| Message ID | 166329936739.2786261.14035402420254589047.stgit@dwillia2-xfh.jf.intel.com (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | Fix the DAX-gup mistake | expand |
On Thu, Sep 15, 2022 at 08:36:07PM -0700, Dan Williams wrote: > The percpu_ref in 'struct dev_pagemap' is used to coordinate active > mappings of device-memory with the device-removal / unbind path. It > enables the semantic that initiating device-removal (or > device-driver-unbind) blocks new mapping and DMA attempts, and waits for > mapping revocation or inflight DMA to complete. This seems strange to me The pagemap should be ref'd as long as the filesystem is mounted over the dax. The ref should be incrd when the filesystem is mounted and decrd when it is unmounted. When the filesystem unmounts it should zap all the mappings (actually I don't think you can even unmount a filesystem while mappings are open) and wait for all page references to go to zero, then put the final pagemap back. The rule is nothing can touch page->pgmap while page->refcount == 0, and if page->refcount != 0 then page->pgmap must be valid, without any refcounting on the page map itself. So, why do we need pgmap refcounting all over the place? It seems like it only existed before because of the abuse of the page->refcount? Jason
Jason Gunthorpe wrote: > On Thu, Sep 15, 2022 at 08:36:07PM -0700, Dan Williams wrote: > > The percpu_ref in 'struct dev_pagemap' is used to coordinate active > > mappings of device-memory with the device-removal / unbind path. It > > enables the semantic that initiating device-removal (or > > device-driver-unbind) blocks new mapping and DMA attempts, and waits for > > mapping revocation or inflight DMA to complete. > > This seems strange to me > > The pagemap should be ref'd as long as the filesystem is mounted over > the dax. The ref should be incrd when the filesystem is mounted and > decrd when it is unmounted. > > When the filesystem unmounts it should zap all the mappings (actually > I don't think you can even unmount a filesystem while mappings are > open) and wait for all page references to go to zero, then put the > final pagemap back. > > The rule is nothing can touch page->pgmap while page->refcount == 0, > and if page->refcount != 0 then page->pgmap must be valid, without any > refcounting on the page map itself. > > So, why do we need pgmap refcounting all over the place? It seems like > it only existed before because of the abuse of the page->refcount? Recall that this percpu_ref is mirroring the same function as blk_queue_enter() whereby every new request is checking to make sure the device is still alive, or whether it has started exiting. So pgmap 'live' reference taking in fs/dax.c allows the core to start failing fault requests once device teardown has started. It is a 'block new, and drain old' semantic.
Dan Williams wrote: > Jason Gunthorpe wrote: > > On Thu, Sep 15, 2022 at 08:36:07PM -0700, Dan Williams wrote: > > > The percpu_ref in 'struct dev_pagemap' is used to coordinate active > > > mappings of device-memory with the device-removal / unbind path. It > > > enables the semantic that initiating device-removal (or > > > device-driver-unbind) blocks new mapping and DMA attempts, and waits for > > > mapping revocation or inflight DMA to complete. > > > > This seems strange to me > > > > The pagemap should be ref'd as long as the filesystem is mounted over > > the dax. The ref should be incrd when the filesystem is mounted and > > decrd when it is unmounted. > > > > When the filesystem unmounts it should zap all the mappings (actually > > I don't think you can even unmount a filesystem while mappings are > > open) and wait for all page references to go to zero, then put the > > final pagemap back. > > > > The rule is nothing can touch page->pgmap while page->refcount == 0, > > and if page->refcount != 0 then page->pgmap must be valid, without any > > refcounting on the page map itself. > > > > So, why do we need pgmap refcounting all over the place? It seems like > > it only existed before because of the abuse of the page->refcount? > > Recall that this percpu_ref is mirroring the same function as > blk_queue_enter() whereby every new request is checking to make sure the > device is still alive, or whether it has started exiting. > > So pgmap 'live' reference taking in fs/dax.c allows the core to start > failing fault requests once device teardown has started. It is a 'block > new, and drain old' semantic. However this line of questioning has me realizing that I have the put_dev_pagemap() in the wrong place. It needs to go in free_zone_device_page(), so that gup extends the lifetime of the device.
On Wed, Sep 21, 2022 at 02:38:56PM -0700, Dan Williams wrote: > Dan Williams wrote: > > Jason Gunthorpe wrote: > > > On Thu, Sep 15, 2022 at 08:36:07PM -0700, Dan Williams wrote: > > > > The percpu_ref in 'struct dev_pagemap' is used to coordinate active > > > > mappings of device-memory with the device-removal / unbind path. It > > > > enables the semantic that initiating device-removal (or > > > > device-driver-unbind) blocks new mapping and DMA attempts, and waits for > > > > mapping revocation or inflight DMA to complete. > > > > > > This seems strange to me > > > > > > The pagemap should be ref'd as long as the filesystem is mounted over > > > the dax. The ref should be incrd when the filesystem is mounted and > > > decrd when it is unmounted. > > > > > > When the filesystem unmounts it should zap all the mappings (actually > > > I don't think you can even unmount a filesystem while mappings are > > > open) and wait for all page references to go to zero, then put the > > > final pagemap back. > > > > > > The rule is nothing can touch page->pgmap while page->refcount == 0, > > > and if page->refcount != 0 then page->pgmap must be valid, without any > > > refcounting on the page map itself. > > > > > > So, why do we need pgmap refcounting all over the place? It seems like > > > it only existed before because of the abuse of the page->refcount? > > > > Recall that this percpu_ref is mirroring the same function as > > blk_queue_enter() whereby every new request is checking to make sure the > > device is still alive, or whether it has started exiting. > > > > So pgmap 'live' reference taking in fs/dax.c allows the core to start > > failing fault requests once device teardown has started. It is a 'block > > new, and drain old' semantic. It is weird this email never arrived for me.. I think that is all fine, but it would be much more logically expressed as a simple 'is pgmap alive' call before doing a new mapping than mucking with the refcount logic. Such a test could simply READ_ONCE a bool value in the pgmap struct. Indeed, you could reasonably put such a liveness test at the moment every driver takes a 0 refcount struct page and turns it into a 1 refcount struct page. Jason
Jason Gunthorpe wrote: > On Wed, Sep 21, 2022 at 02:38:56PM -0700, Dan Williams wrote: > > Dan Williams wrote: > > > Jason Gunthorpe wrote: > > > > On Thu, Sep 15, 2022 at 08:36:07PM -0700, Dan Williams wrote: > > > > > The percpu_ref in 'struct dev_pagemap' is used to coordinate active > > > > > mappings of device-memory with the device-removal / unbind path. It > > > > > enables the semantic that initiating device-removal (or > > > > > device-driver-unbind) blocks new mapping and DMA attempts, and waits for > > > > > mapping revocation or inflight DMA to complete. > > > > > > > > This seems strange to me > > > > > > > > The pagemap should be ref'd as long as the filesystem is mounted over > > > > the dax. The ref should be incrd when the filesystem is mounted and > > > > decrd when it is unmounted. > > > > > > > > When the filesystem unmounts it should zap all the mappings (actually > > > > I don't think you can even unmount a filesystem while mappings are > > > > open) and wait for all page references to go to zero, then put the > > > > final pagemap back. > > > > > > > > The rule is nothing can touch page->pgmap while page->refcount == 0, > > > > and if page->refcount != 0 then page->pgmap must be valid, without any > > > > refcounting on the page map itself. > > > > > > > > So, why do we need pgmap refcounting all over the place? It seems like > > > > it only existed before because of the abuse of the page->refcount? > > > > > > Recall that this percpu_ref is mirroring the same function as > > > blk_queue_enter() whereby every new request is checking to make sure the > > > device is still alive, or whether it has started exiting. > > > > > > So pgmap 'live' reference taking in fs/dax.c allows the core to start > > > failing fault requests once device teardown has started. It is a 'block > > > new, and drain old' semantic. > > It is weird this email never arrived for me.. > > I think that is all fine, but it would be much more logically > expressed as a simple 'is pgmap alive' call before doing a new mapping > than mucking with the refcount logic. Such a test could simply > READ_ONCE a bool value in the pgmap struct. > > Indeed, you could reasonably put such a liveness test at the moment > every driver takes a 0 refcount struct page and turns it into a 1 > refcount struct page. I could do it with a flag, but the reason to have pgmap->ref managed at the page->_refcount 0 -> 1 and 1 -> 0 transitions is so at the end of time memunmap_pages() can look at the one counter rather than scanning and rescanning all the pages to see when they go to final idle.
On Wed, Sep 21, 2022 at 05:14:34PM -0700, Dan Williams wrote: > > Indeed, you could reasonably put such a liveness test at the moment > > every driver takes a 0 refcount struct page and turns it into a 1 > > refcount struct page. > > I could do it with a flag, but the reason to have pgmap->ref managed at > the page->_refcount 0 -> 1 and 1 -> 0 transitions is so at the end of > time memunmap_pages() can look at the one counter rather than scanning > and rescanning all the pages to see when they go to final idle. That makes some sense too, but the logical way to do that is to put some counter along the page_free() path, and establish a 'make a page not free' path that does the other side. ie it should not be in DAX code, it should be all in common pgmap code. The pgmap should never be freed while any page->refcount != 0 and that should be an intrinsic property of pgmap, not relying on external parties. Though I suspect if we were to look at performance it is probably better to scan the memory on the unlikely case of pgmap removal than to put more code in hot paths to keep track of refcounts.. It doesn't need rescanning, just one sweep where it waits on every non-zero page to become zero. Jason
Jason Gunthorpe wrote: > On Wed, Sep 21, 2022 at 05:14:34PM -0700, Dan Williams wrote: > > > > Indeed, you could reasonably put such a liveness test at the moment > > > every driver takes a 0 refcount struct page and turns it into a 1 > > > refcount struct page. > > > > I could do it with a flag, but the reason to have pgmap->ref managed at > > the page->_refcount 0 -> 1 and 1 -> 0 transitions is so at the end of > > time memunmap_pages() can look at the one counter rather than scanning > > and rescanning all the pages to see when they go to final idle. > > That makes some sense too, but the logical way to do that is to put some > counter along the page_free() path, and establish a 'make a page not > free' path that does the other side. > > ie it should not be in DAX code, it should be all in common pgmap > code. The pgmap should never be freed while any page->refcount != 0 > and that should be an intrinsic property of pgmap, not relying on > external parties. I just do not know where to put such intrinsics since there is nothing today that requires going through the pgmap object to discover the pfn and 'allocate' the page. I think you may be asking to unify dax_direct_access() with pgmap management where all dax_direct_access() users are required to take a page reference if the pfn it returns is going to be used outside of dax_read_lock(). In other words make dax_direct_access() the 'allocation' event that pins the pgmap? I might be speaking a foreign language if you're not familiar with the relationship of 'struct dax_device' to 'struct dev_pagemap' instances. This is not the first time I have considered making them one in the same. > Though I suspect if we were to look at performance it is probably > better to scan the memory on the unlikely case of pgmap removal than > to put more code in hot paths to keep track of refcounts.. It doesn't > need rescanning, just one sweep where it waits on every non-zero page > to become zero. True, on the way down nothing should be elevating page references, just waiting for the last one to drain. I am just not sure that pgmap removal is that unlikely going forward with things like the dax_kmem driver and CXL Dynamic Capacity Devices where tearing down DAX devices happens. Perhaps something to revisit if the pgmap percpu_ref ever shows up in profiles.
On Wed, Sep 21, 2022 at 07:17:40PM -0700, Dan Williams wrote: > Jason Gunthorpe wrote: > > On Wed, Sep 21, 2022 at 05:14:34PM -0700, Dan Williams wrote: > > > > > > Indeed, you could reasonably put such a liveness test at the moment > > > > every driver takes a 0 refcount struct page and turns it into a 1 > > > > refcount struct page. > > > > > > I could do it with a flag, but the reason to have pgmap->ref managed at > > > the page->_refcount 0 -> 1 and 1 -> 0 transitions is so at the end of > > > time memunmap_pages() can look at the one counter rather than scanning > > > and rescanning all the pages to see when they go to final idle. > > > > That makes some sense too, but the logical way to do that is to put some > > counter along the page_free() path, and establish a 'make a page not > > free' path that does the other side. > > > > ie it should not be in DAX code, it should be all in common pgmap > > code. The pgmap should never be freed while any page->refcount != 0 > > and that should be an intrinsic property of pgmap, not relying on > > external parties. > > I just do not know where to put such intrinsics since there is nothing > today that requires going through the pgmap object to discover the pfn > and 'allocate' the page. I think that is just a new API that wrappers the set refcount = 1, percpu refcount and maybe building appropriate compound pages too. Eg maybe something like: struct folio *pgmap_alloc_folios(pgmap, start, length) And you get back maximally sized allocated folios with refcount = 1 that span the requested range. > In other words make dax_direct_access() the 'allocation' event that pins > the pgmap? I might be speaking a foreign language if you're not familiar > with the relationship of 'struct dax_device' to 'struct dev_pagemap' > instances. This is not the first time I have considered making them one > in the same. I don't know enough about dax, so yes very foreign :) I'm thinking broadly about how to make pgmap usable to all the other drivers in a safe and robust way that makes some kind of logical sense. Jason
Jason Gunthorpe wrote: > On Wed, Sep 21, 2022 at 07:17:40PM -0700, Dan Williams wrote: > > Jason Gunthorpe wrote: > > > On Wed, Sep 21, 2022 at 05:14:34PM -0700, Dan Williams wrote: > > > > > > > > Indeed, you could reasonably put such a liveness test at the moment > > > > > every driver takes a 0 refcount struct page and turns it into a 1 > > > > > refcount struct page. > > > > > > > > I could do it with a flag, but the reason to have pgmap->ref managed at > > > > the page->_refcount 0 -> 1 and 1 -> 0 transitions is so at the end of > > > > time memunmap_pages() can look at the one counter rather than scanning > > > > and rescanning all the pages to see when they go to final idle. > > > > > > That makes some sense too, but the logical way to do that is to put some > > > counter along the page_free() path, and establish a 'make a page not > > > free' path that does the other side. > > > > > > ie it should not be in DAX code, it should be all in common pgmap > > > code. The pgmap should never be freed while any page->refcount != 0 > > > and that should be an intrinsic property of pgmap, not relying on > > > external parties. > > > > I just do not know where to put such intrinsics since there is nothing > > today that requires going through the pgmap object to discover the pfn > > and 'allocate' the page. > > I think that is just a new API that wrappers the set refcount = 1, > percpu refcount and maybe building appropriate compound pages too. > > Eg maybe something like: > > struct folio *pgmap_alloc_folios(pgmap, start, length) > > And you get back maximally sized allocated folios with refcount = 1 > that span the requested range. > > > In other words make dax_direct_access() the 'allocation' event that pins > > the pgmap? I might be speaking a foreign language if you're not familiar > > with the relationship of 'struct dax_device' to 'struct dev_pagemap' > > instances. This is not the first time I have considered making them one > > in the same. > > I don't know enough about dax, so yes very foreign :) > > I'm thinking broadly about how to make pgmap usable to all the other > drivers in a safe and robust way that makes some kind of logical sense. I think the API should be pgmap_folio_get() because, at least for DAX, the memory is already allocated. The 'allocator' for fsdax is the filesystem block allocator, and pgmap_folio_get() grants access to a folio in the pgmap by a pfn that the block allocator knows about. If the GPU use case wants to wrap an allocator around that they can, but the fundamental requirement is check if the pgmap is dead and if not elevate the page reference. So something like: /** * pgmap_get_folio() - reference a folio in a live @pgmap by @pfn * @pgmap: live pgmap instance, caller ensures this does not race @pgmap death * @pfn: page frame number covered by @pgmap */ struct folio *pgmap_get_folio(struct dev_pagemap *pgmap, unsigned long pfn) { struct page *page; VM_WARN_ONCE(pgmap != xa_load(&pgmap_array, PHYS_PFN(phys))); if (WARN_ONCE(percpu_ref_is_dying(&pgmap->ref))) return NULL; page = pfn_to_page(pfn); return page_folio(page); } This does not create compound folios, that needs to be coordinated with the caller and likely needs an explicit pgmap_construct_folio(pgmap, pfn, order) ...call that can be done while holding locks against operations that will cause the folio to be broken down.
On Thu, Sep 22, 2022 at 02:54:42PM -0700, Dan Williams wrote: > Jason Gunthorpe wrote: > > On Wed, Sep 21, 2022 at 07:17:40PM -0700, Dan Williams wrote: > > > Jason Gunthorpe wrote: > > > > On Wed, Sep 21, 2022 at 05:14:34PM -0700, Dan Williams wrote: > > > > > > > > > > Indeed, you could reasonably put such a liveness test at the moment > > > > > > every driver takes a 0 refcount struct page and turns it into a 1 > > > > > > refcount struct page. > > > > > > > > > > I could do it with a flag, but the reason to have pgmap->ref managed at > > > > > the page->_refcount 0 -> 1 and 1 -> 0 transitions is so at the end of > > > > > time memunmap_pages() can look at the one counter rather than scanning > > > > > and rescanning all the pages to see when they go to final idle. > > > > > > > > That makes some sense too, but the logical way to do that is to put some > > > > counter along the page_free() path, and establish a 'make a page not > > > > free' path that does the other side. > > > > > > > > ie it should not be in DAX code, it should be all in common pgmap > > > > code. The pgmap should never be freed while any page->refcount != 0 > > > > and that should be an intrinsic property of pgmap, not relying on > > > > external parties. > > > > > > I just do not know where to put such intrinsics since there is nothing > > > today that requires going through the pgmap object to discover the pfn > > > and 'allocate' the page. > > > > I think that is just a new API that wrappers the set refcount = 1, > > percpu refcount and maybe building appropriate compound pages too. > > > > Eg maybe something like: > > > > struct folio *pgmap_alloc_folios(pgmap, start, length) > > > > And you get back maximally sized allocated folios with refcount = 1 > > that span the requested range. > > > > > In other words make dax_direct_access() the 'allocation' event that pins > > > the pgmap? I might be speaking a foreign language if you're not familiar > > > with the relationship of 'struct dax_device' to 'struct dev_pagemap' > > > instances. This is not the first time I have considered making them one > > > in the same. > > > > I don't know enough about dax, so yes very foreign :) > > > > I'm thinking broadly about how to make pgmap usable to all the other > > drivers in a safe and robust way that makes some kind of logical sense. > > I think the API should be pgmap_folio_get() because, at least for DAX, > the memory is already allocated. The 'allocator' for fsdax is the > filesystem block allocator, and pgmap_folio_get() grants access to a No, the "allocator" for fsdax is the inode iomap interface, not the filesystem block allocator. The filesystem block allocator is only involved in iomapping if we have to allocate a new mapping for a given file offset. A better name for this is "arbiter", not allocator. To get an active mapping of the DAX pages backing a file, we need to ask the inode iomap subsystem to *map a file offset* and it will return kaddr and/or pfns for the backing store the file offset maps to. IOWs, for FSDAX, access to the backing store (i.e. the physical pages) is arbitrated by the *inode*, not the filesystem allocator or the dax device. Hence if a subsystem needs to pin the backing store for some use, it must first ensure that it holds an inode reference (direct or indirect) for that range of the backing store that will spans the life of the pin. When the pin is done, it can tear down the mappings it was using and then the inode reference can be released. This ensures that any racing unlink of the inode will not result in the backing store being freed from under the application that has a pin. It will prevent the inode from being reclaimed and so potentially accessing stale or freed in-memory structures. And it will prevent the filesytem from being unmounted while the application using FSDAX access is still actively using that functionality even if it's already closed all it's fds.... Cheers, Dave.
Dave Chinner wrote: > On Thu, Sep 22, 2022 at 02:54:42PM -0700, Dan Williams wrote: > > Jason Gunthorpe wrote: > > > On Wed, Sep 21, 2022 at 07:17:40PM -0700, Dan Williams wrote: > > > > Jason Gunthorpe wrote: > > > > > On Wed, Sep 21, 2022 at 05:14:34PM -0700, Dan Williams wrote: > > > > > > > > > > > > Indeed, you could reasonably put such a liveness test at the moment > > > > > > > every driver takes a 0 refcount struct page and turns it into a 1 > > > > > > > refcount struct page. > > > > > > > > > > > > I could do it with a flag, but the reason to have pgmap->ref managed at > > > > > > the page->_refcount 0 -> 1 and 1 -> 0 transitions is so at the end of > > > > > > time memunmap_pages() can look at the one counter rather than scanning > > > > > > and rescanning all the pages to see when they go to final idle. > > > > > > > > > > That makes some sense too, but the logical way to do that is to put some > > > > > counter along the page_free() path, and establish a 'make a page not > > > > > free' path that does the other side. > > > > > > > > > > ie it should not be in DAX code, it should be all in common pgmap > > > > > code. The pgmap should never be freed while any page->refcount != 0 > > > > > and that should be an intrinsic property of pgmap, not relying on > > > > > external parties. > > > > > > > > I just do not know where to put such intrinsics since there is nothing > > > > today that requires going through the pgmap object to discover the pfn > > > > and 'allocate' the page. > > > > > > I think that is just a new API that wrappers the set refcount = 1, > > > percpu refcount and maybe building appropriate compound pages too. > > > > > > Eg maybe something like: > > > > > > struct folio *pgmap_alloc_folios(pgmap, start, length) > > > > > > And you get back maximally sized allocated folios with refcount = 1 > > > that span the requested range. > > > > > > > In other words make dax_direct_access() the 'allocation' event that pins > > > > the pgmap? I might be speaking a foreign language if you're not familiar > > > > with the relationship of 'struct dax_device' to 'struct dev_pagemap' > > > > instances. This is not the first time I have considered making them one > > > > in the same. > > > > > > I don't know enough about dax, so yes very foreign :) > > > > > > I'm thinking broadly about how to make pgmap usable to all the other > > > drivers in a safe and robust way that makes some kind of logical sense. > > > > I think the API should be pgmap_folio_get() because, at least for DAX, > > the memory is already allocated. The 'allocator' for fsdax is the > > filesystem block allocator, and pgmap_folio_get() grants access to a > > No, the "allocator" for fsdax is the inode iomap interface, not the > filesystem block allocator. The filesystem block allocator is only > involved in iomapping if we have to allocate a new mapping for a > given file offset. > > A better name for this is "arbiter", not allocator. To get an > active mapping of the DAX pages backing a file, we need to ask the > inode iomap subsystem to *map a file offset* and it will return > kaddr and/or pfns for the backing store the file offset maps to. > > IOWs, for FSDAX, access to the backing store (i.e. the physical pages) is > arbitrated by the *inode*, not the filesystem allocator or the dax > device. Hence if a subsystem needs to pin the backing store for some > use, it must first ensure that it holds an inode reference (direct > or indirect) for that range of the backing store that will spans the > life of the pin. When the pin is done, it can tear down the mappings > it was using and then the inode reference can be released. > > This ensures that any racing unlink of the inode will not result in > the backing store being freed from under the application that has a > pin. It will prevent the inode from being reclaimed and so > potentially accessing stale or freed in-memory structures. And it > will prevent the filesytem from being unmounted while the > application using FSDAX access is still actively using that > functionality even if it's already closed all it's fds.... Sounds so simple when you put it that way. I'll give it a shot and stop the gymnastics of trying to get in front of truncate_inode_pages_final() with a 'dax break layouts', just hold it off until final unpin.
On Thu, Sep 22, 2022 at 02:54:42PM -0700, Dan Williams wrote: > > I'm thinking broadly about how to make pgmap usable to all the other > > drivers in a safe and robust way that makes some kind of logical sense. > > I think the API should be pgmap_folio_get() because, at least for DAX, > the memory is already allocated. I would pick a name that has some logical connection to ops->page_free() This function is starting a pairing where once it completes page_free will eventually be called. > /** > * pgmap_get_folio() - reference a folio in a live @pgmap by @pfn > * @pgmap: live pgmap instance, caller ensures this does not race @pgmap death > * @pfn: page frame number covered by @pgmap > */ > struct folio *pgmap_get_folio(struct dev_pagemap *pgmap, unsigned long pfn) > { > struct page *page; > > VM_WARN_ONCE(pgmap != xa_load(&pgmap_array, PHYS_PFN(phys))); > > if (WARN_ONCE(percpu_ref_is_dying(&pgmap->ref))) > return NULL; This shouldn't be a WARN? > page = pfn_to_page(pfn); > return page_folio(page); > } Yeah, makes sense to me, but I would do a len as well to amortize the cost of all these checks.. > This does not create compound folios, that needs to be coordinated with > the caller and likely needs an explicit Does it? What situations do you think the caller needs to coordinate the folio size? Caller should call the function for each logical unit of storage it wants to allocate from the pgmap.. Jason
Jason Gunthorpe wrote: > On Thu, Sep 22, 2022 at 02:54:42PM -0700, Dan Williams wrote: > > > > I'm thinking broadly about how to make pgmap usable to all the other > > > drivers in a safe and robust way that makes some kind of logical sense. > > > > I think the API should be pgmap_folio_get() because, at least for DAX, > > the memory is already allocated. > > I would pick a name that has some logical connection to > ops->page_free() > > This function is starting a pairing where once it completes page_free > will eventually be called. Following Dave's note that this is an 'arbitration' mechanism I think request/release is more appropriate than alloc/free for what this is doing. > > > /** > > * pgmap_get_folio() - reference a folio in a live @pgmap by @pfn > > * @pgmap: live pgmap instance, caller ensures this does not race @pgmap death > > * @pfn: page frame number covered by @pgmap > > */ > > struct folio *pgmap_get_folio(struct dev_pagemap *pgmap, unsigned long pfn) > > { > > struct page *page; > > > > VM_WARN_ONCE(pgmap != xa_load(&pgmap_array, PHYS_PFN(phys))); > > > > if (WARN_ONCE(percpu_ref_is_dying(&pgmap->ref))) > > return NULL; > > This shouldn't be a WARN? It's a bug if someone calls this after killing the pgmap. I.e. the expectation is that the caller is synchronzing this. The only reason this isn't a VM_WARN_ONCE is because the sanity check is cheap, but I do not expect it to fire on anything but a development kernel. > > > page = pfn_to_page(pfn); > > return page_folio(page); > > } > > Yeah, makes sense to me, but I would do a len as well to amortize the > cost of all these checks.. > > > This does not create compound folios, that needs to be coordinated with > > the caller and likely needs an explicit > > Does it? What situations do you think the caller needs to coordinate > the folio size? Caller should call the function for each logical unit > of storage it wants to allocate from the pgmap.. The problem for fsdax is that it needs to gather all the PTEs, hold a lock to synchronize against events that would shatter a huge page, and then build up the compound folio metadata before inserting the PMD. So I think that flow is request all pfns, lock, fixup refcounts, build up compound folio, insert huge i_pages entry, unlock and install the pmd.
On Fri, Sep 23, 2022 at 09:29:51AM -0700, Dan Williams wrote: > > > /** > > > * pgmap_get_folio() - reference a folio in a live @pgmap by @pfn > > > * @pgmap: live pgmap instance, caller ensures this does not race @pgmap death > > > * @pfn: page frame number covered by @pgmap > > > */ > > > struct folio *pgmap_get_folio(struct dev_pagemap *pgmap, > > > unsigned long pfn) Maybe should be not be pfn but be 'offset from the first page of the pgmap' ? Then we don't need the xa_load stuff, since it cann't be wrong by definition. > > > { > > > struct page *page; > > > > > > VM_WARN_ONCE(pgmap != xa_load(&pgmap_array, PHYS_PFN(phys))); > > > > > > if (WARN_ONCE(percpu_ref_is_dying(&pgmap->ref))) > > > return NULL; > > > > This shouldn't be a WARN? > > It's a bug if someone calls this after killing the pgmap. I.e. the > expectation is that the caller is synchronzing this. The only reason > this isn't a VM_WARN_ONCE is because the sanity check is cheap, but I do > not expect it to fire on anything but a development kernel. OK, that makes sense But shouldn't this get the pgmap refcount here? The reason we started talking about this was to make all the pgmap logic self contained so that the pgmap doesn't pass its own destroy until all the all the page_free()'s have been done. > > > This does not create compound folios, that needs to be coordinated with > > > the caller and likely needs an explicit > > > > Does it? What situations do you think the caller needs to coordinate > > the folio size? Caller should call the function for each logical unit > > of storage it wants to allocate from the pgmap.. > > The problem for fsdax is that it needs to gather all the PTEs, hold a > lock to synchronize against events that would shatter a huge page, and > then build up the compound folio metadata before inserting the PMD. Er, at this point we are just talking about acquiring virgin pages nobody else is using, not inserting things. There is no possibility of conurrent shattering because, by definition, nothing else can reference these struct pages at this instant. Also, the caller must already be serializating pgmap_get_folio() against concurrent calls on the same pfn (since it is an error to call pgmap_get_folio() on an non-free pfn) So, I would expect the caller must already have all the necessary locking to accept maximally sized folios. eg if it has some reason to punch a hole in the contiguous range (shatter the folio) it must *already* serialize against pgmap_get_folio(), since something like punching a hole must know with certainty if any struct pages are refcount != 0 or not, and must not race with something trying to set their refcount to 1. Jason
Jason Gunthorpe wrote: > On Fri, Sep 23, 2022 at 09:29:51AM -0700, Dan Williams wrote: > > > > /** > > > > * pgmap_get_folio() - reference a folio in a live @pgmap by @pfn > > > > * @pgmap: live pgmap instance, caller ensures this does not race @pgmap death > > > > * @pfn: page frame number covered by @pgmap > > > > */ > > > > struct folio *pgmap_get_folio(struct dev_pagemap *pgmap, > > > > unsigned long pfn) > > Maybe should be not be pfn but be 'offset from the first page of the > pgmap' ? Then we don't need the xa_load stuff, since it cann't be > wrong by definition. > > > > > { > > > > struct page *page; > > > > > > > > VM_WARN_ONCE(pgmap != xa_load(&pgmap_array, PHYS_PFN(phys))); > > > > > > > > if (WARN_ONCE(percpu_ref_is_dying(&pgmap->ref))) > > > > return NULL; > > > > > > This shouldn't be a WARN? > > > > It's a bug if someone calls this after killing the pgmap. I.e. the > > expectation is that the caller is synchronzing this. The only reason > > this isn't a VM_WARN_ONCE is because the sanity check is cheap, but I do > > not expect it to fire on anything but a development kernel. > > OK, that makes sense > > But shouldn't this get the pgmap refcount here? The reason we started > talking about this was to make all the pgmap logic self contained so > that the pgmap doesn't pass its own destroy until all the all the > page_free()'s have been done. > > > > > This does not create compound folios, that needs to be coordinated with > > > > the caller and likely needs an explicit > > > > > > Does it? What situations do you think the caller needs to coordinate > > > the folio size? Caller should call the function for each logical unit > > > of storage it wants to allocate from the pgmap.. > > > > The problem for fsdax is that it needs to gather all the PTEs, hold a > > lock to synchronize against events that would shatter a huge page, and > > then build up the compound folio metadata before inserting the PMD. > > Er, at this point we are just talking about acquiring virgin pages > nobody else is using, not inserting things. There is no possibility of > conurrent shattering because, by definition, nothing else can > reference these struct pages at this instant. > > Also, the caller must already be serializating pgmap_get_folio() > against concurrent calls on the same pfn (since it is an error to call > pgmap_get_folio() on an non-free pfn) > > So, I would expect the caller must already have all the necessary > locking to accept maximally sized folios. > > eg if it has some reason to punch a hole in the contiguous range > (shatter the folio) it must *already* serialize against > pgmap_get_folio(), since something like punching a hole must know with > certainty if any struct pages are refcount != 0 or not, and must not > race with something trying to set their refcount to 1. Perhaps, I'll take a look. The scenario I am more concerned about is processA sets up a VMA of PAGE_SIZE and races processB to fault in the same filesystem block with a VMA of PMD_SIZE. Right now processA gets a PTE mapping and processB gets a PMD mapping, but the refcounting is all handled in small pages. I need to investigate more what is needed for fsdax to support folio_size() > mapping entry size.
On Fri, Sep 23, 2022 at 12:03:53PM -0700, Dan Williams wrote: > Perhaps, I'll take a look. The scenario I am more concerned about is > processA sets up a VMA of PAGE_SIZE and races processB to fault in the > same filesystem block with a VMA of PMD_SIZE. Right now processA gets a > PTE mapping and processB gets a PMD mapping, but the refcounting is all > handled in small pages. I need to investigate more what is needed for > fsdax to support folio_size() > mapping entry size. This is fine actually. The PMD/PTE can hold a tail page. So the page cache will hold a PMD sized folio, procesA will have a PTE pointing to a tail page and processB will have a PMD pointing at the head page. For the immediate instant you can keep accounting for each tail page as you do now, just with folio wrappers. Once you have proper folios you shift the accounting responsibility to the core code and the core will faster with one ref per PMD/PTE. The trick with folios is probably going to be breaking up a folio. THP has some nasty stuff for that, but I think a FS would be better to just revoke the entire folio, bring the refcount to 0, change the underling physical mapping, and then fault will naturally restore a properly sized folio to accomodate the new physical layout. ie you never break up a folio once it is created from the pgmap. What you want is to have largest possibile folios because it optimizes all the handling logic. .. and then you are well positioned to do some kind of trick where the FS asserts at mount time that it never needs a folio less than order X and you can then trigger the devdax optimization of folding struct page memory and significantly reducing the wastage for struct page.. Jason
Dan Williams <dan.j.williams@intel.com> writes: > Jason Gunthorpe wrote: >> On Fri, Sep 23, 2022 at 09:29:51AM -0700, Dan Williams wrote: >> > > > /** >> > > > * pgmap_get_folio() - reference a folio in a live @pgmap by @pfn >> > > > * @pgmap: live pgmap instance, caller ensures this does not race @pgmap death >> > > > * @pfn: page frame number covered by @pgmap >> > > > */ >> > > > struct folio *pgmap_get_folio(struct dev_pagemap *pgmap, >> > > > unsigned long pfn) >> >> Maybe should be not be pfn but be 'offset from the first page of the >> pgmap' ? Then we don't need the xa_load stuff, since it cann't be >> wrong by definition. >> >> > > > { >> > > > struct page *page; >> > > > >> > > > VM_WARN_ONCE(pgmap != xa_load(&pgmap_array, PHYS_PFN(phys))); >> > > > >> > > > if (WARN_ONCE(percpu_ref_is_dying(&pgmap->ref))) >> > > > return NULL; >> > > >> > > This shouldn't be a WARN? >> > >> > It's a bug if someone calls this after killing the pgmap. I.e. the >> > expectation is that the caller is synchronzing this. The only reason >> > this isn't a VM_WARN_ONCE is because the sanity check is cheap, but I do >> > not expect it to fire on anything but a development kernel. >> >> OK, that makes sense >> >> But shouldn't this get the pgmap refcount here? The reason we started >> talking about this was to make all the pgmap logic self contained so >> that the pgmap doesn't pass its own destroy until all the all the >> page_free()'s have been done. That sounds good to me at least. I just noticed we introduced this exact bug for device private/coherent pages when making their refcounts zero based. Nothing currently takes pgmap->ref when a private/coherent page is mapped. Therefore memunmap_pages() will complete and pgmap destroyed while pgmap pages are still mapped. So I think we need to call put_dev_pagemap() as part of free_zone_device_page(). - Alistair >> > > > This does not create compound folios, that needs to be coordinated with >> > > > the caller and likely needs an explicit >> > > >> > > Does it? What situations do you think the caller needs to coordinate >> > > the folio size? Caller should call the function for each logical unit >> > > of storage it wants to allocate from the pgmap.. >> > >> > The problem for fsdax is that it needs to gather all the PTEs, hold a >> > lock to synchronize against events that would shatter a huge page, and >> > then build up the compound folio metadata before inserting the PMD. >> >> Er, at this point we are just talking about acquiring virgin pages >> nobody else is using, not inserting things. There is no possibility of >> conurrent shattering because, by definition, nothing else can >> reference these struct pages at this instant. >> >> Also, the caller must already be serializating pgmap_get_folio() >> against concurrent calls on the same pfn (since it is an error to call >> pgmap_get_folio() on an non-free pfn) >> >> So, I would expect the caller must already have all the necessary >> locking to accept maximally sized folios. >> >> eg if it has some reason to punch a hole in the contiguous range >> (shatter the folio) it must *already* serialize against >> pgmap_get_folio(), since something like punching a hole must know with >> certainty if any struct pages are refcount != 0 or not, and must not >> race with something trying to set their refcount to 1. > > Perhaps, I'll take a look. The scenario I am more concerned about is > processA sets up a VMA of PAGE_SIZE and races processB to fault in the > same filesystem block with a VMA of PMD_SIZE. Right now processA gets a > PTE mapping and processB gets a PMD mapping, but the refcounting is all > handled in small pages. I need to investigate more what is needed for > fsdax to support folio_size() > mapping entry size.
On Tue, Sep 27, 2022 at 04:07:05PM +1000, Alistair Popple wrote: > That sounds good to me at least. I just noticed we introduced this exact > bug for device private/coherent pages when making their refcounts zero > based. Nothing currently takes pgmap->ref when a private/coherent page > is mapped. Therefore memunmap_pages() will complete and pgmap destroyed > while pgmap pages are still mapped. To kind of summarize this thread Either we should get the pgmap reference during the refcount = 1 flow, and put it during page_free() Or we should have the pgmap destroy sweep all the pages and wait for them to become ref == 0 I don't think we should have pgmap references randomly strewn all over the place. A positive refcount on the page alone must be enough to prove that the struct page exists and the pgmap is not destroyed. Every driver using pgmap needs something liek this, so I'd prefer it be in the pgmap code.. Jason
diff --git a/fs/dax.c b/fs/dax.c index 5d9f30105db4..ee2568c8b135 100644 --- a/fs/dax.c +++ b/fs/dax.c @@ -376,14 +376,26 @@ static inline void dax_mapping_set_cow(struct page *page) * whether this entry is shared by multiple files. If so, set the page->mapping * FS_DAX_MAPPING_COW, and use page->index as refcount. */ -static void dax_associate_entry(void *entry, struct address_space *mapping, - struct vm_fault *vmf, unsigned long flags) +static vm_fault_t dax_associate_entry(void *entry, + struct address_space *mapping, + struct vm_fault *vmf, unsigned long flags) { unsigned long size = dax_entry_size(entry), pfn, index; + struct dev_pagemap *pgmap; int i = 0; if (IS_ENABLED(CONFIG_FS_DAX_LIMITED)) - return; + return 0; + + if (!size) + return 0; + + if (!(flags & DAX_COW)) { + pfn = dax_to_pfn(entry); + pgmap = get_dev_pagemap_many(pfn, NULL, PHYS_PFN(size)); + if (!pgmap) + return VM_FAULT_SIGBUS; + } index = linear_page_index(vmf->vma, ALIGN(vmf->address, size)); for_each_mapped_pfn(entry, pfn) { @@ -398,19 +410,24 @@ static void dax_associate_entry(void *entry, struct address_space *mapping, page_ref_inc(page); } } + + return 0; } static void dax_disassociate_entry(void *entry, struct address_space *mapping, bool trunc) { - unsigned long pfn; + unsigned long size = dax_entry_size(entry), pfn; + struct page *page; if (IS_ENABLED(CONFIG_FS_DAX_LIMITED)) return; - for_each_mapped_pfn(entry, pfn) { - struct page *page = pfn_to_page(pfn); + if (!size) + return; + for_each_mapped_pfn(entry, pfn) { + page = pfn_to_page(pfn); if (dax_mapping_is_cow(page->mapping)) { /* keep the CoW flag if this page is still shared */ if (page->index-- > 0) @@ -423,6 +440,11 @@ static void dax_disassociate_entry(void *entry, struct address_space *mapping, page->mapping = NULL; page->index = 0; } + + if (trunc && !dax_mapping_is_cow(page->mapping)) { + page = pfn_to_page(dax_to_pfn(entry)); + put_dev_pagemap_many(page->pgmap, PHYS_PFN(size)); + } } /* diff --git a/include/linux/memremap.h b/include/linux/memremap.h index c3b4cc84877b..fd57407e7f3d 100644 --- a/include/linux/memremap.h +++ b/include/linux/memremap.h @@ -191,8 +191,13 @@ void *memremap_pages(struct dev_pagemap *pgmap, int nid); void memunmap_pages(struct dev_pagemap *pgmap); void *devm_memremap_pages(struct device *dev, struct dev_pagemap *pgmap); void devm_memunmap_pages(struct device *dev, struct dev_pagemap *pgmap); -struct dev_pagemap *get_dev_pagemap(unsigned long pfn, - struct dev_pagemap *pgmap); +struct dev_pagemap *get_dev_pagemap_many(unsigned long pfn, + struct dev_pagemap *pgmap, int refs); +static inline struct dev_pagemap *get_dev_pagemap(unsigned long pfn, + struct dev_pagemap *pgmap) +{ + return get_dev_pagemap_many(pfn, pgmap, 1); +} bool pgmap_pfn_valid(struct dev_pagemap *pgmap, unsigned long pfn); unsigned long vmem_altmap_offset(struct vmem_altmap *altmap); @@ -244,10 +249,15 @@ static inline unsigned long memremap_compat_align(void) } #endif /* CONFIG_ZONE_DEVICE */ -static inline void put_dev_pagemap(struct dev_pagemap *pgmap) +static inline void put_dev_pagemap_many(struct dev_pagemap *pgmap, int refs) { if (pgmap) - percpu_ref_put(&pgmap->ref); + percpu_ref_put_many(&pgmap->ref, refs); +} + +static inline void put_dev_pagemap(struct dev_pagemap *pgmap) +{ + put_dev_pagemap_many(pgmap, 1); } #endif /* _LINUX_MEMREMAP_H_ */ diff --git a/mm/memremap.c b/mm/memremap.c index 95f6ffe9cb0f..83c5e6fafd84 100644 --- a/mm/memremap.c +++ b/mm/memremap.c @@ -430,15 +430,16 @@ void vmem_altmap_free(struct vmem_altmap *altmap, unsigned long nr_pfns) } /** - * get_dev_pagemap() - take a new live reference on the dev_pagemap for @pfn + * get_dev_pagemap_many() - take new live references(s) on the dev_pagemap for @pfn * @pfn: page frame number to lookup page_map * @pgmap: optional known pgmap that already has a reference + * @refs: number of references to take * * If @pgmap is non-NULL and covers @pfn it will be returned as-is. If @pgmap * is non-NULL but does not cover @pfn the reference to it will be released. */ -struct dev_pagemap *get_dev_pagemap(unsigned long pfn, - struct dev_pagemap *pgmap) +struct dev_pagemap *get_dev_pagemap_many(unsigned long pfn, + struct dev_pagemap *pgmap, int refs) { resource_size_t phys = PFN_PHYS(pfn); @@ -454,13 +455,15 @@ struct dev_pagemap *get_dev_pagemap(unsigned long pfn, /* fall back to slow path lookup */ rcu_read_lock(); pgmap = xa_load(&pgmap_array, PHYS_PFN(phys)); - if (pgmap && !percpu_ref_tryget_live(&pgmap->ref)) + if (pgmap && !percpu_ref_tryget_live_rcu(&pgmap->ref)) pgmap = NULL; + if (pgmap && refs > 1) + percpu_ref_get_many(&pgmap->ref, refs - 1); rcu_read_unlock(); return pgmap; } -EXPORT_SYMBOL_GPL(get_dev_pagemap); +EXPORT_SYMBOL_GPL(get_dev_pagemap_many); void free_zone_device_page(struct page *page) {
The percpu_ref in 'struct dev_pagemap' is used to coordinate active mappings of device-memory with the device-removal / unbind path. It enables the semantic that initiating device-removal (or device-driver-unbind) blocks new mapping and DMA attempts, and waits for mapping revocation or inflight DMA to complete. Expand the scope of the reference count to pin the DAX device active at mapping time and not later at the first gup event. With a device reference being held while any page on that device is mapped the need to manage pgmap reference counts in the gup code is eliminated. That cleanup is saved for a follow-on change. For now, teach dax_insert_entry() and dax_delete_mapping_entry() to take and drop pgmap references respectively. Where dax_insert_entry() is called to take the initial reference on the page, and dax_delete_mapping_entry() is called once there are no outstanding references to the given page(s). Cc: Matthew Wilcox <willy@infradead.org> Cc: Jan Kara <jack@suse.cz> Cc: "Darrick J. Wong" <djwong@kernel.org> Cc: Jason Gunthorpe <jgg@nvidia.com> Cc: Christoph Hellwig <hch@lst.de> Cc: John Hubbard <jhubbard@nvidia.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com> --- fs/dax.c | 34 ++++++++++++++++++++++++++++------ include/linux/memremap.h | 18 ++++++++++++++---- mm/memremap.c | 13 ++++++++----- 3 files changed, 50 insertions(+), 15 deletions(-)