Re: [RFC PATCH v2 3/6] cxl/core: add report option for cxl_mem_get_poison()
On Wed, 3 Apr 2024 22:56:58 +0800 Shiyang Ruan wrote: > 在 2024/3/30 9:50, Dan Williams 写道: > > Shiyang Ruan wrote: > >> The GMER only has "Physical Address" field, no such one indicates length. > >> So, when a poison event is received, we could use GET_POISON_LIST command > >> to get the poison list. Now driver has cxl_mem_get_poison(), so > >> reuse it and add a parameter 'bool report', report poison record to MCE > >> if set true. > > > > I am not sure I agree with the rationale here because there is no > > correlation between the event being signaled and the current state of > > the poison list. It also establishes race between multiple GMER events, > > i.e. imagine the hardware sends 4 GMER events to communicate a 256B > > poison discovery event. Does the driver need logic to support GMER event > > 2, 3, and 4 if it already say all 256B of poison after processing GMER > > event 1? > > Yes, I didn't thought about that. > > > > > I think the best the driver can do is assume at least 64B of poison > > per-event and depend on multiple notifications to handle larger poison > > lengths. > > Agree. This also makes things easier. > > And for qemu, I'm thinking of making a patch to limit the length of a > poison record when injecting. The length should between 64B to 4KiB per > GMER. And emit many GMERs if length > 4KiB. I'm not keen on such a restriction in QEMU. QEMU is injecting lengths allowed by the specification. That facility is useful for testing the kernel and the QEMU modeling should not be based on what the kernel supports. When you said this I wondered if we had a clever implementation that fused entries in the list, but we don't (I thought about doing so a long time ago but seems I never bothered :) So if you are using QEMU for testing and you don't want to exceed the kernel supported poison lengths, don't inject poison that big. Jonathan > > > > > Otherwise, the poison list is really only useful for pre-populating > > pages to offline after a reboot, i.e. to catch the kernel up with the > > state of poison pages after a reboot. > > Got it. > > > -- > Thanks, > Ruan.
Re: [RFC PATCH v2 3/6] cxl/core: add report option for cxl_mem_get_poison()
在 2024/3/30 9:50, Dan Williams 写道: Shiyang Ruan wrote: The GMER only has "Physical Address" field, no such one indicates length. So, when a poison event is received, we could use GET_POISON_LIST command to get the poison list. Now driver has cxl_mem_get_poison(), so reuse it and add a parameter 'bool report', report poison record to MCE if set true. I am not sure I agree with the rationale here because there is no correlation between the event being signaled and the current state of the poison list. It also establishes race between multiple GMER events, i.e. imagine the hardware sends 4 GMER events to communicate a 256B poison discovery event. Does the driver need logic to support GMER event 2, 3, and 4 if it already say all 256B of poison after processing GMER event 1? Yes, I didn't thought about that. I think the best the driver can do is assume at least 64B of poison per-event and depend on multiple notifications to handle larger poison lengths. Agree. This also makes things easier. And for qemu, I'm thinking of making a patch to limit the length of a poison record when injecting. The length should between 64B to 4KiB per GMER. And emit many GMERs if length > 4KiB. Otherwise, the poison list is really only useful for pre-populating pages to offline after a reboot, i.e. to catch the kernel up with the state of poison pages after a reboot. Got it. -- Thanks, Ruan.
Re: [RFC PATCH v2 3/6] cxl/core: add report option for cxl_mem_get_poison()
Shiyang Ruan wrote: > The GMER only has "Physical Address" field, no such one indicates length. > So, when a poison event is received, we could use GET_POISON_LIST command > to get the poison list. Now driver has cxl_mem_get_poison(), so > reuse it and add a parameter 'bool report', report poison record to MCE > if set true. I am not sure I agree with the rationale here because there is no correlation between the event being signaled and the current state of the poison list. It also establishes race between multiple GMER events, i.e. imagine the hardware sends 4 GMER events to communicate a 256B poison discovery event. Does the driver need logic to support GMER event 2, 3, and 4 if it already say all 256B of poison after processing GMER event 1? I think the best the driver can do is assume at least 64B of poison per-event and depend on multiple notifications to handle larger poison lengths. Otherwise, the poison list is really only useful for pre-populating pages to offline after a reboot, i.e. to catch the kernel up with the state of poison pages after a reboot.
