On Tue, Feb 06, 2024 at 07:31:26PM +0100, Mikulas Patocka wrote: > > > On Tue, 6 Feb 2024, Ming Lei wrote: > > > On Tue, Feb 06, 2024 at 12:14:14PM +0100, Mikulas Patocka wrote: > > > The __packed macro (expanding to __attribute__((__packed__))) specifies > > > that the structure has an alignment of 1. Therefore, it may be arbitrarily > > > misaligned. On architectures that don't have hardware support for > > > unaligned accesses, gcc generates very inefficient code that accesses the > > > structure fields byte-by-byte and assembles the result using shifts and > > > ors. > > > > > > For example, on PA-RISC, this function is compiled to 23 instructions with > > > the __packed attribute and only 2 instructions without the __packed > > > attribute. > > > > Can you share user visible effects in this way? such as IOPS or CPU > > utilization effect when running typical workload on null_blk or NVMe. > > The patch reduces total kernel size by 4096 bytes. The parisc machine > doesn't have PCIe, so I can't test it with NVMe :) You can run test over null-blk, which is enough to cover this report or change, given this patch is marked as "Fixes: ", we need to understand what it fixes. > > > CPU is supposed to be super fast if the data is in single L1 cacheline, > > but removing '__packed' may introduce one extra L1 cacheline load for > > bio. > > Saving the intruction cache is also important. Removing the __packed > keyword increases the bio structure size by 8 bytes - that is, L1 data > cache consumption will be increased with the probability 8/64. And it > reduces L1 instruction cache consumption by 84 bytes - that is one or two > cachelines. Yes. But the two kinds of caches have different properties, such as: - instruction cache has lower miss rate - instruction cache is read only so I'd suggest to provide null-blk test result at least. Thanks, Ming