Index: /trunk/include/VBox/VBoxHDD.h =================================================================== --- /trunk/include/VBox/VBoxHDD.h (revision 30308) +++ /trunk/include/VBox/VBoxHDD.h (revision 30309) @@ -1060,5 +1060,5 @@ /** - * Send data from a socket. + * Send data to a socket. * * @return iprt status code. @@ -1066,15 +1066,32 @@ * @param pvBuffer Buffer to write data to socket. * @param cbBuffer How much to write. - * @param pcbRead Number of bytes read. */ DECLR3CALLBACKMEMBER(int, pfnWrite, (RTSOCKET Sock, const void *pvBuffer, size_t cbBuffer)); /** - * Flush socket write buffers. + * Send data from scatter/gather buffer to a socket. * * @return iprt status code. * @param Sock Socket descriptor. + * @param pSgBuf Scatter/gather buffer to write data to socket. + */ + DECLR3CALLBACKMEMBER(int, pfnSgWrite, (RTSOCKET Sock, PCRTSGBUF pSgBuf)); + + /** + * Flush socket write buffers. + * + * @return iprt status code. + * @param Sock Socket descriptor. */ DECLR3CALLBACKMEMBER(int, pfnFlush, (RTSOCKET Sock)); + + /** + * Enables or disables delaying sends to coalesce packets. + * + * @return iprt status code. + * @param Sock Socket descriptor. + * @param fEnable When set to true enables coalescing. + */ + DECLR3CALLBACKMEMBER(int, pfnSetSendCoalescing, (RTSOCKET Sock, bool fEnable)); /** Index: /trunk/src/VBox/Devices/Makefile.kmk =================================================================== --- /trunk/src/VBox/Devices/Makefile.kmk (revision 30308) +++ /trunk/src/VBox/Devices/Makefile.kmk (revision 30309) @@ -150,9 +150,6 @@ VBoxDD_DEFS += VBOX_WITH_AHCI endif -ifdef VBOX_WITH_ISCSI - VBoxDD_DEFS += VBOX_WITH_ISCSI - ifdef VBOX_WITH_INIP - VBoxDD_DEFS += VBOX_WITH_INIP - endif +ifdef VBOX_WITH_INIP + VBoxDD_DEFS += VBOX_WITH_INIP endif ifdef VBOX_WITH_E1000 @@ -819,4 +816,7 @@ Drivers_DEFS += VBOX_WITH_NETFLT endif +ifdef VBOX_WITH_INIP +Drivers_DEFS += VBOX_WITH_INIP +endif ifdef VBOX_WITH_DRV_DISK_INTEGRITY Drivers_DEFS += VBOX_WITH_DRV_DISK_INTEGRITY Index: /trunk/src/VBox/Devices/Storage/DrvVD.cpp =================================================================== --- /trunk/src/VBox/Devices/Storage/DrvVD.cpp (revision 30308) +++ /trunk/src/VBox/Devices/Storage/DrvVD.cpp (revision 30309) @@ -586,4 +586,12 @@ *******************************************************************************/ +typedef union INIPSOCKADDRUNION +{ + struct sockaddr Addr; + struct sockaddr_in Ipv4; +} INIPSOCKADDRUNION; + +static DECLCALLBACK(int) drvvdINIPFlush(RTSOCKET Sock); + /** @copydoc VDINTERFACETCPNET::pfnClientConnect */ static DECLCALLBACK(int) drvvdINIPClientConnect(const char *pszAddress, uint32_t uPort, PRTSOCKET pSock) @@ -605,5 +613,5 @@ } /* Create socket and connect. */ - RTSOCKET Sock = lwip_socket(PF_INET, SOCK_STREAM, 0); + int Sock = lwip_socket(PF_INET, SOCK_STREAM, 0); if (Sock != -1) { @@ -614,5 +622,5 @@ if (!lwip_connect(Sock, (struct sockaddr *)&InAddr, sizeof(InAddr))) { - *pSock = Sock; + *pSock = (RTSOCKET)Sock; return VINF_SUCCESS; } @@ -628,5 +636,5 @@ static DECLCALLBACK(int) drvvdINIPClientClose(RTSOCKET Sock) { - lwip_close(Sock); + lwip_close((uintptr_t)Sock); return VINF_SUCCESS; /** @todo real solution needed */ } @@ -637,10 +645,10 @@ fd_set fdsetR; FD_ZERO(&fdsetR); - FD_SET(Sock, &fdsetR); + FD_SET((uintptr_t)Sock, &fdsetR); fd_set fdsetE = fdsetR; int rc; if (cMillies == RT_INDEFINITE_WAIT) - rc = lwip_select(Sock + 1, &fdsetR, NULL, &fdsetE, NULL); + rc = lwip_select((uintptr_t)Sock + 1, &fdsetR, NULL, &fdsetE, NULL); else { @@ -648,5 +656,5 @@ timeout.tv_sec = cMillies / 1000; timeout.tv_usec = (cMillies % 1000) * 1000; - rc = lwip_select(Sock + 1, &fdsetR, NULL, &fdsetE, &timeout); + rc = lwip_select((uintptr_t)Sock + 1, &fdsetR, NULL, &fdsetE, &timeout); } if (rc > 0) @@ -678,5 +686,5 @@ * needed it, so I added it here, too). Didn't investigate if this * really has issues. Better be safe than sorry. */ - ssize_t cbBytesRead = lwip_recv(Sock, (char *)pvBuffer + cbRead, + ssize_t cbBytesRead = lwip_recv((uintptr_t)Sock, (char *)pvBuffer + cbRead, RT_MIN(cbToRead, 32768), 0); if (cbBytesRead < 0) @@ -712,5 +720,5 @@ * don't get any wraparounds. This should be moved to DevINIP * stack interface once that's implemented. */ - ssize_t cbWritten = lwip_send(Sock, (void *)pvBuffer, + ssize_t cbWritten = lwip_send((uintptr_t)Sock, (void *)pvBuffer, RT_MIN(cbBuffer, 32768), 0); if (cbWritten < 0) @@ -725,27 +733,52 @@ } +/** @copydoc VDINTERFACETCPNET::pfnSgWrite */ +static DECLCALLBACK(int) drvvdINIPSgWrite(RTSOCKET Sock, PCRTSGBUF pSgBuf) +{ + int rc = VINF_SUCCESS; + + /* This is an extremely crude emulation, however it's good enough + * for our iSCSI code. INIP has no sendmsg(). */ + for (unsigned i = 0; i < pSgBuf->cSeg; i++) + { + rc = drvvdINIPWrite(Sock, pSgBuf->pcaSeg[i].pvSeg, + pSgBuf->pcaSeg[i].cbSeg); + if (RT_FAILURE(rc)) + break; + } + if (RT_SUCCESS(rc)) + drvvdINIPFlush(Sock); + + return rc; +} + /** @copydoc VDINTERFACETCPNET::pfnFlush */ static DECLCALLBACK(int) drvvdINIPFlush(RTSOCKET Sock) { int fFlag = 1; - lwip_setsockopt(Sock, IPPROTO_TCP, TCP_NODELAY, + lwip_setsockopt((uintptr_t)Sock, IPPROTO_TCP, TCP_NODELAY, (const char *)&fFlag, sizeof(fFlag)); fFlag = 0; - lwip_setsockopt(Sock, IPPROTO_TCP, TCP_NODELAY, + lwip_setsockopt((uintptr_t)Sock, IPPROTO_TCP, TCP_NODELAY, (const char *)&fFlag, sizeof(fFlag)); return VINF_SUCCESS; } +/** @copydoc VDINTERFACETCPNET::pfnSetSendCoalescing */ +static DECLCALLBACK(int) drvvdINIPSetSendCoalescing(RTSOCKET Sock, bool fEnable) +{ + int fFlag = fEnable ? 0 : 1; + lwip_setsockopt((uintptr_t)Sock, IPPROTO_TCP, TCP_NODELAY, + (const char *)&fFlag, sizeof(fFlag)); + return VINF_SUCCESS; +} + /** @copydoc VDINTERFACETCPNET::pfnGetLocalAddress */ static DECLCALLBACK(int) drvvdINIPGetLocalAddress(RTSOCKET Sock, PRTNETADDR pAddr) { - union - { - struct sockaddr Addr; - struct sockaddr_in Ipv4; - } u; - socklen_t cbAddr = sizeof(u); + INIPSOCKADDRUNION u; + socklen_t cbAddr = sizeof(u); RT_ZERO(u); - if (!lwip_getsockname(Sock, &u.Addr, &cbAddr)) + if (!lwip_getsockname((uintptr_t)Sock, &u.Addr, &cbAddr)) { /* @@ -770,12 +803,8 @@ static DECLCALLBACK(int) drvvdINIPGetPeerAddress(RTSOCKET Sock, PRTNETADDR pAddr) { - union - { - struct sockaddr Addr; - struct sockaddr_in Ipv4; - } u; - socklen_t cbAddr = sizeof(u); + INIPSOCKADDRUNION u; + socklen_t cbAddr = sizeof(u); RT_ZERO(u); - if (!lwip_getpeername(Sock, &u.Addr, &cbAddr)) + if (!lwip_getpeername((uintptr_t)Sock, &u.Addr, &cbAddr)) { /* @@ -1386,5 +1415,7 @@ pThis->VDITcpNetCallbacks.pfnRead = RTTcpRead; pThis->VDITcpNetCallbacks.pfnWrite = RTTcpWrite; + pThis->VDITcpNetCallbacks.pfnSgWrite = RTTcpSgWrite; pThis->VDITcpNetCallbacks.pfnFlush = RTTcpFlush; + pThis->VDITcpNetCallbacks.pfnSetSendCoalescing = RTTcpSetSendCoalescing; pThis->VDITcpNetCallbacks.pfnGetLocalAddress = RTTcpGetLocalAddress; pThis->VDITcpNetCallbacks.pfnGetPeerAddress = RTTcpGetPeerAddress; @@ -1403,5 +1434,7 @@ pThis->VDITcpNetCallbacks.pfnRead = drvvdINIPRead; pThis->VDITcpNetCallbacks.pfnWrite = drvvdINIPWrite; + pThis->VDITcpNetCallbacks.pfnSgWrite = drvvdINIPSgWrite; pThis->VDITcpNetCallbacks.pfnFlush = drvvdINIPFlush; + pThis->VDITcpNetCallbacks.pfnSetSendCoalescing = drvvdINIPSetSendCoalescing; pThis->VDITcpNetCallbacks.pfnGetLocalAddress = drvvdINIPGetLocalAddress; pThis->VDITcpNetCallbacks.pfnGetPeerAddress = drvvdINIPGetPeerAddress; Index: /trunk/src/VBox/Devices/Storage/ISCSIHDDCore.cpp =================================================================== --- /trunk/src/VBox/Devices/Storage/ISCSIHDDCore.cpp (revision 30308) +++ /trunk/src/VBox/Devices/Storage/ISCSIHDDCore.cpp (revision 30309) @@ -218,4 +218,7 @@ /** ISCSI BHS word 0: response includes status. */ #define ISCSI_STATUS_BIT 0x00010000 + +/** Maximum number of scatter/gather segments needed to send a PDU. */ +#define ISCSI_SG_SEGMENTS_MAX 4 @@ -548,4 +551,7 @@ return rc; } + + /* Disable Nagle algorithm, we want things to be sent immediately. */ + pImage->pInterfaceNetCallbacks->pfnSetSendCoalescing(pImage->Socket, false); /* Make initiator name and ISID unique on this host. */ @@ -714,25 +720,37 @@ if (RT_SUCCESS(rc)) { + /* Construct scatter/gather buffer for entire request, worst case + * needs twice as many entries to allow for padding. */ + unsigned cBuf = 0; for (i = 0; i < cnRequest; i++) { - /* Write one chunk of data. */ - rc = pImage->pInterfaceNetCallbacks->pfnWrite(pImage->Socket, - paRequest[i].pcvSeg, - paRequest[i].cbSeg); - if (RT_FAILURE(rc)) - break; - /* Insert proper padding before the next chunk us written. */ + cBuf++; + if (paRequest[i].cbSeg & 3) + cBuf++; + } + Assert(cBuf < ISCSI_SG_SEGMENTS_MAX); + RTSGBUF buf; + RTSGSEG aSeg[ISCSI_SG_SEGMENTS_MAX]; + static char aPad[4] = { 0, 0, 0, 0 }; + RTSgBufInit(&buf, &aSeg[0], cBuf); + unsigned iBuf = 0; + for (i = 0; i < cnRequest; i++) + { + /* Actual data chunk. */ + aSeg[iBuf].pvSeg = (void *)paRequest[i].pcvSeg; + aSeg[iBuf].cbSeg = paRequest[i].cbSeg; + iBuf++; + /* Insert proper padding before the next chunk. */ if (paRequest[i].cbSeg & 3) { - rc = pImage->pInterfaceNetCallbacks->pfnWrite(pImage->Socket, - &pad, - 4 - (paRequest[i].cbSeg & 3)); - if (RT_FAILURE(rc)) - break; + aSeg[iBuf].pvSeg = &aPad[0]; + aSeg[iBuf].cbSeg = 4 - (paRequest[i].cbSeg & 3); + iBuf++; } } - /* Send out the request as soon as possible, otherwise the target will - * answer after an unnecessary delay. */ - pImage->pInterfaceNetCallbacks->pfnFlush(pImage->Socket); + /* Send out the request, the socket is set to send data immediately, + * avoiding unnecessary delays. */ + rc = pImage->pInterfaceNetCallbacks->pfnSgWrite(pImage->Socket, &buf); + } Index: /trunk/src/VBox/Main/MediumImpl.cpp =================================================================== --- /trunk/src/VBox/Main/MediumImpl.cpp (revision 30308) +++ /trunk/src/VBox/Main/MediumImpl.cpp (revision 30309) @@ -673,5 +673,7 @@ m->vdIfCallsTcpNet.pfnRead = RTTcpRead; m->vdIfCallsTcpNet.pfnWrite = RTTcpWrite; + m->vdIfCallsTcpNet.pfnSgWrite = RTTcpSgWrite; m->vdIfCallsTcpNet.pfnFlush = RTTcpFlush; + m->vdIfCallsTcpNet.pfnSetSendCoalescing = RTTcpSetSendCoalescing; m->vdIfCallsTcpNet.pfnGetLocalAddress = RTTcpGetLocalAddress; m->vdIfCallsTcpNet.pfnGetPeerAddress = RTTcpGetPeerAddress;