| Message ID | 1456199868-9054-1-git-send-email-shankerd@codeaurora.org (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On 23/02/16 03:57, Shanker Donthineni wrote: > We are not checking whether the requested device identifier fits into > table or not. The ITS MAPD command fails if 'Device ID' is outside of > device table range. > > Add a simple validation check to avoid MAPD failures since we are > not handling ITS command errors. This change also helps to return an > error -ENOMEM instead of success to caller. > In which circumstances do you see this failing? We allocate memory for the whole DevID range (as advertised by the ITS), so anything that comes up outside of that range cannot be handled anyway, and is a HW integration issue. You might as well pretend that these devices do not exist. Thanks, M.
Hi Marc, On 02/23/2016 02:51 AM, Marc Zyngier wrote: > On 23/02/16 03:57, Shanker Donthineni wrote: >> We are not checking whether the requested device identifier fits into >> table or not. The ITS MAPD command fails if 'Device ID' is outside of >> device table range. >> >> Add a simple validation check to avoid MAPD failures since we are >> not handling ITS command errors. This change also helps to return an >> error -ENOMEM instead of success to caller. >> > In which circumstances do you see this failing? We allocate memory for > the whole DevID range (as advertised by the ITS), so anything that comes > up outside of that range cannot be handled anyway, and is a HW > integration issue. You might as well pretend that these devices do not > exist. > > Thanks, > > M. We will try to allocate maximum memory as much as possible to cover whole DevID. sparse. We may not always successful, sometimes we limit memory allocation size for two reasons MAX_ORDER and ITS flat table capacity. 1) According to ARM-GIC spec, ITS hw can access maximum of 256 (ITS-pages) * 64K (ITS-pageszie) bytes. 21bits = 16 MBytes / 8 (flat table size / device table entry size) Assuming: minimum device table entry size 8 Bytes 2) On 4K page size kernel with default arch/arm64/defconfig. 19bits = 4 MBytes / 8 (flat table size / device table entry size) Assuming: minimum device table entry size 8 Bytes This is definitely a problem if DevID range is much more than 19Bits and we fail to allocate enough memory. Our ITS hardware has capable of supporting all 32bits, so it advertises DevID 32. I am preparing an another patch to support ITS-Indirection (two-level table walk) feature.
Hi Marc, On 02/22/2016 09:57 PM, Shanker Donthineni wrote: > We are not checking whether the requested device identifier fits into > table or not. The ITS MAPD command fails if 'Device ID' is outside of > device table range. > > Add a simple validation check to avoid MAPD failures since we are > not handling ITS command errors. This change also helps to return an > error -ENOMEM instead of success to caller. > > Signed-off-by: Shanker Donthineni <shankerd@codeaurora.org> > --- > This patch depends on https://git.kernel.org/cgit/linux/kernel/git/maz/arm-platforms.git/commit/?h=irq/gic-4.5-fixes&id=2eca0d6ceea1f108b2d3ac81fb34698c4fd41006 > > drivers/irqchip/irq-gic-v3-its.c | 18 +++++++++++++++--- > 1 file changed, 15 insertions(+), 3 deletions(-) > > diff --git a/drivers/irqchip/irq-gic-v3-its.c b/drivers/irqchip/irq-gic-v3-its.c > index c0f227a..9bdcdf5 100644 > --- a/drivers/irqchip/irq-gic-v3-its.c > +++ b/drivers/irqchip/irq-gic-v3-its.c > @@ -70,6 +70,7 @@ struct its_node { > struct { > void *base; > u32 order; > + u32 entry_size; > } tables[GITS_BASER_NR_REGS]; > struct its_collection *collections; > struct list_head its_device_list; > @@ -880,6 +881,7 @@ static int its_alloc_tables(const char *node_name, struct its_node *its) > node_name, order); > } > } > + its->tables[type].entry_size = entry_size; > > retry_alloc_baser: > alloc_pages = (PAGE_ORDER_TO_SIZE(order) / psz); > @@ -896,8 +898,8 @@ retry_alloc_baser: > goto out_free; > } > > - its->tables[i].base = base; > - its->tables[i].order = order; > + its->tables[type].base = base; > + its->tables[type].order = order; > > retry_baser: > val = (virt_to_phys(base) | > @@ -947,7 +949,7 @@ retry_baser: > * something is horribly wrong... > */ > free_pages((unsigned long)base, order); > - its->tables[i].base = NULL; > + its->tables[type].base = NULL; > > switch (psz) { > case SZ_16K: > @@ -1152,12 +1154,22 @@ static struct its_device *its_create_device(struct its_node *its, u32 dev_id, > unsigned long *lpi_map; > unsigned long flags; > u16 *col_map = NULL; > + u8 type = GITS_BASER_TYPE_DEVICE; > + u32 entry_size; > + u32 order; > void *itt; > int lpi_base; > int nr_lpis; > int nr_ites; > int sz; > > + entry_size = its->tables[type].entry_size; > + order = its->tables[type].order; > + > + /* Don't allow 'dev_id' that exceeds single, flat table limit */ > + if (dev_id >= (PAGE_ORDER_TO_SIZE(order) / entry_size)) > + return NULL; > + > dev = kzalloc(sizeof(*dev), GFP_KERNEL); > /* > * At least one bit of EventID is being used, hence a minimum ITS probe function tries to allocate maximum memory as much as possible to cover whole DevID. sparse which is reported (ITS_TYPER.Devbits) by HW. We might not be able to allocate enough memory for device table. We reduce memory allocation size for two reasons either because of crossing MAX_ORDER limit or max ITS_BASERn size. 1) According to ARM-GIC spec, we can assign maximum memory size 256 (ITS-Size) * 64K (ITS-PageSzie) Bytes to level-one (flat) device table. DevID = flat_table_size / device_table_entry_size = 16 MBytes / 8 = 21bits (max) 2) Maximum 4Mbytes memory is possible on kernel using PAGE_SIZE=4K with default arch/arm64/defconfig (MAX_ORDER = 12). DevID = flat_table_size / device_table_entry_size = 4 MBytes / 8 = 19bits (max) Assuming: minimum device table entry size 8 Bytes. I think it is a generic problem and this fail scenario happens if DevID range is more than 21bits or failed to allocate required memory size for device table. Our ITS hardware is capable of supporting 32bit DevID space, so it advertises ITS_TYPER.Devbits field with value 31 causing this problem. I think we need some protection in its_create_device() to avoid this scenario. I am planning to post an another patch to add support for ITS-Indirection (two-level) table walk feature.
Hi Marc, Sorry, I have not responded to your email earlier, by mistake replied to original patch. On 02/23/2016 02:51 AM, Marc Zyngier wrote: > On 23/02/16 03:57, Shanker Donthineni wrote: >> We are not checking whether the requested device identifier fits into >> table or not. The ITS MAPD command fails if 'Device ID' is outside of >> device table range. >> >> Add a simple validation check to avoid MAPD failures since we are >> not handling ITS command errors. This change also helps to return an >> error -ENOMEM instead of success to caller. >> > In which circumstances do you see this failing? We allocate memory for > the whole DevID range (as advertised by the ITS), so anything that comes > up outside of that range cannot be handled anyway, and is a HW > integration issue. You might as well pretend that these devices do not > exist. > > Thanks, > > M. ITS probe function tries to allocate maximum memory as much as possible to cover whole DevID. sparse which is reported (ITS_TYPER.Devbits) by HW. We might not be able to allocate enough memory for device table. We reduce memory allocation size for two reasons either because of crossing MAX_ORDER limit or max ITS_BASERn size. 1) According to ARM-GIC spec, we can assign maximum memory size 256 (ITS-Size) * 64K (ITS-PageSzie) Bytes to level-one (flat) device table. DevID = flat_table_size/device_table_entry_size = 16MBytes/8 = 21bits (max) 2) Maximum 4Mbytes memory is possible on kernel using PAGE_SIZE=4K with default arch/arm64/defconfig (MAX_ORDER = 12). DevID = flat_table_size/device_table_entry_size = 4MBytes/8 = 19bits (max) Assuming: minimum device table entry size 8 Bytes. I think it is a generic problem and this fail scenario happens if DevID range is more than 21bits or failed to allocate required memory size for device table. Our ITS hardware is capable of supporting 32bit DevID space, so it advertises ITS_TYPER.Devbits field with value 31 causing this problem. I think we need some protection in its_create_device() to avoid this scenario. I am planning to post an another patch to add support for ITS-Indirection (two-level) table walk feature.
diff --git a/drivers/irqchip/irq-gic-v3-its.c b/drivers/irqchip/irq-gic-v3-its.c index c0f227a..9bdcdf5 100644 --- a/drivers/irqchip/irq-gic-v3-its.c +++ b/drivers/irqchip/irq-gic-v3-its.c @@ -70,6 +70,7 @@ struct its_node { struct { void *base; u32 order; + u32 entry_size; } tables[GITS_BASER_NR_REGS]; struct its_collection *collections; struct list_head its_device_list; @@ -880,6 +881,7 @@ static int its_alloc_tables(const char *node_name, struct its_node *its) node_name, order); } } + its->tables[type].entry_size = entry_size; retry_alloc_baser: alloc_pages = (PAGE_ORDER_TO_SIZE(order) / psz); @@ -896,8 +898,8 @@ retry_alloc_baser: goto out_free; } - its->tables[i].base = base; - its->tables[i].order = order; + its->tables[type].base = base; + its->tables[type].order = order; retry_baser: val = (virt_to_phys(base) | @@ -947,7 +949,7 @@ retry_baser: * something is horribly wrong... */ free_pages((unsigned long)base, order); - its->tables[i].base = NULL; + its->tables[type].base = NULL; switch (psz) { case SZ_16K: @@ -1152,12 +1154,22 @@ static struct its_device *its_create_device(struct its_node *its, u32 dev_id, unsigned long *lpi_map; unsigned long flags; u16 *col_map = NULL; + u8 type = GITS_BASER_TYPE_DEVICE; + u32 entry_size; + u32 order; void *itt; int lpi_base; int nr_lpis; int nr_ites; int sz; + entry_size = its->tables[type].entry_size; + order = its->tables[type].order; + + /* Don't allow 'dev_id' that exceeds single, flat table limit */ + if (dev_id >= (PAGE_ORDER_TO_SIZE(order) / entry_size)) + return NULL; + dev = kzalloc(sizeof(*dev), GFP_KERNEL); /* * At least one bit of EventID is being used, hence a minimum
We are not checking whether the requested device identifier fits into table or not. The ITS MAPD command fails if 'Device ID' is outside of device table range. Add a simple validation check to avoid MAPD failures since we are not handling ITS command errors. This change also helps to return an error -ENOMEM instead of success to caller. Signed-off-by: Shanker Donthineni <shankerd@codeaurora.org> --- This patch depends on https://git.kernel.org/cgit/linux/kernel/git/maz/arm-platforms.git/commit/?h=irq/gic-4.5-fixes&id=2eca0d6ceea1f108b2d3ac81fb34698c4fd41006 drivers/irqchip/irq-gic-v3-its.c | 18 +++++++++++++++--- 1 file changed, 15 insertions(+), 3 deletions(-)