[PATCH] cxl: Prevent adapter reset if an active context exists
Frederic Barrat
fbarrat at linux.vnet.ibm.com
Tue Oct 11 02:09:07 AEDT 2016
Hi Vaibhav,
A few comments below...
Le 10/10/2016 à 16:09, Vaibhav Jain a écrit :
> This patch prevents resetting the cxl adapter via sysfs in presence of
> one or more active cxl_context on it. This protects against an
> unrecoverable error caused by PSL owning a dirty cache line even after
> reset and host tries to touch the same cache line. In case a force reset
> of the card is required irrespective of any active contexts, the int
> value -1 can be stored in the 'reset' sysfs attribute of the card.
>
> The patch introduces a new atomic_t member named contexts_num inside
> struct cxl that holds the number of active context attached to the card
> , which is checked against '0' before proceeding with the reset. To
> prevent against a race condition where a context is activated just after
> reset check is performed, the contexts_num is atomically set to '-1'
> after reset-check to indicate that no more contexts can be activated on
> the card anymore.
>
> Before activating a context we atomically test if contexts_num is
> non-negative and if so, increment its value by one. In case the value of
> contexts_num is negative then it indicates that the card is about to be
> reset and context activation is error-ed out at that point.
>
> Signed-off-by: Vaibhav Jain <vaibhav at linux.vnet.ibm.com>
> Reviewed-by: Andrew Donnellan <andrew.donnellan at au1.ibm.com>
> ---
> Documentation/ABI/testing/sysfs-class-cxl | 7 +++++--
> drivers/misc/cxl/api.c | 9 +++++++++
> drivers/misc/cxl/context.c | 3 +++
> drivers/misc/cxl/cxl.h | 21 +++++++++++++++++++++
> drivers/misc/cxl/file.c | 9 +++++++++
> drivers/misc/cxl/main.c | 24 +++++++++++++++++++++++-
> drivers/misc/cxl/sysfs.c | 18 ++++++++++++++----
> 7 files changed, 84 insertions(+), 7 deletions(-)
>
> diff --git a/Documentation/ABI/testing/sysfs-class-cxl b/Documentation/ABI/testing/sysfs-class-cxl
> index 4ba0a2a..dae2b38 100644
> --- a/Documentation/ABI/testing/sysfs-class-cxl
> +++ b/Documentation/ABI/testing/sysfs-class-cxl
> @@ -220,8 +220,11 @@ What: /sys/class/cxl/<card>/reset
> Date: October 2014
> Contact: linuxppc-dev at lists.ozlabs.org
> Description: write only
> - Writing 1 will issue a PERST to card which may cause the card
> - to reload the FPGA depending on load_image_on_perst.
> + Writing 1 will issue a PERST to card provided there are no
> + contexts active on any one of the card AFUs. This may cause
> + the card to reload the FPGA depending on load_image_on_perst.
> + Writing -1 will do a force PERST irrespective of any active
> + contexts on the card AFUs.
> Users: https://github.com/ibm-capi/libcxl
>
> What: /sys/class/cxl/<card>/perst_reloads_same_image (not in a guest)
> diff --git a/drivers/misc/cxl/api.c b/drivers/misc/cxl/api.c
> index f3d34b9..85bb2d9 100644
> --- a/drivers/misc/cxl/api.c
> +++ b/drivers/misc/cxl/api.c
> @@ -236,10 +236,19 @@ int cxl_start_context(struct cxl_context *ctx, u64 wed,
> ctx->real_mode = false;
> }
>
> + /*
> + * Increment the mapped context count for adapter. This also checks
> + * if adapter_context_lock is taken.
> + */
> + rc = cxl_adapter_context_get(ctx->afu->adapter);
> + if (rc)
> + goto out;
> +
> cxl_ctx_get();
>
> if ((rc = cxl_ops->attach_process(ctx, kernel, wed, 0))) {
> put_pid(ctx->pid);
> + cxl_adapter_context_put(ctx->afu->adapter);
> cxl_ctx_put();
> goto out;
> }
> diff --git a/drivers/misc/cxl/context.c b/drivers/misc/cxl/context.c
> index c466ee2..5e506c1 100644
> --- a/drivers/misc/cxl/context.c
> +++ b/drivers/misc/cxl/context.c
> @@ -238,6 +238,9 @@ int __detach_context(struct cxl_context *ctx)
> put_pid(ctx->glpid);
>
> cxl_ctx_put();
> +
> + /* Decrease the attached context count on the adapter */
> + cxl_adapter_context_put(ctx->afu->adapter);
> return 0;
> }
>
> diff --git a/drivers/misc/cxl/cxl.h b/drivers/misc/cxl/cxl.h
> index 01d372a..ed89c57 100644
> --- a/drivers/misc/cxl/cxl.h
> +++ b/drivers/misc/cxl/cxl.h
> @@ -618,6 +618,14 @@ struct cxl {
> bool perst_select_user;
> bool perst_same_image;
> bool psl_timebase_synced;
> +
> + /*
> + * number of contexts mapped on to this card.
> + * +ve: Number of contexts mapped and new one can be mapped.
> + * 0 : No active contexts and new ones can be mapped.
> + * -ve: No contexts mapped and new ones cannot be mapped.
what does "ve" stand for ?
For the last one, shouldn't it be '-1' ?
> + */
> + atomic_t contexts_num;
> };
>
> int cxl_pci_alloc_one_irq(struct cxl *adapter);
> @@ -944,4 +952,17 @@ bool cxl_pci_is_vphb_device(struct pci_dev *dev);
>
> /* decode AFU error bits in the PSL register PSL_SERR_An */
> void cxl_afu_decode_psl_serr(struct cxl_afu *afu, u64 serr);
> +
> +/*
> + * Increments the number of attached contexts on an adapter.
> + * Incase an adapter_context_lock is taken the return -EBUSY.
typo "In case"
> + */
> +int cxl_adapter_context_get(struct cxl *adapter);
> +
> +/* Decrements the number of attached contexts on an adapter */
> +void cxl_adapter_context_put(struct cxl *adapter);
> +
> +/* If no active contexts then prevents contexts from being attached */
> +int cxl_adapter_context_lock(struct cxl *adapter);
> +
> #endif
> diff --git a/drivers/misc/cxl/file.c b/drivers/misc/cxl/file.c
> index 5fb9894..3b2272a 100644
> --- a/drivers/misc/cxl/file.c
> +++ b/drivers/misc/cxl/file.c
> @@ -205,11 +205,20 @@ static long afu_ioctl_start_work(struct cxl_context *ctx,
> ctx->pid = get_task_pid(current, PIDTYPE_PID);
> ctx->glpid = get_task_pid(current->group_leader, PIDTYPE_PID);
>
> + /*
> + * Increment the mapped context count for adapter. This also checks
> + * if adapter_context_lock is taken.
> + */
> + rc = cxl_adapter_context_get(ctx->afu->adapter);
> + if (rc)
> + goto out;
> +
We are missing a call to afu_release_irqs() in case of error here.
> trace_cxl_attach(ctx, work.work_element_descriptor, work.num_interrupts, amr);
>
> if ((rc = cxl_ops->attach_process(ctx, false, work.work_element_descriptor,
> amr))) {
> afu_release_irqs(ctx, ctx);
> + cxl_adapter_context_put(ctx->afu->adapter);
> goto out;
> }
>
> diff --git a/drivers/misc/cxl/main.c b/drivers/misc/cxl/main.c
> index d9be23b2..5c3c02a 100644
> --- a/drivers/misc/cxl/main.c
> +++ b/drivers/misc/cxl/main.c
> @@ -243,8 +243,9 @@ struct cxl *cxl_alloc_adapter(void)
> if (dev_set_name(&adapter->dev, "card%i", adapter->adapter_num))
> goto err2;
>
> - return adapter;
> + atomic_set(&adapter->contexts_num, 0);
I think the initialization of the counter should be done in
cxl_configure_adapter() instead of cxl_alloc_adapter(). I would expect
most reset scenarios to end up unloading/reloading the driver, but if
you look at cxl_pci_error_detected(), it should be possible to tweak
some settings to just deconfigure/reconfigure the adapter. So it seems
safer to reset the counter to 0 in cxl_configure_adapter().
Also wouldn't it make sense to add a WARN if the counter is not NULL
when unconfiguring the adapter? It seems that it would be a bug.
> + return adapter;
> err2:
> cxl_remove_adapter_nr(adapter);
> err1:
> @@ -286,6 +287,27 @@ int cxl_afu_select_best_mode(struct cxl_afu *afu)
> return 0;
> }
>
> +int cxl_adapter_context_get(struct cxl *adapter)
> +{
> + int rc;
> +
> + rc = atomic_inc_unless_negative(&adapter->contexts_num);
> + return rc >= 0 ? 0 : -EBUSY;
> +}
> +
> +void cxl_adapter_context_put(struct cxl *adapter)
> +{
> + atomic_dec_if_positive(&adapter->contexts_num);
> +}
> +
> +int cxl_adapter_context_lock(struct cxl *adapter)
> +{
> + int rc;
> + /* no active contexts -> contexts_num == 0 */
> + rc = atomic_cmpxchg(&adapter->contexts_num, 0, -1);
> + return rc ? -EBUSY : 0;
> +}
> +
> static int __init init_cxl(void)
> {
> int rc = 0;
> diff --git a/drivers/misc/cxl/sysfs.c b/drivers/misc/cxl/sysfs.c
> index b043c20..592dbf2 100644
> --- a/drivers/misc/cxl/sysfs.c
> +++ b/drivers/misc/cxl/sysfs.c
> @@ -75,12 +75,22 @@ static ssize_t reset_adapter_store(struct device *device,
> int val;
>
> rc = sscanf(buf, "%i", &val);
> - if ((rc != 1) || (val != 1))
> + if ((rc != 1) || (val != 1 && val != -1))
> return -EINVAL;
>
> - if ((rc = cxl_ops->adapter_reset(adapter)))
> - return rc;
> - return count;
> + /*
> + * See if we can lock the context mapping that's only allowed
> + * when there are no contexts attached to the adapter. Once
> + * taken this will also prevent any context being attached.
> + */
> + if (val == 1)
> + rc = cxl_adapter_context_lock(adapter);
> +
> + /* Perform a forced adapter reset */
> + if (rc >= 0)
> + rc = cxl_ops->adapter_reset(adapter);
Readability could be improved, rc meaning seems overloaded.
Also, if the adapter_reset callback fails, we need to reset the count.
Fred
> +
> + return rc ? rc : count;
> }
>
> static ssize_t load_image_on_perst_show(struct device *device,
>
More information about the Linuxppc-dev
mailing list