Skip to contentSkip to conversation

exit_mmap

DISCUSSION Stephane 1 participant 2 messages Mar 27, 2003

lore
Share & tools
lore

Stephane

author
#0 lore raw

I have a question about mmap and the close operation of a vm_area_struct.
Is there a reason why in exit_mmap, when a process dies unexpectedly, the vm_ops->close is called before zap_page_range is called?

The problem is that if you have allocated one or several kernel pages for a vm_area_struct, you can not free them in the vm_ops->close operation since the count field of the pages is not 0 because they are still mapped. The count will be cleared when zap_page_range is called.

This means that exit_mmap calls vm_ops->close and zap_page_range in the reverse order of a normal execution of the process, that is when the process unmap the area before dying.

It would be more deterministic and simple if vm_ops->close was always called when all the pages of the area was unmapped.

PS: please can you CC'ed the answer to [email protected]

Signature · 5 lines
Stephane Tessier
Net Insight AB          [email protected]
Västberga Allé 9        http://www.netinsight.se
SE-126 30 Hägersten     phone:+46-8-685 04 60
Sweden                  fax:  +46-8-685 04 20
author
#1 lore raw

Actually I was talking about 2.4.19 but I saw that this was solved in 2.5,
sorry for the disturbance

Quoted 25 lines · Stephane

I have a question about mmap and the close operation of a vm_area_struct.
Is there a reason why in exit_mmap, when a process dies unexpectedly, the vm_ops->close is called before zap_page_range is called?

The problem is that if you have allocated one or several kernel pages for a vm_area_struct, you can not free them in the vm_ops->close operation since the count field of the pages is not 0 because they are still mapped. The count will be cleared when zap_page_range is called.

This means that exit_mmap calls vm_ops->close and zap_page_range in the reverse order of a normal execution of the process, that is when the process unmap the area before dying.

It would be more deterministic and simple if vm_ops->close was always called when all the pages of the area was unmapped.

PS: please can you CC'ed the answer to [email protected] --
Stephane Tessier
Net Insight AB [email protected]
Västberga Allé 9 http://www.netinsight.se
SE-126 30 Hägersten phone:+46-8-685 04 60
Sweden fax: +46-8-685 04 20

Signature · 5 lines
Stephane Tessier
Net Insight AB          [email protected]
Västberga Allé 9        http://www.netinsight.se
SE-126 30 Hägersten     phone:+46-8-685 04 60
Sweden                  fax:  +46-8-685 04 20