Oracle Data Analytics Accelerator (DAX) --------------------------------------- DAX is a coprocessor which resides on the SPARC M7 (DAX1) and M8 (DAX2) processor chips, and has direct access to the CPU's L3 caches as well as physical memory. It can perform several operations on data streams with various input and output formats. A driver provides a transport mechanism and has limited knowledge of the various opcodes and data formats. A user space library provides high level services and translates these into low level commands which are then passed into the driver and subsequently the Hypervisor and the coprocessor. The library is the recommended way for applications to use the coprocessor, and the driver interface is not intended for general use. This document describes the general flow of the driver, its structures, and its programmatic interface. It also provides example code sufficient to write user or kernel applications that use DAX functionality. The user library is open source and available at: https://oss.oracle.com/git/gitweb.cgi?p=libdax.git The Hypervisor interface to the coprocessor is described in detail in the accompanying document, dax-hv-api.txt, which is a plain text excerpt of the (Oracle internal) "UltraSPARC Virtual Machine Specification" version 3.0.20+15, dated 2017-09-25. High Level Overview ------------------- A coprocessor request is described by a Command Control Block (CCB). The CCB contains an opcode and various parameters. The opcode specifies what operation is to be done, and the parameters specify options, flags, sizes, and addresses. The CCB (or an array of CCBs) is passed to the Hypervisor, which handles queueing and scheduling of requests to the available coprocessor execution units. A status code returned indicates if the request was submitted successfully or if there was an error. One of the addresses given in each CCB is a pointer to a "completion area", which is a 128 byte memory block that is written by the coprocessor to provide execution status. No interrupt is generated upon completion; the completion area must be polled by software to find out when a transaction has finished, but the M7 and later processors provide a mechanism to pause the virtual processor until the completion status has been updated by the coprocessor. This is done using the monitored load and mwait instructions, which are described in more detail later. The DAX coprocessor was designed so that after a request is submitted, the kernel is no longer involved in the processing of it. The polling is done at the user level, which results in almost zero latency between completion of a request and resumption of execution of the requesting thread. Addressing Memory ----------------- The kernel does not have access to physical memory in the Sun4v architecture, as there is an additional level of memory virtualization present. This intermediate level is called "real" memory, and the kernel treats this as if it were physical. The Hypervisor handles the translations between real memory and physical so that each logical domain (LDOM) can have a partition of physical memory that is isolated from that of other LDOMs. When the kernel sets up a virtual mapping, it specifies a virtual address and the real address to which it should be mapped. The DAX coprocessor can only operate on physical memory, so before a request can be fed to the coprocessor, all the addresses in a CCB must be converted into physical addresses. The kernel cannot do this since it has no visibility into physical addresses. So a CCB may contain either the virtual or real addresses of the buffers or a combination of them. An "address type" field is available for each address that may be given in the CCB. In all cases, the Hypervisor will translate all the addresses to physical before dispatching to hardware. Address translations are performed using the context of the process initiating the request. The Driver API -------------- An application makes requests to the driver via the write() system call, and gets results (if any) via read(). The completion areas are made accessible via mmap(), and are read-only for the application. The request may either be an immediate command or an array of CCBs to be submitted to the hardware. Each open instance of the device is exclusive to the thread that opened it, and must be used by that thread for all subsequent operations. The driver open function creates a new context for the thread and initializes it for use. This context contains pointers and values used internally by the driver to keep track of submitted requests. The completion area buffer is also allocated, and this is large enough to contain the completion areas for many concurrent requests. When the device is closed, any outstanding transactions are flushed and the context is cleaned up. On a DAX1 system (M7), the device will be called "oradax1", while on a DAX2 system (M8) it will be "oradax2". If an application requires one or the other, it should simply attempt to open the appropriate device. Only one of the devices will exist on any given system, so the name can be used to determine what the platform supports. The immediate commands are CCB_DEQUEUE, CCB_KILL, and CCB_INFO. For all of these, success is indicated by a return value from write() equal to the number of bytes given in the call. Otherwise -1 is returned and errno is set. CCB_DEQUEUE Tells the driver to clean up resources associated with past requests. Since no interrupt is generated upon the completion of a request, the driver must be told when it may reclaim resources. No further status information is returned, so the user should not subsequently call read(). CCB_KILL Kills a CCB during execution. The CCB is guaranteed to not continue executing once this call returns successfully. On success, read() must be called to retrieve the result of the action. CCB_INFO Retrieves information about a currently executing CCB. Note that some Hypervisors might return 'notfound' when the CCB is in 'inprogress' state. To ensure a CCB in the 'notfound' state will never be executed, CCB_KILL must be invoked on that CCB. Upon success, read() must be called to retrieve the details of the action. Submission of an array of CCBs for execution A write() whose length is a multiple of the CCB size is treated as a submit operation. The file offset is treated as the index of the completion area to use, and may be set via lseek() or using the pwrite() system call. If -1 is returned then errno is set to indicate the error. Otherwise, the return value is the length of the array that was actually accepted by the coprocessor. If the accepted length is equal to the requested length, then the submission was completely successful and there is no further status needed; hence, the user should not subsequently call read(). Partial acceptance of the CCB array is indicated by a return value less than the requested length, and read() must be called to retrieve further status information. The status will reflect the error caused by the first CCB that was not accepted, and status_data will provide additional data in some cases. MMAP The mmap() function provides access to the completion area allocated in the driver. Note that the completion area is not writeable by the user process, and the mmap call must not specify PROT_WRITE. Completion of a Request ----------------------- The first byte in each completion area is the command status which is updated by the coprocessor hardware. Software may take advantage of new M7/M8 processor capabilities to efficiently poll this status byte. First, a "monitored load" is achieved via a Load from Alternate Space (ldxa, lduba, etc.) with ASI 0x84 (ASI_MONITOR_PRIMARY). Second, a "monitored wait" is achieved via the mwait instruction (a write to %asr28). This instruction is like pause in that it suspends execution of the virtual processor for the given number of nanoseconds, but in addition will terminate early when one of several events occur. If the block of data containing the monitored location is modified, then the mwait terminates. This causes software to resume execution immediately (without a context switch or kernel to user transition) after a transaction completes. Thus the latency between transaction completion and resumption of execution may be just a few nanoseconds. Application Life Cycle of a DAX Submission ------------------------------------------ - open dax device - call mmap() to get the completion area address - allocate a CCB and fill in the opcode, flags, parameters, addresses, etc. - submit CCB via write() or pwrite() - go into a loop executing monitored load + monitored wait and terminate when the command status indicates the request is complete (CCB_KILL or CCB_INFO may be used any time as necessary) - perform a CCB_DEQUEUE - call munmap() for completion area - close the dax device Memory Constraints ------------------ The DAX hardware operates only on physical addresses. Therefore, it is not aware of virtual memory mappings and the discontiguities that may exist in the physical memory that a virtual buffer maps to. There is no I/O TLB or any scatter/gather mechanism. All buffers, whether input or output, must reside in a physically contiguous region of memory. The Hypervisor translates all addresses within a CCB to physical before handing off the CCB to DAX. The Hypervisor determines the virtual page size for each virtual address given, and uses this to program a size limit for each address. This prevents the coprocessor from reading or writing beyond the bound of the virtual page, even though it is accessing physical memory directly. A simpler way of saying this is that a DAX operation will never "cross" a virtual page boundary. If an 8k virtual page is used, then the data is strictly limited to 8k. If a user's buffer is larger than 8k, then a larger page size must be used, or the transaction size will be truncated to 8k. Huge pages. A user may allocate huge pages using standard interfaces. Memory buffers residing on huge pages may be used to achieve much larger DAX transaction sizes, but the rules must still be followed, and no transaction will cross a page boundary, even a huge page. A major caveat is that Linux on Sparc presents 8Mb as one of the huge page sizes. Sparc does not actually provide a 8Mb hardware page size, and this size is synthesized by pasting together two 4Mb pages. The reasons for this are historical, and it creates an issue because only half of this 8Mb page can actually be used for any given buffer in a DAX request, and it must be either the first half or the second half; it cannot be a 4Mb chunk in the middle, since that crosses a (hardware) page boundary. Note that this entire issue may be hidden by higher level libraries. CCB Structure ------------- A CCB is an array of 8 64-bit words. Several of these words provide command opcodes, parameters, flags, etc., and the rest are addresses for the completion area, output buffer, and various inputs: struct ccb { u64 control; u64 completion; u64 input0; u64 access; u64 input1; u64 op_data; u64 output; u64 table; }; See libdax/common/sys/dax1/dax1_ccb.h for a detailed description of each of these fields, and see dax-hv-api.txt for a complete description of the Hypervisor API available to the guest OS (ie, Linux kernel). The first word (control) is examined by the driver for the following: - CCB version, which must be consistent with hardware version - Opcode, which must be one of the documented allowable commands - Address types, which must be set to "virtual" for all the addresses given by the user, thereby ensuring that the application can only access memory that it owns Example Code ------------ The DAX is accessible to both user and kernel code. The kernel code can make hypercalls directly while the user code must use wrappers provided by the driver. The setup of the CCB is nearly identical for both; the only difference is in preparation of the completion area. An example of user code is given now, with kernel code afterwards. In order to program using the driver API, the file arch/sparc/include/uapi/asm/oradax.h must be included. First, the proper device must be opened. For M7 it will be /dev/oradax1 and for M8 it will be /dev/oradax2. The simplest procedure is to attempt to open both, as only one will succeed: fd = open("/dev/oradax1", O_RDWR); if (fd < 0) fd = open("/dev/oradax2", O_RDWR); if (fd < 0) /* No DAX found */ Next, the completion area must be mapped: completion_area = mmap(NULL, DAX_MMAP_LEN, PROT_READ, MAP_SHARED, fd, 0); All input and output buffers must be fully contained in one hardware page, since as explained above, the DAX is strictly constrained by virtual page boundaries. In addition, the output buffer must be 64-byte aligned and its size must be a multiple of 64 bytes because the coprocessor writes in units of cache lines. This example demonstrates the DAX Scan command, which takes as input a vector and a match value, and produces a bitmap as the output. For each input element that matches the value, the corresponding bit is set in the output. In this example, the input vector consists of a series of single bits, and the match value is 0. So each 0 bit in the input will produce a 1 in the output, and vice versa, which produces an output bitmap which is the input bitmap inverted. For details of all the parameters and bits used in this CCB, please refer to section 36.2.1.3 of the DAX Hypervisor API document, which describes the Scan command in detail. ccb->control = /* Table 36.1, CCB Header Format */ (2L << 48) /* command = Scan Value */ | (3L << 40) /* output address type = primary virtual */ | (3L << 34) /* primary input address type = primary virtual */ /* Section 36.2.1, Query CCB Command Formats */ | (1 << 28) /* 36.2.1.1.1 primary input format = fixed width bit packed */ | (0 << 23) /* 36.2.1.1.2 primary input element size = 0 (1 bit) */ | (8 << 10) /* 36.2.1.1.6 output format = bit vector */ | (0 << 5) /* 36.2.1.3 First scan criteria size = 0 (1 byte) */ | (31 << 0); /* 36.2.1.3 Disable second scan criteria */ ccb->completion = 0; /* Completion area address, to be filled in by driver */ ccb->input0 = (unsigned long) input; /* primary input address */ ccb->access = /* Section 36.2.1.2, Data Access Control */ (2 << 24) /* Primary input length format = bits */ | (nbits - 1); /* number of bits in primary input stream, minus 1 */ ccb->input1 = 0; /* secondary input address, unused */ ccb->op_data = 0; /* scan criteria (value to be matched) */ ccb->output = (unsigned long) output; /* output address */ ccb->table = 0; /* table address, unused */ The CCB submission is a write() or pwrite() system call to the driver. If the call fails, then a read() must be used to retrieve the status: if (pwrite(fd, ccb, 64, 0) != 64) { struct ccb_exec_result status; read(fd, &status, sizeof(status)); /* bail out */ } After a successful submission of the CCB, the completion area may be polled to determine when the DAX is finished. Detailed information on the contents of the completion area can be found in section 36.2.2 of the DAX HV API document. while (1) { /* Monitored Load */ __asm__ __volatile__("lduba [%1] 0x84, %0\n" : "=r" (status) : "r" (completion_area)); if (status) /* 0 indicates command in progress */ break; /* MWAIT */ __asm__ __volatile__("wr %%g0, 1000, %%asr28\n" ::); /* 1000 ns */ } A completion area status of 1 indicates successful completion of the CCB and validity of the output bitmap, which may be used immediately. All other non-zero values indicate error conditions which are described in section 36.2.2. if (completion_area[0] != 1) { /* section 36.2.2, 1 = command ran and succeeded */ /* completion_area[0] contains the completion status */ /* completion_area[1] contains an error code, see 36.2.2 */ } After the completion area has been processed, the driver must be notified that it can release any resources associated with the request. This is done via the dequeue operation: struct dax_command cmd; cmd.command = CCB_DEQUEUE; if (write(fd, &cmd, sizeof(cmd)) != sizeof(cmd)) { /* bail out */ } Finally, normal program cleanup should be done, i.e., unmapping completion area, closing the dax device, freeing memory etc. [Kernel example] The only difference in using the DAX in kernel code is the treatment of the completion area. Unlike user applications which mmap the completion area allocated by the driver, kernel code must allocate its own memory to use for the completion area, and this address and its type must be given in the CCB: ccb->control |= /* Table 36.1, CCB Header Format */ (3L << 32); /* completion area address type = primary virtual */ ccb->completion = (unsigned long) completion_area; /* Completion area address */ The dax submit hypercall is made directly. The flags used in the ccb_submit call are documented in the DAX HV API in section 36.3.1. #include hv_rv = sun4v_ccb_submit((unsigned long)ccb, 64, HV_CCB_QUERY_CMD | HV_CCB_ARG0_PRIVILEGED | HV_CCB_ARG0_TYPE_PRIMARY | HV_CCB_VA_PRIVILEGED, 0, &bytes_accepted, &status_data); if (hv_rv != HV_EOK) { /* hv_rv is an error code, status_data contains */ /* potential additional status, see 36.3.1.1 */ } After the submission, the completion area polling code is identical to that in user land: while (1) { /* Monitored Load */ __asm__ __volatile__("lduba [%1] 0x84, %0\n" : "=r" (status) : "r" (completion_area)); if (status) /* 0 indicates command in progress */ break; /* MWAIT */ __asm__ __volatile__("wr %%g0, 1000, %%asr28\n" ::); /* 1000 ns */ } if (completion_area[0] != 1) { /* section 36.2.2, 1 = command ran and succeeded */ /* completion_area[0] contains the completion status */ /* completion_area[1] contains an error code, see 36.2.2 */ } The output bitmap is ready for consumption immediately after the completion status indicates success.