3. Copy the patch to the /tmp directory.

	4. Move to the /tmp directory and unshar the patch:

		cd /tmp
		sh PHKL_34926

	5. Run swinstall to install the patch:

		swinstall -x autoreboot=true -x patch_match_target=true \
			  -s /tmp/PHKL_34926.depot

	By default swinstall will archive the original software in 
	/var/adm/sw/save/PHKL_34926.  If you do not wish to retain a
	copy of the original software, include the patch_save_files
	option in the swinstall command above:

		-x patch_save_files=false

	WARNING: If patch_save_files is false when a patch is installed,
		 the patch cannot be deinstalled.  Please be careful
		 when using this feature.

	For future reference, the contents of the PHKL_34926.text file is 
	available in the product readme:

		swlist -l product -a readme -d @ /tmp/PHKL_34926.depot

	To put this patch on a magnetic tape and install from the
	tape drive, use the command:

		dd if=/tmp/PHKL_34926.depot of=/dev/rmt/0m bs=2k

Special Installation Instructions: None