Use abd in dbuf and arc_buf to allow buffer sharing - #18854
Conversation
|
Note, I hit an VERIFY failure in this test
|
|
As you have said, ARC is typically built from scatter, not mapped into KVA buffers, that is good for memory fragmentation, but hard to access. Dbufs same time are linear and KVA-mapped for easy access, but we can't use them for everything, since we quickly get out of KVA due to fragmentation. What you haven't said is where do you expect them to meet. |
|
@amotin It's possible to use scatter abd for some types of dbuf like bp and dnode blocks where the structure is well aligned and won't cross any page boundary. It will be a trade off of memory fragmentation vs extra memcpy for compression and extra code complexity. |
|
The decrypt_fault failure is not caused by this change specifically, but seems to happen when buffer sharing is enabled. |
38b49df to
05f1eaa
Compare
receive_spill will call dbuf_spill_set_blksz, which in turn calls dbuf_new_size when resizing spill block. However, receive_spill holds a raw buf, which can be compressed and less than db_size. This causes an out-of-bound access when dbuf_new_size tries to copy the data using db_size as length. To fix this, we just disable the copy when coming from receive_spill, since we will assign a new arc_buf later anyway. Signed-off-by: Chunwei Chen <david.chen@nutanix.com>
For uncompressed block, we were supposed to allow buffer sharing between hdr->b_pabd and arc_buf->b_data. However, we can only do so for linear buffer as arc_buf and dbuf doesn't work with abd. And in practice, b_pabd will almost always be scatter type abd except for small buffer, so buffer sharing almost never happens. In this change we make arc_buf and dbuf use abd instead of raw buf, and make them use scatter abd for user data. This allows us to share the abd between b_pabd, arc_buf and dbuf for data. This reduces memory copy needed for both read and write, which can be a bottleneck for high performance setup. For metadata, we only use linear abd, so the user of metadata dbuf will need to use abd_to_buf to access the data, but otherwise it's pretty much the same as before. Signed-off-by: Chunwei Chen <david.chen@nutanix.com>
4 participants
For uncompressed block, we were supposed to allow buffer sharing between hdr->b_pabd and arc_buf->b_data. However, we can only do so for linear buffer as arc_buf and dbuf doesn't work with abd. And in practice, b_pabd will almost always be scatter type abd except for small buffer, so buffer sharing almost never happens.
In this change we make arc_buf and dbuf use abd instead of raw buf, and make them use scatter abd for user data. This allows us to share the abd between b_pabd, arc_buf and dbuf for data. This reduces memory copy needed for both read and write, which can be a bottleneck for high performance setup.
For metadata, we only use linear abd, so the user of metadata dbuf will need to use abd_to_buf to access the data, but otherwise it's pretty much the same as before.
Motivation and Context
Description
How Has This Been Tested?
Types of Changes
Checklist
Signed-off-by.