This particular Kioxia device times out and aborts I/O during any load,
but it's more easily observable with discards (fstrim).
The device gets to a state that is also not possible to use
"nvme set-feature" to disable APST.
Booting with nvme_core.default_ps_max_latency=0 solves the issue.
We had a dozen or so of these devices behaving this same way in
customer environments.
Signed-off-by: Enzo Matsumiya <ematsumiya@suse.de>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Pull block fixes from Jens Axboe:
- Set of fixes for the batched tag allocation (Ming, me)
- add_disk() error handling fix (Luis)
- Nested queue quiesce fixes (Ming)
- Shared tags init error handling fix (Ye)
- Misc cleanups (Jean, Ming, me)
* tag 'for-5.16/block-2021-11-09' of git://git.kernel.dk/linux-block:
nvme: wait until quiesce is done
scsi: make sure that request queue queiesce and unquiesce balanced
scsi: avoid to quiesce sdev->request_queue two times
blk-mq: add one API for waiting until quiesce is done
blk-mq: don't free tags if the tag_set is used by other device in queue initialztion
block: fix device_add_disk() kobject_create_and_add() error handling
block: ensure cached plug request matches the current queue
block: move queue enter logic into blk_mq_submit_bio()
block: make bio_queue_enter() fast-path available inline
block: split request allocation components into helpers
block: have plug stored requests hold references to the queue
blk-mq: update hctx->nr_active in blk_mq_end_request_batch()
blk-mq: add RQF_ELV debug entry
blk-mq: only try to run plug merge if request has same queue with incoming bio
block: move RQF_ELV setting into allocators
dm: don't stop request queue after the dm device is suspended
block: replace always false argument with 'false'
block: assign correct tag before doing prefetch of request
blk-mq: fix redundant check of !e expression
Pull block driver updates from Jens Axboe:
- paride driver cleanups (Christoph)
- Remove cryptoloop support (Christoph)
- null_blk poll support (me)
- Now that add_disk() supports proper error handling, add it to various
drivers (Luis)
- Make ataflop actually work again (Michael)
- s390 dasd fixes (Stefan, Heiko)
- nbd fixes (Yu, Ye)
- Remove redundant wq flush in mtip32xx (Christophe)
- NVMe updates
- fix a multipath partition scanning deadlock (Hannes Reinecke)
- generate uevent once a multipath namespace is operational again
(Hannes Reinecke)
- support unique discovery controller NQNs (Hannes Reinecke)
- fix use-after-free when a port is removed (Israel Rukshin)
- clear shadow doorbell memory on resets (Keith Busch)
- use struct_size (Len Baker)
- add error handling support for add_disk (Luis Chamberlain)
- limit the maximal queue size for RDMA controllers (Max Gurtovoy)
- use a few more symbolic names (Max Gurtovoy)
- fix error code in nvme_rdma_setup_ctrl (Max Gurtovoy)
- add support for ->map_queues on FC (Saurav Kashyap)
- support the current discovery subsystem entry (Hannes Reinecke)
- use flex_array_size and struct_size (Len Baker)
- bcache fixes (Christoph, Coly, Chao, Lin, Qing)
- MD updates (Christoph, Guoqing, Xiao)
- Misc fixes (Dan, Ding, Jiapeng, Shin'ichiro, Ye)
* tag 'for-5.16/drivers-2021-10-29' of git://git.kernel.dk/linux-block: (117 commits)
null_blk: Fix handling of submit_queues and poll_queues attributes
block: ataflop: Fix warning comparing pointer to 0
bcache: replace snprintf in show functions with sysfs_emit
bcache: move uapi header bcache.h to bcache code directory
nvmet: use flex_array_size and struct_size
nvmet: register discovery subsystem as 'current'
nvmet: switch check for subsystem type
nvme: add new discovery log page entry definitions
block: ataflop: more blk-mq refactoring fixes
block: remove support for cryptoloop and the xor transfer
mtd: add add_disk() error handling
rnbd: add error handling support for add_disk()
um/drivers/ubd_kern: add error handling support for add_disk()
m68k/emu/nfblock: add error handling support for add_disk()
xen-blkfront: add error handling support for add_disk()
bcache: add error handling support for add_disk()
dm: add add_disk() error handling
block: aoe: fixup coccinelle warnings
nvmet: use struct_size over open coded arithmetic
nvme: drop scan_lock and always kick requeue list when removing namespaces
...
With discovery controllers supporting unique subsystem NQNs the
actual subsystem NQN might be different from that one passed in
via the connect args. So add a helper to display the resulting
subsystem NQN.
Signed-off-by: Hannes Reinecke <hare@suse.de>
Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Add a connect option 'discovery' to specify that the connection
should be made to a discovery controller, not a normal I/O controller.
With discovery controllers supporting unique subsystem NQNs we
cannot easily distinguish by the subsystem NQN if this should be
a discovery connection, but we need this information to blank out
options not supported by discovery controllers.
Signed-off-by: Hannes Reinecke <hare@suse.de>
Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
With unique discovery controller NQNs we cannot distinguish the
subsystem type by the NQN alone, but need to check the subsystem
type, too.
So expose the subsystem type in a new sysfs attribute 'subsystype'.
Signed-off-by: Hannes Reinecke <hare@suse.de>
Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
The current blk_mq_quiesce_queue() and blk_mq_unquiesce_queue() always
stops and starts the queue unconditionally. And there can be concurrent
quiesce/unquiesce coming from different unrelated code paths, so
unquiesce may come unexpectedly and start queue too early.
Prepare for supporting concurrent quiesce/unquiesce from multiple
contexts, so that we can address the above issue.
NVMe has very complicated quiesce/unquiesce use pattern, add one atomic
bit for makeiing sure that blk-mq quiece/unquiesce is always called in
pair.
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Link: https://lore.kernel.org/r/20211014081710.1871747-5-ming.lei@redhat.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
This memset in the fast path costs a lot of cycles on my setup. Here's a
top-of-profile of doing ~6.7M IOPS:
+ 5.90% io_uring [nvme] [k] nvme_queue_rq
+ 5.32% io_uring [nvme_core] [k] nvme_setup_cmd
+ 5.17% io_uring [kernel.vmlinux] [k] io_submit_sqes
+ 4.97% io_uring [kernel.vmlinux] [k] blkdev_direct_IO
and a perf diff with this patch:
0.92% +4.40% [nvme_core] [k] nvme_setup_cmd
reducing it from 5.3% to only 0.9%. This takes it from the 2nd most
cycle consumer to something that's mostly irrelevant.
Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
Reviewed-by: Keith Busch <kbusch@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We don't have to worry about doing extra memsets by moving it outside
the protection of RQF_DONTPREP, as nvme doesn't do partial completions.
This is in preparation for making the read/write fast path not do a full
memset of the command.
Reviewed-by: Keith Busch <kbusch@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Take advantage of struct io_comp_batch, if passed in to the nvme poll
handler. If it's set, rather than complete each request individually
inline, store them in the io_comp_batch list. We only do so for requests
that will complete successfully, anything else will be completed inline as
before.
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Decrease reference count of chardevice during char device deletion in
order to fix a memory leak. Add a release callabck for the device
associated chardev and move ida_simple_remove into the release function.
Fixes: 2637baed78 ("nvme: introduce generic per-namespace chardev")
Reported-by: Yi Zhang <yi.zhang@redhat.com>
Suggested-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Adam Manzanares <a.manzanares@samsung.com>
Reviewed-by: Javier González <javier@javigon.com>
Tested-by: Yi Zhang <yi.zhang@redhat.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Various places in the nvme code that rely on ctrl->namespace to be
ordered. Ensure that the namespae is inserted into the list at the
right position from the start instead of sorting it after the fact.
Fixes: 540c801c65 ("NVMe: Implement namespace list scanning")
Reported-by: Anton Eidelman <anton.eidelman@gmail.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Keith Busch <kbusch@kernel.org>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
Reviewed-by: Damien Le Moal <damien.lemoal@wdc.com>
Pull NVMe fixes from Christoph:
"nvme fixes for Linux 5.15
- fix ANA state updates when a namespace is not present (Anton Eidelman)
- nvmet: fix a width vs precision bug in nvmet_subsys_attr_serial_show
(Dan Carpenter)
- avoid race in shutdown namespace removal (Daniel Wagner)
- fix io_work priority inversion in nvme-tcp (Keith Busch)
- destroy cm id before destroy qp to avoid use after free (Ruozhu Li)"
* tag 'nvme-5.15-2021-09-15' of git://git.infradead.org/nvme:
nvme-tcp: fix io_work priority inversion
nvme-rdma: destroy cm id before destroy qp to avoid use after free
nvme-multipath: fix ANA state updates when a namespace is not present
nvme: avoid race in shutdown namespace removal
nvmet: fix a width vs precision bug in nvmet_subsys_attr_serial_show()
When we remove the siblings entry, we update ns->head->list, hence we
can't separate the removal and test for being empty. They have to be
in the same critical section to avoid a race.
To avoid breaking the refcounting imbalance again, add a list empty
check to nvme_find_ns_head.
Fixes: 5396fdac56 ("nvme: fix refcounting imbalance when all paths are down")
Signed-off-by: Daniel Wagner <dwagner@suse.de>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Tested-by: Hannes Reinecke <hare@suse.de>
Signed-off-by: Christoph Hellwig <hch@lst.de>
We never checked for errors on add_disk() as this function
returned void. Now that this is fixed, use the shiny new
error handling.
Signed-off-by: Luis Chamberlain <mcgrof@kernel.org>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Signed-off-by: Christoph Hellwig <hch@lst.de>
The function nmve_mpath_clear_current_path returns true if the current
path has changed. In this case we have to wait for all concurrent
submissions to finish. But if we didn't change the current path, there
is no point in waiting for another RCU period to finish.
Signed-off-by: Daniel Wagner <dwagner@suse.de>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Currently the connection between host and NVMe-oF target gets
disconnected by keep-alive timeout when a user connects to a target
with a relatively large kato value and then sets the smaller kato
with a set features command (e.g. connects with 60 seconds kato value
and then sets 10 seconds kato value).
The cause is that keep alive command interval on the host, which is
defined as unsigned int kato in nvme_ctrl structure, does not follow
the kato value changes.
This patch updates the keep alive interval in the following steps when
the kato is modified by a set features command: stops the keep alive
work queue, then sets the kato as new timer value and re-start the queue.
Signed-off-by: Tatsuya Sasaki <tatsuya6.sasaki@kioxia.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Preparatory patch in order to reuse nvme_multi_css in the nvme target
code.
Signed-off-by: Adam Manzanares <a.manzanares@samsung.com>
Reviewed-by: Keith Busch <kbusch@kernel.org>
Signed-off-by: Christoph Hellwig <hch@lst.de>
When triggering a rescan due to a namespace resize we will be
receiving AENs on every controller, triggering a rescan of all
attached namespaces. If multipath is active only the current path and
the ns_head disk will be updated, the other paths will still refer to
the old size until AENs for the remaining controllers are received.
If I/O comes in before that it might be routed to one of the old
paths, triggering an I/O failure with 'access beyond end of device'.
With this patch the old paths are skipped from multipath path
selection until the controller serving these paths has been rescanned.
Signed-off-by: Hannes Reinecke <hare@suse.de>
[dwagner: - introduce NVME_NS_READY flag instead of NVME_NS_INVALIDATE
- use 'revalidate' instead of 'invalidate' which
follows the zoned device code path.
- clear NVME_NS_READY before clearing current_path]
Signed-off-by: Daniel Wagner <dwagner@suse.de>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Pull block driver updates from Jens Axboe:
"Sitting on top of the core block changes, here are the driver changes
for the 5.15 merge window:
- NVMe updates via Christoph:
- suspend improvements for devices with an HMB (Keith Busch)
- handle double completions more gacefull (Sagi Grimberg)
- cleanup the selects for the nvme core code a bit (Sagi Grimberg)
- don't update queue count when failing to set io queues (Ruozhu Li)
- various nvmet connect fixes (Amit Engel)
- cleanup lightnvm leftovers (Keith Busch, me)
- small cleanups (Colin Ian King, Hou Pu)
- add tracing for the Set Features command (Hou Pu)
- CMB sysfs cleanups (Keith Busch)
- add a mutex_destroy call (Keith Busch)
- remove lightnvm subsystem. It's served its purpose and ultimately
led to zoned nvme support, we no longer need it (Christoph)
- revert floppy O_NDELAY fix (Denis)
- nbd fixes (Hou, Pavel, Baokun)
- nbd locking fixes (Tetsuo)
- nbd device removal fixes (Christoph)
- raid10 rcu warning fix (Xiao)
- raid1 write behind fix (Guoqing)
- rnbd fixes (Gioh, Md Haris)
- misc fixes (Colin)"
* tag 'for-5.15/drivers-2021-08-30' of git://git.kernel.dk/linux-block: (42 commits)
Revert "floppy: reintroduce O_NDELAY fix"
raid1: ensure write behind bio has less than BIO_MAX_VECS sectors
md/raid10: Remove unnecessary rcu_dereference in raid10_handle_discard
nbd: remove nbd->destroy_complete
nbd: only return usable devices from nbd_find_unused
nbd: set nbd->index before releasing nbd_index_mutex
nbd: prevent IDR lookups from finding partially initialized devices
nbd: reset NBD to NULL when restarting in nbd_genl_connect
nbd: add missing locking to the nbd_dev_add error path
nvme: remove the unused NVME_NS_* enum
nvme: remove nvm_ndev from ns
nvme: Have NVME_FABRICS select NVME_CORE instead of transport drivers
block: nbd: add sanity check for first_minor
nvmet: check that host sqsize does not exceed ctrl MQES
nvmet: avoid duplicate qid in connect cmd
nvmet: pass back cntlid on successful completion
nvme-rdma: don't update queue count when failing to set io queues
nvme-tcp: don't update queue count when failing to set io queues
nvme-tcp: pair send_mutex init with destroy
nvme: allow user toggling hmb usage
...
We cannot detect a (perhaps buggy) controller that is sending us
a completion for a request that was already completed (for example
sending a completion twice), this phenomenon was seen in the wild
a few times.
So to protect against this, we use the upper 4 msbits of the nvme sqe
command_id to use as a 4-bit generation counter and verify it matches
the existing request generation that is incrementing on every execution.
The 16-bit command_id structure now is constructed by:
| xxxx | xxxxxxxxxxxx |
gen request tag
This means that we are giving up some possible queue depth as 12 bits
allow for a maximum queue depth of 4095 instead of 65536, however we
never create such long queues anyways so no real harm done.
Suggested-by: Keith Busch <kbusch@kernel.org>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Acked-by: Keith Busch <kbusch@kernel.org>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Reviewed-by: Daniel Wagner <dwagner@suse.de>
Tested-by: Daniel Wagner <dwagner@suse.de>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Lightnvm supports the OCSSD 1.x and 2.0 specs which were early attempts
to produce Open Channel SSDs and never made it into the NVMe spec
proper. They have since been superceeded by NVMe enhancements such
as ZNS support. Remove the support per the deprecation schedule.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Link: https://lore.kernel.org/r/20210812132308.38486-1-hch@lst.de
Reviewed-by: Matias Bjørling <mb@lightnvm.io>
Reviewed-by: Javier González <javier@javigon.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
When using Write Zeroes on a namespace that has protection
information enabled they behavior without the PRACT bit
counter-intuitive and will generally lead to validation failures
when reading the written blocks. Fix this by always setting the
PRACT bit that generates matching PI data on the fly.
Fixes: 6e02318eae ("nvme: add support for the Write Zeroes command")
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Keith Busch <kbusch@kernel.org>
Reviewed-by: Martin K. Petersen <martin.petersen@oracle.com>
When the last path to a ns_head drops the current code
removes the ns_head from the subsystem list, but will only
delete the disk itself if the last reference to the ns_head
drops. This is causing an refcounting imbalance eg when
applications have a reference to the disk, as then they'll
never get notified that the disk is in fact dead.
This patch moves the call 'del_gendisk' into nvme_mpath_check_last_path(),
ensuring that the disk can be properly removed and applications get the
appropriate notifications.
Signed-off-by: Hannes Reinecke <hare@suse.de>
Reviewed-by: Keith Busch <kbusch@kernel.org>
Signed-off-by: Christoph Hellwig <hch@lst.de>
We don't have an nvme status to report if the driver's .queue_rq()
returns an error without dispatching the requested nvme command. Check
the return value from blk_execute_rq() for all passthrough commands so
the caller may know their command was not successful.
If the command is from the target passthrough interface and fails to
dispatch, synthesize the response back to the host as a internal target
error.
Signed-off-by: Keith Busch <kbusch@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com>
Link: https://lore.kernel.org/r/20210610214437.641245-5-kbusch@kernel.org
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Pull block driver updates from Jens Axboe:
"Pretty calm round, mostly just NVMe and a bit of MD:
- NVMe updates (via Christoph)
- improve the APST configuration algorithm (Alexey Bogoslavsky)
- look for StorageD3Enable on companion ACPI device
(Mario Limonciello)
- allow selecting the network interface for TCP connections
(Martin Belanger)
- misc cleanups (Amit Engel, Chaitanya Kulkarni, Colin Ian King,
Christoph)
- move the ACPI StorageD3 code to drivers/acpi/ and add quirks
for certain AMD CPUs (Mario Limonciello)
- zoned device support for nvmet (Chaitanya Kulkarni)
- fix the rules for changing the serial number in nvmet
(Noam Gottlieb)
- various small fixes and cleanups (Dan Carpenter, JK Kim,
Chaitanya Kulkarni, Hannes Reinecke, Wesley Sheng, Geert
Uytterhoeven, Daniel Wagner)
- MD updates (Via Song)
- iostats rewrite (Guoqing Jiang)
- raid5 lock contention optimization (Gal Ofri)
- Fall through warning fix (Gustavo)
- Misc fixes (Gustavo, Jiapeng)"
* tag 'for-5.14/drivers-2021-06-29' of git://git.kernel.dk/linux-block: (78 commits)
nvmet: use NVMET_MAX_NAMESPACES to set nn value
loop: Fix missing discard support when using LOOP_CONFIGURE
nvme.h: add missing nvme_lba_range_type endianness annotations
nvme: remove zeroout memset call for struct
nvme-pci: remove zeroout memset call for struct
nvmet: remove zeroout memset call for struct
nvmet: add ZBD over ZNS backend support
nvmet: add Command Set Identifier support
nvmet: add nvmet_req_bio put helper for backends
nvmet: add req cns error complete helper
block: export blk_next_bio()
nvmet: remove local variable
nvmet: use nvme status value directly
nvmet: use u32 type for the local variable nsid
nvmet: use u32 for nvmet_subsys max_nsid
nvmet: use req->cmd directly in file-ns fast path
nvmet: use req->cmd directly in bdev-ns fast path
nvmet: make ver stable once connection established
nvmet: allow mn change if subsys not discovered
nvmet: make sn stable once connection was established
...
Declare and initialize structure variables to zero values so that we can
remove zeroout memset calls in the host/core.c.
Signed-off-by: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Now that only one caller is left remove the helpers by restructuring
nvme_pr_command so that it has two helpers for sending a command of to a
given nsid using either the ns_head for multipath, or the namespace
stored in the gendisk.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Keith Busch <kbusch@kernel.org>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Split multipath support out of nvme_report_zones into a separate helper
and simplify the non-multipath version as a result.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Keith Busch <kbusch@kernel.org>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
In our application, we need a way to force TCP connections to go out a
specific IP interface instead of letting Linux select the interface
based on the routing tables.
Add the 'host-iface' option to allow specifying the interface to use.
When the option host-iface is specified, the driver uses the specified
interface to set the option SO_BINDTODEVICE on the TCP socket before
connecting.
This new option is needed in addtion to the existing host-traddr for
the following reasons:
Specifying an IP interface by its associated IP address is less
intuitive than specifying the actual interface name and, in some cases,
simply doesn't work. That's because the association between interfaces
and IP addresses is not predictable. IP addresses can be changed or can
change by themselves over time (e.g. DHCP). Interface names are
predictable [1] and will persist over time. Consider the following
configuration.
1: lo: <LOOPBACK,UP,LOWER_UP> mtu 65536 qdisc noqueue state ...
link/loopback 00:00:00:00:00:00 brd 00:00:00:00:00:00
inet 100.0.0.100/24 scope global lo
valid_lft forever preferred_lft forever
2: enp0s3: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc ...
link/ether 08:00:27:21:65:ec brd ff:ff:ff:ff:ff:ff
inet 100.0.0.100/24 scope global enp0s3
valid_lft forever preferred_lft forever
3: enp0s8: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc ...
link/ether 08:00:27:4f:95:5c brd ff:ff:ff:ff:ff:ff
inet 100.0.0.100/24 scope global enp0s8
valid_lft forever preferred_lft forever
The above is a VM that I configured with the same IP address
(100.0.0.100) on all interfaces. Doing a reverse lookup to identify the
unique interface associated with 100.0.0.100 does not work here. And
this is why the option host_iface is required. I understand that the
above config does not represent a standard host system, but I'm using
this to prove a point: "We can never know how users will configure
their systems". By te way, The above configuration is perfectly fine
by Linux.
The current TCP implementation for host_traddr performs a
bind()-before-connect(). This is a common construct to set the source
IP address on a TCP socket before connecting. This has no effect on how
Linux selects the interface for the connection. That's because Linux
uses the Weak End System model as described in RFC1122 [2]. On the other
hand, setting the Source IP Address has benefits and should be supported
by linux-nvme. In fact, setting the Source IP Address is a mandatory
FedGov requirement (e.g. connection to a RADIUS/TACACS+ server).
Consider the following configuration.
$ ip addr list dev enp0s8
3: enp0s8: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc ...
link/ether 08:00:27:4f:95:5c brd ff:ff:ff:ff:ff:ff
inet 192.168.56.101/24 brd 192.168.56.255 scope global enp0s8
valid_lft 426sec preferred_lft 426sec
inet 192.168.56.102/24 scope global secondary enp0s8
valid_lft forever preferred_lft forever
inet 192.168.56.103/24 scope global secondary enp0s8
valid_lft forever preferred_lft forever
inet 192.168.56.104/24 scope global secondary enp0s8
valid_lft forever preferred_lft forever
Here we can see that several addresses are associated with interface
enp0s8. By default, Linux always selects the default IP address,
192.168.56.101, as the source address when connecting over interface
enp0s8. Some users, however, want the ability to specify a different
source address (e.g., 192.168.56.102, 192.168.56.103, ...). The option
host_traddr can be used as-is to perform this function.
In conclusion, I believe that we need 2 options for TCP connections.
One that can be used to specify an interface (host-iface). And one that
can be used to set the source address (host-traddr). Users should be
allowed to use one or the other, or both, or none. Of course, the
documentation for host_traddr will need some clarification. It should
state that when used for TCP connection, this option only sets the
source address. And the documentation for host_iface should say that
this option is only available for TCP connections.
References:
[1] https://www.freedesktop.org/wiki/Software/systemd/PredictableNetworkInterfaceNames/
[2] https://tools.ietf.org/html/rfc1122
Tested both IPv4 and IPv6 connections.
Signed-off-by: Martin Belanger <martin.belanger@dell.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Signed-off-by: Christoph Hellwig <hch@lst.de>
The algorithm that was used until now for building the APST configuration
table has been found to produce entries with excessively long ITPT
(idle time prior to transition) for devices declaring relatively long
entry and exit latencies for non-operational power states. This leads
to unnecessary waste of power and, as a result, failure to pass
mandatory power consumption tests on Chromebook platforms.
The new algorithm is based on two predefined ITPT values and two
predefined latency tolerances. Based on these values, as well as on
exit and entry latencies reported by the device, the algorithm looks
for up to 2 suitable non-operational power states to use as primary
and secondary APST transition targets. The predefined values are
supplied to the nvme driver as module parameters:
- apst_primary_timeout_ms (default: 100)
- apst_secondary_timeout_ms (default: 2000)
- apst_primary_latency_tol_us (default: 15000)
- apst_secondary_latency_tol_us (default: 100000)
The algorithm echoes the approach used by Intel's and Microsoft's drivers
on Windows. The specific default parameter values are also based on those
drivers. Yet, this patch doesn't introduce the ability to dynamically
regenerate the APST table in the event of switching the power source from
AC to battery and back. Adding this functionality may be considered in the
future. In the meantime, the timeouts and tolerances reflect a compromise
between values used by Microsoft for AC and battery scenarios.
In most NVMe devices the new algorithm causes them to implement a more
aggressive power saving policy. While beneficial in most cases, this
sometimes comes at the price of a higher IO processing latency in certain
scenarios as well as at the price of a potential impact on the drive's
endurance (due to more frequent context saving when entering deep non-
operational states). So in order to provide a fallback for systems where
these regressions cannot be tolerated, the patch allows to revert to
the legacy behavior by setting either apst_primary_timeout_ms or
apst_primary_latency_tol_us parameter to 0. Eventually (and possibly after
fine tuning the default values of the module parameters) the legacy behavior
can be removed.
TESTING.
The new algorithm has been extensively tested. Initially, simulations were
used to compare APST tables generated by old and new algorithms for a wide
range of devices. After that, power consumption, performance and latencies
were measured under different workloads on devices from multiple vendors
(WD, Intel, Samsung, Hynix, Kioxia). Below is the description of the tests
and the findings.
General observations.
The effect the patch has on the APST table varies depending on the entry and
exit latencies advertised by the devices. For some devices, the effect is
negligible (e.g. Kioxia KBG40ZNS), for some significant, making the
transitions to PS3 and PS4 much quicker (e.g. WD SN530, Intel 760P), or making
the sleep deeper, PS4 rather than PS3 after a similar amount of time (e.g.
SK Hynix BC511). For some devices (e.g. Samsung PM991) the effect is mixed:
the initial transition happens after a longer idle time, but takes the device
to a lower power state.
Workflows.
In order to evaluate the patch's effect on the power consumption and latency,
7 workflows were used for each device. The workflows were designed to test
the scenarios where significant differences between the old and new behaviors
are most likely. Each workflow was tested twice: with the new and with the
old APST table generation implementation. Power consumption, performance and
latency were measured in the process. The following workflows were used:
1) Consecutive write at the maximum rate with IO depth of 2, with no pauses
2) Repeated pattern of 1000 consecutive writes of 4K packets followed by 50ms
idle time
3) Repeated pattern of 1000 consecutive writes of 4K packets followed by 150ms
idle time
4) Repeated pattern of 1000 consecutive writes of 4K packets followed by 500ms
idle time
5) Repeated pattern of 1000 consecutive writes of 4K packets followed by 1.5s
idle time
6) Repeated pattern of 1000 consecutive writes of 4K packets followed by 5s
idle time
7) Repeated pattern of a single random read of a 4K packet followed by 150ms
idle time
Power consumption
Actual power consumption measurements produced predictable results in
accordance with the APST mechanism's theory of operation.
Devices with long entry and exit latencies such as WD SN530 showed huge
improvement on scenarios 4,5 and 6 of up to 62%. Devices such as Kioxia
KBG40ZNS where the resulting APST table looks virtually identical with
both legacy and new algorithms, showed little or no change in the average power
consumption on all workflows. Devices with extra short latencies such as
Samsung PM991 showed moderate increase in power consumption of up to 18% in
worst case scenarios.
In addition, on Intel and Samsung devices a more complex impact was observed
on scenarios 3, 4 and 7. Our understanding is that due to longer stay in deep
non-operational states between the writes the devices start performing background
operations leading to an increase of power consumption. With the old APST tables
part of these operations are delayed until the scenario is over and a longer idle
period begins, but eventually this extra power is consumed anyway.
Performance.
In terms of performance measured on sustained write or read scenarios, the
effect of the patch is minimal as in this case the device doesn't enter low power
states.
Latency
As expected, in devices where the patch causes a more aggressive power saving
policy (e.g. WD SN530, Intel 760P), an increase in latency was observed in
certain scenarios. Workflow number 7, specifically designed to simulate the
worst case scenario as far as latency is concerned, indeed shows a sharp
increase in average latency (~2ms -> ~53ms on Intel 760P and 0.6 -> 10ms on
WD SN530). The latency increase on other workloads and other devices is much
milder or non-existent.
Signed-off-by: Alexey Bogoslavsky <alexey.bogoslavsky@wdc.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
nvme_init_identify and thus nvme_mpath_init can be called multiple
times and thus must not overwrite potentially initialized or in-use
fields. Split out a helper for the basic initialization when the
controller is initialized and make sure the init_identify path does
not blindly change in-use data structures.
Fixes: 0d0b660f21 ("nvme: add ANA support")
Reported-by: Martin Wilck <mwilck@suse.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Keith Busch <kbusch@kernel.org>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Hannes Reinecke <hare@suse.de>
