i_mutex questions
Hi All,
I have been trying to find a way to synchronize punch hole with read and write operations with out the use of i_mutex. The concern is that after punch hole has released the pages inside the hole, another process may remap the page to a block before punch has taken i_data_sem. I think putting i_mutex around the punch hole operation would fix this, but since we are trying to avoid further improper use of i_mutex, I am trying to avoid that solution.
I cannot use i_data_sem to protect the pages because it seems most of the code has already established a locking order of pages first, then i_data_sem. So moving i_data_sem up tends to cause a lot of dead locks.
I'm thinking that there probably needs to be a another mutex involved
some where, but I wasnt sure if some one is already working on the idea of introducing a replacement for i_mutex. So I just wanted to know if there are any plans already in motion for this, or if any one else could suggest some ideas for the punch hole issue. Thx all!
Allison Henderson
Quoted 9 lines · Allison Henderson
Hi All,
I have been trying to find a way to synchronize punch hole with read and write operations with out the use of
i_mutex. The concern is that after punch hole has released the pages inside the hole, another process may remap the page to a block before punch has takeni_data_sem. I think puttingi_mutexaround the punch hole operation would fix this, but since we are trying to avoid further improper use ofi_mutex, I am trying to avoid that solution.
Hey Allison,
Actually, i_mutex is the normal way to handle this. ocfs2 takesi_mutex down under its ->fallocate(). Truncate is in the same boat, which is why do_truncate() takes i_mutex before calling notify_change().
The read-write paths grab i_mutex for buffered operation. They
don't for O_DIRECT, which doesn't map to the pagecache. This is where i_data_sem should speed things up.
Joel
Signature · 5 lines
"Anything that is too stupid to be spoken is sung."
- Voltaire
http://www.jlbec.org/
[email protected]Quoted 19 lines · Joel Becker
Quoted 9 lines · Allison Henderson
Hi All,
I have been trying to find a way to synchronize punch hole with read and write operations with out the use of
i_mutex. The concern is that after punch hole has released the pages inside the hole, another process may remap the page to a block before punch has takeni_data_sem. I think puttingi_mutexaround the punch hole operation would fix this, but since we are trying to avoid further improper use ofi_mutex, I am trying to avoid that solution.Hey Allison,
Actually,i_mutexis the normal way to handle this. ocfs2 takesi_mutexdown under its ->fallocate(). Truncate is in the same boat, which is whydo_truncate()takesi_mutexbefore callingnotify_change().
The read-write paths grabi_mutexfor buffered operation. They
don't forO_DIRECT, which doesn't map to the pagecache. This is wherei_data_semshould speed things up.
Quoted 2 lines
Joel
Hi Joel,
Well, I actually already had a patch that was trying to use i_mutex to solve this ([PATCH 4/6 v7] ext4: Lock i_mutex for punch hole). But we decided not to apply it because of plans to reduce the usage of i_mutex in the ext4 code. So I've been trying to figure out a different way to solve this, but so far I haven't had a whole lot of luck finding a solution that doesn't involve introducing a new locking mechanism. So I wanted to check back here for more details on what the plan for i_mutex is so I dont conflict with anything that might already be going on. :)
Ted, would you be able to give us some more details on this topic? Thx!
Allison Henderson
Quoted 9 lines · Allison Henderson
Well, I actually already had a patch that was trying to use
i_mutexto solve this ([PATCH 4/6 v7] ext4: Locki_mutexfor punch hole). But we decided not to apply it because of plans to reduce the usage ofi_mutexin the ext4 code. So I've been trying to figure out a different way to solve this, but so far I haven't had a whole lot of luck finding a solution that doesn't involve introducing a new locking mechanism. So I wanted to check back here for more details on what the plan fori_mutexis so I dont conflict with anything that might already be going on. :)
Sure ;-) If you find another mechanism that reduces contention
but still plays well with read/write et al, please let us all know. Getting i_mutex out of read/write would be interesting.
Joel
Signature · 5 lines
"A narcissist is someone better looking than you are."
- Gore Vidal
http://www.jlbec.org/
[email protected]Quoted 10 lines · Allison Henderson
Hi All,
I have been trying to find a way to synchronize punch hole with read and write operations with out the use of
i_mutex. The concern is that after punch hole has released the pages inside the hole, another process may remap the page to a block before punch has takeni_data_sem. I think puttingi_mutexaround the punch hole operation would fix this, but since we are trying to avoid further improper use ofi_mutex, I am trying to avoid that solution.
Quoted 8 lines
I cannot use
i_data_semto protect the pages because it seems most of the code has already established a locking order of pages first, theni_data_sem. So movingi_data_semup tends to cause a lot of dead locks. I'm thinking that there probably needs to be a another mutex involved some where, but I wasnt sure if some one is already working on the idea of introducing a replacement fori_mutex. So I just wanted to know if there are any plans already in motion for this, or if any one else could suggest some ideas for the punch hole issue. Thx all!
HI,
Lukas sent out a patch ([PATCH] ext4: Make reads/writes atomic with i_rwlock semaphore) which collected some feedbacks suggesting using extent lock instead of a read-write semaphore. If there is extent lock implementation in ext4, then fallocate can use it, maybe dioread-nolock can use it as well, e.g. locking a range and unlocking the range until the extent is converted from unwritten to init.
Yongqiang.
Quoted 8 lines
Allison Henderson
--
To unsubscribe from this list: send the line "unsubscribe linux-ext4" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Signature · 6 lines
Best Wishes Yongqiang Yang -- To unsubscribe from this list: send the line "unsubscribe linux-ext4" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html
Quoted 21 lines · Yongqiang Yang
Quoted 10 lines · Allison Henderson
Hi All,
I have been trying to find a way to synchronize punch hole with read and write operations with out the use of
i_mutex. The concern is that after punch hole has released the pages inside the hole, another process may remap the page to a block before punch has takeni_data_sem. I think puttingi_mutexaround the punch hole operation would fix this, but since we are trying to avoid further improper use ofi_mutex, I am trying to avoid that solution.Quoted 8 lines
I cannot use
i_data_semto protect the pages because it seems most of the code has already established a locking order of pages first, theni_data_sem. So movingi_data_semup tends to cause a lot of dead locks.
I'm thinking that there probably needs to be a another mutex involved some
where, but I wasnt sure if some one is already working on the idea of introducing a replacement fori_mutex. So I just wanted to know if there are any plans already in motion for this, or if any one else could suggest some ideas for the punch hole issue. Thx all!
Quoted 6 lines
Lukas sent out a patch ([PATCH] ext4: Make reads/writes atomic with
i_rwlocksemaphore) which collected some feedbacks suggesting using extent lock instead of a read-write semaphore. If there is extent lock implementation in ext4, then fallocate can use it, maybe dioread-nolock can use it as well, e.g. locking a range and unlocking the range until the extent is converted from unwritten to init.
We have a prototype patch for extent locking for ext4. We are planning to use this for parallel locking of htree directories, but it could potentially be modified to for extent locking of files.
The current patch is below, but it hasn't gone through a lot of testing:
Cheers, Andreas
Quoted 33 lines · Andreas Dilger
Quoted 21 lines · Yongqiang Yang
Quoted 10 lines · Allison Henderson
Hi All,
I have been trying to find a way to synchronize punch hole with read and write operations with out the use of
i_mutex. The concern is that after punch hole has released the pages inside the hole, another process may remap the page to a block before punch has takeni_data_sem. I think puttingi_mutexaround the punch hole operation would fix this, but since we are trying to avoid further improper use ofi_mutex, I am trying to avoid that solution.Quoted 8 lines
I cannot use
i_data_semto protect the pages because it seems most of the code has already established a locking order of pages first, theni_data_sem. So movingi_data_semup tends to cause a lot of dead locks.
I'm thinking that there probably needs to be a another mutex involved some
where, but I wasnt sure if some one is already working on the idea of introducing a replacement fori_mutex. So I just wanted to know if there are any plans already in motion for this, or if any one else could suggest some ideas for the punch hole issue. Thx all!Quoted 6 lines
Lukas sent out a patch ([PATCH] ext4: Make reads/writes atomic with
i_rwlocksemaphore) which collected some feedbacks suggesting using extent lock instead of a read-write semaphore. If there is extent lock implementation in ext4, then fallocate can use it, maybe dioread-nolock can use it as well, e.g. locking a range and unlocking the range until the extent is converted from unwritten to init.We have a prototype patch for extent locking for ext4. We are planning to use this for parallel locking of htree directories, but it could potentially be modified to for extent locking of files.
Quoted 6 lines
The current patch is below, but it hasn't gone through a lot of testing:
Cheers, Andreas
Alrighty then, I will have a look at the existing patches. Thx!
Allison Henderson
Quoted 6 lines
--
To unsubscribe from this list: send the line "unsubscribe linux-ext4" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html