xref: /illumos-gate/usr/src/uts/common/io/mac/mac_sched.c (revision bbf21555)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright 2010 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  * Copyright 2018 Joyent, Inc.
25  * Copyright 2013 Nexenta Systems, Inc. All rights reserved.
26  */
27 
28 /*
29  * MAC data path
30  *
31  * The MAC data path is concerned with the flow of traffic from mac clients --
32  * DLS, IP, etc. -- to various GLDv3 device drivers -- e1000g, vnic, aggr,
33  * ixgbe, etc. -- and from the GLDv3 device drivers back to clients.
34  *
35  * -----------
36  * Terminology
37  * -----------
38  *
39  * MAC uses a lot of different, but related terms that are associated with the
40  * design and structure of the data path. Before we cover other aspects, first
41  * let's review the terminology that MAC uses.
42  *
43  * MAC
44  *
45  *	This driver. It interfaces with device drivers and provides abstractions
46  *	that the rest of the system consumes. All data links -- things managed
47  *	with dladm(8), are accessed through MAC.
48  *
49  * GLDv3 DEVICE DRIVER
50  *
51  *	A GLDv3 device driver refers to a driver, both for pseudo-devices and
52  *	real devices, which implement the GLDv3 driver API. Common examples of
53  *	these are igb and ixgbe, which are drivers for various Intel networking
54  *	cards. These devices may or may not have various features, such as
55  *	hardware rings and checksum offloading. For MAC, a GLDv3 device is the
56  *	final point for the transmission of a packet and the starting point for
57  *	the receipt of a packet.
58  *
59  * FLOWS
60  *
61  *	At a high level, a flow refers to a series of packets that are related.
62  *	Often times the term is used in the context of TCP to indicate a unique
63  *	TCP connection and the traffic over it. However, a flow can exist at
64  *	other levels of the system as well. MAC has a notion of a default flow
65  *	which is used for all unicast traffic addressed to the address of a MAC
66  *	device. For example, when a VNIC is created, a default flow is created
67  *	for the VNIC's MAC address. In addition, flows are created for broadcast
68  *	groups and a user may create a flow with flowadm(8).
69  *
70  * CLASSIFICATION
71  *
72  *	Classification refers to the notion of identifying an incoming frame
73  *	based on its destination address and optionally its source addresses and
74  *	doing different processing based on that information. Classification can
75  *	be done in both hardware and software. In general, we usually only
76  *	classify based on the layer two destination, eg. for Ethernet, the
77  *	destination MAC address.
78  *
79  *	The system also will do classification based on layer three and layer
80  *	four properties. This is used to support things like flowadm(8), which
81  *	allows setting QoS and other properties on a per-flow basis.
82  *
83  * RING
84  *
85  *	Conceptually, a ring represents a series of framed messages, often in a
86  *	contiguous chunk of memory that acts as a circular buffer. Rings come in
87  *	a couple of forms. Generally they are either a hardware construct (hw
88  *	ring) or they are a software construct (sw ring) maintained by MAC.
89  *
90  * HW RING
91  *
92  *	A hardware ring is a set of resources provided by a GLDv3 device driver
93  *	(even if it is a pseudo-device). A hardware ring comes in two different
94  *	forms: receive (rx) rings and transmit (tx) rings. An rx hw ring is
95  *	something that has a unique DMA (direct memory access) region and
96  *	generally supports some form of classification (though it isn't always
97  *	used), as well as a means of generating an interrupt specific to that
98  *	ring. For example, the device may generate a specific MSI-X for a PCI
99  *	express device. A tx ring is similar, except that it is dedicated to
100  *	transmission. It may also be a vector for enabling features such as VLAN
101  *	tagging and large transmit offloading. It usually has its own dedicated
102  *	interrupts for transmit being completed.
103  *
104  * SW RING
105  *
106  *	A software ring is a construction of MAC. It represents the same thing
107  *	that a hardware ring generally does, a collection of frames. However,
108  *	instead of being in a contiguous ring of memory, they're instead linked
109  *	by using the mblk_t's b_next pointer. Each frame may itself be multiple
110  *	mblk_t's linked together by the b_cont pointer. A software ring always
111  *	represents a collection of classified packets; however, it varies as to
112  *	whether it uses only layer two information, or a combination of that and
113  *	additional layer three and layer four data.
114  *
115  * FANOUT
116  *
117  *	Fanout is the idea of spreading out the load of processing frames based
118  *	on the source and destination information contained in the layer two,
119  *	three, and four headers, such that the data can then be processed in
120  *	parallel using multiple hardware threads.
121  *
122  *	A fanout algorithm hashes the headers and uses that to place different
123  *	flows into a bucket. The most important thing is that packets that are
124  *	in the same flow end up in the same bucket. If they do not, performance
125  *	can be adversely affected. Consider the case of TCP.  TCP severely
126  *	penalizes a connection if the data arrives out of order. If a given flow
127  *	is processed on different CPUs, then the data will appear out of order,
128  *	hence the invariant that fanout always hash a given flow to the same
129  *	bucket and thus get processed on the same CPU.
130  *
131  * RECEIVE SIDE SCALING (RSS)
132  *
133  *
134  *	Receive side scaling is a term that isn't common in illumos, but is used
135  *	by vendors and was popularized by Microsoft. It refers to the idea of
136  *	spreading the incoming receive load out across multiple interrupts which
137  *	can be directed to different CPUs. This allows a device to leverage
138  *	hardware rings even when it doesn't support hardware classification. The
139  *	hardware uses an algorithm to perform fanout that ensures the flow
140  *	invariant is maintained.
141  *
142  * SOFT RING SET
143  *
144  *	A soft ring set, commonly abbreviated SRS, is a collection of rings and
145  *	is used for both transmitting and receiving. It is maintained in the
146  *	structure mac_soft_ring_set_t. A soft ring set is usually associated
147  *	with flows, and coordinates both the use of hardware and software rings.
148  *	Because the use of hardware rings can change as devices such as VNICs
149  *	come and go, we always ensure that the set has software classification
150  *	rules that correspond to the hardware classification rules from rings.
151  *
152  *	Soft ring sets are also used for the enforcement of various QoS
153  *	properties. For example, if a bandwidth limit has been placed on a
154  *	specific flow or device, then that will be enforced by the soft ring
155  *	set.
156  *
157  * SERVICE ATTACHMENT POINT (SAP)
158  *
159  *	The service attachment point is a DLPI (Data Link Provider Interface)
160  *	concept; however, it comes up quite often in MAC. Most MAC devices speak
161  *	a protocol that has some notion of different channels or message type
162  *	identifiers. For example, Ethernet defines an EtherType which is a part
163  *	of the Ethernet header and defines the particular protocol of the data
164  *	payload. If the EtherType is set to 0x0800, then it defines that the
165  *	contents of that Ethernet frame is IPv4 traffic. For Ethernet, the
166  *	EtherType is the SAP.
167  *
168  *	In DLPI, a given consumer attaches to a specific SAP. In illumos, the ip
169  *	and arp drivers attach to the EtherTypes for IPv4, IPv6, and ARP. Using
170  *	libdlpi(3LIB) user software can attach to arbitrary SAPs. With the
171  *	exception of 802.1Q VLAN tagged traffic, MAC itself does not directly
172  *	consume the SAP; however, it uses that information as part of hashing
173  *	and it may be used as part of the construction of flows.
174  *
175  * PRIMARY MAC CLIENT
176  *
177  *	The primary mac client refers to a mac client whose unicast address
178  *	matches the address of the device itself. For example, if the system has
179  *	instance of the e1000g driver such as e1000g0, e1000g1, etc., the
180  *	primary mac client is the one named after the device itself. VNICs that
181  *	are created on top of such devices are not the primary client.
182  *
183  * TRANSMIT DESCRIPTORS
184  *
185  *	Transmit descriptors are a resource that most GLDv3 device drivers have.
186  *	Generally, a GLDv3 device driver takes a frame that's meant to be output
187  *	and puts a copy of it into a region of memory. Each region of memory
188  *	usually has an associated descriptor that the device uses to manage
189  *	properties of the frames. Devices have a limited number of such
190  *	descriptors. They get reclaimed once the device finishes putting the
191  *	frame on the wire.
192  *
193  *	If the driver runs out of transmit descriptors, for example, the OS is
194  *	generating more frames than it can put on the wire, then it will return
195  *	them back to the MAC layer.
196  *
197  * ---------------------------------
198  * Rings, Classification, and Fanout
199  * ---------------------------------
200  *
201  * The heart of MAC is made up of rings, and not those that Elven-kings wear.
202  * When receiving a packet, MAC breaks the work into two different, though
203  * interrelated phases. The first phase is generally classification and then the
204  * second phase is generally fanout. When a frame comes in from a GLDv3 Device,
205  * MAC needs to determine where that frame should be delivered. If it's a
206  * unicast frame (say a normal TCP/IP packet), then it will be delivered to a
207  * single MAC client; however, if it's a broadcast or multicast frame, then MAC
208  * may need to deliver it to multiple MAC clients.
209  *
210  * On transmit, classification isn't quite as important, but may still be used.
211  * Unlike with the receive path, the classification is not used to determine
212  * devices that should transmit something, but rather is used for special
213  * properties of a flow, eg. bandwidth limits for a given IP address, device, or
214  * connection.
215  *
216  * MAC employs a software classifier and leverages hardware classification as
217  * well. The software classifier can leverage the full layer two information,
218  * source, destination, VLAN, and SAP. If the SAP indicates that IP traffic is
219  * being sent, it can classify based on the IP header, and finally, it also
220  * knows how to classify based on the local and remote ports of TCP, UDP, and
221  * SCTP.
222  *
223  * Hardware classifiers vary in capability. Generally all hardware classifiers
224  * provide the capability to classify based on the destination MAC address. Some
225  * hardware has additional filters built in for performing more in-depth
226  * classification; however, it often has much more limited resources for these
227  * activities as compared to the layer two destination address classification.
228  *
229  * The modus operandi in MAC is to always ensure that we have software-based
230  * capabilities and rules in place and then to supplement that with hardware
231  * resources when available. In general, simple layer two classification is
232  * sufficient and nothing else is used, unless a specific flow is created with
233  * tools such as flowadm(8) or bandwidth limits are set on a device with
234  * dladm(8).
235  *
236  * RINGS AND GROUPS
237  *
238  * To get into how rings and classification play together, it's first important
239  * to understand how hardware devices commonly associate rings and allow them to
240  * be programmed. Recall that a hardware ring should be thought of as a DMA
241  * buffer and an interrupt resource. Rings are then collected into groups. A
242  * group itself has a series of classification rules. One or more MAC addresses
243  * are assigned to a group.
244  *
245  * Hardware devices vary in terms of what capabilities they provide. Sometimes
246  * they allow for a dynamic assignment of rings to a group and sometimes they
247  * have a static assignment of rings to a group. For example, the ixgbe driver
248  * has a static assignment of rings to groups such that every group has exactly
249  * one ring and the number of groups is equal to the number of rings.
250  *
251  * Classification and receive side scaling both come into play with how a device
252  * advertises itself to MAC and how MAC uses it. If a device supports layer two
253  * classification of frames, then MAC will assign MAC addresses to a group as a
254  * form of primary classification. If a single MAC address is assigned to a
255  * group, a common case, then MAC will consider packets that come in from rings
256  * on that group to be fully classified and will not need to do any software
257  * classification unless a specific flow has been created.
258  *
259  * If a device supports receive side scaling, then it may advertise or support
260  * groups with multiple rings. In those cases, then receive side scaling will
261  * come into play and MAC will use that as a means of fanning out received
262  * frames across multiple CPUs. This can also be combined with groups that
263  * support layer two classification.
264  *
265  * If a device supports dynamic assignments of rings to groups, then MAC will
266  * change around the way that rings are assigned to various groups as devices
267  * come and go from the system. For example, when a VNIC is created, a new flow
268  * will be created for the VNIC's MAC address. If a hardware ring is available,
269  * MAC may opt to reassign it from one group to another.
270  *
271  * ASSIGNMENT OF HARDWARE RINGS
272  *
273  * This is a bit of a complicated subject that varies depending on the device,
274  * the use of aggregations, the special nature of the primary mac client. This
275  * section deserves being fleshed out.
276  *
277  * FANOUT
278  *
279  * illumos uses fanout to help spread out the incoming processing load of chains
280  * of frames away from a single CPU. If a device supports receive side scaling,
281  * then that provides an initial form of fanout; however, what we're concerned
282  * with all happens after the context of a given set of frames being classified
283  * to a soft ring set.
284  *
285  * After frames reach a soft ring set and account for any potential bandwidth
286  * related accounting, they may be fanned out based on one of the following
287  * three modes:
288  *
289  *     o No Fanout
290  *     o Protocol level fanout
291  *     o Full software ring protocol fanout
292  *
293  * MAC makes the determination as to which of these modes a given soft ring set
294  * obtains based on parameters such as whether or not it's the primary mac
295  * client, whether it's on a 10 GbE or faster device, user controlled dladm(8)
296  * properties, and the nature of the hardware and the resources that it has.
297  *
298  * When there is no fanout, MAC does not create any soft rings for a device and
299  * the device has frames delivered directly to the MAC client.
300  *
301  * Otherwise, all fanout is performed by software. MAC divides incoming frames
302  * into one of three buckets -- IPv4 TCP traffic, IPv4 UDP traffic, and
303  * everything else. Regardless of the type of fanout, these three categories
304  * or buckets are always used.
305  *
306  * The difference between protocol level fanout and full software ring protocol
307  * fanout is the number of software rings that end up getting created. The
308  * system always uses the same number of software rings per protocol bucket. So
309  * in the first case when we're just doing protocol level fanout, we just create
310  * one software ring each for IPv4 TCP traffic, IPv4 UDP traffic, and everything
311  * else.
312  *
313  * In the case where we do full software ring protocol fanout, we generally use
314  * mac_compute_soft_ring_count() to determine the number of rings. There are
315  * other combinations of properties and devices that may send us down other
316  * paths, but this is a common starting point. If it's a non-bandwidth enforced
317  * device and we're on at least a 10 GbE link, then we'll use eight soft rings
318  * per protocol bucket as a starting point. See mac_compute_soft_ring_count()
319  * for more information on the total number.
320  *
321  * For each of these rings, we create a mac_soft_ring_t and an associated worker
322  * thread. Particularly when doing full software ring protocol fanout, we bind
323  * each of the worker threads to individual CPUs.
324  *
325  * The other advantage of these software rings is that it allows upper layers to
326  * optionally poll on them. For example, TCP can leverage an squeue to poll on
327  * the software ring, see squeue.c for more information.
328  *
329  * DLS BYPASS
330  *
331  * DLS is the data link services module. It interfaces with DLPI, which is the
332  * primary way that other parts of the system such as IP interface with the MAC
333  * layer. While DLS is traditionally a STREAMS-based interface, it allows for
334  * certain modules such as IP to negotiate various more modern interfaces to be
335  * used, which are useful for higher performance and allow it to use direct
336  * function calls to DLS instead of using STREAMS.
337  *
338  * When we have IPv4 TCP or UDP software rings, then traffic on those rings is
339  * eligible for what we call the dls bypass. In those cases, rather than going
340  * out mac_rx_deliver() to DLS, DLS instead registers them to go directly via
341  * the direct callback registered with DLS, generally ip_input().
342  *
343  * HARDWARE RING POLLING
344  *
345  * GLDv3 devices with hardware rings generally deliver chains of messages
346  * (mblk_t chain) during the context of a single interrupt. However, interrupts
347  * are not the only way that these devices may be used. As part of implementing
348  * ring support, a GLDv3 device driver must have a way to disable the generation
349  * of that interrupt and allow for the operating system to poll on that ring.
350  *
351  * To implement this, every soft ring set has a worker thread and a polling
352  * thread. If a sufficient packet rate comes into the system, MAC will 'blank'
353  * (disable) interrupts on that specific ring and the polling thread will start
354  * consuming packets from the hardware device and deliver them to the soft ring
355  * set, where the worker thread will take over.
356  *
357  * Once the rate of packet intake drops down below a certain threshold, then
358  * polling on the hardware ring will be quiesced and interrupts will be
359  * re-enabled for the given ring. This effectively allows the system to shift
360  * how it handles a ring based on its load. At high packet rates, polling on the
361  * device as opposed to relying on interrupts can actually reduce overall system
362  * load due to the minimization of interrupt activity.
363  *
364  * Note the importance of each ring having its own interrupt source. The whole
365  * idea here is that we do not disable interrupts on the device as a whole, but
366  * rather each ring can be independently toggled.
367  *
368  * USE OF WORKER THREADS
369  *
370  * Both the soft ring set and individual soft rings have a worker thread
371  * associated with them that may be bound to a specific CPU in the system. Any
372  * such assignment will get reassessed as part of dynamic reconfiguration events
373  * in the system such as the onlining and offlining of CPUs and the creation of
374  * CPU partitions.
375  *
376  * In many cases, while in an interrupt, we try to deliver a frame all the way
377  * through the stack in the context of the interrupt itself. However, if the
378  * amount of queued frames has exceeded a threshold, then we instead defer to
379  * the worker thread to do this work and signal it. This is particularly useful
380  * when you have the soft ring set delivering frames into multiple software
381  * rings. If it was only delivering frames into a single software ring then
382  * there'd be no need to have another thread take over. However, if it's
383  * delivering chains of frames to multiple rings, then it's worthwhile to have
384  * the worker for the software ring take over so that the different software
385  * rings can be processed in parallel.
386  *
387  * In a similar fashion to the hardware polling thread, if we don't have a
388  * backlog or there's nothing to do, then the worker thread will go back to
389  * sleep and frames can be delivered all the way from an interrupt. This
390  * behavior is useful as it's designed to minimize latency and the default
391  * disposition of MAC is to optimize for latency.
392  *
393  * MAINTAINING CHAINS
394  *
395  * Another useful idea that MAC uses is to try and maintain frames in chains for
396  * as long as possible. The idea is that all of MAC can handle chains of frames
397  * structured as a series of mblk_t structures linked with the b_next pointer.
398  * When performing software classification and software fanout, MAC does not
399  * simply determine the destination and send the frame along. Instead, in the
400  * case of classification, it tries to maintain a chain for as long as possible
401  * before passing it along and performing additional processing.
402  *
403  * In the case of fanout, MAC first determines what the target software ring is
404  * for every frame in the original chain and constructs a new chain for each
405  * target. MAC then delivers the new chain to each software ring in succession.
406  *
407  * The whole rationale for doing this is that we want to try and maintain the
408  * pipe as much as possible and deliver as many frames through the stack at once
409  * that we can, rather than just pushing a single frame through. This can often
410  * help bring down latency and allows MAC to get a better sense of the overall
411  * activity in the system and properly engage worker threads.
412  *
413  * --------------------
414  * Bandwidth Management
415  * --------------------
416  *
417  * Bandwidth management is something that's built into the soft ring set itself.
418  * When bandwidth limits are placed on a flow, a corresponding soft ring set is
419  * toggled into bandwidth mode. This changes how we transmit and receive the
420  * frames in question.
421  *
422  * Bandwidth management is done on a per-tick basis. We translate the user's
423  * requested bandwidth from a quantity per-second into a quantity per-tick. MAC
424  * cannot process a frame across more than one tick, thus it sets a lower bound
425  * for the bandwidth cap to be a single MTU. This also means that when
426  * hires ticks are enabled (hz is set to 1000), that the minimum amount of
427  * bandwidth is higher, because the number of ticks has increased and MAC has to
428  * go from accepting 100 packets / sec to 1000 / sec.
429  *
430  * The bandwidth counter is reset by either the soft ring set's worker thread or
431  * a thread that is doing an inline transmit or receive if they discover that
432  * the current tick is in the future from the recorded tick.
433  *
434  * Whenever we're receiving or transmitting data, we end up leaving most of the
435  * work to the soft ring set's worker thread. This forces data inserted into the
436  * soft ring set to be effectively serialized and allows us to exhume bandwidth
437  * at a reasonable rate. If there is nothing in the soft ring set at the moment
438  * and the set has available bandwidth, then it may processed inline.
439  * Otherwise, the worker is responsible for taking care of the soft ring set.
440  *
441  * ---------------------
442  * The Receive Data Path
443  * ---------------------
444  *
445  * The following series of ASCII art images breaks apart the way that a frame
446  * comes in and is processed in MAC.
447  *
448  * Part 1 -- Initial frame receipt, SRS classification
449  *
450  * Here, a frame is received by a GLDv3 driver, generally in the context of an
451  * interrupt, and it ends up in mac_rx_common(). A driver calls either mac_rx or
452  * mac_rx_ring, depending on whether or not it supports rings and can identify
453  * the interrupt as having come from a specific ring. Here we determine whether
454  * or not it's fully classified and perform software classification as
455  * appropriate. From here, everything always ends up going to either entry [A]
456  * or entry [B] based on whether or not they have subflow processing needed. We
457  * leave via fanout or delivery.
458  *
459  *           +===========+
460  *           v hardware  v
461  *           v interrupt v
462  *           +===========+
463  *                 |
464  *                 * . . appropriate
465  *                 |     upcall made
466  *                 |     by GLDv3 driver  . . always
467  *                 |                      .
468  *  +--------+     |     +----------+     .    +---------------+
469  *  | GLDv3  |     +---->| mac_rx   |-----*--->| mac_rx_common |
470  *  | Driver |-->--+     +----------+          +---------------+
471  *  +--------+     |        ^                         |
472  *      |          |        ^                         v
473  *      ^          |        * . . always   +----------------------+
474  *      |          |        |              | mac_promisc_dispatch |
475  *      |          |    +-------------+    +----------------------+
476  *      |          +--->| mac_rx_ring |               |
477  *      |               +-------------+               * . . hw classified
478  *      |                                             v     or single flow?
479  *      |                                             |
480  *      |                                   +--------++--------------+
481  *      |                                   |        |               * hw class,
482  *      |                                   |        * hw classified | subflows
483  *      |                 no hw class and . *        | or single     | exist
484  *      |                 subflows          |        | flow          |
485  *      |                                   |        v               v
486  *      |                                   |   +-----------+   +-----------+
487  *      |                                   |   |   goto    |   |  goto     |
488  *      |                                   |   | entry [A] |   | entry [B] |
489  *      |                                   |   +-----------+   +-----------+
490  *      |                                   v          ^
491  *      |                            +-------------+   |
492  *      |                            | mac_rx_flow |   * SRS and flow found,
493  *      |                            +-------------+   | call flow cb
494  *      |                                   |          +------+
495  *      |                                   v                 |
496  *      v                             +==========+    +-----------------+
497  *      |                             v For each v--->| mac_rx_classify |
498  * +----------+                       v  mblk_t  v    +-----------------+
499  * |   srs    |                       +==========+
500  * | pollling |
501  * |  thread  |->------------------------------------------+
502  * +----------+                                            |
503  *                                                         v       . inline
504  *            +--------------------+   +----------+   +---------+  .
505  *    [A]---->| mac_rx_srs_process |-->| check bw |-->| enqueue |--*---------+
506  *            +--------------------+   |  limits  |   | frames  |            |
507  *               ^                     +----------+   | to SRS  |            |
508  *               |                                    +---------+            |
509  *               |  send chain              +--------+    |                  |
510  *               *  when clasified          | signal |    * BW limits,       |
511  *               |  flow changes            |  srs   |<---+ loopback,        |
512  *               |                          | worker |      stack too        |
513  *               |                          +--------+      deep             |
514  *      +-----------------+        +--------+                                |
515  *      | mac_flow_lookup |        |  srs   |     +---------------------+    |
516  *      +-----------------+        | worker |---->| mac_rx_srs_drain    |<---+
517  *               ^                 | thread |     | mac_rx_srs_drain_bw |
518  *               |                 +--------+     +---------------------+
519  *               |                                          |
520  *         +----------------------------+                   * software rings
521  *   [B]-->| mac_rx_srs_subflow_process |                   | for fanout?
522  *         +----------------------------+                   |
523  *                                               +----------+-----------+
524  *                                               |                      |
525  *                                               v                      v
526  *                                          +--------+             +--------+
527  *                                          |  goto  |             |  goto  |
528  *                                          | Part 2 |             | Part 3 |
529  *                                          +--------+             +--------+
530  *
531  * Part 2 -- Fanout
532  *
533  * This part is concerned with using software fanout to assign frames to
534  * software rings and then deliver them to MAC clients or allow those rings to
535  * be polled upon. While there are two different primary fanout entry points,
536  * mac_rx_fanout and mac_rx_proto_fanout, they behave in similar ways, and aside
537  * from some of the individual hashing techniques used, most of the general
538  * flow is the same.
539  *
540  *  +--------+              +-------------------+
541  *  |  From  |---+--------->| mac_rx_srs_fanout |----+
542  *  | Part 1 |   |          +-------------------+    |    +=================+
543  *  +--------+   |                                   |    v for each mblk_t v
544  *               * . . protocol only                 +--->v assign to new   v
545  *               |     fanout                        |    v chain based on  v
546  *               |                                   |    v hash % nrings   v
547  *               |    +-------------------------+    |    +=================+
548  *               +--->| mac_rx_srs_proto_fanout |----+             |
549  *                    +-------------------------+                  |
550  *                                                                 v
551  *    +------------+    +--------------------------+       +================+
552  *    | enqueue in |<---| mac_rx_soft_ring_process |<------v for each chain v
553  *    | soft ring  |    +--------------------------+       +================+
554  *    +------------+
555  *         |                                    +-----------+
556  *         * soft ring set                      | soft ring |
557  *         | empty and no                       |  worker   |
558  *         | worker?                            |  thread   |
559  *         |                                    +-----------+
560  *         +------*----------------+                  |
561  *         |      .                |                  v
562  *    No . *      . Yes            |       +------------------------+
563  *         |                       +----<--| mac_rx_soft_ring_drain |
564  *         |                       |       +------------------------+
565  *         v                       |
566  *   +-----------+                 v
567  *   |   signal  |         +---------------+
568  *   | soft ring |         | Deliver chain |
569  *   |   worker  |         | goto Part 3   |
570  *   +-----------+         +---------------+
571  *
572  *
573  * Part 3 -- Packet Delivery
574  *
575  * Here, we go through and deliver the mblk_t chain directly to a given
576  * processing function. In a lot of cases this is mac_rx_deliver(). In the case
577  * of DLS bypass being used, then instead we end up going ahead and deliver it
578  * to the direct callback registered with DLS, generally ip_input.
579  *
580  *
581  *   +---------+            +----------------+    +------------------+
582  *   |  From   |---+------->| mac_rx_deliver |--->| Off to DLS, or   |
583  *   | Parts 1 |   |        +----------------+    | other MAC client |
584  *   |  and 2  |   * DLS bypass                   +------------------+
585  *   +---------+   | enabled   +----------+    +-------------+
586  *                 +---------->| ip_input |--->|    To IP    |
587  *                             +----------+    | and beyond! |
588  *                                             +-------------+
589  *
590  * ----------------------
591  * The Transmit Data Path
592  * ----------------------
593  *
594  * Before we go into the images, it's worth talking about a problem that is a
595  * bit different from the receive data path. GLDv3 device drivers have a finite
596  * amount of transmit descriptors. When they run out, they return unused frames
597  * back to MAC. MAC, at this point has several options about what it will do,
598  * which vary based upon the settings that the client uses.
599  *
600  * When a device runs out of descriptors, the next thing that MAC does is
601  * enqueue them off of the soft ring set or a software ring, depending on the
602  * configuration of the soft ring set. MAC will enqueue up to a high watermark
603  * of mblk_t chains, at which point it will indicate flow control back to the
604  * client. Once this condition is reached, any mblk_t chains that were not
605  * enqueued will be returned to the caller and they will have to decide what to
606  * do with them. There are various flags that control this behavior that a
607  * client may pass, which are discussed below.
608  *
609  * When this condition is hit, MAC also returns a cookie to the client in
610  * addition to unconsumed frames. Clients can poll on that cookie and register a
611  * callback with MAC to be notified when they are no longer subject to flow
612  * control, at which point they may continue to call mac_tx(). This flow control
613  * actually manages to work itself all the way up the stack, back through dls,
614  * to ip, through the various protocols, and to sockfs.
615  *
616  * While the behavior described above is the default, this behavior can be
617  * modified. There are two alternate modes, described below, which are
618  * controlled with flags.
619  *
620  * DROP MODE
621  *
622  * This mode is controlled by having the client pass the MAC_DROP_ON_NO_DESC
623  * flag. When this is passed, if a device driver runs out of transmit
624  * descriptors, then the MAC layer will drop any unsent traffic. The client in
625  * this case will never have any frames returned to it.
626  *
627  * DON'T ENQUEUE
628  *
629  * This mode is controlled by having the client pass the MAC_TX_NO_ENQUEUE flag.
630  * If the MAC_DROP_ON_NO_DESC flag is also passed, it takes precedence. In this
631  * mode, when we hit a case where a driver runs out of transmit descriptors,
632  * then instead of enqueuing packets in a soft ring set or software ring, we
633  * instead return the mblk_t chain back to the caller and immediately put the
634  * soft ring set into flow control mode.
635  *
636  * The following series of ASCII art images describe the transmit data path that
637  * MAC clients enter into based on calling into mac_tx(). A soft ring set has a
638  * transmission function associated with it. There are seven possible
639  * transmission modes, some of which share function entry points. The one that a
640  * soft ring set gets depends on properties such as whether there are
641  * transmission rings for fanout, whether the device involves aggregations,
642  * whether any bandwidth limits exist, etc.
643  *
644  *
645  * Part 1 -- Initial checks
646  *
647  *      * . called by
648  *      |   MAC clients
649  *      v                     . . No
650  *  +--------+  +-----------+ .   +-------------------+  +====================+
651  *  | mac_tx |->| device    |-*-->| mac_protect_check |->v Is this the simple v
652  *  +--------+  | quiesced? |     +-------------------+  v case? See [1]      v
653  *              +-----------+            |               +====================+
654  *                  * . Yes              * failed                 |
655  *                  v                    | frames                 |
656  *             +--------------+          |                +-------+---------+
657  *             | freemsgchain |<---------+          Yes . *            No . *
658  *             +--------------+                           v                 v
659  *                                                  +-----------+     +--------+
660  *                                                  |   goto    |     |  goto  |
661  *                                                  |  Part 2   |     | SRS TX |
662  *                                                  | Entry [A] |     |  func  |
663  *                                                  +-----------+     +--------+
664  *                                                        |                 |
665  *                                                        |                 v
666  *                                                        |           +--------+
667  *                                                        +---------->| return |
668  *                                                                    | cookie |
669  *                                                                    +--------+
670  *
671  * [1] The simple case refers to the SRS being configured with the
672  * SRS_TX_DEFAULT transmission mode, having a single mblk_t (not a chain), their
673  * being only a single active client, and not having a backlog in the srs.
674  *
675  *
676  * Part 2 -- The SRS transmission functions
677  *
678  * This part is a bit more complicated. The different transmission paths often
679  * leverage one another. In this case, we'll draw out the more common ones
680  * before the parts that depend upon them. Here, we're going to start with the
681  * workings of mac_tx_send() a common function that most of the others end up
682  * calling.
683  *
684  *      +-------------+
685  *      | mac_tx_send |
686  *      +-------------+
687  *            |
688  *            v
689  *      +=============+    +==============+
690  *      v  more than  v--->v    check     v
691  *      v one client? v    v VLAN and add v
692  *      +=============+    v  VLAN tags   v
693  *            |            +==============+
694  *            |                  |
695  *            +------------------+
696  *            |
697  *            |                 [A]
698  *            v                  |
699  *       +============+ . No     v
700  *       v more than  v .     +==========+     +--------------------------+
701  *       v one active v-*---->v for each v---->| mac_promisc_dispatch_one |---+
702  *       v  client?   v       v mblk_t   v     +--------------------------+   |
703  *       +============+       +==========+        ^                           |
704  *            |                                   |       +==========+        |
705  *            * . Yes                             |       v hardware v<-------+
706  *            v                      +------------+       v  rings?  v
707  *       +==========+                |                    +==========+
708  *       v for each v       No . . . *                         |
709  *       v mblk_t   v       specific |                         |
710  *       +==========+       flow     |                   +-----+-----+
711  *            |                      |                   |           |
712  *            v                      |                   v           v
713  *    +-----------------+            |               +-------+  +---------+
714  *    | mac_tx_classify |------------+               | GLDv3 |  |  GLDv3  |
715  *    +-----------------+                            |TX func|  | ring tx |
716  *            |                                      +-------+  |  func   |
717  *            * Specific flow, generally                 |      +---------+
718  *            | bcast, mcast, loopback                   |           |
719  *            v                                          +-----+-----+
720  *      +==========+       +---------+                         |
721  *      v valid L2 v--*--->| freemsg |                         v
722  *      v  header  v  . No +---------+               +-------------------+
723  *      +==========+                                 | return unconsumed |
724  *            * . Yes                                |   frames to the   |
725  *            v                                      |      caller       |
726  *      +===========+                                +-------------------+
727  *      v braodcast v      +----------------+                  ^
728  *      v   flow?   v--*-->| mac_bcast_send |------------------+
729  *      +===========+  .   +----------------+                  |
730  *            |        . . Yes                                 |
731  *       No . *                                                v
732  *            |  +---------------------+  +---------------+  +----------+
733  *            +->|mac_promisc_dispatch |->| mac_fix_cksum |->|   flow   |
734  *               +---------------------+  +---------------+  | callback |
735  *                                                           +----------+
736  *
737  *
738  * In addition, many but not all of the routines, all rely on
739  * mac_tx_softring_process as an entry point.
740  *
741  *
742  *                                           . No             . No
743  * +--------------------------+   +========+ .  +===========+ .  +-------------+
744  * | mac_tx_soft_ring_process |-->v worker v-*->v out of tx v-*->|    goto     |
745  * +--------------------------+   v only?  v    v  descr.?  v    | mac_tx_send |
746  *                                +========+    +===========+    +-------------+
747  *                              Yes . *               * . Yes           |
748  *                   . No             v               |                 v
749  *     v=========+   .          +===========+ . Yes   |     Yes .  +==========+
750  *     v apppend v<--*----------v out of tx v-*-------+---------*--v returned v
751  *     v mblk_t  v              v  descr.?  v         |            v frames?  v
752  *     v chain   v              +===========+         |            +==========+
753  *     +=========+                                    |                 *. No
754  *         |                                          |                 v
755  *         v                                          v           +------------+
756  * +===================+           +----------------------+       |   done     |
757  * v worker scheduled? v           | mac_tx_sring_enqueue |       | processing |
758  * v Out of tx descr?  v           +----------------------+       +------------+
759  * +===================+                      |
760  *    |           |           . Yes           v
761  *    * Yes       * No        .         +============+
762  *    |           v         +-*---------v drop on no v
763  *    |      +========+     v           v  TX desc?  v
764  *    |      v  wake  v  +----------+   +============+
765  *    |      v worker v  | mac_pkt_ |         * . No
766  *    |      +========+  | drop     |         |         . Yes         . No
767  *    |           |      +----------+         v         .             .
768  *    |           |         v   ^     +===============+ .  +========+ .
769  *    +--+--------+---------+   |     v Don't enqueue v-*->v ring   v-*----+
770  *       |                      |     v     Set?      v    v empty? v      |
771  *       |      +---------------+     +===============+    +========+      |
772  *       |      |                            |                |            |
773  *       |      |        +-------------------+                |            |
774  *       |      *. Yes   |                          +---------+            |
775  *       |      |        v                          v                      v
776  *       |      |  +===========+               +========+      +--------------+
777  *       |      +<-v At hiwat? v               v append v      |    return    |
778  *       |         +===========+               v mblk_t v      | mblk_t chain |
779  *       |                  * No               v chain  v      |   and flow   |
780  *       |                  v                  +========+      |    control   |
781  *       |               +=========+                |          |    cookie    |
782  *       |               v  append v                v          +--------------+
783  *       |               v  mblk_t v           +========+
784  *       |               v  chain  v           v  wake  v   +------------+
785  *       |               +=========+           v worker v-->|    done    |
786  *       |                    |                +========+   | processing |
787  *       |                    v       .. Yes                +------------+
788  *       |               +=========+  .   +========+
789  *       |               v  first  v--*-->v  wake  v
790  *       |               v append? v      v worker v
791  *       |               +=========+      +========+
792  *       |                   |                |
793  *       |              No . *                |
794  *       |                   v                |
795  *       |       +--------------+             |
796  *       +------>|   Return     |             |
797  *               | flow control |<------------+
798  *               |   cookie     |
799  *               +--------------+
800  *
801  *
802  * The remaining images are all specific to each of the different transmission
803  * modes.
804  *
805  * SRS TX DEFAULT
806  *
807  *      [ From Part 1 ]
808  *             |
809  *             v
810  * +-------------------------+
811  * | mac_tx_single_ring_mode |
812  * +-------------------------+
813  *            |
814  *            |       . Yes
815  *            v       .
816  *       +==========+ .  +============+
817  *       v   SRS    v-*->v   Try to   v---->---------------------+
818  *       v backlog? v    v enqueue in v                          |
819  *       +==========+    v     SRS    v-->------+                * . . Queue too
820  *            |          +============+         * don't enqueue  |     deep or
821  *            * . No         ^     |            | flag or at     |     drop flag
822  *            |              |     v            | hiwat,         |
823  *            v              |     |            | return    +---------+
824  *     +-------------+       |     |            | cookie    | freemsg |
825  *     |    goto     |-*-----+     |            |           +---------+
826  *     | mac_tx_send | . returned  |            |                |
827  *     +-------------+   mblk_t    |            |                |
828  *            |                    |            |                |
829  *            |                    |            |                |
830  *            * . . all mblk_t     * queued,    |                |
831  *            v     consumed       | may return |                |
832  *     +-------------+             | tx cookie  |                |
833  *     | SRS TX func |<------------+------------+----------------+
834  *     |  completed  |
835  *     +-------------+
836  *
837  * SRS_TX_SERIALIZE
838  *
839  *   +------------------------+
840  *   | mac_tx_serializer_mode |
841  *   +------------------------+
842  *               |
843  *               |        . No
844  *               v        .
845  *         +============+ .  +============+    +-------------+   +============+
846  *         v srs being  v-*->v  set SRS   v--->|    goto     |-->v remove SRS v
847  *         v processed? v    v proc flags v    | mac_tx_send |   v proc flag  v
848  *         +============+    +============+    +-------------+   +============+
849  *               |                                                     |
850  *               * Yes                                                 |
851  *               v                                       . No          v
852  *      +--------------------+                           .        +==========+
853  *      | mac_tx_srs_enqueue |  +------------------------*-----<--v returned v
854  *      +--------------------+  |                                 v frames?  v
855  *               |              |   . Yes                         +==========+
856  *               |              |   .                                  |
857  *               |              |   . +=========+                      v
858  *               v              +-<-*-v queued  v     +--------------------+
859  *        +-------------+       |     v frames? v<----| mac_tx_srs_enqueue |
860  *        | SRS TX func |       |     +=========+     +--------------------+
861  *        | completed,  |<------+         * . Yes
862  *        | may return  |       |         v
863  *        |   cookie    |       |     +========+
864  *        +-------------+       +-<---v  wake  v
865  *                                    v worker v
866  *                                    +========+
867  *
868  *
869  * SRS_TX_FANOUT
870  *
871  *                                             . Yes
872  *   +--------------------+    +=============+ .   +--------------------------+
873  *   | mac_tx_fanout_mode |--->v Have fanout v-*-->|           goto           |
874  *   +--------------------+    v   hint?     v     | mac_rx_soft_ring_process |
875  *                             +=============+     +--------------------------+
876  *                                   * . No                    |
877  *                                   v                         ^
878  *                             +===========+                   |
879  *                        +--->v for each  v           +===============+
880  *                        |    v   mblk_t  v           v pick softring v
881  *                 same   *    +===========+           v   from hash   v
882  *                 hash   |          |                 +===============+
883  *                        |          v                         |
884  *                        |   +--------------+                 |
885  *                        +---| mac_pkt_hash |--->*------------+
886  *                            +--------------+    . different
887  *                                                  hash or
888  *                                                  done proc.
889  * SRS_TX_AGGR                                      chain
890  *
891  *   +------------------+    +================================+
892  *   | mac_tx_aggr_mode |--->v Use aggr capab function to     v
893  *   +------------------+    v find appropriate tx ring.      v
894  *                           v Applies hash based on aggr     v
895  *                           v policy, see mac_tx_aggr_mode() v
896  *                           +================================+
897  *                                          |
898  *                                          v
899  *                           +-------------------------------+
900  *                           |            goto               |
901  *                           |  mac_rx_srs_soft_ring_process |
902  *                           +-------------------------------+
903  *
904  *
905  * SRS_TX_BW, SRS_TX_BW_FANOUT, SRS_TX_BW_AGGR
906  *
907  * Note, all three of these tx functions start from the same place --
908  * mac_tx_bw_mode().
909  *
910  *  +----------------+
911  *  | mac_tx_bw_mode |
912  *  +----------------+
913  *         |
914  *         v          . No               . No               . Yes
915  *  +==============+  .  +============+  .  +=============+ .  +=========+
916  *  v  Out of BW?  v--*->v SRS empty? v--*->v  reset BW   v-*->v Bump BW v
917  *  +==============+     +============+     v tick count? v    v Usage   v
918  *         |                   |            +=============+    +=========+
919  *         |         +---------+                   |                |
920  *         |         |        +--------------------+                |
921  *         |         |        |              +----------------------+
922  *         v         |        v              v
923  * +===============+ |  +==========+   +==========+      +------------------+
924  * v Don't enqueue v |  v  set bw  v   v Is aggr? v--*-->|       goto       |
925  * v   flag set?   v |  v enforced v   +==========+  .   | mac_tx_aggr_mode |-+
926  * +===============+ |  +==========+         |       .   +------------------+ |
927  *   |    Yes .*     |        |         No . *       .                        |
928  *   |         |     |        |              |       . Yes                    |
929  *   * . No    |     |        v              |                                |
930  *   |  +---------+  |   +========+          v              +======+          |
931  *   |  | freemsg |  |   v append v   +============+  . Yes v pick v          |
932  *   |  +---------+  |   v mblk_t v   v Is fanout? v--*---->v ring v          |
933  *   |      |        |   v chain  v   +============+        +======+          |
934  *   +------+        |   +========+          |                  |             |
935  *          v        |        |              v                  v             |
936  *    +---------+    |        v       +-------------+ +--------------------+  |
937  *    | return  |    |   +========+   |    goto     | |       goto         |  |
938  *    |  flow   |    |   v wakeup v   | mac_tx_send | | mac_tx_fanout_mode |  |
939  *    | control |    |   v worker v   +-------------+ +--------------------+  |
940  *    | cookie  |    |   +========+          |                  |             |
941  *    +---------+    |        |              |                  +------+------+
942  *                   |        v              |                         |
943  *                   |   +---------+         |                         v
944  *                   |   | return  |   +============+           +------------+
945  *                   |   |  flow   |   v unconsumed v-------+   |   done     |
946  *                   |   | control |   v   frames?  v       |   | processing |
947  *                   |   | cookie  |   +============+       |   +------------+
948  *                   |   +---------+         |              |
949  *                   |                  Yes  *              |
950  *                   |                       |              |
951  *                   |                 +===========+        |
952  *                   |                 v subtract  v        |
953  *                   |                 v unused bw v        |
954  *                   |                 +===========+        |
955  *                   |                       |              |
956  *                   |                       v              |
957  *                   |              +--------------------+  |
958  *                   +------------->| mac_tx_srs_enqueue |  |
959  *                                  +--------------------+  |
960  *                                           |              |
961  *                                           |              |
962  *                                     +------------+       |
963  *                                     |  return fc |       |
964  *                                     | cookie and |<------+
965  *                                     |    mblk_t  |
966  *                                     +------------+
967  */
968 
969 #include <sys/types.h>
970 #include <sys/callb.h>
971 #include <sys/pattr.h>
972 #include <sys/sdt.h>
973 #include <sys/strsubr.h>
974 #include <sys/strsun.h>
975 #include <sys/vlan.h>
976 #include <sys/stack.h>
977 #include <sys/archsystm.h>
978 #include <inet/ipsec_impl.h>
979 #include <inet/ip_impl.h>
980 #include <inet/sadb.h>
981 #include <inet/ipsecesp.h>
982 #include <inet/ipsecah.h>
983 #include <inet/ip6.h>
984 
985 #include <sys/mac_impl.h>
986 #include <sys/mac_client_impl.h>
987 #include <sys/mac_client_priv.h>
988 #include <sys/mac_soft_ring.h>
989 #include <sys/mac_flow_impl.h>
990 
991 static mac_tx_cookie_t mac_tx_single_ring_mode(mac_soft_ring_set_t *, mblk_t *,
992     uintptr_t, uint16_t, mblk_t **);
993 static mac_tx_cookie_t mac_tx_serializer_mode(mac_soft_ring_set_t *, mblk_t *,
994     uintptr_t, uint16_t, mblk_t **);
995 static mac_tx_cookie_t mac_tx_fanout_mode(mac_soft_ring_set_t *, mblk_t *,
996     uintptr_t, uint16_t, mblk_t **);
997 static mac_tx_cookie_t mac_tx_bw_mode(mac_soft_ring_set_t *, mblk_t *,
998     uintptr_t, uint16_t, mblk_t **);
999 static mac_tx_cookie_t mac_tx_aggr_mode(mac_soft_ring_set_t *, mblk_t *,
1000     uintptr_t, uint16_t, mblk_t **);
1001 
1002 typedef struct mac_tx_mode_s {
1003 	mac_tx_srs_mode_t	mac_tx_mode;
1004 	mac_tx_func_t		mac_tx_func;
1005 } mac_tx_mode_t;
1006 
1007 /*
1008  * There are seven modes of operation on the Tx side. These modes get set
1009  * in mac_tx_srs_setup(). Except for the experimental TX_SERIALIZE mode,
1010  * none of the other modes are user configurable. They get selected by
1011  * the system depending upon whether the link (or flow) has multiple Tx
1012  * rings or a bandwidth configured, or if the link is an aggr, etc.
1013  *
1014  * When the Tx SRS is operating in aggr mode (st_mode) or if there are
1015  * multiple Tx rings owned by Tx SRS, then each Tx ring (pseudo or
1016  * otherwise) will have a soft ring associated with it. These soft rings
1017  * are stored in srs_tx_soft_rings[] array.
1018  *
1019  * Additionally in the case of aggr, there is the st_soft_rings[] array
1020  * in the mac_srs_tx_t structure. This array is used to store the same
1021  * set of soft rings that are present in srs_tx_soft_rings[] array but
1022  * in a different manner. The soft ring associated with the pseudo Tx
1023  * ring is saved at mr_index (of the pseudo ring) in st_soft_rings[]
1024  * array. This helps in quickly getting the soft ring associated with the
1025  * Tx ring when aggr_find_tx_ring() returns the pseudo Tx ring that is to
1026  * be used for transmit.
1027  */
1028 mac_tx_mode_t mac_tx_mode_list[] = {
1029 	{SRS_TX_DEFAULT,	mac_tx_single_ring_mode},
1030 	{SRS_TX_SERIALIZE,	mac_tx_serializer_mode},
1031 	{SRS_TX_FANOUT,		mac_tx_fanout_mode},
1032 	{SRS_TX_BW,		mac_tx_bw_mode},
1033 	{SRS_TX_BW_FANOUT,	mac_tx_bw_mode},
1034 	{SRS_TX_AGGR,		mac_tx_aggr_mode},
1035 	{SRS_TX_BW_AGGR,	mac_tx_bw_mode}
1036 };
1037 
1038 /*
1039  * Soft Ring Set (SRS) - The Run time code that deals with
1040  * dynamic polling from the hardware, bandwidth enforcement,
1041  * fanout etc.
1042  *
1043  * We try to use H/W classification on NIC and assign traffic for
1044  * a MAC address to a particular Rx ring or ring group. There is a
1045  * 1-1 mapping between a SRS and a Rx ring. The SRS dynamically
1046  * switches the underlying Rx ring between interrupt and
1047  * polling mode and enforces any specified B/W control.
1048  *
1049  * There is always a SRS created and tied to each H/W and S/W rule.
1050  * Whenever we create a H/W rule, we always add the the same rule to
1051  * S/W classifier and tie a SRS to it.
1052  *
1053  * In case a B/W control is specified, it is broken into bytes
1054  * per ticks and as soon as the quota for a tick is exhausted,
1055  * the underlying Rx ring is forced into poll mode for remainder of
1056  * the tick. The SRS poll thread only polls for bytes that are
1057  * allowed to come in the SRS. We typically let 4x the configured
1058  * B/W worth of packets to come in the SRS (to prevent unnecessary
1059  * drops due to bursts) but only process the specified amount.
1060  *
1061  * A MAC client (e.g. a VNIC or aggr) can have 1 or more
1062  * Rx rings (and corresponding SRSs) assigned to it. The SRS
1063  * in turn can have softrings to do protocol level fanout or
1064  * softrings to do S/W based fanout or both. In case the NIC
1065  * has no Rx rings, we do S/W classification to respective SRS.
1066  * The S/W classification rule is always setup and ready. This
1067  * allows the MAC layer to reassign Rx rings whenever needed
1068  * but packets still continue to flow via the default path and
1069  * getting S/W classified to correct SRS.
1070  *
1071  * The SRS's are used on both Tx and Rx side. They use the same
1072  * data structure but the processing routines have slightly different
1073  * semantics due to the fact that Rx side needs to do dynamic
1074  * polling etc.
1075  *
1076  * Dynamic Polling Notes
1077  * =====================
1078  *
1079  * Each Soft ring set is capable of switching its Rx ring between
1080  * interrupt and poll mode and actively 'polls' for packets in
1081  * poll mode. If the SRS is implementing a B/W limit, it makes
1082  * sure that only Max allowed packets are pulled in poll mode
1083  * and goes to poll mode as soon as B/W limit is exceeded. As
1084  * such, there are no overheads to implement B/W limits.
1085  *
1086  * In poll mode, its better to keep the pipeline going where the
1087  * SRS worker thread keeps processing packets and poll thread
1088  * keeps bringing more packets (specially if they get to run
1089  * on different CPUs). This also prevents the overheads associated
1090  * by excessive signalling (on NUMA machines, this can be
1091  * pretty devastating). The exception is latency optimized case
1092  * where worker thread does no work and interrupt and poll thread
1093  * are allowed to do their own drain.
1094  *
1095  * We use the following policy to control Dynamic Polling:
1096  * 1) We switch to poll mode anytime the processing
1097  *    thread causes a backlog to build up in SRS and
1098  *    its associated Soft Rings (sr_poll_pkt_cnt > 0).
1099  * 2) As long as the backlog stays under the low water
1100  *    mark (sr_lowat), we poll the H/W for more packets.
1101  * 3) If the backlog (sr_poll_pkt_cnt) exceeds low
1102  *    water mark, we stay in poll mode but don't poll
1103  *    the H/W for more packets.
1104  * 4) Anytime in polling mode, if we poll the H/W for
1105  *    packets and find nothing plus we have an existing
1106  *    backlog (sr_poll_pkt_cnt > 0), we stay in polling
1107  *    mode but don't poll the H/W for packets anymore
1108  *    (let the polling thread go to sleep).
1109  * 5) Once the backlog is relived (packets are processed)
1110  *    we reenable polling (by signalling the poll thread)
1111  *    only when the backlog dips below sr_poll_thres.
1112  * 6) sr_hiwat is used exclusively when we are not
1113  *    polling capable and is used to decide when to
1114  *    drop packets so the SRS queue length doesn't grow
1115  *    infinitely.
1116  *
1117  * NOTE: Also see the block level comment on top of mac_soft_ring.c
1118  */
1119 
1120 /*
1121  * mac_latency_optimize
1122  *
1123  * Controls whether the poll thread can process the packets inline
1124  * or let the SRS worker thread do the processing. This applies if
1125  * the SRS was not being processed. For latency sensitive traffic,
1126  * this needs to be true to allow inline processing. For throughput
1127  * under load, this should be false.
1128  *
1129  * This (and other similar) tunable should be rolled into a link
1130  * or flow specific workload hint that can be set using dladm
1131  * linkprop (instead of multiple such tunables).
1132  */
1133 boolean_t mac_latency_optimize = B_TRUE;
1134 
1135 /*
1136  * MAC_RX_SRS_ENQUEUE_CHAIN and MAC_TX_SRS_ENQUEUE_CHAIN
1137  *
1138  * queue a mp or chain in soft ring set and increment the
1139  * local count (srs_count) for the SRS and the shared counter
1140  * (srs_poll_pkt_cnt - shared between SRS and its soft rings
1141  * to track the total unprocessed packets for polling to work
1142  * correctly).
1143  *
1144  * The size (total bytes queued) counters are incremented only
1145  * if we are doing B/W control.
1146  */
1147 #define	MAC_SRS_ENQUEUE_CHAIN(mac_srs, head, tail, count, sz) {		\
1148 	ASSERT(MUTEX_HELD(&(mac_srs)->srs_lock));			\
1149 	if ((mac_srs)->srs_last != NULL)				\
1150 		(mac_srs)->srs_last->b_next = (head);			\
1151 	else								\
1152 		(mac_srs)->srs_first = (head);				\
1153 	(mac_srs)->srs_last = (tail);					\
1154 	(mac_srs)->srs_count += count;					\
1155 }
1156 
1157 #define	MAC_RX_SRS_ENQUEUE_CHAIN(mac_srs, head, tail, count, sz) {	\
1158 	mac_srs_rx_t	*srs_rx = &(mac_srs)->srs_rx;			\
1159 									\
1160 	MAC_SRS_ENQUEUE_CHAIN(mac_srs, head, tail, count, sz);		\
1161 	srs_rx->sr_poll_pkt_cnt += count;				\
1162 	ASSERT(srs_rx->sr_poll_pkt_cnt > 0);				\
1163 	if ((mac_srs)->srs_type & SRST_BW_CONTROL) {			\
1164 		(mac_srs)->srs_size += (sz);				\
1165 		mutex_enter(&(mac_srs)->srs_bw->mac_bw_lock);		\
1166 		(mac_srs)->srs_bw->mac_bw_sz += (sz);			\
1167 		mutex_exit(&(mac_srs)->srs_bw->mac_bw_lock);		\
1168 	}								\
1169 }
1170 
1171 #define	MAC_TX_SRS_ENQUEUE_CHAIN(mac_srs, head, tail, count, sz) {	\
1172 	mac_srs->srs_state |= SRS_ENQUEUED;				\
1173 	MAC_SRS_ENQUEUE_CHAIN(mac_srs, head, tail, count, sz);		\
1174 	if ((mac_srs)->srs_type & SRST_BW_CONTROL) {			\
1175 		(mac_srs)->srs_size += (sz);				\
1176 		(mac_srs)->srs_bw->mac_bw_sz += (sz);			\
1177 	}								\
1178 }
1179 
1180 /*
1181  * Turn polling on routines
1182  */
1183 #define	MAC_SRS_POLLING_ON(mac_srs) {					\
1184 	ASSERT(MUTEX_HELD(&(mac_srs)->srs_lock));			\
1185 	if (((mac_srs)->srs_state &					\
1186 	    (SRS_POLLING_CAPAB|SRS_POLLING)) == SRS_POLLING_CAPAB) {	\
1187 		(mac_srs)->srs_state |= SRS_POLLING;			\
1188 		(void) mac_hwring_disable_intr((mac_ring_handle_t)	\
1189 		    (mac_srs)->srs_ring);				\
1190 		(mac_srs)->srs_rx.sr_poll_on++;				\
1191 	}								\
1192 }
1193 
1194 #define	MAC_SRS_WORKER_POLLING_ON(mac_srs) {				\
1195 	ASSERT(MUTEX_HELD(&(mac_srs)->srs_lock));			\
1196 	if (((mac_srs)->srs_state &					\
1197 	    (SRS_POLLING_CAPAB|SRS_WORKER|SRS_POLLING)) ==		\
1198 	    (SRS_POLLING_CAPAB|SRS_WORKER)) {				\
1199 		(mac_srs)->srs_state |= SRS_POLLING;			\
1200 		(void) mac_hwring_disable_intr((mac_ring_handle_t)	\
1201 		    (mac_srs)->srs_ring);				\
1202 		(mac_srs)->srs_rx.sr_worker_poll_on++;			\
1203 	}								\
1204 }
1205 
1206 /*
1207  * MAC_SRS_POLL_RING
1208  *
1209  * Signal the SRS poll thread to poll the underlying H/W ring
1210  * provided it wasn't already polling (SRS_GET_PKTS was set).
1211  *
1212  * Poll thread gets to run only from mac_rx_srs_drain() and only
1213  * if the drain was being done by the worker thread.
1214  */
1215 #define	MAC_SRS_POLL_RING(mac_srs) {					\
1216 	mac_srs_rx_t	*srs_rx = &(mac_srs)->srs_rx;			\
1217 									\
1218 	ASSERT(MUTEX_HELD(&(mac_srs)->srs_lock));			\
1219 	srs_rx->sr_poll_thr_sig++;					\
1220 	if (((mac_srs)->srs_state &					\
1221 	    (SRS_POLLING_CAPAB|SRS_WORKER|SRS_GET_PKTS)) ==		\
1222 		(SRS_WORKER|SRS_POLLING_CAPAB)) {			\
1223 		(mac_srs)->srs_state |= SRS_GET_PKTS;			\
1224 		cv_signal(&(mac_srs)->srs_cv);				\
1225 	} else {							\
1226 		srs_rx->sr_poll_thr_busy++;				\
1227 	}								\
1228 }
1229 
1230 /*
1231  * MAC_SRS_CHECK_BW_CONTROL
1232  *
1233  * Check to see if next tick has started so we can reset the
1234  * SRS_BW_ENFORCED flag and allow more packets to come in the
1235  * system.
1236  */
1237 #define	MAC_SRS_CHECK_BW_CONTROL(mac_srs) {				\
1238 	ASSERT(MUTEX_HELD(&(mac_srs)->srs_lock));			\
1239 	ASSERT(((mac_srs)->srs_type & SRST_TX) ||			\
1240 	    MUTEX_HELD(&(mac_srs)->srs_bw->mac_bw_lock));		\
1241 	clock_t now = ddi_get_lbolt();					\
1242 	if ((mac_srs)->srs_bw->mac_bw_curr_time != now) {		\
1243 		(mac_srs)->srs_bw->mac_bw_curr_time = now;		\
1244 		(mac_srs)->srs_bw->mac_bw_used = 0;			\
1245 		if ((mac_srs)->srs_bw->mac_bw_state & SRS_BW_ENFORCED)	\
1246 			(mac_srs)->srs_bw->mac_bw_state &= ~SRS_BW_ENFORCED; \
1247 	}								\
1248 }
1249 
1250 /*
1251  * MAC_SRS_WORKER_WAKEUP
1252  *
1253  * Wake up the SRS worker thread to process the queue as long as
1254  * no one else is processing the queue. If we are optimizing for
1255  * latency, we wake up the worker thread immediately or else we
1256  * wait mac_srs_worker_wakeup_ticks before worker thread gets
1257  * woken up.
1258  */
1259 int mac_srs_worker_wakeup_ticks = 0;
1260 #define	MAC_SRS_WORKER_WAKEUP(mac_srs) {				\
1261 	ASSERT(MUTEX_HELD(&(mac_srs)->srs_lock));			\
1262 	if (!((mac_srs)->srs_state & SRS_PROC) &&			\
1263 		(mac_srs)->srs_tid == NULL) {				\
1264 		if (((mac_srs)->srs_state & SRS_LATENCY_OPT) ||		\
1265 			(mac_srs_worker_wakeup_ticks == 0))		\
1266 			cv_signal(&(mac_srs)->srs_async);		\
1267 		else							\
1268 			(mac_srs)->srs_tid =				\
1269 				timeout(mac_srs_fire, (mac_srs),	\
1270 					mac_srs_worker_wakeup_ticks);	\
1271 	}								\
1272 }
1273 
1274 #define	TX_BANDWIDTH_MODE(mac_srs)				\
1275 	((mac_srs)->srs_tx.st_mode == SRS_TX_BW ||		\
1276 	    (mac_srs)->srs_tx.st_mode == SRS_TX_BW_FANOUT ||	\
1277 	    (mac_srs)->srs_tx.st_mode == SRS_TX_BW_AGGR)
1278 
1279 #define	TX_SRS_TO_SOFT_RING(mac_srs, head, hint) {			\
1280 	if (tx_mode == SRS_TX_BW_FANOUT)				\
1281 		(void) mac_tx_fanout_mode(mac_srs, head, hint, 0, NULL);\
1282 	else								\
1283 		(void) mac_tx_aggr_mode(mac_srs, head, hint, 0, NULL);	\
1284 }
1285 
1286 /*
1287  * MAC_TX_SRS_BLOCK
1288  *
1289  * Always called from mac_tx_srs_drain() function. SRS_TX_BLOCKED
1290  * will be set only if srs_tx_woken_up is FALSE. If
1291  * srs_tx_woken_up is TRUE, it indicates that the wakeup arrived
1292  * before we grabbed srs_lock to set SRS_TX_BLOCKED. We need to
1293  * attempt to transmit again and not setting SRS_TX_BLOCKED does
1294  * that.
1295  */
1296 #define	MAC_TX_SRS_BLOCK(srs, mp)	{			\
1297 	ASSERT(MUTEX_HELD(&(srs)->srs_lock));			\
1298 	if ((srs)->srs_tx.st_woken_up) {			\
1299 		(srs)->srs_tx.st_woken_up = B_FALSE;		\
1300 	} else {						\
1301 		ASSERT(!((srs)->srs_state & SRS_TX_BLOCKED));	\
1302 		(srs)->srs_state |= SRS_TX_BLOCKED;		\
1303 		(srs)->srs_tx.st_stat.mts_blockcnt++;		\
1304 	}							\
1305 }
1306 
1307 /*
1308  * MAC_TX_SRS_TEST_HIWAT
1309  *
1310  * Called before queueing a packet onto Tx SRS to test and set
1311  * SRS_TX_HIWAT if srs_count exceeds srs_tx_hiwat.
1312  */
1313 #define	MAC_TX_SRS_TEST_HIWAT(srs, mp, tail, cnt, sz, cookie) {		\
1314 	boolean_t enqueue = 1;						\
1315 									\
1316 	if ((srs)->srs_count > (srs)->srs_tx.st_hiwat) {		\
1317 		/*							\
1318 		 * flow-controlled. Store srs in cookie so that it	\
1319 		 * can be returned as mac_tx_cookie_t to client		\
1320 		 */							\
1321 		(srs)->srs_state |= SRS_TX_HIWAT;			\
1322 		cookie = (mac_tx_cookie_t)srs;				\
1323 		(srs)->srs_tx.st_hiwat_cnt++;				\
1324 		if ((srs)->srs_count > (srs)->srs_tx.st_max_q_cnt) {	\
1325 			/* increment freed stats */			\
1326 			(srs)->srs_tx.st_stat.mts_sdrops += cnt;	\
1327 			/*						\
1328 			 * b_prev may be set to the fanout hint		\
1329 			 * hence can't use freemsg directly		\
1330 			 */						\
1331 			mac_drop_chain(mp_chain, "SRS Tx max queue");	\
1332 			DTRACE_PROBE1(tx_queued_hiwat,			\
1333 			    mac_soft_ring_set_t *, srs);		\
1334 			enqueue = 0;					\
1335 		}							\
1336 	}								\
1337 	if (enqueue)							\
1338 		MAC_TX_SRS_ENQUEUE_CHAIN(srs, mp, tail, cnt, sz);	\
1339 }
1340 
1341 /* Some utility macros */
1342 #define	MAC_SRS_BW_LOCK(srs)						\
1343 	if (!(srs->srs_type & SRST_TX))					\
1344 		mutex_enter(&srs->srs_bw->mac_bw_lock);
1345 
1346 #define	MAC_SRS_BW_UNLOCK(srs)						\
1347 	if (!(srs->srs_type & SRST_TX))					\
1348 		mutex_exit(&srs->srs_bw->mac_bw_lock);
1349 
1350 #define	MAC_TX_SRS_DROP_MESSAGE(srs, chain, cookie, s) {	\
1351 	mac_drop_chain((chain), (s));				\
1352 	/* increment freed stats */				\
1353 	(srs)->srs_tx.st_stat.mts_sdrops++;			\
1354 	(cookie) = (mac_tx_cookie_t)(srs);			\
1355 }
1356 
1357 #define	MAC_TX_SET_NO_ENQUEUE(srs, mp_chain, ret_mp, cookie) {		\
1358 	mac_srs->srs_state |= SRS_TX_WAKEUP_CLIENT;			\
1359 	cookie = (mac_tx_cookie_t)srs;					\
1360 	*ret_mp = mp_chain;						\
1361 }
1362 
1363 /*
1364  * Threshold used in receive-side processing to determine if handling
1365  * can occur in situ (in the interrupt thread) or if it should be left to a
1366  * worker thread.  Note that the constant used to make this determination is
1367  * not entirely made-up, and is a result of some emprical validation. That
1368  * said, the constant is left as a global variable to allow it to be
1369  * dynamically tuned in the field if and as needed.
1370  */
1371 uintptr_t mac_rx_srs_stack_needed = 14336;
1372 uint_t mac_rx_srs_stack_toodeep;
1373 
1374 #ifndef STACK_GROWTH_DOWN
1375 #error Downward stack growth assumed.
1376 #endif
1377 
1378 /*
1379  * Drop the rx packet and advance to the next one in the chain.
1380  */
1381 static void
mac_rx_drop_pkt(mac_soft_ring_set_t * srs,mblk_t * mp)1382 mac_rx_drop_pkt(mac_soft_ring_set_t *srs, mblk_t *mp)
1383 {
1384 	mac_srs_rx_t	*srs_rx = &srs->srs_rx;
1385 
1386 	ASSERT(mp->b_next == NULL);
1387 	mutex_enter(&srs->srs_lock);
1388 	MAC_UPDATE_SRS_COUNT_LOCKED(srs, 1);
1389 	MAC_UPDATE_SRS_SIZE_LOCKED(srs, msgdsize(mp));
1390 	mutex_exit(&srs->srs_lock);
1391 
1392 	srs_rx->sr_stat.mrs_sdrops++;
1393 	freemsg(mp);
1394 }
1395 
1396 /* DATAPATH RUNTIME ROUTINES */
1397 
1398 /*
1399  * mac_srs_fire
1400  *
1401  * Timer callback routine for waking up the SRS worker thread.
1402  */
1403 static void
mac_srs_fire(void * arg)1404 mac_srs_fire(void *arg)
1405 {
1406 	mac_soft_ring_set_t *mac_srs = (mac_soft_ring_set_t *)arg;
1407 
1408 	mutex_enter(&mac_srs->srs_lock);
1409 	if (mac_srs->srs_tid == NULL) {
1410 		mutex_exit(&mac_srs->srs_lock);
1411 		return;
1412 	}
1413 
1414 	mac_srs->srs_tid = NULL;
1415 	if (!(mac_srs->srs_state & SRS_PROC))
1416 		cv_signal(&mac_srs->srs_async);
1417 
1418 	mutex_exit(&mac_srs->srs_lock);
1419 }
1420 
1421 /*
1422  * 'hint' is fanout_hint (type of uint64_t) which is given by the TCP/IP stack,
1423  * and it is used on the TX path.
1424  */
1425 #define	HASH_HINT(hint)	\
1426 	((hint) ^ ((hint) >> 24) ^ ((hint) >> 16) ^ ((hint) >> 8))
1427 
1428 
1429 /*
1430  * hash based on the src address, dst address and the port information.
1431  */
1432 #define	HASH_ADDR(src, dst, ports)					\
1433 	(ntohl((src) + (dst)) ^ ((ports) >> 24) ^ ((ports) >> 16) ^	\
1434 	((ports) >> 8) ^ (ports))
1435 
1436 #define	COMPUTE_INDEX(key, sz)	(key % sz)
1437 
1438 #define	FANOUT_ENQUEUE_MP(head, tail, cnt, bw_ctl, sz, sz0, mp) {	\
1439 	if ((tail) != NULL) {						\
1440 		ASSERT((tail)->b_next == NULL);				\
1441 		(tail)->b_next = (mp);					\
1442 	} else {							\
1443 		ASSERT((head) == NULL);					\
1444 		(head) = (mp);						\
1445 	}								\
1446 	(tail) = (mp);							\
1447 	(cnt)++;							\
1448 	if ((bw_ctl))							\
1449 		(sz) += (sz0);						\
1450 }
1451 
1452 #define	MAC_FANOUT_DEFAULT	0
1453 #define	MAC_FANOUT_RND_ROBIN	1
1454 int mac_fanout_type = MAC_FANOUT_DEFAULT;
1455 
1456 #define	MAX_SR_TYPES	3
1457 /* fanout types for port based hashing */
1458 enum pkt_type {
1459 	V4_TCP = 0,
1460 	V4_UDP,
1461 	OTH,
1462 	UNDEF
1463 };
1464 
1465 /*
1466  * Pair of local and remote ports in the transport header
1467  */
1468 #define	PORTS_SIZE 4
1469 
1470 /*
1471  * This routine delivers packets destined for an SRS into one of the
1472  * protocol soft rings.
1473  *
1474  * Given a chain of packets we need to split it up into multiple sub
1475  * chains: TCP, UDP or OTH soft ring. Instead of entering the soft
1476  * ring one packet at a time, we want to enter it in the form of a
1477  * chain otherwise we get this start/stop behaviour where the worker
1478  * thread goes to sleep and then next packet comes in forcing it to
1479  * wake up.
1480  */
1481 static void
mac_rx_srs_proto_fanout(mac_soft_ring_set_t * mac_srs,mblk_t * head)1482 mac_rx_srs_proto_fanout(mac_soft_ring_set_t *mac_srs, mblk_t *head)
1483 {
1484 	struct ether_header		*ehp;
1485 	struct ether_vlan_header	*evhp;
1486 	uint32_t			sap;
1487 	ipha_t				*ipha;
1488 	uint8_t				*dstaddr;
1489 	size_t				hdrsize;
1490 	mblk_t				*mp;
1491 	mblk_t				*headmp[MAX_SR_TYPES];
1492 	mblk_t				*tailmp[MAX_SR_TYPES];
1493 	int				cnt[MAX_SR_TYPES];
1494 	size_t				sz[MAX_SR_TYPES];
1495 	size_t				sz1;
1496 	boolean_t			bw_ctl;
1497 	boolean_t			hw_classified;
1498 	boolean_t			dls_bypass;
1499 	boolean_t			is_ether;
1500 	boolean_t			is_unicast;
1501 	enum pkt_type			type;
1502 	mac_client_impl_t		*mcip = mac_srs->srs_mcip;
1503 
1504 	is_ether = (mcip->mci_mip->mi_info.mi_nativemedia == DL_ETHER);
1505 	bw_ctl = ((mac_srs->srs_type & SRST_BW_CONTROL) != 0);
1506 
1507 	/*
1508 	 * If we don't have a Rx ring, S/W classification would have done
1509 	 * its job and its a packet meant for us. If we were polling on
1510 	 * the default ring (i.e. there was a ring assigned to this SRS),
1511 	 * then we need to make sure that the mac address really belongs
1512 	 * to us.
1513 	 */
1514 	hw_classified = mac_srs->srs_ring != NULL &&
1515 	    mac_srs->srs_ring->mr_classify_type == MAC_HW_CLASSIFIER;
1516 
1517 	/*
1518 	 * Some clients, such as non-ethernet, need DLS processing in
1519 	 * the Rx path. Such clients clear the SRST_DLS_BYPASS flag.
1520 	 * DLS bypass may also be disabled via the
1521 	 * MCIS_RX_BYPASS_DISABLE flag.
1522 	 */
1523 	dls_bypass = ((mac_srs->srs_type & SRST_DLS_BYPASS) != 0) &&
1524 	    ((mcip->mci_state_flags & MCIS_RX_BYPASS_DISABLE) == 0);
1525 
1526 	bzero(headmp, MAX_SR_TYPES * sizeof (mblk_t *));
1527 	bzero(tailmp, MAX_SR_TYPES * sizeof (mblk_t *));
1528 	bzero(cnt, MAX_SR_TYPES * sizeof (int));
1529 	bzero(sz, MAX_SR_TYPES * sizeof (size_t));
1530 
1531 	/*
1532 	 * We have a chain from SRS that we need to split across the
1533 	 * soft rings. The squeues for the TCP and IPv4 SAPs use their
1534 	 * own soft rings to allow polling from the squeue. The rest of
1535 	 * the packets are delivered on the OTH soft ring which cannot
1536 	 * be polled.
1537 	 */
1538 	while (head != NULL) {
1539 		mp = head;
1540 		head = head->b_next;
1541 		mp->b_next = NULL;
1542 
1543 		type = OTH;
1544 		sz1 = (mp->b_cont == NULL) ? MBLKL(mp) : msgdsize(mp);
1545 
1546 		if (is_ether) {
1547 			/*
1548 			 * At this point we can be sure the packet at least
1549 			 * has an ether header.
1550 			 */
1551 			if (sz1 < sizeof (struct ether_header)) {
1552 				mac_rx_drop_pkt(mac_srs, mp);
1553 				continue;
1554 			}
1555 			ehp = (struct ether_header *)mp->b_rptr;
1556 
1557 			/*
1558 			 * Determine if this is a VLAN or non-VLAN packet.
1559 			 */
1560 			if ((sap = ntohs(ehp->ether_type)) == VLAN_TPID) {
1561 				evhp = (struct ether_vlan_header *)mp->b_rptr;
1562 				sap = ntohs(evhp->ether_type);
1563 				hdrsize = sizeof (struct ether_vlan_header);
1564 
1565 				/*
1566 				 * Check if the VID of the packet, if
1567 				 * any, belongs to this client.
1568 				 * Technically, if this packet came up
1569 				 * via a HW classified ring then we
1570 				 * don't need to perform this check.
1571 				 * Perhaps a future optimization.
1572 				 */
1573 				if (!mac_client_check_flow_vid(mcip,
1574 				    VLAN_ID(ntohs(evhp->ether_tci)))) {
1575 					mac_rx_drop_pkt(mac_srs, mp);
1576 					continue;
1577 				}
1578 			} else {
1579 				hdrsize = sizeof (struct ether_header);
1580 			}
1581 			is_unicast =
1582 			    ((((uint8_t *)&ehp->ether_dhost)[0] & 0x01) == 0);
1583 			dstaddr = (uint8_t *)&ehp->ether_dhost;
1584 		} else {
1585 			mac_header_info_t		mhi;
1586 
1587 			if (mac_header_info((mac_handle_t)mcip->mci_mip,
1588 			    mp, &mhi) != 0) {
1589 				mac_rx_drop_pkt(mac_srs, mp);
1590 				continue;
1591 			}
1592 			hdrsize = mhi.mhi_hdrsize;
1593 			sap = mhi.mhi_bindsap;
1594 			is_unicast = (mhi.mhi_dsttype == MAC_ADDRTYPE_UNICAST);
1595 			dstaddr = (uint8_t *)mhi.mhi_daddr;
1596 		}
1597 
1598 		if (!dls_bypass) {
1599 			FANOUT_ENQUEUE_MP(headmp[type], tailmp[type],
1600 			    cnt[type], bw_ctl, sz[type], sz1, mp);
1601 			continue;
1602 		}
1603 
1604 		if (sap == ETHERTYPE_IP) {
1605 			/*
1606 			 * If we are H/W classified, but we have promisc
1607 			 * on, then we need to check for the unicast address.
1608 			 */
1609 			if (hw_classified && mcip->mci_promisc_list != NULL) {
1610 				mac_address_t		*map;
1611 
1612 				rw_enter(&mcip->mci_rw_lock, RW_READER);
1613 				map = mcip->mci_unicast;
1614 				if (bcmp(dstaddr, map->ma_addr,
1615 				    map->ma_len) == 0)
1616 					type = UNDEF;
1617 				rw_exit(&mcip->mci_rw_lock);
1618 			} else if (is_unicast) {
1619 				type = UNDEF;
1620 			}
1621 		}
1622 
1623 		/*
1624 		 * This needs to become a contract with the driver for
1625 		 * the fast path.
1626 		 *
1627 		 * In the normal case the packet will have at least the L2
1628 		 * header and the IP + Transport header in the same mblk.
1629 		 * This is usually the case when the NIC driver sends up
1630 		 * the packet. This is also true when the stack generates
1631 		 * a packet that is looped back and when the stack uses the
1632 		 * fastpath mechanism. The normal case is optimized for
1633 		 * performance and may bypass DLS. All other cases go through
1634 		 * the 'OTH' type path without DLS bypass.
1635 		 */
1636 		ipha = (ipha_t *)(mp->b_rptr + hdrsize);
1637 		if ((type != OTH) && MBLK_RX_FANOUT_SLOWPATH(mp, ipha))
1638 			type = OTH;
1639 
1640 		if (type == OTH) {
1641 			FANOUT_ENQUEUE_MP(headmp[type], tailmp[type],
1642 			    cnt[type], bw_ctl, sz[type], sz1, mp);
1643 			continue;
1644 		}
1645 
1646 		ASSERT(type == UNDEF);
1647 
1648 		/*
1649 		 * Determine the type from the IP protocol value. If
1650 		 * classified as TCP or UDP, then update the read
1651 		 * pointer to the beginning of the IP header.
1652 		 * Otherwise leave the message as is for further
1653 		 * processing by DLS.
1654 		 */
1655 		switch (ipha->ipha_protocol) {
1656 		case IPPROTO_TCP:
1657 			type = V4_TCP;
1658 			mp->b_rptr += hdrsize;
1659 			break;
1660 		case IPPROTO_UDP:
1661 			type = V4_UDP;
1662 			mp->b_rptr += hdrsize;
1663 			break;
1664 		default:
1665 			type = OTH;
1666 			break;
1667 		}
1668 
1669 		FANOUT_ENQUEUE_MP(headmp[type], tailmp[type], cnt[type],
1670 		    bw_ctl, sz[type], sz1, mp);
1671 	}
1672 
1673 	for (type = V4_TCP; type < UNDEF; type++) {
1674 		if (headmp[type] != NULL) {
1675 			mac_soft_ring_t			*softring;
1676 
1677 			ASSERT(tailmp[type]->b_next == NULL);
1678 			switch (type) {
1679 			case V4_TCP:
1680 				softring = mac_srs->srs_tcp_soft_rings[0];
1681 				break;
1682 			case V4_UDP:
1683 				softring = mac_srs->srs_udp_soft_rings[0];
1684 				break;
1685 			case OTH:
1686 				softring = mac_srs->srs_oth_soft_rings[0];
1687 			}
1688 			mac_rx_soft_ring_process(mcip, softring,
1689 			    headmp[type], tailmp[type], cnt[type], sz[type]);
1690 		}
1691 	}
1692 }
1693 
1694 int	fanout_unaligned = 0;
1695 
1696 /*
1697  * The fanout routine for any clients with DLS bypass disabled or for
1698  * traffic classified as "other". Returns -1 on an error (drop the
1699  * packet due to a malformed packet), 0 on success, with values
1700  * written in *indx and *type.
1701  */
1702 static int
mac_rx_srs_long_fanout(mac_soft_ring_set_t * mac_srs,mblk_t * mp,uint32_t sap,size_t hdrsize,enum pkt_type * type,uint_t * indx)1703 mac_rx_srs_long_fanout(mac_soft_ring_set_t *mac_srs, mblk_t *mp,
1704     uint32_t sap, size_t hdrsize, enum pkt_type *type, uint_t *indx)
1705 {
1706 	ip6_t		*ip6h;
1707 	ipha_t		*ipha;
1708 	uint8_t		*whereptr;
1709 	uint_t		hash;
1710 	uint16_t	remlen;
1711 	uint8_t		nexthdr;
1712 	uint16_t	hdr_len;
1713 	uint32_t	src_val, dst_val;
1714 	boolean_t	modifiable = B_TRUE;
1715 	boolean_t	v6;
1716 
1717 	ASSERT(MBLKL(mp) >= hdrsize);
1718 
1719 	if (sap == ETHERTYPE_IPV6) {
1720 		v6 = B_TRUE;
1721 		hdr_len = IPV6_HDR_LEN;
1722 	} else if (sap == ETHERTYPE_IP) {
1723 		v6 = B_FALSE;
1724 		hdr_len = IP_SIMPLE_HDR_LENGTH;
1725 	} else {
1726 		*indx = 0;
1727 		*type = OTH;
1728 		return (0);
1729 	}
1730 
1731 	ip6h = (ip6_t *)(mp->b_rptr + hdrsize);
1732 	ipha = (ipha_t *)ip6h;
1733 
1734 	if ((uint8_t *)ip6h == mp->b_wptr) {
1735 		/*
1736 		 * The first mblk_t only includes the mac header.
1737 		 * Note that it is safe to change the mp pointer here,
1738 		 * as the subsequent operation does not assume mp
1739 		 * points to the start of the mac header.
1740 		 */
1741 		mp = mp->b_cont;
1742 
1743 		/*
1744 		 * Make sure the IP header points to an entire one.
1745 		 */
1746 		if (mp == NULL)
1747 			return (-1);
1748 
1749 		if (MBLKL(mp) < hdr_len) {
1750 			modifiable = (DB_REF(mp) == 1);
1751 
1752 			if (modifiable && !pullupmsg(mp, hdr_len))
1753 				return (-1);
1754 		}
1755 
1756 		ip6h = (ip6_t *)mp->b_rptr;
1757 		ipha = (ipha_t *)ip6h;
1758 	}
1759 
1760 	if (!modifiable || !(OK_32PTR((char *)ip6h)) ||
1761 	    ((uint8_t *)ip6h + hdr_len > mp->b_wptr)) {
1762 		/*
1763 		 * If either the IP header is not aligned, or it does not hold
1764 		 * the complete simple structure (a pullupmsg() is not an
1765 		 * option since it would result in an unaligned IP header),
1766 		 * fanout to the default ring.
1767 		 *
1768 		 * Note that this may cause packet reordering.
1769 		 */
1770 		*indx = 0;
1771 		*type = OTH;
1772 		fanout_unaligned++;
1773 		return (0);
1774 	}
1775 
1776 	/*
1777 	 * Extract next-header, full header length, and source-hash value
1778 	 * using v4/v6 specific fields.
1779 	 */
1780 	if (v6) {
1781 		remlen = ntohs(ip6h->ip6_plen);
1782 		nexthdr = ip6h->ip6_nxt;
1783 		src_val = V4_PART_OF_V6(ip6h->ip6_src);
1784 		dst_val = V4_PART_OF_V6(ip6h->ip6_dst);
1785 		/*
1786 		 * Do src based fanout if below tunable is set to B_TRUE or
1787 		 * when mac_ip_hdr_length_v6() fails because of malformed
1788 		 * packets or because mblks need to be concatenated using
1789 		 * pullupmsg().
1790 		 *
1791 		 * Perform a version check to prevent parsing weirdness...
1792 		 */
1793 		if (IPH_HDR_VERSION(ip6h) != IPV6_VERSION ||
1794 		    !mac_ip_hdr_length_v6(ip6h, mp->b_wptr, &hdr_len, &nexthdr,
1795 		    NULL)) {
1796 			goto src_dst_based_fanout;
1797 		}
1798 	} else {
1799 		hdr_len = IPH_HDR_LENGTH(ipha);
1800 		remlen = ntohs(ipha->ipha_length) - hdr_len;
1801 		nexthdr = ipha->ipha_protocol;
1802 		src_val = (uint32_t)ipha->ipha_src;
1803 		dst_val = (uint32_t)ipha->ipha_dst;
1804 		/*
1805 		 * Catch IPv4 fragment case here.  IPv6 has nexthdr == FRAG
1806 		 * for its equivalent case.
1807 		 */
1808 		if ((ntohs(ipha->ipha_fragment_offset_and_flags) &
1809 		    (IPH_MF | IPH_OFFSET)) != 0) {
1810 			goto src_dst_based_fanout;
1811 		}
1812 	}
1813 	if (remlen < MIN_EHDR_LEN)
1814 		return (-1);
1815 	whereptr = (uint8_t *)ip6h + hdr_len;
1816 
1817 	/* If the transport is one of below, we do port/SPI based fanout */
1818 	switch (nexthdr) {
1819 	case IPPROTO_TCP:
1820 	case IPPROTO_UDP:
1821 	case IPPROTO_SCTP:
1822 	case IPPROTO_ESP:
1823 		/*
1824 		 * If the ports or SPI in the transport header is not part of
1825 		 * the mblk, do src_based_fanout, instead of calling
1826 		 * pullupmsg().
1827 		 */
1828 		if (mp->b_cont == NULL || whereptr + PORTS_SIZE <= mp->b_wptr)
1829 			break;	/* out of switch... */
1830 		/* FALLTHRU */
1831 	default:
1832 		goto src_dst_based_fanout;
1833 	}
1834 
1835 	switch (nexthdr) {
1836 	case IPPROTO_TCP:
1837 		hash = HASH_ADDR(src_val, dst_val, *(uint32_t *)whereptr);
1838 		*indx = COMPUTE_INDEX(hash, mac_srs->srs_tcp_ring_count);
1839 		*type = OTH;
1840 		break;
1841 	case IPPROTO_UDP:
1842 	case IPPROTO_SCTP:
1843 	case IPPROTO_ESP:
1844 		if (mac_fanout_type == MAC_FANOUT_DEFAULT) {
1845 			hash = HASH_ADDR(src_val, dst_val,
1846 			    *(uint32_t *)whereptr);
1847 			*indx = COMPUTE_INDEX(hash,
1848 			    mac_srs->srs_udp_ring_count);
1849 		} else {
1850 			*indx = mac_srs->srs_ind % mac_srs->srs_udp_ring_count;
1851 			mac_srs->srs_ind++;
1852 		}
1853 		*type = OTH;
1854 		break;
1855 	}
1856 	return (0);
1857 
1858 src_dst_based_fanout:
1859 	hash = HASH_ADDR(src_val, dst_val, (uint32_t)0);
1860 	*indx = COMPUTE_INDEX(hash, mac_srs->srs_oth_ring_count);
1861 	*type = OTH;
1862 	return (0);
1863 }
1864 
1865 /*
1866  * This routine delivers packets destined for an SRS into a soft ring member
1867  * of the set.
1868  *
1869  * Given a chain of packets we need to split it up into multiple sub
1870  * chains: TCP, UDP or OTH soft ring. Instead of entering the soft
1871  * ring one packet at a time, we want to enter it in the form of a
1872  * chain otherwise we get this start/stop behaviour where the worker
1873  * thread goes to sleep and then next packet comes in forcing it to
1874  * wake up.
1875  *
1876  * Note:
1877  * Since we know what is the maximum fanout possible, we create a 2D array
1878  * of 'softring types * MAX_SR_FANOUT' for the head, tail, cnt and sz
1879  * variables so that we can enter the softrings with chain. We need the
1880  * MAX_SR_FANOUT so we can allocate the arrays on the stack (a kmem_alloc
1881  * for each packet would be expensive). If we ever want to have the
1882  * ability to have unlimited fanout, we should probably declare a head,
1883  * tail, cnt, sz with each soft ring (a data struct which contains a softring
1884  * along with these members) and create an array of this uber struct so we
1885  * don't have to do kmem_alloc.
1886  */
1887 int	fanout_oth1 = 0;
1888 int	fanout_oth2 = 0;
1889 int	fanout_oth3 = 0;
1890 int	fanout_oth4 = 0;
1891 int	fanout_oth5 = 0;
1892 
1893 static void
mac_rx_srs_fanout(mac_soft_ring_set_t * mac_srs,mblk_t * head)1894 mac_rx_srs_fanout(mac_soft_ring_set_t *mac_srs, mblk_t *head)
1895 {
1896 	struct ether_header		*ehp;
1897 	struct ether_vlan_header	*evhp;
1898 	uint32_t			sap;
1899 	ipha_t				*ipha;
1900 	uint8_t				*dstaddr;
1901 	uint_t				indx;
1902 	size_t				ports_offset;
1903 	size_t				ipha_len;
1904 	size_t				hdrsize;
1905 	uint_t				hash;
1906 	mblk_t				*mp;
1907 	mblk_t				*headmp[MAX_SR_TYPES][MAX_SR_FANOUT];
1908 	mblk_t				*tailmp[MAX_SR_TYPES][MAX_SR_FANOUT];
1909 	int				cnt[MAX_SR_TYPES][MAX_SR_FANOUT];
1910 	size_t				sz[MAX_SR_TYPES][MAX_SR_FANOUT];
1911 	size_t				sz1;
1912 	boolean_t			bw_ctl;
1913 	boolean_t			hw_classified;
1914 	boolean_t			dls_bypass;
1915 	boolean_t			is_ether;
1916 	boolean_t			is_unicast;
1917 	int				fanout_cnt;
1918 	enum pkt_type			type;
1919 	mac_client_impl_t		*mcip = mac_srs->srs_mcip;
1920 
1921 	is_ether = (mcip->mci_mip->mi_info.mi_nativemedia == DL_ETHER);
1922 	bw_ctl = ((mac_srs->srs_type & SRST_BW_CONTROL) != 0);
1923 
1924 	/*
1925 	 * If we don't have a Rx ring, S/W classification would have done
1926 	 * its job and its a packet meant for us. If we were polling on
1927 	 * the default ring (i.e. there was a ring assigned to this SRS),
1928 	 * then we need to make sure that the mac address really belongs
1929 	 * to us.
1930 	 */
1931 	hw_classified = mac_srs->srs_ring != NULL &&
1932 	    mac_srs->srs_ring->mr_classify_type == MAC_HW_CLASSIFIER;
1933 
1934 	/*
1935 	 * Some clients, such as non Ethernet, need DLS processing in
1936 	 * the Rx path. Such clients clear the SRST_DLS_BYPASS flag.
1937 	 * DLS bypass may also be disabled via the
1938 	 * MCIS_RX_BYPASS_DISABLE flag, but this is only consumed by
1939 	 * sun4v vsw currently.
1940 	 */
1941 	dls_bypass = ((mac_srs->srs_type & SRST_DLS_BYPASS) != 0) &&
1942 	    ((mcip->mci_state_flags & MCIS_RX_BYPASS_DISABLE) == 0);
1943 
1944 	/*
1945 	 * Since the softrings are never destroyed and we always
1946 	 * create equal number of softrings for TCP, UDP and rest,
1947 	 * its OK to check one of them for count and use it without
1948 	 * any lock. In future, if soft rings get destroyed because
1949 	 * of reduction in fanout, we will need to ensure that happens
1950 	 * behind the SRS_PROC.
1951 	 */
1952 	fanout_cnt = mac_srs->srs_tcp_ring_count;
1953 
1954 	bzero(headmp, MAX_SR_TYPES * MAX_SR_FANOUT * sizeof (mblk_t *));
1955 	bzero(tailmp, MAX_SR_TYPES * MAX_SR_FANOUT * sizeof (mblk_t *));
1956 	bzero(cnt, MAX_SR_TYPES * MAX_SR_FANOUT * sizeof (int));
1957 	bzero(sz, MAX_SR_TYPES * MAX_SR_FANOUT * sizeof (size_t));
1958 
1959 	/*
1960 	 * We got a chain from SRS that we need to send to the soft rings.
1961 	 * Since squeues for TCP & IPv4 SAP poll their soft rings (for
1962 	 * performance reasons), we need to separate out v4_tcp, v4_udp
1963 	 * and the rest goes in other.
1964 	 */
1965 	while (head != NULL) {
1966 		mp = head;
1967 		head = head->b_next;
1968 		mp->b_next = NULL;
1969 
1970 		type = OTH;
1971 		sz1 = (mp->b_cont == NULL) ? MBLKL(mp) : msgdsize(mp);
1972 
1973 		if (is_ether) {
1974 			/*
1975 			 * At this point we can be sure the packet at least
1976 			 * has an ether header.
1977 			 */
1978 			if (sz1 < sizeof (struct ether_header)) {
1979 				mac_rx_drop_pkt(mac_srs, mp);
1980 				continue;
1981 			}
1982 			ehp = (struct ether_header *)mp->b_rptr;
1983 
1984 			/*
1985 			 * Determine if this is a VLAN or non-VLAN packet.
1986 			 */
1987 			if ((sap = ntohs(ehp->ether_type)) == VLAN_TPID) {
1988 				evhp = (struct ether_vlan_header *)mp->b_rptr;
1989 				sap = ntohs(evhp->ether_type);
1990 				hdrsize = sizeof (struct ether_vlan_header);
1991 
1992 				/*
1993 				 * Check if the VID of the packet, if
1994 				 * any, belongs to this client.
1995 				 * Technically, if this packet came up
1996 				 * via a HW classified ring then we
1997 				 * don't need to perform this check.
1998 				 * Perhaps a future optimization.
1999 				 */
2000 				if (!mac_client_check_flow_vid(mcip,
2001 				    VLAN_ID(ntohs(evhp->ether_tci)))) {
2002 					mac_rx_drop_pkt(mac_srs, mp);
2003 					continue;
2004 				}
2005 			} else {
2006 				hdrsize = sizeof (struct ether_header);
2007 			}
2008 			is_unicast =
2009 			    ((((uint8_t *)&ehp->ether_dhost)[0] & 0x01) == 0);
2010 			dstaddr = (uint8_t *)&ehp->ether_dhost;
2011 		} else {
2012 			mac_header_info_t		mhi;
2013 
2014 			if (mac_header_info((mac_handle_t)mcip->mci_mip,
2015 			    mp, &mhi) != 0) {
2016 				mac_rx_drop_pkt(mac_srs, mp);
2017 				continue;
2018 			}
2019 			hdrsize = mhi.mhi_hdrsize;
2020 			sap = mhi.mhi_bindsap;
2021 			is_unicast = (mhi.mhi_dsttype == MAC_ADDRTYPE_UNICAST);
2022 			dstaddr = (uint8_t *)mhi.mhi_daddr;
2023 		}
2024 
2025 		if (!dls_bypass) {
2026 			if (mac_rx_srs_long_fanout(mac_srs, mp, sap,
2027 			    hdrsize, &type, &indx) == -1) {
2028 				mac_rx_drop_pkt(mac_srs, mp);
2029 				continue;
2030 			}
2031 
2032 			FANOUT_ENQUEUE_MP(headmp[type][indx],
2033 			    tailmp[type][indx], cnt[type][indx], bw_ctl,
2034 			    sz[type][indx], sz1, mp);
2035 			continue;
2036 		}
2037 
2038 		/*
2039 		 * If we are using the default Rx ring where H/W or S/W
2040 		 * classification has not happened, we need to verify if
2041 		 * this unicast packet really belongs to us.
2042 		 */
2043 		if (sap == ETHERTYPE_IP) {
2044 			/*
2045 			 * If we are H/W classified, but we have promisc
2046 			 * on, then we need to check for the unicast address.
2047 			 */
2048 			if (hw_classified && mcip->mci_promisc_list != NULL) {
2049 				mac_address_t		*map;
2050 
2051 				rw_enter(&mcip->mci_rw_lock, RW_READER);
2052 				map = mcip->mci_unicast;
2053 				if (bcmp(dstaddr, map->ma_addr,
2054 				    map->ma_len) == 0)
2055 					type = UNDEF;
2056 				rw_exit(&mcip->mci_rw_lock);
2057 			} else if (is_unicast) {
2058 				type = UNDEF;
2059 			}
2060 		}
2061 
2062 		/*
2063 		 * This needs to become a contract with the driver for
2064 		 * the fast path.
2065 		 */
2066 
2067 		ipha = (ipha_t *)(mp->b_rptr + hdrsize);
2068 		if ((type != OTH) && MBLK_RX_FANOUT_SLOWPATH(mp, ipha)) {
2069 			type = OTH;
2070 			fanout_oth1++;
2071 		}
2072 
2073 		if (type != OTH) {
2074 			uint16_t	frag_offset_flags;
2075 
2076 			switch (ipha->ipha_protocol) {
2077 			case IPPROTO_TCP:
2078 			case IPPROTO_UDP:
2079 			case IPPROTO_SCTP:
2080 			case IPPROTO_ESP:
2081 				ipha_len = IPH_HDR_LENGTH(ipha);
2082 				if ((uchar_t *)ipha + ipha_len + PORTS_SIZE >
2083 				    mp->b_wptr) {
2084 					type = OTH;
2085 					break;
2086 				}
2087 				frag_offset_flags =
2088 				    ntohs(ipha->ipha_fragment_offset_and_flags);
2089 				if ((frag_offset_flags &
2090 				    (IPH_MF | IPH_OFFSET)) != 0) {
2091 					type = OTH;
2092 					fanout_oth3++;
2093 					break;
2094 				}
2095 				ports_offset = hdrsize + ipha_len;
2096 				break;
2097 			default:
2098 				type = OTH;
2099 				fanout_oth4++;
2100 				break;
2101 			}
2102 		}
2103 
2104 		if (type == OTH) {
2105 			if (mac_rx_srs_long_fanout(mac_srs, mp, sap,
2106 			    hdrsize, &type, &indx) == -1) {
2107 				mac_rx_drop_pkt(mac_srs, mp);
2108 				continue;
2109 			}
2110 
2111 			FANOUT_ENQUEUE_MP(headmp[type][indx],
2112 			    tailmp[type][indx], cnt[type][indx], bw_ctl,
2113 			    sz[type][indx], sz1, mp);
2114 			continue;
2115 		}
2116 
2117 		ASSERT(type == UNDEF);
2118 
2119 		/*
2120 		 * XXX-Sunay: We should hold srs_lock since ring_count
2121 		 * below can change. But if we are always called from
2122 		 * mac_rx_srs_drain and SRS_PROC is set, then we can
2123 		 * enforce that ring_count can't be changed i.e.
2124 		 * to change fanout type or ring count, the calling
2125 		 * thread needs to be behind SRS_PROC.
2126 		 */
2127 		switch (ipha->ipha_protocol) {
2128 		case IPPROTO_TCP:
2129 			/*
2130 			 * Note that for ESP, we fanout on SPI and it is at the
2131 			 * same offset as the 2x16-bit ports. So it is clumped
2132 			 * along with TCP, UDP and SCTP.
2133 			 */
2134 			hash = HASH_ADDR(ipha->ipha_src, ipha->ipha_dst,
2135 			    *(uint32_t *)(mp->b_rptr + ports_offset));
2136 			indx = COMPUTE_INDEX(hash, mac_srs->srs_tcp_ring_count);
2137 			type = V4_TCP;
2138 			mp->b_rptr += hdrsize;
2139 			break;
2140 		case IPPROTO_UDP:
2141 		case IPPROTO_SCTP:
2142 		case IPPROTO_ESP:
2143 			if (mac_fanout_type == MAC_FANOUT_DEFAULT) {
2144 				hash = HASH_ADDR(ipha->ipha_src, ipha->ipha_dst,
2145 				    *(uint32_t *)(mp->b_rptr + ports_offset));
2146 				indx = COMPUTE_INDEX(hash,
2147 				    mac_srs->srs_udp_ring_count);
2148 			} else {
2149 				indx = mac_srs->srs_ind %
2150 				    mac_srs->srs_udp_ring_count;
2151 				mac_srs->srs_ind++;
2152 			}
2153 			type = V4_UDP;
2154 			mp->b_rptr += hdrsize;
2155 			break;
2156 		default:
2157 			indx = 0;
2158 			type = OTH;
2159 		}
2160 
2161 		FANOUT_ENQUEUE_MP(headmp[type][indx], tailmp[type][indx],
2162 		    cnt[type][indx], bw_ctl, sz[type][indx], sz1, mp);
2163 	}
2164 
2165 	for (type = V4_TCP; type < UNDEF; type++) {
2166 		int	i;
2167 
2168 		for (i = 0; i < fanout_cnt; i++) {
2169 			if (headmp[type][i] != NULL) {
2170 				mac_soft_ring_t	*softring;
2171 
2172 				ASSERT(tailmp[type][i]->b_next == NULL);
2173 				switch (type) {
2174 				case V4_TCP:
2175 					softring =
2176 					    mac_srs->srs_tcp_soft_rings[i];
2177 					break;
2178 				case V4_UDP:
2179 					softring =
2180 					    mac_srs->srs_udp_soft_rings[i];
2181 					break;
2182 				case OTH:
2183 					softring =
2184 					    mac_srs->srs_oth_soft_rings[i];
2185 					break;
2186 				}
2187 				mac_rx_soft_ring_process(mcip,
2188 				    softring, headmp[type][i], tailmp[type][i],
2189 				    cnt[type][i], sz[type][i]);
2190 			}
2191 		}
2192 	}
2193 }
2194 
2195 #define	SRS_BYTES_TO_PICKUP	150000
2196 ssize_t	max_bytes_to_pickup = SRS_BYTES_TO_PICKUP;
2197 
2198 /*
2199  * mac_rx_srs_poll_ring
2200  *
2201  * This SRS Poll thread uses this routine to poll the underlying hardware
2202  * Rx ring to get a chain of packets. It can inline process that chain
2203  * if mac_latency_optimize is set (default) or signal the SRS worker thread
2204  * to do the remaining processing.
2205  *
2206  * Since packets come in the system via interrupt or poll path, we also
2207  * update the stats and deal with promiscous clients here.
2208  */
2209 void
mac_rx_srs_poll_ring(mac_soft_ring_set_t * mac_srs)2210 mac_rx_srs_poll_ring(mac_soft_ring_set_t *mac_srs)
2211 {
2212 	kmutex_t		*lock = &mac_srs->srs_lock;
2213 	kcondvar_t		*async = &mac_srs->srs_cv;
2214 	mac_srs_rx_t		*srs_rx = &mac_srs->srs_rx;
2215 	mblk_t			*head, *tail, *mp;
2216 	callb_cpr_t		cprinfo;
2217 	ssize_t			bytes_to_pickup;
2218 	size_t			sz;
2219 	int			count;
2220 	mac_client_impl_t	*smcip;
2221 
2222 	CALLB_CPR_INIT(&cprinfo, lock, callb_generic_cpr, "mac_srs_poll");
2223 	mutex_enter(lock);
2224 
2225 start:
2226 	for (;;) {
2227 		if (mac_srs->srs_state & SRS_PAUSE)
2228 			goto done;
2229 
2230 		CALLB_CPR_SAFE_BEGIN(&cprinfo);
2231 		cv_wait(async, lock);
2232 		CALLB_CPR_SAFE_END(&cprinfo, lock);
2233 
2234 		if (mac_srs->srs_state & SRS_PAUSE)
2235 			goto done;
2236 
2237 check_again:
2238 		if (mac_srs->srs_type & SRST_BW_CONTROL) {
2239 			/*
2240 			 * We pick as many bytes as we are allowed to queue.
2241 			 * Its possible that we will exceed the total
2242 			 * packets queued in case this SRS is part of the
2243 			 * Rx ring group since > 1 poll thread can be pulling
2244 			 * upto the max allowed packets at the same time
2245 			 * but that should be OK.
2246 			 */
2247 			mutex_enter(&mac_srs->srs_bw->mac_bw_lock);
2248 			bytes_to_pickup =
2249 			    mac_srs->srs_bw->mac_bw_drop_threshold -
2250 			    mac_srs->srs_bw->mac_bw_sz;
2251 			/*
2252 			 * We shouldn't have been signalled if we
2253 			 * have 0 or less bytes to pick but since
2254 			 * some of the bytes accounting is driver
2255 			 * dependant, we do the safety check.
2256 			 */
2257 			if (bytes_to_pickup < 0)
2258 				bytes_to_pickup = 0;
2259 			mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
2260 		} else {
2261 			/*
2262 			 * ToDO: Need to change the polling API
2263 			 * to add a packet count and a flag which
2264 			 * tells the driver whether we want packets
2265 			 * based on a count, or bytes, or all the
2266 			 * packets queued in the driver/HW. This
2267 			 * way, we never have to check the limits
2268 			 * on poll path. We truly let only as many
2269 			 * packets enter the system as we are willing
2270 			 * to process or queue.
2271 			 *
2272 			 * Something along the lines of
2273 			 * pkts_to_pickup = mac_soft_ring_max_q_cnt -
2274 			 *	mac_srs->srs_poll_pkt_cnt
2275 			 */
2276 
2277 			/*
2278 			 * Since we are not doing B/W control, pick
2279 			 * as many packets as allowed.
2280 			 */
2281 			bytes_to_pickup = max_bytes_to_pickup;
2282 		}
2283 
2284 		/* Poll the underlying Hardware */
2285 		mutex_exit(lock);
2286 		head = MAC_HWRING_POLL(mac_srs->srs_ring, (int)bytes_to_pickup);
2287 		mutex_enter(lock);
2288 
2289 		ASSERT((mac_srs->srs_state & SRS_POLL_THR_OWNER) ==
2290 		    SRS_POLL_THR_OWNER);
2291 
2292 		mp = tail = head;
2293 		count = 0;
2294 		sz = 0;
2295 		while (mp != NULL) {
2296 			tail = mp;
2297 			sz += msgdsize(mp);
2298 			mp = mp->b_next;
2299 			count++;
2300 		}
2301 
2302 		if (head != NULL) {
2303 			tail->b_next = NULL;
2304 			smcip = mac_srs->srs_mcip;
2305 
2306 			SRS_RX_STAT_UPDATE(mac_srs, pollbytes, sz);
2307 			SRS_RX_STAT_UPDATE(mac_srs, pollcnt, count);
2308 
2309 			/*
2310 			 * If there are any promiscuous mode callbacks
2311 			 * defined for this MAC client, pass them a copy
2312 			 * if appropriate and also update the counters.
2313 			 */
2314 			if (smcip != NULL) {
2315 				if (smcip->mci_mip->mi_promisc_list != NULL) {
2316 					mutex_exit(lock);
2317 					mac_promisc_dispatch(smcip->mci_mip,
2318 					    head, NULL, B_FALSE);
2319 					mutex_enter(lock);
2320 				}
2321 			}
2322 			if (mac_srs->srs_type & SRST_BW_CONTROL) {
2323 				mutex_enter(&mac_srs->srs_bw->mac_bw_lock);
2324 				mac_srs->srs_bw->mac_bw_polled += sz;
2325 				mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
2326 			}
2327 			MAC_RX_SRS_ENQUEUE_CHAIN(mac_srs, head, tail,
2328 			    count, sz);
2329 			if (count <= 10)
2330 				srs_rx->sr_stat.mrs_chaincntundr10++;
2331 			else if (count > 10 && count <= 50)
2332 				srs_rx->sr_stat.mrs_chaincnt10to50++;
2333 			else
2334 				srs_rx->sr_stat.mrs_chaincntover50++;
2335 		}
2336 
2337 		/*
2338 		 * We are guaranteed that SRS_PROC will be set if we
2339 		 * are here. Also, poll thread gets to run only if
2340 		 * the drain was being done by a worker thread although
2341 		 * its possible that worker thread is still running
2342 		 * and poll thread was sent down to keep the pipeline
2343 		 * going instead of doing a complete drain and then
2344 		 * trying to poll the NIC.
2345 		 *
2346 		 * So we need to check SRS_WORKER flag to make sure
2347 		 * that the worker thread is not processing the queue
2348 		 * in parallel to us. The flags and conditions are
2349 		 * protected by the srs_lock to prevent any race. We
2350 		 * ensure that we don't drop the srs_lock from now
2351 		 * till the end and similarly we don't drop the srs_lock
2352 		 * in mac_rx_srs_drain() till similar condition check
2353 		 * are complete. The mac_rx_srs_drain() needs to ensure
2354 		 * that SRS_WORKER flag remains set as long as its
2355 		 * processing the queue.
2356 		 */
2357 		if (!(mac_srs->srs_state & SRS_WORKER) &&
2358 		    (mac_srs->srs_first != NULL)) {
2359 			/*
2360 			 * We have packets to process and worker thread
2361 			 * is not running. Check to see if poll thread is
2362 			 * allowed to process.
2363 			 */
2364 			if (mac_srs->srs_state & SRS_LATENCY_OPT) {
2365 				mac_srs->srs_drain_func(mac_srs, SRS_POLL_PROC);
2366 				if (!(mac_srs->srs_state & SRS_PAUSE) &&
2367 				    srs_rx->sr_poll_pkt_cnt <=
2368 				    srs_rx->sr_lowat) {
2369 					srs_rx->sr_poll_again++;
2370 					goto check_again;
2371 				}
2372 				/*
2373 				 * We are already above low water mark
2374 				 * so stay in the polling mode but no
2375 				 * need to poll. Once we dip below
2376 				 * the polling threshold, the processing
2377 				 * thread (soft ring) will signal us
2378 				 * to poll again (MAC_UPDATE_SRS_COUNT)
2379 				 */
2380 				srs_rx->sr_poll_drain_no_poll++;
2381 				mac_srs->srs_state &= ~(SRS_PROC|SRS_GET_PKTS);
2382 				/*
2383 				 * In B/W control case, its possible
2384 				 * that the backlog built up due to
2385 				 * B/W limit being reached and packets
2386 				 * are queued only in SRS. In this case,
2387 				 * we should schedule worker thread
2388 				 * since no one else will wake us up.
2389 				 */
2390 				if ((mac_srs->srs_type & SRST_BW_CONTROL) &&
2391 				    (mac_srs->srs_tid == NULL)) {
2392 					mac_srs->srs_tid =
2393 					    timeout(mac_srs_fire, mac_srs, 1);
2394 					srs_rx->sr_poll_worker_wakeup++;
2395 				}
2396 			} else {
2397 				/*
2398 				 * Wakeup the worker thread for more processing.
2399 				 * We optimize for throughput in this case.
2400 				 */
2401 				mac_srs->srs_state &= ~(SRS_PROC|SRS_GET_PKTS);
2402 				MAC_SRS_WORKER_WAKEUP(mac_srs);
2403 				srs_rx->sr_poll_sig_worker++;
2404 			}
2405 		} else if ((mac_srs->srs_first == NULL) &&
2406 		    !(mac_srs->srs_state & SRS_WORKER)) {
2407 			/*
2408 			 * There is nothing queued in SRS and
2409 			 * no worker thread running. Plus we
2410 			 * didn't get anything from the H/W
2411 			 * as well (head == NULL);
2412 			 */
2413 			ASSERT(head == NULL);
2414 			mac_srs->srs_state &=
2415 			    ~(SRS_PROC|SRS_GET_PKTS);
2416 
2417 			/*
2418 			 * If we have a packets in soft ring, don't allow
2419 			 * more packets to come into this SRS by keeping the
2420 			 * interrupts off but not polling the H/W. The
2421 			 * poll thread will get signaled as soon as
2422 			 * srs_poll_pkt_cnt dips below poll threshold.
2423 			 */
2424 			if (srs_rx->sr_poll_pkt_cnt == 0) {
2425 				srs_rx->sr_poll_intr_enable++;
2426 				MAC_SRS_POLLING_OFF(mac_srs);
2427 			} else {
2428 				/*
2429 				 * We know nothing is queued in SRS
2430 				 * since we are here after checking
2431 				 * srs_first is NULL. The backlog
2432 				 * is entirely due to packets queued
2433 				 * in Soft ring which will wake us up
2434 				 * and get the interface out of polling
2435 				 * mode once the backlog dips below
2436 				 * sr_poll_thres.
2437 				 */
2438 				srs_rx->sr_poll_no_poll++;
2439 			}
2440 		} else {
2441 			/*
2442 			 * Worker thread is already running.
2443 			 * Nothing much to do. If the polling
2444 			 * was enabled, worker thread will deal
2445 			 * with that.
2446 			 */
2447 			mac_srs->srs_state &= ~SRS_GET_PKTS;
2448 			srs_rx->sr_poll_goto_sleep++;
2449 		}
2450 	}
2451 done:
2452 	mac_srs->srs_state |= SRS_POLL_THR_QUIESCED;
2453 	cv_signal(&mac_srs->srs_async);
2454 	/*
2455 	 * If this is a temporary quiesce then wait for the restart signal
2456 	 * from the srs worker. Then clear the flags and signal the srs worker
2457 	 * to ensure a positive handshake and go back to start.
2458 	 */
2459 	while (!(mac_srs->srs_state & (SRS_CONDEMNED | SRS_POLL_THR_RESTART)))
2460 		cv_wait(async, lock);
2461 	if (mac_srs->srs_state & SRS_POLL_THR_RESTART) {
2462 		ASSERT(!(mac_srs->srs_state & SRS_CONDEMNED));
2463 		mac_srs->srs_state &=
2464 		    ~(SRS_POLL_THR_QUIESCED | SRS_POLL_THR_RESTART);
2465 		cv_signal(&mac_srs->srs_async);
2466 		goto start;
2467 	} else {
2468 		mac_srs->srs_state |= SRS_POLL_THR_EXITED;
2469 		cv_signal(&mac_srs->srs_async);
2470 		CALLB_CPR_EXIT(&cprinfo);
2471 		thread_exit();
2472 	}
2473 }
2474 
2475 /*
2476  * mac_srs_pick_chain
2477  *
2478  * In Bandwidth control case, checks how many packets can be processed
2479  * and return them in a sub chain.
2480  */
2481 static mblk_t *
mac_srs_pick_chain(mac_soft_ring_set_t * mac_srs,mblk_t ** chain_tail,size_t * chain_sz,int * chain_cnt)2482 mac_srs_pick_chain(mac_soft_ring_set_t *mac_srs, mblk_t **chain_tail,
2483     size_t *chain_sz, int *chain_cnt)
2484 {
2485 	mblk_t			*head = NULL;
2486 	mblk_t			*tail = NULL;
2487 	size_t			sz;
2488 	size_t			tsz = 0;
2489 	int			cnt = 0;
2490 	mblk_t			*mp;
2491 
2492 	ASSERT(MUTEX_HELD(&mac_srs->srs_lock));
2493 	mutex_enter(&mac_srs->srs_bw->mac_bw_lock);
2494 	if (((mac_srs->srs_bw->mac_bw_used + mac_srs->srs_size) <=
2495 	    mac_srs->srs_bw->mac_bw_limit) ||
2496 	    (mac_srs->srs_bw->mac_bw_limit == 0)) {
2497 		mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
2498 		head = mac_srs->srs_first;
2499 		mac_srs->srs_first = NULL;
2500 		*chain_tail = mac_srs->srs_last;
2501 		mac_srs->srs_last = NULL;
2502 		*chain_sz = mac_srs->srs_size;
2503 		*chain_cnt = mac_srs->srs_count;
2504 		mac_srs->srs_count = 0;
2505 		mac_srs->srs_size = 0;
2506 		return (head);
2507 	}
2508 
2509 	/*
2510 	 * Can't clear the entire backlog.
2511 	 * Need to find how many packets to pick
2512 	 */
2513 	ASSERT(MUTEX_HELD(&mac_srs->srs_bw->mac_bw_lock));
2514 	while ((mp = mac_srs->srs_first) != NULL) {
2515 		sz = msgdsize(mp);
2516 		if ((tsz + sz + mac_srs->srs_bw->mac_bw_used) >
2517 		    mac_srs->srs_bw->mac_bw_limit) {
2518 			if (!(mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED))
2519 				mac_srs->srs_bw->mac_bw_state |=
2520 				    SRS_BW_ENFORCED;
2521 			break;
2522 		}
2523 
2524 		/*
2525 		 * The _size & cnt is  decremented from the softrings
2526 		 * when they send up the packet for polling to work
2527 		 * properly.
2528 		 */
2529 		tsz += sz;
2530 		cnt++;
2531 		mac_srs->srs_count--;
2532 		mac_srs->srs_size -= sz;
2533 		if (tail != NULL)
2534 			tail->b_next = mp;
2535 		else
2536 			head = mp;
2537 		tail = mp;
2538 		mac_srs->srs_first = mac_srs->srs_first->b_next;
2539 	}
2540 	mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
2541 	if (mac_srs->srs_first == NULL)
2542 		mac_srs->srs_last = NULL;
2543 
2544 	if (tail != NULL)
2545 		tail->b_next = NULL;
2546 	*chain_tail = tail;
2547 	*chain_cnt = cnt;
2548 	*chain_sz = tsz;
2549 
2550 	return (head);
2551 }
2552 
2553 /*
2554  * mac_rx_srs_drain
2555  *
2556  * The SRS drain routine. Gets to run to clear the queue. Any thread
2557  * (worker, interrupt, poll) can call this based on processing model.
2558  * The first thing we do is disable interrupts if possible and then
2559  * drain the queue. we also try to poll the underlying hardware if
2560  * there is a dedicated hardware Rx ring assigned to this SRS.
2561  *
2562  * There is a equivalent drain routine in bandwidth control mode
2563  * mac_rx_srs_drain_bw. There is some code duplication between the two
2564  * routines but they are highly performance sensitive and are easier
2565  * to read/debug if they stay separate. Any code changes here might
2566  * also apply to mac_rx_srs_drain_bw as well.
2567  */
2568 void
mac_rx_srs_drain(mac_soft_ring_set_t * mac_srs,uint_t proc_type)2569 mac_rx_srs_drain(mac_soft_ring_set_t *mac_srs, uint_t proc_type)
2570 {
2571 	mblk_t			*head;
2572 	mblk_t			*tail;
2573 	timeout_id_t		tid;
2574 	int			cnt = 0;
2575 	mac_client_impl_t	*mcip = mac_srs->srs_mcip;
2576 	mac_srs_rx_t		*srs_rx = &mac_srs->srs_rx;
2577 
2578 	ASSERT(MUTEX_HELD(&mac_srs->srs_lock));
2579 	ASSERT(!(mac_srs->srs_type & SRST_BW_CONTROL));
2580 
2581 	/* If we are blanked i.e. can't do upcalls, then we are done */
2582 	if (mac_srs->srs_state & (SRS_BLANK | SRS_PAUSE)) {
2583 		ASSERT((mac_srs->srs_type & SRST_NO_SOFT_RINGS) ||
2584 		    (mac_srs->srs_state & SRS_PAUSE));
2585 		goto out;
2586 	}
2587 
2588 	if (mac_srs->srs_first == NULL)
2589 		goto out;
2590 
2591 	if (!(mac_srs->srs_state & SRS_LATENCY_OPT) &&
2592 	    (srs_rx->sr_poll_pkt_cnt <= srs_rx->sr_lowat)) {
2593 		/*
2594 		 * In the normal case, the SRS worker thread does no
2595 		 * work and we wait for a backlog to build up before
2596 		 * we switch into polling mode. In case we are
2597 		 * optimizing for throughput, we use the worker thread
2598 		 * as well. The goal is to let worker thread process
2599 		 * the queue and poll thread to feed packets into
2600 		 * the queue. As such, we should signal the poll
2601 		 * thread to try and get more packets.
2602 		 *
2603 		 * We could have pulled this check in the POLL_RING
2604 		 * macro itself but keeping it explicit here makes
2605 		 * the architecture more human understandable.
2606 		 */
2607 		MAC_SRS_POLL_RING(mac_srs);
2608 	}
2609 
2610 again:
2611 	head = mac_srs->srs_first;
2612 	mac_srs->srs_first = NULL;
2613 	tail = mac_srs->srs_last;
2614 	mac_srs->srs_last = NULL;
2615 	cnt = mac_srs->srs_count;
2616 	mac_srs->srs_count = 0;
2617 
2618 	ASSERT(head != NULL);
2619 	ASSERT(tail != NULL);
2620 
2621 	if ((tid = mac_srs->srs_tid) != NULL)
2622 		mac_srs->srs_tid = NULL;
2623 
2624 	mac_srs->srs_state |= (SRS_PROC|proc_type);
2625 
2626 	/*
2627 	 * mcip is NULL for broadcast and multicast flows. The promisc
2628 	 * callbacks for broadcast and multicast packets are delivered from
2629 	 * mac_rx() and we don't need to worry about that case in this path
2630 	 */
2631 	if (mcip != NULL) {
2632 		if (mcip->mci_promisc_list != NULL) {
2633 			mutex_exit(&mac_srs->srs_lock);
2634 			mac_promisc_client_dispatch(mcip, head);
2635 			mutex_enter(&mac_srs->srs_lock);
2636 		}
2637 		if (MAC_PROTECT_ENABLED(mcip, MPT_IPNOSPOOF)) {
2638 			mutex_exit(&mac_srs->srs_lock);
2639 			mac_protect_intercept_dynamic(mcip, head);
2640 			mutex_enter(&mac_srs->srs_lock);
2641 		}
2642 	}
2643 
2644 	/*
2645 	 * Check if SRS itself is doing the processing. This direct
2646 	 * path applies only when subflows are present.
2647 	 */
2648 	if (mac_srs->srs_type & SRST_NO_SOFT_RINGS) {
2649 		mac_direct_rx_t		proc;
2650 		void			*arg1;
2651 		mac_resource_handle_t	arg2;
2652 
2653 		/*
2654 		 * This is the case when a Rx is directly
2655 		 * assigned and we have a fully classified
2656 		 * protocol chain. We can deal with it in
2657 		 * one shot.
2658 		 */
2659 		proc = srs_rx->sr_func;
2660 		arg1 = srs_rx->sr_arg1;
2661 		arg2 = srs_rx->sr_arg2;
2662 
2663 		mac_srs->srs_state |= SRS_CLIENT_PROC;
2664 		mutex_exit(&mac_srs->srs_lock);
2665 		if (tid != NULL) {
2666 			(void) untimeout(tid);
2667 			tid = NULL;
2668 		}
2669 
2670 		proc(arg1, arg2, head, NULL);
2671 		/*
2672 		 * Decrement the size and count here itelf
2673 		 * since the packet has been processed.
2674 		 */
2675 		mutex_enter(&mac_srs->srs_lock);
2676 		MAC_UPDATE_SRS_COUNT_LOCKED(mac_srs, cnt);
2677 		if (mac_srs->srs_state & SRS_CLIENT_WAIT)
2678 			cv_signal(&mac_srs->srs_client_cv);
2679 		mac_srs->srs_state &= ~SRS_CLIENT_PROC;
2680 	} else {
2681 		/* Some kind of softrings based fanout is required */
2682 		mutex_exit(&mac_srs->srs_lock);
2683 		if (tid != NULL) {
2684 			(void) untimeout(tid);
2685 			tid = NULL;
2686 		}
2687 
2688 		/*
2689 		 * Since the fanout routines can deal with chains,
2690 		 * shoot the entire chain up.
2691 		 */
2692 		if (mac_srs->srs_type & SRST_FANOUT_SRC_IP)
2693 			mac_rx_srs_fanout(mac_srs, head);
2694 		else
2695 			mac_rx_srs_proto_fanout(mac_srs, head);
2696 		mutex_enter(&mac_srs->srs_lock);
2697 	}
2698 
2699 	if (!(mac_srs->srs_state & (SRS_BLANK|SRS_PAUSE)) &&
2700 	    (mac_srs->srs_first != NULL)) {
2701 		/*
2702 		 * More packets arrived while we were clearing the
2703 		 * SRS. This can be possible because of one of
2704 		 * three conditions below:
2705 		 * 1) The driver is using multiple worker threads
2706 		 *    to send the packets to us.
2707 		 * 2) The driver has a race in switching
2708 		 *    between interrupt and polling mode or
2709 		 * 3) Packets are arriving in this SRS via the
2710 		 *    S/W classification as well.
2711 		 *
2712 		 * We should switch to polling mode and see if we
2713 		 * need to send the poll thread down. Also, signal
2714 		 * the worker thread to process whats just arrived.
2715 		 */
2716 		MAC_SRS_POLLING_ON(mac_srs);
2717 		if (srs_rx->sr_poll_pkt_cnt <= srs_rx->sr_lowat) {
2718 			srs_rx->sr_drain_poll_sig++;
2719 			MAC_SRS_POLL_RING(mac_srs);
2720 		}
2721 
2722 		/*
2723 		 * If we didn't signal the poll thread, we need
2724 		 * to deal with the pending packets ourselves.
2725 		 */
2726 		if (proc_type == SRS_WORKER) {
2727 			srs_rx->sr_drain_again++;
2728 			goto again;
2729 		} else {
2730 			srs_rx->sr_drain_worker_sig++;
2731 			cv_signal(&mac_srs->srs_async);
2732 		}
2733 	}
2734 
2735 out:
2736 	if (mac_srs->srs_state & SRS_GET_PKTS) {
2737 		/*
2738 		 * Poll thread is already running. Leave the
2739 		 * SRS_RPOC set and hand over the control to
2740 		 * poll thread.
2741 		 */
2742 		mac_srs->srs_state &= ~proc_type;
2743 		srs_rx->sr_drain_poll_running++;
2744 		return;
2745 	}
2746 
2747 	/*
2748 	 * Even if there are no packets queued in SRS, we
2749 	 * need to make sure that the shared counter is
2750 	 * clear and any associated softrings have cleared
2751 	 * all the backlog. Otherwise, leave the interface
2752 	 * in polling mode and the poll thread will get
2753 	 * signalled once the count goes down to zero.
2754 	 *
2755 	 * If someone is already draining the queue (SRS_PROC is
2756 	 * set) when the srs_poll_pkt_cnt goes down to zero,
2757 	 * then it means that drain is already running and we
2758 	 * will turn off polling at that time if there is
2759 	 * no backlog.
2760 	 *
2761 	 * As long as there are packets queued either
2762 	 * in soft ring set or its soft rings, we will leave
2763 	 * the interface in polling mode (even if the drain
2764 	 * was done being the interrupt thread). We signal
2765 	 * the poll thread as well if we have dipped below
2766 	 * low water mark.
2767 	 *
2768 	 * NOTE: We can't use the MAC_SRS_POLLING_ON macro
2769 	 * since that turn polling on only for worker thread.
2770 	 * Its not worth turning polling on for interrupt
2771 	 * thread (since NIC will not issue another interrupt)
2772 	 * unless a backlog builds up.
2773 	 */
2774 	if ((srs_rx->sr_poll_pkt_cnt > 0) &&
2775 	    (mac_srs->srs_state & SRS_POLLING_CAPAB)) {
2776 		mac_srs->srs_state &= ~(SRS_PROC|proc_type);
2777 		srs_rx->sr_drain_keep_polling++;
2778 		MAC_SRS_POLLING_ON(mac_srs);
2779 		if (srs_rx->sr_poll_pkt_cnt <= srs_rx->sr_lowat)
2780 			MAC_SRS_POLL_RING(mac_srs);
2781 		return;
2782 	}
2783 
2784 	/* Nothing else to do. Get out of poll mode */
2785 	MAC_SRS_POLLING_OFF(mac_srs);
2786 	mac_srs->srs_state &= ~(SRS_PROC|proc_type);
2787 	srs_rx->sr_drain_finish_intr++;
2788 }
2789 
2790 /*
2791  * mac_rx_srs_drain_bw
2792  *
2793  * The SRS BW drain routine. Gets to run to clear the queue. Any thread
2794  * (worker, interrupt, poll) can call this based on processing model.
2795  * The first thing we do is disable interrupts if possible and then
2796  * drain the queue. we also try to poll the underlying hardware if
2797  * there is a dedicated hardware Rx ring assigned to this SRS.
2798  *
2799  * There is a equivalent drain routine in non bandwidth control mode
2800  * mac_rx_srs_drain. There is some code duplication between the two
2801  * routines but they are highly performance sensitive and are easier
2802  * to read/debug if they stay separate. Any code changes here might
2803  * also apply to mac_rx_srs_drain as well.
2804  */
2805 void
mac_rx_srs_drain_bw(mac_soft_ring_set_t * mac_srs,uint_t proc_type)2806 mac_rx_srs_drain_bw(mac_soft_ring_set_t *mac_srs, uint_t proc_type)
2807 {
2808 	mblk_t			*head;
2809 	mblk_t			*tail;
2810 	timeout_id_t		tid;
2811 	size_t			sz = 0;
2812 	int			cnt = 0;
2813 	mac_client_impl_t	*mcip = mac_srs->srs_mcip;
2814 	mac_srs_rx_t		*srs_rx = &mac_srs->srs_rx;
2815 	clock_t			now;
2816 
2817 	ASSERT(MUTEX_HELD(&mac_srs->srs_lock));
2818 	ASSERT(mac_srs->srs_type & SRST_BW_CONTROL);
2819 again:
2820 	/* Check if we are doing B/W control */
2821 	mutex_enter(&mac_srs->srs_bw->mac_bw_lock);
2822 	now = ddi_get_lbolt();
2823 	if (mac_srs->srs_bw->mac_bw_curr_time != now) {
2824 		mac_srs->srs_bw->mac_bw_curr_time = now;
2825 		mac_srs->srs_bw->mac_bw_used = 0;
2826 		if (mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED)
2827 			mac_srs->srs_bw->mac_bw_state &= ~SRS_BW_ENFORCED;
2828 	} else if (mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED) {
2829 		mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
2830 		goto done;
2831 	} else if (mac_srs->srs_bw->mac_bw_used >
2832 	    mac_srs->srs_bw->mac_bw_limit) {
2833 		mac_srs->srs_bw->mac_bw_state |= SRS_BW_ENFORCED;
2834 		mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
2835 		goto done;
2836 	}
2837 	mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
2838 
2839 	/* If we are blanked i.e. can't do upcalls, then we are done */
2840 	if (mac_srs->srs_state & (SRS_BLANK | SRS_PAUSE)) {
2841 		ASSERT((mac_srs->srs_type & SRST_NO_SOFT_RINGS) ||
2842 		    (mac_srs->srs_state & SRS_PAUSE));
2843 		goto done;
2844 	}
2845 
2846 	sz = 0;
2847 	cnt = 0;
2848 	if ((head = mac_srs_pick_chain(mac_srs, &tail, &sz, &cnt)) == NULL) {
2849 		/*
2850 		 * We couldn't pick up a single packet.
2851 		 */
2852 		mutex_enter(&mac_srs->srs_bw->mac_bw_lock);
2853 		if ((mac_srs->srs_bw->mac_bw_used == 0) &&
2854 		    (mac_srs->srs_size != 0) &&
2855 		    !(mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED)) {
2856 			/*
2857 			 * Seems like configured B/W doesn't
2858 			 * even allow processing of 1 packet
2859 			 * per tick.
2860 			 *
2861 			 * XXX: raise the limit to processing
2862 			 * at least 1 packet per tick.
2863 			 */
2864 			mac_srs->srs_bw->mac_bw_limit +=
2865 			    mac_srs->srs_bw->mac_bw_limit;
2866 			mac_srs->srs_bw->mac_bw_drop_threshold +=
2867 			    mac_srs->srs_bw->mac_bw_drop_threshold;
2868 			cmn_err(CE_NOTE, "mac_rx_srs_drain: srs(%p) "
2869 			    "raised B/W limit to %d since not even a "
2870 			    "single packet can be processed per "
2871 			    "tick %d\n", (void *)mac_srs,
2872 			    (int)mac_srs->srs_bw->mac_bw_limit,
2873 			    (int)msgdsize(mac_srs->srs_first));
2874 		}
2875 		mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
2876 		goto done;
2877 	}
2878 
2879 	ASSERT(head != NULL);
2880 	ASSERT(tail != NULL);
2881 
2882 	/* zero bandwidth: drop all and return to interrupt mode */
2883 	mutex_enter(&mac_srs->srs_bw->mac_bw_lock);
2884 	if (mac_srs->srs_bw->mac_bw_limit == 0) {
2885 		srs_rx->sr_stat.mrs_sdrops += cnt;
2886 		ASSERT(mac_srs->srs_bw->mac_bw_sz >= sz);
2887 		mac_srs->srs_bw->mac_bw_sz -= sz;
2888 		mac_srs->srs_bw->mac_bw_drop_bytes += sz;
2889 		mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
2890 		mac_drop_chain(head, "Rx no bandwidth");
2891 		goto leave_poll;
2892 	} else {
2893 		mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
2894 	}
2895 
2896 	if ((tid = mac_srs->srs_tid) != NULL)
2897 		mac_srs->srs_tid = NULL;
2898 
2899 	mac_srs->srs_state |= (SRS_PROC|proc_type);
2900 	MAC_SRS_WORKER_POLLING_ON(mac_srs);
2901 
2902 	/*
2903 	 * mcip is NULL for broadcast and multicast flows. The promisc
2904 	 * callbacks for broadcast and multicast packets are delivered from
2905 	 * mac_rx() and we don't need to worry about that case in this path
2906 	 */
2907 	if (mcip != NULL) {
2908 		if (mcip->mci_promisc_list != NULL) {
2909 			mutex_exit(&mac_srs->srs_lock);
2910 			mac_promisc_client_dispatch(mcip, head);
2911 			mutex_enter(&mac_srs->srs_lock);
2912 		}
2913 		if (MAC_PROTECT_ENABLED(mcip, MPT_IPNOSPOOF)) {
2914 			mutex_exit(&mac_srs->srs_lock);
2915 			mac_protect_intercept_dynamic(mcip, head);
2916 			mutex_enter(&mac_srs->srs_lock);
2917 		}
2918 	}
2919 
2920 	/*
2921 	 * Check if SRS itself is doing the processing
2922 	 * This direct path does not apply when subflows are present. In this
2923 	 * case, packets need to be dispatched to a soft ring according to the
2924 	 * flow's bandwidth and other resources contraints.
2925 	 */
2926 	if (mac_srs->srs_type & SRST_NO_SOFT_RINGS) {
2927 		mac_direct_rx_t		proc;
2928 		void			*arg1;
2929 		mac_resource_handle_t	arg2;
2930 
2931 		/*
2932 		 * This is the case when a Rx is directly
2933 		 * assigned and we have a fully classified
2934 		 * protocol chain. We can deal with it in
2935 		 * one shot.
2936 		 */
2937 		proc = srs_rx->sr_func;
2938 		arg1 = srs_rx->sr_arg1;
2939 		arg2 = srs_rx->sr_arg2;
2940 
2941 		mac_srs->srs_state |= SRS_CLIENT_PROC;
2942 		mutex_exit(&mac_srs->srs_lock);
2943 		if (tid != NULL) {
2944 			(void) untimeout(tid);
2945 			tid = NULL;
2946 		}
2947 
2948 		proc(arg1, arg2, head, NULL);
2949 		/*
2950 		 * Decrement the size and count here itelf
2951 		 * since the packet has been processed.
2952 		 */
2953 		mutex_enter(&mac_srs->srs_lock);
2954 		MAC_UPDATE_SRS_COUNT_LOCKED(mac_srs, cnt);
2955 		MAC_UPDATE_SRS_SIZE_LOCKED(mac_srs, sz);
2956 
2957 		if (mac_srs->srs_state & SRS_CLIENT_WAIT)
2958 			cv_signal(&mac_srs->srs_client_cv);
2959 		mac_srs->srs_state &= ~SRS_CLIENT_PROC;
2960 	} else {
2961 		/* Some kind of softrings based fanout is required */
2962 		mutex_exit(&mac_srs->srs_lock);
2963 		if (tid != NULL) {
2964 			(void) untimeout(tid);
2965 			tid = NULL;
2966 		}
2967 
2968 		/*
2969 		 * Since the fanout routines can deal with chains,
2970 		 * shoot the entire chain up.
2971 		 */
2972 		if (mac_srs->srs_type & SRST_FANOUT_SRC_IP)
2973 			mac_rx_srs_fanout(mac_srs, head);
2974 		else
2975 			mac_rx_srs_proto_fanout(mac_srs, head);
2976 		mutex_enter(&mac_srs->srs_lock);
2977 	}
2978 
2979 	/*
2980 	 * Send the poll thread to pick up any packets arrived
2981 	 * so far. This also serves as the last check in case
2982 	 * nothing else is queued in the SRS. The poll thread
2983 	 * is signalled only in the case the drain was done
2984 	 * by the worker thread and SRS_WORKER is set. The
2985 	 * worker thread can run in parallel as long as the
2986 	 * SRS_WORKER flag is set. We we have nothing else to
2987 	 * process, we can exit while leaving SRS_PROC set
2988 	 * which gives the poll thread control to process and
2989 	 * cleanup once it returns from the NIC.
2990 	 *
2991 	 * If we have nothing else to process, we need to
2992 	 * ensure that we keep holding the srs_lock till
2993 	 * all the checks below are done and control is
2994 	 * handed to the poll thread if it was running.
2995 	 */
2996 	mutex_enter(&mac_srs->srs_bw->mac_bw_lock);
2997 	if (!(mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED)) {
2998 		if (mac_srs->srs_first != NULL) {
2999 			if (proc_type == SRS_WORKER) {
3000 				mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
3001 				if (srs_rx->sr_poll_pkt_cnt <=
3002 				    srs_rx->sr_lowat)
3003 					MAC_SRS_POLL_RING(mac_srs);
3004 				goto again;
3005 			} else {
3006 				cv_signal(&mac_srs->srs_async);
3007 			}
3008 		}
3009 	}
3010 	mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
3011 
3012 done:
3013 
3014 	if (mac_srs->srs_state & SRS_GET_PKTS) {
3015 		/*
3016 		 * Poll thread is already running. Leave the
3017 		 * SRS_RPOC set and hand over the control to
3018 		 * poll thread.
3019 		 */
3020 		mac_srs->srs_state &= ~proc_type;
3021 		return;
3022 	}
3023 
3024 	/*
3025 	 * If we can't process packets because we have exceeded
3026 	 * B/W limit for this tick, just set the timeout
3027 	 * and leave.
3028 	 *
3029 	 * Even if there are no packets queued in SRS, we
3030 	 * need to make sure that the shared counter is
3031 	 * clear and any associated softrings have cleared
3032 	 * all the backlog. Otherwise, leave the interface
3033 	 * in polling mode and the poll thread will get
3034 	 * signalled once the count goes down to zero.
3035 	 *
3036 	 * If someone is already draining the queue (SRS_PROC is
3037 	 * set) when the srs_poll_pkt_cnt goes down to zero,
3038 	 * then it means that drain is already running and we
3039 	 * will turn off polling at that time if there is
3040 	 * no backlog. As long as there are packets queued either
3041 	 * is soft ring set or its soft rings, we will leave
3042 	 * the interface in polling mode.
3043 	 */
3044 	mutex_enter(&mac_srs->srs_bw->mac_bw_lock);
3045 	if ((mac_srs->srs_state & SRS_POLLING_CAPAB) &&
3046 	    ((mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED) ||
3047 	    (srs_rx->sr_poll_pkt_cnt > 0))) {
3048 		MAC_SRS_POLLING_ON(mac_srs);
3049 		mac_srs->srs_state &= ~(SRS_PROC|proc_type);
3050 		if ((mac_srs->srs_first != NULL) &&
3051 		    (mac_srs->srs_tid == NULL))
3052 			mac_srs->srs_tid = timeout(mac_srs_fire,
3053 			    mac_srs, 1);
3054 		mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
3055 		return;
3056 	}
3057 	mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
3058 
3059 leave_poll:
3060 
3061 	/* Nothing else to do. Get out of poll mode */
3062 	MAC_SRS_POLLING_OFF(mac_srs);
3063 	mac_srs->srs_state &= ~(SRS_PROC|proc_type);
3064 }
3065 
3066 /*
3067  * mac_srs_worker
3068  *
3069  * The SRS worker routine. Drains the queue when no one else is
3070  * processing it.
3071  */
3072 void
mac_srs_worker(mac_soft_ring_set_t * mac_srs)3073 mac_srs_worker(mac_soft_ring_set_t *mac_srs)
3074 {
3075 	kmutex_t		*lock = &mac_srs->srs_lock;
3076 	kcondvar_t		*async = &mac_srs->srs_async;
3077 	callb_cpr_t		cprinfo;
3078 	boolean_t		bw_ctl_flag;
3079 
3080 	CALLB_CPR_INIT(&cprinfo, lock, callb_generic_cpr, "srs_worker");
3081 	mutex_enter(lock);
3082 
3083 start:
3084 	for (;;) {
3085 		bw_ctl_flag = B_FALSE;
3086 		if (mac_srs->srs_type & SRST_BW_CONTROL) {
3087 			MAC_SRS_BW_LOCK(mac_srs);
3088 			MAC_SRS_CHECK_BW_CONTROL(mac_srs);
3089 			if (mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED)
3090 				bw_ctl_flag = B_TRUE;
3091 			MAC_SRS_BW_UNLOCK(mac_srs);
3092 		}
3093 		/*
3094 		 * The SRS_BW_ENFORCED flag may change since we have dropped
3095 		 * the mac_bw_lock. However the drain function can handle both
3096 		 * a drainable SRS or a bandwidth controlled SRS, and the
3097 		 * effect of scheduling a timeout is to wakeup the worker
3098 		 * thread which in turn will call the drain function. Since
3099 		 * we release the srs_lock atomically only in the cv_wait there
3100 		 * isn't a fear of waiting for ever.
3101 		 */
3102 		while (((mac_srs->srs_state & SRS_PROC) ||
3103 		    (mac_srs->srs_first == NULL) || bw_ctl_flag ||
3104 		    (mac_srs->srs_state & SRS_TX_BLOCKED)) &&
3105 		    !(mac_srs->srs_state & SRS_PAUSE)) {
3106 			/*
3107 			 * If we have packets queued and we are here
3108 			 * because B/W control is in place, we better
3109 			 * schedule the worker wakeup after 1 tick
3110 			 * to see if bandwidth control can be relaxed.
3111 			 */
3112 			if (bw_ctl_flag && mac_srs->srs_tid == NULL) {
3113 				/*
3114 				 * We need to ensure that a timer  is already
3115 				 * scheduled or we force  schedule one for
3116 				 * later so that we can continue processing
3117 				 * after this  quanta is over.
3118 				 */
3119 				mac_srs->srs_tid = timeout(mac_srs_fire,
3120 				    mac_srs, 1);
3121 			}
3122 wait:
3123 			CALLB_CPR_SAFE_BEGIN(&cprinfo);
3124 			cv_wait(async, lock);
3125 			CALLB_CPR_SAFE_END(&cprinfo, lock);
3126 
3127 			if (mac_srs->srs_state & SRS_PAUSE)
3128 				goto done;
3129 			if (mac_srs->srs_state & SRS_PROC)
3130 				goto wait;
3131 
3132 			if (mac_srs->srs_first != NULL &&
3133 			    mac_srs->srs_type & SRST_BW_CONTROL) {
3134 				MAC_SRS_BW_LOCK(mac_srs);
3135 				if (mac_srs->srs_bw->mac_bw_state &
3136 				    SRS_BW_ENFORCED) {
3137 					MAC_SRS_CHECK_BW_CONTROL(mac_srs);
3138 				}
3139 				bw_ctl_flag = mac_srs->srs_bw->mac_bw_state &
3140 				    SRS_BW_ENFORCED;
3141 				MAC_SRS_BW_UNLOCK(mac_srs);
3142 			}
3143 		}
3144 
3145 		if (mac_srs->srs_state & SRS_PAUSE)
3146 			goto done;
3147 		mac_srs->srs_drain_func(mac_srs, SRS_WORKER);
3148 	}
3149 done:
3150 	/*
3151 	 * The Rx SRS quiesce logic first cuts off packet supply to the SRS
3152 	 * from both hard and soft classifications and waits for such threads
3153 	 * to finish before signaling the worker. So at this point the only
3154 	 * thread left that could be competing with the worker is the poll
3155 	 * thread. In the case of Tx, there shouldn't be any thread holding
3156 	 * SRS_PROC at this point.
3157 	 */
3158 	if (!(mac_srs->srs_state & SRS_PROC)) {
3159 		mac_srs->srs_state |= SRS_PROC;
3160 	} else {
3161 		ASSERT((mac_srs->srs_type & SRST_TX) == 0);
3162 		/*
3163 		 * Poll thread still owns the SRS and is still running
3164 		 */
3165 		ASSERT((mac_srs->srs_poll_thr == NULL) ||
3166 		    ((mac_srs->srs_state & SRS_POLL_THR_OWNER) ==
3167 		    SRS_POLL_THR_OWNER));
3168 	}
3169 	mac_srs_worker_quiesce(mac_srs);
3170 	/*
3171 	 * Wait for the SRS_RESTART or SRS_CONDEMNED signal from the initiator
3172 	 * of the quiesce operation
3173 	 */
3174 	while (!(mac_srs->srs_state & (SRS_CONDEMNED | SRS_RESTART)))
3175 		cv_wait(&mac_srs->srs_async, &mac_srs->srs_lock);
3176 
3177 	if (mac_srs->srs_state & SRS_RESTART) {
3178 		ASSERT(!(mac_srs->srs_state & SRS_CONDEMNED));
3179 		mac_srs_worker_restart(mac_srs);
3180 		mac_srs->srs_state &= ~SRS_PROC;
3181 		goto start;
3182 	}
3183 
3184 	if (!(mac_srs->srs_state & SRS_CONDEMNED_DONE))
3185 		mac_srs_worker_quiesce(mac_srs);
3186 
3187 	mac_srs->srs_state &= ~SRS_PROC;
3188 	/* The macro drops the srs_lock */
3189 	CALLB_CPR_EXIT(&cprinfo);
3190 	thread_exit();
3191 }
3192 
3193 /*
3194  * mac_rx_srs_subflow_process
3195  *
3196  * Receive side routine called from interrupt path when there are
3197  * sub flows present on this SRS.
3198  */
3199 /* ARGSUSED */
3200 void
mac_rx_srs_subflow_process(void * arg,mac_resource_handle_t srs,mblk_t * mp_chain,boolean_t loopback)3201 mac_rx_srs_subflow_process(void *arg, mac_resource_handle_t srs,
3202     mblk_t *mp_chain, boolean_t loopback)
3203 {
3204 	flow_entry_t		*flent = NULL;
3205 	flow_entry_t		*prev_flent = NULL;
3206 	mblk_t			*mp = NULL;
3207 	mblk_t			*tail = NULL;
3208 	mac_soft_ring_set_t	*mac_srs = (mac_soft_ring_set_t *)srs;
3209 	mac_client_impl_t	*mcip;
3210 
3211 	mcip = mac_srs->srs_mcip;
3212 	ASSERT(mcip != NULL);
3213 
3214 	/*
3215 	 * We need to determine the SRS for every packet
3216 	 * by walking the flow table, if we don't get any,
3217 	 * then we proceed using the SRS we came with.
3218 	 */
3219 	mp = tail = mp_chain;
3220 	while (mp != NULL) {
3221 
3222 		/*
3223 		 * We will increment the stats for the mactching subflow.
3224 		 * when we get the bytes/pkt count for the classified packets
3225 		 * later in mac_rx_srs_process.
3226 		 */
3227 		(void) mac_flow_lookup(mcip->mci_subflow_tab, mp,
3228 		    FLOW_INBOUND, &flent);
3229 
3230 		if (mp == mp_chain || flent == prev_flent) {
3231 			if (prev_flent != NULL)
3232 				FLOW_REFRELE(prev_flent);
3233 			prev_flent = flent;
3234 			flent = NULL;
3235 			tail = mp;
3236 			mp = mp->b_next;
3237 			continue;
3238 		}
3239 		tail->b_next = NULL;
3240 		/*
3241 		 * A null indicates, this is for the mac_srs itself.
3242 		 * XXX-venu : probably assert for fe_rx_srs_cnt == 0.
3243 		 */
3244 		if (prev_flent == NULL || prev_flent->fe_rx_srs_cnt == 0) {
3245 			mac_rx_srs_process(arg,
3246 			    (mac_resource_handle_t)mac_srs, mp_chain,
3247 			    loopback);
3248 		} else {
3249 			(prev_flent->fe_cb_fn)(prev_flent->fe_cb_arg1,
3250 			    prev_flent->fe_cb_arg2, mp_chain, loopback);
3251 			FLOW_REFRELE(prev_flent);
3252 		}
3253 		prev_flent = flent;
3254 		flent = NULL;
3255 		mp_chain = mp;
3256 		tail = mp;
3257 		mp = mp->b_next;
3258 	}
3259 	/* Last chain */
3260 	ASSERT(mp_chain != NULL);
3261 	if (prev_flent == NULL || prev_flent->fe_rx_srs_cnt == 0) {
3262 		mac_rx_srs_process(arg,
3263 		    (mac_resource_handle_t)mac_srs, mp_chain, loopback);
3264 	} else {
3265 		(prev_flent->fe_cb_fn)(prev_flent->fe_cb_arg1,
3266 		    prev_flent->fe_cb_arg2, mp_chain, loopback);
3267 		FLOW_REFRELE(prev_flent);
3268 	}
3269 }
3270 
3271 /*
3272  * MAC SRS receive side routine. If the data is coming from the
3273  * network (i.e. from a NIC) then this is called in interrupt context.
3274  * If the data is coming from a local sender (e.g. mac_tx_send() or
3275  * bridge_forward()) then this is not called in interrupt context.
3276  *
3277  * loopback is set to force a context switch on the loopback
3278  * path between MAC clients.
3279  */
3280 /* ARGSUSED */
3281 void
mac_rx_srs_process(void * arg,mac_resource_handle_t srs,mblk_t * mp_chain,boolean_t loopback)3282 mac_rx_srs_process(void *arg, mac_resource_handle_t srs, mblk_t *mp_chain,
3283     boolean_t loopback)
3284 {
3285 	mac_soft_ring_set_t	*mac_srs = (mac_soft_ring_set_t *)srs;
3286 	mblk_t			*mp, *tail, *head;
3287 	int			count = 0;
3288 	int			count1;
3289 	size_t			sz = 0;
3290 	size_t			chain_sz, sz1;
3291 	mac_bw_ctl_t		*mac_bw;
3292 	mac_srs_rx_t		*srs_rx = &mac_srs->srs_rx;
3293 
3294 	/*
3295 	 * Set the tail, count and sz. We set the sz irrespective
3296 	 * of whether we are doing B/W control or not for the
3297 	 * purpose of updating the stats.
3298 	 */
3299 	mp = tail = mp_chain;
3300 	while (mp != NULL) {
3301 		tail = mp;
3302 		count++;
3303 		sz += msgdsize(mp);
3304 		mp = mp->b_next;
3305 	}
3306 
3307 	mutex_enter(&mac_srs->srs_lock);
3308 
3309 	if (loopback) {
3310 		SRS_RX_STAT_UPDATE(mac_srs, lclbytes, sz);
3311 		SRS_RX_STAT_UPDATE(mac_srs, lclcnt, count);
3312 
3313 	} else {
3314 		SRS_RX_STAT_UPDATE(mac_srs, intrbytes, sz);
3315 		SRS_RX_STAT_UPDATE(mac_srs, intrcnt, count);
3316 	}
3317 
3318 	/*
3319 	 * If the SRS in already being processed; has been blanked;
3320 	 * can be processed by worker thread only; or the B/W limit
3321 	 * has been reached, then queue the chain and check if
3322 	 * worker thread needs to be awakend.
3323 	 */
3324 	if (mac_srs->srs_type & SRST_BW_CONTROL) {
3325 		mac_bw = mac_srs->srs_bw;
3326 		ASSERT(mac_bw != NULL);
3327 		mutex_enter(&mac_bw->mac_bw_lock);
3328 		mac_bw->mac_bw_intr += sz;
3329 		if (mac_bw->mac_bw_limit == 0) {
3330 			/* zero bandwidth: drop all */
3331 			srs_rx->sr_stat.mrs_sdrops += count;
3332 			mac_bw->mac_bw_drop_bytes += sz;
3333 			mutex_exit(&mac_bw->mac_bw_lock);
3334 			mutex_exit(&mac_srs->srs_lock);
3335 			mac_drop_chain(mp_chain, "Rx no bandwidth");
3336 			return;
3337 		} else {
3338 			if ((mac_bw->mac_bw_sz + sz) <=
3339 			    mac_bw->mac_bw_drop_threshold) {
3340 				mutex_exit(&mac_bw->mac_bw_lock);
3341 				MAC_RX_SRS_ENQUEUE_CHAIN(mac_srs, mp_chain,
3342 				    tail, count, sz);
3343 			} else {
3344 				mp = mp_chain;
3345 				chain_sz = 0;
3346 				count1 = 0;
3347 				tail = NULL;
3348 				head = NULL;
3349 				while (mp != NULL) {
3350 					sz1 = msgdsize(mp);
3351 					if (mac_bw->mac_bw_sz + chain_sz + sz1 >
3352 					    mac_bw->mac_bw_drop_threshold)
3353 						break;
3354 					chain_sz += sz1;
3355 					count1++;
3356 					tail = mp;
3357 					mp = mp->b_next;
3358 				}
3359 				mutex_exit(&mac_bw->mac_bw_lock);
3360 				if (tail != NULL) {
3361 					head = tail->b_next;
3362 					tail->b_next = NULL;
3363 					MAC_RX_SRS_ENQUEUE_CHAIN(mac_srs,
3364 					    mp_chain, tail, count1, chain_sz);
3365 					sz -= chain_sz;
3366 					count -= count1;
3367 				} else {
3368 					/* Can't pick up any */
3369 					head = mp_chain;
3370 				}
3371 				if (head != NULL) {
3372 					/* Drop any packet over the threshold */
3373 					srs_rx->sr_stat.mrs_sdrops += count;
3374 					mutex_enter(&mac_bw->mac_bw_lock);
3375 					mac_bw->mac_bw_drop_bytes += sz;
3376 					mutex_exit(&mac_bw->mac_bw_lock);
3377 					freemsgchain(head);
3378 				}
3379 			}
3380 			MAC_SRS_WORKER_WAKEUP(mac_srs);
3381 			mutex_exit(&mac_srs->srs_lock);
3382 			return;
3383 		}
3384 	}
3385 
3386 	/*
3387 	 * If the total number of packets queued in the SRS and
3388 	 * its associated soft rings exceeds the max allowed,
3389 	 * then drop the chain. If we are polling capable, this
3390 	 * shouldn't be happening.
3391 	 */
3392 	if (!(mac_srs->srs_type & SRST_BW_CONTROL) &&
3393 	    (srs_rx->sr_poll_pkt_cnt > srs_rx->sr_hiwat)) {
3394 		mac_bw = mac_srs->srs_bw;
3395 		srs_rx->sr_stat.mrs_sdrops += count;
3396 		mutex_enter(&mac_bw->mac_bw_lock);
3397 		mac_bw->mac_bw_drop_bytes += sz;
3398 		mutex_exit(&mac_bw->mac_bw_lock);
3399 		freemsgchain(mp_chain);
3400 		mutex_exit(&mac_srs->srs_lock);
3401 		return;
3402 	}
3403 
3404 	MAC_RX_SRS_ENQUEUE_CHAIN(mac_srs, mp_chain, tail, count, sz);
3405 
3406 	if (!(mac_srs->srs_state & SRS_PROC)) {
3407 		/*
3408 		 * If we are coming via loopback, if we are not optimizing for
3409 		 * latency, or if our stack is running deep, we should signal
3410 		 * the worker thread.
3411 		 */
3412 		if (loopback || !(mac_srs->srs_state & SRS_LATENCY_OPT)) {
3413 			/*
3414 			 * For loopback, We need to let the worker take
3415 			 * over as we don't want to continue in the same
3416 			 * thread even if we can. This could lead to stack
3417 			 * overflows and may also end up using
3418 			 * resources (cpu) incorrectly.
3419 			 */
3420 			cv_signal(&mac_srs->srs_async);
3421 		} else if (STACK_BIAS + (uintptr_t)getfp() -
3422 		    (uintptr_t)curthread->t_stkbase < mac_rx_srs_stack_needed) {
3423 			if (++mac_rx_srs_stack_toodeep == 0)
3424 				mac_rx_srs_stack_toodeep = 1;
3425 			cv_signal(&mac_srs->srs_async);
3426 		} else {
3427 			/*
3428 			 * Seems like no one is processing the SRS and
3429 			 * there is no backlog. We also inline process
3430 			 * our packet if its a single packet in non
3431 			 * latency optimized case (in latency optimized
3432 			 * case, we inline process chains of any size).
3433 			 */
3434 			mac_srs->srs_drain_func(mac_srs, SRS_PROC_FAST);
3435 		}
3436 	}
3437 	mutex_exit(&mac_srs->srs_lock);
3438 }
3439 
3440 /* TX SIDE ROUTINES (RUNTIME) */
3441 
3442 /*
3443  * mac_tx_srs_no_desc
3444  *
3445  * This routine is called by Tx single ring default mode
3446  * when Tx ring runs out of descs.
3447  */
3448 mac_tx_cookie_t
mac_tx_srs_no_desc(mac_soft_ring_set_t * mac_srs,mblk_t * mp_chain,uint16_t flag,mblk_t ** ret_mp)3449 mac_tx_srs_no_desc(mac_soft_ring_set_t *mac_srs, mblk_t *mp_chain,
3450     uint16_t flag, mblk_t **ret_mp)
3451 {
3452 	mac_tx_cookie_t cookie = 0;
3453 	mac_srs_tx_t *srs_tx = &mac_srs->srs_tx;
3454 	boolean_t wakeup_worker = B_TRUE;
3455 	uint32_t tx_mode = srs_tx->st_mode;
3456 	int cnt, sz;
3457 	mblk_t *tail;
3458 
3459 	ASSERT(tx_mode == SRS_TX_DEFAULT || tx_mode == SRS_TX_BW);
3460 	if (flag & MAC_DROP_ON_NO_DESC) {
3461 		MAC_TX_SRS_DROP_MESSAGE(mac_srs, mp_chain, cookie,
3462 		    "Tx no desc");
3463 	} else {
3464 		if (mac_srs->srs_first != NULL)
3465 			wakeup_worker = B_FALSE;
3466 		MAC_COUNT_CHAIN(mac_srs, mp_chain, tail, cnt, sz);
3467 		if (flag & MAC_TX_NO_ENQUEUE) {
3468 			/*
3469 			 * If TX_QUEUED is not set, queue the
3470 			 * packet and let mac_tx_srs_drain()
3471 			 * set the TX_BLOCKED bit for the
3472 			 * reasons explained above. Otherwise,
3473 			 * return the mblks.
3474 			 */
3475 			if (wakeup_worker) {
3476 				MAC_TX_SRS_ENQUEUE_CHAIN(mac_srs,
3477 				    mp_chain, tail, cnt, sz);
3478 			} else {
3479 				MAC_TX_SET_NO_ENQUEUE(mac_srs,
3480 				    mp_chain, ret_mp, cookie);
3481 			}
3482 		} else {
3483 			MAC_TX_SRS_TEST_HIWAT(mac_srs, mp_chain,
3484 			    tail, cnt, sz, cookie);
3485 		}
3486 		if (wakeup_worker)
3487 			cv_signal(&mac_srs->srs_async);
3488 	}
3489 	return (cookie);
3490 }
3491 
3492 /*
3493  * mac_tx_srs_enqueue
3494  *
3495  * This routine is called when Tx SRS is operating in either serializer
3496  * or bandwidth mode. In serializer mode, a packet will get enqueued
3497  * when a thread cannot enter SRS exclusively. In bandwidth mode,
3498  * packets gets queued if allowed byte-count limit for a tick is
3499  * exceeded. The action that gets taken when MAC_DROP_ON_NO_DESC and
3500  * MAC_TX_NO_ENQUEUE is set is different than when operaing in either
3501  * the default mode or fanout mode. Here packets get dropped or
3502  * returned back to the caller only after hi-watermark worth of data
3503  * is queued.
3504  */
3505 static mac_tx_cookie_t
mac_tx_srs_enqueue(mac_soft_ring_set_t * mac_srs,mblk_t * mp_chain,uint16_t flag,uintptr_t fanout_hint,mblk_t ** ret_mp)3506 mac_tx_srs_enqueue(mac_soft_ring_set_t *mac_srs, mblk_t *mp_chain,
3507     uint16_t flag, uintptr_t fanout_hint, mblk_t **ret_mp)
3508 {
3509 	mac_tx_cookie_t cookie = 0;
3510 	int cnt, sz;
3511 	mblk_t *tail;
3512 	boolean_t wakeup_worker = B_TRUE;
3513 
3514 	/*
3515 	 * Ignore fanout hint if we don't have multiple tx rings.
3516 	 */
3517 	if (!MAC_TX_SOFT_RINGS(mac_srs))
3518 		fanout_hint = 0;
3519 
3520 	if (mac_srs->srs_first != NULL)
3521 		wakeup_worker = B_FALSE;
3522 	MAC_COUNT_CHAIN(mac_srs, mp_chain, tail, cnt, sz);
3523 	if (flag & MAC_DROP_ON_NO_DESC) {
3524 		if (mac_srs->srs_count > mac_srs->srs_tx.st_hiwat) {
3525 			MAC_TX_SRS_DROP_MESSAGE(mac_srs, mp_chain, cookie,
3526 			    "Tx SRS hiwat");
3527 		} else {
3528 			MAC_TX_SRS_ENQUEUE_CHAIN(mac_srs,
3529 			    mp_chain, tail, cnt, sz);
3530 		}
3531 	} else if (flag & MAC_TX_NO_ENQUEUE) {
3532 		if ((mac_srs->srs_count > mac_srs->srs_tx.st_hiwat) ||
3533 		    (mac_srs->srs_state & SRS_TX_WAKEUP_CLIENT)) {
3534 			MAC_TX_SET_NO_ENQUEUE(mac_srs, mp_chain,
3535 			    ret_mp, cookie);
3536 		} else {
3537 			mp_chain->b_prev = (mblk_t *)fanout_hint;
3538 			MAC_TX_SRS_ENQUEUE_CHAIN(mac_srs,
3539 			    mp_chain, tail, cnt, sz);
3540 		}
3541 	} else {
3542 		/*
3543 		 * If you are BW_ENFORCED, just enqueue the
3544 		 * packet. srs_worker will drain it at the
3545 		 * prescribed rate. Before enqueueing, save
3546 		 * the fanout hint.
3547 		 */
3548 		mp_chain->b_prev = (mblk_t *)fanout_hint;
3549 		MAC_TX_SRS_TEST_HIWAT(mac_srs, mp_chain,
3550 		    tail, cnt, sz, cookie);
3551 	}
3552 	if (wakeup_worker)
3553 		cv_signal(&mac_srs->srs_async);
3554 	return (cookie);
3555 }
3556 
3557 /*
3558  * There are seven tx modes:
3559  *
3560  * 1) Default mode (SRS_TX_DEFAULT)
3561  * 2) Serialization mode (SRS_TX_SERIALIZE)
3562  * 3) Fanout mode (SRS_TX_FANOUT)
3563  * 4) Bandwdith mode (SRS_TX_BW)
3564  * 5) Fanout and Bandwidth mode (SRS_TX_BW_FANOUT)
3565  * 6) aggr Tx mode (SRS_TX_AGGR)
3566  * 7) aggr Tx bw mode (SRS_TX_BW_AGGR)
3567  *
3568  * The tx mode in which an SRS operates is decided in mac_tx_srs_setup()
3569  * based on the number of Tx rings requested for an SRS and whether
3570  * bandwidth control is requested or not.
3571  *
3572  * The default mode (i.e., no fanout/no bandwidth) is used when the
3573  * underlying NIC does not have Tx rings or just one Tx ring. In this mode,
3574  * the SRS acts as a pass-thru. Packets will go directly to mac_tx_send().
3575  * When the underlying Tx ring runs out of Tx descs, it starts queueing up
3576  * packets in SRS. When flow-control is relieved, the srs_worker drains
3577  * the queued packets and informs blocked clients to restart sending
3578  * packets.
3579  *
3580  * In the SRS_TX_SERIALIZE mode, all calls to mac_tx() are serialized. This
3581  * mode is used when the link has no Tx rings or only one Tx ring.
3582  *
3583  * In the SRS_TX_FANOUT mode, packets will be fanned out to multiple
3584  * Tx rings. Each Tx ring will have a soft ring associated with it.
3585  * These soft rings will be hung off the Tx SRS. Queueing if it happens
3586  * due to lack of Tx desc will be in individual soft ring (and not srs)
3587  * associated with Tx ring.
3588  *
3589  * In the TX_BW mode, tx srs will allow packets to go down to Tx ring
3590  * only if bw is available. Otherwise the packets will be queued in
3591  * SRS. If fanout to multiple Tx rings is configured, the packets will
3592  * be fanned out among the soft rings associated with the Tx rings.
3593  *
3594  * In SRS_TX_AGGR mode, mac_tx_aggr_mode() routine is called. This routine
3595  * invokes an aggr function, aggr_find_tx_ring(), to find a pseudo Tx ring
3596  * belonging to a port on which the packet has to be sent. Aggr will
3597  * always have a pseudo Tx ring associated with it even when it is an
3598  * aggregation over a single NIC that has no Tx rings. Even in such a
3599  * case, the single pseudo Tx ring will have a soft ring associated with
3600  * it and the soft ring will hang off the SRS.
3601  *
3602  * If a bandwidth is specified for an aggr, SRS_TX_BW_AGGR mode is used.
3603  * In this mode, the bandwidth is first applied on the outgoing packets
3604  * and later mac_tx_addr_mode() function is called to send the packet out
3605  * of one of the pseudo Tx rings.
3606  *
3607  * Four flags are used in srs_state for indicating flow control
3608  * conditions : SRS_TX_BLOCKED, SRS_TX_HIWAT, SRS_TX_WAKEUP_CLIENT.
3609  * SRS_TX_BLOCKED indicates out of Tx descs. SRS expects a wakeup from the
3610  * driver below.
3611  * SRS_TX_HIWAT indicates packet count enqueued in Tx SRS exceeded Tx hiwat
3612  * and flow-control pressure is applied back to clients. The clients expect
3613  * wakeup when flow-control is relieved.
3614  * SRS_TX_WAKEUP_CLIENT get set when (flag == MAC_TX_NO_ENQUEUE) and mblk
3615  * got returned back to client either due to lack of Tx descs or due to bw
3616  * control reasons. The clients expect a wakeup when condition is relieved.
3617  *
3618  * The fourth argument to mac_tx() is the flag. Normally it will be 0 but
3619  * some clients set the following values too: MAC_DROP_ON_NO_DESC,
3620  * MAC_TX_NO_ENQUEUE
3621  * Mac clients that do not want packets to be enqueued in the mac layer set
3622  * MAC_DROP_ON_NO_DESC value. The packets won't be queued in the Tx SRS or
3623  * Tx soft rings but instead get dropped when the NIC runs out of desc. The
3624  * behaviour of this flag is different when the Tx is running in serializer
3625  * or bandwidth mode. Under these (Serializer, bandwidth) modes, the packet
3626  * get dropped when Tx high watermark is reached.
3627  * There are some mac clients like vsw, aggr that want the mblks to be
3628  * returned back to clients instead of being queued in Tx SRS (or Tx soft
3629  * rings) under flow-control (i.e., out of desc or exceeding bw limits)
3630  * conditions. These clients call mac_tx() with MAC_TX_NO_ENQUEUE flag set.
3631  * In the default and Tx fanout mode, the un-transmitted mblks will be
3632  * returned back to the clients when the driver runs out of Tx descs.
3633  * SRS_TX_WAKEUP_CLIENT (or S_RING_WAKEUP_CLIENT) will be set in SRS (or
3634  * soft ring) so that the clients can be woken up when Tx desc become
3635  * available. When running in serializer or bandwidth mode mode,
3636  * SRS_TX_WAKEUP_CLIENT will be set when tx hi-watermark is reached.
3637  */
3638 
3639 mac_tx_func_t
mac_tx_get_func(uint32_t mode)3640 mac_tx_get_func(uint32_t mode)
3641 {
3642 	return (mac_tx_mode_list[mode].mac_tx_func);
3643 }
3644 
3645 /* ARGSUSED */
3646 static mac_tx_cookie_t
mac_tx_single_ring_mode(mac_soft_ring_set_t * mac_srs,mblk_t * mp_chain,uintptr_t fanout_hint,uint16_t flag,mblk_t ** ret_mp)3647 mac_tx_single_ring_mode(mac_soft_ring_set_t *mac_srs, mblk_t *mp_chain,
3648     uintptr_t fanout_hint, uint16_t flag, mblk_t **ret_mp)
3649 {
3650 	mac_srs_tx_t		*srs_tx = &mac_srs->srs_tx;
3651 	mac_tx_stats_t		stats;
3652 	mac_tx_cookie_t		cookie = 0;
3653 
3654 	ASSERT(srs_tx->st_mode == SRS_TX_DEFAULT);
3655 
3656 	/* Regular case with a single Tx ring */
3657 	/*
3658 	 * SRS_TX_BLOCKED is set when underlying NIC runs
3659 	 * out of Tx descs and messages start getting
3660 	 * queued. It won't get reset until
3661 	 * tx_srs_drain() completely drains out the
3662 	 * messages.
3663 	 */
3664 	if ((mac_srs->srs_state & SRS_ENQUEUED) != 0) {
3665 		/* Tx descs/resources not available */
3666 		mutex_enter(&mac_srs->srs_lock);
3667 		if ((mac_srs->srs_state & SRS_ENQUEUED) != 0) {
3668 			cookie = mac_tx_srs_no_desc(mac_srs, mp_chain,
3669 			    flag, ret_mp);
3670 			mutex_exit(&mac_srs->srs_lock);
3671 			return (cookie);
3672 		}
3673 		/*
3674 		 * While we were computing mblk count, the
3675 		 * flow control condition got relieved.
3676 		 * Continue with the transmission.
3677 		 */
3678 		mutex_exit(&mac_srs->srs_lock);
3679 	}
3680 
3681 	mp_chain = mac_tx_send(srs_tx->st_arg1, srs_tx->st_arg2,
3682 	    mp_chain, &stats);
3683 
3684 	/*
3685 	 * Multiple threads could be here sending packets.
3686 	 * Under such conditions, it is not possible to
3687 	 * automically set SRS_TX_BLOCKED bit to indicate
3688 	 * out of tx desc condition. To atomically set
3689 	 * this, we queue the returned packet and do
3690 	 * the setting of SRS_TX_BLOCKED in
3691 	 * mac_tx_srs_drain().
3692 	 */
3693 	if (mp_chain != NULL) {
3694 		mutex_enter(&mac_srs->srs_lock);
3695 		cookie = mac_tx_srs_no_desc(mac_srs, mp_chain, flag, ret_mp);
3696 		mutex_exit(&mac_srs->srs_lock);
3697 		return (cookie);
3698 	}
3699 	SRS_TX_STATS_UPDATE(mac_srs, &stats);
3700 
3701 	return (0);
3702 }
3703 
3704 /*
3705  * mac_tx_serialize_mode
3706  *
3707  * This is an experimental mode implemented as per the request of PAE.
3708  * In this mode, all callers attempting to send a packet to the NIC
3709  * will get serialized. Only one thread at any time will access the
3710  * NIC to send the packet out.
3711  */
3712 /* ARGSUSED */
3713 static mac_tx_cookie_t
mac_tx_serializer_mode(mac_soft_ring_set_t * mac_srs,mblk_t * mp_chain,uintptr_t fanout_hint,uint16_t flag,mblk_t ** ret_mp)3714 mac_tx_serializer_mode(mac_soft_ring_set_t *mac_srs, mblk_t *mp_chain,
3715     uintptr_t fanout_hint, uint16_t flag, mblk_t **ret_mp)
3716 {
3717 	mac_tx_stats_t		stats;
3718 	mac_tx_cookie_t		cookie = 0;
3719 	mac_srs_tx_t		*srs_tx = &mac_srs->srs_tx;
3720 
3721 	/* Single ring, serialize below */
3722 	ASSERT(srs_tx->st_mode == SRS_TX_SERIALIZE);
3723 	mutex_enter(&mac_srs->srs_lock);
3724 	if ((mac_srs->srs_first != NULL) ||
3725 	    (mac_srs->srs_state & SRS_PROC)) {
3726 		/*
3727 		 * In serialization mode, queue all packets until
3728 		 * TX_HIWAT is set.
3729 		 * If drop bit is set, drop if TX_HIWAT is set.
3730 		 * If no_enqueue is set, still enqueue until hiwat
3731 		 * is set and return mblks after TX_HIWAT is set.
3732 		 */
3733 		cookie = mac_tx_srs_enqueue(mac_srs, mp_chain,
3734 		    flag, 0, ret_mp);
3735 		mutex_exit(&mac_srs->srs_lock);
3736 		return (cookie);
3737 	}
3738 	/*
3739 	 * No packets queued, nothing on proc and no flow
3740 	 * control condition. Fast-path, ok. Do inline
3741 	 * processing.
3742 	 */
3743 	mac_srs->srs_state |= SRS_PROC;
3744 	mutex_exit(&mac_srs->srs_lock);
3745 
3746 	mp_chain = mac_tx_send(srs_tx->st_arg1, srs_tx->st_arg2,
3747 	    mp_chain, &stats);
3748 
3749 	mutex_enter(&mac_srs->srs_lock);
3750 	mac_srs->srs_state &= ~SRS_PROC;
3751 	if (mp_chain != NULL) {
3752 		cookie = mac_tx_srs_enqueue(mac_srs,
3753 		    mp_chain, flag, 0, ret_mp);
3754 	}
3755 	if (mac_srs->srs_first != NULL) {
3756 		/*
3757 		 * We processed inline our packet and a new
3758 		 * packet/s got queued while we were
3759 		 * processing. Wakeup srs worker
3760 		 */
3761 		cv_signal(&mac_srs->srs_async);
3762 	}
3763 	mutex_exit(&mac_srs->srs_lock);
3764 
3765 	if (cookie == 0)
3766 		SRS_TX_STATS_UPDATE(mac_srs, &stats);
3767 
3768 	return (cookie);
3769 }
3770 
3771 /*
3772  * mac_tx_fanout_mode
3773  *
3774  * In this mode, the SRS will have access to multiple Tx rings to send
3775  * the packet out. The fanout hint that is passed as an argument is
3776  * used to find an appropriate ring to fanout the traffic. Each Tx
3777  * ring, in turn,  will have a soft ring associated with it. If a Tx
3778  * ring runs out of Tx desc's the returned packet will be queued in
3779  * the soft ring associated with that Tx ring. The srs itself will not
3780  * queue any packets.
3781  */
3782 
3783 #define	MAC_TX_SOFT_RING_PROCESS(chain) {				\
3784 	index = COMPUTE_INDEX(hash, mac_srs->srs_tx_ring_count),	\
3785 	softring = mac_srs->srs_tx_soft_rings[index];			\
3786 	cookie = mac_tx_soft_ring_process(softring, chain, flag, ret_mp); \
3787 	DTRACE_PROBE2(tx__fanout, uint64_t, hash, uint_t, index);	\
3788 }
3789 
3790 static mac_tx_cookie_t
mac_tx_fanout_mode(mac_soft_ring_set_t * mac_srs,mblk_t * mp_chain,uintptr_t fanout_hint,uint16_t flag,mblk_t ** ret_mp)3791 mac_tx_fanout_mode(mac_soft_ring_set_t *mac_srs, mblk_t *mp_chain,
3792     uintptr_t fanout_hint, uint16_t flag, mblk_t **ret_mp)
3793 {
3794 	mac_soft_ring_t		*softring;
3795 	uint64_t		hash;
3796 	uint_t			index;
3797 	mac_tx_cookie_t		cookie = 0;
3798 
3799 	ASSERT(mac_srs->srs_tx.st_mode == SRS_TX_FANOUT ||
3800 	    mac_srs->srs_tx.st_mode == SRS_TX_BW_FANOUT);
3801 	if (fanout_hint != 0) {
3802 		/*
3803 		 * The hint is specified by the caller, simply pass the
3804 		 * whole chain to the soft ring.
3805 		 */
3806 		hash = HASH_HINT(fanout_hint);
3807 		MAC_TX_SOFT_RING_PROCESS(mp_chain);
3808 	} else {
3809 		mblk_t *last_mp, *cur_mp, *sub_chain;
3810 		uint64_t last_hash = 0;
3811 		uint_t media = mac_srs->srs_mcip->mci_mip->mi_info.mi_media;
3812 
3813 		/*
3814 		 * Compute the hash from the contents (headers) of the
3815 		 * packets of the mblk chain. Split the chains into
3816 		 * subchains of the same conversation.
3817 		 *
3818 		 * Since there may be more than one ring used for
3819 		 * sub-chains of the same call, and since the caller
3820 		 * does not maintain per conversation state since it
3821 		 * passed a zero hint, unsent subchains will be
3822 		 * dropped.
3823 		 */
3824 
3825 		flag |= MAC_DROP_ON_NO_DESC;
3826 		ret_mp = NULL;
3827 
3828 		ASSERT(ret_mp == NULL);
3829 
3830 		sub_chain = NULL;
3831 		last_mp = NULL;
3832 
3833 		for (cur_mp = mp_chain; cur_mp != NULL;
3834 		    cur_mp = cur_mp->b_next) {
3835 			hash = mac_pkt_hash(media, cur_mp, MAC_PKT_HASH_L4,
3836 			    B_TRUE);
3837 			if (last_hash != 0 && hash != last_hash) {
3838 				/*
3839 				 * Starting a different subchain, send current
3840 				 * chain out.
3841 				 */
3842 				ASSERT(last_mp != NULL);
3843 				last_mp->b_next = NULL;
3844 				MAC_TX_SOFT_RING_PROCESS(sub_chain);
3845 				sub_chain = NULL;
3846 			}
3847 
3848 			/* add packet to subchain */
3849 			if (sub_chain == NULL)
3850 				sub_chain = cur_mp;
3851 			last_mp = cur_mp;
3852 			last_hash = hash;
3853 		}
3854 
3855 		if (sub_chain != NULL) {
3856 			/* send last subchain */
3857 			ASSERT(last_mp != NULL);
3858 			last_mp->b_next = NULL;
3859 			MAC_TX_SOFT_RING_PROCESS(sub_chain);
3860 		}
3861 
3862 		cookie = 0;
3863 	}
3864 
3865 	return (cookie);
3866 }
3867 
3868 /*
3869  * mac_tx_bw_mode
3870  *
3871  * In the bandwidth mode, Tx srs will allow packets to go down to Tx ring
3872  * only if bw is available. Otherwise the packets will be queued in
3873  * SRS. If the SRS has multiple Tx rings, then packets will get fanned
3874  * out to a Tx rings.
3875  */
3876 static mac_tx_cookie_t
mac_tx_bw_mode(mac_soft_ring_set_t * mac_srs,mblk_t * mp_chain,uintptr_t fanout_hint,uint16_t flag,mblk_t ** ret_mp)3877 mac_tx_bw_mode(mac_soft_ring_set_t *mac_srs, mblk_t *mp_chain,
3878     uintptr_t fanout_hint, uint16_t flag, mblk_t **ret_mp)
3879 {
3880 	int			cnt, sz;
3881 	mblk_t			*tail;
3882 	mac_tx_cookie_t		cookie = 0;
3883 	mac_srs_tx_t		*srs_tx = &mac_srs->srs_tx;
3884 	clock_t			now;
3885 
3886 	ASSERT(TX_BANDWIDTH_MODE(mac_srs));
3887 	ASSERT(mac_srs->srs_type & SRST_BW_CONTROL);
3888 	mutex_enter(&mac_srs->srs_lock);
3889 	if (mac_srs->srs_bw->mac_bw_limit == 0) {
3890 		/*
3891 		 * zero bandwidth, no traffic is sent: drop the packets,
3892 		 * or return the whole chain if the caller requests all
3893 		 * unsent packets back.
3894 		 */
3895 		if (flag & MAC_TX_NO_ENQUEUE) {
3896 			cookie = (mac_tx_cookie_t)mac_srs;
3897 			*ret_mp = mp_chain;
3898 		} else {
3899 			MAC_TX_SRS_DROP_MESSAGE(mac_srs, mp_chain, cookie,
3900 			    "Tx no bandwidth");
3901 		}
3902 		mutex_exit(&mac_srs->srs_lock);
3903 		return (cookie);
3904 	} else if ((mac_srs->srs_first != NULL) ||
3905 	    (mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED)) {
3906 		cookie = mac_tx_srs_enqueue(mac_srs, mp_chain, flag,
3907 		    fanout_hint, ret_mp);
3908 		mutex_exit(&mac_srs->srs_lock);
3909 		return (cookie);
3910 	}
3911 	MAC_COUNT_CHAIN(mac_srs, mp_chain, tail, cnt, sz);
3912 	now = ddi_get_lbolt();
3913 	if (mac_srs->srs_bw->mac_bw_curr_time != now) {
3914 		mac_srs->srs_bw->mac_bw_curr_time = now;
3915 		mac_srs->srs_bw->mac_bw_used = 0;
3916 	} else if (mac_srs->srs_bw->mac_bw_used >
3917 	    mac_srs->srs_bw->mac_bw_limit) {
3918 		mac_srs->srs_bw->mac_bw_state |= SRS_BW_ENFORCED;
3919 		MAC_TX_SRS_ENQUEUE_CHAIN(mac_srs,
3920 		    mp_chain, tail, cnt, sz);
3921 		/*
3922 		 * Wakeup worker thread. Note that worker
3923 		 * thread has to be woken up so that it
3924 		 * can fire up the timer to be woken up
3925 		 * on the next tick. Also once
3926 		 * BW_ENFORCED is set, it can only be
3927 		 * reset by srs_worker thread. Until then
3928 		 * all packets will get queued up in SRS
3929 		 * and hence this this code path won't be
3930 		 * entered until BW_ENFORCED is reset.
3931 		 */
3932 		cv_signal(&mac_srs->srs_async);
3933 		mutex_exit(&mac_srs->srs_lock);
3934 		return (cookie);
3935 	}
3936 
3937 	mac_srs->srs_bw->mac_bw_used += sz;
3938 	mutex_exit(&mac_srs->srs_lock);
3939 
3940 	if (srs_tx->st_mode == SRS_TX_BW_FANOUT) {
3941 		mac_soft_ring_t *softring;
3942 		uint_t indx, hash;
3943 
3944 		hash = HASH_HINT(fanout_hint);
3945 		indx = COMPUTE_INDEX(hash,
3946 		    mac_srs->srs_tx_ring_count);
3947 		softring = mac_srs->srs_tx_soft_rings[indx];
3948 		return (mac_tx_soft_ring_process(softring, mp_chain, flag,
3949 		    ret_mp));
3950 	} else if (srs_tx->st_mode == SRS_TX_BW_AGGR) {
3951 		return (mac_tx_aggr_mode(mac_srs, mp_chain,
3952 		    fanout_hint, flag, ret_mp));
3953 	} else {
3954 		mac_tx_stats_t		stats;
3955 
3956 		mp_chain = mac_tx_send(srs_tx->st_arg1, srs_tx->st_arg2,
3957 		    mp_chain, &stats);
3958 
3959 		if (mp_chain != NULL) {
3960 			mutex_enter(&mac_srs->srs_lock);
3961 			MAC_COUNT_CHAIN(mac_srs, mp_chain, tail, cnt, sz);
3962 			if (mac_srs->srs_bw->mac_bw_used > sz)
3963 				mac_srs->srs_bw->mac_bw_used -= sz;
3964 			else
3965 				mac_srs->srs_bw->mac_bw_used = 0;
3966 			cookie = mac_tx_srs_enqueue(mac_srs, mp_chain, flag,
3967 			    fanout_hint, ret_mp);
3968 			mutex_exit(&mac_srs->srs_lock);
3969 			return (cookie);
3970 		}
3971 		SRS_TX_STATS_UPDATE(mac_srs, &stats);
3972 
3973 		return (0);
3974 	}
3975 }
3976 
3977 /*
3978  * mac_tx_aggr_mode
3979  *
3980  * This routine invokes an aggr function, aggr_find_tx_ring(), to find
3981  * a (pseudo) Tx ring belonging to a port on which the packet has to
3982  * be sent. aggr_find_tx_ring() first finds the outgoing port based on
3983  * L2/L3/L4 policy and then uses the fanout_hint passed to it to pick
3984  * a Tx ring from the selected port.
3985  *
3986  * Note that a port can be deleted from the aggregation. In such a case,
3987  * the aggregation layer first separates the port from the rest of the
3988  * ports making sure that port (and thus any Tx rings associated with
3989  * it) won't get selected in the call to aggr_find_tx_ring() function.
3990  * Later calls are made to mac_group_rem_ring() passing pseudo Tx ring
3991  * handles one by one which in turn will quiesce the Tx SRS and remove
3992  * the soft ring associated with the pseudo Tx ring. Unlike Rx side
3993  * where a cookie is used to protect against mac_rx_ring() calls on
3994  * rings that have been removed, no such cookie is needed on the Tx
3995  * side as the pseudo Tx ring won't be available anymore to
3996  * aggr_find_tx_ring() once the port has been removed.
3997  */
3998 static mac_tx_cookie_t
mac_tx_aggr_mode(mac_soft_ring_set_t * mac_srs,mblk_t * mp_chain,uintptr_t fanout_hint,uint16_t flag,mblk_t ** ret_mp)3999 mac_tx_aggr_mode(mac_soft_ring_set_t *mac_srs, mblk_t *mp_chain,
4000     uintptr_t fanout_hint, uint16_t flag, mblk_t **ret_mp)
4001 {
4002 	mac_srs_tx_t		*srs_tx = &mac_srs->srs_tx;
4003 	mac_tx_ring_fn_t	find_tx_ring_fn;
4004 	mac_ring_handle_t	ring = NULL;
4005 	void			*arg;
4006 	mac_soft_ring_t		*sringp;
4007 
4008 	find_tx_ring_fn = srs_tx->st_capab_aggr.mca_find_tx_ring_fn;
4009 	arg = srs_tx->st_capab_aggr.mca_arg;
4010 	if (find_tx_ring_fn(arg, mp_chain, fanout_hint, &ring) == NULL)
4011 		return (0);
4012 	sringp = srs_tx->st_soft_rings[((mac_ring_t *)ring)->mr_index];
4013 	return (mac_tx_soft_ring_process(sringp, mp_chain, flag, ret_mp));
4014 }
4015 
4016 void
mac_tx_invoke_callbacks(mac_client_impl_t * mcip,mac_tx_cookie_t cookie)4017 mac_tx_invoke_callbacks(mac_client_impl_t *mcip, mac_tx_cookie_t cookie)
4018 {
4019 	mac_cb_t *mcb;
4020 	mac_tx_notify_cb_t *mtnfp;
4021 
4022 	/* Wakeup callback registered clients */
4023 	MAC_CALLBACK_WALKER_INC(&mcip->mci_tx_notify_cb_info);
4024 	for (mcb = mcip->mci_tx_notify_cb_list; mcb != NULL;
4025 	    mcb = mcb->mcb_nextp) {
4026 		mtnfp = (mac_tx_notify_cb_t *)mcb->mcb_objp;
4027 		mtnfp->mtnf_fn(mtnfp->mtnf_arg, cookie);
4028 	}
4029 	MAC_CALLBACK_WALKER_DCR(&mcip->mci_tx_notify_cb_info,
4030 	    &mcip->mci_tx_notify_cb_list);
4031 }
4032 
4033 /* ARGSUSED */
4034 void
mac_tx_srs_drain(mac_soft_ring_set_t * mac_srs,uint_t proc_type)4035 mac_tx_srs_drain(mac_soft_ring_set_t *mac_srs, uint_t proc_type)
4036 {
4037 	mblk_t			*head, *tail;
4038 	size_t			sz;
4039 	uint32_t		tx_mode;
4040 	uint_t			saved_pkt_count;
4041 	mac_tx_stats_t		stats;
4042 	mac_srs_tx_t		*srs_tx = &mac_srs->srs_tx;
4043 	clock_t			now;
4044 
4045 	saved_pkt_count = 0;
4046 	ASSERT(mutex_owned(&mac_srs->srs_lock));
4047 	ASSERT(!(mac_srs->srs_state & SRS_PROC));
4048 
4049 	mac_srs->srs_state |= SRS_PROC;
4050 
4051 	tx_mode = srs_tx->st_mode;
4052 	if (tx_mode == SRS_TX_DEFAULT || tx_mode == SRS_TX_SERIALIZE) {
4053 		if (mac_srs->srs_first != NULL) {
4054 			head = mac_srs->srs_first;
4055 			tail = mac_srs->srs_last;
4056 			saved_pkt_count = mac_srs->srs_count;
4057 			mac_srs->srs_first = NULL;
4058 			mac_srs->srs_last = NULL;
4059 			mac_srs->srs_count = 0;
4060 			mutex_exit(&mac_srs->srs_lock);
4061 
4062 			head = mac_tx_send(srs_tx->st_arg1,