sch_cake targets the home router use case and is intended to squeeze the
most bandwidth and latency out of even the slowest ISP links and routers,
while presenting an API simple enough that even an ISP can configure it.
Example of use on a cable ISP uplink:
tc qdisc add dev eth0 cake bandwidth 20Mbit nat docsis ack-filter
To shape a cable download link (ifb and tc-mirred setup elided)
tc qdisc add dev ifb0 cake bandwidth 200mbit nat docsis ingress wash
CAKE is filled with:
* A hybrid Codel/Blue AQM algorithm, "Cobalt", tied to an FQ_Codel
derived Flow Queuing system, which autoconfigures based on the bandwidth.
* A novel "triple-isolate" mode (the default) which balances per-host
and per-flow FQ even through NAT.
* An deficit based shaper, that can also be used in an unlimited mode.
* 8 way set associative hashing to reduce flow collisions to a minimum.
* A reasonable interpretation of various diffserv latency/loss tradeoffs.
* Support for zeroing diffserv markings for entering and exiting traffic.
* Support for interacting well with Docsis 3.0 shaper framing.
* Extensive support for DSL framing types.
* Support for ack filtering.
* Extensive statistics for measuring, loss, ecn markings, latency
variation.
A paper describing the design of CAKE is available at
https://arxiv.org/abs/1804.07617, and will be published at the 2018 IEEE
International Symposium on Local and Metropolitan Area Networks (LANMAN).
This patch adds the base shaper and packet scheduler, while subsequent
commits add the optional (configurable) features. The full userspace API
and most data structures are included in this commit, but options not
understood in the base version will be ignored.
Various versions baking have been available as an out of tree build for
kernel versions going back to 3.10, as the embedded router world has been
running a few years behind mainline Linux. A stable version has been
generally available on lede-17.01 and later.
sch_cake replaces a combination of iptables, tc filter, htb and fq_codel
in the sqm-scripts, with sane defaults and vastly simpler configuration.
CAKE's principal author is Jonathan Morton, with contributions from
Kevin Darbyshire-Bryant, Toke Høiland-Jørgensen, Sebastian Moeller,
Ryan Mounce, Tony Ambardar, Dean Scarff, Nils Andreas Svee, Dave Täht,
and Loganaden Velvindron.
Testing from Pete Heist, Georgios Amanakis, and the many other members of
the cake@lists.bufferbloat.net mailing list.
tc -s qdisc show dev eth2
qdisc cake 8017: root refcnt 2 bandwidth 1Gbit diffserv3 triple-isolate split-gso rtt 100.0ms noatm overhead 38 mpu 84
Sent 51504294511 bytes 37724591 pkt (dropped 6, overlimits 64958695 requeues 12)
backlog 0b 0p requeues 12
memory used: 1053008b of 15140Kb
capacity estimate: 970Mbit
min/max network layer size: 28 / 1500
min/max overhead-adjusted size: 84 / 1538
average network hdr offset: 14
Bulk Best Effort Voice
thresh 62500Kbit 1Gbit 250Mbit
target 5.0ms 5.0ms 5.0ms
interval 100.0ms 100.0ms 100.0ms
pk_delay 5us 5us 6us
av_delay 3us 2us 2us
sp_delay 2us 1us 1us
backlog 0b 0b 0b
pkts 3164050 25030267 9530280
bytes 3227519915 35396974782 12879808898
way_inds 0 8 0
way_miss 21 366 25
way_cols 0 0 0
drops 5 0 1
marks 0 0 0
ack_drop 0 0 0
sp_flows 1 3 0
bk_flows 0 1 1
un_flows 0 0 0
max_len 68130 68130 68130
Tested-by: Pete Heist <peteheist@gmail.com>
Tested-by: Georgios Amanakis <gamanakis@gmail.com>
Signed-off-by: Dave Taht <dave.taht@gmail.com>
Signed-off-by: Toke Høiland-Jørgensen <toke@toke.dk>
Signed-off-by: David S. Miller <davem@davemloft.net>
The target core runs into a warning in the linux/sbitmap.h
file in some configurations:
In file included from include/target/target_core_base.h:7,
from drivers/target/target_core_fabric_lib.c:41:
include/linux/sbitmap.h:331:46: error: 'struct seq_file' declared inside parameter list will not be visible outside of this definition or declaration [-Werror]
void sbitmap_show(struct sbitmap *sb, struct seq_file *m);
^~~~~~~~
In general, headers should not depend on others being included first,
so this fixes it with a forward declaration for that struct name, but
we probably want to merge the patch through the scsi tree to help
bisection.
Fixes: 10e9cbb6b5 ("scsi: target: Convert target drivers to use sbitmap")
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Reviewed-by: Bart Van Assche <bart.vanassche@wdc.com>
Reviewed-by: Omar Sandoval <osandov@fb.com>
Reviewed-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
The TTM buffer-object interface provides ttm_bo_reference() and
ttm_bo_unref() for managing reference counts. Replacing them with
ttm_bo_get() and ttm_bo_put() aligns the API with conventions used
throughout the Linux kernel.
The implementation of ttm_bo_unref() clears the supplied pointer
to NULL. This leads to workarounds where the caller saves the
pointer's value before de-referencing the BO. ttm_bo_put() does
not clear the supplied pointer.
Signed-off-by: Thomas Zimmermann <contact@tzimmermann.org>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Enable uverbs_destroy_def_handler to be used by drivers and replace
current code to use it.
Signed-off-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Each PMU has a set of 32bit event counters. But in some
special cases, the events could be counted using counters
which are effectively 64bit wide.
e.g, Arm V8 PMUv3 has a 64 bit cycle counter which can count
only the CPU cycles. Also, the PMU can chain the event counters
to effectively count as a 64bit counter.
Add support for tracking the events that uses 64bit counters.
This only affects the periods set for each counter in the core
driver.
Cc: Will Deacon <will.deacon@arm.com>
Reviewed-by: Julien Thierry <julien.thierry@arm.com>
Acked-by: Mark Rutland <mark.rutland@arm.com>
Signed-off-by: Suzuki K Poulose <suzuki.poulose@arm.com>
Signed-off-by: Will Deacon <will.deacon@arm.com>
Convert the {read/write}_counter APIs to handle 64bit values
to enable supporting chained event counters. The backends still
use 32bit values and we pass them 32bit values only. So in effect
there are no functional changes.
Cc: Will Deacon <will.deacon@arm.com>
Acked-by: Mark Rutland <mark.rutland@arm.com>
Reviewed-by: Julien Thierry <julien.thierry@arm.com>
Signed-off-by: Suzuki K Poulose <suzuki.poulose@arm.com>
Signed-off-by: Will Deacon <will.deacon@arm.com>
Each PMU defines their max_period of the counter as the maximum
value that can be counted. Since all the PMU backends support
32bit counters by default, let us remove the redundant field.
No functional changes.
Cc: Will Deacon <will.deacon@arm.com>
Acked-by: Mark Rutland <mark.rutland@arm.com>
Reviewed-by: Julien Thierry <julien.thierry@arm.com>
Signed-off-by: Suzuki K Poulose <suzuki.poulose@arm.com>
Signed-off-by: Will Deacon <will.deacon@arm.com>
Userspace also needs to know if the port requires GRHs to properly form
the AVs it creates.
Signed-off-by: Artemy Kovalyov <artemyko@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Extend the existing grh_required flag to check when AV's are handled that
a GRH is present.
Since we don't want to do query_port during the AV checks for performance
reasons move the flag into the immutable_data.
Signed-off-by: Artemy Kovalyov <artemyko@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
The internal flag IP_BASED_GIDS was added to a field that was being used
to hold the port Info CapabilityMask without considering the effects this
will have. Since most drivers just use the value from the HW MAD it means
IP_BASED_GIDS will also become set on any HW that sets the IBA flag
IsOtherLocalChangesNoticeSupported - which is not intended.
Fix this by keeping port_cap_flags only for the IBA CapabilityMask value
and store unrelated flags externally. Move the bit definitions for this to
ib_mad.h to make it clear what is happening.
To keep the uAPI unchanged define a new set of flags in the uapi header
that are only used by ib_uverbs_query_port_resp.port_cap_flags which match
the current flags supported in rdma-core, and the values exposed by the
current kernel.
Fixes: b4a26a2728 ("IB: Report using RoCE IP based gids in port caps")
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Artemy Kovalyov <artemyko@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Right now, the DRM panel logic returns NULL when a panel pointing to
the passed OF node is not present in the list of registered panels.
Most drivers interpret this NULL value as -EPROBE_DEFER, but we are
about to modify the semantic of of_drm_find_panel() and let the
framework return -ENODEV when the device node we're pointing to has
a status property that is not equal to "okay" or "ok".
Let's first patch the of_drm_find_panel() implementation to return
ERR_PTR(-EPROBE_DEFER) instead of NULL and patch all callers to replace
the '!panel' check by an 'IS_ERR(panel)' one.
Signed-off-by: Boris Brezillon <boris.brezillon@bootlin.com>
Signed-off-by: Thierry Reding <treding@nvidia.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20180509130042.9435-2-boris.brezillon@bootlin.com
Deferred probe will currently wait forever on dependent devices to probe,
but sometimes a driver will never exist. It's also not always critical for
a driver to exist. Platforms can rely on default configuration from the
bootloader or reset defaults for things such as pinctrl and power domains.
This is often the case with initial platform support until various drivers
get enabled. There's at least 2 scenarios where deferred probe can render
a platform broken. Both involve using a DT which has more devices and
dependencies than the kernel supports. The 1st case is a driver may be
disabled in the kernel config. The 2nd case is the kernel version may
simply not have the dependent driver. This can happen if using a newer DT
(provided by firmware perhaps) with a stable kernel version. Deferred
probe issues can be difficult to debug especially if the console has
dependencies or userspace fails to boot to a shell.
There are also cases like IOMMUs where only built-in drivers are
supported, so deferring probe after initcalls is not needed. The IOMMU
subsystem implemented its own mechanism to handle this using OF_DECLARE
linker sections.
This commit adds makes ending deferred probe conditional on initcalls
being completed or a debug timeout. Subsystems or drivers may opt-in by
calling driver_deferred_probe_check_init_done() instead of
unconditionally returning -EPROBE_DEFER. They may use additional
information from DT or kernel's config to decide whether to continue to
defer probe or not.
The timeout mechanism is intended for debug purposes and WARNs loudly.
The remaining deferred probe pending list will also be dumped after the
timeout. Not that this timeout won't work for the console which needs
to be enabled before userspace starts. However, if the console's
dependencies are resolved, then the kernel log will be printed (as
opposed to no output).
Cc: Alexander Graf <agraf@suse.de>
Signed-off-by: Rob Herring <robh@kernel.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
This switches the CMA helper drivers that use its fbdev emulation over
to the generic fbdev emulation. It's the first phase of using generic
fbdev. A later phase will use DRM client callbacks for the
lastclose/hotplug/remove callbacks.
There are currently 2 fbdev init/fini functions:
- drm_fb_cma_fbdev_init/drm_fb_cma_fbdev_fini
- drm_fbdev_cma_init/drm_fbdev_cma_fini
This is because the work on generic fbdev came up during a fbdev
refactoring and thus wasn't completed. No point in completing that
refactoring when drivers will soon move to drm_fb_helper_generic_probe().
tinydrm uses drm_fb_cma_fbdev_init_with_funcs().
Cc: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Signed-off-by: Noralf Trønnes <noralf@tronnes.org>
Acked-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Link: https://patchwork.freedesktop.org/patch/msgid/20180703160354.59955-5-noralf@tronnes.org
This the beginning of an API for in-kernel clients.
First out is a way to get a framebuffer backed by a dumb buffer.
Only GEM drivers are supported.
The original idea of using an exported dma-buf was dropped because it
also creates an anonomous file descriptor which doesn't work when the
buffer is created from a kernel thread. The easy way out is to use
drm_driver.gem_prime_vmap to get the virtual address, which requires a
GEM object. This excludes the vmwgfx driver which is the only non-GEM
driver apart from the legacy ones. A solution for vmwgfx will have to be
worked out later if it wants to support the client API which it probably
will when we have a bootsplash client.
Suggested-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Noralf Trønnes <noralf@tronnes.org>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Link: https://patchwork.freedesktop.org/patch/msgid/20180703160354.59955-2-noralf@tronnes.org
All users now register platform devices using the ti-aemif driver.
Remove the handcrafted aemif API.
Signed-off-by: Bartosz Golaszewski <bgolaszewski@baylibre.com>
Signed-off-by: Sekhar Nori <nsekhar@ti.com>
Add support for RTC mode to low level suspend code. This includes
providing the rtc base address for the assembly code to configuring the
PMIC_PWR_EN line late in suspend to enter RTC+DDR mode.
Note: This patch also fold in left out space parameter for
am33xx_emif_sram_table and am43xx_emif_sram_table
Signed-off-by: Dave Gerlach <d-gerlach@ti.com>
Signed-off-by: Keerthy <j-keerthy@ti.com>
Signed-off-by: Tony Lindgren <tony@atomide.com>
Add an argument to the sleep33xx and sleep43xx code to allow us to set
flags to determine which portions of the code get called in order to use
the same code for multiple power saving modes. This patch allows us to
decide whether or not we flush and disable caches, save EMIF context,
put the memory into self refresh and disable the EMIF, and/or invoke
the wkup_m3 when entering into WFI.
Signed-off-by: Dave Gerlach <d-gerlach@ti.com>
Signed-off-by: Tero Kristo <t-kristo@ti.com>
Signed-off-by: Keerthy <j-keerthy@ti.com>
Signed-off-by: Tony Lindgren <tony@atomide.com>
The dra76x MCAN generic interconnect module has a its own
format for the bits in the control registers.
Therefore add a new module type, new regbits and new capabilities
specific to the MCAN module.
Acked-by: Rob Herring <robh@kernel.org>
CC: Tony Lindgren <tony@atomide.com>
Signed-off-by: Faiz Abbas <faiz_abbas@ti.com>
Signed-off-by: Tony Lindgren <tony@atomide.com>
More features for 4.19:
- Use core pcie functionality rather than duplicating our own for pcie
gens and lanes
- Scheduler function naming cleanups
- More documentation
- Reworked DC/Powerplay interfaces to improve power savings
- Initial stutter mode support for RV (power feature)
- Vega12 powerplay updates
- GFXOFF fixes
- Misc fixes
Signed-off-by: Dave Airlie <airlied@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20180705221447.2807-1-alexander.deucher@amd.com
Pull HID fixes from Jiri Kosina:
- spectrev1 pattern fix in hiddev from Gustavo A. R. Silva
- bounds check fix for hid-debug from Daniel Rosenberg
- regression fix for HID autobinding from Benjamin Tissoires
- removal of excessive logging from i2c-hid driver from Jason Andryuk
- fix specific to 2nd generation of Wacom Intuos devices from Jason
Gerecke
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/jikos/hid:
HID: hiddev: fix potential Spectre v1
HID: i2c-hid: Fix "incomplete report" noise
HID: wacom: Correct touch maximum XY of 2nd-gen Intuos
HID: debug: check length before copy_to_user()
HID: core: allow concurrent registration of drivers
We are not supposed to use u32 in uapi, so change the flags member of
struct sock_txtime from u32 to __u32 instead.
Fixes: 80b14dee2b ("net: Add a new socket option for a future transmit time")
Reported-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Jesus Sanchez-Palencia <jesus.sanchez-palencia@intel.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
nf_hook() can free the skb, so we need to remove it from the list before
calling, and add passed skbs to a sublist afterwards.
Fixes: 17266ee939 ("net: ipv4: listified version of ip_rcv")
Reported-by: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: Edward Cree <ecree@solarflare.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Pablo Neira Ayuso says:
====================
Netfilter fixes for net
The following patchset contains Netfilter fixes for your net tree:
1) Missing module autoloadfor icmp and icmpv6 x_tables matches,
from Florian Westphal.
2) Possible non-linear access to TCP header from tproxy, from
Mate Eckl.
3) Do not allow rbtree to be used for single elements, this patch
moves all set backend into one single module since such thing
can only happen if hashtable module is explicitly blacklisted,
which should not ever be done.
4) Reject error and standard targets from nft_compat for sanity
reasons, they are never used from there.
5) Don't crash on double hashsize module parameter, from Andrey
Ryabinin.
6) Drop dst on skb before placing it in the fragmentation
reassembly queue, from Florian Westphal.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
For most of these calls we can just pass NULL through to the fallback
function as the sb_dev. The only cases where we cannot are the cases where
we might be dealing with either an upper device or a driver that would
have configured things to support an sb_dev itself.
The only driver that has any significant change in this patch set should be
ixgbe as we can drop the redundant functionality that existed in both the
ndo_select_queue function and the fallback function that was passed through
to us.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
This patch makes it so that instead of passing a void pointer as the
accel_priv we instead pass a net_device pointer as sb_dev. Making this
change allows us to pass the subordinate device through to the fallback
function eventually so that we can keep the actual code in the
ndo_select_queue call as focused on possible on the exception cases.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
This patch adds a generic version of the ndo_select_queue functions for
either returning 0 or selecting a queue based on the processor ID. This is
generally meant to just reduce the number of functions we have to change
in the future when we have to deal with ndo_select_queue changes.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
This change makes it so that we can support the concept of subordinate
device traffic classes to the core networking code. In doing this we can
start pulling out the driver specific bits needed to support selecting a
queue based on an upper device.
The solution at is currently stands is only partially implemented. I have
the start of some XPS bits in here, but I would still need to allow for
configuration of the XPS maps on the queues reserved for the subordinate
devices. For now I am using the reference to the sb_dev XPS map as just a
way to skip the lookup of the lower device XPS map for now as that would
result in the wrong queue being picked.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
This patch is meant to provide the basic tools needed to allow us to create
subordinate device traffic classes. The general idea here is to allow
subdividing the queues of a device into queue groups accessible through an
upper device such as a macvlan.
The idea here is to enforce the idea that an upper device has to be a
single queue device, ideally with IFF_NO_QUQUE set. With that being the
case we can pretty much guarantee that the tc_to_txq mappings and XPS maps
for the upper device are unused. As such we could reuse those in order to
support subdividing the lower device and distributing those queues between
the subordinate devices.
In order to distinguish between a regular set of traffic classes and if a
device is carrying subordinate traffic classes I changed num_tc from a u8
to a s16 value and use the negative values to represent the subordinate
pool values. So starting at -1 and running to -32768 we can encode those as
pool values, and the existing values of 0 to 15 can be maintained.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
'cmd_pause' DMA channel capability means that respective DMA engine
supports both pausing and resuming given DMA channel. However, in some
cases it is important to know if DMA channel can be paused without the
need to resume it. This is a typical requirement for proper residue
reading on transfer timeout in UART drivers. There are also some DMA
engines with limited hardware, which doesn't really support resuming.
Reporting pause and resume capabilities separately allows UART drivers to
properly check for the really required capabilities and operate in DMA
mode also in systems with limited DMA hardware. On the other hand drivers,
which rely on full channel suspend/resume support, should now check for
both 'pause' and 'resume' features.
Existing clients of dma_get_slave_caps() have been checked and the only
driver which rely on proper channel resuming is soc-generic-dmaengine-pcm
driver, which has been updated to check the newly added capability.
Existing 'cmd_pause' now only indicates that DMA engine support pausing
given DMA channel.
Signed-off-by: Marek Szyprowski <m.szyprowski@samsung.com>
Acked-by: Mark Brown <broonie@kernel.org>
Signed-off-by: Vinod Koul <vkoul@kernel.org>
The only purpose for this structure was to hold the ib_uobject_file
pointer, but now that is part of the standard ib_uobject the structure
no longer makes any sense, so get rid of it.
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
The ucontext isn't needed any more, just pass the uverbs_file directly.
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
The correct handle to refer to the idr/etc is ib_uverbs_file, revise all
the core APIs to use this instead. The user API are left as wrappers
that automatically convert a ucontext to a ufile for now.
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
The IDR is part of the ib_ufile so all the machinery to lock it, handle
closing and disassociation rightly belongs to the ufile not the ucontext.
This changes the lifetime of that data to match the lifetime of the file
descriptor which is always strictly longer than the lifetime of the
ucontext.
We need the entire locking machinery to continue to exist after ucontext
destruction to allow us to return the destroy data after a device has been
disassociated.
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
This consolidates a bunch of repeated code patterns into a helper.
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Current IO controllers for the block layer are less than ideal for our
use case. The io.max controller is great at hard limiting, but it is
not work conserving. This patch introduces io.latency. You provide a
latency target for your group and we monitor the io in short windows to
make sure we are not exceeding those latency targets. This makes use of
the rq-qos infrastructure and works much like the wbt stuff. There are
a few differences from wbt
- It's bio based, so the latency covers the whole block layer in addition to
the actual io.
- We will throttle all IO types that comes in here if we need to.
- We use the mean latency over the 100ms window. This is because writes can
be particularly fast, which could give us a false sense of the impact of
other workloads on our protected workload.
- By default there's no throttling, we set the queue_depth to INT_MAX so that
we can have as many outstanding bio's as we're allowed to. Only at
throttle time do we pay attention to the actual queue depth.
- We backcharge cgroups for root cg issued IO and induce artificial
delays in order to deal with cases like metadata only or swap heavy
workloads.
In testing this has worked out relatively well. Protected workloads
will throttle noisy workloads down to 1 io at time if they are doing
normal IO on their own, or induce up to a 1 second delay per syscall if
they are doing a lot of root issued IO (metadata/swap IO).
Our testing has revolved mostly around our production web servers where
we have hhvm (the web server application) in a protected group and
everything else in another group. We see slightly higher requests per
second (RPS) on the test tier vs the control tier, and much more stable
RPS across all machines in the test tier vs the control tier.
Another test we run is a slow memory allocator in the unprotected group.
Before this would eventually push us into swap and cause the whole box
to die and not recover at all. With these patches we see slight RPS
drops (usually 10-15%) before the memory consumer is properly killed and
things recover within seconds.
Signed-off-by: Josef Bacik <jbacik@fb.com>
Acked-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
