MC_TX(9E) | Driver Entry Points | MC_TX(9E) |
mc_tx
, mri_tx
— transmit a message block chain
#include
<sys/mac_provider.h>
mblk_t *
prefix_m_tx
(void *driver,
mblk_t *mp_chain);
mblk_t *
prefix_ring_tx
(void *driver_rh,
mblk_t *mp_chain);
illumos DDI specific
The
mri_tx
()
entry point is
Uncommitted
- API and ABI stability is not guaranteed.
The
mc_tx
()
entry point is called when the system requires a device driver to transmit
data. The device driver will receive a chain of message blocks. The
mp_chain argument represents the first frame. The
frame may be spread out across one or more
mblk(9S) structures that are linked
together by the b_cont member. There may be multiple
frames, linked together by the b_next pointer of the
mblk(9S).
For each frame, the driver should allocate the required resources and prepare it for being transmitted on the wire. The driver may opt to copy those resources to a DMA buffer or it may bind them. For more information on these options, see the MBLKS AND DMA section of mac(9E).
As it processes each frame in the chain, if the device driver has
advertised either of the MAC_CAPAB_HCKSUM
or
MAC_CAPAB_LSO
flags, it must check whether either
apply for the given frame using the
mac_hcksum_get(9F) and
mac_lso_get(9F) functions
respectively. If either is enabled for the given frame, the hardware must
arrange for that to be taken care of.
For each frame that the device driver processes it is responsible for doing one of three things with it:
The device driver is in charge of the memory associated with mp_chain. If the device driver does not return the message blocks to the MAC framework, then it must call freemsg(9F) on the frames. If it does not, the memory associated with them will be leaked. When a frame is being transmitted, if the device driver performed DMA binding, it should not free the message block until after it is guaranteed that the frame has been transmitted. If the message block was copied to a DMA buffer, then it is allowed to call freemsg(9F) at any point.
In general, the device driver should not drop frames without transmitting them unless it has no other choice. Times when this happens may include the device driver being in a state where it can't transmit, an error was found in the frame while trying to establish the checksum or LSO state, or some other kind of error that represents an issue with the passed frame.
The device driver should not free the chain when it does not have enough resources. For example, if entries in a device's descriptor ring fill up, then it should not drop those frames and instead should return all of the frames that were not transmitted. This indicates to the stack that the device is full and that flow control should be asserted. Back pressure will be applied to the rest of the stack, allowing most systems to behave better.
Once a device driver has returned unprocessed frames
from its
mc_tx
()
entry point, then the device driver will not receive any additional calls to
its mc_tx
() entry point until it calls the
mac_tx_update(9F) function to
indicate that resources are available again. Note that because it is the
device driver that is calling this function to indicate resources are
available, it is very important that it only return frames in cases where
the device driver itself will be notified that resources are available
again. For example, when it receives an interrupt indicating that the data
that it transmitted has been completed so it can use entries in its
descriptor ring or other data structures again.
The device driver can obtain access to its soft state through the driver member. It should cast it to the appropriate structure. The device driver should employ any necessary locking to access the transmit related data structures. Note that the device driver should expect that it may have its transmit endpoints called into from other threads while it's servicing device interrupts related to them.
The
mri_tx
()
entry point is similar to the mc_tx
() entry point,
except that it is used by device drivers that have negotiated the
MAC_CAPAB_RINGS
capability with transmit rings. The
driver should follow all of the same rules described earlier, except that it
will access a ring-specific data structure through
driver_rh and when it needs to update that there is
additional space available, it must use
mac_tx_ring_update(9F)
and not mac_tx_update(9F).
When the
mri_tx
()
entry point is called, the ring that should be used has been specified. The
driver must not attempt to use any other ring than the one specified by
driver_rh for any reason, including a lack of
resources or an attempt to perform its own hashing.
The mc_tx
() entry point may be called from
kernel
or
interrupt
context.
Upon successful completion, the device driver should return
NULL
. Otherwise, it should return all unprocessed
message blocks and ensure that it calls either
mac_tx_update(9F) or
mac_tx_ring_update(9F)
some time in the future.
mac(9E), mac_capab_rings(9E), mr_rget(9E), freemsg(9F), mac_hcksum_get(9F), mac_lso_get(9F), mac_register(9F), mac_tx_ring_update(9F), mac_tx_update(9F), mac_register(9S), mac_ring_info(9S), mblk(9S)
December 11, 2022 | OmniOS |