[Bug 211852] Unsafe shutdowns on Intel 750 SSD

[bugzilla-noreply at freebsd.org]

https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=211852

--- Comment #2 from Ravi Pokala <rpokala at panasas.com> ---
Neat, I didn't know `smartctl' had been extended to understand NVMe! :-)

In any case, it the code for handling power-down looks grossly correct:

sys/dev/nvme/nvme_ctrlr.c (r308431)
1184 void
1185 nvme_ctrlr_shutdown(struct nvme_controller *ctrlr)
1186 {
1187         union cc_register       cc;
1188         union csts_register     csts;
1189         int                     ticks = 0;
1190 
1191         cc.raw = nvme_mmio_read_4(ctrlr, cc);
1192         cc.bits.shn = NVME_SHN_NORMAL;
1193         nvme_mmio_write_4(ctrlr, cc, cc.raw);
1194         csts.raw = nvme_mmio_read_4(ctrlr, csts);
1195         while ((csts.bits.shst != NVME_SHST_COMPLETE) && (ticks++ < 5*hz))
{
1196                 pause("nvme shn", 1);
1197                 csts.raw = nvme_mmio_read_4(ctrlr, csts);
1198         }
1199         if (csts.bits.shst != NVME_SHST_COMPLETE)
1200                 nvme_printf(ctrlr, "did not complete shutdown within 5
seconds "
1201                     "of notification\n");
1202 }

In English, that's roughly: notify the controller about a normal shutdown (as
opposed to an "abrupt" shutdown), then wait until the controller status
indicates that shutdown is complete; if the controller doesn't indicate
complete shutdown within 5 seconds, print a log message and continue anyway.

It has been in that state since r254302 (2013-08-13). (That's in -HEAD, but the
same code is in 10.3-RELEASE.)

Hmmm... In NVMe-1.2.1, section 7.6.2:

"It is recommended that the host wait a minimum of the RTD3 Entry Latency
reported in the Identify Controller data structure for the shutdown operations
to complete; if the value reported in RTD3 Entry Latency is 0h, then the host
should wait for a minimum of one second."

The "RTD3 Entry Latency" is described in section 5.11, Figure 90:

"Bytes 91:88: RTD3 Entry Latency (RTD3E): This field indicates the typical
latency in microseconds to enter Runtime D3 (RTD3). Refer to section 8.4.4 for
test conditions. A value of 0h indicates RTD3 Entry Latency is not reported."

So, that hard-coded 5 seconds might not be correct. It looks like (struct
nvme_controller_data) treats the part of the "Identify Controller" data
structure which contains RTD3E as reserved. It looks like it was in fact
reserved in NVMe-1.1, but was defined later.

-- 
You are receiving this mail because:
You are the assignee for the bug.