[Intel-gfx] [PATCH 4/7] drm/i915/guc: Add non blocking CTB send function
Matthew Brost
matthew.brost at intel.com
Thu Jul 1 00:19:01 UTC 2021
On Thu, Jul 01, 2021 at 12:52:58AM +0200, Michal Wajdeczko wrote:
>
>
> On 28.06.2021 01:14, Matthew Brost wrote:
> > Add non blocking CTB send function, intel_guc_send_nb. GuC submission
> > will send CTBs in the critical path and does not need to wait for these
> > CTBs to complete before moving on, hence the need for this new function.
> >
> > The non-blocking CTB now must have a flow control mechanism to ensure
> > the buffer isn't overrun. A lazy spin wait is used as we believe the
> > flow control condition should be rare with a properly sized buffer.
> >
> > The function, intel_guc_send_nb, is exported in this patch but unused.
> > Several patches later in the series make use of this function.
> >
> > v2:
> > (Michal)
> > - Use define for H2G room calculations
> > - Move INTEL_GUC_SEND_NB define
> > (Daniel Vetter)
> > - Use msleep_interruptible rather than cond_resched
> >
> > Signed-off-by: John Harrison <John.C.Harrison at Intel.com>
> > Signed-off-by: Matthew Brost <matthew.brost at intel.com>
> > ---
> > .../gt/uc/abi/guc_communication_ctb_abi.h | 3 +-
> > drivers/gpu/drm/i915/gt/uc/intel_guc.h | 11 ++-
> > drivers/gpu/drm/i915/gt/uc/intel_guc_ct.c | 90 +++++++++++++++++--
> > drivers/gpu/drm/i915/gt/uc/intel_guc_ct.h | 4 +-
> > 4 files changed, 94 insertions(+), 14 deletions(-)
> >
> > diff --git a/drivers/gpu/drm/i915/gt/uc/abi/guc_communication_ctb_abi.h b/drivers/gpu/drm/i915/gt/uc/abi/guc_communication_ctb_abi.h
> > index e933ca02d0eb..99e1fad5ca20 100644
> > --- a/drivers/gpu/drm/i915/gt/uc/abi/guc_communication_ctb_abi.h
> > +++ b/drivers/gpu/drm/i915/gt/uc/abi/guc_communication_ctb_abi.h
> > @@ -79,7 +79,8 @@ static_assert(sizeof(struct guc_ct_buffer_desc) == 64);
> > * +---+-------+--------------------------------------------------------------+
> > */
> >
> > -#define GUC_CTB_MSG_MIN_LEN 1u
> > +#define GUC_CTB_HDR_LEN 1u
> > +#define GUC_CTB_MSG_MIN_LEN GUC_CTB_HDR_LEN
>
> if you insist to use dedicated macro for the CTB header len then to be
> consistent you need rename header bitfield macros as well, thus
> sections/tables will look like:
>
Kernel doc can't link to defines, right? So what does it matter?
This example is also a reason why I think the kernel doc included is too
verbose to hand generate. Either we auto-gen this or just don't include
it.
> /**
> * DOC: CTB Message
> *
> * +---+-------+-------------------------------------+
> * | | Bits | Description |
> * +===+=======+=====================================+
> * | 0 | 31:0 | `CTB Header`_ |
> * +---+-------+-------------------------------------+
> * | 1 | 31:0 | +-------------------------------+ |
> * +---+-------+ | | |
> * |...| | | CTB Payload | |
> * +---+-------+ | | |
> * | n | 31:0 | +-------------------------------+ |
> * +---+-------+-------------------------------------+
> */
>
> /**
> * DOC: CTB Header
> *
> * +---+-------+-------------------------------------+
> * | | Bits | Description |
> * +===+=======+=====================================+
> * | 0 | 31:16 | **FENCE** - ... |
> * | +-------+-------------------------------------+
> * | | 15:12 | **FORMAT** - ... |
> * | +-------+-------------------------------------+
> * | | 11:8 | **RESERVED** |
> * | +-------+-------------------------------------+
> * | | 7:0 | **NUM_DWORDS** - ... |
> * +---+-------+-------------------------------------+
> */
>
> #define GUC_CTB_HDR_0_FENCE (0xffff << 16)
> #define GUC_CTB_HDR_0_FORMAT (0xf << 12)
> #define GUC_CTB_FORMAT_HXG 0u
> #define GUC_CTB_HDR_0_RESERVED (0xf << 8)
> #define GUC_CTB_HDR_0_NUM_DWORDS (0xff << 0)
> #define GUC_CTB_MAX_DWORDS 255u
>
> #define GUC_CTB_MSG_MIN_LEN GUC_CTB_HDR_LEN
> #define GUC_CTB_MSG_MAX_LEN (GUC_CTB_HDR_LEN + GUC_CTB_MAX_DWORDS)
>
>
> alternatively leave ABI defs as-as and just move your HDR definition out
> of ABI headers to inteL_guc_fwif.h as:
>
> #define GUC_CTB_HDR_LEN GUC_CTB_MSG_MIN_LEN
This is backwards. The minimum length of a message is the header length.
>
>
> > #define GUC_CTB_MSG_MAX_LEN 256u
> > #define GUC_CTB_MSG_0_FENCE (0xffff << 16)
> > #define GUC_CTB_MSG_0_FORMAT (0xf << 12)
> > diff --git a/drivers/gpu/drm/i915/gt/uc/intel_guc.h b/drivers/gpu/drm/i915/gt/uc/intel_guc.h
> > index 4abc59f6f3cd..efc690fc8fb1 100644
> > --- a/drivers/gpu/drm/i915/gt/uc/intel_guc.h
> > +++ b/drivers/gpu/drm/i915/gt/uc/intel_guc.h
> > @@ -74,7 +74,14 @@ static inline struct intel_guc *log_to_guc(struct intel_guc_log *log)
> > static
> > inline int intel_guc_send(struct intel_guc *guc, const u32 *action, u32 len)
> > {
> > - return intel_guc_ct_send(&guc->ct, action, len, NULL, 0);
> > + return intel_guc_ct_send(&guc->ct, action, len, NULL, 0, 0);
> > +}
> > +
> > +static
> > +inline int intel_guc_send_nb(struct intel_guc *guc, const u32 *action, u32 len)
> > +{
> > + return intel_guc_ct_send(&guc->ct, action, len, NULL, 0,
> > + INTEL_GUC_SEND_NB);
> > }
> >
> > static inline int
> > @@ -82,7 +89,7 @@ intel_guc_send_and_receive(struct intel_guc *guc, const u32 *action, u32 len,
> > u32 *response_buf, u32 response_buf_size)
> > {
> > return intel_guc_ct_send(&guc->ct, action, len,
> > - response_buf, response_buf_size);
> > + response_buf, response_buf_size, 0);
> > }
> >
> > static inline void intel_guc_to_host_event_handler(struct intel_guc *guc)
> > diff --git a/drivers/gpu/drm/i915/gt/uc/intel_guc_ct.c b/drivers/gpu/drm/i915/gt/uc/intel_guc_ct.c
> > index 43e03aa2dde8..90ee95a240e8 100644
> > --- a/drivers/gpu/drm/i915/gt/uc/intel_guc_ct.c
> > +++ b/drivers/gpu/drm/i915/gt/uc/intel_guc_ct.c
> > @@ -3,6 +3,11 @@
> > * Copyright © 2016-2019 Intel Corporation
> > */
> >
> > +#include <linux/circ_buf.h>
>
> > +#include <linux/ktime.h>
> > +#include <linux/time64.h>
> > +#include <linux/timekeeping.h>
>
> these headers should likely be part of patch 5/7
>
Yes, will fix.
> > +
> > #include "i915_drv.h"
> > #include "intel_guc_ct.h"
> > #include "gt/intel_gt.h"
> > @@ -373,7 +378,7 @@ static void write_barrier(struct intel_guc_ct *ct)
> > static int ct_write(struct intel_guc_ct *ct,
> > const u32 *action,
> > u32 len /* in dwords */,
> > - u32 fence)
> > + u32 fence, u32 flags)
> > {
> > struct intel_guc_ct_buffer *ctb = &ct->ctbs.send;
> > struct guc_ct_buffer_desc *desc = ctb->desc;
> > @@ -409,7 +414,7 @@ static int ct_write(struct intel_guc_ct *ct,
> > used = tail - head;
> ^^^^
> code above for blocking path is likely calculating free space in old
> fashion, while below you are using CIRC_BUF ...
>
> so maybe switch to use circ_buf could be done as separate earlier step
> (not part of intro of non-blocking send) so then below use of CIRC_BUF
> in non-blocking path will look natural
>
This is existing code that deleted 2 patches from now. I don't think we
need to rework it to just delete it.
> >
> > /* make sure there is a space including extra dw for the fence */
>
> this extra DW is now "header" not just fence
>
Correct but existing code, right? We should've fixed this with the CTB
interface change boondoggle. I suppose I'll fix this if I respin.
> > - if (unlikely(used + len + 1 >= size))
> > + if (unlikely(used + len + GUC_CTB_HDR_LEN >= size))
> > return -ENOSPC;
> >
> > /*
> > @@ -421,9 +426,13 @@ static int ct_write(struct intel_guc_ct *ct,
> > FIELD_PREP(GUC_CTB_MSG_0_NUM_DWORDS, len) |
> > FIELD_PREP(GUC_CTB_MSG_0_FENCE, fence);
> >
> > - hxg = FIELD_PREP(GUC_HXG_MSG_0_TYPE, GUC_HXG_TYPE_REQUEST) |
> > - FIELD_PREP(GUC_HXG_REQUEST_MSG_0_ACTION |
> > - GUC_HXG_REQUEST_MSG_0_DATA0, action[0]);
> > + hxg = (flags & INTEL_GUC_SEND_NB) ?
> > + (FIELD_PREP(GUC_HXG_MSG_0_TYPE, GUC_HXG_TYPE_EVENT) |
> > + FIELD_PREP(GUC_HXG_EVENT_MSG_0_ACTION |
> > + GUC_HXG_EVENT_MSG_0_DATA0, action[0])) :
> > + (FIELD_PREP(GUC_HXG_MSG_0_TYPE, GUC_HXG_TYPE_REQUEST) |
> > + FIELD_PREP(GUC_HXG_REQUEST_MSG_0_ACTION |
> > + GUC_HXG_REQUEST_MSG_0_DATA0, action[0]));
> >
> > CT_DEBUG(ct, "writing (tail %u) %*ph %*ph %*ph\n",
> > tail, 4, &header, 4, &hxg, 4 * (len - 1), &action[1]);
> > @@ -500,6 +509,48 @@ static int wait_for_ct_request_update(struct ct_request *req, u32 *status)
> > return err;
> > }
> >
> > +static inline bool h2g_has_room(struct intel_guc_ct_buffer *ctb, u32 len_dw)
>
> why h2g prefix ? function is taking pure *ctb, so maybe:
>
This was called ctb_has_room in the prior to this rev. You literally
suggested this change in previous revs comments. A follow up patch
basically does what you suggest below once we start checking the g2h
credits too.
>
> static bool ctb_has_room(struct intel_guc_ct_buffer *ctb, u32 len_dw)
> { ... }
>
> static bool ct_can_send(struct intel_guc_ct *ct, u32 len_dw)
> {
> return ctb_has_room(&ct->ctbs.send, len_dw + CTB_HDR_LEN);
> }
>
>
> as we use send/recv, not h2g/g2h and to avoid mistakes add extra header
> len here (not by caller - same as in call to ct_write)
>
Don't agree. I prefer h2g / g2h here. Send and recv and both relative to
the observer while h2g / g2h are not.
> > +{
> > + struct guc_ct_buffer_desc *desc = ctb->desc;
> > + u32 head = READ_ONCE(desc->head);
> > + u32 space;
> > +
> > + space = CIRC_SPACE(desc->tail, head, ctb->size);
> > +
> > + return space >= len_dw;
> > +}
> > +
> > +static int ct_send_nb(struct intel_guc_ct *ct,
> > + const u32 *action,
> > + u32 len,
> > + u32 flags)
> > +{
> > + struct intel_guc_ct_buffer *ctb = &ct->ctbs.send;
> > + unsigned long spin_flags;
> > + u32 fence;
> > + int ret;
> > +
> > + spin_lock_irqsave(&ctb->lock, spin_flags);
> > +
> > + ret = h2g_has_room(ctb, len + GUC_CTB_HDR_LEN);
> > + if (unlikely(!ret)) {
> > + ret = -EBUSY;
> > + goto out;
> > + }
> > +
> > + fence = ct_get_next_fence(ct);
> > + ret = ct_write(ct, action, len, fence, flags);
> > + if (unlikely(ret))
> > + goto out;
> > +
> > + intel_guc_notify(ct_to_guc(ct));
> > +
> > +out:
> > + spin_unlock_irqrestore(&ctb->lock, spin_flags);
> > +
> > + return ret;
> > +}
> > +
> > static int ct_send(struct intel_guc_ct *ct,
> > const u32 *action,
> > u32 len,
> > @@ -507,8 +558,10 @@ static int ct_send(struct intel_guc_ct *ct,
> > u32 response_buf_size,
> > u32 *status)
> > {
> > + struct intel_guc_ct_buffer *ctb = &ct->ctbs.send;
> > struct ct_request request;
> > unsigned long flags;
> > + unsigned int sleep_period_ms = 1;
> > u32 fence;
> > int err;
> >
> > @@ -516,8 +569,24 @@ static int ct_send(struct intel_guc_ct *ct,
> > GEM_BUG_ON(!len);
> > GEM_BUG_ON(len & ~GUC_CT_MSG_LEN_MASK);
> > GEM_BUG_ON(!response_buf && response_buf_size);
> > + might_sleep();
> > +
> > + /*
> > + * We use a lazy spin wait loop here as we believe that if the CT
> > + * buffers are sized correctly the flow control condition should be
> > + * rare.
> > + */
> > +retry:
> > + spin_lock_irqsave(&ctb->lock, flags);
> > + if (unlikely(!h2g_has_room(ctb, len + GUC_CTB_HDR_LEN))) {
> > + spin_unlock_irqrestore(&ctb->lock, flags);
> >
> > - spin_lock_irqsave(&ct->ctbs.send.lock, flags);
> > + if (msleep_interruptible(sleep_period_ms))
> > + return -EINTR;
> > + sleep_period_ms = sleep_period_ms << 1;
> > +
> > + goto retry;
> > + }
>
> I'm still not convinced that this chunk should be exactly in the same
> patch that adds non-blocking path, it's not that with very first use of
> one NB call we will hit such problem
>
Once you have non-blocking calls the credits can be exhausted.
> so IMHO it can be done earlier, in by anticipating such problem, or
> later as fix for potential problem due to stress
>
What?
> note that in the commit message you admitted that new NB variant is not
> used yet, so we can't be hit
>
If we introduce a function that can cause a problem for another function
we likely should make both work in a single patch.
> in other words, try to make patches focused on one item at the time.
>
This patch is quite simple and all the changes follow together. For
example if you did a 'git blame' on the above lines and the changes were
only by themselves it wouldn't make any sense at all. With the patch as
is it makes sense.
> >
> > fence = ct_get_next_fence(ct);
> > request.fence = fence;
> > @@ -529,9 +598,9 @@ static int ct_send(struct intel_guc_ct *ct,
> > list_add_tail(&request.link, &ct->requests.pending);
> > spin_unlock(&ct->requests.lock);
> >
> > - err = ct_write(ct, action, len, fence);
> > + err = ct_write(ct, action, len, fence, 0);
> >
> > - spin_unlock_irqrestore(&ct->ctbs.send.lock, flags);
> > + spin_unlock_irqrestore(&ctb->lock, flags);
> >
> > if (unlikely(err))
> > goto unlink;
> > @@ -571,7 +640,7 @@ static int ct_send(struct intel_guc_ct *ct,
> > * Command Transport (CT) buffer based GuC send function.
> > */
> > int intel_guc_ct_send(struct intel_guc_ct *ct, const u32 *action, u32 len,
> > - u32 *response_buf, u32 response_buf_size)
> > + u32 *response_buf, u32 response_buf_size, u32 flags)
> > {
> > u32 status = ~0; /* undefined */
> > int ret;
> > @@ -581,6 +650,9 @@ int intel_guc_ct_send(struct intel_guc_ct *ct, const u32 *action, u32 len,
> > return -ENODEV;
> > }
> >
> > + if (flags & INTEL_GUC_SEND_NB)
> > + return ct_send_nb(ct, action, len, flags);
> > +
> > ret = ct_send(ct, action, len, response_buf, response_buf_size, &status);
> > if (unlikely(ret < 0)) {
> > CT_ERROR(ct, "Sending action %#x failed (err=%d status=%#X)\n",
> > diff --git a/drivers/gpu/drm/i915/gt/uc/intel_guc_ct.h b/drivers/gpu/drm/i915/gt/uc/intel_guc_ct.h
> > index 1ae2dde6db93..f6a4d5b33467 100644
> > --- a/drivers/gpu/drm/i915/gt/uc/intel_guc_ct.h
> > +++ b/drivers/gpu/drm/i915/gt/uc/intel_guc_ct.h
> > @@ -42,7 +42,6 @@ struct intel_guc_ct_buffer {
> > bool broken;
> > };
> >
> > -
> > /** Top-level structure for Command Transport related data
> > *
> > * Includes a pair of CT buffers for bi-directional communication and tracking
> > @@ -87,8 +86,9 @@ static inline bool intel_guc_ct_enabled(struct intel_guc_ct *ct)
> > return ct->enabled;
> > }
> >
> > +#define INTEL_GUC_SEND_NB BIT(31)
>
> INTEL_GUC_CT_SEND_NB ?
> ^^^^
>
I guess, but does it really matter? I could see it either way.
Again please sign off with your name when done with your comments so I
know if it is your last comment.
Matt
> > int intel_guc_ct_send(struct intel_guc_ct *ct, const u32 *action, u32 len,
> > - u32 *response_buf, u32 response_buf_size);
> > + u32 *response_buf, u32 response_buf_size, u32 flags);
> > void intel_guc_ct_event_handler(struct intel_guc_ct *ct);
> >
> > #endif /* _INTEL_GUC_CT_H_ */
> >
More information about the Intel-gfx
mailing list