`FF_ZC_SEND` is same as `FF_USE_PAGE_ARRAY`, it will improve performance slightly in some scenarios, need to be tested in combination with real applications.
You can enable both compilation options at the same time or separately.
`FF_ZC_SEND` is same as `FF_USE_PAGE_ARRAY`, it will improve performance slightly in some scenarios, need to be tested in combination with real applications.
You can enable both compilation options at the same time or separately.
Two kinds of mbuf are used in f-stack: freebsd mbuf and dpdk mbuf.
freebsd mbufs are metadata used in freebsd stack, and their data
pointers (m_data) point to dpdk mbuf's data (buf_addr). And they have
their own chain, like this:
bsd_mbuf1 -> bsd_mbuf2 -> bsd_mbuf3
\ \ \
dpdk_mbuf1 -> dpdk_mbuf2 -> dpdk_mbuf3
Considering the map relationship,
- m_freem() is corresponding to rte_pktmbuf_free(), is to free the whole
chain of mbufs.
- m_free() is corresponding to rte_pktmbuf_free_seg(), is to free the
specified mbuf segment.
The current implementation in f-stack uses rte_pktmbuf_free() for
m_free(). This leads to mbufs, which are still in use, be freed
unexpectedly. For example, if the bsd_mbuf1 is trimed into zero length,
bsd will invoke m_free() to free the specified segment, however, the
whole mbuf chain is freed by calling rte_pktmbuf_free().
#0 rte_pktmbuf_free (m=0x22006fb480)
#1 in ff_dpdk_pktmbuf_free (m=0x22006fb480)
#2 in ff_mbuf_ext_free (m=0x7ffff7f82800, arg1=0x22006fb480, arg2=0x0)
#3 in mb_free_ext (m=0x7ffff7f82800)
#4 in m_free (m=0x7ffff7f82800)
#5 in sbcompress (sb=, m=0x7ffff7f82800, n=)
#6 in sbappendstream_locked (sb=, m=0x7ffff7f82800, flags=0)
The fix is straightforward. Use the correct API for segment free.
Reported-by: Yong-Hao Zou <yonghaoz1994@gmail.com>
Signed-off-by: Jianfeng Tan <henry.tjf@antgroup.com>
Two kinds of mbuf are used in f-stack: freebsd mbuf and dpdk mbuf.
freebsd mbufs are metadata used in freebsd stack, and their data
pointers (m_data) point to dpdk mbuf's data (buf_addr). And they have
their own chain, like this:
bsd_mbuf1 -> bsd_mbuf2 -> bsd_mbuf3
\ \ \
dpdk_mbuf1 -> dpdk_mbuf2 -> dpdk_mbuf3
Considering the map relationship,
- m_freem() is corresponding to rte_pktmbuf_free(), is to free the whole
chain of mbufs.
- m_free() is corresponding to rte_pktmbuf_free_seg(), is to free the
specified mbuf segment.
The current implementation in f-stack uses rte_pktmbuf_free() for
m_free(). This leads to mbufs, which are still in use, be freed
unexpectedly. For example, if the bsd_mbuf1 is trimed into zero length,
bsd will invoke m_free() to free the specified segment, however, the
whole mbuf chain is freed by calling rte_pktmbuf_free().
#0 rte_pktmbuf_free (m=0x22006fb480)
#1 in ff_dpdk_pktmbuf_free (m=0x22006fb480)
#2 in ff_mbuf_ext_free (m=0x7ffff7f82800, arg1=0x22006fb480, arg2=0x0)
#3 in mb_free_ext (m=0x7ffff7f82800)
#4 in m_free (m=0x7ffff7f82800)
#5 in sbcompress (sb=, m=0x7ffff7f82800, n=)
#6 in sbappendstream_locked (sb=, m=0x7ffff7f82800, flags=0)
The fix is straightforward. Use the correct API for segment free.
Reported-by: Yong-Hao Zou <yonghaoz1994@gmail.com>
Signed-off-by: Jianfeng Tan <henry.tjf@antgroup.com>