Re: [PATCH 00/10] iov_iter: Add new iters and use with AFS
On Thu, Sep 13, 2018 at 04:51:35PM +0100, David Howells wrote: > > Hi Al, > > Here's a set of patches that adds two new iov_iter types and then makes AFS > use them to do I/O. The iov_iter changes are: > > (1) Separate the type from the direction in the iov_iter struct and > provide accessor functions to wrap type checking. > > (2) Renumber the type constants to be contiguous small unsigned integers, > starting from 0 and then use switch-statements rather than if-else > chains using bit-testing. > > Note that the compiler can optimise this better by using CMP rather > than AND/TEST, say, as comparing integers requires fewer CMP > instructions or can use jump tables. > > (3) Change iov_offset from size_t to loff_t. This allows iov_offset to be > then used as a byte offset with ITER_MAPPING and allows 4GiB and larger > reads and writes to be proposed. > > This makes no difference on a 64-bit system, but does make a 32-bit > compilation a bit larger. > > (4) Add an ITER_MAPPING iterator type. This provides an iterator that > directly accesses an address_space, and assumes that the target pages > are in some way locked (eg. PG_lock or PG_writeback). > > (5) Add an ITER_DISCARD iterator type. This provides an iterator that > simply discards anything written to it. It cannot be used as a data > source. Hmm... I'll take a look in about an hour (need to finish the sodding tty-ioctl stuff first), will post review then. Sorry, I remember seeing it posted several weeks ago, slipped through the cracks then ;-/
Re: [PATCH 00/10] iov_iter: Add new iters and use with AFS
On Thu, Sep 13, 2018 at 04:51:35PM +0100, David Howells wrote: > > Hi Al, > > Here's a set of patches that adds two new iov_iter types and then makes AFS > use them to do I/O. The iov_iter changes are: > > (1) Separate the type from the direction in the iov_iter struct and > provide accessor functions to wrap type checking. > > (2) Renumber the type constants to be contiguous small unsigned integers, > starting from 0 and then use switch-statements rather than if-else > chains using bit-testing. > > Note that the compiler can optimise this better by using CMP rather > than AND/TEST, say, as comparing integers requires fewer CMP > instructions or can use jump tables. > > (3) Change iov_offset from size_t to loff_t. This allows iov_offset to be > then used as a byte offset with ITER_MAPPING and allows 4GiB and larger > reads and writes to be proposed. > > This makes no difference on a 64-bit system, but does make a 32-bit > compilation a bit larger. > > (4) Add an ITER_MAPPING iterator type. This provides an iterator that > directly accesses an address_space, and assumes that the target pages > are in some way locked (eg. PG_lock or PG_writeback). > > (5) Add an ITER_DISCARD iterator type. This provides an iterator that > simply discards anything written to it. It cannot be used as a data > source. Hmm... I'll take a look in about an hour (need to finish the sodding tty-ioctl stuff first), will post review then. Sorry, I remember seeing it posted several weeks ago, slipped through the cracks then ;-/
Re: [PATCH 00/10] iov_iter: Add new iters and use with AFS
Matthew Wilcox wrote: > It might serve to replace iov_iter_zero(). I don't think it can work like that. iov_iter_zero() writes zeros into an iterator someone else set up. To use ITER_ZERO, you'd be setting up the iterator and passing it to someone else to comsume. I think these are at opposite ends of the concept. David
Re: [PATCH 00/10] iov_iter: Add new iters and use with AFS
Matthew Wilcox wrote: > It might serve to replace iov_iter_zero(). I don't think it can work like that. iov_iter_zero() writes zeros into an iterator someone else set up. To use ITER_ZERO, you'd be setting up the iterator and passing it to someone else to comsume. I think these are at opposite ends of the concept. David
Re: [PATCH 00/10] iov_iter: Add new iters and use with AFS
On Thu, Sep 13, 2018 at 05:18:23PM +0100, David Howells wrote: > Matthew Wilcox wrote: > > > > (5) Add an ITER_DISCARD iterator type. This provides an iterator that > > > simply discards anything written to it. It cannot be used as a data > > > source. > > > > May I suggest an ITER_ZERO iterator type instead? It acts like /dev/zero; > > writes are discarded (as you have here) and reads return zeroes. > > I've wanted such a thing in the past, but got distracted away from > > that project. > > I thought about that, but the zero-filling is not as easy to implement as the > discard side - plus I don't have any use case to test it with. > > Do you have a use case in mind? At the time I did it, it was useful in dax, but I don't know if it is any more. It might serve to replace iov_iter_zero().
Re: [PATCH 00/10] iov_iter: Add new iters and use with AFS
On Thu, Sep 13, 2018 at 05:18:23PM +0100, David Howells wrote: > Matthew Wilcox wrote: > > > > (5) Add an ITER_DISCARD iterator type. This provides an iterator that > > > simply discards anything written to it. It cannot be used as a data > > > source. > > > > May I suggest an ITER_ZERO iterator type instead? It acts like /dev/zero; > > writes are discarded (as you have here) and reads return zeroes. > > I've wanted such a thing in the past, but got distracted away from > > that project. > > I thought about that, but the zero-filling is not as easy to implement as the > discard side - plus I don't have any use case to test it with. > > Do you have a use case in mind? At the time I did it, it was useful in dax, but I don't know if it is any more. It might serve to replace iov_iter_zero().
Re: [PATCH 00/10] iov_iter: Add new iters and use with AFS
Matthew Wilcox wrote: > > (5) Add an ITER_DISCARD iterator type. This provides an iterator that > > simply discards anything written to it. It cannot be used as a data > > source. > > May I suggest an ITER_ZERO iterator type instead? It acts like /dev/zero; > writes are discarded (as you have here) and reads return zeroes. > I've wanted such a thing in the past, but got distracted away from > that project. I thought about that, but the zero-filling is not as easy to implement as the discard side - plus I don't have any use case to test it with. Do you have a use case in mind? David
Re: [PATCH 00/10] iov_iter: Add new iters and use with AFS
Matthew Wilcox wrote: > > (5) Add an ITER_DISCARD iterator type. This provides an iterator that > > simply discards anything written to it. It cannot be used as a data > > source. > > May I suggest an ITER_ZERO iterator type instead? It acts like /dev/zero; > writes are discarded (as you have here) and reads return zeroes. > I've wanted such a thing in the past, but got distracted away from > that project. I thought about that, but the zero-filling is not as easy to implement as the discard side - plus I don't have any use case to test it with. Do you have a use case in mind? David
Re: [PATCH 00/10] iov_iter: Add new iters and use with AFS
On Thu, Sep 13, 2018 at 04:51:35PM +0100, David Howells wrote: > (5) Add an ITER_DISCARD iterator type. This provides an iterator that > simply discards anything written to it. It cannot be used as a data > source. May I suggest an ITER_ZERO iterator type instead? It acts like /dev/zero; writes are discarded (as you have here) and reads return zeroes. I've wanted such a thing in the past, but got distracted away from that project.
Re: [PATCH 00/10] iov_iter: Add new iters and use with AFS
On Thu, Sep 13, 2018 at 04:51:35PM +0100, David Howells wrote: > (5) Add an ITER_DISCARD iterator type. This provides an iterator that > simply discards anything written to it. It cannot be used as a data > source. May I suggest an ITER_ZERO iterator type instead? It acts like /dev/zero; writes are discarded (as you have here) and reads return zeroes. I've wanted such a thing in the past, but got distracted away from that project.