[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Xen-devel] [PATCH RFC 17/18] OvmfPkg/XenPvBlkDxe: Add BlockFront client.



On Wed, Jul 16, 2014 at 03:41:17PM -0400, Konrad Rzeszutek Wilk wrote:
> On Wed, Jul 16, 2014 at 04:15:46PM +0100, Anthony PERARD wrote:
> > This is the code that will do the actual communication between OVMF and
> > a PV block backend, where the block device lives.
> > 
> > This implementation originally comes from Mini-OS, a part of the Xen
> > Project.
> > 
> > Signed-off-by: Samuel Thibault <samuel.thibault@xxxxxxxxxxxxx>
> > Contributed-under: TianoCore Contribution Agreement 1.0
> > Signed-off-by: Anthony PERARD <anthony.perard@xxxxxxxxxx>
> > ---
> >  OvmfPkg/XenPvBlkDxe/BlockFront.c    | 583 
> > ++++++++++++++++++++++++++++++++++++
> >  OvmfPkg/XenPvBlkDxe/BlockFront.h    |  88 ++++++
> >  OvmfPkg/XenPvBlkDxe/XenPvBlkDxe.c   |   8 +
> >  OvmfPkg/XenPvBlkDxe/XenPvBlkDxe.inf |   2 +
> >  4 files changed, 681 insertions(+)
> >  create mode 100644 OvmfPkg/XenPvBlkDxe/BlockFront.c
> >  create mode 100644 OvmfPkg/XenPvBlkDxe/BlockFront.h
> > 
> > diff --git a/OvmfPkg/XenPvBlkDxe/BlockFront.c 
> > b/OvmfPkg/XenPvBlkDxe/BlockFront.c
> > new file mode 100644
> > index 0000000..ff5641c
> > --- /dev/null
> > +++ b/OvmfPkg/XenPvBlkDxe/BlockFront.c
> > @@ -0,0 +1,583 @@
> > +/* Minimal block driver for Mini-OS.
> > + * Copyright (c) 2007-2008 Samuel Thibault.
> > + * Based on netfront.c.
> > + */
> > +
> > +#include <Library/PrintLib.h>
> > +#include <Library/DebugLib.h>
> > +
> > +#include "BlockFront.h"
> > +
> > +#include <IndustryStandard/Xen/io/protocols.h>
> > +#include <IndustryStandard/Xen/io/xenbus.h>
> > +
> > +#include "inttypes.h"
> > +
> > +STATIC
> > +XENSTORE_STATUS
> > +XenbusReadInteger (
> > +  IN  XENBUS_PROTOCOL *This,
> > +  IN  CONST CHAR8 *Node,
> > +  IN  BOOLEAN FromBackend,
> > +  OUT UINT64 *ValuePtr
> > +  )
> > +{
> > +  XENSTORE_STATUS Status;
> > +  CHAR8 *p;
> > +
> > +  if (!FromBackend) {
> > +    Status = This->XsRead (This, XST_NIL, Node, (VOID**)&p);
> > +  } else {
> > +    Status = This->XsBackendRead (This, XST_NIL, Node, (VOID**)&p);
> > +  }
> > +  if (Status != XENSTORE_STATUS_SUCCESS) {
> > +    return Status;
> > +  }
> > +  // XXX Will assert if p overflow UINT64 ...
> > +  *ValuePtr = AsciiStrDecimalToUint64 (p);
> > +  FreePool (p);
> > +  return Status;
> > +}
> > +
> > +STATIC
> > +VOID
> > +XenPvBlockFree (
> > +  IN XEN_BLOCK_FRONT_DEVICE *Dev
> > +  )
> > +{
> > +  XENBUS_PROTOCOL *XenbusIo = Dev->XenbusIo;
> > +
> > +  if (Dev->RingRef != 0) {
> > +    XenbusIo->GrantEndAccess (XenbusIo, Dev->RingRef);
> > +  }
> > +  if (Dev->Ring.sring != NULL) {
> > +    FreePages (Dev->Ring.sring, 1);
> > +  }
> > +  if (Dev->EventChannel != 0) {
> > +    XenbusIo->EventChannelClose (XenbusIo, Dev->EventChannel);
> > +  }
> > +  FreePool (Dev);
> > +}
> > +
> > +XENSTORE_STATUS
> > +XenPvBlkWaitForBackendState (
> > +  IN  XEN_BLOCK_FRONT_DEVICE *Dev,
> > +  IN  XenbusState            WaitedState,
> 
> ExpectedState
> 
> ?

Yes, will change.

> > +  OUT XenbusState            *LastStatePtr OPTIONAL
> > +  )
> > +{
> > +  XENBUS_PROTOCOL *XenbusIo = Dev->XenbusIo;
> > +  XenbusState State;
> > +  UINT64 Value;
> > +  XENSTORE_STATUS Status;
> > +
> > +  while (TRUE) {
> > +    Status = XenbusReadInteger (XenbusIo, "state", TRUE, &Value);
> > +    if (Status != XENSTORE_STATUS_SUCCESS) {
> > +      return Status;
> > +    }
> > +    if (Value > XenbusStateReconfigured) {
> > +      return XENSTORE_STATUS_EIO;
> > +    }
> > +    State = Value;
> > +    if (State >= WaitedState) {
> 
> Not State != WaitedState ?

I don't think so, the function is actually waiting for the State to be
equal to WaitedState. At least, that how I wrote it. There might be a
better algorithm.

> > +      break;
> > +    }
> > +    DEBUG ((EFI_D_INFO,
> > +            "XenPvBlk: waiting backend state %d, current: %d\n",
> > +            WaitedState, State));
> > +    XenbusIo->WaitForWatch (XenbusIo, Dev->StateWatchToken);
> > +  }
> > +
> > +  if (LastStatePtr != NULL) {
> > +    *LastStatePtr = State;
> > +  }
> > +
> > +  return XENSTORE_STATUS_SUCCESS;
> > +}
> > +
> > +EFI_STATUS
> > +XenPvBlockFrontInitialization (
> > +  IN  XENBUS_PROTOCOL         *XenbusIo,
> > +  IN  CONST CHAR8             *NodeName,
> > +  OUT XEN_BLOCK_FRONT_DEVICE  **DevPtr
> > +  )
> > +{
> > +  XENSTORE_TRANSACTION xbt;
> > +  CHAR8 *DeviceType;
> > +  blkif_sring_t *SharedRing;
> > +  XENSTORE_STATUS Status;
> > +  XEN_BLOCK_FRONT_DEVICE *Dev;
> > +  XenbusState State;
> > +  UINT64 Value;
> > +
> > +  ASSERT (NodeName != NULL);
> > +
> > +  Dev = AllocateZeroPool (sizeof (XEN_BLOCK_FRONT_DEVICE));
> > +  Dev->Signature = XEN_BLOCK_FRONT_SIGNATURE;
> > +  Dev->NodeName = NodeName;
> > +  Dev->XenbusIo = XenbusIo;
> > +  Dev->DeviceId = XenbusIo->DeviceId;
> > +
> > +  XenbusIo->XsRead (XenbusIo, XST_NIL, "device-type", (VOID**)&DeviceType);
> > +  if (AsciiStrCmp (DeviceType, "cdrom") == 0) {
> > +    Dev->MediaInfo.CdRom = TRUE;
> > +  } else {
> > +    Dev->MediaInfo.CdRom = FALSE;
> > +  }
> > +  FreePool (DeviceType);
> > +
> > +  Status = XenbusReadInteger (XenbusIo, "backend-id", FALSE, &Value);
> > +  if (Status != XENSTORE_STATUS_SUCCESS || Value > UINT16_MAX) {
> > +    DEBUG ((EFI_D_ERROR, "XenPvBlk: Failed to get backend-id (%d)\n",
> > +            Status));
> > +    goto Error;
> > +  }
> > +  Dev->DomainId = Value;
> > +  XenbusIo->EventChannelAllocate (XenbusIo, Dev->DomainId, 
> > &Dev->EventChannel);
> > +
> > +  SharedRing = (blkif_sring_t*) AllocatePages (1);
> > +  SHARED_RING_INIT (SharedRing);
> > +  FRONT_RING_INIT (&Dev->Ring, SharedRing, EFI_PAGE_SIZE);
> > +  XenbusIo->GrantAccess (XenbusIo,
> > +                         Dev->DomainId,
> > +                         (INTN) SharedRing >> EFI_PAGE_SHIFT,
> > +                         FALSE,
> > +                         &Dev->RingRef);
> > +
> > +Again:
> > +  Status = XenbusIo->XsTransactionStart (XenbusIo, &xbt);
> > +  if (Status != XENSTORE_STATUS_SUCCESS) {
> > +    DEBUG ((EFI_D_WARN, "XenPvBlk: Failed to start transaction, %d\n", 
> > Status));
> 
> .. and do you want to error out?

The original code (mini-os) is continuing in thes case, so I thought it
was possible to do the initialisation without transaction. But maybe I'm
wrong and this should be an error.

> > +  }
> > +
> > +  Status = XenbusIo->XsPrintf (XenbusIo, xbt, NodeName, "ring-ref", "%d",
> > +                               Dev->RingRef);
> > +  if (Status != XENSTORE_STATUS_SUCCESS) {
> > +    DEBUG ((EFI_D_ERROR, "XenPvBlk: Failed to write ring-ref.\n"));
> > +    goto AbortTransaction;
> > +  }
> > +  Status = XenbusIo->XsPrintf (XenbusIo, xbt, NodeName,
> > +                               "event-channel", "%d", Dev->EventChannel);
> > +  if (Status != XENSTORE_STATUS_SUCCESS) {
> > +    DEBUG ((EFI_D_ERROR, "XenPvBlk: Failed to write event-channel.\n"));
> > +    goto AbortTransaction;
> > +  }
> > +  Status = XenbusIo->XsPrintf (XenbusIo, xbt, NodeName,
> > +                               "protocol", "%a", XEN_IO_PROTO_ABI_NATIVE);
> > +  if (Status != XENSTORE_STATUS_SUCCESS) {
> > +    DEBUG ((EFI_D_ERROR, "XenPvBlk: Failed to write protocol.\n"));
> > +    goto AbortTransaction;
> > +  }
> > +
> > +  Status = XenbusIo->SetState (XenbusIo, xbt, XenbusStateConnected);
> > +  if (Status != XENSTORE_STATUS_SUCCESS) {
> > +    DEBUG ((EFI_D_ERROR, "XenPvBlk: Failed switch state.\n"));
> > +    goto AbortTransaction;
> > +  }
> > +
> > +  Status = XenbusIo->XsTransactionEnd (XenbusIo, xbt, FALSE);
> > +  if (Status == XENSTORE_STATUS_EAGAIN) {
> > +    goto Again;
> > +  }
> > +
> > +  XenbusIo->RegisterWatchBackend (XenbusIo, "state", 
> > &Dev->StateWatchToken);
> > +
> > +  // Waiting backend
> > +  Status = XenPvBlkWaitForBackendState (Dev, XenbusStateConnected, &State);
> > +  if (Status != XENSTORE_STATUS_SUCCESS || State != XenbusStateConnected) {
> > +    DEBUG ((EFI_D_ERROR,
> > +            "XenPvBlk: backend for %a/%d not available, rc=%d state=%d\n",
> > +            XenbusIo->Type, XenbusIo->DeviceId, Status, State));
> > +    goto Error2;
> > +  }
> > +
> > +  Status = XenbusReadInteger (XenbusIo, "info", TRUE, &Value);
> > +  if (Status != XENSTORE_STATUS_SUCCESS || Value > UINT32_MAX) {
> > +    goto Error2;
> > +  }
> > +  Dev->MediaInfo.VDiskInfo = Value;
> > +  if (Dev->MediaInfo.VDiskInfo & VDISK_READONLY) {
> > +    Dev->MediaInfo.ReadWrite = FALSE;
> > +  } else {
> > +    Dev->MediaInfo.ReadWrite = TRUE;
> > +  }
> > +
> > +  Status = XenbusReadInteger (XenbusIo, "sectors", TRUE, 
> > &Dev->MediaInfo.Sectors);
> > +  if (Status != XENSTORE_STATUS_SUCCESS) {
> > +    goto Error2;
> > +  }
> > +
> > +  Status = XenbusReadInteger (XenbusIo, "sector-size", TRUE, &Value);
> > +  if (Status != XENSTORE_STATUS_SUCCESS || Value > UINT32_MAX) {
> > +    goto Error2;
> > +  }
> > +  Dev->MediaInfo.SectorSize = Value;
> 
> Why not &Dev->MediaInfo.SectorSize in XenbusReadInteger?
> 
> Is it because of types? If so, please add a comment stating that.

Yes, because of the type, uint64 vs uint32.

Maybe I could rename the XenbusReadInteger to XenbusReadUint64, and it
would be more obvious that the function require a specific type.

> > +
> > +  // Default value
> > +  Value = 0;
> > +  Status = XenbusReadInteger (XenbusIo, "feature-barrier", TRUE, &Value);
> > +  if (Value == 1) {
> > +    Dev->MediaInfo.FeatureBarrier = TRUE;
> > +  } else {
> > +    Dev->MediaInfo.FeatureBarrier = FALSE;
> > +  }
> > +
> > +  // Default value
> > +  Value = 0;
> > +  XenbusReadInteger (XenbusIo, "feature-flush-cache", TRUE, &Value);
> > +  if (Value == 1) {
> > +    Dev->MediaInfo.FeatureFlushCache = TRUE;
> > +  } else {
> > +    Dev->MediaInfo.FeatureFlushCache = FALSE;
> > +  }
> > +
> > +  DEBUG ((EFI_D_INFO, "XenPvBlk: New disk with %d sectors of %d bytes\n",
> > +          Dev->MediaInfo.Sectors, Dev->MediaInfo.SectorSize));
> > +
> > +  *DevPtr = Dev;
> > +  return EFI_SUCCESS;
> > +
> > +Error2:
> > +  XenbusIo->UnregisterWatch (XenbusIo, Dev->StateWatchToken);
> > +  XenbusIo->XsRemove (XenbusIo, XST_NIL, "ring-ref");
> > +  XenbusIo->XsRemove (XenbusIo, XST_NIL, "event-channel");
> 
> XenbusIo->XsRemove (XenbusIo, XST_NIL, "protocol");
> 
> ?

I don't know, but I can do that.

> > +  goto Error;
> > +AbortTransaction:
> > +  XenbusIo->XsTransactionEnd (XenbusIo, xbt, TRUE);
> > +Error:
> > +  XenPvBlockFree (Dev);
> > +  return EFI_DEVICE_ERROR;
> > +}
> > +
> > +VOID
> > +XenPvBlockFrontShutdown (
> > +  IN XEN_BLOCK_FRONT_DEVICE *Dev
> > +  )
> > +{
> > +  XENBUS_PROTOCOL *XenbusIo = Dev->XenbusIo;
> > +  XENSTORE_STATUS Status;
> > +  UINT64 Value;
> > +
> > +  XenPvBlockSync (Dev);
> > +
> > +  Status = XenbusIo->SetState (XenbusIo, XST_NIL, XenbusStateClosing);
> > +  if (Status != XENSTORE_STATUS_SUCCESS) {
> > +    DEBUG ((EFI_D_ERROR,
> > +            "XenPvBlk: error while changing state to Closing: %d\n",
> > +            Status));
> > +    goto Close;
> > +  }
> > +
> > +  Status = XenPvBlkWaitForBackendState (Dev, XenbusStateClosing, NULL);
> > +  if (Status != XENSTORE_STATUS_SUCCESS) {
> > +    DEBUG ((EFI_D_ERROR,
> > +            "XenPvBlk: error while waiting for closing backend state: 
> > %d\n",
> > +            Status));
> > +    goto Close;
> > +  }
> > +
> > +  Status = XenbusIo->SetState (XenbusIo, XST_NIL, XenbusStateClosed);
> > +  if (Status != XENSTORE_STATUS_SUCCESS) {
> > +    DEBUG ((EFI_D_ERROR,
> > +            "XenPvBlk: error while changing state to Closed: %d\n",
> > +            Status));
> > +    goto Close;
> > +  }
> > +
> > +  Status = XenPvBlkWaitForBackendState (Dev, XenbusStateClosed, NULL);
> > +  if (Status != XENSTORE_STATUS_SUCCESS) {
> > +    DEBUG ((EFI_D_ERROR,
> > +            "XenPvBlk: error while waiting for closed backend state: %d\n",
> > +            Status));
> > +    goto Close;
> > +  }
> > +
> > +  Status = XenbusIo->SetState (XenbusIo, XST_NIL, XenbusStateInitialising);
> > +  if (Status != XENSTORE_STATUS_SUCCESS) {
> > +    DEBUG ((EFI_D_ERROR,
> > +            "XenPvBlk: error while changing state to initialising: %d\n",
> > +            Status));
> > +    goto Close;
> > +  }
> > +
> > +  while (TRUE) {
> > +    Status = XenbusReadInteger (XenbusIo, "state", TRUE, &Value);
> > +    if (Status != XENSTORE_STATUS_SUCCESS) {
> > +      DEBUG ((EFI_D_ERROR,
> > +              "XenPvBlk: error while waiting for new backend state: %d\n",
> > +              Status));
> > +      goto Close;
> > +    }
> > +    if (Value < XenbusStateInitWait || Value >= XenbusStateClosed) {
> > +      break;
> > +    }
> > +    DEBUG ((EFI_D_INFO,
> > +            "XenPvBlk: waiting backend state %d, current: %d\n",
> > +            XenbusStateInitWait, Value));
> > +    XenbusIo->WaitForWatch (XenbusIo, Dev->StateWatchToken);
> > +  }
> > +
> > +Close:
> > +  XenbusIo->UnregisterWatch (XenbusIo, Dev->StateWatchToken);
> > +  XenbusIo->XsRemove (XenbusIo, XST_NIL, "ring-ref");
> > +  XenbusIo->XsRemove (XenbusIo, XST_NIL, "event-channel");
> 
> "protocol" 
> > +
> > +  XenPvBlockFree (Dev);
> > +}
> > +
> > +STATIC
> > +VOID
> > +XenPvBlockWaitSlot (
> > +  IN XEN_BLOCK_FRONT_DEVICE *Dev
> > +  )
> > +{
> > +  /* Wait for a slot */
> > +  if (RING_FULL (&Dev->Ring)) {
> > +    while (TRUE) {
> > +      XenPvBlockAsyncIoPoll (Dev);
> > +      if (!RING_FULL(&Dev->Ring)) {
> > +        break;
> > +      }
> > +      /* Really no slot, go to sleep. */
> > +      // XXX wait for an event on Dev->EventChannel
> > +      DEBUG ((EFI_D_INFO, "%a %d, waiting\n", __func__, __LINE__));
> > +    }
> > +  }
> > +}
> > +
> > +/* Issue an aio */
> > +VOID
> > +XenPvBlockAsyncIo (
> > +  IN OUT XEN_BLOCK_FRONT_IO *IoData,
> > +  IN     BOOLEAN            IsWrite
> > +  )
> > +{
> > +  XEN_BLOCK_FRONT_DEVICE *Dev = IoData->Dev;
> > +  XENBUS_PROTOCOL *XenbusIo = Dev->XenbusIo;
> > +  blkif_request_t *Request;
> > +  RING_IDX RingIndex;
> > +  BOOLEAN Notify;
> > +  INT32 NumSegments, Index;
> > +  UINTN Start, End;
> > +
> > +  // Can't io at non-sector-aligned location
> > +  ASSERT(!(IoData->Offset & (Dev->MediaInfo.SectorSize - 1)));
> > +  // Can't io non-sector-sized amounts
> > +  ASSERT(!(IoData->Size & (Dev->MediaInfo.SectorSize - 1)));
> > +  // Can't io non-sector-aligned buffer
> > +  ASSERT(!((UINTN) IoData->Buffer & (Dev->MediaInfo.SectorSize - 1)));
> > +
> > +  Start = (UINTN) IoData->Buffer & ~EFI_PAGE_MASK;
> > +  End = ((UINTN) IoData->Buffer + IoData->Size + EFI_PAGE_SIZE - 1) & 
> > ~EFI_PAGE_MASK;
> > +  IoData->NumRef = NumSegments = (End - Start) / EFI_PAGE_SIZE;
> > +
> > +  // XXX is this comment still true ?
> 
> I think it is outdated, as Linux would do max DMA up to the segment
> size that is advertised. Ditto for the SCSI stack.

OK, thanks.

> > +  /* qemu's IDE max multsect is 16 (8KB) and SCSI max DMA was set to 32KB,
> > +   * so max 44KB can't happen */
> > +  ASSERT (NumSegments <= BLKIF_MAX_SEGMENTS_PER_REQUEST);
> 
> But this check is OK as we can't truly fit more on the ring.

OK.

> > +
> > +  XenPvBlockWaitSlot (Dev);
> 
> What happens if there are multiple threads calling this
> function? Don't we need a mutex to only allow
> one thread to grab an RingIndex?

So, OVMF does not have threads but the are "events", called by a timer.
This code does not support it, yet, so I don't need any form of locking,
yet.

> > +  RingIndex = Dev->Ring.req_prod_pvt;
> > +  Request = RING_GET_REQUEST (&Dev->Ring, RingIndex);
> > +
> > +  Request->operation = IsWrite ? BLKIF_OP_WRITE : BLKIF_OP_READ;
> > +  Request->nr_segments = NumSegments;
> > +  Request->handle = Dev->DeviceId;
> > +  Request->id = (UINTN) IoData;
> > +  Request->sector_number = IoData->Offset / 512;
> > +
> > +  for (Index = 0; Index < NumSegments; Index++) {
> > +    Request->seg[Index].first_sect = 0;
> 
> Uh? 
> > +    Request->seg[Index].last_sect = EFI_PAGE_SIZE / 512 - 1;
> 
> Uh? I thought you wanted to use the real sector values?
> > +  }
> > +  Request->seg[0].first_sect = ((UINTN) IoData->Buffer & EFI_PAGE_MASK) / 
> > 512;
> > +  Request->seg[NumSegments - 1].last_sect =
> > +      (((UINTN) IoData->Buffer + IoData->Size - 1) & EFI_PAGE_MASK) / 512;
> 
> What about in between 0 and NumSegments? You should also
> fill out those values with valid entries.

It's done by the previous for loop. Maybe this part is hard to
understand without the interface explain.
include io/blkif.h says:
  @first_sect: first sector in frame to transfer (inclusive).
  @last_sect: last sector in frame to transfer (inclusive).

but does not explain what a frame is, beyond this comment on:
  grant_ref_t gref; /* reference to I/O buffer frame */
(all this from the struct blkif_request_segment)
Or maybe frame == page.

Anyway, the way the segments are initialized feel weirds now, but I
don't think it's incorrect, otherwise, I would not be able to boot any
guest.

> > +  for (Index = 0; Index < NumSegments; Index++) {
> > +    UINTN Data = Start + Index * EFI_PAGE_SIZE;
> > +    if (!IsWrite) {
> > +      /* Trigger CoW if needed */
> 
> Not sure I understand that reference?

Maybe a reference to the original commit will help to understand:
minios: Fix bug when blkfront reading into zero-mapped buffer by just poking 
the page.
No need to use virtual_to_mfn() for the ring since that is a real page.
7b1257946080b63421a8dc59c7b3dfd67997c643 (xen.git)

But I still don't undertand this.

> > +      *(CHAR8 *)(Data + (Request->seg[Index].first_sect << 9)) = 0;
> > +      MemoryFence ();
> > +    }
> > +    XenbusIo->GrantAccess (XenbusIo, Dev->DomainId,
> > +                           Data >> EFI_PAGE_SHIFT, IsWrite,
> > +                           &Request->seg[Index].gref);
> > +    IoData->GrantRef[Index] = Request->seg[Index].gref;
> > +  }
> > +
> > +  Dev->Ring.req_prod_pvt = RingIndex + 1;
> > +
> > +  MemoryFence ();
> > +  RING_PUSH_REQUESTS_AND_CHECK_NOTIFY (&Dev->Ring, Notify);
> > +
> > +  if (Notify) {
> > +    XenbusIo->EventChannelNotify (XenbusIo, Dev->EventChannel);
> > +  }
> > +}
> > +
> > +STATIC
> > +VOID
> > +XenPvBlockAsyncIoCallback (
> > +  IN OUT XEN_BLOCK_FRONT_IO *IoData,
> > +  IN     EFI_STATUS         Status
> > +  )
> > +{
> > +  IoData->Data = (VOID *) 1;
> > +  IoData->Callback = NULL;
> > +}
> > +
> > +VOID
> > +XenPvBlockIo (
> > +  IN OUT XEN_BLOCK_FRONT_IO *IoData,
> > +  IN     BOOLEAN            IsWrite
> > +  )
> > +{
> > +  ASSERT (IoData->Callback == NULL);
> > +  // XXX Signal/Event instead ?
> 
> Maybe for later versions. I think this is OK for right now, thought
> it will mean we can only do one I/O .. and have then to wait
> until it is completed.
> 
> That limititation should be mentioned in the commit description.

That right, right now, it's only one IO at a time. I think my comment is
more to tell me that Callback is not really one, and it could be replace
by something more Uefi, an event.

> > +  IoData->Callback = XenPvBlockAsyncIoCallback;
> > +  XenPvBlockAsyncIo (IoData, IsWrite);
> > +  IoData->Data = NULL;
> > +
> > +  while (TRUE) {
> > +    XenPvBlockAsyncIoPoll (IoData->Dev);
> > +    if (IoData->Data) {
> > +      break;
> > +    }
> > +  }
> > +}
> > +
> > +STATIC
> > +VOID
> > +XenPvBlockPushOperation (
> > +  IN XEN_BLOCK_FRONT_DEVICE *Dev,
> > +  IN UINT8                  Operation,
> > +  IN UINT64                 Id
> > +  )
> > +{
> > +  INT32 Index;
> > +  blkif_request_t *Request;
> > +  BOOLEAN Notify;
> > +
> > +  XenPvBlockWaitSlot (Dev);
> 
> How do we make sure that there aren't multiple threads
> calling this?

Not possible yet.

> > +  Index = Dev->Ring.req_prod_pvt;
> > +  Request = RING_GET_REQUEST(&Dev->Ring, Index);
> > +  Request->operation = Operation;
> > +  Request->nr_segments = 0;
> > +  Request->handle = Dev->DeviceId;
> > +  Request->id = Id;
> > +  /* Not needed anyway, but the backend will check it */
> > +  Request->sector_number = 0;
> > +  Dev->Ring.req_prod_pvt = Index + 1;
> > +  MemoryFence ();
> > +  RING_PUSH_REQUESTS_AND_CHECK_NOTIFY (&Dev->Ring, Notify);
> > +  if (Notify) {
> > +    XENBUS_PROTOCOL *XenbusIo = Dev->XenbusIo;
> > +    XenbusIo->EventChannelNotify (XenbusIo, Dev->EventChannel);
> > +  }
> > +}
> > +
> > +VOID
> > +XenPvBlockSync (
> > +  IN XEN_BLOCK_FRONT_DEVICE *Dev
> > +  )
> > +{
> > +  if (Dev->MediaInfo.ReadWrite) {
> > +    if (Dev->MediaInfo.FeatureBarrier) {
> > +      XenPvBlockPushOperation (Dev, BLKIF_OP_WRITE_BARRIER, 0);
> > +    }
> > +
> > +    if (Dev->MediaInfo.FeatureFlushCache) {
> > +      XenPvBlockPushOperation (Dev, BLKIF_OP_FLUSH_DISKCACHE, 0);
> > +    }
> > +  }
> > +
> > +  /* Note: This won't finish if another thread enqueues requests.  */
> > +  while (TRUE) {
> > +    XenPvBlockAsyncIoPoll (Dev);
> > +    if (RING_FREE_REQUESTS (&Dev->Ring) == RING_SIZE (&Dev->Ring)) {
> > +      break;
> > +    }
> > +  }
> > +}
> > +
> > +VOID
> > +XenPvBlockAsyncIoPoll (
> > +  IN XEN_BLOCK_FRONT_DEVICE *Dev
> > +  )
> > +{
> > +  RING_IDX ProducerIndex, ConsumerIndex;
> > +  blkif_response_t *Response;
> > +  INT32 More;
> > +
> > +  do {
> > +    ProducerIndex = Dev->Ring.sring->rsp_prod;
> > +    /* Ensure we see queued responses up to 'ProducerIndex'. */
> > +    MemoryFence ();
> > +    ConsumerIndex = Dev->Ring.rsp_cons;
> > +
> > +    while (ConsumerIndex != ProducerIndex) {
> > +      XEN_BLOCK_FRONT_IO *IoData;
> > +      INT16 Status;
> > +
> > +      Response = RING_GET_RESPONSE (&Dev->Ring, ConsumerIndex);
> > +
> > +      IoData = (VOID *) (UINTN) Response->id;
> > +      Status = Response->status;
> > +
> > +      switch (Response->operation) {
> > +      case BLKIF_OP_READ:
> > +      case BLKIF_OP_WRITE:
> > +        {
> > +          INT32 Index;
> > +
> > +          if (Status != BLKIF_RSP_OKAY) {
> > +            DEBUG ((EFI_D_ERROR,
> > +                    "XenPvBlk: "
> > +                    "%a error %d on %a at offset %Ld, num bytes %Ld\n",
> > +                    Response->operation == BLKIF_OP_READ ? "read" : 
> > "write",
> > +                    Status, IoData->Dev->NodeName,
> > +                    IoData->Offset,
> > +                    IoData->Size));
> > +          }
> > +
> > +          for (Index = 0; Index < IoData->NumRef; Index++) {
> > +            Dev->XenbusIo->GrantEndAccess (Dev->XenbusIo, 
> > IoData->GrantRef[Index]);
> > +          }
> > +
> > +          break;
> > +        }
> > +
> > +      case BLKIF_OP_WRITE_BARRIER:
> > +        if (Status != BLKIF_RSP_OKAY) {
> > +          DEBUG ((EFI_D_ERROR, "XenPvBlk: write barrier error %d\n", 
> > Status));
> > +        }
> > +        break;
> > +      case BLKIF_OP_FLUSH_DISKCACHE:
> > +        if (Status != BLKIF_RSP_OKAY) {
> > +          DEBUG ((EFI_D_ERROR, "XenPvBlk: flush error %d\n", Status));
> > +        }
> > +        break;
> > +
> > +      default:
> > +        DEBUG ((EFI_D_ERROR,
> > +                "XenPvBlk: unrecognized block operation %d response 
> > (status %d)\n",
> > +                Response->operation, Status));
> > +        break;
> > +      }
> > +
> > +      Dev->Ring.rsp_cons = ++ConsumerIndex;
> > +      /* Nota: callback frees IoData itself */
> > +      if (IoData && IoData->Callback) {
> > +        IoData->Callback (IoData, Status ? EFI_DEVICE_ERROR : EFI_SUCCESS);
> > +      }
> > +      if (Dev->Ring.rsp_cons != ConsumerIndex) {
> > +        /* We reentered, we must not continue here */
> 
> Uh, how does that happen?

That an original comment, I don't think it's possible yet in ovmf.

> > +        break;
> > +      }
> > +    }
> > +
> > +    RING_FINAL_CHECK_FOR_RESPONSES (&Dev->Ring, More);
> > +  } while (More != 0);
> > +}

-- 
Anthony PERARD

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


 


Rackspace

Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.