Also use num_online_cpus that faithfully represents the number of online cpus.
Add a new pool for skb container
Signed-off-by: Vikram Narayanan <vikram186@gmail.com>

Abhiram Balasubramanian authored Apr 06, 2017 and Vikram Narayanan committed Jun 09, 2017

Signed-off-by: Abhiram Balasubramanian <abhiram@cs.utah.edu>

Abhiram Balasubramanian authored Mar 17, 2017 and Vikram Narayanan committed Jun 09, 2017

Needed to differentiate between lcd and klcd in `ps`
output (use case - mangle thread priority/cpu affinity)
Signed-off-by: Abhiram Balasubramanian <abhiram@cs.utah.edu>

Signed-off-by: Vikram Narayanan <vikram186@gmail.com>

Instead of sharing the same async channel with dispatch loop, create a separate
channel for xmit. This channel is exclusively used only for xmit
Signed-off-by: Vikram Narayanan <vikram186@gmail.com>

Signed-off-by: Vikram Narayanan <vikram186@gmail.com>

For now, grant the sync endpoint and never revert it. Set a flag for non LCD
context and handle it inside the original xmit function.
Signed-off-by: Vikram Narayanan <vikram186@gmail.com>

Signed-off-by: Vikram Narayanan <vikram186@gmail.com>

Make our device use private pool for skb->data by setting the feature flag.
Initialize shared data pool and pass it on to the other side during probe.
Refactor caller/callee functions to transfer shared data skb's. Marshal the
necessary skb members inside the shared data region.
Signed-off-by: Vikram Narayanan <vikram186@gmail.com>

Signed-off-by: Vikram Narayanan <vikram186@gmail.com>

Rx packets > 256 requires page fragments. (256 size limit is imposed by the
IXGBE driver according to their internal buffer size). Add skb_rx_fragto add a
page fragment to the skb. This is not an optimized solution. Just add this to
make the Rxpath complete.
Signed-off-by: Vikram Narayanan <vikram186@gmail.com>

In the Rx path, page fragments are mapped inside the isolated domain and passed
onto the kernel for consumption by upper layers. Let's save the lcd_struct so
that we can easily do a lookup.
Signed-off-by: Vikram Narayanan <vikram186@gmail.com>

When memory is volunteered, it needs to be unvolunteered when the packet is
freed. For that we need to access the cspace of the task struct which
volunteered the memory. To make sure whether or not klcd volunteered it, let's
save its current pointer.
Signed-off-by: Vikram Narayanan <vikram186@gmail.com>

skb->data needs to be shared across LCD and KLCD. To facilitate this, create a
shared area and an allocator on top and use this for allocating skb->data for
sk_buffs. Use NETIF flags to enable priv pool allocation.
Signed-off-by: Vikram Narayanan <vikram186@gmail.com>

Based on per thread userspace gary's allocator
Signed-off-by: Vikram Narayanan <vikram186@gmail.com>

Signed-off-by: Vikram Narayanan <vikram186@gmail.com>

Use the native driver's frequency. This would indirectly facilicate frequent
polling, as the LCD's main loop calls the poll function only when it gets an
IPC message (ipc_poll_recv is a blocking call)
Signed-off-by: Vikram Narayanan <vikram186@gmail.com>

Add support for the following APIs
* netif_napi_add
* netif_napi_del
* napi_consume_skb
* netif_wake_subqueue
* netif_receive_skb
* napi_gro_receive
* __napi_alloc_skb
* eth_type_trans

Enable interrupt for only one tx/rx queue and enable a sort of napi polling at
LCD end. In contrast to napi polling, this polling never switches to interrupt
mode. Polling is shutdown when the device is removed
Signed-off-by: Vikram Narayanan <vikram186@gmail.com>

skbs are used extensively in tx/rx path. Use a simple hashtable instead of
polluting the kernel sources with sk_buff_containers
Signed-off-by: Vikram Narayanan <vikram186@gmail.com>

Introduce hashtable for storing containers. This needs to be unified and
genaralized for all datastructures.
Signed-off-by: Vikram Narayanan <vikram186@gmail.com>

Handle the case if lcd has not started at all. In that case, current->lcd would
be null.
Signed-off-by: Vikram Narayanan <vikram186@gmail.com>

Replace native api with LCD variant
Signed-off-by: Vikram Narayanan <vikram186@gmail.com>

Signed-off-by: Vikram Narayanan <vikram186@gmail.com>

Add stub functions for timer and dev_trans_start
Signed-off-by: Vikram Narayanan <vikram186@gmail.com>

dma_map_single uses the default helpers such as virt_to_page which is not valid
for pages mapped in the rammap region. As every page is mapped during xmit,
this page needs to be programmed in the iommu before transmission.
Signed-off-by: Vikram Narayanan <vikram186@gmail.com>

Using iommu apis, assign/deassign the pci device during probe and remove inside
the LCD. Also map appropriate number of pages during xmit.
Signed-off-by: Vikram Narayanan <vikram186@gmail.com>

The helper macros phys_to_page doesn't work inside LCD. Fix that by changing
those calls to appropriate LCD helpers.
Signed-off-by: Vikram Narayanan <vikram186@gmail.com>

Signed-off-by: Vikram Narayanan <vikram186@gmail.com>

Introduce four lcd_syscalls for programming iommu for a pci device from LCD.
PCI devices can now be assigned to an LCD. During probe, one can assign a
device to a particular LCD. During device assignment, all pages inside the heap
and rammap regions are mapped in the iommu. During data plane operation, one
can map/unmap a page by means of lcd_syscall's.
Signed-off-by: Vikram Narayanan <vikram186@gmail.com>

During ept walk, not always do we need to know about missing mappings. Add
verbose flag to print a warning for the missig mapping. This is just to prevent
printk flood during iommu map.
Signed-off-by: Vikram Narayanan <vikram186@gmail.com>

netdev->state is updated by many kernel functions. Synchronize the states
across RPCs to have a sane state inside the LCD. Also call the local liblcd
functions for dev_addr helpers to keep the LCD copy of the datastructure sane.
Signed-off-by: Vikram Narayanan <vikram186@gmail.com>

mdio45_probe calls the read callback multiple times. So handle it inside the
LCD.
Signed-off-by: Vikram Narayanan <vikram186@gmail.com>

Timers can't run inside LCD as there is no support. So, start a timer in KLCD
(but with a lower frequency compared to the original one) and trigger a handler
inside LCD when the timer expires.

In case of ixgbe driver, a service_event is triggered by the timer.

Also add support for asynchronous register dump via sysfs module params.
Signed-off-by: Vikram Narayanan <vikram186@gmail.com>

Add the following functions
* __hw_addr_sync_dev
* __hw_addr_unsync_dev
* pci_disable_msix
* pci_enable_msix_range
* netif_tx_wake_all_queues

pci helpers are required for manipulating msi interrupts. __hw_addr helpers
provide functionality for adding/removing hw address to the netdev address
lists.

Also add print and dump stack in all the stub functions to know where this
function originates from.
Signed-off-by: Vikram Narayanan <vikram186@gmail.com>

struct net_device instance of the driver is heavily modified by the kernel. In
order to keep things in sync, we need some internal helpers so as to avoid
volunteering memory for various structures and marshalling lists.These files
pull in a minimal set of helper functions from the kernel to keep things
intact.
Signed-off-by: Vikram Narayanan <vikram186@gmail.com>

Signed-off-by: Vikram Narayanan <vikram186@gmail.com>

Enable the following function pointers:
* ndo_open
* ndo_stop
* ndo_start_xmit
* ndo_set_rx_mode
* ndo_validate_addr
* ndo_set_mac_address
* ndo_change_mtu
* ndo_tx_timeout
* ndo_set_tx_maxrate
* ndo_get_stats64

Add missing dev_addr_add function as well. Remove VLAN features from ixgbe (to
be explored in the future)
Signed-off-by: Vikram Narayanan <vikram186@gmail.com>

Signed-off-by: Vikram Narayanan <vikram186@gmail.com>

In the native kernel, usleep_range leans on the hrtimer interface. As we don't
have any timer support inside LCDs, make this api perform an udelay instead.
Both usleep_range and msleep needs to be fixed in the future
Signed-off-by: Vikram Narayanan <vikram186@gmail.com>

Pass dma_mask and other ref pointer as well
Signed-off-by: Vikram Narayanan <vikram186@gmail.com>