Skip to contentSkip to conversation

Rebase aio dma cancellation work

PATCH In review Ryan Harper 5 participants 17 messages Apr 20, 2009 – May 14, 2009

lore
Share & tools
lore
#1 lore raw

PATCH 1/3 Refactor aio callback allocation to use an aiocb pool

Re-based against qemu-stable 0.10 branch

Move the AIOCB allocation code to use a dedicate structure, AIOPool. AIOCB specific information, such as the AIOCB size and cancellation routine, is moved into the pool.

At present, there is exactly one pool per block format driver, maintaining the status quo.

block.c +27 -15
index 3f262fa..1aa3de5 100644
@@ -141,6 +141,7 @@ static void bdrv_register(BlockDriver *bdrv)
         bdrv->bdrv_read = bdrv_read_em;
         bdrv->bdrv_write = bdrv_write_em;
     }
+    aio_pool_init(&bdrv->aio_pool, bdrv->aiocb_size, bdrv->bdrv_aio_cancel);
     bdrv->next = first_drv;
     first_drv = bdrv;
 }
@@ -1443,14 +1444,12 @@ BlockDriverAIOCB *bdrv_aio_write(BlockDriverState *bs, int64_t sector_num,
 
 void bdrv_aio_cancel(BlockDriverAIOCB *acb)
 {
-    BlockDriver *drv = acb->bs->drv;
-
     if (acb->cb == bdrv_aio_rw_vector_cb) {
         VectorTranslationState *s = acb->opaque;
         acb = s->aiocb;
     }
 
-    drv->bdrv_aio_cancel(acb);
+    acb->pool->cancel(acb);
 }
 
 
@@ -1568,18 +1567,25 @@ void bdrv_init(void)
     bdrv_register(&bdrv_nbd);
 }
 
-void *qemu_aio_get(BlockDriverState *bs, BlockDriverCompletionFunc *cb,
-                   void *opaque)
+void aio_pool_init(AIOPool *pool, int aiocb_size,
+                   void (*cancel)(BlockDriverAIOCB *acb))
+{
+    pool->aiocb_size = aiocb_size;
+    pool->cancel = cancel;
+    pool->free_aiocb = NULL;
+}
+
+void *qemu_aio_get_pool(AIOPool *pool, BlockDriverState *bs,
+                        BlockDriverCompletionFunc *cb, void *opaque)
 {
-    BlockDriver *drv;
     BlockDriverAIOCB *acb;
 
-    drv = bs->drv;
-    if (drv->free_aiocb) {
-        acb = drv->free_aiocb;
-        drv->free_aiocb = acb->next;
+    if (pool->free_aiocb) {
+        acb = pool->free_aiocb;
+        pool->free_aiocb = acb->next;
     } else {
-        acb = qemu_mallocz(drv->aiocb_size);
+        acb = qemu_mallocz(pool->aiocb_size);
+        acb->pool = pool;
     }
     acb->bs = bs;
     acb->cb = cb;
@@ -1587,12 +1593,18 @@ void *qemu_aio_get(BlockDriverState *bs, BlockDriverCompletionFunc *cb,
     return acb;
 }
 
+void *qemu_aio_get(BlockDriverState *bs, BlockDriverCompletionFunc *cb,
+                   void *opaque)
+{
+    return qemu_aio_get_pool(&bs->drv->aio_pool, bs, cb, opaque);
+}
+
 void qemu_aio_release(void *p)
 {
-    BlockDriverAIOCB *acb = p;
-    BlockDriver *drv = acb->bs->drv;
-    acb->next = drv->free_aiocb;
-    drv->free_aiocb = acb;
+    BlockDriverAIOCB *acb = (BlockDriverAIOCB *)p;
+    AIOPool *pool = acb->pool;
+    acb->next = pool->free_aiocb;
+    pool->free_aiocb = acb;
 }
 
 /**************************************************************/
block_int.h +13 -1
index c6af1c6..e581637 100644
@@ -30,6 +30,12 @@
 #define BLOCK_FLAG_COMPRESS	2
 #define BLOCK_FLAG_COMPAT6	4
 
+typedef struct AIOPool {
+    void (*cancel)(BlockDriverAIOCB *acb);
+    int aiocb_size;
+    BlockDriverAIOCB *free_aiocb;
+} AIOPool;
+
 struct BlockDriver {
     const char *format_name;
     int instance_size;
@@ -90,7 +96,7 @@ struct BlockDriver {
     /* to control generic scsi devices */
     int (*bdrv_ioctl)(BlockDriverState *bs, unsigned long int req, void *buf);
 
-    BlockDriverAIOCB *free_aiocb;
+    AIOPool aio_pool;
     struct BlockDriver *next;
 };
 
@@ -140,6 +146,7 @@ struct BlockDriverState {
 };
 
 struct BlockDriverAIOCB {
+    AIOPool *pool;
     BlockDriverState *bs;
     BlockDriverCompletionFunc *cb;
     void *opaque;
@@ -148,8 +155,13 @@ struct BlockDriverAIOCB {
 
 void get_tmp_filename(char *filename, int size);
 
+void aio_pool_init(AIOPool *pool, int aiocb_size,
+                   void (*cancel)(BlockDriverAIOCB *acb));
+
 void *qemu_aio_get(BlockDriverState *bs, BlockDriverCompletionFunc *cb,
                    void *opaque);
+void *qemu_aio_get_pool(AIOPool *pool, BlockDriverState *bs,
+                        BlockDriverCompletionFunc *cb, void *opaque);
 void qemu_aio_release(void *p);
 
 extern BlockDriverState *bdrv_first;
#2 lore raw

PATCH 2/3 Convert vectored aio emulation to use a dedicated pool

In reply to Ryan Harper

Rebased on qemu-stable.

This allows us to remove a hack in the vectored aio cancellation code.

block.c +14 -6
index 1aa3de5..465a09f 100644
@@ -41,6 +41,8 @@
 #define SECTOR_BITS 9
 #define SECTOR_SIZE (1 << SECTOR_BITS)
 
+static AIOPool vectored_aio_pool;
+
 typedef struct BlockDriverAIOCBSync {
     BlockDriverAIOCB common;
     QEMUBH *bh;
@@ -1333,6 +1335,13 @@ typedef struct VectorTranslationState {
     BlockDriverAIOCB *this_aiocb;
 } VectorTranslationState;
 
+static void bdrv_aio_cancel_vector(BlockDriverAIOCB *acb)
+{
+    VectorTranslationState *s = acb->opaque;
+
+    bdrv_aio_cancel(s->aiocb);
+}
+
 static void bdrv_aio_rw_vector_cb(void *opaque, int ret)
 {
     VectorTranslationState *s = opaque;
@@ -1355,7 +1364,8 @@ static BlockDriverAIOCB *bdrv_aio_rw_vector(BlockDriverState *bs,
 
 {
     VectorTranslationState *s = qemu_mallocz(sizeof(*s));
-    BlockDriverAIOCB *aiocb = qemu_aio_get(bs, cb, opaque);
+    BlockDriverAIOCB *aiocb = qemu_aio_get_pool(&vectored_aio_pool, bs,
+                                                cb, opaque);
 
     s->this_aiocb = aiocb;
     s->iov = iov;
@@ -1444,11 +1454,6 @@ BlockDriverAIOCB *bdrv_aio_write(BlockDriverState *bs, int64_t sector_num,
 
 void bdrv_aio_cancel(BlockDriverAIOCB *acb)
 {
-    if (acb->cb == bdrv_aio_rw_vector_cb) {
-        VectorTranslationState *s = acb->opaque;
-        acb = s->aiocb;
-    }
-
     acb->pool->cancel(acb);
 }
 
@@ -1550,6 +1555,9 @@ static int bdrv_write_em(BlockDriverState *bs, int64_t sector_num,
 
 void bdrv_init(void)
 {
+    aio_pool_init(&vectored_aio_pool, sizeof(BlockDriverAIOCB),
+                  bdrv_aio_cancel_vector);
+
     bdrv_register(&bdrv_raw);
     bdrv_register(&bdrv_host_device);
 #ifndef _WIN32
#3 lore raw

PATCH 3/3 Implement cancellation method for dma async I/O

In reply to Ryan Harper

Rebased on qemu-stable.

Move the dma helpers to a private aio pool, and implement a cancellation method for them. Should prevent issues when cancelling I/O while dma is in progress.

dma-helpers.c +14 -1
index b2ade19..19fa4f0 100644
@@ -10,6 +10,8 @@
 #include "dma.h"
 #include "block_int.h"
 
+static AIOPool dma_aio_pool;
+
 void qemu_sglist_init(QEMUSGList *qsg, int alloc_hint)
 {
     qsg->sg = qemu_malloc(alloc_hint * sizeof(ScatterGatherEntry));
@@ -126,7 +128,7 @@ static BlockDriverAIOCB *dma_bdrv_io(
     DMABlockState *dbs = qemu_malloc(sizeof(*dbs));
 
     dbs->bs = bs;
-    dbs->acb = qemu_aio_get(bs, cb, opaque);
+    dbs->acb = qemu_aio_get_pool(&dma_aio_pool, bs, cb, opaque);
     dbs->sg = sg;
     dbs->sector_num = sector_num;
     dbs->sg_cur_index = 0;
@@ -153,3 +155,14 @@ BlockDriverAIOCB *dma_bdrv_write(BlockDriverState *bs,
     return dma_bdrv_io(bs, sg, sector, cb, opaque, 1);
 }
 
+static void dma_aio_cancel(BlockDriverAIOCB *acb)
+{
+    DMABlockState *dbs = (DMABlockState *)acb->opaque;
+
+    bdrv_aio_cancel(dbs->acb);
+}
+
+void dma_helper_init(void)
+{
+    aio_pool_init(&dma_aio_pool, sizeof(BlockDriverAIOCB), dma_aio_cancel);
+}
dma.h +1 -0
index d596717..7d2ab45 100644
@@ -37,5 +37,6 @@ BlockDriverAIOCB *dma_bdrv_read(BlockDriverState *bs,
 BlockDriverAIOCB *dma_bdrv_write(BlockDriverState *bs,
                                  QEMUSGList *sg, uint64_t sector,
                                  BlockDriverCompletionFunc *cb, void *opaque);
+void dma_helper_init(void);
 
 #endif
vl.c +1 -0
index 54ae45b..d6d8307 100644
@@ -5485,6 +5485,7 @@ int main(int argc, char **argv, char **envp)
     cpu_exec_init_all(tb_size * 1024 * 1024);
 
     bdrv_init();
+    dma_helper_init();
 
     /* we always create the cdrom drive, even if no disk is there */
 
#5 lore raw
Quoted 11 lines · Anthony Liguori
Quoted 5 lines · Ryan Harper

Rebased Avi's aio patch series for dma cancellation to stable branch. Cursory tested Linux 64-bit guest and WinXP SP2.

Avi,

Are the first 3 of your original 5 patches what's suitable for stable?

Yes.

Of course, the intermediate state wasn't tested as well as the complete patchset, but I don't think there's any issue lurking in there.

Signature · 1 line
error compiling committee.c: too many arguments to function
#6 lore raw

cover Rebase more aio dma cancellation work

In reply to Ryan Harper

Hi,
We've a report of F11 qemu segfaulting when using read-only
images here:

https://bugzilla.redhat.com/497170

Looking at the current code in stable, there is clearly some
borkage:

- No error handling in bdrv_aio_rw_vector()

- No error handling in dma_bdrv_cb()/dma_bdrv_io()

- VectorTranslationState is leaked by bdrv_aio_rw_vector_cb()

- dma_aio_cancel() is cancelling the wrong aio

The rest of Avi's aio cancellation patches actually fixes all
these problems. Please apply to stable.

Cheers,
Mark.

#7 lore raw

PATCH 1/4 Use vectored aiocb storage to store vector translation state

From: Avi Kivity <[email protected]>

Now that we have a dedicated acb pool for vector translation acbs, we can store the vector translation state in the acbs instead of in an external structure.

---
 block.c |   29 ++++++++++++++---------------
 1 files changed, 14 insertions(+), 15 deletions(-)
block.c +14 -15
index b12318f..689ea37 100644
@@ -1332,31 +1332,32 @@ char *bdrv_snapshot_dump(char *buf, int buf_size, QEMUSnapshotInfo *sn)
 /**************************************************************/
 /* async I/Os */
 
-typedef struct VectorTranslationState {
+typedef struct VectorTranslationAIOCB {
+    BlockDriverAIOCB common;
     QEMUIOVector *iov;
     uint8_t *bounce;
     int is_write;
     BlockDriverAIOCB *aiocb;
-    BlockDriverAIOCB *this_aiocb;
-} VectorTranslationState;
+} VectorTranslationAIOCB;
 
-static void bdrv_aio_cancel_vector(BlockDriverAIOCB *acb)
+static void bdrv_aio_cancel_vector(BlockDriverAIOCB *_acb)
 {
-    VectorTranslationState *s = acb->opaque;
+    VectorTranslationAIOCB *acb
+        = container_of(_acb, VectorTranslationAIOCB, common);
 
-    bdrv_aio_cancel(s->aiocb);
+    bdrv_aio_cancel(acb->aiocb);
 }
 
 static void bdrv_aio_rw_vector_cb(void *opaque, int ret)
 {
-    VectorTranslationState *s = opaque;
+    VectorTranslationAIOCB *s = (VectorTranslationAIOCB *)opaque;
 
     if (!s->is_write) {
         qemu_iovec_from_buffer(s->iov, s->bounce, s->iov->size);
     }
     qemu_vfree(s->bounce);
-    s->this_aiocb->cb(s->this_aiocb->opaque, ret);
-    qemu_aio_release(s->this_aiocb);
+    s->common.cb(s->common.opaque, ret);
+    qemu_aio_release(s);
 }
 
 static BlockDriverAIOCB *bdrv_aio_rw_vector(BlockDriverState *bs,
@@ -1368,11 +1369,9 @@ static BlockDriverAIOCB *bdrv_aio_rw_vector(BlockDriverState *bs,
                                             int is_write)
 
 {
-    VectorTranslationState *s = qemu_mallocz(sizeof(*s));
-    BlockDriverAIOCB *aiocb = qemu_aio_get_pool(&vectored_aio_pool, bs,
-                                                cb, opaque);
+    VectorTranslationAIOCB *s = qemu_aio_get_pool(&vectored_aio_pool, bs,
+                                                  cb, opaque);
 
-    s->this_aiocb = aiocb;
     s->iov = iov;
     s->bounce = qemu_memalign(512, nb_sectors * 512);
     s->is_write = is_write;
@@ -1384,7 +1383,7 @@ static BlockDriverAIOCB *bdrv_aio_rw_vector(BlockDriverState *bs,
         s->aiocb = bdrv_aio_read(bs, sector_num, s->bounce, nb_sectors,
                                  bdrv_aio_rw_vector_cb, s);
     }
-    return aiocb;
+    return &s->common;
 }
 
 BlockDriverAIOCB *bdrv_aio_readv(BlockDriverState *bs, int64_t sector_num,
@@ -1560,7 +1559,7 @@ static int bdrv_write_em(BlockDriverState *bs, int64_t sector_num,
 
 void bdrv_init(void)
 {
-    aio_pool_init(&vectored_aio_pool, sizeof(BlockDriverAIOCB),
+    aio_pool_init(&vectored_aio_pool, sizeof(VectorTranslationAIOCB),
                   bdrv_aio_cancel_vector);
 
     bdrv_register(&bdrv_raw);
Signature · 1 line
1.6.0.6
#8 lore raw

PATCH 2/4 Move block dma helpers aiocb to store dma state

From: Avi Kivity <[email protected]>

Use the dedicated dma aiocb to store intermediate state for dma block transactions.

---
 dma-helpers.c |   37 ++++++++++++++++++++-----------------
 1 files changed, 20 insertions(+), 17 deletions(-)
dma-helpers.c +20 -17
index 19fa4f0..96a120c 100644
@@ -39,6 +39,7 @@ void qemu_sglist_destroy(QEMUSGList *qsg)
 }
 
 typedef struct {
+    BlockDriverAIOCB common;
     BlockDriverState *bs;
     BlockDriverAIOCB *acb;
     QEMUSGList *sg;
@@ -48,13 +49,13 @@ typedef struct {
     target_phys_addr_t sg_cur_byte;
     QEMUIOVector iov;
     QEMUBH *bh;
-} DMABlockState;
+} DMAAIOCB;
 
 static void dma_bdrv_cb(void *opaque, int ret);
 
 static void reschedule_dma(void *opaque)
 {
-    DMABlockState *dbs = (DMABlockState *)opaque;
+    DMAAIOCB *dbs = (DMAAIOCB *)opaque;
 
     qemu_bh_delete(dbs->bh);
     dbs->bh = NULL;
@@ -63,7 +64,7 @@ static void reschedule_dma(void *opaque)
 
 static void continue_after_map_failure(void *opaque)
 {
-    DMABlockState *dbs = (DMABlockState *)opaque;
+    DMAAIOCB *dbs = (DMAAIOCB *)opaque;
 
     dbs->bh = qemu_bh_new(reschedule_dma, dbs);
     qemu_bh_schedule(dbs->bh);
@@ -71,11 +72,12 @@ static void continue_after_map_failure(void *opaque)
 
 static void dma_bdrv_cb(void *opaque, int ret)
 {
-    DMABlockState *dbs = (DMABlockState *)opaque;
+    DMAAIOCB *dbs = (DMAAIOCB *)opaque;
     target_phys_addr_t cur_addr, cur_len;
     void *mem;
     int i;
 
+    dbs->acb = NULL;
     dbs->sector_num += dbs->iov.size / 512;
     for (i = 0; i < dbs->iov.niov; ++i) {
         cpu_physical_memory_unmap(dbs->iov.iov[i].iov_base,
@@ -85,10 +87,9 @@ static void dma_bdrv_cb(void *opaque, int ret)
     qemu_iovec_reset(&dbs->iov);
 
     if (dbs->sg_cur_index == dbs->sg->nsg || ret < 0) {
-        dbs->acb->cb(dbs->acb->opaque, ret);
+        dbs->common.cb(dbs->common.opaque, ret);
         qemu_iovec_destroy(&dbs->iov);
-        qemu_aio_release(dbs->acb);
-        qemu_free(dbs);
+        qemu_aio_release(dbs);
         return;
     }
 
@@ -112,11 +113,11 @@ static void dma_bdrv_cb(void *opaque, int ret)
     }
 
     if (dbs->is_write) {
-        bdrv_aio_writev(dbs->bs, dbs->sector_num, &dbs->iov,
-                        dbs->iov.size / 512, dma_bdrv_cb, dbs);
+        dbs->acb = bdrv_aio_writev(dbs->bs, dbs->sector_num, &dbs->iov,
+                                   dbs->iov.size / 512, dma_bdrv_cb, dbs);
     } else {
-        bdrv_aio_readv(dbs->bs, dbs->sector_num, &dbs->iov,
-                       dbs->iov.size / 512, dma_bdrv_cb, dbs);
+        dbs->acb = bdrv_aio_readv(dbs->bs, dbs->sector_num, &dbs->iov,
+                                  dbs->iov.size / 512, dma_bdrv_cb, dbs);
     }
 }
 
@@ -125,10 +126,10 @@ static BlockDriverAIOCB *dma_bdrv_io(
     BlockDriverCompletionFunc *cb, void *opaque,
     int is_write)
 {
-    DMABlockState *dbs = qemu_malloc(sizeof(*dbs));
+    DMAAIOCB *dbs =  qemu_aio_get_pool(&dma_aio_pool, bs, cb, opaque);
 
+    dbs->acb = NULL;
     dbs->bs = bs;
-    dbs->acb = qemu_aio_get_pool(&dma_aio_pool, bs, cb, opaque);
     dbs->sg = sg;
     dbs->sector_num = sector_num;
     dbs->sg_cur_index = 0;
@@ -137,7 +138,7 @@ static BlockDriverAIOCB *dma_bdrv_io(
     dbs->bh = NULL;
     qemu_iovec_init(&dbs->iov, sg->nsg);
     dma_bdrv_cb(dbs, 0);
-    return dbs->acb;
+    return &dbs->common;
 }
 
 
@@ -157,12 +158,14 @@ BlockDriverAIOCB *dma_bdrv_write(BlockDriverState *bs,
 
 static void dma_aio_cancel(BlockDriverAIOCB *acb)
 {
-    DMABlockState *dbs = (DMABlockState *)acb->opaque;
+    DMAAIOCB *dbs = container_of(acb, DMAAIOCB, common);
 
-    bdrv_aio_cancel(dbs->acb);
+    if (dbs->acb) {
+        bdrv_aio_cancel(dbs->acb);
+    }
 }
 
 void dma_helper_init(void)
 {
-    aio_pool_init(&dma_aio_pool, sizeof(BlockDriverAIOCB), dma_aio_cancel);
+    aio_pool_init(&dma_aio_pool, sizeof(DMAAIOCB), dma_aio_cancel);
 }
Signature · 1 line
1.6.0.6
#9 lore raw

PATCH 3/4 Fix vectored aio bounce handling immediate errors

From: Avi Kivity <[email protected]>

If a bounced vectored aio fails immediately (the inner aio submission returning NULL) then the bounce handler erronously returns an aio request which will never be completed (and which crashes when cancelled).

Fix by detecting that the inner request has failed and propagating the error.

---
 block.c |    5 +++++
 1 files changed, 5 insertions(+), 0 deletions(-)
block.c +5 -0
index 689ea37..b68a8da 100644
@@ -1383,6 +1383,11 @@ static BlockDriverAIOCB *bdrv_aio_rw_vector(BlockDriverState *bs,
         s->aiocb = bdrv_aio_read(bs, sector_num, s->bounce, nb_sectors,
                                  bdrv_aio_rw_vector_cb, s);
     }
+    if (!s->aiocb) {
+        qemu_vfree(s->bounce);
+        qemu_aio_release(s);
+        return NULL;
+    }
     return &s->common;
 }
 
Signature · 1 line
1.6.0.6
#10 lore raw

PATCH 4/4 Fix DMA API when handling an immediate error from block layer

From: Avi Kivity <[email protected]>

The block layer may signal an immediate error on an asynchronous request by returning NULL. The DMA API did not handle this correctly, returning an AIO request which would never complete (and which would crash if cancelled).

Fix by detecting the failure and propagating it.

---
 dma-helpers.c |   27 +++++++++++++++++++++------
 1 files changed, 21 insertions(+), 6 deletions(-)
dma-helpers.c +21 -6
index 96a120c..1469e34 100644
@@ -70,20 +70,26 @@ static void continue_after_map_failure(void *opaque)
     qemu_bh_schedule(dbs->bh);
 }
 
-static void dma_bdrv_cb(void *opaque, int ret)
+static void dma_bdrv_unmap(DMAAIOCB *dbs)
 {
-    DMAAIOCB *dbs = (DMAAIOCB *)opaque;
-    target_phys_addr_t cur_addr, cur_len;
-    void *mem;
     int i;
 
-    dbs->acb = NULL;
-    dbs->sector_num += dbs->iov.size / 512;
     for (i = 0; i < dbs->iov.niov; ++i) {
         cpu_physical_memory_unmap(dbs->iov.iov[i].iov_base,
                                   dbs->iov.iov[i].iov_len, !dbs->is_write,
                                   dbs->iov.iov[i].iov_len);
     }
+}
+
+void dma_bdrv_cb(void *opaque, int ret)
+{
+    DMAAIOCB *dbs = (DMAAIOCB *)opaque;
+    target_phys_addr_t cur_addr, cur_len;
+    void *mem;
+
+    dbs->acb = NULL;
+    dbs->sector_num += dbs->iov.size / 512;
+    dma_bdrv_unmap(dbs);
     qemu_iovec_reset(&dbs->iov);
 
     if (dbs->sg_cur_index == dbs->sg->nsg || ret < 0) {
@@ -119,6 +125,11 @@ static void dma_bdrv_cb(void *opaque, int ret)
         dbs->acb = bdrv_aio_readv(dbs->bs, dbs->sector_num, &dbs->iov,
                                   dbs->iov.size / 512, dma_bdrv_cb, dbs);
     }
+    if (!dbs->acb) {
+        dma_bdrv_unmap(dbs);
+        qemu_iovec_destroy(&dbs->iov);
+        return;
+    }
 }
 
 static BlockDriverAIOCB *dma_bdrv_io(
@@ -138,6 +149,10 @@ static BlockDriverAIOCB *dma_bdrv_io(
     dbs->bh = NULL;
     qemu_iovec_init(&dbs->iov, sg->nsg);
     dma_bdrv_cb(dbs, 0);
+    if (!dbs->acb) {
+        qemu_aio_release(dbs);
+        return NULL;
+    }
     return &dbs->common;
 }
 
Signature · 1 line
1.6.0.6
#11 lore raw

Re: [Qemu-devel] [STABLE][PATCH 4/4] Fix DMA API when handling an immediate error from block layer

Quoted 10 lines · Mark McLoughlin

From: Avi Kivity <[email protected]>

The block layer may signal an immediate error on an asynchronous request by returning NULL. The DMA API did not handle this correctly, returning an AIO request which would never complete (and which would crash if cancelled).

Fix by detecting the failure and propagating it.

Quoted 64 lines
---
 dma-helpers.c |   27 +++++++++++++++++++++------
 1 files changed, 21 insertions(+), 6 deletions(-)
dma-helpers.c +21 -6
index 96a120c..1469e34 100644
@@ -70,20 +70,26 @@ static void continue_after_map_failure(void *opaque)
     qemu_bh_schedule(dbs->bh);
 }
 
-static void dma_bdrv_cb(void *opaque, int ret)
+static void dma_bdrv_unmap(DMAAIOCB *dbs)
 {
-    DMAAIOCB *dbs = (DMAAIOCB *)opaque;
-    target_phys_addr_t cur_addr, cur_len;
-    void *mem;
     int i;
 
-    dbs->acb = NULL;
-    dbs->sector_num += dbs->iov.size / 512;
     for (i = 0; i < dbs->iov.niov; ++i) {
         cpu_physical_memory_unmap(dbs->iov.iov[i].iov_base,
                                   dbs->iov.iov[i].iov_len, !dbs->is_write,
                                   dbs->iov.iov[i].iov_len);
     }
+}
+
+void dma_bdrv_cb(void *opaque, int ret)
+{
+    DMAAIOCB *dbs = (DMAAIOCB *)opaque;
+    target_phys_addr_t cur_addr, cur_len;
+    void *mem;
+
+    dbs->acb = NULL;
+    dbs->sector_num += dbs->iov.size / 512;
+    dma_bdrv_unmap(dbs);
     qemu_iovec_reset(&dbs->iov);
 
     if (dbs->sg_cur_index == dbs->sg->nsg || ret < 0) {
@@ -119,6 +125,11 @@ static void dma_bdrv_cb(void *opaque, int ret)
         dbs->acb = bdrv_aio_readv(dbs->bs, dbs->sector_num, &dbs->iov,
                                   dbs->iov.size / 512, dma_bdrv_cb, dbs);
     }
+    if (!dbs->acb) {
+        dma_bdrv_unmap(dbs);
+        qemu_iovec_destroy(&dbs->iov);
+        return;
+    }
 }
 
 static BlockDriverAIOCB *dma_bdrv_io(
@@ -138,6 +149,10 @@ static BlockDriverAIOCB *dma_bdrv_io(
     dbs->bh = NULL;
     qemu_iovec_init(&dbs->iov, sg->nsg);
     dma_bdrv_cb(dbs, 0);
+    if (!dbs->acb) {
+        qemu_aio_release(dbs);
+        return NULL;
+    }
     return &dbs->common;
 }
 
#12 lore raw

Re: [Qemu-devel] [STABLE][PATCH 3/4] Fix vectored aio bounce handling immediate errors

In reply to Mark McLoughlin
Quoted 10 lines · Mark McLoughlin

From: Avi Kivity <[email protected]>

If a bounced vectored aio fails immediately (the inner aio submission returning NULL) then the bounce handler erronously returns an aio request which will never be completed (and which crashes when cancelled).

Fix by detecting that the inner request has failed and propagating the error.

Quoted 20 lines
---
 block.c |    5 +++++
 1 files changed, 5 insertions(+), 0 deletions(-)
block.c +5 -0
index 689ea37..b68a8da 100644
@@ -1383,6 +1383,11 @@ static BlockDriverAIOCB *bdrv_aio_rw_vector(BlockDriverState *bs,
         s->aiocb = bdrv_aio_read(bs, sector_num, s->bounce, nb_sectors,
                                  bdrv_aio_rw_vector_cb, s);
     }
+    if (!s->aiocb) {
+        qemu_vfree(s->bounce);
+        qemu_aio_release(s);
+        return NULL;
+    }
     return &s->common;
 }
 
#13 lore raw

Re: [Qemu-devel] [STABLE][PATCH 2/4] Move block dma helpers aiocb to store dma state

In reply to Mark McLoughlin
Quoted 6 lines · Mark McLoughlin

From: Avi Kivity <[email protected]>

Use the dedicated dma aiocb to store intermediate state for dma block transactions.

Quoted 124 lines
---
 dma-helpers.c |   37 ++++++++++++++++++++-----------------
 1 files changed, 20 insertions(+), 17 deletions(-)
dma-helpers.c +20 -17
index 19fa4f0..96a120c 100644
@@ -39,6 +39,7 @@ void qemu_sglist_destroy(QEMUSGList *qsg)
 }
 
 typedef struct {
+    BlockDriverAIOCB common;
     BlockDriverState *bs;
     BlockDriverAIOCB *acb;
     QEMUSGList *sg;
@@ -48,13 +49,13 @@ typedef struct {
     target_phys_addr_t sg_cur_byte;
     QEMUIOVector iov;
     QEMUBH *bh;
-} DMABlockState;
+} DMAAIOCB;
 
 static void dma_bdrv_cb(void *opaque, int ret);
 
 static void reschedule_dma(void *opaque)
 {
-    DMABlockState *dbs = (DMABlockState *)opaque;
+    DMAAIOCB *dbs = (DMAAIOCB *)opaque;
 
     qemu_bh_delete(dbs->bh);
     dbs->bh = NULL;
@@ -63,7 +64,7 @@ static void reschedule_dma(void *opaque)
 
 static void continue_after_map_failure(void *opaque)
 {
-    DMABlockState *dbs = (DMABlockState *)opaque;
+    DMAAIOCB *dbs = (DMAAIOCB *)opaque;
 
     dbs->bh = qemu_bh_new(reschedule_dma, dbs);
     qemu_bh_schedule(dbs->bh);
@@ -71,11 +72,12 @@ static void continue_after_map_failure(void *opaque)
 
 static void dma_bdrv_cb(void *opaque, int ret)
 {
-    DMABlockState *dbs = (DMABlockState *)opaque;
+    DMAAIOCB *dbs = (DMAAIOCB *)opaque;
     target_phys_addr_t cur_addr, cur_len;
     void *mem;
     int i;
 
+    dbs->acb = NULL;
     dbs->sector_num += dbs->iov.size / 512;
     for (i = 0; i < dbs->iov.niov; ++i) {
         cpu_physical_memory_unmap(dbs->iov.iov[i].iov_base,
@@ -85,10 +87,9 @@ static void dma_bdrv_cb(void *opaque, int ret)
     qemu_iovec_reset(&dbs->iov);
 
     if (dbs->sg_cur_index == dbs->sg->nsg || ret < 0) {
-        dbs->acb->cb(dbs->acb->opaque, ret);
+        dbs->common.cb(dbs->common.opaque, ret);
         qemu_iovec_destroy(&dbs->iov);
-        qemu_aio_release(dbs->acb);
-        qemu_free(dbs);
+        qemu_aio_release(dbs);
         return;
     }
 
@@ -112,11 +113,11 @@ static void dma_bdrv_cb(void *opaque, int ret)
     }
 
     if (dbs->is_write) {
-        bdrv_aio_writev(dbs->bs, dbs->sector_num, &dbs->iov,
-                        dbs->iov.size / 512, dma_bdrv_cb, dbs);
+        dbs->acb = bdrv_aio_writev(dbs->bs, dbs->sector_num, &dbs->iov,
+                                   dbs->iov.size / 512, dma_bdrv_cb, dbs);
     } else {
-        bdrv_aio_readv(dbs->bs, dbs->sector_num, &dbs->iov,
-                       dbs->iov.size / 512, dma_bdrv_cb, dbs);
+        dbs->acb = bdrv_aio_readv(dbs->bs, dbs->sector_num, &dbs->iov,
+                                  dbs->iov.size / 512, dma_bdrv_cb, dbs);
     }
 }
 
@@ -125,10 +126,10 @@ static BlockDriverAIOCB *dma_bdrv_io(
     BlockDriverCompletionFunc *cb, void *opaque,
     int is_write)
 {
-    DMABlockState *dbs = qemu_malloc(sizeof(*dbs));
+    DMAAIOCB *dbs =  qemu_aio_get_pool(&dma_aio_pool, bs, cb, opaque);
 
+    dbs->acb = NULL;
     dbs->bs = bs;
-    dbs->acb = qemu_aio_get_pool(&dma_aio_pool, bs, cb, opaque);
     dbs->sg = sg;
     dbs->sector_num = sector_num;
     dbs->sg_cur_index = 0;
@@ -137,7 +138,7 @@ static BlockDriverAIOCB *dma_bdrv_io(
     dbs->bh = NULL;
     qemu_iovec_init(&dbs->iov, sg->nsg);
     dma_bdrv_cb(dbs, 0);
-    return dbs->acb;
+    return &dbs->common;
 }
 
 
@@ -157,12 +158,14 @@ BlockDriverAIOCB *dma_bdrv_write(BlockDriverState *bs,
 
 static void dma_aio_cancel(BlockDriverAIOCB *acb)
 {
-    DMABlockState *dbs = (DMABlockState *)acb->opaque;
+    DMAAIOCB *dbs = container_of(acb, DMAAIOCB, common);
 
-    bdrv_aio_cancel(dbs->acb);
+    if (dbs->acb) {
+        bdrv_aio_cancel(dbs->acb);
+    }
 }
 
 void dma_helper_init(void)
 {
-    aio_pool_init(&dma_aio_pool, sizeof(BlockDriverAIOCB), dma_aio_cancel);
+    aio_pool_init(&dma_aio_pool, sizeof(DMAAIOCB), dma_aio_cancel);
 }
#14 lore raw

Re: [Qemu-devel] [STABLE][PATCH 1/4] Use vectored aiocb storage to store vector translation state

In reply to Mark McLoughlin
Quoted 7 lines · Mark McLoughlin

From: Avi Kivity <[email protected]>

Now that we have a dedicated acb pool for vector translation acbs, we can store the vector translation state in the acbs instead of in an external structure.

Quoted 82 lines
---
 block.c |   29 ++++++++++++++---------------
 1 files changed, 14 insertions(+), 15 deletions(-)
block.c +14 -15
index b12318f..689ea37 100644
@@ -1332,31 +1332,32 @@ char *bdrv_snapshot_dump(char *buf, int buf_size, QEMUSnapshotInfo *sn)
 /**************************************************************/
 /* async I/Os */
 
-typedef struct VectorTranslationState {
+typedef struct VectorTranslationAIOCB {
+    BlockDriverAIOCB common;
     QEMUIOVector *iov;
     uint8_t *bounce;
     int is_write;
     BlockDriverAIOCB *aiocb;
-    BlockDriverAIOCB *this_aiocb;
-} VectorTranslationState;
+} VectorTranslationAIOCB;
 
-static void bdrv_aio_cancel_vector(BlockDriverAIOCB *acb)
+static void bdrv_aio_cancel_vector(BlockDriverAIOCB *_acb)
 {
-    VectorTranslationState *s = acb->opaque;
+    VectorTranslationAIOCB *acb
+        = container_of(_acb, VectorTranslationAIOCB, common);
 
-    bdrv_aio_cancel(s->aiocb);
+    bdrv_aio_cancel(acb->aiocb);
 }
 
 static void bdrv_aio_rw_vector_cb(void *opaque, int ret)
 {
-    VectorTranslationState *s = opaque;
+    VectorTranslationAIOCB *s = (VectorTranslationAIOCB *)opaque;
 
     if (!s->is_write) {
         qemu_iovec_from_buffer(s->iov, s->bounce, s->iov->size);
     }
     qemu_vfree(s->bounce);
-    s->this_aiocb->cb(s->this_aiocb->opaque, ret);
-    qemu_aio_release(s->this_aiocb);
+    s->common.cb(s->common.opaque, ret);
+    qemu_aio_release(s);
 }
 
 static BlockDriverAIOCB *bdrv_aio_rw_vector(BlockDriverState *bs,
@@ -1368,11 +1369,9 @@ static BlockDriverAIOCB *bdrv_aio_rw_vector(BlockDriverState *bs,
                                             int is_write)
 
 {
-    VectorTranslationState *s = qemu_mallocz(sizeof(*s));
-    BlockDriverAIOCB *aiocb = qemu_aio_get_pool(&vectored_aio_pool, bs,
-                                                cb, opaque);
+    VectorTranslationAIOCB *s = qemu_aio_get_pool(&vectored_aio_pool, bs,
+                                                  cb, opaque);
 
-    s->this_aiocb = aiocb;
     s->iov = iov;
     s->bounce = qemu_memalign(512, nb_sectors * 512);
     s->is_write = is_write;
@@ -1384,7 +1383,7 @@ static BlockDriverAIOCB *bdrv_aio_rw_vector(BlockDriverState *bs,
         s->aiocb = bdrv_aio_read(bs, sector_num, s->bounce, nb_sectors,
                                  bdrv_aio_rw_vector_cb, s);
     }
-    return aiocb;
+    return &s->common;
 }
 
 BlockDriverAIOCB *bdrv_aio_readv(BlockDriverState *bs, int64_t sector_num,
@@ -1560,7 +1559,7 @@ static int bdrv_write_em(BlockDriverState *bs, int64_t sector_num,
 
 void bdrv_init(void)
 {
-    aio_pool_init(&vectored_aio_pool, sizeof(BlockDriverAIOCB),
+    aio_pool_init(&vectored_aio_pool, sizeof(VectorTranslationAIOCB),
                   bdrv_aio_cancel_vector);
 
     bdrv_register(&bdrv_raw);
#15 lore raw

[Qemu-devel] Re: [STABLE][PATCH 0/4] Rebase more aio dma cancellation work

In reply to Mark McLoughlin
Quoted 11 lines · Mark McLoughlin

Hi,
We've a report of F11 qemu segfaulting when using read-only
images here:

https://bugzilla.redhat.com/497170

Looking at the current code in stable, there is clearly some
borkage:

- No error handling in bdrv_aio_rw_vector()

Quoted 9 lines

- No error handling in dma_bdrv_cb()/dma_bdrv_io()

- VectorTranslationState is leaked by bdrv_aio_rw_vector_cb()

- dma_aio_cancel() is cancelling the wrong aio

The rest of Avi's aio cancellation patches actually fixes all
these problems. Please apply to stable.

It's a little sad that we have to apply such a non trivial patchset to stable, but I see no way around it. At least it has seen extensive testing in the main branch and in qemu-kvm.git.

(Mark, you have to add your signoff if you're submitting a patch, even if you've done no changes to it at all)

Signature · 1 line
error compiling committee.c: too many arguments to function
#16 lore raw

Re: [Qemu-devel] [STABLE][PATCH 0/4] Rebase more aio dma cancellation work

In reply to Mark McLoughlin
Quoted 11 lines · Mark McLoughlin

Hi,
We've a report of F11 qemu segfaulting when using read-only
images here:

https://bugzilla.redhat.com/497170

Looking at the current code in stable, there is clearly some
borkage:

- No error handling in bdrv_aio_rw_vector()

Quoted 9 lines

- No error handling in dma_bdrv_cb()/dma_bdrv_io()

- VectorTranslationState is leaked by bdrv_aio_rw_vector_cb()

- dma_aio_cancel() is cancelling the wrong aio

The rest of Avi's aio cancellation patches actually fixes all
these problems. Please apply to stable.

Thanks Mark.

I actually cherry picked the patches directly. This is less error prone.

Regards,

Anthony Liguori