Since the introduction of 'soft-rc6', we aim to park the device quickly
and that results in frequent idling of the whole device. Currently upon
idling we free the batch buffer pool, and so this renders the cache
ineffective for many workloads. If we want to have an effective cache of
recently allocated buffers available for reuse, we need to decouple that
cache from the engine powermanagement and make it timer based. As there
is no reason then to keep it within the engine (where it once made
retirement order easier to track), we can move it up the hierarchy to the
owner of the memory allocations.
v2: Hook up to debugfs/drop_caches to clear the cache on demand.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20200430111819.10262-2-chris@chris-wilson.co.uk
We treat parking as a manual RPS timeout event, and downclock the GPU
for the next unpark and batch execution. However, having restored the
aggressive downclocking and observed that we have very light workloads
whose only interaction is through the manual parking events, carry over
the aggressive downclocking to the fake RPS events.
References: 21abf0bf16 ("drm/i915/gt: Treat idling as a RPS downclock event")
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Andi Shyti <andi.shyti@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20200429205446.3259-5-chris@chris-wilson.co.uk
As with the realisation for soft-rc6, we respond to idling the engines
within microseconds, far faster than the response times for HW RC6 and
RPS. Furthermore, our fast parking upon idle, prevents HW RPS from
running for many desktop workloads, as the RPS evaluation intervals are
on the order of tens of milliseconds, but the typical workload is just a
couple of milliseconds, but yet we still need to determine the best
frequency for user latency versus power.
Recognising that the HW evaluation intervals are a poor fit, and that
they were deprecated [in bspec at least] from gen10, start to wean
ourselves off them and replace the EI with a timer and our accurate
busy-stats. The principle benefit of manually evaluating RPS intervals
is that we can be more responsive for better performance and powersaving
for both spiky workloads and steady-state.
Closes: https://gitlab.freedesktop.org/drm/intel/-/issues/1698
Fixes: 98479ada42 ("drm/i915/gt: Treat idling as a RPS downclock event")
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Cc: Andi Shyti <andi.shyti@intel.com>
Reviewed-by: Andi Shyti <andi.shyti@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20200429205446.3259-4-chris@chris-wilson.co.uk
In the near future, we will utilize the busy-stats on each engine to
approximate the C0 cycles of each, and use that as an input to a manual
RPS mechanism. That entails having busy-stats always enabled and so we
can remove the enable/disable routines and simplify the pmu setup. As a
consequence of always having the stats enabled, we can also show the
current active time via sysfs/engine/xcs/active_time_ns.
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: https://patchwork.freedesktop.org/patch/msgid/20200429205446.3259-1-chris@chris-wilson.co.uk
We need to keep the default context state around to instantiate new
contexts (aka golden rendercontext), and we also keep it pinned while
the engine is active so that we can quickly reset a hanging context.
However, the default contexts are large enough to merit keeping in
swappable memory as opposed to kernel memory, so we store them inside
shmemfs. Currently, we use the normal GEM objects to create the default
context image, but we can throw away all but the shmemfs file.
This greatly simplifies the tricky power management code which wants to
run underneath the normal GT locking, and we definitely do not want to
use any high level objects that may appear to recurse back into the GT.
Though perhaps the primary advantage of the complex GEM object is that
we aggressively cache the mapping, but here we are recreating the
vm_area everytime time we unpark. At the worst, we add a lightweight
cache, but first find a microbenchmark that is impacted.
Having started to create some utility functions to make working with
shmemfs objects easier, we can start putting them to wider use, where
GEM objects are overkill, such as storing persistent error state.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Matthew Auld <matthew.auld@intel.com>
Cc: Ramalingam C <ramalingam.c@intel.com>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Matthew Auld <matthew.auld@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20200429172429.6054-1-chris@chris-wilson.co.uk
When building with clang + -Wuninitialized:
drivers/gpu/drm/i915/gt/debugfs_gt_pm.c:407:7: warning: variable
'rpcurupei' is uninitialized when used here [-Wuninitialized]
rpcurupei,
^~~~~~~~~
drivers/gpu/drm/i915/gt/debugfs_gt_pm.c:304:16: note: initialize the
variable 'rpcurupei' to silence this warning
u32 rpcurupei, rpcurup, rpprevup;
^
= 0
1 warning generated.
rpupei is assigned twice; based on the second argument to
intel_uncore_read, it seems this one should have been assigned to
rpcurupei.
Fixes: 9c878557b1 ("drm/i915/gt: Use the RPM config register to determine clk frequencies")
Link: https://github.com/ClangBuiltLinux/linux/issues/1016
Signed-off-by: Nathan Chancellor <natechancellor@gmail.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Link: https://patchwork.freedesktop.org/patch/msgid/20200429030051.920203-1-natechancellor@gmail.com
The bspec is confusing on the nature of the upper 32bits of the LRC
descriptor. Once upon a time, it said that it uses the upper 32b to
decide if it should perform a lite-restore, and so we must ensure that
each unique context submitted to HW is given a unique CCID [for the
duration of it being on the HW]. Currently, this is achieved by using
a small circular tag, and assigning every context submitted to HW a
new id. However, this tag is being cleared on repinning an inflight
context such that we end up re-using the 0 tag for multiple contexts.
To avoid accidentally clearing the CCID in the upper 32bits of the LRC
descriptor, split the descriptor into two dwords so we can update the
GGTT address separately from the CCID.
Closes: https://gitlab.freedesktop.org/drm/intel/-/issues/1796
Fixes: 2935ed5339 ("drm/i915: Remove logical HW ID")
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Cc: <stable@vger.kernel.org> # v5.5+
Reviewed-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20200428184751.11257-1-chris@chris-wilson.co.uk
While the ggtt vma are protected by their object lifetime, the list
continues until it hits a non-ggtt vma, and that vma is not protected
and may be freed as we inspect it. Hence, we require the obj->vma.lock
to protect the list as we iterate.
An example of forgetting to hold the obj->vma.lock is
[1642834.464973] general protection fault, probably for non-canonical address 0xdead000000000122: 0000 [#1] SMP PTI
[1642834.464977] CPU: 3 PID: 1954 Comm: Xorg Not tainted 5.6.0-300.fc32.x86_64 #1
[1642834.464979] Hardware name: LENOVO 20ARS25701/20ARS25701, BIOS GJET94WW (2.44 ) 09/14/2017
[1642834.465021] RIP: 0010:i915_gem_object_set_tiling+0x2c0/0x3e0 [i915]
[1642834.465024] Code: 8b 84 24 18 01 00 00 f6 c4 80 74 59 49 8b 94 24 a0 00 00 00 49 8b 84 24 e0 00 00 00 49 8b 74 24 10 48 8b 92 30 01 00 00 89 c7 <80> ba 0a 06 00 00 03 0f 87 86 00 00 00 ba 00 00 08 00 b9 00 00 10
[1642834.465025] RSP: 0018:ffffa98780c77d60 EFLAGS: 00010282
[1642834.465028] RAX: ffff8d232bfb2578 RBX: 0000000000000002 RCX: ffff8d25873a0000
[1642834.465029] RDX: dead000000000122 RSI: fffff0af8ac6e408 RDI: 000000002bfb2578
[1642834.465030] RBP: ffff8d25873a0000 R08: ffff8d252bfb5638 R09: 0000000000000000
[1642834.465031] R10: 0000000000000000 R11: ffff8d252bfb5640 R12: ffffa987801cb8f8
[1642834.465032] R13: 0000000000001000 R14: ffff8d233e972e50 R15: ffff8d233e972d00
[1642834.465034] FS: 00007f6a3d327f00(0000) GS:ffff8d25926c0000(0000) knlGS:0000000000000000
[1642834.465036] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[1642834.465037] CR2: 00007f6a2064d000 CR3: 00000002fb57c001 CR4: 00000000001606e0
[1642834.465038] Call Trace:
[1642834.465083] i915_gem_set_tiling_ioctl+0x122/0x230 [i915]
[1642834.465121] ? i915_gem_object_set_tiling+0x3e0/0x3e0 [i915]
[1642834.465151] drm_ioctl_kernel+0x86/0xd0 [drm]
[1642834.465156] ? avc_has_perm+0x3b/0x160
[1642834.465178] drm_ioctl+0x206/0x390 [drm]
[1642834.465216] ? i915_gem_object_set_tiling+0x3e0/0x3e0 [i915]
[1642834.465221] ? selinux_file_ioctl+0x122/0x1c0
[1642834.465226] ? __do_munmap+0x24b/0x4d0
[1642834.465231] ksys_ioctl+0x82/0xc0
[1642834.465235] __x64_sys_ioctl+0x16/0x20
[1642834.465238] do_syscall_64+0x5b/0xf0
[1642834.465243] entry_SYSCALL_64_after_hwframe+0x44/0xa9
[1642834.465245] RIP: 0033:0x7f6a3d7b047b
[1642834.465247] Code: 0f 1e fa 48 8b 05 1d aa 0c 00 64 c7 00 26 00 00 00 48 c7 c0 ff ff ff ff c3 66 0f 1f 44 00 00 f3 0f 1e fa b8 10 00 00 00 0f 05 <48> 3d 01 f0 ff ff 73 01 c3 48 8b 0d ed a9 0c 00 f7 d8 64 89 01 48
[1642834.465249] RSP: 002b:00007ffe71adba28 EFLAGS: 00000246 ORIG_RAX: 0000000000000010
[1642834.465251] RAX: ffffffffffffffda RBX: 000055f99048fa40 RCX: 00007f6a3d7b047b
[1642834.465253] RDX: 00007ffe71adba30 RSI: 00000000c0106461 RDI: 000000000000000e
[1642834.465254] RBP: 0000000000000002 R08: 000055f98f3f1798 R09: 0000000000000002
[1642834.465255] R10: 0000000000001000 R11: 0000000000000246 R12: 0000000000000080
[1642834.465257] R13: 000055f98f3f1690 R14: 00000000c0106461 R15: 00007ffe71adba30
Now to take the spinlock during the list iteration, we need to break it
down into two phases. In the first phase under the lock, we cannot sleep
and so must defer the actual work to a second list, protected by the
ggtt->mutex.
We also need to hold the spinlock during creation of a new vma to
serialise with updates of the tiling on the object.
Reported-by: Dave Airlie <airlied@redhat.com>
Fixes: 2850748ef8 ("drm/i915: Pull i915_vma_pin under the vm->mutex")
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: Dave Airlie <airlied@redhat.com>
Cc: <stable@vger.kernel.org> # v5.5+
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20200422072805.17340-1-chris@chris-wilson.co.uk
(cherry picked from commit cb593e5d2b)
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
igt_ppgtt_pin_update() invokes i915_gem_context_get_vm_rcu(), which
returns a reference of the i915_address_space object to "vm" with
increased refcount.
When igt_ppgtt_pin_update() returns, "vm" becomes invalid, so the
refcount should be decreased to keep refcount balanced.
The reference counting issue happens in two exception handling paths of
igt_ppgtt_pin_update(). When i915_gem_object_create_internal() returns
IS_ERR, the refcnt increased by i915_gem_context_get_vm_rcu() is not
decreased, causing a refcnt leak.
Fix this issue by jumping to "out_vm" label when
i915_gem_object_create_internal() returns IS_ERR.
Fixes: a4e7ccdac3 ("drm/i915: Move context management under GEM")
Signed-off-by: Xiyu Yang <xiyuyang19@fudan.edu.cn>
Signed-off-by: Xin Tan <tanxin.ctf@gmail.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Link: https://patchwork.freedesktop.org/patch/msgid/1587361342-83494-1-git-send-email-xiyuyang19@fudan.edu.cn
(cherry picked from commit e07c7606a0)
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
We see that if the HW doesn't actually sleep, the HW may eat the poison
we set in its write-only HWSP during sanitize:
intel_gt_resume.part.8: 0000:00:02.0
__gt_unpark: 0000:00:02.0
gt_sanitize: 0000:00:02.0 force:yes
process_csb: 0000:00:02.0 vcs0: cs-irq head=5, tail=90
process_csb: 0000:00:02.0 vcs0: csb[0]: status=0x5a5a5a5a:0x5a5a5a5a
assert_pending_valid: Nothing pending for promotion!
The CS TAIL pointer should have been reset by reset_csb_pointers(), so
in this case it is likely that we have read back from the CPU cache and
so we must clflush our control over that page. In doing so, push the
sanitisation to the start of the GT sequence so that our poisoning is
assuredly before we start talking to the HW.
References: https://gitlab.freedesktop.org/drm/intel/-/issues/1794
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20200427084000.10999-1-chris@chris-wilson.co.uk
We evaluate *active, which is a pointer into execlists->inflight[]
during dequeue to decide how long a preempt-timeout we need to apply.
However, as soon as we do the submit_ports, the HW may send its ACK
interrupt causing us to promote execlists->pending[] tp
execlists->inflight[], overwriting the value of *active. We know *active
is only stable until we submit (as we only submit when there is no
pending promotion).
[ 16.102328] BUG: KCSAN: data-race in execlists_dequeue+0x1449/0x1600 [i915]
[ 16.102356]
[ 16.102375] race at unknown origin, with read to 0xffff8881e9500488 of 8 bytes by task 429 on cpu 1:
[ 16.102780] execlists_dequeue+0x1449/0x1600 [i915]
[ 16.103160] __execlists_submission_tasklet+0x48/0x60 [i915]
[ 16.103540] execlists_submit_request+0x38e/0x3c0 [i915]
[ 16.103940] submit_notify+0x8f/0xc0 [i915]
[ 16.104308] __i915_sw_fence_complete+0x61/0x420 [i915]
[ 16.104683] i915_sw_fence_complete+0x58/0x80 [i915]
[ 16.105054] i915_sw_fence_commit+0x16/0x20 [i915]
[ 16.105457] __i915_request_queue+0x60/0x70 [i915]
[ 16.105843] i915_gem_do_execbuffer+0x2d6b/0x4230 [i915]
[ 16.106227] i915_gem_execbuffer2_ioctl+0x2b0/0x580 [i915]
[ 16.106257] drm_ioctl_kernel+0xe9/0x130
[ 16.106279] drm_ioctl+0x27d/0x45e
[ 16.106311] ksys_ioctl+0x89/0xb0
[ 16.106336] __x64_sys_ioctl+0x42/0x60
[ 16.106370] do_syscall_64+0x6e/0x2c0
[ 16.106397] entry_SYSCALL_64_after_hwframe+0x44/0xa9
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: https://patchwork.freedesktop.org/patch/msgid/20200426094231.21995-1-chris@chris-wilson.co.uk
Indirect ctx batchbuffers are a hw feature of which
batch can be run, by hardware, during context restoration stage.
Driver can setup a batchbuffer and also an offset into the
context image. When context image is marshalled from
memory to registers, and when the offset from the start of
context register state is equal of what driver pre-determined,
batch will run. So one can manipulate context restoration
process at cacheline granularity, given some limitations,
as you need to have rudimentaries in place before you can
run a batch.
Add selftest which will write the ring start register
to a canary spot. This will test that hardware will run a
batchbuffer for the context in question.
v2: request wait fix, naming (Chris)
v3: test order (Chris)
v4: rebase
Signed-off-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Acked-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Link: https://patchwork.freedesktop.org/patch/msgid/20200424214841.28076-3-mika.kuoppala@linux.intel.com
Restoration of a previous timestamp can collide
with updating the timestamp, causing a value corruption.
Combat this issue by using indirect ctx bb to
modify the context image during restoring process.
We can preload value into scratch register. From which
we then do the actual write with LRR. LRR is faster and
thus less error prone as probability of race drops.
v2: tidying (Chris)
v3: lrr for all engines
v4: grp
v5: reg bit
v6: wa_bb_offset, virtual engines (Chris)
References: HSDES#16010904313
Testcase: igt/i915_selftest/gt_lrc
Suggested-by: Joseph Koston <joseph.koston@intel.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Acked-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Link: https://patchwork.freedesktop.org/patch/msgid/20200424230546.30271-1-mika.kuoppala@linux.intel.com
The RPS DOWN_TIMEOUT interrupt is signaled after a period of rc6, and
upon receipt of that interrupt we reprogram the GPU clocks down to the
next idle notch [to help convserve power during rc6]. However, on
execlists, we benefit from soft-rc6 immediately parking the GPU and
setting idle frequencies upon idling [within a jiffie], and here the
interrupt prevents us from restarting from our last frequency.
In the process, we can simply opt for a static pm_events mask and rely
on the enable/disable interrupts to flush the worker on parking.
This will reduce the amount of oscillation observed during steady
workloads with microsleeps, as each time the rc6 timeout occurs we
immediately follow with a waitboost for a dropped frame.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Andi Shyti <andi.shyti@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20200422001703.1697-1-chris@chris-wilson.co.uk
