Patent Grant
12135646
Some memory systems include a multi-level cache system, in which a hierarchy of memories (e.g., caches) provides varying access speeds to cache data. A first level (L1) cache is closely coupled to a central processing unit (CPU) core and provides the CPU core with relatively fast access to cache data. A second level (L2) cache is also coupled to the CPU core and, in some examples, is larger and thus holds more data than the L1 cache, although the L2 cache provides relatively slower access to cache data than the L1 cache. Additional memory levels of the hierarchy are possible.
In accordance with at least one example of the disclosure, a method includes determining, by a level one (L1) controller, to change a size of a L1 main cache; servicing, by the L1 controller, pending read requests and pending write requests from a central processing unit (CPU) core; stalling, by the L1 controller, new read requests and new write requests from the CPU core; writing back and invalidating, by the L1 controller, the L1 main cache. The method also includes receiving, by a level two (L2) controller, an indication that the L1 main cache has been invalidated and, in response, flushing a pipeline of the L2 controller; in response to the pipeline being flushed, stalling, by the L2 controller, requests received from any master; reinitializing, by the L2 controller, a shadow L1 main cache. Reinitializing includes clearing previous contents of the shadow L1 main cache and changing the size of the shadow L1 main cache.
In accordance with at least one example of the disclosure, an apparatus, includes a central processing unit (CPU) core and a level one (L1) cache subsystem coupled to the CPU core. The L1 cache subsystem includes a L1 main cache and a L1 controller. The L1 controller is configured to determine to change a size of the L1 main cache, service pending read requests and pending write requests from the CPU core, stall new read requests and new write requests from the CPU core, and write back and invalidate the L1 main cache. The apparatus also includes a level two (L2) cache subsystem coupled to the L1 cache subsystem. The L2 cache subsystem includes a L2 main cache, a shadow L1 main cache, and a L2 controller. The L2 controller is configured to receive an indication that the L1 main cache has been invalidated and, in response, flush a pipeline of the L2 controller; in response to the pipeline being flushed, stall requests received from any master; and reinitialize the shadow L1 main cache. Reinitializing includes clearing previous contents of the shadow L1 main cache and changing the size of the shadow L1 main cache.
In accordance with another example of the disclosure, a method includes receiving, by a level two (L2) controller, an indication that the a level one (L1) main cache has been invalidated and, in response, flushing a pipeline of the L2 controller; in response to the pipeline being flushed, stalling, by the L2 controller, requests received from any master; and reinitializing, by the L2 controller, a shadow L1 main cache. Reinitializing includes clearing previous contents of the shadow L1 main cache and changing the size of the shadow L1 main cache.
In accordance with at least one example of the disclosure, an apparatus includes a CPU core and a L1 cache subsystem coupled to the CPU core. The L1 cache subsystem includes a L1 main cache, a L1 victim cache, and a L1 controller. The apparatus includes a L2 cache subsystem coupled to the L1 cache subsystem. The L2 cache subsystem includes a L2 main cache, a shadow L1 main cache, a shadow L1 victim cache, and a L2 controller. The L2 controller receives an indication from the L1 controller that a cache line A is being relocated from the L1 main cache to the L1 victim cache; in response to the indication, update the shadow L1 main cache to reflect that the cache line A is no longer located in the L1 main cache; and in response to the indication, update the shadow L1 victim cache to reflect that the cache line A is located in the L1 victim cache.
In accordance with at least one example of the disclosure, a method includes receiving, by a level two (L2) controller of a L2 cache subsystem, an indication from a level one (L1) cache subsystem that a cache line A is being relocated from a L1 main cache to a L1 victim cache; in response to the indication, updating, by the L2 controller, a shadow L1 main cache of the L2 cache subsystem to reflect that the cache line A is no longer located in the L1 main cache; and in response to the indication, updating a shadow L1 victim cache of the L2 cache subsystem to reflect that the cache line A is located in the L1 victim cache.
In accordance with at least one example of the disclosure, a level two (L2) cache subsystem includes a L2 main cache; a shadow level one (L1) main cache; a shadow L1 victim cache; and a L2 controller. The L2 controller is configured to: receive an indication from a level one (L1) controller that a cache line A is being relocated from a L1 main cache to a L1 victim cache; in response to the indication, update the shadow L1 main cache to reflect that the cache line A is no longer located in the L1 main cache; and in response to the indication, update the shadow L1 victim cache to reflect that the cache line A is located in the L1 victim cache.
In accordance with at least one example of the disclosure, a system includes a non-coherent component; a coherent, non-caching component; a coherent, caching component; and a level two (L2) cache subsystem coupled to the non-coherent component, the coherent, non-caching component, and the coherent, caching component. The L2 cache subsystem includes a L2 cache; a shadow level one (L1) main cache; a shadow L1 victim cache; and a L2 controller. The L2 controller is configured to receive and process a first transaction from the non-coherent component; receive and process a second transaction from the coherent, non-caching component; and receive and process a third transaction from the coherent, caching component.
In accordance with another example of the disclosure, a method includes maintaining, by a level two (L2) cache controller, a L2 cache, a shadow level one (L1) main cache and a shadow L1 victim cache; receiving and processing, by the L2 cache controller, a first transaction from a non-coherent component; receiving and processing, by the L2 cache controller, a second transaction from a coherent, non-caching component; and receiving and processing, by the L2 cache controller, a third transaction from a coherent, caching component.
In accordance with at least one example of the disclosure, an apparatus includes a CPU core and a L1 cache subsystem including a L1 main cache, a L1 victim cache, and a L1 controller. The apparatus includes a L2 cache subsystem including a L2 main cache, a shadow L1 main cache, a shadow L1 victim cache, and a L2 controller configured to receive a read request from the L1 controller as a single transaction. Read request includes a read address, a first indication of an address and a coherence state of a cache line A to be moved from the L1 main cache to the L1 victim cache to allocate space for data returned in response to the read request, and a second indication of an address and a coherence state of a cache line B to be removed from the L1 victim cache in response to the cache line A being moved to the L1 victim cache.
In accordance with at least one example of the disclosure, a method includes receiving, by a level two (L2) controller, a read request from a level one (L1) controller as a single transaction. The read request includes a read address, a first indication of an address and a coherence state of a cache line A to be moved from a L1 main cache to a L1 victim cache to allocate space for data returned in response to the read request, and a second indication of an address and a coherence state of a cache line B to be removed from the L1 victim cache in response to the cache line A being moved to the L1 victim cache.
In accordance with at least one example of the disclosure, a level two (L2) cache subsystem, includes a L2 main cache, a shadow L1 main cache, a shadow L1 victim cache, and a L2 controller configured to receive a read request in a single transaction from a level one (L1) controller. The read request includes a read address, a first indication of an address and a coherence state of a cache line A to be moved from a L1 main cache to a L1 victim cache to allocate space for data returned in response to the read request, and a second indication of an address and a coherence state of a cache line B to be removed from the L1 victim cache in response to the cache line A being moved to the L1 victim cache.
In accordance with at least one example of the disclosure, an apparatus includes a CPU core and a L1 cache subsystem including a L1 main cache, a L1 victim cache, and a L1 controller. The apparatus includes a L2 cache subsystem coupled to the L1 cache subsystem by a transaction bus and a tag update bus. The L2 cache subsystem includes a L2 main cache, a shadow L1 main cache, a shadow L1 victim cache, and a L2 controller. The L2 controller receives a message from the L1 controller over the tag update bus, including a valid signal, an address, and a coherence state. In response to the valid signal being asserted, the L2 controller identifies an entry in the shadow L1 main cache or the shadow L1 victim cache having an address corresponding to the address of the message and updates a coherence state of the identified entry to be the coherence state of the message.
In accordance with at least one example of the disclosure, a method includes receiving, by a level two (L2) controller, a message from a level one (L1) controller over a tag update bus separate from a transaction bus between the L2 controller and the L1 controller, the message comprising a valid signal, an address, and a coherence state. The method also includes, in response to the valid signal being asserted, identifying, by the L2 controller, an entry in a shadow L1 main cache or a shadow L1 victim cache having an address corresponding to the address of the message; and updating, by the L2 controller, a coherence state of the identified entry to be the coherence state of the message.
In accordance with at least one example of the disclosure, an apparatus includes a central processing unit (CPU) core and a level one (L1) cache subsystem coupled to the CPU core. The L1 cache subsystem includes a L1 main cache, a L1 victim cache, and a L1 controller. The apparatus includes a level two (L2) cache subsystem coupled to the L1 cache subsystem by a transaction bus and a tag update bus. The L2 cache subsystem includes a L2 main cache, a shadow L1 main cache, a shadow L1 victim cache, and a L2 controller. The L2 controller is configured to receive a message from the L1 controller over the tag update bus, the message comprising a valid signal and an address. In response to the valid signal being asserted, the L2 controller identifies an entry in the shadow L1 victim cache having an address corresponding to the address of the message and updates a coherence state of the identified entry to be invalid.
In accordance with at least one example of the disclosure, a method includes receiving, by a level two (L2) controller, a first request for a cache line in a shared cache coherence state; mapping, by the L2 controller, the first request to a second request for a cache line in an exclusive cache coherence state; and responding, by the L2 controller, to the second request.
In accordance with at least one example of the disclosure, a method includes receiving, by a level two (L2) controller, a first request for a cache line in a shared cache coherence state; determining, by the L2 controller, that the cache line is not present in a L2 cache; mapping, by the L2 controller, the first request to a second request for a cache line in an exclusive cache coherence state; and forwarding, by the L2 controller, the second request.
In accordance with at least one example of the disclosure, an apparatus includes a central processing unit (CPU) core, a level one (L1) cache subsystem coupled to the CPU core, and a level two (L2) cache subsystem coupled to the L1 cache subsystem. The L2 cache subsystem includes a L2 main cache and a L2 controller. The L2 controller is configured to receive a first request for a cache line in a shared cache coherence state, map the first request to a second request for a cache line in an exclusive cache coherence state, and respond to the second request.
For a detailed description of various examples, reference will now be made to the accompanying drawings in which:
In some examples, the CPU cores 102a-102n include a register file, an integer arithmetic logic unit, an integer multiplier, and program flow control units. In an example, the L1 caches 104a-104n associated with each CPU core 102a-102n include a separate level one program cache (L1 P) and level one data cache (L1 D). The L2 caches 106a-106n are combined instruction/data caches that hold both instructions and data. In certain examples, a CPU core 102a and its associated L1 cache 104a and L2 cache 106a are formed on a single integrated circuit.
The CPU cores 102a-102n operate under program control to perform data processing operations upon data. Instructions are fetched before decoding and execution. In the example of
The unified L2 cache 106 is further coupled to a third level (L3) cache 108, which is shared by the L2 caches 106a-106n in the example of
The L2 cache subsystem 206 includes L2 tag ram 212, L2 coherence (e.g., MESI) data 214, shadow L1 tag ram 216, and L1 coherence (e.g., MESI) data 218. Each of the blocks 212, 214, 216, 218 are alternately referred to as a memory or a RAM. The L2 cache subsystem 206 also includes tag ram error correcting code (ECC) data 220. In an example, the ECC data 220 is maintained for each of the memories 212, 214, 216, 218.
The L2 cache subsystem 206 includes L2 controller 222, the functionality of which will be described in further detail below. In the example of
In the example of
The L2 tag ram 212 includes a list of the physical addresses whose contents (e.g., data or program instructions) have been cached to the L2 cache 224. In an example, an address translator translates virtual addresses to physical addresses. In one example, the address translator generates the physical address directly from the virtual address. For example, the lower n bits of the virtual address are used as the least significant n bits of the physical address, with the most significant bits of the physical address (above the lower n bits) being generated based on a set of tables configured in main memory. In this example, the L2 cache 224 is addressable using physical addresses. In certain examples, a hit/miss indicator from a tag ram 212 look-up is stored.
The L2 MESI memory 214 maintains coherence data to implement full MESI coherence with L2 SRAM 224, external shared memories, and data cached in L2 cache from other places in the system 200. The functionalities of system 200 coherence are explained in further detail below.
The L2 cache subsystem 206 also shadows L1D tags in the L1D shadow tag ram 216 and L1D MESI memory 218. The tag ram ECC data 220 provides error detection and correction for the tag memories and, additionally, for one or both of the L2 MESI memory 214 and the L1D MESI memory 218. The L2 cache controller 222 generally controls the operations of the L2 cache subsystem 206, including handling coherency operations both internal to the L2 cache subsystem 206 and among the other components of the system 200.
The L1 controller 310 is coupled to a L2 controller 320 of the L2 cache subsystem 306. The L2 controller 320 also couples to L2 SRAM 322. The L2 controller 320 couples to a L2 cache 324 and to a shadow of the L1 main cache 326 as well as a shadow of the L1 victim cache 328. L2 cache 324 and L2 SRAM 322 are shown separately for ease of discussion, although may be implemented physically together (e.g., as part of L2 SRAM 224, including in a banked configuration, as described above. Similarly, the shadow L1 main cache 326 and the shadow L1 victim cache 328 may be implemented physically together, and are similar to the L1D shadow tag ram 216 and the L1 D MESI 218, described above. The L2 controller 320 is also coupled to a L3 controller 309 of the L3 cache subsystem 308. L3 cache and main memory (e.g., DDR 110 described above) are not shown for simplicity.
Cache coherence is a technique that allows data and program caches, as well as different requestors (including requestors that do not have caches) to determine the most current data value for a given address in memory. Cache coherence enables this coherent data value to be accurately reflected to observers (e.g., a cache or requestor that issues commands to read a given memory location) present in the system 300. Certain examples of this disclosure refer to an exemplary MESI coherence scheme, in which a cache line is set to one of four cache coherence states: modified, exclusive, shared, or invalid. Other examples of this disclosure refer to a subset of the MESI coherence scheme, while still other examples include more coherence states than the MESI coherence scheme. Regardless of the coherence scheme, cache coherence states for a given cache line are stored in, for example, the L2 MESI memory 214 described above.
A cache line having a cache coherence state of modified indicates that the cache line is modified with respect to main memory (e.g., DDR 110), and the cache line is held exclusively in the current cache (e.g., the L2 cache 324). A modified cache coherence state also indicates that the cache line is explicitly not present in any other caches (e.g., L1 or L3 caches).
A cache line having a cache coherence state of exclusive indicates that the cache line is not modified with respect to main memory (e.g., DDR 110), but the cache line is held exclusively in the current cache (e.g., the L2 cache 324). An exclusive cache coherence state also indicates that the cache line is explicitly not present in any other caches (e.g., L1 or L3 caches).
A cache line having a cache coherence state of shared indicates that the cache line is not modified with respect to main memory (e.g., DDR 110). A shared cache state also indicates that the cache line may be present in multiple caches (e.g., caches in addition to the L2 cache 324).
A cache line having a cache coherence state of invalid indicates that the cache line is not present in the cache (e.g., the L2 cache 324).
Examples of this disclosure leverage hardware techniques, control logic, and/or state information to implement a coherent system. Each observer can issue read requests—and certain observers are able to issue write requests—to memory locations that are marked shareable. Caches in particular can also have snoop requests issued to them, requiring their cache state to be read, returned, or even updated, depending on the type of the snoop operation. In the exemplary multi-level cache hierarchy described above, the L2 cache subsystem 306 is configured to both send and receive snoop operations. The L1 cache subsystem 304 receives snoop operations, but does not send snoop operations. The L3 cache subsystem 308 sends snoop operations, but does not receive snoop operations. In examples of this disclosure, the L2 cache controller 320 maintains state information (e.g., in the form of hardware buffers, memories, and logic) to additionally track the state of coherent cache lines present in both the L1 main cache 314 and the L1 victim cache 316. Tracking the state of coherent cache lines enables the implementation of a coherent hardware cache system.
Examples of this disclosure refer to various types of coherent transactions, including read transactions, write transactions, snoop transactions, victim transactions, and cache maintenance operations (CMO). These transactions are at times referred to as reads, writes, snoops, victims, and CMOs, respectively.
Reads return the current value for a given address, whether that value is stored at the endpoint (e.g., DDR 110), or in one of the caches in the coherent system 300. Writes update the current value for a given address, and invalidate other copies for the given address stored in caches in the coherent system 300. Snoops read or invalidate (or both) copies of data stored in caches. Snoops are initiated from a numerically-higher level of the hierarchy to a cache at the next, numerically-lower level of the hierarchy (e.g., from the L2 controller 320 to the L1 controller 310), and are able be further propagated to even lower levels of the hierarchy as needed. Victims are initiated from a numerically-lower level cache in the hierarchy to the next, numerically-higher level of the cache hierarchy (e.g., from the L1 controller 310 to the L2 controller 320). Victims transfer modified data to the next level of the hierarchy. In some cases, victims are further propagated to numerically-higher levels of the cache hierarchy (e.g., if the L2 controller 310 sends a victim to the L2 controller 320 for an address in the DDR 110, and the line is not present in the L2 cache 324, the L2 controller 320 forwards the victim to the L3 controller 309). Finally, CMOs cause an action to be taken in one of the caches for a given address.
Still referring to
In an example, the L1 victim cache 316 is a fully associative cache that holds cache lines that have been removed from the L1 main cache 314, for example due to replacement. The L1 victim cache 316 holds both clean and dirty lines. The L1 victim cache 316 services read and write hits and snoops. The L1 victim cache 316 also keeps track of cache coherence state information (e.g., MESI state) for its cache lines. When a cache line in the modified state is replaced from the L1 victim cache 316, that cache line is sent to the L2 cache subsystem 306 as a victim.
As explained above, the L2 cache subsystem 306 includes a unified L2 cache 324 that is used to service requests from multiple requestor types, including L1 D and L1 P (through the L1 controller 310), the streaming engine 210, a memory management unit (MMU 207), and the L3 cache (through the L3 controller 309). In an example, the L2 cache 324 is non-inclusive with the L1 cache subsystem 304, which means that the L2 cache 324 is not required to include all cache lines stored in the L1 caches 314, 316, but that some lines may be cached in both levels. Continuing this example, the L2 cache 324 is also non-exclusive, which means that cache lines are not explicitly prevented from being cached in both the L1 and L2 caches 314, 316, 324. For example, due to allocation and random replacement, cache lines may be present in one, both, or neither of the L1 and L2 caches. The combination of non-inclusive and non-exclusive cache policies enables the L2 controller 320 to manage its cache contents without requiring the L1 controller 310 to invalidate or remove cache lines. This simplifies processing in the L2 cache subsystem 306 and enables increased performance for the CPU core 302 by allowing critical data to remain cached in the L1 cache subsystem 304 even if it has been evicted from the L2 cache 324.
Still referring to
Local coherence requests are received by the L2 controller 320 from, for example, the CPU core 302 or as a direct memory access (DMA) request from another CPU core or a master associated with another CPU core. External coherence requests are received by the L2 controller 320 from, for example, the CPU core 302 or L3 controller 309. Thus, the single L2 controller 320 is configured to address both local and external coherence.
In accordance with various examples, the L2 controller 320 manages the CPU core 302 coherent view of three endpoints: L2 SRAM 322, L3 SRAM (part of the L3 cache subsystem 308, not shown for simplicity), and main memory or DDR 110, described above. For ease of discussion, L3 SRAM and DDR 110 are grouped together and referred to as an “external” memory or endpoint, which distinguishes them from the L2 SRAM 322 as a “local” (e.g., to the L2 controller 320) memory or endpoint.
A master refers to a requestor that issues read and write accesses to an endpoint. In some examples, a master stores the results of these read and write accesses in a cache, although the master does not necessarily store such results in a cache. Coherent masters (e.g., masters for whom coherence must be handled by L2 controller 320) are classified as either caching or non-caching. Non-coherent masters (e.g., masters that do not require coherent data) are not distinguished as caching or non-caching due to their being non-coherent. Referring briefly back to
The L2 controller 320 is configured to provide coherent access to both internal and external endpoints for coherent masters, while also providing access to those internal and external endpoints for non-coherent masters. As will be explained in further detail below, the L2 controller manages coherent state information, issues coherence transactions (e.g., snoop, victim) to maintain proper coherence states, and propagates information as needed to the downstream controllers such as the L3 controller 309 to provide a coherent view of the memory stored in the L2 cache subsystem 306.
As will be explained further below, the L2 controller 320 is configured to perform normal cache allocation, replacement, and victimization operations, while also sending coherent transactions to communicate the storage of coherent locations within the L2 cache subsystem 306 or L1 cache subsystem 304. As a result, downstream cache controllers such as the L3 controller 309 are able to maintain the directory information, if so enabled, about what addresses are held in the L1 and L2 cache subsystems 304, 306.
In accordance with examples of this disclosure, the L2 controller 320 is part of a system that includes a non-coherent master; a non-caching, coherent master; and a caching, coherent master. The L2 controller 320 is configured to receive and process transactions from each of these masters, while maintaining global coherence (e.g., with respect to external memories) and local coherence (e.g., with respect to its local memory) as required by the particular master. Thus, the L2 controller 320 also enables interleaving of coherent and non-coherent traffic among the various masters.
The following table summarizes interactions between various masters and the L2 controller 320 in accordance with various examples. In particular, Table 1 indicates for a particular master what transaction types that master can initiate to the L2 controller 320, what transaction types the L2 controller 320 can initiate to that master, and whether global and/or local coherence is supported by the L2 controller 320 for that master.
The method 410 begins in block 412 with the L2 controller 320 receiving a read request from a coherent, non-caching master, and continues in block 414 with the L2 controller 320 determining whether the read request hits in the shadow L1 main cache 326 or the shadow L1 victim cache 328, which indicates that the requested data may be present in the L1 cache subsystem 304.
If, in block 414, the read request does not hit in the shadow L1 caches 326, 328, the method 410 continues in block 416 in which the L2 controller 320 reads the data from an endpoint and returns the data as a read response. However, if in block 414 the read request hits one of the shadow L1 caches 326, 328, the method 410 continues in block 418 with the L2 controller 320 generating a snoop read to the L1 controller 310. If a snoop response from the L1 controller 310 contains valid data in block 420, then the L2 controller 320 returns the snoop response data as a read response to the requesting master in block 422. If the snoop response from the L1 controller 310 contains invalid data in block 420, then the L2 controller 320 returns endpoint data as the read response to the requesting master in block 416.
In particular, the method 430 begins in block 432 with the L2 controller 320 receiving an allocating read request from a coherent, caching master, which in this examples is the L1 cache subsystem 304. This read request includes sideband signals that indicate it is an allocating request (e.g., alloc==1). The method 430 then proceeds to block 434 in which the L2 controller 320 writes an address and, optionally, a secure bit, indicated by the sideband signals to the shadow L1 main cache 326, which now indicates the address that is being allocated to the L1 main cache 314 as a result of this read request.
The method 430 continues in block 436 with determining whether a main_valid sideband signal is asserted, which indicates that a cache line is moving from the L1 main cache 314 to the L1 victim cache 316 as a result of this read request. If the main_valid signal is asserted, the method 430 continues to block 438 in which the L2 controller 320 updates its shadow L1 victim cache 328 to include an address specified by main_address, a coherence state specified by main_mesi, and optionally a secure bit specified by main_secure. As a result, the shadow L1 victim cache 328 now includes the address and coherence state information of the line that is being moved from the L1 main cache 314 to the L1 victim cache 316 as a result of this read request.
If the main_valid signal is de-asserted, then a line is not being moved from the L1 main cache 314 to the L1 victim cache 316 as a result of this read request, and the method 430 continues to block 440 with determining whether a victim_valid sideband signal is asserted, which indicates that a cache line is moving out of the L1 victim cache 316 as a result of this read request (e.g., is being displaced by the L1 main cache 314 to L1 victim cache 316 movement described above). If the victim_valid signal is asserted, the method 430 continues in block 442 with determining whether the coherence state specified by victim_mesi (e.g., the coherence state of the line being moved out of the L1 victim cache 316) is invalid, modified, or shared/exclusive.
If victim_mesi is invalid, the method 430 proceeds to block 448 in which the L2 controller 320 returns read response data from an endpoint, or the L2 cache 324.
If victim_mesi is shared/exclusive, the method 430 continues to block 444 where the L2 controller 320 removes an entry from its shadow L1 victim cache 328 having an address that matches victim_address and, optionally, victim_secure. As explained further below, the L2 controller 320 removes the entry in this case because a subsequent victim transaction from the L1 controller 310 does not result when the line evicted from L1 victim cache 316 is in the shared/exclusive state, and thus it is safe to also remove from the shadow L1 victim cache 328. The method 430 then proceeds to block 448 in which the L2 controller 320 returns read response data from an endpoint, or the L2 cache 324.
If victim_mesi is modified, the method 430 continues to block 446 where the L2 controller 320 retains an entry from its shadow L1 victim cache 328 having an address that matches victim_address and, optionally, victim_secure. As explained further below, the L2 controller 320 retains the entry in this case because a subsequent victim transaction from the L1 controller 310 is expected when the line evicted from L1 victim cache 316 is in the modified state. The method 430 then proceeds to block 448 in which the L2 controller 320 returns read response data from an endpoint, or the L2 cache 324.
However, if the write request hits in the shadow L1 main or victim caches 326, 328, then the method 450 proceeds to block 458 in which the L2 controller 320 issues a snoop read and invalidate request to the L1 cache subsystem 304. If the snoop response has dirty (e.g., modified) data in block 460, then the L2 controller 320 merges the write data over the snoop response data and writes to an endpoint in block 462. If the snoop response contains unmodified data in block 460, then the L2 controller 320 writes the write data to the endpoint in block 456.
If the snoop response from the L1 controller 310 contains invalid data (or if the snoop request did not hit in the shadow L1 main or victim caches 326, 328 in block 484), the method 480 continues to block 490 in which the L2 controller 320 determines whether the snoop read hits in the L2 cache 324. If the snoop read does not hit in the L2 cache 324, the method 480 continues to block 492 and the L2 controller 320 issues a snoop miss to the L3 controller 309. However, if the snoop read hits in the L2 cache 324, the method 480 continues to block 493 in which the L2 controller 320 reads the data from the L2 cache 324 and to block 494 in which the L2 controller 320 updates a coherence state as needed. Then the L2 controller 320 returns the data from the L2 cache 324 as snoop response data to the L3 controller 309 in block 495.
If the snoop response from the L1 controller 310 contains valid data in block 488, the method 480 continues to block 496 in which the L2 controller 320 determines whether the snoop response from the L1 controller 310 hits in the L2 cache 324. If the snoop response from the L1 controller 310 hits in the L2 cache 324, the method 480 continues to block 497 in which the L2 controller 320 updates a coherence state of the L2 cache 324 as needed. Then, the L2 controller 320 returns the snoop response data from the L1 controller 310 as a snoop response to the L3 controller 309 in block 498. If the snoop response from the L1 controller 310 does not hit in the L2 cache 324, the method 480 proceeds directly to block 498, in which the L2 controller 320 returns the snoop response data from the L1 controller 310 as a snoop response to the L3 controller 309.
The foregoing are examples of ways in which the L2 controller 320 receives and processes various types of transactions from various types of masters, including non-coherent masters; coherent, non-caching masters; and coherent, caching masters. By handling such diverse combinations of transactions and master requirements in a single, unified controller, overall system flexibility is enhanced.
As explained, there is a need for the L2 cache subsystem 306 to include hardware, control logic, and/or state information to allow the L2 controller at 320 to accurately track and process the state of coherent, cache lines in the lower-level L1 cache subsystem 304. In this example, the L1 cache subsystem 304 is utilizing a heterogeneous cache system, including the L1 main cache 314 and the L1 victim cache 316. Examples of this disclosure allow the L2 controller 320 to maintain appropriate state information to accurately track the state of all coherent cache lines present in both the L1 main cache 314 and L1 victim cache 316.
The shadow L1 victim cache 328 is a shadow copy of the address tag and MESI state information for the cache lines held in the L1 victim cache 316. As above with respect to the shadow L1 main cache 326, the maintenance of the shadow L1 victim cache 328 enables the L2 controller 320 to accurately determine when to send snoop transactions to the L1 controller 310. For example, if the shadow tags were not maintained in the L2 cache subsystem 306, then the L2 controller 320 would need to snoop the L1 cache subsystem 304 for each request that could possibly be held in the L1 main or victim caches 314, 316, which could reduce performance due to the resulting snoop traffic bandwidth. In this example, the shadow L1 victim cache 328 includes one set (Set 0) and ways 0 through X, along with floating entries, which render the shadow L1 victim cache 328 to reflect more entries than can be stored in the L1 victim cache 316. The floating entries are explained in further detail below.
In both the shadow L1 main cache 326 and the shadow L1 victim cache 328, only the tag (e.g., address) and coherence state information is shadowed. That is, in at least this example, it is not necessary to shadow the cached data itself.
When the L2 controller 320 receives a snoop transaction or a read or write transaction occurs from the L3 controller 310 to the L2 controller 320, the L2 controller 320 first checks the shadow L1 main and shadow L1 victim caches 326, 328. If a match is found (e.g., a hit), then the L2 controller 320 initiates a snoop transaction to the L1 controller 310. When the snoop transaction returns, the L2 controller 320 uses the snoop response to update the shadow L1 main and shadow L1 victim caches 326, 328, if necessary.
Similarly, when the L1 controller 310 allocates a line in its L1 main cache 314, or moves or relocates a line from the L1 main cache 314 to the L1 victim cache 316, the L1 controller 310 communicates such movement to the L2 controller 320 to enable the L2 controller 320 to update the shadow L1 main and shadow L1 victim caches 326, 328. When the L1 controller 310 evicts a line from either the L1 main cache 314 or the L1 victim cache 316, the line is either modified (e.g., dirty) or unmodified (e.g., clean) with respect to main memory (e.g., DDR 110). The L1 controller 310 is configured to communicate both clean line evictions and dirty line victims to the L2 controller 320, which enables the L2 controller 320 to accurately update its shadow L1 main and shadow L1 victim caches 326, 328. The signaling protocol to communicate such movement, relocation, and evictions between the L1 controller 310 and the L2 controller 320 is discussed in further detail below.
In an example, the L2 controller 320 learns that the L1 controller 310 is kicking a line out of its L1 victim cache 316 (e.g., to make room for a line coming from the L1 main cache 314) before the L2 controller 320 receives the displaced victim from the L1 victim cache 316. The line kicked out of the L1 victim cache 316 is held in a victim buffer 702 (e.g., as shown in
The floating entries in the shadow L1 victim cache 328 address this issue. These floating entries extend the size of the shadow L1 victim cache 328 to include at least the number of victim buffers in the L1 cache subsystem 304. In one example, the floating entries result in the shadow L1 victim cache 328 having twice the number of entries as the L1 victim cache 316. In an example, the exact location of entries in the L1 victim cache 316 does not need to match the location of the same cache line as it is shadowed in the shadow L1 victim cache 328. Decoupling the locations between the L1 victim cache 316 and the shadow L1 victim cache 328 improves the safety of the protocol, as a full address comparison is performed when the L2 controller 320 looks for an entry in the L1 victim cache 316. Subsequently, when the L2 controller 320 receives the displaced victim across the interface from the victim buffer, the L2 controller 320 causes the line to be removed from its shadow L1 victim cache 328.
When the L1 controller 310 decides to allocate line C, the L1 controller 310 conveys this allocation to the L2 controller (e.g., as part of a read request issued by the L1 controller 310). In this example, the address of line C maps to the same location in the L1 main cache 314 as the line A, and thus the L1 controller 310 relocates line A to the L1 victim cache 316, in a location occupied by the line B. As a result of the line B being modified, the L1 controller 310 determines to send line B to the L2 cache subsystem 306 as a victim and moves the line B to the victim buffer 702. After the read allocate for the line C, the L1 main cache 314 contains the line C in the location that formerly held the line A, the L1 victim cache 316 contains the cache line A that was relocated from the L1 main cache 314, and the victim buffer 702 contains the cache line B that was evicted from the L1 victim cache 316.
Similarly, after the read allocate for the line C (e.g., communicated by the L1 controller 310 to the L2 controller 320 as part of the read request for the line C), the shadow L1 main cache 326 contains the line C in the location that formerly held the line A and the shadow L1 victim cache 328 contains the relocated line A in one of its floating entries, while the line B also remains in the shadow L1 victim cache 328. As explained above, there is a period of time in which the L2 controller 320 is aware that the L1 controller is moving the line A from the L1 main cache 314 to the L1 victim cache 316, but the L2 controller 320 has not yet received the line B as a victim (e.g., the line B is still in the victim buffer 702). The floating entries of the shadow L1 victim cache 328 provide an additional storage buffer, and the L2 controller 320 is configured to remove the line B from the shadow L1 victim cache 328 when the line B is received as a victim on the interface between the L2 cache subsystem 306 and the L1 cache subsystem 304.
In general and as explained above, the L2 controller 320 is configured to receive an indication from the L1 controller 310 that a cache line is being relocated from the L1 main cache 314 to the L1 victim cache 316 (e.g., the cache line A in the example of
These examples, in particular the floating entries of the shadow L1 victim cache 328, enable cleaner handoff of a victim line from the L1 cache subsystem 304 to the L2 cache subsystem 306 by removing the timing window where a line is removed from the L1 victim cache 316, but has not yet been received by the L2 cache subsystem 306 as a victim Additionally, the L2 controller 320 maintaining accurate shadows of the L1 main cache 314 and the L1 victim cache 316 allows the L2 controller to only generate snoop transactions when necessary (e.g., when the L2 controller 320 is aware that a line is held in one of the L1 caches 314, 316).
As explained above, the L1 controller 310 communicates movement of cache lines between its L1 main cache 314, L1 victim cache 316, and victim buffer 702 to the L2 controller 320. In some examples, this communication occurs in conjunction with a response to a snoop request from the L2 cache subsystem 306 to the L1 cache subsystem 304. In other examples, this communication occurs in conjunction with a read request from the L1 cache subsystem 304 to the L2 cache subsystem 306.
Referring back to
In particular, the L1 controller 310 is configured, in some examples, to send sideband signals in conjunction with a functional read transaction to the L2 controller 320. The sideband signals contain information related to cache line movement (e.g., as described above with respect to the example of
Examples of this disclosure reduce bandwidth on the transaction bus by avoiding the need for multiple messages to communicate both the functional read transaction and movements of cache lines within the L1 cache subsystem 304 that will result from that read transaction. Further, examples of this disclosure reduce timing dependencies and implementation complexity by avoiding the use of a separate asynchronous interface to communicate cache line movement information.
In the table 900, the main_valid and victim_valid signals indicate whether the other main* and victim* signals, respectively, are valid. For example, the L1 controller 310 is configured to de-assert the valid signals when transmitted in parallel with a transaction that does not result in cache line movement(s) in the L1 main cache 314 and the L1 victim cache, respectively. The main_mesi and victim_mesi signals indicate the cache coherence state (e.g., MESI state) for a cache line moving from the L1 main cache 314 to the L1 victim cache 316 and for a cache line moving out of the L1 victim cache 316, respectively. The main_secure and victim_secure signals indicate whether the cache line moving from the L1 main cache 314 to the L1 victim cache 316 and the cache line moving out of the L1 victim cache 316, respectively, is secure. The main_address and victim_address signals indicate the addresses for the cache line moving from the L1 main cache 314 to the L1 victim cache 316 and for the cache line moving out of the L1 victim cache 316, respectively.
The L2 controller 320 is thus configured, in this example, to receive, in a single transaction, a read request in parallel with the aforementioned sideband signals that detail the cache line movement(s) that will occur in the L1 cache subsystem 304 as a result of the read request. In order for the L1 controller 310 to allocate space for data returned in response to the read request, the sideband signals indicate an address and coherence state of the cache line moving from the L1 main cache 314 to the L1 victim cache 316 and for the cache line moving out of the L1 victim cache 316.
The L2 controller 320 is configured to update the shadow L1 main cache 326 to reflect that the cache line moving from the L1 main cache 314 to the L1 victim cache 316 is no longer present in the L1 main cache 314. Similarly, the L2 controller 320 is configured to update the shadow L1 victim cache 328 to reflect that the cache line moving from the L1 main cache 314 to the L1 victim cache 316 is now present in the L1 victim cache 316. If one or more of the valid bits in the sideband signals 900 are de-asserted, the L2 controller 320 is configured not to update its shadow L1 main cache 326 (main_valid de-asserted) or its shadow L1 victim cache 328 (victim_valid de-asserted).
In some examples, the L2 controller 320 is also configured to update the shadow L1 victim cache 328 to reflect that a cache line is no longer located in the L1 victim cache 316. In particular, if the victim_mesi signal indicates that the cache line moving out of the L1 victim cache 316 has a coherence state other than modified (e.g., exclusive or shared), then the L2 controller 320 does not expect to receive a corresponding victim transaction because it is not necessary to write back a cache line that is not dirty. On the other hand, if the victim_mesi signal indicates that the cache line moving out of the L1 victim cache 316 has a modified coherence state, then the L2 controller 320 waits to receive a victim transaction (e.g., from the victim buffer 702). Upon receiving the victim transaction, the L2 controller 320 is configured to update the shadow L1 victim cache 328 to reflect that a cache line is no longer located in the L1 victim cache 316.
The foregoing examples reduce bandwidth on the transaction bus between the L1 cache subsystem 304 and the L2 cache subsystem 306 by avoiding the need for multiple messages to communicate both the functional read transaction and the movements of cache lines within the L1 cache subsystem 304 that will result from that read transaction.
The sideband signaling protocol discussed above leverages unused bandwidth on a transaction bus to facilitate communication of both the functional read transaction and the movements of cache lines within the L1 cache subsystem 304 that will result from that read transaction. However, in certain cases, the L1 controller 310 makes changes to the L1 main cache 314 and/or L1 victim cache 316 that are not coupled to a transaction that would be communicated to the L2 controller 320. In these cases, the L2 controller 320 needs to be made aware of the changes to L1 main cache 314 and/or L1 victim cache 316 in another way.
In particular, for accurate coherent behavior, the L2 controller 320 maintains an accurate directory of the lines held in the L1 main cache 314 and L1 victim cache 316 (e.g., as shadow copies). This enables the L2 controller 320 to send snoop transactions to the L1 controller 320 to get the most up to date copy of the data when the L2 controller 320 knows the line is cached in the L1 cache subsystem 304.
When the L1 controller 310 determines it must evict a non-modified line from the L1 victim cache 316 (e.g., for various reasons dependent on workload), the L1 controller 310 is configured in an example to inform the L2 controller 320 that the line is no longer held in the L1 cache subsystem 304. In an example, the L1 controller 310 does not inform the L2 controller 320 that the line is no longer held in the L1 cache subsystem 304. If the L1 controller 310 did not notify the L2 controller 320 that the line is no longer present, the L2 controller 320 may send at least one more snoop transaction to the address at a later time, believing that the line is still held in the L1 cache subsystem 304. When the line is not found, the L1 controller will return a snoop response indicating that the line was not present. This concept is described as a snoop miss, and results in unnecessary delays when the line was evicted knowingly by the L1 controller.
Examples of this disclosure address the foregoing by utilizing a tag update bus to employ a single cycle, pulsed protocol that enables the L1 controller 310 to communicate with the L2 controller 320 outside of the transaction-based sideband signaling protocol explained above. The tag update bus is separate from the transaction bus described above. Similarly to the transaction bus, the tag update bus is represented schematically by the coupling between the L1 cache subsystem 304 and the L2 cache subsystem 306 (or similar couplings between L1 and L2 structures in
By communicating the invalidations to the L2 controller 320, unnecessary snoop transactions can be avoided, resulting in shorter latencies for processing transactions in the L2 cache subsystem 306. Additionally, power savings may be realized by reducing the number of RAM accesses required by multiple arbitrations for the command that resulted in a snoop miss.
In certain examples, t0_address and t1_address comprises an address in either the L1 main cache 314 or the L1 victim cache 316, while in other examples the tag update bus is used solely to communicate updates to coherence state information for cache lines in the L1 victim cache 316. In some examples, t0_mesi and t1_mesi could indicate any final cache coherence state for the cache line identified by t0_address and t1_address. The tag update bus provides the L1 controller 310 a means to communicate the cache line invalidations that result from the L1 controller 310, while avoiding the snoop miss scenario described above.
The L2 controller 320 is thus configured to receive, over the tag update bus separate from a transaction bus, a message from the L1 controller 310 that includes a valid signal (e.g., t0_req), an address (e.g., t0_address), and a cache coherence state (e.g., t0_mesi). The message thus details an update to cache line coherence state(s) that will occur in the L1 cache subsystem 304.
The L2 controller, in response to the valid signal being asserted, is configured to identify an entry in the shadow L1 main cache 326 or the shadow L1 victim cache 328 having an address corresponding to the address of the message and update a coherence state of the identified entry to be the coherence state of the message. In one example, the L2 controller 320 is configured only to identify an entry in the shadow L1 victim cache 328 having the address of the message. Concurrently, the L2 controller 320 may receive transactions over the transaction bus from the L1 controller 310. These transactions are separate from the message received over the tag update bus.
In some examples, the L2 cache subsystem 306 includes a transaction first-in, first-out buffer (FIFO, not shown for simplicity) coupled to the transaction bus that stores transactions received from the L1 cache subsystem 304 pending processing of those transactions by the L2 controller 320. Messages received by the L2 controller 320 over the tag update bus are not stored in the transaction FIFO, and are instead processed by the L2 controller 320 upon receipt of an asserted valid signal (e.g., t0_req).
In accordance with some examples of this disclosure, the L2 controller 320 exists in a system-wide MESI cache coherence implementation as described above. However, the L2 controller 320 is configured to remap certain transactions from certain masters to implement a local MEI protocol between the L2 controller 320 and the L1 controller 310 or the L3 controller 309. In certain circumstances, remapping from MESI to MEI by the L2 controller 320 enables higher performance on read/write software workloads where memory locations are frequently read before being written to. For example, in a multi-core coherence solution, multiple caches can hold a cache line in the shared state. When one cache needs to modify the line, it must first initiate messaging to a downstream (e.g., numerically higher) controller which results in each of the other caches receiving an invalidating snoop to remove their copy. Finally, once snoop responses have been received, the initiating cache updates the cache coherence state of the line from shared to exclusive. The initiating cache then performs its cache line write and transitions the cache line to the modified state. Thus, when a line is actively shared between multiple different caches, and modified frequently, the number of coherence messages (read, write, victim, snoop) that are required can become large, negatively impacting the performance of software executing on the CPU core 302. Suppression of the shared state by the L2 controller 320 causes each cache line allocation to bring the line in the exclusive state, reducing the need for future coherent messaging when a modification of the cache line occurs
In accordance with examples of this disclosure, the L2 controller 320 suppresses the shared state by applying opcode mapping logic 1106 to the transaction 1104. As will be explained further below, the opcode mapping logic 1106 maps a transaction opcode to a subset of opcodes for the final coherent cache state of the cache line comprising the modified, exclusive, or invalid states. In particular, opcodes that would have resulted in a final cache line coherence state of shared are remapped to one of this subset of opcodes. The opcode mapping logic 1106 need not map opcodes that would have resulted in a final cache state of modified, exclusive, or invalid.
The first request, or transaction 1104, is thus mapped to a second request demonstrated by block 1108, which avoids invoking the shared coherence state. The second request is then arbitrated as normal by L2 arbitration logic 1110 and enters the L2 instruction pipeline 1112, to be subsequently processed by the L2 controller 320.
In some examples, the L2 cache subsystem 306 includes a configuration register shown as block 1107, which includes a shared field. The shared field allows the L2 cache subsystem 306 to be programmatically configured to either suppress the shared state, or not to suppress the shared state (e.g., not perform opcode mapping and function as a part of the larger MESI-based coherence system, described above). For example, if the shared field in configuration register 1107 is asserted, then the opcode mapping logic 1106 is not employed to map transaction opcodes to that would have resulted in a final cache line coherence state of shared. Thus, if a transaction 1104 is received as a third request when the shared field is asserted, the third request is processed by the L2 controller 320 without having its opcode mapped by the opcode mapping logic 1106.
The method 1200 continues in block 1204 with the L2 controller 320 mapping the first request to a second request for a cache line in an exclusive cache coherence state, as explained above. For example, opcode mapping logic 1106 is applied to the opcode of the first request (e.g., invoking the shared coherence state) to map to the opcode of the second request (e.g., invoking the exclusive coherence state). As explained, the opcode mapping logic 1106 is carried out prior to the request entering the L2 arbitration logic 1110 and the L2 instruction pipeline 1112, and thus being processed by the L2 controller 320. In an example, read requests are either allocating or non-allocating, and either non-coherent or coherent. In this example, the opcode mapping logic 1106 maps non-coherent allocating reads to a read command without snoop, because no snooping is required for a non-coherent read and reading without snooping does not invoke the shared state. Similarly, the opcode mapping logic 1106 also maps non-coherent non-allocating reads to a read command without snoop. The opcode mapping logic 1106 maps coherent allocating reads to a read exclusive command, which guarantees that the line will be allocated in the exclusive state. The opcode mapping logic 1106 maps coherent non-allocating reads to a read once command, since these only need to sample the coherent data (e.g., not allocate), and thus the current owner can keep the line without invoking the shared state. In another example, certain snoop commands (e.g., from the L3 controller 309) or CMOs have an opcode that would normally require a line to transition to the shared state. In this example, the opcode mapping logic 1106 maps such snoop commands and CMOs to a snoop command or CMO, respectively, that requires the line to instead transition to the invalid state. Additionally, if the L2 controller 320 determines to send a snoop command to the L1 controller 310, the opcode mapping logic 1106 maps such a snoop command to a snoop command that requires the L1 controller 310 to instead transition the line to the invalid state.
The method 1200 then continues in block 1206 with the L2 controller 320 responding to the second request, if the second request is of a type that warrants a response (e.g., if the second request is a read response, a read response is warranted). In the event that the requested cache line is stored in the L2 cache subsystem 306, as part of its response, the L2 controller 320 transitions a coherence state of the cache line to invalid rather than shared. Alternately, the method 1200 continues in block 1208 with forwarding the second request. For example, where the first request results in a snoop being issued by the L2 controller 320, the L2 controller forwards the second request for the cache line in the exclusive state, rather than the shared state.
In some examples, the L1 controller 310 determines to change a size of the L1 main cache 314. For example, the L1 main cache 314 may be an allocated region of the larger L1 SRAM 312 that can grow (e.g., from 32 KB to 64 KB) or shrink (e.g., from 32 KB to 16 KB) in size over time, depending on requirements communicated to the L1 controller 310, for example from the CPU core 302 or software executing thereon. The L2 controller 320 needs to be aware of changes in size to the L1 main cache 314, so that the L2 controller 320 can properly maintain (e.g., change the size of) its shadow L1 main cache 326.
The following protocol enables the L2 controller 320, in an example, to change the size of its shadow L1 main cache 326 while avoiding data corruption and/or transaction deadlocks (e.g., where a first transaction relies on a second transaction, which is pending resolution of the first transaction). In one example, sideband signals of the transaction bus (explained above) are used by the L1 controller 310 to communicate the size change of the L1 main cache 314 to the L2 controller 320. In this example, reference is made to certain ones of the sideband signals of the transaction bus, in particular referred to as: global_on, global_coh_type, and cache_size. The global_on signal indicates that the L1 controller 310 is performing a global operation on its L1 main cache 314. The global_coh_type signal indicates the type of global coherence operation being performed on the L1 main cache 314. In the specific example of a size change of the L1 main cache 314, the global_coh_type will be a writeback invalidate operation. During a cache size change, coherence is maintained by writing the data to the endpoint and by invalidating the cacheline. The cache_size signal indicates the size to which the L1 main cache 314 is transitioning.
The method 1300 continues in block 1304 with the L1 controller 310 servicing pending read and write requests from a CPU core, such as the CPU core 302. The method 1300 then continues in block 1306 in which the L1 controller 310 stalls new read and write requests from the CPU core 302. This allows the L1 controller 310 to work through pending requests but restrict new requests so that it may perform the global operation (e.g., writeback invalidate and cache size change) on the L1 main cache 314.
The method 1300 continues in block 1308 with the L1 controller 310 writing back and invalidating the L1 main cache 314. At this point in the method 1300, the L1 controller 310 asserts the global_on signal to indicate it is performing a global operation, and the global_coh_type indicates a writeback invalidate as explained above. The L1 controller 310 is configured to send victims to the L2 controller 320 during this stage, which enables the L2 controller 320 to update the shadow L1 main and victim caches 326, 328. If the victim hits in the L2 cache 324, the L2 controller 320 is also configured to update that cache line with the victim data. If the victim is not present in L2 cache 324, the L2 controller 320 forwards the victim (e.g., to the L3 controller 309). During the size change of the L1 main cache 314, coherence is maintained writing the data back to the endpoint and invalidating the cache line. While the L1 controller 310 writes back and invalidates the L1 main cache 314, the L1 controller 310 is also configured to accept and stall a snoop request from the L2 controller 320.
While the L1 controller 310 asserts the global_on signal (e.g., during a global operation), the L1 controller 310 also de-asserts a ready signal, which indicates to the CPU core 302 not to send the L1 controller 310 additional requests for a cache size change or other global coherence operations. The ready signal remains de-asserted until the global operation is completed (e.g., the global_on signal is de-asserted).
Once the global_on signal is de-asserted, the L1 controller 310 responds to any pending snoop transactions that were received from the L2 controller 320 and stalled by the L1 controller 310 during the writeback invalidate (e.g., the global coherence operation for L1 main cache 314 size change). In an example, the L1 controller 310 responds to the pending snoop transactions with a response indicating a cache miss because the L1 main cache 314 is invalidated as part of the size change protocol. Once the global_on signal is de-asserted, the L1 controller 310 also begins accepting read and write requests from the CPU core 302 using the new cache size for the L1 main cache 314. At this point the L1 controller 310 has implemented the functionality to change the size of its L1 main cache 314.
The method 1300 then continues to block 1310 in which the L2 controller 320 receives an indication that the L1 main cache 314 has been invalidated and had its size changed. In an example, the L1 controller 310 sends such an indication to the L2 controller in response to the L1 controller 310 having received write responses for all victims written back by the L1 controller 310, while no further victims are pending to be written back by the L1 controller 310.
In this example, the L1 controller 310 uses sideband signals of global_on, global_coh_type, and cache_size to communicate that the L1 main cache 314 has been invalidated and had its size changed. For example, when global_coh_type indicates a writeback invalidate and the cache_size signal has changed, the L1 controller 310 de-asserting global_on indicates to the L2 controller 320 that the L1 main cache 314 has been invalidated and had its size changed. This indication allows the L2 controller 320 to begin the process of resizing its shadow L1 main cache 326. To begin resizing the shadow L1 main cache 326, the L2 controller 320 flushes its pipeline, or completes all transactions that are present in its pipeline while stalling transactions from other masters.
In some examples, the L2 controller 320 flushes its pipeline in separate phases, which include a blocking soft stall phase, a non-blocking soft stall phase, and a hard stall phase. In general, blocking transactions include read requests and write requests that are not victims, which have the potential to create a secondary transaction (e.g., a snoop), while non-blocking transactions include victims, snoops, and all responses.
In an example, during the blocking soft stall phase, the L2 controller 320 stalls all blocking transactions, such as fetches, read requests, and write requests from the CPU core 302 and DMA read/write accesses (e.g., from another CPU core) but allows response transactions, non-blocking snoop and victim transactions to be accepted and arbitrated. In some examples the L2 controller 320 flushes its pipeline over several cycles. Following the blocking soft stall phase, the L2 controller 320 enters the non-blocking soft stall phase, in which the L2 controller 320 allows response transactions and victims but stalls snoop transactions, in addition to the blocking transactions already stalled in the blocking soft stall phase. As a result, the L2 controller 320 does not initiate new snoops to the L1 main cache 314 for lines previously cached in in the L1 main cache 314.
After the L2 controller 320 pipeline is flushed, the method 1300 continues to block 1312 in which the L2 controller 320 stalls requests received from any master. This phase is the hard stall phase referred to above. In particular, the L2 controller 320 pipeline is flushed, the L2 controller 320 enforce a hard stall where all transactions, including response transactions, are stalled from all masters.
In some examples, the L2 controller 320 also de-asserts or causes the ready signal (explained above with respect to the L1 controller) to be de-asserted. By de-asserting the ready signal, the L2 controller 320 prevents the CPU core 302 from sending requests for a cache size change or other global coherence operation until the L2 controller 320 has completed the currently-pending request (e.g., a cache size change). In one example, the ready signal provided to the CPU core 302 comprises a logical AND of a ready signal from the L1 controller 310 and the L2 controller 320. That is, the CPU core 302 only receives an asserted ready signal when both the L1 controller 310 and the L2 controller 320 assert their ready signals (e.g., when the cache size change operation is complete).
When the hard stall is enforced in block 1312, the method 1300 then continues to block 1314 in which the L2 controller 320 reinitializes the shadow L1 main cache 326 to clear its previous contents (e.g., invalidate cache lines previously held in the shadow L1 main cache 326) and change a size of the shadow L1 main cache 326. In some examples, reinitializing the shadow L1 main cache 326 takes several cycles, during which the L2 controller 320 continues to enforce the hard stall on other masters. Once the shadow L1 main cache 326 is reinitialized, the L2 controller 320 unstalls the masters and asserts its ready signal. The L2 controller 320 then begins to process pending transactions from one or more holding buffers, and accepts new transactions. At this point the size change protocol execution is complete. In some cases, the L1 controller 310 sends a transaction (e.g., a read request) to the L2 controller 320 while the L2 controller 320 is flushing its pipeline in block 1310 or stalled in block 1312, and thus the transaction from the L1 controller 310 is stalled as well. The L2 controller 320 responds to such transactions after reinitializing the shadow L1 main cache 326.
In the foregoing discussion and in the claims, the terms “including” and “comprising” are used in an open-ended fashion, and thus should be interpreted to mean “including, but not limited to . . . .” Also, the term “couple” or “couples” is intended to mean either an indirect or direct connection. Thus, if a first device couples to a second device, that connection may be through a direct connection or through an indirect connection via other devices and connections. Similarly, a device that is coupled between a first component or location and a second component or location may be through a direct connection or through an indirect connection via other devices and connections. An element or feature that is “configured to” perform a task or function may be configured (e.g., programmed or structurally designed) at a time of manufacturing by a manufacturer to perform the function and/or may be configurable (or re-configurable) by a user after manufacturing to perform the function and/or other additional or alternative functions. The configuring may be through firmware and/or software programming of the device, through a construction and/or layout of hardware components and interconnections of the device, or a combination thereof. Additionally, uses of the phrases “ground” or similar in the foregoing discussion are intended to include a chassis ground, an Earth ground, a floating ground, a virtual ground, a digital ground, a common ground, and/or any other form of ground connection applicable to, or suitable for, the teachings of the present disclosure. Unless otherwise stated, “about,” “approximately,” or “substantially” preceding a value means +/−10 percent of the stated value.
The above discussion is meant to be illustrative of the principles and various embodiments of the present disclosure. Numerous variations and modifications will become apparent to those skilled in the art once the above disclosure is fully appreciated. It is intended that the following claims be interpreted to embrace all such variations and modifications.
The present application is a continuation of U.S. patent application Ser. No. 17/666,196, filed Feb. 7, 2022, which is a continuation of U.S. patent application Ser. No. 16/882,257, filed May 22, 2020, now U.S. Pat. No. 11,243,883, which claims priority to U.S. Provisional Patent Application No. 62/852,416, filed May 24, 2019, each of which is incorporated by reference herein in its entirety.
| Number | Name | Date | Kind |
|---|---|---|---|
| 5860114 | Sell | Jan 1999 | A |
| 5926830 | Feiste | Jul 1999 | A |
| 5940858 | Green | Aug 1999 | A |
| 6148416 | Masubuchi | Nov 2000 | A |
| 8266383 | Minkin | Sep 2012 | B1 |
| 9720833 | Drapala | Aug 2017 | B2 |
| 20020147889 | Kruckemyer | Oct 2002 | A1 |
| 20040117561 | Quach | Jun 2004 | A1 |
| 20110113196 | Hooker | May 2011 | A1 |
| 20170123987 | Cheng | May 2017 | A1 |
| 20180293693 | Ray | Oct 2018 | A1 |
| Number | Date | Country |
|---|---|---|
| 101681423 | Nov 2016 | KR |
| Entry |
|---|
| International Search Report for PCT/US2020/034471 mailed Aug. 20, 2020. |
| Machine Translation for KR101681423B1. |
| Number | Date | Country | |
|---|---|---|---|
| 20230297506 A1 | Sep 2023 | US |
| Number | Date | Country | |
|---|---|---|---|
| 62852416 | May 2019 | US |
| Number | Date | Country | |
|---|---|---|---|
| Parent | 17666196 | Feb 2022 | US |
| Child | 18325190 | US | |
| Parent | 16882257 | May 2020 | US |
| Child | 17666196 | US |
