Commit ab9da627906a ("drm/i915: make context status notifier head be
per engine") gives us a chance to inspect every single request. Then
we can eliminate unnecessary mmio switching for same vGPU. We only
need mmio switching for different VMs (including host).
This patch introduced a new general API intel_gvt_switch_mmio() to
replace the old intel_gvt_load/restore_render_mmio(). This function
can be further optimized for vGPU to vGPU switching.
To support individual ring switch, we track the owner who occupy
each ring. When another VM or host request a ring we do the mmio
context switching. Otherwise no need to switch the ring.
This optimization is very useful if only one guest has plenty of
workloads and the host is mostly idle. The best case is no mmio
switching will happen.
v2:
o fix missing ring switch issue. (chuanxiao)
o support individual ring switch.
Signed-off-by: Changbin Du <changbin.du@intel.com>
Reviewed-by: Chuanxiao Dong <chuanxiao.dong@intel.com>
Signed-off-by: Zhenyu Wang <zhenyuw@linux.intel.com>
The function intel_vgpu_submit_execlist could be more simpler. It
actually does:
1) validate the submission. The first context must be valid,
and all two must be privilege_access.
2) submit valid contexts. The first one need emulate schedule_in.
We do not need a bitmap, valid desc copy valid_desc. Local variable
emulate_schedule_in also can be optimized out.
v2: dump desc content in err msg (Zhi Wang)
Signed-off-by: Changbin Du <changbin.du@intel.com>
Reviewed-by: Zhi Wang <zhi.a.wang@intel.com>
Signed-off-by: Zhenyu Wang <zhenyuw@linux.intel.com>
The gvt:gvt_command trace involve unnecessary overhead even this trace is
not enabled. We need improve it.
The kernel trace infrastructure provide a full api to define a trace event.
We should leverage them if possible. And one important thing is that a trace
point should store raw data but not format string.
This patch include two part work:
1) Refactor the gvt_command trace definition, including:
o only store raw trace data.
o use __dynamic_array() to declare a variable size buffer.
o use __print_array() to format raw cmd data.
o rename vm_id as vgpu_id.
2) Improve the trace invoking, including:
o remove the cycles calculation for handler. We can get this data
by any perf tool.
o do not make a backup for raw cmd data which just doesn't make sense.
With this patch, this trace has no overhead if it is not enabled. And we are
trace style now.
The final output example:
gvt workload 0-211 [000] ...1 120.555964: gvt_command: vgpu1 ring 0: buf_type 0, ip_gma e161e880, raw cmd {0x4000000}
gvt workload 0-211 [000] ...1 120.556014: gvt_command: vgpu1 ring 0: buf_type 0, ip_gma e161e884, raw cmd {0x7a000004,0x1004000,0xe1511018,0x0,0x7d,0x0}
gvt workload 0-211 [000] ...1 120.556062: gvt_command: vgpu1 ring 0: buf_type 0, ip_gma e161e89c, raw cmd {0x7a000004,0x140000,0x0,0x0,0x0,0x0}
gvt workload 0-211 [000] ...1 120.556110: gvt_command: vgpu1 ring 0: buf_type 0, ip_gma e161e8b4, raw cmd {0x10400002,0xe1511018,0x0,0x7d}
Signed-off-by: Changbin Du <changbin.du@intel.com>
Signed-off-by: Zhenyu Wang <zhenyuw@linux.intel.com>
Currently the timer is armed for 1ms after the first use and is killed
immediately, dropping the forcewake as early as possible. However, for
very frequent operations the forcewake dance has a large impact on
latency and keeping the timer alive until we are idle is preferred. To
achieve this, if we call intel_uncore_forcewake_get whilst the timer is
alive (repeated use), then set a flag to restart the timer on expiry
rather than drop the forcewake usage count. The timer is racy, the
consequence of the race is to expire the timer earlier than is now
desired but does not impact on correct behaviour. The offset the race
slightly, we set the active flag again on intel_uncore_forcewake_put.
The effect should be to reduce the jitter of reacquiring the fw every
1ms on a busy system. However, the cost is to keep the timer alive for
an extra 1ms on a nearly idle system. We chose to incur the jitter
previously to keep the timer off for as much as possible.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Mika Kuoppala <mika.kuoppala@intel.com>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20170526132209.14640-1-chris@chris-wilson.co.uk
We're currently deleting the GuC logs if the FW fails to load, but those
are still useful to understand why the loading failed. Keeping the
object around allows us to access them after driver load is completed.
v2: keep the object around instead of using kernel memory (chris)
don't store the logs in the gpu_error struct (Chris)
add a check on guc_log_level to avoid snapshotting empty logs
v3: use separate debugfs for error log (Chris)
v4: rebased
v5: clean up obj selection, move err_load inside guc_log, move err_load
cleanup, rename functions (Michal)
v6: move obj back to intel_guc, move functions to intel_uc.c, don't
clear obj on new GuC load, free object only if enable_guc_loading
is set (Michal)
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Oscar Mateo <oscar.mateo@intel.com>
Cc: Michal Wajdeczko <michal.wajdeczko@intel.com>
Signed-off-by: Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/1495475428-19295-1-git-send-email-daniele.ceraolospurio@intel.com
Reviewed-by: Michal Wajdeczko <michal.wajdeczko@intel.com>
Tested-by: Michel Thierry <michel.thierry@intel.com>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
BXT has a H/W issue with IOMMU which can lead to system hangs when
Aperture accesses are queued within the GAM behind GTT Accesses.
This patch avoids the condition by wrapping all GTT updates in stop_machine
and using a flushing read prior to restarting the machine.
The stop_machine guarantees no new Aperture accesses can begin while
the PTE writes are being emmitted. The flushing read ensures that
any following Aperture accesses cannot begin until the PTE writes
have been cleared out of the GAM's fifo.
Only FOLLOWING Aperture accesses need to be separated from in flight
PTE updates. PTE Writes may follow tightly behind already in flight
Aperture accesses, so no flushing read is required at the start of
a PTE update sequence.
This issue was reproduced by running
igt/gem_readwrite and
igt/gem_render_copy
simultaneously from different processes, each in a tight loop,
with INTEL_IOMMU enabled.
This patch was originally published as:
drm/i915: Serialize GTT Updates on BXT
v2: Move bxt/iommu detection into static function
Remove #ifdef CONFIG_INTEL_IOMMU protection
Make function names more reflective of purpose
Move flushing read into static function
v3: Tidy up for checkpatch.pl
Testcase: igt/gem_concurrent_blit
Signed-off-by: Jon Bloomfield <jon.bloomfield@intel.com>
Cc: John Harrison <john.C.Harrison@intel.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Daniel Vetter <daniel.vetter@intel.com>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/1495641251-30022-1-git-send-email-jon.bloomfield@intel.com
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
ELK seems to very picky about the preconditions to reset.
Evidence on Eaglelake (8086:2e12 (rev 03)) shows that it does
not like if reset occurs when there is active ring.
Ville found out that there is workaround with name
'WaMediaResetMainRingCleanup' which suggests that we need to
cleanup rings before resetting. It is unclear what cleanup
exactly means but evidence shows that stopping the ring
does have an effect on reset reliability. This patch makes
reset successful on hangs induced by chained batches (the igt ones).
Note that if the hang is inside a shader, it is possible
that our attempts to stop the ring achieves anything.
v2: zero ctl,head,tail also. bug ref. use driver debugs (Chris)
v3: specify platform on testcases, comment tidyup (Chris)
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=100942
Testcase: igt/gem_busy/*-hang #elk
Testcase: igt/gem_ringfill/hang-* #elk
Suggested-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tomi Sarvela <tomi.p.sarvela@intel.com>
Signed-off-by: Mika Kuoppala <mika.kuoppala@intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Link: http://patchwork.freedesktop.org/patch/msgid/20170519091340.21439-1-mika.kuoppala@intel.com
This patch make changes to use linetime latency if allocated
DDB size during plane watermark calculation is not available.
linetime is the time, display engine takes to fetch one line worth of
pixels with given pixel clock rate.
This is required to implement new DDB allocation algorithm.
In New Algorithm DDB is allocated based on WM values, because of which
number of DDB blocks will not be available during WM calculation,
So this "linetime latency" is suggested by SV/HW team to be used during
switch-case for WM blocks selection.
linetime latency us = pipe horizontal total pixels/adjusted pixel rate MHz
Changes since v1:
- Rebase on top of Paulo's patch series
Changes since v2:
- Fix if-else condition (pointed by Maarten)
Changes since v3:
- Use common function for timetime_us calculation (Paulo)
- rebase on drm-tip
Changes since v4:
- Use consistent name for fixed_point operation
Changes since v5:
- Improve commit message
- rename skl_get_linetime_us to intel_get_linetime_us
- fix watermark result selection (Matt)
Signed-off-by: "Mahesh Kumar" <mahesh1.kumar@intel.com>
Reviewed-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
Reviewed-by: Matt Roper <matthew.d.roper@intel.com>
Signed-off-by: Matt Roper <matthew.d.roper@intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20170517115831.13830-11-mahesh1.kumar@intel.com
DDB minimum requirement of crtc configuration (cumulative of all the
enabled planes in crtc) may exceed the allocated DDB for crtc/pipe.
This patch make changes to fail the flip/ioctl if minimum requirement
for pipe exceeds the total ddb allocated to the pipe.
Previously it succeeded but making alloc_size a negative value. Which
will make subsequent calculations for plane ddb allocation bogus & may
lead to screen corruption or system hang.
Changes from V1:
- Improve commit message as per Ander's comment
- Remove extra parentheses (Ander)
Signed-off-by: Mahesh Kumar <mahesh1.kumar@intel.com>
Reviewed-by: Matt Roper <matthew.d.roper@intel.com>
Reviewed-by: Ander Conselvan de Oliveira <conselvan2@gmail.com>
Reviewed-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
Signed-off-by: Matt Roper <matthew.d.roper@intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20170517115831.13830-8-mahesh1.kumar@intel.com
This patch adds few wrapper to perform fixed_point_16_16 operations
mul_round_up_u32_fixed16 : Multiplies u32 and fixed_16_16_t variables
& returns u32 result with rounding-up.
mul_fixed16 : Multiplies two fixed_16_16_t variable & returns fixed_16_16
div_round_up_fixed16 : Perform division operation on fixed_16_16_t
variables & return u32 result with round-off
div_round_up_u32_fixed16 : devide uint32_t variable by fixed_16_16 variable
and round_up the result to uint32_t.
These wrappers will be used by later patches in the series.
Changes from V1:
- Rename wrapper as per Matt's comment
Changes from V2:
- Fix indentation
Signed-off-by: Mahesh Kumar <mahesh1.kumar@intel.com>
Reviewed-by: Matt Roper <matthew.d.roper@intel.com>
Reviewed-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
Signed-off-by: Matt Roper <matthew.d.roper@intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20170517115831.13830-3-mahesh1.kumar@intel.com
The i915_priolist are allocated within an atomic context on a path where
we wish to minimise latency. If we use a dedicated kmem_cache, we have
the advantage of a local freelist from which to service new requests
that should keep the latency impact of an allocation small. Though
currently we expect the majority of requests to be at default priority
(and so hit the preallocate priolist), once userspace starts using
priorities they are likely to use many fine grained policies improving
the utilisation of a private slab.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20170517121007.27224-9-chris@chris-wilson.co.uk
All the requests at the same priority are executed in FIFO order. They
do not need to be stored in the rbtree themselves, as they are a simple
list within a level. If we move the requests at one priority into a list,
we can then reduce the rbtree to the set of priorities. This should keep
the height of the rbtree small, as the number of active priorities can not
exceed the number of active requests and should be typically only a few.
Currently, we have ~2k possible different priority levels, that may
increase to allow even more fine grained selection. Allocating those in
advance seems a waste (and may be impossible), so we opt for allocating
upon first use, and freeing after its requests are depleted. To avoid
the possibility of an allocation failure causing us to lose a request,
we preallocate the default priority (0) and bump any request to that
priority if we fail to allocate it the appropriate plist. Having a
request (that is ready to run, so not leading to corruption) execute
out-of-order is better than leaking the request (and its dependency
tree) entirely.
There should be a benefit to reducing execlists_dequeue() to principally
using a simple list (and reducing the frequency of both rbtree iteration
and balancing on erase) but for typical workloads, request coalescing
should be small enough that we don't notice any change. The main gain is
from improving PI calls to schedule, and the explicit list within a
level should make request unwinding simpler (we just need to insert at
the head of the list rather than the tail and not have to make the
rbtree search more complicated).
v2: Avoid use-after-free when deleting a depleted priolist
v3: Michał found the solution to handling the allocation failure
gracefully. If we disable all priority scheduling following the
allocation failure, those requests will be executed in fifo and we will
ensure that this request and its dependencies are in strict fifo (even
when it doesn't realise it is only a single list). Normal scheduling is
restored once we know the device is idle, until the next failure!
Suggested-by: Michał Wajdeczko <michal.wajdeczko@intel.com>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Michał Winiarski <michal.winiarski@intel.com>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Reviewed-by: Michał Winiarski <michal.winiarski@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20170517121007.27224-8-chris@chris-wilson.co.uk
add/remove: 1/1 grow/shrink: 5/4 up/down: 391/-578 (-187)
function old new delta
execlists_submit_ports 262 471 +209
port_assign.isra - 136 +136
capture 6344 6359 +15
reset_common_ring 438 452 +14
execlists_submit_request 228 238 +10
gen8_init_common_ring 334 341 +7
intel_engine_is_idle 106 105 -1
i915_engine_info 2314 2290 -24
__i915_gem_set_wedged_BKL 485 411 -74
intel_lrc_irq_handler 1789 1604 -185
execlists_update_context 294 - -294
The most important change there is the improve to the
intel_lrc_irq_handler and excclist_submit_ports (net improvement since
execlists_update_context is now inlined).
v2: Use the port_api() for guc as well (even though currently we do not
pack any counters in there, yet) and hide all port->request_count inside
the helpers.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Mika Kuoppala <mika.kuoppala@intel.com>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20170517121007.27224-5-chris@chris-wilson.co.uk
ptr_unpack_bits() is a function-like macro, as such it is meant to be
replaceable by a function. In this case, we should be passing in the
out-param as a pointer.
Bizarrely this does affect code generation:
function old new delta
i915_gem_object_pin_map 409 389 -20
An improvement(?) in this case, but one can't help wonder what
strict-aliasing optimisations we are preventing.
The generated code looks identical in using ptr_unpack_bits (no extra
motions to stack, the pointer and bits appear to be kept in registers),
the difference appears to be code ordering and with a reorder it is able
to use smaller forward jumps.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20170517121007.27224-3-chris@chris-wilson.co.uk
