platform/kernel/linux-amlogic.git
6 years agoscsi: zfcp: fix missing trace records for early returns in TMF eh handlers
Steffen Maier [Fri, 28 Jul 2017 10:30:55 +0000 (12:30 +0200)]
scsi: zfcp: fix missing trace records for early returns in TMF eh handlers

commit 1a5d999ebfc7bfe28deb48931bb57faa8e4102b6 upstream.

For problem determination we need to see that we were in scsi_eh
as well as whether and why we were successful or not.

The following commits introduced new early returns without adding
a trace record:

v2.6.35 commit a1dbfddd02d2
("[SCSI] zfcp: Pass return code from fc_block_scsi_eh to scsi eh")
on fc_block_scsi_eh() returning != 0 which is FAST_IO_FAIL,

v2.6.30 commit 63caf367e1c9
("[SCSI] zfcp: Improve reliability of SCSI eh handlers in zfcp")
on not having gotten an FSF request after the maximum number of retry
attempts and thus could not issue a TMF and has to return FAILED.

Signed-off-by: Steffen Maier <maier@linux.vnet.ibm.com>
Fixes: a1dbfddd02d2 ("[SCSI] zfcp: Pass return code from fc_block_scsi_eh to scsi eh")
Fixes: 63caf367e1c9 ("[SCSI] zfcp: Improve reliability of SCSI eh handlers in zfcp")
Reviewed-by: Benjamin Block <bblock@linux.vnet.ibm.com>
Signed-off-by: Benjamin Block <bblock@linux.vnet.ibm.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoscsi: zfcp: fix passing fsf_req to SCSI trace on TMF to correlate with HBA
Steffen Maier [Fri, 28 Jul 2017 10:30:54 +0000 (12:30 +0200)]
scsi: zfcp: fix passing fsf_req to SCSI trace on TMF to correlate with HBA

commit 9fe5d2b2fd30aa8c7827ec62cbbe6d30df4fe3e3 upstream.

Without this fix we get SCSI trace records on task management functions
which cannot be correlated to HBA trace records because all fields
related to the FSF request are empty (zero).
Also, the FCP_RSP_IU is missing as well as any sense data if available.

This was caused by v2.6.14 commit 8a36e4532ea1 ("[SCSI] zfcp: enhancement
of zfcp debug features") introducing trace records for TMFs but
hard coding NULL for a possibly existing TMF FSF request.
The scsi_cmnd scribble is also zero or unrelated for the TMF request
so it also could not lookup a suitable FSF request from there.

A broken example trace record formatted with zfcpdbf from the s390-tools
package:

Timestamp      : ...
Area           : SCSI
Subarea        : 00
Level          : 1
Exception      : -
CPU ID         : ..
Caller         : 0x...
Record ID      : 1
Tag            : lr_fail
Request ID     : 0x0000000000000000
                   ^^^^^^^^^^^^^^^^ no correlation to HBA record
SCSI ID        : 0x<scsitarget>
SCSI LUN       : 0x<scsilun>
SCSI result    : 0x000e0000
SCSI retries   : 0x00
SCSI allowed   : 0x05
SCSI scribble  : 0x0000000000000000
SCSI opcode    : 2a000017 3bb80000 08000000 00000000
FCP rsp inf cod: 0x00
                   ^^ no TMF response
FCP rsp IU     : 00000000 00000000 00000000 00000000
                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                 00000000 00000000
                 ^^^^^^^^^^^^^^^^^ no interesting FCP_RSP_IU
Sense len      : ...
^^^^^^^^^^^^^^^^^^^^ no sense data length
Sense info     : ...
^^^^^^^^^^^^^^^^^^^^ no sense data content, even if present

There are some true cases where we really do not have an FSF request:
"rsl_fai" from zfcp_dbf_scsi_fail_send() called for early
returns / completions in zfcp_scsi_queuecommand(),
"abrt_or", "abrt_bl", "abrt_ru", "abrt_ar" from
zfcp_scsi_eh_abort_handler() where we did not get as far,
"lr_nres", "tr_nres" from zfcp_task_mgmt_function() where we're
successful and do not need to do anything because adapter stopped.
For these cases it's correct to pass NULL for fsf_req to _zfcp_dbf_scsi().

Signed-off-by: Steffen Maier <maier@linux.vnet.ibm.com>
Fixes: 8a36e4532ea1 ("[SCSI] zfcp: enhancement of zfcp debug features")
Reviewed-by: Benjamin Block <bblock@linux.vnet.ibm.com>
Signed-off-by: Benjamin Block <bblock@linux.vnet.ibm.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoscsi: zfcp: fix capping of unsuccessful GPN_FT SAN response trace records
Steffen Maier [Fri, 28 Jul 2017 10:30:53 +0000 (12:30 +0200)]
scsi: zfcp: fix capping of unsuccessful GPN_FT SAN response trace records

commit 975171b4461be296a35e83ebd748946b81cf0635 upstream.

v4.9 commit aceeffbb59bb ("zfcp: trace full payload of all SAN records
(req,resp,iels)") fixed trace data loss of 2.6.38 commit 2c55b750a884
("[SCSI] zfcp: Redesign of the debug tracing for SAN records.")
necessary for problem determination, e.g. to see the
currently active zone set during automatic port scan.

While it already saves space by not dumping any empty residual entries
of the large successful GPN_FT response (4 pages), there are seldom cases
where the GPN_FT response is unsuccessful and likely does not have
FC_NS_FID_LAST set in fp_flags so we did not cap the trace record.
We typically see such case for an initiator WWPN, which is not in any zone.

Cap unsuccessful responses to at least the actual basic CT_IU response
plus whatever fits the SAN trace record built-in "payload" buffer
just in case there's trailing information
of which we would at least see the existence and its beginning.

In order not to erroneously cap successful responses, we need to swap
calling the trace function and setting the CT / ELS status to success (0).

Example trace record pair formatted with zfcpdbf:

Timestamp      : ...
Area           : SAN
Subarea        : 00
Level          : 1
Exception      : -
CPU ID         : ..
Caller         : 0x...
Record ID      : 1
Tag            : fssct_1
Request ID     : 0x<request_id>
Destination ID : 0x00fffffc
SAN req short  : 01000000 fc020000 01720ffc 00000000
                 00000008
SAN req length : 20
|
Timestamp      : ...
Area           : SAN
Subarea        : 00
Level          : 1
Exception      : -
CPU ID         : ..
Caller         : 0x...
Record ID      : 2
Tag            : fsscth2
Request ID     : 0x<request_id>
Destination ID : 0x00fffffc
SAN resp short : 01000000 fc020000 80010000 00090700
                 00000000 00000000 00000000 00000000 [trailing info]
                 00000000 00000000 00000000 00000000 [trailing info]
SAN resp length: 16384
San resp info  : 01000000 fc020000 80010000 00090700
                 00000000 00000000 00000000 00000000 [trailing info]
                 00000000 00000000 00000000 00000000 [trailing info]
                 00000000 00000000 00000000 00000000 [trailing info]
                 00000000 00000000 00000000 00000000 [trailing info]
                 00000000 00000000 00000000 00000000 [trailing info]
                 00000000 00000000 00000000 00000000 [trailing info]
                 00000000 00000000 00000000 00000000 [trailing info]
                 00000000 00000000 00000000 00000000 [trailing info]
                 00000000 00000000 00000000 00000000 [trailing info]
                 00000000 00000000 00000000 00000000 [trailing info]
                 00000000 00000000 00000000 00000000 [trailing info]
                 00000000 00000000 00000000 00000000 [trailing info]
                 00000000 00000000 00000000 00000000 [trailing info]
                 00000000 00000000 00000000 00000000 [trailing info]
                 00000000 00000000 00000000 00000000 [trailing info]

The fix saves all but one of the previously associated 64 PAYload trace
record chunks of size 256 bytes each.

Signed-off-by: Steffen Maier <maier@linux.vnet.ibm.com>
Fixes: aceeffbb59bb ("zfcp: trace full payload of all SAN records (req,resp,iels)")
Fixes: 2c55b750a884 ("[SCSI] zfcp: Redesign of the debug tracing for SAN records.")
Reviewed-by: Benjamin Block <bblock@linux.vnet.ibm.com>
Signed-off-by: Benjamin Block <bblock@linux.vnet.ibm.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoscsi: zfcp: add handling for FCP_RESID_OVER to the fcp ingress path
Benjamin Block [Fri, 28 Jul 2017 10:30:52 +0000 (12:30 +0200)]
scsi: zfcp: add handling for FCP_RESID_OVER to the fcp ingress path

commit a099b7b1fc1f0418ab8d79ecf98153e1e134656e upstream.

Up until now zfcp would just ignore the FCP_RESID_OVER flag in the FCP
response IU. When this flag is set, it is possible, in regards to the
FCP standard, that the storage-server processes the command normally, up
to the point where data is missing and simply ignores those.

In this case no CHECK CONDITION would be set, and because we ignored the
FCP_RESID_OVER flag we resulted in at least a data loss or even
-corruption as a follow-up error, depending on how the
applications/layers on top behave. To prevent this, we now set the
host-byte of the corresponding scsi_cmnd to DID_ERROR.

Other storage-behaviors, where the same condition results in a CHECK
CONDITION set in the answer, don't need to be changed as they are
handled in the mid-layer already.

Following is an example trace record decoded with zfcpdbf from the
s390-tools package. We forcefully injected a fc_dl which is one byte too
small:

Timestamp      : ...
Area           : SCSI
Subarea        : 00
Level          : 3
Exception      : -
CPU ID         : ..
Caller         : 0x...
Record ID      : 1
Tag            : rsl_err
Request ID     : 0x...
SCSI ID        : 0x...
SCSI LUN       : 0x...
SCSI result    : 0x00070000
                     ^^DID_ERROR
SCSI retries   : 0x..
SCSI allowed   : 0x..
SCSI scribble  : 0x...
SCSI opcode    : 2a000000 00000000 08000000 00000000
FCP rsp inf cod: 0x00
FCP rsp IU     : 00000000 00000000 00000400 00000001
                                       ^^fr_flags==FCP_RESID_OVER
                                         ^^fr_status==SAM_STAT_GOOD
                                            ^^^^^^^^fr_resid
                 00000000 00000000

As of now, we don't actively handle to possibility that a response IU
has both flags - FCP_RESID_OVER and FCP_RESID_UNDER - set at once.

Reported-by: Luke M. Hopkins <lmhopkin@us.ibm.com>
Reviewed-by: Steffen Maier <maier@linux.vnet.ibm.com>
Fixes: 553448f6c483 ("[SCSI] zfcp: Message cleanup")
Fixes: ea127f975424 ("[PATCH] s390 (7/7): zfcp host adapter.") (tglx/history.git)
Signed-off-by: Benjamin Block <bblock@linux.vnet.ibm.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoscsi: zfcp: fix queuecommand for scsi_eh commands when DIX enabled
Steffen Maier [Fri, 28 Jul 2017 10:30:51 +0000 (12:30 +0200)]
scsi: zfcp: fix queuecommand for scsi_eh commands when DIX enabled

commit 71b8e45da51a7b64a23378221c0a5868bd79da4f upstream.

Since commit db007fc5e20c ("[SCSI] Command protection operation"),
scsi_eh_prep_cmnd() saves scmd->prot_op and temporarily resets it to
SCSI_PROT_NORMAL.
Other FCP LLDDs such as qla2xxx and lpfc shield their queuecommand()
to only access any of scsi_prot_sg...() if
(scsi_get_prot_op(cmd) != SCSI_PROT_NORMAL).

Do the same thing for zfcp, which introduced DIX support with
commit ef3eb71d8ba4 ("[SCSI] zfcp: Introduce experimental support for
DIF/DIX").

Otherwise, TUR SCSI commands as part of scsi_eh likely fail in zfcp,
because the regular SCSI command with DIX protection data, that scsi_eh
re-uses in scsi_send_eh_cmnd(), of course still has
(scsi_prot_sg_count() != 0) and so zfcp sends down bogus requests to the
FCP channel hardware.

This causes scsi_eh_test_devices() to have (finish_cmds == 0)
[not SCSI device is online or not scsi_eh_tur() failed]
so regular SCSI commands, that caused / were affected by scsi_eh,
are moved to work_q and scsi_eh_test_devices() itself returns false.
In turn, it unnecessarily escalates in our case in scsi_eh_ready_devs()
beyond host reset to finally scsi_eh_offline_sdevs()
which sets affected SCSI devices offline with the following kernel message:

"kernel: sd H:0:T:L: Device offlined - not ready after error recovery"

Signed-off-by: Steffen Maier <maier@linux.vnet.ibm.com>
Fixes: ef3eb71d8ba4 ("[SCSI] zfcp: Introduce experimental support for DIF/DIX")
Reviewed-by: Benjamin Block <bblock@linux.vnet.ibm.com>
Signed-off-by: Benjamin Block <bblock@linux.vnet.ibm.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoskd: Submit requests to firmware before triggering the doorbell
Bart Van Assche [Thu, 17 Aug 2017 20:12:46 +0000 (13:12 -0700)]
skd: Submit requests to firmware before triggering the doorbell

commit 5fbd545cd3fd311ea1d6e8be4cedddd0ee5684c7 upstream.

Ensure that the members of struct skd_msg_buf have been transferred
to the PCIe adapter before the doorbell is triggered. This patch
avoids that I/O fails sporadically and that the following error
message is reported:

(skd0:STM000196603:[0000:00:09.0]): Completion mismatch comp_id=0x0000 skreq=0x0400 new=0x0000

Signed-off-by: Bart Van Assche <bart.vanassche@wdc.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Hannes Reinecke <hare@suse.de>
Cc: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoskd: Avoid that module unloading triggers a use-after-free
Bart Van Assche [Thu, 17 Aug 2017 20:12:45 +0000 (13:12 -0700)]
skd: Avoid that module unloading triggers a use-after-free

commit 7277cc67b3916eed47558c64f9c9c0de00a35cda upstream.

Since put_disk() triggers a disk_release() call and since that
last function calls blk_put_queue() if disk->queue != NULL, clear
the disk->queue pointer before calling put_disk(). This avoids
that unloading the skd kernel module triggers the following
use-after-free:

WARNING: CPU: 8 PID: 297 at lib/refcount.c:128 refcount_sub_and_test+0x70/0x80
refcount_t: underflow; use-after-free.
CPU: 8 PID: 297 Comm: kworker/8:1 Not tainted 4.11.10-300.fc26.x86_64 #1
Workqueue: events work_for_cpu_fn
Call Trace:
 dump_stack+0x63/0x84
 __warn+0xcb/0xf0
 warn_slowpath_fmt+0x5a/0x80
 refcount_sub_and_test+0x70/0x80
 refcount_dec_and_test+0x11/0x20
 kobject_put+0x1f/0x50
 blk_put_queue+0x15/0x20
 disk_release+0xae/0xf0
 device_release+0x32/0x90
 kobject_release+0x67/0x170
 kobject_put+0x2b/0x50
 put_disk+0x17/0x20
 skd_destruct+0x5c/0x890 [skd]
 skd_pci_probe+0x124d/0x13a0 [skd]
 local_pci_probe+0x42/0xa0
 work_for_cpu_fn+0x14/0x20
 process_one_work+0x19e/0x470
 worker_thread+0x1dc/0x4a0
 kthread+0x125/0x140
 ret_from_fork+0x25/0x30

Signed-off-by: Bart Van Assche <bart.vanassche@wdc.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Hannes Reinecke <hare@suse.de>
Cc: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agomd/bitmap: disable bitmap_resize for file-backed bitmaps.
NeilBrown [Thu, 31 Aug 2017 00:23:25 +0000 (10:23 +1000)]
md/bitmap: disable bitmap_resize for file-backed bitmaps.

commit e8a27f836f165c26f867ece7f31eb5c811692319 upstream.

bitmap_resize() does not work for file-backed bitmaps.
The buffer_heads are allocated and initialized when
the bitmap is read from the file, but resize doesn't
read from the file, it loads from the internal bitmap.
When it comes time to write the new bitmap, the bh is
non-existent and we crash.

The common case when growing an array involves making the array larger,
and that normally means making the bitmap larger.  Doing
that inside the kernel is possible, but would need more code.
It is probably easier to require people who use file-backed
bitmaps to remove them and re-add after a reshape.

So this patch disables the resizing of arrays which have
file-backed bitmaps.  This is better than crashing.

Reported-by: Zhilong Liu <zlliu@suse.com>
Fixes: d60b479d177a ("md/bitmap: add bitmap_resize function to allow bitmap resizing.")
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Shaohua Li <shli@fb.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoblock: Relax a check in blk_start_queue()
Bart Van Assche [Thu, 17 Aug 2017 20:12:44 +0000 (13:12 -0700)]
block: Relax a check in blk_start_queue()

commit 4ddd56b003f251091a67c15ae3fe4a5c5c5e390a upstream.

Calling blk_start_queue() from interrupt context with the queue
lock held and without disabling IRQs, as the skd driver does, is
safe. This patch avoids that loading the skd driver triggers the
following warning:

WARNING: CPU: 11 PID: 1348 at block/blk-core.c:283 blk_start_queue+0x84/0xa0
RIP: 0010:blk_start_queue+0x84/0xa0
Call Trace:
 skd_unquiesce_dev+0x12a/0x1d0 [skd]
 skd_complete_internal+0x1e7/0x5a0 [skd]
 skd_complete_other+0xc2/0xd0 [skd]
 skd_isr_completion_posted.isra.30+0x2a5/0x470 [skd]
 skd_isr+0x14f/0x180 [skd]
 irq_forced_thread_fn+0x2a/0x70
 irq_thread+0x144/0x1a0
 kthread+0x125/0x140
 ret_from_fork+0x2a/0x40

Fixes: commit a038e2536472 ("[PATCH] blk_start_queue() must be called with irq disabled - add warning")
Signed-off-by: Bart Van Assche <bart.vanassche@wdc.com>
Cc: Paolo 'Blaisorblade' Giarrusso <blaisorblade@yahoo.it>
Cc: Andrew Morton <akpm@osdl.org>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Hannes Reinecke <hare@suse.de>
Cc: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agopowerpc: Fix DAR reporting when alignment handler faults
Michael Ellerman [Thu, 24 Aug 2017 10:49:57 +0000 (20:49 +1000)]
powerpc: Fix DAR reporting when alignment handler faults

commit f9effe925039cf54489b5c04e0d40073bb3a123d upstream.

Anton noticed that if we fault part way through emulating an unaligned
instruction, we don't update the DAR to reflect that.

The DAR value is eventually reported back to userspace as the address
in the SEGV signal, and if userspace is using that value to demand
fault then it can be confused by us not setting the value correctly.

This patch is ugly as hell, but is intended to be the minimal fix and
back ports easily.

Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Reviewed-by: Paul Mackerras <paulus@ozlabs.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoext4: fix quota inconsistency during orphan cleanup for read-only mounts
zhangyi (F) [Thu, 24 Aug 2017 19:21:50 +0000 (15:21 -0400)]
ext4: fix quota inconsistency during orphan cleanup for read-only mounts

commit 95f1fda47c9d8738f858c3861add7bf0a36a7c0b upstream.

Quota does not get enabled for read-only mounts if filesystem
has quota feature, so that quotas cannot updated during orphan
cleanup, which will lead to quota inconsistency.

This patch turn on quotas during orphan cleanup for this case,
make sure quotas can be updated correctly.

Reported-by: Jan Kara <jack@suse.cz>
Signed-off-by: zhangyi (F) <yi.zhang@huawei.com>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoext4: fix incorrect quotaoff if the quota feature is enabled
zhangyi (F) [Thu, 24 Aug 2017 19:19:39 +0000 (15:19 -0400)]
ext4: fix incorrect quotaoff if the quota feature is enabled

commit b0a5a9589decd07db755d6a8d9c0910d96ff7992 upstream.

Current ext4 quota should always "usage enabled" if the
quota feautre is enabled. But in ext4_orphan_cleanup(), it
turn quotas off directly (used for the older journaled
quota), so we cannot turn it on again via "quotaon" unless
umount and remount ext4.

Simple reproduce:

  mkfs.ext4 -O project,quota /dev/vdb1
  mount -o prjquota /dev/vdb1 /mnt
  chattr -p 123 /mnt
  chattr +P /mnt
  touch /mnt/aa /mnt/bb
  exec 100<>/mnt/aa
  rm -f /mnt/aa
  sync
  echo c > /proc/sysrq-trigger

  #reboot and mount
  mount -o prjquota /dev/vdb1 /mnt
  #query status
  quotaon -Ppv /dev/vdb1
  #output
  quotaon: Cannot find mountpoint for device /dev/vdb1
  quotaon: No correct mountpoint specified.

This patch add check for journaled quotas to avoid incorrect
quotaoff when ext4 has quota feautre.

Signed-off-by: zhangyi (F) <yi.zhang@huawei.com>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agocrypto: AF_ALG - remove SGL terminator indicator when chaining
Stephan Mueller [Thu, 21 Sep 2017 08:16:53 +0000 (10:16 +0200)]
crypto: AF_ALG - remove SGL terminator indicator when chaining

Fixed differently upstream as commit 2d97591ef43d ("crypto: af_alg - consolidation of duplicate code")

The SGL is MAX_SGL_ENTS + 1 in size. The last SG entry is used for the
chaining and is properly updated with the sg_chain invocation. During
the filling-in of the initial SG entries, sg_mark_end is called for each
SG entry. This is appropriate as long as no additional SGL is chained
with the current SGL. However, when a new SGL is chained and the last
SG entry is updated with sg_chain, the last but one entry still contains
the end marker from the sg_mark_end. This end marker must be removed as
otherwise a walk of the chained SGLs will cause a NULL pointer
dereference at the last but one SG entry, because sg_next will return
NULL.

The patch only applies to all kernels up to and including 4.13. The
patch 2d97591ef43d0587be22ad1b0d758d6df4999a0b added to 4.14-rc1
introduced a complete new code base which addresses this bug in
a different way. Yet, that patch is too invasive for stable kernels
and was therefore not marked for stable.

Fixes: 8ff590903d5fc ("crypto: algif_skcipher - User-space interface for skcipher operations")
Signed-off-by: Stephan Mueller <smueller@chronox.de>
Acked-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agocrypto: ccp - Fix XTS-AES-128 support on v5 CCPs
Gary R Hook [Tue, 25 Jul 2017 19:12:11 +0000 (14:12 -0500)]
crypto: ccp - Fix XTS-AES-128 support on v5 CCPs

commit e652399edba99a5497f0d80f240c9075d3b43493 upstream.

Version 5 CCPs have some new requirements for XTS-AES: the type field
must be specified, and the key requires 512 bits, with each part
occupying 256 bits and padded with zeroes.

Signed-off-by: Gary R Hook <ghook@amd.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoMIPS: math-emu: <MADDF|MSUBF>.D: Fix accuracy (64-bit case)
Douglas Leung [Thu, 27 Jul 2017 16:08:59 +0000 (18:08 +0200)]
MIPS: math-emu: <MADDF|MSUBF>.D: Fix accuracy (64-bit case)

commit 2cfa58259f4b65b33ebe8f167019a1f89c6c3289 upstream.

Implement fused multiply-add with correct accuracy.

Fused multiply-add operation has better accuracy than respective
sequential execution of multiply and add operations applied on the
same inputs. This is because accuracy errors accumulate in latter
case.

This patch implements fused multiply-add with the same accuracy
as it is implemented in hardware, using 128-bit intermediate
calculations.

One test case example (raw bits) that this patch fixes:

MADDF.D fd,fs,ft:
  fd = 0x00000ca000000000
  fs = ft = 0x3f40624dd2f1a9fc

Fixes: e24c3bec3e8e ("MIPS: math-emu: Add support for the MIPS R6 MADDF FPU instruction")
Fixes: 83d43305a1df ("MIPS: math-emu: Add support for the MIPS R6 MSUBF FPU instruction")

Signed-off-by: Douglas Leung <douglas.leung@imgtec.com>
Signed-off-by: Miodrag Dinic <miodrag.dinic@imgtec.com>
Signed-off-by: Goran Ferenc <goran.ferenc@imgtec.com>
Signed-off-by: Aleksandar Markovic <aleksandar.markovic@imgtec.com>
Cc: Douglas Leung <douglas.leung@imgtec.com>
Cc: Bo Hu <bohu@google.com>
Cc: James Hogan <james.hogan@imgtec.com>
Cc: Jin Qian <jinqian@google.com>
Cc: Paul Burton <paul.burton@imgtec.com>
Cc: Petar Jovanovic <petar.jovanovic@imgtec.com>
Cc: Raghu Gandham <raghu.gandham@imgtec.com>
Cc: linux-mips@linux-mips.org
Cc: linux-kernel@vger.kernel.org
Patchwork: https://patchwork.linux-mips.org/patch/16891/
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoMIPS: math-emu: <MADDF|MSUBF>.S: Fix accuracy (32-bit case)
Douglas Leung [Thu, 27 Jul 2017 16:08:58 +0000 (18:08 +0200)]
MIPS: math-emu: <MADDF|MSUBF>.S: Fix accuracy (32-bit case)

commit b3b8e1eb27c523e32b6a8aa7ec8ac4754456af57 upstream.

Implement fused multiply-add with correct accuracy.

Fused multiply-add operation has better accuracy than respective
sequential execution of multiply and add operations applied on the
same inputs. This is because accuracy errors accumulate in latter
case.

This patch implements fused multiply-add with the same accuracy
as it is implemented in hardware, using 64-bit intermediate
calculations.

One test case example (raw bits) that this patch fixes:

MADDF.S fd,fs,ft:
  fd = 0x22575225
  fs = ft = 0x3727c5ac

Fixes: e24c3bec3e8e ("MIPS: math-emu: Add support for the MIPS R6 MADDF FPU instruction")
Fixes: 83d43305a1df ("MIPS: math-emu: Add support for the MIPS R6 MSUBF FPU instruction")

Signed-off-by: Douglas Leung <douglas.leung@imgtec.com>
Signed-off-by: Miodrag Dinic <miodrag.dinic@imgtec.com>
Signed-off-by: Goran Ferenc <goran.ferenc@imgtec.com>
Signed-off-by: Aleksandar Markovic <aleksandar.markovic@imgtec.com>
Cc: Douglas Leung <douglas.leung@imgtec.com>
Cc: Bo Hu <bohu@google.com>
Cc: James Hogan <james.hogan@imgtec.com>
Cc: Jin Qian <jinqian@google.com>
Cc: Paul Burton <paul.burton@imgtec.com>
Cc: Petar Jovanovic <petar.jovanovic@imgtec.com>
Cc: Raghu Gandham <raghu.gandham@imgtec.com>
Cc: linux-mips@linux-mips.org
Cc: linux-kernel@vger.kernel.org
Patchwork: https://patchwork.linux-mips.org/patch/16890/
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoMIPS: math-emu: <MADDF|MSUBF>.<D|S>: Clean up "maddf_flags" enumeration
Aleksandar Markovic [Thu, 27 Jul 2017 16:08:57 +0000 (18:08 +0200)]
MIPS: math-emu: <MADDF|MSUBF>.<D|S>: Clean up "maddf_flags" enumeration

commit ae11c0619973ffd73a496308d8a1cb5e1a353737 upstream.

Fix definition and usage of "maddf_flags" enumeration. Avoid duplicate
definition and apply more common capitalization.

This patch does not change any scenario. It just makes MADDF and
MSUBF emulation code more readable and easier to maintain, and
hopefully prevents future bugs as well.

Signed-off-by: Miodrag Dinic <miodrag.dinic@imgtec.com>
Signed-off-by: Goran Ferenc <goran.ferenc@imgtec.com>
Signed-off-by: Aleksandar Markovic <aleksandar.markovic@imgtec.com>
Reviewed-by: James Hogan <james.hogan@imgtec.com>
Cc: Bo Hu <bohu@google.com>
Cc: Douglas Leung <douglas.leung@imgtec.com>
Cc: Jin Qian <jinqian@google.com>
Cc: Paul Burton <paul.burton@imgtec.com>
Cc: Petar Jovanovic <petar.jovanovic@imgtec.com>
Cc: Raghu Gandham <raghu.gandham@imgtec.com>
Cc: linux-mips@linux-mips.org
Cc: linux-kernel@vger.kernel.org
Patchwork: https://patchwork.linux-mips.org/patch/16889/
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoMIPS: math-emu: <MADDF|MSUBF>.<D|S>: Fix some cases of zero inputs
Aleksandar Markovic [Thu, 27 Jul 2017 16:08:56 +0000 (18:08 +0200)]
MIPS: math-emu: <MADDF|MSUBF>.<D|S>: Fix some cases of zero inputs

commit 7cf64ce4d37f1b4f44365fcf77f565d523819dcd upstream.

Fix the cases of <MADDF|MSUBF>.<D|S> when any of two multiplicands is
+0 or -0, and the third input is also +0 or -0. Depending on the signs
of inputs, certain special cases must be handled.

A relevant example:

MADDF.S fd,fs,ft:
  If fs contains +0.0, ft contains -0.0, and fd contains 0.0, fd is
  going to contain +0.0 (without this patch, it used to contain -0.0).

Fixes: e24c3bec3e8e ("MIPS: math-emu: Add support for the MIPS R6 MADDF FPU instruction")
Fixes: 83d43305a1df ("MIPS: math-emu: Add support for the MIPS R6 MSUBF FPU instruction")

Signed-off-by: Miodrag Dinic <miodrag.dinic@imgtec.com>
Signed-off-by: Goran Ferenc <goran.ferenc@imgtec.com>
Signed-off-by: Aleksandar Markovic <aleksandar.markovic@imgtec.com>
Reviewed-by: James Hogan <james.hogan@imgtec.com>
Cc: Bo Hu <bohu@google.com>
Cc: Douglas Leung <douglas.leung@imgtec.com>
Cc: Jin Qian <jinqian@google.com>
Cc: Paul Burton <paul.burton@imgtec.com>
Cc: Petar Jovanovic <petar.jovanovic@imgtec.com>
Cc: Raghu Gandham <raghu.gandham@imgtec.com>
Cc: linux-mips@linux-mips.org
Cc: linux-kernel@vger.kernel.org
Patchwork: https://patchwork.linux-mips.org/patch/16888/
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoMIPS: math-emu: <MADDF|MSUBF>.<D|S>: Fix some cases of infinite inputs
Aleksandar Markovic [Thu, 27 Jul 2017 16:08:55 +0000 (18:08 +0200)]
MIPS: math-emu: <MADDF|MSUBF>.<D|S>: Fix some cases of infinite inputs

commit 0c64fe6348687f0e1cea9a608eae9d351124a73a upstream.

Fix the cases of <MADDF|MSUBF>.<D|S> when any of two multiplicands is
infinity. The correct behavior in such cases is affected by the nature
of third input. Cases of addition of infinities with opposite signs
and subtraction of infinities with same signs may arise and must be
handles separately. Also, the value od flags argument (that determines
whether the instruction is MADDF or MSUBF) affects the outcome.

Relevant examples:

MADDF.S fd,fs,ft:
  If fs contains +inf, ft contains +inf, and fd contains -inf, fd is
  going to contain indef (without this patch, it used to contain
  -inf).

MSUBF.S fd,fs,ft:
  If fs contains +inf, ft contains 1.0, and fd contains +0.0, fd is
  going to contain -inf (without this patch, it used to contain +inf).

Fixes: e24c3bec3e8e ("MIPS: math-emu: Add support for the MIPS R6 MADDF FPU instruction")
Fixes: 83d43305a1df ("MIPS: math-emu: Add support for the MIPS R6 MSUBF FPU instruction")

Signed-off-by: Douglas Leung <douglas.leung@imgtec.com>
Signed-off-by: Miodrag Dinic <miodrag.dinic@imgtec.com>
Signed-off-by: Goran Ferenc <goran.ferenc@imgtec.com>
Signed-off-by: Aleksandar Markovic <aleksandar.markovic@imgtec.com>
Reviewed-by: James Hogan <james.hogan@imgtec.com>
Cc: Douglas Leung <douglas.leung@imgtec.com>
Cc: Bo Hu <bohu@google.com>
Cc: Jin Qian <jinqian@google.com>
Cc: Paul Burton <paul.burton@imgtec.com>
Cc: Petar Jovanovic <petar.jovanovic@imgtec.com>
Cc: Raghu Gandham <raghu.gandham@imgtec.com>
Cc: linux-mips@linux-mips.org
Cc: linux-kernel@vger.kernel.org
Patchwork: https://patchwork.linux-mips.org/patch/16887/
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoMIPS: math-emu: <MADDF|MSUBF>.<D|S>: Fix NaN propagation
Aleksandar Markovic [Thu, 27 Jul 2017 16:08:54 +0000 (18:08 +0200)]
MIPS: math-emu: <MADDF|MSUBF>.<D|S>: Fix NaN propagation

commit e840be6e7057757befc3581e1699e30fe7f0dd51 upstream.

Fix the cases of <MADDF|MSUBF>.<D|S> when any of three inputs is any
NaN. Correct behavior of <MADDF|MSUBF>.<D|S> fd, fs, ft is following:

  - if any of inputs is sNaN, return a sNaN using following rules: if
    only one input is sNaN, return that one; if more than one input is
    sNaN, order of precedence for return value is fd, fs, ft
  - if no input is sNaN, but at least one of inputs is qNaN, return a
    qNaN using following rules: if only one input is qNaN, return that
    one; if more than one input is qNaN, order of precedence for
    return value is fd, fs, ft

The previous code contained correct handling of some above cases, but
not all. Also, such handling was scattered into various cases of
"switch (CLPAIR(xc, yc))" statement, and elsewhere. With this patch,
this logic is placed in one place, and "switch (CLPAIR(xc, yc))" is
significantly simplified.

A relevant example:

MADDF.S fd,fs,ft:
  If fs contains qNaN1, ft contains qNaN2, and fd contains qNaN3, fd
  is going to contain qNaN3 (without this patch, it used to contain
  qNaN1).

Fixes: e24c3bec3e8e ("MIPS: math-emu: Add support for the MIPS R6 MADDF FPU instruction")
Fixes: 83d43305a1df ("MIPS: math-emu: Add support for the MIPS R6 MSUBF FPU instruction")

Signed-off-by: Miodrag Dinic <miodrag.dinic@imgtec.com>
Signed-off-by: Goran Ferenc <goran.ferenc@imgtec.com>
Signed-off-by: Aleksandar Markovic <aleksandar.markovic@imgtec.com>
Reviewed-by: James Hogan <james.hogan@imgtec.com>
Cc: Bo Hu <bohu@google.com>
Cc: Douglas Leung <douglas.leung@imgtec.com>
Cc: Jin Qian <jinqian@google.com>
Cc: Paul Burton <paul.burton@imgtec.com>
Cc: Petar Jovanovic <petar.jovanovic@imgtec.com>
Cc: Raghu Gandham <raghu.gandham@imgtec.com>
Cc: linux-mips@linux-mips.org
Cc: linux-kernel@vger.kernel.org
Patchwork: https://patchwork.linux-mips.org/patch/16886/
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoMIPS: math-emu: Handle zero accumulator case in MADDF and MSUBF separately
Aleksandar Markovic [Mon, 19 Jun 2017 15:50:12 +0000 (17:50 +0200)]
MIPS: math-emu: Handle zero accumulator case in MADDF and MSUBF separately

commit ddbfff7429a75d954bf5bdff9f2222bceb4c236a upstream.

If accumulator value is zero, just return the value of previously
calculated product. This brings logic in MADDF/MSUBF implementation
closer to the logic in ADD/SUB case.

Signed-off-by: Miodrag Dinic <miodrag.dinic@imgtec.com>
Signed-off-by: Goran Ferenc <goran.ferenc@imgtec.com>
Signed-off-by: Aleksandar Markovic <aleksandar.markovic@imgtec.com>
Cc: James.Hogan@imgtec.com
Cc: Paul.Burton@imgtec.com
Cc: Raghu.Gandham@imgtec.com
Cc: Leonid.Yegoshin@imgtec.com
Cc: Douglas.Leung@imgtec.com
Cc: Petar.Jovanovic@imgtec.com
Cc: linux-mips@linux-mips.org
Patchwork: https://patchwork.linux-mips.org/patch/16512/
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoMIPS: math-emu: MINA.<D|S>: Fix some cases of infinity and zero inputs
Aleksandar Markovic [Thu, 27 Jul 2017 16:08:53 +0000 (18:08 +0200)]
MIPS: math-emu: MINA.<D|S>: Fix some cases of infinity and zero inputs

commit 304bfe473e70523e591fb1c9223289d355e0bdcb upstream.

Fix following special cases for MINA>.<D|S>:

  - if one of the inputs is zero, and the other is subnormal, normal,
    or infinity, the  value of the former should be returned (that is,
    a zero).
  - if one of the inputs is infinity, and the other input is normal,
    or subnormal, the value of the latter should be returned.

The previous implementation's logic for such cases was incorrect - it
appears as if it implements MAXA, and not MINA instruction.

A relevant example:

MINA.S fd,fs,ft:
  If fs contains 100.0, and ft contains 0.0, fd is going to contain
  0.0 (without this patch, it used to contain 100.0).

Fixes: a79f5f9ba508 ("MIPS: math-emu: Add support for the MIPS R6 MAX{, A} FPU instruction")
Fixes: 4e9561b20e2f ("MIPS: math-emu: Add support for the MIPS R6 MIN{, A} FPU instruction")

Signed-off-by: Miodrag Dinic <miodrag.dinic@imgtec.com>
Signed-off-by: Goran Ferenc <goran.ferenc@imgtec.com>
Signed-off-by: Aleksandar Markovic <aleksandar.markovic@imgtec.com>
Reviewed-by: James Hogan <james.hogan@imgtec.com>
Cc: Bo Hu <bohu@google.com>
Cc: Douglas Leung <douglas.leung@imgtec.com>
Cc: Jin Qian <jinqian@google.com>
Cc: Paul Burton <paul.burton@imgtec.com>
Cc: Petar Jovanovic <petar.jovanovic@imgtec.com>
Cc: Raghu Gandham <raghu.gandham@imgtec.com>
Cc: linux-mips@linux-mips.org
Cc: linux-kernel@vger.kernel.org
Patchwork: https://patchwork.linux-mips.org/patch/16885/
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoMIPS: math-emu: <MAXA|MINA>.<D|S>: Fix cases of both infinite inputs
Aleksandar Markovic [Thu, 27 Jul 2017 16:08:52 +0000 (18:08 +0200)]
MIPS: math-emu: <MAXA|MINA>.<D|S>: Fix cases of both infinite inputs

commit 3444c4eb534c20e44f0d6670b34263efaf8b531f upstream.

Fix the value returned by <MAXA|MINA>.<D|S> fd,fs,ft, if both inputs
are infinite. The previous implementation returned always the value
contained in ft in such cases. The correct behavior is specified
in Mips instruction set manual and is as follows:

    fs    ft        MAXA     MINA
  ---------------------------------
    inf   inf        inf      inf
    inf  -inf        inf     -inf
   -inf   inf        inf     -inf
   -inf  -inf       -inf     -inf

A relevant example:

MAXA.S fd,fs,ft:
  If fs contains +inf, and ft contains -inf, fd is going to contain
  +inf (without this patch, it used to contain -inf).

Fixes: a79f5f9ba508 ("MIPS: math-emu: Add support for the MIPS R6 MAX{, A} FPU instruction")
Fixes: 4e9561b20e2f ("MIPS: math-emu: Add support for the MIPS R6 MIN{, A} FPU instruction")

Signed-off-by: Miodrag Dinic <miodrag.dinic@imgtec.com>
Signed-off-by: Goran Ferenc <goran.ferenc@imgtec.com>
Signed-off-by: Aleksandar Markovic <aleksandar.markovic@imgtec.com>
Reviewed-by: James Hogan <james.hogan@imgtec.com>
Cc: Bo Hu <bohu@google.com>
Cc: Douglas Leung <douglas.leung@imgtec.com>
Cc: Jin Qian <jinqian@google.com>
Cc: Paul Burton <paul.burton@imgtec.com>
Cc: Petar Jovanovic <petar.jovanovic@imgtec.com>
Cc: Raghu Gandham <raghu.gandham@imgtec.com>
Cc: linux-mips@linux-mips.org
Cc: linux-kernel@vger.kernel.org
Patchwork: https://patchwork.linux-mips.org/patch/16884/
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoMIPS: math-emu: <MAXA|MINA>.<D|S>: Fix cases of input values with opposite signs
Aleksandar Markovic [Thu, 27 Jul 2017 16:08:51 +0000 (18:08 +0200)]
MIPS: math-emu: <MAXA|MINA>.<D|S>: Fix cases of input values with opposite signs

commit 1a41b3b441508ae63b1a9ec699ec94065739eb60 upstream.

Fix the value returned by <MAXA|MINA>.<D|S>, if the inputs are normal
fp numbers of the same absolute value, but opposite signs.

A relevant example:

MAXA.S fd,fs,ft:
  If fs contains -3.0, and ft contains +3.0, fd is going to contain
  +3.0 (without this patch, it used to contain -3.0).

Fixes: a79f5f9ba508 ("MIPS: math-emu: Add support for the MIPS R6 MAX{, A} FPU instruction")
Fixes: 4e9561b20e2f ("MIPS: math-emu: Add support for the MIPS R6 MIN{, A} FPU instruction")

Signed-off-by: Miodrag Dinic <miodrag.dinic@imgtec.com>
Signed-off-by: Goran Ferenc <goran.ferenc@imgtec.com>
Signed-off-by: Aleksandar Markovic <aleksandar.markovic@imgtec.com>
Reviewed-by: James Hogan <james.hogan@imgtec.com>
Cc: Bo Hu <bohu@google.com>
Cc: Douglas Leung <douglas.leung@imgtec.com>
Cc: Jin Qian <jinqian@google.com>
Cc: Paul Burton <paul.burton@imgtec.com>
Cc: Petar Jovanovic <petar.jovanovic@imgtec.com>
Cc: Raghu Gandham <raghu.gandham@imgtec.com>
Cc: linux-mips@linux-mips.org
Cc: linux-kernel@vger.kernel.org
Patchwork: https://patchwork.linux-mips.org/patch/16883/
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoMIPS: math-emu: <MAX|MIN>.<D|S>: Fix cases of both inputs negative
Aleksandar Markovic [Thu, 27 Jul 2017 16:08:50 +0000 (18:08 +0200)]
MIPS: math-emu: <MAX|MIN>.<D|S>: Fix cases of both inputs negative

commit aabf5cf02e22ebc4e541adf835910f388b6c3e65 upstream.

Fix the value returned by <MAX|MIN>.<D|S>, if both inputs are negative
normal fp numbers. The previous logic did not take into account that
if both inputs have the same sign, there should be separate treatment
of the cases when both inputs are negative and when both inputs are
positive.

A relevant example:

MAX.S fd,fs,ft:
  If fs contains -5.0, and ft contains -7.0, fd is going to contain
  -5.0 (without this patch, it used to contain -7.0).

Fixes: a79f5f9ba508 ("MIPS: math-emu: Add support for the MIPS R6 MAX{, A} FPU instruction")
Fixes: 4e9561b20e2f ("MIPS: math-emu: Add support for the MIPS R6 MIN{, A} FPU instruction")

Signed-off-by: Miodrag Dinic <miodrag.dinic@imgtec.com>
Signed-off-by: Goran Ferenc <goran.ferenc@imgtec.com>
Signed-off-by: Aleksandar Markovic <aleksandar.markovic@imgtec.com>
Reviewed-by: James Hogan <james.hogan@imgtec.com>
Cc: Bo Hu <bohu@google.com>
Cc: Douglas Leung <douglas.leung@imgtec.com>
Cc: Jin Qian <jinqian@google.com>
Cc: Paul Burton <paul.burton@imgtec.com>
Cc: Petar Jovanovic <petar.jovanovic@imgtec.com>
Cc: Raghu Gandham <raghu.gandham@imgtec.com>
Cc: linux-mips@linux-mips.org
Cc: linux-kernel@vger.kernel.org
Patchwork: https://patchwork.linux-mips.org/patch/16882/
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoMIPS: math-emu: <MAX|MAXA|MIN|MINA>.<D|S>: Fix cases of both inputs zero
Aleksandar Markovic [Thu, 27 Jul 2017 16:08:49 +0000 (18:08 +0200)]
MIPS: math-emu: <MAX|MAXA|MIN|MINA>.<D|S>: Fix cases of both inputs zero

commit 15560a58bfd4ff82cdd16b2270d4ef9b06d2cc4d upstream.

Fix the value returned by <MAX|MAXA|MIN|MINA>.<D|S>, if both inputs
are zeros. The right behavior in such cases is stated in instruction
reference manual and is as follows:

   fs  ft       MAX     MIN       MAXA    MINA
  ---------------------------------------------
    0   0        0       0         0       0
    0  -0        0      -0         0      -0
   -0   0        0      -0         0      -0
   -0  -0       -0      -0        -0      -0

Prior to this patch, some of the above cases were yielding correct
results. However, for the sake of code consistency, all such cases
are rewritten in this patch.

A relevant example:

MAX.S fd,fs,ft:
  If fs contains +0.0, and ft contains -0.0, fd is going to contain
  +0.0 (without this patch, it used to contain -0.0).

Fixes: a79f5f9ba508 ("MIPS: math-emu: Add support for the MIPS R6 MAX{, A} FPU instruction")
Fixes: 4e9561b20e2f ("MIPS: math-emu: Add support for the MIPS R6 MIN{, A} FPU instruction")

Signed-off-by: Miodrag Dinic <miodrag.dinic@imgtec.com>
Signed-off-by: Goran Ferenc <goran.ferenc@imgtec.com>
Signed-off-by: Aleksandar Markovic <aleksandar.markovic@imgtec.com>
Reviewed-by: James Hogan <james.hogan@imgtec.com>
Cc: Bo Hu <bohu@google.com>
Cc: Douglas Leung <douglas.leung@imgtec.com>
Cc: Jin Qian <jinqian@google.com>
Cc: Paul Burton <paul.burton@imgtec.com>
Cc: Petar Jovanovic <petar.jovanovic@imgtec.com>
Cc: Raghu Gandham <raghu.gandham@imgtec.com>
Cc: linux-mips@linux-mips.org
Cc: linux-kernel@vger.kernel.org
Patchwork: https://patchwork.linux-mips.org/patch/16881/
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoMIPS: math-emu: <MAX|MAXA|MIN|MINA>.<D|S>: Fix quiet NaN propagation
Aleksandar Markovic [Thu, 27 Jul 2017 16:08:48 +0000 (18:08 +0200)]
MIPS: math-emu: <MAX|MAXA|MIN|MINA>.<D|S>: Fix quiet NaN propagation

commit e78bf0dc4789bdea1453595ae89e8db65918e22e upstream.

Fix the value returned by <MAX|MAXA|MIN|MINA>.<D|S> fd,fs,ft, if both
inputs are quiet NaNs. The <MAX|MAXA|MIN|MINA>.<D|S> specifications
state that the returned value in such cases should be the quiet NaN
contained in register fs.

A relevant example:

MAX.S fd,fs,ft:
  If fs contains qNaN1, and ft contains qNaN2, fd is going to contain
  qNaN1 (without this patch, it used to contain qNaN2).

Fixes: a79f5f9ba508 ("MIPS: math-emu: Add support for the MIPS R6 MAX{, A} FPU instruction")
Fixes: 4e9561b20e2f ("MIPS: math-emu: Add support for the MIPS R6 MIN{, A} FPU instruction")

Signed-off-by: Miodrag Dinic <miodrag.dinic@imgtec.com>
Signed-off-by: Goran Ferenc <goran.ferenc@imgtec.com>
Signed-off-by: Aleksandar Markovic <aleksandar.markovic@imgtec.com>
Reviewed-by: James Hogan <james.hogan@imgtec.com>
Cc: Bo Hu <bohu@google.com>
Cc: Douglas Leung <douglas.leung@imgtec.com>
Cc: Jin Qian <jinqian@google.com>
Cc: Paul Burton <paul.burton@imgtec.com>
Cc: Petar Jovanovic <petar.jovanovic@imgtec.com>
Cc: Raghu Gandham <raghu.gandham@imgtec.com>
Cc: linux-mips@linux-mips.org
Cc: linux-kernel@vger.kernel.org
Patchwork: https://patchwork.linux-mips.org/patch/16880/
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoInput: i8042 - add Gigabyte P57 to the keyboard reset table
Kai-Heng Feng [Fri, 15 Sep 2017 16:36:16 +0000 (09:36 -0700)]
Input: i8042 - add Gigabyte P57 to the keyboard reset table

commit 697c5d8a36768b36729533fb44622b35d56d6ad0 upstream.

Similar to other Gigabyte laptops, the touchpad on P57 requires a
keyboard reset to detect Elantech touchpad correctly.

BugLink: https://bugs.launchpad.net/bugs/1594214
Signed-off-by: Kai-Heng Feng <kai.heng.feng@canonical.com>
Signed-off-by: Dmitry Torokhov <dmitry.torokhov@gmail.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agopinctrl/amd: save pin registers over suspend/resume
Daniel Drake [Mon, 11 Sep 2017 06:11:56 +0000 (14:11 +0800)]
pinctrl/amd: save pin registers over suspend/resume

commit 79d2c8bede2c93f9432d7da0bc2f76a195c90fc0 upstream.

The touchpad in the Asus laptop models X505BA/BP and X542BA/BP is
unresponsive after suspend/resume. The following error appears during
resume:

  i2c_hid i2c-ELAN1300:00: failed to reset device.

The problem here is that i2c_hid does not notice the interrupt being
generated at this point, because the GPIO is no longer configured
for interrupts.

Fix this by saving pinctrl-amd pin registers during suspend and
restoring them at resume time.

Based on code from pinctrl-intel.

Signed-off-by: Daniel Drake <drake@endlessm.com>
Signed-off-by: Linus Walleij <linus.walleij@linaro.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agotty: fix __tty_insert_flip_char regression
Arnd Bergmann [Wed, 2 Aug 2017 11:11:39 +0000 (13:11 +0200)]
tty: fix __tty_insert_flip_char regression

commit 8a5a90a2a477b86a3dc2eaa5a706db9bfdd647ca upstream.

Sergey noticed a small but fatal mistake in __tty_insert_flip_char,
leading to an oops in an interrupt handler when using any serial
port.

The problem is that I accidentally took the tty_buffer pointer
before calling __tty_buffer_request_room(), which replaces the
buffer. This moves the pointer lookup to the right place after
allocating the new buffer space.

Fixes: 979990c62848 ("tty: improve tty_insert_flip_char() fast path")
Reported-by: Sergey Senozhatsky <sergey.senozhatsky@gmail.com>
Tested-by: Sergey Senozhatsky <sergey.senozhatsky@gmail.com>
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agotty: improve tty_insert_flip_char() slow path
Arnd Bergmann [Tue, 20 Jun 2017 21:10:42 +0000 (23:10 +0200)]
tty: improve tty_insert_flip_char() slow path

commit 065ea0a7afd64d6cf3464bdd1d8cd227527e2045 upstream.

While working on improving the fast path of tty_insert_flip_char(),
I noticed that by calling tty_buffer_request_room(), we needlessly
move to the separate flag buffer mode for the tty, even when all
characters use TTY_NORMAL as the flag.

This changes the code to call __tty_buffer_request_room() with the
correct flag, which will then allocate a regular buffer when it rounds
out of space but no special flags have been used. I'm guessing that
this is the behavior that Peter Hurley intended when he introduced
the compacted flip buffers.

Fixes: acc0f67f307f ("tty: Halve flip buffer GFP_ATOMIC memory consumption")
Cc: Peter Hurley <peter@hurleysoftware.com>
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agotty: improve tty_insert_flip_char() fast path
Arnd Bergmann [Tue, 20 Jun 2017 21:10:41 +0000 (23:10 +0200)]
tty: improve tty_insert_flip_char() fast path

commit 979990c6284814617d8f2179d197f72ff62b5d85 upstream.

kernelci.org reports a crazy stack usage for the VT code when CONFIG_KASAN
is enabled:

drivers/tty/vt/keyboard.c: In function 'kbd_keycode':
drivers/tty/vt/keyboard.c:1452:1: error: the frame size of 2240 bytes is larger than 2048 bytes [-Werror=frame-larger-than=]

The problem is that tty_insert_flip_char() gets inlined many times into
kbd_keycode(), and also into other functions, and each copy requires 128
bytes for stack redzone to check for a possible out-of-bounds access on
the 'ch' and 'flags' arguments that are passed into
tty_insert_flip_string_flags as a variable-length string.

This introduces a new __tty_insert_flip_char() function for the slow
path, which receives the two arguments by value. This completely avoids
the problem and the stack usage goes back down to around 100 bytes.

Without KASAN, this is also slightly better, as we don't have to
spill the arguments to the stack but can simply pass 'ch' and 'flag'
in registers, saving a few bytes in .text for each call site.

This should be backported to linux-4.0 or later, which first introduced
the stack sanitizer in the kernel.

Fixes: c420f167db8c ("kasan: enable stack instrumentation")
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoIB/addr: Fix setting source address in addr6_resolve()
Roland Dreier [Tue, 6 Jun 2017 16:22:00 +0000 (09:22 -0700)]
IB/addr: Fix setting source address in addr6_resolve()

commit 79e25959403e6a79552db28a87abed34de32a1df upstream.

Commit eea40b8f624f ("infiniband: call ipv6 route lookup via the stub
interface") introduced a regression in address resolution when connecting
to IPv6 destination addresses.  The old code called ip6_route_output(),
while the new code calls ipv6_stub->ipv6_dst_lookup().  The two are almost
the same, except that ipv6_dst_lookup() also calls ip6_route_get_saddr()
if the source address is in6addr_any.

This means that the test of ipv6_addr_any(&fl6.saddr) now never succeeds,
and so we never copy the source address out.  This ends up causing
rdma_resolve_addr() to fail, because without a resolved source address,
cma_acquire_dev() will fail to find an RDMA device to use.  For me, this
causes connecting to an NVMe over Fabrics target via RoCE / IPv6 to fail.

Fix this by copying out fl6.saddr if ipv6_addr_any() is true for the original
source address passed into addr6_resolve().  We can drop our call to
ipv6_dev_get_saddr() because ipv6_dst_lookup() already does that work.

Fixes: eea40b8f624 ("infiniband: call ipv6 route lookup via the stub interface")
Signed-off-by: Roland Dreier <roland@purestorage.com>
Acked-by: Paolo Abeni <pabeni@redhat.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Signed-off-by: Raju Rangoju <rajur@chelsio.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agodrm/sun4i: Implement drm_driver lastclose to restore fbdev console
Jonathan Liu [Mon, 10 Jul 2017 06:55:04 +0000 (16:55 +1000)]
drm/sun4i: Implement drm_driver lastclose to restore fbdev console

commit 2a596fc9d974bb040eda9ab70bf8756fcaaa6afe upstream.

The drm_driver lastclose callback is called when the last userspace
DRM client has closed. Call drm_fbdev_cma_restore_mode to restore
the fbdev console otherwise the fbdev console will stop working.

Fixes: 9026e0d122ac ("drm: Add Allwinner A10 Display Engine support")
Tested-by: Olliver Schinagl <oliver@schinagl.nl>
Reviewed-by: Chen-Yu Tsai <wens@csie.org>
Signed-off-by: Jonathan Liu <net147@gmail.com>
Signed-off-by: Maxime Ripard <maxime.ripard@free-electrons.com>
[net147@gmail.com: Backport to 4.9, minor context change]
Signed-off-by: Jonathan Liu <net147@gmail.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoIB/{qib, hfi1}: Avoid flow control testing for RDMA write operation
Mike Marciniszyn [Tue, 22 Aug 2017 01:26:20 +0000 (18:26 -0700)]
IB/{qib, hfi1}: Avoid flow control testing for RDMA write operation

commit 5b0ef650bd0f820e922fcc42f1985d4621ae19cf upstream.

Section 9.7.7.2.5 of the 1.3 IBTA spec clearly says that receive
credits should never apply to RDMA write.

qib and hfi1 were doing that.  The following situation will result
in a QP hang:
- A prior SEND or RDMA_WRITE with immmediate consumed the last
  credit for a QP using RC receive buffer credits
- The prior op is acked so there are no more acks
- The peer ULP fails to post receive for some reason
- An RDMA write sees that the credits are exhausted and waits
- The peer ULP posts receive buffers
- The ULP posts a send or RDMA write that will be hung

The fix is to avoid the credit test for the RDMA write operation.

Reviewed-by: Kaike Wan <kaike.wan@intel.com>
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoorangefs: Don't clear SGID when inheriting ACLs
Jan Kara [Thu, 22 Jun 2017 13:31:13 +0000 (15:31 +0200)]
orangefs: Don't clear SGID when inheriting ACLs

commit b5accbb0dfae36d8d36cd882096943c98d5ede15 upstream.

When new directory 'DIR1' is created in a directory 'DIR0' with SGID bit
set, DIR1 is expected to have SGID bit set (and owning group equal to
the owning group of 'DIR0'). However when 'DIR0' also has some default
ACLs that 'DIR1' inherits, setting these ACLs will result in SGID bit on
'DIR1' to get cleared if user is not member of the owning group.

Fix the problem by creating __orangefs_set_acl() function that does not
call posix_acl_update_mode() and use it when inheriting ACLs. That
prevents SGID bit clearing and the mode has been properly set by
posix_acl_create() anyway.

Fixes: 073931017b49d9458aa351605b43a7e34598caef
CC: stable@vger.kernel.org
CC: Mike Marshall <hubcap@omnibond.com>
CC: pvfs2-developers@beowulf-underground.org
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Mike Marshall <hubcap@omnibond.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agomm: prevent double decrease of nr_reserved_highatomic
Minchan Kim [Tue, 13 Dec 2016 00:42:08 +0000 (16:42 -0800)]
mm: prevent double decrease of nr_reserved_highatomic

commit 4855e4a7f29d6d10b0b9c84e189c770c9a94e91e upstream.

There is race between page freeing and unreserved highatomic.

 CPU 0     CPU 1

    free_hot_cold_page
      mt = get_pfnblock_migratetype
      set_pcppage_migratetype(page, mt)
         unreserve_highatomic_pageblock
         spin_lock_irqsave(&zone->lock)
         move_freepages_block
         set_pageblock_migratetype(page)
         spin_unlock_irqrestore(&zone->lock)
      free_pcppages_bulk
        __free_one_page(mt) <- mt is stale

By above race, a page on CPU 0 could go non-highorderatomic free list
since the pageblock's type is changed.  By that, unreserve logic of
highorderatomic can decrease reserved count on a same pageblock severak
times and then it will make mismatch between nr_reserved_highatomic and
the number of reserved pageblock.

So, this patch verifies whether the pageblock is highatomic or not and
decrease the count only if the pageblock is highatomic.

Link: http://lkml.kernel.org/r/1476259429-18279-3-git-send-email-minchan@kernel.org
Signed-off-by: Minchan Kim <minchan@kernel.org>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Sangseok Lee <sangseok.lee@lge.com>
Cc: Michal Hocko <mhocko@suse.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Miles Chen <miles.chen@mediatek.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoNFSv4: Fix callback server shutdown
Trond Myklebust [Wed, 26 Apr 2017 15:55:27 +0000 (11:55 -0400)]
NFSv4: Fix callback server shutdown

commit ed6473ddc704a2005b9900ca08e236ebb2d8540a upstream.

We want to use kthread_stop() in order to ensure the threads are
shut down before we tear down the nfs_callback_info in nfs_callback_down.

Tested-and-reviewed-by: Kinglong Mee <kinglongmee@gmail.com>
Reported-by: Kinglong Mee <kinglongmee@gmail.com>
Fixes: bb6aeba736ba9 ("NFSv4.x: Switch to using svc_set_num_threads()...")
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
Cc: Jan Hudoba <kernel@jahu.sk>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoSUNRPC: Refactor svc_set_num_threads()
Trond Myklebust [Wed, 26 Apr 2017 15:55:26 +0000 (11:55 -0400)]
SUNRPC: Refactor svc_set_num_threads()

commit 9e0d87680d689f1758185851c3da6eafb16e71e1 upstream.

Refactor to separate out the functions of starting and stopping threads
so that they can be used in other helpers.

Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Tested-and-reviewed-by: Kinglong Mee <kinglongmee@gmail.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
Cc: Jan Hudoba <kernel@jahu.sk>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoLinux 4.9.51 v4.9.51
Greg Kroah-Hartman [Wed, 20 Sep 2017 06:20:15 +0000 (08:20 +0200)]
Linux 4.9.51

6 years agoipv6: Fix may be used uninitialized warning in rt6_check
Steffen Klassert [Fri, 25 Aug 2017 07:05:42 +0000 (09:05 +0200)]
ipv6: Fix may be used uninitialized warning in rt6_check

commit 3614364527daa870264f6dde77f02853cdecd02c upstream.

rt_cookie might be used uninitialized, fix this by
initializing it.

Fixes: c5cff8561d2d ("ipv6: add rcu grace period before freeing fib6_node")
Signed-off-by: Steffen Klassert <steffen.klassert@secunet.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Cc: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoxfs: fix compiler warnings
Darrick J. Wong [Thu, 31 Aug 2017 22:11:06 +0000 (15:11 -0700)]
xfs: fix compiler warnings

commit 7bf7a193a90cadccaad21c5970435c665c40fe27 upstream.

Fix up all the compiler warnings that have crept in.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Cc: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agomd/raid5: release/flush io in raid5_do_work()
Song Liu [Thu, 24 Aug 2017 16:53:59 +0000 (09:53 -0700)]
md/raid5: release/flush io in raid5_do_work()

commit 9c72a18e46ebe0f09484cce8ebf847abdab58498 upstream.

In raid5, there are scenarios where some ios are deferred to a later
time, and some IO need a flush to complete. To make sure we make
progress with these IOs, we need to call the following functions:

    flush_deferred_bios(conf);
    r5l_flush_stripe_to_raid(conf->log);

Both of these functions are called in raid5d(), but missing in
raid5_do_work(). As a result, these functions are not called
when multi-threading (group_thread_cnt > 0) is enabled. This patch
adds calls to these function to raid5_do_work().

Note for stable branches:

  r5l_flush_stripe_to_raid(conf->log) is need for 4.4+
  flush_deferred_bios(conf) is only needed for 4.11+

Signed-off-by: Song Liu <songliubraving@fb.com>
Signed-off-by: Shaohua Li <shli@fb.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoxfs: use kmem_free to free return value of kmem_zalloc
Pan Bian [Sun, 17 Sep 2017 21:07:12 +0000 (14:07 -0700)]
xfs: use kmem_free to free return value of kmem_zalloc

commit 6c370590cfe0c36bcd62d548148aa65c984540b7 upstream.

In function xfs_test_remount_options(), kfree() is used to free memory
allocated by kmem_zalloc(). But it is better to use kmem_free().

Signed-off-by: Pan Bian <bianpan2016@163.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoxfs: open code end_buffer_async_write in xfs_finish_page_writeback
Christoph Hellwig [Sun, 17 Sep 2017 21:07:11 +0000 (14:07 -0700)]
xfs: open code end_buffer_async_write in xfs_finish_page_writeback

commit 8353a814f2518dcfa79a5bb77afd0e7dfa391bb1 upstream.

Our loop in xfs_finish_page_writeback, which iterates over all buffer
heads in a page and then calls end_buffer_async_write, which also
iterates over all buffers in the page to check if any I/O is in flight
is not only inefficient, but also potentially dangerous as
end_buffer_async_write can cause the page and all buffers to be freed.

Replace it with a single loop that does the work of end_buffer_async_write
on a per-page basis.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoxfs: don't set v3 xflags for v2 inodes
Christoph Hellwig [Sun, 17 Sep 2017 21:07:10 +0000 (14:07 -0700)]
xfs: don't set v3 xflags for v2 inodes

commit dd60687ee541ca3f6df8758f38e6f22f57c42a37 upstream.

Reject attempts to set XFLAGS that correspond to di_flags2 inode flags
if the inode isn't a v3 inode, because di_flags2 only exists on v3.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoxfs: fix incorrect log_flushed on fsync
Amir Goldstein [Sun, 17 Sep 2017 21:07:09 +0000 (14:07 -0700)]
xfs: fix incorrect log_flushed on fsync

commit 47c7d0b19502583120c3f396c7559e7a77288a68 upstream.

When calling into _xfs_log_force{,_lsn}() with a pointer
to log_flushed variable, log_flushed will be set to 1 if:
1. xlog_sync() is called to flush the active log buffer
AND/OR
2. xlog_wait() is called to wait on a syncing log buffers

xfs_file_fsync() checks the value of log_flushed after
_xfs_log_force_lsn() call to optimize away an explicit
PREFLUSH request to the data block device after writing
out all the file's pages to disk.

This optimization is incorrect in the following sequence of events:

 Task A                    Task B
 -------------------------------------------------------
 xfs_file_fsync()
   _xfs_log_force_lsn()
     xlog_sync()
        [submit PREFLUSH]
                           xfs_file_fsync()
                             file_write_and_wait_range()
                               [submit WRITE X]
                               [endio  WRITE X]
                             _xfs_log_force_lsn()
                               xlog_wait()
        [endio  PREFLUSH]

The write X is not guarantied to be on persistent storage
when PREFLUSH request in completed, because write A was submitted
after the PREFLUSH request, but xfs_file_fsync() of task A will
be notified of log_flushed=1 and will skip explicit flush.

If the system crashes after fsync of task A, write X may not be
present on disk after reboot.

This bug was discovered and demonstrated using Josef Bacik's
dm-log-writes target, which can be used to record block io operations
and then replay a subset of these operations onto the target device.
The test goes something like this:
- Use fsx to execute ops of a file and record ops on log device
- Every now and then fsync the file, store md5 of file and mark
  the location in the log
- Then replay log onto device for each mark, mount fs and compare
  md5 of file to stored value

Cc: Christoph Hellwig <hch@lst.de>
Cc: Josef Bacik <jbacik@fb.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Amir Goldstein <amir73il@gmail.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoxfs: disable per-inode DAX flag
Christoph Hellwig [Sun, 17 Sep 2017 21:07:08 +0000 (14:07 -0700)]
xfs: disable per-inode DAX flag

commit 742d84290739ae908f1b61b7d17ea382c8c0073a upstream.

Currently flag switching can be used to easily crash the kernel.  Disable
the per-inode DAX flag until that is sorted out.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoxfs: relog dirty buffers during swapext bmbt owner change
Brian Foster [Sun, 17 Sep 2017 21:07:07 +0000 (14:07 -0700)]
xfs: relog dirty buffers during swapext bmbt owner change

commit 2dd3d709fc4338681a3aa61658122fa8faa5a437 upstream.

The owner change bmbt scan that occurs during extent swap operations
does not handle ordered buffer failures. Buffers that cannot be
marked ordered must be physically logged so previously dirty ranges
of the buffer can be relogged in the transaction.

Since the bmbt scan may need to process and potentially log a large
number of blocks, we can't expect to complete this operation in a
single transaction. Update extent swap to use a permanent
transaction with enough log reservation to physically log a buffer.
Update the bmbt scan to physically log any buffers that cannot be
ordered and to terminate the scan with -EAGAIN. On -EAGAIN, the
caller rolls the transaction and restarts the scan. Finally, update
the bmbt scan helper function to skip bmbt blocks that already match
the expected owner so they are not reprocessed after scan restarts.

Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
[darrick: fix the xfs_trans_roll call]
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoxfs: disallow marking previously dirty buffers as ordered
Brian Foster [Sun, 17 Sep 2017 21:07:06 +0000 (14:07 -0700)]
xfs: disallow marking previously dirty buffers as ordered

commit a5814bceea48ee1c57c4db2bd54b0c0246daf54a upstream.

Ordered buffers are used in situations where the buffer is not
physically logged but must pass through the transaction/logging
pipeline for a particular transaction. As a result, ordered buffers
are not unpinned and written back until the transaction commits to
the log. Ordered buffers have a strict requirement that the target
buffer must not be currently dirty and resident in the log pipeline
at the time it is marked ordered. If a dirty+ordered buffer is
committed, the buffer is reinserted to the AIL but not physically
relogged at the LSN of the associated checkpoint. The buffer log
item is assigned the LSN of the latest checkpoint and the AIL
effectively releases the previously logged buffer content from the
active log before the buffer has been written back. If the tail
pushes forward and a filesystem crash occurs while in this state, an
inconsistent filesystem could result.

It is currently the caller responsibility to ensure an ordered
buffer is not already dirty from a previous modification. This is
unclear and error prone when not used in situations where it is
guaranteed a buffer has not been previously modified (such as new
metadata allocations).

To facilitate general purpose use of ordered buffers, update
xfs_trans_ordered_buf() to conditionally order the buffer based on
state of the log item and return the status of the result. If the
bli is dirty, do not order the buffer and return false. The caller
must either physically log the buffer (having acquired the
appropriate log reservation) or push it from the AIL to clean it
before it can be marked ordered in the current transaction.

Note that ordered buffers are currently only used in two situations:
1.) inode chunk allocation where previously logged buffers are not
possible and 2.) extent swap which will be updated to handle ordered
buffer failures in a separate patch.

Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoxfs: move bmbt owner change to last step of extent swap
Brian Foster [Sun, 17 Sep 2017 21:07:05 +0000 (14:07 -0700)]
xfs: move bmbt owner change to last step of extent swap

commit 6fb10d6d22094bc4062f92b9ccbcee2f54033d04 upstream.

The extent swap operation currently resets bmbt block owners before
the inode forks are swapped. The bmbt buffers are marked as ordered
so they do not have to be physically logged in the transaction.

This use of ordered buffers is not safe as bmbt buffers may have
been previously physically logged. The bmbt owner change algorithm
needs to be updated to physically log buffers that are already dirty
when/if they are encountered. This means that an extent swap will
eventually require multiple rolling transactions to handle large
btrees. In addition, all inode related changes must be logged before
the bmbt owner change scan begins and can roll the transaction for
the first time to preserve fs consistency via log recovery.

In preparation for such fixes to the bmbt owner change algorithm,
refactor the bmbt scan out of the extent fork swap code to the last
operation before the transaction is committed. Update
xfs_swap_extent_forks() to only set the inode log flags when an
owner change scan is necessary. Update xfs_swap_extents() to trigger
the owner change based on the inode log flags. Note that since the
owner change now occurs after the extent fork swap, the inode btrees
must be fixed up with the inode number of the current inode (similar
to log recovery).

Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoxfs: skip bmbt block ino validation during owner change
Brian Foster [Sun, 17 Sep 2017 21:07:04 +0000 (14:07 -0700)]
xfs: skip bmbt block ino validation during owner change

commit 99c794c639a65cc7b74f30a674048fd100fe9ac8 upstream.

Extent swap uses xfs_btree_visit_blocks() to fix up bmbt block
owners on v5 (!rmapbt) filesystems. The bmbt scan uses
xfs_btree_lookup_get_block() to read bmbt blocks which verifies the
current owner of the block against the parent inode of the bmbt.
This works during extent swap because the bmbt owners are updated to
the opposite inode number before the inode extent forks are swapped.

The modified bmbt blocks are marked as ordered buffers which allows
everything to commit in a single transaction. If the transaction
commits to the log and the system crashes such that recovery of the
extent swap is required, log recovery restarts the bmbt scan to fix
up any bmbt blocks that may have not been written back before the
crash. The log recovery bmbt scan occurs after the inode forks have
been swapped, however. This causes the bmbt block owner verification
to fail, leads to log recovery failure and requires xfs_repair to
zap the log to recover.

Define a new invalid inode owner flag to inform the btree block
lookup mechanism that the current inode may be invalid with respect
to the current owner of the bmbt block. Set this flag on the cursor
used for change owner scans to allow this operation to work at
runtime and during log recovery.

Signed-off-by: Brian Foster <bfoster@redhat.com>
Fixes: bb3be7e7c ("xfs: check for bogus values in btree block headers")
Cc: stable@vger.kernel.org
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoxfs: don't log dirty ranges for ordered buffers
Brian Foster [Sun, 17 Sep 2017 21:07:03 +0000 (14:07 -0700)]
xfs: don't log dirty ranges for ordered buffers

commit 8dc518dfa7dbd079581269e51074b3c55a65a880 upstream.

Ordered buffers are attached to transactions and pushed through the
logging infrastructure just like normal buffers with the exception
that they are not actually written to the log. Therefore, we don't
need to log dirty ranges of ordered buffers. xfs_trans_log_buf() is
called on ordered buffers to set up all of the dirty state on the
transaction, buffer and log item and prepare the buffer for I/O.

Now that xfs_trans_dirty_buf() is available, call it from
xfs_trans_ordered_buf() so the latter is now mutually exclusive with
xfs_trans_log_buf(). This reflects the implementation of ordered
buffers and helps eliminate confusion over the need to log ranges of
ordered buffers just to set up internal log state.

Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Allison Henderson <allison.henderson@oracle.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoxfs: refactor buffer logging into buffer dirtying helper
Brian Foster [Sun, 17 Sep 2017 21:07:02 +0000 (14:07 -0700)]
xfs: refactor buffer logging into buffer dirtying helper

commit 9684010d38eccda733b61106765e9357cf436f65 upstream.

xfs_trans_log_buf() is responsible for logging the dirty segments of
a buffer along with setting all of the necessary state on the
transaction, buffer, bli, etc., to ensure that the associated items
are marked as dirty and prepared for I/O. We have a couple use cases
that need to to dirty a buffer in a transaction without actually
logging dirty ranges of the buffer.  One existing use case is
ordered buffers, which are currently logged with arbitrary ranges to
accomplish this even though the content of ordered buffers is never
written to the log. Another pending use case is to relog an already
dirty buffer across rolled transactions within the deferred
operations infrastructure. This is required to prevent a held
(XFS_BLI_HOLD) buffer from pinning the tail of the log.

Refactor xfs_trans_log_buf() into a new function that contains all
of the logic responsible to dirty the transaction, lidp, buffer and
bli. This new function can be used in the future for the use cases
outlined above. This patch does not introduce functional changes.

Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Allison Henderson <allison.henderson@oracle.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoxfs: ordered buffer log items are never formatted
Brian Foster [Sun, 17 Sep 2017 21:07:01 +0000 (14:07 -0700)]
xfs: ordered buffer log items are never formatted

commit e9385cc6fb7edf23702de33a2dc82965d92d9392 upstream.

Ordered buffers pass through the logging infrastructure without ever
being written to the log. The way this works is that the ordered
buffer status is transferred to the log vector at commit time via
the ->iop_size() callback. In xlog_cil_insert_format_items(),
ordered log vectors bypass ->iop_format() processing altogether.

Therefore it is unnecessary for xfs_buf_item_format() to handle
ordered buffers. Remove the unnecessary logic and assert that an
ordered buffer never reaches this point.

Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoxfs: remove unnecessary dirty bli format check for ordered bufs
Brian Foster [Sun, 17 Sep 2017 21:07:00 +0000 (14:07 -0700)]
xfs: remove unnecessary dirty bli format check for ordered bufs

commit 6453c65d3576bc3e602abb5add15f112755c08ca upstream.

xfs_buf_item_unlock() historically checked the dirty state of the
buffer by manually checking the buffer log formats for dirty
segments. The introduction of ordered buffers invalidated this check
because ordered buffers have dirty bli's but no dirty (logged)
segments. The check was updated to accommodate ordered buffers by
looking at the bli state first and considering the blf only if the
bli is clean.

This logic is safe but unnecessary. There is no valid case where the
bli is clean yet the blf has dirty segments. The bli is set dirty
whenever the blf is logged (via xfs_trans_log_buf()) and the blf is
cleared in the only place BLI_DIRTY is cleared (xfs_trans_binval()).

Remove the conditional blf dirty checks and replace with an assert
that should catch any discrepencies between bli and blf dirty
states. Refactor the old blf dirty check into a helper function to
be used by the assert.

Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoxfs: open-code xfs_buf_item_dirty()
Brian Foster [Sun, 17 Sep 2017 21:06:59 +0000 (14:06 -0700)]
xfs: open-code xfs_buf_item_dirty()

commit a4f6cf6b2b6b60ec2a05a33a32e65caa4149aa2b upstream.

It checks a single flag and has one caller. It probably isn't worth
its own function.

Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoxfs: check for race with xfs_reclaim_inode() in xfs_ifree_cluster()
Omar Sandoval [Sun, 17 Sep 2017 21:06:58 +0000 (14:06 -0700)]
xfs: check for race with xfs_reclaim_inode() in xfs_ifree_cluster()

commit f2e9ad212def50bcf4c098c6288779dd97fff0f0 upstream.

After xfs_ifree_cluster() finds an inode in the radix tree and verifies
that the inode number is what it expected, xfs_reclaim_inode() can swoop
in and free it. xfs_ifree_cluster() will then happily continue working
on the freed inode. Most importantly, it will mark the inode stale,
which will probably be overwritten when the inode slab object is
reallocated, but if it has already been reallocated then we can end up
with an inode spuriously marked stale.

In 8a17d7ddedb4 ("xfs: mark reclaimed inodes invalid earlier") we added
a second check to xfs_iflush_cluster() to detect this race, but the
similar RCU lookup in xfs_ifree_cluster() needs the same treatment.

Signed-off-by: Omar Sandoval <osandov@fb.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoxfs: evict all inodes involved with log redo item
Darrick J. Wong [Sun, 17 Sep 2017 21:06:57 +0000 (14:06 -0700)]
xfs: evict all inodes involved with log redo item

commit 799ea9e9c59949008770aab4e1da87f10e99dbe4 upstream.

When we introduced the bmap redo log items, we set MS_ACTIVE on the
mountpoint and XFS_IRECOVERY on the inode to prevent unlinked inodes
from being truncated prematurely during log recovery.  This also had the
effect of putting linked inodes on the lru instead of evicting them.

Unfortunately, we neglected to find all those unreferenced lru inodes
and evict them after finishing log recovery, which means that we leak
them if anything goes wrong in the rest of xfs_mountfs, because the lru
is only cleaned out on unmount.

Therefore, evict unreferenced inodes in the lru list immediately
after clearing MS_ACTIVE.

Fixes: 17c12bcd30 ("xfs: when replaying bmap operations, don't let unlinked inodes get reaped")
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Cc: viro@ZenIV.linux.org.uk
Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoxfs: stop searching for free slots in an inode chunk when there are none
Carlos Maiolino [Sun, 17 Sep 2017 21:06:56 +0000 (14:06 -0700)]
xfs: stop searching for free slots in an inode chunk when there are none

commit 2d32311cf19bfb8c1d2b4601974ddd951f9cfd0b upstream.

In a filesystem without finobt, the Space manager selects an AG to alloc a new
inode, where xfs_dialloc_ag_inobt() will search the AG for the free slot chunk.

When the new inode is in the same AG as its parent, the btree will be searched
starting on the parent's record, and then retried from the top if no slot is
available beyond the parent's record.

To exit this loop though, xfs_dialloc_ag_inobt() relies on the fact that the
btree must have a free slot available, once its callers relied on the
agi->freecount when deciding how/where to allocate this new inode.

In the case when the agi->freecount is corrupted, showing available inodes in an
AG, when in fact there is none, this becomes an infinite loop.

Add a way to stop the loop when a free slot is not found in the btree, making
the function to fall into the whole AG scan which will then, be able to detect
the corruption and shut the filesystem down.

As pointed by Brian, this might impact performance, giving the fact we
don't reset the search distance anymore when we reach the end of the
tree, giving it fewer tries before falling back to the whole AG search, but
it will only affect searches that start within 10 records to the end of the tree.

Signed-off-by: Carlos Maiolino <cmaiolino@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoxfs: add log recovery tracepoint for head/tail
Brian Foster [Sun, 17 Sep 2017 21:06:55 +0000 (14:06 -0700)]
xfs: add log recovery tracepoint for head/tail

commit e67d3d4246e5fbb0c7c700426d11241ca9c6f473 upstream.

Torn write detection and tail overwrite detection can shift the log
head and tail respectively in the event of CRC mismatch or
corruption errors. Add a high-level log recovery tracepoint to dump
the final log head/tail and make those values easily attainable in
debug/diagnostic situations.

Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoxfs: handle -EFSCORRUPTED during head/tail verification
Brian Foster [Sun, 17 Sep 2017 21:06:54 +0000 (14:06 -0700)]
xfs: handle -EFSCORRUPTED during head/tail verification

commit a4c9b34d6a17081005ec459b57b8effc08f4c731 upstream.

Torn write and tail overwrite detection both trigger only on
-EFSBADCRC errors. While this is the most likely failure scenario
for each condition, -EFSCORRUPTED is still possible in certain cases
depending on what ends up on disk when a torn write or partial tail
overwrite occurs. For example, an invalid log record h_len can lead
to an -EFSCORRUPTED error when running the log recovery CRC pass.

Therefore, update log head and tail verification to trigger the
associated head/tail fixups in the event of -EFSCORRUPTED errors
along with -EFSBADCRC. Also, -EFSCORRUPTED can currently be returned
from xlog_do_recovery_pass() before rhead_blk is initialized if the
first record encountered happens to be corrupted. This leads to an
incorrect 'first_bad' return value. Initialize rhead_blk earlier in
the function to address that problem as well.

Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoxfs: fix log recovery corruption error due to tail overwrite
Brian Foster [Sun, 17 Sep 2017 21:06:53 +0000 (14:06 -0700)]
xfs: fix log recovery corruption error due to tail overwrite

commit 4a4f66eac4681378996a1837ad1ffec3a2e2981f upstream.

If we consider the case where the tail (T) of the log is pinned long
enough for the head (H) to push and block behind the tail, we can
end up blocked in the following state without enough free space (f)
in the log to satisfy a transaction reservation:

0 phys. log N
[-------HffT---H'--T'---]

The last good record in the log (before H) refers to T. The tail
eventually pushes forward (T') leaving more free space in the log
for writes to H. At this point, suppose space frees up in the log
for the maximum of 8 in-core log buffers to start flushing out to
the log. If this pushes the head from H to H', these next writes
overwrite the previous tail T. This is safe because the items logged
from T to T' have been written back and removed from the AIL.

If the next log writes (H -> H') happen to fail and result in
partial records in the log, the filesystem shuts down having
overwritten T with invalid data. Log recovery correctly locates H on
the subsequent mount, but H still refers to the now corrupted tail
T. This results in log corruption errors and recovery failure.

Since the tail overwrite results from otherwise correct runtime
behavior, it is up to log recovery to try and deal with this
situation. Update log recovery tail verification to run a CRC pass
from the first record past the tail to the head. This facilitates
error detection at T and moves the recovery tail to the first good
record past H' (similar to truncating the head on torn write
detection). If corruption is detected beyond the range possibly
affected by the max number of iclogs, the log is legitimately
corrupted and log recovery failure is expected.

Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoxfs: always verify the log tail during recovery
Brian Foster [Sun, 17 Sep 2017 21:06:52 +0000 (14:06 -0700)]
xfs: always verify the log tail during recovery

commit 5297ac1f6d7cbf45464a49b9558831f271dfc559 upstream.

Log tail verification currently only occurs when torn writes are
detected at the head of the log. This was introduced because a
change in the head block due to torn writes can lead to a change in
the tail block (each log record header references the current tail)
and the tail block should be verified before log recovery proceeds.

Tail corruption is possible outside of torn write scenarios,
however. For example, partial log writes can be detected and cleared
during the initial head/tail block discovery process. If the partial
write coincides with a tail overwrite, the log tail is corrupted and
recovery fails.

To facilitate correct handling of log tail overwites, update log
recovery to always perform tail verification. This is necessary to
detect potential tail overwrite conditions when torn writes may not
have occurred. This changes normal (i.e., no torn writes) recovery
behavior slightly to detect and return CRC related errors near the
tail before actual recovery starts.

Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoxfs: fix recovery failure when log record header wraps log end
Brian Foster [Sun, 17 Sep 2017 21:06:51 +0000 (14:06 -0700)]
xfs: fix recovery failure when log record header wraps log end

commit 284f1c2c9bebf871861184b0e2c40fa921dd380b upstream.

The high-level log recovery algorithm consists of two loops that
walk the physical log and process log records from the tail to the
head. The first loop handles the case where the tail is beyond the
head and processes records up to the end of the physical log. The
subsequent loop processes records from the beginning of the physical
log to the head.

Because log records can wrap around the end of the physical log, the
first loop mentioned above must handle this case appropriately.
Records are processed from in-core buffers, which means that this
algorithm must split the reads of such records into two partial
I/Os: 1.) from the beginning of the record to the end of the log and
2.) from the beginning of the log to the end of the record. This is
further complicated by the fact that the log record header and log
record data are read into independent buffers.

The current handling of each buffer correctly splits the reads when
either the header or data starts before the end of the log and wraps
around the end. The data read does not correctly handle the case
where the prior header read wrapped or ends on the physical log end
boundary. blk_no is incremented to or beyond the log end after the
header read to point to the record data, but the split data read
logic triggers, attempts to read from an invalid log block and
ultimately causes log recovery to fail. This can be reproduced
fairly reliably via xfstests tests generic/047 and generic/388 with
large iclog sizes (256k) and small (10M) logs.

If the record header read has pushed beyond the end of the physical
log, the subsequent data read is actually contiguous. Update the
data read logic to detect the case where blk_no has wrapped, mod it
against the log size to read from the correct address and issue one
contiguous read for the log data buffer. The log record is processed
as normal from the buffer(s), the loop exits after the current
iteration and the subsequent loop picks up with the first new record
after the start of the log.

Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoxfs: Properly retry failed inode items in case of error during buffer writeback
Carlos Maiolino [Sun, 17 Sep 2017 21:06:50 +0000 (14:06 -0700)]
xfs: Properly retry failed inode items in case of error during buffer writeback

commit d3a304b6292168b83b45d624784f973fdc1ca674 upstream.

When a buffer has been failed during writeback, the inode items into it
are kept flush locked, and are never resubmitted due the flush lock, so,
if any buffer fails to be written, the items in AIL are never written to
disk and never unlocked.

This causes unmount operation to hang due these items flush locked in AIL,
but this also causes the items in AIL to never be written back, even when
the IO device comes back to normal.

I've been testing this patch with a DM-thin device, creating a
filesystem larger than the real device.

When writing enough data to fill the DM-thin device, XFS receives ENOSPC
errors from the device, and keep spinning on xfsaild (when 'retry
forever' configuration is set).

At this point, the filesystem can not be unmounted because of the flush locked
items in AIL, but worse, the items in AIL are never retried at all
(once xfs_inode_item_push() will skip the items that are flush locked),
even if the underlying DM-thin device is expanded to the proper size.

This patch fixes both cases, retrying any item that has been failed
previously, using the infra-structure provided by the previous patch.

Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Carlos Maiolino <cmaiolino@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoxfs: Add infrastructure needed for error propagation during buffer IO failure
Carlos Maiolino [Sun, 17 Sep 2017 21:06:49 +0000 (14:06 -0700)]
xfs: Add infrastructure needed for error propagation during buffer IO failure

commit 0b80ae6ed13169bd3a244e71169f2cc020b0c57a upstream.

With the current code, XFS never re-submit a failed buffer for IO,
because the failed item in the buffer is kept in the flush locked state
forever.

To be able to resubmit an log item for IO, we need a way to mark an item
as failed, if, for any reason the buffer which the item belonged to
failed during writeback.

Add a new log item callback to be used after an IO completion failure
and make the needed clean ups.

Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Carlos Maiolino <cmaiolino@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoxfs: remove xfs_trans_ail_delete_bulk
Christoph Hellwig [Sun, 17 Sep 2017 21:06:48 +0000 (14:06 -0700)]
xfs: remove xfs_trans_ail_delete_bulk

commit 27af1bbf524459962d1477a38ac6e0b7f79aaecc upstream.

xfs_iflush_done uses an on-stack variable length array to pass the log
items to be deleted to xfs_trans_ail_delete_bulk.  On-stack VLAs are a
nasty gcc extension that can lead to unbounded stack allocations, but
fortunately we can easily avoid them by simply open coding
xfs_trans_ail_delete_bulk in xfs_iflush_done, which is the only caller
of it except for the single-item xfs_trans_ail_delete.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoxfs: toggle readonly state around xfs_log_mount_finish
Eric Sandeen [Sun, 17 Sep 2017 21:06:47 +0000 (14:06 -0700)]
xfs: toggle readonly state around xfs_log_mount_finish

commit 6f4a1eefdd0ad4561543270a7fceadabcca075dd upstream.

When we do log recovery on a readonly mount, unlinked inode
processing does not happen due to the readonly checks in
xfs_inactive(), which are trying to prevent any I/O on a
readonly mount.

This is misguided - we do I/O on readonly mounts all the time,
for consistency; for example, log recovery.  So do the same
RDONLY flag twiddling around xfs_log_mount_finish() as we
do around xfs_log_mount(), for the same reason.

This all cries out for a big rework but for now this is a
simple fix to an obvious problem.

Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoxfs: write unmount record for ro mounts
Eric Sandeen [Sun, 17 Sep 2017 21:06:46 +0000 (14:06 -0700)]
xfs: write unmount record for ro mounts

commit 757a69ef6cf2bf839bd4088e5609ddddd663b0c4 upstream.

There are dueling comments in the xfs code about intent
for log writes when unmounting a readonly filesystem.

In xfs_mountfs, we see the intent:

/*
 * Now the log is fully replayed, we can transition to full read-only
 * mode for read-only mounts. This will sync all the metadata and clean
 * the log so that the recovery we just performed does not have to be
 * replayed again on the next mount.
 */

and it calls xfs_quiesce_attr(), but by the time we get to
xfs_log_unmount_write(), it returns early for a RDONLY mount:

 * Don't write out unmount record on read-only mounts.

Because of this, sequential ro mounts of a filesystem with
a dirty log will replay the log each time, which seems odd.

Fix this by writing an unmount record even for RO mounts, as long
as norecovery wasn't specified (don't write a clean log record
if a dirty log may still be there!) and the log device is
writable.

Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoiomap: fix integer truncation issues in the zeroing and dirtying helpers
Christoph Hellwig [Sun, 17 Sep 2017 21:06:45 +0000 (14:06 -0700)]
iomap: fix integer truncation issues in the zeroing and dirtying helpers

commit e28ae8e428fefe2facd72cea9f29906ecb9c861d upstream.

Fix the min_t calls in the zeroing and dirtying helpers to perform the
comparisms on 64-bit types, which prevents them from incorrectly
being truncated, and larger zeroing operations being stuck in a never
ending loop.

Special thanks to Markus Stockhausen for spotting the bug.

Reported-by: Paul Menzel <pmenzel@molgen.mpg.de>
Tested-by: Paul Menzel <pmenzel@molgen.mpg.de>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoxfs: don't leak quotacheck dquots when cow recovery
Darrick J. Wong [Sun, 17 Sep 2017 21:06:44 +0000 (14:06 -0700)]
xfs: don't leak quotacheck dquots when cow recovery

commit 77aff8c76425c8f49b50d0b9009915066739e7d2 upstream.

If we fail a mount on account of cow recovery errors, it's possible that
a previous quotacheck left some dquots in memory.  The bailout clause of
xfs_mountfs forgets to purge these, and so we leak them.  Fix that.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoxfs: clear MS_ACTIVE after finishing log recovery
Darrick J. Wong [Sun, 17 Sep 2017 21:06:43 +0000 (14:06 -0700)]
xfs: clear MS_ACTIVE after finishing log recovery

commit 8204f8ddaafafcae074746fcf2a05a45e6827603 upstream.

Way back when we established inode block-map redo log items, it was
discovered that we needed to prevent the VFS from evicting inodes during
log recovery because any given inode might be have bmap redo items to
replay even if the inode has no link count and is ultimately deleted,
and any eviction of an unlinked inode causes the inode to be truncated
and freed too early.

To make this possible, we set MS_ACTIVE so that inodes would not be torn
down immediately upon release.  Unfortunately, this also results in the
quota inodes not being released at all if a later part of the mount
process should fail, because we never reclaim the inodes.  So, set
MS_ACTIVE right before we do the last part of log recovery and clear it
immediately after we finish the log recovery so that everything
will be torn down properly if we abort the mount.

Fixes: 17c12bcd30 ("xfs: when replaying bmap operations, don't let unlinked inodes get reaped")
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoxfs: fix inobt inode allocation search optimization
Omar Sandoval [Sun, 17 Sep 2017 21:06:42 +0000 (14:06 -0700)]
xfs: fix inobt inode allocation search optimization

commit c44245b3d5435f533ca8346ece65918f84c057f9 upstream.

When we try to allocate a free inode by searching the inobt, we try to
find the inode nearest the parent inode by searching chunks both left
and right of the chunk containing the parent. As an optimization, we
cache the leftmost and rightmost records that we previously searched; if
we do another allocation with the same parent inode, we'll pick up the
search where it last left off.

There's a bug in the case where we found a free inode to the left of the
parent's chunk: we need to update the cached left and right records, but
because we already reassigned the right record to point to the left, we
end up assigning the left record to both the cached left and right
records.

This isn't a correctness problem strictly, but it can result in the next
allocation rechecking chunks unnecessarily or allocating inodes further
away from the parent than it needs to. Fix it by swapping the record
pointer after we update the cached left and right records.

Fixes: bd169565993b ("xfs: speed up free inode search")
Signed-off-by: Omar Sandoval <osandov@fb.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoxfs: Fix per-inode DAX flag inheritance
Lukas Czerner [Sun, 17 Sep 2017 21:06:41 +0000 (14:06 -0700)]
xfs: Fix per-inode DAX flag inheritance

commit 56bdf855e676f1f2ed7033f288f57dfd315725ba upstream.

According to the commit that implemented per-inode DAX flag:
commit 58f88ca2df72 ("xfs: introduce per-inode DAX enablement")
the flag is supposed to act as "inherit flag".

Currently this only works in the situations where parent directory
already has a flag in di_flags set, otherwise inheritance does not
work. This is because setting the XFS_DIFLAG2_DAX flag is done in a
wrong branch designated for di_flags, not di_flags2.

Fix this by moving the code to branch designated for setting di_flags2,
which does test for flags in di_flags2.

Fixes: 58f88ca2df72 ("xfs: introduce per-inode DAX enablement")
Signed-off-by: Lukas Czerner <lczerner@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoxfs: fix multi-AG deadlock in xfs_bunmapi
Christoph Hellwig [Sun, 17 Sep 2017 21:06:40 +0000 (14:06 -0700)]
xfs: fix multi-AG deadlock in xfs_bunmapi

commit 5b094d6dac0451ad89b1dc088395c7b399b7e9e8 upstream.

Just like in the allocator we must avoid touching multiple AGs out of
order when freeing blocks, as freeing still locks the AGF and can cause
the same AB-BA deadlocks as in the allocation path.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reported-by: Nikolay Borisov <n.borisov.lkml@gmail.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoxfs: fix quotacheck dquot id overflow infinite loop
Brian Foster [Sun, 17 Sep 2017 21:06:39 +0000 (14:06 -0700)]
xfs: fix quotacheck dquot id overflow infinite loop

commit cfaf2d034360166e569a4929dd83ae9698bed856 upstream.

If a dquot has an id of U32_MAX, the next lookup index increment
overflows the uint32_t back to 0. This starts the lookup sequence
over from the beginning, repeats indefinitely and results in a
livelock.

Update xfs_qm_dquot_walk() to explicitly check for the lookup
overflow and exit the loop.

Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoxfs: check _alloc_read_agf buffer pointer before using
Darrick J. Wong [Sun, 17 Sep 2017 21:06:38 +0000 (14:06 -0700)]
xfs: check _alloc_read_agf buffer pointer before using

commit 10479e2dea83d4c421ad05dfc55d918aa8dfc0cd upstream.

In some circumstances, _alloc_read_agf can return an error code of zero
but also a null AGF buffer pointer.  Check for this and jump out.

Fixes-coverity-id: 1415250
Fixes-coverity-id: 1415320
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoxfs: set firstfsb to NULLFSBLOCK before feeding it to _bmapi_write
Darrick J. Wong [Sun, 17 Sep 2017 21:06:37 +0000 (14:06 -0700)]
xfs: set firstfsb to NULLFSBLOCK before feeding it to _bmapi_write

commit 4c1a67bd3606540b9b42caff34a1d5cd94b1cf65 upstream.

We must initialize the firstfsb parameter to _bmapi_write so that it
doesn't incorrectly treat stack garbage as a restriction on which AGs
it can search for free space.

Fixes-coverity-id: 1402025
Fixes-coverity-id: 1415167
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoxfs: check _btree_check_block value
Darrick J. Wong [Sun, 17 Sep 2017 21:06:36 +0000 (14:06 -0700)]
xfs: check _btree_check_block value

commit 1e86eabe73b73c82e1110c746ed3ec6d5e1c0a0d upstream.

Check the _btree_check_block return value for the firstrec and lastrec
functions, since we have the ability to signal that the repositioning
did not succeed.

Fixes-coverity-id: 114067
Fixes-coverity-id: 114068
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoxfs: don't crash on unexpected holes in dir/attr btrees
Darrick J. Wong [Sun, 17 Sep 2017 21:06:35 +0000 (14:06 -0700)]
xfs: don't crash on unexpected holes in dir/attr btrees

commit cd87d867920155911d0d2e6485b769d853547750 upstream.

In quite a few places we call xfs_da_read_buf with a mappedbno that we
don't control, then assume that the function passes back either an error
code or a buffer pointer.  Unfortunately, if mappedbno == -2 and bno
maps to a hole, we get a return code of zero and a NULL buffer, which
means that we crash if we actually try to use that buffer pointer.  This
happens immediately when we set the buffer type for transaction context.

Therefore, check that we have no error code and a non-NULL bp before
trying to use bp.  This patch is a follow-up to an incomplete fix in
96a3aefb8ffde231 ("xfs: don't crash if reading a directory results in an
unexpected hole").

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoxfs: free cowblocks and retry on buffered write ENOSPC
Brian Foster [Sun, 17 Sep 2017 21:06:34 +0000 (14:06 -0700)]
xfs: free cowblocks and retry on buffered write ENOSPC

commit cf2cb7845d6e101cb17bd62f8aa08cd514fc8988 upstream.

XFS runs an eofblocks reclaim scan before returning an ENOSPC error to
userspace for buffered writes. This facilitates aggressive speculative
preallocation without causing user visible side effects such as
premature ENOSPC.

Run a cowblocks scan in the same situation to reclaim lingering COW fork
preallocation throughout the filesystem.

Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoxfs: free uncommitted transactions during log recovery
Brian Foster [Sun, 17 Sep 2017 21:06:33 +0000 (14:06 -0700)]
xfs: free uncommitted transactions during log recovery

commit 39775431f82f890f4aaa08860a30883d081bffc7 upstream.

Log recovery allocates in-core transaction and member item data
structures on-demand as it processes the on-disk log. Transactions
are allocated on first encounter on-disk and stored in a hash table
structure where they are easily accessible for subsequent lookups.
Transaction items are also allocated on demand and are attached to
the associated transactions.

When a commit record is encountered in the log, the transaction is
committed to the fs and the in-core structures are freed. If a
filesystem crashes or shuts down before all in-core log buffers are
flushed to the log, however, not all transactions may have commit
records in the log. As expected, the modifications in such an
incomplete transaction are not replayed to the fs. The in-core data
structures for the partial transaction are never freed, however,
resulting in a memory leak.

Update xlog_do_recovery_pass() to first correctly initialize the
hash table array so empty lists can be distinguished from populated
lists on function exit. Update xlog_recover_free_trans() to always
remove the transaction from the list prior to freeing the associated
memory. Finally, walk the hash table of transaction lists as the
last step before it goes out of scope and free any transactions that
may remain on the lists. This prevents a memory leak of partial
transactions in the log.

Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoxfs: don't allow bmap on rt files
Darrick J. Wong [Mon, 19 Jun 2017 20:19:08 +0000 (13:19 -0700)]
xfs: don't allow bmap on rt files

commit 61d819e7bcb7f33da710bf3f5dcb2bcf1e48203c upstream.

bmap returns a dumb LBA address but not the block device that goes with
that LBA.  Swapfiles don't care about this and will blindly assume that
the data volume is the correct blockdev, which is totally bogus for
files on the rt subvolume.  This results in the swap code doing IOs to
arbitrary locations on the data device(!) if the passed in mapping is a
realtime file, so just turn off bmap for rt files.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoxfs: remove bli from AIL before release on transaction abort
Brian Foster [Thu, 15 Jun 2017 04:35:35 +0000 (21:35 -0700)]
xfs: remove bli from AIL before release on transaction abort

commit 3d4b4a3e30ae7a949c31e1e10268a3da4723d290 upstream.

When a buffer is modified, logged and committed, it ultimately ends
up sitting on the AIL with a dirty bli waiting for metadata
writeback. If another transaction locks and invalidates the buffer
(freeing an inode chunk, for example) in the meantime, the bli is
flagged as stale, the dirty state is cleared and the bli remains in
the AIL.

If a shutdown occurs before the transaction that has invalidated the
buffer is committed, the transaction is ultimately aborted. The log
items are flagged as such and ->iop_unlock() handles the aborted
items. Because the bli is clean (due to the invalidation),
->iop_unlock() unconditionally releases it. The log item may still
reside in the AIL, however, which means the I/O completion handler
may still run and attempt to access it. This results in assert
failure due to the release of the bli while still present in the AIL
and a subsequent NULL dereference and panic in the buffer I/O
completion handling. This can be reproduced by running generic/388
in repetition.

To avoid this problem, update xfs_buf_item_unlock() to first check
whether the bli is aborted and if so, remove it from the AIL before
it is released. This ensures that the bli is no longer accessed
during the shutdown sequence after it has been freed.

Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Carlos Maiolino <cmaiolino@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoxfs: release bli from transaction properly on fs shutdown
Brian Foster [Thu, 15 Jun 2017 04:35:35 +0000 (21:35 -0700)]
xfs: release bli from transaction properly on fs shutdown

commit 79e641ce29cfae5b8fc55fb77ac62d11d2d849c0 upstream.

If a filesystem shutdown occurs with a buffer log item in the CIL
and a log force occurs, the ->iop_unpin() handler is generally
expected to tear down the bli properly. This entails freeing the bli
memory and releasing the associated hold on the buffer so it can be
released and the filesystem unmounted.

If this sequence occurs while ->bli_refcount is elevated (i.e.,
another transaction is open and attempting to modify the buffer),
however, ->iop_unpin() may not be responsible for releasing the bli.
Instead, the transaction may release the final ->bli_refcount
reference and thus xfs_trans_brelse() is responsible for tearing
down the bli.

While xfs_trans_brelse() does drop the reference count, it only
attempts to release the bli if it is clean (i.e., not in the
CIL/AIL). If the filesystem is shutdown and the bli is sitting dirty
in the CIL as noted above, this ends up skipping the last
opportunity to release the bli. In turn, this leaves the hold on the
buffer and causes an unmount hang. This can be reproduced by running
generic/388 in repetition.

Update xfs_trans_brelse() to handle this shutdown corner case
correctly. If the final bli reference is dropped and the filesystem
is shutdown, remove the bli from the AIL (if necessary) and release
the bli to drop the buffer hold and ensure an unmount does not hang.

Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Carlos Maiolino <cmaiolino@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoxfs: try to avoid blowing out the transaction reservation when bunmaping a shared...
Darrick J. Wong [Thu, 15 Jun 2017 04:25:57 +0000 (21:25 -0700)]
xfs: try to avoid blowing out the transaction reservation when bunmaping a shared extent

commit e1a4e37cc7b665b6804fba812aca2f4d7402c249 upstream.

In a pathological scenario where we are trying to bunmapi a single
extent in which every other block is shared, it's possible that trying
to unmap the entire large extent in a single transaction can generate so
many EFIs that we overflow the transaction reservation.

Therefore, use a heuristic to guess at the number of blocks we can
safely unmap from a reflink file's data fork in an single transaction.
This should prevent problems such as the log head slamming into the tail
and ASSERTs that trigger because we've exceeded the transaction
reservation.

Note that since bunmapi can fail to unmap the entire range, we must also
teach the deferred unmap code to roll into a new transaction whenever we
get low on reservation.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
[hch: random edits, all bugs are my fault]
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoxfs: push buffer of flush locked dquot to avoid quotacheck deadlock
Brian Foster [Thu, 15 Jun 2017 04:21:45 +0000 (21:21 -0700)]
xfs: push buffer of flush locked dquot to avoid quotacheck deadlock

commit 7912e7fef2aebe577f0b46d3cba261f2783c5695 upstream.

Reclaim during quotacheck can lead to deadlocks on the dquot flush
lock:

 - Quotacheck populates a local delwri queue with the physical dquot
   buffers.
 - Quotacheck performs the xfs_qm_dqusage_adjust() bulkstat and
   dirties all of the dquots.
 - Reclaim kicks in and attempts to flush a dquot whose buffer is
   already queud on the quotacheck queue. The flush succeeds but
   queueing to the reclaim delwri queue fails as the backing buffer is
   already queued. The flush unlock is now deferred to I/O completion
   of the buffer from the quotacheck queue.
 - The dqadjust bulkstat continues and dirties the recently flushed
   dquot once again.
 - Quotacheck proceeds to the xfs_qm_flush_one() walk which requires
   the flush lock to update the backing buffers with the in-core
   recalculated values. It deadlocks on the redirtied dquot as the
   flush lock was already acquired by reclaim, but the buffer resides
   on the local delwri queue which isn't submitted until the end of
   quotacheck.

This is reproduced by running quotacheck on a filesystem with a
couple million inodes in low memory (512MB-1GB) situations. This is
a regression as of commit 43ff2122e6 ("xfs: on-stack delayed write
buffer lists"), which removed a trylock and buffer I/O submission
from the quotacheck dquot flush sequence.

Quotacheck first resets and collects the physical dquot buffers in a
delwri queue. Then, it traverses the filesystem inodes via bulkstat,
updates the in-core dquots, flushes the corrected dquots to the
backing buffers and finally submits the delwri queue for I/O. Since
the backing buffers are queued across the entire quotacheck
operation, dquot reclaim cannot possibly complete a dquot flush
before quotacheck completes.

Therefore, quotacheck must submit the buffer for I/O in order to
cycle the flush lock and flush the dirty in-core dquot to the
buffer. Add a delwri queue buffer push mechanism to submit an
individual buffer for I/O without losing the delwri queue status and
use it from quotacheck to avoid the deadlock. This restores
quotacheck behavior to as before the regression was introduced.

Reported-by: Martin Svec <martin.svec@zoner.cz>
Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoxfs: fix spurious spin_is_locked() assert failures on non-smp kernels
Brian Foster [Thu, 8 Jun 2017 15:23:07 +0000 (08:23 -0700)]
xfs: fix spurious spin_is_locked() assert failures on non-smp kernels

commit 95989c46d2a156365867b1d795fdefce71bce378 upstream.

The 0-day kernel test robot reports assertion failures on
!CONFIG_SMP kernels due to failed spin_is_locked() checks. As it
turns out, spin_is_locked() is hardcoded to return zero on
!CONFIG_SMP kernels and so this function cannot be relied on to
verify spinlock state in this configuration.

To avoid this problem, replace the associated asserts with lockdep
variants that do the right thing regardless of kernel configuration.
Drop the one assert that checks for an unlocked lock as there is no
suitable lockdep variant for that case. This moves the spinlock
checks from XFS debug code to lockdep, but generally provides the
same level of protection.

Reported-by: kbuild test robot <fengguang.wu@intel.com>
Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoxfs: Move handling of missing page into one place in xfs_find_get_desired_pgoff()
Jan Kara [Thu, 18 May 2017 23:36:24 +0000 (16:36 -0700)]
xfs: Move handling of missing page into one place in xfs_find_get_desired_pgoff()

commit a54fba8f5a0dc36161cacdf2aa90f007f702ec1a upstream.

Currently several places in xfs_find_get_desired_pgoff() handle the case
of a missing page. Make them all handled in one place after the loop has
terminated.

Signed-off-by: Jan Kara <jack@suse.cz>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agox86/switch_to/64: Rewrite FS/GS switching yet again to fix AMD CPUs
Andy Lutomirski [Tue, 1 Aug 2017 14:11:37 +0000 (07:11 -0700)]
x86/switch_to/64: Rewrite FS/GS switching yet again to fix AMD CPUs

commit e137a4d8f4dd2e277e355495b6b2cb241a8693c3 upstream.

Switching FS and GS is a mess, and the current code is still subtly
wrong: it assumes that "Loading a nonzero value into FS sets the
index and base", which is false on AMD CPUs if the value being
loaded is 1, 2, or 3.

(The current code came from commit 3e2b68d752c9 ("x86/asm,
sched/x86: Rewrite the FS and GS context switch code"), which made
it better but didn't fully fix it.)

Rewrite it to be much simpler and more obviously correct.  This
should fix it fully on AMD CPUs and shouldn't adversely affect
performance.

Signed-off-by: Andy Lutomirski <luto@kernel.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Borislav Petkov <bpetkov@suse.de>
Cc: Brian Gerst <brgerst@gmail.com>
Cc: Chang Seok <chang.seok.bae@intel.com>
Cc: Denys Vlasenko <dvlasenk@redhat.com>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Josh Poimboeuf <jpoimboe@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agox86/fsgsbase/64: Report FSBASE and GSBASE correctly in core dumps
Andy Lutomirski [Tue, 1 Aug 2017 14:11:35 +0000 (07:11 -0700)]
x86/fsgsbase/64: Report FSBASE and GSBASE correctly in core dumps

commit 9584d98bed7a7a904d0702ad06bbcc94703cb5b4 upstream.

In ELF_COPY_CORE_REGS, we're copying from the current task, so
accessing thread.fsbase and thread.gsbase makes no sense.  Just read
the values from the CPU registers.

In practice, the old code would have been correct most of the time
simply because thread.fsbase and thread.gsbase usually matched the
CPU registers.

Signed-off-by: Andy Lutomirski <luto@kernel.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Borislav Petkov <bpetkov@suse.de>
Cc: Brian Gerst <brgerst@gmail.com>
Cc: Chang Seok <chang.seok.bae@intel.com>
Cc: Denys Vlasenko <dvlasenk@redhat.com>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Josh Poimboeuf <jpoimboe@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agox86/fsgsbase/64: Fully initialize FS and GS state in start_thread_common
Andy Lutomirski [Tue, 1 Aug 2017 14:11:34 +0000 (07:11 -0700)]
x86/fsgsbase/64: Fully initialize FS and GS state in start_thread_common

commit 767d035d838f4fd6b5a5bbd7a3f6d293b7f65a49 upstream.

execve used to leak FSBASE and GSBASE on AMD CPUs.  Fix it.

The security impact of this bug is small but not quite zero -- it
could weaken ASLR when a privileged task execs a less privileged
program, but only if program changed bitness across the exec, or the
child binary was highly unusual or actively malicious.  A child
program that was compromised after the exec would not have access to
the leaked base.

Signed-off-by: Andy Lutomirski <luto@kernel.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Borislav Petkov <bpetkov@suse.de>
Cc: Brian Gerst <brgerst@gmail.com>
Cc: Chang Seok <chang.seok.bae@intel.com>
Cc: Denys Vlasenko <dvlasenk@redhat.com>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Josh Poimboeuf <jpoimboe@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agof2fs: check hot_data for roll-forward recovery
Jaegeuk Kim [Sun, 13 Aug 2017 04:33:23 +0000 (21:33 -0700)]
f2fs: check hot_data for roll-forward recovery

commit 125c9fb1ccb53eb2ea9380df40f3c743f3fb2fed upstream.

We need to check HOT_DATA to truncate any previous data block when doing
roll-forward recovery.

Reviewed-by: Chao Yu <yuchao0@huawei.com>
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agof2fs: let fill_super handle roll-forward errors
Jaegeuk Kim [Fri, 11 Aug 2017 00:35:04 +0000 (17:35 -0700)]
f2fs: let fill_super handle roll-forward errors

commit afd2b4da40b3b567ef8d8e6881479345a2312a03 upstream.

If we set CP_ERROR_FLAG in roll-forward error, f2fs is no longer to proceed
any IOs due to f2fs_cp_error(). But, for example, if some stale data is involved
on roll-forward process, we're able to get -ENOENT, getting fs stuck.
If we get any error, let fill_super set SBI_NEED_FSCK and try to recover back
to stable point.

Reviewed-by: Chao Yu <yuchao0@huawei.com>
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoip_tunnel: fix setting ttl and tos value in collect_md mode
Haishuang Yan [Thu, 7 Sep 2017 06:08:34 +0000 (14:08 +0800)]
ip_tunnel: fix setting ttl and tos value in collect_md mode

[ Upstream commit 0f693f1995cf002432b70f43ce73f79bf8d0b6c9 ]

ttl and tos variables are declared and assigned, but are not used in
iptunnel_xmit() function.

Fixes: cfc7381b3002 ("ip_tunnel: add collect_md mode to IPIP tunnel")
Cc: Alexei Starovoitov <ast@fb.com>
Signed-off-by: Haishuang Yan <yanhaishuang@cmss.chinamobile.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agosctp: fix missing wake ups in some situations
Marcelo Ricardo Leitner [Fri, 8 Sep 2017 14:35:21 +0000 (11:35 -0300)]
sctp: fix missing wake ups in some situations

[ Upstream commit 7906b00f5cd1cd484fced7fcda892176e3202c8a ]

Commit fb586f25300f ("sctp: delay calls to sk_data_ready() as much as
possible") minimized the number of wake ups that are triggered in case
the association receives a packet with multiple data chunks on it and/or
when io_events are enabled and then commit 0970f5b36659 ("sctp: signal
sk_data_ready earlier on data chunks reception") moved the wake up to as
soon as possible. It thus relies on the state machine running later to
clean the flag that the event was already generated.

The issue is that there are 2 call paths that calls
sctp_ulpq_tail_event() outside of the state machine, causing the flag to
linger and possibly omitting a needed wake up in the sequence.

One of the call paths is when enabling SCTP_SENDER_DRY_EVENTS via
setsockopt(SCTP_EVENTS), as noticed by Harald Welte. The other is when
partial reliability triggers removal of chunks from the send queue when
the application calls sendmsg().

This commit fixes it by not setting the flag in case the socket is not
owned by the user, as it won't be cleaned later. This works for
user-initiated calls and also for rx path processing.

Fixes: fb586f25300f ("sctp: delay calls to sk_data_ready() as much as possible")
Reported-by: Harald Welte <laforge@gnumonks.org>
Signed-off-by: Marcelo Ricardo Leitner <marcelo.leitner@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoipv6: fix typo in fib6_net_exit()
Eric Dumazet [Fri, 8 Sep 2017 22:48:47 +0000 (15:48 -0700)]
ipv6: fix typo in fib6_net_exit()

[ Upstream commit 32a805baf0fb70b6dbedefcd7249ac7f580f9e3b ]

IPv6 FIB should use FIB6_TABLE_HASHSZ, not FIB_TABLE_HASHSZ.

Fixes: ba1cc08d9488 ("ipv6: fix memory leak with multiple tables during netns destruction")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoipv6: fix memory leak with multiple tables during netns destruction
Sabrina Dubroca [Fri, 8 Sep 2017 08:26:19 +0000 (10:26 +0200)]
ipv6: fix memory leak with multiple tables during netns destruction

[ Upstream commit ba1cc08d9488c94cb8d94f545305688b72a2a300 ]

fib6_net_exit only frees the main and local tables. If another table was
created with fib6_alloc_table, we leak it when the netns is destroyed.

Fix this in the same way ip_fib_net_exit cleans up tables, by walking
through the whole hashtable of fib6_table's. We can get rid of the
special cases for local and main, since they're also part of the
hashtable.

Reproducer:
    ip netns add x
    ip -net x -6 rule add from 6003:1::/64 table 100
    ip netns del x

Reported-by: Jianlin Shi <jishi@redhat.com>
Fixes: 58f09b78b730 ("[NETNS][IPV6] ip6_fib - make it per network namespace")
Signed-off-by: Sabrina Dubroca <sd@queasysnail.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
6 years agoip6_gre: update mtu properly in ip6gre_err
Xin Long [Tue, 5 Sep 2017 09:26:33 +0000 (17:26 +0800)]
ip6_gre: update mtu properly in ip6gre_err

[ Upstream commit 5c25f30c93fdc5bf25e62101aeaae7a4f9b421b3 ]

Now when probessing ICMPV6_PKT_TOOBIG, ip6gre_err only subtracts the
offset of gre header from mtu info. The expected mtu of gre device
should also subtract gre header. Otherwise, the next packets still
can't be sent out.

Jianlin found this issue when using the topo:
  client(ip6gre)<---->(nic1)route(nic2)<----->(ip6gre)server

and reducing nic2's mtu, then both tcp and sctp's performance with
big size data became 0.

This patch is to fix it by also subtracting grehdr (tun->tun_hlen)
from mtu info when updating gre device's mtu in ip6gre_err(). It
also needs to subtract ETH_HLEN if gre dev'type is ARPHRD_ETHER.

Reported-by: Jianlin Shi <jishi@redhat.com>
Signed-off-by: Xin Long <lucien.xin@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>