We can use bit fields to store boolean values and when the
bit fields are next to each other, the compiler will combine them
(as long as the size holds enough).
Signed-off-by: Jesse Brandeburg <jesse.brandeburg@intel.com>
Signed-off-by: Anirudh Venkataramanan <anirudh.venkataramanan@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Use more efficient structure ordering by using the pahole tool
and a lot of code inspection to get hot cache lines to have
packed data (no holes if possible) and adjacent warm data.
ice_ring prior to this change:
/* size: 192, cachelines: 3, members: 23 */
/* sum members: 158, holes: 4, sum holes: 12 */
/* padding: 22 */
ice_ring after this change:
/* size: 192, cachelines: 3, members: 25 */
/* sum members: 162, holes: 1, sum holes: 1 */
/* padding: 29 */
ice_tx_buf prior to this change:
/* size: 48, cachelines: 1, members: 7 */
/* sum members: 38, holes: 2, sum holes: 6 */
/* padding: 4 */
/* last cacheline: 48 bytes */
ice_tx_buf after this change:
/* size: 40, cachelines: 1, members: 7 */
/* sum members: 38, holes: 1, sum holes: 2 */
/* last cacheline: 40 bytes */
Signed-off-by: Jesse Brandeburg <jesse.brandeburg@intel.com>
Signed-off-by: Anirudh Venkataramanan <anirudh.venkataramanan@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Currently the driver allows rx-usecs-high values to be set,
but when querying the device for rx-usecs-high the value
does not stick. This is because it was not yet implemented.
Add code to allow the user to change rx-usecs-high and
use this to set the q_vector's intrl value.
Signed-off-by: Brett Creeley <brett.creeley@intel.com>
Signed-off-by: Anirudh Venkataramanan <anirudh.venkataramanan@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
This patch introduces a new function ice_tx_prepare_vlan_flags_dcb to
insert 802.1p priority information into the VLAN header
Signed-off-by: Anirudh Venkataramanan <anirudh.venkataramanan@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
This patch adds a new function ice_vsi_cfg_dcb_rings which updates a
VSI's rings based on DCB traffic class information.
Signed-off-by: Anirudh Venkataramanan <anirudh.venkataramanan@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Since the driver now hard codes the ITR granularity to 2 us in the
GLINT_CTL register the comment next to ITR_GRAN_S needs to be updated.
Signed-off-by: Brett Creeley <brett.creeley@intel.com>
Signed-off-by: Anirudh Venkataramanan <anirudh.venkataramanan@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Currently the ice_q_vector structure and ice_ring_container structure
are taking up more space than necessary due to cache alignment holes
and unnecessary variables respectively. This is not helping the
driver's performance. The following fixes were done to improve cache
alignment, reduce wasted space, and increase performance.
1. Remove the ice_latency_range enum as it is unused.
2. Remove the latency_range variable in the ice_ring_container structure.
3. Change the size of the itr_idx in the ice_ring_container structure
from an int to an u16. This reduced the size of ice_ring_container
structure to 32 Bytes so it has no holes or padding.
4. Re-arrange the ice_q_vector structure using pahole to align
members as best as possible in regards to 64 Byte cache line size.
Signed-off-by: Brett Creeley <brett.creeley@intel.com>
Signed-off-by: Anirudh Venkataramanan <anirudh.venkataramanan@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
commit 63f545ed12 ("ice: Add support for adaptive interrupt moderation")
was meant to add support for adaptive interrupt moderation but there was
an error on my part while formatting the patch, and thus only part of the
patch ended up being submitted.
This patch rectifies the error by adding the rest of the code.
Fixes: 63f545ed12 ("ice: Add support for adaptive interrupt moderation")
Signed-off-by: Anirudh Venkataramanan <anirudh.venkataramanan@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Provide DMA_ATTR_WEAK_ORDERING and DMA_ATTR_SKIP_CPU_SYNC attributes to
the DMA API during the mapping operations on Rx side. With this change
the non-x86 platforms will be able to sync only with what is being used
(2k buffer) instead of entire page. This should yield a slight
performance improvement.
Furthermore, DMA unmap may destroy the changes that were made to the
buffer by CPU when platform is not a x86 one. DMA_ATTR_SKIP_CPU_SYNC
attribute usage fixes this issue.
Also add a sync_single_for_device call during the Rx buffer assignment,
to make sure that the cache lines are cleared before device attempting
to write to the buffer.
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Anirudh Venkataramanan <anirudh.venkataramanan@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
{get,put}_page are atomic operations which we use for page count
handling. The current logic for refcount handling is that we increment
it when passing a skb with the data from the first half of page up to
netstack and recycle the second half of page. This operation protects us
from losing a page since the network stack can decrement the refcount of
page from skb.
The performance can be gently improved by doing the bulk updates of
refcount instead of doing it one by one. During the buffer initialization,
maximize the page's refcount and don't allow the refcount to become
less than two.
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Anirudh Venkataramanan <anirudh.venkataramanan@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Instead of hoping that our ITR granularity will be 2 usec program the
GLINT_CTL register to make sure the ITR granularity is always 2 usecs.
Now that we know what the ITR granularity will be get rid of the check
in ice_probe() to verify our previous assumption.
Signed-off-by: Brett Creeley <brett.creeley@intel.com>
Signed-off-by: Anirudh Venkataramanan <anirudh.venkataramanan@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
This patch includes the following ethtool operations:
1. get_coalesce
2. set_coalesce
3. get_per_q_coalesce
4. set_per_q_coalesce
Each ITR value (current_itr/target_itr) are stored on a per
ice_ring_container basis. This is because each valid ice_ring_container
can have 1 or more rings that are tied to the same q_vector ITR index.
Signed-off-by: Brett Creeley <brett.creeley@intel.com>
Signed-off-by: Anirudh Venkataramanan <anirudh.venkataramanan@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Currently the driver does not support adaptive/dynamic interrupt
moderation. This patch adds support for this. Also, adaptive/dynamic
interrupt moderation is turned on by default upon driver load.
In order to support adaptive interrupt moderation, two functions were
added, ice_update_itr() and ice_itr_divisor(). These are used to
determine the current packet load and to determine a divisor based
on link speed respectively.
This patch also adds the ICE_ITR_GRAN_S define that is used in the
hot-path when setting a new ITR value. The shift is used to pet two
birds with one hand, set the ITR value while re-enabling the
interrupt. Also, the ICE_ITR_GRAN_S is defined as 1 because the device
has a ITR granularity of 2usecs.
Signed-off-by: Brett Creeley <brett.creeley@intel.com>
Signed-off-by: Anirudh Venkataramanan <anirudh.venkataramanan@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Prior to this commit the driver was running into tx_timeouts when a
queue was stressed enough. This was happening because the HW tail
and SW tail (NTU) were incorrectly out of sync. Consequently this was
causing the HW head to collide with the HW tail, which to the hardware
means that all descriptors posted for Tx have been processed.
Due to the Tx logic used in the driver SW tail and HW tail are allowed
to be out of sync. This is done as an optimization because it allows the
driver to write HW tail as infrequently as possible, while still
updating the SW tail index to keep track. However, there are situations
where this results in the tail never getting updated, resulting in Tx
timeouts.
Tx HW tail write condition:
if (netif_xmit_stopped(txring_txq(tx_ring) || !skb->xmit_more)
writel(sw_tail, tx_ring->tail);
An issue was found in the Tx logic that was causing the afore mentioned
condition for updating HW tail to never happen, causing tx_timeouts.
In ice_xmit_frame_ring we calculate how many descriptors we need for the
Tx transaction based on the skb the kernel hands us. This is then passed
into ice_maybe_stop_tx along with some extra padding to determine if we
have enough descriptors available for this transaction. If we don't then
we return -EBUSY to the stack, otherwise we move on and eventually
prepare the Tx descriptors accordingly in ice_tx_map and set
next_to_watch. In ice_tx_map we make another call to ice_maybe_stop_tx
with a value of MAX_SKB_FRAGS + 4. The key here is that this value is
possibly less than the value we sent in the first call to
ice_maybe_stop_tx in ice_xmit_frame_ring. Now, if the number of unused
descriptors is between MAX_SKB_FRAGS + 4 and the value used in the first
call to ice_maybe_stop_tx in ice_xmit_frame_ring then we do not update
the HW tail because of the "Tx HW tail write condition" above. This is
because in ice_maybe_stop_tx we return success from ice_maybe_stop_tx
instead of calling __ice_maybe_stop_tx and subsequently calling
netif_stop_subqueue, which sets the __QUEUE_STATE_DEV_XOFF bit. This
bit is then checked in the "Tx HW tail write condition" by calling
netif_xmit_stopped and subsequently updating HW tail if the
afore mentioned bit is set.
In ice_clean_tx_irq, if next_to_watch is not NULL, we end up cleaning
the descriptors that HW sets the DD bit on and we have the budget. The
HW head will eventually run into the HW tail in response to the
description in the paragraph above.
The next time through ice_xmit_frame_ring we make the initial call to
ice_maybe_stop_tx with another skb from the stack. This time we do not
have enough descriptors available and we return NETDEV_TX_BUSY to the
stack and end up setting next_to_watch to NULL.
This is where we are stuck. In ice_clean_tx_irq we never clean anything
because next_to_watch is always NULL and in ice_xmit_frame_ring we never
update HW tail because we already return NETDEV_TX_BUSY to the stack and
eventually we hit a tx_timeout.
This issue was fixed by making sure that the second call to
ice_maybe_stop_tx in ice_tx_map is passed a value that is >= the value
that was used on the initial call to ice_maybe_stop_tx in
ice_xmit_frame_ring. This was done by adding the following defines to
make the logic more clear and to reduce the chance of mucking this up
again:
ICE_CACHE_LINE_BYTES 64
ICE_DESCS_PER_CACHE_LINE (ICE_CACHE_LINE_BYTES / \
sizeof(struct ice_tx_desc))
ICE_DESCS_FOR_CTX_DESC 1
ICE_DESCS_FOR_SKB_DATA_PTR 1
The ICE_CACHE_LINE_BYTES being 64 is an assumption being made so we
don't have to figure this out on every pass through the Tx path. Instead
I added a sanity check in ice_probe to verify cache line size and print
a message if it's not 64 Bytes. This will make it easier to file issues
if they are seen when the cache line size is not 64 Bytes when reading
from the GLPCI_CNF2 register.
Signed-off-by: Brett Creeley <brett.creeley@intel.com>
Signed-off-by: Anirudh Venkataramanan <anirudh.venkataramanan@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
This issue came about when looking at the VF function
ice_vc_cfg_irq_map_msg. Currently we are assigning the itr_setting value
to the itr_idx received from the AVF driver, which is not correct and is
not used for the VF flow anyway. Currently the only way we set the ITR
index for both the PF and VF driver is by hard coding ICE_TX_ITR or
ICE_RX_ITR for the ITR index on each q_vector.
To fix this, add the member itr_idx in struct ice_ring_container. This
can then be used to dynamically program the correct ITR index. This change
also affected the PF driver so make the necessary changes there as well.
Also, removed the itr_setting member in struct ice_ring because it is not
being used meaningfully and is going to be removed in a future patch that
includes dynamic ITR.
On another note, this will be useful moving forward if we decide to split
Rx/Tx rings on different q_vectors instead of sharing them as queue pairs.
Signed-off-by: Brett Creeley <brett.creeley@intel.com>
Signed-off-by: Anirudh Venkataramanan <anirudh.venkataramanan@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Currently there is no support for dynamic interrupt moderation. This
patch adds some initial code to support this. The following changes
were made:
1. Currently we are using multiple members to store the interrupt
granularity (itr_gran_25/50/100/200). This is not necessary because
we can query the device to determine what the interrupt granularity
should be set to, done by a new function ice_get_itr_intrl_gran.
2. Added intrl to ice_q_vector structure to support interrupt rate
limiting.
3. Added the function ice_intrl_usecs_to_reg for converting to a value
in usecs that the device understands.
4. Added call to write to the GLINT_RATE register. Disable intrl by
default for now.
5. Changed rx/tx_itr_setting to itr_setting because having both seems
redundant because a ring is either Tx or Rx.
6. Initialize itr_setting for both Tx/Rx rings in ice_vsi_alloc_rings()
Signed-off-by: Brett Creeley <brett.creeley@intel.com>
Signed-off-by: Anirudh Venkataramanan <anirudh.venkataramanan@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
When a malicious operation is detected, the firmware triggers an
interrupt, which is then picked up by the service task (specifically by
ice_handle_mdd_event). A reset is scheduled if required.
Tx hang detection works in a similar way, except the logic here monitors
the VSI's Tx queues and tries to revive them if stalled. If the hang is
not resolved, the kernel eventually calls ndo_tx_timeout, which is
handled by ice_tx_timeout.
Signed-off-by: Sudheer Mogilappagari <sudheer.mogilappagari@intel.com>
Signed-off-by: Anirudh Venkataramanan <anirudh.venkataramanan@intel.com>
Tested-by: Tony Brelinski <tonyx.brelinski@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Recent versions of checkpatch have a new warning based on a documented
preference of Linus to not use bool in structures due to wasted space and
the size of bool is implementation dependent. For more information, see
the email thread at https://lkml.org/lkml/2017/11/21/384.
Signed-off-by: Bruce Allan <bruce.w.allan@intel.com>
Signed-off-by: Anirudh Venkataramanan <anirudh.venkataramanan@intel.com>
Tested-by: Tony Brelinski <tonyx.brelinski@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
This patch adds support for VLANs. When a VLAN is created a switch filter
is added to direct the VLAN traffic to the corresponding VSI. When a VLAN
is deleted, the filter is deleted as well.
This patch also adds support for the following hardware offloads.
1) VLAN tag insertion/stripping
2) Receive Side Scaling (RSS)
3) Tx checksum and TCP segmentation
4) Rx checksum
Signed-off-by: Anirudh Venkataramanan <anirudh.venkataramanan@intel.com>
Tested-by: Tony Brelinski <tonyx.brelinski@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
This patch implements ice_start_xmit (the handler for ndo_start_xmit) and
related functions. ice_start_xmit ultimately calls ice_tx_map, where the
Tx descriptor is built and posted to the hardware by bumping the ring tail.
This patch also implements ice_napi_poll, which is invoked when there's an
interrupt on the VSI's queues. The interrupt can be due to either a
completed Tx or an Rx event. In case of a completed Tx/Rx event, resources
are reclaimed. Additionally, in case of an Rx event, the skb is fetched
and passed up to the network stack.
Signed-off-by: Anirudh Venkataramanan <anirudh.venkataramanan@intel.com>
Tested-by: Tony Brelinski <tonyx.brelinski@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
This patch configures the VSIs to be able to send and receive
packets by doing the following:
1) Initialize flexible parser to extract and include certain
fields in the Rx descriptor.
2) Add Tx queues by programming the Tx queue context (implemented in
ice_vsi_cfg_txqs). Note that adding the queues also enables (starts)
the queues.
3) Add Rx queues by programming Rx queue context (implemented in
ice_vsi_cfg_rxqs). Note that this only adds queues but doesn't start
them. The rings will be started by calling ice_vsi_start_rx_rings on
interface up.
4) Configure interrupts for VSI queues.
5) Implement ice_open and ice_stop.
Signed-off-by: Anirudh Venkataramanan <anirudh.venkataramanan@intel.com>
Tested-by: Tony Brelinski <tonyx.brelinski@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
This patch introduces data structures and functions to alloc/free
VSIs. The driver represents a VSI using the ice_vsi structure.
Some noteworthy points about VSI allocation:
1) A VSI is allocated in the firmware using the "add VSI" admin queue
command (implemented as ice_aq_add_vsi). The firmware returns an
identifier for the allocated VSI. The VSI context is used to program
certain aspects (loopback, queue map, etc.) of the VSI's configuration.
2) A VSI is deleted using the "free VSI" admin queue command (implemented
as ice_aq_free_vsi).
3) The driver represents a VSI using struct ice_vsi. This is allocated
and initialized as part of the ice_vsi_alloc flow, and deallocated
as part of the ice_vsi_delete flow.
4) Once the VSI is created, a netdev is allocated and associated with it.
The VSI's ring and vector related data structures are also allocated
and initialized.
5) A VSI's queues can either be contiguous or scattered. To do this, the
driver maintains a bitmap (vsi->avail_txqs) which is kept in sync with
the firmware's VSI queue allocation imap. If the VSI can't get a
contiguous queue allocation, it will fallback to scatter. This is
implemented in ice_vsi_get_qs which is called as part of the VSI setup
flow. In the release flow, the VSI's queues are released and the bitmap
is updated to reflect this by ice_vsi_put_qs.
CC: Shannon Nelson <shannon.nelson@oracle.com>
Signed-off-by: Anirudh Venkataramanan <anirudh.venkataramanan@intel.com>
Acked-by: Shannon Nelson <shannon.nelson@oracle.com>
Tested-by: Tony Brelinski <tonyx.brelinski@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
This patch continues the initialization flow as follows:
1) Allocate and initialize necessary fields (like vsi, num_alloc_vsi,
irq_tracker, etc) in the ice_pf instance.
2) Setup the miscellaneous interrupt handler. This also known as the
"other interrupt causes" (OIC) handler and is used to handle non
hotpath interrupts (like control queue events, link events,
exceptions, etc.
3) Implement a background task to process admin queue receive (ARQ)
events received by the driver.
CC: Shannon Nelson <shannon.nelson@oracle.com>
Signed-off-by: Anirudh Venkataramanan <anirudh.venkataramanan@intel.com>
Acked-by: Shannon Nelson <shannon.nelson@oracle.com>
Tested-by: Tony Brelinski <tonyx.brelinski@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Jesse Brandeburg