Skip to contentSkip to conversation

ide: avoid main-loop hang on CDROM/NFS failure

PATCH Superseded Peter Lieven 3 participants 18 messages Oct 12, 2015 – Nov 3, 2015

v2 of 4 Superseded Latest is v4 →

lore
Share & tools
lore
#0 lore raw

v2 cover ide: avoid main-loop hang on CDROM/NFS failure

This series aims at avoiding a hanging main-loop if a vserver has a CDROM image mounted from a NFS share and that NFS share goes down. Typical situation is that users mount an CDROM ISO to install something and then forget to eject that CDROM afterwards.
As a consequence this mounted CD is able to bring down the whole vserver if the backend NFS share is unreachable. This is bad especially if the CDROM itself is not needed anymore at this point.

This series aims at fixing 2 blocking I/O operations that would hang if the NFS server is unavailable:
- ATAPI PIO read requests used sync calls to blk_read, convert them to an async variant where possible.
- If a busmaster DMA request is cancelled all requests are drained. Convert the drain to an async request canceling.

v1->v2: - fix offset for 2352 byte sector size [Kevin]
- use a sync request if we continue an elementary transfer. As John pointed out we enter a race condition between next IDE command and async transfer otherwise. This is sill not optimal, but it fixes the NFS down problems for all cases where the NFS server goes down while there is no PIO CD activity. Of course, it could still happen during a PIO transfer, but I expect this to be the unlikelier case.
I spent some effort trying to read more sectors at once and avoiding continuation of elementary transfers, but with whatever I came up it was destroying migration between different Qemu versions. I have a quite hackish patch that works and should survive migration, but I am not happy with it. So I would like to start with this version as it is a big improvement already.
- Dropped Patch 5 because it is upstream meanwhile.

Peter Lieven (4):
ide/atapi: make PIO read requests async
ide/atapi: blk_aio_readv may return NULL
ide: add support for cancelable read requests
ide/atapi: enable cancelable requests

 hw/ide/atapi.c    | 99 +++++++++++++++++++++++++++++++++++++++++++++++++------
 hw/ide/core.c     | 55 +++++++++++++++++++++++++++++++
 hw/ide/internal.h | 16 +++++++++
 hw/ide/pci.c      | 42 +++++++++++++++--------
 4 files changed, 188 insertions(+), 24 deletions(-)
Signature · 1 line
1.9.1
#1 lore raw

PATCH v2 1/4 ide/atapi: make PIO read requests async

PIO read requests on the ATAPI interface used to be sync blk requests. This has two significant drawbacks. First the main loop hangs util an I/O request is completed and secondly if the I/O request does not complete (e.g. due to an unresponsive storage) Qemu hangs completely.

---
 hw/ide/atapi.c | 93 ++++++++++++++++++++++++++++++++++++++++++++++++++++------
 1 file changed, 84 insertions(+), 9 deletions(-)
hw/ide/atapi.c +84 -9
index 747f466..2271ea2 100644
@@ -105,11 +105,16 @@ static void cd_data_to_raw(uint8_t *buf, int lba)
     memset(buf, 0, 288);
 }
 
-static int cd_read_sector(IDEState *s, int lba, uint8_t *buf, int sector_size)
+static int
+cd_read_sector_sync(IDEState *s, int lba, uint8_t *buf)
 {
     int ret;
 
-    switch(sector_size) {
+#ifdef DEBUG_IDE_ATAPI
+    printf("cd_read_sector_sync: lba=%d\n", lba);
+#endif
+
+    switch (s->cd_sector_size) {
     case 2048:
         block_acct_start(blk_get_stats(s->blk), &s->acct,
                          4 * BDRV_SECTOR_SIZE, BLOCK_ACCT_READ);
@@ -129,9 +134,71 @@ static int cd_read_sector(IDEState *s, int lba, uint8_t *buf, int sector_size)
         ret = -EIO;
         break;
     }
+
+    if (!ret) {
+        s->lba++;
+        s->io_buffer_index = 0;
+    }
+
     return ret;
 }
 
+static void cd_read_sector_cb(void *opaque, int ret)
+{
+    IDEState *s = opaque;
+
+    block_acct_done(blk_get_stats(s->blk), &s->acct);
+
+#ifdef DEBUG_IDE_ATAPI
+    printf("cd_read_sector_cb: lba=%d ret=%d\n", s->lba, ret);
+#endif
+
+    if (ret < 0) {
+        ide_atapi_io_error(s, ret);
+        return;
+    }
+
+    if (s->cd_sector_size == 2352) {
+        cd_data_to_raw(s->io_buffer, s->lba);
+    }
+
+    s->lba++;
+    s->io_buffer_index = 0;
+    s->status &= ~BUSY_STAT;
+
+    ide_atapi_cmd_reply_end(s);
+}
+
+static int cd_read_sector(IDEState *s, int lba, void *buf)
+{
+    if (s->cd_sector_size != 2048 && s->cd_sector_size != 2352) {
+        return -EINVAL;
+    }
+
+    s->iov.iov_base = buf;
+    if (s->cd_sector_size == 2352) {
+        buf += 16;
+    }
+
+    s->iov.iov_len = 4 * BDRV_SECTOR_SIZE;
+    qemu_iovec_init_external(&s->qiov, &s->iov, 1);
+
+#ifdef DEBUG_IDE_ATAPI
+    printf("cd_read_sector: lba=%d\n", lba);
+#endif
+
+    if (blk_aio_readv(s->blk, (int64_t)lba << 2, &s->qiov, 4,
+                      cd_read_sector_cb, s) == NULL) {
+        return -EIO;
+    }
+
+    block_acct_start(blk_get_stats(s->blk), &s->acct,
+                     4 * BDRV_SECTOR_SIZE, BLOCK_ACCT_READ);
+
+    s->status |= BUSY_STAT;
+    return 0;
+}
+
 void ide_atapi_cmd_ok(IDEState *s)
 {
     s->error = 0;
@@ -182,18 +249,27 @@ void ide_atapi_cmd_reply_end(IDEState *s)
         ide_atapi_cmd_ok(s);
         ide_set_irq(s->bus);
 #ifdef DEBUG_IDE_ATAPI
-        printf("status=0x%x\n", s->status);
+        printf("end of transfer, status=0x%x\n", s->status);
 #endif
     } else {
         /* see if a new sector must be read */
         if (s->lba != -1 && s->io_buffer_index >= s->cd_sector_size) {
-            ret = cd_read_sector(s, s->lba, s->io_buffer, s->cd_sector_size);
-            if (ret < 0) {
-                ide_atapi_io_error(s, ret);
+            if (!s->elementary_transfer_size) {
+                ret = cd_read_sector(s, s->lba, s->io_buffer);
+                if (ret < 0) {
+                    ide_atapi_io_error(s, ret);
+                }
                 return;
+            } else {
+                /* rebuffering within an elementary transfer is
+                 * only possible with a sync request because we
+                 * end up with a race condition otherwise */
+                ret = cd_read_sector_sync(s, s->lba, s->io_buffer);
+                if (ret < 0) {
+                    ide_atapi_io_error(s, ret);
+                    return;
+                }
             }
-            s->lba++;
-            s->io_buffer_index = 0;
         }
         if (s->elementary_transfer_size > 0) {
             /* there are some data left to transmit in this elementary
@@ -275,7 +351,6 @@ static void ide_atapi_cmd_read_pio(IDEState *s, int lba, int nb_sectors,
     s->io_buffer_index = sector_size;
     s->cd_sector_size = sector_size;
 
-    s->status = READY_STAT | SEEK_STAT;
     ide_atapi_cmd_reply_end(s);
 }
 
Signature · 1 line
1.9.1
#2 lore raw

Re: [Qemu-devel] [PATCH 1/4] ide/atapi: make PIO read requests async

hw/ide/atapi.c static int cd_read_sector(IDEState *s, int lba, uint8_t *buf, int sector_size) · line 139

Quoted 7 lines · Peter Lieven
hw/ide/atapi.c +3 -0
@@ -129,9 +134,71 @@ static int cd_read_sector(IDEState *s, int lba, uint8_t *buf, int sector_size)
         ret = -EIO;
         break;
     }
+
+    if (!ret) {
+        s->lba++;

This function probably shouldn't take the lba argument if it modifies s->lba. You dropped the sector_size argument and I think the lba argument should be dropped for the same reason.

Quoted 10 lines
hw/ide/atapi.c +10 -0
+static int cd_read_sector(IDEState *s, int lba, void *buf)
+{
+    if (s->cd_sector_size != 2048 && s->cd_sector_size != 2352) {
+        return -EINVAL;
+    }
+
+    s->iov.iov_base = buf;
+    if (s->cd_sector_size == 2352) {
+        buf += 16;
+    }

hw/ide/atapi.c static int cd_read_sector(IDEState *s, int lba, uint8_t *buf, int sector_size) · line 191

Quoted 10 lines
hw/ide/atapi.c +10 -0
+
+    s->iov.iov_len = 4 * BDRV_SECTOR_SIZE;
+    qemu_iovec_init_external(&s->qiov, &s->iov, 1);
+
+#ifdef DEBUG_IDE_ATAPI
+    printf("cd_read_sector: lba=%d\n", lba);
+#endif
+
+    if (blk_aio_readv(s->blk, (int64_t)lba << 2, &s->qiov, 4,
+                      cd_read_sector_cb, s) == NULL) {

This function never returns NULL, the if statement can be removed.

Doesn't the aiocb need to be stored for cancellation (e.g. device reset)?

hw/ide/atapi.c static int cd_read_sector(IDEState *s, int lba, uint8_t *buf, int sector_size) · line 196

Quoted 5 lines
hw/ide/atapi.c +5 -0
+        return -EIO;
+    }
+
+    block_acct_start(blk_get_stats(s->blk), &s->acct,
+                     4 * BDRV_SECTOR_SIZE, BLOCK_ACCT_READ);

Why does accounting start *after* the read request has been submitted?

#3 lore raw

Re: [Qemu-devel] [PATCH 1/4] ide/atapi: make PIO read requests async

Am 22.10.2015 um 18:17 schrieb Stefan Hajnoczi:

Quoted 25 lines · Peter Lieven

hw/ide/atapi.c static int cd_read_sector(IDEState *s, int lba, uint8_t *buf, int sector_size) · line 139

Quoted 7 lines
hw/ide/atapi.c +3 -0
@@ -129,9 +134,71 @@ static int cd_read_sector(IDEState *s, int lba, uint8_t *buf, int sector_size)
         ret = -EIO;
         break;
     }
+
+    if (!ret) {
+        s->lba++;

This function probably shouldn't take the lba argument if it modifies s->lba. You dropped the sector_size argument and I think the lba argument should be dropped for the same reason.

Quoted 13 lines
hw/ide/atapi.c +13 -0
+static int cd_read_sector(IDEState *s, int lba, void *buf)
+{
+    if (s->cd_sector_size != 2048 && s->cd_sector_size != 2352) {
+        return -EINVAL;
+    }
+
+    s->iov.iov_base = buf;
+    if (s->cd_sector_size == 2352) {
+        buf += 16;
+    }
+
+    s->iov.iov_len = 4 * BDRV_SECTOR_SIZE;
+    qemu_iovec_init_external(&s->qiov, &s->iov, 1);
Quoted 8 lines

hw/ide/atapi.c static int cd_read_sector(IDEState *s, int lba, uint8_t *buf, int sector_size) · line 191

Quoted 7 lines
hw/ide/atapi.c +7 -0
+
+#ifdef DEBUG_IDE_ATAPI
+    printf("cd_read_sector: lba=%d\n", lba);
+#endif
+
+    if (blk_aio_readv(s->blk, (int64_t)lba << 2, &s->qiov, 4,
+                      cd_read_sector_cb, s) == NULL) {

This function never returns NULL, the if statement can be removed.

Okay, that was my fault. I was believing it could return NULL.

Quoted 2 lines

Doesn't the aiocb need to be stored for cancellation (e.g. device reset)?

The ide_readv_cancelable function introduced in Patch 3 will store the aiocb. At this point there is blk_drain_all called when the device is reset.

Quoted 7 lines

hw/ide/atapi.c static int cd_read_sector(IDEState *s, int lba, uint8_t *buf, int sector_size) · line 196

Quoted 6 lines
hw/ide/atapi.c +5 -0
+        return -EIO;
+    }
+
+    block_acct_start(blk_get_stats(s->blk), &s->acct,
+                     4 * BDRV_SECTOR_SIZE, BLOCK_ACCT_READ);

Why does accounting start *after* the read request has been submitted?

You are right it has to start before the request.

Peter

#4 lore raw

Re: [Qemu-devel] [PATCH 1/4] ide/atapi: make PIO read requests async

In reply to Peter Lieven
Quoted 141 lines · Peter Lieven

PIO read requests on the ATAPI interface used to be sync blk requests. This has two significant drawbacks. First the main loop hangs util an I/O request is completed and secondly if the I/O request does not complete (e.g. due to an unresponsive storage) Qemu hangs completely.

---
 hw/ide/atapi.c | 93 ++++++++++++++++++++++++++++++++++++++++++++++++++++------
 1 file changed, 84 insertions(+), 9 deletions(-)
hw/ide/atapi.c +84 -8
index 747f466..2271ea2 100644
@@ -105,11 +105,16 @@ static void cd_data_to_raw(uint8_t *buf, int lba)
     memset(buf, 0, 288);
 }
 
-static int cd_read_sector(IDEState *s, int lba, uint8_t *buf, int sector_size)
+static int
+cd_read_sector_sync(IDEState *s, int lba, uint8_t *buf)
 {
     int ret;
 
-    switch(sector_size) {
+#ifdef DEBUG_IDE_ATAPI
+    printf("cd_read_sector_sync: lba=%d\n", lba);
+#endif
+
+    switch (s->cd_sector_size) {
     case 2048:
         block_acct_start(blk_get_stats(s->blk), &s->acct,
                          4 * BDRV_SECTOR_SIZE, BLOCK_ACCT_READ);
@@ -129,9 +134,71 @@ static int cd_read_sector(IDEState *s, int lba, uint8_t *buf, int sector_size)
         ret = -EIO;
         break;
     }
+
+    if (!ret) {
+        s->lba++;
+        s->io_buffer_index = 0;
+    }
+
     return ret;
 }
 
+static void cd_read_sector_cb(void *opaque, int ret)
+{
+    IDEState *s = opaque;
+
+    block_acct_done(blk_get_stats(s->blk), &s->acct);
+
+#ifdef DEBUG_IDE_ATAPI
+    printf("cd_read_sector_cb: lba=%d ret=%d\n", s->lba, ret);
+#endif
+
+    if (ret < 0) {
+        ide_atapi_io_error(s, ret);
+        return;
+    }
+
+    if (s->cd_sector_size == 2352) {
+        cd_data_to_raw(s->io_buffer, s->lba);
+    }
+
+    s->lba++;
+    s->io_buffer_index = 0;
+    s->status &= ~BUSY_STAT;
+
+    ide_atapi_cmd_reply_end(s);
+}
+
+static int cd_read_sector(IDEState *s, int lba, void *buf)
+{
+    if (s->cd_sector_size != 2048 && s->cd_sector_size != 2352) {
+        return -EINVAL;
+    }
+
+    s->iov.iov_base = buf;
+    if (s->cd_sector_size == 2352) {
+        buf += 16;
+    }
+
+    s->iov.iov_len = 4 * BDRV_SECTOR_SIZE;
+    qemu_iovec_init_external(&s->qiov, &s->iov, 1);
+
+#ifdef DEBUG_IDE_ATAPI
+    printf("cd_read_sector: lba=%d\n", lba);
+#endif
+
+    if (blk_aio_readv(s->blk, (int64_t)lba << 2, &s->qiov, 4,
+                      cd_read_sector_cb, s) == NULL) {
+        return -EIO;
+    }
+
+    block_acct_start(blk_get_stats(s->blk), &s->acct,
+                     4 * BDRV_SECTOR_SIZE, BLOCK_ACCT_READ);
+
+    s->status |= BUSY_STAT;
+    return 0;
+}
+
 void ide_atapi_cmd_ok(IDEState *s)
 {
     s->error = 0;
@@ -182,18 +249,27 @@ void ide_atapi_cmd_reply_end(IDEState *s)
         ide_atapi_cmd_ok(s);
         ide_set_irq(s->bus);
 #ifdef DEBUG_IDE_ATAPI
-        printf("status=0x%x\n", s->status);
+        printf("end of transfer, status=0x%x\n", s->status);
 #endif
     } else {
         /* see if a new sector must be read */
         if (s->lba != -1 && s->io_buffer_index >= s->cd_sector_size) {
-            ret = cd_read_sector(s, s->lba, s->io_buffer, s->cd_sector_size);
-            if (ret < 0) {
-                ide_atapi_io_error(s, ret);
+            if (!s->elementary_transfer_size) {
+                ret = cd_read_sector(s, s->lba, s->io_buffer);
+                if (ret < 0) {
+                    ide_atapi_io_error(s, ret);
+                }
                 return;
+            } else {
+                /* rebuffering within an elementary transfer is
+                 * only possible with a sync request because we
+                 * end up with a race condition otherwise */
+                ret = cd_read_sector_sync(s, s->lba, s->io_buffer);
+                if (ret < 0) {
+                    ide_atapi_io_error(s, ret);
+                    return;
+                }
             }
-            s->lba++;
-            s->io_buffer_index = 0;
         }
         if (s->elementary_transfer_size > 0) {
             /* there are some data left to transmit in this elementary

hw/ide/atapi.c static void ide_atapi_cmd_read_pio(IDEState *s, int lba, int nb_sectors, · line 356

Quoted 7 lines
hw/ide/atapi.c +0 -1
@@ -275,7 +351,6 @@ static void ide_atapi_cmd_read_pio(IDEState *s, int lba, int nb_sectors,
     s->io_buffer_index = sector_size;
     s->cd_sector_size = sector_size;
 
-    s->status = READY_STAT | SEEK_STAT;
     ide_atapi_cmd_reply_end(s);
 }
 

This patch looks good to me, apart from Stefan's other comments. Will you be sending a V3? I can try to merge it for this week before the hard freeze hits.

--js

#5 lore raw

Re: [Qemu-devel] [PATCH 1/4] ide/atapi: make PIO read requests async

Am 03.11.2015 um 01:48 schrieb John Snow:

Quoted 145 lines · Peter Lieven
Quoted 143 lines

PIO read requests on the ATAPI interface used to be sync blk requests. This has two significant drawbacks. First the main loop hangs util an I/O request is completed and secondly if the I/O request does not complete (e.g. due to an unresponsive storage) Qemu hangs completely.

---
  hw/ide/atapi.c | 93 ++++++++++++++++++++++++++++++++++++++++++++++++++++------
  1 file changed, 84 insertions(+), 9 deletions(-)
hw/ide/atapi.c +84 -8
index 747f466..2271ea2 100644
@@ -105,11 +105,16 @@ static void cd_data_to_raw(uint8_t *buf, int lba)
      memset(buf, 0, 288);
  }
 
-static int cd_read_sector(IDEState *s, int lba, uint8_t *buf, int sector_size)
+static int
+cd_read_sector_sync(IDEState *s, int lba, uint8_t *buf)
  {
      int ret;
 
-    switch(sector_size) {
+#ifdef DEBUG_IDE_ATAPI
+    printf("cd_read_sector_sync: lba=%d\n", lba);
+#endif
+
+    switch (s->cd_sector_size) {
      case 2048:
          block_acct_start(blk_get_stats(s->blk), &s->acct,
                           4 * BDRV_SECTOR_SIZE, BLOCK_ACCT_READ);
@@ -129,9 +134,71 @@ static int cd_read_sector(IDEState *s, int lba, uint8_t *buf, int sector_size)
          ret = -EIO;
          break;
      }
+
+    if (!ret) {
+        s->lba++;
+        s->io_buffer_index = 0;
+    }
+
      return ret;
  }
 
+static void cd_read_sector_cb(void *opaque, int ret)
+{
+    IDEState *s = opaque;
+
+    block_acct_done(blk_get_stats(s->blk), &s->acct);
+
+#ifdef DEBUG_IDE_ATAPI
+    printf("cd_read_sector_cb: lba=%d ret=%d\n", s->lba, ret);
+#endif
+
+    if (ret < 0) {
+        ide_atapi_io_error(s, ret);
+        return;
+    }
+
+    if (s->cd_sector_size == 2352) {
+        cd_data_to_raw(s->io_buffer, s->lba);
+    }
+
+    s->lba++;
+    s->io_buffer_index = 0;
+    s->status &= ~BUSY_STAT;
+
+    ide_atapi_cmd_reply_end(s);
+}
+
+static int cd_read_sector(IDEState *s, int lba, void *buf)
+{
+    if (s->cd_sector_size != 2048 && s->cd_sector_size != 2352) {
+        return -EINVAL;
+    }
+
+    s->iov.iov_base = buf;
+    if (s->cd_sector_size == 2352) {
+        buf += 16;
+    }
+
+    s->iov.iov_len = 4 * BDRV_SECTOR_SIZE;
+    qemu_iovec_init_external(&s->qiov, &s->iov, 1);
+
+#ifdef DEBUG_IDE_ATAPI
+    printf("cd_read_sector: lba=%d\n", lba);
+#endif
+
+    if (blk_aio_readv(s->blk, (int64_t)lba << 2, &s->qiov, 4,
+                      cd_read_sector_cb, s) == NULL) {
+        return -EIO;
+    }
+
+    block_acct_start(blk_get_stats(s->blk), &s->acct,
+                     4 * BDRV_SECTOR_SIZE, BLOCK_ACCT_READ);
+
+    s->status |= BUSY_STAT;
+    return 0;
+}
+
  void ide_atapi_cmd_ok(IDEState *s)
  {
      s->error = 0;
@@ -182,18 +249,27 @@ void ide_atapi_cmd_reply_end(IDEState *s)
          ide_atapi_cmd_ok(s);
          ide_set_irq(s->bus);
  #ifdef DEBUG_IDE_ATAPI
-        printf("status=0x%x\n", s->status);
+        printf("end of transfer, status=0x%x\n", s->status);
  #endif
      } else {
          /* see if a new sector must be read */
          if (s->lba != -1 && s->io_buffer_index >= s->cd_sector_size) {
-            ret = cd_read_sector(s, s->lba, s->io_buffer, s->cd_sector_size);
-            if (ret < 0) {
-                ide_atapi_io_error(s, ret);
+            if (!s->elementary_transfer_size) {
+                ret = cd_read_sector(s, s->lba, s->io_buffer);
+                if (ret < 0) {
+                    ide_atapi_io_error(s, ret);
+                }
                  return;
+            } else {
+                /* rebuffering within an elementary transfer is
+                 * only possible with a sync request because we
+                 * end up with a race condition otherwise */
+                ret = cd_read_sector_sync(s, s->lba, s->io_buffer);
+                if (ret < 0) {
+                    ide_atapi_io_error(s, ret);
+                    return;
+                }
              }
-            s->lba++;
-            s->io_buffer_index = 0;
          }
          if (s->elementary_transfer_size > 0) {
              /* there are some data left to transmit in this elementary
@@ -275,7 +351,6 @@ static void ide_atapi_cmd_read_pio(IDEState *s, int lba, int nb_sectors,
      s->io_buffer_index = sector_size;
      s->cd_sector_size = sector_size;
 
Quoted 8 lines

hw/ide/atapi.c static void ide_atapi_cmd_read_pio(IDEState *s, int lba, int nb_sectors, · line 356

Quoted 5 lines
hw/ide/atapi.c +0 -1
-    s->status = READY_STAT | SEEK_STAT;
      ide_atapi_cmd_reply_end(s);
  }
 

This patch looks good to me, apart from Stefan's other comments. Will you be sending a V3? I can try to merge it for this week before the hard freeze hits.

I will look at this today.

Peter

#6 lore raw

PATCH v2 2/4 ide/atapi: blk_aio_readv may return NULL

In reply to Peter Lieven
---
 hw/ide/atapi.c | 4 ++++
 1 file changed, 4 insertions(+)
hw/ide/atapi.c +4 -0
index 2271ea2..e0cf066 100644
@@ -429,6 +429,10 @@ static void ide_atapi_cmd_read_dma_cb(void *opaque, int ret)
     s->bus->dma->aiocb = blk_aio_readv(s->blk, (int64_t)s->lba << 2,
                                        &s->bus->dma->qiov, n * 4,
                                        ide_atapi_cmd_read_dma_cb, s);
+    if (s->bus->dma->aiocb == NULL) {
+        ide_atapi_io_error(s, -EIO);
+        goto eot;
+    }
     return;
 
 eot:
Signature · 1 line
1.9.1
#7 lore raw

Re: [Qemu-devel] [PATCH 2/4] ide/atapi: blk_aio_readv may return NULL

Quoted 10 lines · Peter Lieven
---
 hw/ide/atapi.c | 4 ++++
 1 file changed, 4 insertions(+)
hw/ide/atapi.c +0 -0
index 2271ea2..e0cf066 100644

hw/ide/atapi.c static void ide_atapi_cmd_read_dma_cb(void *opaque, int ret) · line 435

Quoted 7 lines
hw/ide/atapi.c +4 -0
@@ -429,6 +429,10 @@ static void ide_atapi_cmd_read_dma_cb(void *opaque, int ret)
     s->bus->dma->aiocb = blk_aio_readv(s->blk, (int64_t)s->lba << 2,
                                        &s->bus->dma->qiov, n * 4,
                                        ide_atapi_cmd_read_dma_cb, s);
+    if (s->bus->dma->aiocb == NULL) {
+        ide_atapi_io_error(s, -EIO);
+        goto eot;
+    }

Where does blk_aio_readv() return NULL?

#8 lore raw

Re: [Qemu-devel] [PATCH 2/4] ide/atapi: blk_aio_readv may return NULL

Am 22.10.2015 um 18:20 schrieb Stefan Hajnoczi:

Quoted 6 lines · Peter Lieven
Quoted 5 lines
---
 hw/ide/atapi.c | 4 ++++
 1 file changed, 4 insertions(+)
Quoted 13 lines

hw/ide/atapi.c static void ide_atapi_cmd_read_dma_cb(void *opaque, int ret) · line 435

Quoted 12 lines
hw/ide/atapi.c +4 -0
index 2271ea2..e0cf066 100644
@@ -429,6 +429,10 @@ static void ide_atapi_cmd_read_dma_cb(void *opaque, int ret)
     s->bus->dma->aiocb = blk_aio_readv(s->blk, (int64_t)s->lba << 2,
                                        &s->bus->dma->qiov, n * 4,
                                        ide_atapi_cmd_read_dma_cb, s);
+    if (s->bus->dma->aiocb == NULL) {
+        ide_atapi_io_error(s, -EIO);
+        goto eot;
+    }

Where does blk_aio_readv() return NULL?

Never. My fault.

Peter

#9 lore raw

PATCH v2 3/4 ide: add support for cancelable read requests

In reply to Peter Lieven

this patch adds a new aio readv compatible function which copies all data through a bounce buffer. The benefit is that these requests can be flagged as canceled to avoid guest memory corruption when a canceled request is completed by the backend at a later stage.

If an IDE protocol wants to use this function it has to pipe all read requests through ide_readv_cancelable and it may then enable requests_cancelable in the IDEState.

If this state is enable we can avoid the blocking blk_drain_all in case of a BMDMA reset.

Currently only read operations are cancelable thus we can only use this logic for read-only devices.

---
 hw/ide/core.c     | 54 ++++++++++++++++++++++++++++++++++++++++++++++++++++++
 hw/ide/internal.h | 16 ++++++++++++++++
 hw/ide/pci.c      | 42 ++++++++++++++++++++++++++++--------------
 3 files changed, 98 insertions(+), 14 deletions(-)
hw/ide/core.c +54 -0
index 317406d..24547ce 100644
@@ -561,6 +561,59 @@ static bool ide_sect_range_ok(IDEState *s,
     return true;
 }
 
+static void ide_readv_cancelable_cb(void *opaque, int ret)
+{
+    IDECancelableRequest *req = opaque;
+    if (!req->canceled) {
+        if (!ret) {
+            qemu_iovec_from_buf(req->org_qiov, 0, req->buf, req->org_qiov->size);
+        }
+        req->org_cb(req->org_opaque, ret);
+    }
+    QLIST_REMOVE(req, list);
+    qemu_vfree(req->buf);
+    qemu_iovec_destroy(&req->qiov);
+    g_free(req);
+}
+
+#define MAX_CANCELABLE_REQS 16
+
+BlockAIOCB *ide_readv_cancelable(IDEState *s, int64_t sector_num,
+                                 QEMUIOVector *iov, int nb_sectors,
+                                 BlockCompletionFunc *cb, void *opaque)
+{
+    BlockAIOCB *aioreq;
+    IDECancelableRequest *req;
+    int c = 0;
+
+    QLIST_FOREACH(req, &s->cancelable_requests, list) {
+        c++;
+    }
+    if (c > MAX_CANCELABLE_REQS) {
+        return NULL;
+    }
+
+    req = g_new0(IDECancelableRequest, 1);
+    qemu_iovec_init(&req->qiov, 1);
+    req->buf = qemu_blockalign(blk_bs(s->blk), iov->size);
+    qemu_iovec_add(&req->qiov, req->buf, iov->size);
+    req->org_qiov = iov;
+    req->org_cb = cb;
+    req->org_opaque = opaque;
+
+    aioreq = blk_aio_readv(s->blk, sector_num, &req->qiov, nb_sectors,
+                           ide_readv_cancelable_cb, req);
+    if (aioreq == NULL) {
+        qemu_vfree(req->buf);
+        qemu_iovec_destroy(&req->qiov);
+        g_free(req);
+    } else {
+        QLIST_INSERT_HEAD(&s->cancelable_requests, req, list);
+    }
+
+    return aioreq;
+}
+
 static void ide_sector_read(IDEState *s);
 
 static void ide_sector_read_cb(void *opaque, int ret)
@@ -805,6 +858,7 @@ void ide_start_dma(IDEState *s, BlockCompletionFunc *cb)
     s->bus->retry_unit = s->unit;
     s->bus->retry_sector_num = ide_get_sector(s);
     s->bus->retry_nsector = s->nsector;
+    s->bus->s = s;
     if (s->bus->dma->ops->start_dma) {
         s->bus->dma->ops->start_dma(s->bus->dma, s, cb);
     }
hw/ide/internal.h +16 -0
index 05e93ff..ad188c2 100644
@@ -343,6 +343,16 @@ enum ide_dma_cmd {
 #define ide_cmd_is_read(s) \
 	((s)->dma_cmd == IDE_DMA_READ)
 
+typedef struct IDECancelableRequest {
+    QLIST_ENTRY(IDECancelableRequest) list;
+    QEMUIOVector qiov;
+    uint8_t *buf;
+    QEMUIOVector *org_qiov;
+    BlockCompletionFunc *org_cb;
+    void *org_opaque;
+    bool canceled;
+} IDECancelableRequest;
+
 /* NOTE: IDEState represents in fact one drive */
 struct IDEState {
     IDEBus *bus;
@@ -396,6 +406,8 @@ struct IDEState {
     BlockAIOCB *pio_aiocb;
     struct iovec iov;
     QEMUIOVector qiov;
+    QLIST_HEAD(, IDECancelableRequest) cancelable_requests;
+    bool requests_cancelable;
     /* ATA DMA state */
     int32_t io_buffer_offset;
     int32_t io_buffer_size;
@@ -468,6 +480,7 @@ struct IDEBus {
     uint8_t retry_unit;
     int64_t retry_sector_num;
     uint32_t retry_nsector;
+    IDEState *s;
 };
 
 #define TYPE_IDE_DEVICE "ide-device"
@@ -572,6 +585,9 @@ void ide_set_inactive(IDEState *s, bool more);
 BlockAIOCB *ide_issue_trim(BlockBackend *blk,
         int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
         BlockCompletionFunc *cb, void *opaque);
+BlockAIOCB *ide_readv_cancelable(IDEState *s, int64_t sector_num,
+                                 QEMUIOVector *iov, int nb_sectors,
+                                 BlockCompletionFunc *cb, void *opaque);
 
 /* hw/ide/atapi.c */
 void ide_atapi_cmd(IDEState *s);
hw/ide/pci.c +28 -14

49 lines not on this page · raw message

Signature · 1 line
1.9.1
#10 lore raw

Re: [Qemu-devel] [PATCH 3/4] ide: add support for cancelable read requests

Quoted 9 lines · Peter Lieven

this patch adds a new aio readv compatible function which copies all data through a bounce buffer. The benefit is that these requests can be flagged as canceled to avoid guest memory corruption when a canceled request is completed by the backend at a later stage.

If an IDE protocol wants to use this function it has to pipe all read requests through ide_readv_cancelable and it may then enable requests_cancelable in the IDEState.

Quoted 5 lines

If this state is enable we can avoid the blocking blk_drain_all in case of a BMDMA reset.

Currently only read operations are cancelable thus we can only use this logic for read-only devices.

Naming is confusing here. Requests are already "cancelable" using bdv_aio_cancel().

Please use a different name, for example "orphan" requests. These are requests that QEMU still knows about but the guest believes are complete. Or maybe "IDEBufferedRequest" since data is transferred through a bounce buffer.

Quoted 36 lines
---
 hw/ide/core.c     | 54 ++++++++++++++++++++++++++++++++++++++++++++++++++++++
 hw/ide/internal.h | 16 ++++++++++++++++
 hw/ide/pci.c      | 42 ++++++++++++++++++++++++++++--------------
 3 files changed, 98 insertions(+), 14 deletions(-)
hw/ide/core.c +21 -0
index 317406d..24547ce 100644
@@ -561,6 +561,59 @@ static bool ide_sect_range_ok(IDEState *s,
     return true;
 }
 
+static void ide_readv_cancelable_cb(void *opaque, int ret)
+{
+    IDECancelableRequest *req = opaque;
+    if (!req->canceled) {
+        if (!ret) {
+            qemu_iovec_from_buf(req->org_qiov, 0, req->buf, req->org_qiov->size);
+        }
+        req->org_cb(req->org_opaque, ret);
+    }
+    QLIST_REMOVE(req, list);
+    qemu_vfree(req->buf);
+    qemu_iovec_destroy(&req->qiov);
+    g_free(req);
+}
+
+#define MAX_CANCELABLE_REQS 16
+
+BlockAIOCB *ide_readv_cancelable(IDEState *s, int64_t sector_num,
+                                 QEMUIOVector *iov, int nb_sectors,
+                                 BlockCompletionFunc *cb, void *opaque)
+{

hw/ide/core.c static bool ide_sect_range_ok(IDEState *s, · line 594

Quoted 10 lines
hw/ide/core.c +10 -0
+    BlockAIOCB *aioreq;
+    IDECancelableRequest *req;
+    int c = 0;
+
+    QLIST_FOREACH(req, &s->cancelable_requests, list) {
+        c++;
+    }
+    if (c > MAX_CANCELABLE_REQS) {
+        return NULL;
+    }

A BH is probably needed here to schedule an cb(-EIO) call since this function isn't supposed to return NULL if it's a direct replacement for blk_aio_readv().

hw/ide/core.c static bool ide_sect_range_ok(IDEState *s, · line 597

Quoted 3 lines
hw/ide/core.c +3 -0
+
+    req = g_new0(IDECancelableRequest, 1);
+    qemu_iovec_init(&req->qiov, 1);

It saves a g_new() call if you add a struct iovec field to IDECancelableRequest and use qemu_iovec_init_external() instead of qemu_iovec_init().

The qemu_iovec_destroy() calls must be dropped when an external struct iovec is used.

The qemu_iovec_init_external() call must be moved after the qemu_blockalign() and struct iovec setup below.

Quoted 23 lines
hw/ide/core.c +19 -0
+    req->buf = qemu_blockalign(blk_bs(s->blk), iov->size);
+    qemu_iovec_add(&req->qiov, req->buf, iov->size);
+    req->org_qiov = iov;
+    req->org_cb = cb;
+    req->org_opaque = opaque;
+
+    aioreq = blk_aio_readv(s->blk, sector_num, &req->qiov, nb_sectors,
+                           ide_readv_cancelable_cb, req);
+    if (aioreq == NULL) {
+        qemu_vfree(req->buf);
+        qemu_iovec_destroy(&req->qiov);
+        g_free(req);
+    } else {
+        QLIST_INSERT_HEAD(&s->cancelable_requests, req, list);
+    }
+
+    return aioreq;
+}
+
 static void ide_sector_read(IDEState *s);
 
 static void ide_sector_read_cb(void *opaque, int ret)

hw/ide/core.c void ide_start_dma(IDEState *s, BlockCompletionFunc *cb) · line 861

Quoted 4 lines
hw/ide/core.c +1 -0
@@ -805,6 +858,7 @@ void ide_start_dma(IDEState *s, BlockCompletionFunc *cb)
     s->bus->retry_unit = s->unit;
     s->bus->retry_sector_num = ide_get_sector(s);
     s->bus->retry_nsector = s->nsector;
+    s->bus->s = s;

How is 's' different from 'unit' and 'retry_unit'?

The logic for switching between units is already a little tricky since the guest can write to the hardware registers while requests are in-flight.

Please don't duplicate "active unit" state, that increases the risk of inconsistencies.

Can you use idebus_active_if() to get an equivalent IDEState pointer without storing s?

Quoted 8 lines
hw/ide/core.c +0 -0
     if (s->bus->dma->ops->start_dma) {
         s->bus->dma->ops->start_dma(s->bus->dma, s, cb);
     }
hw/ide/internal.h +0 -0
index 05e93ff..ad188c2 100644

hw/ide/internal.h enum ide_dma_cmd { · line 352

Quoted 10 lines
hw/ide/internal.h +7 -0
@@ -343,6 +343,16 @@ enum ide_dma_cmd {
 #define ide_cmd_is_read(s) \
 	((s)->dma_cmd == IDE_DMA_READ)
 
+typedef struct IDECancelableRequest {
+    QLIST_ENTRY(IDECancelableRequest) list;
+    QEMUIOVector qiov;
+    uint8_t *buf;
+    QEMUIOVector *org_qiov;
+    BlockCompletionFunc *org_cb;
+    void *org_opaque;

Please don't shorten names, original_* is clearer than org_*.

Quoted 73 lines
hw/ide/internal.h +9 -0
+    bool canceled;
+} IDECancelableRequest;
+
 /* NOTE: IDEState represents in fact one drive */
 struct IDEState {
     IDEBus *bus;
@@ -396,6 +406,8 @@ struct IDEState {
     BlockAIOCB *pio_aiocb;
     struct iovec iov;
     QEMUIOVector qiov;
+    QLIST_HEAD(, IDECancelableRequest) cancelable_requests;
+    bool requests_cancelable;
     /* ATA DMA state */
     int32_t io_buffer_offset;
     int32_t io_buffer_size;
@@ -468,6 +480,7 @@ struct IDEBus {
     uint8_t retry_unit;
     int64_t retry_sector_num;
     uint32_t retry_nsector;
+    IDEState *s;
 };
 
 #define TYPE_IDE_DEVICE "ide-device"
@@ -572,6 +585,9 @@ void ide_set_inactive(IDEState *s, bool more);
 BlockAIOCB *ide_issue_trim(BlockBackend *blk,
         int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
         BlockCompletionFunc *cb, void *opaque);
+BlockAIOCB *ide_readv_cancelable(IDEState *s, int64_t sector_num,
+                                 QEMUIOVector *iov, int nb_sectors,
+                                 BlockCompletionFunc *cb, void *opaque);
 
 /* hw/ide/atapi.c */
 void ide_atapi_cmd(IDEState *s);
hw/ide/pci.c +17 -14
index d31ff88..5587183 100644
@@ -240,21 +240,35 @@ void bmdma_cmd_writeb(BMDMAState *bm, uint32_t val)
     /* Ignore writes to SSBM if it keeps the old value */
     if ((val & BM_CMD_START) != (bm->cmd & BM_CMD_START)) {
         if (!(val & BM_CMD_START)) {
-            /*
-             * We can't cancel Scatter Gather DMA in the middle of the
-             * operation or a partial (not full) DMA transfer would reach
-             * the storage so we wait for completion instead (we beahve
-             * like if the DMA was completed by the time the guest trying
-             * to cancel dma with bmdma_cmd_writeb with BM_CMD_START not
-             * set).
-             *
-             * In the future we'll be able to safely cancel the I/O if the
-             * whole DMA operation will be submitted to disk with a single
-             * aio operation with preadv/pwritev.
-             */
             if (bm->bus->dma->aiocb) {
-                blk_drain_all();
-                assert(bm->bus->dma->aiocb == NULL);
+                if (bm->bus->s && bm->bus->s->requests_cancelable) {
+                    /*
+                     * If the used IDE protocol supports request cancelation we
+                     * can flag requests as canceled here and disable DMA.
+                     * The IDE protocol used MUST use ide_readv_cancelable for all
+                     * read operations and then subsequently can enable this code
+                     * path. Currently this is only supported for read-only
+                     * devices.
+                    */
+                    IDECancelableRequest *req;
+                    QLIST_FOREACH(req, &bm->bus->s->cancelable_requests, list) {
+                        if (!req->canceled) {
+                            req->org_cb(req->org_opaque, -ECANCELED);
+                        }
+                        req->canceled = true;
+                    }
+                } else {

hw/ide/pci.c void bmdma_cmd_writeb(BMDMAState *bm, uint32_t val) · line 270

Quoted 10 lines
hw/ide/pci.c +10 -0
+                    /*
+                     * We can't cancel Scatter Gather DMA in the middle of the
+                     * operation or a partial (not full) DMA transfer would reach
+                     * the storage so we wait for completion instead (we beahve
+                     * like if the DMA was completed by the time the guest trying
+                     * to cancel dma with bmdma_cmd_writeb with BM_CMD_START not
+                     * set).
+                     */
+                    blk_drain_all();
+                    assert(bm->bus->dma->aiocb == NULL);

This assertion applies is both branches of the if statement, it could be moved after the if statement.

Quoted 7 lines
hw/ide/pci.c +1 -0
+                }
             }
             bm->status &= ~BM_STATUS_DMAING;
         } else {
Signature · 1 line
1.9.1
#11 lore raw

Re: [Qemu-devel] [PATCH 3/4] ide: add support for cancelable read requests

Am 26.10.2015 um 11:39 schrieb Stefan Hajnoczi:

Quoted 64 lines · Peter Lieven
Quoted 9 lines

this patch adds a new aio readv compatible function which copies all data through a bounce buffer. The benefit is that these requests can be flagged as canceled to avoid guest memory corruption when a canceled request is completed by the backend at a later stage.

If an IDE protocol wants to use this function it has to pipe all read requests through ide_readv_cancelable and it may then enable requests_cancelable in the IDEState.

Quoted 5 lines

If this state is enable we can avoid the blocking blk_drain_all in case of a BMDMA reset.

Currently only read operations are cancelable thus we can only use this logic for read-only devices.

Naming is confusing here. Requests are already "cancelable" using bdv_aio_cancel().

Please use a different name, for example "orphan" requests. These are requests that QEMU still knows about but the guest believes are complete. Or maybe "IDEBufferedRequest" since data is transferred through a bounce buffer.

Quoted 41 lines
---
  hw/ide/core.c     | 54 ++++++++++++++++++++++++++++++++++++++++++++++++++++++
  hw/ide/internal.h | 16 ++++++++++++++++
  hw/ide/pci.c      | 42 ++++++++++++++++++++++++++++--------------
  3 files changed, 98 insertions(+), 14 deletions(-)
hw/ide/core.c +26 -0
index 317406d..24547ce 100644
@@ -561,6 +561,59 @@ static bool ide_sect_range_ok(IDEState *s,
      return true;
  }
 
+static void ide_readv_cancelable_cb(void *opaque, int ret)
+{
+    IDECancelableRequest *req = opaque;
+    if (!req->canceled) {
+        if (!ret) {
+            qemu_iovec_from_buf(req->org_qiov, 0, req->buf, req->org_qiov->size);
+        }
+        req->org_cb(req->org_opaque, ret);
+    }
+    QLIST_REMOVE(req, list);
+    qemu_vfree(req->buf);
+    qemu_iovec_destroy(&req->qiov);
+    g_free(req);
+}
+
+#define MAX_CANCELABLE_REQS 16
+
+BlockAIOCB *ide_readv_cancelable(IDEState *s, int64_t sector_num,
+                                 QEMUIOVector *iov, int nb_sectors,
+                                 BlockCompletionFunc *cb, void *opaque)
+{
+    BlockAIOCB *aioreq;
+    IDECancelableRequest *req;
+    int c = 0;
+
+    QLIST_FOREACH(req, &s->cancelable_requests, list) {
Quoted 8 lines

hw/ide/core.c static bool ide_sect_range_ok(IDEState *s, · line 594

Quoted 5 lines
hw/ide/core.c +5 -0
+        c++;
+    }
+    if (c > MAX_CANCELABLE_REQS) {
+        return NULL;
+    }

A BH is probably needed here to schedule an cb(-EIO) call since this function isn't supposed to return NULL if it's a direct replacement for blk_aio_readv().

You mean sth like:

acb = qemu_aio_get(&bdrv_em_aiocb_info, bs, cb, opaque); acb->bh = aio_bh_new(bdrv_get_aio_context(bs), bdrv_aio_bh_cb, acb); acb->ret = -EIO;
qemu_bh_schedule(acb->bh);

return &acb->common;

Quoted 8 lines

hw/ide/core.c static bool ide_sect_range_ok(IDEState *s, · line 597

Quoted 4 lines
hw/ide/core.c +3 -0
+
+    req = g_new0(IDECancelableRequest, 1);
+    qemu_iovec_init(&req->qiov, 1);

It saves a g_new() call if you add a struct iovec field to IDECancelableRequest and use qemu_iovec_init_external() instead of qemu_iovec_init().

Quoted 5 lines

The qemu_iovec_destroy() calls must be dropped when an external struct iovec is used.

The qemu_iovec_init_external() call must be moved after the qemu_blockalign() and struct iovec setup below.

okay

Quoted 34 lines
Quoted 24 lines
hw/ide/core.c +19 -0
+    req->buf = qemu_blockalign(blk_bs(s->blk), iov->size);
+    qemu_iovec_add(&req->qiov, req->buf, iov->size);
+    req->org_qiov = iov;
+    req->org_cb = cb;
+    req->org_opaque = opaque;
+
+    aioreq = blk_aio_readv(s->blk, sector_num, &req->qiov, nb_sectors,
+                           ide_readv_cancelable_cb, req);
+    if (aioreq == NULL) {
+        qemu_vfree(req->buf);
+        qemu_iovec_destroy(&req->qiov);
+        g_free(req);
+    } else {
+        QLIST_INSERT_HEAD(&s->cancelable_requests, req, list);
+    }
+
+    return aioreq;
+}
+
  static void ide_sector_read(IDEState *s);
 
  static void ide_sector_read_cb(void *opaque, int ret)

hw/ide/core.c void ide_start_dma(IDEState *s, BlockCompletionFunc *cb) · line 861

Quoted 4 lines
hw/ide/core.c +1 -0
@@ -805,6 +858,7 @@ void ide_start_dma(IDEState *s, BlockCompletionFunc *cb)
      s->bus->retry_unit = s->unit;
      s->bus->retry_sector_num = ide_get_sector(s);
      s->bus->retry_nsector = s->nsector;
+    s->bus->s = s;

How is 's' different from 'unit' and 'retry_unit'?

The logic for switching between units is already a little tricky since the guest can write to the hardware registers while requests are in-flight.

Quoted 5 lines

Please don't duplicate "active unit" state, that increases the risk of inconsistencies.

Can you use idebus_active_if() to get an equivalent IDEState pointer without storing s?

That should be possible.

Quoted 12 lines
Quoted 12 lines
hw/ide/core.c +0 -0
      if (s->bus->dma->ops->start_dma) {
          s->bus->dma->ops->start_dma(s->bus->dma, s, cb);
      }
hw/ide/internal.h +0 -0
index 05e93ff..ad188c2 100644
@@ -343,6 +343,16 @@ enum ide_dma_cmd {
  #define ide_cmd_is_read(s) \
  	((s)->dma_cmd == IDE_DMA_READ)
 
Quoted 8 lines

hw/ide/internal.h enum ide_dma_cmd { · line 352

Quoted 7 lines
hw/ide/internal.h +7 -0
+typedef struct IDECancelableRequest {
+    QLIST_ENTRY(IDECancelableRequest) list;
+    QEMUIOVector qiov;
+    uint8_t *buf;
+    QEMUIOVector *org_qiov;
+    BlockCompletionFunc *org_cb;
+    void *org_opaque;

Please don't shorten names, original_* is clearer than org_*.

Ok.

Quoted 78 lines
Quoted 78 lines
hw/ide/internal.h +9 -0
+    bool canceled;
+} IDECancelableRequest;
+
  /* NOTE: IDEState represents in fact one drive */
  struct IDEState {
      IDEBus *bus;
@@ -396,6 +406,8 @@ struct IDEState {
      BlockAIOCB *pio_aiocb;
      struct iovec iov;
      QEMUIOVector qiov;
+    QLIST_HEAD(, IDECancelableRequest) cancelable_requests;
+    bool requests_cancelable;
      /* ATA DMA state */
      int32_t io_buffer_offset;
      int32_t io_buffer_size;
@@ -468,6 +480,7 @@ struct IDEBus {
      uint8_t retry_unit;
      int64_t retry_sector_num;
      uint32_t retry_nsector;
+    IDEState *s;
  };
 
  #define TYPE_IDE_DEVICE "ide-device"
@@ -572,6 +585,9 @@ void ide_set_inactive(IDEState *s, bool more);
  BlockAIOCB *ide_issue_trim(BlockBackend *blk,
          int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
          BlockCompletionFunc *cb, void *opaque);
+BlockAIOCB *ide_readv_cancelable(IDEState *s, int64_t sector_num,
+                                 QEMUIOVector *iov, int nb_sectors,
+                                 BlockCompletionFunc *cb, void *opaque);
 
  /* hw/ide/atapi.c */
  void ide_atapi_cmd(IDEState *s);
hw/ide/pci.c +21 -14
index d31ff88..5587183 100644
@@ -240,21 +240,35 @@ void bmdma_cmd_writeb(BMDMAState *bm, uint32_t val)
      /* Ignore writes to SSBM if it keeps the old value */
      if ((val & BM_CMD_START) != (bm->cmd & BM_CMD_START)) {
          if (!(val & BM_CMD_START)) {
-            /*
-             * We can't cancel Scatter Gather DMA in the middle of the
-             * operation or a partial (not full) DMA transfer would reach
-             * the storage so we wait for completion instead (we beahve
-             * like if the DMA was completed by the time the guest trying
-             * to cancel dma with bmdma_cmd_writeb with BM_CMD_START not
-             * set).
-             *
-             * In the future we'll be able to safely cancel the I/O if the
-             * whole DMA operation will be submitted to disk with a single
-             * aio operation with preadv/pwritev.
-             */
              if (bm->bus->dma->aiocb) {
-                blk_drain_all();
-                assert(bm->bus->dma->aiocb == NULL);
+                if (bm->bus->s && bm->bus->s->requests_cancelable) {
+                    /*
+                     * If the used IDE protocol supports request cancelation we
+                     * can flag requests as canceled here and disable DMA.
+                     * The IDE protocol used MUST use ide_readv_cancelable for all
+                     * read operations and then subsequently can enable this code
+                     * path. Currently this is only supported for read-only
+                     * devices.
+                    */
+                    IDECancelableRequest *req;
+                    QLIST_FOREACH(req, &bm->bus->s->cancelable_requests, list) {
+                        if (!req->canceled) {
+                            req->org_cb(req->org_opaque, -ECANCELED);
+                        }
+                        req->canceled = true;
+                    }
+                } else {
+                    /*
+                     * We can't cancel Scatter Gather DMA in the middle of the
+                     * operation or a partial (not full) DMA transfer would reach
+                     * the storage so we wait for completion instead (we beahve
Quoted 8 lines

hw/ide/pci.c void bmdma_cmd_writeb(BMDMAState *bm, uint32_t val) · line 270

Quoted 6 lines
hw/ide/pci.c +6 -0
+                     * like if the DMA was completed by the time the guest trying
+                     * to cancel dma with bmdma_cmd_writeb with BM_CMD_START not
+                     * set).
+                     */
+                    blk_drain_all();
+                    assert(bm->bus->dma->aiocb == NULL);

This assertion applies is both branches of the if statement, it could be moved after the if statement.

Right.

As pointed out in my comment to your requestion about write/discard I think it should be feasible to use buffered readv requests for all read-only IDE devices. Only thing I'm unsure about is reopening. A reopen seems to only flush the device not drain all requests.

Peter

#12 lore raw

Re: [Qemu-devel] [PATCH 3/4] ide: add support for cancelable read requests

Quoted 20 lines · Peter Lieven

Am 26.10.2015 um 11:39 schrieb Stefan Hajnoczi:

Quoted 10 lines · Peter Lieven
Quoted 9 lines
hw/ide/core.c +9 -0
+BlockAIOCB *ide_readv_cancelable(IDEState *s, int64_t sector_num,
+                                 QEMUIOVector *iov, int nb_sectors,
+                                 BlockCompletionFunc *cb, void *opaque)
+{
+    BlockAIOCB *aioreq;
+    IDECancelableRequest *req;
+    int c = 0;
+
+    QLIST_FOREACH(req, &s->cancelable_requests, list) {
Quoted 8 lines

hw/ide/core.c static bool ide_sect_range_ok(IDEState *s, · line 594

Quoted 5 lines
hw/ide/core.c +5 -0
+        c++;
+    }
+    if (c > MAX_CANCELABLE_REQS) {
+        return NULL;
+    }

A BH is probably needed here to schedule an cb(-EIO) call since this function isn't supposed to return NULL if it's a direct replacement for blk_aio_readv().

Quoted 8 lines

You mean sth like:

acb = qemu_aio_get(&bdrv_em_aiocb_info, bs, cb, opaque); acb->bh = aio_bh_new(bdrv_get_aio_context(bs), bdrv_aio_bh_cb, acb); acb->ret = -EIO;
qemu_bh_schedule(acb->bh);

return &acb->common;

Yes.

Quoted 4 lines

As pointed out in my comment to your requestion about write/discard I think it should be feasible to use buffered readv requests for all read-only IDE devices. Only thing I'm unsure about is reopening. A reopen seems to only flush the device not drain all requests.

bdrv_reopen_prepare() callers should drain requests. For example, bdrv_reopen_multiple() (and indirectly bdrv_reopen()) call bdrv_drain_all(). Is this what you mean?

#13 lore raw

Re: [Qemu-devel] [PATCH 3/4] ide: add support for cancelable read requests

Am 28.10.2015 um 12:26 schrieb Stefan Hajnoczi:

Quoted 30 lines · Peter Lieven
Quoted 19 lines

Am 26.10.2015 um 11:39 schrieb Stefan Hajnoczi:

Quoted 10 lines · Peter Lieven
Quoted 9 lines
hw/ide/core.c +9 -0
+BlockAIOCB *ide_readv_cancelable(IDEState *s, int64_t sector_num,
+                                 QEMUIOVector *iov, int nb_sectors,
+                                 BlockCompletionFunc *cb, void *opaque)
+{
+    BlockAIOCB *aioreq;
+    IDECancelableRequest *req;
+    int c = 0;
+
+    QLIST_FOREACH(req, &s->cancelable_requests, list) {
Quoted 8 lines

hw/ide/core.c static bool ide_sect_range_ok(IDEState *s, · line 594

Quoted 5 lines
hw/ide/core.c +5 -0
+        c++;
+    }
+    if (c > MAX_CANCELABLE_REQS) {
+        return NULL;
+    }

A BH is probably needed here to schedule an cb(-EIO) call since this function isn't supposed to return NULL if it's a direct replacement for blk_aio_readv().

Quoted 8 lines

You mean sth like:

acb = qemu_aio_get(&bdrv_em_aiocb_info, bs, cb, opaque); acb->bh = aio_bh_new(bdrv_get_aio_context(bs), bdrv_aio_bh_cb, acb); acb->ret = -EIO;
qemu_bh_schedule(acb->bh);

return &acb->common;

Yes.

Quoted 7 lines
Quoted 4 lines

As pointed out in my comment to your requestion about write/discard I think it should be feasible to use buffered readv requests for all read-only IDE devices. Only thing I'm unsure about is reopening. A reopen seems to only flush the device not drain all requests.

bdrv_reopen_prepare() callers should drain requests. For example, bdrv_reopen_multiple() (and indirectly bdrv_reopen()) call bdrv_drain_all(). Is this what you mean?

Yes, I have only found a flush in bdrv_reopen_prepare, but if you say they need to drain before I think it is safe to use the buffered_ide_readv for all read-only IDE devices not only CDROMs.

Peter

#14 lore raw

PATCH v2 4/4 ide/atapi: enable cancelable requests

In reply to Peter Lieven
---
 hw/ide/atapi.c | 4 ++--
 hw/ide/core.c  | 1 +
 2 files changed, 3 insertions(+), 2 deletions(-)
hw/ide/atapi.c +2 -2
index e0cf066..8d38b1d 100644
@@ -187,7 +187,7 @@ static int cd_read_sector(IDEState *s, int lba, void *buf)
     printf("cd_read_sector: lba=%d\n", lba);
 #endif
 
-    if (blk_aio_readv(s->blk, (int64_t)lba << 2, &s->qiov, 4,
+    if (ide_readv_cancelable(s, (int64_t)lba << 2, &s->qiov, 4,
                       cd_read_sector_cb, s) == NULL) {
         return -EIO;
     }
@@ -426,7 +426,7 @@ static void ide_atapi_cmd_read_dma_cb(void *opaque, int ret)
     s->bus->dma->iov.iov_len = n * 4 * 512;
     qemu_iovec_init_external(&s->bus->dma->qiov, &s->bus->dma->iov, 1);
 
-    s->bus->dma->aiocb = blk_aio_readv(s->blk, (int64_t)s->lba << 2,
+    s->bus->dma->aiocb = ide_readv_cancelable(s, (int64_t)s->lba << 2,
                                        &s->bus->dma->qiov, n * 4,
                                        ide_atapi_cmd_read_dma_cb, s);
     if (s->bus->dma->aiocb == NULL) {
hw/ide/core.c +1 -0
index 24547ce..5c7a346 100644
@@ -2330,6 +2330,7 @@ int ide_init_drive(IDEState *s, BlockBackend *blk, IDEDriveKind kind,
     if (kind == IDE_CD) {
         blk_set_dev_ops(blk, &ide_cd_block_ops, s);
         blk_set_guest_block_size(blk, 2048);
+        s->requests_cancelable = true;
     } else {
         if (!blk_is_inserted(s->blk)) {
             error_report("Device needs media, but drive is empty");
Signature · 1 line
1.9.1
#15 lore raw
In reply to Peter Lieven
Quoted 38 lines · Peter Lieven

This series aims at avoiding a hanging main-loop if a vserver has a CDROM image mounted from a NFS share and that NFS share goes down. Typical situation is that users mount an CDROM ISO to install something and then forget to eject that CDROM afterwards.
As a consequence this mounted CD is able to bring down the whole vserver if the backend NFS share is unreachable. This is bad especially if the CDROM itself is not needed anymore at this point.

This series aims at fixing 2 blocking I/O operations that would hang if the NFS server is unavailable:
- ATAPI PIO read requests used sync calls to blk_read, convert them to an async variant where possible.
- If a busmaster DMA request is cancelled all requests are drained. Convert the drain to an async request canceling.

v1->v2: - fix offset for 2352 byte sector size [Kevin]
- use a sync request if we continue an elementary transfer. As John pointed out we enter a race condition between next IDE command and async transfer otherwise. This is sill not optimal, but it fixes the NFS down problems for all cases where the NFS server goes down while there is no PIO CD activity. Of course, it could still happen during a PIO transfer, but I expect this to be the unlikelier case.
I spent some effort trying to read more sectors at once and avoiding continuation of elementary transfers, but with whatever I came up it was destroying migration between different Qemu versions. I have a quite hackish patch that works and should survive migration, but I am not happy with it. So I would like to start with this version as it is a big improvement already.
- Dropped Patch 5 because it is upstream meanwhile.

Peter Lieven (4):
ide/atapi: make PIO read requests async
ide/atapi: blk_aio_readv may return NULL
ide: add support for cancelable read requests
ide/atapi: enable cancelable requests

Quoted 5 lines
 hw/ide/atapi.c    | 99 +++++++++++++++++++++++++++++++++++++++++++++++++------
 hw/ide/core.c     | 55 +++++++++++++++++++++++++++++++
 hw/ide/internal.h | 16 +++++++++
 hw/ide/pci.c      | 42 +++++++++++++++--------
 4 files changed, 188 insertions(+), 24 deletions(-)

Any reason why write and discard requests aren't covered in this series?

If this is a good idea for CD-ROM it should be a good idea for all PCI IDE devices.

Having a specialized code path is often a sign that it hasn't been tested enough. Can we get confident enough to enable this everywhere?

#16 lore raw

Am 26.10.2015 um 11:42 schrieb Stefan Hajnoczi:

Quoted 44 lines · Peter Lieven
Quoted 38 lines

This series aims at avoiding a hanging main-loop if a vserver has a CDROM image mounted from a NFS share and that NFS share goes down. Typical situation is that users mount an CDROM ISO to install something and then forget to eject that CDROM afterwards.
As a consequence this mounted CD is able to bring down the whole vserver if the backend NFS share is unreachable. This is bad especially if the CDROM itself is not needed anymore at this point.

This series aims at fixing 2 blocking I/O operations that would hang if the NFS server is unavailable:
- ATAPI PIO read requests used sync calls to blk_read, convert them to an async variant where possible.
- If a busmaster DMA request is cancelled all requests are drained. Convert the drain to an async request canceling.

v1->v2: - fix offset for 2352 byte sector size [Kevin]
- use a sync request if we continue an elementary transfer. As John pointed out we enter a race condition between next IDE command and async transfer otherwise. This is sill not optimal, but it fixes the NFS down problems for all cases where the NFS server goes down while there is no PIO CD activity. Of course, it could still happen during a PIO transfer, but I expect this to be the unlikelier case.
I spent some effort trying to read more sectors at once and avoiding continuation of elementary transfers, but with whatever I came up it was destroying migration between different Qemu versions. I have a quite hackish patch that works and should survive migration, but I am not happy with it. So I would like to start with this version as it is a big improvement already.
- Dropped Patch 5 because it is upstream meanwhile.

Peter Lieven (4):
ide/atapi: make PIO read requests async
ide/atapi: blk_aio_readv may return NULL
ide: add support for cancelable read requests
ide/atapi: enable cancelable requests

Quoted 5 lines
  hw/ide/atapi.c    | 99 +++++++++++++++++++++++++++++++++++++++++++++++++------
  hw/ide/core.c     | 55 +++++++++++++++++++++++++++++++
  hw/ide/internal.h | 16 +++++++++
  hw/ide/pci.c      | 42 +++++++++++++++--------
  4 files changed, 188 insertions(+), 24 deletions(-)
Quoted 7 lines

Any reason why write and discard requests aren't covered in this series?

If this is a good idea for CD-ROM it should be a good idea for all PCI IDE devices.

Having a specialized code path is often a sign that it hasn't been tested enough. Can we get confident enough to enable this everywhere?

The reason is that the buffered request trick does only work for read-only devices (like a CDROM). A write request that is completed on the backend storage at a later point (after the OS thinks the request is canceled) can cause damage to the filesystem.

Peter

#17 lore raw
Quoted 50 lines · Peter Lieven

Am 26.10.2015 um 11:42 schrieb Stefan Hajnoczi:

Quoted 49 lines · Peter Lieven
Quoted 38 lines

This series aims at avoiding a hanging main-loop if a vserver has a CDROM image mounted from a NFS share and that NFS share goes down. Typical situation is that users mount an CDROM ISO to install something and then forget to eject that CDROM afterwards.
As a consequence this mounted CD is able to bring down the whole vserver if the backend NFS share is unreachable. This is bad especially if the CDROM itself is not needed anymore at this point.

This series aims at fixing 2 blocking I/O operations that would hang if the NFS server is unavailable:
- ATAPI PIO read requests used sync calls to blk_read, convert them to an async variant where possible.
- If a busmaster DMA request is cancelled all requests are drained. Convert the drain to an async request canceling.

v1->v2: - fix offset for 2352 byte sector size [Kevin]
- use a sync request if we continue an elementary transfer. As John pointed out we enter a race condition between next IDE command and async transfer otherwise. This is sill not optimal, but it fixes the NFS down problems for all cases where the NFS server goes down while there is no PIO CD activity. Of course, it could still happen during a PIO transfer, but I expect this to be the unlikelier case.
I spent some effort trying to read more sectors at once and avoiding continuation of elementary transfers, but with whatever I came up it was destroying migration between different Qemu versions. I have a quite hackish patch that works and should survive migration, but I am not happy with it. So I would like to start with this version as it is a big improvement already.
- Dropped Patch 5 because it is upstream meanwhile.

Peter Lieven (4):
ide/atapi: make PIO read requests async
ide/atapi: blk_aio_readv may return NULL
ide: add support for cancelable read requests
ide/atapi: enable cancelable requests

Quoted 5 lines
 hw/ide/atapi.c    | 99 +++++++++++++++++++++++++++++++++++++++++++++++++------
 hw/ide/core.c     | 55 +++++++++++++++++++++++++++++++
 hw/ide/internal.h | 16 +++++++++
 hw/ide/pci.c      | 42 +++++++++++++++--------
 4 files changed, 188 insertions(+), 24 deletions(-)

Any reason why write and discard requests aren't covered in this series?

If this is a good idea for CD-ROM it should be a good idea for all PCI IDE devices.

Quoted 7 lines
Quoted 2 lines

Having a specialized code path is often a sign that it hasn't been tested enough. Can we get confident enough to enable this everywhere?

The reason is that the buffered request trick does only work for read-only devices (like a CDROM). A write request that is completed on the backend storage at a later point (after the OS thinks the request is canceled) can cause damage to the filesystem.

Of course, you are right.

This is really annoying because it means a guest cannot reboot if writes are pending...

Stefan