On Sun, Aug 18, 2013 at 11:48:56AM +0300, Dan Aloni wrote:
> On 3.10.7 and x86_64, as a result of sizeof(struct vhost_scsi) being
> 2152960 bytes the allocation failed once on my development machine.
>
> Saw it would be prudent to split the bulk of it, which is the vqs array
> into separately allocated parts. sizeof(struct vhost_virtqueue) is
> currently 16816 bytes.
>
> Signed-off-by: Dan Aloni <[email protected]>
This extra indirection is likely to have measureable cost though.
net core saw a similar problem, it was fixed in patch
net: allow large number of tx queues
So let's do it in a similar way: try to allocate with
GFP_KERNEL | __GFP_NOWARN | __GFP_REPEAT
and if that fails, do vmalloc.
To free, we can do
if (is_vmalloc_addr())
vfree();
else
kfree();
> ---
> drivers/vhost/scsi.c | 68
> +++++++++++++++++++++++++++++++++-------------------
> 1 file changed, 44 insertions(+), 24 deletions(-)
>
> diff --git a/drivers/vhost/scsi.c b/drivers/vhost/scsi.c
> index 7014202..7f605b6 100644
> --- a/drivers/vhost/scsi.c
> +++ b/drivers/vhost/scsi.c
> @@ -171,6 +171,7 @@ enum {
> #define VHOST_SCSI_MAX_EVENT 128
>
> struct vhost_scsi_virtqueue {
> + int index;
> struct vhost_virtqueue vq;
> /*
> * Reference counting for inflight reqs, used for flush operation. At
> @@ -191,7 +192,7 @@ struct vhost_scsi {
> char vs_vhost_wwpn[TRANSPORT_IQN_LEN];
>
> struct vhost_dev dev;
> - struct vhost_scsi_virtqueue vqs[VHOST_SCSI_MAX_VQ];
> + struct vhost_scsi_virtqueue *vqs[VHOST_SCSI_MAX_VQ];
>
> struct vhost_work vs_completion_work; /* cmd completion work item */
> struct llist_head vs_completion_list; /* cmd completion queue */
> @@ -234,18 +235,18 @@ static void tcm_vhost_init_inflight(struct vhost_scsi
> *vs,
> int idx, i;
>
> for (i = 0; i < VHOST_SCSI_MAX_VQ; i++) {
> - vq = &vs->vqs[i].vq;
> + vq = &vs->vqs[i]->vq;
>
> mutex_lock(&vq->mutex);
>
> /* store old infight */
> - idx = vs->vqs[i].inflight_idx;
> + idx = vs->vqs[i]->inflight_idx;
> if (old_inflight)
> - old_inflight[i] = &vs->vqs[i].inflights[idx];
> + old_inflight[i] = &vs->vqs[i]->inflights[idx];
>
> /* setup new infight */
> - vs->vqs[i].inflight_idx = idx ^ 1;
> - new_inflight = &vs->vqs[i].inflights[idx ^ 1];
> + vs->vqs[i]->inflight_idx = idx ^ 1;
> + new_inflight = &vs->vqs[i]->inflights[idx ^ 1];
> kref_init(&new_inflight->kref);
> init_completion(&new_inflight->comp);
>
> @@ -530,7 +531,7 @@ static void tcm_vhost_free_evt(struct vhost_scsi *vs,
> struct tcm_vhost_evt *evt)
> static struct tcm_vhost_evt *tcm_vhost_allocate_evt(struct vhost_scsi *vs,
> u32 event, u32 reason)
> {
> - struct vhost_virtqueue *vq = &vs->vqs[VHOST_SCSI_VQ_EVT].vq;
> + struct vhost_virtqueue *vq = &vs->vqs[VHOST_SCSI_VQ_EVT]->vq;
> struct tcm_vhost_evt *evt;
>
> if (vs->vs_events_nr > VHOST_SCSI_MAX_EVENT) {
> @@ -575,7 +576,7 @@ static void vhost_scsi_free_cmd(struct tcm_vhost_cmd
> *tv_cmd)
> static void tcm_vhost_do_evt_work(struct vhost_scsi *vs,
> struct tcm_vhost_evt *evt)
> {
> - struct vhost_virtqueue *vq = &vs->vqs[VHOST_SCSI_VQ_EVT].vq;
> + struct vhost_virtqueue *vq = &vs->vqs[VHOST_SCSI_VQ_EVT]->vq;
> struct virtio_scsi_event *event = &evt->event;
> struct virtio_scsi_event __user *eventp;
> unsigned out, in;
> @@ -626,7 +627,7 @@ static void tcm_vhost_evt_work(struct vhost_work *work)
> {
> struct vhost_scsi *vs = container_of(work, struct vhost_scsi,
> vs_event_work);
> - struct vhost_virtqueue *vq = &vs->vqs[VHOST_SCSI_VQ_EVT].vq;
> + struct vhost_virtqueue *vq = &vs->vqs[VHOST_SCSI_VQ_EVT]->vq;
> struct tcm_vhost_evt *evt;
> struct llist_node *llnode;
>
> @@ -680,7 +681,7 @@ static void vhost_scsi_complete_cmd_work(struct
> vhost_work *work)
> struct vhost_scsi_virtqueue *q;
> vhost_add_used(tv_cmd->tvc_vq, tv_cmd->tvc_vq_desc, 0);
> q = container_of(tv_cmd->tvc_vq, struct
> vhost_scsi_virtqueue, vq);
> - vq = q - vs->vqs;
> + vq = q->index;
> __set_bit(vq, signal);
> } else
> pr_err("Faulted on virtio_scsi_cmd_resp\n");
> @@ -691,7 +692,7 @@ static void vhost_scsi_complete_cmd_work(struct
> vhost_work *work)
> vq = -1;
> while ((vq = find_next_bit(signal, VHOST_SCSI_MAX_VQ, vq + 1))
> < VHOST_SCSI_MAX_VQ)
> - vhost_signal(&vs->dev, &vs->vqs[vq].vq);
> + vhost_signal(&vs->dev, &vs->vqs[vq]->vq);
> }
>
> static struct tcm_vhost_cmd *vhost_scsi_allocate_cmd(
> @@ -1108,7 +1109,7 @@ static void vhost_scsi_handle_kick(struct vhost_work
> *work)
>
> static void vhost_scsi_flush_vq(struct vhost_scsi *vs, int index)
> {
> - vhost_poll_flush(&vs->vqs[index].vq.poll);
> + vhost_poll_flush(&vs->vqs[index]->vq.poll);
> }
>
> /* Callers must hold dev mutex */
> @@ -1163,7 +1164,7 @@ static int vhost_scsi_set_endpoint(
> /* Verify that ring has been setup correctly. */
> for (index = 0; index < vs->dev.nvqs; ++index) {
> /* Verify that ring has been setup correctly. */
> - if (!vhost_vq_access_ok(&vs->vqs[index].vq)) {
> + if (!vhost_vq_access_ok(&vs->vqs[index]->vq)) {
> ret = -EFAULT;
> goto out;
> }
> @@ -1210,7 +1211,7 @@ static int vhost_scsi_set_endpoint(
> memcpy(vs->vs_vhost_wwpn, t->vhost_wwpn,
> sizeof(vs->vs_vhost_wwpn));
> for (i = 0; i < VHOST_SCSI_MAX_VQ; i++) {
> - vq = &vs->vqs[i].vq;
> + vq = &vs->vqs[i]->vq;
> /* Flushing the vhost_work acts as synchronize_rcu */
> mutex_lock(&vq->mutex);
> rcu_assign_pointer(vq->private_data, vs_tpg);
> @@ -1251,7 +1252,7 @@ static int vhost_scsi_clear_endpoint(
> mutex_lock(&vs->dev.mutex);
> /* Verify that ring has been setup correctly. */
> for (index = 0; index < vs->dev.nvqs; ++index) {
> - if (!vhost_vq_access_ok(&vs->vqs[index].vq)) {
> + if (!vhost_vq_access_ok(&vs->vqs[index]->vq)) {
> ret = -EFAULT;
> goto err_dev;
> }
> @@ -1291,7 +1292,7 @@ static int vhost_scsi_clear_endpoint(
> }
> if (match) {
> for (i = 0; i < VHOST_SCSI_MAX_VQ; i++) {
> - vq = &vs->vqs[i].vq;
> + vq = &vs->vqs[i]->vq;
> /* Flushing the vhost_work acts as synchronize_rcu */
> mutex_lock(&vq->mutex);
> rcu_assign_pointer(vq->private_data, NULL);
> @@ -1346,6 +1347,18 @@ static int vhost_scsi_open(struct inode *inode, struct
> file *f)
> if (!s)
> return -ENOMEM;
>
> + for (i = 0; i < VHOST_SCSI_MAX_VQ; i++) {
> + s->vqs[i] = kzalloc(sizeof(*s->vqs[0]), GFP_KERNEL);;
> + if (!s->vqs[i]) {
> + while (i >= 1)
> + kfree(s->vqs[--i]);
> + kfree(s);
> + return -ENOMEM;
> + }
> +
> + s->vqs[i]->index = i;
> + }
> +
> vqs = kmalloc(VHOST_SCSI_MAX_VQ * sizeof(*vqs), GFP_KERNEL);
> if (!vqs) {
> kfree(s);
> @@ -1358,19 +1371,22 @@ static int vhost_scsi_open(struct inode *inode,
> struct file *f)
> s->vs_events_nr = 0;
> s->vs_events_missed = false;
>
> - vqs[VHOST_SCSI_VQ_CTL] = &s->vqs[VHOST_SCSI_VQ_CTL].vq;
> - vqs[VHOST_SCSI_VQ_EVT] = &s->vqs[VHOST_SCSI_VQ_EVT].vq;
> - s->vqs[VHOST_SCSI_VQ_CTL].vq.handle_kick = vhost_scsi_ctl_handle_kick;
> - s->vqs[VHOST_SCSI_VQ_EVT].vq.handle_kick = vhost_scsi_evt_handle_kick;
> + vqs[VHOST_SCSI_VQ_CTL] = &s->vqs[VHOST_SCSI_VQ_CTL]->vq;
> + vqs[VHOST_SCSI_VQ_EVT] = &s->vqs[VHOST_SCSI_VQ_EVT]->vq;
> + s->vqs[VHOST_SCSI_VQ_CTL]->vq.handle_kick = vhost_scsi_ctl_handle_kick;
> + s->vqs[VHOST_SCSI_VQ_EVT]->vq.handle_kick = vhost_scsi_evt_handle_kick;
> for (i = VHOST_SCSI_VQ_IO; i < VHOST_SCSI_MAX_VQ; i++) {
> - vqs[i] = &s->vqs[i].vq;
> - s->vqs[i].vq.handle_kick = vhost_scsi_handle_kick;
> + vqs[i] = &s->vqs[i]->vq;
> + s->vqs[i]->vq.handle_kick = vhost_scsi_handle_kick;
> }
> r = vhost_dev_init(&s->dev, vqs, VHOST_SCSI_MAX_VQ);
>
> tcm_vhost_init_inflight(s, NULL);
>
> if (r < 0) {
> + for (i = 0; i < VHOST_SCSI_MAX_VQ; i++) {
> + kfree(s->vqs[i]);
> + }
> kfree(vqs);
> kfree(s);
> return r;
> @@ -1384,6 +1400,7 @@ static int vhost_scsi_release(struct inode *inode,
> struct file *f)
> {
> struct vhost_scsi *s = f->private_data;
> struct vhost_scsi_target t;
> + int i;
>
> mutex_lock(&s->dev.mutex);
> memcpy(t.vhost_wwpn, s->vs_vhost_wwpn, sizeof(t.vhost_wwpn));
> @@ -1393,6 +1410,9 @@ static int vhost_scsi_release(struct inode *inode,
> struct file *f)
> vhost_dev_cleanup(&s->dev, false);
> /* Jobs can re-queue themselves in evt kick handler. Do extra flush. */
> vhost_scsi_flush(s);
> + for (i = 0; i < VHOST_SCSI_MAX_VQ; i++) {
> + kfree(s->vqs[i]);
> + }
> kfree(s->dev.vqs);
> kfree(s);
> return 0;
> @@ -1409,7 +1429,7 @@ static long vhost_scsi_ioctl(struct file *f, unsigned
> int ioctl,
> u32 events_missed;
> u64 features;
> int r, abi_version = VHOST_SCSI_ABI_VERSION;
> - struct vhost_virtqueue *vq = &vs->vqs[VHOST_SCSI_VQ_EVT].vq;
> + struct vhost_virtqueue *vq = &vs->vqs[VHOST_SCSI_VQ_EVT]->vq;
>
> switch (ioctl) {
> case VHOST_SCSI_SET_ENDPOINT:
> @@ -1537,7 +1557,7 @@ static void tcm_vhost_do_plug(struct tcm_vhost_tpg *tpg,
> else
> reason = VIRTIO_SCSI_EVT_RESET_REMOVED;
>
> - vq = &vs->vqs[VHOST_SCSI_VQ_EVT].vq;
> + vq = &vs->vqs[VHOST_SCSI_VQ_EVT]->vq;
> mutex_lock(&vq->mutex);
> tcm_vhost_send_evt(vs, tpg, lun,
> VIRTIO_SCSI_T_TRANSPORT_RESET, reason);
> --
> 1.8.1.4
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
