Chris Metcalf | e5a0693 | 2010-11-01 17:00:37 -0400 | [diff] [blame] | 1 | /* |
| 2 | * Copyright 2010 Tilera Corporation. All Rights Reserved. |
| 3 | * |
| 4 | * This program is free software; you can redistribute it and/or |
| 5 | * modify it under the terms of the GNU General Public License |
| 6 | * as published by the Free Software Foundation, version 2. |
| 7 | * |
| 8 | * This program is distributed in the hope that it will be useful, but |
| 9 | * WITHOUT ANY WARRANTY; without even the implied warranty of |
| 10 | * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or |
| 11 | * NON INFRINGEMENT. See the GNU General Public License for |
| 12 | * more details. |
| 13 | */ |
| 14 | |
| 15 | /** |
| 16 | * @file drv_xgbe_intf.h |
| 17 | * Interface to the hypervisor XGBE driver. |
| 18 | */ |
| 19 | |
| 20 | #ifndef __DRV_XGBE_INTF_H__ |
| 21 | #define __DRV_XGBE_INTF_H__ |
| 22 | |
| 23 | /** |
| 24 | * An object for forwarding VAs and PAs to the hypervisor. |
| 25 | * @ingroup types |
| 26 | * |
| 27 | * This allows the supervisor to specify a number of areas of memory to |
| 28 | * store packet buffers. |
| 29 | */ |
| 30 | typedef struct |
| 31 | { |
| 32 | /** The physical address of the memory. */ |
| 33 | HV_PhysAddr pa; |
| 34 | /** Page table entry for the memory. This is only used to derive the |
| 35 | * memory's caching mode; the PA bits are ignored. */ |
| 36 | HV_PTE pte; |
| 37 | /** The virtual address of the memory. */ |
| 38 | HV_VirtAddr va; |
| 39 | /** Size (in bytes) of the memory area. */ |
| 40 | int size; |
| 41 | |
| 42 | } |
| 43 | netio_ipp_address_t; |
| 44 | |
| 45 | /** The various pread/pwrite offsets into the hypervisor-level driver. |
| 46 | * @ingroup types |
| 47 | */ |
| 48 | typedef enum |
| 49 | { |
| 50 | /** Inform the Linux driver of the address of the NetIO arena memory. |
| 51 | * This offset is actually only used to convey information from netio |
| 52 | * to the Linux driver; it never makes it from there to the hypervisor. |
| 53 | * Write-only; takes a uint32_t specifying the VA address. */ |
| 54 | NETIO_FIXED_ADDR = 0x5000000000000000ULL, |
| 55 | |
| 56 | /** Inform the Linux driver of the size of the NetIO arena memory. |
| 57 | * This offset is actually only used to convey information from netio |
| 58 | * to the Linux driver; it never makes it from there to the hypervisor. |
| 59 | * Write-only; takes a uint32_t specifying the VA size. */ |
| 60 | NETIO_FIXED_SIZE = 0x5100000000000000ULL, |
| 61 | |
| 62 | /** Register current tile with IPP. Write then read: write, takes a |
| 63 | * netio_input_config_t, read returns a pointer to a netio_queue_impl_t. */ |
| 64 | NETIO_IPP_INPUT_REGISTER_OFF = 0x6000000000000000ULL, |
| 65 | |
| 66 | /** Unregister current tile from IPP. Write-only, takes a dummy argument. */ |
| 67 | NETIO_IPP_INPUT_UNREGISTER_OFF = 0x6100000000000000ULL, |
| 68 | |
| 69 | /** Start packets flowing. Write-only, takes a dummy argument. */ |
| 70 | NETIO_IPP_INPUT_INIT_OFF = 0x6200000000000000ULL, |
| 71 | |
| 72 | /** Stop packets flowing. Write-only, takes a dummy argument. */ |
| 73 | NETIO_IPP_INPUT_UNINIT_OFF = 0x6300000000000000ULL, |
| 74 | |
| 75 | /** Configure group (typically we group on VLAN). Write-only: takes an |
| 76 | * array of netio_group_t's, low 24 bits of the offset is the base group |
| 77 | * number times the size of a netio_group_t. */ |
| 78 | NETIO_IPP_INPUT_GROUP_CFG_OFF = 0x6400000000000000ULL, |
| 79 | |
| 80 | /** Configure bucket. Write-only: takes an array of netio_bucket_t's, low |
| 81 | * 24 bits of the offset is the base bucket number times the size of a |
| 82 | * netio_bucket_t. */ |
| 83 | NETIO_IPP_INPUT_BUCKET_CFG_OFF = 0x6500000000000000ULL, |
| 84 | |
| 85 | /** Get/set a parameter. Read or write: read or write data is the parameter |
| 86 | * value, low 32 bits of the offset is a __netio_getset_offset_t. */ |
| 87 | NETIO_IPP_PARAM_OFF = 0x6600000000000000ULL, |
| 88 | |
| 89 | /** Get fast I/O index. Read-only; returns a 4-byte base index value. */ |
| 90 | NETIO_IPP_GET_FASTIO_OFF = 0x6700000000000000ULL, |
| 91 | |
| 92 | /** Configure hijack IP address. Packets with this IPv4 dest address |
| 93 | * go to bucket NETIO_NUM_BUCKETS - 1. Write-only: takes an IP address |
| 94 | * in some standard form. FIXME: Define the form! */ |
| 95 | NETIO_IPP_INPUT_HIJACK_CFG_OFF = 0x6800000000000000ULL, |
| 96 | |
| 97 | /** |
| 98 | * Offsets beyond this point are reserved for the supervisor (although that |
| 99 | * enforcement must be done by the supervisor driver itself). |
| 100 | */ |
| 101 | NETIO_IPP_USER_MAX_OFF = 0x6FFFFFFFFFFFFFFFULL, |
| 102 | |
| 103 | /** Register I/O memory. Write-only, takes a netio_ipp_address_t. */ |
| 104 | NETIO_IPP_IOMEM_REGISTER_OFF = 0x7000000000000000ULL, |
| 105 | |
| 106 | /** Unregister I/O memory. Write-only, takes a netio_ipp_address_t. */ |
| 107 | NETIO_IPP_IOMEM_UNREGISTER_OFF = 0x7100000000000000ULL, |
| 108 | |
| 109 | /* Offsets greater than 0x7FFFFFFF can't be used directly from Linux |
| 110 | * userspace code due to limitations in the pread/pwrite syscalls. */ |
| 111 | |
| 112 | /** Drain LIPP buffers. */ |
| 113 | NETIO_IPP_DRAIN_OFF = 0xFA00000000000000ULL, |
| 114 | |
| 115 | /** Supply a netio_ipp_address_t to be used as shared memory for the |
| 116 | * LEPP command queue. */ |
| 117 | NETIO_EPP_SHM_OFF = 0xFB00000000000000ULL, |
| 118 | |
| 119 | /* 0xFC... is currently unused. */ |
| 120 | |
| 121 | /** Stop IPP/EPP tiles. Write-only, takes a dummy argument. */ |
| 122 | NETIO_IPP_STOP_SHIM_OFF = 0xFD00000000000000ULL, |
| 123 | |
| 124 | /** Start IPP/EPP tiles. Write-only, takes a dummy argument. */ |
| 125 | NETIO_IPP_START_SHIM_OFF = 0xFE00000000000000ULL, |
| 126 | |
| 127 | /** Supply packet arena. Write-only, takes an array of |
| 128 | * netio_ipp_address_t values. */ |
| 129 | NETIO_IPP_ADDRESS_OFF = 0xFF00000000000000ULL, |
| 130 | } netio_hv_offset_t; |
| 131 | |
| 132 | /** Extract the base offset from an offset */ |
| 133 | #define NETIO_BASE_OFFSET(off) ((off) & 0xFF00000000000000ULL) |
| 134 | /** Extract the local offset from an offset */ |
| 135 | #define NETIO_LOCAL_OFFSET(off) ((off) & 0x00FFFFFFFFFFFFFFULL) |
| 136 | |
| 137 | |
| 138 | /** |
| 139 | * Get/set offset. |
| 140 | */ |
| 141 | typedef union |
| 142 | { |
| 143 | struct |
| 144 | { |
| 145 | uint64_t addr:48; /**< Class-specific address */ |
| 146 | unsigned int class:8; /**< Class (e.g., NETIO_PARAM) */ |
| 147 | unsigned int opcode:8; /**< High 8 bits of NETIO_IPP_PARAM_OFF */ |
| 148 | } |
| 149 | bits; /**< Bitfields */ |
| 150 | uint64_t word; /**< Aggregated value to use as the offset */ |
| 151 | } |
| 152 | __netio_getset_offset_t; |
| 153 | |
| 154 | /** |
| 155 | * Fast I/O index offsets (must be contiguous). |
| 156 | */ |
| 157 | typedef enum |
| 158 | { |
| 159 | NETIO_FASTIO_ALLOCATE = 0, /**< Get empty packet buffer */ |
| 160 | NETIO_FASTIO_FREE_BUFFER = 1, /**< Give buffer back to IPP */ |
| 161 | NETIO_FASTIO_RETURN_CREDITS = 2, /**< Give credits to IPP */ |
| 162 | NETIO_FASTIO_SEND_PKT_NOCK = 3, /**< Send a packet, no checksum */ |
| 163 | NETIO_FASTIO_SEND_PKT_CK = 4, /**< Send a packet, with checksum */ |
| 164 | NETIO_FASTIO_SEND_PKT_VEC = 5, /**< Send a vector of packets */ |
| 165 | NETIO_FASTIO_SENDV_PKT = 6, /**< Sendv one packet */ |
| 166 | NETIO_FASTIO_NUM_INDEX = 7, /**< Total number of fast I/O indices */ |
| 167 | } netio_fastio_index_t; |
| 168 | |
| 169 | /** 3-word return type for Fast I/O call. */ |
| 170 | typedef struct |
| 171 | { |
| 172 | int err; /**< Error code. */ |
| 173 | uint32_t val0; /**< Value. Meaning depends upon the specific call. */ |
| 174 | uint32_t val1; /**< Value. Meaning depends upon the specific call. */ |
| 175 | } netio_fastio_rv3_t; |
| 176 | |
| 177 | /** 0-argument fast I/O call */ |
| 178 | int __netio_fastio0(uint32_t fastio_index); |
| 179 | /** 1-argument fast I/O call */ |
| 180 | int __netio_fastio1(uint32_t fastio_index, uint32_t arg0); |
| 181 | /** 3-argument fast I/O call, 2-word return value */ |
| 182 | netio_fastio_rv3_t __netio_fastio3_rv3(uint32_t fastio_index, uint32_t arg0, |
| 183 | uint32_t arg1, uint32_t arg2); |
| 184 | /** 4-argument fast I/O call */ |
| 185 | int __netio_fastio4(uint32_t fastio_index, uint32_t arg0, uint32_t arg1, |
| 186 | uint32_t arg2, uint32_t arg3); |
| 187 | /** 6-argument fast I/O call */ |
| 188 | int __netio_fastio6(uint32_t fastio_index, uint32_t arg0, uint32_t arg1, |
| 189 | uint32_t arg2, uint32_t arg3, uint32_t arg4, uint32_t arg5); |
| 190 | /** 9-argument fast I/O call */ |
| 191 | int __netio_fastio9(uint32_t fastio_index, uint32_t arg0, uint32_t arg1, |
| 192 | uint32_t arg2, uint32_t arg3, uint32_t arg4, uint32_t arg5, |
| 193 | uint32_t arg6, uint32_t arg7, uint32_t arg8); |
| 194 | |
| 195 | /** Allocate an empty packet. |
| 196 | * @param fastio_index Fast I/O index. |
| 197 | * @param size Size of the packet to allocate. |
| 198 | */ |
| 199 | #define __netio_fastio_allocate(fastio_index, size) \ |
| 200 | __netio_fastio1((fastio_index) + NETIO_FASTIO_ALLOCATE, size) |
| 201 | |
| 202 | /** Free a buffer. |
| 203 | * @param fastio_index Fast I/O index. |
| 204 | * @param handle Handle for the packet to free. |
| 205 | */ |
| 206 | #define __netio_fastio_free_buffer(fastio_index, handle) \ |
| 207 | __netio_fastio1((fastio_index) + NETIO_FASTIO_FREE_BUFFER, handle) |
| 208 | |
| 209 | /** Increment our receive credits. |
| 210 | * @param fastio_index Fast I/O index. |
| 211 | * @param credits Number of credits to add. |
| 212 | */ |
| 213 | #define __netio_fastio_return_credits(fastio_index, credits) \ |
| 214 | __netio_fastio1((fastio_index) + NETIO_FASTIO_RETURN_CREDITS, credits) |
| 215 | |
| 216 | /** Send packet, no checksum. |
| 217 | * @param fastio_index Fast I/O index. |
| 218 | * @param ackflag Nonzero if we want an ack. |
| 219 | * @param size Size of the packet. |
| 220 | * @param va Virtual address of start of packet. |
| 221 | * @param handle Packet handle. |
| 222 | */ |
| 223 | #define __netio_fastio_send_pkt_nock(fastio_index, ackflag, size, va, handle) \ |
| 224 | __netio_fastio4((fastio_index) + NETIO_FASTIO_SEND_PKT_NOCK, ackflag, \ |
| 225 | size, va, handle) |
| 226 | |
| 227 | /** Send packet, calculate checksum. |
| 228 | * @param fastio_index Fast I/O index. |
| 229 | * @param ackflag Nonzero if we want an ack. |
| 230 | * @param size Size of the packet. |
| 231 | * @param va Virtual address of start of packet. |
| 232 | * @param handle Packet handle. |
| 233 | * @param csum0 Shim checksum header. |
| 234 | * @param csum1 Checksum seed. |
| 235 | */ |
| 236 | #define __netio_fastio_send_pkt_ck(fastio_index, ackflag, size, va, handle, \ |
| 237 | csum0, csum1) \ |
| 238 | __netio_fastio6((fastio_index) + NETIO_FASTIO_SEND_PKT_CK, ackflag, \ |
| 239 | size, va, handle, csum0, csum1) |
| 240 | |
| 241 | |
| 242 | /** Format for the "csum0" argument to the __netio_fastio_send routines |
| 243 | * and LEPP. Note that this is currently exactly identical to the |
| 244 | * ShimProtocolOffloadHeader. |
| 245 | */ |
| 246 | typedef union |
| 247 | { |
| 248 | struct |
| 249 | { |
| 250 | unsigned int start_byte:7; /**< The first byte to be checksummed */ |
| 251 | unsigned int count:14; /**< Number of bytes to be checksummed. */ |
| 252 | unsigned int destination_byte:7; /**< The byte to write the checksum to. */ |
| 253 | unsigned int reserved:4; /**< Reserved. */ |
| 254 | } bits; /**< Decomposed method of access. */ |
| 255 | unsigned int word; /**< To send out the IDN. */ |
| 256 | } __netio_checksum_header_t; |
| 257 | |
| 258 | |
| 259 | /** Sendv packet with 1 or 2 segments. |
| 260 | * @param fastio_index Fast I/O index. |
| 261 | * @param flags Ack/csum/notify flags in low 3 bits; number of segments minus |
| 262 | * 1 in next 2 bits; expected checksum in high 16 bits. |
| 263 | * @param confno Confirmation number to request, if notify flag set. |
| 264 | * @param csum0 Checksum descriptor; if zero, no checksum. |
| 265 | * @param va_F Virtual address of first segment. |
| 266 | * @param va_L Virtual address of last segment, if 2 segments. |
| 267 | * @param len_F_L Length of first segment in low 16 bits; length of last |
| 268 | * segment, if 2 segments, in high 16 bits. |
| 269 | */ |
| 270 | #define __netio_fastio_sendv_pkt_1_2(fastio_index, flags, confno, csum0, \ |
| 271 | va_F, va_L, len_F_L) \ |
| 272 | __netio_fastio6((fastio_index) + NETIO_FASTIO_SENDV_PKT, flags, confno, \ |
| 273 | csum0, va_F, va_L, len_F_L) |
| 274 | |
| 275 | /** Send packet on PCIe interface. |
| 276 | * @param fastio_index Fast I/O index. |
| 277 | * @param flags Ack/csum/notify flags in low 3 bits. |
| 278 | * @param confno Confirmation number to request, if notify flag set. |
| 279 | * @param csum0 Checksum descriptor; Hard wired 0, not needed for PCIe. |
| 280 | * @param va_F Virtual address of the packet buffer. |
| 281 | * @param va_L Virtual address of last segment, if 2 segments. Hard wired 0. |
| 282 | * @param len_F_L Length of the packet buffer in low 16 bits. |
| 283 | */ |
| 284 | #define __netio_fastio_send_pcie_pkt(fastio_index, flags, confno, csum0, \ |
| 285 | va_F, va_L, len_F_L) \ |
| 286 | __netio_fastio6((fastio_index) + PCIE_FASTIO_SENDV_PKT, flags, confno, \ |
| 287 | csum0, va_F, va_L, len_F_L) |
| 288 | |
| 289 | /** Sendv packet with 3 or 4 segments. |
| 290 | * @param fastio_index Fast I/O index. |
| 291 | * @param flags Ack/csum/notify flags in low 3 bits; number of segments minus |
| 292 | * 1 in next 2 bits; expected checksum in high 16 bits. |
| 293 | * @param confno Confirmation number to request, if notify flag set. |
| 294 | * @param csum0 Checksum descriptor; if zero, no checksum. |
| 295 | * @param va_F Virtual address of first segment. |
| 296 | * @param va_L Virtual address of last segment (third segment if 3 segments, |
| 297 | * fourth segment if 4 segments). |
| 298 | * @param len_F_L Length of first segment in low 16 bits; length of last |
| 299 | * segment in high 16 bits. |
| 300 | * @param va_M0 Virtual address of "middle 0" segment; this segment is sent |
| 301 | * second when there are three segments, and third if there are four. |
| 302 | * @param va_M1 Virtual address of "middle 1" segment; this segment is sent |
| 303 | * second when there are four segments. |
| 304 | * @param len_M0_M1 Length of middle 0 segment in low 16 bits; length of middle |
| 305 | * 1 segment, if 4 segments, in high 16 bits. |
| 306 | */ |
| 307 | #define __netio_fastio_sendv_pkt_3_4(fastio_index, flags, confno, csum0, va_F, \ |
| 308 | va_L, len_F_L, va_M0, va_M1, len_M0_M1) \ |
| 309 | __netio_fastio9((fastio_index) + NETIO_FASTIO_SENDV_PKT, flags, confno, \ |
| 310 | csum0, va_F, va_L, len_F_L, va_M0, va_M1, len_M0_M1) |
| 311 | |
| 312 | /** Send vector of packets. |
| 313 | * @param fastio_index Fast I/O index. |
| 314 | * @param seqno Number of packets transmitted so far on this interface; |
| 315 | * used to decide which packets should be acknowledged. |
| 316 | * @param nentries Number of entries in vector. |
| 317 | * @param va Virtual address of start of vector entry array. |
| 318 | * @return 3-word netio_fastio_rv3_t structure. The structure's err member |
| 319 | * is an error code, or zero if no error. The val0 member is the |
| 320 | * updated value of seqno; it has been incremented by 1 for each |
| 321 | * packet sent. That increment may be less than nentries if an |
Lucas De Marchi | 25985ed | 2011-03-30 22:57:33 -0300 | [diff] [blame] | 322 | * error occurred, or if some of the entries in the vector contain |
Chris Metcalf | e5a0693 | 2010-11-01 17:00:37 -0400 | [diff] [blame] | 323 | * handles equal to NETIO_PKT_HANDLE_NONE. The val1 member is the |
| 324 | * updated value of nentries; it has been decremented by 1 for each |
| 325 | * vector entry processed. Again, that decrement may be less than |
| 326 | * nentries (leaving the returned value positive) if an error |
| 327 | * occurred. |
| 328 | */ |
| 329 | #define __netio_fastio_send_pkt_vec(fastio_index, seqno, nentries, va) \ |
| 330 | __netio_fastio3_rv3((fastio_index) + NETIO_FASTIO_SEND_PKT_VEC, seqno, \ |
| 331 | nentries, va) |
| 332 | |
| 333 | |
| 334 | /** An egress DMA command for LEPP. */ |
| 335 | typedef struct |
| 336 | { |
| 337 | /** Is this a TSO transfer? |
| 338 | * |
| 339 | * NOTE: This field is always 0, to distinguish it from |
| 340 | * lepp_tso_cmd_t. It must come first! |
| 341 | */ |
| 342 | uint8_t tso : 1; |
| 343 | |
| 344 | /** Unused padding bits. */ |
| 345 | uint8_t _unused : 3; |
| 346 | |
| 347 | /** Should this packet be sent directly from caches instead of DRAM, |
| 348 | * using hash-for-home to locate the packet data? |
| 349 | */ |
| 350 | uint8_t hash_for_home : 1; |
| 351 | |
| 352 | /** Should we compute a checksum? */ |
| 353 | uint8_t compute_checksum : 1; |
| 354 | |
| 355 | /** Is this the final buffer for this packet? |
| 356 | * |
| 357 | * A single packet can be split over several input buffers (a "gather" |
| 358 | * operation). This flag indicates that this is the last buffer |
| 359 | * in a packet. |
| 360 | */ |
| 361 | uint8_t end_of_packet : 1; |
| 362 | |
| 363 | /** Should LEPP advance 'comp_busy' when this DMA is fully finished? */ |
| 364 | uint8_t send_completion : 1; |
| 365 | |
| 366 | /** High bits of Client Physical Address of the start of the buffer |
| 367 | * to be egressed. |
| 368 | * |
| 369 | * NOTE: Only 6 bits are actually needed here, as CPAs are |
| 370 | * currently 38 bits. So two bits could be scavenged from this. |
| 371 | */ |
| 372 | uint8_t cpa_hi; |
| 373 | |
| 374 | /** The number of bytes to be egressed. */ |
| 375 | uint16_t length; |
| 376 | |
| 377 | /** Low 32 bits of Client Physical Address of the start of the buffer |
| 378 | * to be egressed. |
| 379 | */ |
| 380 | uint32_t cpa_lo; |
| 381 | |
| 382 | /** Checksum information (only used if 'compute_checksum'). */ |
| 383 | __netio_checksum_header_t checksum_data; |
| 384 | |
| 385 | } lepp_cmd_t; |
| 386 | |
| 387 | |
| 388 | /** A chunk of physical memory for a TSO egress. */ |
| 389 | typedef struct |
| 390 | { |
| 391 | /** The low bits of the CPA. */ |
| 392 | uint32_t cpa_lo; |
| 393 | /** The high bits of the CPA. */ |
| 394 | uint16_t cpa_hi : 15; |
| 395 | /** Should this packet be sent directly from caches instead of DRAM, |
| 396 | * using hash-for-home to locate the packet data? |
| 397 | */ |
| 398 | uint16_t hash_for_home : 1; |
| 399 | /** The length in bytes. */ |
| 400 | uint16_t length; |
| 401 | } lepp_frag_t; |
| 402 | |
| 403 | |
| 404 | /** An LEPP command that handles TSO. */ |
| 405 | typedef struct |
| 406 | { |
| 407 | /** Is this a TSO transfer? |
| 408 | * |
| 409 | * NOTE: This field is always 1, to distinguish it from |
| 410 | * lepp_cmd_t. It must come first! |
| 411 | */ |
| 412 | uint8_t tso : 1; |
| 413 | |
| 414 | /** Unused padding bits. */ |
| 415 | uint8_t _unused : 7; |
| 416 | |
| 417 | /** Size of the header[] array in bytes. It must be in the range |
| 418 | * [40, 127], which are the smallest header for a TCP packet over |
| 419 | * Ethernet and the maximum possible prepend size supported by |
| 420 | * hardware, respectively. Note that the array storage must be |
| 421 | * padded out to a multiple of four bytes so that the following |
| 422 | * LEPP command is aligned properly. |
| 423 | */ |
| 424 | uint8_t header_size; |
| 425 | |
| 426 | /** Byte offset of the IP header in header[]. */ |
| 427 | uint8_t ip_offset; |
| 428 | |
| 429 | /** Byte offset of the TCP header in header[]. */ |
| 430 | uint8_t tcp_offset; |
| 431 | |
| 432 | /** The number of bytes to use for the payload of each packet, |
| 433 | * except of course the last one, which may not have enough bytes. |
| 434 | * This means that each Ethernet packet except the last will have a |
| 435 | * size of header_size + payload_size. |
| 436 | */ |
| 437 | uint16_t payload_size; |
| 438 | |
| 439 | /** The length of the 'frags' array that follows this struct. */ |
| 440 | uint16_t num_frags; |
| 441 | |
| 442 | /** The actual frags. */ |
| 443 | lepp_frag_t frags[0 /* Variable-sized; num_frags entries. */]; |
| 444 | |
| 445 | /* |
| 446 | * The packet header template logically follows frags[], |
| 447 | * but you can't declare that in C. |
| 448 | * |
| 449 | * uint32_t header[header_size_in_words_rounded_up]; |
| 450 | */ |
| 451 | |
| 452 | } lepp_tso_cmd_t; |
| 453 | |
| 454 | |
| 455 | /** An LEPP completion ring entry. */ |
| 456 | typedef void* lepp_comp_t; |
| 457 | |
| 458 | |
| 459 | /** Maximum number of frags for one TSO command. This is adapted from |
| 460 | * linux's "MAX_SKB_FRAGS", and presumably over-estimates by one, for |
| 461 | * our page size of exactly 65536. We add one for a "body" fragment. |
| 462 | */ |
| 463 | #define LEPP_MAX_FRAGS (65536 / HV_PAGE_SIZE_SMALL + 2 + 1) |
| 464 | |
| 465 | /** Total number of bytes needed for an lepp_tso_cmd_t. */ |
| 466 | #define LEPP_TSO_CMD_SIZE(num_frags, header_size) \ |
| 467 | (sizeof(lepp_tso_cmd_t) + \ |
| 468 | (num_frags) * sizeof(lepp_frag_t) + \ |
| 469 | (((header_size) + 3) & -4)) |
| 470 | |
| 471 | /** The size of the lepp "cmd" queue. */ |
| 472 | #define LEPP_CMD_QUEUE_BYTES \ |
| 473 | (((CHIP_L2_CACHE_SIZE() - 2 * CHIP_L2_LINE_SIZE()) / \ |
| 474 | (sizeof(lepp_cmd_t) + sizeof(lepp_comp_t))) * sizeof(lepp_cmd_t)) |
| 475 | |
| 476 | /** The largest possible command that can go in lepp_queue_t::cmds[]. */ |
| 477 | #define LEPP_MAX_CMD_SIZE LEPP_TSO_CMD_SIZE(LEPP_MAX_FRAGS, 128) |
| 478 | |
| 479 | /** The largest possible value of lepp_queue_t::cmd_{head, tail} (inclusive). |
| 480 | */ |
| 481 | #define LEPP_CMD_LIMIT \ |
| 482 | (LEPP_CMD_QUEUE_BYTES - LEPP_MAX_CMD_SIZE) |
| 483 | |
| 484 | /** The maximum number of completions in an LEPP queue. */ |
| 485 | #define LEPP_COMP_QUEUE_SIZE \ |
| 486 | ((LEPP_CMD_LIMIT + sizeof(lepp_cmd_t) - 1) / sizeof(lepp_cmd_t)) |
| 487 | |
| 488 | /** Increment an index modulo the queue size. */ |
| 489 | #define LEPP_QINC(var) \ |
| 490 | (var = __insn_mnz(var - (LEPP_COMP_QUEUE_SIZE - 1), var + 1)) |
| 491 | |
| 492 | /** A queue used to convey egress commands from the client to LEPP. */ |
| 493 | typedef struct |
| 494 | { |
| 495 | /** Index of first completion not yet processed by user code. |
| 496 | * If this is equal to comp_busy, there are no such completions. |
| 497 | * |
| 498 | * NOTE: This is only read/written by the user. |
| 499 | */ |
| 500 | unsigned int comp_head; |
| 501 | |
| 502 | /** Index of first completion record not yet completed. |
| 503 | * If this is equal to comp_tail, there are no such completions. |
| 504 | * This index gets advanced (modulo LEPP_QUEUE_SIZE) whenever |
| 505 | * a command with the 'completion' bit set is finished. |
| 506 | * |
| 507 | * NOTE: This is only written by LEPP, only read by the user. |
| 508 | */ |
| 509 | volatile unsigned int comp_busy; |
| 510 | |
| 511 | /** Index of the first empty slot in the completion ring. |
| 512 | * Entries from this up to but not including comp_head (in ring order) |
| 513 | * can be filled in with completion data. |
| 514 | * |
| 515 | * NOTE: This is only read/written by the user. |
| 516 | */ |
| 517 | unsigned int comp_tail; |
| 518 | |
| 519 | /** Byte index of first command enqueued for LEPP but not yet processed. |
| 520 | * |
| 521 | * This is always divisible by sizeof(void*) and always <= LEPP_CMD_LIMIT. |
| 522 | * |
| 523 | * NOTE: LEPP advances this counter as soon as it no longer needs |
| 524 | * the cmds[] storage for this entry, but the transfer is not actually |
| 525 | * complete (i.e. the buffer pointed to by the command is no longer |
| 526 | * needed) until comp_busy advances. |
| 527 | * |
| 528 | * If this is equal to cmd_tail, the ring is empty. |
| 529 | * |
| 530 | * NOTE: This is only written by LEPP, only read by the user. |
| 531 | */ |
| 532 | volatile unsigned int cmd_head; |
| 533 | |
| 534 | /** Byte index of first empty slot in the command ring. This field can |
| 535 | * be incremented up to but not equal to cmd_head (because that would |
| 536 | * mean the ring is empty). |
| 537 | * |
| 538 | * This is always divisible by sizeof(void*) and always <= LEPP_CMD_LIMIT. |
| 539 | * |
| 540 | * NOTE: This is read/written by the user, only read by LEPP. |
| 541 | */ |
| 542 | volatile unsigned int cmd_tail; |
| 543 | |
| 544 | /** A ring of variable-sized egress DMA commands. |
| 545 | * |
| 546 | * NOTE: Only written by the user, only read by LEPP. |
| 547 | */ |
| 548 | char cmds[LEPP_CMD_QUEUE_BYTES] |
| 549 | __attribute__((aligned(CHIP_L2_LINE_SIZE()))); |
| 550 | |
| 551 | /** A ring of user completion data. |
| 552 | * NOTE: Only read/written by the user. |
| 553 | */ |
| 554 | lepp_comp_t comps[LEPP_COMP_QUEUE_SIZE] |
| 555 | __attribute__((aligned(CHIP_L2_LINE_SIZE()))); |
| 556 | } lepp_queue_t; |
| 557 | |
| 558 | |
| 559 | /** An internal helper function for determining the number of entries |
| 560 | * available in a ring buffer, given that there is one sentinel. |
| 561 | */ |
| 562 | static inline unsigned int |
| 563 | _lepp_num_free_slots(unsigned int head, unsigned int tail) |
| 564 | { |
| 565 | /* |
| 566 | * One entry is reserved for use as a sentinel, to distinguish |
| 567 | * "empty" from "full". So we compute |
| 568 | * (head - tail - 1) % LEPP_QUEUE_SIZE, but without using a slow % operation. |
| 569 | */ |
| 570 | return (head - tail - 1) + ((head <= tail) ? LEPP_COMP_QUEUE_SIZE : 0); |
| 571 | } |
| 572 | |
| 573 | |
| 574 | /** Returns how many new comp entries can be enqueued. */ |
| 575 | static inline unsigned int |
| 576 | lepp_num_free_comp_slots(const lepp_queue_t* q) |
| 577 | { |
| 578 | return _lepp_num_free_slots(q->comp_head, q->comp_tail); |
| 579 | } |
| 580 | |
| 581 | static inline int |
| 582 | lepp_qsub(int v1, int v2) |
| 583 | { |
| 584 | int delta = v1 - v2; |
| 585 | return delta + ((delta >> 31) & LEPP_COMP_QUEUE_SIZE); |
| 586 | } |
| 587 | |
| 588 | |
| 589 | /** FIXME: Check this from linux, via a new "pwrite()" call. */ |
| 590 | #define LIPP_VERSION 1 |
| 591 | |
| 592 | |
| 593 | /** We use exactly two bytes of alignment padding. */ |
| 594 | #define LIPP_PACKET_PADDING 2 |
| 595 | |
| 596 | /** The minimum size of a "small" buffer (including the padding). */ |
| 597 | #define LIPP_SMALL_PACKET_SIZE 128 |
| 598 | |
| 599 | /* |
| 600 | * NOTE: The following two values should total to less than around |
| 601 | * 13582, to keep the total size used for "lipp_state_t" below 64K. |
| 602 | */ |
| 603 | |
| 604 | /** The maximum number of "small" buffers. |
| 605 | * This is enough for 53 network cpus with 128 credits. Note that |
| 606 | * if these are exhausted, we will fall back to using large buffers. |
| 607 | */ |
| 608 | #define LIPP_SMALL_BUFFERS 6785 |
| 609 | |
| 610 | /** The maximum number of "large" buffers. |
| 611 | * This is enough for 53 network cpus with 128 credits. |
| 612 | */ |
| 613 | #define LIPP_LARGE_BUFFERS 6785 |
| 614 | |
| 615 | #endif /* __DRV_XGBE_INTF_H__ */ |