I believe that there is a sound reason for not reporting the unmodified size of the device. Specifically, the swap pager can't use the first, if I recall correctly, 8KB of a swap device because it might contain a disk label. However, I can't explain why we use dmmax to calculate the modified size.

Skipping is about https://svnweb.freebsd.org/base/head/sys/vm/swap_pager.c?revision=352407&view=markup#l2313

But sysctl describes totally different:

% sysctl -d vm.dmmax
vm.dmmax: Maximum size of a swap block in pages
% sysctl vm.dmmax
vm.dmmax: 32

After this removal, dmmax sysctl will be totally eliminated (I just realized).

% grep -R dmmax /usr/src/*
/usr/src/lib/libkvm/kvm_getswapinfo.c: { .n_name = "_dmmax" }, /* maximum size of a swap block */
/usr/src/lib/libkvm/kvm_getswapinfo.c:static int dmmax;
/usr/src/lib/libkvm/kvm_getswapinfo.c: ttl = swinfo.sw_nblks - dmmax;
/usr/src/lib/libkvm/kvm_getswapinfo.c: if (!GETSYSCTL(kd, "vm.dmmax", dmmax))
/usr/src/lib/libkvm/kvm_getswapinfo.c: ttl = xsd.xsw_nblks - dmmax;
/usr/src/lib/libkvm/kvm_getswapinfo.c: _kvm_err(kd, kd->program, "unable to find dmmax");
/usr/src/lib/libkvm/kvm_getswapinfo.c: KGET(NL_DMMAX, &dmmax);
/usr/src/sys/vm/swap_pager.c:SYSCTL_INT(_vm, OID_AUTO, dmmax, CTLFLAG_RD, &nsw_cluster_max, 0,
/usr/src/tools/tools/sysdoc/tunables.mdoc:vm.dmmax

In D20674#476657, @alc wrote:

I believe that there is a sound reason for not reporting the unmodified size of the device. Specifically, the swap pager can't use the first, if I recall correctly, 8KB of a swap device because it might contain a disk label. However, I can't explain why we use dmmax to calculate the modified size.

I think reporting the total size minus reserved blocks (for any definition of the block size, correct or not), is more confusing to users than reporting raw size and used. At worst, user would observe that some amount of swap volume is not filled, instead of seeing mismatch between e.g. geom size and vm size for the swap volume.

That said, I suspect that we no longer need to avoid initial blocks. At least not for the label preservation.

In D20674#476783, @kib wrote:

In D20674#476657, @alc wrote:

I believe that there is a sound reason for not reporting the unmodified size of the device. Specifically, the swap pager can't use the first, if I recall correctly, 8KB of a swap device because it might contain a disk label. However, I can't explain why we use dmmax to calculate the modified size.

I think reporting the total size minus reserved blocks (for any definition of the block size, correct or not), is more confusing to users than reporting raw size and used. At worst, user would observe that some amount of swap volume is not filled, instead of seeing mismatch between e.g. geom size and vm size for the swap volume.

I'm sympathetic to that argument. In any case, the current code is wrong, in that, it subtracts too much.

That said, I suspect that we no longer need to avoid initial blocks. At least not for the label preservation.

I'm afraid that is not true. I found out the hard way that it is needed. I lost a disk label, swap area, and file system when I first enabled "trimonce". However, the configuration of that storage device is usual. The first partition is a swap area, not a file system.

In D20674#476856, @alc wrote:

In D20674#476783, @kib wrote:

That said, I suspect that we no longer need to avoid initial blocks. At least not for the label preservation.

I'm afraid that is not true. I found out the hard way that it is needed. I lost a disk label, swap area, and file system when I first enabled "trimonce". However, the configuration of that storage device is usual. The first partition is a swap area, not a file system.

Was it for bsd label or gpt partitions ? Even then, if for bsd you did not used c, it should be fine.

In D20674#476856, @alc wrote:

In D20674#476783, @kib wrote:

In D20674#476657, @alc wrote:

I believe that there is a sound reason for not reporting the unmodified size of the device. Specifically, the swap pager can't use the first, if I recall correctly, 8KB of a swap device because it might contain a disk label. However, I can't explain why we use dmmax to calculate the modified size.

I think reporting the total size minus reserved blocks (for any definition of the block size, correct or not), is more confusing to users than reporting raw size and used. At worst, user would observe that some amount of swap volume is not filled, instead of seeing mismatch between e.g. geom size and vm size for the swap volume.

I'm sympathetic to that argument. In any case, the current code is wrong, in that, it subtracts too much.

That said, I suspect that we no longer need to avoid initial blocks. At least not for the label preservation.

I'm afraid that is not true. I found out the hard way that it is needed. I lost a disk label, swap area, and file system when I first enabled "trimonce". However, the configuration of that storage device is usual. The first partition is a swap area, not a file system.

We can't stop skipping. For the bsd label, we have an area at the front of the first partition that is both part of the partition as well as being part of the metadata and/or boot blocks. If we don't skip, we will overwrite this data as alc found out. Normally, these days, this isn't done, but since we have no way of asking the lower layers how much should be reserved, we need to skip. It might give a performance boost if we skip to the reported stripe size, since NAND devices use this to report optimal LBA alignment. I haven't checked lately to see if we do that and/or try to set the swap out block size to that, clustering pages if needed. For things like SD cards, it can be a huge win. For SATA SSD it's usually a moderate win. Nvme is more variable since enterprise drives don't really care, while super cheap QLC drives care rather a lot.

In D20674#476860, @kib wrote:

In D20674#476856, @alc wrote:

In D20674#476783, @kib wrote:

That said, I suspect that we no longer need to avoid initial blocks. At least not for the label preservation.

I'm afraid that is not true. I found out the hard way that it is needed. I lost a disk label, swap area, and file system when I first enabled "trimonce". However, the configuration of that storage device is usual. The first partition is a swap area, not a file system.

Was it for bsd label or gpt partitions ? Even then, if for bsd you did not used c, it should be fine.

It was a bsd label, and I did not use 'c'.

In D20674#476861, @imp wrote:

In D20674#476856, @alc wrote:

In D20674#476783, @kib wrote:

In D20674#476657, @alc wrote:

I believe that there is a sound reason for not reporting the unmodified size of the device. Specifically, the swap pager can't use the first, if I recall correctly, 8KB of a swap device because it might contain a disk label. However, I can't explain why we use dmmax to calculate the modified size.

I think reporting the total size minus reserved blocks (for any definition of the block size, correct or not), is more confusing to users than reporting raw size and used. At worst, user would observe that some amount of swap volume is not filled, instead of seeing mismatch between e.g. geom size and vm size for the swap volume.

I'm sympathetic to that argument. In any case, the current code is wrong, in that, it subtracts too much.

That said, I suspect that we no longer need to avoid initial blocks. At least not for the label preservation.

I'm afraid that is not true. I found out the hard way that it is needed. I lost a disk label, swap area, and file system when I first enabled "trimonce". However, the configuration of that storage device is usual. The first partition is a swap area, not a file system.

We can't stop skipping. For the bsd label, we have an area at the front of the first partition that is both part of the partition as well as being part of the metadata and/or boot blocks. If we don't skip, we will overwrite this data as alc found out. Normally, these days, this isn't done, but since we have no way of asking the lower layers how much should be reserved, we need to skip. It might give a performance boost if we skip to the reported stripe size, since NAND devices use this to report optimal LBA alignment. I haven't checked lately to see if we do that and/or try to set the swap out block size to that, clustering pages if needed. For things like SD cards, it can be a huge win. For SATA SSD it's usually a moderate win. Nvme is more variable since enterprise drives don't really care, while super cheap QLC drives care rather a lot.

We do clustering, but that clustering is in no way related to the device's stripe size. What is a typical stripe size?

In D20674#476867, @alc wrote:

I see, even gpart, by default, if not explicitly managed to specify the start offset, creates the slice 'a' at the start of the volume. OTOH I agree that it is too late to fix it, anyway most people migrated or should migrate to GPT.

v2% sudo mdconfig -a -t swap -s 32m
md0
v2% sudo gpart create -s bsd md0
md0 created
v2% sudo gpart show md0
=>    0  65536  md0  BSD  (32M)
      0  65536       - free -  (32M)

v2% sudo gpart add -t freebsd-swap -i 0
gpart: Invalid number of arguments.
v2% sudo gpart add -t freebsd-swap -i 0 md0
md0a added
v2% gpart show md0
=>    0  65536  md0  BSD  (32M)
      0  65536    1  freebsd-swap  (32M)

I'm wondering if this change can be merged.
I have this change in local tree and haven't exposed to an issue since review was requested.

In D20674#768930, @ota_j.email.ne.jp wrote:

I'm wondering if this change can be merged.
I have this change in local tree and haven't exposed to an issue since review was requested.

Did you read the conversation above? In short, it indicates that swap on bsd label still has the wart that the initial blocks of the 'a' partition overlap the label, and this is the reason for the calculation.