1 .. SPDX-License-Identifier: GPL-2.0
4 ===================================================
5 Virtual Accelerator Switchboard (VAS) userspace API
6 ===================================================
11 Power9 processor introduced Virtual Accelerator Switchboard (VAS) which
12 allows both userspace and kernel communicate to co-processor
13 (hardware accelerator) referred to as the Nest Accelerator (NX). The NX
14 unit comprises of one or more hardware engines or co-processor types
15 such as 842 compression, GZIP compression and encryption. On power9,
16 userspace applications will have access to only GZIP Compression engine
17 which supports ZLIB and GZIP compression algorithms in the hardware.
19 To communicate with NX, kernel has to establish a channel or window and
20 then requests can be submitted directly without kernel involvement.
21 Requests to the GZIP engine must be formatted as a co-processor Request
22 Block (CRB) and these CRBs must be submitted to the NX using COPY/PASTE
23 instructions to paste the CRB to hardware address that is associated with
24 the engine's request queue.
26 The GZIP engine provides two priority levels of requests: Normal and
27 High. Only Normal requests are supported from userspace right now.
29 This document explains userspace API that is used to interact with
30 kernel to setup channel / window which can be used to send compression
31 requests directly to NX accelerator.
37 Application access to the GZIP engine is provided through
38 /dev/crypto/nx-gzip device node implemented by the VAS/NX device driver.
39 An application must open the /dev/crypto/nx-gzip device to obtain a file
40 descriptor (fd). Then should issue VAS_TX_WIN_OPEN ioctl with this fd to
41 establish connection to the engine. It means send window is opened on GZIP
42 engine for this process. Once a connection is established, the application
43 should use the mmap() system call to map the hardware address of engine's
44 request queue into the application's virtual address space.
46 The application can then submit one or more requests to the the engine by
47 using copy/paste instructions and pasting the CRBs to the virtual address
48 (aka paste_address) returned by mmap(). User space can close the
49 established connection or send window by closing the file descriptior
50 (close(fd)) or upon the process exit.
52 Note that applications can send several requests with the same window or
53 can establish multiple windows, but one window for each file descriptor.
55 Following sections provide additional details and references about the
61 There is one /dev/crypto/nx-gzip node in the system and it provides
62 access to all GZIP engines in the system. The only valid operations on
63 /dev/crypto/nx-gzip are:
65 * open() the device for read and write.
66 * issue VAS_TX_WIN_OPEN ioctl
67 * mmap() the engine's request queue into application's virtual
68 address space (i.e. get a paste_address for the co-processor
70 * close the device node.
72 Other file operations on this device node are undefined.
74 Note that the copy and paste operations go directly to the hardware and
75 do not go through this device. Refer COPY/PASTE document for more
78 Although a system may have several instances of the NX co-processor
79 engines (typically, one per P9 chip) there is just one
80 /dev/crypto/nx-gzip device node in the system. When the nx-gzip device
81 node is opened, Kernel opens send window on a suitable instance of NX
82 accelerator. It finds CPU on which the user process is executing and
83 determine the NX instance for the corresponding chip on which this CPU
86 Applications may chose a specific instance of the NX co-processor using
87 the vas_id field in the VAS_TX_WIN_OPEN ioctl as detailed below.
89 A userspace library libnxz is available here but still in development:
90 https://github.com/abalib/power-gzip
92 Applications that use inflate / deflate calls can link with libnxz
93 instead of libz and use NX GZIP compression without any modification.
95 Open /dev/crypto/nx-gzip
96 ========================
98 The nx-gzip device should be opened for read and write. No special
99 privileges are needed to open the device. Each window corresponds to one
100 file descriptor. So if the userspace process needs multiple windows,
101 several open calls have to be issued.
103 See open(2) system call man pages for other details such as return values,
104 error codes and restrictions.
106 VAS_TX_WIN_OPEN ioctl
107 =====================
109 Applications should use the VAS_TX_WIN_OPEN ioctl as follows to establish
110 a connection with NX co-processor engine:
113 struct vas_tx_win_open_attr {
115 __s16 vas_id; /* specific instance of vas or -1
118 __u64 flags; /* For future use */
122 version: The version field must be currently set to 1.
123 vas_id: If '-1' is passed, kernel will make a best-effort attempt
124 to assign an optimal instance of NX for the process. To
125 select the specific VAS instance, refer
126 "Discovery of available VAS engines" section below.
128 flags, reserved1 and reserved2[6] fields are for future extension
129 and must be set to 0.
131 The attributes attr for the VAS_TX_WIN_OPEN ioctl are defined as
133 #define VAS_MAGIC 'v'
134 #define VAS_TX_WIN_OPEN _IOW(VAS_MAGIC, 1,
135 struct vas_tx_win_open_attr)
137 struct vas_tx_win_open_attr attr;
138 rc = ioctl(fd, VAS_TX_WIN_OPEN, &attr);
140 The VAS_TX_WIN_OPEN ioctl returns 0 on success. On errors, it
141 returns -1 and sets the errno variable to indicate the error.
144 EINVAL fd does not refer to a valid VAS device.
145 EINVAL Invalid vas ID
146 EINVAL version is not set with proper value
147 EEXIST Window is already opened for the given fd
148 ENOMEM Memory is not available to allocate window
149 ENOSPC System has too many active windows (connections)
151 EINVAL reserved fields are not set to 0.
153 See the ioctl(2) man page for more details, error codes and
156 mmap() NX-GZIP device
157 =====================
159 The mmap() system call for a NX-GZIP device fd returns a paste_address
160 that the application can use to copy/paste its CRB to the hardware engines.
163 paste_addr = mmap(addr, size, prot, flags, fd, offset);
165 Only restrictions on mmap for a NX-GZIP device fd are:
166 * size should be PAGE_SIZE
167 * offset parameter should be 0ULL
169 Refer to mmap(2) man page for additional details/restrictions.
170 In addition to the error conditions listed on the mmap(2) man
171 page, can also fail with one of the following error codes:
173 EINVAL fd is not associated with an open window
174 (i.e mmap() does not follow a successful call
175 to the VAS_TX_WIN_OPEN ioctl).
176 EINVAL offset field is not 0ULL.
178 Discovery of available VAS engines
179 ==================================
181 Each available VAS instance in the system will have a device tree node
182 like /proc/device-tree/vas@* or /proc/device-tree/xscom@*/vas@*.
183 Determine the chip or VAS instance and use the corresponding ibm,vas-id
184 property value in this node to select specific VAS instance.
186 Copy/Paste operations
187 =====================
189 Applications should use the copy and paste instructions to send CRB to NX.
190 Refer section 4.4 in PowerISA for Copy/Paste instructions:
191 https://openpowerfoundation.org/?resource_lib=power-isa-version-3-0
193 CRB Specification and use NX
194 ============================
196 Applications should format requests to the co-processor using the
197 co-processor Request Block (CRBs). Refer NX-GZIP user's manual for the format
198 of CRB and use NX from userspace such as sending requests and checking
204 Applications send requests to NX and wait for the status by polling on
205 co-processor Status Block (CSB) flags. NX updates status in CSB after each
206 request is processed. Refer NX-GZIP user's manual for the format of CSB and
209 In case if NX encounters translation error (called NX page fault) on CSB
210 address or any request buffer, raises an interrupt on the CPU to handle the
211 fault. Page fault can happen if an application passes invalid addresses or
212 request buffers are not in memory. The operating system handles the fault by
213 updating CSB with the following data:
216 csb.cc = CSB_CC_FAULT_ADDRESS;
217 csb.ce = CSB_CE_TERMINATION;
218 csb.address = fault_address;
220 When an application receives translation error, it can touch or access
221 the page that has a fault address so that this page will be in memory. Then
222 the application can resend this request to NX.
224 If the OS can not update CSB due to invalid CSB address, sends SEGV signal
225 to the process who opened the send window on which the original request was
226 issued. This signal returns with the following siginfo struct:
228 siginfo.si_signo = SIGSEGV;
229 siginfo.si_errno = EFAULT;
230 siginfo.si_code = SEGV_MAPERR;
231 siginfo.si_addr = CSB adress;
233 In the case of multi-thread applications, NX send windows can be shared
234 across all threads. For example, a child thread can open a send window,
235 but other threads can send requests to NX using this window. These
236 requests will be successful even in the case of OS handling faults as long
237 as CSB address is valid. If the NX request contains an invalid CSB address,
238 the signal will be sent to the child thread that opened the window. But if
239 the thread is exited without closing the window and the request is issued
240 using this window. the signal will be issued to the thread group leader
241 (tgid). It is up to the application whether to ignore or handle these
244 NX-GZIP User's Manual:
245 https://github.com/libnxz/power-gzip/blob/master/power_nx_gzip_um.pdf
255 struct vas_setup_attr txattr;
257 fd = open("/dev/crypto/nx-gzip", O_RDWR);
259 fprintf(stderr, "open nx-gzip failed\n");
262 memset(&txattr, 0, sizeof(txattr));
265 rc = ioctl(fd, VAS_TX_WIN_OPEN,
266 (unsigned long)&txattr);
268 fprintf(stderr, "ioctl() n %d, error %d\n",
272 addr = mmap(NULL, 4096, PROT_READ|PROT_WRITE,
273 MAP_SHARED, fd, 0ULL);
274 if (addr == MAP_FAILED) {
275 fprintf(stderr, "mmap() failed, errno %d\n",
280 //Format CRB request with compression or
282 // Refer tests for vas_copy/vas_paste
283 vas_copy((&crb, 0, 1);
284 vas_paste(addr, 0, 1);
285 // Poll on csb.flags with timeout
286 // csb address is listed in CRB
288 close(fd) or window can be closed upon process exit
291 Refer https://github.com/abalib/power-gzip for tests or more