Patent Application
20050053060
This invention relates in general to the field of computer network architecture, and more specifically to an architecture to allow sharing and/or partitioning of network input/output (I/O) endpoint devices in a load/store fabric, particularly a shared network interface controller.
Although the eight above referenced pending patent applications have been incorporated by reference, to assist the reader in appreciating the problem to which the present invention is directed, the Background of those applications is substantially repeated below.
Moderm computer architecture may be viewed as having three distinct subsystems which when combined, form what most think of when they hear the term computer. These subsystems are: 1) a processing complex; 2) an interface between the processing complex and I/O controllers or devices; and 3) the I/O (i.e., input/output) controllers or devices themselves.
A processing complex may be as simple as a single microprocessor, such as a Pentium microprocessor, coupled to memory. Or, it might be as complex as two or more processors which share memory.
The interface between the processing complex and I/O is commonly known as the chipset. On the north side of the chipset (i.e., between the processing complex and the chipset) is a bus referred to as the HOST bus. The HOST bus is usually a proprietary bus designed to interface to memory, to one or more microprocessors within the processing complex, and to the chipset. On the south side of the chipset are a number of buses which connect the chipset to I/O devices. Examples of such buses include: ISA, EISA, PCI, PCI-X, and AGP.
I/O devices are devices that allow data to be transferred to or from the processing complex through the chipset, on one or more of the buses supported by the chipset. Examples of I/O devices include: graphics cards coupled to a computer display; disk controllers, such as Serial ATA (SATA) or Fiber Channel controllers (which are coupled to hard disk drives or other data storage systems); network controllers (to interface to networks such as Ethernet); USB and Firewire controllers which interface to a variety of devices from digital cameras to external data storage to digital music systems, etc.; and PS/2 controllers for interfacing to keyboards/mice. The I/O devices are designed to connect to the chipset via one of its supported interface buses. For example, modern computers typically couple graphic cards to the chipset via an AGP bus. Ethernet cards, SATA, Fiber Channel, and SCSI (data storage) cards, USB and Firewire controllers all connect to a PCI bus, and PS/2 devices connect to an ISA bus.
One skilled in the art will appreciate that the above description is general. However, what should be appreciated is that regardless of the type of computer, it will include a processing complex for executing instructions, an interface to I/O, and I/O devices to allow the processing complex to communicate with the world outside of itself. This is true whether the computer is an inexpensive desktop in a home, a high-end workstation used for graphics and video editing, or a clustered server which provides database support to hundreds within a large organization.
Also, although not yet referenced, a processing complex typically executes one or more operating systems (e.g., Microsoft Windows, Windows Server, Unix, Linux, Macintosh, etc.). This application therefore refers to the combination of a processing complex with one or more operating systems as an operating system domain (OSD). An OS domain, within the present context, is a system load-store memory map that is associated with one or more processing complexes. Typically, present day operating systems such as Windows, Unix, Linux, VxWorks, Macintosh, etc., must comport with a specific load-store memory map that corresponds to the processing complex upon which they execute. For example, a typical x86 load-store memory map provides for both memory space and I/O space. Conventional memory is mapped to the lower 640 kilobytes (KB) of memory. The next higher 128 KB of memory are employed by legacy video devices. Above that is another 128 KB block of addresses mapped to expansion ROM. And the 128 KB block of addresses below the 1 megabyte (MB) boundary is mapped to boot ROM (i.e., BIOS). Both DRAM space and PCI memory are mapped above the 1 MB boundary. Accordingly, two separate processing complexes may be executing within two distinct OS domains, which typically means that the two processing complexes are executing either two instances of the same operating system or that they are executing two distinct operating systems. However, in a symmetrical multi-processing environment, a plurality of processing complexes may together be executing a single instance of an SMP operating system, in which case the plurality of processing complexes would be associated with a single OS domain.
A problem that has been recognized by the present inventor is that the requirement to place a processing complex, interface and I/O within every computer is costly, and lacks modularity. That is, once a computer is purchased, all of the subsystems are static from the standpoint of the user. The ability to change a processing complex while still utilizing the interface and I/O is extremely difficult. The interface or chipset is typically so tied to the processing complex that swapping one without the other doesn't make sense. And, the I/O is typically integrated within the computer, at least for servers and business desktops, such that upgrade or modification of the I/O is either impossible or cost prohibitive.
An example of the above limitations is considered helpful. A popular network server designed by Dell Computer Corporation is the Dell PowerEdge 1750. This server includes one or more microprocessors designed by Intel (Xeon processors), along with memory (e.g., the processing complex). It has a server class chipset for interfacing the processing complex to I/O (e.g., the interface). And, it has onboard graphics for connecting to a display, onboard PS/2 for connecting a mouse/keyboard, onboard RAID control for connecting to data storage, onboard network interface controllers for connecting to 10/100 and 1 gig Ethernet; and a PCI bus for adding other I/O such as SCSI or Fiber Channel controllers. It is believed that none of the onboard features are upgradeable.
So, as mentioned above, one of the problems with this architecture is that if another I/O demand emerges, it is difficult, or cost prohibitive to implement the upgrade. For example, 10 gigabit Ethernet is on the horizon. How can this be easily added to this server? Well, perhaps a 10 gig Ethernet controller could be purchased and inserted onto the PCI bus. Consider a technology infrastructure that included tens or hundreds of these servers. To move to a faster network architecture requires an upgrade to each of the existing servers. This is an extremely cost prohibitive scenario, which is why it is very difficult to upgrade existing network infrastructures.
This one-to-one correspondence between the processing complex, the interface, and the I/O is also costly to the manufacturer. That is, in the example above, much of the I/O is manufactured on the motherboard of the server. To include the I/O on the motherboard is costly to the manufacturer, and ultimately to the end user. If the end user utilizes all of the I/O provided, then s/he is happy. But, if the end user does not wish to utilize the onboard RAID, or the 10/100 Ethernet, then s/he is still required to pay for its inclusion. This is not optimal.
Consider another emerging platform, the blade server. A blade server is essentially a processing complex, an interface, and I/O together on a relatively small printed circuit board that has a backplane connector. The blade is made to be inserted with other blades into a chassis that has a form factor similar to a rack server today. The benefit is that many blades can be located in the same rack space previously required by just one or two rack servers. While blades have seen market growth in some areas, where processing density is a real issue, they have yet to gain significant market share, for many reasons. One of the reasons is cost. That is, blade servers still must provide all of the features of a pedestal or rack server, including a processing complex, an interface to I/O, and I/O. Further, the blade servers must integrate all necessary I/O because they do not have an external bus which would allow them to add other I/O on to them. So, each blade must include such I/O as Ethernet (10/100, and/or 1 gig), and data storage control (SCSI, Fiber Channel, etc.).
One recent development to try and allow multiple processing complexes to separate themselves from I/O devices was introduced by Intel and other vendors. It is called Infiniband. Infiniband is a high-speed serial interconnect designed to provide for multiple, out of the box interconnects. However, it is a switched, channel-based architecture that is not part of the load-store architecture of the processing complex. That is, it uses message passing where the processing complex communicates with a Host-Channel-Adapter (HCA) which then communicates with all downstream devices, such as I/O devices. It is the HCA that handles all the transport to the Infiniband fabric rather than the processing complex. That is, the only device that is within the load/store domain of the processing complex is the HCA. What this means is that you have to leave the processing complex domain to get to your I/O devices. This jump out of processing complex domain (the load/store domain) is one of the things that contributed to Infinibands failure as a solution to shared I/O. According to one industry analyst referring to Infiniband, “[i]t was overbilled, overhyped to be the nirvana for everything server, everything I/O, the solution to every problem you can imagine in the data center . . . but turned out to be more complex and expensive to deploy . . . because it required installing a new cabling system and significant investments in yet another switched high speed serial interconnect”.
Thus, the inventor has recognized that separation between the processing complex and its interface, and I/O, should occur, but the separation must not impact either existing operating systems, software, or existing hardware or hardware infrastructures. By breaking apart the processing complex from the I/O, more cost effective and flexible solutions can be introduced.
Further, the inventor has recognized that the solution must not be a channel-based architecture, performed outside of the box. Rather, the solution should use a load-store architecture, where the processing complex sends data directly to (or at least architecturally directly) or receives data directly from an I/O device (such as a network controller, or data storage controller). This allows the separation to be accomplished without affecting a network infrastructure or disrupting the operating system.
Therefore, what is needed is an apparatus and method which separates the processing complex and its interface to I/O from the I/O devices.
Further, what is needed is an apparatus and method which allows processing complexes and their interfaces to be designed, manufactured, and sold, without requiring I/O to be included within them.
Additionally, what is needed is an apparatus and method which allows a single I/O device to be shared by multiple processing complexes.
Further, what is needed is an apparatus and method that allows multiple processing complexes to share one or more I/O devices through a common load-store fabric.
Additionally, what is needed is an apparatus and method that provides switching between multiple processing complexes and shared I/O.
Further, what is needed is an apparatus and method that allows multiple processing complexes, each operating independently, and having their own operating system domain, to view shared I/O devices as if the I/O devices were dedicated to them.
And, what is needed is an apparatus and method which allows shared I/O devices to be utilized by different processing complexes without requiring modification to the processing complexes existing operating systems or other software. Of course, one skilled in the art will appreciate that modification of driver software may allow for increased functionality within the shared environment.
The previously filed applications from which this application depends address each of these needs. However, in addition to the above, what is further needed is an I/O device that can be shared by two or more processing complexes using a common load-store fabric.
Further, what is needed is a network interface controller which can be shared, or mapped, to one or more processing complexes (or OSD's) using a common load-store fabric. Network interface controllers, Ethernet controllers (10/100, 1 gig, and 10 gig) are all implementations of a network interface controller (NIC).
The present invention provides a method and apparatus for allowing a network interface controller to be shared by one or more operating system domains within the load-store domains of processing complexes.
In one aspect, the present invention provides a shareable network interface controller to be shared within the load-store architecture of a number of operating system domains. The controller includes transmit/receive logic, a bus interface, and association logic. The transmit/receive logic couples the controller to a network to transmit/receive network communication. The bus interface includes operating system domain identification logic, and interfaces the controller to the operating system domains (OSD's), and for associates OSD communication to and from the bus interface with ones of the OSD's. The association logic is coupled to the transmit/receive logic, and associates the network communication with ones of the OSD's. By associating the network communication with the OSD's, and associating the OSD communication with the OSD's, the shared controller provides communication between the OSD's and the network.
An additional feature of this invention includes a number of control status register sets (CSR's), coupled to the bus interface, and available to be associated with a unique one of the operating system domains. That is, for each OSD to be connected to the controller, a CSR set is made available to allow the controller to be mapped within the load-store domain of each OSD.
Another feature of this invention includes a descriptor cache coupled to the bus interface to store descriptors retrieved from the plurality of operating system domains.
An additional feature of this invention includes a number of direct memory access (DMA) engines, the engines used by the controller to DMA OSD communication from the OSD's, and to DMA the network communication to the OSD's.
In another aspect, the present invention provides a network interface controller (NIC) which provides an interface between a network and a number of operating system domains, where the controller exists within the load-store domain of each of the operating system domains (OSD's). The controller includes a bus interface, a network interface and association logic. The bus interface is located between the operating system domains and the controller. The network interface is located between the network and the controller. The association logic is coupled to both the bus interface and the network interface and associates data coming from the network with at least one of the plurality of operating system domains.
In a further aspect, the present invention provides a shared network interface controller (NIC) which allows a number of computer servers to interface to an Ethernet network. The controller includes: a bus interface to couple the controller to a packet based load-store serial link, the load-store serial link coupling each of the computer servers to the controller; server identification logic, coupled to the bus interface, to identify packets received by the bus interface with ones of the computer servers; buffering logic, coupled to the server identification logic, to store the packets received by the bus interface, along with tags to associate the packets with ones of the computer servers; association logic, coupled to the buffering logic, to determine MAC addresses for the packets utilizing the tags; and transmit/receive logic, coupled to the Ethernet network, and to the association logic, to transmit the packets to the Ethernet network, and to receive packets from the Ethernet network. By associating packets with ones of the computer servers, and buffering the packets between the computer servers and the Ethernet network, the shared network interface controller provides an Ethernet interface for all of the computer servers via the packet based load-store serial link.
In yet another aspect, the present invention provides a blade server environment including: a number of blade servers; a shared switch to couple to the blade servers, and to provide communication between the blade servers and a downstream endpoint, the downstream endpoint being a shared network interface controller. The controller includes: a bus interface to couple the controller to the shared switch; identification logic, to determine which of the blade servers is associated with data packets received by the controller; transmit/receive logic, to couple the controller to an Ethernet network; buffer logic, coupled to the bus interface and to the transmit/receive logic, to buffer the data packets received by the controller, either from the shared switch, or from the Ethernet network; and association logic, coupled to the buffer logic, to determine from a MAC address, which of the blade servers should be associated with the data packets.
In a further aspect, the present invention provides a method for sharing a network interface controller by a number of operating system domains. The method includes: providing a load-store domain link between the controller and the operating system domains; providing OSD identification for data transmitted over the load-store domain link; providing an interface between the controller and a network; buffering data between the load-store domain link and the network; and associating the buffered data with the operating system domains, the associating utilizing association logic to associate MAC addresses within the buffered data with ones of the operating system domains.
In yet another aspect, the present invention provides a method for correlating Ethernet packets within a shared network interface controller with a number of upstream operating system domains. The method includes: receiving the packets into the shared network interface controller; associating the packets with the upstream operating system domains by correlating destination MAC addresses within the packets with ones of the upstream operating system domains; caching descriptors for each of the upstream operating system domains which define where in their memories the packets are to be transferred; and transferring the packets, per the descriptors, to the upstream operating system domains via a load-store link which allows the packets to be tagged with information which associates the packets with their upstream operating system domain.
In a further aspect, the present invention provides a method for transmitting packets from a number of operating system domains to an Ethernet network thru a shared network interface controller. The method includes: requesting a packet transmit from one of the operating system domains to the controller; tagging the request with an OSD header to associate the request with its operating system domain; transmitting the request to the shared network interface controller; within the shared network interface controller, correlating the request with its operating system domain; utilizing the OSD header for the request to determine associated descriptors for the request; transferring data from the operating system domain, corresponding to the request utilizing the associated descriptors; buffering the transferred data; and transmitting the buffered data to the Ethernet network.
In another aspect, the present invention provides a shared network interface controller to interface a number of operating system domains to a network. The controller includes: a bus interface, to interface the controller to a number of operating system domains; transmit/receive logic, to interface the controller to a network; and packet replication logic, coupled to the bus interface, to detect whether packets received from a first operating system domain is destined for a second operating system domain, and if the packets are destined for the second operating system domain, the logic causing the packets to be transferred to the second operating system domain.
In a further aspect, the present invention provides a method for determining a loopback condition within a shared network interface controller. The method includes: determining whether a packet transmitted by a first operating system domain is destined for a second operating system domain, the first and second operating system domains interfaced to each other, and to a network, through the shared network interface controller; if the packet is destined for the second operating system domain, forwarding the packet to the second operating system domain, and not forwarding the packet to the network; and if the packet is not destined for the second operating system domain, forwarding the packet to the network.
Other features and advantages of the present invention will become apparent upon study of the remaining portions of the specification and drawings.
Although the present invention may be implemented in any of a number of load-store fabrics, the below discussion is provided with particular reference to PCI-Express. One skilled in the art will appreciate that although embodiments of the present invention will be described within the context of PCI Express, a number of alternative, or yet to be developed load/store protocols might be used without departing from the spirit and scope of the present invention.
By way of background, Peripheral Component Interconnect (PCI) was developed in the early 1990's by Intel Corporation as a general I/O architecture to transfer data and instructions faster than the ISA architecture of the time. PCI has gone thru several improvements since that time, with the latest proposal being PCI Express. In a nutshell, PCI Express is a replacement of the PCI and PCI-X bus specification to provide platforms with much greater performance, while using a much lower pin count (Note: PCI and PCI-X are parallel bus architectures, PCI Express is a serial architecture). A complete discussion of PCI Express is beyond the scope of this specification, but a thorough background and description can be found in the following books which are incorporated herein by reference for all purposes: Introduction to PCI Express, A Hardware and Software Developer's Guide, by Adam Wilen, Justin Schade, Ron Thornburg; The Complete PCI Express Reference, Design Insights for Hardware and Software Developers, by Edward Solari and Brad Congdon; and PCI Express System Architecture, by Ravi Budruk, Don Anderson, Tom Shanley; all of which are available at www.amazon.com. In addition, the PCI Express specification is managed and disseminated through the Special Interest Group (SIG) for PCI found at www.pcisig.com.
This invention is also directed at describing a shared network interface controller. Interface controllers have existed to connect computers to a variety of networks, such as Ethernet, Token Ring, etc. However, Applicant's are unaware of any network interface controller that may be shared by multiple processing complexes as part of their load-store domain. While the present invention will be described with reference to interfacing to an Ethernet network, one skilled in the art will appreciate that the teachings of the present invention are applicable to any type of computer network.
Referring now to
Referring now to
As mentioned above, a processing complex may be as simple as a single microprocessor, such as a Pentium microprocessor, coupled to memory, or it might be as complex as two or more processors which share memory. The processing complex may execute a single operating system, or may execute multiple operating systems which share memory. In either case, applicant intends that from the viewpoint of the shared I/O switch 214, that whatever configuration of the processing complex, each load-store bus 208, 210, 212 be considered a separate operating system domain (OSD). At this point, it is sufficient that the reader understand that in the environment described with reference to
In the reverse, when information flows from the network interface controller 220 to the shared I/O switch 214, the information is associated with the appropriate upstream link 208, 210, 212 by embedding (or providing out of band), OSD association for each piece of information (e.g., packet) transmitted over the link 216. The shared I/O switch 214 receives the OSD aware information via the link 216, determines which upstream port the information should be transmitted on, and then transmits the information on the associated link 208, 210, 212.
What should be appreciated by reference to
Referring now to
A detailed description of one embodiment of the shared network interface controller of the present invention will be described below with respect to
Each of the processing complexes 302, 304, 306 are coupled to the shared I/O switch 314 via links 308, 310, 312. The links, in one embodiment, utilize PCI-Express. The shared I/O switch 314 couples each of the links 308, 310, 312 to downstream devices such as the shared network interface controller 320. In addition, the shared I/O switch 314 tags communication from each of the processing complexes 302, 304, 306 with an operating system domain header (OSD header) to indicate to the downstream devices, which of the processing complexes 302, 304, 306 is associated with the communication. Thus, when the shared network interface controller 320 receives a communication from the shared I/O switch 314, included in the communication is an OSD header. The controller 320 can utilize this header to determine which of the processing complexes 302, 304, 306 sent the communication, so that the controller 320 can deal with communication from each of the complexes 302, 304, 306 distinctly. In reverse, communication from the controller 320 to the processing complexes 302, 304, 306 gets tagged by the controller 320 with an OSD header, so that the shared I/O switch 314 can determine which of the processing complexes 302, 304, 306 the communication should be passed to. Thus, by tagging communication between the processing complexes 302, 304, 306 and the shared network interface controller 320 with an OSD header (or any other type of identifier), the controller 320 can distinguish communication between the different complexes it supports.
Referring now to
Referring now to
Referring now to
The controller 600 includes bus interface/OS ID logic 602 for interfacing the controller 600 to an upstream load/store shared I/O link such as described above with reference to
The bus interface/OS ID logic 602 is coupled to a data path mux 604. The mux 604 is coupled to packet replication logic 605. In one embodiment, the packet replication logic 605 is used for loopback, multi-cast and broadcast operations. More specifically, since packets originating from one of the processing complexes may be destined for one or more of the other processing complexes for which the shared network interface controller 600 is coupled, the packet replication logic 605 performs the function of determining whether such packets should be transmitted to the Ethernet network, or alternatively, should be replicated and presented to one or more of the other processing complexes to which the controller 600 is coupled. Details of a multicast operation will be described below with reference to
The mux 604 is also coupled to a plurality of CSR blocks 606. As mentioned above, to establish communication to an operating system domain, a controller must have control status registers which are addressable by the operating system domain. These control status registers 606 have been duplicated in
Referring to
In one embodiment, the CSR's 606 contain the Control and Status Registers used by device drivers in the OSD's to interface to the controller 600. The CSR's 606 are responsible for generating interrupts to the interface between the OSD's and the controller 600. The CSR's 606 also include any generic timers or system functions specific to a given OSD. In one embodiment, there is one CSR set, with several registers replicated per each OSD. The following table describes some of the CSR registers 606 of an embodiment. Mirrored registers map a single or global function/register into all OSD's. Note that in some cases the registers may be located in separate address locations to ensure that an OSD does not have to do Byte accesses or RMW.
Referring back to
The controller 600 further includes DMA logic having DMA arbitration 610 coupled to a number of DMA engines 611. Since the controller 600 will be supporting more than one OSD, additional DMA engines 611 allow increased performance for the controller 600, although additional DMA engines 611 are not required. Thus, one DMA engine 611 could be handling communication from a first OSD, while a second DMA engine 611 could be handling communication from a second OSD. Or, one DMA engine 611 could be handling transmit communication from a first OSD, while a second DMA engine 611 could be handling receive communication for the first OSD. Thus, it is not intended to necessarily provide a DMA engine 611 per supported OSD. Rather, the manufacturer may provide any number of DMA engines 611, according to the performance desired. Further, the DMA arbitration 610 may be configured to select/control utilization of the DMA engines 611 according to predefined criteria. One simple criteria would simply be a round robin selection of engines 611 by the supported OSD's. Another criteria would designate a DMA engine per OSD. Yet another criteria would associate particular DMA engines with either transmit or receive operations. Specifics associated with DMA arbitration are beyond the scope of the present application. However, one skilled in the art should appreciate that it is not the arbitration schemes which are important to the present application, but rather, the provision of 1-N DMA engines, along with appropriate arbitration, to allow for desired performance to be obtained for a desired number of supported OSD's.
The controller 600 further includes descriptor logic having descriptor arbitration 613, a plurality of descriptor caches 615, and in one embodiment descriptor tags 617. One skilled in the art will appreciate that present non shared network interface controllers contain a descriptor cache for storing transmit/receive descriptors. The transmit/receive descriptors are associated with the OSD to which the non shared controller is attached. The descriptors are retrieved by the non shared controller from the memory system of the OSD, and are used to receive/transmit data from/to the OSD. With the shared network interface controller 600 of the present invention, descriptors must be available within the controller 600 for each of the supported OSD's. And, each of the descriptors must be associated with their specific OSD. Applicant has envisioned a number of embodiments for providing descriptors for multiple OSD's, and has illustrated these embodiments in
Embodiment (b) includes a virtual descriptor cache 806 having tags 810. The virtual descriptor cache 806 may be used to store descriptors for any of the supported OSD's. But, when a descriptor is retrieved from a particular OSD, that OSD's header (or some other identifier) is placed as a tag which is associated with that descriptor. Thus, the controller can readily identify which of the descriptors in the virtual descriptor cache 806 are associated with which one of the supported OSD's. In this embodiment, descriptor arbitration 808 is used to insure that each supported OSD is adequately supported by the virtual descriptor cache 806. For example, the virtual descriptor cache 806 caches both transmit and receive descriptors for all of the supported OSD's. One scenario would allocate equal memory space to transmit descriptors and receive descriptors (such as shown in embodiment (c) discussed below. An alternative scenario would allocate a greater portion of the memory to transmit descriptors. Further, the allocation of memory to either transmit or receive descriptors could be made dynamic, so that a greater portion of the memory is used to store transmit descriptors, until the OSD's begin receiving a greater portion of receive packets, at which time a greater portion of the memory would be allocated for receive descriptors. And, the allocation of transmit receiver cache could be equal across all supported OSD's, or alternatively, could be based on pre-defined criteria. For example, it may be established that one or more of the OSD's should be given higher priority (or rights) to the descriptor cache. That is, OSD “0” might be allocated 30% of the transmit descriptor cache, while the other OSD's compete for the other 70%. Or, rights to the cache 806 may be made in a pure round-robin fashion, giving each OSD essentially equal rights to the cache for its descriptors. Thus, whether the allocation of fifo cache between transmit and receive descriptors, and/or between OSDs is made equal, or is made unequal based on static criteria, or is allowed to fluctuate based on dynamic criteria (e.g., statistics, timing, etc.), all such configurations are anticipated by the inventor.
One skilled in the art will appreciate that the design choices made with respect to descriptor size, and arbitration, is a result of trying to provide ready access to descriptors, both transmit and receive, for each supported OSD, while also trying to keep the cost of the controller 600 close to the cost of a non shared controller. Increasing the descriptor cache size impacts cost. Thus, descriptor arbitration schemes are used to best allocate the memory used to store the descriptors in a manner that optimizes performance. For example, if all of the descriptor memory is taken, and an OSD needs to obtain transmit descriptors to perform a transmit, a decision must be made to flush certain active descriptors in the cache. Which descriptors should be flushed? For which OSD? What has been described above are a number of descriptor arbitration models, which allow a designer to utilize static or dynamic criteria in allocating descriptor space, based on the type of descriptor and the OSD.
In embodiment (c), a virtual transmit descriptor cache 812 is provided to store transmit descriptors for the supported OSD's, and a virtual receive descriptor 814 is provided to store receive descriptors for the supported OSD's. This embodiment is essentially a specific implementation of embodiment (b) that prevents transmit descriptors for one OSD from overwriting active receive descriptors. Although not shown, it should be appreciated that tags for each of the descriptors are also stored within the transmit/received caches 812, 814, respectively.
What should be appreciated from the above is that for the shared network interface controller 600 to support multiple OSD's, memory/storage must be provided on the controller 600 for storing descriptors, and some mechanism should exist for associating the descriptors with their OSD. Three embodiments for accomplishing the association have been shown but others are possible without departing from the scope of the present invention.
Referring back to
A number of embodiments for accomplishing such buffering are envisioned by the applicant, and are illustrated in
Embodiment (b) provides a virtual transmit fifo 712 and a virtual receive fifo 716, coupled to OSD management 714, 718, respectively. In addition, the transmit fifo 712 includes tag logic 713 for storing origin OSD tags (or destination MAC address information) for each packet within the fifo 712, and the receive fifo 716 includes tag logic 715 for storing destination OSD tags (or destination MAC address information) for each packet within the fifo 716. The virtual fifo's are capable of storing communication from/to any of the supported OSD's as long as the communication is tagged or associated with its origin/destination OSD. The purpose of the OSD management 714, 718 is to insure such association. Details of how communication gets associated with its OSD will be described below with reference back to
Embodiment (c) provides a single virtual fifo 720, for buffering both transmit and receive communication for all of the supported OSD's, and tag logic 721 for storing tag information to associate transmit and receive communication with the supported OSD's, as explained with reference to embodiment (b). The single virtual fifo is coupled to OSD management 722, as above. The OSD management 722 tags each of the communications with their associated OSD, and indicates whether the communication is transmit or receive. One skilled in the art will appreciate that although three embodiments of transmit/receive fifo's are shown, others are possible. What is important is that the controller 600 provide buffering for transmit/receive packets for multiple OSD's, which associates each of the transmit/receive packets with their origin or destination OSD(s).
Referring back to
The controller 600 further includes statistics logic 624. The statistics logic provides statistics, locally per OSD, and globally for the controller 600, for packets transmitted and received by the controller 600. For example, local statistics may include the number of packets transmitted and/or received per OSD, per network port. Global statistics may included the number of packets transmitted and/or received per network port, without regard to OSD. Further, as will be explained further below, it is important for loopback, broadcast, and multicast packets, to consider the statistics locally per OSD, and globally, as if such packets were being transmitted/received through non shared interface controllers. That is, a server to server communication through the shared network interface controller should have local statistics that look like X packets transmitted by a first OSD, and X packets received by a second OSD, even though as described below with reference to
What has been described above is one embodiment of a shared network interface controller 600, having a number of logical blocks which provide support for transmitting/receiving packets to/from a network for multiple OSD's. To accomplish the support necessary for sharing the controller 600 among multiple OSD's, blocks which are considered OSD specific have been replicated or vitualized with tags to associate data with its OSD. Association logic has also been provided for mapping an OSD to one (or more) MAC addresses. Other embodiments which accomplish these purposes are also envisioned.
Further, one skilled in the art will appreciate that the logical blocks described with reference to
Referring now to
Referring now to
At some point in time, processing complex 1102 alerts the controller 1101 that it has packet “0” in its memory, and requires that it be transferred to the network. Typically, this is accomplished by writing into a head pointer within the CSR 1120 associated with that processing complex 1102. The controller 1101 will arbitrate for one of the dma engines 1124 to dma the descriptors associated with the packet into its descriptor cache 1122. The controller will then use the descriptors, and initiates a dma of the packet into its virtual transmit fifo 1130. When the packet is placed into the fifo 1130, a tag indicating the OSD origin of the packet is placed into the fifo 1130 along with the packet.
At another point in time, processing complex 1106 alerts the controller 1101 that it has packet “N” in its memory, and requires that it be transferred to the network. The controller 1101 obtains the descriptors for packet “N” similar to above, and then dma's the packet into the fifo 1130.
As shown, the packets arrive in the order “N”, then “0”, and are placed into the fifo 1130 in that order. The packets are then transmitted to the network 1140.
Also, at some point in time, packet “1” is received from the network 1140 and is placed into the receive fifo 1132. Upon receipt, the destination MAC address of the packet is looked up in the association logic 1128 to determine which OSD corresponds to the packet. In this case, processing complex 1104 (“1”) is associated with the packet, and the packet is tagged as such within the fifo 1132. Once the packet is in the fifo 1132, the controller 1101 determines whether receive descriptors exist in the descriptor cache 1122 for processing complex 1104. If so, it uses these descriptors to initiate a dma of the packet from the controller 1101 to processing complex 1104. If the descriptors do not exist, the controller 1101 obtains receive descriptors from processing complex 1104, then dma's the packet to processing complex 1104 to the memory locations specified by the descriptors. Communication to the processing complex 1104 from the controller 1101 contains OSD header information, specifically designating to the shared I/O switch 1110 which of its upstream processing complexes 1102, 1104, 1106 the communication is intended.
The description above with respect to
Referring now to
In the example illustrated in
The above example is provided to illustrate that packets transmitted by any one of the supported processing complexes may be destined for one of the other processing complexes connected to the shared controller 1201. If this is the case, it would be inappropriate (at least within an Ethernet network) to present such a packet onto the network 1140, since it will not be returned. Thus, the controller 1201 has been designed to detect, using the destination MAC address, and the association logic 1228, whether any transmit packet is destined for one of the other processing complexes. And, if such is the case, packet replication logic causes the packet to be placed into the receive fifo 1232, to get the packet to the correct processing complex(es).
Referring now to
Each of the above packet flows, with respect to
Referring now to
In an alternative embodiment (b), the replication logic 1818 is placed between the transmit/receive fifo's 1830, 1832 and the transmit/receive logic. Thus, a loopback packet is allowed to be transferred from an OSD into the transmit fifo 1830. Once it is in the transmit fifo 1830, a determination is made that its destination MAC address corresponds to one of the OSD's supported by the controller. Thus, packet replication logic 1818 causes the packet to be transferred into the receive fifo 1832 for later transfer to the destination OSD.
In yet another embodiment (c), the replication logic 1818 is placed either between the fifo's and the transmit/receive logic, or between the bus interface 1814 and the fifo's 1830, 1832. In either case, a loopback fifo 1833 is provided as a separate buffer for loopback packets. The loopback fifo 1833 can be used to store loopback packets, regardless of when the loopback condition is determined (i.e., before transfer from the OSD; or after transfer into the transmit fifo 1830).
What should be appreciated from the above discussion is that a number of implementations exist to detect whether a transmit packet from one OSD has as its destination any of the other OSD's supported by the shared controller. As long as the controller detects such an event (a “loopback”), and forwards the packet to the appropriate destination OSD(s), the shared controller has efficiently, and effectively communicated the packet accurately.
Referring now to
At decision block 1404, a determination is made as to whether a packet has been received. If not, flow proceeds back to decision block 1404. If a packet has been received, flow proceeds to decision block 1406. In an alternative embodiment, a determination is made as to whether the header portion of a packet has been received. That is, once the header portion of a packet is received, it is possible to associate the destination MAC address with one (or more) OSD's, without waiting for the packet to be completely received.
At decision block 1406, a determination is made as to whether the destination MAC address of the packet matches any of the MAC addresses for which the controller is responsible. If not, flow proceeds to block 1408 where the packet is dropped. However, if a match exists, flow proceeds to block 1410.
At block 1410, association logic is consulted to determine which OSD's correspond to the destination MAC addresses referenced in the received packet. A further determination is made as to whether the MAC addresses correspond to particular virtual lans (VLAN's) for a particular OSD. Flow then proceeds to block 1412.
At block 1412, the packet is stored in the receive fifo, and designating with its appropriate OSD(s). Flow then proceeds to decision block 1414.
At decision block 1414, a determination is made as to whether the controller contains a valid receive descriptor for the designated OSD. If not, flow proceeds to block 1416 where the controller retrieves a valid receive descriptor from the designated OSD, and returns flow to block 1418. If the controller already has a valid receive descriptor for the designated OSD, flow proceeds to block 1418.
At block 1418, the packet begins transfer to the designated OSD (via the shared I/O switch). Flow then proceeds to block 1420.
At block 1420, packet transfer is completed. Flow then proceeds to decision block 1422.
At decision block 1422, a determination is made as to whether the packet is destined for another OSD. If not, flow proceeds to block 1424 where the method completes. But, if the packet is destined for another OSD, flow returns to decision block 1414 for that designated OSD. This flow continues for all designated OSD's.
Referring now to
Flow begins at block 1502 and proceeds to block 1504.
At block 1504, a determination is made as to which OSD is transmitting the packet. Flow then proceeds to block 1506.
At block 1506, a valid transmit descriptor for the transmit OSD is obtained from the OSD. Flow then proceeds to block 1507.
At block 1507, the packet is dma'ed into the transmit fifo. Flow then proceeds to decision block 1508. Note, as discussed above, in one embodiment, the OSD places the destination MAC address within the descriptor to allow the packet replication logic to determine whether a loopback condition exists, prior to transferring the packet into the transmit fifo. In an alternative embodiment, the OSD does not do the copy, so the shared controller does not associate a packet with loopback until the first part of the header has been read from the OSD. In either case, the loopback condition is determined prior to block 1520. If the destination MAC address (and/or an indication of broadcast or multicast) is sent with the descriptor, the packet replication logic can determine whether a loopback condition exists, and can therefore steer the dma engine to transfer the packet directly into the receive fifo. Alternatively, if the descriptor does not contain the destination MAC address (for loopback determination), then a determination of loopback cannot be made until the packet header comes into the controller. In this instance, the packet header could be examined while in the bus interface, to alert the packet replication logic whether to steer the packet into the transmit fifo, or into the receive fifo. Alternatively, the packet could simply be stored into the transmit fifo, and await for packet replication logic to determine whether a loopback condition exists.
At decision block 1508 a determination is made as to whether the transmit packet is either a broadcast or a multicast packet. If the packet is either a broadcast or multicast packet, flow proceeds to block 1510 where packet replication is notified. In one embodiment, packet replication is responsible for managing packet transfer to multiple MAC addresses by tagging the packet with information corresponding to each destination OSD, and for insuring that the packet is transmitted to each destination OSD. While not shown, one implementation utilizes a bit-wise OSD tag (i.e., one bit per supported OSD), such that an eight bit tag could reference eight possible OSD destinations for a packet. Of course, any manner of designating OSD destinations for a packet may be used without departing from the scope of the present invention. Once the tagging of the packet for destination OSD's is performed, flow proceeds to decision block 1512.
At decision block 1512, a determination is made as to whether the transmit packet is a loopback packet. As mentioned above, on an Ethernet network, a network interface controller may not transmit a packet which is ultimately destined for one of the devices it supports. In non shared controllers, this is never the case (unless an OSD is trying to transmit packets to itself). But, in a shared controller, it is likely that for server to server communications, a transfer packet is presented to the controller for a destination MAC address that is within the realm of responsibility of the controller. This is called a loopback packet. Thus, the controller examines the destination MAC address of the packet to determine whether the destination is for one of the OSD's for which the controller is responsible. If not, flow proceeds to block 1520. However, if the packet is a loopback packet, flow proceeds to block 1514.
At block 1514, the packet is transferred to the receive fifo rather than the transmit fifo. Flow then proceeds to block 1516.
At block 1516, the destination OSD is notified that a packet has been received for it. In one embodiment this requires CSR's for the destination OSD to be updated. Flow then proceeds to block 1518.
At block 1518, flow proceeds to the flow chart of
At block 1520, the packet is transferred to the transmit fifo. Flow then proceeds to block 1522.
At block 1522, the packet is transmitted out to the network. Flow then proceeds to block 1524.
At block 1524, packet transmit is completed. Flow then proceeds to block 1526 where the method completes.
Referring now to
Also shown in
While not particularly shown, one skilled in the art will appreciate that many alternative embodiments may be implemented which differ from the above description, while not departing from the scope of the invention as claimed. For example, the context of the processing complexes, i.e., the environment in which they are placed has not been described because such discussion is exhaustively provided in the parent application(s). However, one skilled in the art will appreciate that the processing complexes (or operating system domains) of the present application should be read to include at least one or more processor cores within a SOC, or one or more processors within a board level system, whether the system is a desktop, server or blade. Moreover, the location of the shared I/O switch, whether placed within an SOC, on the backplane of a blade enclosure, or within a shared network interface controller should not be controlling. Rather, it is the provision of a network interfacecontroller which can process transmits/receives for multiple processing complexes, as part of their load-store domain, to which the present invention is directed. This is true whether the OSD ID logic is within the shared network interface controller, or whether the shared network interface controller provides multiple upstream OSD aware (or non OSD aware) ports. Further, it is the tracking of outstanding transmits/receives such that the transmits/receives are accurately associated with their upstream links (or OSD's) that is important.
Additionally, the above discussion has described the present invention within the context of three processing complexes communicating with the shared network interface controller. The choice of three processing complexes was simply for purposes of illustration. The present invention could be utilized in any environment that has one or more processing complexes (servers, CPU's, etc.) that require access to a network.
Further, the present invention has utilized a shared I/O switch to associate and route packets from processing complexes to the shared network interface controller. It is within the scope of the present invention to incorporate the features of the present invention within a processing complex (or chipset) such that everything downstream of the processing complex is shared I/O aware (e.g., PCI Express+). If this were the case, the shared network interface controller could be coupled directly to ports on a processing complex, as long as the ports on the processing complex provided shared I/O information to the shared network interface controller, such as OS Domain information. What is important is that the shared network interface controller be able to recognize and associate packets with origin or upstream OS Domains, whether or not a shared I/O switch is placed external to the processing complexes, or resides within the processing complexes themselves.
And, if the shared I/O switch were incorporated within the processing complex, it is also possible to incorporate one or more shared network interface controllers into the processing complex. This would allow a single processing complex to support multiple upstream OS Domains while packaging everything necessary to talk to fabrics outside of the load/store domain (Ethernet, Fiber Channel, SATA, etc.) within the processing complex. Further, if the upstream OS Domains were made shared I/O aware, it is also possible to couple the domains directly to the network interface controllers, all within the processing complex.
And, it is envisioned that multiple shared I/O switches according to the present invention be cascaded to allow many variations of interconnecting processing complexes with downstream I/O devices such as the shared network interface controller. In such a cascaded scenario, an OS Header may be global, or it might be local. That is, it is possible that a local ID be placed within an OS Header, the local ID particularly identifying a packet, within a given link (e.g., between a processing complex and a switch, between a switch and a switch, and/or between a switch and an endpoint). So, a local ID may exist between a downstream shared I/O switch and an endpoint, while a different local ID may be used between an upstream shared I/O switch and the downstream shared I/O switch, and yet another local ID between an upstream shared I/O switch and a root complex. In this scenario, each of the switches would be responsible for mapping packets from one port to another, and rebuilding packets to appropriately identify the packets with their associating upstream/downstream port.
It is also envisioned that the addition of an OSD header within a load-store fabric, as described above, could be further encapsulated within another load-store fabric yet to be developed, or could be further encapsulated, tunneled, or embedded within a channel-based fabric such as Advanced Switching (AS) or Ethernet. AS is a multi-point, peer-to-peer switched interconnect architecture that is governed by a core AS specification along with a series of companion specifications that define protocol encapsulations that are to be tunneled through AS fabrics. These specifications are controlled by the Advanced Switching Interface Special Interest Group (ASI-SIG), 5440 SW Westgate Drive, Suite 217, Portland, Oreg. 97221 (Phone: 503-291-2566). For example, within an AS embodiment, the present invention contemplates employing an existing AS header that specifically defines a packet path through a I/O switch according to the present invention. Regardless of the fabric used downstream from the OS domain (or root complex), the inventors consider any utilization of the method of associating a shared I/O endpoint with an OS domain to be within the scope of their invention, as long as the shared I/O endpoint is considered to be within the load-store fabric of the OS domain.
Further, the above discussion has been directed at an embodiment of the present invention within the context of the Ethernet network protocol. This was chosen to illustrate the novelty of the present invention with respect to providing a shareable controller for access to a network. One skilled in the art should appreciate that other network protocols such as Infiniband, OC48/OC192, ATM, SONET, 802.11 are encompassed within the above discussion to allow for sharing controllers for such protocols among multiple processing complexes. Further, Ethernet should be understood as including the general class of IEEE Ethernet protocols, including various wired and wireless media. It is not the specific protocol to which this invention is directed. Rather, it is the sharing of a controller by multiple processing complexes which is of interest. Further, although the term MAC address should be appreciated by one skilled in the art, it should be understood as an address which is used by the Media Access Control sublayer of the Data-Link Layer (DLC) of telecommunication protocols. There is a different MAC sublayer for each physical device type. The other sublayer level in the DLC layer is the Logical Link Control sublayer.
Although the present invention and its objects, features and advantages have been described in detail, other embodiments are encompassed by the invention. In addition to implementations of the invention using hardware, the invention can be implemented in computer readable code (e.g., computer readable program code, data, etc.) embodied in a computer usable (e.g., readable) medium. The computer code causes the enablement of the functions or fabrication or both of the invention disclosed herein. For example, this can be accomplished through the use of general programming languages (e.g., C, C++, JAVA, and the like); GDSII databases; hardware description languages (HDL) including Verilog HDL, VHDL, Altera HDL (AHDL), and so on; or other programming and/or circuit (i.e., schematic) capture tools available in the art. The computer code can be disposed in any known computer usable (e.g., readable) medium including semiconductor memory, magnetic disk, optical disk (e.g., CD-ROM, DVD-ROM, and the like), and as a computer data signal embodied in a computer usable (e.g., readable) transmission medium (e.g., carrier wave or any other medium including digital, optical or analog-based medium). As such, the computer code can be transmitted over communication networks, including Intemets and intranets. It is understood that the invention can be embodied in computer code (e.g., as part of an IP (intellectual property) core, such as a microprocessor core, or as a system-level design, such as a System on Chip (SOC)) and transformed to hardware as part of the production of integrated circuits. Also, the invention may be embodied as a combination of hardware and computer code.
Finally, those skilled in the art should appreciate that they can readily use the disclosed conception and specific embodiments as a basis for designing or modifying other structures for carrying out the same purposes of the present invention without departing from the spirit and scope of the invention as defined by the appended claims.
This application claims the benefit of the following U.S. Provisional Applications which are hereby incorporated by reference for all purposes: Serial No.Filing DateTitle60/491314Jul. 30, 2003SHARED NIC BLOCK DIAGRAM(NEXTIO.0104)60/515558Oct. 29, 2003NEXSIS(NEXTIO.0105)60/523522Nov. 19, 2003SWITCH FOR SHARED I/O(NEXTIO.0106)FABRIC60/541673Feb. 4, 2004PCI SHARED I/O WIRE LINE(NEXTIO.0107)PROTOCOL60/555127Mar. 22, 2004PCI EXPRESS SHARED IO(NEXTIO.0108)WIRELINE PROTOCOLSPECIFICATION60/575005May 27, 2004NEXSIS SWITCH(NEXTIO.0109)60/588941Jul. 19, 2004SHARED I/O DEVICE(NEXTIO.0110)60/589174Jul, 19, 2004ARCHITECTURE(NEXTIO.0111) This application is a Continuation-in-Part (CIP) of the below referenced pending U.S. Non-Provisional Patent Applications: Serial No.Filing DateTitle10/757714Jan. 14, 2004METHOD AND APPARATUS FOR(NEXTIO.0300)SHARED I/O IN A LOAD/STOREFABRIC10/757713Jan. 14, 2004METHOD AND APPARATUS FOR(NEXTIO.0301)SHARED I/O IN A LOAD/STOREFABRIC10/757711Jan. 14, 2004METHOD AND APPARATUS FOR(NEXTIO.0302)SHARED I/O IN A LOAD/STOREFABRIC10/802532Mar. 16, 2004SHARED INPUT/OUTPUT LOAD-(NEXTIO.0200)STORE ARCHITECTURE10/864766Jun. 9, 2004METHOD AND APPARATUS FOR(NEXTIO.0310)A SHARED I/O SERIAL ATACONTROLLER10/827622Apr. 19, 2004SWITCHING APPARATUS AND(NEXTIO.0400)METHOD FOR PROVIDINGSHARED I/O WITHIN A LOAD-STORE FABRIC10/827620Apr. 19, 2004SWITCHING APPARATUS AND(NEXTIO.0401)METHOD FOR PROVIDINGSHARED I/O WITHIN A LOAD-STORE FABRIC10/827117Apr. 19, 2004SWITCHING APPARATUS AND(NEXTIO.0402)METHOD FOR PROVIDINGSHARED I/O WITHIN A LOAD-STORE FABRIC each of which are assigned to a common assignee (NextIO Inc.), and each of which are hereby incorporated by reference herein for all purposes. All of the above referenced applications claim priority from the above referenced provisional applications which have a provisional filing date earlier than their non-provisional filing date. In addition, all of the above applications claim priority from the below referenced provisional applications which were not expired prior to their non-provisional filing date: Serial No.Filing DateTitle60/440788Jan. 15, 2003SHARED IO ARCHITECTURE(NEXTIO.0101)60/440789Jan. 21, 20033GIO-XAUI COMBINED SWITCH(NEXTIO.0102)60/464382Apr. 18, 2003SHARED-IO PCI COMPLIANT(NEXTIO.0103)SWITCH each of which are hereby incorporated by reference herein for all purposes.
| Number | Date | Country | |
|---|---|---|---|
| 60491314 | Jul 2003 | US | |
| 60515558 | Oct 2003 | US | |
| 60523522 | Nov 2003 | US | |
| 60541673 | Feb 2004 | US | |
| 60555127 | Mar 2004 | US | |
| 60575005 | May 2004 | US | |
| 60588941 | Jul 2004 | US | |
| 60589174 | Jul 2004 | US |
