Re: [Xen-devel] [PATCH RESEND 3/4] qdisk, hw/block/xen_disk: Perform grant copy instead of grant map.

[David Vrabel <dvrabel@xxxxxxxxxx>]

On 31/05/2016 05:44, Paulina Szubarczyk wrote:
> Grant copy operation is divided into two phases different for
> 'read' and 'write' operation.
> 
> For a 'read' operation the flow is as follow:
>     1. allocate local buffers for all the segments contained in
>        a request.

Allocating buffers page-by-page looks sub-optimal to me.  Why not
allocate one large buffer for the whole request?

>     2. fill the request io vectors with the buffers' addresses
>     3. invoke read operation by qemu device
>     4. in the completition call grant copy
>     5. free the buffers
> 
> Function 'ioreq_read_init' implements 1. and 2. step. It is called
> instead of 'ioreq_map' in 'ioreq_runio_qemu_aio'. Then the function
> 'ioreq_runio_qemu_aio' continues withouth changes performing step 3.
> Steps 4. and 5. are called in the callback function
> 'qemu_aio_complete'. The ioreq_read' function is implemented for
> step 4 which calls the new function 'xc_gnttab_copy_grant' presented
> in the other part of the patch.
> 
> For a 'write' operation steps 4. happens before step 2.. First data
> are copied from calling guest domains and then qemu operates on
> them.
> ---
>  hw/block/xen_disk.c | 185 
> ++++++++++++++++++++++++++++++++++++++++++++++++++++
>  1 file changed, 185 insertions(+)
> 
> diff --git a/hw/block/xen_disk.c b/hw/block/xen_disk.c
> index 3b7882e..43cd9c9 100644
> --- a/hw/block/xen_disk.c
> +++ b/hw/block/xen_disk.c
> @@ -284,6 +284,154 @@ err:
>      return -1;
>  }
>  
> +
> +static void* get_buffer(void) {
> +    void *buf;
> +
> +    buf = mmap(NULL, 1 << XC_PAGE_SHIFT, PROT_READ | PROT_WRITE, 
> +               MAP_SHARED | MAP_ANONYMOUS, -1, 0);
> +
> +    if (unlikely(buf == MAP_FAILED))
> +        return NULL;
> +
> +    return buf;
> +}
> +
> +static int free_buffer(void* buf) {
> +    return munmap(buf, 1 << XC_PAGE_SHIFT);

I would make this void and assert() the munmap is successful since if
buf is valid the munmap() cannot fail.  This means...

> +}
> +
> +static int free_buffers(void** page, int count) 

This can be void and...

> +{
> +    int i, r = 0;
> +
> +    for (i = 0; i < count; i++) { 
> +        
> +        if(free_buffer(page[i])) 
> +            r = 1;
> +        
> +        page[i] = NULL;
> +    }
> +
> +    return r;
> +}
> +
> +static int ioreq_write(struct ioreq *ioreq) 
> +{
> +    XenGnttab gnt = ioreq->blkdev->xendev.gnttabdev;
> +    uint16_t domids[BLKIF_MAX_SEGMENTS_PER_REQUEST];
> +    uint32_t refs[BLKIF_MAX_SEGMENTS_PER_REQUEST];
> +    uint32_t offset[BLKIF_MAX_SEGMENTS_PER_REQUEST];
> +    uint32_t len[BLKIF_MAX_SEGMENTS_PER_REQUEST];
> +    void *pages[BLKIF_MAX_SEGMENTS_PER_REQUEST];
> +    int i, count = 0, rc, r;
> +    /* Copy the data for write operation from guest grant pages addressed by 
> +     * domids, refs, offset, len to local buffers.
> +     * 
> +     * Bufferes are then mapped to the pending request for further 
> +     * completition.
> +     */
> +
> +    if (ioreq->v.niov == 0) {
> +        r = 0; goto out;
> +    }
> +
> +    count = ioreq->v.niov;
> +    for (i = 0; i < count; i++) {
> +        domids[i] = ioreq->domids[i];
> +        refs[i]   = ioreq->refs[i];
> +        offset[i] = ioreq->req.seg[i].first_sect * ioreq->blkdev->file_blk;
> +        len[i] = (ioreq->req.seg[i].last_sect - ioreq->req.seg[i].first_sect 
> + 1) 
> +                  * ioreq->blkdev->file_blk;
> +        pages[i]  = get_buffer();
> +
> +        if(!pages[i]) {
> +            xen_be_printf(&ioreq->blkdev->xendev, 0, 
> +                          "failed to alloc page, errno %d \n", errno);
> +            r = 1; goto out;
> +        }
> +    }
> +    rc = xc_gnttab_copy_grant(gnt, count, domids, refs, pages, offset, len, 
> 1);
> +
> +    if(rc) {
> +        xen_be_printf(&ioreq->blkdev->xendev, 0, 
> +                      "failed to copy data for write %d \n", rc);
> +
> +        if(free_buffers(ioreq->page, ioreq->v.niov)) {
> +            xen_be_printf(&ioreq->blkdev->xendev, 0, 
> +                          "failed to free page, errno %d \n", errno);
> +        }
> +        r = 1; goto out;
> +    }
> +
> +    for (i = 0; i < ioreq->v.niov; i++) {
> +        ioreq->page[i] = pages[i];
> +        ioreq->v.iov[i].iov_base += (uintptr_t)pages[i];
> +    }
> +
> +    r = 0;
> +out:
> +    return r;
> +}
> +
> +static int ioreq_read_init(struct ioreq *ioreq) 
> +{
> +    int i;
> +
> +    if (ioreq->v.niov == 0) {
> +        return 0;
> +    }
> +
> +    for (i = 0; i < ioreq->v.niov; i++) {
> +        ioreq->page[i] = get_buffer();
> +        if(!ioreq->page[i]) {
> +            return -1;
> +        }
> +        ioreq->v.iov[i].iov_base += (uintptr_t)ioreq->page[i];
> +    }
> +
> +    return 0;
> +}
> +
> +static int ioreq_read(struct ioreq *ioreq) 
> +{
> +    XenGnttab gnt = ioreq->blkdev->xendev.gnttabdev;
> +    uint16_t domids[BLKIF_MAX_SEGMENTS_PER_REQUEST];
> +    uint32_t refs[BLKIF_MAX_SEGMENTS_PER_REQUEST];
> +    uint32_t offset[BLKIF_MAX_SEGMENTS_PER_REQUEST];
> +    uint32_t len[BLKIF_MAX_SEGMENTS_PER_REQUEST];
> +    void *pages[BLKIF_MAX_SEGMENTS_PER_REQUEST];
> +    int i, count = 0, rc;
> +
> +    /* Copy the data from local buffers to guest grant pages addressed by 
> +     * domids, refs, offset on the completition of read operation.
> +     */
> +
> +    if (ioreq->v.niov == 0) {
> +        return 0;
> +    }
> +
> +    count = ioreq->v.niov;
> +    for (i = 0; i < count; i++) {
> +        domids[i] = ioreq->domids[i];
> +        refs[i]   = ioreq->refs[i];
> +        offset[i] = ioreq->req.seg[i].first_sect * ioreq->blkdev->file_blk;
> +        len[i] = (ioreq->req.seg[i].last_sect - ioreq->req.seg[i].first_sect 
> + 1) 
> +                 * ioreq->blkdev->file_blk;
> +        pages[i]   = ioreq->v.iov[i].iov_base;
> +    }

You can build the ops for read/write at the same time using the same
code as the only difference is the direction.

> +
> +    rc = xc_gnttab_copy_grant(gnt, count, domids, refs, pages, offset, len, 
> 0);
> +    
> +    if(rc) {
> +        xen_be_printf(&ioreq->blkdev->xendev, 0, 
> +                      "failed to copy data to guest %d \n", rc);
> +        return -1;
> +    }
> +
> +    return 0;
> +}
> +
>  static int ioreq_runio_qemu_aio(struct ioreq *ioreq);
>  
>  static void qemu_aio_complete(void *opaque, int ret)
> @@ -313,6 +461,22 @@ static void qemu_aio_complete(void *opaque, int ret)
>      }
>  
>      ioreq->status = ioreq->aio_errors ? BLKIF_RSP_ERROR : BLKIF_RSP_OKAY;
> +    
> +    switch(ioreq->req.operation) {
> +    case BLKIF_OP_READ: 
> +        if(ioreq_read(ioreq)) {
> +            xen_be_printf(&ioreq->blkdev->xendev, 0, 
> +                          "failed to copy read data to guest\n");

You need to report the failure back to the frontend.

> +        }

Need a comment here since you're deliberating missing the "break".

> +    case BLKIF_OP_WRITE:
> +        if(free_buffers(ioreq->page, ioreq->v.niov)) {

...you don't need to consider errors here (see comment on free_buffer()
above).

> +            xen_be_printf(&ioreq->blkdev->xendev, 0, 
> +                          "failed to free page, errno %d \n", errno);
> +        }
> +        break;
> +    default:
> +        break;
> +    }
>  
>      ioreq_finish(ioreq);
>      switch (ioreq->req.operation) {
> @@ -335,6 +499,27 @@ static int ioreq_runio_qemu_aio(struct ioreq *ioreq)
>  {
>      struct XenBlkDev *blkdev = ioreq->blkdev;
>  
> +    switch (ioreq->req.operation) {
> +    case BLKIF_OP_READ:
> +        if (ioreq_read_init(ioreq)) {
> +            xen_be_printf(&ioreq->blkdev->xendev, 0, 
> +                          "failed to initialize buffers for"
> +                          "copy data to guest %d \n", errno);
> +            goto out;
> +        }
> +            break;
> +    case BLKIF_OP_WRITE:
> +    case BLKIF_OP_FLUSH_DISKCACHE:
> +        if (ioreq_write(ioreq)) {
> +            xen_be_printf(&ioreq->blkdev->xendev, 0, 
> +                          "failed to write data from guest\n");
> +            goto out;
> +        }
> +        break;
> +    default:
> +        break;
> +    }
> +
>      ioreq->aio_inflight++;
>      if (ioreq->presync) {
>          blk_aio_flush(ioreq->blkdev->blk, qemu_aio_complete, ioreq);
> 

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel