16e790746SPaolo Bonzini /* 26e790746SPaolo Bonzini * Virtio Block Device 36e790746SPaolo Bonzini * 46e790746SPaolo Bonzini * Copyright IBM, Corp. 2007 56e790746SPaolo Bonzini * 66e790746SPaolo Bonzini * Authors: 76e790746SPaolo Bonzini * Anthony Liguori <aliguori@us.ibm.com> 86e790746SPaolo Bonzini * 96e790746SPaolo Bonzini * This work is licensed under the terms of the GNU GPL, version 2. See 106e790746SPaolo Bonzini * the COPYING file in the top-level directory. 116e790746SPaolo Bonzini * 126e790746SPaolo Bonzini */ 136e790746SPaolo Bonzini 1480c71a24SPeter Maydell #include "qemu/osdep.h" 15da34e65cSMarkus Armbruster #include "qapi/error.h" 16827805a2SFam Zheng #include "qemu/iov.h" 170b8fa32fSMarkus Armbruster #include "qemu/module.h" 186e790746SPaolo Bonzini #include "qemu/error-report.h" 199b92fbcfSSergio Lopez #include "qemu/main-loop.h" 206e790746SPaolo Bonzini #include "trace.h" 216e790746SPaolo Bonzini #include "hw/block/block.h" 22a27bd6c7SMarkus Armbruster #include "hw/qdev-properties.h" 236e790746SPaolo Bonzini #include "sysemu/blockdev.h" 242f780b6aSMarkus Armbruster #include "sysemu/sysemu.h" 2554d31236SMarkus Armbruster #include "sysemu/runstate.h" 266e790746SPaolo Bonzini #include "hw/virtio/virtio-blk.h" 276e790746SPaolo Bonzini #include "dataplane/virtio-blk.h" 2808e2c9f1SPaolo Bonzini #include "scsi/constants.h" 296e790746SPaolo Bonzini #ifdef __linux__ 306e790746SPaolo Bonzini # include <scsi/sg.h> 316e790746SPaolo Bonzini #endif 326e790746SPaolo Bonzini #include "hw/virtio/virtio-bus.h" 33ca77ee28SMarkus Armbruster #include "migration/qemu-file-types.h" 34783d1897SRusty Russell #include "hw/virtio/virtio-access.h" 356e790746SPaolo Bonzini 3620764be0SStefano Garzarella /* Config size before the discard support (hide associated config fields) */ 3742824b4dSChangpeng Liu #define VIRTIO_BLK_CFG_SIZE offsetof(struct virtio_blk_config, \ 3842824b4dSChangpeng Liu max_discard_sectors) 3920764be0SStefano Garzarella /* 4020764be0SStefano Garzarella * Starting from the discard feature, we can use this array to properly 4120764be0SStefano Garzarella * set the config size depending on the features enabled. 4220764be0SStefano Garzarella */ 4320764be0SStefano Garzarella static VirtIOFeature feature_sizes[] = { 4420764be0SStefano Garzarella {.flags = 1ULL << VIRTIO_BLK_F_DISCARD, 455d5b33c0SMax Reitz .end = endof(struct virtio_blk_config, discard_sector_alignment)}, 4620764be0SStefano Garzarella {.flags = 1ULL << VIRTIO_BLK_F_WRITE_ZEROES, 475d5b33c0SMax Reitz .end = endof(struct virtio_blk_config, write_zeroes_may_unmap)}, 4820764be0SStefano Garzarella {} 4920764be0SStefano Garzarella }; 5020764be0SStefano Garzarella 5120764be0SStefano Garzarella static void virtio_blk_set_config_size(VirtIOBlock *s, uint64_t host_features) 5220764be0SStefano Garzarella { 5320764be0SStefano Garzarella s->config_size = MAX(VIRTIO_BLK_CFG_SIZE, 5420764be0SStefano Garzarella virtio_feature_get_config_size(feature_sizes, host_features)); 5520764be0SStefano Garzarella 5620764be0SStefano Garzarella assert(s->config_size <= sizeof(struct virtio_blk_config)); 5720764be0SStefano Garzarella } 5842824b4dSChangpeng Liu 59d14dde5eSGreg Kurz static void virtio_blk_init_request(VirtIOBlock *s, VirtQueue *vq, 60edaffd9fSStefan Hajnoczi VirtIOBlockReq *req) 61671ec3f0SFam Zheng { 62671ec3f0SFam Zheng req->dev = s; 63edaffd9fSStefan Hajnoczi req->vq = vq; 64869d66afSStefan Hajnoczi req->qiov.size = 0; 652a6cdd6dSPaolo Bonzini req->in_len = 0; 66869d66afSStefan Hajnoczi req->next = NULL; 6795f7142aSPeter Lieven req->mr_next = NULL; 68671ec3f0SFam Zheng } 69671ec3f0SFam Zheng 70d14dde5eSGreg Kurz static void virtio_blk_free_request(VirtIOBlockReq *req) 71671ec3f0SFam Zheng { 72c84b3192SPaolo Bonzini g_free(req); 73671ec3f0SFam Zheng } 74671ec3f0SFam Zheng 7503de2f52SPaolo Bonzini static void virtio_blk_req_complete(VirtIOBlockReq *req, unsigned char status) 766e790746SPaolo Bonzini { 776e790746SPaolo Bonzini VirtIOBlock *s = req->dev; 786e790746SPaolo Bonzini VirtIODevice *vdev = VIRTIO_DEVICE(s); 796e790746SPaolo Bonzini 80a576ceacSStefan Hajnoczi trace_virtio_blk_req_complete(vdev, req, status); 816e790746SPaolo Bonzini 826e790746SPaolo Bonzini stb_p(&req->in->status, status); 83edaffd9fSStefan Hajnoczi virtqueue_push(req->vq, &req->elem, req->in_len); 84eb41cf78SPaolo Bonzini if (s->dataplane_started && !s->dataplane_disabled) { 85edaffd9fSStefan Hajnoczi virtio_blk_data_plane_notify(s->dataplane, req->vq); 8603de2f52SPaolo Bonzini } else { 87edaffd9fSStefan Hajnoczi virtio_notify(vdev, req->vq); 886e790746SPaolo Bonzini } 89bf4bd461SFam Zheng } 90bf4bd461SFam Zheng 916e790746SPaolo Bonzini static int virtio_blk_handle_rw_error(VirtIOBlockReq *req, int error, 9200f639fbSStefano Garzarella bool is_read, bool acct_failed) 936e790746SPaolo Bonzini { 946e790746SPaolo Bonzini VirtIOBlock *s = req->dev; 959a6719d5SStefano Garzarella BlockErrorAction action = blk_get_error_action(s->blk, is_read, error); 966e790746SPaolo Bonzini 97a589569fSWenchao Xia if (action == BLOCK_ERROR_ACTION_STOP) { 98466138dcSFam Zheng /* Break the link as the next request is going to be parsed from the 99466138dcSFam Zheng * ring again. Otherwise we may end up doing a double completion! */ 100466138dcSFam Zheng req->mr_next = NULL; 1016e790746SPaolo Bonzini req->next = s->rq; 1026e790746SPaolo Bonzini s->rq = req; 103a589569fSWenchao Xia } else if (action == BLOCK_ERROR_ACTION_REPORT) { 1046e790746SPaolo Bonzini virtio_blk_req_complete(req, VIRTIO_BLK_S_IOERR); 10500f639fbSStefano Garzarella if (acct_failed) { 10601762e03SAlberto Garcia block_acct_failed(blk_get_stats(s->blk), &req->acct); 10700f639fbSStefano Garzarella } 108671ec3f0SFam Zheng virtio_blk_free_request(req); 1096e790746SPaolo Bonzini } 1106e790746SPaolo Bonzini 1114be74634SMarkus Armbruster blk_error_action(s->blk, action, is_read, error); 112a589569fSWenchao Xia return action != BLOCK_ERROR_ACTION_IGNORE; 1136e790746SPaolo Bonzini } 1146e790746SPaolo Bonzini 1156e790746SPaolo Bonzini static void virtio_blk_rw_complete(void *opaque, int ret) 1166e790746SPaolo Bonzini { 11795f7142aSPeter Lieven VirtIOBlockReq *next = opaque; 118b9e413ddSPaolo Bonzini VirtIOBlock *s = next->dev; 119a576ceacSStefan Hajnoczi VirtIODevice *vdev = VIRTIO_DEVICE(s); 1206e790746SPaolo Bonzini 121b9e413ddSPaolo Bonzini aio_context_acquire(blk_get_aio_context(s->conf.conf.blk)); 12295f7142aSPeter Lieven while (next) { 12395f7142aSPeter Lieven VirtIOBlockReq *req = next; 12495f7142aSPeter Lieven next = req->mr_next; 125a576ceacSStefan Hajnoczi trace_virtio_blk_rw_complete(vdev, req, ret); 1266e790746SPaolo Bonzini 12795f7142aSPeter Lieven if (req->qiov.nalloc != -1) { 128e61809edSDongli Zhang /* If nalloc is != -1 req->qiov is a local copy of the original 1299bb192a4SYaowei Bai * external iovec. It was allocated in submit_requests to be 1309bb192a4SYaowei Bai * able to merge requests. */ 13195f7142aSPeter Lieven qemu_iovec_destroy(&req->qiov); 13295f7142aSPeter Lieven } 13395f7142aSPeter Lieven 1346e790746SPaolo Bonzini if (ret) { 135bf4069fbSAnastasiia Rusakova int p = virtio_ldl_p(VIRTIO_DEVICE(s), &req->out.type); 136783d1897SRusty Russell bool is_read = !(p & VIRTIO_BLK_T_OUT); 1372a6cdd6dSPaolo Bonzini /* Note that memory may be dirtied on read failure. If the 1382a6cdd6dSPaolo Bonzini * virtio request is not completed here, as is the case for 1392a6cdd6dSPaolo Bonzini * BLOCK_ERROR_ACTION_STOP, the memory may not be copied 1402a6cdd6dSPaolo Bonzini * correctly during live migration. While this is ugly, 1412a6cdd6dSPaolo Bonzini * it is acceptable because the device is free to write to 1422a6cdd6dSPaolo Bonzini * the memory until the request is completed (which will 1432a6cdd6dSPaolo Bonzini * happen on the other side of the migration). 1442a6cdd6dSPaolo Bonzini */ 14500f639fbSStefano Garzarella if (virtio_blk_handle_rw_error(req, -ret, is_read, true)) { 14695f7142aSPeter Lieven continue; 14795f7142aSPeter Lieven } 1486e790746SPaolo Bonzini } 1496e790746SPaolo Bonzini 1506e790746SPaolo Bonzini virtio_blk_req_complete(req, VIRTIO_BLK_S_OK); 151bf4069fbSAnastasiia Rusakova block_acct_done(blk_get_stats(s->blk), &req->acct); 152671ec3f0SFam Zheng virtio_blk_free_request(req); 1536e790746SPaolo Bonzini } 154b9e413ddSPaolo Bonzini aio_context_release(blk_get_aio_context(s->conf.conf.blk)); 15595f7142aSPeter Lieven } 1566e790746SPaolo Bonzini 1576e790746SPaolo Bonzini static void virtio_blk_flush_complete(void *opaque, int ret) 1586e790746SPaolo Bonzini { 1596e790746SPaolo Bonzini VirtIOBlockReq *req = opaque; 160b9e413ddSPaolo Bonzini VirtIOBlock *s = req->dev; 1616e790746SPaolo Bonzini 162b9e413ddSPaolo Bonzini aio_context_acquire(blk_get_aio_context(s->conf.conf.blk)); 1636e790746SPaolo Bonzini if (ret) { 16400f639fbSStefano Garzarella if (virtio_blk_handle_rw_error(req, -ret, 0, true)) { 165b9e413ddSPaolo Bonzini goto out; 1666e790746SPaolo Bonzini } 1676e790746SPaolo Bonzini } 1686e790746SPaolo Bonzini 1696e790746SPaolo Bonzini virtio_blk_req_complete(req, VIRTIO_BLK_S_OK); 1709a6719d5SStefano Garzarella block_acct_done(blk_get_stats(s->blk), &req->acct); 171671ec3f0SFam Zheng virtio_blk_free_request(req); 172b9e413ddSPaolo Bonzini 173b9e413ddSPaolo Bonzini out: 174b9e413ddSPaolo Bonzini aio_context_release(blk_get_aio_context(s->conf.conf.blk)); 1756e790746SPaolo Bonzini } 1766e790746SPaolo Bonzini 17737b06f8dSStefano Garzarella static void virtio_blk_discard_write_zeroes_complete(void *opaque, int ret) 17837b06f8dSStefano Garzarella { 17937b06f8dSStefano Garzarella VirtIOBlockReq *req = opaque; 18037b06f8dSStefano Garzarella VirtIOBlock *s = req->dev; 18137b06f8dSStefano Garzarella bool is_write_zeroes = (virtio_ldl_p(VIRTIO_DEVICE(s), &req->out.type) & 18237b06f8dSStefano Garzarella ~VIRTIO_BLK_T_BARRIER) == VIRTIO_BLK_T_WRITE_ZEROES; 18337b06f8dSStefano Garzarella 18437b06f8dSStefano Garzarella aio_context_acquire(blk_get_aio_context(s->conf.conf.blk)); 18537b06f8dSStefano Garzarella if (ret) { 18637b06f8dSStefano Garzarella if (virtio_blk_handle_rw_error(req, -ret, false, is_write_zeroes)) { 18737b06f8dSStefano Garzarella goto out; 18837b06f8dSStefano Garzarella } 18937b06f8dSStefano Garzarella } 19037b06f8dSStefano Garzarella 19137b06f8dSStefano Garzarella virtio_blk_req_complete(req, VIRTIO_BLK_S_OK); 19237b06f8dSStefano Garzarella if (is_write_zeroes) { 19337b06f8dSStefano Garzarella block_acct_done(blk_get_stats(s->blk), &req->acct); 19437b06f8dSStefano Garzarella } 19537b06f8dSStefano Garzarella virtio_blk_free_request(req); 19637b06f8dSStefano Garzarella 19737b06f8dSStefano Garzarella out: 19837b06f8dSStefano Garzarella aio_context_release(blk_get_aio_context(s->conf.conf.blk)); 19937b06f8dSStefano Garzarella } 20037b06f8dSStefano Garzarella 2011dc936aaSFam Zheng #ifdef __linux__ 2021dc936aaSFam Zheng 2031dc936aaSFam Zheng typedef struct { 2041dc936aaSFam Zheng VirtIOBlockReq *req; 2051dc936aaSFam Zheng struct sg_io_hdr hdr; 2061dc936aaSFam Zheng } VirtIOBlockIoctlReq; 2071dc936aaSFam Zheng 2081dc936aaSFam Zheng static void virtio_blk_ioctl_complete(void *opaque, int status) 2091dc936aaSFam Zheng { 2101dc936aaSFam Zheng VirtIOBlockIoctlReq *ioctl_req = opaque; 2111dc936aaSFam Zheng VirtIOBlockReq *req = ioctl_req->req; 2129d456654SPaolo Bonzini VirtIOBlock *s = req->dev; 2139d456654SPaolo Bonzini VirtIODevice *vdev = VIRTIO_DEVICE(s); 2141dc936aaSFam Zheng struct virtio_scsi_inhdr *scsi; 2151dc936aaSFam Zheng struct sg_io_hdr *hdr; 2161dc936aaSFam Zheng 2171dc936aaSFam Zheng scsi = (void *)req->elem.in_sg[req->elem.in_num - 2].iov_base; 2181dc936aaSFam Zheng 2191dc936aaSFam Zheng if (status) { 2201dc936aaSFam Zheng status = VIRTIO_BLK_S_UNSUPP; 2211dc936aaSFam Zheng virtio_stl_p(vdev, &scsi->errors, 255); 2221dc936aaSFam Zheng goto out; 2231dc936aaSFam Zheng } 2241dc936aaSFam Zheng 2251dc936aaSFam Zheng hdr = &ioctl_req->hdr; 2261dc936aaSFam Zheng /* 2271dc936aaSFam Zheng * From SCSI-Generic-HOWTO: "Some lower level drivers (e.g. ide-scsi) 2281dc936aaSFam Zheng * clear the masked_status field [hence status gets cleared too, see 2291dc936aaSFam Zheng * block/scsi_ioctl.c] even when a CHECK_CONDITION or COMMAND_TERMINATED 2301dc936aaSFam Zheng * status has occurred. However they do set DRIVER_SENSE in driver_status 2311dc936aaSFam Zheng * field. Also a (sb_len_wr > 0) indicates there is a sense buffer. 2321dc936aaSFam Zheng */ 2331dc936aaSFam Zheng if (hdr->status == 0 && hdr->sb_len_wr > 0) { 2341dc936aaSFam Zheng hdr->status = CHECK_CONDITION; 2351dc936aaSFam Zheng } 2361dc936aaSFam Zheng 2371dc936aaSFam Zheng virtio_stl_p(vdev, &scsi->errors, 2381dc936aaSFam Zheng hdr->status | (hdr->msg_status << 8) | 2391dc936aaSFam Zheng (hdr->host_status << 16) | (hdr->driver_status << 24)); 2401dc936aaSFam Zheng virtio_stl_p(vdev, &scsi->residual, hdr->resid); 2411dc936aaSFam Zheng virtio_stl_p(vdev, &scsi->sense_len, hdr->sb_len_wr); 2421dc936aaSFam Zheng virtio_stl_p(vdev, &scsi->data_len, hdr->dxfer_len); 2431dc936aaSFam Zheng 2441dc936aaSFam Zheng out: 245b9e413ddSPaolo Bonzini aio_context_acquire(blk_get_aio_context(s->conf.conf.blk)); 2461dc936aaSFam Zheng virtio_blk_req_complete(req, status); 2471dc936aaSFam Zheng virtio_blk_free_request(req); 248b9e413ddSPaolo Bonzini aio_context_release(blk_get_aio_context(s->conf.conf.blk)); 2491dc936aaSFam Zheng g_free(ioctl_req); 2501dc936aaSFam Zheng } 2511dc936aaSFam Zheng 2521dc936aaSFam Zheng #endif 2531dc936aaSFam Zheng 254edaffd9fSStefan Hajnoczi static VirtIOBlockReq *virtio_blk_get_request(VirtIOBlock *s, VirtQueue *vq) 2556e790746SPaolo Bonzini { 256edaffd9fSStefan Hajnoczi VirtIOBlockReq *req = virtqueue_pop(vq, sizeof(VirtIOBlockReq)); 2576e790746SPaolo Bonzini 25851b19ebeSPaolo Bonzini if (req) { 259edaffd9fSStefan Hajnoczi virtio_blk_init_request(s, vq, req); 2606e790746SPaolo Bonzini } 2616e790746SPaolo Bonzini return req; 2626e790746SPaolo Bonzini } 2636e790746SPaolo Bonzini 26475344fa4SFam Zheng static int virtio_blk_handle_scsi_req(VirtIOBlockReq *req) 2656e790746SPaolo Bonzini { 2666e790746SPaolo Bonzini int status = VIRTIO_BLK_S_OK; 2675a05cbeeSFam Zheng struct virtio_scsi_inhdr *scsi = NULL; 26875344fa4SFam Zheng VirtIOBlock *blk = req->dev; 269bf4069fbSAnastasiia Rusakova VirtIODevice *vdev = VIRTIO_DEVICE(blk); 270bf4069fbSAnastasiia Rusakova VirtQueueElement *elem = &req->elem; 271783d1897SRusty Russell 2725a05cbeeSFam Zheng #ifdef __linux__ 2735a05cbeeSFam Zheng int i; 2741dc936aaSFam Zheng VirtIOBlockIoctlReq *ioctl_req; 275a209f461SFam Zheng BlockAIOCB *acb; 2765a05cbeeSFam Zheng #endif 2776e790746SPaolo Bonzini 2786e790746SPaolo Bonzini /* 2796e790746SPaolo Bonzini * We require at least one output segment each for the virtio_blk_outhdr 2806e790746SPaolo Bonzini * and the SCSI command block. 2816e790746SPaolo Bonzini * 2826e790746SPaolo Bonzini * We also at least require the virtio_blk_inhdr, the virtio_scsi_inhdr 2836e790746SPaolo Bonzini * and the sense buffer pointer in the input segments. 2846e790746SPaolo Bonzini */ 2855a05cbeeSFam Zheng if (elem->out_num < 2 || elem->in_num < 3) { 2865a05cbeeSFam Zheng status = VIRTIO_BLK_S_IOERR; 2875a05cbeeSFam Zheng goto fail; 2886e790746SPaolo Bonzini } 2896e790746SPaolo Bonzini 2906e790746SPaolo Bonzini /* 2916e790746SPaolo Bonzini * The scsi inhdr is placed in the second-to-last input segment, just 2926e790746SPaolo Bonzini * before the regular inhdr. 2936e790746SPaolo Bonzini */ 2945a05cbeeSFam Zheng scsi = (void *)elem->in_sg[elem->in_num - 2].iov_base; 2956e790746SPaolo Bonzini 296bbe8bd4dSStefano Garzarella if (!virtio_has_feature(blk->host_features, VIRTIO_BLK_F_SCSI)) { 2976e790746SPaolo Bonzini status = VIRTIO_BLK_S_UNSUPP; 2986e790746SPaolo Bonzini goto fail; 2996e790746SPaolo Bonzini } 3006e790746SPaolo Bonzini 3016e790746SPaolo Bonzini /* 3026e790746SPaolo Bonzini * No support for bidirection commands yet. 3036e790746SPaolo Bonzini */ 3045a05cbeeSFam Zheng if (elem->out_num > 2 && elem->in_num > 3) { 3056e790746SPaolo Bonzini status = VIRTIO_BLK_S_UNSUPP; 3066e790746SPaolo Bonzini goto fail; 3076e790746SPaolo Bonzini } 3086e790746SPaolo Bonzini 3096e790746SPaolo Bonzini #ifdef __linux__ 3101dc936aaSFam Zheng ioctl_req = g_new0(VirtIOBlockIoctlReq, 1); 3111dc936aaSFam Zheng ioctl_req->req = req; 3121dc936aaSFam Zheng ioctl_req->hdr.interface_id = 'S'; 3131dc936aaSFam Zheng ioctl_req->hdr.cmd_len = elem->out_sg[1].iov_len; 3141dc936aaSFam Zheng ioctl_req->hdr.cmdp = elem->out_sg[1].iov_base; 3151dc936aaSFam Zheng ioctl_req->hdr.dxfer_len = 0; 3166e790746SPaolo Bonzini 3175a05cbeeSFam Zheng if (elem->out_num > 2) { 3186e790746SPaolo Bonzini /* 3196e790746SPaolo Bonzini * If there are more than the minimally required 2 output segments 3206e790746SPaolo Bonzini * there is write payload starting from the third iovec. 3216e790746SPaolo Bonzini */ 3221dc936aaSFam Zheng ioctl_req->hdr.dxfer_direction = SG_DXFER_TO_DEV; 3231dc936aaSFam Zheng ioctl_req->hdr.iovec_count = elem->out_num - 2; 3246e790746SPaolo Bonzini 3251dc936aaSFam Zheng for (i = 0; i < ioctl_req->hdr.iovec_count; i++) { 3261dc936aaSFam Zheng ioctl_req->hdr.dxfer_len += elem->out_sg[i + 2].iov_len; 3271dc936aaSFam Zheng } 3286e790746SPaolo Bonzini 3291dc936aaSFam Zheng ioctl_req->hdr.dxferp = elem->out_sg + 2; 3306e790746SPaolo Bonzini 3315a05cbeeSFam Zheng } else if (elem->in_num > 3) { 3326e790746SPaolo Bonzini /* 3336e790746SPaolo Bonzini * If we have more than 3 input segments the guest wants to actually 3346e790746SPaolo Bonzini * read data. 3356e790746SPaolo Bonzini */ 3361dc936aaSFam Zheng ioctl_req->hdr.dxfer_direction = SG_DXFER_FROM_DEV; 3371dc936aaSFam Zheng ioctl_req->hdr.iovec_count = elem->in_num - 3; 3381dc936aaSFam Zheng for (i = 0; i < ioctl_req->hdr.iovec_count; i++) { 3391dc936aaSFam Zheng ioctl_req->hdr.dxfer_len += elem->in_sg[i].iov_len; 3401dc936aaSFam Zheng } 3416e790746SPaolo Bonzini 3421dc936aaSFam Zheng ioctl_req->hdr.dxferp = elem->in_sg; 3436e790746SPaolo Bonzini } else { 3446e790746SPaolo Bonzini /* 3456e790746SPaolo Bonzini * Some SCSI commands don't actually transfer any data. 3466e790746SPaolo Bonzini */ 3471dc936aaSFam Zheng ioctl_req->hdr.dxfer_direction = SG_DXFER_NONE; 3486e790746SPaolo Bonzini } 3496e790746SPaolo Bonzini 3501dc936aaSFam Zheng ioctl_req->hdr.sbp = elem->in_sg[elem->in_num - 3].iov_base; 3511dc936aaSFam Zheng ioctl_req->hdr.mx_sb_len = elem->in_sg[elem->in_num - 3].iov_len; 3526e790746SPaolo Bonzini 353a209f461SFam Zheng acb = blk_aio_ioctl(blk->blk, SG_IO, &ioctl_req->hdr, 3541dc936aaSFam Zheng virtio_blk_ioctl_complete, ioctl_req); 355a209f461SFam Zheng if (!acb) { 356a209f461SFam Zheng g_free(ioctl_req); 357a209f461SFam Zheng status = VIRTIO_BLK_S_UNSUPP; 358a209f461SFam Zheng goto fail; 359a209f461SFam Zheng } 3601dc936aaSFam Zheng return -EINPROGRESS; 3616e790746SPaolo Bonzini #else 3626e790746SPaolo Bonzini abort(); 3636e790746SPaolo Bonzini #endif 3646e790746SPaolo Bonzini 3656e790746SPaolo Bonzini fail: 3666e790746SPaolo Bonzini /* Just put anything nonzero so that the ioctl fails in the guest. */ 3675a05cbeeSFam Zheng if (scsi) { 368783d1897SRusty Russell virtio_stl_p(vdev, &scsi->errors, 255); 3695a05cbeeSFam Zheng } 3705a05cbeeSFam Zheng return status; 3715a05cbeeSFam Zheng } 3725a05cbeeSFam Zheng 3735a05cbeeSFam Zheng static void virtio_blk_handle_scsi(VirtIOBlockReq *req) 3745a05cbeeSFam Zheng { 3755a05cbeeSFam Zheng int status; 3765a05cbeeSFam Zheng 37775344fa4SFam Zheng status = virtio_blk_handle_scsi_req(req); 3781dc936aaSFam Zheng if (status != -EINPROGRESS) { 3796e790746SPaolo Bonzini virtio_blk_req_complete(req, status); 380671ec3f0SFam Zheng virtio_blk_free_request(req); 3816e790746SPaolo Bonzini } 3821dc936aaSFam Zheng } 3836e790746SPaolo Bonzini 38495f7142aSPeter Lieven static inline void submit_requests(BlockBackend *blk, MultiReqBuffer *mrb, 38595f7142aSPeter Lieven int start, int num_reqs, int niov) 3866e790746SPaolo Bonzini { 38795f7142aSPeter Lieven QEMUIOVector *qiov = &mrb->reqs[start]->qiov; 38895f7142aSPeter Lieven int64_t sector_num = mrb->reqs[start]->sector_num; 38995f7142aSPeter Lieven bool is_write = mrb->is_write; 3906e790746SPaolo Bonzini 39195f7142aSPeter Lieven if (num_reqs > 1) { 39295f7142aSPeter Lieven int i; 39395f7142aSPeter Lieven struct iovec *tmp_iov = qiov->iov; 39495f7142aSPeter Lieven int tmp_niov = qiov->niov; 39595f7142aSPeter Lieven 39695f7142aSPeter Lieven /* mrb->reqs[start]->qiov was initialized from external so we can't 397b5772fddSEric Blake * modify it here. We need to initialize it locally and then add the 39895f7142aSPeter Lieven * external iovecs. */ 39995f7142aSPeter Lieven qemu_iovec_init(qiov, niov); 40095f7142aSPeter Lieven 40195f7142aSPeter Lieven for (i = 0; i < tmp_niov; i++) { 40295f7142aSPeter Lieven qemu_iovec_add(qiov, tmp_iov[i].iov_base, tmp_iov[i].iov_len); 40395f7142aSPeter Lieven } 40495f7142aSPeter Lieven 40595f7142aSPeter Lieven for (i = start + 1; i < start + num_reqs; i++) { 40695f7142aSPeter Lieven qemu_iovec_concat(qiov, &mrb->reqs[i]->qiov, 0, 40795f7142aSPeter Lieven mrb->reqs[i]->qiov.size); 40895f7142aSPeter Lieven mrb->reqs[i - 1]->mr_next = mrb->reqs[i]; 40995f7142aSPeter Lieven } 41095f7142aSPeter Lieven 411a576ceacSStefan Hajnoczi trace_virtio_blk_submit_multireq(VIRTIO_DEVICE(mrb->reqs[start]->dev), 412a576ceacSStefan Hajnoczi mrb, start, num_reqs, 413b5772fddSEric Blake sector_num << BDRV_SECTOR_BITS, 414b5772fddSEric Blake qiov->size, is_write); 41595f7142aSPeter Lieven block_acct_merge_done(blk_get_stats(blk), 41695f7142aSPeter Lieven is_write ? BLOCK_ACCT_WRITE : BLOCK_ACCT_READ, 41795f7142aSPeter Lieven num_reqs - 1); 41895f7142aSPeter Lieven } 41995f7142aSPeter Lieven 42095f7142aSPeter Lieven if (is_write) { 421b5772fddSEric Blake blk_aio_pwritev(blk, sector_num << BDRV_SECTOR_BITS, qiov, 0, 42295f7142aSPeter Lieven virtio_blk_rw_complete, mrb->reqs[start]); 42395f7142aSPeter Lieven } else { 424b5772fddSEric Blake blk_aio_preadv(blk, sector_num << BDRV_SECTOR_BITS, qiov, 0, 42595f7142aSPeter Lieven virtio_blk_rw_complete, mrb->reqs[start]); 42695f7142aSPeter Lieven } 42795f7142aSPeter Lieven } 42895f7142aSPeter Lieven 42995f7142aSPeter Lieven static int multireq_compare(const void *a, const void *b) 43095f7142aSPeter Lieven { 43195f7142aSPeter Lieven const VirtIOBlockReq *req1 = *(VirtIOBlockReq **)a, 43295f7142aSPeter Lieven *req2 = *(VirtIOBlockReq **)b; 43395f7142aSPeter Lieven 43495f7142aSPeter Lieven /* 43595f7142aSPeter Lieven * Note that we can't simply subtract sector_num1 from sector_num2 43695f7142aSPeter Lieven * here as that could overflow the return value. 43795f7142aSPeter Lieven */ 43895f7142aSPeter Lieven if (req1->sector_num > req2->sector_num) { 43995f7142aSPeter Lieven return 1; 44095f7142aSPeter Lieven } else if (req1->sector_num < req2->sector_num) { 44195f7142aSPeter Lieven return -1; 44295f7142aSPeter Lieven } else { 44395f7142aSPeter Lieven return 0; 44495f7142aSPeter Lieven } 44595f7142aSPeter Lieven } 44695f7142aSPeter Lieven 447d14dde5eSGreg Kurz static void virtio_blk_submit_multireq(BlockBackend *blk, MultiReqBuffer *mrb) 44895f7142aSPeter Lieven { 44995f7142aSPeter Lieven int i = 0, start = 0, num_reqs = 0, niov = 0, nb_sectors = 0; 4505def6b80SEric Blake uint32_t max_transfer; 45195f7142aSPeter Lieven int64_t sector_num = 0; 45295f7142aSPeter Lieven 45395f7142aSPeter Lieven if (mrb->num_reqs == 1) { 45495f7142aSPeter Lieven submit_requests(blk, mrb, 0, 1, -1); 45595f7142aSPeter Lieven mrb->num_reqs = 0; 4566e790746SPaolo Bonzini return; 4576e790746SPaolo Bonzini } 4586e790746SPaolo Bonzini 4595def6b80SEric Blake max_transfer = blk_get_max_transfer(mrb->reqs[0]->dev->blk); 46095f7142aSPeter Lieven 46195f7142aSPeter Lieven qsort(mrb->reqs, mrb->num_reqs, sizeof(*mrb->reqs), 46295f7142aSPeter Lieven &multireq_compare); 46395f7142aSPeter Lieven 46495f7142aSPeter Lieven for (i = 0; i < mrb->num_reqs; i++) { 46595f7142aSPeter Lieven VirtIOBlockReq *req = mrb->reqs[i]; 46695f7142aSPeter Lieven if (num_reqs > 0) { 46749cffbc6SGonglei /* 46849cffbc6SGonglei * NOTE: We cannot merge the requests in below situations: 46949cffbc6SGonglei * 1. requests are not sequential 47049cffbc6SGonglei * 2. merge would exceed maximum number of IOVs 47149cffbc6SGonglei * 3. merge would exceed maximum transfer length of backend device 47249cffbc6SGonglei */ 47349cffbc6SGonglei if (sector_num + nb_sectors != req->sector_num || 474222565f6SStefan Hajnoczi niov > blk_get_max_iov(blk) - req->qiov.niov || 4755def6b80SEric Blake req->qiov.size > max_transfer || 4765def6b80SEric Blake nb_sectors > (max_transfer - 4775def6b80SEric Blake req->qiov.size) / BDRV_SECTOR_SIZE) { 47895f7142aSPeter Lieven submit_requests(blk, mrb, start, num_reqs, niov); 47995f7142aSPeter Lieven num_reqs = 0; 4806e790746SPaolo Bonzini } 4816e790746SPaolo Bonzini } 4826e790746SPaolo Bonzini 48395f7142aSPeter Lieven if (num_reqs == 0) { 48495f7142aSPeter Lieven sector_num = req->sector_num; 48595f7142aSPeter Lieven nb_sectors = niov = 0; 48695f7142aSPeter Lieven start = i; 48795f7142aSPeter Lieven } 48895f7142aSPeter Lieven 48995f7142aSPeter Lieven nb_sectors += req->qiov.size / BDRV_SECTOR_SIZE; 49095f7142aSPeter Lieven niov += req->qiov.niov; 49195f7142aSPeter Lieven num_reqs++; 49295f7142aSPeter Lieven } 49395f7142aSPeter Lieven 49495f7142aSPeter Lieven submit_requests(blk, mrb, start, num_reqs, niov); 49595f7142aSPeter Lieven mrb->num_reqs = 0; 4966e790746SPaolo Bonzini } 4976e790746SPaolo Bonzini 4986e790746SPaolo Bonzini static void virtio_blk_handle_flush(VirtIOBlockReq *req, MultiReqBuffer *mrb) 4996e790746SPaolo Bonzini { 500bf4069fbSAnastasiia Rusakova VirtIOBlock *s = req->dev; 501bf4069fbSAnastasiia Rusakova 502bf4069fbSAnastasiia Rusakova block_acct_start(blk_get_stats(s->blk), &req->acct, 0, 5035366d0c8SBenoît Canet BLOCK_ACCT_FLUSH); 5046e790746SPaolo Bonzini 5056e790746SPaolo Bonzini /* 5066e790746SPaolo Bonzini * Make sure all outstanding writes are posted to the backing device. 5076e790746SPaolo Bonzini */ 50895f7142aSPeter Lieven if (mrb->is_write && mrb->num_reqs > 0) { 509bf4069fbSAnastasiia Rusakova virtio_blk_submit_multireq(s->blk, mrb); 51095f7142aSPeter Lieven } 511bf4069fbSAnastasiia Rusakova blk_aio_flush(s->blk, virtio_blk_flush_complete, req); 5126e790746SPaolo Bonzini } 5136e790746SPaolo Bonzini 514d0e14376SMarkus Armbruster static bool virtio_blk_sect_range_ok(VirtIOBlock *dev, 515d0e14376SMarkus Armbruster uint64_t sector, size_t size) 516d0e14376SMarkus Armbruster { 5173c2daac0SMarkus Armbruster uint64_t nb_sectors = size >> BDRV_SECTOR_BITS; 5183c2daac0SMarkus Armbruster uint64_t total_sectors; 5193c2daac0SMarkus Armbruster 52075af1f34SPeter Lieven if (nb_sectors > BDRV_REQUEST_MAX_SECTORS) { 52195f7142aSPeter Lieven return false; 52295f7142aSPeter Lieven } 523d0e14376SMarkus Armbruster if (sector & dev->sector_mask) { 524d0e14376SMarkus Armbruster return false; 525d0e14376SMarkus Armbruster } 5262a30307fSMarkus Armbruster if (size % dev->conf.conf.logical_block_size) { 527d0e14376SMarkus Armbruster return false; 528d0e14376SMarkus Armbruster } 5294be74634SMarkus Armbruster blk_get_geometry(dev->blk, &total_sectors); 5303c2daac0SMarkus Armbruster if (sector > total_sectors || nb_sectors > total_sectors - sector) { 5313c2daac0SMarkus Armbruster return false; 5323c2daac0SMarkus Armbruster } 533d0e14376SMarkus Armbruster return true; 534d0e14376SMarkus Armbruster } 535d0e14376SMarkus Armbruster 53637b06f8dSStefano Garzarella static uint8_t virtio_blk_handle_discard_write_zeroes(VirtIOBlockReq *req, 53737b06f8dSStefano Garzarella struct virtio_blk_discard_write_zeroes *dwz_hdr, bool is_write_zeroes) 53837b06f8dSStefano Garzarella { 53937b06f8dSStefano Garzarella VirtIOBlock *s = req->dev; 54037b06f8dSStefano Garzarella VirtIODevice *vdev = VIRTIO_DEVICE(s); 54137b06f8dSStefano Garzarella uint64_t sector; 54237b06f8dSStefano Garzarella uint32_t num_sectors, flags, max_sectors; 54337b06f8dSStefano Garzarella uint8_t err_status; 54437b06f8dSStefano Garzarella int bytes; 54537b06f8dSStefano Garzarella 54637b06f8dSStefano Garzarella sector = virtio_ldq_p(vdev, &dwz_hdr->sector); 54737b06f8dSStefano Garzarella num_sectors = virtio_ldl_p(vdev, &dwz_hdr->num_sectors); 54837b06f8dSStefano Garzarella flags = virtio_ldl_p(vdev, &dwz_hdr->flags); 54937b06f8dSStefano Garzarella max_sectors = is_write_zeroes ? s->conf.max_write_zeroes_sectors : 55037b06f8dSStefano Garzarella s->conf.max_discard_sectors; 55137b06f8dSStefano Garzarella 55237b06f8dSStefano Garzarella /* 55337b06f8dSStefano Garzarella * max_sectors is at most BDRV_REQUEST_MAX_SECTORS, this check 55437b06f8dSStefano Garzarella * make us sure that "num_sectors << BDRV_SECTOR_BITS" can fit in 55537b06f8dSStefano Garzarella * the integer variable. 55637b06f8dSStefano Garzarella */ 55737b06f8dSStefano Garzarella if (unlikely(num_sectors > max_sectors)) { 55837b06f8dSStefano Garzarella err_status = VIRTIO_BLK_S_IOERR; 55937b06f8dSStefano Garzarella goto err; 56037b06f8dSStefano Garzarella } 56137b06f8dSStefano Garzarella 56237b06f8dSStefano Garzarella bytes = num_sectors << BDRV_SECTOR_BITS; 56337b06f8dSStefano Garzarella 56437b06f8dSStefano Garzarella if (unlikely(!virtio_blk_sect_range_ok(s, sector, bytes))) { 56537b06f8dSStefano Garzarella err_status = VIRTIO_BLK_S_IOERR; 56637b06f8dSStefano Garzarella goto err; 56737b06f8dSStefano Garzarella } 56837b06f8dSStefano Garzarella 56937b06f8dSStefano Garzarella /* 57037b06f8dSStefano Garzarella * The device MUST set the status byte to VIRTIO_BLK_S_UNSUPP for discard 57137b06f8dSStefano Garzarella * and write zeroes commands if any unknown flag is set. 57237b06f8dSStefano Garzarella */ 57337b06f8dSStefano Garzarella if (unlikely(flags & ~VIRTIO_BLK_WRITE_ZEROES_FLAG_UNMAP)) { 57437b06f8dSStefano Garzarella err_status = VIRTIO_BLK_S_UNSUPP; 57537b06f8dSStefano Garzarella goto err; 57637b06f8dSStefano Garzarella } 57737b06f8dSStefano Garzarella 57837b06f8dSStefano Garzarella if (is_write_zeroes) { /* VIRTIO_BLK_T_WRITE_ZEROES */ 57937b06f8dSStefano Garzarella int blk_aio_flags = 0; 58037b06f8dSStefano Garzarella 58137b06f8dSStefano Garzarella if (flags & VIRTIO_BLK_WRITE_ZEROES_FLAG_UNMAP) { 58237b06f8dSStefano Garzarella blk_aio_flags |= BDRV_REQ_MAY_UNMAP; 58337b06f8dSStefano Garzarella } 58437b06f8dSStefano Garzarella 58537b06f8dSStefano Garzarella block_acct_start(blk_get_stats(s->blk), &req->acct, bytes, 58637b06f8dSStefano Garzarella BLOCK_ACCT_WRITE); 58737b06f8dSStefano Garzarella 58837b06f8dSStefano Garzarella blk_aio_pwrite_zeroes(s->blk, sector << BDRV_SECTOR_BITS, 58937b06f8dSStefano Garzarella bytes, blk_aio_flags, 59037b06f8dSStefano Garzarella virtio_blk_discard_write_zeroes_complete, req); 59137b06f8dSStefano Garzarella } else { /* VIRTIO_BLK_T_DISCARD */ 59237b06f8dSStefano Garzarella /* 59337b06f8dSStefano Garzarella * The device MUST set the status byte to VIRTIO_BLK_S_UNSUPP for 59437b06f8dSStefano Garzarella * discard commands if the unmap flag is set. 59537b06f8dSStefano Garzarella */ 59637b06f8dSStefano Garzarella if (unlikely(flags & VIRTIO_BLK_WRITE_ZEROES_FLAG_UNMAP)) { 59737b06f8dSStefano Garzarella err_status = VIRTIO_BLK_S_UNSUPP; 59837b06f8dSStefano Garzarella goto err; 59937b06f8dSStefano Garzarella } 60037b06f8dSStefano Garzarella 60137b06f8dSStefano Garzarella blk_aio_pdiscard(s->blk, sector << BDRV_SECTOR_BITS, bytes, 60237b06f8dSStefano Garzarella virtio_blk_discard_write_zeroes_complete, req); 60337b06f8dSStefano Garzarella } 60437b06f8dSStefano Garzarella 60537b06f8dSStefano Garzarella return VIRTIO_BLK_S_OK; 60637b06f8dSStefano Garzarella 60737b06f8dSStefano Garzarella err: 60837b06f8dSStefano Garzarella if (is_write_zeroes) { 60937b06f8dSStefano Garzarella block_acct_invalid(blk_get_stats(s->blk), BLOCK_ACCT_WRITE); 61037b06f8dSStefano Garzarella } 61137b06f8dSStefano Garzarella return err_status; 61237b06f8dSStefano Garzarella } 61337b06f8dSStefano Garzarella 61420ea686aSGreg Kurz static int virtio_blk_handle_request(VirtIOBlockReq *req, MultiReqBuffer *mrb) 6156e790746SPaolo Bonzini { 6166e790746SPaolo Bonzini uint32_t type; 617f897bf75SStefan Hajnoczi struct iovec *in_iov = req->elem.in_sg; 6185636da76SDongli Zhang struct iovec *out_iov = req->elem.out_sg; 619f897bf75SStefan Hajnoczi unsigned in_num = req->elem.in_num; 620f897bf75SStefan Hajnoczi unsigned out_num = req->elem.out_num; 62120ea686aSGreg Kurz VirtIOBlock *s = req->dev; 62220ea686aSGreg Kurz VirtIODevice *vdev = VIRTIO_DEVICE(s); 6236e790746SPaolo Bonzini 624f897bf75SStefan Hajnoczi if (req->elem.out_num < 1 || req->elem.in_num < 1) { 62520ea686aSGreg Kurz virtio_error(vdev, "virtio-blk missing headers"); 62620ea686aSGreg Kurz return -1; 6276e790746SPaolo Bonzini } 6286e790746SPaolo Bonzini 6295636da76SDongli Zhang if (unlikely(iov_to_buf(out_iov, out_num, 0, &req->out, 630827805a2SFam Zheng sizeof(req->out)) != sizeof(req->out))) { 63120ea686aSGreg Kurz virtio_error(vdev, "virtio-blk request outhdr too short"); 63220ea686aSGreg Kurz return -1; 633827805a2SFam Zheng } 634ee17e848SFam Zheng 6355636da76SDongli Zhang iov_discard_front(&out_iov, &out_num, sizeof(req->out)); 636ee17e848SFam Zheng 63712048545SGonglei if (in_iov[in_num - 1].iov_len < sizeof(struct virtio_blk_inhdr)) { 63820ea686aSGreg Kurz virtio_error(vdev, "virtio-blk request inhdr too short"); 63920ea686aSGreg Kurz return -1; 640ee17e848SFam Zheng } 641ee17e848SFam Zheng 6422a6cdd6dSPaolo Bonzini /* We always touch the last byte, so just see how big in_iov is. */ 6432a6cdd6dSPaolo Bonzini req->in_len = iov_size(in_iov, in_num); 644ee17e848SFam Zheng req->in = (void *)in_iov[in_num - 1].iov_base 645ee17e848SFam Zheng + in_iov[in_num - 1].iov_len 646ee17e848SFam Zheng - sizeof(struct virtio_blk_inhdr); 647ee17e848SFam Zheng iov_discard_back(in_iov, &in_num, sizeof(struct virtio_blk_inhdr)); 6486e790746SPaolo Bonzini 6499a6719d5SStefano Garzarella type = virtio_ldl_p(vdev, &req->out.type); 6506e790746SPaolo Bonzini 65195f7142aSPeter Lieven /* VIRTIO_BLK_T_OUT defines the command direction. VIRTIO_BLK_T_BARRIER 652631b22eaSStefan Weil * is an optional flag. Although a guest should not send this flag if 65395f7142aSPeter Lieven * not negotiated we ignored it in the past. So keep ignoring it. */ 65495f7142aSPeter Lieven switch (type & ~(VIRTIO_BLK_T_OUT | VIRTIO_BLK_T_BARRIER)) { 65595f7142aSPeter Lieven case VIRTIO_BLK_T_IN: 65695f7142aSPeter Lieven { 65795f7142aSPeter Lieven bool is_write = type & VIRTIO_BLK_T_OUT; 6589a6719d5SStefano Garzarella req->sector_num = virtio_ldq_p(vdev, &req->out.sector); 65995f7142aSPeter Lieven 66095f7142aSPeter Lieven if (is_write) { 6615636da76SDongli Zhang qemu_iovec_init_external(&req->qiov, out_iov, out_num); 662a576ceacSStefan Hajnoczi trace_virtio_blk_handle_write(vdev, req, req->sector_num, 66395f7142aSPeter Lieven req->qiov.size / BDRV_SECTOR_SIZE); 66495f7142aSPeter Lieven } else { 66595f7142aSPeter Lieven qemu_iovec_init_external(&req->qiov, in_iov, in_num); 666a576ceacSStefan Hajnoczi trace_virtio_blk_handle_read(vdev, req, req->sector_num, 66795f7142aSPeter Lieven req->qiov.size / BDRV_SECTOR_SIZE); 66895f7142aSPeter Lieven } 66995f7142aSPeter Lieven 6709a6719d5SStefano Garzarella if (!virtio_blk_sect_range_ok(s, req->sector_num, req->qiov.size)) { 67195f7142aSPeter Lieven virtio_blk_req_complete(req, VIRTIO_BLK_S_IOERR); 6729a6719d5SStefano Garzarella block_acct_invalid(blk_get_stats(s->blk), 67301762e03SAlberto Garcia is_write ? BLOCK_ACCT_WRITE : BLOCK_ACCT_READ); 67495f7142aSPeter Lieven virtio_blk_free_request(req); 67520ea686aSGreg Kurz return 0; 67695f7142aSPeter Lieven } 67795f7142aSPeter Lieven 6789a6719d5SStefano Garzarella block_acct_start(blk_get_stats(s->blk), &req->acct, req->qiov.size, 67995f7142aSPeter Lieven is_write ? BLOCK_ACCT_WRITE : BLOCK_ACCT_READ); 68095f7142aSPeter Lieven 68195f7142aSPeter Lieven /* merge would exceed maximum number of requests or IO direction 68295f7142aSPeter Lieven * changes */ 68395f7142aSPeter Lieven if (mrb->num_reqs > 0 && (mrb->num_reqs == VIRTIO_BLK_MAX_MERGE_REQS || 684c99495acSPeter Lieven is_write != mrb->is_write || 6859a6719d5SStefano Garzarella !s->conf.request_merging)) { 6869a6719d5SStefano Garzarella virtio_blk_submit_multireq(s->blk, mrb); 68795f7142aSPeter Lieven } 68895f7142aSPeter Lieven 68995f7142aSPeter Lieven assert(mrb->num_reqs < VIRTIO_BLK_MAX_MERGE_REQS); 69095f7142aSPeter Lieven mrb->reqs[mrb->num_reqs++] = req; 69195f7142aSPeter Lieven mrb->is_write = is_write; 69295f7142aSPeter Lieven break; 69395f7142aSPeter Lieven } 69495f7142aSPeter Lieven case VIRTIO_BLK_T_FLUSH: 6956e790746SPaolo Bonzini virtio_blk_handle_flush(req, mrb); 69695f7142aSPeter Lieven break; 69795f7142aSPeter Lieven case VIRTIO_BLK_T_SCSI_CMD: 6986e790746SPaolo Bonzini virtio_blk_handle_scsi(req); 69995f7142aSPeter Lieven break; 70095f7142aSPeter Lieven case VIRTIO_BLK_T_GET_ID: 70195f7142aSPeter Lieven { 7026e790746SPaolo Bonzini /* 7036e790746SPaolo Bonzini * NB: per existing s/n string convention the string is 7046e790746SPaolo Bonzini * terminated by '\0' only when shorter than buffer. 7056e790746SPaolo Bonzini */ 7062a30307fSMarkus Armbruster const char *serial = s->conf.serial ? s->conf.serial : ""; 707a83ceea8SMarc Marí size_t size = MIN(strlen(serial) + 1, 708a83ceea8SMarc Marí MIN(iov_size(in_iov, in_num), 709a83ceea8SMarc Marí VIRTIO_BLK_ID_BYTES)); 710a83ceea8SMarc Marí iov_from_buf(in_iov, in_num, 0, serial, size); 7116e790746SPaolo Bonzini virtio_blk_req_complete(req, VIRTIO_BLK_S_OK); 712671ec3f0SFam Zheng virtio_blk_free_request(req); 71395f7142aSPeter Lieven break; 71495f7142aSPeter Lieven } 71537b06f8dSStefano Garzarella /* 71637b06f8dSStefano Garzarella * VIRTIO_BLK_T_DISCARD and VIRTIO_BLK_T_WRITE_ZEROES are defined with 71737b06f8dSStefano Garzarella * VIRTIO_BLK_T_OUT flag set. We masked this flag in the switch statement, 71837b06f8dSStefano Garzarella * so we must mask it for these requests, then we will check if it is set. 71937b06f8dSStefano Garzarella */ 72037b06f8dSStefano Garzarella case VIRTIO_BLK_T_DISCARD & ~VIRTIO_BLK_T_OUT: 72137b06f8dSStefano Garzarella case VIRTIO_BLK_T_WRITE_ZEROES & ~VIRTIO_BLK_T_OUT: 72237b06f8dSStefano Garzarella { 72337b06f8dSStefano Garzarella struct virtio_blk_discard_write_zeroes dwz_hdr; 72437b06f8dSStefano Garzarella size_t out_len = iov_size(out_iov, out_num); 72537b06f8dSStefano Garzarella bool is_write_zeroes = (type & ~VIRTIO_BLK_T_BARRIER) == 72637b06f8dSStefano Garzarella VIRTIO_BLK_T_WRITE_ZEROES; 72737b06f8dSStefano Garzarella uint8_t err_status; 72837b06f8dSStefano Garzarella 72937b06f8dSStefano Garzarella /* 73037b06f8dSStefano Garzarella * Unsupported if VIRTIO_BLK_T_OUT is not set or the request contains 73137b06f8dSStefano Garzarella * more than one segment. 73237b06f8dSStefano Garzarella */ 73337b06f8dSStefano Garzarella if (unlikely(!(type & VIRTIO_BLK_T_OUT) || 73437b06f8dSStefano Garzarella out_len > sizeof(dwz_hdr))) { 73537b06f8dSStefano Garzarella virtio_blk_req_complete(req, VIRTIO_BLK_S_UNSUPP); 73637b06f8dSStefano Garzarella virtio_blk_free_request(req); 73737b06f8dSStefano Garzarella return 0; 73837b06f8dSStefano Garzarella } 73937b06f8dSStefano Garzarella 74037b06f8dSStefano Garzarella if (unlikely(iov_to_buf(out_iov, out_num, 0, &dwz_hdr, 74137b06f8dSStefano Garzarella sizeof(dwz_hdr)) != sizeof(dwz_hdr))) { 74237b06f8dSStefano Garzarella virtio_error(vdev, "virtio-blk discard/write_zeroes header" 74337b06f8dSStefano Garzarella " too short"); 74437b06f8dSStefano Garzarella return -1; 74537b06f8dSStefano Garzarella } 74637b06f8dSStefano Garzarella 74737b06f8dSStefano Garzarella err_status = virtio_blk_handle_discard_write_zeroes(req, &dwz_hdr, 74837b06f8dSStefano Garzarella is_write_zeroes); 74937b06f8dSStefano Garzarella if (err_status != VIRTIO_BLK_S_OK) { 75037b06f8dSStefano Garzarella virtio_blk_req_complete(req, err_status); 75137b06f8dSStefano Garzarella virtio_blk_free_request(req); 75237b06f8dSStefano Garzarella } 75337b06f8dSStefano Garzarella 75437b06f8dSStefano Garzarella break; 75537b06f8dSStefano Garzarella } 75695f7142aSPeter Lieven default: 7576e790746SPaolo Bonzini virtio_blk_req_complete(req, VIRTIO_BLK_S_UNSUPP); 758671ec3f0SFam Zheng virtio_blk_free_request(req); 7596e790746SPaolo Bonzini } 76020ea686aSGreg Kurz return 0; 7616e790746SPaolo Bonzini } 7626e790746SPaolo Bonzini 76307931698SFam Zheng bool virtio_blk_handle_vq(VirtIOBlock *s, VirtQueue *vq) 7646e790746SPaolo Bonzini { 7656e790746SPaolo Bonzini VirtIOBlockReq *req; 76695f7142aSPeter Lieven MultiReqBuffer mrb = {}; 767d0435bc5SStefan Hajnoczi bool suppress_notifications = virtio_queue_get_notification(vq); 76807931698SFam Zheng bool progress = false; 7696e790746SPaolo Bonzini 7709d456654SPaolo Bonzini aio_context_acquire(blk_get_aio_context(s->blk)); 771fc73548eSStefan Hajnoczi blk_io_plug(s->blk); 772fc73548eSStefan Hajnoczi 7739ef9d402SStefan Hajnoczi do { 774d0435bc5SStefan Hajnoczi if (suppress_notifications) { 7759ef9d402SStefan Hajnoczi virtio_queue_set_notification(vq, 0); 776d0435bc5SStefan Hajnoczi } 7779ef9d402SStefan Hajnoczi 778edaffd9fSStefan Hajnoczi while ((req = virtio_blk_get_request(s, vq))) { 77907931698SFam Zheng progress = true; 78020ea686aSGreg Kurz if (virtio_blk_handle_request(req, &mrb)) { 78120ea686aSGreg Kurz virtqueue_detach_element(req->vq, &req->elem, 0); 78220ea686aSGreg Kurz virtio_blk_free_request(req); 78320ea686aSGreg Kurz break; 78420ea686aSGreg Kurz } 7856e790746SPaolo Bonzini } 7866e790746SPaolo Bonzini 787d0435bc5SStefan Hajnoczi if (suppress_notifications) { 7889ef9d402SStefan Hajnoczi virtio_queue_set_notification(vq, 1); 789d0435bc5SStefan Hajnoczi } 7909ef9d402SStefan Hajnoczi } while (!virtio_queue_empty(vq)); 7919ef9d402SStefan Hajnoczi 79295f7142aSPeter Lieven if (mrb.num_reqs) { 79395f7142aSPeter Lieven virtio_blk_submit_multireq(s->blk, &mrb); 79495f7142aSPeter Lieven } 795fc73548eSStefan Hajnoczi 796fc73548eSStefan Hajnoczi blk_io_unplug(s->blk); 7979d456654SPaolo Bonzini aio_context_release(blk_get_aio_context(s->blk)); 79807931698SFam Zheng return progress; 79907931698SFam Zheng } 80007931698SFam Zheng 80107931698SFam Zheng static void virtio_blk_handle_output_do(VirtIOBlock *s, VirtQueue *vq) 80207931698SFam Zheng { 80307931698SFam Zheng virtio_blk_handle_vq(s, vq); 8046e790746SPaolo Bonzini } 8056e790746SPaolo Bonzini 8068a2fad57SMichael S. Tsirkin static void virtio_blk_handle_output(VirtIODevice *vdev, VirtQueue *vq) 8078a2fad57SMichael S. Tsirkin { 8088a2fad57SMichael S. Tsirkin VirtIOBlock *s = (VirtIOBlock *)vdev; 8098a2fad57SMichael S. Tsirkin 8108a2fad57SMichael S. Tsirkin if (s->dataplane) { 8118a2fad57SMichael S. Tsirkin /* Some guests kick before setting VIRTIO_CONFIG_S_DRIVER_OK so start 8128a2fad57SMichael S. Tsirkin * dataplane here instead of waiting for .set_status(). 8138a2fad57SMichael S. Tsirkin */ 8149ffe337cSPaolo Bonzini virtio_device_start_ioeventfd(vdev); 8158a2fad57SMichael S. Tsirkin if (!s->dataplane_disabled) { 8168a2fad57SMichael S. Tsirkin return; 8178a2fad57SMichael S. Tsirkin } 8188a2fad57SMichael S. Tsirkin } 81907931698SFam Zheng virtio_blk_handle_output_do(s, vq); 8208a2fad57SMichael S. Tsirkin } 8218a2fad57SMichael S. Tsirkin 8226e790746SPaolo Bonzini static void virtio_blk_dma_restart_bh(void *opaque) 8236e790746SPaolo Bonzini { 8246e790746SPaolo Bonzini VirtIOBlock *s = opaque; 8256e790746SPaolo Bonzini VirtIOBlockReq *req = s->rq; 82695f7142aSPeter Lieven MultiReqBuffer mrb = {}; 8276e790746SPaolo Bonzini 8286e790746SPaolo Bonzini qemu_bh_delete(s->bh); 8296e790746SPaolo Bonzini s->bh = NULL; 8306e790746SPaolo Bonzini 8316e790746SPaolo Bonzini s->rq = NULL; 8326e790746SPaolo Bonzini 8331919631eSPaolo Bonzini aio_context_acquire(blk_get_aio_context(s->conf.conf.blk)); 8346e790746SPaolo Bonzini while (req) { 8351bdb176aSzhanghailiang VirtIOBlockReq *next = req->next; 83620ea686aSGreg Kurz if (virtio_blk_handle_request(req, &mrb)) { 83720ea686aSGreg Kurz /* Device is now broken and won't do any processing until it gets 83820ea686aSGreg Kurz * reset. Already queued requests will be lost: let's purge them. 83920ea686aSGreg Kurz */ 84020ea686aSGreg Kurz while (req) { 84120ea686aSGreg Kurz next = req->next; 84220ea686aSGreg Kurz virtqueue_detach_element(req->vq, &req->elem, 0); 84320ea686aSGreg Kurz virtio_blk_free_request(req); 84420ea686aSGreg Kurz req = next; 84520ea686aSGreg Kurz } 84620ea686aSGreg Kurz break; 84720ea686aSGreg Kurz } 8481bdb176aSzhanghailiang req = next; 8496e790746SPaolo Bonzini } 8506e790746SPaolo Bonzini 85195f7142aSPeter Lieven if (mrb.num_reqs) { 85295f7142aSPeter Lieven virtio_blk_submit_multireq(s->blk, &mrb); 85395f7142aSPeter Lieven } 854680f2002SKevin Wolf blk_dec_in_flight(s->conf.conf.blk); 8551919631eSPaolo Bonzini aio_context_release(blk_get_aio_context(s->conf.conf.blk)); 8566e790746SPaolo Bonzini } 8576e790746SPaolo Bonzini 8586e790746SPaolo Bonzini static void virtio_blk_dma_restart_cb(void *opaque, int running, 8596e790746SPaolo Bonzini RunState state) 8606e790746SPaolo Bonzini { 8616e790746SPaolo Bonzini VirtIOBlock *s = opaque; 8626e790746SPaolo Bonzini 8636e790746SPaolo Bonzini if (!running) { 8646e790746SPaolo Bonzini return; 8656e790746SPaolo Bonzini } 8666e790746SPaolo Bonzini 8676e790746SPaolo Bonzini if (!s->bh) { 868680f2002SKevin Wolf /* FIXME The data plane is not started yet, so these requests are 869680f2002SKevin Wolf * processed in the main thread. */ 8704be74634SMarkus Armbruster s->bh = aio_bh_new(blk_get_aio_context(s->conf.conf.blk), 8714407c1c5SFam Zheng virtio_blk_dma_restart_bh, s); 872680f2002SKevin Wolf blk_inc_in_flight(s->conf.conf.blk); 8736e790746SPaolo Bonzini qemu_bh_schedule(s->bh); 8746e790746SPaolo Bonzini } 8756e790746SPaolo Bonzini } 8766e790746SPaolo Bonzini 8776e790746SPaolo Bonzini static void virtio_blk_reset(VirtIODevice *vdev) 8786e790746SPaolo Bonzini { 8796e790746SPaolo Bonzini VirtIOBlock *s = VIRTIO_BLK(vdev); 8806e40b3bfSAlexander Yarygin AioContext *ctx; 88126307f6aSFam Zheng VirtIOBlockReq *req; 8826e790746SPaolo Bonzini 8836e40b3bfSAlexander Yarygin ctx = blk_get_aio_context(s->blk); 8846e40b3bfSAlexander Yarygin aio_context_acquire(ctx); 8856e40b3bfSAlexander Yarygin blk_drain(s->blk); 8866e40b3bfSAlexander Yarygin 88726307f6aSFam Zheng /* We drop queued requests after blk_drain() because blk_drain() itself can 88826307f6aSFam Zheng * produce them. */ 88926307f6aSFam Zheng while (s->rq) { 89026307f6aSFam Zheng req = s->rq; 89126307f6aSFam Zheng s->rq = req->next; 89297b93c8aSStefan Hajnoczi virtqueue_detach_element(req->vq, &req->elem, 0); 89326307f6aSFam Zheng virtio_blk_free_request(req); 89426307f6aSFam Zheng } 89526307f6aSFam Zheng 8966e40b3bfSAlexander Yarygin aio_context_release(ctx); 8976e40b3bfSAlexander Yarygin 8989ffe337cSPaolo Bonzini assert(!s->dataplane_started); 8994be74634SMarkus Armbruster blk_set_enable_write_cache(s->blk, s->original_wce); 9006e790746SPaolo Bonzini } 9016e790746SPaolo Bonzini 9026e790746SPaolo Bonzini /* coalesce internal state, copy to pci i/o region 0 9036e790746SPaolo Bonzini */ 9046e790746SPaolo Bonzini static void virtio_blk_update_config(VirtIODevice *vdev, uint8_t *config) 9056e790746SPaolo Bonzini { 9066e790746SPaolo Bonzini VirtIOBlock *s = VIRTIO_BLK(vdev); 9072a30307fSMarkus Armbruster BlockConf *conf = &s->conf.conf; 9086e790746SPaolo Bonzini struct virtio_blk_config blkcfg; 9096e790746SPaolo Bonzini uint64_t capacity; 91017d0bc01SStefan Hajnoczi int64_t length; 911f7516731SMarkus Armbruster int blk_size = conf->logical_block_size; 9126e790746SPaolo Bonzini 9134be74634SMarkus Armbruster blk_get_geometry(s->blk, &capacity); 9146e790746SPaolo Bonzini memset(&blkcfg, 0, sizeof(blkcfg)); 915783d1897SRusty Russell virtio_stq_p(vdev, &blkcfg.capacity, capacity); 916*1bf8a989SDenis Plotnikov virtio_stl_p(vdev, &blkcfg.seg_max, 917*1bf8a989SDenis Plotnikov s->conf.seg_max_adjust ? s->conf.queue_size - 2 : 128 - 2); 918907eb3e5SMichael S. Tsirkin virtio_stw_p(vdev, &blkcfg.geometry.cylinders, conf->cyls); 919783d1897SRusty Russell virtio_stl_p(vdev, &blkcfg.blk_size, blk_size); 920f7516731SMarkus Armbruster virtio_stw_p(vdev, &blkcfg.min_io_size, conf->min_io_size / blk_size); 921f7516731SMarkus Armbruster virtio_stw_p(vdev, &blkcfg.opt_io_size, conf->opt_io_size / blk_size); 922907eb3e5SMichael S. Tsirkin blkcfg.geometry.heads = conf->heads; 9236e790746SPaolo Bonzini /* 9246e790746SPaolo Bonzini * We must ensure that the block device capacity is a multiple of 925e03ba136SPeter Maydell * the logical block size. If that is not the case, let's use 9266e790746SPaolo Bonzini * sector_mask to adopt the geometry to have a correct picture. 9276e790746SPaolo Bonzini * For those devices where the capacity is ok for the given geometry 928e03ba136SPeter Maydell * we don't touch the sector value of the geometry, since some devices 9296e790746SPaolo Bonzini * (like s390 dasd) need a specific value. Here the capacity is already 9306e790746SPaolo Bonzini * cyls*heads*secs*blk_size and the sector value is not block size 9316e790746SPaolo Bonzini * divided by 512 - instead it is the amount of blk_size blocks 9326e790746SPaolo Bonzini * per track (cylinder). 9336e790746SPaolo Bonzini */ 93417d0bc01SStefan Hajnoczi length = blk_getlength(s->blk); 93517d0bc01SStefan Hajnoczi if (length > 0 && length / conf->heads / conf->secs % blk_size) { 936907eb3e5SMichael S. Tsirkin blkcfg.geometry.sectors = conf->secs & ~s->sector_mask; 9376e790746SPaolo Bonzini } else { 938907eb3e5SMichael S. Tsirkin blkcfg.geometry.sectors = conf->secs; 9396e790746SPaolo Bonzini } 9406e790746SPaolo Bonzini blkcfg.size_max = 0; 941f7516731SMarkus Armbruster blkcfg.physical_block_exp = get_physical_block_exp(conf); 9426e790746SPaolo Bonzini blkcfg.alignment_offset = 0; 9434be74634SMarkus Armbruster blkcfg.wce = blk_enable_write_cache(s->blk); 9442f270590SStefan Hajnoczi virtio_stw_p(vdev, &blkcfg.num_queues, s->conf.num_queues); 94537b06f8dSStefano Garzarella if (virtio_has_feature(s->host_features, VIRTIO_BLK_F_DISCARD)) { 94637b06f8dSStefano Garzarella virtio_stl_p(vdev, &blkcfg.max_discard_sectors, 94737b06f8dSStefano Garzarella s->conf.max_discard_sectors); 94837b06f8dSStefano Garzarella virtio_stl_p(vdev, &blkcfg.discard_sector_alignment, 94937b06f8dSStefano Garzarella blk_size >> BDRV_SECTOR_BITS); 95037b06f8dSStefano Garzarella /* 95137b06f8dSStefano Garzarella * We support only one segment per request since multiple segments 95237b06f8dSStefano Garzarella * are not widely used and there are no userspace APIs that allow 95337b06f8dSStefano Garzarella * applications to submit multiple segments in a single call. 95437b06f8dSStefano Garzarella */ 95537b06f8dSStefano Garzarella virtio_stl_p(vdev, &blkcfg.max_discard_seg, 1); 95637b06f8dSStefano Garzarella } 95737b06f8dSStefano Garzarella if (virtio_has_feature(s->host_features, VIRTIO_BLK_F_WRITE_ZEROES)) { 95837b06f8dSStefano Garzarella virtio_stl_p(vdev, &blkcfg.max_write_zeroes_sectors, 95937b06f8dSStefano Garzarella s->conf.max_write_zeroes_sectors); 96037b06f8dSStefano Garzarella blkcfg.write_zeroes_may_unmap = 1; 96137b06f8dSStefano Garzarella virtio_stl_p(vdev, &blkcfg.max_write_zeroes_seg, 1); 96237b06f8dSStefano Garzarella } 96320764be0SStefano Garzarella memcpy(config, &blkcfg, s->config_size); 9646e790746SPaolo Bonzini } 9656e790746SPaolo Bonzini 9666e790746SPaolo Bonzini static void virtio_blk_set_config(VirtIODevice *vdev, const uint8_t *config) 9676e790746SPaolo Bonzini { 9686e790746SPaolo Bonzini VirtIOBlock *s = VIRTIO_BLK(vdev); 9696e790746SPaolo Bonzini struct virtio_blk_config blkcfg; 9706e790746SPaolo Bonzini 97120764be0SStefano Garzarella memcpy(&blkcfg, config, s->config_size); 9726d7e73d6SFam Zheng 9734be74634SMarkus Armbruster aio_context_acquire(blk_get_aio_context(s->blk)); 9744be74634SMarkus Armbruster blk_set_enable_write_cache(s->blk, blkcfg.wce != 0); 9754be74634SMarkus Armbruster aio_context_release(blk_get_aio_context(s->blk)); 9766e790746SPaolo Bonzini } 9776e790746SPaolo Bonzini 9789d5b731dSJason Wang static uint64_t virtio_blk_get_features(VirtIODevice *vdev, uint64_t features, 9799d5b731dSJason Wang Error **errp) 9806e790746SPaolo Bonzini { 9816e790746SPaolo Bonzini VirtIOBlock *s = VIRTIO_BLK(vdev); 9826e790746SPaolo Bonzini 983bbe8bd4dSStefano Garzarella /* Firstly sync all virtio-blk possible supported features */ 984bbe8bd4dSStefano Garzarella features |= s->host_features; 985bbe8bd4dSStefano Garzarella 9860cd09c3aSCornelia Huck virtio_add_feature(&features, VIRTIO_BLK_F_SEG_MAX); 9870cd09c3aSCornelia Huck virtio_add_feature(&features, VIRTIO_BLK_F_GEOMETRY); 9880cd09c3aSCornelia Huck virtio_add_feature(&features, VIRTIO_BLK_F_TOPOLOGY); 9890cd09c3aSCornelia Huck virtio_add_feature(&features, VIRTIO_BLK_F_BLK_SIZE); 99095129d6fSCornelia Huck if (virtio_has_feature(features, VIRTIO_F_VERSION_1)) { 991bbe8bd4dSStefano Garzarella if (virtio_has_feature(s->host_features, VIRTIO_BLK_F_SCSI)) { 992efb8206cSJason Wang error_setg(errp, "Please set scsi=off for virtio-blk devices in order to use virtio 1.0"); 993efb8206cSJason Wang return 0; 994efb8206cSJason Wang } 995efb8206cSJason Wang } else { 996c9b11f97SJason Wang virtio_clear_feature(&features, VIRTIO_F_ANY_LAYOUT); 997efb8206cSJason Wang virtio_add_feature(&features, VIRTIO_BLK_F_SCSI); 998efb8206cSJason Wang } 9996e790746SPaolo Bonzini 10005f258577SEvgeny Yakovlev if (blk_enable_write_cache(s->blk) || 10015f258577SEvgeny Yakovlev (s->conf.x_enable_wce_if_config_wce && 10025f258577SEvgeny Yakovlev virtio_has_feature(features, VIRTIO_BLK_F_CONFIG_WCE))) { 10030cd09c3aSCornelia Huck virtio_add_feature(&features, VIRTIO_BLK_F_WCE); 10044be74634SMarkus Armbruster } 10054be74634SMarkus Armbruster if (blk_is_read_only(s->blk)) { 10060cd09c3aSCornelia Huck virtio_add_feature(&features, VIRTIO_BLK_F_RO); 10074be74634SMarkus Armbruster } 10082f270590SStefan Hajnoczi if (s->conf.num_queues > 1) { 10092f270590SStefan Hajnoczi virtio_add_feature(&features, VIRTIO_BLK_F_MQ); 10102f270590SStefan Hajnoczi } 10116e790746SPaolo Bonzini 10126e790746SPaolo Bonzini return features; 10136e790746SPaolo Bonzini } 10146e790746SPaolo Bonzini 10156e790746SPaolo Bonzini static void virtio_blk_set_status(VirtIODevice *vdev, uint8_t status) 10166e790746SPaolo Bonzini { 10176e790746SPaolo Bonzini VirtIOBlock *s = VIRTIO_BLK(vdev); 10186e790746SPaolo Bonzini 10199ffe337cSPaolo Bonzini if (!(status & (VIRTIO_CONFIG_S_DRIVER | VIRTIO_CONFIG_S_DRIVER_OK))) { 10209ffe337cSPaolo Bonzini assert(!s->dataplane_started); 10216e790746SPaolo Bonzini } 10226e790746SPaolo Bonzini 10236e790746SPaolo Bonzini if (!(status & VIRTIO_CONFIG_S_DRIVER_OK)) { 10246e790746SPaolo Bonzini return; 10256e790746SPaolo Bonzini } 10266e790746SPaolo Bonzini 1027ef5bc962SPaolo Bonzini /* A guest that supports VIRTIO_BLK_F_CONFIG_WCE must be able to send 1028ef5bc962SPaolo Bonzini * cache flushes. Thus, the "auto writethrough" behavior is never 1029ef5bc962SPaolo Bonzini * necessary for guests that support the VIRTIO_BLK_F_CONFIG_WCE feature. 1030ef5bc962SPaolo Bonzini * Leaving it enabled would break the following sequence: 1031ef5bc962SPaolo Bonzini * 1032ef5bc962SPaolo Bonzini * Guest started with "-drive cache=writethrough" 1033ef5bc962SPaolo Bonzini * Guest sets status to 0 1034ef5bc962SPaolo Bonzini * Guest sets DRIVER bit in status field 1035ef5bc962SPaolo Bonzini * Guest reads host features (WCE=0, CONFIG_WCE=1) 1036ef5bc962SPaolo Bonzini * Guest writes guest features (WCE=0, CONFIG_WCE=1) 1037ef5bc962SPaolo Bonzini * Guest writes 1 to the WCE configuration field (writeback mode) 1038ef5bc962SPaolo Bonzini * Guest sets DRIVER_OK bit in status field 1039ef5bc962SPaolo Bonzini * 10404be74634SMarkus Armbruster * s->blk would erroneously be placed in writethrough mode. 1041ef5bc962SPaolo Bonzini */ 104295129d6fSCornelia Huck if (!virtio_vdev_has_feature(vdev, VIRTIO_BLK_F_CONFIG_WCE)) { 10434be74634SMarkus Armbruster aio_context_acquire(blk_get_aio_context(s->blk)); 10444be74634SMarkus Armbruster blk_set_enable_write_cache(s->blk, 104595129d6fSCornelia Huck virtio_vdev_has_feature(vdev, 104695129d6fSCornelia Huck VIRTIO_BLK_F_WCE)); 10474be74634SMarkus Armbruster aio_context_release(blk_get_aio_context(s->blk)); 10486e790746SPaolo Bonzini } 1049ef5bc962SPaolo Bonzini } 10506e790746SPaolo Bonzini 1051b2b295a7SGreg Kurz static void virtio_blk_save_device(VirtIODevice *vdev, QEMUFile *f) 1052b2b295a7SGreg Kurz { 1053b2b295a7SGreg Kurz VirtIOBlock *s = VIRTIO_BLK(vdev); 1054b2b295a7SGreg Kurz VirtIOBlockReq *req = s->rq; 10556e790746SPaolo Bonzini 10566e790746SPaolo Bonzini while (req) { 10576e790746SPaolo Bonzini qemu_put_sbyte(f, 1); 105830d8bf6dSStefan Hajnoczi 105930d8bf6dSStefan Hajnoczi if (s->conf.num_queues > 1) { 106030d8bf6dSStefan Hajnoczi qemu_put_be32(f, virtio_get_queue_index(req->vq)); 106130d8bf6dSStefan Hajnoczi } 106230d8bf6dSStefan Hajnoczi 106386044b24SJason Wang qemu_put_virtqueue_element(vdev, f, &req->elem); 10646e790746SPaolo Bonzini req = req->next; 10656e790746SPaolo Bonzini } 10666e790746SPaolo Bonzini qemu_put_sbyte(f, 0); 10676e790746SPaolo Bonzini } 10686e790746SPaolo Bonzini 1069b2b295a7SGreg Kurz static int virtio_blk_load_device(VirtIODevice *vdev, QEMUFile *f, 1070b2b295a7SGreg Kurz int version_id) 1071b2b295a7SGreg Kurz { 1072b2b295a7SGreg Kurz VirtIOBlock *s = VIRTIO_BLK(vdev); 1073b2b295a7SGreg Kurz 10746e790746SPaolo Bonzini while (qemu_get_sbyte(f)) { 107530d8bf6dSStefan Hajnoczi unsigned nvqs = s->conf.num_queues; 107630d8bf6dSStefan Hajnoczi unsigned vq_idx = 0; 1077ab281c17SPaolo Bonzini VirtIOBlockReq *req; 107830d8bf6dSStefan Hajnoczi 107930d8bf6dSStefan Hajnoczi if (nvqs > 1) { 108030d8bf6dSStefan Hajnoczi vq_idx = qemu_get_be32(f); 108130d8bf6dSStefan Hajnoczi 108230d8bf6dSStefan Hajnoczi if (vq_idx >= nvqs) { 108330d8bf6dSStefan Hajnoczi error_report("Invalid virtqueue index in request list: %#x", 108430d8bf6dSStefan Hajnoczi vq_idx); 108530d8bf6dSStefan Hajnoczi return -EINVAL; 108630d8bf6dSStefan Hajnoczi } 108730d8bf6dSStefan Hajnoczi } 108830d8bf6dSStefan Hajnoczi 10898607f5c3SJason Wang req = qemu_get_virtqueue_element(vdev, f, sizeof(VirtIOBlockReq)); 109030d8bf6dSStefan Hajnoczi virtio_blk_init_request(s, virtio_get_queue(vdev, vq_idx), req); 10916e790746SPaolo Bonzini req->next = s->rq; 10926e790746SPaolo Bonzini s->rq = req; 10936e790746SPaolo Bonzini } 10946e790746SPaolo Bonzini 10956e790746SPaolo Bonzini return 0; 10966e790746SPaolo Bonzini } 10976e790746SPaolo Bonzini 10989b92fbcfSSergio Lopez static void virtio_resize_cb(void *opaque) 10999b92fbcfSSergio Lopez { 11009b92fbcfSSergio Lopez VirtIODevice *vdev = opaque; 11019b92fbcfSSergio Lopez 11029b92fbcfSSergio Lopez assert(qemu_get_current_aio_context() == qemu_get_aio_context()); 11039b92fbcfSSergio Lopez virtio_notify_config(vdev); 11049b92fbcfSSergio Lopez } 11059b92fbcfSSergio Lopez 11066e790746SPaolo Bonzini static void virtio_blk_resize(void *opaque) 11076e790746SPaolo Bonzini { 11086e790746SPaolo Bonzini VirtIODevice *vdev = VIRTIO_DEVICE(opaque); 11096e790746SPaolo Bonzini 11109b92fbcfSSergio Lopez /* 11119b92fbcfSSergio Lopez * virtio_notify_config() needs to acquire the global mutex, 11129b92fbcfSSergio Lopez * so it can't be called from an iothread. Instead, schedule 11139b92fbcfSSergio Lopez * it to be run in the main context BH. 11149b92fbcfSSergio Lopez */ 11159b92fbcfSSergio Lopez aio_bh_schedule_oneshot(qemu_get_aio_context(), virtio_resize_cb, vdev); 11166e790746SPaolo Bonzini } 11176e790746SPaolo Bonzini 11186e790746SPaolo Bonzini static const BlockDevOps virtio_block_ops = { 11196e790746SPaolo Bonzini .resize_cb = virtio_blk_resize, 11206e790746SPaolo Bonzini }; 11216e790746SPaolo Bonzini 112275884afdSAndreas Färber static void virtio_blk_device_realize(DeviceState *dev, Error **errp) 11236e790746SPaolo Bonzini { 112475884afdSAndreas Färber VirtIODevice *vdev = VIRTIO_DEVICE(dev); 1125179b417eSAndreas Färber VirtIOBlock *s = VIRTIO_BLK(dev); 11262a30307fSMarkus Armbruster VirtIOBlkConf *conf = &s->conf; 11273ffeeef7SAndreas Färber Error *err = NULL; 11282f270590SStefan Hajnoczi unsigned i; 11296e790746SPaolo Bonzini 11304be74634SMarkus Armbruster if (!conf->conf.blk) { 113175884afdSAndreas Färber error_setg(errp, "drive property not set"); 113275884afdSAndreas Färber return; 11336e790746SPaolo Bonzini } 11344be74634SMarkus Armbruster if (!blk_is_inserted(conf->conf.blk)) { 113575884afdSAndreas Färber error_setg(errp, "Device needs media, but drive is empty"); 113675884afdSAndreas Färber return; 11376e790746SPaolo Bonzini } 11382f270590SStefan Hajnoczi if (!conf->num_queues) { 11392f270590SStefan Hajnoczi error_setg(errp, "num-queues property must be larger than 0"); 11402f270590SStefan Hajnoczi return; 11412f270590SStefan Hajnoczi } 1142*1bf8a989SDenis Plotnikov if (conf->queue_size <= 2) { 1143*1bf8a989SDenis Plotnikov error_setg(errp, "invalid queue-size property (%" PRIu16 "), " 1144*1bf8a989SDenis Plotnikov "must be > 2", conf->queue_size); 1145*1bf8a989SDenis Plotnikov return; 1146*1bf8a989SDenis Plotnikov } 11476040aeddSMark Kanda if (!is_power_of_2(conf->queue_size) || 11486040aeddSMark Kanda conf->queue_size > VIRTQUEUE_MAX_SIZE) { 11496040aeddSMark Kanda error_setg(errp, "invalid queue-size property (%" PRIu16 "), " 11506040aeddSMark Kanda "must be a power of 2 (max %d)", 11516040aeddSMark Kanda conf->queue_size, VIRTQUEUE_MAX_SIZE); 11526040aeddSMark Kanda return; 11536040aeddSMark Kanda } 11546e790746SPaolo Bonzini 1155ceff3e1fSMao Zhongyi if (!blkconf_apply_backend_options(&conf->conf, 1156a17c17a2SKevin Wolf blk_is_read_only(conf->conf.blk), true, 1157ceff3e1fSMao Zhongyi errp)) { 1158a17c17a2SKevin Wolf return; 1159a17c17a2SKevin Wolf } 11604be74634SMarkus Armbruster s->original_wce = blk_enable_write_cache(conf->conf.blk); 1161ceff3e1fSMao Zhongyi if (!blkconf_geometry(&conf->conf, NULL, 65535, 255, 255, errp)) { 116275884afdSAndreas Färber return; 11636e790746SPaolo Bonzini } 1164ceff3e1fSMao Zhongyi 11650eb28a42SEkaterina Tumanova blkconf_blocksizes(&conf->conf); 11666e790746SPaolo Bonzini 11670a75b60cSMark Kanda if (conf->conf.logical_block_size > 11680a75b60cSMark Kanda conf->conf.physical_block_size) { 11690a75b60cSMark Kanda error_setg(errp, 11700a75b60cSMark Kanda "logical_block_size > physical_block_size not supported"); 11710a75b60cSMark Kanda return; 11720a75b60cSMark Kanda } 11730a75b60cSMark Kanda 117437b06f8dSStefano Garzarella if (virtio_has_feature(s->host_features, VIRTIO_BLK_F_DISCARD) && 117537b06f8dSStefano Garzarella (!conf->max_discard_sectors || 117637b06f8dSStefano Garzarella conf->max_discard_sectors > BDRV_REQUEST_MAX_SECTORS)) { 117737b06f8dSStefano Garzarella error_setg(errp, "invalid max-discard-sectors property (%" PRIu32 ")" 117837b06f8dSStefano Garzarella ", must be between 1 and %d", 117937b06f8dSStefano Garzarella conf->max_discard_sectors, (int)BDRV_REQUEST_MAX_SECTORS); 118037b06f8dSStefano Garzarella return; 118137b06f8dSStefano Garzarella } 118237b06f8dSStefano Garzarella 118337b06f8dSStefano Garzarella if (virtio_has_feature(s->host_features, VIRTIO_BLK_F_WRITE_ZEROES) && 118437b06f8dSStefano Garzarella (!conf->max_write_zeroes_sectors || 118537b06f8dSStefano Garzarella conf->max_write_zeroes_sectors > BDRV_REQUEST_MAX_SECTORS)) { 118637b06f8dSStefano Garzarella error_setg(errp, "invalid max-write-zeroes-sectors property (%" PRIu32 118737b06f8dSStefano Garzarella "), must be between 1 and %d", 118837b06f8dSStefano Garzarella conf->max_write_zeroes_sectors, 118937b06f8dSStefano Garzarella (int)BDRV_REQUEST_MAX_SECTORS); 119037b06f8dSStefano Garzarella return; 119137b06f8dSStefano Garzarella } 119237b06f8dSStefano Garzarella 119320764be0SStefano Garzarella virtio_blk_set_config_size(s, s->host_features); 119420764be0SStefano Garzarella 119520764be0SStefano Garzarella virtio_init(vdev, "virtio-blk", VIRTIO_ID_BLOCK, s->config_size); 11966e790746SPaolo Bonzini 11974be74634SMarkus Armbruster s->blk = conf->conf.blk; 11986e790746SPaolo Bonzini s->rq = NULL; 11992a30307fSMarkus Armbruster s->sector_mask = (s->conf.conf.logical_block_size / BDRV_SECTOR_SIZE) - 1; 12006e790746SPaolo Bonzini 12012f270590SStefan Hajnoczi for (i = 0; i < conf->num_queues; i++) { 12026040aeddSMark Kanda virtio_add_queue(vdev, conf->queue_size, virtio_blk_handle_output); 12032f270590SStefan Hajnoczi } 12042a30307fSMarkus Armbruster virtio_blk_data_plane_create(vdev, conf, &s->dataplane, &err); 12053ffeeef7SAndreas Färber if (err != NULL) { 120675884afdSAndreas Färber error_propagate(errp, err); 12076a1a8cc7SKONRAD Frederic virtio_cleanup(vdev); 120875884afdSAndreas Färber return; 12096e790746SPaolo Bonzini } 12106e790746SPaolo Bonzini 12116e790746SPaolo Bonzini s->change = qemu_add_vm_change_state_handler(virtio_blk_dma_restart_cb, s); 12124be74634SMarkus Armbruster blk_set_dev_ops(s->blk, &virtio_block_ops, s); 12134be74634SMarkus Armbruster blk_set_guest_block_size(s->blk, s->conf.conf.logical_block_size); 12146e790746SPaolo Bonzini 12154be74634SMarkus Armbruster blk_iostatus_enable(s->blk); 121671f571a2SSam Eiderman 121771f571a2SSam Eiderman add_boot_device_lchs(dev, "/disk@0,0", 121871f571a2SSam Eiderman conf->conf.lcyls, 121971f571a2SSam Eiderman conf->conf.lheads, 122071f571a2SSam Eiderman conf->conf.lsecs); 12216e790746SPaolo Bonzini } 12226e790746SPaolo Bonzini 1223306ec6c3SAndreas Färber static void virtio_blk_device_unrealize(DeviceState *dev, Error **errp) 12246e790746SPaolo Bonzini { 1225306ec6c3SAndreas Färber VirtIODevice *vdev = VIRTIO_DEVICE(dev); 1226306ec6c3SAndreas Färber VirtIOBlock *s = VIRTIO_BLK(dev); 12274a0117cfSEugenio Pérez VirtIOBlkConf *conf = &s->conf; 12284a0117cfSEugenio Pérez unsigned i; 1229306ec6c3SAndreas Färber 12307bfde688SJulia Suvorova blk_drain(s->blk); 123171f571a2SSam Eiderman del_boot_device_lchs(dev, "/disk@0,0"); 12326e790746SPaolo Bonzini virtio_blk_data_plane_destroy(s->dataplane); 12336e790746SPaolo Bonzini s->dataplane = NULL; 12344a0117cfSEugenio Pérez for (i = 0; i < conf->num_queues; i++) { 12354a0117cfSEugenio Pérez virtio_del_queue(vdev, i); 12364a0117cfSEugenio Pérez } 12376e790746SPaolo Bonzini qemu_del_vm_change_state_handler(s->change); 12384be74634SMarkus Armbruster blockdev_mark_auto_del(s->blk); 12396a1a8cc7SKONRAD Frederic virtio_cleanup(vdev); 12406e790746SPaolo Bonzini } 12416e790746SPaolo Bonzini 1242467b3f33SStefan Hajnoczi static void virtio_blk_instance_init(Object *obj) 1243467b3f33SStefan Hajnoczi { 1244467b3f33SStefan Hajnoczi VirtIOBlock *s = VIRTIO_BLK(obj); 1245467b3f33SStefan Hajnoczi 12462a30307fSMarkus Armbruster device_add_bootindex_property(obj, &s->conf.conf.bootindex, 12473342ec32SGonglei "bootindex", "/disk@0,0", 12483342ec32SGonglei DEVICE(obj), NULL); 1249467b3f33SStefan Hajnoczi } 1250467b3f33SStefan Hajnoczi 1251977a117fSHalil Pasic static const VMStateDescription vmstate_virtio_blk = { 1252977a117fSHalil Pasic .name = "virtio-blk", 1253977a117fSHalil Pasic .minimum_version_id = 2, 1254977a117fSHalil Pasic .version_id = 2, 1255977a117fSHalil Pasic .fields = (VMStateField[]) { 1256977a117fSHalil Pasic VMSTATE_VIRTIO_DEVICE, 1257977a117fSHalil Pasic VMSTATE_END_OF_LIST() 1258977a117fSHalil Pasic }, 1259977a117fSHalil Pasic }; 1260bbded32cSDr. David Alan Gilbert 12616e790746SPaolo Bonzini static Property virtio_blk_properties[] = { 12622a30307fSMarkus Armbruster DEFINE_BLOCK_PROPERTIES(VirtIOBlock, conf.conf), 12638c398252SKevin Wolf DEFINE_BLOCK_ERROR_PROPERTIES(VirtIOBlock, conf.conf), 12642a30307fSMarkus Armbruster DEFINE_BLOCK_CHS_PROPERTIES(VirtIOBlock, conf.conf), 12652a30307fSMarkus Armbruster DEFINE_PROP_STRING("serial", VirtIOBlock, conf.serial), 1266bbe8bd4dSStefano Garzarella DEFINE_PROP_BIT64("config-wce", VirtIOBlock, host_features, 1267bbe8bd4dSStefano Garzarella VIRTIO_BLK_F_CONFIG_WCE, true), 126832a877e4SStefan Hajnoczi #ifdef __linux__ 1269bbe8bd4dSStefano Garzarella DEFINE_PROP_BIT64("scsi", VirtIOBlock, host_features, 1270bbe8bd4dSStefano Garzarella VIRTIO_BLK_F_SCSI, false), 127132a877e4SStefan Hajnoczi #endif 1272c99495acSPeter Lieven DEFINE_PROP_BIT("request-merging", VirtIOBlock, conf.request_merging, 0, 1273c99495acSPeter Lieven true), 12742f270590SStefan Hajnoczi DEFINE_PROP_UINT16("num-queues", VirtIOBlock, conf.num_queues, 1), 12756040aeddSMark Kanda DEFINE_PROP_UINT16("queue-size", VirtIOBlock, conf.queue_size, 128), 1276*1bf8a989SDenis Plotnikov DEFINE_PROP_BOOL("seg-max-adjust", VirtIOBlock, conf.seg_max_adjust, true), 1277d679ac09SFam Zheng DEFINE_PROP_LINK("iothread", VirtIOBlock, conf.iothread, TYPE_IOTHREAD, 1278d679ac09SFam Zheng IOThread *), 12795c81161fSStefano Garzarella DEFINE_PROP_BIT64("discard", VirtIOBlock, host_features, 12805c81161fSStefano Garzarella VIRTIO_BLK_F_DISCARD, true), 12815c81161fSStefano Garzarella DEFINE_PROP_BIT64("write-zeroes", VirtIOBlock, host_features, 12825c81161fSStefano Garzarella VIRTIO_BLK_F_WRITE_ZEROES, true), 128337b06f8dSStefano Garzarella DEFINE_PROP_UINT32("max-discard-sectors", VirtIOBlock, 128437b06f8dSStefano Garzarella conf.max_discard_sectors, BDRV_REQUEST_MAX_SECTORS), 128537b06f8dSStefano Garzarella DEFINE_PROP_UINT32("max-write-zeroes-sectors", VirtIOBlock, 128637b06f8dSStefano Garzarella conf.max_write_zeroes_sectors, BDRV_REQUEST_MAX_SECTORS), 12875f258577SEvgeny Yakovlev DEFINE_PROP_BOOL("x-enable-wce-if-config-wce", VirtIOBlock, 12885f258577SEvgeny Yakovlev conf.x_enable_wce_if_config_wce, true), 12896e790746SPaolo Bonzini DEFINE_PROP_END_OF_LIST(), 12906e790746SPaolo Bonzini }; 12916e790746SPaolo Bonzini 12926e790746SPaolo Bonzini static void virtio_blk_class_init(ObjectClass *klass, void *data) 12936e790746SPaolo Bonzini { 12946e790746SPaolo Bonzini DeviceClass *dc = DEVICE_CLASS(klass); 12956e790746SPaolo Bonzini VirtioDeviceClass *vdc = VIRTIO_DEVICE_CLASS(klass); 129675884afdSAndreas Färber 12976e790746SPaolo Bonzini dc->props = virtio_blk_properties; 1298bbded32cSDr. David Alan Gilbert dc->vmsd = &vmstate_virtio_blk; 1299125ee0edSMarcel Apfelbaum set_bit(DEVICE_CATEGORY_STORAGE, dc->categories); 130075884afdSAndreas Färber vdc->realize = virtio_blk_device_realize; 1301306ec6c3SAndreas Färber vdc->unrealize = virtio_blk_device_unrealize; 13026e790746SPaolo Bonzini vdc->get_config = virtio_blk_update_config; 13036e790746SPaolo Bonzini vdc->set_config = virtio_blk_set_config; 13046e790746SPaolo Bonzini vdc->get_features = virtio_blk_get_features; 13056e790746SPaolo Bonzini vdc->set_status = virtio_blk_set_status; 13066e790746SPaolo Bonzini vdc->reset = virtio_blk_reset; 1307b2b295a7SGreg Kurz vdc->save = virtio_blk_save_device; 1308b2b295a7SGreg Kurz vdc->load = virtio_blk_load_device; 13099ffe337cSPaolo Bonzini vdc->start_ioeventfd = virtio_blk_data_plane_start; 13109ffe337cSPaolo Bonzini vdc->stop_ioeventfd = virtio_blk_data_plane_stop; 13116e790746SPaolo Bonzini } 13126e790746SPaolo Bonzini 1313b5c7ceafSChanglong Xie static const TypeInfo virtio_blk_info = { 13146e790746SPaolo Bonzini .name = TYPE_VIRTIO_BLK, 13156e790746SPaolo Bonzini .parent = TYPE_VIRTIO_DEVICE, 13166e790746SPaolo Bonzini .instance_size = sizeof(VirtIOBlock), 1317467b3f33SStefan Hajnoczi .instance_init = virtio_blk_instance_init, 13186e790746SPaolo Bonzini .class_init = virtio_blk_class_init, 13196e790746SPaolo Bonzini }; 13206e790746SPaolo Bonzini 13216e790746SPaolo Bonzini static void virtio_register_types(void) 13226e790746SPaolo Bonzini { 1323b5c7ceafSChanglong Xie type_register_static(&virtio_blk_info); 13246e790746SPaolo Bonzini } 13256e790746SPaolo Bonzini 13266e790746SPaolo Bonzini type_init(virtio_register_types) 1327