Re: [PATCH v7 1/4] file-posix: add tracking of the zone write pointers

[Sam Li]

Stefan Hajnoczi <stefanha@redhat.com> 于2023年4月4日周二 01:04写道：
>
> On Thu, Mar 23, 2023 at 01:19:04PM +0800, Sam Li wrote:
> > Since Linux doesn't have a user API to issue zone append operations to
> > zoned devices from user space, the file-posix driver is modified to add
> > zone append emulation using regular writes. To do this, the file-posix
> > driver tracks the wp location of all zones of the device. It uses an
> > array of uint64_t. The most significant bit of each wp location indicates
> > if the zone type is conventional zones.
> >
> > The zones wp can be changed due to the following operations issued:
> > - zone reset: change the wp to the start offset of that zone
> > - zone finish: change to the end location of that zone
> > - write to a zone
> > - zone append
> >
> > Signed-off-by: Sam Li <faithilikerun@gmail.com>
> > ---
> >  block/file-posix.c               | 168 ++++++++++++++++++++++++++++++-
> >  include/block/block-common.h     |  14 +++
> >  include/block/block_int-common.h |   5 +
> >  3 files changed, 183 insertions(+), 4 deletions(-)
> >
> > diff --git a/block/file-posix.c b/block/file-posix.c
> > index 65efe5147e..0fb425dcae 100644
> > --- a/block/file-posix.c
> > +++ b/block/file-posix.c
> > @@ -1324,6 +1324,85 @@ static int hdev_get_max_segments(int fd, struct stat 
> > *st)
> >  #endif
> >  }
> >
> > +#if defined(CONFIG_BLKZONED)
> > +/*
> > + * If the ra (reset_all) flag > 0, then the wp of that zone should be 
> > reset to
> > + * the start sector. Else, take the real wp of the device.
> > + */
> > +static int get_zones_wp(int fd, BlockZoneWps *wps, int64_t offset,
> > +                        unsigned int nrz, int ra) {
>
> Please use bool for true/false and use clear variable names:
> int ra -> bool reset_all
>
> > +    struct blk_zone *blkz;
> > +    size_t rep_size;
> > +    uint64_t sector = offset >> BDRV_SECTOR_BITS;
> > +    int ret, n = 0, i = 0;
> > +    rep_size = sizeof(struct blk_zone_report) + nrz * sizeof(struct 
> > blk_zone);
> > +    g_autofree struct blk_zone_report *rep = NULL;
> > +
> > +    rep = g_malloc(rep_size);
> > +    blkz = (struct blk_zone *)(rep + 1);
> > +    while (n < nrz) {
> > +        memset(rep, 0, rep_size);
> > +        rep->sector = sector;
> > +        rep->nr_zones = nrz - n;
> > +
> > +        do {
> > +            ret = ioctl(fd, BLKREPORTZONE, rep);
> > +        } while (ret != 0 && errno == EINTR);
> > +        if (ret != 0) {
> > +            error_report("%d: ioctl BLKREPORTZONE at %" PRId64 " failed 
> > %d",
> > +                    fd, offset, errno);
> > +            return -errno;
> > +        }
> > +
> > +        if (!rep->nr_zones) {
> > +            break;
> > +        }
> > +
> > +        for (i = 0; i < rep->nr_zones; i++, n++) {
> > +            /*
> > +             * The wp tracking cares only about sequential writes required 
> > and
> > +             * sequential write preferred zones so that the wp can advance 
> > to
> > +             * the right location.
> > +             * Use the most significant bit of the wp location to indicate 
> > the
> > +             * zone type: 0 for SWR/SWP zones and 1 for conventional zones.
> > +             */
> > +            if (blkz[i].type == BLK_ZONE_TYPE_CONVENTIONAL) {
> > +                wps->wp[i] &= 1ULL << 63;
> > +            } else {
> > +                switch(blkz[i].cond) {
> > +                case BLK_ZONE_COND_FULL:
> > +                case BLK_ZONE_COND_READONLY:
> > +                    /* Zone not writable */
> > +                    wps->wp[i] = (blkz[i].start + blkz[i].len) << 
> > BDRV_SECTOR_BITS;
> > +                    break;
> > +                case BLK_ZONE_COND_OFFLINE:
> > +                    /* Zone not writable nor readable */
> > +                    wps->wp[i] = (blkz[i].start) << BDRV_SECTOR_BITS;
> > +                    break;
> > +                default:
> > +                    if (ra > 0) {
> > +                        wps->wp[i] = blkz[i].start << BDRV_SECTOR_BITS;
> > +                    } else {
> > +                        wps->wp[i] = blkz[i].wp << BDRV_SECTOR_BITS;
> > +                    }
> > +                    break;
> > +                }
> > +            }
> > +        }
> > +        sector = blkz[i - 1].start + blkz[i - 1].len;
> > +    }
> > +
> > +    return 0;
> > +}
> > +
> > +static void update_zones_wp(int fd, BlockZoneWps *wps, int64_t offset,
> > +                            unsigned int nrz) {
>
> QEMU coding style puts the opening curly bracket on a new line:
>
>   static void update_zones_wp(int fd, BlockZoneWps *wps, int64_t offset,
>                               unsigned int nrz)
>   {
>
> > +    if (get_zones_wp(fd, wps, offset, nrz, 0) < 0) {
> > +        error_report("update zone wp failed");
> > +    }
> > +}
> > +#endif
> > +
> >  static void raw_refresh_limits(BlockDriverState *bs, Error **errp)
> >  {
> >      BDRVRawState *s = bs->opaque;
> > @@ -1413,6 +1492,21 @@ static void raw_refresh_limits(BlockDriverState *bs, 
> > Error **errp)
> >          if (ret >= 0) {
> >              bs->bl.max_active_zones = ret;
> >          }
> > +
> > +        ret = get_sysfs_long_val(&st, "physical_block_size");
> > +        if (ret >= 0) {
> > +            bs->bl.write_granularity = ret;
> > +        }
> > +
> > +        bs->bl.wps = g_malloc(sizeof(BlockZoneWps) +
> > +                sizeof(int64_t) * bs->bl.nr_zones);
>
> This function can be called multiple times, so the old bs->bl.wps needs
> to be freed to avoid a memory leak here.
>
> > +        ret = get_zones_wp(s->fd, bs->bl.wps, 0, bs->bl.nr_zones, 0);
> > +        if (ret < 0) {
> > +            error_setg_errno(errp, -ret, "report wps failed");
> > +            g_free(bs->bl.wps);
>
> Please set it to NULL to reduce the risk of a double-free.
>
> > +            return;
> > +        }
> > +        qemu_co_mutex_init(&bs->bl.wps->colock);
>
> I just noticed there is a problem with keeping the mutex and
> heap-allocated wps inside bs.bl. bdrv_refresh_limits does this:
>
>   memset(&bs->bl, 0, sizeof(bs->bl));
>
> It would be possible to exclude the wps and mutex from the memset, but
> maybe they should be BlockDriverState fields instead. They are not
> really limits.
>
> >          return;
> >      }
> >  out:
> > @@ -2338,9 +2432,15 @@ static int coroutine_fn raw_co_prw(BlockDriverState 
> > *bs, uint64_t offset,
> >  {
> >      BDRVRawState *s = bs->opaque;
> >      RawPosixAIOData acb;
> > +    int ret;
> >
> >      if (fd_open(bs) < 0)
> >          return -EIO;
> > +#if defined(CONFIG_BLKZONED)
> > +    if (type & QEMU_AIO_WRITE && bs->bl.wps) {
> > +        qemu_co_mutex_lock(&bs->bl.wps->colock);
> > +    }
> > +#endif
> >
> >      /*
> >       * When using O_DIRECT, the request must be aligned to be able to use
> > @@ -2354,14 +2454,16 @@ static int coroutine_fn raw_co_prw(BlockDriverState 
> > *bs, uint64_t offset,
> >      } else if (s->use_linux_io_uring) {
> >          LuringState *aio = 
> > aio_get_linux_io_uring(bdrv_get_aio_context(bs));
> >          assert(qiov->size == bytes);
> > -        return luring_co_submit(bs, aio, s->fd, offset, qiov, type);
> > +        ret = luring_co_submit(bs, aio, s->fd, offset, qiov, type);
> > +        goto out;
> >  #endif
> >  #ifdef CONFIG_LINUX_AIO
> >      } else if (s->use_linux_aio) {
> >          LinuxAioState *aio = aio_get_linux_aio(bdrv_get_aio_context(bs));
> >          assert(qiov->size == bytes);
> > -        return laio_co_submit(bs, aio, s->fd, offset, qiov, type,
> > +        ret = laio_co_submit(bs, aio, s->fd, offset, qiov, type,
> >                                s->aio_max_batch);
> > +        goto out;
> >  #endif
> >      }
> >
> > @@ -2378,7 +2480,32 @@ static int coroutine_fn raw_co_prw(BlockDriverState 
> > *bs, uint64_t offset,
> >      };
> >
> >      assert(qiov->size == bytes);
> > -    return raw_thread_pool_submit(bs, handle_aiocb_rw, &acb);
> > +    ret = raw_thread_pool_submit(bs, handle_aiocb_rw, &acb);
> > +
> > +out:
> > +#if defined(CONFIG_BLKZONED)
> > +    BlockZoneWps *wps = bs->bl.wps;
> > +    if (ret == 0) {
> > +        if (type & QEMU_AIO_WRITE && wps && bs->bl.zone_size) {
> > +            uint64_t *wp = &wps->wp[offset / bs->bl.zone_size];
> > +            if (!BDRV_ZT_IS_CONV(*wp)) {
> > +                /* Advance the wp if needed */
> > +                if (offset + bytes > *wp) {
> > +                    *wp = offset + bytes;
> > +                }
> > +            }
> > +        }
> > +    } else {
> > +        if (type & QEMU_AIO_WRITE) {
> > +            update_zones_wp(s->fd, bs->bl.wps, 0, 1);
> > +        }
> > +    }
> > +
> > +    if (type & QEMU_AIO_WRITE && wps) {
> > +        qemu_co_mutex_unlock(&wps->colock);
> > +    }
> > +#endif
> > +    return ret;
> >  }
> >
> >  static int coroutine_fn raw_co_preadv(BlockDriverState *bs, int64_t offset,
> > @@ -2486,6 +2613,11 @@ static void raw_close(BlockDriverState *bs)
> >      BDRVRawState *s = bs->opaque;
> >
> >      if (s->fd >= 0) {
> > +#if defined(CONFIG_BLKZONED)
> > +        if (bs->bl.wps) {
> > +            g_free(bs->bl.wps);
> > +        }
>
> The if statement can be replaced with an unconditional
> g_free(bs->bl.wps) call. g_free(NULL) is valid and just returns
> immediately.
>
> > +#endif
> >          qemu_close(s->fd);
> >          s->fd = -1;
> >      }
> > @@ -3283,6 +3415,7 @@ static int coroutine_fn 
> > raw_co_zone_mgmt(BlockDriverState *bs, BlockZoneOp op,
> >      const char *op_name;
> >      unsigned long zo;
> >      int ret;
> > +    BlockZoneWps *wps = bs->bl.wps;
> >      int64_t capacity = bs->total_sectors << BDRV_SECTOR_BITS;
> >
> >      zone_size = bs->bl.zone_size;
> > @@ -3300,6 +3433,15 @@ static int coroutine_fn 
> > raw_co_zone_mgmt(BlockDriverState *bs, BlockZoneOp op,
> >          return -EINVAL;
> >      }
> >
> > +    qemu_co_mutex_lock(&wps->colock);
>
> I suggest using:
>
>   QEMU_LOCK_GUARD(&wps->colock);
>   ...
>
> or:
>
>   WITH_QEMU_LOCK_GUARD(&wps->colock) {
>       ...
>   }
>
> instead of qemu_co_mutex_lock/unlock().
>
> That way the lock is guaranteed to be unlocked when the function returns
> and you don't need to convert the error code paths to use goto.
>
> > +    uint32_t i = offset / bs->bl.zone_size;
> > +    uint64_t *wp = &wps->wp[i];
> > +    if (BDRV_ZT_IS_CONV(*wp) && len != capacity) {
> > +        error_report("zone mgmt operations are not allowed for 
> > conventional zones");
> > +        ret = -EIO;
> > +        goto out;
> > +    }
> > +
> >      switch (op) {
> >      case BLK_ZO_OPEN:
> >          op_name = "BLKOPENZONE";
> > @@ -3319,7 +3461,8 @@ static int coroutine_fn 
> > raw_co_zone_mgmt(BlockDriverState *bs, BlockZoneOp op,
> >          break;
> >      default:
> >          error_report("Unsupported zone op: 0x%x", op);
> > -        return -ENOTSUP;
> > +        ret = -ENOTSUP;
> > +        goto out;
> >      }
> >
> >      acb = (RawPosixAIOData) {
> > @@ -3337,10 +3480,27 @@ static int coroutine_fn 
> > raw_co_zone_mgmt(BlockDriverState *bs, BlockZoneOp op,
> >                          len >> BDRV_SECTOR_BITS);
> >      ret = raw_thread_pool_submit(bs, handle_aiocb_zone_mgmt, &acb);
> >      if (ret != 0) {
> > +        update_zones_wp(s->fd, wps, offset, i);
>
>
> >          ret = -errno;
> >          error_report("ioctl %s failed %d", op_name, ret);
> > +        goto out;
> > +    }
> > +
> > +    if (zo == BLKRESETZONE && len == capacity) {
> > +        ret = get_zones_wp(s->fd, wps, 0, bs->bl.nr_zones, 1);
> > +        if (ret < 0) {
> > +            error_report("reporting single wp failed");
> > +            return ret;
> > +        }
> > +    } else if (zo == BLKRESETZONE) {
> > +        *wp = offset;
> > +    } else if (zo == BLKFINISHZONE) {
> > +        /* The zoned device allows the last zone smaller that the zone 
> > size. */
> > +        *wp = offset + len;
> >      }
>
> The BLKRESETZONE and BLKFINISHZONE only update one zone's wp, but
> [offset, offset+len) can cover multiple zones.
>
> A loop is needed to update wps for multiple zones:
>
>   } else if (zo == BLKRESETZONE) {
>       for each zone {
>           wp[i] = offset + i * zone_size;
>       }
>   } else if (zo == BLKFINISHZONE) {
>       for each zone {
>           /* The last zone may be short */
>           wp[i] = MIN(offset + (i + 1) * zone_size, offset + len);
>       }
>   }

Thanks! Indeed, only zone_append and write can not advance to the next zone.

>
> >
> > +out:
> > +    qemu_co_mutex_unlock(&wps->colock);
> >      return ret;
> >  }
> >  #endif
> > diff --git a/include/block/block-common.h b/include/block/block-common.h
> > index 1576fcf2ed..93196229ac 100644
> > --- a/include/block/block-common.h
> > +++ b/include/block/block-common.h
> > @@ -118,6 +118,14 @@ typedef struct BlockZoneDescriptor {
> >      BlockZoneState state;
> >  } BlockZoneDescriptor;
> >
> > +/*
> > + * Track write pointers of a zone in bytes.
> > + */
> > +typedef struct BlockZoneWps {
> > +    CoMutex colock;
> > +    uint64_t wp[];
> > +} BlockZoneWps;
> > +
> >  typedef struct BlockDriverInfo {
> >      /* in bytes, 0 if irrelevant */
> >      int cluster_size;
> > @@ -240,6 +248,12 @@ typedef enum {
> >  #define BDRV_SECTOR_BITS   9
> >  #define BDRV_SECTOR_SIZE   (1ULL << BDRV_SECTOR_BITS)
> >
> > +/*
> > + * Get the first most significant bit of wp. If it is zero, then
> > + * the zone type is SWR.
> > + */
> > +#define BDRV_ZT_IS_CONV(wp)    (wp & (1ULL << 63))
> > +
> >  #define BDRV_REQUEST_MAX_SECTORS MIN_CONST(SIZE_MAX >> BDRV_SECTOR_BITS, \
> >                                             INT_MAX >> BDRV_SECTOR_BITS)
> >  #define BDRV_REQUEST_MAX_BYTES (BDRV_REQUEST_MAX_SECTORS << 
> > BDRV_SECTOR_BITS)
> > diff --git a/include/block/block_int-common.h 
> > b/include/block/block_int-common.h
> > index 1bd2aef4d5..69d1c3e6dd 100644
> > --- a/include/block/block_int-common.h
> > +++ b/include/block/block_int-common.h
> > @@ -884,6 +884,11 @@ typedef struct BlockLimits {
> >
> >      /* maximum number of active zones */
> >      int64_t max_active_zones;
> > +
> > +    /* array of write pointers' location of each zone in the zoned device. 
> > */
> > +    BlockZoneWps *wps;
> > +
> > +    int64_t write_granularity;
>
> What is this limit? Is it specific to zones?

It's the physical block size. Zone append write needs to check its io
size aligned with it.

>
> >  } BlockLimits;
> >
> >  typedef struct BdrvOpBlocker BdrvOpBlocker;
> > --
> > 2.39.2
> >