[dpdk-dev] eal/bsd: don't zero the pages during mmap in contigmem
Checks
Commit Message
Don't zero the pages during mmap in contigmem. Instead, zero the
pages after mmap in primary process. Otherwise, the multi-process
support will be broken, as the pages will be zeroed when secondary
processes map the memory.
Fixes: 82f931805506 ("contigmem: zero all pages during mmap")
Cc: stable@dpdk.org
Signed-off-by: Tiwei Bie <tiwei.bie@intel.com>
---
lib/librte_eal/bsdapp/contigmem/contigmem.c | 1 -
lib/librte_eal/bsdapp/eal/eal_memory.c | 3 +++
2 files changed, 3 insertions(+), 1 deletion(-)
Comments
On Mon, May 08, 2017 at 08:09:16AM +0000, Tiwei Bie wrote:
> Don't zero the pages during mmap in contigmem. Instead, zero the
> pages after mmap in primary process. Otherwise, the multi-process
> support will be broken, as the pages will be zeroed when secondary
> processes map the memory.
>
> Fixes: 82f931805506 ("contigmem: zero all pages during mmap")
> Cc: stable@dpdk.org
>
> Signed-off-by: Tiwei Bie <tiwei.bie@intel.com>
> ---
I agree there is a problem here, but I'm not sure about the solution to
it. I still think that the kernel should zero the pages before they get
given to userspace. Is there any way to keep that working e.g
* have them zeroed on mmap only when they are not already mmaped into
another process?
* have them zeroed on init, and again on unmap by the last process to
have them mapped?
/Bruce
On Mon, May 08, 2017 at 09:53:57AM +0100, Bruce Richardson wrote:
> On Mon, May 08, 2017 at 08:09:16AM +0000, Tiwei Bie wrote:
> > Don't zero the pages during mmap in contigmem. Instead, zero the
> > pages after mmap in primary process. Otherwise, the multi-process
> > support will be broken, as the pages will be zeroed when secondary
> > processes map the memory.
> >
> > Fixes: 82f931805506 ("contigmem: zero all pages during mmap")
> > Cc: stable@dpdk.org
> >
> > Signed-off-by: Tiwei Bie <tiwei.bie@intel.com>
> > ---
> I agree there is a problem here, but I'm not sure about the solution to
> it. I still think that the kernel should zero the pages before they get
> given to userspace. Is there any way to keep that working e.g
>
> * have them zeroed on mmap only when they are not already mmaped into
> another process?
> * have them zeroed on init, and again on unmap by the last process to
> have them mapped?
>
I think it's the simplest way to fix it in userspace, so I just did it.
I'd like to fix it in kernel if you also prefer this.
Best regards,
Tiwei Bie
@@ -227,7 +227,6 @@ contigmem_mmap_single(struct cdev *cdev, vm_ooffset_t *offset, vm_size_t size,
if (buffer_index >= contigmem_num_buffers)
return EINVAL;
- memset(contigmem_buffers[buffer_index], 0, contigmem_buffer_size);
*offset = (vm_ooffset_t)vtophys(contigmem_buffers[buffer_index]);
*obj = vm_pager_allocate(OBJT_DEVICE, cdev, size, nprot, *offset,
curthread->td_ucred);
@@ -36,6 +36,7 @@
#include <sys/sysctl.h>
#include <inttypes.h>
#include <fcntl.h>
+#include <string.h>
#include <rte_eal.h>
#include <rte_eal_memconfig.h>
@@ -121,6 +122,8 @@ rte_eal_hugepage_init(void)
seg->nrank = mcfg->nrank;
seg->socket_id = 0;
+ memset(seg->addr, 0, seg->len);
+
RTE_LOG(INFO, EAL, "Mapped memory segment %u @ %p: physaddr:0x%"
PRIx64", len %zu\n",
seg_idx, addr, physaddr, hpi->hugepage_sz);